content
stringlengths 1
15.9M
|
|---|
\section{Conclusion Remarks}
In this paper, we propose a deep land-use configuration generation framework, namely CLUVAE, which can generate land-use configurations according to human guidance and considering surrounding contexts' socioeconomic features.
To produce personalized and reasonable land-use configuration, and make our generation model robust, we formulate the automated urban planning process
into a deep variational autoencoder framework.
We implement the framework based on an encoder-decoder paradigm.
Specifically, we first utilize the encoder part to learn a distribution based on condition embeddings (\textit{i.e., surrounding contexts' feature embedding + human guidance embedding}) and land-use configurations.
Then, a variational Gaussian embedding mechanism is used to produce latent embeddings by sampling from the learned distribution.
After that, we input condition embeddings and latent embeddings into the decoder part for reconstructing land-use configurations and the corresponding urban functional zones.
Finally, the well-trained decoder is our expected land-use configuration generator.
Extensive experiments validate the superiority and effectiveness of CLUVAE.
Although CLUVAE could obtain desirable results, it is also subject to certain limitations.
In the future, we will focus on utilizing sophisticated human guidance to improve generative performance.
Because complex human guidance can provide more information for generating reasonable land-use configurations.
Thus, how to deal with sophisticated human guidance in planning generation is our next focus.
\section{Experimental Results}
We presented extensive experiments to answer:
\textbf{Q1.} Does our method (CLUVAE) outperform other baseline models?
\textbf{Q2.} Does the variational Gaussian embedding mechanism make the model learning process stable?
\textbf{Q3.} Is each technical component of CLUVAE necessary to generate better land-use configurations?
\textbf{Q4.}
We split the target area into $N\times N$ squares to collect land-use configurations.
How does the square size influence configuration generation?
\textbf{Q5.} What do the generated land-use configurations look like? What are the differences between generated configurations and real configurations?
\subsection{Experimental Setup}
\subsubsection{Data Description}
The datasets used in the experiments include:
\textbf{(1) Residential Communities:}
are crawled from soufun.com and contains 2990 residential communities in Beijing.
These communities are regarded as target areas to collect data samples for model learning and each community has its green rate record.
\textbf{(2) POIs:} include 328,668 POIs in Beijing. Each POI is associated with latitude, longitude, POI categories. Table \ref{poi_lists} shows the 20 POI categories.
\textbf{(3) Taxi Trajectories:} are collected from a taxi company in Beijing. Each trajectory contains a trip ID, a travel distance, travel times, average speed, pick-up and drop-off times, pick-up and drop-off points.
\textbf{(4) Public Transportation:} includes bus transactions in Beijing from 2012 to 2013 and contains 718 bus lines, 1734247 bus trips.
\textbf{(5) Housing Prices:} are collected from soufun.com.
\textbf{(6) Mobile Checkins: }which are collected from weibo.com. Each check-in record includes longitude, latitude, check-in time and place.
\begin{table}[htbp]
\vspace{-0.2cm}
\small
\centering
\setlength{\abovecaptionskip}{0.cm}
\caption{POI categories}
\setlength{\tabcolsep}{1mm}{
\begin{tabular}{cccccc}
\toprule
code & POI category & code & POI category \\
\midrule
0 & road & 10 & tourist attraction \\
1 & car service & 11 & real estate \\
2 & car repair & 12 & government place \\
3 & motorbike service & 13 & education \\
4 & food service & 14 & transportation \\
5 & shopping & 15 & finance\\
6 & daily life service & 16 & company\\
7 & recreation service & 17 & road furniture\\
8 & medical service & 18 & specific address \\
9 & lodging & 19 & public service\\
\bottomrule
\end{tabular}}
\label{poi_lists}
\vspace{-0.2cm}
\end{table}
\subsubsection{Evaluation Metrics}
We categorized the collected land-use configurations into five levels according to their greenery rate.
Green level labels are regarded as human guidance in our experiments.
The land-use configurations of different green levels come from different distributions.
To evaluate the model performance, we measured the distance between the distribution of generated configurations and the distribution of original configurations under different green level settings.
Let $w_j$ denote the weight of the green level $j$, which is the number of land-use configurations belonging to the green level $j$;
$Y_j$ denote the distribution of original land-use configurations of green level $j$;
$\hat{Y}_j$ denote the distribution of generated land-use configurations of green level $j$;
there are four evaluation metrics in our work:
1) \textbf{Average Kullback-Leibler (KL) Divergence\footnote{https://en.wikipedia.org/wiki/Kullback-Leibler\_divergence}:}
$\text{AVG\_KL} = \frac{\sum_{j=1}^{5}w_j\cdot KL(Y_j,\hat{Y}_j) }{\sum_{j=1}^{5}w_j}. $
2) \textbf{Average Jensen-Shannon (JS) Divergence\footnote{https://en.wikipedia.org/wiki/Jensen-Shannon\_divergence}:}
$\text{AVG\_JS} = \frac{\sum_{j=1}^{5}w_j\cdot JS(Y_j,\hat{Y}_j) }{\sum_{j=1}^{5}w_j}.$
3) \textbf{Average Hellinger Distance (HD) \footnote{https://en.wikipedia.org/wiki/Hellinger\_distance}:} $\text{AVG\_HD} = \frac{\sum_{j=1}^{5}w_j\cdot HD(Y_j,\hat{Y}_j) }{\sum_{j=1}^{5}w_j}.$
4) \textbf{ Average Cosine Distance (Cos)\footnote{https://en.wikipedia.org/wiki/Cosine\_similarity}:}
$ \text{AVG\_Cos} = \frac{\sum_{j=1}^{5}w_j\cdot Cos(Y_j,\hat{Y}_j) }{\sum_{j=1}^{5}w_j}.$
The lower the metric value is, the better performance the model exhibits.
\begin{figure*}[hbtp!]
\subfigure[AVG\_KL]{\includegraphics[width=4.35cm]{img/table1_AVG_KL.pdf} \label{fig:avg_kl}}
\subfigure[AVG\_JS]{\includegraphics[width=4.35cm]{img/table1_AVG_JS.pdf}\label{fig:avg_js}}
\subfigure[AVG\_HD]{\includegraphics[width=4.35cm]{img/table1_AVG_HD.pdf}\label{fig:avg_hd}}
\subfigure[AVG\_Cos]{\includegraphics[width=4.35cm]{img/table1_AVG_Cos.pdf}\label{fig:avg_Cos}}
\vspace{-0.1cm}
\caption{Overall Performance in terms of all evaluation metrics.}
\label{fig:overall performance}
\vspace{-0.3cm}
\end{figure*}
\begin{figure*}[hbtp!]
\subfigure[AVG\_KL]{\includegraphics[width=4.35cm]{img/robust_avg_kl.pdf} \label{fig:rb_avg_kl}}
\subfigure[AVG\_JS]{\includegraphics[width=4.35cm]{img/robust_avg_js.pdf}\label{fig:rb_avg_js}}
\subfigure[AVG\_HD]{\includegraphics[width=4.35cm]{img/robust_avg_hd.pdf}\label{fig:rb_avg_hd}}
\subfigure[AVG\_Cos]{\includegraphics[width=4.35cm]{img/robust_avg_cos.pdf}\label{fig:rb_avg_Cos}}
\vspace{-0.1cm}
\caption{Effectiveness check for variational module of CLUVAE in terms of all evaluation metrics.}
\label{fig:effect_check}
\vspace{-0.6cm}
\end{figure*}
\subsubsection{Baseline Methods}
We compared our proposed framework (CLUVAE) against the following baseline models:
\textbf{CGAN:}~\cite{mirza2014conditional} is the conditional version of generative adversarial nets (GAN).
Compared with classical GAN, CGAN adds a conditional input to the generator and discriminator respectively.
\textbf{CVAE:}~\cite{sohn2015learning} is another conditional generative model, which is an extension of variational auto encoder.
\textbf{DCGAN:}~\cite{radford2015unsupervised} is a classical image generative model, which utilizes convolutional and convolutional transpose layers in the generator and discriminator respectively.
\textbf{WGAN:}~\cite{pmlr-v70-arjovsky17a} improves the stability of learning of traditional GAN, and makes the learning curve of GAN become meaningful.
\textbf{WGAN-GP:}~\cite{gulrajani2017improved} is an enhanced WGAN, which utilizes gradient penalty to replace clipping weights of WGAN.
\textbf{LUCGAN:}~\cite{wang2020reimagining} is designed for land-use configurations specifically, which can generate configurations automatically based on surrounding contexts' socioeconomic activities.
Besides, to validate the necessity of each component of CLUVAE, we developed four internal variants of CLUVAE:
\textbf{CLUVAE*}, which removes the embedding of human guidance in condition embedding;
\textbf{CLUVAE$'$}, which removes the embedding of surrounding contexts' features in condition embedding.
\textbf{CLUVAE-}, which removes the generated constraints of CLUVAE;
\textbf{CLUVAE\#}, which removes the variational component and only uses a single auto-encoder as the basic framework of CLUVAE.
\subsubsection{Environmental Setting} We conducted all experiments in the Ubuntu 18.04.3 LTS operating system, plus Intel(R) Core(TM) i9-9920X CPU@ 3.50GHz, 1 way SLI Titan RTX and 128GB of RAM, with the framework of Python 3.7.4, Tensorflow 2.0.0.
\subsubsection{Hyperparameters and Reproducibility}
In our experiments, for the land-use configuration model as described in Section \ref{cluvae}, the encoder part has two fully connected layers; the decoder part has two branches and each branch is composed of two fully connected layers.
We randomly shuffle our processed dataset and split it into two independent sets according to the index of the target area: the prior 90$\%$ is the training set and the remaining 10$\%$ is the testing set.
During the model training process, the training epochs is set to 50, the Adaptive Moment Estimation (Adam) optimizer is used to optimize the model with learning rate 0.0001, and the value of $\lambda$ used in total loss $\mathcal{L}$ is set to 0.55.
To get the objective experimental results, we set the same value for these basic hyperparameters in other baseline models.
\subsection{Overall Performance \textbf{Q1}}
To validate the effectiveness of CLUVAE, we compared the generated performance of CLUVAE against the baseline methods in terms of AVG\_KL, AVG\_JS, AVG\_HD, and AVG\_Cos.
Figure \ref{fig:overall performance} shows the CLUVAE outperforms other baseline methods in terms of all evaluation metrics.
This observation indicates that CLUVAE is capable of generating land-use configurations automatically based on the features of surrounding contexts and human instructions.
Meanwhile, compared with other models, CLUVAE captures more inner characteristics of land-use configurations with different green levels.
In summary, this experiments proves the effectiveness of CLUVAE compared with other baseline models.
\subsection{Effectiveness check of variational component \textbf{Q2}}
To resolve data sparsity and improve the robustness of our framework, we added a variational module into CLUVAE.
Thus, we designed the experiment to validate that whether or not the variational module achieves our expectations.
We ran CLUVAE and CLUVAE\# six times respectively with random initialization and recorded the change of each evaluation metric.
Figure \ref{fig:effect_check} shows the changes of all metrics under different models.
Meanwhile, to observe the performance and stability more clearly, we provided the mean and variance of the change of each metric in each subfigure.
Figure \ref{fig:effect_check} shows CLUVAE outperforms CLUVAE\# in terms of all metrics.
A possible explanation for the observation is that the variational module helps CLUVAE capture the characteristics of the distribution of land-use configurations further.
In addition, compared with mean and variance, another observation is that CLUVAE is more stable and robust than CLUVAE\#.
A potential interpretation for the observation is that the function of the variational module is equivalent to data augmentation, which enlarges data samples.
In summary, this experiment proves the variational module of CLUVAE is vital and effective.
\begin{figure*}[hbtp!]
\vspace{-0.3cm}
\subfigure[AVG\_KL]{\includegraphics[width=4.35cm]{img/table3_AVG_KL.pdf} \label{fig:ab_avg_kl}
}
\subfigure[AVG\_JS]{\includegraphics[width=4.35cm]{img/table3_AVG_JS.pdf}\label{fig:ab_avg_js}}
\subfigure[AVG\_HD]{\includegraphics[width=4.35cm]{img/table3_AVG_HD.pdf}\label{fig:ab_avg_hd}}
\subfigure[AVG\_Cos]{\includegraphics[width=4.35cm]{img/table3_AVG_Cos.pdf}\label{fig:ab_avg_Cos}}
\vspace{-0.25cm}
\caption{Ablation study for CLUVAE in terms of all evaluation metrics.}
\label{fig:ablation_study}
\vspace{-0.45cm}
\end{figure*}
\begin{figure*}[hbtp!]
\vspace{-0.3cm}
\subfigure[AVG\_KL]{\includegraphics[width=4.35cm]{img/table2_AVG_KL.pdf} \label{fig:ss_avg_kl}}
\subfigure[AVG\_JS]{\includegraphics[width=4.35cm]{img/table2_AVG_JS.pdf}\label{fig:ss_avg_js}}
\subfigure[AVG\_HD]{\includegraphics[width=4.35cm]{img/table2_AVG_HD.pdf}\label{fig:ss_avg_hd}}
\subfigure[AVG\_Cos]{\includegraphics[width=4.35cm]{img/table2_AVG_Cos.pdf}\label{fig:ss_avg_Cos}}
\vspace{-0.25cm}
\caption{The influence of different square sizes for land-use configuration generation.}
\label{fig:square_size}
\vspace{-0.4cm}
\end{figure*}
\subsection{Ablation study for CLUVAE \textbf{Q3}}
We conducted an ablation study of CLUVAE to validate the necessity of each part of CLUVAE.
Figure \ref{fig:ablation_study} shows the performances of CLUVAE, CLUVAE*, CLUVAE', and CLUVAE- in terms of all evaluation metrics.
We can find that CLUVAE outperforms other variants with great improvements, which indicates the effectiveness of the complete CLUVAE.
In addition, another interesting observation that the performance of CLUVAE* is worse than the performance of CLUVAE'.
From another perspective, this observation reflects that in the land-use configuration generation task, human instructions are more important than the surrounding contexts.
This conclusion is reasonable because the surrounding contexts' features only provide environmental constraints to the generative model, but human instructions determine that the generated land-use configuration belongs to which green level.
Moreover, a careful inspection of Figure \ref{fig:ablation_study} shows that CLUVAE- has a big gap to CLUVAE in terms of all metrics.
The observation reveals that the hierarchical relationship between urban functional zones and land-use configurations plays an important role in the configuration generation task.
In summary, this experiment validates that each part of CLUVAE is indispensable and necessary.
\subsection{The study of the influence of square sizes \textbf{Q4}}
To study the influence of the square size on configuration generation, we set $N=5$, $N=10$, $N=25$, $N=50$, $N=100$ respectively to compare the generative performance.
Here, the smaller value of $N$ is, the larger size of square is.
Figure \ref{fig:square_size} shows the comparison results.
We can find that with the increase of the square size, the value of all metrics decreases.
A possible explanation for the observation is that the larger square size makes the information of land-use configuration contain more coarse-grained and simpler.
The generative model can capture the characteristics of the kind of configurations more easily, thus, the model performance becomes better.
In addition, another interesting observation
is that CLUVAE beats other baseline models in terms of all metrics no matter the square size is.
Meanwhile, a potential interpretation for the observation is that the learned distribution by CLUVAE really captures the characteristics of land-use configurations belongs to different green levels.
In summary, the results show the robustness and superiority of CLUVAE from a square size perspective compared with other baseline models.
\begin{figure*}[htbp]
\centering
\subfigure[Original land-use configuration under different human guidance]{
\begin{minipage}[t]{1.0\linewidth}
\centering
\includegraphics[width=7in]{img/original_land.pdf}
\label{fig:original_land}
\end{minipage}%
}%
\vspace{-0.1cm}
\subfigure[Generated land-use configuration under different human guidance]{
\begin{minipage}[t]{1.0\linewidth}
\centering
\includegraphics[width=7in]{img/generate_land.pdf}
\label{fig:generate_land}
\end{minipage}
}%
\centering
\vspace{-0.3cm}
\caption{Visualization comparison between original land-use configurations and generated land-use configurations }
\label{fig:visual_land}
\vspace{-0.6cm}
\end{figure*}
\subsection{Comparison study between original and generated land-use configurations under different human guidance \textbf{Q5}}
To understand the function of human instructions in land-use configuration generation and observe the visual differences between the original and generated land-use configurations,
we selected the most representative original and generated land-use configurations under different human instructions, and displayed these configurations in a 3-dimensional space.
Figure \ref{fig:visual_land} shows the visualization and comparison results, in which the left color legend indicates the mapping relations between the number of POI categories and colors;
the right 3D part reflects the POI distribution of land-use configurations;
the height of each bar indicates the number of POIs at the corresponding position;
the label of each subfigure is the corresponding human instructions (\textit{i.e., green level label}).
A careful inspection for Figure \ref{fig:original_land} and Figure \ref{fig:generate_land} shows that the generated configurations are organized and contain enough planning information for implementation in realistic.
A possible explanation for the observation is that CLUVAE captures the pros of land-use configurations under different green levels.
Thus, the CLUVAE brings these characteristics into land-use configuration generation.
In addition, for the green level of land-use configurations, Green0 $<$ Green1 $<$Green2 $<$ Green3 $<$ Green4.
Through comparing the inner subfigures in Figure \ref{fig:original_land} and Figure \ref{fig:generate_land}, we can find that with the increase of the green level, the POI distribution of land-use configurations becomes more sparse.
A potential interpretation for the observation is that sparse POI distribution indicates that the corresponding land-use configuration has empty geographical spaces.
Thus, planners can utilize these spaces to plant green plants for improving the green level of the land-use configuration.
In summary, this experiment shows that CLUVAE is capable of capturing the characteristics of land-use configurations belonging to different green levels.
Thus, urban planners can use CLUVAE to customize land-use configurations based on their requirements.
\section{Introduction}
Urban planning aims to generate land-use configuration plans, in order to boost commercial activities, enhance public security, foster social interaction, and thus yield livable, sustainable, and viable environments.
Urban planning is critical because poor planning could directly or indirectly lead to high crimes, traffic congestion and accidents, air pollution, depression and anxiety.
Recently, inspired by image generation, deep urban planing, where deep learning is leveraged to automate the generation of land-use configuration, is emerging.
For example, the study in \cite{wang2020reimagining} proposes a Generative Adversarial Neural network (GAN) method to automatically generate land-use configuration for a region given surrounding contexts. Although it is appealing to teach a machine to plan city configuration, there are certain limitations with existing literature.
First, urban planning indeed is a complex process that involves with public policy, social science, engineering, architecture, landscape, and other related field. In real world urban planning practice, civil and regional planning experts participate to guide the entire planning process. Some communities emphasize more accessibility to transportation and groceries, while other communities expect higher ratio in landscape design. Existing literature in deep automated urban planning lacks human machine interaction to provide personalized planning.
Second, many existing methods directly generate POI distribution at the community level, and lacks the in-depth consideration of spatial hierarchical correlations from functional zones to residential communities.
Third, city configurations are usually sparse and even imperfect, and, thus, has imposed a high demand for robust learning algorithm in sparse data.
\begin{figure}[htbp]
\vspace{-0.5cm}
\centering
\includegraphics[width=0.4\textwidth]{img/AI_generation.pdf}
\vspace{-0.1cm}
\caption{Human input textual instructions to guide AI to generate land-use configurations}
\label{landuse image}
\vspace{-0.3cm}
\end{figure}
This paper focuses on the problem of deep human-guided urban planning using imperfect urban geographic and human mobility data.
The basic idea is to formulate the problem into a novel machine learning task of robust generative modeling with spatial hierarchy awareness for sparse data.
The new machine learning model should have the ability to generate a POI distribution tensor as land-use configuration, to overcome data sparsity, to capture spatial hierarchy structure from functional zones to residential communities in the planning process.
There are three technical challenges in addressing such machine learning task: 1) \textbf{Challenge 1: human guidance }: in deep urban planing, the guidance from domain experts is needed. So, how can we provide a mechanism to allow humans to interact with and guide deep learning models to generate personalized planning?
2) \textbf{Challenge 2: spatial hierarchy}:
how can we teach a machine to learn the spatial hierarchy by first planing urban functions at the zone level, then planning POI distribution at the community level? The underlying idea of this strategy is to add structured regularization into the planning generation process.
3) \textbf{Challenge 3: robustness modeling against data sparsity}: community-level configuration data are expensive, imperfect, and sparse. Thus, there is limited quality training data available. As a result, to overcome such practical data challenge, we need to answer: how can we equip the planning generative model with robustness against data sparsity?
As a result, we propose a deep conditional variational encoder-decoder framework with awareness of spatial hierarchy to address these challenges jointly.
The main contributions of our study are summarized as follows:
1) \textbf{Formulating the automated urban planning into a novel deep encoder-decoder learning task.} The encoder is used to learn the relationship between condition inputs and the corresponding land-use configurations.
The decoder is regarded as a generator to generate the best land-use configurations for planning based on condition inputs.
2) \textbf{Condition embedding as a bridge to integrate human guidance.} To allow humans to guide the generation process and provide an interface for human-planner interaction, a human can describe personalized planning needs as texts.
In addition, many factors such as government policy, traffic conditions, etc should be considered for planning reasonable configurations.
Thus, we convert human description texts and the factors in surrounding contexts into embedding respectively, and concatenate the two embeddings as the condition embedding that conditions on both encoder and decoder.
In this way, human textual inputs are involved in the model learning process.
3) \textbf{Variational additive Gaussian as an augmented tool to fight data sparsity.} To overcome the sparsity issue of land-use configuration data, we make the encoder return a distribution over the latent embedding space instead of a single embedding point. Such distribution can better approximate the data spray pattern in the embedding space, then we can keep drawing samples from the distribution, instead of requiring large-scale quality training configuration data. In addition, this strategy will add randomness in embedding, diversify representations, and improve model robustness.
4) \textbf{Dual reconstruction to capture spatial hierarchy.} We force the decoder to reconstruct not just the urban function labels of various zones, but also the POI distribution tensor of various communities, in order to capture the zone-community hierarchical correlation.
5) \textbf{Comprehensive experiments to validate the effectiveness of our framework.}
We evaluated the proposed methods using Beijing's urban geography, human mobility, land-use configuration, POI checkin data. For comparison we implemented a broad range of other algorithms. Results show that the proposed methods consistently outperformed 6 competing methods. We also performed ablation study, sensitivity study, human guided conditional generation study to justify the effectiveness and superiority of our technical insights.
\section*{Acknowledgment}
This research was partially supported by the National Science Foundation (NSF) via the grant numbers: 1755946, 2040950, 2006889, 2045567.
\bibliographystyle{IEEEtran}
\section{Proposed Method}
In this section, we introduce the technical details of our proposed framework CLUVAE.
\subsection{Framework Overview}
Figure \ref{fig:framework} shows our proposed framework, namely CLUVAE.
The proposed framework includes four steps:
Firstly, we collect the land-use configurations and the corresponding urban functional zones of a series of target areas;
Secondly, we obtain the embedding of surrounding contexts' features and human guidance respectively.
For the embedding of surrounding contexts' features, we utilize a spatial attributed graph to organize them and employ a graph embedding model to obtain the corresponding embedding.
For the embedding of human guidance, we regard the one-hot vector of the texts of human guidance as the embedding.
Thirdly, we build up a deep robust configuration generation model.
Specifically, we first concatenate the surrounding contexts' embedding and human guidance embedding as the condition embedding.
Then, the encoder of the generation model learns the distribution that reflects the correlation between the condition embedding and the corresponding land-use configuration.
Next, the decoder of the generation model reconstructs land-use configurations and urban functional zones based on the condition embedding and the latent embedding sampled from the previously learned distribution.
Finally, the well-trained decoder is our desired land-use configuration generator.
\subsection{Finding Target Area and Surrounding Contexts}
How to find target area and surrounding contexts is the cornerstone of our framework.
Specifically, we first collect a list of geographical points that are composed of longitude and latitude coordinates.
Then, we regard each geographical point as the central point to draw the corresponding target area (i.e. one square shape with 1 km$^2$ area).
Later, we find all surrounding contexts of each target area according to the spatial relationships depicted in section \ref{target}.
After that, we collect all target areas and surrounding contexts, and pair them together.
In the following paragraphs, to be convenient, we employ the $k$-th target area and the corresponding surrounding contexts to introduce denotations and all calculation processes.
\subsection{Collecting Land-use Configurations and Discovering Urban Functional Zones}
To build up an AI-based land-use configuration generation model, we need to collect a series of land-use configuration samples.
The collecting process of the $k$-th target area as follows:
we first divide the $k$-th target area into $N \times N$ squares based on latitude and longitude.
Next, we count the number of POIs of each POI category in each square, and organize the numbers of POIs belonging to the same POI category as a matrix according to the squares' position.
After that, we stack different matrices together to form a latitude-longitude-channel tensor.
The format of the tensor as shown in Figure \ref{landuse image} and the $k$-th land-use configuration denoted by $\mathbf{\hat{X}}^{(k)} \in \mathbb{R}^{N \times N \times M}$, where $M$ is the number of POI category.
In addition, urban functional zones provide a rough skeleton for urban planning, which stimulates us to employ them in our generation model for improving generative performance.
Owing to lacking of standard urban functional zones data samples, we utilize the method in \cite{yuan2014discovering} to discover urban functional zones of target areas based on spatio-temporal data.
Limited by space, we only provide the data format of the urban functional zones.
Figure \ref{urban_func} has provided an example of urban functional zones.
For the $k$-th target area, the urban functional zones denoted by $\mathbf{F}^{(k)} \in \mathbb{R}^{N \times N}$, where each value in $\mathbf{F}^{(k)}$ indicates that the corresponding geographical square belongs to which urban functional zone.
\subsection{Extracting Features of the Surrounding Contexts}
The socioeconomic features of surrounding contexts affect the urban planning design of the target area.
For instance, if surrounding contexts have many apartments for people living, the target area should be constructed entertainment POIs (\textit{e.g. Gym, Club}) to improve the vibrancy of all geographical area.
In this paper, we utilize a spatial attributed graph to embrace all features of surrounding contexts, and employ a graph embedding model to convert these features into an embedding.
Firstly, from section \ref{target}, we can find that
the spatial structural relationship among the surrounding contexts can be regarded as a ring that is a special graph.
For the surrounding contexts of the $k$-th target area, the special graph denoted by $\mathcal{G}^{(k)} = (\mathcal{V}^{(k)},\mathcal{E}^{(k)})$, where $\mathcal{V}^{(k)}$ indicates all vertices of the graph, and one vertex is one surrounding context;
$\mathcal{E}^{(k)}$ represents all edges of the graph, and each edge indicates the connectivity between any two vertices.
Secondly, we extract explicit features of each vertex from three perspectives:
(1) \textbf{house price change}, which reflects people's living preference for a geographical area.
Specifically, for each vertex, we first collect the house price in $T$ months.
Then, we obtain the house price change by utilizing the house price of each month to deduct the previous one.
Finally, we collect the house price change of all contexts, denoted by $\mathbf{V} = \{\mathbf{v}_1,\mathbf{v}_2,\dots,\mathbf{v}_8\}$, where $\mathbf{V} \in \mathbb{R}^{8 \times (T-1)}$ and $\mathbf{v}$ is the house price change of a context.
(2) \textbf{POI Ratio}, which describes the distribution of urban functions of one geographical area from the overall perspective.
Specifically, for each vertex, we first count the number of POIs belonging to each POI category.
We then scale the value of each POI category by dividing the total number of POIs in the context, and organize these scaled values as a POI ratio vector.
Finally, we collect the POI ratio vector of all contexts, denoted by $\mathbf{R}=\{\mathbf{r}_1,\mathbf{r}_2,\dots,\mathbf{r}_8\}$, where $\mathbf{R} \in \mathbb{R}^{8 \times M}$
and $\mathbf{r}$ is the POI ratio vector of a context.
(3)\textbf{Transportation}, which demonstrates the traffic condition of a geographical area.
Here, we extract features related to transportation from two sides: public transportation and private transportation.
For \textbf{public transportation}, in each context, we collect 5 values: (a) the leaving volume of buses per day; (b) the arriving volume of buses per day; (c) the transition volume of buses per day; (d) the number of bus stops in per square meters; (e) the average price of each bus trip.
The five values are organized as the public transportation feature vector, denoted by $\mathbf{o}$.
Finally, we collect the vectors of all contexts, denoted by $\mathbf{O} = \{ \mathbf{o}_1, \mathbf{o}_2, \dots, \mathbf{o}_8 \}$, where $\mathbf{O} \in \mathbb{R}^{8 \times 5}$.
For \textbf{private transportation},
in each context, we also collect 5 values: (a) the leaving volume of taxis per day; (b) the arriving volume of taxis per day; (c) the transition volume of taxis per day; (d) the average velocity of taxis per day; (e) the average commute distance of taxis per day.
The five values are organized as the private transportation feature vector, denoted by $\mathbf{u}$.
Finally, we collect the vectors of all contexts, denoted by $\mathbf{U}=\{ \mathbf{u}_1, \mathbf{u}_2, \dots, \mathbf{u}_8\}$, where $\mathbf{U} \in \mathbb{R}^{8\times 5}$.
Thirdly, we horizontally concatenate the feature vectors of all contexts together to form the feature set, denoted by $\mathcal{F}^{(k)}=[\mathbf{V},\mathbf{R},\mathbf{O},\mathbf{U}]$.
$\mathcal{F}^{(k)} \in R^{8\times (M+T+9)}$, where one row indicates all kinds of feature of one vertex.
Then, we add attributes to the graph $\mathcal{G}^{(k)}$ by matching each row of $\mathcal{F}^{(k)}$ to the corresponding vertex for forming the spatial attributed graph, denoted by $\mathcal{G}^{(k)} = (\mathcal{V}^{(k)},\mathcal{E}^{(k)},\mathcal{F}^{(k)})$.
Finally, a variational graph auto encoder model \cite{kipf2016variational} is used to convert $\mathcal{G}^{(k)}$ into the graph embedding $\mathbf{s}^{(k)}$ that contains not only the spatial structural information among surrounding contexts but also all socioeconomic features.
\begin{figure*}[!t]
\centering
\includegraphics[width=0.75\linewidth]{img/model_structure.pdf}
\captionsetup{justification=centering}
\caption{Land-use Configuration Generation Model.}
\vspace{-0.5cm}
\label{fig:model_structure}
\end{figure*}
\subsection{Embedding Human Guidance}
Human guidance describes humans' requirements for land-use configuration generation.
Urban planning is a highly personalized task, because different urban planners have different planning ideas,
meanwhile, different urban planning projects are constrained by different realistic factors.
So how to make our generated model perceive human guidance is a big challenge.
To overcome this challenge, we simplify the human guidance into options that can be selected by urban planners according to their needs.
For instance, if urban planners have greenery rate requirements for producing a land-use configuration, they will have different greenery levels to select.
After the selection, our model can produce the corresponding result based on planners' choice.
Here, we use the one hot vector of the selected level label as the embedding of human guidance.
The human guidance for the $k$-th target area, denoted by $\mathbf{i}^{(k)}$.
Ultimately, for the $k$-th target area, we have obtained the land-use configuration $\mathbf{\hat{X}}^{(k)}$, the urban functional zones $\mathbf{F}^{(k)}$, the embedding of surrounding contexts' features $\mathbf{s}^{(k)}$, and the embedding of the human guidance $\mathbf{i}^{(k)}$.
Then, we organize them as a group $[\mathbf{\hat{X}}^{(k)},\mathbf{F}^{(k)},\mathbf{s}^{(k)}, \mathbf{i}^{(k)}]$, which is one data sample learned by our following model structure.
\subsection{Conditional Land-use Variational Autoencoder}
\label{cluvae}
Our model aims to find the mapping function between land-use configurations and the corresponding conditions (\textit{i.e. surrounding contexts and human instructions}).
To achieve this goal, our framework is implemented through an encoder-decoder paradigm.
The encoder part firstly converts the land-use configuration $\mathbf{\hat{X}}^{(k)}$ and the conditions ($\mathbf{s}^{(k)}, \mathbf{i}^{(k)}$) into a latent embedding vector $\mathbf{z}^{(k)}$.
During the conversion process,
we utilize a variational Gaussian embedding component to find the conditional distribution $p(\mathbf{z}^{(k)} | \mathbf{\hat{X}}^{(k)}, \mathbf{s}^{(k)}, \mathbf{i}^{(k)})$.
Owing to we assume the conditional distribution belongs to the family of normal distribution, the encoder part only needs to estimate the parameters (\textit{i.e. mean $\bm{\mu}^{(k)}$, variance $\bm{\delta}^{(k)}$}) of the distribution and produce the latent embedding $\mathbf{z}^{(k)}$.
Specifically, we first concatenate the embedding of surrounding contexts' features $\mathbf{s}^{(k)}$ and the embedding of human instructions $\mathbf{i}^{(k)}$ as the condition embedding $\mathbf{c}^{(k)}$.
Then we flatten the land-use configuration $\mathbf{\hat{X}}^{(k)}$ into a one-dimensional vector, denoted by $\mathbf{x}^{(k)}$.
Later, we input $\mathbf{c}^{(k)}$ and $\mathbf{x}^{(k)}$ into two fully connected layers respectively to learn the parameters of the conditional distribution $p(\mathbf{z}^{(k)} | \mathbf{x}^{(k)} , \mathbf{c}^{(k)}) $, where $p(\mathbf{z}^{(k)} | \mathbf{x}^{(k)} , \mathbf{c}^{(k)}) $ equals $p(\mathbf{z}^{(k)} | \mathbf{\hat{X}}^{(k)}, \mathbf{s}^{(k)}, \mathbf{i}^{(k)})$.
After that, we utilize a reparametrization technique to sample latent embedding $\mathbf{z}^{(k)}$ from the learned distribution.
The whole calculation process can be formulated as follows:
\begin{equation}
\left\{
\begin{array}{lr}
\mathbf{c}^{(k)} = \text{Concat}(\mathbf{s}^{(k)},\mathbf{i}^{(k)}),\\
\mathbf{x}^{(k)} = \text{Flatten}(\mathbf{\hat{X}}^{(k)}),\\
\bm{\mu}^{(k)} = \text{Fully}_1(\mathbf{x}^{(k)},\mathbf{c}^{(k)}),
\\
\bm{\delta}^{(k)} = \text{Fully}_2(\mathbf{x}^{(k)},\mathbf{c}^{(k)}),
\\
\mathbf{z}^{(k)}=\bm{\mu}^{(k)}+\bm{\delta}^{(k)} \times \epsilon.
&
\end{array}
\right.
\label{equ:encoder}
\end{equation}
The last line of equation \ref{equ:encoder} is the reparameterization trick, where $\epsilon$ sampled from standard normal distribution $N(0,1)$.
Next, we input $\mathbf{z}^{(k)}$ and $\mathbf{c}^{(k)}$ into the decoder part to conduct the decoding step.
The decoder part reconstructs $\mathbf{\hat{X}}^{(k)}$ and constructs the corresponding urban functional zones $\mathbf{F}^{(k)}$ based on $\mathbf{z}^{(k)}$ and $\mathbf{c}^{(k)}$.
Inspired by \cite{ding2020guided}, the multi-head decoder in VAE can improve model performance by adding certain constraints.
Here, we regard urban functional zones as the generated constraints in the decoder for producing reasonable land-use configurations.
Specifically, we first input $\mathbf{z}^{(k)}$ and $\mathbf{c}^{(k)}$ a fully connected layer activated by Relu function to get the reconstructed land-use configuration $\mathbf{\hat{\ddot{X}}^{(k)}}$.
Next, we input $\mathbf{z}^{(k)}$ and $\mathbf{c}^{(k)}$ into another fully connected layer activated by sigmoid function $\sigma$ to construct urban functional zones $\mathbf{\ddot{F}}^{(k)}$.
The calculation process can be formulated as:
\begin{equation}
\left\{
\begin{array}{lr}
\mathbf{\hat{\ddot{X}}^{(k)}} = \text{Relu}(\text{Fully}_3(\mathbf{z}^{(k)},\mathbf{c}^{(k)})),\\
\mathbf{\ddot{F}^{(k)}} =\sigma(\text{Fully}_4(\mathbf{z}^{(k)},\mathbf{c}^{(k)})).
&
\end{array}
\right.
\label{equ:decoder}
\end{equation}
There are three learning objectives during the training phase:
(1) minimizing the difference between $\mathbf{\hat{\ddot{X}}^{(k)}}$ and $\mathbf{\hat{X}^{(k)}}$, denoted by $\mathcal{L}_X$;
(2) owing to we assume the prior distribution of $p(\mathbf{z}^{(k)} | \mathbf{x}^{(k)} , \mathbf{c}^{(k)}) $ is $q(\mathbf{z}^{(k)})$ that is a standard normal distribution $N(0,1)$, we need to minimize the gap between the two distributions, denoted by $\mathcal{L}_p$;
(3) minimizing the difference between $\mathbf{\ddot{F}^{(k)}}$ and $\mathbf{F^{(k)}}$, denoted by $\mathcal{L}_F$;
The final loss $\mathcal{L}$ is the combination of $\mathcal{L}_X$, $\mathcal{L}_p$, and $\mathcal{L}_F$, which can be formulated as follows:
\begin{equation}
\left\{
\begin{array}{lr}
\mathcal{L}_X= \frac{\sum^{K}_{k=1}( \mathbf{\hat{X}^{(k)}} - \mathbf{\hat{\ddot{X}}^{(k)}})^2}{K}
\\
\mathcal{L}_p= \frac{\sum^{K}_{k=1}\text{KL}[p(\mathbf{z}^{(k)}|\mathbf{x}^{(k)} , \mathbf{c}^{(k)}) || q(\mathbf{z}^{(k)})]}{K}
\\
\mathcal{L}_F= - \frac{\sum_{k=1}^{K} \mathbf{F}^{(k)}\times \text{log}(\mathbf{\ddot{F}}^{(k)})}{K}
\\
\mathcal{L} = \mathcal{L}_X + \mathcal{L}_p + \lambda \times \mathcal{L}_F
&
\end{array}
\right.
\label{equ:decoder}
\end{equation}
where KL represents the Kullback-Leibler divergence between $p(.)$ and $q(.)$;
$\lambda$ controls the strength of the generated constraints (\textit{i.e. urban functional zones}) during the model learning process.
The value range of $\lambda$ is $[0 \sim 1]$.
After the generation model converges, the decoder is the land-use configuration generator.
During the testing phase, we first sample a latent embedding from the learned distribution based on the condition embedding.
Then, the decoder takes the sampled latent embedding and the condition embedding as input, and outputs the generated land-use configuration.
\subsection{Difference from Prior Literature}
With the development of artificial intelligence (AI), how to utilize AI models to improve and accelerate urban planning attracts many researchers.
However, there are certain limitations in prior literature: 1) the automated urban planning process cannot be personalized based on human's requirements; 2) the generated urban planning solutions are unreasonable;
3) the available standard urban planning solution is sparse.
Our study aims to overcome the three limitations.
In our framework, the human guidance for urban planning is regarded as a condition input of the land-use configuration generation model.
In addition, We add generated constraints (\textit{i.e., urban functional zones}) in the generation model for improving the reasonability of generated results.
Moreover, we employ a variational Gaussian embedding mechanism to make the model learning process better.
This is how this study differentiates from advanced prior literature.
\section{Definitions and Problem Statement}
\subsection{Definitions}
\subsubsection{Target Area and Surrounding Contexts}
\label{target}
Target area refers to a geographical square region, in which is empty and waiting for being planned.
Surrounding contexts possess the same square shape as the target area and surround the target area from different directions.
Figure \ref{residential_context} illustrates the spatial relationship between the target area and the surrounding contexts.
There are many socioeconomic activities such as commute, real estate transaction, entertainment, and etc in surrounding contexts.
These socioeconomic characteristics of surrounding contexts affect the design of the land-use configuration of the target area.
\begin{figure}[htbp]
\vspace{-0.3cm}
\centering
\includegraphics[width=0.3\textwidth]{img/target_area.pdf}
\caption{A target area and the surrounding contexts.}
\label{residential_context}
\vspace{-0.3cm}
\end{figure}
\subsubsection{Land-use Configuration}
\label{landuse}
refers to the POI distribution in a geographical region.
To make our model perceive the land-use configuration easily, we adopt the quantitative definition in \cite{wang2020reimagining}.
Figure \ref{landuse image} shows an example of a land-use configuration that has three dimensions: latitude, longitude, and POI category.
To see the details of the configuration further, we pick up the purple layer (\textit{i.e store}) to exhibit.
The latitude and longitude of this layer are divided into $N \times N$ parts.
The value of each little square represents the number of stores in the corresponding geographical location.
Other layers in the configuration have the same data structure as the purple layer.
Thus, the whole tensor reflects the POI distribution of all POI categories in the geographical region.
\begin{figure}[htbp]
\vspace{-0.3cm}
\centering
\includegraphics[width=0.5\textwidth]{img/landuse.pdf}
\caption{An example of land-use configuration.}
\label{landuse image}
\vspace{-0.3cm}
\end{figure}
\subsubsection{Urban Functional Zone} refers to a geographical cluster that owns a specific land-use function, such as industrial zone, educational zone, and etc.
Compared with land-use configuration, the urban functional zone is more coarse-grained in the urban planning domain.
Figure \ref{urban_func} show an example of urban functional zones, in which different colors indicate different functional zones.
The same to land-use configuration, we also divide a geographical region into $N \times N$.
Each functional zone consists of certain little squares existing in land-use configurations.
Thus, compared with land-use configuration, urban functional zones are more coarse-grained and provide a rough sketch for urban planning.
\begin{figure}[htbp]
\vspace{-0.3cm}
\centering
\includegraphics[width=0.3\textwidth]{img/Func.pdf}
\vspace{-0.2cm}
\caption{An example of urban functional zones.}
\label{urban_func}
\vspace{-0.3cm}
\end{figure}
\subsubsection{Human Guidance}
\label{human}
refers to the humans' requirements for generating a land-use configuration.
Here, these requirements are described by corresponding texts.
And, we regard the embeddings of these texts as conditional inputs of our land-use configuration generation model.
\begin{figure*}[!t]
\centering
\includegraphics[width=1.0\linewidth]{img/framework.pdf}
\vspace{-0.2cm}
\captionsetup{justification=centering}
\caption{Framework Overview.}
\vspace{-0.6cm}
\label{fig:framework}
\end{figure*}
\subsection{Problem Statement}
In this paper, we aim to build up an automated land-use configuration generator that can generate configurations via considering the socioeconomic characteristics of surrounding contexts and human guidance.
Formally, given a list of land-use configurations $\mathbf{\tilde{X}} = \{ \mathbf{\hat{X}}^{(1)},\mathbf{\hat{X}}^{(2)},\dots,\mathbf{\hat{X}}^{(K)} \}$,
socioeconomic features of surrounding contexts $\mathbf{S} = \{ \mathbf{s}^{(1)},\mathbf{s}^{(2)},\dots,\mathbf{s}^{(K)}\}$,
human guidance $\mathbf{I}=\{ \mathbf{i}^{(1)},\mathbf{i}^{(2)},\dots,\mathbf{i}^{(K)}\}$.
Our purpose is to find the mapping function $f:(\mathbf{S}, \mathbf{I}) \rightarrow \mathbf{\tilde{X}}$ that takes socioeconomic features $\mathbf{S}$ and human guidance $\mathbf{I}$ as input, and outputs the corresponding land-use configurations $\mathbf{\tilde{X}}$.
\section{Related Work}
\textbf{Data Fusion}
refers to integrating multiple data sources for capturing more accurate and valuable information ~\cite{meng2020survey}.
In reality, there are many application scenarios of artificial intelligence that need the data fusion process to capture more informative features for better modeling.
For instance, Wang {\it et al.} ~\cite{wang2021reinforced} proposed a user profiling framework based on reinforcement learning via analyzing the check-in data and geographical POI data.
Liu {\it et al.} ~\cite{liu2018modeling} modeled the interaction coupling of multi-view spatiotemporal contexts to predict shared-bike destination.
Wang {\it et al.} ~\cite{wang2021measuring} studied the urban vibrancy of residential communities via analyzing the multiple data resources such as check-ins, urban geography data, and human mobility.
Wang {\it et al.} ~\cite{wang2020defending} fused the monitoring value of different kinds of sensors together to detect the anomaly status of water treatment plants.
Compared with these works, in this paper, we face more complicated data resources.
Considering the complexity of urban planning, we employ a spatial attributed graph to organize all socioeconomic characteristics of surrounding contexts together, and utilize a graph embedding method to preserve these characteristics into an embedding vector.
Moreover, for integrating the information of human guidance and surrounding context, we concatenate the one-hot vector of human guidance and the embedding of surrounding contexts as our model input.
\textbf{Deep Variational Auto Encoder (VAE)} is a classical kind of deep generative model ~\cite{oussidi2018deep}.
Traditional VAE models achieve great success in many generated applications such as image generation, text generation, etc ~\cite{zhu2020s3vae,luo2020mg,akbari2018semi}.
For instance,
Zhang {\it et al.} ~\cite{zhang2021image} proposed a framework, namely VAE-AGAN, which integrates VAE and generative network (GAN) together to augment regions of interest (ROIs) image data for improving the ROIs detection of the model.
Chen {\it et al.} ~\cite{chen2021trajvae} utilized a VAE model to generate trajectories based on the characteristic of trajectories learned by LSTM for fulfilling the needs of self-driving simulation and other traffic analysis tasks.
However, although VAE can generate data samples by simulating the data distribution, the output of VAE cannot be controlled.
Conditional VAE is proposed to resolve the limitation, and obtain remarkable achievements ~\cite{jiang2020transformer,zhang2020disentangling}.
For example, Yonekura {\it et al.} \cite{yonekura2021data} developed a new turbine designer
based on the conditional VAE technique for improving the design efficiency.
Compared with these works, our CLUVAE is also based on the conditional VAE setting.
But we add a regularization item to the loss function during the optimization process for producing more reasonable results.
\textbf{Urban Planning} is vital for the future development of a geographical area, which designing the land use and environment of the area according to human requirements and surrounding contexts ~\cite{oliveira2010evaluation,naess2001urban}.
However, to attain effective urban planning solutions, human experts have to spend much attention to consider lots of constraints such as government policy, environmental protection, etc.
For instance, Ratcliffe {\it et al.} ~\cite{ratcliffe2009urban} studied the relationship between urban planning and real estate development.
Papa {\it et al.} ~\cite{papa2016smart} studied how to build up a sustainable smart city via urban planning.
Recently, with the development of deep learning, many researchers bring deep models into the urban planning domain ~\cite{wang2020reimagining,nguyen2020simulating}.
For instance, Wang {\it et al.} ~\cite{wang2020reimagining} utilized GAN-based framework to generate land-use configuration based on surrounding contexts.
Compared with these works, CLUVAE has higher customization ability and generates more reasonable configurations.
|
\section{Introduction}
Surfaces and interfaces, i.e., the outermost layer of atoms
of a material,
define how it interacts with its surroundings \cite{Lueth2015}. It is therefore
critically important to study the physical and chemical properties of surfaces.
For more than three decades,
atomic force microscopy (AFM) \cite{PhysRevLett.56.930,Gavara2017} has
been an indispensable tool to explore the topography of surfaces
and their electrical, magnetic and elastic properties \cite{Magonov2008}.
It has been applied to study a plethora of phenomena
at the nano-scale, including the measurement
of forces acting on individual molecules \cite{Hinterdorfer2006},
the study of biological samples \cite{Kreplak2016}, and
nanoparticles \cite{Pyrgiotakis2014}.
Conventional AFM consists of a cantilever with a
nanometer-sized tip that interacts with the sample
while being dragged over its surface (contact mode
\cite{PhysRevLett.56.930}). The properties
of the sample, and in particular the force between tip and sample, are
measured by monitoring the motion of the cantilever.
In order to minimize the damage to the sample, the cantilever can be
made to rapidly tap over the sample surface while scanning.
The distance to the surface is
manipulated by applying an oscillating,
single-frequency piezoelectric force.
This mode is known as dynamic AFM \cite{Garcia2002}.
Our work focuses on intermodulation AFM \cite{Platz2008}, where
the cantilever is driven by a two-frequency force.
We pursue several goals in this work, which
highlight the advantages of dynamic AFM
and shed light on the conditions and parameter settings required for it.
First, we provide a theoretical
description of the cantilever dynamics within the context of
statistical mechanics. This enables us to study tip
fluctuations, opening a path to monitor additional
properties of the tip-surface interaction. We focus in particular on
squeezing of thermal fluctuations, which could increase the resolution of
AFM. Squeezing in AFM has been studied before
within a quantum-mechanical model by Passian and Siopsis
\cite{PhysRevA.94.023812, Passian2017}.
Our second goal is to find a driving scheme that enhances the
generation of squeezing. Since this essentially amounts to maximizing
the effect of the tip-surface interaction, the principle of this
scheme -- maximizing the interaction time -- can be applied to the
measurement of other parameters as well.
The generation of squeezing in AFM can be compared
to classical nonlinear optics, where a
coupling between different harmonics may trigger
phase squeezing through the
Kerr effect \cite{MandelWolf}.
Our third goal is to explore under which circumstances quantum effects
may be measurable in dynamic AFM. There is a growing interest in
quantum effects with
micromechanical resonators
\cite{Aspelmeyer2014Book,Bowen2016Book,Riedinger2018}
or even larger objects \cite{MercierdeLepinay2021,Yu2020}.
A mature experimental design such as AFM may help to progress this
field. In normal operations, AFM works well within the classical regime.
For this reason, we have used classical statistical
mechanics for most of our results. However, we also
describe the cantilever as an
open quantum system \cite{GardinerZollerQuantumNoise} to
distinguish quantum and classical dynamics. Of particular interest is the
role of dissipative surface forces, which appear to be particularly
suitable to generate quantum effects.
The paper is organized as follows. In Sec.~\ref{sec:classicalPicture},
we present the driving scheme to maximize the interaction time.
Secs.~\ref{sec:dynamicalEqns} and \ref{sec:skew} are devoted to the
description of the cantilever as a statistical system. In
Sec.~\ref{sec:results}, we discuss our predictions for
the generation of squeezing and skewness.
We expound on the
physics behind squeezing generation in AFM
and the influence of quantum fluctuations in
Sec.~\ref{sec:discussion}. This is followed by a conclusion
\ref{sec:conclusions}.
Several appendices contain the details of our theoretical methods.
\section{Driving scheme to maximize surface interaction}\label{sec:classicalPicture}
Cantilever and tip of AFM form a complex mechanical system that may
include bending motion and torsion \cite{Passian2017}, but a simple
harmonic oscillator model often suffices and will be used here.
The tip is described in phase space
as a point particle with position $x(t)$ and
momentum $p(t)$ and its dynamics is governed
by Newton's second law,
\begin{align}
\dot{p} &= -k x-\gamma_Q p +F_\text{dr}(t) + F_\text{sf} +
F_\text{dis}.
\end{align}
Here, $x$ denotes the position of the tip above the sample surface.
The tip is in the engaged position with the sample
for $x\approx -h$. Position $x=0$ corresponds to the
equilibrium position of the free tip, and $k$ denotes the spring constant
of the cantilever. Rate $\gamma_Q$ describes internal mechanical
losses of the cantilever. $F_\text{dr} (t)$
denotes a driving force that puts the cantilever in motion. The
tip-surface interaction is decomposed into a conservative part
$F_\text{sf}$ and a dissipative part $F_\text{dis}$. The dissipative
part arises from the deformation of the sample surface due to the
interaction with the tip. It can be modelled in different ways, for
instance through
Kelvin-Voigt viscoelastic dissipation \cite{Benaglia2019},
a hysteretic force \cite{Johnson1971,Dankowicz2008},
a convolution integral \cite{Lee1960}, or a
retarded response to the sample \cite{Ting1966}.
In this
paper, we follow Ref.~\cite{Platz2013} and model the dissipate surface
force by a position-dependent drag force, $F_\text{dis}=-\gamma(x) p$.
\begin{figure}
\begin{center}
\includegraphics[width=7cm]{drivingScheme.png}
\caption{\label{fig:drivingScheme} Proposed driving scheme. For
a periodic motion, the tip would remain close to the
sample (red rectangle near $x=-h$) if the motion would take the form
of a square wave (orange
dotted curve). In single-frequency AFM (gray, dashed), the tip spends little
time near the sample, but the suggested two-frequency driving scheme
(blue, solid) approximates a square wave.}
\end{center}
\end{figure}
In dynamic AFM, the driving force typically takes the form of a
homogeneous force with one or two frequency components. Here,
we consider two angular frequency components $\omega_1, \omega_2$,
\begin{align}
F_\text{dr}(t) &= F_1 \sin(\omega_1 t+\phi_1)
+ F_2 \sin(\omega_2 t+\phi_2),
\label{eq:twoFreqDrivingForce}\end{align}
with $\phi_1,\phi_2$ controlling the relative phase of the force components.
In absence of the sample, this model corresponds to an elementary
driven damped harmonic oscillator. After a transient period, the
oscillator will settle to a steady-state motion of the form
\begin{align}
x_0(t) &= \frac{ F_1}{M \omega_1} \text{Im}
\left (
\frac{ e^{i\omega_1 t+i\phi_1}}{Z(\omega_1)}
\right )+
\frac{ F_2}{M \omega_2} \text{Im}
\left (
\frac{ e^{i\omega_2 t+i\phi_2}}{Z(\omega_2)}
\right ),
\label{eq:xUnperturbed}\end{align}
where
\begin{align}
Z(\omega) &= \frac{ \omega_0^2-\omega^2}{\omega} + i \gamma_Q
\label{eq:defImpedance}\end{align}
is the mechanical impedance of the cantilever.
$M$ denotes the reduced mass
of cantilever and tip, and $\omega_0 = 2\pi f_0 =\sqrt{k/M}$ the
resonance angular frequency.
The principle of the proposed driving scheme is explained in
Fig.~\ref{fig:drivingScheme}. In dynamic AFM, the driving force
induces a periodic motion of the cantilever. If the force could be
arbitrarily strong, the interaction time with the surface would be
maximized for a square wave motion, where the tip spends half the
time period close to the surface during each cycle. In a two-frequency driving scheme,
one can maximize the time the tip spends near the surface if the two
frequency components of the motion correspond to the first two Fourier
coefficients of a square wave. Eq.~(\ref{eq:xUnperturbed}) implies
that this can be achieved if we pick the driving frequencies as
$\omega_2=3\omega_1$, and the force amplitudes
so that they produce the Fourier coefficients, $\frac{ F_1}{M \omega_1}
\text{Im}e^{i\phi_1} Z^{-1}(\omega_1)=\frac{4}{\pi}h$ and
$\frac{ F_2}{M \omega_2}
\text{Im}e^{i\phi_2} Z^{-1}(\omega_2)=-\frac{4}{3\pi}h$.
For the choice $\omega_1=\omega_0$,
$Z(\omega_0)$ is imaginary and $Z(3\omega_0)$ is a negative real
number for $\gamma_Q\ll \omega_0$. The phases then can be chosen as
$\phi_1=0$ and $\phi_2=\frac{ \pi}{2}$. However, we will see below
that the optimal choice of phases will depend on the surface interaction when
the tip interacts with the sample.
\section{Dynamical equations for probability moments}\label{sec:dynamicalEqns}
A real tip is not a point particle, and its motion is generally
subject to thermal or quantum fluctuations, which require
a probabilistic description. The position of the tip is then replaced
by the mean position $\bar{x}(t) = \langle x \rangle $. Here, angle
brackets denote averaging with respect to a probability distribution,
which can be of classical or quantum nature. Variances are described
through mean values of quadratic expressions, such as
$\Delta x^2 = \langle (x-\bar{x})^2 \rangle $. More generally, one can
characterize a probability distribution through
its moments \cite{SpanosProbability},
\begin{align}
\Delta_{n,m} &= \frac{ 1}{2}
\left \langle (x-\bar{x})^n (p-\bar{p})^m +
(p-\bar{p})^m (x-\bar{x})^n
\right \rangle .
\label{def:DeltanmMainPart}\end{align}
In a classical description, the ordering of position and momentum
terms is irrelevant, but we keep a symmetric ordering so that our
formalism can be used for a quantum description as well.
Second-order moments describe position variance $\Delta x^2 =\Delta_{2,0}$,
momentum variance $\Delta p^2 =\Delta_{0,2}$, and cross-correlation
$\Delta_{1,1}$.
Higher order moments describe a non-Gaussian
structure of the distribution. For instance, $\Delta_{3,0}$ describes
the skewness in position, which is a measure for how much the maximum
of the distribution differs from mean value $\bar{x}$ (see section \ref{sec:skew}).
The driving scheme introduced above has the potential to induce a
large amount of squeezing and a small amount of skewness in the
probability distribution. Squeezing refers to the reduction of the
variance of one observable, say $\Delta x$, at the expense of
increasing the variance of its conjugate variable $\Delta p$. It is
most often considered in the quantum description of light, and it is
sometimes assumed that the coupling of stretching $\Delta p$ while
squeezing $\Delta x$ is a consequence of Heisenberg's uncertainty
principle. However, squeezing also occurs in classical probability
distributions, where the coupling of stretching and squeezing is a
consequence of the conservation of phase-space volume for conservative
forces because of Liouville's theorem \cite{Liouville1838}.
For a point particle, the dynamical equations take the form
$\dot{p}=F(x)$, but for an object that is described through a
probability distribution, we have
$\dot{\bar{p}} = \langle F(x) \rangle \neq F(\bar{x})$. This implies
that the dynamics of mean position and momentum is generally coupled
to higher-order moments. For instance, $F(x) = c x^3$ would couple
mean momentum $\bar{p}$ to skewness $\Delta_{3,0}$.
For a force that is not given by a power law, the dynamical equation
$\dot{\bar{p}} = \langle F(x) \rangle $, and the equivalent equations
for all moments, will not generate a closed set of equations for mean
values, variances, and skewness (see App.~\ref{sec:FokkerPlanck}
and \ref{sec:basics}).
To overcome this problem, we
make a {\em localization approximation}: we assume that the
probability distribution of the oscillator is so narrow that the force
varies little over the width of the distribution. In this case, the
mean force can be expanded as a Taylor series around the mean
position,
\begin{align}
\langle F(x) \rangle &\approx
\sum_{n=1}^{n_\text{max}}
\frac{ F^{(n)}(\bar{x})}{n!}
\langle (x-\bar{x})^n \rangle
= \sum_{n=1}^{n_\text{max}}
\frac{ F^{(n)}(\bar{x})}{n!} \Delta_{n,0}.
\label{eq:locApproxMainPart}\end{align}
We have used the localization approximation to derive a coupled set of
equations for mean position $\bar{x}$ and momentum $\bar{p}$, as well as all moments $\Delta_{n,m} $ up to
third order, $(n,m) \in \{ (2,0),(1,1),(0,2),(3,0),(2,1),(1,2),(0,3)\} $.
The result is rather lengthy and given by
Eqs.~(\ref{eq:dglX})-(\ref{eq:dD03dt}).
As an example, the dynamical equation for mean momentum
takes the form
\begin{align}
\dot{\bar{p}} &= F
-(\gamma_Q +\gamma )\bar{p}
+ \frac{ \Delta_{2,0}}{2} \left ( F''
- \bar{p} \gamma '' \right )
- \gamma ' \Delta_{1,1}
\nonumber \\ &\hspace{4mm}
+ \frac{ 1}{6} \Delta_{3,0} \left (
F^{(3)} - \bar{p} \gamma^{(3)}
\right )
- \frac{1}{2} \gamma ''\Delta_{2,1} .
\label{eq:pbarDot}\end{align}
In this expression, functions $F$ and $\gamma$ depend on the mean
position, e.g., $F=F(\bar{x})$. We have derived these equations from a
classical stochastic theory (Fokker-Planck equation, see App.~\ref{sec:FokkerPlanck}) and from a
Lindblad-type master equation for open quantum systems (App.~\ref{sec:basics}).
The results differ by genuine quantum terms, which are proportional to $\hbar^2$
and are highlighted in blue in
Eqs.~(\ref{eq:dglX})-(\ref{eq:dD03dt}). In addition, there is a
technical difference between both derivations that only concerns the
position-dependent dissipative surface force. This difference is
discussed in App.~\ref{sec:basics}.
\section{Tip fluctuations: squeezing and skewness}\label{sec:skew}
Suppose we know the moments $\Delta_{n,m}$, either by measuring it or
by solving the dynamical equations. Can we find the probability
distribution that describes the stochastic state of the oscillator?
The answer is yes, and follows from the reconstruction theorem in
quantum physics \cite{StreaterWightman}, for instance. In
App.~\ref{sec:reconstruction}, we derive an approximate
expression for a classical phase-space probability distribution
$\rho(\boldsymbol{r})$ in terms of moments, where bold symbols
$ \boldsymbol{r}=(x,p)$
denote phase-space vectors. A multivariate skew-normal probability distribution
$\rho(\boldsymbol{r})$ can be expressed as the product of a Gaussian
distribution $\rho_0(\boldsymbol{r}) $ and a second function $\Phi$
\cite{Azzalini1996},
\begin{align}
\rho(\boldsymbol{r}) &= \rho_0(\boldsymbol{r}) \Phi(\boldsymbol{r})
\label{eq:rhoReconstructed}\\
\rho_0(\boldsymbol{r}) &= \frac{1}{2\pi |C|^{\frac{ 1}{2}}}
e^{-\frac{ 1}{2}
(\boldsymbol{r}-\bar{\boldsymbol{r}})^T\cdot C^{-1} \cdot (\boldsymbol{r}-\bar{\boldsymbol{r}})
}
\label{eq:GaussianMultiVar} \\
C &= \left ( \begin{array}{cc}
\Delta_{2,0} & \Delta_{1,1}\\ \Delta_{1,1}& \Delta_{0,2}
\end{array} \right ) .
\label{eq:corrMat} \end{align}
The eigenvalues of correlation matrix $C$ are denoted by
$\sigma_1^2$ and $\sigma_2^2$, where $\sigma_1$ corresponds to the
larger and $\sigma_2$ to the smaller phase-space variance. If the
eigenvectors of $C$ are aligned with position and momentum, position
squeezing corresponds to $\Delta x = \sigma_2$ and $\Delta p=\sigma_1>\sigma_2$
\footnote{In all our calculations, we have expressed $x$ in units of
the harmonic oscillator ground state width
$L=\sqrt{\hbar/(M\omega_0)}$ and momentum in units of $\hbar/L$.}.
The Gaussian factor $\rho_0(\boldsymbol{r}) $ would then have an ellipsoid
form that is stretched along the $p$-axis and squeezed along the
$x$-axis. In general, however, squeezing can occur along any direction
in phase space.
If only information about second-order moments is available, the
Gaussian part is all that can be known about $\rho$. Knowledge about
third-order moments enables us to find
the following expansion of $\Phi(\boldsymbol{r})$ around the mean
position $\bar{\boldsymbol{r}} =(\bar{x},\bar{p})$,
\begin{align}
\Phi(\boldsymbol{r}) &= 1 + \boldsymbol{S}\cdot \boldsymbol{R}
+ \frac{ 1}{6} \sum_{i,j,k}\mu_{ijk} R_i R_j R_k
+ {\cal O}(\boldsymbol{R}^4)
\label{eq:PhiResult}\\
S_i &= -\frac{ 1}{2} \sum_{j,k} \mu_{ijk}( C^{-1})_{jk}
\label{eq:shiftVector}\\
\boldsymbol{R} &= C^{-1}\cdot (\boldsymbol{r}-\bar{\boldsymbol{r}}),
\end{align}
where $\mu_{111}=\Delta_{30}$, $\mu_{222}=\Delta_{03}$, and
$\mu_{112}=\mu_{121}=\mu_{211}=\Delta_{21}$, as well as
$\mu_{221}=\mu_{212}=\mu_{122}=\Delta_{12}$.
Vector $\boldsymbol{S}$ corresponds to the shift of the maximum of
$\rho$ relative to the mean position $\bar{\boldsymbol{r}}$, as long as
$|\boldsymbol{S}|$ is much smaller than the variances of $\rho$.
The third-order term (the triple sum) generates a roughly triangular
distortion of the Gaussian profile.
In the following, we will solve the dynamical equations of motion and
use the reconstructed classical probability distribution to
visualize the effect of the surface force on variances.
\section{Results}\label{sec:results}
To analyze the evolution of the probability distribution
$\rho(\boldsymbol{r})$, we have solved the dynamical equations
(\ref{eq:dglX})-(\ref{eq:dD03dt}) in two different ways.
A numerical
solution was found using the software package Mathematica, with
details given in App.~\ref{sec:numDetails}. In
addition, we evaluated the effect of the surface forces using
first-order perturbation theory, with details provided
in App.~\ref{sec:drivenPT}.
The results of both methods indicate that a classical description
is perfectly adequate for typical AFM oscillators. In the discussion below,
we will describe under which circumstances quantum effects may become
relevant, and how one can isolate surface-induced quantum effects from
classical surface-induced effects. Furthermore, both methods predict
that the evolution of mean position and momentum is only very weakly
affected by the coupling to probability moments. For this reason, we
will concentrate on discussing squeezing and skewness.
In our numerical simulations, we have considered a cantilever with
a resonance frequency of 300 kHz and studied the time evolution for
up to 200 cycles. This is a typical duration for many AFM experiments,
and it is sufficiently long to induce strong squeezing. We have
considered several different cases, including single-frequency on- and
off-resonant driving, as well as two-frequency driving with different
sets of frequencies $\omega_i$ and phase factors $\phi_i$ in
driving force (\ref{eq:twoFreqDrivingForce}). Figures \ref{fig:rho102}
and \ref{fig:rho198} show the reconstructed probability distribution
at the time when the tip is close to the sample during cycle 102 and
198, respectively \footnote{The momentum units used in the graphs
may appear unusual. Squeezing is best observed if position and
momentum are measured in units of $L$ and $\hbar/L$,
respectively. In this case, position and momentum uncertainties are
equal, both for the quantized ground state of the oscillator, and
for a thermal classical distribution. We have used these natural
squeezing units in our calculations. For the graphs, we have used
the same rescaling factor for position and momentum to represent
position in pm. As a consequence, momentum is then measured
in $\hbar$ over atto meters.}.
\begin{figure}
\begin{center}
\includegraphics[width=7cm]{rhoReconstructed-102cycles.png}
\caption{\label{fig:rho102} Contour plot of the reconstructed
classical probability distribution after 102 cycles. A contour
labeled with $n$ corresponds to a reduction of the probability density by an
amount $e^{-n}$ compared to the maximum value.}
\end{center}
\end{figure}
For this specific case, we have used
$\omega_1=\omega_0, \omega_2=3\omega_0$, as well as $F_1 = 1.21 $ nN
and $F_2=950$ nN. The second force component needs to be much stronger
to compensate for the fact that it drives the cantilever off-resonantly.
The phases of the force were chosen as $\phi_1=\pi$ rad
and $\phi_2 = 2.67$ rad. This choice of phases is rather different
from the phases presented at the end of
Sec.~\ref{sec:classicalPicture} for reasons we will discuss below.
We have chosen to display the reconstructed probability distribution
after 102 cycles, because at this time triangular distortions of the
Gaussian distribution, i.e., the triple sum in
Eq.~(\ref{eq:PhiResult}), are strongest. Since these distortions grow
with the third power of the distance vector $R_i$ from the mean
position of the tip, their effect is more pronounced far away from the
center of the distribution. However, even on the outermost contour
in Fig.~\ref{fig:rho102}, the deviation from a Gaussian shape only
amounts to about 2\% and is barely visible. Shift vector $\mathbf{S}$
of Eq.~(\ref{eq:shiftVector}), which quantifies skewness,
is negligibly small at this time, about 0.01 pm.
\begin{figure}
\begin{center}
a)\includegraphics[width=7cm]{rhoReconstructed-198cycles.png}
\\
b) \includegraphics[width=7cm]{rhoReconstructed-198cycles-defaultPhaseChoice.png}
\\
c) \includegraphics[width=7cm]{rhoReconstructed-198cycles-reducedDissipation.png}
\caption{\label{fig:rho198} a) Contour plot of the reconstructed
classical probability distribution after 198 cycles, for
$\phi_1=\pi$ rad and $\phi_2 = 2.67$ rad.
b) The same for $\phi_1=0$ rad and $\phi_2 = \frac{ \pi}{2}$.
c) The same as a) but with strongly reduced dissipative surface force.}
\end{center}
\end{figure}
After 198 cycles, strong squeezing can be observed, see
Fig.~\ref{fig:rho198} a). For our choice of parameters, we observe
almost a pure position squeezing (i.e, the ellipse is almost
vertical). Skewness has grown
by a factor of 10 to about 0.15 pm. However, like triangular
distortions, the effect is too small to be visible in our plots.
We found that the amount of squeezing critically depends on the choice
of phase factors for the driving force. For instance,
Fig.~\ref{fig:rho198} b) shows the same situation, but with phase
factors chosen as described in Sec.~\ref{sec:classicalPicture}. The
reason for the strong reduction in squeezing is that the
surface force has a strong influence on the
double-peak structure in Fig.~\ref{fig:drivingScheme}, which is
key to increasing the tip-surface interaction time.
In the presence of
the surface, the double peak tends to become asymmetric, so that the
tip is not in close contact with the surface for a longer time
anymore, see Fig.~\ref{fig:xMean}.
\begin{figure}
\begin{center}
a)\includegraphics[width=7cm]{xMean.png}
\caption{\label{fig:xMean} Evolution of the tip's mean position over
several periods $T=1/f_0$ for a driving force with phase factors
$\phi_1=\pi$ rad and $\phi_2 = 2.67$ rad.}
\end{center}
\end{figure}
The reason for this strong influence is that the relative sign of the
two Fourier components of the approximate square wave in
Fig.~\ref{fig:drivingScheme} matters a lot for its overall shape. In
the figure, the two Fourier components have opposite signs and thus
cancel each other out close to the central peak. However,
if they have the same sign, they add up to a more pronounced single
peak.
In Fig.~\ref{fig:xMean}, one can see that the double-peak structure
varies over time. The reason is that the relative phase of the two
frequency components is affected by the surface force over time. The
double peak
therefore varies between an approximate square wave and a single peak.
By adjusting the phase factors of the two driving force components,
one can mitigate this effect to some extent and thus optimize
squeezing. Another possibility would be to adjust the driving
frequencies slightly, to compensate for the temporal variation of the
double peak.
Finally, we have studied the effect of the dissipative surface force
on squeezing by comparing the result of our full simulation
Fig.~\ref{fig:rho198} a) to a simulation in which the dissipate
surface force is reduced by a factor of $10^{-3}$. The result is shown in
Fig.~\ref{fig:rho198} c) and indicates that $F_\text{dis}$ both
suppresses squeezing and has a strong influence on the axis of the
ellipse. We generally found that, without $F_\text{dis}$, one almost
always obtains momentum squeezing rather than position squeezing
when the tip is close to the sample.
\section{Discussion}\label{sec:discussion}
Our results show several general trends, which are discussed in
this section.
The role of the {\em dissipative surface force} is two-fold: it reduces
the amount of squeezing, and it rotates the squeezing axis so that one
can have (mostly) position squeezing instead of momentum
squeezing.
Since the fluctuation-dissipation theorem
\cite{PhysRev.83.34,Kubo1966} predicts an
increase of fluctuations if a dissipative force acts on a system,
it is natural that squeezing is reduced;
reducing variances simply becomes more difficult.
In our dynamical equations, fluctuations are
described by terms proportional to $p_\text{th}$ in
Eq.~(\ref{dgl:dp2}). A dissipative force can change the orientation of
the squeezing axis through its momentum dependence, which enables it
to counteract features of the conservative surface force that will be
discussed in the next paragraph.
The {\em conservative surface force} alone tends to create momentum
squeezing. We conjecture that the reason for this is that, for most of
the time during one oscillation cycle, the tip is moving in the long-range van der Waals tail of
the surface force. As can be seen in Fig.~\ref{fig:forcePlots}, this
tail has negative curvature, i.e., the force would pull particles that
are closer to the sample stronger towards the surface than particles
that are further away. Consequently, a phase-space probability distribution
would then tend to be stretched along the spatial axis. Since $F_\text{sf}$ is a
conservative force, Liouville's theorem then implies squeezing in
momentum direction.
\begin{figure}
\begin{center}
\includegraphics[width=8cm]{surfaceForcePlots.png}
\caption{\label{fig:forcePlots} Conservative surface force
$F_\text{sf}$ as a function of cantilever position $x$. The blue curve
shows the modified DMT model (\ref{eq:WDMTmodelSmooth}) used in this work.
The orange curve displays the original DMT model (\ref{eq:WDMTmodelOriginal}).}
\end{center}
\end{figure}
Conversely, a conservative force with positive curvature would tend to
squeeze the spatial direction of a probability distribution
and stretch it in momentum direction. For the DMT force shown in Fig.~\ref{fig:forcePlots},
this is the case when the tip is in close contact with the surface
($x<-10 $ nm).
Furthermore, our numerical results suggest that this may
also be the
most effective way to achieve large squeezing.
The reason is that the coupling between
$\Delta_{1,1}$ and $\Delta_{0,2}$
(the first term on the right-hand side of Eq.~(\ref{dgl:dp2})), which
is responsible for the generation of squeezing, is
proportional to the gradient of the force. The gradient
of the surface force is much
larger in the contact region than in the van der Waals
region, so that a larger amount of squeezing can be achieved.
An alternative way to achieve position squeezing is the
interplay between a conservative and a non-conservative
force, such as $F_\text{dis}$.
The details of such an interplay are involved due to the overall
dynamics of the tip during a cycle, which is shown in Fig.~\ref{fig:fullCycle198}.
For a two-frequency driving force, the tip's mean position follows a
non-circular path. In addition, the axes of the ellipse oscillate at twice the
resonance frequency. The latter effect is known from optical squeezing
and follows from our perturbative treatment in App.~\ref{sec:drivenPT}.
\begin{figure}
\begin{center}
\includegraphics[width=8cm]{fullCycle198.png}
\caption{\label{fig:fullCycle198} Time evolution of the reconstructed
probability distribution during the full 198th cycle. Each ellipse
corresponds to the probability distribution at one particular time
during the cycle, and is centered around the mean position of the
tip at that time.
The size of the ellipse corresponds to the mean variances of the distribution in phase
space. For instance, the red ellipse in Fig.~\ref{fig:rho198} a)
corresponds to the leftmost ellipse in this figure.
For better visibility, the size of each ellipse has been
increased by a factor of 5.}
\end{center}
\end{figure}
{\em Quantum effects} are generally negligible in our numerical
simulations. To understand how they could be increased, we
used first-order
perturbation theory to study the influence of a dissipative surface force $F_\text{dis}$ of form
(\ref{eq:PlatzDissip}). We concentrate on the dissipative part because
it generates terms proportional to $\hbar^2$ in the dynamical equations for
variances $\Delta_{2,0}, \Delta_{1,1},\Delta_{0,2}$, whereas the
conservative surface force only introduces quantum terms for
third-order moments. The first-order correction to the position
variance is given by
\begin{align}
\Delta_{20}^{(1)} &= \gamma_0 \int_0^t dt' e^{-\gamma _Q (t-t')}
e^{-\frac{h }{x_{\gamma }}
\left(1+\cos (\omega _0 t'+\phi_1)\right) }
\times
\nonumber \\ &\hspace{4mm}
\Bigg [
\frac{\left(L^4-16 \Delta x_{\text{th}}^4\right)
}{8 \omega_0 x_\gamma^2}
- \frac{ \gamma _Q }{ \omega _0}
\frac{ L^4
}{8 \Delta x_{\text{th}}^2}
\sin \left(2 \omega _0 (t-t')\right)
\nonumber \\ &\hspace{4mm}
-\frac{
\left(L^4 \left(\Delta x_{\text{th}}^2+x_{\gamma }^2\right)
-16\Delta x_{\text{th}}^6\right)
}{8 \Delta x_{\text{th}}^2 x_{\gamma }^2}
\cos \left(2 \omega _0 (t-t')\right)
\Bigg ].
\label{eq:ptResult}\end{align}
This result applies to the case of a single-frequency driving force
($F_2=0$ in Eq.(\ref{eq:twoFreqDrivingForce})). In this case,
the unperturbed mean position (\ref{eq:xUnperturbed}) varies like
$x_0(t) = h \cos (\omega _0 t+\phi_1)$.
In Eq.~(\ref{eq:ptResult}), all quantum
terms are proportional to $L^4$. Compared to terms that are
proportional to thermal fluctuations
$\Delta x_{\text{th}}^4$, quantum terms are suppressed by a factor
$L^4/\Delta x_{\text{th}}^4 =(\hbar \omega_0/(k_B T))^2$.
For current AFM designs and at room temperature, this is only about
$10^{-15}$, but for a 10-fold increased resonance frequency and cooled
to liquid Helium temperatures, this ratio could be increased to $10^{-8}$. While
this is still small, physicists have developed powerful methods to
eliminate the effect of noise from a signal. This includes
spin echoes \cite{PhysRev.80.580},
Doppler-free spectroscopy \cite{PhysRevLett.34.307},
force-insensitivity in atom interferometry \cite{PRA53:312},
and correlation measurements in AFM \cite{Pottier2017}.
Here, we propose a procedure to separate quantum fluctuations
from thermal fluctuations in AFM by adjusting the phase of the driving force.
We start with the observation that, if a term proportional to
$L^4$ appears together with $\Delta x_{\text{th}}^4$ in one algebraic
expression, it is not possible to separate these two terms.
This applies to the first and third term inside the square brackets
in Eq.~(\ref{eq:ptResult}).
Hence, the
only quantum term that can potentially be separated is the term proportional to
$\sin \left(2 \omega _0 (t-t')\right)$.
To isolate this term, we
first note that the first term in square brackets is constant, while the
other two terms are oscillating at frequency $2\omega_0$. Hence, the
first term can be eliminated through spectral analysis.
To eliminate the second classical term, which is
proportional to $\cos \left(2 \omega _0 (t-t')\right)$,
we note that the exponential factor in Eq.~(\ref{eq:ptResult})
suppresses the integrand unless the tip is close to the sample, i.e.,
when $\cos (\omega _0 t'+\phi_1)\approx -1$, or
$\omega _0 t'+\phi_1 \approx \pi (2n+1)$, with $n \in \mathds{Z}$.
This is a direct consequence of $F_\text{dis}$ decreasing
exponentially with the distance from the surface.
Let us now for simplicity
assume that we observe the signal periodically at times $t$ that are multiples of
the AFM period, $\omega_0 t = 2\pi m$ with $m \in \mathds{Z}$.
We then have $\cos \left(2 \omega _0 (t-t')\right)
= \cos \left(2 \omega _0 t'\right) $, or when the tip is close to the surface,
$\cos \left(2 \omega _0 (t-t')\right)\approx
\cos(2\pi(2n+1)-2\phi_1) =\cos(2\phi_1)$.
Hence, if we set $\phi_1 =\frac{ \pi}{4} + n' \frac{ \pi}{2}$, with
$n'\in \mathds{Z}$,
we ensure that $\cos \left(2 \omega _0 (t-t')\right) \approx 0$ when
the tip is close to the surface. As a consequence, the thermal noise
contribution will be strongly suppressed.
The above argument is supported by a numerical evaluation of the
cosine integral as a function of $\phi$ for parameter settings
$Q=400, h=6.7x_\gamma$, and $t=400\pi/\omega_0$. We found that the integral
then varies like $\cos(2\phi_1)$ and indeed vanishes for specific
choices of the driving force's phase. If the signal is monitored
with period $1/f_0$ at other times $t$, the phase $\phi_1$ of the driving force
can be adjusted so that the thermal signal is suppressed as well.
\section{Conclusion}\label{sec:conclusions}
We have proposed a driving scheme to increase the interaction time between
tip and sample in intermodulation AFM.
The tip is described as a stochastic system that can exhibit fluctuations.
We derived a set of coupled dynamical equations for probability
moments, which have some similarity to
the BBGKY hierarchy in statistical mechanics \cite{Cercignani97}.
The solution to these equations enables us to
reconstruct the tip's phase-space probability
distribution.
We use the driving scheme to investigate the generation of
squeezing in the classical phase space of the tip. We predict that
strong position squeezing is possible if the tip is in close contact
with the surface, and if the phases of the two frequency components
in intermodulation AFM are chosen appropriately.
In the weakly interacting van der Waals regime of
the tip-surface interaction, momentum squeezing is predominant. The
dissipative part of the surface force has a strong influence on amount
and orientation of phase-space squeezing.
To distinguish between classical and quantum effects, we
derived the dynamical equations both
using a classical Fokker-Planck equation and
a quantum master equation. We found that
AFM is generally
well described by a classical model. Quantum effects tend to be
much smaller than thermally induced fluctuations, but it is possible
to separate both effects by adjusting the phase of the driving
force in dynamic AFM.
A particularly interesting result is that dissipative surface forces
can introduce quantum dynamics already at the level of variances,
whereas quantum effects induced by a conservative force
are tied to third-order (or higher) probability moments. The derivation
of our dynamical equations in App.~\ref{sec:basics} indicates that the
reason is a particular feature of the dissipative force.
In the model we studied, it is a
position-dependent drag force that depends on two non-commuting
observables: position and momentum.
We conjecture that similar
effect would also occur with other models for dissipative surface
forces. For instance, in the hysteretic JKR model \cite{Johnson1971}, the
surface force changes depending on whether the tip moves towards or
away from the sample; i.e., it depends on position and on
the sign of the tip's
momentum. In a retarded model \cite{Ting1966}, the dissipative
surface force when the tip is at position $x(t)$ depends on an earlier
position $x(t-t_r)$, where $t_r$ is the retardation time of the
force. In Heisenberg picture, $x(t)$ and $x(t-t_r)$ are generally
non-commuting.
By probing the probability distribution of the tip, one can examine
different Fourier components of the surface force, study its
fluctuations,
and enhance specific
effects, similar to methods used in nonlinear optics. Future
work may address the question whether our driving scheme may
also be helpful to measure magnetic surface forces, or if the separation scheme
for quantum effects could be useful to isolate specific classical
effects as well.
\acknowledgements
We would like to thank the Natural Sciences and Engineering Research Council of Canada (NSERC) for financial support.
\newpage
\begin{widetext}
\begin{appendix}
\section{Classical derivation of dynamical equations based on the Fokker-Planck equation}\label{sec:FokkerPlanck}
For a position-independent drag force $-\gamma_Q p$, the Fokker-Planck
equation \cite{Fokker1914,Planck1917}
for a classical
phase-space probability distribution $f(x,p,t)$ can be written as
\begin{align}
\partial_t f &= - \frac{ p}{M} \partial_x f - \partial_p
((F(x,t)-\gamma_Q p)f)
+ \frac{g}{2}\partial_p^2 f,
\end{align}
with $g=2M \gamma_Q k_B T $.
The mean value of a classical observable
$A(x,p,t)$, where we admit an explicit time dependence, is given by
\begin{align}
\langle A \rangle &= \int dx\, dp\, f(x,p,t) \, A(x,p,t) .
\end{align}
As above, we use the notation $ \bar{x}(t) := \langle x \rangle$ and
$ \bar{p}(t) := \langle p \rangle$.
The time derivative of a mean value can be expressed as
\begin{align}
\frac{ d}{dt}\langle A \rangle &=
\langle \partial_t A \rangle
+\frac{ 1}{M} \langle p \partial_x A \rangle
+ \langle (F-\gamma_Q p) \partial_p A \rangle
+ M \gamma_Q k_B T \langle \partial_p^2A \rangle .
\label{eq:dMeanA}\end{align}
To add a position-dependent drag force $-\gamma(x) p$, we start with
the Langevin equation for such a force,
\begin{align}
\dot{x} &= \frac{ p}{M}
\\
\dot{p} &= -\gamma(x) p+ \zeta,
\end{align}
where $\zeta$ is a noise force. We assume $\delta$-correlated noise,
for which $ \langle \zeta(t) \rangle =0$ and
$\langle \zeta(t)\zeta(t') \rangle = g(t) \delta(t-t')$. The unknown
function
$g(t)$ can be found using the
solution for the momentum,
\begin{align}
p(t)&=p(0) e^{-\bar{\Gamma}(t)}
+\int_0^t dt'\,
e^{-\bar{\Gamma}(t)+\bar{\Gamma}(t')} \zeta(t')
\\
\bar{\Gamma}(t) &= \int_0^t dt' \, \gamma(x(t')).
\end{align}
The mean kinetic energy of the particle is given by $E=\langle p^2
\rangle /(2M)$. For large times, for which
$e^{-\bar{\Gamma}(t)} \approx 0$, the
particle should reach its equilibrium energy $E= \frac{ 1}{2} k_B
T$. Hence,
\begin{align}
M k_B T &= \int_0^t dt'\, \int_0^t dt''\,
e^{-2\bar{\Gamma}(t)+\bar{\Gamma}(t') +\bar{\Gamma}(t'')}
\langle \zeta(t') \zeta(t'') \rangle
\\ &=
\int_0^t dt'\,
e^{-2\bar{\Gamma}(t)+2\bar{\Gamma}(t') } g(t')
\\ &=
\int_0^t d\tau\,
e^{-2\bar{\Gamma}(t)+2\bar{\Gamma}(t-\tau) } g(t-\tau).
\end{align}
If the particle does not move too far during the relaxation process,
we can use the approximation $g(t-\tau)\approx g(t)$
and $\bar{\Gamma}(t-\tau)\approx \bar{\Gamma}(t)-\tau
\gamma(x(t))$. We can also extend the integration interval to infinity.
We then obtain
\begin{align}
M k_B T \approx \frac{ g(t)}{2\gamma(x(t))}.
\end{align}
This implies that the factor $g$ takes the same form as for the constant drag
force, with $\gamma_Q$ replaced by $\gamma(x(t))$. This is the
approximation we will use below, but we emphasize that this is
model-dependent. For instance, if we use instead the approximation
$g(t-\tau)\approx g(t)-\tau \dot{g}(t) $, we instead obtain the equation
\begin{align}
M k_B T \approx \frac{ g(t)}{2\gamma(x(t))} - \frac{ \dot{g}(t)}{4\gamma^2(x(t))},
\end{align}
which has the solution
\begin{align}
g(t) &\approx \frac{ p_\text{th}^2}{2}
e^{
2\int_0^t \gamma(x(t')) dt'
}
\left \{
\gamma(0)-2 \int_0^t dt''\,\gamma^2(x(t''))
e^{
-2\int_0^{t''} \gamma(x(t')) dt'
}
\right \}.
\end{align}
Obviously, the precise form of $g(t)$ is difficult to find in a
classical model. However, with the simple approximation, we obtain a
reasonable agreement with the quantum model.
Putting everything
together, we obtain the Fokker-Planck equation
\begin{align}
\frac{ d}{dt}\langle A \rangle &=
\langle \partial_t A \rangle
+\frac{ 1}{M} \langle p \partial_x A \rangle
+ \langle (F-(\gamma_Q+\gamma(x)) p) \partial_p A \rangle
+ M k_B T
\langle (\gamma_Q+\gamma(x)) \partial_p^2A \rangle .
\end{align}
To derive the classical equations of motion for the correlation
functions (\ref{def:DeltanmMainPart}), we consider observables of the form
$A=(x-\bar{x})^n(p-\bar{p})^m$, so that
\begin{align}
\langle \partial_t A \rangle &=-n \dot{\bar{x}} \Delta_{n-1,m} -m
\dot{\bar{p}} \Delta_{n,m-1}
\\ &= -n \frac{ \bar{p}}{M} \Delta_{n-1,m} -m
\left\langle F-(\gamma_Q+\gamma(x))p\right \rangle \Delta_{n,m-1}
\\
\langle p \partial_x A \rangle &= n\left (
\bar{p}\Delta_{n-1,m} +\Delta_{n-1,m+1}
\right )
\\
\langle \partial_p^2A \rangle &= m(m-1) \Delta_{n,m-2}.
\end{align}
We then obtain
\begin{align}
\dot{\Delta}_{n,m} &= \frac{n}{M} \Delta_{n-1,m+1}
- m \gamma_Q \Delta_{n,m}
+m \left \{
\langle F-\gamma(x)p (x-\bar{x})^n(p-\bar{p})^{m-1} \rangle
-\langle F-\gamma(x)p \rangle \Delta_{n,m-1}
\right \}
\nonumber \\ &\hspace{4mm}
+ M k_B T m(m-1)
\left \{ \gamma_Q \Delta_{n,m-2} +
\langle \gamma(x) (x-\bar{x})^n(p-\bar{p})^{m-2} \rangle
\right \}.
\label{eq:FPbeforeLA}\end{align}
To obtain a closed set of coupled dynamical equations,
we assume that
uncertainties in position and momentum in the tip motion will remain small, and the tip remains localized.
If the forces acting on the system vary little across the extension of
the probability distribution, we can expand $F(x)$ and $\gamma(x)$ in a Taylor series
around mean position and momentum,
\begin{align}
F(x) &\approx
\sum_{n=1}^{n_\text{max}} \frac{ F^{(n)}(\bar{x})}{n!}
(x-\bar{x})^n
\label{eq:locApprox}\\
\gamma(x) &\approx
\sum_{n=1}^{n_\text{max}} \frac{ \gamma^{(n)}(\bar{x})}{n!}
(x-\bar{x})^n.
\end{align}
Applying this approximation to Eq.~(\ref{eq:FPbeforeLA}) leads to
\begin{align}
\dot{\Delta}_{n,m} &\approx \frac{n}{M} \Delta_{n-1,m+1}
- m \gamma_Q \Delta_{n,m}
+ M k_B T m(m-1)
\left \{ \gamma_Q \Delta_{n,m-2} +
\sum_{l=0}^{l_\text{max}} \frac{ \gamma^{(l)}(\bar{x})}{l!} \Delta_{n+l,m-2}
\right \}
\nonumber \\ &\hspace{4mm}
+m \sum_{l=0}^{l_\text{max}} \frac{1}{l!} \left \{
\left( F^{(l)}(\bar{x})- \gamma^{(l)}(\bar{x})\bar{p}\right )
(\Delta_{n+l,m-1}-\Delta_{l,0}\Delta_{n,m-1})
- \gamma^{(l)}(\bar{x}) (\Delta_{n+l,m}-\Delta_{l,1}\Delta_{n,m-1})
\right \},
\label{eq:FokkerPlanck}\end{align}
which corresponds to the classical part ($\hbar \rightarrow 0$) of the set of dynamical
equations (\ref{eq:dglX})-(\ref{eq:dD03dt}).
\section{Derivation of dynamical equations from a quantum Lindblad master equation}\label{sec:basics}
To describe the cantilever as a quantum system,
we model it as a 1D quantum harmonic oscillator of mass $M$ that moves in a potential
$V(\hat{x})$ and is subject to noise. The density matrix $\rho$ obeys the
Master equation in Lindblad form \cite{linblad:cmp76},
\begin{align}
\dot{\rho} &= -\frac{ i}{\hbar} [\hat{H},\rho] - \sum_k
\left (
\hat{J}_k^\dagger \hat{J}_k \rho + \rho \hat{J}_k^\dagger
\hat{J}_k
-2 \hat{J}_k \rho \hat{J}_k^\dagger
\right )
\\
\hat{H} &= \frac{ \hat{p}^2}{2M} + V(\hat{x}),
\end{align}
where $ \hat{J}_k$ are jump operators. Force
\begin{align}
F(\hat{x}) =-\nabla V \quad &=
-k \hat{x} + F_\text{dr}(t) + F_\text{sf}(\hat{x})
\label{eq:Fcons}\end{align}
contains all conservative
forces that act on the cantilever. This includes the elastic force $-k x$ of the cantilever
itself, the homogeneous two-frequency driving force $F_\text{dr}(t)$,
and the conservative part $F_\text{sf}$ of the force
exerted by the sample surface.
The time evolution of the expectation value
of an operator $\hat{A}(t)$, for which we admit an explicit time dependence, is given by
\begin{align}
\frac{ d}{dt}\langle A \rangle &=
\langle \partial_t \hat{A} \rangle
-\frac{ i}{\hbar} \langle [\hat{A},\hat{H}] \rangle
- \sum_k
\left \langle
[ \hat{A}, \hat{J}_k^\dagger] \hat{J}_k
+ \hat{J}_k^\dagger [ \hat{J}_k , \hat{A}] \right \rangle .
\label{eq:masterEq}\end{align}
Our goal is to derive the equations of motion for mean
Position $\bar{x}(t) = \langle \hat{x} \rangle $, mean momentum
$\bar{p}(t) = \langle \hat{p} \rangle $, as well as for their
variances $\Delta x$ and $\Delta p$.
We define a family of symmetric
correlation functions $\Delta_{n,m}=\langle \hat{\Delta}_{n,m} \rangle $, with
\begin{align}
\hat{\Delta}_{n,m} &= \frac{ 1}{2}
(\delta\hat{x}^n \delta\hat{p}^m +
\delta\hat{p}^m \delta\hat{x}^n )
\label{def:Deltanm}\\
\delta\hat{x} &= \hat{x}-\bar{x}
\label{eq:defDeltaX}\\
\delta\hat{p} &= \hat{p}-\bar{p} .
\label{eq:defDeltaP}\end{align}
There are two prominent
noise sources: (i) the damping force due to the finite quality factor
$Q$ of the cantilever, and (ii) the dissipative force associated with
the surface interaction.
Damping of the
cantilever can be modeled by adding a force $F_Q=-\gamma_Q p$
with damping constant $\gamma_Q=\frac{ \omega_0}{Q}$ \cite{Platz2013}.
In open quantum systems, such a friction force is described through a
Caldeira-Leggett model \cite{Caldeira1983}
in Lindblad form, see Eq.~(5.117) of
Ref.~\cite{HornbergerDecoherence}, for instance. In this model,
the jump operator takes the form
\begin{align}
\hat{J}_Q &= \frac{ \sqrt{\gamma_Q}}{2} \left ( \frac{ p_\text{th}}{\hbar} \hat{x} +
\frac{ i}{p_\text{th}} \hat{p} \right ),
\end{align}
and the Hamiltonian is
modified by adding a term
\begin{align}
\hat{H}_Q &=
\frac{ \gamma_Q}{4} (\hat{x}\hat{p}+\hat{p}\hat{x}).
\end{align}
Here, $p_\text{th}=2\sqrt{M k_B T}$ corresponds to the momentum
uncertainty in thermal equilibrium. Putting this together, the full
action of the Caldeira-Leggett model can be written in form of a super-operator,
\begin{align}
{\cal L}_Q\hat{A} &= -\frac{i}{\hbar}[\hat{A},\hat{H}_Q]
- [ \hat{A}, \hat{J}_Q^\dagger] \hat{J}_Q
- \hat{J}_Q^\dagger [ \hat{J}_Q , \hat{A}]
\\ &=
-\frac{ \gamma_Q}{4p_\text{th}^2}
\left [\hat{p}, [\hat{p},\hat{A} ] \right ]
-\frac{ \gamma_Q p_\text{th}^2}{4\hbar^2}
\left [\hat{x}, [\hat{x},\hat{A} ] \right ]
- \frac{ i\gamma_Q}{2\hbar} \left (
[\hat{A},\hat{x}]\hat{p} +\hat{p} [\hat{A},\hat{x}]
\right ).
\end{align}
To describe the dissipative part of the surface-tip interaction, we
adopt the model of Ref.~\cite{Platz2013}, where the dissipative part
takes the form of a drag force that depends on the distance from the
sample, $F_{\text{dis}} =- p \gamma(x)$. For most analytical calculations, we
will
keep the position-dependent dissipation rate $ \gamma(x)$ general.
However, for numerical evaluations, we will follow
Ref.~\cite{Platz2013} and use an exponential-decay model,
\begin{align}
\gamma(x) &= \gamma_0\exp \left(
-\frac{ x+h}{x_\gamma} \right ),
\label{eq:PlatzDissip}\end{align}
where $\gamma_0$ is the dissipation rate at the sample surface and
$x_\gamma$ the length scale on which the dissipative force drops off.
In a quantum treatment,
$\dot{x}$ has to be replaced by $\hat{p}/m$. One also has to write the
operator product in a symmetric way to ensure that the force operator
is hermitian. We therefore seek to generate a dissipative force operator
of the form
\begin{align}
\hat{F}_{\text{dis}} &=- \frac{1}{2}\left (
\hat{p} \, \gamma(\hat{x}) + \gamma (\hat{x})\hat{p}
\right ).
\label{eq:Fdis}\end{align}
We remark that this reduces to the previous case for
$\gamma(x)=\gamma_Q $.
We found that it is possible to describe this process through a
Lindblad master equation with
\begin{align}
\hat{J}_\text{dis} &=
\frac{p_\text{th}\sqrt{\tau_0}}{4 \hbar} \Gamma(\hat{x})+
\frac{ i}{p_\text{th}\sqrt{\tau_0}} \hat{p}
\label{eq:Jdis}\\
\hat{H}_\text{dis} &=
\frac{ 1 }{4}
\left ( \hat{p}
\Gamma (\hat{x}) + \Gamma (\hat{x})\hat{p}
\right ),
\end{align}
where $\Gamma(x)$ is the anti-derivative of $\gamma(x)$ and $\tau_0$
a time scale that will be discussed below. Note that $\tau_0$
may be a function of time. We can then again describe the full action
of the dissipative force in terms of a super-operator,
\begin{align}
{\cal L}_\text{dis}\hat{A} &= -\frac{i}{\hbar}[\hat{A},\hat{H}_\text{dis}]
- [ \hat{A}, \hat{J}_\text{dis}^\dagger] \hat{J}_\text{dis}
- \hat{J}_\text{dis}^\dagger [ \hat{J}_\text{dis} , \hat{A}]
\\ &=
-\frac{ 1}{\tau_0 p_\text{th}^2}
\left [\hat{p}, [\hat{p},\hat{A} ] \right ]
-\frac{ \tau_0 p_\text{th}^2}{16\hbar^2}
\left [\hat{\Gamma}, [\hat{\Gamma},\hat{A} ] \right ]
- \frac{ i}{2\hbar} \left (
[\hat{A},\hat{\Gamma}]\hat{p} +\hat{p} [\hat{A},\hat{\Gamma}]
\right ).
\end{align}
We now have to evaluate the equations of motion,
\begin{align}
\frac{ d}{dt}\langle A \rangle &=
\langle \partial_t \hat{A} \rangle
-\frac{ i}{\hbar} \langle [\hat{A},\hat{H}] \rangle
+ \left \langle
{\cal L}_Q\hat{A} +{\cal L}_\text{dis}\hat{A}
\right \rangle ,
\label{eq:masterEq2}\end{align}
for the correlation functions (\ref{def:Deltanm}), which include an
explicit time dependence in their definition through $\bar{x}(t)$
and $\bar{p}(t)$.
We start with mean position and momentum, which obey
\begin{align}
\dot{\bar{x}} &= \frac{ \bar{p}}{M}
\\
\dot{\bar{p}} &= \langle \hat{F}_\text{tot} \rangle -\gamma_Q \bar{p},
\end{align}
where a dot denotes the time derivative $d/dt$, and
$\hat{F}_\text{tot} = F(\hat{x})+ \hat{F}_\text{dis}$.
For the correlation functions, we find
\begin{align}
\dot{\Delta}_{n,m} &=
\frac{ n}{M}
\Delta_{n-1,m+1}
-\frac{i\hbar}{4M}n(n-1)
\left \langle
[ \delta\hat{x}^{n-2}, \delta\hat{p}^{m}] \right \rangle
+ \frac{i}{2\hbar} \big \langle
\delta\hat{x}^{n} [V, \delta\hat{p}^{m}]
+ [V, \delta\hat{p}^{m}] \delta \hat{x}^{n} \big \rangle
-m \langle \hat{F}_\text{tot} \rangle \Delta_{n,m-1}
\nonumber \\ &\hspace{4mm}
-m\gamma_Q \Delta_{n,m}
+ \frac{\gamma_Q p_\text{th}^2}{4} m(m-1)\Delta_{n,m-2}
+\frac{\hbar^2}{4 p_\text{th}^2}
\left (\gamma_Q +\frac{4}{\tau_0}\right) n(n-1) \Delta_{n-2,m}
\nonumber \\ &\hspace{4mm}
+\frac{ i\hbar\gamma_Q}{4}nm
\left\langle [\delta\hat{x}^{n-1},\delta\hat{p}^{m-1}]
\right \rangle
-\frac{ \tau_0 p_\text{th}^2}{16\hbar^2}
\left \langle
\left [\hat{\Gamma}, [\hat{\Gamma}, \hat{\Delta}_{n,m}] \right ]
\right \rangle
+ \frac{ i}{2\hbar} \left \langle
[\hat{\Gamma},\hat{\Delta}_{n,m}]\hat{p} +\hat{p} [\hat{\Gamma},\hat{\Delta}_{n,m}]
\right \rangle .
\end{align}
This result is exact, but its usefulness hinges on our ability to
evaluate commutators of powers of $ \hat{p}$ and functions of
$ \hat{x}$. This is accomplished in appendix
\ref{sec:commutator}. Result~(\ref{eq:VpnCommutator})
states that a commutator of the form
$ [V(\hat{x}), \hat{p}^n ]$ can be written in terms of Euler polynomials
with an argument that contains $\hat{p} $ and the derivative
operator $\partial/\partial \hat{x}$, which acts on $V(\hat{x})$.
In particular, if the function $V(\hat{x})$ is a polynomial, the
commutator corresponds to
\begin{align}
\langle [\delta \hat{x}^m, \delta\hat{p}^n] \rangle &=
i \sum_{l=1}^{\text{min}(m,n)}
C_{n,m,l} \hbar^l
\Delta_{m-l,n-l},
\label{eq:xPowerpPowerCommSimplified}\end{align}
where coefficients $C_{n,m,l}$ are defined through
Eq.~(\ref{eq:xPowerpPowerComm}). Hence, it can be expressed in terms of
correlation functions (\ref{def:Deltanm}).
However, potential $V(\hat{x})$ and factor $\Gamma(\hat{x})$ will
generally not be of a polynomial form, so that the equations of motion will not
form a closed set of equations. It is therefore necessary to make
approximations.
With localization approximation (\ref{eq:locApprox}), the expectation value of all quantities
appearing in the equations of motion can be expressed in terms of
correlation functions (\ref{def:Deltanm}). For instance,
\begin{align}
\langle F(\hat{x}) \rangle &=
\sum_{n=1}^{n_\text{max}} \frac{ F^{(n)}(\bar{x})}{n!}
\Delta_{n,0}
\\
\langle F_\text{dis}(\hat{x}) \rangle &= - \frac{ 1}{2}
\sum_{n=1}^{n_\text{max}} \frac{ \gamma^{(n)}(\bar{x})}{n!}
\left \langle
\hat{p} (\hat{x}-\bar{x})^n+ (\hat{x}-\bar{x})^n\hat{p}
\right \rangle
\\ &= -\sum_{n=1}^{n_\text{max}} \frac{ \gamma^{(n)}(\bar{x})}{n!}
\left (
\Delta_{n,1} + \bar{p}\Delta_{n,0}
\right ).
\end{align}
We note that elastic force $-kx$,
cantilever damping $\gamma_Q$, and the
homogeneous driving force $F_\text{dr}(t)$ are taken into account
exactly. This is because their Taylor series terminates after the
first term, so that the above approximation leaves these forces unaffected.
Using similar expansions up to third order ($n_\text{max}=3$) for all terms in the equation of motions, and
treating terms $\Delta_{n,m}$ as of order $\epsilon^{n+m}$ in the
deviations from mean values, we arrive at the following equations,
\begin{align}
\frac{ d\bar{x}}{dt} &= \frac{ \bar{p}}{M}
\label{eq:dglX}\\
\frac{d\bar{p}}{dt} &: \text{ see Eq.~(\ref{eq:pbarDot})}
\\
\frac{ d\Delta_{2,0}}{dt} &=
\frac{2}{M}\Delta_{1,1} + \frac{{\color{blue} \hbar}^2}{2p_\text{th}^2}
\left (\gamma_Q + \frac{ 4}{\tau_0 }\right )
\label{eq:delxEqn}\\
\frac{ d \Delta_{1,1} }{dt} &=
\frac{1}{M}\Delta_{0,2}
-(\gamma_Q +\gamma(\bar{x}) )\Delta_{1,1}
+ \Delta_{2,0}\left ( F'(\bar{x})
- \bar{p}\, \gamma '(\bar{x}) \right )
- \gamma '(\bar{x}) \Delta_{2,1} +
\frac{ 1}{2}\Delta_{3,0} \left ( F''(\bar{x})
- \bar{p}\, \gamma ''(\bar{x}) \right )
\\
\frac{d \Delta_{0,2} }{dt}&=
2\left ( F'(\bar{x}) -\bar{p} \gamma'(\bar{x})
\right ) \Delta_{1,1}
-2\gamma_Q \left (
\Delta_{0,2} - \frac{ p_\text{th}^2}{4}
\right )
- 2 \gamma (\bar{x}) \Delta_{0,2}
- \frac{ {\color{blue} \hbar}^2}{2} \gamma''(\bar{x})
+ \left ( F''(\bar{x}) -\bar{p} \gamma''(\bar{x})
\right ) \Delta_{2,1}
\nonumber \\ &\hspace{4mm}
- 2 \gamma'(\bar{x}) \Delta_{1,2}
+ \frac{ p_\text{th}^2 \tau_0}{8}
\left (\gamma^2(\bar{x})+
\Delta_{2,0} ( \gamma '(\bar{x})^2 + \gamma (\bar{x}) \gamma
''(\bar{x}) )
+ \left (
\gamma '(\bar{x}) \gamma ''(\bar{x})
+ \frac{ 1}{3} \gamma(\bar{x}) \gamma '''(\bar{x})
\right ) \Delta_{3,0}
\right )
\label{dgl:dp2}\\
\frac{ d\Delta_{3,0} }{dt}&= \frac{3}{M} \Delta_{2,1}
\\
\frac{d \Delta_{2,1} }{dt} &=
\frac{2 \Delta _{1,2}}{M}
-(\gamma_Q+ \gamma(\bar{x}))\Delta_{2,1}
+\Delta _{3,0} \left(F'(\bar{x})
-\bar{p} \gamma '(\bar{x})\right)
\\
\frac{d \Delta_{1,2} }{dt} &=
\frac{\Delta _{0,3}}{M}
-2 (\gamma _Q+\gamma (\bar{x}))\Delta _{1,2}
+ 2 \Delta _{2,1}
\left(F'(\bar{x})-\bar{p} \gamma '(\bar{x})\right)
+ \Delta _{2,0}
\left(\frac{ p_\text{th}^2\tau_0}{4}\gamma' (\bar{x})
\gamma(\bar{x}) - {\color{blue} \hbar}^2 \gamma'''(\bar{x})
\right)
- {\color{blue} \hbar}^2 \gamma '(\bar{x})
\nonumber \\ &\hspace{4mm}
+ \Delta _{3,0}
\frac{ p_\text{th}^2\tau_0}{8}
\left( \gamma(\bar{x}) \gamma''(\bar{x})
+ \gamma'^2(\bar{x})\right)
\\
\frac{ d\Delta_{0,3}}{dt} &=
-3 \Delta _{0,3} (\gamma_Q+\gamma(\bar{x}))
+3 \Delta _{1,2}
\left(F'(\bar{x})-\bar{p} \gamma'(\bar{x})\right)
+\frac{1}{2} {\color{blue} \hbar}^2
\left(F''(\bar{x})-\bar{p} \gamma''(\bar{x})\right)
-2 {\color{blue} \hbar}^2 \Delta _{1,1} \gamma'''(\bar{x})
\nonumber \\ &\hspace{4mm}
+\frac{3}{8} \tau _0 p_{\text{th}}^2
\left (
2 \Delta_{1,1} \gamma (\bar{x}) \gamma'(\bar{x})
+\Delta _{2,1} \left(\gamma '(\bar{x})^2
+\gamma(\bar{x}) \gamma''(\bar{x})\right)
\right ).
\label{eq:dD03dt}\end{align}
These equations represent the main theoretical result of this work. We
have verified that, except for two types of terms, they agree with the corresponding equations for a
classical model based on the Fokker-Planck equation, which is derived
in appendix \ref{sec:FokkerPlanck}. The two types of terms in which
the two models differ are (i) terms that depend on $\hbar$, and (ii)
terms that depend on derivatives of $\gamma(x)$. We will now discuss these differences,
For better identification terms of type (i), we have printed all occurrences of $\hbar$
in blue. These terms correspond to genuine quantum dynamics.
The first occurence in Eq.~(\ref{eq:delxEqn}) corresponds to
the position uncertainty relation $\Delta x \sim \hbar/\Delta p$ if
the momentum uncertainty is equal to the thermal momentum
$p_\text{th}$.
The only other occurence in the equations for the second-order
variances depends on the curvature $\gamma''(\bar{x})$ of the
dissipative part of the surface force. It is interesting to observe
that applying a position-dependent dissipative force may an effective
way to observe differences between classical and quantum dynamics of
a localized system. By
comparison, quantum dynamics induced by the conservative part of the
force only
appear in the
dynamical equation for the momentum skewness $\Delta_{0,3}$.
We now turn to terms of type (ii) and the
role of time scale $\tau_0$. We start by considering the
stationary solution for the case that $F=0=\gamma_Q$. The equations of
motion then have a quasi-stationary solution of the form
$\bar{p}=0$, $\bar{x}$ constant, as well as
\begin{align}
\Delta p &= \frac{ p_\text{th} }{4} \sqrt{\tau_0\gamma(\bar{x})}
\\
\Delta_{1,1} &= \frac{ p_\text{th}^2 \tau_0 }{16m}
\\
\Delta x &= \sqrt{
t \left ( \frac{ p_\text{th}^2 \tau_0 }{8M^2}
+ \frac{ 2\hbar^2}{p_\text{th}^2 \tau_0}
\right )
},
\end{align}
and $\Delta_{n,m}=0$ for the third-order correlation functions.
If we want to ensure that the momentum uncertainty $\Delta p$ is equal to
$ p_\text{th}/2$ in this case, we have to set $\tau_0=4 /\gamma(\bar{x})$.
A second reason why $\tau_0$ should be equal to $4 /\gamma(\bar{x})$
is the comparison with the classical Fokker-Planck equation
(\ref{eq:FokkerPlanck}). In the classical limit ($\hbar=0$) and for
constant $\bar{x}$, the two sets of equations only agree if $\tau_0$ is chosen
in this way.
However, even for $\tau_0=4 /\gamma(\bar{x})$ and in the classical
limit, the Fokker-Planck equation differs from the results above if
$\gamma(\bar{x}(t))$ varies with $\bar{x}(t)$, and these are the terms
of type (ii). In the quantum derivation, these terms are a direct
consequence of model (\ref{eq:Jdis}) for inhomogeneous quantum
dissipative forces. This model appears to be the only one where the
jump operator $\hat{J}_\text{dis}$ is linear in the momentum
operator. Hence, as long as the dissipative force is created by
systems that are Markovian (this is the underlying assumption of the
Lindblad form), the derivative terms are needed for
consistency. Furthermore, as discussed in appendix
\ref{sec:FokkerPlanck}, the absence of derivatives of $\gamma(x)$ in
the classical equations (\ref{eq:FokkerPlanck}) is merely a
consequence of an approximation. A refined Fokker-Planck model would
likely generate similar terms, so that we believe that the presence
of type (ii) terms is physically well justified.
\section{Reconstructing a classical probability distribution from
correlation functions}\label{sec:reconstruction}
We consider an $N$-dimensional phase space $\mathds{R}^N$ with $\boldsymbol{r}$ denoting an
element of this space. In our case, $\boldsymbol{r} = (x,p)$ is the 2D phase
space of a particle with one degree of freedom. A probability
distribution $\rho(\boldsymbol{ r})$ is a real positive function that is
normalized to unity,
\begin{align}
1 &= \int d^N r\, \rho(\boldsymbol{r}).
\end{align}
The $n$th moment around a point $\bar{\boldsymbol{r}} \in \mathds{R}^N$ is defined as
\begin{align}
\mu_{n} &= \int d^N r\, (\boldsymbol{ r}-\bar{\boldsymbol{r}})^n \rho(\boldsymbol{r}).
\end{align}
Here and in the following, we will employ a multi-index notation: for
a tuple $n=(n_1, \cdots , n_N)$ of $k$ integer numbers, we set
\begin{align}
|n| &= n_1+n_2+\cdots +n_N
\\
n! &= n_1! \, n_2! \cdots n_N!
\\
\boldsymbol{r}^{\, n} &= r_1^{n_1} r_2^{n_2}\cdots r_N^{n_k}
\\
D^n f(\boldsymbol{ r}) &=
\frac{ \partial^{|n|}f}{\partial_1^{n_1}\cdots \partial_N^{n_N}}
\end{align}
In this appendix, we answer the following question: Given the moments $\mu_n$ of a
probability distribution $\rho(\boldsymbol{ r})$, can we reconstruct
$\rho$?
To describe asymmetric probability distributions, Azzalini and Valle
\cite{Azzalini1996} introduced a multi-variable
extension of the skew-normal distribution of the form
\begin{align}
\rho(\boldsymbol{r})&= \rho_0(\boldsymbol{r}) \Phi(\boldsymbol{r}),
\end{align}
where $\rho_0$
corresponds to a Gaussian distribution (\ref{eq:GaussianMultiVar}),
with $C$ a positive symmetric $N\times N$ matrix.
In Ref.~\cite{Azzalini1996}, $\Phi$ is a specific
function, but we consider $\Phi$ as an unknown function that needs to be reconstructed using
the moments. We denote the moments of the Gaussian probability distribution
by $\mu_n^{(0)}$ and assume that $\Phi$ possesses a Taylor
expansion around $\bar{\boldsymbol{r}}$, which in multi-index notation takes the form
\begin{align}
\Phi(\boldsymbol{ r}) &= \sum_k \frac{ 1}{k!} D^k\Phi(\bar{\boldsymbol{r}}) \,
(\boldsymbol{r}-\bar{\boldsymbol{r}})^{\, k}.
\end{align}
We then can express the moments around point $\bar{\boldsymbol{r}}$ as
\begin{align}
\mu_n &= \int d^Nr\, (\boldsymbol{ r}-\bar{\boldsymbol{r}})^{\,n} \, \rho_0(\boldsymbol{ r})
\sum_k \frac{ 1}{k!} D^k\Phi(\bar{\boldsymbol{r}}) \, (\boldsymbol{r}-\bar{\boldsymbol{r}})^{\, k}
\\ &=
\sum_k \frac{ 1}{k!} D^k\Phi(\bar{\boldsymbol{r}}) \mu_{n+k}^{(0)}.
\label{eq:reconstr}\end{align}
This corresponds to a set of linear equations that enable us to
express, up to a given order, $D^k\Phi(\bar{\boldsymbol{r}})$ in terms of $\mu_n$.
Of particular interest is the case when $\bar{\boldsymbol{r}}$ corresponds to the
measured mean value. In this situation, moments $\mu_k$ directly
correspond to the correlation function $\Delta_{n,m}$ that we study in
the main part.
For a Gaussian distribution, all odd moments vanish,
$\mu^{(0)}_{2n+1}=0$. The even moments are given by
\begin{align}
\mu^{(0)}_{i_1,i_2,\cdots,i_{2n}} &=
\int d^Nr\, \rho_0(\boldsymbol{r}) (\boldsymbol{ r}-\bar{\boldsymbol{r}})_{i_1}
\cdots (\boldsymbol{ r}-\bar{\boldsymbol{r}})_{i_{2n}}
\\ &=
\frac{ 1}{(2\pi)^{N/2}|C|^{\frac{ 1}{2}}}
\int d^Nx\,
x_{i_1} \cdots x_{i_{2n}}
e^{-\frac{ 1}{2}
\boldsymbol{x}^T\cdot C^{-1} \cdot \boldsymbol{x}}
\\ &=
\frac{1}{2^n n!} \sum_p
C_{i_{p(1})i_{p(2)}} \cdots
C_{i_{p(2n-1)} i_{p(2n)}},
\end{align}
where the sum runs over all permutations $p$ of the indices
$i_1, \cdots i_{2n}$. Explicitly,
moments 2, 4, and 6 are given by
\begin{align}
\mu^{(0)}_{i_1,i_2} &= C _{i_1,i_2}
\\
\mu^{(0)}_{i_1,i_2,i_3,i_4} &=
C _{i_1,i_4} C _{i_2,i_3}+
C_{i_1,i_3} C _{i_2,i_4}
+C _{i_1,i_2} C _{i_3,i_4}
\\
\mu^{(0)}_{i_1,i_2,i_3,i_4,i_5,i_6} &=
C _{i_1,i_6} C _{i_2,i_5} C_{i_3,i_4}
+C _{i_1,i_5} C_{i_2,i_6} C _{i_3,i_4}
+C_{i_1,i_2} C _{i_5,i_6} C_{i_3,i_4}
+C _{i_1,i_6} C_{i_2,i_4} C _{i_3,i_5}
+C _{i_1,i_4} C _{i_2,i_6} C_{i_3,i_5}
\nonumber \\ &\hspace{4mm}
+C _{i_1,i_5} C_{i_2,i_4} C _{i_3,i_6}
+C_{i_1,i_4} C _{i_2,i_5} C_{i_3,i_6}
+C _{i_1,i_6} C_{i_2,i_3} C _{i_4,i_5}
+C_{i_1,i_3} C _{i_2,i_6} C_{i_4,i_5}
+C _{i_1,i_2} C_{i_3,i_6} C _{i_4,i_5}
\nonumber \\ &\hspace{4mm}
+C_{i_1,i_5} C _{i_2,i_3} C_{i_4,i_6}
+C _{i_1,i_3} C_{i_2,i_5} C _{i_4,i_6}
+C_{i_1,i_2} C _{i_3,i_5} C_{i_4,i_6}
+C _{i_1,i_4} C_{i_2,i_3} C _{i_5,i_6}
+C_{i_1,i_3} C _{i_2,i_4} C_{i_5,i_6}.
\label{eq:mu06}\end{align}
We now use these results and
Eq.~(\ref{eq:reconstr}) to find the Taylor coefficients
$D^k\Phi(\bar{\boldsymbol{r}})$ up to order $k=3$. We start by writing down
explicitly the first four moments. $\mu_0=1$ expresses normalization
of probability,
\begin{align}
1 &= \mu_0
\\ &= \Phi(\bar{\boldsymbol{r}}) \mu^{(0)}_0 + \frac{ 1}{(2)!}D^2\Phi(\bar{\boldsymbol{r}})
\mu^{(0)}_2
\label{eq:moment0}\\ &=\Phi(\bar{\boldsymbol{r}}) + \sum_{i_1,i_2}
\frac{ 1}{(i_1 i_2)!} (\partial_{i_1}\partial_{i_2}\Phi(\bar{\boldsymbol{r}})) C_{i_1i_2}.
\end{align}
Here, notation $(i_1 i_2)!$ is a multi-index notation that is equal to
2! if both indices are equal, and 1 otherwise. Below, we will also use
$(i_1 i_2 i_3)!$ equal to 3! for all three indices equal, 2! if only
two are equal, and 1 otherwise.
Since $\bar{\boldsymbol{r}}$ is equal
to the mean position, $\mu_1=0$ by definition. Hence, the second
equation becomes
\begin{align}
0 &= D^1\Phi\, \mu^{(0)}_2 + \frac{ 1}{(3)!}D^3\Phi \, \mu^{(0)}_4 ,
\end{align}
or with explicit indices,
\begin{align}
0 &= \sum_{i_2}(\partial_{i_2}\Phi)\, \mu^{(0)}_{i_1 i_2}
+\sum_{i_2,i_3,i_4} \frac{ 1}{(i_2 i_3 i_4)!}
( \partial_{i_2}\partial_{i_3}\partial_{i_4}\Phi )\,
\mu^{(0)}_{i_1 i_2 i_3 i_4}
\\ &=
\sum_{i_2}(\partial_{i_2}\Phi)\, C_{i_1 i_2}
+\sum_{i_2,i_3,i_4} \frac{ 1}{(i_2 i_3 i_4)!}
( \partial_{i_2}\partial_{i_3}\partial_{i_4}\Phi )\,
(C _{i_1 i_4} C _{i_2 i_3}+
C_{i_1 i_3} C _{i_2 i_4}
+C _{i_1 i_2} C _{i_3 i_4}).
\end{align}
Since the multi-factorial and the derivatives of $\Phi$ are totally
symmetric under exchange of indices $i_2,i_3,i_4$, we can reduce the
second factor to a sum over a single term,
\begin{align}
0 &=
\sum_{i_2}(\partial_{i_2}\Phi)\, C_{i_1 i_2}
+ 3 \sum_{i_2,i_3,i_4} \frac{ 1}{(i_2 i_3 i_4)!}
( \partial_{i_2}\partial_{i_3}\partial_{i_4}\Phi )\,
C _{i_1 i_4} C _{i_2 i_3} .
\end{align}
Multiplying this with $(C^{-1})_{i i_1}$ and summing over $i_1$ yields
\begin{align}
\partial_{i}\Phi(\bar{\boldsymbol{r}}) &= -3
\sum_{i_2,i_3} \frac{ 1}{(i \, i_2 i_3 )!}
( \partial_{i}\partial_{i_2}\partial_{i_3}\Phi )\, C _{i_2 i_3} .
\label{eq:D1phiIntermediate}\end{align}
The second-order moment equals the correlation matrix,
$\mu_{i_1 i_2} = \langle (\boldsymbol{r}-\bar{\boldsymbol{r}})_{i_1}
(\boldsymbol{r}-\bar{\boldsymbol{r}})_{i_2} \rangle $. The equation for this moment reads
\begin{align}
\mu_2 &= \Phi(\bar{\boldsymbol{r}}) \mu^{(0)}_2 + \frac{ 1}{(2)!}D^2\Phi(\bar{\boldsymbol{r}})
\mu^{(0)}_4 .
\label{eq:moment2}\end{align}
Together with Eq.~(\ref{eq:moment0}), this provides us with a linear
set of equations for the $N^2+1$ coefficients $ \Phi(\bar{\boldsymbol{r}})$ and
$D^2\Phi(\bar{\boldsymbol{r}})$. This is easy to solve if we pick matrix $C$ in the
Gaussian distribution (\ref{eq:GaussianMultiVar}) such that
$\mu^{(0)}_2=\mu_2$. This is accomplished for the choice
$C_{i_1 i_2} = \mu_{i_1 i_2}$.
Eqs.~(\ref{eq:moment0}) and (\ref{eq:moment2}) are then easily solved
by $ \Phi(\bar{\boldsymbol{r}})=1$ and $D^2\Phi(\bar{\boldsymbol{r}}) = 0$.
The equation for the third-order moment is given by
\begin{align}
\mu_3 &=
D^1\Phi\, \mu^{(0)}_4 + \frac{ 1}{(3)!}D^3\Phi \, \mu^{(0)}_6,
\end{align}
or with explicit indices,
\begin{align}
\mu_{i_1 i_2 i_3} &=
\sum_{i_4} (\partial_{i_4}\Phi)
\mu^{(0)}_{i_1 i_2 i_3 i_4}
+ \sum_{i_4, i_5, i_6} \frac{1}{(i_4 i_5 i_6)!}
(\partial_{i_4}\partial_{i_5}\partial_{i_6}\Phi)
\mu^{(0)}_{i_1 i_2 i_3 i_4 i_5 i_6}
\end{align}
We can exploit Eq.~(\ref{eq:D1phiIntermediate}) to turn this equation
into one that only contains the third-order derivatives
$(\partial_{i_4}\partial_{i_5}\partial_{i_6}\Phi)$. The triple sum
contains factors like Eq.~(\ref{eq:mu06}), which are very
lengthy. However, the high degree of symmetry of all factors involved
enables us to recude it to
\begin{align}
\mu_{i_1 i_2 i_3} &= 6
\sum_{i_4, i_5, i_6} \frac{1}{(i_4 i_5 i_6)!}
(\partial_{i_4}\partial_{i_5}\partial_{i_6}\Phi)
C_{i_1 i_4}C_{i_2 i_5}C_{i_3 i_6} .
\end{align}
This equation is easily solved and lets us determine all Taylor
coefficients up to order 3,
\begin{align}
\partial_{i}\partial_{j}\partial_{k}\Phi(\bar{\boldsymbol{r}})
&= \frac{ (i\, j\, k)!}{6}
\sum_{i_1, i_2, i_3} \mu_{i_1 i_2 i_3} (C^{-1})_{i\, i_1}
(C^{-1})_{j \, i_2} (C^{-1})_{k\, i_3}
\\
\partial_{i}\Phi(\bar{\boldsymbol{r}})
&= -\frac{ 1}{2} \sum_{i_1, i_2, i_3}
\mu_{i_1 i_2 i_3} (C^{-1})_{i\, i_1}
(C^{-1})_{i_2 \, i_3}.
\end{align}
The full expression for $\Phi(\boldsymbol{r})$ is then given by Eq.~(\ref{eq:PhiResult}).
Turning to the two-dimensional case that is the subject of this study,
we use correlation matrix (\ref{eq:corrMat}), with
\begin{align}
C^{-1} &= \sigma_1^{-2}\sigma_2^{-2} \left ( \begin{array}{cc}
\Delta_{02} & -\Delta_{11}\\ -\Delta_{11}& \Delta_{20}
\end{array} \right ) .
\end{align}
Matrix $C$ has eigenvectors and eigenvalues
\begin{align}
\boldsymbol{e}_{\underline{1}} &= \frac{
1}{\sqrt{2W(W-\Delta_{02}+\Delta_{20})}}
\left ( \begin{array}{c}
\Delta_{20}-\Delta_{02} +W \\ 2\Delta_{11}
\end{array} \right )
, \quad C\cdot \boldsymbol{e}_{\underline{1}} =
\sigma_1^2 \boldsymbol{e}_{\underline{1}} =
\frac{1}{2}(\Delta_{20}+\Delta_{02}+W) \boldsymbol{e}_{\underline{1}}
\\ \boldsymbol{e}_{\underline{2}} &= \frac{
1}{\sqrt{2W(W+\Delta_{02}-\Delta_{20})}}
\left ( \begin{array}{c}
\Delta_{20}-\Delta_{02} -W \\ 2\Delta_{11}
\end{array} \right )
, \quad C\cdot \boldsymbol{e}_{\underline{2}} =
\sigma_2^2 \boldsymbol{e}_{\underline{2}} =
\frac{1}{2}(\Delta_{20}+\Delta_{02}-W)
\boldsymbol{e}_{\underline{2}}
\\
W &= \sqrt{4\Delta_{11}^2 +(\Delta_{20}-\Delta_{02})^2}.
\end{align}
Here, $\sigma_1, \sigma_2$ are the variances along the directions of
the eigenvectors of $C$. The third-order moment has components
$\mu_{111}=\Delta_{30}$, $\mu_{222}=\Delta_{03}$, and
$\mu_{112}=\mu_{121}=\mu_{211}=\Delta_{21}$, as well as
$\mu_{221}=\mu_{212}=\mu_{122}=\Delta_{12}$.
Introducing tensor components
\begin{align}
T^{(1)}_1 &= \Delta_{30}+\Delta_{12}
\\
T^{(1)}_2 &= \Delta_{03}+\Delta_{21}
\\
T^{(3)}_1 &= \frac{1}{3} \Delta_{30}-\Delta_{12}
\\
T^{(3)}_2 &=-\frac{1}{3} \Delta_{03}+\Delta_{21},
\end{align}
we can express function $\Phi$ as
\begin{align}
\Phi(\boldsymbol{r}) &= 1 + \boldsymbol{S}\cdot \boldsymbol{R}
+ \frac{|\boldsymbol{R}|^2}{8} \boldsymbol{T}^{(1)}\cdot
\boldsymbol{R}
+ \frac{1}{8} T^{(3)}_1 (R_1^3-3R_1 R_2^2)
+ \frac{1}{8} T^{(3)}_2 (3R_2 R_1^2-R_2^3).
\end{align}
In polar coordinates,
$R_1=R\cos\phi$ and $R_2=R\sin\phi$, tensors
$\boldsymbol{T}^{(1)}$ and $\boldsymbol{T}^{(3)}$ describe terms that
vary like $\cos\phi$ and $\sin\phi$, or $\cos3\phi$ and $\sin3\phi$,
respectively. Hence, $\boldsymbol{T}^{(1)}$ describes the direction in
which the extended tail of $\rho$ points, while $\boldsymbol{T}^{(3)}$
describes deformations of a triangular shape.
\section{Numerical Simulations} \label{sec:numDetails}
To solve the dynamical equations
(\ref{eq:dglX})-(\ref{eq:dD03dt})
numerically, we
consider a cantilever with a resonance frequency
$f_0=\omega_0/(2\pi) = 300$ kHz, a spring constant of
$M\omega_0^2 = 40.0$ N/m, and a quality factor of $Q=400$.
The cantilever starts from its equilibrium
position and from thermal equilibrium at room temperature (300 K).
To model the surface force, we follow
Platz {\em et al.} \cite{Platz2013} and assume
an exponentially decreasing dissipative force
$F_\text{dis}=-\gamma(x) p$, where $\gamma(x)$ given by
Eq.~(\ref{eq:PlatzDissip}), with
$x_\gamma=1.5$nm and $2\pi\gamma_0/\omega_0 \approx 0.065$.
For the conservative surface force, we employ a modification of the
van der Waals-Derjaguin-Muller-Toropov (DMT) model. In its original form,
the DMT force is given by
\begin{align}
F_\text{DMT}(x) &= \left \{ \begin{array}{cc}
-\frac{H R}{6(a_0+x+h)^2} & x>-h
\\ -\frac{H R}{6a_0^2} + \frac{4}{3} E^* \sqrt{R}
(-h-x)^{\frac{3}{2}}
& x< -h
\end{array} \right . ,
\label{eq:WDMTmodelOriginal}\end{align}
where $H =3.28\times 10^{-17} \text{kg m}^2\text{s}^{-2}$ is the
Hamaker constant, which is a measure for the van der Waals interaction
energy between tip and surface. $R=10$ nm denotes the tip radius, and
$a_0 =2.7\, $nm represents the intermolecular distance.
$E^*= 1.5\,$GPa is the effective stiffness of the tip-sample
system. The piecewise definition of this force makes it unsuitable for
our purposes, since derivatives of Eq.~(\ref{eq:WDMTmodelOriginal})
are not well-defined at $x=-h$. We therefore employ a modified model,
which is continuously differentiable,
\begin{align}
F_\text{sf}(x) &=
-\frac{1.15 H R}{12}\frac{
\left (1+\tanh\left (\frac{ x+h}{L_f} \right ) \right )
}{L_f^2+ (a_0+x+h)^2}
+\frac{ 1}{2} \left (1-\tanh\left (\frac{ x+h}{L_f} \right ) \right )
\left ( -\frac{1.15 H R}{6(L_f^2 +a_0^2)}
+\frac{4}{3} E^* \sqrt{R}
(L_f^2+(h+x)^2)^{\frac{3}{4}}
\right ) .
\label{eq:WDMTmodelSmooth}\end{align}
Here, $L_f=a_0/4$ controls the smoothness of the transition between
van der Waals and surface region. A plot of model force
(\ref{eq:WDMTmodelSmooth}) and the original DMT model
(\ref{eq:WDMTmodelOriginal}) is shown in Fig.~\ref{fig:forcePlots}.
We have performed a series of numerical simulations of the full
dynamical equations (\ref{eq:dglX})-(\ref{eq:dD03dt}) with the
parameters for the surface-tip interaction as given above.
In all
simulations, we have considered several special cases:
\begin{itemize}
\item
{\em Full equations}: The full set of equations
(\ref{eq:dglX})-(\ref{eq:dD03dt}) is simulated
\item
{\em Variance limit}: Only second order correlation functions
$\Delta_{20},\Delta_{11},\Delta_{02}$ are taken into account;
third-order variances ($\Delta_{n,m} $ with $n+m=3$) are set to zero.
\item
{\em Point-particle limit}: all correlation functions $\Delta_{nm}$
are assumed to vanish.
\item
{\em Reduced dissipative force}: To study the influence of the
dissipative force, we have run the simulations in a situation where
the dissipative surface force is reduced by a factor of $10^{-3}$.
\item
{\em No quantum terms}: All quantum terms (the blue terms in
Eqns.~(\ref{eq:dglX})-(\ref{eq:dD03dt})) are set to zero.
\end{itemize}
In addition, we have performed numerical simulations that include
fourth-order correlation functions ($\Delta_{n,m} $ with $n+m=4$) to
verify that these terms can be ignored. These results
were affirmative and are not
presented in this paper.
In agreement with first-order perturbation theory
(see appendix \ref{sec:drivenPT}), we found two general
results in our simulations.
(i) If one is only interested in studying mean position and momentum
of the cantilever, the point-particle limit is appropriate.
Variances only have a small effect on their dynamics.
(ii) For standard AFM parameters, quantum terms can safely be
neglected. Since thermal variances are generally several orders of
magnitude larger than quantum uncertainties, our results do not
support claims in the literature
that AFM is quantum-limited.
The simulation supports the findings that we found in perturbation
theory: significant squeezing is only generated when the tip is in
contact with the sample.
\section{Perturbation theory of the driven cantilever}\label{sec:drivenPT}
If the tip-surface interaction is sufficiently small, the surface
forces can be treated as a perturbation.
To derive a solution of Eqs.~(\ref{eq:dglX}) -
(\ref{eq:dD03dt}) to first order in perturbation theory,
we assume that the
unperturbed system is initially thermalized, i.e., mean position and
momentum follow a stationary trajectory, and the variances correspond
to a thermal equilibrium. The full unperturbed solution for driving
force (\ref{eq:twoFreqDrivingForce}) is then given by
the unperturbed mean position $x_0(t)$ of
Eq.~(\ref{eq:xUnperturbed}),
unperturbed mean momentum $p_0(t)=M \dot{x}_0$, as well as
\begin{align}
\Delta x_\text{th}^2 &:=\Delta_{20}^{\text{(no srfc)}} = \frac{ k_BT}{M\omega_0^2}
\left ( 1 + \left ( 1+ \frac{ 1}{Q^2} \right )
\left ( \frac{\hbar\omega_0}{4 k_B T }
\right )^2 \right )
\\
\Delta_{02}^{\text{(no srfc)}} &= M k_BT \left ( 1 +
\left ( \frac{\hbar\omega_0}{4 k_B T }
\right )^2 \right ),
\end{align}
and $\Delta_{n,m}=0$ else. For a typical cantilever, the ratio of ground state energy $\hbar
\omega_0$ and thermal energy $k_B T$ is in the order of
$10^{-8}$. Here and in the following, we will therefore neglect terms
of order $\hbar^2$ and only keep lowest-order quantum
contributions. We will also neglect terms of order $Q^{-1}$ since the
quality factor is typically in the order of $10^2$.
With this approximation, we find the usual result for
the thermal uncertainty of a classical oscillator,
$ \Delta x_\text{th} = \sqrt{k_BT/M\omega_0^2}$.
To include the effect of surface forces,
we consider the following
dimensionless 9-component vector of first-order corrections,
\begin{align}
\vec{V} &= \left (
\frac{ \Delta_{20}^{(1)} }{L^2},
\frac{ \Delta_{11}^{(1)} }{\hbar} ,
\frac{ \Delta_{02}^{(1)} L^2}{\hbar^2},
\frac{ \Delta_{30}^{(1)} }{L^3},
\frac{ \Delta_{21}^{(1)} }{L \hbar} ,
\frac{ \Delta_{12}^{(1)} L}{\hbar^2},
\frac{ \Delta_{03}^{(1)} L^3}{\hbar^3},
\frac{ x^{(1)} }{L} , \frac{p^{(1)} L}{\hbar}
\right ),
\end{align}
where a superscript $(1)$ indicates a first-order perturbation term
and $L=\sqrt{\hbar/(M\omega_0)}$ is the ground state width.
The perturbative dynamical equations for this vector can be written as
\begin{align}
\partial_t \vec{V} &= M\cdot \vec{V} + \vec{ J},
\label{eq:pt2}\end{align}
with matrix
\begin{align}
M &=
\left(
\begin{array}{ccccccccc}
0 & 2 \omega _0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\
-\omega _0 & -\gamma _Q & \omega _0 & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & -2 \omega _0 & -2 \gamma _Q & 0 & 0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & 0 & 3 \omega _0 & 0 & 0 & 0 & 0 \\
0 & 0 & 0 & -\omega _0 & -\gamma _Q & 2 \omega _0 & 0 & 0 & 0
\\
0 & 0 & 0 & 0 & -2 \omega _0 & -2 \gamma _Q & \omega _0 & 0 & 0
\\
0 & 0 & 0 & 0 & 0 & -3 \omega _0 & -3 \gamma _Q & 0 & 0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \omega _0 \\
0 & 0 & 0 & 0 & 0 & 0 & 0 & -\omega _0 & -\gamma _Q \\
\end{array}
\right),
\end{align}
and inhomogeneity components $J_4=J_5=J_8=0$, and
\begin{align}
J_1 &= -\frac{\hbar \gamma (\bar{x})}{8M\omega_0 \Delta x_{\text{th}}^2}
\\
J_2 &= -\frac{\Delta x_{\text{th}}^2 \left(
F_{\text{sf}}'(x_0) -p_0 \gamma'(x_0)
\right ) }{\hbar }
-\frac{\hbar \left(F_{\text{sf}}'(x_0)
-p_0 \gamma '(x_0)
+M \gamma _Q \gamma (x_0)
\right)}{16 M^2 \omega
_0^2 \Delta x_{\text{th}}^2}
\\
J_3 &=
-\frac{2 M \omega _0 \Delta x_{\text{th}}^4
\left(\gamma '(x_0)^2
+\gamma (x_0) \gamma''(x_0)\right)
}{\hbar \gamma (x_0)}
+ \frac{\hbar \left(\gamma_Q
\left(F_{\text{sf}}'(x_0)-p_0 \gamma'(x_0)\right)
+M \omega _0^2 \gamma (x_0)\right)
}{8 M^2 \omega_0^3 \Delta x_{\text{th}}^2}
-\frac{\hbar \left(\gamma '(x_0)^2
-3 \gamma(x_0) \gamma''(x_0)\right)
}{8 M \omega _0 \gamma (x_0)}
\\
J_6 &=
-\frac{4 \Delta x_{\text{th}}^4 \gamma '(x_0)}{L^3}
+L \Delta x_{\text{th}}^2 \gamma ^{(3)}(x_0)
+\frac{3}{4} L \gamma'(x_0)
\\
J_7 &=
\frac{3 L \gamma_Q \gamma '(x_0)}{4 \omega _0}
-\frac{L \left(F_{\text{sf}}''(x_0)-p_0 \gamma ''(x_0)\right)
}{2 M\omega _0}
\\
J_9 &=
-\frac{L }{\hbar } \left (
F_{\text{sf}}(x_0) -p_0 \gamma (x_0)
+ \Delta x_{\text{th}}^2\left(F_{\text{sf}}''(x_0) - p_0 \gamma''(x_0)\right)
\right ).
\end{align}
It is worthwhile to note that matrix $M$ is block-diagonal and only
couples correlation functions $\Delta_{nm}$ of the same order
$n+m$. Hence, squeezing and other modifications of correlation
functions must be generated through the inhomogeneity $\vec{J}(t)$.
The solution of Eq.~(\ref{eq:pt2}) for $\vec{V}(0)=0$ is given by
\begin{align}
\vec{V}(t) &= \int_0^t dt'\,
e^{M(t-t')}\cdot \vec{J}(t').
\end{align}
This is best evaluated by using the eigenvalues of $M$. This matrix is
not hermitian, but it is not singular. We can therefore express any
vector in the form
\begin{align}
\vec{J}(t') &= \sum_{\alpha=1}^9 \vec{e}_\alpha \tilde{J}_\alpha(t'),
\end{align}
where $\vec{ e}_\alpha$ are the eigenvectors of $M$.
Specifically, the relationship between the original components
$J_i$ and the expansion coefficients $\tilde{J}_\alpha$ is given by
\begin{align}
\\
\tilde{J}_1 &= \frac{J_9 \left(2 \omega _0-i \gamma _Q\right)}{2
\sqrt{2} \omega _0}-\frac{i J_8}{\sqrt{2}}
\\ &=
-\frac{L}{\sqrt{2} \hbar }
\left ( 1- i \frac{ \gamma_Q}{2\omega_0}
\right )
\left ( F_{\text{sf}}(x_0)-p_0 \gamma (x_0)
+ \Delta x_{\text{th}}^2
\left(F_{\text{sf}}''(x_0)-p_0 \gamma ''(x_0)\right)
\right )
\label{eq:Jtilde1} \\
\tilde{J}_3 &= \frac{J_2 \gamma _Q}{\sqrt{2} \omega_0}
+\frac{J_1}{\sqrt{2}}+\frac{J_3}{\sqrt{2}}
\\ &=
-\frac{\sqrt{2} M \omega _0 \Delta x_{\text{th}}^4
\left(\gamma '(x_0)^2+\gamma (x_0)\gamma''(x_0)\right)
}{\hbar \gamma (x_0)}
- \frac{ \gamma _Q
\left ( \Delta x_{\text{th}}^2
-\frac{ L^4}{ 16\Delta x_{\text{th}}^2} \right )
\left( F_{\text{sf}}'(x_0)-p_0 \gamma'(x_0)\right)
}{\sqrt{2} \hbar \omega _0 }
-\frac{\hbar
\left(\gamma '(x_0)^2-3 \gamma (x_0) \gamma'(x_0)\right)
}{8 \sqrt{2} M \omega _0 \gamma (x_0)}
\\
\tilde{J}_4 &=
\frac{\sqrt{3} J_3
\left(\omega _0-i \gamma _Q\right)}{4 \omega
_0}-\frac{\sqrt{3} J_2 \left(\gamma _Q+2 i \omega
_0\right)}{4 \omega _0}-\frac{1}{4} \sqrt{3} J_1
\\
&= \frac{\sqrt{3} \hbar
\left(4 M \omega _0^2 \gamma (x_0)
+\left(3 \gamma _Q+2 i \omega _0\right)
\left(F_{\text{sf}}'(x_0)-p_0 \gamma'(x_0)\right)\right)
}{64 M^2 \omega _0^3 \Delta x_{\text{th}}^2}
+\frac{\sqrt{3}\Delta x_{\text{th}}^2
\left(\gamma_Q+2 i \omega _0\right)
\left(F_{\text{sf}}'(x_0)-p_0 \gamma '(x_0)\right)
}{4 \omega _0 \hbar }
\nonumber \\ &\hspace{4mm}
+\frac{i \sqrt{3} M \Delta x_{\text{th}}^4
\left(\gamma _Q+i \omega _0\right)
\left(\gamma'(x_0)^2+\gamma (x_0) \gamma''(x_0)\right)
}{2 \hbar \gamma (x_0)}
+\frac{i \sqrt{3} \hbar \left(\gamma _Q+i \omega _0\right)
\left(\gamma '(x_0)^2-3\gamma (x_0) \gamma ''(x_0)\right)
}{32 M \omega _0^2\gamma \left(x_0\right)}
\\
\tilde{J}_6 &= \frac{\sqrt{5} J_7
\left(2 \omega _0-i \gamma _Q\right)}{8 \omega_0}
+\frac{\sqrt{5} J_5
\left(2 \omega _0-3 i \gamma_Q\right)}{8 \omega _0}
+\frac{\sqrt{5} J_6 \left(\gamma _Q-i\omega _0\right)}{4 \omega _0}-\frac{1}{4} i \sqrt{5}
J_4
\\ &=
-\frac{\sqrt{5}(\gamma_Q-i\omega_0) }{L^3\omega_0}\Delta x_{\text{th}}^4 \gamma '(x_0)
+\frac{\sqrt{5}(\gamma_Q-i\omega_0)}{4\omega_0} L \Delta x_{\text{th}}^2 \gamma ^{(3)}(x_0)
\nonumber \\ &\hspace{4mm}
+ \frac{\sqrt{5} L }{16 M \omega _0^2}
\left( (i\gamma_Q-2\omega_0)(F_{\text{sf}}''(x_0) - p_0 \gamma ''(x_0))
+3 (2\gamma_q-i\omega_0)M \omega _0 \gamma'(x_0)\right)
\\
\tilde{J}_8 &= \frac{J_7
\left(2 \omega _0-3 i \gamma _Q\right)}{8 \omega _0}-\frac{3
J_6 \left(\gamma _Q+i \omega _0\right)}{4 \omega _0}+\frac{3
i J_5 \left(\gamma _Q+2 i \omega _0\right)}{8 \omega
_0}+\frac{i J_4}{4}
\\ &=
\frac{3 \Delta x_{\text{th}}^4 \left(\gamma _Q+i \omega _0\right) \gamma '(x_0)
}{L^3 \omega _0}
-\frac{3 L \Delta x_{\text{th}}^2\gamma ^{(3)}(x_0)
\left(\gamma _Q+i \omega _0\right)}{4\omega _0}
\nonumber \\ &\hspace{4mm}
-\frac{L }{16 M \omega _0^2}
\left(\left(2 \omega _0-3 i \gamma _Q\right)
\left(F_{\text{sf}}''(x_0)-p_0 \gamma ''(x_0)\right)
+3 M\omega _0 \left(2 \gamma _Q+3 i \omega _0\right)\gamma '(x_0)\right),
\end{align}
and $\tilde{J}_2 = \tilde{J}_1^*$, $\tilde{J}_5 = \tilde{J}_4^*$,
$\tilde{J}_7 = \tilde{J}_6^*$, and $\tilde{J}_9 = \tilde{J}_8^*$.
For $\gamma_Q\ll \omega_0$, the eigenvalues of $M$ are
approximately given by
\begin{align}
\lambda_\alpha \in
\left \{
-\frac{\gamma _Q}{2}-i \omega _0,
-\frac{\gamma _Q}{2}+i\omega _0,
-\gamma _Q,
-\gamma _Q-2 i \omega_0,
-\gamma _Q+2 i \omega _0,
-\frac{3 \gamma _Q}{2}-i \omega_0,
-\frac{3 \gamma _Q}{2}+i \omega _0,
-\frac{3 \gamma _Q}{2}-3 i\omega _0,
-\frac{3 \gamma _Q}{2}+3 i \omega _0
\right \},
\end{align}
which leads to
\begin{align}
\vec{V}(t) &= \sum_{\alpha=1}^9 \vec{e}_\alpha
\int_0^t dt'\, e^{\lambda_\alpha (t-t')} \tilde{J}_\alpha(t').
\label{eq:pertSolution}\end{align}
From this expression, we can draw several conclusions.\\[2mm]
(i) First-order perturbative effects on {\em Mean position and momentum} is
described through terms involving $\tilde{J}_1$ and $\tilde{J}_2$ in
solution (\ref{eq:pertSolution}).
These terms are not affected by quantum effects. They are affected by thermal fluctuations through
terms proportional to $\Delta x_\text{th}$ in Eq.~(\ref{eq:Jtilde1}),
but a numerical estimate shows that this influence is small, roughly
in the order of $10^{-4}$. Therefore, in agreement with numerical
simulations, we conclude that the point-particle approximation is
appropriate if only the position of the tip is measured.
\\[2mm]
(ii) Terms involving $\tilde{J}_6$ to $\tilde{J}_9$ describe the
influence of the surface force and quantum effects on
{\em skewness}. Numerical simulations show that the overall size of skewness
remains small, so that we do not discuss the details of this case.
\\[2mm]
(iii) The third-order expansion presented above is sufficient
to describe squeezing and skewness for up to 300 cycles of the
cantilever. For longer times, fourth-order terms (coupling to
variances $\Delta_{nm}$ with $n+m=4$) can have a strong influence on
squeezing and skewness. We have analyzed the corresponding coupling
numerically, but since it is not relevant for normal AFM time scales,
we do not discuss it here.
\\[2mm]
(iv) {\em Squeezing} of second-order variances
is the most interesting case, since it may be
observable and exhibits the largest contributions due to quantum dynamics.
Squeezing is introduced through terms involving $\tilde{J}_3$ to $\tilde{J}_5$.
To first order in $\gamma_Q$, the corresponding eigenvectors of $M$
are given by
\begin{align}
\vec{ e}_3 &= \left (\frac{1}{\sqrt{2}},-\frac{\gamma _Q}{2 \sqrt{2} \text{$\omega
$0}},\frac{1}{\sqrt{2}},0,0,0,0,0,0\right)
\\
\vec{ e}_4 &= \left (-\frac{1}{\sqrt{3}}-\frac{i \gamma _Q}{\sqrt{3} \omega _0},-\frac{\gamma _Q}{2 \sqrt{3}
\omega _0}+\frac{i}{\sqrt{3}},\frac{1}{\sqrt{3}},0,0,0,0,0,0\right),
\end{align}
and $\vec{e}_5 = \vec{ e}_4^*$. In this expression, the first three
components correspond to position variance, $\Delta_{11}$, and
momentum variance, respectively. For brevity, we will only discuss the
position variance, for which Eq.~(\ref{eq:pertSolution}) yields
\begin{align}
\Delta_{20}^{(1)} &= \int_0^t dt' e^{\gamma _Q (t'-t)} \Bigg [
\frac{\gamma _Q
\left(\frac{ L^4}{16}-\Delta x_{\text{th}}^4\right)
}{2 M \omega _0^2 \Delta x_{\text{th}}^2}
\left(F_\text{sf}'(x_0(t'))-p_0(t') \gamma'\left(x_0(t')\right)\right)
\nonumber \\ &\hspace{4mm}
-
\left(\Delta x_{\text{th}}^4-\frac{ 3 L^4}{16}\right)
\gamma''(x_0(t'))
-\left(\frac{ L^4}{16}+ \Delta x_{\text{th}}^4\right)
\frac{ \gamma '(x_0(t'))^2 }{ \gamma \left(x_0(t')\right)}
\Bigg ]
\nonumber \\ &\hspace{4mm}
+ \int_0^t dt'
\frac{ e^{\gamma _Q (t'-t)} \cos (2 \omega _0(t-t'))
}{\Delta x_{\text{th}}^2}
\Bigg [
\Delta x_{\text{th}}^2
\left(\Delta x_{\text{th}}^4-\frac{ 3 L^4}{16}\right)
\gamma ''(x_0(t'))
- \frac{ L^4 }{8}\gamma(x_0(t'))
\nonumber \\ &\hspace{4mm}
-\frac{ \gamma_Q}{2M \omega _0^2} \left(\frac{ L^4}{16}- \Delta x_{\text{th}}^4\right)
\left(F_\text{sf}'(x_0(t'))-p_0(t') \gamma'(x_0(t'))\right)
+\Delta x_{\text{th}}^2
\left(\frac{ L^4}{16}+ \Delta x_{\text{th}}^4\right)
\frac{ \gamma'(x_0(t'))^2}{\gamma(x_0(t')) }
\Bigg ]
\nonumber \\ &\hspace{4mm}
-\int_0^t dt'
\frac{ e^{\gamma _Q \left(t'-t\right)}
\sin (2\omega _0 (t-t')) }{ M \omega _0 \Delta x_{\text{th}}^2}
\left [
\left(\frac{ L^4}{16}+ \Delta x_{\text{th}}^4\right)
\left(F_\text{sf}'(x_0(t'))-p_0(t')
\gamma'(x_0(t'))\right)
+\frac{ L^4}{8} M \gamma _Q \gamma(x_0(t'))
\right ]
\end{align}
This expression shows that quantum effects are generally very
small. They enter through the ground state width $L$, which, at room
temperature, is about a factor of $10^{-4}$ smaller than the thermal
variance $\Delta x_{\text{th}}$ of the tip position.
To gain a better understanding of quantum effects, we have analyzed
this expression for the special case of a single-frequency driving
force ($F_2=0$ in Eq.~(\ref{eq:twoFreqDrivingForce})) oscillating at
resonance frequency, $\omega_1=\omega_0$. Furthermore, we concentrate on the
effect of a dissipative surface force of the form
(\ref{eq:PlatzDissip}). The integral then reduces to
Eq.~(\ref{eq:ptResult}). The implications of this result are discussed
in the main text.
\section{Commutators of functions of position and momentum}\label{sec:commutator}
We consider functions of operator
$ \hat{x}$ and want to evaluate commutators of
the form
\begin{align}
X_n &= [V(\hat{x}), \hat{p}^n ] .
\end{align}
We note that, since the commutator between $\delta\hat{x}$ and
$\delta\hat{p}$ of Eqs.~(\ref{eq:defDeltaX}) and (\ref{eq:defDeltaP})
is the same as that of $\hat{x}$ and $\hat{p}$, our
results are also valid for commutators of the form
$ [V(\delta\hat{x}), \delta\hat{p}^n ] $.
{\bf Lemma 1: }
\begin{align}
X_n &= i \sum_{l=1}^{n} c_{n,l} \hbar^l R^{(n-l)}(V^{(l)})
\label{eq:XnForm}\\
R^{(m)}(f(\hat{x})) &= \hat{p}^m f(\hat{x})
+f(\hat{x})\hat{p}^m,
\end{align}
with coefficients $c_{n,l}$ that need to be determined.
{\em Proof:} For $n=1$ we have
\begin{align}
X_1 &= i\hbar c_{1,1} R^{(0)}(V^{(1)}),
\end{align}
with $c_{1,1}=\frac{ 1}{2}$. Assuming relation (\ref{eq:XnForm}) holds
for $n-1$, we obtain
\begin{align}
X_n &=
i \left (
\frac{\hbar}{2} R^{(n-1)}(V') +
\frac{ 1}{2} \sum_{r=1}^{n-1} c_{n-1,r} \hbar^r \left ( \hat{p} R^{(n-1-r)}(V^{(r)})
+ R^{(n-1-r)}(V^{(r)})\hat{p} \right )
\right ).
\end{align}
Now,
\begin{align}
\hat{p} \, R^{(m)}(f)
+ R^{(m)}(f)\hat{p} &=
2R^{(m+1)}(f)
+
\sum_{k=1}^{m} c_{m,k} \hbar^{k+1}R^{(m-k)}(f^{(k+1)}),
\end{align}
so that,
\begin{align}
X_n &= i \left (
\frac{\hbar}{2} R^{(n-1)}(V') +
\sum_{l=1}^{n-1} c_{n-1,l} \hbar^l
R^{(n-l)}(V^{(l)})
+\frac{1 }{2}
\sum_{r=1}^{n-2} c_{n-1,r}
\sum_{k=1}^{n-1-r} c_{n-1-r,k}
\hbar^{r+k+1} R^{(n-1-r-k)}(V^{(r+k+1)})
\right ).
\end{align}
Introducing the new summation index $l=r+k+1$, we find
\begin{align}
X_n &= i \left (
\frac{\hbar}{2} R^{(n-1)}(V') +
\sum_{l=1}^{n-1} c_{n-1,l} \hbar^l
R^{(n-l)}(V^{(l)})
+\frac{1}{2}
\sum_{l=3}^{n} \hbar^l R^{(n-l)}(V^{(l)}) \sum_{r=1}^{l-2} c_{n-1,r}
c_{n-1-r,l-r-1}
\right )
\\ &=
i \sum_{l=1}^{n} c_{n,l} \hbar^l
R^{(n-l)}(V^{(l)}),
\end{align}
with
\begin{align}
c_{n,1} &=\frac{ 1}{2} + c_{n-1,1}
\\
c_{n,2} &= c_{n-1,2}
\\
c_{n,l} &= c_{n-1,l} +\frac{1}{2}
\sum_{r=1}^{l-2} c_{n-1,r}
c_{n-1-r,l-r-1}
\quad
\text{ for } 3 \leq l \leq n.
\label{eq:CnRecursion}\end{align}
This completes the proof of lemma 1.
Eq.~(\ref{eq:CnRecursion}) provides us with a recursion relation that
can be used to determine all factors $c_{n,l}$. We have verified that,
up to $n=15$, these factors correspond to coefficients of Euler
polynomials $E_n(x)$. More specifically, we found that
\begin{align}
\sum_{l=1}^n c_{n,l} x^{n-l} &=
i\left (
i^n E_n(-ix) -x^n
\right ).
\end{align}
We can now introduce super-operators defined by
\begin{align}
\overset{\rightarrow}{{\cal P}}^m V(\hat{x}) &=
\left ( \frac{\hat{p}^m}{\hbar
\frac{ \partial }{\partial \hat{x}}}
\right )^m V(\hat{x})
\\
\overset{\leftarrow}{{\cal P}}^m V(\hat{x}) &=
\left ( \frac{1}{\hbar
\frac{ \partial }{\partial \hat{x}}}
\right )^m V(\hat{x}) \hat{p}^m ,
\end{align}
to write the commutation relations in a compact form,
\begin{align}
[V(\hat{x}), \hat{p}^n ] &=
- \left ( \hbar \frac{ \partial }{\partial \hat{x}}
\right )^n \left (
i^n E_n (-i \overset{\rightarrow}{{\cal P}})
- \overset{\rightarrow}{{\cal P}}^n
+ i^n E_n (-i \overset{\leftarrow}{{\cal P}})
- \overset{\leftarrow}{{\cal P}}^n
\right ) V(\hat{x}).
\label{eq:VpnCommutator}\end{align}
It may appear strange that a derivative operator
appears in the denominator of super-operators
$\overset{\leftrightarrow}{{\cal P}}$. However, no negative powers of
derivative operators appear in result (\ref{eq:VpnCommutator}).
We remark that a similar result for quadratic potentials has been
proven by De Angelis and Vignat \cite{DeAngelis2015}.
An important special case is when the function is a power, $V(
\hat{x}) =\hat{x}^m$. It is easy to see that the (mean value of
the) commutator then
reduces to
\begin{align}
\langle [ \delta\hat{x}^m, \delta\hat{p}^n] \rangle &=
2i \sum_{l=1}^{\text{min}(m,n)} \binom{m}{l}
\frac{c_{n,l}\hbar^l}{l!}
\Delta_{m-l,n-l}.
\label{eq:xPowerpPowerComm}\end{align}
\end{appendix}
\end{widetext}
\bibliographystyle{apsrev4-1}
|
\section{INTRODUCTION} \label{sec:I}
A great success of the energy frontier is the discovery of the Higgs particle by ATLAS and CMS collaborations at the Large Hadron Collider (LHC) on 4th July 2012~\cite{ATLAS:2012yve,CMS:2012qbp}. After that discovery, no new particle has been found so far by the experiments at
LHC with $13\operatorname{TeV}$ proton-proton centre of mass energy. This highlights the fact that
not just the energy frontier but also the luminosity (intensity) frontier should be considered as of equal importance
in the search for physics beyond the Standard Model (SM).
For example, one may consider observables mediated by flavour-changing-neutral-currents (FCNCs), which are quite sensitive to new physics, since such FCNC observables are extremely suppressed in the Standard Model
(SM) due to the well-known Glashow-Illiopoulos-Maiani (GIM) mechanism. Another example of a highly suppressed process is provided by
the branching ratio of $\mu \rightarrow e \gamma$ decay mediated by
massive neutrinos at the one-loop level~\cite{Calibbi:2017uvl}:
\begin{equation}
\operatorname{BR}\left( \mu \rightarrow e \gamma \right) \approx 10^{-55}.
\end{equation}
The experimentally known sensitivity for the branching ratio of $\mu \rightarrow e \gamma$ is
\begin{equation}
\operatorname{BR}\left( \mu \rightarrow e \gamma \right)_{\operatorname{EXP}} = 4.2 \times 10^{-13}.
\end{equation}
The large gap between the tiny rates of the flavour violating decays predicted by the SM
and their experimental upper limits has motivated the construction of many flavour models with extended scalar, quark and leptonic spectrum aimed at enhancing those rates by several orders of magnitude up to an observable level within the reach of the sensitivity of the future experiments. A similar situation occurs for other rare FCNC decays such as, for instance $Z\to\mu\tau$ and $t\rightarrow cZ$, which are very suppressed in the SM, \Antonio{but in extensions of the SM,} can acquire sizeable values, within the reach of the future experimental sensitivity.
Although various models with a heavy $Z^\prime$ boson have also
received a lot of attention by the particle physics community as a new source of FCNCs, its properties, being not fully constrained, do not lead to definite predictions. For this reason we shall restrict ourselves to the SM $Z$ couplings in this paper.
In this paper we focus on the SM $Z$ FCNC interactions induced by tree-level gauge boson exchange in a model in which the fermion sector of the SM is enlarged with a fourth vector-like family. An interesting feature of this approach is all coupling constants of $Z$ interactions in this work are fixed by the known values of the SM $Z$ gauge boson interactions, together with mixing parameters.
Our main motivation for adding a fourth vector-like family is to explain quark and lepton mass hierarchies. We first forbid the SM Yukawa couplings with a global
$U(1)'$ symmetry, then allow them to be generated effectively via mixing with the fourth vector-like family, a mechanism somewhat analogous to the seesaw mechanism for neutrino masses.
Consequently, the SM charged fermion masses are inversely proportional to the masses of the heavy vector-like leptons and directly proportional to the product of the couplings of Yukawa interactions that mix SM charged fermions with vector-like fermions. This implies that a small hierarchy in those couplings can yield a quadratically larger hierarchy in the
effective couplings. Combined with a moderate hierarchy in the vector-like masses, this allows us to naturally explain the SM charged fermion mass hierarchy and to predict the mass scale of vector-like fermions.\\~\\
A similar model was discussed in
our previous works~\cite{Hernandez:2021tii}, although with two vector-like families, but the effect on the $Z$ and $W$ boson couplings was not studied.
\Steve{In our previous work~\cite{Hernandez:2021tii}, whose purpose was to explain the muon and electron anomalous magnetic moments simultaneously, the main focus was on the 2HDM scalar sector, and the FCNCs arising from the $Z$ and $W$ boson couplings were not considered,
since the full mass matrices were not accurately diagonalised, and hence such effects were beyond the approximations used there.
By contrast, the main goal of this work is to study the SM $Z$ and $W$ contributions to the FCNC observables at leading order to constrain the masses of vector-like fermions, and to explore other possible phenomenological signatures. The SM $W$ contributions to the CKM mixing matrix with the extended quark sector are also studied for the first time in this work. In order for these effects to be considered reliably, the mass matrices of each fermion sector are accurately diagonalized, both numerically and analytically, unlike the previous work where simple approximations were used which masked the
effects we consider here. The results in this work are sufficiently accurate to enable the contribution of the $Z$ and $W$ boson couplings
to physics beyond the SM to be reliably considered for the first time.}
In order to make the results completely transparent, we shall study the $Z$ and $W$ boson couplings in the presence of only one vector-like family
so that mass matrices of this work can be straightforwardly diagonalized using both analytical and numerical methods. Since only one vector-like fermion family is used, the first generation of SM charged fermions do not acquire masses, which nonetheless is
a very good approximation considering
the SM fermions belonging to the first family
are very light. Consequently, we restrict our attention to
the second and third generations of SM fermions,
as well as to several observables related to FCNC processes involving the second and third SM families.
In our approach, then, the SM is a low effective energy theory arising after
integrating out a single heavy fourth vector-like family.
In order to dynamically generate the hierarchical structure of SM fermion masses, the fermionic mass matrices given in \cite{Hernandez:2021tii} as well as the ones obtained in this work must be accurately and completely diagonalized, which, as mentioned above, has not been done previously.
The mass matrices for the charged lepton and up-quark sectors share the same structure, whereas the one for
the down-type quark sector involves
an additional non-zero element in a particular basis, although we later show that the results are basis independent.
This reasoning does not apply to the neutrino sector, since this sector is treated independently.
This different
feature of the down-type mass matrix, in the preferred basis, allows us to achieve
all mixings among the three generations of SM fermions even though the first one remains massless, and
this leads to a prediction for the Cabbibo-Kobayashi-Maskawa (CKM) mixing matrix. In addition, due to the mixings between the SM quarks and the vector-like quarks, the CKM quark mixing matrix originating from the $W$ couplings
is not unitary, thus implying the need of relaxing the unitarity condition of the CKM mixing matrix, and we also study this feature.
\\~\\
This paper is organized as follows. In Section~\ref{sec:II} we introduce our model to explain the origin of the SM fermion's mass with a fourth vector-like family. In Section~\ref{sec:III} the mass matrices in both quark and lepton sectors are constructed and diagonalized using the mixing formalism. \Antonio{In Section~\ref{sec:IV} the
$Z$ gauge boson interactions with fermions
are determined from the mixing matrices used in the mass matrix diagonalization.
Several FCNC observables for both lepton and quark sectors are analyzed in detail in Sections~\ref{sec:V} and \ref{sec:VI}, respectively.
We state our conclusions in Section~\ref{sec:VII}. Several technical details are relegated to the Appendices. The perturbative analytical diagonalization of the mass matrices for the charged lepton, up type quark and down type quark sectors are discussed in detail in Appendices \ref{app:A}, \ref{app:B} and \ref{app:C}, respectively. The comparison between the numerical and approximate analytic diagonalization of the mass matrices for charged leptons and quarks is made in Appendices \ref{app:D} and \ref{app:E}, respectively.}
\section{AN EXTENDED MODEL
WITH A FOURTH VECTOR-LIKE FAMILY} \label{sec:II}
The origin of the pattern of SM fermion masses is interesting open question, not addressed by the SM. The mass parameters of the SM have been experimentally determined with good precision,
and these experimentally observed mass parameters show a
strong hierarchical structure of the SM fermion masses. The most extreme hierarchy is exhibited between the SM neutrino Yukawa coupling of about $10^{-12}$ and the top quark Yukawa coupling of about $1$. Regarding the tiny neutrino masses, many particle physicists regard their masses as most likely explained by the see-saw mechanism rather than by the Yukawa interactions, thus predicting
the presence of the heavy right-handed neutrinos. The reason why the see-saw mechanism has received a large amount of attention by the particle physics community is that
it provides a dynamical explanation of the tiny active neutrino masses. For a similar reason, it is interesting to speculate about the existence of a dynamical mechanism that produces the masses of all SM fermions via the exchange of heavy fermionic degrees of freedom thus implying that
the SM is an effective low energy theory arising from some spontaneous breaking at higher energy scales
of a more complete underlying theory.
In order to specify a possible candidate of an underlying theory responsible for the generation of the SM fermion mass hierarchy,
we shall consider a minimal extension
of the SM
consistent with the SM current experimental bounds. With this motivation in mind, we enlarge the SM fermion and scalar sectors by including a fourth vector-like family and an extra $SU(2)$ scalar doublet as well as a scalar singlet, respectively. Furthermore, we extend the SM gauge symmetry by adding a $U(1)^\prime$ global symmetry. The particle content of the proposed model is shown in Table \ref{tab:BSM_model}.
\begin{table}[H]
\resizebox{\textwidth}{!}{
\centering\renewcommand{\arraystretch}{1.3}
\begin{tabular}{*{21}{c}}
\toprule
\toprule
Field & $Q_{iL}$ & $u_{iR}$ & $d_{iR}$ & $L_{iL}$ & $e_{iR}$ & $Q_{kL}$ & $%
u_{kR}$ & $d_{kR}$ & $L_{kL}$ & $e_{kR} $ & $\nu_{kR}$ & $\widetilde{Q}_{kR}$
& $\widetilde{u}_{kL}$ & $\widetilde{d}_{kL}$ & $\widetilde{L}_{kR}$ & $%
\widetilde{e}_{kL}$ & $\widetilde{\nu}_{kR}$ & $\phi$ & $H_u$ & $H_d$ \\
\midrule
$SU(3)_C$ & $\mathbf{3}$ & $\mathbf{3}$ & $\mathbf{3}$ & $\mathbf{1}$
& $\mathbf{1}$ & $\mathbf{3}$ & $\mathbf{3}$ & $\mathbf{3}$ & $\mathbf{1}$ &
$\mathbf{1}$ & $\mathbf{1}$ & $\mathbf{3}$ & $\mathbf{3}$ & $\mathbf{3}$ & $%
\mathbf{1}$ & $\mathbf{1}$ & $\mathbf{1}$ & $\mathbf{1}$ & $\mathbf{1}$ & $%
\mathbf{1}$ \\
$SU(2)_L$ & $\mathbf{2}$ & $\mathbf{1}$ & $\mathbf{1}$ & $\mathbf{2}$
& $\mathbf{1}$ & $\mathbf{2}$ & $\mathbf{1}$ & $\mathbf{1}$ & $\mathbf{2}$ &
$\mathbf{1}$ & $\mathbf{1}$ & $\mathbf{2}$ & $\mathbf{1}$ & $\mathbf{1}$ & $%
\mathbf{2}$ & $\mathbf{1}$ & $\mathbf{1}$ & $\mathbf{1}$ & $\mathbf{2}$ & $%
\mathbf{2}$ \\
$U(1)_Y$ & $\frac{1}{6}$ & $\frac{2}{3}$ & $-\frac{1}{3}$ & $-\frac{1%
}{2}$ & $1$ & $\frac{1}{6}$ & $\frac{2}{3}$ & $-\frac{1}{3}$ & $-\frac{1}{2}$
& $-1$ & $0$ & $\frac{1}{6}$ & $\frac{2}{3}$ & $-\frac{1}{3}$ & $-\frac{1}{2}
$ & $-1$ & $0$ & $0$ & $\frac{1}{2}$ & $-\frac{1}{2}$ \\
$U(1)^\prime$ & $0$ & $0$ & $0$ & $0$ & $0$ & $1$ & $-1$ & $-1$ & $1$
& $-1$ & $-1$ & $1$ & $-1$ & $-1$ & $1$ & $-1$ & $-1$ & $1$ & $-1$ & $-1$ \\
\bottomrule
\bottomrule
\end{tabular}}%
\caption{Particle assigments under the $SU(3)_C\times SU(2)_L\times U(1)_Y\times U(1)^\prime$ symmetry of the
extended 2HDM theory
with fourth vector-like family.
The index $i=1,2,3$ denotes the
the $i$th SM fermion generation and $k=4$ stands for the
fourth vector-like family. }
\label{tab:BSM_model}
\end{table}
Our proposed theory is a minimal extended 2 Higgs Doublet Model (2HDM) where the SM fermion sector is enlarged by the inclusion of a fourth vector-like family and the scalar sector is augmented by an extra $SU(2)$ scalar doublet
and a singlet flavon and lastly the SM gauge symmetry is extended by the $U(1)^\prime$ global symmetry. As this model features the global $U(1)^\prime$ symmetry, there is no a
neutral $Z^\prime$ gauge boson in the particle spectrum.
Furthermore,
the up-type quarks feature Yukawa interactions with the
up-type SM Higgs $H_u$, whereas the extra scalar doublet $H_d$ couples with the SM down-type quarks and charged leptons.
Our proposed model is especially motivated by the hierarchical structure of the SM and, in order to implement
this hierarchy,
we forbid the SM-type Yukawa interactions
by appropiate $U(1)^\prime$ charge assignments of the scalar and fermionic fields. Then, for the above specified particle content, the following effective Yukawa interactions arise:
\begin{equation}
\mathcal{L}_{\operatorname{eff}}^{\operatorname{Yukawa}} =y_{ik}^\psi (M_{\psi^\prime}^{-1})_{kl} {x_{lj}^{\psi^\prime}
\left\langle \phi \right\rangle} \overline{%
\psi}_{iL} \widetilde{H} \psi_{jR} + {x_{ik}^{\psi^\prime} \left\langle \phi
\right\rangle}(M_{\psi}^{-1} )_{kl} y_{lj}^\psi \overline{\psi}_{iL} \widetilde{H}
\psi_{jR} + \operatorname{h.c.} \label{eqn:the_effective_Yukawa_Lagrangian}
\end{equation}
where the indices $i,j=1,2,3$ and $k,l=4$ whereas $\psi,\psi^\prime = Q, u, d, L, e$ and $M$ means heavy vector-like mass. The masses of all SM fermions can be explained by this effective Lagrangian of Equation~\ref{eqn:the_effective_Yukawa_Lagrangian}, emphasizing their relative different masses are explained by the factor $\langle \phi \rangle/M \ll 1 \RIfM@\expandafter\text@\else\expandafter\mbox\fi{ ( apart from top quark )}$, except for the neutrinos which requires an independent approach to their mass. Feynman diagrams corresponding for the effective Lagrangian are shown in Figure \ref{fig:mass_insertion_diagrams}:
\begin{figure}[H]
\centering
\begin{subfigure}{0.48\textwidth}
\includegraphics[width=1.0\textwidth]{Effective_Yukawa_Interactions1}
\end{subfigure} \hspace{0.1cm}
\begin{subfigure}{0.48\textwidth}
\includegraphics[width=1.0\textwidth]{Effective_Yukawa_Interactions2}
\end{subfigure}
\caption{Feynman diagrams leading to the effective Yukawa
interactions, where $\protect\psi,\protect\psi^\prime = Q,u,d,L,e$ (neutrinos
will be treated separately), $i,j=1,2,3$, $k,l=4$, $M_{lk}$ is vector-like
mass and $\widetilde{H} = i\protect\sigma_2 H^*, H = H_{u,d}$}
\label{fig:mass_insertion_diagrams}
\end{figure}
The theory considered in this paper corresponds to the one given in
one of our previous works~\cite{Hernandez:2021tii}, however one vector-like family is used instead of two so that the mass matrices for both quark and lepton sectors can be diagonalized much more economically than in our previous model of \cite{Hernandez:2021tii} at cost of having massless
the first generation SM charged fermions (One of our main purposes is to diagonalize mass matrices for the quark and lepton sectors without any assumptions) and this is actually a good approximation taking into account that the first generation of SM charged fermions are very light.
\subsection{Effective Yukawa interactions for the SM fermions} \label{sec:II_1}
The renormalizable interactions of the quark sector in this model are given by:
\begin{equation}
\begin{split}
\mathcal{L}_{q}^{\operatorname{Yukawa+Mass}} &= y_{ik}^{u} \overline{Q}_{iL}
\widetilde{H}_u u_{kR} + x_{ki}^{u} \phi \overline{\widetilde{u}}_{kL}
u_{iR} + x_{ik}^Q \phi \overline{Q}_{iL} \widetilde{Q}_{kR} + y_{ki}^u
\overline{Q}_{kL} \widetilde{H}_u u_{iR} \\
&+ y_{ik}^{d} \overline{Q}_{iL} \widetilde{H}_d d_{kR} + x_{ki}^{d} \phi
\overline{\widetilde{d}}_{kL} d_{iR} + y_{ki}^d \overline{Q}_{kL} \widetilde{%
H}_d d_{iR} \\
&+ M_{kl}^{u} \overline{\widetilde{u}}_{lL} u_{kR} + M_{kl}^{d} \overline{%
\widetilde{d}}_{lL} d_{kR} + M_{kl}^Q \overline{Q}_{kL} \widetilde{Q}_{lR} +
\operatorname{h.c.} \label{eqn:general_Quark_Yukawa__Mass_Lagrangian}
\end{split}%
\end{equation}
where $i,j=1,2,3$, $k,l=4$ and $\widetilde{H}=i\sigma _{2}H^{\ast }$. After the $U(1)^\prime$ symmetry is spontaneously broken by the vacuum expectation value (vev) of the
the singlet flavon $\phi$, and the heavy vector-like fermions are integrated out, the renormalizable Yukawa terms at higher energy scale give rise to
the effective Yukawa interactions which explain the current SM fermion mass hierarchy. The Feynman diagrams corresponding to the effective Yukawa interactions of the quark sector are shown in Figure~\ref{fig:diagrams_quark_mass_insertion}:
\begin{figure}[H]
\centering
\begin{subfigure}{0.48\textwidth}
\includegraphics[width=1.0\textwidth]{EYI_up_quarks_1}
\end{subfigure} \hspace{0.1cm}
\begin{subfigure}{0.48\textwidth}
\includegraphics[width=1.0\textwidth]{EYI_up_quarks_2}
\end{subfigure}
\begin{subfigure}{0.48\textwidth}
\includegraphics[width=1.0\textwidth]{EYI_down_quarks_1}
\end{subfigure} \hspace{0.1cm}
\begin{subfigure}{0.48\textwidth}
\includegraphics[width=1.0\textwidth]{EYI_down_quarks_2}
\end{subfigure}
\caption{Feynman diagrams contributing to the up and down type quark's effective Yukawa
interactions in the mass insertion formalism. Here $i,j=1,2,3$ and $k,l=4$ and $M_{lk}$ is vector-like mass.}
\label{fig:diagrams_quark_mass_insertion}
\end{figure}
The same approach can be applied to the SM charged lepton sector and the renormalizable charged lepton Yukawa interactions are given by:
\begin{equation}
\begin{split}
\mathcal{L}_{e}^{\operatorname{Yukawa+Mass}} &= y_{ik}^{e} \overline{L}_{iL}
\widetilde{H}_{d} e_{kR} + x_{ki}^{e} \phi \overline{\widetilde{e}}_{kL}
e_{iR} + x_{ik}^L \phi \overline{L}_{iL} \widetilde{L}_{kR} + y_{ki}^e
\overline{L}_{kL} \widetilde{H}_{d} e_{iR}
\\
&+ M_{kl}^{e} \overline{\widetilde{%
e}}_{lL} e_{kR} + M_{kl}^L \overline{L}_{kL} \widetilde{L}_{lR} + \operatorname{h.c.},
\end{split}
\label{eqn:general_charged_lepton_Yukawa_Mass_Lagrangian}
\end{equation}
and its following effective Yukawa interactions read off in Figure~\ref{fig:diagrams_charged_leptons_mass_insertion}.
\begin{figure}[H]
\centering
\begin{subfigure}{0.48\textwidth}
\includegraphics[width=1.0\textwidth]{EYI_charged_leptons_1}
\end{subfigure} \hspace{0.1cm}
\begin{subfigure}{0.48\textwidth}
\includegraphics[width=1.0\textwidth]{EYI_charged_leptons_2}
\end{subfigure}
\caption{Feynman diagrams contributing to the charged lepton's effective Yukawa interactions in the mass insertion formalism. Here $i,j=1,2,3$ and $k,l=4$ and $M_{lk}$ is vector-like mass.}
\label{fig:diagrams_charged_leptons_mass_insertion}
\end{figure}
It is possible to generate the masses of all SM charged fermions by the same method relying on
effective Yukawa interactions. However, this is not the case for the SM light active neutrinos as they need to be independently treated since the simplest mechanism responsible for generating their tiny masses requires the inclusion of
Majorana particles in the leptonic spectrum. In order to make the SM neutrinos massive, we made use of
two important assumptions, one of which is that the SM neutrinos are Majorana particles and the other is they get masses via the type 1b seesaw mechanism~\cite{Hernandez:2021tii,Hernandez-Garcia:2019uof} mediated by the heavy vector-like neutrinos without considering the right-handed neutrinos $\nu_{iR}$. The renormalizable Yukawa interactions for the neutrino sector are given by:
\begin{equation}
\begin{split}
\mathcal{L}_{\nu}^{\operatorname{Yukawa+Mass}} = y_{ik}^{\nu} \overline{L}_{iL}
\widetilde{H}_u \nu_{kR} + x_{ik}^L \overline{L}_{iL} H_d \overline{%
\widetilde{\nu}}_{kR} + M_{kl}^{M} \overline{\widetilde{\nu}}_{lR} \nu_{kR}
+ \operatorname{h.c.}
\end{split}
\label{eqn:general_neutrinos_Yukawa_Mass_Lagrangian}
\end{equation}
It is worth mentioning that
the nature of the vector-like mass appearing in Equation~\ref{eqn:general_neutrinos_Yukawa_Mass_Lagrangian} is that the vector-like mass is different than the Majorana mass since the particles involved in vector-like mass terms
are different, whereas the ones appearing in a Majorana mass terms does not. However, they share
the common feature that both break the lepton number, which is confirmed
by checking each lepton number of $\nu_{kR}$ and $\widetilde{\nu}_{kR}$ in the two Yukawa interactions of Equation~\ref{eqn:general_neutrinos_Yukawa_Mass_Lagrangian}. We call this mechanism ``type 1b seesaw mechanism" and it can allow a different Yukawa interaction at each vertex as seen in Figure~\ref{fig:diagrams_neutrinos_mass_insertion}.
\begin{figure}[H]
\centering
\begin{subfigure}{0.48\textwidth}
\includegraphics[width=1.0\textwidth]{Type_1b_seesaw_mechanism}
\end{subfigure}
\caption{Type Ib seesaw diagram~\cite{Hernandez:2021tii,Hernandez-Garcia:2019uof} which leads to the effective Yukawa
interactions for the Majorana neutrinos in mass insertion formalism,
where $i,j=1,2,3$ and $k,l=4$ and $M_{lk}$ is vector-like mass.}
\label{fig:diagrams_neutrinos_mass_insertion}
\end{figure}
Allowing a different Yukawa interaction at each vertex of Figure~\ref{fig:diagrams_neutrinos_mass_insertion} means that one of the Yukawa interactions can have a very suppressed coupling constant, which can lower the expected order of magnitude of the right-handed Majorana neutrinos masses of the usual type I seesaw mechanism from $10^{14}\operatorname{GeV}$ up to the $\operatorname{TeV}$ scale. The most relevant
features of the
model considered in this paper are:
\begin{enumerate}
\item
It allows a dynamical and natural explanation of the origin of the observed SM fermion mass hierarchy
\item The model under consideration
is economical in the sense that it includes a common mechanism for generating the masses of the SM charged fermions via
effective Yukawa interactions resulting after integrating out the heavy vector-like fermions.
\item The expected right-handed neutrinos can have a much smaller mass compared to the ones mediating the usual type I seesaw mechanism, thus allowing to test our model at colliders as well as via charged lepton flavor violating processes.
\end{enumerate}
Now that we have discussed how the SM fermions get massive via the effective Yukawa interactions, the next task is to construct their mass matrices in the flavor basis and then to diagonalize those and this will be discussed in the next section.
\section{EFFECTIVE YUKAWA MATRICES USING A MIXING FORMALISM}
\label{sec:III}
The effective Yukawa interactions discussed in section~\ref{sec:II} give rise to the following
mass matrix for fermions written in the flavor basis:
\begin{equation}
M^{\psi }=\left(
\begin{array}{c|ccccc}
& \psi _{1R} & \psi _{2R} & \psi _{3R} & \psi _{4R} & \widetilde{\psi }_{4R} \\[0.5ex] \hline
\overline{\psi }_{1L} & 0 & 0 & 0 & y_{14}^{\psi }\langle \widetilde{H}%
^{0}\rangle & x_{14}^{\psi }\langle \phi \rangle \\[1ex]
\overline{\psi }_{2L} & 0 & 0 & 0 & y_{24}^{\psi }\langle \widetilde{H}%
^{0}\rangle & x_{24}^{\psi }\langle \phi \rangle \\[1ex]
\overline{\psi }_{3L} & 0 & 0 & 0 & y_{34}^{\psi }\langle \widetilde{H}%
^{0}\rangle & x_{34}^{\psi }\langle \phi \rangle \\[1ex]
\overline{\psi }_{4L} & y_{41}^{\psi }\langle \widetilde{H}%
^{0}\rangle & y_{42}^{\psi }\langle \widetilde{H}^{0}\rangle
& y_{43}^{\psi }\langle \widetilde{H}^{0}\rangle & 0 &
M_{44}^{\psi } \\[1ex]
\overline{\widetilde{\psi }}_{4L} & x_{41}^{\psi ^{\prime }}\langle
\phi \rangle & x_{42}^{\psi ^{\prime }}\langle \phi \rangle
& x_{43}^{\psi ^{\prime }}\langle \phi \rangle & M_{44}^{\psi
^{\prime }} & 0 \\
\end{array}%
\right) ,
\label{eqn:general_55_mass_matrix}
\end{equation}
where $\psi,\psi^{\prime} = Q,u,d,L,e$ and the zeros in the upper-left $3 \times 3$ block mean that the SM fermions
acquire masses only via their mixing
with the fourth vector-like family. The other zeros appearing in the diagonal positions are forbidden by the $U(1)^{\prime}$ charge conservation. This mass matrix was obtained for the first time in
\cite{King:2018fcg} and it can reveal the hierarchical structure of the SM since this mass matrix involves three different mass scales $\langle H^0 \rangle, \langle \phi \rangle$ and $M$. In order to dynamically reproduce the hierarchical structure of the SM fermion masses,
we need to maximally rotate this mass matrix and the resulting maximally rotated mass matrix should be a starting point for our analysis in this work. For the fully rotated mass matrix, the up-quark and charged lepton sectors share the same structure, whereas the down-type quark mass matrix has an
additional element since one of the quark doublet rotations was already used in the up quark sector, as it will be shown below.
\Huchan{\Antonio{Regarding} the diagonalization for each fermion sector, we will employ two methods for comparison; one of which is the numerical SVD diagonalization and the \Antonio{other is the analytical perturbative
step-by-step diagonalization}. We will make use of the numerical SVD diagonalization for the \Antonio{exact} diagonalization as well as for our numerical scans in main body of this work, however it is important to look at the analytical approximated step-by-step diagonalization since
\Antonio{it provides an analytical understanding
on how the SM $Z$ gauge boson} can induce the flavor violating interactions at tree-level and this analytic diagonalization will be covered in Appendix~\ref{app:A} to \ref{app:C}. Lastly, we have found that the analytic diagoanlization for each fermion sector is quite close to its numerical result with very small differences and this feature will be discussed in detail \Antonio{in} Appendix~\ref{app:D} to \ref{app:E}.}
\subsection{Diagonalizing the charged lepton mass matrix} \label{sec:III_1}
After all scalars of our proposed model acquire
their vevs ($v_{d} = \langle H_d^0 \rangle$ and $v_{\phi} = \langle \phi \rangle$) from Equation~\ref{eqn:general_55_mass_matrix}, we otain the following
fully rotated mass matrix for the charged lepton sector
\begin{equation}
M^{e }=\left(
\begin{array}{c|ccccc}
& e _{1R} & e _{2R} & e _{3R} & e _{4R} & \widetilde{L }_{4R} \\[0.5ex] \hline
\overline{L }_{1L} & 0 & 0 & 0 & 0 & 0 \\[1ex]
\overline{L }_{2L} & 0 & 0 & 0 & y_{24}^{e } v_{d} & 0 \\[1ex]
\overline{L }_{3L} & 0 & 0 & 0 & y_{34}^{e } v_{d} & x_{34}^{L } v_{\phi} \\[1ex]
\overline{L }_{4L} & 0 & 0 & y_{43}^{e } v_{d} & 0 & M_{44}^{L } \\[1ex]
\overline{\widetilde{e }}_{4L} & 0 & x_{42}^{e} v_{\phi}
& x_{43}^{e} v_{\phi} & M_{44}^{e} & 0 \\
\end{array}%
\right),
\label{eqn:cl_1}
\end{equation}
where we use this fully rotated basis as a starting point in order to easily explain the observed SM fermion mass hierarchy.
This rotated basis is exactly consistent with the one given in~\cite{King:2018fcg} and we need to explain how the mass matrix of Equation~\ref{eqn:cl_1} is fully rotated. First of all, we rotate the left-handed leptonic fields $L_{1L}$ and $L_{3L}$ to turn off the entry $x_{14}^{L} v_{\phi}$ and then rotate $L_{2L}$ and $L_{3L}$ to trun off the next $x_{24}^{L} v_{\phi}$ entry. Next, we can rotate again the leptonic fields $L_{1L}$ and $L_{2L}$ to turn off the $y_{14}^{e} v_d$ entry. These rotations can be applied to the right-handed leptonic fields $e_{1,2,3R}$ in order to make the zeros appearing in the lower-left $2 \times 3$ block. This fully rotated mass matrix of Equation~\ref{eqn:cl_1} is our starting point to implement both the hierarchical structure of the SM fermion masses and to analyze the flavor violating interactions mediated by the SM $Z$ gauge boson. \Huchan{Before diagonalizing the mass matrix of Equation~\ref{eqn:cl_1}, it is convenient to rearrange the mass matrix by switching the Yukawa terms by mass parameters and then by swapping the fourth and fifth column in order to make the heavy vector-like masses locate in the diagonal positions as given in Equation~\ref{eqn:cl_2}}.
\begin{equation}
M^{e }
=
\left(
\begin{array}{c|ccccc}
& e _{1R} & e _{2R} & e _{3R} & e _{4R} & \widetilde{L }_{4R} \\[0.5ex] \hline
\overline{L }_{1L} & 0 & 0 & 0 & 0 & 0 \\[1ex]
\overline{L }_{2L} & 0 & 0 & 0 & m_{24} & 0 \\[1ex]
\overline{L }_{3L} & 0 & 0 & 0 & m_{34} & m_{35} \\[1ex]
\overline{L }_{4L} & 0 & 0 & m_{43} & 0 & M_{45}^{L } \\[1ex]
\overline{\widetilde{e }}_{4L} & 0 & m_{52} & m_{53} & M_{54}^{e} & 0 \\
\end{array}%
\right)
=
\left(
\begin{array}{c|ccccc}
& e _{1R} & e _{2R} & e _{3R} & \widetilde{L }_{4R} & e _{4R} \\[0.5ex] \hline
\overline{L }_{1L} & 0 & 0 & 0 & 0 & 0 \\[1ex]
\overline{L }_{2L} & 0 & 0 & 0 & 0 & m_{24} \\[1ex]
\overline{L }_{3L} & 0 & 0 & 0 & m_{35} & m_{34} \\[1ex]
\overline{L }_{4L} & 0 & 0 & m_{43} & M_{45}^{L } & 0 \\[1ex]
\overline{\widetilde{e }}_{4L} & 0 & m_{52} & m_{53} & 0 & M_{54}^{e} \\
\end{array}%
\right),
\label{eqn:cl_2}
\end{equation}
\Huchan{\Antonio{We use two methods for diagonalizing
the rotated mass matrix of Equation~\ref{eqn:cl_2},
one of which corresponds to the
numerical diagonalization carried out by the singular value decomposition (SVD) and the other} is an approximated analytical step-by-step diagonalization. We make use of the numerical SVD diagonalization for the \Antonio{exact diagonalization and perform numerical scans in main body of this work}, however it is worth discussing the analytic step-by-step diagonalization as it gives an \Antonio{analytical understanding}
on how the SM $Z$ gauge boson can induce the flavor violating interactions at tree-level with the $SU(2)$ violating mixings, which will be defined in the analytic diagonalization covered in Appendix~\ref{app:A}. \Antonio{From the comparison between the analytic and numerical computations, we found that the former
works quite well and yields resuls close the ones obtained from the latter. The comparisons between the analytic and numerical computations for both lepton and quark sectors will be discussed in detail in
Appendices \ref{app:D} and \ref{app:E}, respectively}. The charged lepton sector can be diagonalized by \Antonio{peforming} the SVD diagonalization as follows:}
\begin{equation}
M^{e \prime} = \operatorname{diag}\left( 0, m_{\mu}, m_{\tau}, M_{E_4}, M_{\widetilde{E}_4} \right) = V^{L} M^{e} (V^{e})^{\dagger},
\end{equation}
\Huchan{where $V^{L} (V^{e})$ is the mixing matrix for the left-handed (right-handed) leptonic fields, \Antonio{defined} as follows:}
\begin{equation}
\begin{split}
\begin{pmatrix}
e_{L} \\[0.5ex]
\mu_{L} \\[0.5ex]
\tau_{L} \\[0.5ex]
E_{4L} \\[0.5ex]
\widetilde{E}_{4L}
\end{pmatrix}
=
V^L
\begin{pmatrix}
e_{1L} \\[0.5ex]
e_{2L} \\[0.5ex]
e_{3L} \\[0.5ex]
e_{4L} \\[0.5ex]
\widetilde{e}_{4L}
\end{pmatrix},
\qquad
\begin{pmatrix}
e_{R} \\[0.5ex]
\mu_{R} \\[0.5ex]
\tau_{R} \\[0.5ex]
\widetilde{E}_{4R} \\[0.5ex]
E_{4R}
\end{pmatrix}
=
V^e
\begin{pmatrix}
e_{1R} \\[0.5ex]
e_{2R} \\[0.5ex]
e_{3R} \\[0.5ex]
\widetilde{e}_{4R} \\[0.5ex]
e_{4R}
\end{pmatrix},
\label{eqn:cl_mixing}
\end{split}
\end{equation}
\Huchan{and the numerical mixing matrices $V^{L,e}$ can be expressed by an analytic expression consisting of a series of $V_{ij}^{L,e}$ which describes mixing between $i$th and $j$th fermion where $i,j=1,2,3,4,5$ and this will be discussed in Appendix~\ref{app:A}.}
\subsection{Diagonalizing the up-type quark mass matrix} \label{sec:III_2}
The initial mass matrix for the up-type quark sector in the flavor basis is given by:
\begin{equation}
M^{u }
=
\left(
\begin{array}{c|ccccc}
& u _{1R} & u _{2R} & u _{3R} & u _{4R} & \widetilde{Q }_{4R} \\[0.5ex] \hline
\overline{Q }_{1L} & 0 & 0 & 0 & 0 & 0 \\[1ex]
\overline{Q }_{2L} & 0 & 0 & 0 & y_{24}^{u } v_{u} & 0 \\[1ex]
\overline{Q }_{3L} & 0 & 0 & 0 & y_{34}^{u } v_{u} & x_{34}^{Q } v_{\phi} \\[1ex]
\overline{Q }_{4L} & 0 & 0 & y_{43}^{u } v_{u} & 0 & M_{44}^{Q } \\[1ex]
\overline{\widetilde{u }}_{4L} & 0 & x_{42}^{u} v_{\phi}
& x_{43}^{u} v_{\phi} & M_{44}^{u} & 0 \\
\end{array}%
\right)
=
\left(
\begin{array}{c|ccccc}
& u _{1R} & u _{2R} & u _{3R} & \widetilde{Q }_{4R} & u _{4R} \\[0.5ex] \hline
\overline{Q }_{1L} & 0 & 0 & 0 & 0 & 0 \\[1ex]
\overline{Q }_{2L} & 0 & 0 & 0 & 0 & m_{24}^{u} \\[1ex]
\overline{Q }_{3L} & 0 & 0 & 0 & m_{35}^{u} & m_{34}^{u} \\[1ex]
\overline{Q }_{4L} & 0 & 0 & m_{43}^{u} & M_{44}^{Q } & 0 \\[1ex]
\overline{\widetilde{u }}_{4L} & 0 & m_{52}^{u}
& m_{53}^{u} & 0 & M_{44}^{u} \\
\end{array}%
\right)
, \label{eqn:uq_1}
\end{equation}
The mass matrix of Equation~\ref{eqn:uq_1} in the flavor basis is exactly consistent with the one corresponding to the charged lepton sector excepting for a few substitutions $y^{e} \rightarrow y^{u}$, $v_{d} \rightarrow v_{u}$, $x^{L} \rightarrow x^{Q}$ and $x^{e} \rightarrow x^{u}$.
\Huchan{The analytic mixing matrix for the up-quark sector is exactly same as the one for the charged lepton sector, however \Antonio{unlike the charged lepton sector,
diagonalizing the mass matrix for the up-quark sector requires} more caution as some numerical off-diagonal elements \Antonio{of order unity.}
This feature is resulted from the
some off-diagonal elements arising as a result of mixing between the heavy top quark mass and the other heavy exotic up-type quark masses are of order unitiry, not small enough to obtain precise results in the perturbative diagonalization when compared to the charged lepton sector. We discuss this feature in Appendix~\ref{app:E} by comparing a numerical mixing matrix \Antonio{obtained from the SVD with the one resulting from the analytic diagonalization.} Then we can numerically diagonalize the up-type mass matrix by using the SVD diagonalization as follows:}
\begin{equation}
\begin{split}
\begin{pmatrix}
u_{L} \\[0.5ex]
c_{L} \\[0.5ex]
t_{L} \\[0.5ex]
U_{4L} \\[0.5ex]
\widetilde{U}_{4L}
\end{pmatrix}
=
V_{L}^{u}
\begin{pmatrix}
u_{1L} \\[0.5ex]
u_{2L} \\[0.5ex]
u_{3L} \\[0.5ex]
u_{4L} \\[0.5ex]
\widetilde{u}_{4L}
\end{pmatrix}
, \qquad
\begin{pmatrix}
u_{R} \\[0.5ex]
c_{R} \\[0.5ex]
t_{R} \\[0.5ex]
\widetilde{U}_{4R} \\[0.5ex]
U_{4R}
\end{pmatrix}
=
V_{R}^{u}
\begin{pmatrix}
u_{1R} \\[0.5ex]
u_{2R} \\[0.5ex]
u_{3R} \\[0.5ex]
\widetilde{u}_{4R} \\[0.5ex]
u_{4R}
\end{pmatrix}
\label{eqn:up_mixing}
\end{split}
\end{equation}
\Antonio{where
the symbol $L$ means left-handed doublet and $e$ denotes right-handed singlet in the charged lepton sector, however
it is worth mentioning that the above described notation used in the lepton sector} becomes complicated in the quark sector since the mass matrices for the up- and down-type quark sectors have a different form, so we change the mixing notation by $V_{L(R)}^{u,d}$ instead of $V^{Q}$.
\Huchan{The analytic diagonalizations for the up-quark sector \Antonio{will be} discussed in Appendix~\ref{app:B}.}
\subsection{Diagonalizing the down-type quark mass matrix} \label{sec:III_3}
A nice
feature of the model under consideration is that it can predict
the CKM mixing matrix and this feature is mainly based on the mixings derived from the down-type quark mass matrix as we will see soon. A quite encouraging feature is that all the mixings among the three SM generations in the down-type quark sector can be accessible even though the first generation of the down-type quark sector remains massless and this feature is quite naturally attributed to this model with the vector-like family. We start from two mass matrices, one of which is for the up-quark sector whereas the another one is for the down-quark sector.
\begin{equation}
M^{u }=\left(
\begin{array}{c|ccccc}
& u _{1R} & u _{2R} & u _{3R} & u _{4R} & \widetilde{Q }_{4R} \\[0.5ex] \hline
\overline{Q }_{1L} & 0 & 0 & 0 & 0 & 0 \\[1ex]
\overline{Q }_{2L} & 0 & 0 & 0 & y_{24}^{u } v_{u} & 0 \\[1ex]
\overline{Q }_{3L} & 0 & 0 & 0 & y_{34}^{u } v_{u} & x_{34}^{Q } v_{\phi} \\[1ex]
\overline{Q }_{4L} & 0 & 0 & y_{43}^{u } v_{u} & 0 & M_{44}^{Q } \\[1ex]
\overline{\widetilde{u }}_{4L} & 0 & x_{42}^{u} v_{\phi}
& x_{43}^{u} v_{\phi} & M_{44}^{u} & 0 \\
\end{array}%
\right),
\quad
M^{d }=\left(
\begin{array}{c|ccccc}
& d _{1R} & d _{2R} & d _{3R} & d _{4R} & \widetilde{Q }_{4R} \\[0.5ex] \hline
\overline{Q }_{1L} & 0 & 0 & 0 & y_{14}^{d } v_{d} & 0 \\[1ex]
\overline{Q }_{2L} & 0 & 0 & 0 & y_{24}^{d } v_{d} & 0 \\[1ex]
\overline{Q }_{3L} & 0 & 0 & 0 & y_{34}^{d } v_{d} & x_{34}^{Q } v_{\phi} \\[1ex]
\overline{Q }_{4L} & 0 & 0 & y_{43}^{d } v_{d} & 0 & M_{44}^{Q } \\[1ex]
\overline{\widetilde{d }}_{4L} & 0 & x_{42}^{d} v_{\phi}
& x_{43}^{d} v_{\phi} & M_{44}^{d} & 0 \\
\end{array}%
\right),
\label{eqn:diff_up_down}
\end{equation}
This difference between the mass matrices for the up-type and down-type quark sectors was noticed for the first time in~\cite{King:2018fcg}.
The first property we need to focus on is the fifth column of both mass matrices is exactly same. The zeros appearing in the fifth column of both are the common elements shared by both up- and down-type quark mass matrices since the quark doublets as well as the fourth vector-like quark doublets contribute equally to both sectors. For the up-type quark sector, we were able to rotate further between $Q_{1L}$ and $Q_{2L}$ to vanish $y_{14}^{u} v_{u}$, however this rotation simply remixes $y_{14}^{d} v_{d}$ and $y_{24}^{d} v_{d}$, so both the down-type Yukawa terms survive. For the lower-left $2 \times 3$ block of the down-type quark mass matrix, the same zeros can appear since the down-type quarks $d_{1,2,3R}$ have a different mixing angle against that for the up-type quarks $u_{1,2,3R}$. The down-type mass matrix $M^{d}$ of Equation~\ref{eqn:diff_up_down} can be diagonalized by the numerical SVD diagonalization as follows:
\begin{equation}
\begin{split}
\begin{pmatrix}
d_{L} \\
s_{L} \\
b_{L} \\
D_{4L} \\
\widetilde{D}_{4L}
\end{pmatrix}
=
V_{L}^{d}
\begin{pmatrix}
d_{1L} \\
d_{2L} \\
d_{3L} \\
d_{4L} \\
\widetilde{d}_{4L}
\end{pmatrix}
, \qquad
\begin{pmatrix}
d_{R} \\
s_{R} \\
b_{R} \\
D_{4R} \\
\widetilde{D}_{4R}
\end{pmatrix}
=
V_{R}^{d}
\begin{pmatrix}
d_{1R} \\
d_{2R} \\
d_{3R} \\
\widetilde{d}_{4R} \\
d_{4R}
\end{pmatrix}.
\label{eqn:down_mixing}
\end{split}
\end{equation}
\Huchan{The analytic diagonalization for the down-quark sector is discussed in Appendix~\ref{app:C}.}
We will see that a numerical mixing matrix derived by the SVD is quite close to one by the analytic diagonalization in Appendix~\ref{app:E}. In Appendix~\ref{app:E} we confirm that even though the numerical mixing matrix $V_{L}^{d}$ can have all mixings among the three SM generations, the $Z$ coupling constants $D_{L}^{d \prime}$ in the mass basis will have zeros in the first column and row due to some internal cancellations. Therefore the whole structure of $D_{L}^{d \prime}$ is exactly same as the other $Z$ coupling constants $D_{L,R}^{u \prime}$ and $D_{R}^{d \prime}$ in the mass basis, so we verify that the SM $Z$ physics does not get affected by any specific basis we choose. This feature will be discussed again in the next section as well as in Appendix~\ref{app:E}.
\section{THE SM $Z$ GAUGE BOSON INTERACTIONS WITH THE VECTOR-LIKE FAMILY}
\label{sec:IV}
One of our main motivations of this work is to study flavor violating
processes mediated by the $Z$ gauge boson in order to constrain the mass range of the vector-like fermions. It is worth reminding that the neutral $Z^{\prime}$ gauge boson does not appear in the particle spectrum of this model due to the global $U(1)^{\prime}$ symmetry of the theory under consideration. It is worth mentioning that the tree-level flavor violating $Z$ decays are absent in the SM, indepently of the fermion mixings, as can be seen from Equation~\ref{eqn:SM_Z}
shown below:
\begin{equation}
\mathcal{L}_{\operatorname{SM}}^{Z} = g Z_{\mu} J_{Z}^{\mu} = \frac{g}{c_w} Z_{\mu} \sum_{f=e,\mu,\tau} \overline{f} \gamma^{\mu} \left( T^3 - \sin^2\theta_w Q \right) f
\label{eqn:SM_Z}
\end{equation}
Factoring out the prefactor $g/c_{w}$, we can find matrices $D_{L,R}^{e}$, which determine the magnitude of the coupling constant for the $Z$ interactions to either the left-handed or right-handed SM fermions.
\begin{equation}
\begin{split}
D_{L}^{e}&=
\left(
\begin{array}{c|ccc}
& e_{1L} & e_{2L} & e_{3L} \\[0.5ex]
\hline
\overline{e}_{1L} & \left( -\frac{1}{2}+\sin^2\theta_w \right) & 0 & 0 \\[1ex]
\overline{e}_{2L} & 0 & \left( -\frac{1}{2}+\sin^2\theta_w \right) & 0 \\[1ex]
\overline{e}_{3L} & 0 & 0 & \left( -\frac{1}{2}+\sin^2\theta_w \right) \\[1ex]
\end{array}
\right)
\\
D_{R}^{e}&=
\left(
\begin{array}{c|ccc}
& e_{1R} & e_{2R} & e_{3R} \\[0.5ex]
\hline
\overline{e}_{1R} & \left(\sin^2\theta_w \right) & 0 & 0 \\[1ex]
\overline{e}_{2R} & 0 & \left(\sin^2\theta_w \right) & 0 \\[1ex]
\overline{e}_{3R} & 0 & 0 & \left(\sin^2\theta_w \right) \\[1ex]
\end{array}
\right)
\end{split}
\end{equation}
However, this SM $Z$ gauge boson can cause the renormalizable flavor violating interactions with the SM fermions by extending the SM fermion sector by the vector-like fermions as well as by considering the $SU(2)$ violating mixings defined in Appendix~\ref{app:A} together. This features will be discussed in detail in the following subsections.
\subsection{FCNC mediated by the SM $Z$ gauge boson in the charged lepton sector with the fourth vector-like charged leptons} \label{sec:IV_1}
We can construct an extended the SM $Z$ coupling constants in the charged lepton sector with the vector-like charged leptons in the flavor basis.
\begin{equation}
\begin{split}
D_{L}^{e}
&=
\left(
\begin{array}{c|ccccc}
& e_{1L} & e_{2L} & e_{3L} & e_{4L} & \widetilde{e}_{4L} \\[1ex]
\hline
\overline{e}_{1L} & \left( -\frac{1}{2}+\sin^2\theta_w \right) & 0 & 0 & 0 & 0 \\[1ex]
\overline{e}_{2L} & 0 & \left( -\frac{1}{2}+\sin^2\theta_w \right) & 0 & 0 & 0 \\[1ex]
\overline{e}_{3L} & 0 & 0 & \left( -\frac{1}{2}+\sin^2\theta_w \right) & 0 & 0 \\[1ex]
\overline{e}_{4L} & 0 & 0 & 0 & \left( -\frac{1}{2}+\sin^2\theta_w \right) & 0 \\[1ex]
\overline{\widetilde{e}}_{4L} & 0 & 0 & 0 & 0 & \left( \sin^2\theta_w \right)
\end{array}
\right)
\\
D_{R}^{e}
&=
\left(
\begin{array}{c|ccccc}
& e_{1R} & e_{2R} & e_{3R} & \widetilde{e}_{4R} & e_{4R} \\[1ex]
\hline
\overline{e}_{1R} & \left( \sin^2\theta_w \right) & 0 & 0 & 0 & 0 \\[1ex]
\overline{e}_{2R} & 0 & \left( \sin^2\theta_w \right) & 0 & 0 & 0 \\[1ex]
\overline{e}_{3R} & 0 & 0 & \left( \sin^2\theta_w \right) & 0 & 0 \\[1ex]
\overline{\widetilde{e}}_{4R} & 0 & 0 & 0 & \left( -\frac{1}{2}+\sin^2\theta_w \right) & 0 \\[1ex]
\overline{e}_{4R} & 0 & 0 & 0 & 0 & \left( \sin^2\theta_w \right)
\end{array}
\right)
\label{eqn:SMZ_DLe_DRe}
\end{split}
\end{equation}
where it is worth reminding that the order of the left-handed fermions is 12345, whereas that of the right-handed fermions is 12354 (This ordering is stressed in Appendix~\ref{app:A}). An important feature of this SM $Z$ coupling constants of Equation~\ref{eqn:SMZ_DLe_DRe} is the couplings constants are naturally determined, based on the nature of the vector-like charged leptons, without imposing neither any other symmetry nor other constraints. Therefore, the SM $Z$ coupling constants of Equation~\ref{eqn:SMZ_DLe_DRe} are not the identity matrix anymore unlike the case for
the SM charged leptons. From these considerations, it follows that there can exist non-zero
off-diagonal elements in the mass basis by operating the $SU(2)$ violating mixings. Reminding the mixing matrices for the left- or right-handed charged leptons of Equation~\ref{eqn:cl_umm_LH_RH}, the coupling constant in the mass basis ($D_{L,R}^{e \prime}$) can be written as follows:
\begin{equation}
\begin{split}
D_{L}^{e \prime} &= V^{L} D_{L}^{e} (V^{L})^{\dagger}
\\[1ex]
&= V_{45}^{L} V_{23}^{L} V_{25}^{L} V_{35}^{L} V_{34}^{L} D_{L}^{e} (V_{34}^{L})^{\dagger} (V_{35}^{L})^{\dagger} (V_{25}^{L})^{\dagger} (V_{23}^{L})^{\dagger} (V_{45}^{L})^{\dagger}
\\[1ex]
D_{R}^{e \prime} &= V^{e} D_{R}^{e} (V^{e})^{\dagger}
\\[1ex]
&= V_{54}^{e} V_{23}^{e} V_{25}^{e} V_{35}^{e} V_{24}^{e} V_{34}^{e} D_{R}^{e} (V_{34}^{e})^{\dagger} (V_{24}^{e})^{\dagger} (V_{35}^{e})^{\dagger} (V_{25}^{e})^{\dagger} (V_{23}^{e})^{\dagger} (V_{54}^{e})^{\dagger}
\label{eqn:SMZ_cl_mb_LR_1}
\end{split}
\end{equation}
It is possible to make the SM $Z$ coupling constants of Equation~\ref{eqn:SMZ_cl_mb_LR_1} simpler by using the $SU(2)$ conserving mixing as it just remixes an identity matrix.
\begin{equation}
\begin{split}
D_{L}^{e \prime} &= V_{45}^{L} V_{23}^{L} V_{25}^{L} V_{35}^{L} D_{L}^{e} (V_{35}^{L})^{\dagger} (V_{25}^{L})^{\dagger} (V_{23}^{L})^{\dagger} (V_{45}^{L})^{\dagger}
\\[1ex]
D_{R}^{e \prime} &= V_{54}^{e} V_{23}^{e} V_{25}^{e} V_{35}^{e} D_{R}^{e} (V_{35}^{e})^{\dagger} (V_{25}^{e})^{\dagger} (V_{23}^{e})^{\dagger} (V_{54}^{e})^{\dagger}
\label{eqn:SMZ_cl_mb_LR_2}
\end{split}
\end{equation}
However, the following mixing matrices after the $SU(2)$ violating mixings $V_{35}^{L,e}$ must be conserved since all of them contribute to the off-diagonal elements of the coupling constants in the mass basis. It is insightful to look at the coupling constants $D_{L,R}^{e \prime}$ in the mass basis (We substitute $\left( -1/2+\sin^2\theta_w \right)$ by $a$ and $\left( \sin^2\theta_w \right)$ by $b$ and assume the mixing angles $\theta_{35,25,23,45(54)}^{L,e}$ are small enough to approximate for simplicity).
\begin{equation}
\begin{split}
D_{L}^{e \prime} &\approx
\begin{pmatrix}
a & 0 & 0 & 0 & 0 \\[1ex]
0 & a(1 + \theta_{23}^{L2}) + b \theta_{25}^{L2} & b \theta_{25}^{L} \theta_{35}^{L} & b \theta_{25}^{L} \theta_{45}^{L} & (a-b) \theta_{25}^{L} \\[1ex]
0 & b \theta_{25}^{L} \theta_{35}^{L} & a(1+\theta_{23}^{L2}) + b \theta_{35}^{L2} & b \theta_{35}^{L} \theta_{45}^{L} & (a-b) \theta_{35}^{L} \\[1ex]
0 & b \theta_{25}^{L} \theta_{45}^{L} & b \theta_{35}^{L} \theta_{45}^{L} & a + b \theta_{45}^{L2} & (a-b) \theta_{45}^{L} \\[1ex]
0 & (a-b) \theta_{25}^{L} & (a-b) \theta_{35}^{L} & (a-b) \theta_{45}^{L} & b + a(\theta_{25}^{L2} + \theta_{35}^{L2} + \theta_{45}^{L2})
\end{pmatrix}
\\[1ex]
D_{R}^{e \prime} &\approx
\begin{pmatrix}
b & 0 & 0 & 0 & 0 \\[1ex]
0 & b(1 + \theta_{23}^{e2}) + a \theta_{25}^{e2} & a \theta_{25}^{e} \theta_{35}^{e} & (b-a) \theta_{25}^{e} & -a \theta_{25}^{e} \theta_{54}^{e} \\[1ex]
0 & a \theta_{25}^{e} \theta_{35}^{e} & b(1+\theta_{23}^{e2}) + a \theta_{35}^{e2} & (b-a) \theta_{35}^{e} & -a \theta_{35}^{e} \theta_{54}^{e} \\[1ex]
0 & (b-a) \theta_{25}^{e} & (b-a) \theta_{35}^{e} & a + b (\theta_{25}^{e2}+\theta_{35}^{e2}+\theta_{54}^{e2}) & (a-b) \theta_{54}^{e} \\[1ex]
0 & -a \theta_{25}^{e} \theta_{54}^{e} & -a \theta_{35}^{e} \theta_{54}^{e} & (a-b) \theta_{54}^{e} & b + a \theta_{54}^{e2}
\end{pmatrix}
\label{eqn:simple_DLep_DRep}
\end{split}
\end{equation}
There are two important features we can read off from the SM $Z$ coupling constants of Equation~\ref{eqn:simple_DLep_DRep}; the first is the diagonal elements $(a,a,a,a,b)$ and $(b,b,b,a,b)$ get hardly affected by the small mixing angles and the second is magnitude of the off-diagonal elements are very weak as the mixing angles are defined by the ratio between the Yukawa and the vector-like masses. Therefore, we can imply the flavor violating mixing mediated by the SM $Z$ gauge boson in the mass basis. Using the SM $Z$ coupling constants in the mass basis, we can draw the Feynman diagram for $\tau \rightarrow \mu \mu \mu$ and $Z \rightarrow \mu \tau$ decay at tree-level and this will be discussed in the next section.
\subsection{FCNC mediated by the SM $Z$ gauge boson in the quark sector with the fourth vector-like quarks} \label{sec:IV_2}
The quark sector have two different mass matrices for the up- and down-type quark sector. We start from the up-type quark sector first. The SM $Z$ coupling constants for the up-type quark sector are given by:
\begin{equation}
\begin{split}
D_{L}^{u} &=
\left(
\begin{array}{c|ccccc}
& u_{1L} & u_{2L} & u_{3L} & u_{4L} & \widetilde{u}_{4L} \\[0.5ex]
\hline
\overline{u}_{1L} & \left( \frac{1}{2}-\frac{2}{3}\sin^2\theta_w \right) & 0 & 0 & 0 & 0 \\[1ex]
\overline{u}_{2L} & 0 & \left( \frac{1}{2}-\frac{2}{3}\sin^2\theta_w \right) & 0 & 0 & 0 \\[1ex]
\overline{u}_{3L} & 0 & 0 & \left( \frac{1}{2}-\frac{2}{3}\sin^2\theta_w \right) & 0 & 0 \\[1ex]
\overline{u}_{4L} & 0 & 0 & 0 & \left( \frac{1}{2}-\frac{2}{3}\sin^2\theta_w \right) & 0 \\[1ex]
\overline{\widetilde{u}}_{4L} & 0 & 0 & 0 & 0 & \left( -\frac{2}{3}\sin^2\theta_w \right) \\[1ex]
\end{array}
\right) \\[1ex]
D_{R}^{u} &=
\left(
\begin{array}{c|ccccc}
& u_{1R} & u_{2R} & u_{3R} & \widetilde{u}_{4R} & u_{4R} \\[0.5ex]
\hline
\overline{u}_{1R} & \left( -\frac{2}{3}\sin^2\theta_w \right) & 0 & 0 & 0 & 0 \\[1ex]
\overline{u}_{2R} & 0 & \left( -\frac{2}{3}\sin^2\theta_w \right) & 0 & 0 & 0 \\[1ex]
\overline{u}_{3R} & 0 & 0 & \left( -\frac{2}{3}\sin^2\theta_w \right) & 0 & 0 \\[1ex]
\overline{\widetilde{u}}_{4R} & 0 & 0 & 0 & \left( \frac{1}{2}-\frac{2}{3}\sin^2\theta_w \right) & 0 \\[1ex]
\overline{u}_{4R} & 0 & 0 & 0 & 0 & \left( -\frac{2}{3}\sin^2\theta_w \right) \\[1ex]
\end{array}
\right)
\label{eqn:SMZ_DLu_DRu}
\end{split}
\end{equation}
The up-type quark mass matrix is exactly same as the one for the charged lepton, so we can simply follow the mixing matrices given in Equation~\ref{eqn:ana_up_mixing}. Then the SM $Z$ gauge coupling constants in the mass basis are defined by:
\begin{equation}
\begin{split}
D_{L}^{u \prime} &= V_{L}^{u} D_{L}^{u} V_{L}^{u \dagger}
\\[1ex]
&= (V_{L}^{u})_{45} (V_{L}^{u})_{23} (V_{L}^{u})_{25} (V_{L}^{u})_{35} (V_{L}^{u})_{34} D_{L}^{u} (V_{L}^{u})_{34}^{\dagger} (V_{L}^{u})_{35}^{\dagger} (V_{L}^{u})_{25}^{\dagger} (V_{L}^{u})_{23}^{\dagger} (V_{L}^{u})_{45}^{\dagger}
\\[1ex]
&= (V_{L}^{u})_{45} (V_{L}^{u})_{23} (V_{L}^{u})_{25} (V_{L}^{u})_{35} D_{L}^{u} (V_{L}^{u})_{35}^{\dagger} (V_{L}^{u})_{25}^{\dagger} (V_{L}^{u})_{23}^{\dagger} (V_{L}^{u})_{45}^{\dagger}
\\[1ex]
D_{R}^{u \prime} &= V_{R}^{u} D_{R}^{u} V_{R}^{u \dagger}
\\[1ex]
&= (V_{R}^{u})_{54} (V_{R}^{u})_{23} (V_{R}^{u})_{25} (V_{R}^{u})_{35} (V_{R}^{u})_{24} (V_{R}^{u})_{34} D_{R}^{u} (V_{R}^{u})_{34}^{\dagger} (V_{R}^{u})_{24}^{\dagger} (V_{R}^{u})_{35}^{\dagger} (V_{R}^{u})_{25}^{\dagger} (V_{R}^{u})_{23}^{\dagger} (V_{R}^{u})_{54}^{\dagger}
\\
&= (V_{R}^{u})_{54} (V_{R}^{u})_{23} (V_{R}^{u})_{25} (V_{R}^{u})_{35} D_{R}^{u} (V_{R}^{u})_{35}^{\dagger} (V_{R}^{u})_{25}^{\dagger} (V_{R}^{u})_{23}^{\dagger} (V_{R}^{u})_{54}^{\dagger}
\label{eqn:SMZ_uq_mb_LR_1}
\end{split}
\end{equation}
The SM $Z$ gauge coupling constants for the up-quark sector in the mass basis can be seen by (We substitute $\left( 1/2-2/3\sin^2\theta_w \right)$ by $c$ and $\left( -2/3\sin^2\theta_w \right)$ by $d$ and assume the mixing angles $\theta_{35,25,23,45(54)L,R}^{u}$ are small enough to approximate for simplicity):
\begin{equation}
\begin{split}
D_{L}^{u \prime} &\approx
\begin{pmatrix}
c & 0 & 0 & 0 & 0 \\[1ex]
0 & c & d \theta_{25L}^{u} \theta_{35L}^{u} & d \theta_{25L}^{u} \theta_{45L}^{u} & (c-d) \theta_{25L}^{u} \\[1ex]
0 & d \theta_{25L}^{u} \theta_{35L}^{u} & c & d \theta_{35L}^{u} \theta_{45L}^{u} & (c-d) \theta_{35L}^{u} \\[1ex]
0 & d \theta_{25L}^{u} \theta_{45L}^{u} & d \theta_{35L}^{u} \theta_{45L}^{u} & c & (c-d) \theta_{45L}^{u} \\[1ex]
0 & (c-d) \theta_{25L}^{u} & (c-d) \theta_{35L}^{u} & (c-d) \theta_{45L}^{u} & d
\end{pmatrix}
\\[1ex]
D_{R}^{u \prime} &\approx
\begin{pmatrix}
d & 0 & 0 & 0 & 0 \\[1ex]
0 & d & c \theta_{25R}^{u} \theta_{35R}^{u} & (d-c) \theta_{25R}^{u} & -c \theta_{25R}^{u} \theta_{54R}^{u} \\[1ex]
0 & c \theta_{25R}^{u} \theta_{35R}^{u} & d & (d-c) \theta_{35R}^{u} & -c \theta_{35R}^{u} \theta_{54R}^{u} \\[1ex]
0 & (d-c) \theta_{25R}^{u} & (d-c) \theta_{35R}^{u} & c & (c-d) \theta_{54R}^{u} \\[1ex]
0 & -c \theta_{25R}^{u} \theta_{54R}^{u} & -c \theta_{35R}^{u} \theta_{54R}^{u} & (c-d) \theta_{54R}^{u} & d
\end{pmatrix}
\label{eqn:simple_DLup_DRup}
\end{split}
\end{equation}
Next, we focus on the down-type quark mass matrix and the left-handed mixing matrices for that is different when compared to other left-handed mixing matrices for the up-type or charged lepton mass matrix in that it can reach to all mixings among the three SM generations. Keeping that in mind, we start from the SM $Z$ coupling constants for the down-type quarks.
\begin{equation}
\begin{split}
D_{L}^{d}
&=
\left(
\begin{array}{c|ccccc}
& d_{1L} & d_{2L} & d_{3L} & d_{4L} & \widetilde{d}_{4L} \\[0.5ex]
\hline
\overline{d}_{1L} & \left( -\frac{1}{2}+\frac{1}{3}\sin^2\theta_w \right) & 0 & 0 & 0 & 0 \\[1ex]
\overline{d}_{2L} & 0 & \left( -\frac{1}{2}+\frac{1}{3}\sin^2\theta_w \right) & 0 & 0 & 0 \\[1ex]
\overline{d}_{3L} & 0 & 0 & \left( -\frac{1}{2}+\frac{1}{3}\sin^2\theta_w \right) & 0 & 0 \\[1ex]
\overline{d}_{4L} & 0 & 0 & 0 & \left( -\frac{1}{2}+\frac{1}{3}\sin^2\theta_w \right) & 0 \\[1ex]
\overline{\widetilde{d}}_{4L} & 0 & 0 & 0 & 0 & \left( \frac{1}{3}\sin^2\theta_w \right)
\end{array}
\right)
\\
D_{R}^{d}
&=
\left(
\begin{array}{c|ccccc}
& d_{1R} & d_{2R} & d_{3R} & \widetilde{d}_{4R} & d_{4R} \\[0.5ex]
\hline
\overline{d}_{1R} & \left( \frac{1}{3}\sin^2\theta_w \right) & 0 & 0 & 0 & 0 \\[1ex]
\overline{d}_{2R} & 0 & \left( \frac{1}{3}\sin^2\theta_w \right) & 0 & 0 & 0 \\[1ex]
\overline{d}_{3R} & 0 & 0 & \left( \frac{1}{3}\sin^2\theta_w \right) & 0 & 0 \\[1ex]
\overline{\widetilde{d}}_{4R} & 0 & 0 & 0 & \left( -\frac{1}{2}+\frac{1}{3}\sin^2\theta_w \right) & 0 \\[1ex]
\overline{d}_{4R} & 0 & 0 & 0 & 0 & \left( \frac{1}{3}\sin^2\theta_w \right)
\end{array}
\right)
\label{eqn:SMZ_DLd_DRd}
\end{split}
\end{equation}
After simplifying the whole left-handed (right-handed) mixing matrices of Equation~\ref{eqn:down_mix2},
\Antonio{we obtain the following matrices of $Z$ couplings with quarks
\begin{equation}
\begin{split}
D_{L}^{d \prime} = V_{L}^{d} D_{L}^{d} (V_{L}^{d})^{\dagger} \\[1ex]
D_{R}^{d \prime} = V_{R}^{d} D_{R}^{d} (V_{R}^{d})^{\dagger}
\label{eqn:DLdp_DRdp}
\end{split}
\end{equation}}
The SM $Z$ coupling constants for the right-handed down-type quarks in the mass basis have the same form given in Equation~\ref{eqn:simple_DLep_DRep}, whereas those for the left-handed down-type quarks involves $12$ and $13$ mixings more, so it is worthwhile to look at its complete form (We again substitute $\left( -1/2+1/3 \sin^2\theta_w \right)$ by $e$ and $\left( 1/3 \sin^2\theta_w \right)$ by $f$ and assume all relevant mixing angles are small enough to approximate for simplicity)
\begin{equation}
\begin{split}
D_{L}^{d \prime} &\approx
\begin{pmatrix}
e & 0 & 0 & 0 & 0 \\[1ex]
0 & e & e \theta_{12L}^{d} \theta_{13L}^{d} + f \theta_{25L}^{d} \theta_{35L}^{d} & -e \theta_{23L}^{d} \theta_{34L}^{d} + f \theta_{25L}^{d} \theta_{45L}^{d} & (e-f) \theta_{25L}^{d} \\[1ex]
0 & e \theta_{12L}^{d} \theta_{13L}^{d} + f \theta_{25L}^{d} \theta_{35L}^{d} & e & f \theta_{35L}^{d} \theta_{45L}^{d} & (e-f) \theta_{35L}^{d} \\[1ex]
0 & -e \theta_{23L}^{d} \theta_{34L}^{d} + f \theta_{25L}^{d} \theta_{45L}^{d} & f \theta_{35L}^{d} \theta_{45L}^{d} & e & (e-f) \theta_{45L}^{d} \\[1ex]
0 & (e-f) \theta_{25L}^{d} & (e-f) \theta_{35L}^{d} & (e-f) \theta_{45L}^{d} & f
\end{pmatrix}
\\[1ex]
D_{R}^{d \prime} &\approx
\begin{pmatrix}
f & 0 & 0 & 0 & 0 \\[1ex]
0 & f & e \theta_{25R}^{d} \theta_{35R}^{d} & (f-e) \theta_{25R}^{d} & -e \theta_{25R}^{d} \theta_{54R}^{d} \\[1ex]
0 & e \theta_{25R}^{d} \theta_{35R}^{d} & f & (f-e) \theta_{35R}^{d} & -e \theta_{35R}^{d} \theta_{54R}^{d} \\[1ex]
0 & (f-e) \theta_{25R}^{d} & (f-e) \theta_{35R}^{d} & e & (e-f) \theta_{54R}^{d} \\[1ex]
0 & -e \theta_{25R}^{d} \theta_{54R}^{d} & -e \theta_{35R}^{d} \theta_{54R}^{d} & (e-f) \theta_{54R}^{d} & f
\end{pmatrix},
\label{eqn:simple_DLdp_DRdp}
\end{split}
\end{equation}
where it can confirm two relations hold for the zeros appearing in the first row and column of $D_{L}^{d \prime}$: $\theta_{13L}^{d} \simeq \theta_{12L}^{d} \theta_{23L}^{d}$ and $\theta_{15L}^{d} \simeq \theta_{12L}^{d} \theta_{25L}^{d}$. Following the analytic mixings given in Equation~\ref{eqn:down_mix2}, the left-handed down type quark sector can access to all mixings among the three SM generations and this feature is also reflected on a numerical mixing matrix $V_{L}^{d}$ of Equation~\ref{eqn:a_num_VCKM}. What this implies is the SM $Z$ physics does not get affected by any specific basis we choose and this will be verified again numerically in Appendix~\ref{app:E}.
\section{PHENOMENOLOGY IN THE CHARGED LEPTON SECTOR OF THE SM} \label{sec:V}
Now that we have defined all required mixings and coupling constants in both quark and charged lepton sectors, it is time to discuss the relevant phenomenology. As mentioned in the introduction, one of our main goals is to study the FCNC observables to constrain the possible mass range of the vector-like fermions.
Given that the second and third generations of SM charged leptons
acquire masses through their mixings with the fourth vector-like charged leptons, we will restrict our analysis
to the constraints imposed on the flavor violating decays involving
the second and third charged lepton generations such as $\tau \rightarrow \mu \gamma, \tau \rightarrow \mu \mu \mu$ and lastly $Z \rightarrow \mu \tau$ decay.
\subsection{Analytic expression for $\tau \rightarrow \mu \gamma$ decay}
The most important FCNC constraint corresponds to the charged lepton flavor violating (CLFV) $\mu \rightarrow e \gamma$ decay, however we can not make an appropriate prediction for the constraint as the electron does not acquire a mass in the model under consideration. This is due to the fact that the fermion sector of the model includes two heavy fermionic seesaw mediators. As previously mentioned, adding a fifth vector-like family to the fermion sector of the model will generate a nonvanishing electron. However in order to keep our model as economical as possible and to simplify our analysis corresponding to the FCNC constraints on vector-like masses, we restrict to the case of a fourth vector-like family in the fermionic spectrum.
Therefore, in view of the aforementioned considerations, we first consider the CLFV decay $\tau \rightarrow \mu \gamma$ in order to determine how the model parameter space gets affected by the experimental constraint arising from this decay.
For the $\tau \rightarrow \mu \gamma$ decay, the
leading order contribution appears in the one-loop diagrams since there is no possible contribution at tree-level. Then, all possible Feynman diagrams contributing to
the $\tau \rightarrow \mu \gamma$ decay are given in Figure~\ref{fig:tmgdecay},
\begin{figure}[H]
\centering
\includegraphics[keepaspectratio,width=\textwidth]{taumugamma_Feynman1.png}
\caption{Diagrams contributing to the charged lepton flavor violation (CLFV) $\tau \rightarrow \mu \gamma$ decay at one-loop level in the mass basis. The cross notation in each diagram means the helicity flip process.}
\label{fig:tmgdecay}
\end{figure}
where $g_{LL(RR)}$ are the LH-LH (RH-RH) coupling constants defined in the mass basis for the charged lepton sector. These one-loop contributions mediated by the
$Z^{\prime}$ boson were studied in one of our previous works~\cite{CarcamoHernandez:2019ydc} and their corresponding analytic prediction for the branching ratio of $\tau \rightarrow \mu \gamma$ decay is given in Equation~\ref{eqn:analytic_tmgdecay}~\cite{CarcamoHernandez:2019ydc,Raby:2017igl,Lindner:2016bgg,Lavoura:2003xp,Chiang:2011cv} shown below:
\begin{equation}
\begin{split}
\operatorname{BR} \left( \tau \rightarrow \mu \gamma \right) &= \frac{\alpha_{\operatorname{em}}}{1024\pi^4} \frac{m_{\tau}^5}{M_Z^4 \Gamma_{\tau}} \Big( \lvert g_{\tau \mu}^L g_{\mu \mu}^L F(x_\mu) + g_{\tau \tau}^L g_{\tau \mu}^L F(x_\tau) + g_{\tau E_4}^L g_{E_4 \mu}^L F(x_{E_4}) + g_{\tau \widetilde{E}_4}^L g_{\widetilde{E}_4 \mu}^L F(x_{\widetilde{E}_4}) \\
&+ \frac{m_{\mu}}{m_{\tau}} g_{\tau \mu}^L g_{\mu \mu}^L F(x_\mu) + \frac{m_{\mu}}{m_{\tau}} g_{\tau \tau}^L g_{\tau \mu}^L F(x_\tau) + \frac{m_{\mu}}{m_{\tau}} g_{\tau E_4}^L g_{E_4 \mu}^L F(x_{E_4}) + \frac{m_{\mu}}{m_{\tau}} g_{\tau \widetilde{E}_4}^L g_{\widetilde{E}_4 \mu}^L F(x_{\widetilde{E}_4}) \\
&+ \frac{m_{\mu}}{m_{\tau}} g_{\tau \mu}^L g_{\mu \mu}^R G(x_{\mu}) + \frac{m_{\tau}}{m_{\tau}} g_{\tau \tau}^L g_{\tau \mu}^R G(x_{\mu}) + \frac{M_{E_4}}{m_{\tau}} g_{\tau E_4}^L g_{\widetilde{E}_4 \mu}^R G(x_{E_4}) + \frac{M_{\widetilde{E}_4}}{m_{\tau}} g_{\tau \widetilde{E}_4}^L g_{E_4 \mu}^R G(x_{\widetilde{E}_4}) \rvert^2 \\
&+ \lvert g_{\tau \mu}^R g_{\mu \mu}^R F(x_\mu) + g_{\tau \tau}^R g_{\tau \mu}^R F(x_\tau) + g_{\tau E_4}^R g_{E_4 \mu}^R F(x_{E_4}) + g_{\tau \widetilde{E}_4}^R g_{\widetilde{E}_4 \mu}^R F(x_{\widetilde{E}_4}) \\
&+ \frac{m_{\mu}}{m_{\tau}} g_{\tau \mu}^R g_{\mu \mu}^R F(x_\mu) + \frac{m_{\mu}}{m_{\tau}} g_{\tau \tau}^R g_{\tau \mu}^R F(x_\tau) + \frac{m_{\mu}}{m_{\tau}} g_{\tau E_4}^R g_{E_4 \mu}^R F(x_{E_4}) + \frac{m_{\mu}}{m_{\tau}} g_{\tau \widetilde{E}_4}^R g_{\widetilde{E}_4 \mu}^R F(x_{\widetilde{E}_4}) \\
&+ \frac{m_{\mu}}{m_{\tau}} g_{\tau \mu}^R g_{\mu \mu}^L G(x_{\mu}) + \frac{m_{\tau}}{m_{\tau}} g_{\tau \tau}^R g_{\tau \mu}^L G(x_{\mu}) + \frac{M_{\widetilde{E}_4}}{m_{\tau}} g_{\tau E_4}^R g_{\widetilde{E}_4 \mu}^L G(x_{\widetilde{E}_4}) + \frac{M_{E_4}}{m_{\tau}} g_{\tau \widetilde{E}_4}^R g_{E_4 \mu}^L G(x_{E_4}) \rvert^2 \Big),
\label{eqn:analytic_tmgdecay}
\end{split}
\end{equation}
where $\alpha_{\operatorname{em}}$ is the fine structure constant, $\Gamma_{\tau}$ is the total decay width of the tau lepton ($\Gamma_{\tau} = 5 \times \Gamma(\tau_L^- \rightarrow \nu_\tau e_L^- \overline{\nu}_e) = 2.0 \times 10^{-12}$) and $F$ and $G$ are the loop functions defined by:
\begin{equation}
\begin{split}
F(x) &= \frac{5x^4 - 14x^3 + 39x^2 - 38x - 18x^2 \operatorname{ln}x + 8}{12(1-x)^4} \\[1ex]
G(x) &= \frac{x^3 + 3x - 6x \operatorname{ln}x - 4}{2(1-x)^3}, \qquad x = \frac{m_{\operatorname{loop}}^2}{M_{Z}^2}
\end{split}
\end{equation}
where $m_{\operatorname{loop}}$ is the propagating mass of the charged leptons in the loop.
The most dominant contributions to the $\tau \rightarrow \mu \gamma$ branching ratio correspond to the terms proportional to $M_{E_4(\widetilde{E}_4)}/m_{\tau}$ because charged vector-like leptons are heavier than $200$ GeV, thus implying that the enhancement factor $M_{E_4(\widetilde{E}_4)}/m_{\tau}$ makes those contributions much larger than the ones not involving this factor.
However, the contributions to the $\tau \rightarrow \mu \gamma$ decay rate involving the terms having the aforementioned proportionality factor do not keep increasing
as the vector-like fermions get heavier
since their flavor violating coupling constants get more suppressed at the same time by the small mixing angles, defined by the ratio between Yukawa and vector-like masses.
Therefore, these compensations provide some balanced relation between the vector-like mass and the coupling of the $Z$ gauge boson with a SM charged antilepton (lepton) and heavy charged vector-like (antilepton) lepton.
\Steve{The experimental bound for the branching ratio of $\tau \rightarrow \mu \gamma$ decay is given by~\cite{Crivellin:2020ebi,MEG:2016leq,BaBar:2009hkt}}:
\begin{equation}
\operatorname{BR}\left( \tau \rightarrow \mu \gamma \right)_{\operatorname{EXP}} = 4.4 \times 10^{-8}
\end{equation}
\subsection{Analytic expression for $\tau \rightarrow \mu \mu \mu$ decay}
The other interesting flavor violating decay mode is the $\tau \rightarrow \mu \mu \mu$ decay mediated by the SM $Z$ gauge boson. As the model under consideration
has
$Z$ mediated renormalizable flavor violating interactions, we can draw the Feynman diagrams for the $\tau \rightarrow \mu \mu \mu$ decay at tree-level as given in Figure~\ref{fig:diagrams_tau3mu}.
\begin{figure}[H]
\centering
\includegraphics[keepaspectratio,width=\textwidth]{tau3mu_Feynman1.png}
\caption{Diagrams contributing to the charged lepton flavor violation (CLFV) $\tau \rightarrow 3\mu$ decay at tree-level. We refer to the top-left diagram as $\mathcal{M}_{LL}$ and the top-right as $\mathcal{M}_{RR}$ and similarly for the two below diagrams as $\mathcal{M}_{LR,RL}$, respectively.}
\label{fig:diagrams_tau3mu}
\end{figure}
The contributions shown in Figure~\ref{fig:diagrams_tau3mu} are
beyond Standard Model (BSM)
effects, thus they need to be computed to set constraints on the model parameter space.
In order to derive an analytic expression for the CLFV $\tau \rightarrow 3\mu$ decay rate mediated by the SM $Z$ gauge boson,
we start by writting down its definition as follows:
\begin{equation}
d\Gamma\left( \tau \rightarrow 3\mu \right) = \frac{1}{2m_{\tau}} \frac{d^3 p_2}{(2\pi)^3 2E_2} \frac{d^3 p_3}{(2\pi)^3 2E_3} \frac{d^3 p_4}{(2\pi)^3 2E_4} \lvert \mathcal{M} \rvert^2 \left( 2\pi \right)^4 \delta^{4} \left( p_1 - p_2 - p_3 - p_4 \right)
\end{equation}
Evaluating each polarized diagram in Figure~\ref{fig:diagrams_tau3mu}, it yields the following result:
\begin{equation}
\begin{aligned}
\lvert \mathcal{M}_{LL} \rvert^2 &= \left( \frac{g_{\mu \tau}^L g_{\mu \mu}^L}{4M_Z^2} \right)^2 256\left( p_1 \cdot p_3 \right) \left( p_2 \cdot p_4 \right), \quad \lvert \mathcal{M}_{RR} \rvert^2 &= \left( \frac{g_{\mu \tau}^R g_{\mu \mu}^R}{4M_Z^2} \right)^2 256\left( p_2 \cdot p_4 \right) \left( p_1 \cdot p_3 \right) \\
\lvert \mathcal{M}_{LR} \rvert^2 &= \left( \frac{g_{\mu \tau}^L g_{\mu \mu}^R}{4M_Z^2} \right)^2 256\left( p_1 \cdot p_4 \right) \left( p_2 \cdot p_3 \right), \quad \lvert \mathcal{M}_{RL} \rvert^2 &= \left( \frac{g_{\mu \tau}^R g_{\mu \mu}^L}{4M_Z^2} \right)^2 256\left( p_1 \cdot p_4 \right) \left( p_2 \cdot p_3 \right)
\end{aligned}
\end{equation}
We are now ready to determine the squared amplitude averaged and summed over the initial and final spin states.
\begin{equation}
\begin{split}
\frac{1}{2} \sum_{\operatorname{spin}} \lvert \mathcal{M} \rvert^2 &= \frac{1}{2} \left( \lvert \mathcal{M}_{LL} \rvert^2 + \lvert \mathcal{M}_{RR} \rvert^2 + \lvert \mathcal{M}_{LR} \rvert^2 + \lvert \mathcal{M}_{RL} \rvert^2 \right) \\
&= \frac{8}{M_Z^4} \left[ \left( g_{\mu \tau}^{2L} g_{\mu \mu}^{2L} + g_{\mu \tau}^{2R} g_{\mu \mu}^{2R} \right) \left( p_1 \cdot p_3 \right) \left( p_2 \cdot p_4 \right) + \left( g_{\mu \tau}^{2L} g_{\mu \mu}^{2R} + g_{\mu \tau}^{2R} g_{\mu \mu}^{2L} \right) \left( p_1 \cdot p_4 \right) \left( p_2 \cdot p_3 \right) \right]
\end{split}
\end{equation}
The momentum of the particles involved in the $\tau \rightarrow 3\mu$ are written in the rest frame as follows:
\begin{equation}
\begin{split}
p_1 &= ( m_{\tau}, \vec{0} ) \\
p_2 &= \left( E_2, \vec{p}_2 \right) \\
p_3 &= \left( m_{\tau} - E_2 - E_4, -\vec{p}_2 -\vec{p}_4 \right) \\
p_4 &= \left( E_4, \vec{p}_4 \right) \\
\end{split}
\end{equation}
Then, we can carry out the inner products of momenta taking into account
the momentum conservation ($p_1 = p_2 + p_3 + p_4$).
\begin{equation}
\begin{split}
p_1 \cdot p_3 &= m_{\tau} \left( m_{\tau} - E_2 - E_4 \right) \\
p_2 \cdot p_4 &= \frac{1}{2} \left( -m_{\tau}^2 - m_{\mu}^2 + 2m_{\tau} (E_2 + E_4) \right) \\
p_1 \cdot p_4 &= m_{\tau} E_4 \\
p_2 \cdot p_3 &= \frac{1}{2} \left( m_{\tau}^2 - m_{\mu}^2 - 2m_{\tau} E_4 \right)
\end{split}
\end{equation}
We can rewrite the squared amplitude in terms of the mass parameters after simplifying the summing over the diverse coupling constants to $g_{1,2}$, respectively ($g_1 = g_{\mu \tau}^{2L} g_{\mu \mu}^{2L} + g_{\mu \tau}^{2R} g_{\mu \mu}^{2R}, g_2 = g_{\mu \tau}^{2L} g_{\mu \mu}^{2R} + g_{\mu \tau}^{2R} g_{\mu \mu}^{2L}$).
\begin{equation}
\begin{split}
\frac{1}{2}\sum_{\operatorname{spin}} \lvert \mathcal{M} \left( g_1, g_2, E_2, E_4 \right) \rvert^2 &= \frac{4}{M_Z^4} \Big[ g_1 \left( m_{\tau}^2 - m_{\tau} (E_2 + E_4) \right) \left( -m_{\tau}^2 - m_{\mu}^2 + 2m_{\tau} (E_2 + E_4) \right) \\
&+ g_2 \left( m_{\tau} E_4 \right) \left( m_{\tau}^2 - m_{\mu}^2 - 2m_{\tau} E_4 \right) \Big]
\end{split}
\end{equation}
Now it is time to evaluate the three body phase space integral by turning it into an effective two-body phase integral as follows (we simply drop out the prefactor $1/(2\pi)^5$ for simplicity in this derivation while keeping
the prefactor
in the computation of the aforementioned partial decay width).
\begin{equation}
\begin{split}
&\frac{d^3p_2}{2E_2} \frac{d^3p_3}{2E_3} \frac{d^3p_4}{2E_4} \delta^{4} \left( p_1 - p_2 - p_3 - p_4 \right) = \frac{d^3p_2}{2E_2} d^4p_3 \Theta \left( p_3^0 \right) \delta \left( p_3^2 \right) \frac{d^3p_4}{2E_4} \delta^{4} \left( p_1 - p_2 - p_3 - p_4 \right) \\
&= \frac{d^3p_2}{2E_2} \frac{d^3p_4}{2E_4} \Theta \left( p_1^0 - p_2^0 - p_4^0 \right) \delta \left( (p_1 - p_2 - p_4)^2 \right) \\
&=\frac{d^3p_2}{2E_2} \frac{d^3p_4}{2E_4} \Theta \left( p_1^0 - p_2^0 - p_4^0 \right) \delta \left( p_1^2 -2p_1 \cdot (p_2 + p_4) + (p_2^2 + 2p_2 \cdot p_4 + p_4^2) \right) \\
&= \frac{d^3p_2}{2E_2} \frac{d^3p_4}{2E_4} \Theta \left( p_1^0 - p_2^0 - p_4^0 \right) \delta \left( m_{\tau}^2 +2m_{\mu}^2 -2m_{\tau} (E_2 + E_4) + 2(E_2 E_4 - \lvert \vec{p}_2 \rvert \lvert \vec{p}_4 \rvert \cos\theta) \right) \\
&= \frac{d^3p_2}{2E_2} \frac{d^3p_4}{2E_4} \Theta \left( p_1^0 - p_2^0 - p_4^0 \right) \frac{1}{2\lvert \vec{p}_2 \rvert \lvert \vec{p}_4 \rvert} \delta \left( \frac{m_{\tau}^2 +2m_{\mu}^2 -2m_{\tau} (E_2 + E_4) + 2 E_2 E_4}{2\lvert \vec{p}_2 \rvert \lvert \vec{p}_4 \rvert} - \cos\theta \right)
\end{split}
\end{equation}
From the delta function, we can determine the integration range by assuming $E_2 \approx \lvert \vec{p}_2 \rvert, E_4 \approx \lvert \vec{p}_4 \rvert$. When $\cos\theta = 1$, the obtained result is
\begin{equation}
m_{\tau}^2 + 2m_{\mu}^2 - 2m_{\tau} (E_2 + E_4) = 0
\label{eqn:inside_deltafunction}
\end{equation}
From Equation~\ref{eqn:inside_deltafunction}, the integration range can be read off as follows:
\begin{equation}
\frac{m_{\mu}^2}{m_{\tau}} \leq E_2 \leq \frac{1}{2}m_{\tau}, \quad \frac{1}{2}m_{\tau} + \frac{m_{\mu}^2}{m_{\tau}} - E_2 \leq E_4 \leq \frac{1}{2}m_{\tau}
\end{equation}
It can be easily understood that once one mass parameter $E_2$ is set up by $\frac{1}{2}m_{\tau}$, the energy of the other mass parameters $E_4, E_3$ must be given by $\frac{m_{\mu}^2}{m_{\tau}}, \frac{1}{2}m_{\tau} - \frac{m_{\mu}^2}{m_{\tau}}$, respectively. Then, it remains
to simplify the effective two body phase space integral as follows.
\begin{equation}
d^3p_2 d^3p_4 = 4\pi \lvert \vec{p}_2 \rvert^2 d\lvert \vec{p}_2 \rvert 2\pi \lvert \vec{p}_4 \rvert d\lvert \vec{p}_4 \rvert d\cos\theta, \quad \lvert \vec{p}_2 \rvert d\lvert \vec{p}_2 \rvert = E_2 dE_2, \quad \lvert \vec{p}_4 \rvert d\lvert \vec{p}_4 \rvert = E_4 dE_4
\end{equation}
Putting all pieces
together, the decay width for the CLFV $\tau \rightarrow 3\mu$ decay at tree-level after carrying out the $\cos\theta$ integration is given by:
\begin{equation}
\Gamma\left( \tau \rightarrow 3 \mu \right) = \frac{1}{64 m_{\tau} \pi^3} \int_{m_{\mu}^2/m_{\tau}}^{\frac{1}{2}m_{\tau}} \int_{\frac{1}{2}m_{\tau}+\frac{m_{\mu}^2}{m_{\tau}}-E_2}^{\frac{1}{2}m_{\tau}} \left( \frac{1}{2}\sum_{\operatorname{spin}} \lvert \mathcal{M} \left( g_1, g_2, E_2, E_4 \right) \rvert^2 \right) dE_4 dE_2
\end{equation}
The experimental bound for the $\tau \rightarrow 3\mu$ decay is given by~\Steve{\cite{Hayasaka:2010np}}:
\begin{equation}
\operatorname{BR}\left( \tau \rightarrow 3\mu \right)_{\operatorname{EXP}} = 2.1 \times 10^{-8}
\end{equation}
\subsection{Analytic expression for $Z \rightarrow \mu \tau$ decay}
The last FCNC constraint we discuss is the $Z \rightarrow \mu \tau$ decay and diagrams contributing to the $Z \rightarrow \mu \tau$ decay are given in Figure~\ref{fig:diagrams_Zmutau}.
\begin{figure}[H]
\centering
\includegraphics[keepaspectratio,width=\textwidth]{Zmutau_Feynman1.png}
\caption{Diagrams contributing to the charged lepton flavor violation (CLFV) $Z\tau\mu$ decay at tree-level}
\label{fig:diagrams_Zmutau}
\end{figure}
As in the CLFV $\tau \rightarrow 3\mu$ decay mediated by the SM $Z$ gauge boson, this CLFV $Z\rightarrow \mu\tau$ decay is also a new effect and it requires to derive its appropriate prediction from the ground. We can read off the invariant amplitude for each diagram given in Figure~\ref{fig:diagrams_Zmutau}. We refer to the left diagram as $\mathcal{M}_L$ and the right as $\mathcal{M}_R$. Then, the amplitudes are written as follows:
\begin{equation}
\begin{split}
i\mathcal{M}_L &= i (g_L)_{\tau\mu} \epsilon_\mu(p_1) \overline{u} (p_2) \gamma^\mu P_L v (p_3), \\[1ex]
i\mathcal{M}_R &= i (g_R)_{\tau\mu} \epsilon_\mu(p_1) \overline{u} (p_2) \gamma^\mu P_R v (p_3).
\label{eqn:amplitude_for_each_Zmutau}
\end{split}
\end{equation}
In order to have the squared amplitude averaged and summed, we square the amplitude given in Equation~\ref{eqn:amplitude_for_each_Zmutau} as follows:
\begin{equation}
\begin{split}
\frac{1}{3}\sum_{\operatorname{spin}} \lvert \mathcal{M}_L \rvert^2 &= \frac{1}{3} (g_L)_{\tau\mu}^2 \left( -g_{\mu\nu} + \frac{p_{1\mu}p_{1\nu}}{M_Z^2} \right) \operatorname{Tr} \left[ (\slashed{p}_2+m_{\mu}) \gamma^\mu P_L (\slashed{p}_3-m_{\tau}) \gamma^\nu P_L) \right] \\
&\simeq \frac{2}{3} (g_L)_{\tau\mu}^2 M_Z^2 \\
\frac{1}{3}\sum_{\operatorname{spin}} \lvert \mathcal{M}_R \rvert^2 &= \frac{1}{3} (g_R)_{\tau\mu}^2 \left( -g_{\mu\nu} + \frac{p_{1\mu}p_{1\nu}}{M_Z^2} \right) \operatorname{Tr} \left[ (\slashed{p}_2+m_{\mu}) \gamma^\mu P_R (\slashed{p}_3-m_{\tau}) \gamma^\nu P_R) \right] \\
&\simeq \frac{2}{3} (g_R)_{\tau\mu}^2 M_Z^2 \\
\frac{1}{3} \sum_{\operatorname{spin}} \lvert \mathcal{M} \rvert^2 &= \frac{1}{3} \sum_{\operatorname{spin}} \left( \lvert \mathcal{M}_L \rvert^2 + \lvert \mathcal{M}_R \rvert^2 \right) = \frac{2}{3}\left( (g_L)_{\tau\mu}^2 + (g_R)_{\tau\mu}^2 \right) M_Z^2
\end{split}
\end{equation}
Then, the decay rate equation is given by:
\begin{equation}
\begin{split}
d\Gamma\left( Z\rightarrow\mu\tau \right) &= \frac{1}{2M_Z} \frac{d^3p_2}{(2\pi)^3 2E_2} \frac{d^3p_3}{(2\pi)^3 2E_3} \lvert \mathcal{M} \rvert^2 (2\pi)^4 \delta^{(4)} \left( p_1 - p_2 - p_3 \right) \\
\Gamma\left( Z\rightarrow\mu\tau \right) &= \frac{\lvert p^* \rvert}{32\pi^2 M_Z^2} \int \lvert \mathcal{M} \rvert^2 d\Omega \\
&= \frac{M_Z}{24\pi} \left( (g_L)_{\tau\mu}^2 + (g_R)_{\tau\mu}^2 \right)
\end{split}
\end{equation}
where $p^* \simeq M_Z/2$. Then, we are ready to write our prediction for the branching ratio of $Z\mu\tau$ decay at tree-level
\begin{equation}
\operatorname{BR}\left(Z \rightarrow \mu\tau \right) = \frac{\Gamma\left( Z \rightarrow \mu\tau \right)}{\Gamma_{Z}} \simeq \frac{1}{2.5} \frac{M_Z}{24\pi} \left( (g_L)_{\tau\mu}^2 + (g_R)_{\tau\mu}^2 \right),
\end{equation}
where $\Gamma_{Z}$ is the total decay width of the SM $Z$ gauge boson ($\Gamma_{Z} \simeq 2.5\operatorname{GeV}$). The experimental bound of the CLFV $Z \rightarrow \mu\tau$ decay is known as~\Steve{\cite{DELPHI:1996iox}}:
\begin{equation}
\operatorname{BR}\left( Z \rightarrow \mu\tau \right)_{\operatorname{EXP}} = 1.2 \times 10^{-5}
\end{equation}
\subsection{Numerical analysis for each prediction in the charged lepton sector}
We have discussed some
relevant CLFV decay modes such as the $\tau \rightarrow \mu \gamma, \tau \rightarrow 3\mu$ and $Z \rightarrow \mu\tau$ from a theoretical point of view.
By defining the renormalizable flavor violating interactions
we showed that it is possible for the new physics to arise in a simple scenario thanks to the presence of vector-like charged leptons, which play a crucial role for these CLFV decay modes to happen. It is an encouraging feature that the mass range of the vector-like charged leptons can be constrained by the experimental bound of the CLFV decays and numerical scans for this feature will be discussed
in detail in the following subsection.
\subsubsection{Free parameter setup}
For the numerical scan for the charged lepton sector, we first proceed to
set up a possible mass range of the mass parameters of Equation~\ref{eqn:cl_2} \Steve{in Table~\ref{tab:parameter_region_initial_scan}}.
\begin{center}
{\renewcommand{\arraystretch}{1.5}
\begin{tabular}{cc}
\toprule
\toprule
\textbf{Mass parameter} & \textbf{Scanned Region($\operatorname{GeV}$)} \\
\midrule
$y_{24}^{e} v_{d} = m_{24}$ & $\pm [1,10]$ \\
$y_{34}^{e} v_{d} = m_{34}$ & $\pm [1,10]$ \\
$y_{43}^{e} v_{d} = m_{43}$ & $\pm [1,10]$ \\
$x_{34}^{L} v_{\phi} = m_{35}$ & $\pm [50,200]$ \\
$x_{42}^{e} v_{\phi} = m_{52}$ & $\pm [50,200]$ \\
$x_{43}^{e} v_{\phi} = m_{53}$ & $\pm [50,200]$ \\
$M_{45}^{L}$ & $\pm [150,2000]$ \\
$M_{54}^{e}$ & $\pm [150,2000]$ \\
\bottomrule
\bottomrule
\end{tabular}
\captionof{table}{Initial parameter setup for scanning the mass of the vector-like charged leptons} \label%
{tab:parameter_region_initial_scan}}
\end{center}
There are a few of features to be noticed before we start the numerical scan.
\begin{enumerate}
\item We assumed a vev for the SM up-type Higgs $H_u$
very close to $246\operatorname{GeV}$, whereas the one
of the SM down-type Higgs $H_d$ is assumed to be very small compared to the $v_u = \langle H_u \rangle$ and is ranged from $1$ to $10\operatorname{GeV}$. The two vevs hold the relation $v_u^2 + v_d^2 = (246\operatorname{GeV})^2$. We made that assumption since we are considering an scenario close to the decoupling limit where the neutral CP even part of $H_u$ is mostly identified with the $126$ GeV SM like Higgs boson.
\item As the vev of the singlet flavon $\phi$ is a free parameter, we varied it in the
range $[50,200]\operatorname{GeV}$ whereas the mass parameters $m_{35,52,53}$ were varied in a range of values consistent
with the observed hierarchical structure of the charged lepton masses. Furthermore,
the vector-like masses are also other free parameters assumed to be larger or equal than
$150\operatorname{GeV}$ in order to successfully fulfill the
experimental bounds on exotic charged lepton masses.
\item What we need to constrain in this numerical scan is the predicted muon and tau masses as well as the $23$ mixing angle. For the muon and tau masses, we required that the
obtained values of the muon and tau masses to be in the range
$[1 \pm 0.1] \times m_{\mu,\tau}$. Considering that the sizeable off-diagonal elements of the PMNS mixing matrix mainly arise from the neutrino sector, all mixing angles in the charged lepton sector are required to be
as small as possible and
thus we limit them to be lower than $0.2$.
\end{enumerate}
\subsubsection{Numerical scan result for the charged lepton sector}
The scanned mass range of the vector-like charged leptons are shown
in Figure~\ref{fig:cl_CLFV_decays}.
\begin{figure}[H]
\centering
\begin{subfigure}{0.48\textwidth}
\includegraphics[keepaspectratio,width=0.95\textwidth]{ME4BRtmgMtE47.jpg}
\end{subfigure}
\begin{subfigure}{0.48\textwidth}
\includegraphics[keepaspectratio,width=0.95\textwidth]{MtE4BRtmgME47.jpg}
\end{subfigure}
\begin{subfigure}{0.48\textwidth}
\includegraphics[keepaspectratio,width=0.95\textwidth]{ME4MtE4BRtmg3.jpg}
\end{subfigure}
\begin{subfigure}{0.48\textwidth}
\includegraphics[keepaspectratio,width=0.95\textwidth]{ME4MtE4BRt3m3.jpg}
\end{subfigure}
\begin{subfigure}{0.48\textwidth}
\includegraphics[keepaspectratio,width=0.95\textwidth]{ME4MtE4BRZmt3.jpg}
\end{subfigure}
\caption{Scanned mass region of the vector-like charged leptons and contributions of the flavor violating interactions with the SM $Z$ gauge boson to the diverse CLFV decays $\tau \rightarrow \mu \gamma, \tau \rightarrow 3\mu$ and $Z \rightarrow \mu \tau$. The used constraints are the predicted muon and tau mass to be put between $[1 \pm 0.1] \times m_{\mu,\tau}$ and the $23$ mixing angle to be less than $0.2$. The darker blue region appearing in each diagram means either the singlet or doublet vector-like masses $M_{\widetilde{E}_4}, M_{E_4}$ are excluded up to $200\operatorname{GeV}$ by \Steve{reference~\cite{Xu:2018pnq,Guedes:2021oqx}}. The brighter blue region means the doublet vector-like mass $M_{E_4}$ is excluded up to $790\operatorname{GeV}$ by the CMS~\cite{Bhattiprolu:2019vdu,CMS:2019hsm}. The brighter red region appearing in above two plots is the region excluded by the experimental bound for the $\operatorname{BR}\left( \tau \rightarrow \mu \gamma \right)_{\operatorname{EXP}} = 4.4 \times 10^{-8}$.}
\label{fig:cl_CLFV_decays}
\end{figure}
The first we need to discuss is the experimental bounds for the vector-like charged leptons appearing in Figure~\ref{fig:cl_CLFV_decays}. The darker blue region is the excluded region for the vector-like charged leptons by $200\operatorname{GeV}$~\Steve{\cite{Xu:2018pnq,Guedes:2021oqx}}. The vector-like mass $M_{E_4}$ consists of the doublet vector-like charged leptons $E_{4L},\widetilde{E}_{4R}$, whereas the other vector-like mass $M_{\widetilde{E}_4}$ consists of the singlet vector-like charged leptons $\widetilde{E}_{4L}, E_{4R}$. Therefore, $M_{E_4}$ is the doublet vector-like mass, whereas $M_{\widetilde{E}_4}$ is the singlet vector-like mass, and the doublet vector-like mass is excluded by CMS up to $790\operatorname{GeV}$~\cite{Bhattiprolu:2019vdu,CMS:2019hsm}, expressed by the brighter blue region of Figure~\ref{fig:cl_CLFV_decays}. The second is our predictions for the branching ratio of $\tau \rightarrow \mu \gamma$ in Figure~\ref{fig:cl_CLFV_decays}. The relevant experimental bound for each CLFV decay is given by:
\begin{equation}
\begin{split}
&\operatorname{BR}\left( \tau \rightarrow \mu \gamma \right)_{\operatorname{EXP}} = 4.4 \times 10^{-8}
\\[1ex]
&-
\operatorname{Log}_{10} \operatorname{BR} \left( \tau \rightarrow \mu \gamma \right)_{\operatorname{EXP}} \simeq -7.4
\\[1ex]
&\operatorname{BR}\left( \tau \rightarrow 3\mu \right)_{\operatorname{EXP}} = 2.1 \times 10^{-8}
\\[1ex]
&- \operatorname{Log}_{10} \operatorname{BR} \left( \tau \rightarrow 3\mu \right)_{\operatorname{EXP}} \simeq -7.7
\\[1ex]
&\operatorname{BR}\left( Z \rightarrow \mu \tau \right)_{\operatorname{EXP}} = 1.2 \times 10^{-5}
\\[1ex]
&- \operatorname{Log}_{10} \operatorname{BR} \left( Z \rightarrow \mu \tau \right)_{\operatorname{EXP}} \simeq -4.9
\end{split}
\end{equation}
Our predictions for the CLFV $\tau \rightarrow 3\mu$ and $Z \rightarrow \mu \tau$ decays are not excluded by the experimental bound, however those are not the case for the CLFV decay $\tau \rightarrow \mu \gamma$,
which exceed its upper experimental bound in some parts of the parameter space. This is due to, in some parts of the parameter space, the dominant contributions to the $\tau \rightarrow \mu \gamma$ decay involving a charged exotic lepton as well as chirality flip in the internal line and proportional to $M/m_{\tau}$ because of the sizeable large value of the charged exotic lepton - SM charged lepton mass ratio.
After removing all excluded points by the experimental bound of the $\tau \rightarrow \mu \gamma$ decay,
we obtain Figure~\ref{fig:cl_CLFV_second}.
\begin{figure}[H]
\centering
\begin{subfigure}{0.48\textwidth}
\includegraphics[keepaspectratio,width=0.95\textwidth]{ME4BRtmgMtE49.jpg}
\end{subfigure}
\begin{subfigure}{0.48\textwidth}
\includegraphics[keepaspectratio,width=0.95\textwidth]{MtE4BRtmgME49.jpg}
\end{subfigure}
\begin{subfigure}{0.48\textwidth}
\includegraphics[keepaspectratio,width=0.95\textwidth]{ME4MtE4BRtmg5.jpg}
\end{subfigure}
\begin{subfigure}{0.48\textwidth}
\includegraphics[keepaspectratio,width=0.95\textwidth]{ME4MtE4BRt3m5.jpg}
\end{subfigure}
\begin{subfigure}{0.48\textwidth}
\includegraphics[keepaspectratio,width=0.95\textwidth]{ME4MtE4BRZmt5.jpg}
\end{subfigure}
\caption{Reduced number of numerical predictions. The numerical predictions are constrained by the experimental value of the branching ratio of $\tau \rightarrow \mu \gamma$ decay. None of them are constrained by the branching ratio of $\tau \rightarrow 3\mu$ and $Z \rightarrow \mu \tau$ experimental bounds.}
\label{fig:cl_CLFV_second}
\end{figure}
Looking at our numerical predictions for the branching ratio of the $\tau \rightarrow \mu \gamma$ decay shown in Figure~\ref{fig:cl_CLFV_second}, some
of them are constrained by the experimental limit of this
branching ratio, however most of them survive, which implies that
our numerical predictions for the branching ratio of $\tau \rightarrow \mu \gamma$ are not significantly constrained by its experimental bound. Furthermore,
none of the numerical predictions for the branching ratios of the $\tau \rightarrow 3\mu$ and $Z \rightarrow \mu \tau$ decays
are constrained by its experimental bound. However, our numerical predictions for the different
CLFV decays can significantly be constrained by the future LHC upgrades having higher center of mass energy and luminosity than the ones of the current LHC, which will allow to set tighest constraints on charged exotic vector-like masses, thus leading to stronger constraints on the model parameter space.
Regarding the CLFV $Z \rightarrow \mu \tau$ decay, the FCC-ee experiment has planned to generate $10^{12}$ SM $Z$ gauge bosons, which will allow to probe our model since
the branching ratio of the $Z \rightarrow \mu \tau$ decay can reach values of the order of $10^{-10}$ in the allowed region of the parameter space. Thus, the $Z \rightarrow \mu \tau$ decay is within the reach of the FCC-ee experiment, whose
$Z$ factory
\cite{DeRomeri:2016gum,Calibbi:2021pyh} will be crucial to verify or rule out this model. Concluding this subsection, our numerical predictions are not significantly constrained by any of the CLFV $\tau \rightarrow \mu \gamma, \tau \rightarrow 3\mu$ and $Z \rightarrow \mu \tau$ decays and might be able to be seen from the $Z$ factory for the first time, predicting the doublet vector-like charged lepton mass which is ranged from $790$ to $1600\operatorname{GeV}$ whereas the singlet vector-like charged lepton mass is ranged from $500$ to $2000\operatorname{GeV}$ or above than that.
\section{QUARK SECTOR PHENOMENOLOGY}
\label{sec:VI}
We have discussed the up- and down-type quark mass matrices pointed out that they have a different form, since the quark doublet rotation used in the up-type quark sector can not remove the down-type Yukawa term. This difference between up- and down-type quark mass matrices cause a distinct feature for each sector as follows:
\begin{itemize}
\item The up-type quark mass matrix can reach to the $23$ left (right)-handed mixing.
\item The down-type quark mass matrix can access to all left-handed mixings among the three SM generations, whereas the right-handed mixing can only have the $23$ mixing.
\end{itemize}
The interesting feature of the down-type quark mass matrix allows for flavor changing $Z$ interactions with down type quarks which yield
neutral meson oscillations such as $K, B_d$ and $B_s$. Furthermore, an important feature to be mentioned is that the
first generation of SM charged fermions do not acquire masses.
Due to this property, our predictions for the neutral meson oscillations including the $d$ quark give a very suppressed energy difference corresponding to $10^{-40} \operatorname{GeV}$, which is impossible to reach with the current experimental sensitivity. Then, the rest of the neutral meson oscillation $B_s$ is possible and an encouraging feature of the $B_s$ meson oscillation in our proposed model
is the $B_s$ meson oscillation mediated by the SM $Z$ gauge boson can be calculated at tree-level as given in Figure~\ref{fig:BsBbars_mixing}.
\begin{figure}[H]
\centering
\includegraphics[keepaspectratio,width=\textwidth]{BsBbars_Feynman1.png}
\caption{Feynman diagrams contributing to the
the $B_s-\bar{B}_s$ meson mixing involving the
tree-level exchange of the $Z$ gauge boson in the polarized basis}
\label{fig:BsBbars_mixing}
\end{figure}
From our numerical analysis we have found that the BSM contributions to the
$B_s$ meson oscillation arising from
the tree-level exchange of the
$Z$ gauge boson yield the meson mass splitting of the order of
$10^{-15}$ GeV or even less than that value, which is quite negligible compared to its corresponding
experimental bound of $10^{-11} \operatorname{GeV}$. The very suppressed
new physics effect for the $B_s$ meson oscillation can be explained by considering
the flavor violating coupling constants at each vertex of each diagram, whose value is about $10^{-6,-7}$ and this values are determined by
the ratio between Yukawa and vector-like masses. For this reason, in the study of the phenomenological implications of our model in the flavor changing neutral interactions in the quark sector, we do not consider
the neutral meson oscillations as well as the $B_s \rightarrow \mu^+\mu^-$ decay. It is worth mentioning that the $B_s \rightarrow \mu^+\mu^-$ decay gives weaker effects than the neutral meson oscillations. Considering these facts, we conclude that the rare $t \rightarrow c Z$ decay and the CKM mixing matrix can constrain the quark sector of our
model, and thus
we discuss these two phenomenological aspects in the following subsections.
\subsection{Analytic expression for the $t \rightarrow c Z$ decay}
The $t \rightarrow c Z$ decay, which only appears at one-loop level in the SM,
can take place at tree-level in our proposed model, thanks to the $Z$ mediated flavor changing neutral current interactions in the quark sector. In our proposed model, the $t \rightarrow cZ$ receives tree-level contributions which are depicted in
Figure~\ref{fig:tcZ}.
\begin{figure}[H]
\centering
\includegraphics[keepaspectratio,width=\textwidth]{tcZ_Feynman1.png}
\caption{tree-level Feynman diagrams contributing to the rare $t \rightarrow c Z$ decay}
\label{fig:tcZ}
\end{figure}
Denoting the invariant amplitudes for the Feynman diagrams of the left and right panels of Figure~\ref{fig:tcZ} as $\mathcal{M}_L$ and $\mathcal{M}_R$, respectively, we find that they can be written as:
\begin{equation}
\begin{split}
i\mathcal{M}_L &= i (g_L)_{tc} \epsilon_\mu^* (p_3) \overline{u} (p_2) \gamma^\mu P_L u (p_1) \\
i\mathcal{M}_R &= i (g_R)_{tc} \epsilon_\mu^* (p_3) \overline{u} (p_2) \gamma^\mu P_R u (p_1)
\label{eqn:amplitude_for_each_tcZ}
\end{split}
\end{equation}
In order to have the squared amplitude averaged and summed, we square the amplitudes given in Equation~\ref{eqn:amplitude_for_each_tcZ} and then sum over the different spin states, as follows:
\begin{equation}
\begin{gathered}
\frac{1}{2}\sum_{\operatorname{spin}} \lvert \mathcal{M}_L \rvert^2 = (g_L)_{ct}^2 \left( (p_2 \cdot p_1) + \frac{2}{M_Z^2} (p_2 \cdot p_3)(p_1 \cdot p_3) \right) \\
\frac{1}{2}\sum_{\operatorname{spin}} \lvert \mathcal{M}_R \rvert^2 = (g_R)_{ct}^2 \left( (p_2 \cdot p_1) + \frac{2}{M_Z^2} (p_2 \cdot p_3)(p_1 \cdot p_3) \right) \\
\frac{1}{2} \sum_{\operatorname{spin}} \lvert \mathcal{M} \rvert^2 = \frac{1}{2} \sum_{\operatorname{spin}} \left( \lvert \mathcal{M}_L \rvert^2 + \lvert \mathcal{M}_R \rvert^2 \right) = \frac{1}{2}\left( (g_L)_{ct}^2 + (g_R)_{ct}^2 \right) \left( (p_2 \cdot p_1) + \frac{2}{M_Z^2} (p_2 \cdot p_3)(p_1 \cdot p_3) \right)
\end{gathered}
\end{equation}
Then, the decay rate equation is given by:
\begin{equation}
\begin{split}
d\Gamma\left( t \rightarrow c Z \right) &= \frac{1}{2m_t} \frac{d^3p_2}{(2\pi)^3 2E_2} \frac{d^3p_3}{(2\pi)^3 2E_3} \lvert \mathcal{M} \rvert^2 (2\pi)^4 \delta^{(4)} \left( p_1 - p_2 - p_3 \right) \\
\Gamma\left( t \rightarrow c Z \right) &= \frac{\lvert p^* \rvert}{32\pi^2 m_t^2} \int \lvert \mathcal{M} \rvert^2 d\Omega \\
&= \frac{1}{8\pi m_t^2} \frac{1}{2m_t}(m_t^2 - M_Z^2) ( (g_L)_{ct}^2 + (g_R)_{ct}^2 ) \\
&\times \left[ \frac{m_t^2 + m_c^2 - M_Z^2}{2} + \frac{2}{M_Z^2} (\frac{m_t^2 - m_c^2 - M_Z^2}{2}) (\frac{m_t^2 - m_c^2 + M_Z^2}{2}) \right]
\end{split}
\end{equation}
where $p^* \simeq \frac{1}{2m_t} \left( m_t^2 - M_Z^2 \right)$.
Then, we are ready to write down our prediction for the branching ratio of the tree-level $t \rightarrow c Z$ decay~\Steve{\cite{ATLAS:2018zsq}}:
\begin{equation}
\operatorname{BR}\left(t \rightarrow c Z \right) = \frac{\Gamma\left( t \rightarrow c Z \right)}{\Gamma_{t}} < \operatorname{BR} \left( t \rightarrow c Z \right)_{\operatorname{EXP}} = 2.4 \times 10^{-4} \quad (95\% \operatorname{CL})
\label{eqn:BRtcZ_exp}
\end{equation}
where $\Gamma_t = 1.32 \operatorname{GeV}$.
\subsection{Analytic expression for the CKM mixing matrix}
In order to discuss the CKM mixing matrix, the first task we need to investigate is the $W$ current of the SM in order to see how the CKM mixing matrix can take place (we only consider the three SM generations at the moment).
\begin{equation}
\begin{split}
\mathcal{L}_{\operatorname{SM}}^{W} &= g j_{\mu}^{W+} W^{\mu +} = \frac{g}{\sqrt{2}} \left( \overline{u}_{L}^{i} \gamma_{\mu} d_{L}^{i} \right) W^{\mu +}
\\
&= \frac{g}{\sqrt{2}} \left( \overline{u}_{L}^{i} (V_{L}^{u\dagger} V_{L}^{u}) \gamma_{\mu} (V_{L}^{d\dagger} V_{L}^{d}) d_{L}^{i} \right) W^{\mu +} = \frac{g}{\sqrt{2}} \left( \overline{u}_{L}^{i \prime} \gamma_{\mu} (V_{\operatorname{CKM}}) d_{L}^{i \prime} \right) W^{\mu +}
\end{split}
\end{equation}
where $u_{L}^{i \prime}, d_{L}^{i \prime}$ are the up- and down-type quarks of the SM in the mass basis and $V_{\operatorname{CKM}}$ is the CKM mixing matrix defined as $V_{L}^{u} V_{L}^{d \dagger}$. Now we extend the quark spectrum by considering the vector-like quarks, thus implying that
the $W$ current takes the form:
\begin{equation}
\begin{split}
\mathcal{L}^{W} &=
\frac{g}{\sqrt{2}}
\begin{pmatrix}
\overline{u}_{1L} & \overline{u}_{2L} & \overline{u}_{3L} & \overline{u}_{4L} & \overline{\widetilde{u}}_{4L}
\end{pmatrix}
\gamma_{\mu}
\begin{pmatrix}
1 & 0 & 0 & 0 & 0 \\
0 & 1 & 0 & 0 & 0 \\
0 & 0 & 1 & 0 & 0 \\
0 & 0 & 0 & 1 & 0 \\
0 & 0 & 0 & 0 & 0
\end{pmatrix}
\begin{pmatrix}
d_{1L} \\
d_{2L} \\
d_{3L} \\
d_{4L} \\
\widetilde{d}_{4L}
\end{pmatrix}
W^{\mu +}
\\
&=
\frac{g}{\sqrt{2}}
\begin{pmatrix}
\overline{u}_{1L} & \overline{u}_{2L} & \overline{u}_{3L} & \overline{u}_{4L} & \overline{\widetilde{u}}_{4L}
\end{pmatrix}
V_{L}^{u \dagger} V_{L}^{u}
\gamma_{\mu}
\begin{pmatrix}
1 & 0 & 0 & 0 & 0 \\
0 & 1 & 0 & 0 & 0 \\
0 & 0 & 1 & 0 & 0 \\
0 & 0 & 0 & 1 & 0 \\
0 & 0 & 0 & 0 & 0
\end{pmatrix}
V_{L}^{d \dagger} V_{L}^{d}
\begin{pmatrix}
d_{1L} \\
d_{2L} \\
d_{3L} \\
d_{4L} \\
\widetilde{d}_{4L}
\end{pmatrix}
W^{\mu +}
\\
&=
\frac{g}{\sqrt{2}}
\begin{pmatrix}
\overline{u}_{L} & \overline{c}_{L} & \overline{t}_{L} & \overline{U}_{4L} & \overline{\widetilde{U}}_{4L}
\end{pmatrix}
\gamma_{\mu}
V_{L}^{u}
\begin{pmatrix}
1 & 0 & 0 & 0 & 0 \\
0 & 1 & 0 & 0 & 0 \\
0 & 0 & 1 & 0 & 0 \\
0 & 0 & 0 & 1 & 0 \\
0 & 0 & 0 & 0 & 0
\end{pmatrix}
V_{L}^{d \dagger}
\begin{pmatrix}
d_{L} \\
s_{L} \\
b_{L} \\
D_{4L} \\
\widetilde{D}_{4L}
\end{pmatrix}
W^{\mu +}
\end{split}
\label{eqn:CKM_analytic}
\end{equation}
where the CKM mixing prediction in our model is given by
\begin{equation}
V_{\operatorname{CKM}} =
V_{L}^{u}
\begin{pmatrix}
1 & 0 & 0 & 0 & 0 \\
0 & 1 & 0 & 0 & 0 \\
0 & 0 & 1 & 0 & 0 \\
0 & 0 & 0 & 1 & 0 \\
0 & 0 & 0 & 0 & 0
\end{pmatrix}
V_{L}^{d \dagger}, \quad \RIfM@\expandafter\text@\else\expandafter\mbox\fi{for the upper-left $3 \times 3$ block}
\label{eqn:CKM_prediction}
\end{equation}
where $V_{L}^{u}$ is the mixing matrix for the up-type quarks defined in Equation~\ref{eqn:up_mixing} and $V_{L}^{d}$ is the one corresponding to
the down-type quarks given in Equation~\ref{eqn:down_mixing}. The zero appearing in the middle matrix between $V_{L}^{u}$ and $V_{L}^{d \dagger}$ arises from the left-handed vector-like quark singlets $\widetilde{U}_{4L}$ and $\widetilde{D}_{4L}$ which do not interact with the $W$ currents, so our prediction for the CKM mixing matrix does not feature the unitarity requirement and this leads to the need of relaxing the unitarity constraint of the CKM quark mixing matrix. That deviation of unitarity of the CKM quark mixing matrix is due to the presence of heavy vector-like quarks and this aspect was studied in
\cite{Branco:2021vhs} in the context of theory with different particle spectrum and symmetry than ours.
Few sigma of SM deviations from the first row of the CKM mixing matrix without unitarity were analyzed in \cite{Branco:2021vhs}.
The deviation from the Unitarity of the CKM will also be discussed in our numerical result and the experimental CKM mixing matrix without unitarity is given by~\cite{Branco:2021vhs,ParticleDataGroup:2020ssz}:
\begingroup
\setlength\arraycolsep{5pt}
\begin{equation}
\lvert K_{\operatorname{CKM}} \rvert
=
\begin{pmatrix}
0.97370 \pm 0.00014 & 0.22450 \pm 0.00080 & 0.00382 \pm 0.00024 \\[1.5ex]
0.22100 \pm 0.00400 & 0.98700 \pm 0.01100 & 0.04100 \pm 0.00140 \\[1.5ex]
0.00800 \pm 0.00030 & 0.03880 \pm 0.00110 & 1.01300 \pm 0.03000
\end{pmatrix}
\end{equation}
\endgroup
\subsection{Numerical analysis for each prediction in the quark sector}
When compared to the charged lepton sector simulation, the quark sector becomes much more complicated
since we need
to fit the masses of the $c,t,s$ and $b$ quarks simultaneously as well as the CKM mixing matrix without imposing the unitarity requirement. Therefore, we fit the masses of the four quarks first by using a fitting function $\chi_{\operatorname{mass}}^{2}$ and then we start a second fitting procedure by using another fitting function $\chi_{\operatorname{CKM}}^{2}$ and this will be discussed in detail
in the following subsections.
\subsubsection{The fitting function $\chi^2$ and free parameter setup}
We set up our parameter region as follows:
\begin{center}
{\renewcommand{\arraystretch}{1.5}
\begin{tabular}{cc}
\toprule
\toprule
\textbf{Mass parameter} & \textbf{Scanned Region($\operatorname{GeV}$)} \\
\midrule
$y_{24}^{u} v_{u} = m_{24}^{u}$ & $\pm [10,50]$ \\[0.5ex]
$y_{34}^{u} v_{u} = m_{34}^{u}$ & $\pm [200,400]$ \\[0.5ex]
$y_{43}^{u} v_{u} = m_{43}^{u}$ & $\pm [200,400]$ \\[0.5ex]
$x_{34}^{Q} v_{\phi} = m_{35}^{Q}$ & $m_{35}^{Q}$ \\[0.5ex]
$x_{42}^{u} v_{\phi} = m_{52}^{u}$ & $\pm [500,700]$ \\[0.5ex]
$x_{43}^{u} v_{\phi} = m_{53}^{u}$ & $\pm [50,500]$ \\[0.5ex]
$M_{45}^{Q}$ & $M_{45}^{Q}$ \\[0.5ex]
$M_{54}^{u}$ & $\pm [1000,3000]$ \\[0.5ex]
\midrule
$y_{14}^{d} v_{d} = m_{14}^{d}$ & $\pm [1,10]$ \\[0.5ex]
$y_{24}^{d} v_{d} = m_{24}^{d}$ & $\pm [5,20]$ \\[0.5ex]
$y_{34}^{d} v_{d} = m_{34}^{d}$ & $\pm [10,30]$ \\[0.5ex]
$y_{43}^{d} v_{d} = m_{43}^{d}$ & $\pm [5,10]$ \\[0.5ex]
$x_{34}^{Q} v_{\phi} = m_{35}^{Q}$ & $\pm [10,100]$ \\[0.5ex]
$x_{42}^{d} v_{\phi} = m_{52}^{d}$ & $\pm [10,100]$ \\[0.5ex]
$x_{43}^{d} v_{\phi} = m_{53}^{d}$ & $\pm [10,100]$ \\[0.5ex]
$M_{45}^{Q}$ & $\pm [1000,3000]$ \\[0.5ex]
$M_{54}^{d}$ & $\pm [1000,3000]$ \\[0.5ex]
\bottomrule
\bottomrule
\end{tabular}
\captionof{table}{Initial parameter setup for scanning mass of the vector-like quarks}
\label{tab:parameter_region_initial_quarks}}
\end{center}
There are a few things to be noticed as in the charged lepton case.
\begin{enumerate}
\item The relation $v_{u}^{2} + v_{d}^{2} = (246\operatorname{GeV})^2$ still holds and the mass parameters $m_{24,34,43}^{u}$ can not exceed the upper perturbative limit on
the Yukawa constant $\sqrt{4\pi} \simeq 3.54$ multiplied by the vev
$\approx 240\operatorname{GeV}$,
of the $H_{u}$ Higgs, thus yielding the bound of $850\operatorname{GeV}$ for these mass parameters. These restrictions have been taken into account through the whole fitting process.
\item The down-type Higgs $H_{d}$ has a very small vev, which is about order of $10\operatorname{GeV}$, based on our previous analysis
~\cite{Hernandez:2021tii}, and the range of values of the mass parameters $m_{14,24,34,43}^{d}$ are considered under this assumption.
\item Since we do not know the correct scale of $v_{\phi}$, we considered
$m_{35}^{Q}, m_{52,53}^{u}$ and $m_{52,53}^{d}$ as free parameters. For the same reason, the vector-like masses $M_{45}^{Q}$ and $M_{54}^{u,d}$ are considered free parameters as well.
\item The mass parameters $m_{35}^{Q}$ and $M_{45}^{Q}$ appear in a
common term shared by both up- and down-type quark sector mass matrices and this feature has been discussed in the paragraph below
Equation~\ref{eqn:diff_up_down}.
\end{enumerate}
The next thing to do is to set up the two fitting functions $\chi_{\operatorname{mass}}^{2}$ and $\chi_{\operatorname{CKM}}^{2}$ as follows:
\begin{equation}
\chi_{\operatorname{mass}}^2 = \sum_{f=c,t,s,b} = \frac{(m_{f}^{\operatorname{pred}}-m_{f}^{\operatorname{EXP}})^2}{(\delta m_{f}^{\operatorname{EXP}})^2}, \quad \chi_{\operatorname{CKM}}^2 = \sum_{i,j=1,2,3} \frac{((V_{\operatorname{CKM}}^{\operatorname{pred}})_{ij} - (V_{\operatorname{CKM}}^{\operatorname{EXP}})_{ij})^2}{((\delta V_{\operatorname{CKM}}^{\operatorname{EXP}})_{ij})^2},
\end{equation}
where the superscript $\operatorname{pred}$ means our prediction to its experimental value and the delta means error bar of the physical quantity at $1\sigma$. Our first goal is to fit the masses of the four quarks simultaneously. For the charged lepton case, we require that our obtained muon and tau masses to be in the range
$[1 \pm 0.1] \times m_{\mu,\tau}$ and this requirement is also imposed for
the $c,s$ and $b$ quarks excepting for the $t$ quark since the $t$ quark is too heavy. Besides that, we require that the obtained top quark mass
to be in the range
$[1 \pm 0.01] \times m_{t}$ instead of $[1 \pm 0.1] \times m_{t}$. After the mass parameters have been converged to be put between the arranged range of each quark mass,
we use the other fitting function $\chi_{\operatorname{CKM}}^{2}$ to fit our prediction for the CKM mixing matrix. Using one of the defined fitting functions, we need to vary the mass parameters of Table~\ref{tab:parameter_region_initial_quarks} by a factor of $\left[ 1 \pm \kappa \right]$ where $\kappa = 0.1$ in order to find better mass parameters. We rename the given parameters from the initial parameter setup by adding an subscript $r$ to the mass parameters. The varied mass parameters are given in Table~\ref{tab:parameter_region_initial_q_second}.
\begin{center}
{\renewcommand{\arraystretch}{1.5}
\begin{tabular}{cc}
\toprule
\toprule
\textbf{Mass parameter} & \textbf{Scanned Region($\operatorname{GeV}$)} \\
\midrule
$y_{24}^{u} v_{u} = m_{24}^{u}$ & $\left[ 1 \pm \kappa \right] \times m_{24r}^{u}$ \\[0.5ex]
$y_{34}^{u} v_{u} = m_{34}^{u}$ & $\left[ 1 \pm \kappa \right] \times m_{34r}^{u}$ \\[0.5ex]
$y_{43}^{u} v_{u} = m_{43}^{u}$ & $\left[ 1 \pm \kappa \right] \times m_{43r}^{u}$ \\[0.5ex]
$x_{34}^{Q} v_{\phi} = m_{35}^{Q}$ & $m_{35}^{Q}$ \\[0.5ex]
$x_{42}^{u} v_{\phi} = m_{52}^{u}$ & $\left[ 1 \pm \kappa \right] \times m_{42r}^{u}$ \\[0.5ex]
$x_{43}^{u} v_{\phi} = m_{53}^{u}$ & $\left[ 1 \pm \kappa \right] \times m_{43r}^{u}$ \\[0.5ex]
$M_{45}^{Q}$ & $M_{45}^{Q}$ \\[0.5ex]
$M_{54}^{u}$ & $\left[ 1 \pm \kappa \right] \times M_{54r}^{u}$ \\[0.5ex]
\midrule
$y_{14}^{d} v_{d} = m_{14}^{d}$ & $\left[ 1 \pm \kappa \right] \times m_{14r}^{d}$ \\[0.5ex]
$y_{24}^{d} v_{d} = m_{24}^{d}$ & $\left[ 1 \pm \kappa \right] \times m_{24r}^{d}$ \\[0.5ex]
$y_{34}^{d} v_{d} = m_{34}^{d}$ & $\left[ 1 \pm \kappa \right] \times m_{34r}^{d}$ \\[0.5ex]
$y_{43}^{d} v_{d} = m_{43}^{d}$ & $\left[ 1 \pm \kappa \right] \times m_{43r}^{d}$ \\[0.5ex]
$x_{34}^{Q} v_{\phi} = m_{35}^{Q}$ & $\left[ 1 \pm \kappa \right] \times m_{35r}^{Q}$ \\[0.5ex]
$x_{42}^{d} v_{\phi} = m_{52}^{d}$ & $\left[ 1 \pm \kappa \right] \times m_{42r}^{d}$ \\[0.5ex]
$x_{43}^{d} v_{\phi} = m_{53}^{d}$ & $\left[ 1 \pm \kappa \right] \times m_{43r}^{d}$ \\[0.5ex]
$M_{45}^{Q}$ & $\left[ 1 \pm \kappa \right] \times M_{45r}^{Q}$ \\[0.5ex]
$M_{54}^{d}$ & $\left[ 1 \pm \kappa \right] \times M_{54r}^{d}$ \\[0.5ex]
\midrule
$\kappa$ & $0.1$ \\
\bottomrule
\bottomrule
\end{tabular}
\captionof{table}{Next parameter setup after the initial parameter setup to find better mass parameters}
\label{tab:parameter_region_initial_q_second}}
\end{center}
We vary the parameter space given in Table~\ref{tab:parameter_region_initial_q_second} by first using the fitting function $\chi_{\operatorname{mass}}^{2}$
in order to find a suitable mass prediction for the four quarks $t,b,c$ and $b$. Once the obtained masses of these quarks are allocated in the ranges
($\left[ 1 \pm 0.1 \right] \times m_{c,s,b}$ and $\left[ 1 \pm 0.01 \right] \times m_{t}$), we proceed
to fit the CKM quark mixing matrix once more by using the other fitting function $\chi_{\operatorname{CKM}}^{2}$ and it is worth mentioning that
fitting the CKM mixing matrix is much more challenging due to the very small experimental errors of the CKM matrix elements. We display a benchmark point most converged for the CKM mixing matrix at the next subsection to discuss the possible deviation from the SM result arising from the first row of the CKM mixing matrix.
\subsubsection{Numerical scan result for the quark sector}
We start with the most converged benchmark point ($\chi_{\operatorname{CKM}}^{2} = 956.828$) after repeating the varying many times
\begin{equation}
\begin{split}
M^{u} &=
\begin{pmatrix}
0 & 0 & 0 & 0 & 0 \\[0.5ex]
0 & 0 & 0 & 0 & 14.474 \\[0.5ex]
0 & 0 & 0 & 1206.340 & 277.563 \\[0.5ex]
0 & 0 & 273.503 & -1775.200 & 0 \\[0.5ex]
0 & 550.990 & 434.462 & 0 & -5624.050
\end{pmatrix}
\quad
M^{d} =
\begin{pmatrix}
0 & 0 & 0 & 0 & -0.938 \\[0.5ex]
0 & 0 & 0 & 0 & -4.041 \\[0.5ex]
0 & 0 & 0 & 1206.340 & -27.427 \\[0.5ex]
0 & 0 & -5.636 & -1775.200 & 0 \\[0.5ex]
0 & 72.915 & -75.760 & 0 & 2623.620
\end{pmatrix}
\\[2ex]
M_{\operatorname{diag}}^{u} &=
\begin{pmatrix}
0 & 0 & 0 & 0 & 0 \\[0.5ex]
0 & 1.255 & 0 & 0 & 0 \\[0.5ex]
0 & 0 & 171.303 & 0 & 0 \\[0.5ex]
0 & 0 & 0 & 2155.890 & 0 \\[0.5ex]
0 & 0 & 0 & 0 & 5674.840
\end{pmatrix}
\quad
M_{\operatorname{diag}}^{d} =
\begin{pmatrix}
0 & 0 & 0 & 0 & 0 \\[0.5ex]
0 & 0.094 & 0 & 0 & 0 \\[0.5ex]
0 & 0 & 3.875 & 0 & 0 \\[0.5ex]
0 & 0 & 0 & 2146.190 & 0 \\[0.5ex]
0 & 0 & 0 & 0 & 2625.960
\end{pmatrix},
\label{eqn:MuMd_diagMuMd}
\end{split}
\end{equation}
where the above two mass matrices of Equation~\ref{eqn:MuMd_diagMuMd} are the mass matrices for the up- and down-type quarks in the flavor basis, whereas the below two mass matrices are ones fully diagonalized, so revealing all propagating quark mass. From the mass matrices of Equation~\ref{eqn:MuMd_diagMuMd}, we have the mixing matrices $V_{L}^{u}$ and $V_{L}^{d}$ and arrive to our CKM prediction using the formula of Equation~\ref{eqn:CKM_prediction}.
\begin{equation}
V_{\operatorname{CKM}}^{\operatorname{pred}}
=
\begin{pmatrix}
0.97409 & 0.22602 & 0.00799 & -6.38471 \times 10^{-6} & -0.00036 \\[0.5ex]
0.22615 & -0.97372 & -0.02697 & 3.80102 \times 10^{-5} & 0.00147 \\[0.5ex]
0.00166 & 0.02815 & -0.99880 & -0.00766 & 0.00874 \\[0.5ex]
0.00003 & 0.00019 & -0.00916 & 0.99919 & -0.01773 \\[0.5ex]
-0.00057 & 0.00112 & 0.03812 & 0.03539 & -0.00096
\label{eqn:CKM_bestfit}
\end{pmatrix}
\end{equation}
where a feature we should remember is the left-handed down-quark sector are able to reach to all mixings among the three SM generations, whereas the only left-handed $23$ mixing is allowed for the up-quark sector in this BSM model. The experimental CKM mixing matrix without unitarity is given in Equation~\ref{eqn:CKM_exp_wx_uni}.
\begingroup
\setlength\arraycolsep{5pt}
\begin{equation}
\lvert K_{\operatorname{CKM}} \rvert
=
\begin{pmatrix}
0.97370 \pm 0.00014 & 0.22450 \pm 0.00080 & 0.00382 \pm 0.00024 \\[1.5ex]
0.22100 \pm 0.00400 & 0.98700 \pm 0.01100 & 0.04100 \pm 0.00140 \\[1.5ex]
0.00800 \pm 0.00030 & 0.03880 \pm 0.00110 & 1.01300 \pm 0.03000
\end{pmatrix}
\label{eqn:CKM_exp_wx_uni}
\end{equation}
\endgroup
Restricting our attention to the upper-left $3\times 3$ block of Equation~\ref{eqn:CKM_bestfit}, it can be compared to its experimental bound given in Equation~\ref{eqn:CKM_exp_wx_uni}. In order to confirm that our prediction for the CKM mixing matrix is consistent with the experimental data,
it requires for the upper-left $3 \times 3$ block of Equation~\ref{eqn:CKM_bestfit} to be inside the $3\sigma$ experimentally allowed range
as follows:
\begin{equation}
\lvert (K_{\operatorname{CKM}})_{ij} \rvert - 3\lvert (\delta K_{\operatorname{CKM}})_{ij} \rvert < \lvert (V_{\operatorname{CKM}}^{\operatorname{pred}})_{ij} \rvert < \lvert (K_{\operatorname{CKM}})_{ij} \rvert + 3\lvert (\delta K_{\operatorname{CKM}} )_{ij}, \rvert, \quad \RIfM@\expandafter\text@\else\expandafter\mbox\fi{for $i,j=1,2,3$}
\end{equation}
and we confirm that the $13, 23, 31, 32$ elements in the CKM prediction of Equation~\ref{eqn:CKM_bestfit} cannot be fitted within the $3\sigma$ range with a small difference.
From our numerical analysis we find that in our model the
CKM quark mixing matrix mainly arises from the down type quark sector and has a subleading correction coming from the up type quark sector.
It is worth mentioning that the inclusion of an additional
vector-like family in our proposed model to provide masses for the
first generation of SM charged fermions
will lead to an improvement of our predictions related to the CKM quark mixing matrix.
However, that approach of having a fifth vector-like fermion family goes beyond the
scope in this work and is deferred for a future publication.
Furthermore, in this section we also discuss the possible deviation of the first row of the CKM mixing matrix without unitarity and this study is also covered in \Steve{references~\cite{Branco:2021vhs,Belfatto:2021jhf}} with an isosinglet vector-like quark in a model different than the one considered in this paper. According to
\Steve{\cite{Branco:2021vhs,Belfatto:2021jhf}},
the deviation $\Delta$ of unitarity is defined as follows:
\begin{equation}
\Delta = 1 - \lvert V_{ud} \rvert^2 - \lvert V_{us} \rvert^2 - \lvert V_{ub} \rvert^2,
\end{equation}
and its experimental value is given by
\cite{Belfatto:2019swo}.
\begin{equation}
\sqrt{\Delta} \sim 0.04
\label{eqn:dev_exp}
\end{equation}
Calculating the deviation of unitarity $\Delta$ from the best fitted CKM prediction of Equation~\ref{eqn:CKM_bestfit}, the result is
\begin{equation}
\begin{split}
\sqrt{\Delta} &\simeq 0.00035
\end{split}
\end{equation}
Therefore, the deviation of unitarity derived from \Antonio{the model under consideration}
is too small to be observed compared to its experimental bound given in Equation~\ref{eqn:dev_exp}. Lastly, we discuss the rare $t \rightarrow c Z$ decay and collect all benchmark points satisfying $\chi_{\operatorname{CKM}}^{2} < 980$ (notice that the most converged point reports $\chi^2_{\operatorname{CKM}} = 956.828$).
\begin{figure}[H]
\centering
\begin{subfigure}{0.48\textwidth}
\includegraphics[keepaspectratio,width=0.9\textwidth]{MU4MtU4BRtcZ1.jpg}
\end{subfigure}
\hspace{0.1cm}
\begin{subfigure}{0.48\textwidth}
\includegraphics[keepaspectratio,width=0.9\textwidth]{MD4MtD4BRtcZ1.jpg}
\end{subfigure}
\caption{Scanned mass region of the vector-like quarks and contributions of the flavor violating interactions with the SM $Z$ gauge boson to the rare $t \rightarrow c Z$ decay. The used constraints are the predicted $c,s,b$ and $t$ quark mass to be put between $[1 \pm 0.1] \times m_{c,s,b}$ and $[1 \pm 0.01] \times m_{t}$ and the CKM mixing matrix.}
\label{fig:tcZ_decay}
\end{figure}
Figure~\ref{fig:tcZ_decay} displays the allowed values of vector-like quark masses consistent with the constraints arising from the rare $t \rightarrow c Z$ decay. Our obtained values for the vector-like quark masses are consistent with their lower experimental bound of $1000\operatorname{GeV}$ arising from collider searches. In our numerical analysis the vector-like doublet up-type quark mass $M_{U_4}$ is ranged from $1850\operatorname{GeV}$ up to about
$2250\operatorname{GeV}$ and the vector-like singlet up-type quark mass $M_{\widetilde{U}_4}$ is varied from $4750\operatorname{GeV}$ to $5800\operatorname{GeV}$. Regarding the exotic down type quark sector, we have varied
the vector-like doublet down-type quark mass $M_{D_4}$
from $1850\operatorname{GeV}$ up to $2250\operatorname{GeV}$ and the vector-like singlet down-type quark mass $M_{\widetilde{D}_4}$
from $2450\operatorname{GeV}$ up to $3000\operatorname{GeV}$. As seen from Figure~\ref{fig:tcZ_decay} the order of magnitude of the obtained values for the branching ratio of the rare $t \rightarrow c Z$ decay range from $10^{-6}$ up to $10^{-5}$, which is consistent with
its current experimental bound, whose logarithmic value is about $-3.6$ as indicated by
Equation~\ref{eqn:BRtcZ_exp}.
\section{CONCLUSION} \label{sec:VII}
In this work we have considered a model where the SM fermion sector is extended by the inclusion of a
fourth vector-like family and the scalar sector is augmented by the incorporation of an extra scalar doublet and a gauge singlet scalar. In addition, we have assumed a global $U(1)^{\prime}$ symmetry under which all particles are charged except the SM chiral quark and lepton fields.
The model explains the hierarchical structure of the SM quark and lepton masses
by assuming that the SM Yukawa interactions are forbidden by the $U(1)^{\prime}$ symmetry and arise effectively after it is spontaneously broken, due to
induced mixing with the fourth vector-like family. This mixing also results in non-standard couplings of the $W$ and $Z$ gauge bosons which have been studied here for the first time.
This setup leads to sizeable \Antonio{branching fractions for the FCNC decays such as} $\mu \rightarrow e \gamma$, $Z\to\mu\tau$ and $t\rightarrow cZ$, within the reach of the future experimental sensitivity. These FCNC decays are studied in detail in this work, in order to set constraints on the model parameter space.
A great advantage of the approach taken in this work with respect to the ones considered in extensions of the SM having a $Z^{\prime}$ gauge boson
is that it makes the study of the FCNC observables simpler than in the latter since in the former \Antonio{we can avoid assuming specific values for the unknown $U(1)^{\prime}$ coupling and
$Z^{\prime}$ gauge boson mass.}
This makes the present phenomenology based on $W$ and $Z$ gauge boson couplings
more predictive than if the $U(1)^{\prime}$ were a spontaneously broken gauge symmetry, leading to a massive $Z'$.
Given that the hierarchical structure of the SM is implemented in our proposed model, the extended mass matrices for the charged lepton and quark sectors need to be completely and accurately diagonalised, as
the starting point of our analytical and numerical analysis.
Since we only consider a fourth vector-like family, the model
cannot provide masses for the first generation of SM charged fermions,
nevertheless this is a good approximation given that the first generation of the SM fermions are very light. For this reason, we mainly focus on the study of FCNC observables
involving the second and third generations of SM fermions in both quark and lepton sectors.
In the chosen convenient basis, the different shape of the down-type quark mass matrix allows all left-handed mixings between the three SM generations, whereas the up type quark sector can have only the $23$ left-handed mixing, while all quarks and charged lepton have the only $23$ right-handed mixing, and we have checked that the results are basis independent.
This feature implies that we can obtain
a prediction for the CKM mixing matrix and this is one of main phenomenological aspects analyzed in this work.
In order to diagonalize the fermionic mass matrices, in an analytic approximation,
we have defined the $SU(2)$ conserving and $SU(2)$ violating mixings and we have shown that the $SU(2)$ violating mixing plays a crucial role for generating the \Antonio{$Z$ mediated flavor violating interactions.
Furthermore, the
extension of the SM fermion sector by the inclusion of a vector-like family makes the matrices of $Z$ couplings with fermions
different than the identity matrix
due to the appearance of non-zero off-diagonal matrix elements of the $Z$ coupling matrices which will give rise to flavor violating $Z$ decays. The non-zero off-diagonal SM $Z$ gauge coupling constants are generally proportional to two of the small mixing angles, which are defined by the ratio between the SM fermion and vector-like masses, thus leading to small values.}
\\~
\Antonio{Defining all the required $Z$ gauge coupling constants with fermions in the mass basis, as discussed above, we began by
analyzing the FCNC processes of the charged lepton sector.}
We have found that in the lepton sector,
the following three
FCNC decays are allowed:
$\tau \rightarrow \mu \gamma, \tau \rightarrow 3 \mu$ and $Z \rightarrow \mu \tau$. Regarding
the $\tau \rightarrow \mu \gamma$ decay, we discussed its leading contribution, which
arises from the Feynman diagrams having a chirality flip in the internal fermionic lines and being proportional to
$M/m_{\tau}$, where $M$ is the mass scale of the heavy charged vector-like leptons. However, the dominant terms cannot be as big as the vector-like masses get heavier since their coupling constants get suppressed at the same time, thus providing a
balanced relation between the vector-like masses and their coupling constants. We have found that
our predictions for the vector-like charged lepton masses are not severely constrained by the $\tau \rightarrow \mu \gamma$ decay since most of the obtained values for the $\tau \rightarrow \mu \gamma$ decay are consistent with its experimental upper bound. In the concerning to the
$\tau \rightarrow 3\mu$ and $Z \rightarrow \mu \tau$ decays, we have derived an analytic expression for their corresponding rates at tree-level finding that
none of
our predictions is constrained by the experimental bounds of these decays. Considering the FCC-ee experiment which have planned to generate $10^{12}$ the $Z$ gauge bosons and our numerical prediction for the $Z \rightarrow \mu \tau$ branching ratio is of the
order of $10^{-9}$ at most, thus implying that our model can be tested at the $Z$ factory via the $Z \rightarrow \mu \tau$ decay.
However, the CMS provided that the doublet vector-like mass can be constrained up to $790\operatorname{GeV}$~\cite{Hernandez:2021tii,Xu:2018pnq} and our numerical predictions for the vector-like charged leptons are severely constrained by the CMS result. Therefore, we can expect that the vector-like charged lepton doublet mass is ranged from $790$ to nearly $1600 \operatorname{GeV}$, whereas the vector-like charged lepton singlet mass is ranged from $500$ to $2000 \operatorname{GeV}$ or above than that.
Turning to the quark sector phenomenology, we have analyzed the rare $t \rightarrow c Z$ decay as well as the CKM mixing
to set constraints on the quark sector parameters. It is worth mentioning that the
neutral meson $K, B_{d}, B_{s}$ oscillations do not set constraints on the quark sector parameters of our model since their new physics effects are quite negligible compared to the SM expectation.
We have derived analytic expressions for the rare $t \rightarrow c Z$ decay as well as for the CKM mixing matrix. \Antonio{Due to the mixings between SM fermions and vector-like fermions, the CKM quark mixing matrix is not unitary, thus implying that the unitarity requirement has to be relaxed}
~\cite{Branco:2021vhs,ParticleDataGroup:2020ssz}. \Antonio{Using the most converged benchmark point, we showed how dominant the down-type quark mixing matrix plays a crucial role in the CKM mixing matrix and we discussed the deviation of unitarity arisen from the first row of the CKM mixing matrix, whose value is too small to be experimentally measured.
Finally, we investigate the branching ratio for the $t \rightarrow c Z$ decay and found that our numerical predictions are not excluded by its experimental bound, for
vector-like doublet up-type and down type quark masses $M_{U_4}$ and $M_{D_4}$ in the window \mbox{$1850$\RIfM@\expandafter\text@\else\expandafter\mbox\fi{GeV}$\leqslant M_{U_4},M_{D_4}\leqslant 2250\RIfM@\expandafter\text@\else\expandafter\mbox\fi{GeV}$} as well as
vector-like singlet up and down type quark masses $M_{\widetilde{U}_4}$ and $M_{\widetilde{D}_4}$ in the ranges \mbox{$4750$\RIfM@\expandafter\text@\else\expandafter\mbox\fi{GeV}$\leqslant M_{\widetilde{U}_4}\leqslant 5800\RIfM@\expandafter\text@\else\expandafter\mbox\fi{GeV}$},
and \mbox{$2450$\RIfM@\expandafter\text@\else\expandafter\mbox\fi{GeV}$\leqslant M_{\widetilde{D}_4}\leqslant 3000\RIfM@\expandafter\text@\else\expandafter\mbox\fi{GeV}$}, respectively.}
In conclusion, we have analysed a range of FCNCs arising from non-standard $W$ and $Z$ gauge boson couplings in an extension of the SM with a fourth vector-like family, which can also address the hierarchy of quark and lepton masses, leading to several interesting rare decays which may be probed in future high luminosity experiments.
\section*{Acknowledgements}
This research has received funding from Fondecyt (Chile), Grant No.~1210378, ANID PIA/APOYO AFB180002 and Milenio-ANID-ICN2019\_044. SFK acknowledges the STFC Consolidated Grant ST/L000296/1 and the European Union's Horizon 2020 Research and Innovation programme under Marie Sk\l {}odowska-Curie grant agreement HIDDeN European ITN project (H2020-MSCA-ITN-2019//860881-HIDDeN).
\section*{Abstract (Not appropriate in this style!)}%
\else \small
\begin{center}{\bf Abstract\vspace{-.5em}\vspace{\z@}}\end{center}%
\quotation
\fi
}%
}{%
}%
\@ifundefined{endabstract}{\def\endabstract
{\if@twocolumn\else\endquotation\fi}}{}%
\@ifundefined{maketitle}{\def\maketitle#1{}}{}%
\@ifundefined{affiliation}{\def\affiliation#1{}}{}%
\@ifundefined{proof}{\def\proof{\noindent{\bfseries Proof. }}}{}%
\@ifundefined{endproof}{\def\endproof{\mbox{\ \rule{.1in}{.1in}}}}{}%
\@ifundefined{newfield}{\def\newfield#1#2{}}{}%
\@ifundefined{chapter}{\def\chapter#1{\par(Chapter head:)#1\par }%
\newcount\c@chapter}{}%
\@ifundefined{part}{\def\part#1{\par(Part head:)#1\par }}{}%
\@ifundefined{section}{\def\section#1{\par(Section head:)#1\par }}{}%
\@ifundefined{subsection}{\def\subsection#1%
{\par(Subsection head:)#1\par }}{}%
\@ifundefined{subsubsection}{\def\subsubsection#1%
{\par(Subsubsection head:)#1\par }}{}%
\@ifundefined{paragraph}{\def\paragraph#1%
{\par(Subsubsubsection head:)#1\par }}{}%
\@ifundefined{subparagraph}{\def\subparagraph#1%
{\par(Subsubsubsubsection head:)#1\par }}{}%
\@ifundefined{therefore}{\def\therefore{}}{}%
\@ifundefined{backepsilon}{\def\backepsilon{}}{}%
\@ifundefined{yen}{\def\yen{\hbox{\rm\rlap=Y}}}{}%
\@ifundefined{registered}{%
\def\registered{\relax\ifmmode{}\r@gistered
\else$\m@th\r@gistered$\fi}%
\def\r@gistered{^{\ooalign
{\hfil\raise.07ex\hbox{$\scriptstyle\rm\RIfM@\expandafter\text@\else\expandafter\mbox\fi{R}$}\hfil\crcr
\mathhexbox20D}}}}{}%
\@ifundefined{Eth}{\def\Eth{}}{}%
\@ifundefined{eth}{\def\eth{}}{}%
\@ifundefined{Thorn}{\def\Thorn{}}{}%
\@ifundefined{thorn}{\def\thorn{}}{}%
\def\TEXTsymbol#1{\mbox{$#1$}}%
\@ifundefined{degree}{\def\degree{{}^{\circ}}}{}%
\newdimen\theight
\@ifundefined{Column}{\def\Column{%
\vadjust{\setbox\z@=\hbox{\scriptsize\quad\quad tcol}%
\theight=\ht\z@\advance\theight by \dp\z@\advance\theight by \lineskip
\kern -\theight \vbox to \theight{%
\rightline{\rlap{\box\z@}}%
\vss
}%
}%
}}{}%
\@ifundefined{qed}{\def\qed{%
\ifhmode\unskip\nobreak\fi\ifmmode\ifinner\else\hskip5\p@\fi\fi
\hbox{\hskip5\p@\vrule width4\p@ height6\p@ depth1.5\p@\hskip\p@}%
}}{}%
\@ifundefined{cents}{\def\cents{\hbox{\rm\rlap c/}}}{}%
\@ifundefined{tciLaplace}{\def\tciLaplace{\ensuremath{\mathcal{L}}}}{}%
\@ifundefined{tciFourier}{\def\tciFourier{\ensuremath{\mathcal{F}}}}{}%
\@ifundefined{textcurrency}{\def\textcurrency{\hbox{\rm\rlap xo}}}{}%
\@ifundefined{texteuro}{\def\texteuro{\hbox{\rm\rlap C=}}}{}%
\@ifundefined{euro}{\def\euro{\hbox{\rm\rlap C=}}}{}%
\@ifundefined{textfranc}{\def\textfranc{\hbox{\rm\rlap-F}}}{}%
\@ifundefined{textlira}{\def\textlira{\hbox{\rm\rlap L=}}}{}%
\@ifundefined{textpeseta}{\def\textpeseta{\hbox{\rm P\negthinspace s}}}{}%
\@ifundefined{miss}{\def\miss{\hbox{\vrule height2\p@ width 2\p@ depth\z@}}}{}%
\@ifundefined{vvert}{\def\vvert{\Vert}}{
\@ifundefined{tcol}{\def\tcol#1{{\baselineskip=6\p@ \vcenter{#1}} \Column}}{}%
\@ifundefined{dB}{\def\dB{\hbox{{}}}}{
\@ifundefined{mB}{\def\mB#1{\hbox{$#1$}}}{
\@ifundefined{nB}{\def\nB#1{\hbox{#1}}}{
\@ifundefined{note}{\def\note{$^{\dag}}}{}%
\defLaTeX2e{LaTeX2e}
\ifx\fmtnameLaTeX2e
\DeclareOldFontCommand{\rm}{\normalfont\rmfamily}{\mathrm}
\DeclareOldFontCommand{\sf}{\normalfont\sffamily}{\mathsf}
\DeclareOldFontCommand{\tt}{\normalfont\ttfamily}{\mathtt}
\DeclareOldFontCommand{\bf}{\normalfont\bfseries}{\mathbf}
\DeclareOldFontCommand{\it}{\normalfont\itshape}{\mathit}
\DeclareOldFontCommand{\sl}{\normalfont\slshape}{\@nomath\sl}
\DeclareOldFontCommand{\sc}{\normalfont\scshape}{\@nomath\sc}
\fi
\def\alpha{{\Greekmath 010B}}%
\def\beta{{\Greekmath 010C}}%
\def\gamma{{\Greekmath 010D}}%
\def\delta{{\Greekmath 010E}}%
\def\epsilon{{\Greekmath 010F}}%
\def\zeta{{\Greekmath 0110}}%
\def\eta{{\Greekmath 0111}}%
\def\theta{{\Greekmath 0112}}%
\def\iota{{\Greekmath 0113}}%
\def\kappa{{\Greekmath 0114}}%
\def\lambda{{\Greekmath 0115}}%
\def\mu{{\Greekmath 0116}}%
\def\nu{{\Greekmath 0117}}%
\def\xi{{\Greekmath 0118}}%
\def\pi{{\Greekmath 0119}}%
\def\rho{{\Greekmath 011A}}%
\def\sigma{{\Greekmath 011B}}%
\def\tau{{\Greekmath 011C}}%
\def\upsilon{{\Greekmath 011D}}%
\def\phi{{\Greekmath 011E}}%
\def\chi{{\Greekmath 011F}}%
\def\psi{{\Greekmath 0120}}%
\def\omega{{\Greekmath 0121}}%
\def\varepsilon{{\Greekmath 0122}}%
\def\vartheta{{\Greekmath 0123}}%
\def\varpi{{\Greekmath 0124}}%
\def\varrho{{\Greekmath 0125}}%
\def\varsigma{{\Greekmath 0126}}%
\def\varphi{{\Greekmath 0127}}%
\def{\Greekmath 0272}{{\Greekmath 0272}}
\def\FindBoldGroup{%
{\setbox0=\hbox{$\mathbf{x\global\edef\theboldgroup{\the\mathgroup}}$}}%
}
\def\Greekmath#1#2#3#4{%
\if@compatibility
\ifnum\mathgroup=\symbold
\mathchoice{\mbox{\boldmath$\displaystyle\mathchar"#1#2#3#4$}}%
{\mbox{\boldmath$\textstyle\mathchar"#1#2#3#4$}}%
{\mbox{\boldmath$\scriptstyle\mathchar"#1#2#3#4$}}%
{\mbox{\boldmath$\scriptscriptstyle\mathchar"#1#2#3#4$}}%
\else
\mathchar"#1#2#3#
\fi
\else
\FindBoldGroup
\ifnum\mathgroup=\theboldgroup
\mathchoice{\mbox{\boldmath$\displaystyle\mathchar"#1#2#3#4$}}%
{\mbox{\boldmath$\textstyle\mathchar"#1#2#3#4$}}%
{\mbox{\boldmath$\scriptstyle\mathchar"#1#2#3#4$}}%
{\mbox{\boldmath$\scriptscriptstyle\mathchar"#1#2#3#4$}}%
\else
\mathchar"#1#2#3#
\fi
\fi}
\newif\ifGreekBold \GreekBoldfalse
\let\SAVEPBF=\pbf
\def\pbf{\GreekBoldtrue\SAVEPBF}%
\@ifundefined{theorem}{\newtheorem{theorem}{Theorem}}{}
\@ifundefined{lemma}{\newtheorem{lemma}[theorem]{Lemma}}{}
\@ifundefined{corollary}{\newtheorem{corollary}[theorem]{Corollary}}{}
\@ifundefined{conjecture}{\newtheorem{conjecture}[theorem]{Conjecture}}{}
\@ifundefined{proposition}{\newtheorem{proposition}[theorem]{Proposition}}{}
\@ifundefined{axiom}{\newtheorem{axiom}{Axiom}}{}
\@ifundefined{remark}{\newtheorem{remark}{Remark}}{}
\@ifundefined{example}{\newtheorem{example}{Example}}{}
\@ifundefined{exercise}{\newtheorem{exercise}{Exercise}}{}
\@ifundefined{definition}{\newtheorem{definition}{Definition}}{}
\@ifundefined{mathletters}{%
\newcounter{equationnumber}
\def\mathletters{%
\addtocounter{equation}{1}
\edef\@currentlabel{\arabic{equation}}%
\setcounter{equationnumber}{\c@equation}
\setcounter{equation}{0}%
\edef\arabic{equation}{\@currentlabel\noexpand\alph{equation}}%
}
\def\endmathletters{%
\setcounter{equation}{\value{equationnumber}}%
}
}{}
\@ifundefined{BibTeX}{%
\def\BibTeX{{\rm B\kern-.05em{\sc i\kern-.025em b}\kern-.08em
T\kern-.1667em\lower.7ex\hbox{E}\kern-.125emX}}}{}%
\@ifundefined{AmS}%
{\def\AmS{{\protect\usefont{OMS}{cmsy}{m}{n}%
A\kern-.1667em\lower.5ex\hbox{M}\kern-.125emS}}}{}%
\@ifundefined{AmSTeX}{\def\AmSTeX{\protect\AmS-\protect\TeX\@}}{}%
\def\@@eqncr{\let\@tempa\relax
\ifcase\@eqcnt \def\@tempa{& & &}\or \def\@tempa{& &}%
\else \def\@tempa{&}\fi
\@tempa
\if@eqnsw
\iftag@
\@taggnum
\else
\@eqnnum\stepcounter{equation}%
\fi
\fi
\global\@ifnextchar*{\@tagstar}{\@tag}@false
\global\@eqnswtrue
\global\@eqcnt\z@\cr}
\def\@ifnextchar*{\@TCItagstar}{\@TCItag}{\@ifnextchar*{\@TCItagstar}{\@TCItag}}
\def\@TCItag#1{%
\global\@ifnextchar*{\@tagstar}{\@tag}@true
\global\def\@taggnum{(#1)}%
\global\def\@currentlabel{#1}}
\def\@TCItagstar*#1{%
\global\@ifnextchar*{\@tagstar}{\@tag}@true
\global\def\@taggnum{#1}%
\global\def\@currentlabel{#1}}
\def\QATOP#1#2{{#1 \atop #2}}%
\def\QTATOP#1#2{{\textstyle {#1 \atop #2}}}%
\def\QDATOP#1#2{{\displaystyle {#1 \atop #2}}}%
\def\QABOVE#1#2#3{{#2 \above#1 #3}}%
\def\QTABOVE#1#2#3{{\textstyle {#2 \above#1 #3}}}%
\def\QDABOVE#1#2#3{{\displaystyle {#2 \above#1 #3}}}%
\def\QOVERD#1#2#3#4{{#3 \overwithdelims#1#2 #4}}%
\def\QTOVERD#1#2#3#4{{\textstyle {#3 \overwithdelims#1#2 #4}}}%
\def\QDOVERD#1#2#3#4{{\displaystyle {#3 \overwithdelims#1#2 #4}}}%
\def\QATOPD#1#2#3#4{{#3 \atopwithdelims#1#2 #4}}%
\def\QTATOPD#1#2#3#4{{\textstyle {#3 \atopwithdelims#1#2 #4}}}%
\def\QDATOPD#1#2#3#4{{\displaystyle {#3 \atopwithdelims#1#2 #4}}}%
\def\QABOVED#1#2#3#4#5{{#4 \abovewithdelims#1#2#3 #5}}%
\def\QTABOVED#1#2#3#4#5{{\textstyle
{#4 \abovewithdelims#1#2#3 #5}}}%
\def\QDABOVED#1#2#3#4#5{{\displaystyle
{#4 \abovewithdelims#1#2#3 #5}}}%
\def\tint{\mathop{\textstyle \int}}%
\def\tiint{\mathop{\textstyle \iint }}%
\def\tiiint{\mathop{\textstyle \iiint }}%
\def\tiiiint{\mathop{\textstyle \iiiint }}%
\def\tidotsint{\mathop{\textstyle \idotsint }}%
\def\toint{\mathop{\textstyle \oint}}%
\def\tsum{\mathop{\textstyle \sum }}%
\def\tprod{\mathop{\textstyle \prod }}%
\def\tbigcap{\mathop{\textstyle \bigcap }}%
\def\tbigwedge{\mathop{\textstyle \bigwedge }}%
\def\tbigoplus{\mathop{\textstyle \bigoplus }}%
\def\tbigodot{\mathop{\textstyle \bigodot }}%
\def\tbigsqcup{\mathop{\textstyle \bigsqcup }}%
\def\tcoprod{\mathop{\textstyle \coprod }}%
\def\tbigcup{\mathop{\textstyle \bigcup }}%
\def\tbigvee{\mathop{\textstyle \bigvee }}%
\def\tbigotimes{\mathop{\textstyle \bigotimes }}%
\def\tbiguplus{\mathop{\textstyle \biguplus }}%
\def\dint{\mathop{\displaystyle \int}}%
\def\diint{\mathop{\displaystyle \iint}}%
\def\diiint{\mathop{\displaystyle \iiint}}%
\def\diiiint{\mathop{\displaystyle \iiiint }}%
\def\didotsint{\mathop{\displaystyle \idotsint }}%
\def\doint{\mathop{\displaystyle \oint}}%
\def\dsum{\mathop{\displaystyle \sum }}%
\def\dprod{\mathop{\displaystyle \prod }}%
\def\dbigcap{\mathop{\displaystyle \bigcap }}%
\def\dbigwedge{\mathop{\displaystyle \bigwedge }}%
\def\dbigoplus{\mathop{\displaystyle \bigoplus }}%
\def\dbigodot{\mathop{\displaystyle \bigodot }}%
\def\dbigsqcup{\mathop{\displaystyle \bigsqcup }}%
\def\dcoprod{\mathop{\displaystyle \coprod }}%
\def\dbigcup{\mathop{\displaystyle \bigcup }}%
\def\dbigvee{\mathop{\displaystyle \bigvee }}%
\def\dbigotimes{\mathop{\displaystyle \bigotimes }}%
\def\dbiguplus{\mathop{\displaystyle \biguplus }}%
\if@compatibility\else
\RequirePackage{amsmath}
\fi
\def\makeatother\endinput{\makeatother\endinput}
\bgroup
\ifx\ds@amstex\relax
\message{amstex already loaded}\aftergroup\makeatother\endinput
\else
\@ifpackageloaded{amsmath}%
{\if@compatibility\message{amsmath already loaded}\fi\aftergroup\makeatother\endinput}
{}
\@ifpackageloaded{amstex}%
{\if@compatibility\message{amstex already loaded}\fi\aftergroup\makeatother\endinput}
{}
\@ifpackageloaded{amsgen}%
{\if@compatibility\message{amsgen already loaded}\fi\aftergroup\makeatother\endinput}
{}
\fi
\egroup
\typeout{TCILATEX defining AMS-like constructs in LaTeX 2.09 COMPATIBILITY MODE}
\let\DOTSI\relax
\def\RIfM@{\relax\ifmmode}%
\def\FN@{\futurelet\next}%
\newcount\intno@
\def\iint{\DOTSI\intno@\tw@\FN@\ints@}%
\def\iiint{\DOTSI\intno@\thr@@\FN@\ints@}%
\def\iiiint{\DOTSI\intno@4 \FN@\ints@}%
\def\idotsint{\DOTSI\intno@\z@\FN@\ints@}%
\def\ints@{\findlimits@\ints@@}%
\newif\iflimtoken@
\newif\iflimits@
\def\findlimits@{\limtoken@true\ifx\next\limits\limits@true
\else\ifx\next\nolimits\limits@false\else
\limtoken@false\ifx\ilimits@\nolimits\limits@false\else
\ifinner\limits@false\else\limits@true\fi\fi\fi\fi}%
\def\multint@{\int\ifnum\intno@=\z@\intdots@
\else\intkern@\fi
\ifnum\intno@>\tw@\int\intkern@\fi
\ifnum\intno@>\thr@@\int\intkern@\fi
\int
\def\multintlimits@{\intop\ifnum\intno@=\z@\intdots@\else\intkern@\fi
\ifnum\intno@>\tw@\intop\intkern@\fi
\ifnum\intno@>\thr@@\intop\intkern@\fi\intop}%
\def\intic@{%
\mathchoice{\hskip.5em}{\hskip.4em}{\hskip.4em}{\hskip.4em}}%
\def\negintic@{\mathchoice
{\hskip-.5em}{\hskip-.4em}{\hskip-.4em}{\hskip-.4em}}%
\def\ints@@{\iflimtoken@
\def\ints@@@{\iflimits@\negintic@
\mathop{\intic@\multintlimits@}\limits
\else\multint@\nolimits\fi
\eat@
\else
\def\ints@@@{\iflimits@\negintic@
\mathop{\intic@\multintlimits@}\limits\else
\multint@\nolimits\fi}\fi\ints@@@}%
\def\intkern@{\mathchoice{\!\!\!}{\!\!}{\!\!}{\!\!}}%
\def\plaincdots@{\mathinner{\cdotp\cdotp\cdotp}}%
\def\intdots@{\mathchoice{\plaincdots@}%
{{\cdotp}\mkern1.5mu{\cdotp}\mkern1.5mu{\cdotp}}%
{{\cdotp}\mkern1mu{\cdotp}\mkern1mu{\cdotp}}%
{{\cdotp}\mkern1mu{\cdotp}\mkern1mu{\cdotp}}}%
\def\RIfM@{\relax\protect\ifmmode}
\def\RIfM@\expandafter\text@\else\expandafter\mbox\fi{\RIfM@\expandafter\RIfM@\expandafter\text@\else\expandafter\mbox\fi@\else\expandafter\mbox\fi}
\let\nfss@text\RIfM@\expandafter\text@\else\expandafter\mbox\fi
\def\RIfM@\expandafter\text@\else\expandafter\mbox\fi@#1{\mathchoice
{\textdef@\displaystyle\f@size{#1}}%
{\textdef@\textstyle\tf@size{\firstchoice@false #1}}%
{\textdef@\textstyle\sf@size{\firstchoice@false #1}}%
{\textdef@\textstyle \ssf@size{\firstchoice@false #1}}%
\glb@settings}
\def\textdef@#1#2#3{\hbox{{%
\everymath{#1}%
\let\f@size#2\selectfont
#3}}}
\newif\iffirstchoice@
\firstchoice@true
\def\Let@{\relax\iffalse{\fi\let\\=\cr\iffalse}\fi}%
\def\vspace@{\def\vspace##1{\crcr\noalign{\vskip##1\relax}}}%
\def\multilimits@{\bgroup\vspace@\Let@
\baselineskip\fontdimen10 \scriptfont\tw@
\advance\baselineskip\fontdimen12 \scriptfont\tw@
\lineskip\thr@@\fontdimen8 \scriptfont\thr@@
\lineskiplimit\lineskip
\vbox\bgroup\ialign\bgroup\hfil$\m@th\scriptstyle{##}$\hfil\crcr}%
\def\Sb{_\multilimits@}%
\def\endSb{\crcr\egroup\egroup\egroup}%
\def\Sp{^\multilimits@}%
\let\endSp\endSb
\newdimen\ex@
\ex@.2326ex
\def\rightarrowfill@#1{$#1\m@th\mathord-\mkern-6mu\cleaders
\hbox{$#1\mkern-2mu\mathord-\mkern-2mu$}\hfill
\mkern-6mu\mathord\rightarrow$}%
\def\leftarrowfill@#1{$#1\m@th\mathord\leftarrow\mkern-6mu\cleaders
\hbox{$#1\mkern-2mu\mathord-\mkern-2mu$}\hfill\mkern-6mu\mathord-$}%
\def\leftrightarrowfill@#1{$#1\m@th\mathord\leftarrow
\mkern-6mu\cleaders
\hbox{$#1\mkern-2mu\mathord-\mkern-2mu$}\hfill
\mkern-6mu\mathord\rightarrow$}%
\def\overrightarrow{\mathpalette\overrightarrow@}%
\def\overrightarrow@#1#2{\vbox{\ialign{##\crcr\rightarrowfill@#1\crcr
\noalign{\kern-\ex@\nointerlineskip}$\m@th\hfil#1#2\hfil$\crcr}}}%
\let\overarrow\overrightarrow
\def\overleftarrow{\mathpalette\overleftarrow@}%
\def\overleftarrow@#1#2{\vbox{\ialign{##\crcr\leftarrowfill@#1\crcr
\noalign{\kern-\ex@\nointerlineskip}$\m@th\hfil#1#2\hfil$\crcr}}}%
\def\overleftrightarrow{\mathpalette\overleftrightarrow@}%
\def\overleftrightarrow@#1#2{\vbox{\ialign{##\crcr
\leftrightarrowfill@#1\crcr
\noalign{\kern-\ex@\nointerlineskip}$\m@th\hfil#1#2\hfil$\crcr}}}%
\def\underrightarrow{\mathpalette\underrightarrow@}%
\def\underrightarrow@#1#2{\vtop{\ialign{##\crcr$\m@th\hfil#1#2\hfil
$\crcr\noalign{\nointerlineskip}\rightarrowfill@#1\crcr}}}%
\let\underarrow\underrightarrow
\def\underleftarrow{\mathpalette\underleftarrow@}%
\def\underleftarrow@#1#2{\vtop{\ialign{##\crcr$\m@th\hfil#1#2\hfil
$\crcr\noalign{\nointerlineskip}\leftarrowfill@#1\crcr}}}%
\def\underleftrightarrow{\mathpalette\underleftrightarrow@}%
\def\underleftrightarrow@#1#2{\vtop{\ialign{##\crcr$\m@th
\hfil#1#2\hfil$\crcr
\noalign{\nointerlineskip}\leftrightarrowfill@#1\crcr}}}%
\def\qopnamewl@#1{\mathop{\operator@font#1}\nlimits@}
\let\nlimits@\displaylimits
\def\setboxz@h{\setbox\z@\hbox}
\def\varlim@#1#2{\mathop{\vtop{\ialign{##\crcr
\hfil$#1\m@th\operator@font lim$\hfil\crcr
\noalign{\nointerlineskip}#2#1\crcr
\noalign{\nointerlineskip\kern-\ex@}\crcr}}}}
\def\rightarrowfill@#1{\m@th\setboxz@h{$#1-$}\ht\z@\z@
$#1\copy\z@\mkern-6mu\cleaders
\hbox{$#1\mkern-2mu\box\z@\mkern-2mu$}\hfill
\mkern-6mu\mathord\rightarrow$}
\def\leftarrowfill@#1{\m@th\setboxz@h{$#1-$}\ht\z@\z@
$#1\mathord\leftarrow\mkern-6mu\cleaders
\hbox{$#1\mkern-2mu\copy\z@\mkern-2mu$}\hfill
\mkern-6mu\box\z@$}
\def\qopnamewl@{proj\,lim}{\qopnamewl@{proj\,lim}}
\def\qopnamewl@{inj\,lim}{\qopnamewl@{inj\,lim}}
\def\mathpalette\varlim@\rightarrowfill@{\mathpalette\varlim@\rightarrowfill@}
\def\mathpalette\varlim@\leftarrowfill@{\mathpalette\varlim@\leftarrowfill@}
\def\mathpalette\varliminf@{}{\mathpalette\mathpalette\varliminf@{}@{}}
\def\mathpalette\varliminf@{}@#1{\mathop{\underline{\vrule\@depth.2\ex@\@width\z@
\hbox{$#1\m@th\operator@font lim$}}}}
\def\mathpalette\varlimsup@{}{\mathpalette\mathpalette\varlimsup@{}@{}}
\def\mathpalette\varlimsup@{}@#1{\mathop{\overline
{\hbox{$#1\m@th\operator@font lim$}}}}
\def\stackunder#1#2{\mathrel{\mathop{#2}\limits_{#1}}}%
\begingroup \catcode `|=0 \catcode `[= 1
\catcode`]=2 \catcode `\{=12 \catcode `\}=12
\catcode`\\=12
|gdef|@alignverbatim#1\end{align}[#1|end[align]]
|gdef|@salignverbatim#1\end{align*}[#1|end[align*]]
|gdef|@alignatverbatim#1\end{alignat}[#1|end[alignat]]
|gdef|@salignatverbatim#1\end{alignat*}[#1|end[alignat*]]
|gdef|@xalignatverbatim#1\end{xalignat}[#1|end[xalignat]]
|gdef|@sxalignatverbatim#1\end{xalignat*}[#1|end[xalignat*]]
|gdef|@gatherverbatim#1\end{gather}[#1|end[gather]]
|gdef|@sgatherverbatim#1\end{gather*}[#1|end[gather*]]
|gdef|@gatherverbatim#1\end{gather}[#1|end[gather]]
|gdef|@sgatherverbatim#1\end{gather*}[#1|end[gather*]]
|gdef|@multilineverbatim#1\end{multiline}[#1|end[multiline]]
|gdef|@smultilineverbatim#1\end{multiline*}[#1|end[multiline*]]
|gdef|@arraxverbatim#1\end{arrax}[#1|end[arrax]]
|gdef|@sarraxverbatim#1\end{arrax*}[#1|end[arrax*]]
|gdef|@tabulaxverbatim#1\end{tabulax}[#1|end[tabulax]]
|gdef|@stabulaxverbatim#1\end{tabulax*}[#1|end[tabulax*]]
|endgroup
\def\align{\@verbatim \frenchspacing\@vobeyspaces \@alignverbatim
You are using the "align" environment in a style in which it is not defined.}
\let\endalign=\endtrivlist
\@namedef{align*}{\@verbatim\@salignverbatim
You are using the "align*" environment in a style in which it is not defined.}
\expandafter\let\csname endalign*\endcsname =\endtrivlist
\def\alignat{\@verbatim \frenchspacing\@vobeyspaces \@alignatverbatim
You are using the "alignat" environment in a style in which it is not defined.}
\let\endalignat=\endtrivlist
\@namedef{alignat*}{\@verbatim\@salignatverbatim
You are using the "alignat*" environment in a style in which it is not defined.}
\expandafter\let\csname endalignat*\endcsname =\endtrivlist
\def\xalignat{\@verbatim \frenchspacing\@vobeyspaces \@xalignatverbatim
You are using the "xalignat" environment in a style in which it is not defined.}
\let\endxalignat=\endtrivlist
\@namedef{xalignat*}{\@verbatim\@sxalignatverbatim
You are using the "xalignat*" environment in a style in which it is not defined.}
\expandafter\let\csname endxalignat*\endcsname =\endtrivlist
\def\gather{\@verbatim \frenchspacing\@vobeyspaces \@gatherverbatim
You are using the "gather" environment in a style in which it is not defined.}
\let\endgather=\endtrivlist
\@namedef{gather*}{\@verbatim\@sgatherverbatim
You are using the "gather*" environment in a style in which it is not defined.}
\expandafter\let\csname endgather*\endcsname =\endtrivlist
\def\multiline{\@verbatim \frenchspacing\@vobeyspaces \@multilineverbatim
You are using the "multiline" environment in a style in which it is not defined.}
\let\endmultiline=\endtrivlist
\@namedef{multiline*}{\@verbatim\@smultilineverbatim
You are using the "multiline*" environment in a style in which it is not defined.}
\expandafter\let\csname endmultiline*\endcsname =\endtrivlist
\def\arrax{\@verbatim \frenchspacing\@vobeyspaces \@arraxverbatim
You are using a type of "array" construct that is only allowed in AmS-LaTeX.}
\let\endarrax=\endtrivlist
\def\tabulax{\@verbatim \frenchspacing\@vobeyspaces \@tabulaxverbatim
You are using a type of "tabular" construct that is only allowed in AmS-LaTeX.}
\let\endtabulax=\endtrivlist
\@namedef{arrax*}{\@verbatim\@sarraxverbatim
You are using a type of "array*" construct that is only allowed in AmS-LaTeX.}
\expandafter\let\csname endarrax*\endcsname =\endtrivlist
\@namedef{tabulax*}{\@verbatim\@stabulaxverbatim
You are using a type of "tabular*" construct that is only allowed in AmS-LaTeX.}
\expandafter\let\csname endtabulax*\endcsname =\endtrivlist
\def\endequation{%
\ifmmode\ifinner
\iftag@
\addtocounter{equation}{-1}
$\hfil
\displaywidth\linewidth\@taggnum\egroup \endtrivlist
\global\@ifnextchar*{\@tagstar}{\@tag}@false
\global\@ignoretrue
\else
$\hfil
\displaywidth\linewidth\@eqnnum\egroup \endtrivlist
\global\@ifnextchar*{\@tagstar}{\@tag}@false
\global\@ignoretrue
\fi
\else
\iftag@
\addtocounter{equation}{-1}
\eqno \hbox{\@taggnum}
\global\@ifnextchar*{\@tagstar}{\@tag}@false%
$$\global\@ignoretrue
\else
\eqno \hbox{\@eqnnum
$$\global\@ignoretrue
\fi
\fi\fi
}
\newif\iftag@ \@ifnextchar*{\@tagstar}{\@tag}@false
\def\@ifnextchar*{\@TCItagstar}{\@TCItag}{\@ifnextchar*{\@TCItagstar}{\@TCItag}}
\def\@TCItag#1{%
\global\@ifnextchar*{\@tagstar}{\@tag}@true
\global\def\@taggnum{(#1)}%
\global\def\@currentlabel{#1}}
\def\@TCItagstar*#1{%
\global\@ifnextchar*{\@tagstar}{\@tag}@true
\global\def\@taggnum{#1}%
\global\def\@currentlabel{#1}}
\@ifundefined{tag}{
\def\@ifnextchar*{\@tagstar}{\@tag}{\@ifnextchar*{\@tagstar}{\@tag}}
\def\@tag#1{%
\global\@ifnextchar*{\@tagstar}{\@tag}@true
\global\def\@taggnum{(#1)}}
\def\@tagstar*#1{%
\global\@ifnextchar*{\@tagstar}{\@tag}@true
\global\def\@taggnum{#1}}
}{}
\def\tfrac#1#2{{\textstyle {#1 \over #2}}}%
\def\dfrac#1#2{{\displaystyle {#1 \over #2}}}%
\def\binom#1#2{{#1 \choose #2}}%
\def\tbinom#1#2{{\textstyle {#1 \choose #2}}}%
\def\dbinom#1#2{{\displaystyle {#1 \choose #2}}}%
\makeatother
\endinput
|
\section{introduction}
Let $\mathcal{A}$ be a Banach algebra and $a,b\in\mathcal A$.
The map $M_{a,b}:\mathcal A\rightarrow\mathcal A$ given by $M_{a,b}(x)=axb$ is called a \emph{multiplication operator}.
Properties of compact
multiplication operators
have been investigated since 1964 when Vala published his work ``On compact sets of compact operators'' \cite{vala64}. Let
$\mathcal{X}$ be a normed space and $\mathcal{B}(\mathcal{X})$ the algebra of all bounded linear maps from $\mathcal{X}$ into
$\mathcal{X}$. Vala proved that a non-zero multiplication operator
$M_{a,b}:\mathcal{B}(\mathcal{X})\rightarrow\mathcal{B}(\mathcal{X})$
is
compact if and only if the operators $a, b \in\mathcal{B}(\mathcal{X})$ are both
compact.
Also, in \cite{vala67} Vala defines an element $a$ of a normed algebra to be \emph{compact}
if the mapping $x\mapsto axa$ is compact.
This concept enabled the study of compactness properties of elements of abstract normed algebras.
Ylinen in \cite{1972} studied compact elements for abstract C*-algebras and
showed that $a$ is a compact element of a $C^*$-algebra $\mathcal{A}$ if and only if there
exists an isometric $*$-representation $\pi$ of $\mathcal{A}$ on a Hilbert
space $\mathcal H$ such
that the operator $\pi(a)$ is compact.
Compactness questions have also been considered in the more general framework
of elementary operators. A map
$\Phi:\mathcal{A}\rightarrow\mathcal{A}$, where $\mathcal A$ is a Banach algebra, is called \textit{elementary} if $\Phi=\sum_{i=1}^m M_{a_i,b_i}$ for some
$a_i,b_i\in\mathcal{A}$, $i=1,\ldots,m$. Fong and Sourour showed that an elementary operator
$\Phi:\mathcal{B}(\mathcal H)\rightarrow\mathcal{B}(\mathcal H)$, where $\mathcal{B}(\mathcal H)$ is the algebra of bounded linear operators
on a Hilbert space $\mathcal H$, is compact if and only if there exist compact operators
$c_i,d_i\in\mathcal{B}(\mathcal H)$, $i=1,\ldots,m$ such that $\Phi=\sum_{i=1}^m M_{c_i,d_i}$ \cite{fs}. This result was expanded by
Mathieu on prime C*-algebras \cite{m} and later on general C*-algebras by Timoney \cite{tim}.
Akemann and Wright
\cite{ake} characterized the weakly compact multiplication operators on $\mathcal{B}(\mathcal H)$,
where $\mathcal H$ is a Hilbert space.
Saksman and Tylli \cite{st, st1}
and
Johnson and Schechtman \cite{js} studied weak compactness of multiplication operators in a Banach space setting.
Moreover, strictly singular multiplication operators are studied by
Lindstr\"{o}m, Saksman and Tylli \cite{lst} and
Mathieu and Tradacete \cite{mt}.
Compactness properties of multiplication operators on nest algebras, a class of non selfadjoint operator algebras,
are studied by Andreolas and Anoussis
in \cite{anan}.
In particular they characterized the compact multiplication operators, the compact elements and the ideal generated by the compact
elements.
In the present paper we study multiplication operators on a semicrossed product $C_0(X)\times_\phi\mathbb Z_+$ where $X$
is a locally compact metrizable space, and $\phi: X\rightarrow X$ a homeomorhism.
We characterize the compact
multiplication operators in terms of
the corresponding dynamical system. As a consequence, we obtain a characterization of the compact
elements of the semicrossed product.
We also characterize the ideal generated by the compact elements.
We would like to note that the equicontinuity condition appearing in the characterization of
the compact multiplication operators on the semicrossed product, follows from the other conditions
if $X$ is discrete
or has no isolated points. However,
in the general case this does not hold and thus
the proof is considerably more elaborated.
\section{Compact multiplication operators on semicrossed products}
Throughout this paper, $X$ will be a locally compact metrisable space and $\phi:X\rightarrow X$ a homeomorphism. The pair $(X, \phi)$
is called a dynamical system.
An action
of $\mathbb{Z}_+$ on $C_0(X)$ by isometric $*$-automorphisms $\alpha_n$, $n\in\mathbb{Z}_+$ is obtained by defining
$\alpha_n(f)=f\circ\phi^n$.
We write the elements of the Banach space $\ell^1(\mathbb
Z_+,C_0(X))$ as formal series $A=\sum_{n\in\mathbb Z_+}U^nf_n$ with the norm given by
$\|A\|_1=\sum\|f_n\|_{C_0(X)}$. The multiplication on $\ell^1(\mathbb Z_+,C_0(X))$ is
defined by setting
$$U^nfU^mg=U^{n+m}(\alpha^m(f)g)$$
and extending by linearity and continuity. With this multiplication, $\ell^1(\mathbb Z_+,C_0(X))$ is a Banach algebra.
The Banach algebra $\ell^1(\mathbb Z_+,C_0(X))$ can be faithfully represented as a (concrete) operator algebra on a
Hilbert space. This is achieved by assuming a faithful action of $C_0(X)$ on a Hilbert space
$\mathcal{H}_0$. Then, we can define a faithful contractive representation $\pi$ of $\ell_1(\mathbb Z_+,C_0(X))$ on the Hilbert
space $\mathcal H=\mathcal{H}_0\otimes \ell^2(\mathbb Z_+)$ by defining $\pi(U^nf)$ as
$$\pi(U^nf)(\xi\otimes e_k)=\alpha^k(f)\xi\otimes e_{k+n}.$$
The \emph{semicrossed product} $C_0(X)\times_{\phi}\mathbb Z_+$ is the closure of the image of
$\ell^1(\mathbb Z_+,C_0(X))$ in $\mathcal{B(H)}$ in the representation just defined, where $\mathcal{B(H)}$ is the algebra of bounded linear operators on $\mathcal{H}$. We will denote the semicrossed product $C_0(X)\times_{\phi}\mathbb Z_+$ by $\mathcal{A}$ and an element $\pi(U^nf)$ of $\mathcal{A}$ by $U^nf$ to simplify the notation. The closed unit ball of $\mathcal{A}$ will be denoted by $\mathcal{A}_1$. We refer to \cite{dfk} and \cite{dkm}, for more information about the semicrossed product.
For $A=\sum_{n\in\mathbb Z_+}U^nf_n\in \ell^1(\mathbb Z_+,C_0(X))$ we call $f_n\equiv
E_n(A)$ the $n$th \emph{Fourier coefficient} of $A$. The maps $E_n:\ell^1(\mathbb N_+,C_0(X))\rightarrow C_0(X)$ are contractive in the (operator) norm of $\mathcal A$, and therefore they extend to
contractions $E_n:\mathcal A \rightarrow C_0 (X)$.
Let $A\in\mathcal A$. If the set $\{m\in\mathbb Z_+:E_m(A)\neq0\}$ is finite, then $A$ is called a \emph{polynomial}. If there exists a unique $m\in\mathbb Z_+$, such that $E_m(A)\neq0$, then $A$ is called \emph{monomial}.
Let $(X,\phi)$ be a dynamical system. Then, a point $x\in X$ is called \textit{recurrent} if
there exists a strictly increasing sequence $(n_k)_{k\in\mathbb{N}}\subseteq \mathbb{N}$, such
that $\lim_{k\rightarrow\infty}\phi^{n_k}(x)=x$. The set of the recurrent points of $(X,\phi)$ will be denoted by $X_r$.
We will denote by
$X_i$ the set of the isolated points of $X$, by
$X_a$ the set of the accumulation points of $X$ and we set
$X_{a,i}=\{x\in X_{a}:\exists (x_j)\subseteq X_i,\ \lim_{j\rightarrow\infty}x_j= x\}$. If $f\in C_0(X)$, we set
$\mathrm \mathrm D(f)=\{x\in X:|f(x)|>0\}$.
\begin{lemma}\label{cmo1}
Let $M_{A,B}:\mathcal A\rightarrow \mathcal A$ be a
compact multiplication operator, where $A,B\in\mathcal A_1$ and $E_m(A)=f_m$, $E_m(B)=g_m$,
for all $m\in\mathbb Z_+$. Then, $(f_m\circ\phi^{n+l}g_n)(X_a)=\{0\}$, for all $m, n, l \in\mathbb{Z_+}$.
\end{lemma}
\begin{proof}
We suppose that there exist $m, n, l\in\mathbb Z_+$ such that
$\left(f_{m}\circ\phi^{n+l}g_{n}\right)(X_a)\neq \{0\}.
$ We define the following indices.
\begin{eqnarray*}
n_0 & = & \min\left\{n\in\mathbb Z_+|\;\exists\; m,l\in\mathbb Z_+:\left(f_m\circ\phi^{n+l}g_n\right)(X_a)\neq \{0\}\right\},\\
m_0 & = & \min\left\{m\in\mathbb Z_+|\;\exists\; l\in\mathbb Z_+:\left(f_m\circ\phi^{n_0+l}g_{n_0}\right)(X_a)\neq \{0\}\right\},\\
l_0 & = & \min\left\{l\in\mathbb Z_+|\left(f_{m_0}\circ\phi^{n_0+l}g_{n_0}\right)(X_a)\neq \{0\}\right\}.
\end{eqnarray*}
Let $x_0\in X_a$, such that $\left(f_{m_0}\circ\phi^{n_0+l_0}g_{n_0}\right)(x_0)\neq 0$.
Then, there exist an $\epsilon>0$ and an open neighborhood $U_0$ of $x_0$ such that $\left|\left(f_{m_0}\circ\phi^{n_0+l_0}g_{n_0}\right)(x)\right|>2\epsilon$, for all $x\in U_0$. Now, we consider the quantity
\begin{eqnarray*}
\bar g=\inf_{x\in U_0}\{|g_{n_0}(x)|\}>0.
\end{eqnarray*}
We consider the function $J:X\rightarrow \mathbb{C}$, defined as follows.
\begin{eqnarray*}
J(x)=\left\{\begin{matrix} 0& \hspace{-1em}, & \mathrm {if}\;\; m_0+n_0=0 \\
\displaystyle \sum_{n=0,n\neq n_0}^{n_0+m_0}\left|(f_{m_0+n_0-n}\circ\phi^{n+l_0}g_ng_{n_0})(x)\right| & \hspace{-1em}, &\mathrm{if}\;\; m_0+n_0>0. \end{matrix} \right.
\end{eqnarray*}
If $m_0+n_0>0$, we claim that $J(x_0) = 0$.
Indeed, if $n<n_0$, then it follows from the definition of $n_0$ that $(f_{m_0+n_0-n}\circ\phi^{n+l_0}g_n)(X_a)=\{0\}$. Otherwise,
if $n>n_0$ and $n\leq m_0+n_0$, it follows from the definition of $m_0$ that $(f_{m_0+n_0-n}\circ\phi^{l_0+n}g_{n_0})(X_a)=\{0\}$, since $m_0+n_0-n<m_0$. Therefore, there exists an open neighborhood $V_0$ of $x_0$, such that
$J(x) < \epsilon\bar g$, for all $x\in V_0,$ by the continuity of $J$.
If $m_0+n_0>0$, we set $W_0=U_0\cap V_0$, otherwise, we set $W_0=U_0$.
Since $x_0\in W_0\cap X_a$, there exist a sequence of points $\{x_i\}_{i\in\mathbb{N}}\subseteq W_0$,
a sequence of open subsets $\{W_i\}_{i\in\mathbb{N}}\subset W_0$ with $x_i\in W_i$ and $W_i\cap W_j=\emptyset$,
for $i\neq j$ and a sequence of norm one
functions $\{h_i\}_{i\in\mathbb{N}}\subseteq C_{0}(X)$ with $\mathrm D(h_i)\subseteq W_i$
and $h_i(x_i)=1$, for all $i\in\mathbb{N}$.
To complete the proof, we consider the sequence $\{U^{l_0}h_i\circ\phi^{-n_0}\}_{i\in\mathbb{N}}$ and we will prove that the sequence
$\{M_{A, B}(U^{l_0}h_i\circ\phi^{-n_0})\}_{i\in\mathbb{N}}$ has no convergent subsequence.
We estimate the quantity $\|M_{A,B}\left(U^{l_0}h_u\circ\phi^{-n_0}\right)-M_{A,B}\left(U^{l_0}h_v\circ\phi^{-n_0}\right)\|_{\mathcal A}$, for $u,v\in\mathbb{N}$, $u\neq v$.
\begin{eqnarray*}
\left\|M_{A,B}\left(U^{l_0}h_u\circ\phi^{-n_0}\right)-M_{A,B}\left(U^{l_0}h_v\circ\phi^{-n_0}\right)\right\|_{\mathcal A} & \ge & \\
\left\|E_{m_0+l_0+n_0}\left(M_{A,B}\left(U^{l_0}h_u\circ\phi^{-n_0}-U^{l_0}h_v\circ\phi^{-n_0}\right)\right)\right\|_{C_0(X)} & = &\\
\left\|\sum_{n=0}^{m_0+n_0}f_{m_0+n_0-n}\circ\phi^{n+l_0}g_n(h_u\circ\phi^{n-n_0}-h_v\circ\phi^{n-n_0})\right\|_{C_0(X)}&\ge&\\
\left\|\sum_{n=0}^{m_0+n_0}f_{m_0+n_0-n}\circ\phi^{n+l_0}g_ng_{n_0}(h_u\circ\phi^{n-n_0}-h_v\circ\phi^{n-n_0})\right\|_{C_0(X)} & \ge &\\
\left|\sum_{n=0}^{m_0+n_0}\left(f_{m_0+n_0-n}\circ\phi^{n+l_0}g_ng_{n_0}(h_u\circ\phi^{n-n_0}-h_v\circ\phi^{n-n_0})\right)(x_u)\right| & \ge &\\
\left|\left(f_{m_0}\circ\phi^{n_0+l_0}g^2_{n_0}\right)(x_u)\right|& &\\
-\sum_{n=0,n\neq n_0}^{m_0+n_0}\left|\left(f_{m_0+n_0-n}\circ\phi^{n+l_0}g_ng_{n_0}(h_u\circ\phi^{n-n_0}-h_v\circ\phi^{n-n_0})\right)(x_u)\right|.
\end{eqnarray*}
We note that $|(h_u\circ\phi^{n-n_0}-h_v\circ\phi^{n-n_0})(x_u)|\leq 1$, since $\mathrm D(h_u)\cap \mathrm D(h_v)=\emptyset$. Therefore, we obtain
\begin{eqnarray*}
\left\|M_{A,B}\left(U^{l_0}h_u\circ\phi^{-n_0}\right)-M_{A,B}\left(U^{l_0}h_v\circ\phi^{-n_0}\right)\right\|_{\mathcal A} & \ge & \\
\left|\left(f_{m_0}\circ\phi^{n_0+l_0}g^2_{n_0}\right)(x_u)\right|-\sum_{n=0,n\neq n_0}^{m_0+n_0}\left|(f_{m_0+n_0-n}\circ\phi^{n+l_0}g_ng_{n_0})(x_u)\right| & = &\\
\left|\left(f_{m_0}\circ\phi^{n_0+l_0}g^2_{n_0}\right)(x_u)\right|-J(x_u) & > & \epsilon\bar g,
\end{eqnarray*}
which concludes the proof.
\end{proof}
\begin{lemma}\label{cmo2}
Let $M_{A,B}:\mathcal A\rightarrow \mathcal A$ be a compact multiplication operator, where $A,B\in \mathcal A_1$ and $E_m(A)=f_m$, $E_m(B)=g_m$, for all $m\in\mathbb Z_+$.
Then, $\lim_{l\rightarrow\infty}(f_m\circ\phi^{n+l}g_n)(x)= 0$, for all $m,n\in\mathbb Z_+$ and $x\in X_i$.
\end{lemma}
\begin{proof}
We suppose that there exist $n,m\in\mathbb Z_+$ and $x\in X_i$, such that $\lim_{l\rightarrow\infty}(f_m\circ\phi^{n+l}g_n)(x)\neq 0$. We define the following indices.
\begin{eqnarray*}
n_0 & = & \min\left\{n\in\mathbb Z_+ | \exists m\in\mathbb Z_+,\ x\in X_i:\lim_{l\rightarrow\infty}(f_m\circ\phi^{n+l}g_n)(x)\neq 0 \right\},\\
m_0 & = & \min\left\{m\in\mathbb Z_+ | \exists x\in X_i:\lim_{l\rightarrow\infty}(f_m\circ\phi^{n_0+l}g_{n_0})(x)\neq 0 \right\}.
\end{eqnarray*}
Then, there exist an element $x_0\in X_i$, an $\epsilon>0$ and a strictly increasing sequence $\{l_i\}_{i\in\mathbb{N}}\subset\mathbb Z_+$, such that
$\left|(f_{m_0}\circ\phi^{n_0+l_i}g_{n_0})(x_{0})\right| > 2\epsilon.$ We consider the sequence $\{J(l)\}_{l\in\mathbb{Z}_+}$,
\begin{eqnarray*}
J(l)=\left\{\begin{matrix} 0& \hspace{-1em}, &\mathrm{if} \;\; m_0+n_0=0 \\
\displaystyle \sum_{n=0,n\neq n_0}^{n_0+m_0}\left|(f_{m_0+n_0-n}\circ\phi^{n+l}g_ng_{n_0})(x_0)\right| & \hspace{-1em}, &\mathrm{if} \;\; m_0+n_0>0. \end{matrix} \right.
\end{eqnarray*}
If $m_0+n_0>0$, we claim that there exists an $L\in\mathbb N$, such that:
\begin{eqnarray*}
J(l)=\sum_{n=0 ,n\neq n_0}^{n_0+m_0}\left|(f_{m_0+n_0-n}\circ\phi^{n+l}g_ng_{n_0})(x_{0})\right| < \epsilon g_{n_0}(x_0), & \forall\; l>L.
\end{eqnarray*}
Indeed, if $n<n_0$, it follows from the definition of $n_0$, that $\lim_{l\rightarrow\infty}(f_{m_0+n_0-n}\circ\phi^{n+l}g_n)(x_{0})= 0$.
On the other hand, if $n>n_0$ and $n\leq m_0+n_0$,
it follows from the definition of $m_0$, that $\lim_{l\rightarrow\infty}(f_{m_0+n_0-n}\circ\phi^{n+l}g_{n_0})(x_{0})= 0$, since $m_0+n_0-n<m_0$.
We choose a subsequence $\{l_{i_k}\}_{k\in\mathbb{N}}\subseteq\{l_i\}_{i\in\mathbb{N}}\subseteq\mathbb Z_+$, such
that $l_{i_1}>L$ and $l_{i_{k+1}}-l_{i_{k}}>m_0+n_0+1$, for all $k\in\mathbb{N}$.
We consider the sequence $\left\{U^{l_{i_k}}\chi\right\}_{k\in\mathbb{N}}$, where $\chi$
is the characteristic function of the singleton $\{\phi^{n_0}(x_0)\}$. To complete the proof we will prove that the
sequence $\left\{M_{A,B}\left(U^{l_{i_{k}}}\chi \right)\right\}_{k\in\mathbb{N}}$ has no convergent subsequence. We estimate the quantity $\left\|M_{A,B}\left(U^{l_{i_{u}}}\chi\right)-M_{A,B}\left(U^{l_{i_{v}}}\chi\right)\right\|_{\mathcal A}$, for $u,v\in\mathbb{N}$, $u< v$.
\begin{eqnarray*}
\left\|M_{A,B}\left(U^{l_{i_{u}}}\chi\right)-M_{A,B}\left(U^{l_{i_{v}}}\chi\right)\right\|_{\mathcal A} & \ge &\\
\left\|E_{m_0+n_0+l_{i_{u}}}\left(M_{A,B}\left(U^{l_{i_{u}}}\chi\right)-M_{A,B}\left(U^{l_{i_{v}}}\chi\right)\right)\right\|_{C_0(X)} & = &\\
\left\|E_{m_0+n_0+l_{i_{u}}}\left(M_{A,B}\left(U^{l_{i_{u}}}\chi\right)\right)\right\|_{C_0(X)},
\end{eqnarray*}
since by the assumption $l_{i_{v}}-l_{i_{u}}>m_0+n_0+1$, the $(m_0+n_0+l_{i_{u}}){\mathrm {th}}$ Fourier coefficient of
$M_{A,B}\left(U^{l_{i_{v}}}\chi\right)$ is $0$.
We thus obtain
\begin{eqnarray*}
\left\|M_{A,B}\left(U^{l_{i_{u}}}\chi\right)-M_{A,B}\left(U^{l_{i_{v}}}\chi\right)\right\|_{\mathcal A} & \ge &\\
\left\|\sum_{n=0}^{m_0+n_0}f_{m_0+n_0-n}\circ\phi^{n+l_{i_{u}}}\chi\circ\phi^{n}g_n\right\|_{C_0(X)}& \ge &\\
\left\|\sum_{n=0}^{m_0+n_0}f_{m_0+n_0-n}\circ\phi^{n+l_{i_{u}}}\chi\circ\phi^{n}g_ng_{n_0}\right\|_{C_0(X)}& \ge &\\
\left|\sum_{n=0}^{m_0+n_0}(f_{m_0+n_0-n}\circ\phi^{n+l_{i_{u}}}\chi\circ\phi^{n}g_ng_{n_0})(x_0)\right|& \ge &\\
|(f_{m_0}\circ\phi^{n_0+l_{i_{u}}}g^2_{n_0})(x_{0})|-\sum_{n=0, n\neq n_0}^{m_0+n_0}|(f_{m_0+n_0-n}\circ\phi^{n+l_{i_{u}}}\chi\circ\phi^{n}g_ng_{n_0})(x_0)|& \ge &\\
\left|(f_{m_0}\circ\phi^{n_0+l_{i_{u}}}g^2_{n_0})(x_{0})\right|-J(l_{i_{u}}) & > &\\\epsilon g_{n_0}(x_0),&&
\end{eqnarray*}
which concludes the proof.
\end{proof}
\begin{lemma}\label{rl2}
Let $M_{A,B}:\mathcal A\rightarrow \mathcal A$ be a compact multiplication operator, where $A,B\in \mathcal A_1$ and $E_m(A)=f_m$, $E_m(B)=g_m$, for all $m\in\mathbb Z_+$.
Then, $(f_m\circ\phi^{n+l}g_n)(x)= 0$, for all $m,n,l\in\mathbb Z_+$ and $x\in X_r$.
\end{lemma}
\begin{proof}
If $x\in X_i\cap X_r$, there exists a $k_0\in\mathbb Z_+$ such that $\phi^{k_0}(x)=x$.
It follows from Lemma \ref{cmo2} that
$\lim_{i\rightarrow\infty}(f_m\circ\phi^{n+l+ik_0}g_n)(x)= 0$, for $m, n, l\in\mathbb Z_+$. We note that $(f_m\circ\phi^{n+l+ik_0}g_n)(x)=
(f_m\circ\phi^{n+l}g_n)(x)$, for all $ i \in \mathbb Z_+$, and hence
$(f_m\circ\phi^{n+l}g_n)(x)=0.$ If $x \in X_a\cap X_r$, the assertion follows from Lemma \ref{cmo1}.
\end{proof}
\begin{lemma}\label{cmo3}
Let $M_{A,B}:\mathcal A\rightarrow \mathcal A$ be a compact multiplication operator, where
$A,B\in \mathcal A_1$ and $E_m(A)=f_m$, $E_m(B)=g_m$, for all $m\in\mathbb Z_+$.
Then, the sequence $\{f_m\circ\phi^{n+l}g_n\}_{l\in\mathbb{Z}_+}$ is pointwise equicontinuous, for all $m,n\in\mathbb Z_+$.
\end{lemma}
\begin{proof}
It follows from Lemma \ref{cmo1} that it is sufficient to prove that
$\{f_m\circ\phi^{n+l}g_n\}_{l\in\mathbb{Z}_+}$ is pointwise equicontinuous on $X_{a, i}$.
We suppose that there exist some $n_0, m_0 \in\mathbb Z_+$ and a point $x_0\in X$, such that the sequence $\{f_{m_0}\circ\phi^{n_0+l}g_{n_0}\}_{l\in\mathbb{Z}_+}$ is not equicontinuous at $x_0$.
We note that $\left(f_{m_0}\circ\phi^{n_0+l}g_{n_0}\right)(x_0)=0$, for all $l\in\mathbb{Z}_+$,
by Lemma \ref{cmo1}. Therefore, there exist an $\epsilon>0$,
a strictly increasing sequence $\{l_i\}_{i\in\mathbb{N}}\subset\mathbb Z_+$ and a
sequence $\{x_i\}_{i\in\mathbb{N}}\subseteq X$, such that $\lim_{i\rightarrow\infty} x_i= x_0$ and
$\left|(f_{m_0}\circ\phi^{n_0+l_i}g_{n_0})(x_{i})\right|>\epsilon$, for all $\; i\in\mathbb{N}.$ We note that the inclusion $\{x_i\}_{i\in\mathbb{N}}\subseteq X_i$ holds by Lemma \ref{cmo1}.
Furthermore, we may assume that $l_{i+1}-l_i>m_0+n_0+1$, for all $i\in\mathbb{N}$. It follows from
Lemma \ref{rl2} that the elements $\{x_i\}_{i\in\mathbb{N}}$ are not periodic. Therefore, if $m_0+n_0>0$,
we have that $\phi^{n_0}(x_i)\neq \phi^{n}(x_i)$, for all $n\in\{0,\dots,m_0+n_0\}\smallsetminus\{n_0\}$ and $i\in\mathbb{N}$.
We consider the sequence $\left\{U^{l_{i}}\chi_i\right\}_{i\in\mathbb{N}}$, where $\chi_i$ is the characteristic function of the set
$\{\phi^{n_0}(x_i)\}$. Then, for $u< v$ we obtain,
\begin{eqnarray*}
\left\|M_{A,B}\left(U^{l_{u}}\chi_{u}\right)-M_{A,B}\left(U^{l_{v}}\chi_v\right)\right\|_{\mathcal A} & \ge &\\
\left\|E_{m_0+n_0+l_{u}}\left(M_{A,B}\left(U^{l_{u}}\chi_u\right)-M_{A,B}\left(U^{l_{v}}\chi_v\right)\right)\right\|_{C_0(X)} & = &\\
\left\|E_{m_0+n_0+l_{u}}\left(M_{A,B}\left(U^{l_{u}}\chi_u\right)\right)\right\|_{C_0(X)} & \ge &\\
\left|E_{m_0+n_0+l_{u}}\left(M_{A,B}\left(U^{l_{u}}\chi_u\right)\right)(x_{u})\right| & = &\\
\left|\sum_{n=0}^{m_0+n_0}(f_{m_0+n_0-n}\circ\phi^{n+l_{u}}\chi_{u} \circ \phi^{n}g_n)(x_{u})\right|& = &\\
\left|(f_{m_0}\circ\phi^{n_0+l_{u}}g_{n_0})(x_{u})\right| & > &\epsilon,
\end{eqnarray*}
since $x_u$ is not periodic.
Therefore, the sequence $\left\{M_{A,B}\left(U^{l_{i}}\chi_i\right)\right\}_{k\in\mathbb{N}}$ has
no Cauchy subsequence.
\end{proof}
\begin{proposition}\label{compactelements}
Let $m, n \in \mathbb Z_+$ and $A=U^mf,\;B=U^ng\in\mathcal A_1$. Then,
the multiplication operator $M_{A,B}:\mathcal A\rightarrow \mathcal A$ is compact if and only if the following assertions are valid.
\begin{enumerate}
\item $(f\circ\phi^{n+l}g)(X_a)=\{0\}$, for all $l\in\mathbb Z_+$,
\item $\lim_{l\rightarrow\infty}(f\circ\phi^{n+l}g)(x)= 0$, for all $x\in X_i$,
\item The sequence $\{f\circ\phi^{n+l}g\}_{l\in\mathbb{Z}_+}$ is pointwise equicontinuous.
\end{enumerate}
\end{proposition}
\begin{proof}
The forward direction is immediate by Lemmas \ref{cmo1}, \ref{cmo2} and \ref{cmo3}. We will show the opposite direction.
We will divide the proof in three steps:
\vspace{0.5em}
\textbf{1st step}
In this step we construct an approximation of $M_{A,B}$ by multiplication operators $M_{A_k, B_k}$
with the property that the Fourier coefficients
of $A_k, B_k$ are compactly supported.
We define the following sets, for $h\in C_0(X)$ and $k\in\mathbb N$.
$$D_{k}(h)=\left\{x\in X:|h(x)|\ge\frac{1}{k}\right\},\
U_{k}(h)=\left\{x\in X:|h(x)|>\frac{2}{3k}\right\}.$$ It is obvious that $D_{k}(h)\subseteq U_{k}(h)\subseteq D_{2k}(h)$ and that the set $\overline{U_{k}(h)}$ is compact.
If $\|f\|\|g\|=0$, the proof is trivial. Otherwise, we choose a natural number $k_s$, such that $\frac{1}{k_s}<\min\{\|f\|,\|g\|\}$. By Urysohn's lemma,
there are norm one functions $v_{f_k}$ and $v_{g_k}$ in $C_0(X)$, for $k>k_s$, such that
\begin{eqnarray*}
v_{f_k}(x) = \left\{\begin{tabular}{l} $1$ , $x\in D_k(f)$ \\ $0$ , $x\in X\setminus U_k(f)$\end{tabular}\right.
&\text{and} &
v_{g_k}(x) = \left\{\begin{tabular}{l} $1$ , $x\in D_k(g)$ \\ $0$ , $x\in X\setminus U_k(g)$\end{tabular}\right..
\end{eqnarray*}
We define the functions
$f_{k}=v_{f_k}f ,\; g_{k}=v_{g_k}g.$
It is immediate that $\|f-f_{k}\|\leq\frac{2}{k}$ and $\|g-g_{k}\|\leq\frac{2}{k}$. It follows that, if $A_k=U^mf_{k}$
and $B_k=U^ng_{k}$, we have that $\|A-A_k\|<\frac{2}{k}$ and $\|B-B_k\|<\frac{2}{k}$. Then, we can see that
\begin{equation*}
\sup_{T\in\mathcal A_1}\|M_{A,B}(T)-M_{A_k,B_k}(T)\|<\frac{4}{k}.
\end{equation*}
Hence, to prove that $M_{A,B}$ is compact,
it suffices to show that there exists a natural number $k_0$ such that $M_{A_k,B_k}$ is compact, for all $k>k_0$.
\vspace{0.5em}
\textbf{2nd step}
1st case
Firstly, we assume that $\overline{U_{k'}(g)}\cap X_{a,i}\neq\emptyset$,
for some $k' \in \mathbb Z_+$.
It follows that $\overline{U_{k}(g)}\cap X_{a,i}\neq\emptyset$, for all $k>k'$.
Let $k_0=\max\{k_s,k'\}$ and $k>k_0$. If $\tilde x\in \overline{U_{k}(g)}\cap X_{a,i}$, there exists an open neighbourhood
$V_{\tilde x}$ of $\tilde x$, such that $|g(x)|>\frac{1}{2k}$, for all $x\in V_{\tilde x}$.
Furthermore, $\tilde x$ is an accumulation point which in turn means that
$(f\circ\phi^{n+l}g)(\tilde x)=0$, for all $l\in\mathbb{Z}_+$. Moreover, we recall that the family
$\{f\circ\phi^{n+l}g\}_{l\in\mathbb{Z}_+}$ is equicontinuous at $\tilde x$.
Therefore, there exists an open neighbourhood $V_{\tilde x}'$ of $\tilde x$, such that
$|(f\circ\phi^{n+l}g)(x)|<\frac{1}{4k^2}$, for all $x\in V_{\tilde x}'$ and $l\in\mathbb{Z}_+$.
We set $W_{\tilde x}=V_{\tilde x}\cap V_{\tilde x}'$. It follows that
\begin{eqnarray*}
\frac{1}{4k^2} & > & |(f\circ\phi^{n+l}g)(x)|\\
&=& |f\circ\phi^{n+l}(x)||g(x)|\\
&\ge& |f\circ\phi^{n+l}(x)|\frac{1}{2k},
\end{eqnarray*}
and therefore, $|f\circ\phi^{n+l}(x)|\leq\frac{1}{2k}$, for all $x\in W_{\tilde x}$
and $l\in\mathbb{Z}_+$. We denote by $\cup_{\tilde x} W_{\tilde x}$ the set $\cup\{W_{\tilde{x}}\;|\; \tilde{x}\in\overline{U_k(g)}\cap X_{a,i}\}$. It follows that
$\phi^{n+l}\left(\cup_{\tilde x} W_{\tilde x}\right)\subseteq X\setminus \overline{U_k(f)}$, for all $l\in\mathbb{Z}_+$ and hence $(f_{k}\circ\phi^{n+l}g_{k})(\cup_{\tilde x}W_{\tilde x})=\{0\}$.
Moreover, the set $\overline{U_{k}(g)}\setminus(\cup_{\tilde x}W_{\tilde x})$ is compact and
$\left(\overline{U_{k}(g)}\setminus(\cup_{\tilde x}W_{\tilde x})\right)\cap X_{a,i}=\emptyset$,
which in turn implies that the set $(\overline{U_{k}(g)}\setminus(\cup_{\tilde x}W_{\tilde x}))\cap X_i$ is finite.
We denote by $I_k$ the set $(\overline{U_{k}(g)}\setminus(\cup_{\tilde x}W_{\tilde x}))\cap X_i$ and
by $\chi_{I_k}$ the characteristic function of $I_k$.
We set $\tilde g_{k}=g_k\chi_{I_k}$.
We note that $(f\circ\phi^{n+l}g)(X_a)=\{0\}$, for all $l\in\mathbb Z_+$ and hence
$(f_k\circ\phi^{n+l}g_k)(X_a)=\{0\}$, for all $l\in\mathbb Z_+$. Furthermore,
we have proved that $(f_{k}\circ\phi^{n+l}g_{k})(\cup_{\tilde x}W_{\tilde x})=\{0\}$. Since the function $g_{k}$ is supported in
$U_k(g)$, we conclude that
\begin{equation*}
f_{k}\circ\phi^{n+l}g_{k}=f_{k}\circ\phi^{n+l}\tilde g_{k},
\end{equation*}
for all $l\in\mathbb Z_+$.
2nd case
We assume now that
$ \overline{U_{k}(g)}\cap X_{a,i}=\emptyset,\forall k\in\mathbb N$.
The set $\overline{U_{k}(g)}\cap X_i$ is finite, since the set $\overline{U_{k}(g)}$ is compact and $\overline{U_{k}(g)}\cap X_{a,i}=\emptyset$.
We denote by the same letter as in the 1st case, $I_k$ the set $\overline{U_{k}(g)}\cap X_i$ and
by $\chi_{I_k}$ the characteristic function of $I_k$.
We set $\tilde g_{k}=g_k\chi_{I_k}$.
We note that $(f\circ\phi^{n+l}g)(X_a)=\{0\}$, for all $l\in\mathbb Z_+$ and hence
$(f_k\circ\phi^{n+l}g_k)(X_a)=\{0\}$, for all $l\in\mathbb Z_+$. Since $g_{k}$ is supported in
$U_k(g)$, we conclude that
\begin{equation*}
f_{k}\circ\phi^{n+l}g_{k}=f_{k}\circ\phi^{n+l}\tilde g_{k},
\end{equation*}
for all $l\in\mathbb Z_+$.
It follows that $M_{A_k,B_k}= M_{A_k,\tilde B_k}$, where
$\tilde B_k=U^n\tilde g_{k}$.
\vspace{0.5em}
\textbf{3rd step}
It follows that
$\lim_{l\rightarrow\infty}(f_k\circ\phi^{n+l}g_k)(y)=0$, for all $y \in I_k$, by assumption.
We observe that $\lim_{l\rightarrow\infty}(f_k\circ\phi^{n+l})(y)=0$, since $g_{k}(y)\neq0$, which follows from the inclusion $I_k\subseteq \overline{U_k(g_k)}$.
Therefore, there exists an $L_{0}\in\mathbb N$, such that
$\phi^{n+l}\left(y\right)\in (X\setminus \overline{U_{k}(f)})$, for all $l\ge L_{0}$ and for all $y \in I_k$, since the set $I_k$ is finite.
Since $f_k$ is supported in $\overline{U_{k}(f)}$, we obtain that $f_{k}\circ\phi^{n+l}\tilde g_{k}=0$, for all $l\ge L_0$.
It follows that
\begin{equation*}
M_{A_k,\tilde B_k}(U^lh)=0,
\end{equation*}
for all $l \ge L_0$ and $h \in C_0(X)$.
Since $\tilde g_{k}$ has finite support, the operator
$
M_{A_k,\tilde B_k},$ is a finite rank operator and hence compact.
\end{proof}
Let $A$ be an element of the semicrossed product $\mathcal{A}$.
We consider the sequence $\{U^nf_n\}_{n\in\mathbb{Z}_+}\subseteq\mathcal A$, where $f_n=E_n(A)$, for $n\in\mathbb{Z}_+$.
We note that the series $\sum_{n\in\mathbb{Z}_+} U^nf_n$ does not converge to $A$ in general.
The $k$th arithmetic mean of $A$ is defined to be the element $A_k=\frac{1}{k+1}\sum_{l=0}^k S_l$,
where $S_l=\sum_{n=0}^l U^nf_n$. Then, the sequence $\{A_k\}_{k\in\mathbb{Z}_+}$ is norm convergent to $A$ \cite[p. 524]{peters}.
\begin{theorem} \label{main2}
Let $A,B\in \mathcal A$ and $E_m(A)=f_m\in C_0(X)$, $E_m(B)=g_m\in C_0(X)$, for all $m\in\mathbb Z_+$.
The following statements are equivalent.
\begin{enumerate}
\item The multiplication operator $M_{A,B}:\mathcal A\rightarrow \mathcal A$ is compact.
\item The following assertions are valid, for all $m,n\in\mathbb Z_+$.
\begin{enumerate}
\item $(f_m\circ\phi^{n+l}g_n)(X_a)=\{0\}$, for all $l\in\mathbb Z_+$.
\item $\lim_{l\rightarrow\infty}(f_m\circ\phi^{n+l}g_n)(x)= 0$, for all $x\in X_i$.
\item The sequence $\{f_m\circ\phi^{n+l}g_n\}_{l\in\mathbb{Z_+}}$ is pointwise equicontinuous.
\end{enumerate}
\end{enumerate}
\end{theorem}
\begin{proof}
It is sufficient to prove the theorem for $A, B \in \mathcal A_1$.
The condition (1) implies the condition (2) by Lemmas \ref{cmo1}, \ref{cmo2} and \ref{cmo3}. We will show the opposite direction.
If $A=\sum_{m=0}^pU^mf_m$ and $B=\sum_{n=0}^qU^ng_n$, for $ p, q \in \mathbb Z_+$, we have
$$M_{A, B}=\sum_{m=0}^p\sum_{n=0}^qM_{U^mf_m, U^ng_n}$$ and the assertion folows
from Proposition \ref{compactelements}.
If $A, B \in \mathcal A$ and
$k\in\mathbb{Z}_+$, we denote by
$A_k$ and $B_k$ the $k$th arithmetic mean of $A$ and $B$ respectively. Since the Fourier coefficients of $A$ and $B$
satisfy the condition
$(2)$, the Fourier coefficients of $A_k$ and $B_k$ satisfy the condition $(2)$ as well.
Thus, the operator $M_{A_k, B_k}$
is compact, for all $k \in \mathbb Z_+$.
The operator $M_{A, B}$ is the norm limit of the sequence $\{M_{A_k, B_k}\}_{k\in\mathbb{Z}_+}$ and hence it is compact.
\end{proof}
As a corollary of the above theorem, we obtain the following characterization of the compact elements of the algebra $\mathcal A$.
\begin{corollary} \label{compelements}
Let $A\in\mathcal{A}$ and $E_m(A)=f_m\in C_0(X)$, for all $m\in Z_+$. Then,
$A$ is compact element of $\mathcal{A}$, if and only if the following conditions are satisfied, for all $m,n\in\mathbb{Z}_+$.
\begin{enumerate}
\item $(f_m\circ\phi^{n+l}f_n)(X_a)=\{0\}$, for all $l\in\mathbb Z_+$.
\item $f_m(X_r)=\{0\}$.
\item The sequence $\{f_m\circ\phi^{n+l}f_n\}_{l\in\mathbb{Z}_+}$ is pointwise equicontinuous.
\end{enumerate}
\end{corollary}
\begin{proof}
It is sufficient to prove the corollary for $A \in \mathcal A_1$.
Firstly, we will show the forward direction. The conditions (1) and (3) are satisfied by Theorem \ref{main2}.
Let $x$ be a recurrent point. Then,
there exists a strictly increasing sequence $\{l_i\}_{i\in\mathbb{N}}\subseteq\mathbb Z_+$, such that
$\lim_{i\rightarrow\infty}\phi^{m+l_i}( x)= x$, which implies that
$\lim_{i\rightarrow\infty}(f_m\circ\phi^{m+l_i}f_m)(x)=f_m^2(x). $ It follows from Lemma \ref{rl2} that $f_m(x)=0$.
Now, we will show the opposite direction. In view of Theorem \ref{main2}, it suffices to prove that
$\lim_{l\rightarrow\infty}(f_m\circ\phi^{n+l}f_n)(x)=0$, for all $m,n\in\mathbb Z_+$ and $x\in X_i$.
Let $m_0,n_0\in\mathbb Z_+$ and $x_0\in X_i$, such that $\lim_{l\rightarrow\infty}(f_{m_0}\circ\phi^{n_0+l}f_{n_0})(x_0)\neq 0$.
Then, there exists a natural number $k_0$ and a strictly increasing sequence $\{l_i\}_{i\in\mathbb{N}}\subseteq\mathbb Z_+$, such
that $|(f_{m_0}\circ\phi^{n_0+l_i}f_{n_0})(x_0)|\ge \frac{1}{k_0}$, for all $i\in\mathbb N$.
Hence, $|(f_{m_0}\circ\phi^{n_0+l_i})(x_0)|\ge \frac{1}{k_0}$, for all $i\in\mathbb N$, which implies that
$\phi^{n_0+l_i}(x_0)\in D_{k_0}(f_{m_0})=\{x\in X : f_{m_0}(x)\geq\frac{1}{k_0}\}$, for all $i\in\mathbb N$. By the condition (2), we obtain that $x_0\in X_i\setminus X_r$
and hence, $\phi^{n_0+l_i}(x_0)\neq \phi^{n_0+l_j}(x_0)$, for $i\neq j$. Moreover, the set $D_{k_0}(f_{m_0})$ is compact and
hence, there exists a point $\tilde x\in X_{a,i}\cap D_{k_0}(f_{m_0})$ and a subsequence
$\{l_{i_j}\}_{j\in\mathbb{N}}$, such that $\lim_{j\rightarrow\infty}\phi^{n_0+l_{i_j}}(x_0)= \tilde x$.
Let $W_{\tilde x}$ be an open neighbourhood of $\tilde x$,
such that $|f_{m_0}(x)|>\frac{1}{2k_0}$, for all $x\in W_{\tilde x}$.
By the condition (1), we have that $(f_{m_0}\circ\phi^{n_0+l}f_{m_0})(\tilde x)=0$, for all $l \in \mathbb N$ and
by (c) we have that the sequence $\{f_{m_0}\circ\phi^{n_0+l}f_{m_0}\}_{l\in\mathbb{Z}_+}$ is equicontinuous at $\tilde x$.
Therefore, there is an open neighbourhood $W_{\tilde x}'$ of $\tilde x$, such
that $|(f_{m_0}\circ\phi^{n_0+l}f_{m_0})(x)|<\frac{1}{5k_0^2}$, for all $l \in \mathbb Z_+$ and $x\in W_{\tilde x}'$.
Let $W=W_{\tilde x}\cap W_{\tilde x}'$ and $j_1,j_2\in\mathbb N$ be
such that $\phi^{n_0+l_{i_{j_1}}}(x_0),\phi^{n_0+l_{i_{j_2}}}(x_0)\in W$ and $l_{i_{j_2}}>l_{i_{j_1}}+n_0$. Then,
\begin{eqnarray*}
\frac{1}{5k_0^2}&\ge&|(f_{m_0}\circ\phi^{l_{i_{j_2}}-l_{i_{j_1}}}f_{m_0})(\phi^{n_0+l_{i_{j_1}}}(x_0))|\\
&=&|f_{m_0}(\phi^{n_0+l_{i_{j_2}}}(x_0))||f_{m_0}(\phi^{n_0+l_{i_{j_1}}}(x_0))|\ge\frac{1}{4k_0^2},
\end{eqnarray*}
which is a contradiction.
\end{proof}
Let us see now how Theorem \ref{main2} applies to two special cases. If $X$ is a discrete space or it has no isolated points, we obtain the following characterizations.
\begin{corollary}
Let $X$ be a discrete space, $A,B\in \mathcal A$ and $E_m(A)=f_m,E_m(B)=g_m$, for all $m\in\mathbb Z_+$.
Then, the following are equivalent.
\begin{enumerate}
\item The multiplication operator $M_{A,B}:\mathcal A\rightarrow \mathcal A$ is compact.
\item $(f_m\circ\phi^{n+l}g_n)(X_r)=\{0\}$, for all $m,n,l\in\mathbb Z_+$.
\item
$\lim_{l\rightarrow \infty}(f_m\circ\phi^{n+l}g_n)(x)= 0$, for all $x\in X .$
\end{enumerate}
\end{corollary}
\begin{proof}
We will show the implication $(2)\Rightarrow (3)$.
Assume that there exist $m, n \in \mathbb Z_+$ and
$x\in X\setminus X_r$, such that $\lim_{l\rightarrow\infty}(f_m\circ\phi^{n+l}g_n)(x)\neq0$.
Then, there exist an $\epsilon >0$ and a strictly increasing sequence $\{l_i\}_{i\in\mathbb{N}}\subseteq\mathbb Z_+$, such that
$|(f_m\circ\phi^{n+l_i}g_n)( x)|\ge\epsilon$, for all $i\in\mathbb N$.
Moreover, $ x\notin X_r$ and therefore
$\phi^{n+l_i}( x)\neq \phi^{n+l_j}( x)$, for $i\neq j$. This is a
contradiction, since $f_m \in C_0(X)$.
The implication $(3)\Rightarrow (1)$ follows from Theorem \ref{main2} and
the implication $(1)\Rightarrow (2)$ follows from Lemma \ref{rl2}.
\end{proof}
\begin{corollary}
Let $X$ be a space without isolated points, $A,B\in \mathcal A$ and $E_m(A)=f_m,E_m(B)=g_m$, for all $m\in\mathbb Z_+$.
Then, the following are equivalent.
\begin{enumerate}
\item The multiplication operator $M_{A,B}:\mathcal A\rightarrow \mathcal A$ is compact.
\item $(f_m\circ\phi^{n+l}g_n)(X)=\{0\}$ for all $l,m,n\in\mathbb N$.
\item $M_{A,B}=0$.
\end{enumerate}
\end{corollary}
\begin{proof}
It follows immediately from Theorem \ref{main2}.
\end{proof}
The next example shows that the equicontinuity condition can not be omitted in general.
\begin{example}
We consider the dynamical system $(X,\phi)$ where
\begin{eqnarray*}
X=\{0\}\cup\{x_n\}_{n\in\mathbb Z}\cup \{2\}, & & x_n=\left\{\begin{matrix}\frac{1}{|n|+1} & \hspace{-1em}, & n<0 \\ 2-\frac{1}{n+1} & \hspace{-1em}, & n\ge 0\end{matrix}\right.
\end{eqnarray*}
and $\phi$ is the homeomorphism
\begin{eqnarray*}
\phi(0)=0 , & \phi(x_n)=x_{n-1} , & \phi(2)=2.
\end{eqnarray*}
We define the monomials $A=U^1f$ and $B=U^1g$ of the semicrossed product $\mathcal{A}$ by the following formulae.
\begin{eqnarray*}
f(x)=\left\{\begin{matrix} 1, & &\hspace{-0.8em} x=1 \\ 0, & &\hspace{-1.5em} \text{else}\end{matrix}\right. & \text{and} & g(x)=\left\{\begin{matrix} 1, & &\hspace{-1em} x> 1 \\ 0, & &\hspace{-1em} x\leq 1\end{matrix}\right..
\end{eqnarray*}
We observe that, $(f\circ\phi^{1+l}g)(X_a)=\{0\}$, for all $l\in\mathbb Z_+$ and
$\lim_{l\rightarrow\infty}(f\circ\phi^{1+l}g)(x)=0$, for all $x\in X_i$.
However, the sequence $\{f\circ\phi^{1+l}g\}_{l\in\mathbb{Z}_+}$ is not equicontinuous at $x=2$ and the multiplication operator $M_{A,B}:\mathcal{A}\rightarrow \mathcal{A}$ is not compact.
\end{example}
In the following theorem, we characterize the ideal generated by the set of
compact elements of the semicrossed product $\mathcal{A}$. Recall that $Y\subseteq X$ is called wandering if the sets
$\phi^{-1}(Y), \phi^{-2}(Y), ...$ are pairwise disjoint. If $\phi$ is a homeomorphism, this condition is equivalent to
the condition that $\phi^m(Y)\cap \phi^n(Y)=\emptyset$ for all $m, n \in \mathbb Z_+, m\neq n$.
A point $x \in X$ is called wandering if it possesses an open wandering neighborhood. Otherwise it is called non wandering. If $x$
is a non wandering point and $W$ is an open neighbourhood of $x$, then there exists $m \in \mathbb N$, such
that $W\cap \phi^m(W)\neq \emptyset$. Note that we may assume that $m$ is arbitrarily large \cite[p. 129]{kh}.
We will denote by $X_w$ the set of wandering points of $X$. It is clear that $X_w$ is
the the union of all open wandering subsets of $X$.
\begin{theorem}\label{cmo31}
The ideal generated by the compact elements of $\mathcal{A}$ is
the set
$$
\{A\in\mathcal{A}\; |\; E_n(A)(X\setminus X_w)=\{0\},\; \forall n\in\mathbb{Z_+}\ \text{and } E_0(A)(X_a)=\{0\}\}.$$
\end{theorem}
\begin{proof}
Let $A$ be a compact element of $\mathcal{A}$.
We will show that
$E_n(A)(X\setminus X_w)=\{0\}$, for all $n\in\mathbb{Z_+}$ and $
E_0(A)(X_a)=\{0\}$.
Let us denote by $f_n$ the $n$th Fourier coefficient, $E_n(A)$, of $A$, for all $n\in\mathbb{Z}_+$.
We suppose that there exists an $n_0\in\mathbb Z_+$ and a point $x_0\in X\setminus X_{w}$, such that $f_{n_0}(x_0)\neq 0$.
Without loss of generality, we assume that $|f_{n_0}(x_0)|=1$.
We note that by Corollary \ref{compelements}, $x_0$ cannot be isolated, since a non wandering isolated
point is periodic and hence recurrent.
Thus, $x_0$ is an accumulation point, and we have $(f_{n_0}\circ\phi^{n_0+l}f_{n_0})(x_0)=0$,
for all $l\in\mathbb Z_+$. Let $0<\epsilon<1/4$. The sequence $\{f_{n_0}\circ\phi^{n_0+l}f_{n_0}\}_{l\in\mathbb{Z}}$
is equicontinuous at
$x_0$ and therefore, there exists an open neighborhood $U$ of $x_0$, such that $|(f_{n_0}\circ\phi^{n_0+l}f_{n_0})(x)|<\epsilon$, for all $x\in U$ and $l\in\mathbb Z_+$.
There exists an open neighborhood $V$ of $x_0$, such that $|f_{n_0}(x)|>1-\epsilon$, for all $x\in V$.
Let $W=U\cap V$. Since $x_0$ is a non wandering point, there exists an $l_0\in\mathbb Z_+$ such that $l_0\ge n_0$
and $W\cap\phi^{l_0}(W)\neq\emptyset$.
Let $y\in W$, such that $\phi^{l_0}(y)\in W$. Then,
\begin{eqnarray*}
\epsilon>|(f_n\circ\phi^{l_0}f_n)(y)|=|f_n(\phi^{l_0}(y))||f_n(y)|>(1-\epsilon)^2
\end{eqnarray*}
which is a contradiction. Furthermore, it is evident from Corollary \ref{compelements}, that $f_0(X_a)=\{0\}$.
We observe that the conditions $f_n(X\setminus X_w)=\{0\}$, for all $n\in\mathbb{Z}_+$ and $f_0(X_a)=\{0\}$,
we have already proved for a compact element $A$, are satisfied by the elements of the ideal generated by the
compact elements as well.
To complete the proof, we will prove that if $A \in \mathcal A$ satisfies
$E_n(A)(X\setminus X_w)=\{0\}$, for all $n\in\mathbb{Z_+}$\ \text{and } $E_0(A)(X_a)=\{0\}$, then $A $ belongs to
the ideal generated by the
compact elements.
We denote by $f_n$ the $n$th Fourier coefficient $E_n(A)$ of $A$, for all $n\in\mathbb{Z}_+$.
It is sufficient to show that $U^nf_n$
belongs to the ideal generated by the
compact elements, $\forall n\in\mathbb{Z_+}$.
First, we consider the case $n>0$.
Let $\mathcal C$ be the set
$\{h\in C_0(X): \mathrm D(h)=h^{-1}(\mathbb C\smallsetminus\{0\}) \textrm{ is wandering}\}$.
Let $h \in \mathcal C$.
Since $\mathrm D(h)$ is wandering and $n>0$, the functions
$(h\circ\phi^{n+l}h)$ are identically $0$, for every $l\in\mathbb Z_+$. Hence, $U^nh$ satisfies the conditions (a) and (c)
of Corollary \ref{compelements}. Since $X_r$ is contained in $X\setminus X_w$ condition (c) is also satisfied and $U^nh$
is a compact element of $\mathcal A$.
The norm closed algebra generated by $\mathcal C$,
is the ideal
$\{f\in C_0(X): f(x)=0,\; \forall x\in X\setminus X_w\}$ of $C_0(X)$. In particular, $f_n$ belongs to this algebra.
We conclude that $U^nf_n $ belongs to the ideal generated by the
compact elements.
We consider now the case $n=0$. The proof is the similar to the proof in the case $n>0$, considering the set
$\mathcal F=\{h\in C_0(X):h(X_a)=\{0\}, \mathrm D(h)=h^{-1}(\mathbb C\smallsetminus\{0\}) \textrm{ is wandering}\}$ instead of $\mathcal C$.
\end{proof}
\bibliographystyle{amsplain}
|
\section{Introduction}
Our understanding of ultraluminous X-ray sources (ULXs) -- off-nuclear X-ray sources
with luminosities in excess of 10$^{39}$\,\hbox{erg~s$^{-1}$}\ -- has evolved significantly over the
past few years. Historically, the debate regarding the nature of these sources has
focused on whether they represent a population of sub-Eddington `intermediate mass'
black holes (IMBHs with $M_{\rm{BH}} \sim 10^{2-5}$\,\hbox{$M_{\odot}$}; {\it e.g.}\ \citealt{Colbert99,
Miller04, Strohmayer09a}) or a population of super-Eddington but otherwise normal
stellar remnants ({\it e.g.}\ \citealt{King01, Poutanen07, Middleton15}); see \cite{Kaaret17rev}
for a recent review. Although evidence for large black holes is now being seen by LIGO
(most notably the recent detection of a BH--BH merger resulting in a $\sim$150\,\hbox{$M_{\odot}$}\
remnant; \citealt{LIGO20_IMBH}), the general consensus is now that the majority of
ULXs represent a population of super-Eddington accretors, thanks in particular to the
broadband spectral and timing studies possible in the {\it NuSTAR}\ era (\citealt{NuSTAR})
and the high-resolution spectra provided by {\it XMM-Newton}\ (\citealt{XMM}).
The broadband spectra obtained early in the {\it NuSTAR}\ mission demonstrated that
ULX spectra are clearly distinct from standard modes of sub-Eddington accretion ({\it e.g.}\
\citealt{Bachetti13, Walton13culx, Walton14hoIX, Walton15hoII, Rana15,
Mukherjee15}), confirming prior indications from {\it XMM-Newton}\ ({\it e.g.}\ \citealt{Stobbart06,
Gladstone09}), and instead revealed high-energy spectra consistent with broad
expectations for super-Eddington accretion ({\it i.e.~\/} spectra that appear to show a strong
contribution from hot, luminous accretion discs, {\it e.g.}\ \citealt{Shakura73, Abram88,
Poutanen07}). The super-Eddington nature of at least some ULXs was then
spectacularly confirmed with the discovery that the ULX M82 X-2 ($L_{\rm{X,peak}}
\sim 2 \times 10^{40}$\,\hbox{erg~s$^{-1}$}) is actually powered by a highly super-Eddington
neutron star, following the detection of coherent X-ray pulsations (\citealt{Bachetti14nat}).
Five more ULX pulsars have since been discovered (\citealt{Fuerst16p13, Israel17p13,
Israel17, Carpano18, Sathyaprakash19, Rodriguez20}), revealing an accretion regime that
extends up to $L/L_{\rm{E}} \sim 500$. In addition to the broadband spectra and the
discovery of ULX pulsars, we now have evidence in ULX data for the powerful outflows
ubiquitously predicted by models of super-Eddington accretion via the detection of
blueshifted atomic features. These have been seen primarily in the low-energy {\it XMM-Newton}\
RGS data, but also in the iron K band in a couple of cases (\citealt{Pinto16nat, Pinto17,
Pinto20, Walton16ufo, Kosec18ulx, Kosec18}). These outflows exhibit extreme velocities
($\sim$0.1--0.3$c$), implying they carry a significant additional energetic output from
these already extreme X-ray binary systems.
Nevertheless, important questions still remain regarding the ULX population. Although it
is now speculated that ULX pulsars could actually make up a significant fraction of these
sources ({\it e.g.}\ \citealt{Pintore17, Koliopanos17, Walton18ulxBB}), their exact contribution
is still highly uncertain. Is there also a significant population of black hole ULXs, and if
so could these be the progenitors of the BH--BH mergers now regularly being seen by
LIGO (\citealt{Inoue16, Mondal20})? Given the history of the field, it is easy to forget that
we still do not have a single ULX with a well-constrained mass function that
unambiguously requires the accretor to be a black hole. Can black hole ULXs (assuming
they exist) reach the same extreme Eddington ratios as ULX pulsars, or is this somehow
related to the magnetic nature of these objects (as suggested by \citealt{DallOsso15,
Mushtukov15})? What fraction of the total energetic output is radiative, and what fraction
is kinetic ({\it i.e.~\/} carried by winds/outflows) at these extreme accretion rates? Understanding
this last issue may in turn be critical for understanding early-universe SMBH growth (and
associated feedback), given that $\sim$10$^{9}$\,\hbox{$M_{\odot}$}\ black holes are now being
observed when the universe was only $\sim$0.7\,Gyr old ({\it e.g.}\ \citealt{Banados18nat}).
Furthermore, although the overall population is now expected to be dominated by
super-Eddington accretors, there are still rare individual sources among the ULX
population that remain good IMBH candidates. Most notable among these is the case of
ESO\,243--49 HLX1, which reaches an astonishing luminosity of $L_{\rm{X,peak}} \sim
10^{42}$\,\hbox{erg~s$^{-1}$}\ (\citealt{Farrell09}). In contrast to the vast majority of the ULX
population, this source does behave as expected for a scaled-up sub-Eddington X-ray
binary (\citealt{Servillat11, Webb12}). Furthermore, M82 X-1 has long been thought of
as an IMBH candidate because of its X-ray properties ({\it e.g.}\ \citealt{Feng10,
Pasham14nat}, although see \citealt{Brightman20m82} for caveats), and NGC\,2276--3c
has also been suggested as an IMBH candidate owing to its position on the radio--X-ray
fundamental plane (\citealt{Mezcua15}). Identifying further IMBH candidates remains of
significant interest, given the scarcity of compelling cases among the ULX population.
Key to advancing all of these areas are efforts to grow the broader ULX population and
provide larger samples with which to undertake statistical studies of ULXs and identify
notable individual sources for follow-up study. Most previous efforts have focused on
searching for ULXs in individual mission archives, using in particular {\it ROSAT}\
(\citealt{Roberts00, Colbert02, LiuBregman05}), {\it Chandra}\ (\citealt{Swartz04, Liu11,
Gong16, Kovlakas20}) and {\it XMM-Newton}\ (\citealt{WaltonULXCat, EarnshawULXcat}). Focusing
on data from a single mission has the advantage that everything (source selection,
energy bands) can be treated in a uniform manner, which is important for performing
population-based studies where selection biases need to be carefully controlled.
However, this comes at the expense of limiting the sky area/temporal coverage utilized
relative to that available in the full, multi-mission X-ray archive, both of which are key
factors in terms of identifying individual sources that may be of particular interest.
Here, we present the results of a search for new ULX candidates, combining data from all
of the public archives from the major soft X-ray imaging observatories currently
operational: {\it XMM-Newton}\ (\citealt{XMM}), {\it Chandra}\ (\citealt{Chandra}) and the {\it Neil Gehrels Swift Observatory}\
(hereafter {\it Swift}; \citealt{SWIFT}). In particular, we make use of the 4XMM-DR10, CSC2
and 2SXPS source catalogues (\citealt{4XMM, CSC2temp, 2SXPS}). Although combining
the data from these facilities does formally introduce some non-uniformity to the
selection, our primary aim is to compile the largest raw sample of ULX candidates to date,
facilitating searches for sources that are bright enough for detailed follow-up with current
and future X-ray facilities, as well as searches for sources with multi-wavelength
counterparts. This is of particular interest with both {\it XRISM}\ (\citealt{XRISM_tmp}) and
{\it Athena}\ (\citealt{Athena}) on the horizon, as well as the new facilities due to come online
at longer wavelengths ({\it e.g.}\ thirty-metre class optical telescopes, \textit{JWST}, the SKA,
{\it etc.}).
The paper is structured as follows: in Section \ref{sec_gals} we outline the galaxy sample
within which we search for ULX candidates, and in Section \ref{sec_search} we discuss
our procedure for identifying ULX candidates from the individual archives. Section
\ref{sec_sample} presents our final, merged sample of ULX candidates, and we highlight
the case of NGC\,3044 ULX1 -- a new extreme ULX discovered here -- in Section
\ref{sec_ulxs}. Finally, we summarise our findings in Section \ref{sec_conc}.
\section{Galaxy Sample}
\label{sec_gals}
In addition to the various X-ray archives considered here, the other major input required
for this work is a catalogue of galaxies within which to search for ULXs. Here, we primarily
use the HyperLEDA database (\citealt{HYPERLEDA}), initially selecting everything labelled
as a galaxy. We focus on HyperLEDA because this is one of the largest homogenized
compilations of known galaxies available in the literature. However, we further supplement
these galaxies with the latest version of the Catalogue of Nearby Galaxies (CNG;
\citealt{Karachentsev18}).
For our work here, we need to be able to define the sky area subtended by the galaxy (in
order to positionally match X-ray sources) as well as the distance to the galaxy (in order
to compute source luminosities). For the galaxy areas, we assume the extent of each
galaxy is determined by the elliptical region defined by its D25 isophote ({\it i.e.~\/} the best
elliptical fit to the area over which $B$-band surface brightness of the galaxy exceeds
25 mag arcsec$^{-2}$), which is given in HyperLEDA (where this information is available).
However, CNG uses the Holmberg radius to define the semi-major axis of the galaxy
ellipse instead, which corresponds to a surface brightness of 26.5 mag arcsec$^{-2}$.
For the subset of galaxies included in both HyperLEDA and CNG, we therefore calculate
an empirical correction between the D25 semi-major axes ($R_{\rm{D25}}$) and the
Holmberg radii ($R_{\rm{Holm}}$), and then apply this to any remaining galaxies that are
only included in CNG in an attempt to normalise these to the D25 definition. On average,
we find $R_{\rm{Holm}} = (1.26 \pm 0.02) R_{\rm{D25}}$ (where the uncertainty quoted
here is the 1$\sigma$ standard error on the mean). Although the full set of D25
information (semi-major axis, semi-minor axis, position angle) is obviously required to
search the full sky area subtended by the galaxy, in cases where the position angle is
missing it is still possible to search for ULX candidates within a circular region defined
by the semi-minor axis, as this will always be within the galaxy area regardless of the
orientation of its full elliptical region. We therefore also retain these galaxies in our input
sample. However, any galaxies in the HyperLEDA/CNG catalogues that do not have
sufficient information that we can compute at least their D25 semi-minor axes are
discarded.
For the majority of the galaxies considered, we compute distances based on their
measured redshifts ($z$) assuming they adhere to the Hubble flow. However, where
redshift-independent distance estimates are available, we prioritise these
measurements. HyperLEDA and CNG both include these based on a variety of different
methods (via {\it e.g.}\ Cepheid variables, the tip of the red giant branch, the Tully-Fischer
relationship), and we further supplement these with distance measurements from the
latest version of the Cosmicflows galaxy catalogue (\citealt{Tully16}). Such
measurements are particularly critical for very nearby galaxies (recession velocities
$cz < 1000$\,\hbox{$\rm\thinspace km~s^{-1}$}), where peculiar motions can dominate over the Hubble flow. For
these galaxies, we therefore also collected further redshift-independent distance
from the NASA Extragalactic Database where such measurements were not available in
any of the HyperLEDA/CNG/Cosmicflows catalogues. Where there are multiple distance
estimates available among these catalogues, we prioritise them in the following order:
Cosmicflows $>$ CNG $>$ Hyperleda $>$ NED $>$ Hubble flow, but we stress that in
the majority of cases there is generally good agreement between the different
catalogues regarding the redshift-independent distance estimates. However, since a
reasonably reliable distance estimate is in turn critical for a reliable luminosity
calculation, we therefore discard galaxies with recession velocities $cz < 1000$\,\hbox{$\rm\thinspace km~s^{-1}$}\
where there is no redshift-independent distance estimate available in any of the above
(similar to both \citealt{WaltonULXCat} and \citealt{EarnshawULXcat}).
The final galaxy sample utilized here consists of 966,010 entries, after accounting
for the requirements outlined above, the vast majority of which come from HyperLEDA
(only 215 of these galaxies are found in CNG but not HyperLEDA). Just under half of
these galaxies have morphology estimates available in the form of the Hubble type, $T$.
Following \cite{WaltonULXCat}, for these galaxies we make the distinction between spiral
galaxies ($T \geq 1$, including irregular galaxies) and elliptical galaxies ($T < 1$,
including lenticular galaxies). We show the distance distributions for the full galaxy
sample utilized here, as well as some of these subsets, in Figure \ref{fig_galdist}. The
majority of the galaxies considered are within a Gpc, although the galaxies for which
morphology information is not available do have larger distances on average than those
where the morphology has been identified.
\section{Selection of ULX Candidates}
\label{sec_search}
\subsection{Basic Approach}
We take the same basic approach to selecting ULXs for each of the three X-ray source
catalogues utilized here (4XMM-DR10, 2SXPS and CSC2). Our initial analysis of these
individual archives can be broadly summarised into 5 main steps, as described below.
Many of the specific details differ for the different catalogues utilized, owing to the
differences between the different X-ray observatories they are derived from and the
different formats in which the data are provided; these details are discussed in Section
\ref{sec_Xcats}.
\textit{\textbf{Step 1 -- positional match:}}
First, we perform a positional cross-match between our input galaxy list and each X-ray
catalogue; as noted above, for galaxies where the full set of spatial information is available
(both the major and minor axes, and the position angle) we perform a standard elliptical
match around the position of the galaxy ({\it i.e.~\/}\ utilizing the full sky area it subtends), while
for galaxies where the position angle is missing we perform a circular match within the
radius set by the semi-minor axis (and thus potentially only utilize a fraction of its sky
area). Where relevant, only X-ray sources listed as being point-like are retained (both
CSC2 and 4XMM-DR10 also contain extended sources, while in principle 2SXPS only
includes point sources).
\begin{figure}
\begin{center}
\hspace*{-0.35cm}
\rotatebox{0}{
{\includegraphics[width=235pt]{./figs/galaxy_distances.eps}}
}
\end{center}
\vspace*{-0.3cm}
\caption{
Distance distributions for the full galaxy sample utilized here (top panel), galaxies
identified as spiral ($T \geq 1$) and elliptical ($T < 1$) galaxies (upper-middle and
lower-middle panels, respectively), and for galaxies where morphology information is
not available (bottom panel).}
\label{fig_galdist}
\end{figure}
There are inevitably a small subset of X-ray sources that are consistent with being
associated with more than one galaxy in our initial matched source lists. In these cases,
we make sure to retain only one of the repeat entries in order to avoid individual X-ray
sources/detections appearing more than once. To do so, we initially associate the
X-ray source with the galaxy for which it is closest to the centre. This is because we
expect the majority of these cases to be interacting galaxies, which by definition will
be at essentially the same distance, and on average the density of ULXs is known to
increase as you approach the galaxy centre (\citealt{Swartz11}). However, we re-visit
this assumption at the end of our analysis of the individual archives (i.e. after step 5),
and assess whether the different galaxy distances really are similar for any remaining
sources that are potentially associated with more than one galaxy. Here, we treat
galaxies with distances that differ by $<$15\% as having similar distances. Where this
is not the case, such that the potential host galaxies appear to be un-associated
galaxies that happen to overlap when viewed in projection, we switch the association
of the X-ray source to the less distant galaxy. This is both a conservative approach,
resulting in lower X-ray luminosities, and also probably a more realistic assumption in
these situations, as the enhanced absorption by gas and dust in the foreground galaxy
would mean sources in the background galaxy are less likely to be seen as ULXs
based on their observed luminosities. For any sources where the association is
changed at this stage, the luminosities are re-evaluated (see step 2), and any sources
that no longer meet the ULX criterion are excluded.
\textit{\textbf{Step 2 -- luminosity cut:}}
With our positionally matched source lists, we then compute X-ray luminosities using
our preferred galaxy distance. Here, we use the full band fluxes available in each of the
individual X-ray catalogues (see Section \ref{sec_Xcats}). These bandpasses are not
precisely identical, but are sufficiently similar that we consider this a reasonable
compromise for the sake of simplicity, particularly in light of the simple spectral forms
assumed when estimating these fluxes (see Section \ref{sec_Xcats}). Attempting to
correct all of the fluxes to have a common treatment is non-trivial, particularly given
the time-dependent nature of the {\it Chandra}\ instrumental responses (owing to the
long-term build-up of the ACIS contamination layer; \citealt{Plucinsky18}). With the
luminosities in hand, we retain only sources with a full band luminosity that exceeds
$10^{39}$\,\hbox{erg~s$^{-1}$}, the standard definition of a ULX. In particular, we select sources
that have exhibited luminosities in excess of $10^{39}$\,\hbox{erg~s$^{-1}$}\ during any individual
observation of the source, allowing us to select both persistent and transient/highly
variable ULXs. This is a key consideration here, since the latter are now being detected
in increasing numbers as our X-ray archives continue to grow ({\it e.g.}\ \citealt{Soria12,
Middleton12, Middleton13nat, Pintore18, Pintore20, Earnshaw19, Earnshaw20,
vanHaaften19, Brightman20m51, Walton21}), and may be of particular interest in the
context of identifying good ULX pulsar candidates ({\it e.g.}\ \citealt{Tsygankov16,
Earnshaw18, Song20}).
\textit{\textbf{Step 3 -- quality flag cut:}}
Each of the X-ray catalogues utilized here contain a variety of information that relate to
the robustness of the X-ray detection included and the source properties derived. In
each case, we make use of this information to ensure that we only consider sources for
which the available X-ray information is reliable, further discarding sources where there
are concerns that this may not be the case. The approach taken here is necessarily
specific to each of the individual catalogues considered, and is detailed below for each
in turn (Section \ref{sec_Xcats}).
\textit{\textbf{Step 4 -- nuclear exclusion:}}
By definition, ULXs are off-nuclear sources, and so we attempt to exclude sources that
may be associated with the nuclei of their host galaxies. However, this is made
challenging by the fact that low-luminosity AGN can exhibit similar luminosities to ULXs
(\citealt{Ghosh08, Zhang09}). We therefore exclude potential AGN through their position
relative to the centre of the galaxy instead, following the approach taken in
\cite{EarnshawULXcat}. In brief, for each X-ray source we compute the minimum
separation from the central galaxy position, $R_{\rm{min}}$, based on its 3$\sigma$
positional uncertainty ({\it i.e.~\/}\ we define $R_{\rm{min}}$ = nuclear separation $-$
3$\sigma$ position error). We then select sources with $L_{\rm{X}} > 10^{42}$\,\hbox{erg~s$^{-1}$},
as these sources are almost certainly AGN (only one ULX, ESO\,243--49 HLX1, has
exhibited such luminosities to date; \citealt{Farrell09}), and calculate the cumulative
distribution of their $R_{\rm{min}}$ values. Unsurprisingly, these typically exhibit very
small minimum nuclear separations, and we determine the value of $R_{\rm{min}}$ that
contains $>$99\% of these sources, which we take to be our exclusion criterion for
nuclear sources, $R_{\rm{min,excl}}$ (see Section \ref{sec_Xcats}). All sources with
$R_{\rm{min}} < R_{\rm{min,excl}}$ are subsequently excluded from our analysis. We
repeat the assessment of $R_{\rm{min}}$ with both our input galaxy catalogue, where
we have a requirement for a minimum amount of information regarding the extent of
the galaxy (Section \ref{sec_gals}), and also with the full HyperLEDA/CNG/Cosmicflows
galaxy catalogues (as for this stage, only the position of the galaxy is required), such
that sources with $R_{\rm{min}} < R_{\rm{min,excl}}$ for \textit{any} galaxy included in
these databases are excluded from our source lists. This empirical approach allows us
to conservatively account for the uncertainty associated with the fact that for some
galaxies it can be difficult to precisely identify its central/nuclear position ({\it e.g.}\ irregular
or merging galaxies, and/or offset nuclei).
\textit{\textbf{Step 5 -- removal of other known contaminants:}}
In addition to the nuclei of the apparent host galaxies for our sources, we also attempt to
remove other known contaminants. At this stage, we particularly focus on background
AGN and foreground stars that coincidentally appear to be associated with a host galaxy
in projection. We therefore positionally match our remaining source lists against lists of
known stars and quasars. For the former, we use the Tycho2 catalogue (\citealt{tycho2})
while for the latter we use the GAIAunWISE quasar catalogue (\citealt{GAIAunWISE}) and
the quasar catalogue of \cite{VeronCetty10}. The search radii we use vary depending on
the X-ray archive, as detailed in Section \ref{sec_Xcats}. Any source that matches with
either a known star or a known quasar is excluded from our analysis.
\subsection{Specific Catalogue Details}
\label{sec_Xcats}
\subsubsection{4XMM-DR10}
The 4XMM catalogue (\citealt{4XMM}) is formatted such that every row entry represents
a unique detection of an X-ray source by the EPIC detectors (pn, MOS1, MOS2;
\citealt{XMM_PN, XMM_MOS}), meaning that the observation-by-observation information
needed to determine the peak flux for sources that have been observed on multiple
occasions is already incorporated. For the initial position match (step 1), we specifically
use the \textsc{ra\_sc} and \textsc{dec\_sc} columns in the 4XMM catalogue for the X-ray
position, which give the catalogue-averaged position for sources that have been detected
on multiple occasions. 4XMM includes both point sources and extended sources, and we
exclude observations in which the detection is marked as extended\footnote{Note that we
primarily use the observation-by-observation measures of source extent here, as
opposed to the mission averaged measure (\textsc{sc\_extent}), as variable/transient
point sources embedded in diffuse emission can still mistakenly be flagged as extended
in the latter. As such, there are some sources included that have non-zero values for
\textsc{sc\_extent}, but we stress that we have inspected the {\it XMM-Newton}\ images from the
observations listed as being point-like for these sources that are included in our final
sample, and visually confirmed the presence of a point-like source.}. When computing
detection luminosities (step 2), we use the full band flux provided in the catalogue (i.e.
4XMM band 8, spanning 0.2--12.0\,keV); these fluxes are computed by summing the
fluxes of the 4XMM sub bands, which are themselves computed assuming a standard
spectral shape (an absorbed powerlaw continuum with $\Gamma = 1.7$ and $N_{\rm{H}}
= 3 \times 10^{20}$\,\hbox{$\rm\thinspace cm^{-2}$}).
For the quality flags (step 3), we largely follow the approach taken in
\cite{EarnshawULXcat}. In brief, detections with a summary flag $\geq$ 2 are excluded
to reduce spurious detections in general, and sources with the out-of-time event flag
(Flag 10) set and a total count rate $<$ 0.05\,\hbox{$\rm\thinspace ct~s^{-1}$}\ are also excluded as these are likely
to be artefacts of out-of-time events that are associated with a nearby bright source.
However, in addition to these cuts, we also filter out entries where the \textsc{maskfrac}
flag (Flag 1) is set to be true for each of the EPIC detectors that registered the detection.
This helps to further limit spurious detections seen at chip edges, and also spurious
`new' detections at the edge of the field-of-view (FoV) that are really associated with
known bright sources just outside the FoV. When filtering out sources consistent with
being the nuclei of the host galaxies (step 4) and identifying likely matches with known
foreground stars/background quasars, we again make sure to use the \textsc{ra\_sc}
and \textsc{dec\_sc} columns for the X-ray positions. For the former, we find
$R_{\rm{min, excl}} = 9''$ following the empirical approach described above. This is a
pretty conservative cut, compared to previous works involving {\it XMM-Newton}\ data
(\citealt{WaltonULXCat, EarnshawULXcat}). In the latter case, since we are simply
matching point-source positions, we use a matching radius for the various star/quasar
catalogues of 5$''$, roughly corresponding to the typical $3\sigma$ positional accuracy
for point sources in 4XMM-DR10 (\citealt{4XMM}).
\subsubsection{2SXPS}
By definition the 2SXPS catalogue (\citealt{2SXPS}) only includes point sources detected
by the XRT (\citealt{SWIFT_XRT}), and the main table of the catalogue is formatted such
that every row entry represents a unique X-ray source, with the
observation-by-observation detection information contained in a separate table. However,
the primary source table includes information on the peak flux seen by the XRT for each
source included, and so we mainly use this table for our analysis. As such, for the
initial position match (step 1), we are naturally using the best-fit position determined from
all of the available observations of a given source. When computing the relevant source
luminosity (step 2), we primarily select sources based on the peak flux given for each
source in the full XRT band (spanning 0.3--10.0\,keV) assuming again an absorbed
powerlaw continuum (fluxes for various potential spectral models are provided, but of
these the absorbed powerlaw is the most appropriate choice for ULXs below 10\,keV).
Here, the powerlaw parameters adopted when computing the catalogued fluxes are
either fit directly, derived from the 2SXPS hardness ratios, or a photon index of
$\Gamma = 1.7$ and the Galactic column in the direction of the source are assumed
(see \citealt{2SXPS} for details).
In contrast to both {\it XMM-Newton}\ and {\it Chandra}, typical {\it Swift}\ observations are very short
exposures ($\sim$1--2\,ks). Furthermore, these observations themselves are often split
up into several shorter `snapshots', and the peak flux included in the catalogue can in
principle be drawn from the count rate seen during one of these snapshots instead of
the full observation. As such, the peak flux often has large uncertainties, being based on
only a handful of counts. For sources where the peak luminosity has a fractional error of
$>$40\% (averaging the positive and negative errors quoted), corresponding to a
detection with $\sim$10 counts based on the approximation for the Poisson distribution
presented in \cite{Gehrels86}, we therefore also require that the source meet at least
one of three additional criterion for inclusion. Either: 1) the average luminosity is also
consistent with the ULX regime, assuming that the average and peak luminosities are not
identical, or 2) there are two or more independent detections of the source in the ULX
regime, based on the observation-by-observation data, or 3) the source has also
previously been detected in the ULX regime by some other facility ({\it i.e.~\/}\ the detection is
spatially consistent with an entry in one of the archival ULX catalogues we compare our
new dataset against; see Section \ref{sec_prev_ulxcat} for further discussion). For the
quality flag cut (step 3), we exclude sources with the summary flag set to $\geq$1 ({\it i.e.~\/}
the ``clean" criterion defined by the {\it Swift}\ team). When filtering out sources consistent
with being the host galaxy nuclei (step 4), we also find $R_{\rm{min, excl}} = 9''$, similar
to our analysis of 4XMM. Finally, when identifying likely matches with known foreground
stars/background quasars (step 5), we use a matching radius of 10$''$ for the {\it Swift}\
data, again corresponding to the typical $3\sigma$ positional accuracy for sources in
2SXPS (\citealt{2SXPS}).
\subsubsection{CSC2}
\label{sec_csc2}
Similar to 2SXPS, the main table of the CSC2 catalogue (\citealt{CSC2temp}) is formatted
such that every row entry represents a unique X-ray source, with the
observation-by-observation detection information presented in a separate table, and
similar to 4XMM both point-like and extended sources are included. We therefore use the
primary source table when performing the initial position match with our galaxy catalogue
(step 1), such that we are again using the best-fit position determined from all of the
available observations of a given source, but we then compile the
observation-by-observation information for each matched source from the secondary
table so that we can be sure to account for the peak flux seen by {\it Chandra}\ for each
source in our analysis. Sources listed as being extended are discarded. When computing
detection luminosities (step 2), we use the broadband CSC2 fluxes, {\it i.e.~\/} `broad' fluxes for
the ACIS detectors (\citealt{CHANDRA_ACIS}), spanning 0.5--8.0\,keV, or `wide' fluxes
for the HRC (\citealt{CHANDRA_HRC}), spanning 0.2--10.0\,keV. Where possible we
again use fluxes derived assuming a powerlaw spectral form (as with 2SXPS, fluxes for
a variety of spectral models are provided, see the CSC2 documentation for details). Here
the catalogued powerlaw fluxes we use are computed assuming $\Gamma = 2$ and the
appropriate Galactic column. However, if this powerlaw flux is not available then we use
the raw aperture flux instead. For the quality flag cut, we only consider sources which
are flagged as `true' detections in the primary source table ({\it i.e.~\/} sources flagged as
`marginal' detections are excluded), and we also exclude source detections at the
observation level for which the `streak' flag is set.
In addition to the standard filtering steps outlined above, one further issue that is of
relevance for the observation-by-observation {\it Chandra}\ data is the fact that the
{\it Chandra}\ PSF degrades rapidly with off-axis angle (in a relative sense, much more
severely than is the case for either {\it XMM-Newton}\ or {\it Swift}). As such, the typical extraction radii
used in CSC2 also increase with increasing off-axis angle; for example, sources with
off-axis angles of $\sim$8$'$ often have extraction radii of $\sim$9--10$''$,
significantly larger than the on-axis PSF. Unfortunately, for point sources that are either
in crowded regions or are embedded in extended regions of diffuse emission, this can
result in spurious fluxes for any significantly off-axis observations, as these off-axis
detections can occasionally be blends of multiple point sources, and/or include
significant diffuse flux not actually associated with the point source in question. This
is particularly an issue for observations of nearby giant elliptical galaxies; {\it Chandra}\
has undertaken significant programs tiling a number of these galaxies ({\it e.g.}\ M87)
resulting in a combination of on- and off-axis observations of the same crowded fields.
As such, there are a number of sources in these galaxies which have very modest
luminosities when viewed on-axis ($L_{\rm{X}} < 10^{38}$\,\hbox{erg~s$^{-1}$}) but which all appear
to share the same ULX-level off-axis detection. Although some ULXs can be highly
variable, as noted above, in many of these cases the ULX-level detections are
unfortunately spurious. We therefore manually inspect cases where the only ULX-level
detection is taken significantly off-axis, and there is also an on-axis observation that
shows the source to have a significantly lower luminosity. Where these are clearly
cases relating to source confusion, we exclude these sources from our analysis. We
also exclude cases in which the size of the aperture increases to the point where it
covers the nominal position of the galaxy centre. In cases where the higher flux could
plausibly be due to variability ({\it i.e.~\/}\ the on-axis observations show no evidence for large
numbers of sources or diffuse emission whose integrated flux could explain that seen
in the off-axis observation) we retain these detections, but stress that they should be
considered high-priority for further (triggered) follow-up to confirm their nature. We
also retain cases in which the aperture marginally overlaps with the edge of the nuclear
exclusion zone (but not the nominal nuclear position).
When filtering out sources consistent with being the host galaxy nuclei (step 4), we find
$R_{\rm{min, excl}} = 6.1''$ for CSC2, smaller than the equivalent value for both 4XMM
and 2SXPS. Although in a qualitative sense this is not surprising, given the superior
imaging capabilities, it is still worth noting that this value is still significantly larger than
the {\it Chandra}\ point spread function. This likely reflects the fact that for more complex
galaxy morphologies it can be difficult to accurately identify the position of the true
galaxy centre. Finally, when identifying likely matches with known foreground
stars/background quasars (step 5), we use a matching radius of 3$''$.
\subsection{Merging and Further Filtering}
Once all of the individual catalogues of ULX candidates from each of the 4XMM-DR10,
CSC2 and 2SXPS archives have been produced, we merge them all into a final
`master' ULX catalogue. To do so, we sequentially match our individual {\it XMM-Newton}, {\it Swift}\
and {\it Chandra}\ catalogues of ULX candidates by position. We begin by matching the
{\it XMM-Newton}\ and {\it Swift}\ catalogues. As {\it Swift}\ is typically the limiting factor regarding position
uncertainties, we match the two within a radius of 10$''$, corresponding to the typical
3$\sigma$ 2SXPS position uncertainty. There are three main outcomes from this initial
match: sources with only {\it XMM-Newton}\ data, sources with only {\it Swift}\ data, and matched
sources with both. Each of these lists are then matched with the {\it Chandra}\ catalogue.
For the sources with only {\it Swift}\ data, we again use a matching radius of 10$''$ here,
and for the sources with only {\it XMM-Newton}\ data we use a matching radius of 5$''$ (again, the
typical 3$\sigma$ 4XMM-DR10 position error, as {\it XMM-Newton}\ is the limiting factor regarding
position uncertainties here). For sources with both {\it XMM-Newton}\ and {\it Swift}\ data, we assume
the {\it XMM-Newton}\ position to be more accurate, and so use this to match to the {\it Chandra}\
catalogue, again using a matching radius of 5$''$.
At each of these matching stages, there is the possibility that there are multiple
matches for the same source. This is particularly the case when matching either the
{\it XMM-Newton}\ or the {\it Swift}\ data with {\it Chandra}, given the potential for source confusion and
the superior imaging capabilities of the latter; a famous example is the case of
NGC\,2276, in which a source perceived to be an extremely luminous ULX by {\it XMM-Newton}\
is actually resolved into three distinct point sources by {\it Chandra}\ (\citealt{Sutton12}).
In that case, all of the resolved sources are themselves ULXs, but it is also possible
that a source that appears as a ULX to {\it XMM-Newton}\ or {\it Swift}\ will actually be resolved into
multiple sub-ULX sources (this is conceptually similar to the issue regarding the
degradation of the off-axis {\it Chandra}\ PSF discussed in Section \ref{sec_csc2}). In
addition to matching them against our {\it Chandra}\ catalogue of ULX candidates, we
therefore also match our {\it XMM-Newton}\ and {\it Swift}\ ULX candidates against the set of
{\it Chandra}\ sources that did not make our luminosity cut, and again manually inspect
all cases of multiple matches in order to identify sources that only appear to be ULXs
because of a detection that is actually likely the blend of several point sources,
artificially inflating its apparent flux. As before, these sources are removed from our
analysis. We note, however, that we still retain cases where {\it e.g.}\ {\it Chandra}\ resolves an
{\it XMM-Newton}\ detection into two discrete sources, but that the {\it XMM-Newton}\ data imply that at least
one of these must have varied into the ULX regime (for example, a scenario in which
{\it Chandra}\ sees two sources at $L_{\rm{X}} \sim 10^{38}$\,\hbox{erg~s$^{-1}$}, but the {\it XMM-Newton}/{\it Swift}\
detection that is consistent with both of these sources implies $L_{\rm{X}} \sim 2
\times 10^{39}$\,\hbox{erg~s$^{-1}$}, meaning that at least one of these sources must have been in
the ULX regime during the {\it XMM-Newton}/{\it Swift}\ observation). In these cases, we add a flag to
the final version of the master catalogue noting that this issue exists (a separate flag is
added for each of the matched catalogue pairs, see Table \ref{tab_matchflag} for the
definitions of the different values these flags can take).
On occasion, where there are multiple matches it is possible to determine with
reasonable confidence which of the resolved sources the unresolved detection is
actually associated with (for example, cases where {\it Chandra}\ sees two sources, one
with $L_{\rm{X}} \sim 10^{38}$\,\hbox{erg~s$^{-1}$} and one with $L_{\rm{X}} \sim 10^{40}$\,\hbox{erg~s$^{-1}$}\
and {\it XMM-Newton}/{\it Swift}\ sees one source that also has $L_{\rm{X}} \sim 10^{40}$\,\hbox{erg~s$^{-1}$}, or
alternatively cases in which the position of the first {\it Chandra}\ source is in outstanding
agreement with the position of the {\it XMM-Newton}/{\it Swift}\ detection, while the second {\it Chandra}\
source is right at the edge of the 3$\sigma$ uncertainty range). In these cases, we
make a judgement call ourselves and assign the unresolved detection to the resolved
source we feel is most appropriate. Where we feel unable to make a judgement, but
there are multiple ULX candidates among the resolved sources (similar to the case of
NGC\,2276 highlighted above), we we retain all of the potential resolved matches
within the master catalogue. Both of these scenarios are also indicated by the
matching flags highlighted above (again, see Table \ref{tab_matchflag}).
\begin{table}
\caption{Definitions for the flags detailing the decision taken for any complex matches
between the individual ULX catalogues}
\begin{center}
\begin{tabular}{c p{6.5cm}}
\hline
\hline
\\[-0.25cm]
Value & Description \\
\\[-0.25cm]
\hline
\hline
\\[-0.2cm]
NULL & No match between the catalogues \\
\\[-0.3cm]
0 & Unique match between ULX candidates \\
\\[-0.3cm]
1 & Formally more than one potential match between ULX candidates, but one is clearly preferred and assumed to be the correct match; only this match is reported \\
\\[-0.3cm]
2 & Formally more than one potential match between ULX candidates, and it is unclear which is the correct association; all potential matches are given \\
\\[-0.3cm]
3 & ULX detection consistent with several lower luminosity sources seen by the other mission in question, but their combined flux is not sufficient to explain that seen of the ULX detection, so the source is still retained \\
\\[-0.2cm]
\hline
\hline
\end{tabular}
\end{center}
\vspace*{-0.1cm}
\label{tab_matchflag}
\end{table}
\begin{figure*}
\begin{center}
\hspace*{-0.2cm}
\rotatebox{0}{
{\includegraphics[width=235pt,trim={0cm -1cm 0cm -1cm},clip]{./figs/image_ngc6946_xmm.eps}}
}
\hspace*{0.6cm}
\rotatebox{0}{
{\includegraphics[width=235pt,trim={0cm -1cm 0cm -1cm},clip]{./figs/image_ngc6946_optical.eps}}
}
\hspace*{-0.2cm}
\rotatebox{0}{
{\includegraphics[width=235pt,trim={0cm -1cm 0cm -1cm},clip]{./figs/image_ngc1097_swift.eps}}
}
\hspace*{0.6cm}
\rotatebox{0}{
{\includegraphics[width=235pt,trim={0cm -1cm 0cm -1cm},clip]{./figs/image_ngc1097_optical.eps}}
}
\hspace*{-0.2cm}
\rotatebox{0}{
{\includegraphics[width=235pt,trim={0cm -1cm 0cm -1cm},clip]{./figs/image_ngc720_chandra.eps}}
}
\hspace*{0.6cm}
\rotatebox{0}{
{\includegraphics[width=235pt,trim={0cm -1cm 0cm -1cm},clip]{./figs/image_ngc720_optical.eps}}
}
\end{center}
\vspace*{-0.3cm}
\caption{
X-ray (\textit{left}) and accompanying optical (\textit{right}) images for three example
galaxies that demonstrate some of the main stages of our source selection. For the
X-ray images, from top to bottom, we show an {\it XMM-Newton}\ EPIC image of NGC\,6946
(OBSID 0691570101), the integrated {\it Swift}\ XRT image of NGC\,1097 (generated with
the standard online XRT pipeline; \citealt{Evans09}) and a {\it Chandra}\ ACIS image of
NGC\,720 (stack ID acisfJ0153056m134345). All are smoothed with a Gaussian of
width 3 pixels. The optical images are from the Digitized Sky Survey. In all panels, the
D25 extent of the galaxy in question is indicated with the blue dashed ellipse, and the
nuclear position with a cross (note that in the case of NGC\,1097, the small companion
galaxy NGC\,1097A is also shown). `Field' sources ({\it i.e.~\/} outside of the D25 extent) are
shown with squares, point sources within the D25 extent that do not qualify as ULX
candidates with circles, and finally sources selected as ULX candidates are shown
with diamonds, respectively; all markers are shown in either black or white simply so
that they can most easily be seen against the background images, there is no further
significance to the choice of colour. The {\it XMM-Newton}\ and {\it Chandra}\ images represent the
deepest observation/stack of NGC\,6946 and NGC\,720 that are included in
4XMM-DR10 and CSC2, respectively. However, owing to the variable nature of the ULX
population (see {\it e.g.}\ \citealt{Earnshaw19} for NGC\,6946 in particular), not all sources
identified as ULX candidates are necessarily visible in these X-ray images.
}
\label{fig_image}
\end{figure*}
Having merged the {\it XMM-Newton}, {\it Swift}\ and {\it Chandra}\ data as best we can, we now address
the presence of one more class of known contaminants, X-ray transients associated
with one-off explosive events ({\it i.e.~\/}\ supernovae). Although certainly not all do, these
events can reach ULX luminosities, and would then be selected by our process (given
our interest in genuinely transient ULXs) even though they are clearly not
accretion-powered X-ray binaries. This is particularly relevant here given our use of
{\it Swift}\ data, since one of {\it Swift}'s main focuses is rapid follow-up of transient events.
We therefore correlate our master catalogue with the positions of known supernovae
recorded in the Open Supernova Catalogue (\citealt{OpenSN}; note that this includes
both supernovae that have occurred since {\it XMM-Newton}, {\it Chandra}\ and {\it Swift}\ have been
observing and more historic supernovae). To do so, we prioritise X-ray source
positions from {\it Chandra}, {\it XMM-Newton}\ and {\it Swift}\ in that order ({\it i.e.~\/}\ in cases where a source
is detected by all three observatories, we use the {\it Chandra}\ position for this match),
and perform the match using search radii of 3$''$, 5$''$ and 10$''$ for {\it Chandra},
{\it XMM-Newton}\ and {\it Swift}\ positions, respectively. However, in order to determine whether
the X-ray source is really associated with the transient in question we also examine
the relative timing of the event and the first detection of the X-ray source (hence our
decision to only apply this filter to the merged dataset, where we can most robustly
determine when the source was first detected). X-ray sources that are positionally
coincident with supernovae, but which were detected as ULXs significantly before
the event occurred are deemed to be unrelated to the supernova and retained in our
sample. However, sources positionally coincident with known transients that have
only been detected after the event occurred are assumed to be associated with the
supernova, and so are excluded from our final sample.
We also match our remaining sample against both the NED and SIMBAD databases
in order to identify and remove any further non-ULX contaminants that have been
identified in the literature (uncatalogued AGN, stars, supernovae). We adopt the
same spatial matching procedure as for the Open Supernova Catalogue, prioritising
{\it Chandra}, {\it XMM-Newton}\ and {\it Swift}\ positions in that order and using matching radii of 3$''$,
5$''$ and 10$''$. For any further supernovae identified, we also again consider the
date of the first X-ray detection when deciding whether the X-ray source should be
removed. We then remove any remaining sources obviously associated with the
host-galaxy AGN that have been missed by our nuclear cut ({\it e.g.}\ sources with
$L_{\rm{X}} \geq 10^{42}$\,\hbox{erg~s$^{-1}$}\ that lie just outside our nuclear exclusion radii or,
in the case of Centaurus A, are located in the X-ray emission from the AGN jet;
\citealt{Hardcastle07}) as well as a number of sources for which we are aware of
follow-up studies that have previously found the ULX candidate to be an
uncatalogued background quasar/foreground star (\citealt{Dadina13, Heida13,
Sutton15, Gou16}).
Finally, after all of the above steps, we find that the remaining sample contains a
number of highly clustered sources which only appear in 2SXPS and actually seem
to be associated with the bright diffuse emission known to be present in the M82
galaxy ({\it e.g.}\ \citealt{Griffiths00, Lopez20}), even though 2SXPS is intended to be a
dedicated point source catalogue. This is likely related to the typical snapsnot
nature of {\it Swift}\ XRT observations; with such short exposures random Poisson
fluctuations from the diffuse emission may more easily be mistaken as point
sources. 2SXPS notes all of the potential aliases for each entry, and many of these
M82 sources are listed as potentially being aliased with each other. We therefore
also manually inspect X-ray images -- both the images integrated over the duration
of the {\it Swift}\ mission and specifically taken from the observation corresponding to
the reported best detection for the XRT, and, where available, any CSC2 {\it Chandra}\
images as well -- for all of the 2SXPS sources which have not also been identified
as a ULX candidate in either of our 4XMM-DR10 or CSC2 subsamples and are
listed as having other potential 2SXPS aliases. Any sources which we judge to be
likely associated with diffuse emission (similar to the M82 case) are removed from
the final sample. During this process, if a source is aliased with another genuine
point source (as opposed to being part of a cluster of sources associated with
extended emission), we also make a judgement over whether these are likely the
same source, and retain only one entry in these cases.
\begin{table*}
\caption{The final sample of ULX candidates compiled from the 4XMM-DR10, 2SXPS and
CSC2 catalogues}
\begin{center}
\begin{tabular}{l c c c c}
\hline
\hline
\\[-0.2cm]
& 4XMM-DR10 & 2SXPS & CSC2 & Combined Sample \\
\\[-0.25cm]
\hline
\hline
\\[-0.2cm]
Number of ULX Candidates & 641 & 501 & 1031 & 1843 \\
\\[-0.3cm]
\hspace{0.5cm} (with multiple ULX detections in the parent catalogue) & 177 & 291 & 246 & 702 \\
\\[-0.3cm]
\hspace{0.5cm} (seen as a ULX by multiple observatories) & 241 & 173 & 209 & 293 \\
\\[-0.3cm]
\hspace{0.5cm} (HLX candidates) & 22 & 36 & 17 & 71 \\
\\
Host Galaxies & 403 & 269 & 548 & 951 \\
\\[-0.3cm]
\hspace{0.5cm} (average distance, Mpc) & $62.3 \pm 3.5$ & $34.8 \pm 2.7$ & $83.8 \pm 3.8$ & $74.7 \pm 2.7$ \\
\\[-0.3cm]
\hspace{0.5cm} (containing multiple ULX candidates) & 130 & 89 & 190 & 333 \\
\\[-0.2cm]
\hline
\hline
\end{tabular}
\end{center}
\vspace*{-0.1cm}
\label{tab_sample}
\end{table*}
\section{The Final Sample}
\label{sec_sample}
Our final sample of ULX candidates consists of 1843\ individual sources residing
in 951\ host galaxies. The catalogue will be made available to the public,
and will be comprised of four tables. The first is a `master' list formatted to have
one row entry per source, summarising some key information and detailing which
combination of {\it XMM-Newton}, {\it Swift}\ and {\it Chandra}\ have reported the source as a ULX.
We stress that we are focused only on the detections of these sources that meet
the ULX luminosity threshold here ({\it i.e.~\/} $L_{\rm{X}} \geq 10^{39}$\,\hbox{erg~s$^{-1}$}); if an
{\it XMM-Newton}\ ULX candidate does not have a {\it Chandra}\ counterpart reported, for
example, this does not necessarily mean that {\it Chandra}\ has not detected that
source, only that {\it Chandra}\ has not seen it at a flux that would correspond to the
ULX regime. The other three tables provide the full details of the 4XMM-DR10,
2SXPS and CSC2 entries for the ULX-level detections of these sources. These
follow the formats of the data used to compile these subsamples of ULX
candidates in the first place (i.e. the {\it XMM-Newton}\ and {\it Chandra}\ tables have one row
entry per observation of a ULX candidate, while the {\it Swift}\ table just has one row
entry per ULX candidate).
Some statistics for the full sample and the individual 4XMM-DR10, 2SXPS and
CSC2 sub-samples are given in Table \ref{tab_sample}, and we show examples of
our source selection in Figure \ref{fig_image} for each of the {\it XMM-Newton}, {\it Swift}\ and
{\it Chandra}\ observatories. By number, the CSC2 component contributes the most
sources to our final sample, followed by 4XMM-DR10 and then 2SXPS. The latter
still makes a very significant contribution though. There is obviously notable
overlap between the individual subsamples ({\it i.e.~\/} some sources are detected as
ULXs by multiple missions), as also detailed in Table \ref{tab_sample} and in the
master table provided. Of our 1843\ individual sources, 50\ are
detected at ULX luminosities in all three of our contributing source catalogues.
\begin{figure}
\begin{center}
\hspace*{-0.35cm}
\rotatebox{0}{
{\includegraphics[width=235pt]{./figs/host_galaxy_distances.eps}}
}
\end{center}
\vspace*{-0.3cm}
\caption{
Distance distributions for ULX host galaxies, showing the full multi-mission sample
(top), and the CSC2, 4XMM-DR10 and 2SXPS subsets, respectively (lower panels).}
\label{fig_ulx_galdist}
\end{figure}
As expected, given the known connection between ULXs and recent star formation
(\citealt{Swartz09, Mineo12, Lehmer19}), the majority of our ULX host galaxies
with morphology information available are spiral galaxies ($\sim$60\%, using the
T-type ranges defined above). We also plot the distribution of host galaxy distances
in Figure \ref{fig_ulx_galdist} for the full sample and each of the individual catalogue
subsamples. There is significant overlap in the individual distributions, but typical
host galaxy distances are lowest for the 2SXPS subsample, and largest for the
CSC2 subsample, as the latter has the best sensitivity to faint point sources among
the X-ray catalogues considered owing to both the low background and superior
imaging capabilities of {\it Chandra}. This allows {\it Chandra}\ to more easily detect ULX
candidates out to larger distances than either {\it XMM-Newton}\ or {\it Swift}, and thus the CSC2
subsample ends up making the largest contribution to our final sample.
Of our 951\ host galaxies, 333\ are found to host multiple
ULX candidates. As our primary interest is focused on individual sources, and
our sample selection is highly non-uniform, we do not make any attempt to
correct for (in)completeness in any galaxies observed with insufficient depth to
reach luminosities of 10$^{39}$\,\hbox{erg~s$^{-1}$}, so this should likely be considered a
lower limit for ULX multiplicity in these hosts. The most extreme example is
NGC\,2207 -- one half of an interacting galaxy pair (the other being IC\,2163;
\citealt{Eskridge02}) -- which appears to host an astonishing 34 ULX
candidates, the majority of which (31) are contributed by the CSC2 catalogue.
This is notably larger than the 21 ULXs reported to reside in NGC\,2207/IC\,2163
by \cite{Mineo13}, likely due to additional {\it Chandra}\ observations being included
in CSC2 and our explicit consideration of long-term variability in selecting our
ULX sample. Owing to the interacting nature of these galaxies, it is not surprising
that there should be a large number of ULXs. It is nevertheless worth noting that
there seems to be some disagreement over the distance to NGC\,2207 in the
literature. The distance we have adopted here is $D = 36.4$\,Mpc, which is
based on the recession velocity reported in HyperLEDA. This distance is very
similar to that reported based on the supernova SN1975A which occurred in
NGC\,2207 ($D = 39.6$\,Mpc; \citealt{Arnett82}), which is adopted by
\cite{Mineo13}. However, the more recent estimates from the Tully-Fisher
method typically seem to imply a distance of $D \sim 17$\,Mpc
(\citealt{Russell02, Theureau07}). Should this be correct, only 7 of our sources
in NGC\,2207 would still be considered ULXs. However, our assumption is that
the supernova-based distance is the most reliable here, and so our luminosity
estimates should be reasonable.
We also note that among the 1843\ ULX candidates, our catalogue contains 71\
`hyperluminous' X-ray source (HLX) candidates.\footnote{Note that for in order for
sources detected in 2SXPS to be considered good HLX candidates we apply a
similar criterion to our initial source selection procedure, such that if the peak
luminosity is not well constrained (average fractional uncertainty of $>$40\%) then
the source has to either have a better-constrained average luminosity that is also in
the HLX regime, or at least two separate XRT observations that place it in the HLX
regime.} These are the most extreme members of the ULX population, exhibiting
luminosities of $L_{\rm{X}} \geq 10^{41}$\,\hbox{erg~s$^{-1}$}. Owing to their astonishing
luminosities, such sources are often considered the best candidates for IMBH
accretors. Indeed, two of the sources discussed as the leading IMBH candidates in
the literature, M82 X-1 and ESO\,243--49 ULX1, are found among this population.
However, it is also worth noting that one of the known ULX pulsars, NGC\,5907
ULX1, also reaches luminosities of $L_{\rm{X}} \sim 10^{41}$\,\hbox{erg~s$^{-1}$}\ (\citealt{Israel17,
Fuerst17ngc5907}), despite being powered by a neutron star. Nevertheless, these
sources are still of particular interest, and our new HLX candidates will be discussed
in more detail in future work (A. D. Mackenzie, \textit{in prep.}).
\subsection{Comparison with Other ULX Catalogues}
\label{sec_prev_ulxcat}
The first major effort to search for ULXs among any of the X-ray source catalogues
considered here was presented by \cite{Kovlakas20}, who also searched CSC2 for
ULX candidates. Although both the approach taken and the input galaxy sample
used are quite similar in both cases, there are also a couple of notable differences.
First and foremost, we have considered the {\it Chandra}\ data down to the
observation-by-observation level, in order to select sources based on their peak
flux and specifically include transient ULXs in our sample, while \cite{Kovlakas20}
base their luminosity selection on the flux recorded in the longest uninterrupted
segment of {\it Chandra}\ data (which is not necessarily the peak flux exhibited by the
source). Second, we have taken a much more conservative approach to excluding
potential nuclear sources associated with our host galaxies. \cite{Kovlakas20} flag
a source as `nuclear' if it is within 3$''$ of the nominal galaxy centre, while we both
consider the position error on the X-ray detection and utilize a much larger
minimum exclusion radius (6.1$''$). This is based on our empirical assessment of
the separation between the nominal centre of the host galaxies and sources that
we consider likely to be their AGN (those that appear to have $L_{\rm{X}} \geq
10^{42}$\,\hbox{erg~s$^{-1}$}). Our more conservative approach does mean that our catalogue
is likely cleaner with regards to any remaining contamination from AGN in our host
galaxies, but this will come at the cost of excluding a larger number of \textit{bona
fide} ULXs from our sample, particularly given that the spatial density of ULXs is
seen to increase towards the galaxy centres (\citealt{Swartz11, Wang16,
Kovlakas20}). Nevertheless, this is a more appropriate approach given that our
primary motivation is to find individual sources that are of interest for follow-up
studies; detailed follow-up of ULXs within a few arcseconds of the nuclear
position will not realistically be feasible for the majority of our current and planned
X-ray facilities if the nuclear black hole is even reasonably active. Despite these
differences, though, there is naturally a fairly significant degree of overlap (754
sources) between our sample and sources that would qualify as ULXs in
\cite{Kovlakas20}.
More recently, \cite{Inoue21} have also searched for CSC2 for ULX candidates.
However, a major difference between these works is that \cite{Inoue21} have
used a much smaller catalogue of input galaxies than that used here, derived by
combining IRAS galaxies with the CNG catalogue. Furthermore, while they do
consider the observation-by-observation data provided in CSC2, they use the
flux from the longest individual observation when computing luminosities, which
again is not necessarily the peak flux exhibited by the sources in question (which
is what we are interested in here), and we have again been more conservative in
our treatment of nuclear sources. Although their work primarily focuses on
CSC2, the final catalogue does also include sources selected from 4XMM-DR9
and 2SXPS, and so is therefore conceptually similar to our multi-mission
approach. There is not a lot of specific detail provided for these latter analyses,
unfortunately, but the approach is stated to be broadly similar to that used for
CSC2, and so similar differences between the two works are presumably present
here too. In addition, another key difference with regard to the 2SXPS analysis is
that they appear to have only made use of the average fluxes from {\it Swift}, while
we have considered the peak flux (where this is considered reliable). Furthermore,
we have used an even more recent release of the 4XMM survey here.
Nevertheless, despite these differences, there is again some notable overlap of
357 sources in total (251, 107 and 149 from {\it Chandra}, {\it XMM-Newton}\ and {\it Swift},
respectively).
Finally, \cite{Barrows19} also utilize CSC2, but only search specifically for HLXs
within the SDSS-DR7 galaxy sample. However, the spatial offsets relative to the
central galaxy positions would result in the majority of their HLX candidates being
considered as nuclear sources with our more empirical approach to this stage
of the catalogue production. Only one of our sources is also present in the
\cite{Barrows19} catalogue, 2CXO\,J155910.3$+$204619, and we have assigned
this to a different (and closer) host galaxy, giving it a much more modest
luminosity of $L_{\rm{X}} \sim 2.5 \times 10^{39}$\,\hbox{erg~s$^{-1}$}. As such, our sample
of HLX candidates differs entirely to that presented by \cite{Barrows19}.
In addition to these more recent works, we also match our new catalogue against
a series of other archival ULX catalogues, which have been derived using
previous generations of X-ray surveys. In these cases we match by position, as
they are not drawn from any of the exact X-ray catalogues used here (and thus
do not have identical naming conventions). Similar to our final matching against
the NED and SIMBAD for remaining contaminants, we split our ULX catalogue
into sources where the best position comes from {\it Chandra}, from {\it XMM-Newton}\ and from
{\it Swift}, and then individually match these sub-sections against each of the
archival ULX catalogues in turn. The matching radius used for each comparison
depends on the origins of the data being compared, and always corresponds to
the larger of the typical positional uncertainties associated with the two input
tables. As before, positions from {\it Chandra}, {\it XMM-Newton}, {\it Swift}\ and {\it ROSAT}\ are
considered to have typical uncertainties of 3$''$, 5$''$ and 10$''$, respectively,
and {\it ROSAT}\ positions are considered to have a typical uncertainty of 20$''$. For
example, when comparing the subset of our catalogue with {\it Chandra}\ positions
against another catalogue derived from {\it Chandra}\ data, we would use a
matching radius of 3$''$, but comparing the same subset against a catalogue
derived from {\it ROSAT}\ observations, we would use a matching radius of 20$''$
instead. For this analysis, we simply note all potential matches. The catalogues
we match against are listed in Table \ref{tab_oldcat}.
Based on all of these matches, we find that 689\ of the ULX candidates
presented here are completely new, {\it i.e.~\/}\ do not seem to appear in any of the
other ULX catalogues considered, and 1318\ have only recently been
catalogued as a ULX, {\it i.e.~\/} they only appear in catalogues based on the latest
generation of X-ray source catalogues (this work, \citealt{Barrows19},
\citealt{Kovlakas20} and \citealt{Inoue21}). Of these `new' and `recent' ULX
candidates, 48\ and 59, respectively, are HLX candidates. We
stress that even if a source is considered `new' in this respect, this does not
necessarily mean the sources are completely unknown, only that it has not
been formally catalogued as a ULX previously. For example, NGC\,300 ULX1 is
considered `new' here, even though this source is one of the few known ULX
pulsars (\citealt{Carpano18}), and as such has received significant individual
attention (\citealt{Walton18crsf, Kosec18, Vasilopoulos18, Vasilopoulos19,
Heida19}).
\begin{table}
\caption{Details of the archival ULX catalogues against which our new archive is compared}
\begin{center}
\begin{tabular}{c p{5cm}}
\hline
\hline
\\[-0.25cm]
Catalogue & Primary Source \& Notes \\
\\[-0.25cm]
\hline
\hline
\\[-0.2cm]
\cite{Colbert02} & {\it ROSAT}\ \\
\\[-0.3cm]
\cite{Swartz04} & {\it Chandra}\ \\
\\[-0.3cm]
\cite{LiuBregman05} & {\it ROSAT}\ \\
\\[-0.3cm]
\cite{LiuMirabel05} & Literature (incl. {\it ROSAT}, so positions treated as having {\it ROSAT}\ accuracy) \\
\\[-0.3cm]
\cite{Liu11} & {\it Chandra}\ \\
\\[-0.3cm]
\cite{Swartz11} & {\it Chandra}\ \\
\\[-0.3cm]
\cite{WaltonULXCat} & {\it XMM-Newton}\ (specifically 2XMM) \\
\\[-0.3cm]
\cite{Gong16} & {\it Chandra}\ (only $L_{\rm{X}} \geq 3 \times 10^{40}$\,\hbox{erg~s$^{-1}$}) \\
\\[-0.3cm]
\cite{EarnshawULXcat} & {\it XMM-Newton}\ (specifically 3XMM-DR4) \\
\\[-0.3cm]
\cite{Barrows19} & {\it Chandra}\ (specifically CSC2 HLXs) \\
\\[-0.3cm]
\cite{Kovlakas20} & {\it Chandra}\ (specifically CSC2) \\
\\[-0.3cm]
\cite{Inoue21} & Mainly {\it Chandra}\ (specifically CSC2), but also includes {\it XMM-Newton}\ and {\it Swift}\ (specifically 4XMM-DR9 and 2SXPS) \\
\\[-0.2cm]
\hline
\hline
\end{tabular}
\end{center}
\vspace*{-0.1cm}
\label{tab_oldcat}
\end{table}
\subsection{Estimation of Unknown Contaminants}
\label{sec_contam}
Although we have taken significant measures to try and remove known contaminants,
these processes can never be perfect, and so we stress that there will still be a
significant contribution of sources that are not actually ULXs in our final sample of ULX
candidates. By far the majority of these will be foreground/background sources that
coincidentally appear to be associated with the host galaxies in question in projection,
but have just not been formally identified/catalogued as such in the databases we have
utilized (and thus have not been removed by our effort to identify and exclude these
sources). Although we cannot remove these sources, it is still important to quantify
their likely contribution.
In order to do so, we broadly follow the approach taken in \cite{WaltonULXCat}, and
subsequently \cite{Sutton12} and \cite{EarnshawULXcat}. This involves a calculation of
the total expected number of sources that would be resolved from the cosmic X-ray
background (CXB) given our selection criterion, the sensitivity of the observations from
which the 4XMM, 2SXPS and CSC2 X-ray catalogues have been generated, and
the full set of galaxies in our catalogue that have been covered by the observations
that contribute to these X-ray catalogues (not just those galaxies that have ULX
detections). These estimates are then compared to the number of sources remaining in
our catalogue, after accounting for the number of identified foreground/background
sources that have already been filtered out, in order to estimate the remaining fractional
contribution from these contaminants.
In order to estimate the total expected number of contaminants, we make use of
empirically determined forms of the $N(>$\textit{S}$)$\ curves which quantify the number of sources
per square degree ($N$) resolved from the CXB as a function of flux sensitivity ($S$).
These are combined with observational sensitivity maps in order to estimate the
number of background sources each galaxy in our input sample that has been
observed should contribute. Sensitivity maps for the observations from which the
4XMM and CSC2 catalogues are compiled are provided as part of these data releases,
but are not available for the 2SXPS catalogue at the time of writing. We therefore
focus our calculations on the 4XMM and CSC2 data, performing this calculation for
each dataset separately; as will be clear below, the expected level of contamination
for these datasets are very similar, and so we still expect these results to hold overall.
For CSC2, since the initial source detection is performed using `stacks' of observations
(a stack is defined as a group of observations for which the aimpoints are all within 1
arcminute; see the CSC2 documentation), we use the sensitivity maps generated for
these stacks in our analysis. These are provided for all of the energy bands considered
in the CSC2 catalogue. However, as noted by \cite{WaltonULXCat}, owing to absorption
in the apparent host galaxies (which lie between us and any background AGN) these
calculations are most robust at higher energies, and so we limit ourselves to the hard
band (2--7\,keV) ACIS maps that correspond to the `true' detection threshold
to match our data selection (the HRC makes a negligible overall contribution here). We
also make use of the $N(>$\textit{S}$)$\ curves recently published by \cite{Masini20}, who present an
expression for the same 2--7\,keV band.
There are two limiting fluxes to consider here. The first is set by our selection of sources
that appear to have $L_{\rm{X}} \geq 10^{39}$\,\hbox{erg~s$^{-1}$}. For each galaxy we work out the
hard band flux that would correspond to a broadband luminosity of $10^{39}$\,\hbox{erg~s$^{-1}$},
$S_{\rm{ulx}}$, based on the distance to the galaxy and the fraction of the broadband
flux that would appear in the hard band. We use a coarse representation of the average
spectral shape for ULXs below 10\,keV ({\it e.g.}\ \citealt{Stobbart06, Gladstone09,
Pintore17}): an absorbed powerlaw spectrum with $\langle N_{\rm{H}} \rangle = 3 \times
10^{21}$\,\hbox{$\rm\thinspace cm^{-2}$}\ and $\langle \Gamma \rangle = 2.1$. The second flux is the limiting
sensitivity of the stack in question, $S_{\rm{obs}}$, provided by the sensitivity maps.
The relevant limiting sensitivity for use with the $N(>$\textit{S}$)$\ curve is then the larger of these
two values, such that if an observation is sensitive enough to detect sources at lower
luminosities, these are not included in our estimated number of contaminants. For each
of the galaxies covered by CSC2 we use the $N(>$\textit{S}$)$\ curve and the appropriate limiting
sensitivity to compute maps of the number of expected background sources per pixel,
and integrate these over the area of the galaxy covered by each relevant {\it Chandra}\
stack (excluding the typical area excised around the central galaxy location by our
nuclear cut). For each of the galaxies covered by CSC2 data (again, not just those with
ULX detections), we perform this calculation for every available stack. We then select
the stack that would give the largest number of contaminants, and sum these values
over all of the galaxies covered by CSC2 stacks to compute the total number of
expected contaminants prior to the removal of any known foreground/background
sources. From this, we compute the remaining fractional contamination among the
CSC2 ULX candidates by comparing the expected number of remaining contaminants
to the number of ULX candidates detected in the hard band for self-consistency ({\it i.e.~\/}
excluding sources that only have upper limits).
For 4XMM, the sensitivity maps are only provided for the full band (0.2--12.0\,keV)
and are based on the combined sensitivity for all of the EPIC detectors (see Section 9
in \citealt{4XMM}). However, as stated above, it is preferable to work in the hard band
here. Furthermore, suitable $N(>$\textit{S}$)$\ curves are not currently available for the full {\it XMM-Newton}\
bandpass; aside from work focusing specifically on {\it Chandra}, $N(>$\textit{S}$)$\ curves are
determined almost exclusively for the 0.5--2.0 and 2--10\,keV bands. It is therefore
necessary to correct the results from the available broadband maps to one of these
bands, and we again choose the harder band, but this is not a trivial process. To do
so, we also make use of the hard band (2--12\,keV) sensitivity maps computed as
part of the \cite{EarnshawULXcat} ULX catalogue for the majority of observations
that make up 4XMM-DR10 (specifically those that make up 3XMM-DR4). However,
these are computed using a different approach (see \citealt{Carrera07} and
\citealt{Mateos08}), consider each of the EPIC detectors separately, and adopt a
different detection threshold (the hard band maps are computed for a
$\sim$4$\sigma$ detection in a single detector, while the broadband maps are
computed for a $\sim$3$\sigma$ detection combining all the EPIC detectors), further
complicating the situation.
For each galaxy covered by these earlier observations, we therefore process the
available hard band sensitivity maps for each of the detectors in a similar manner as
above, using the appropriate $N(>$\textit{S}$)$\ curve published by \cite{Cappelluti09} but only
considering for the limiting observational sensitivity ($S_{\rm{obs}}$) for the time
being, and note the results for the detector that predicts the largest number of
contaminants. We also process the broadband maps for the same observations by
computing the fraction of the broadband flux in both the softer (0.5-2.0\,keV) and
harder (2--10\,keV) bands using the spectral form assumed in their generation (an
absorbed powerlaw with $N_{\rm{H}} = 1.7 \times 10^{20}$\,\hbox{$\rm\thinspace cm^{-2}$}\ and $\Gamma
= 1.42$, typical for CXB sources), processing these updated maps in turn using the
relevant $N(>$\textit{S}$)$\ curves, again only considering $S_{\rm{obs}}$, and averaging the
final results to obtain an estimate for the number of contaminating sources the
broadband maps would imply. For each galaxy covered by these earlier
observations, we then compare the results from the broadband and the hard band
maps to compute an empirical correction for the former; we find this correction to
be a factor of 9. We then process the full set of 4XMM-DR10 broadband sensitivity
maps using this correction to produce maps of the expected number of hard band
CXB sources. At this point, we also consider the number of contaminants implied
by the second limiting sensitivity, $S_{\rm{ulx}}$, and update the maps accordingly.
Similar to before, we then integrate these maps over the galaxy area covered by
every observation of that galaxy (again excluding the typical area around the
central position excised by our nuclear cut, and again considering all galaxies
covered by 4XMM-DR10) and pick the observation that gives the largest number of
hard-band contaminants for each galaxy. We then sum these values to compute the
total number of expected contaminants prior to the removal of any known
foreground/background sources, and finally compute the remaining fractional
contamination among the 4XMM ULX candidates (comparing the expected number
of remaining contaminants against the number of ULX candidates that are detected
at the 4$\sigma$ level in any of the EPIC detectors for self-consistency).
Based on these approaches, and the numbers of known foreground/background
contaminants already removed, we estimate fractional contaminations of $(23 \pm
2)$\%, and $(18 \pm 3)$\% for our CSC2 and 4XMM-DR10 ULX candidates,
respectively (quoted uncertainties are due to counting statistics, and are
1$\sigma$). These values are sufficiently similar that, even though the relevant
sensitivity maps are not yet available for 2SXPS, we still expect that an overall
fractional contamination of $\sim$20\% is relevant for the whole catalogue.
\section{NGC\,3044 ULX1 -- A New Extreme ULX}
\label{sec_ulxs}
The non-uniform selection means the full ULX sample presented here is not necessarily
well suited for detailed statistical studies of the ULX population similar to those
presented by \cite{Kovlakas20} and \cite{Inoue21}. Indeed, our intention in compiling
this sample is to facilitate follow-up studies of interesting individual sources. As such,
in order to highlight the potential of our catalogue, we instead present a case study of
a new extreme ULX candidate with $L_{\rm{X,peak}} \sim 10^{40}$\,\hbox{erg~s$^{-1}$}\ in the
edge-on spiral galaxy NGC\,3044 by both {\it Swift}\ and {\it XMM-Newton}\ (see Figure
\ref{fig_ngc3044_img}). Although we have found several new HLX candidates in our
analysis, we highlight this new source in particular both because of its luminosity is still
very extreme, but also because it already has high signal-to-noise (S/N) {\it XMM-Newton}\ data
(several thousand counts) available in the archive; as noted above, our new HLX
candidates will instead be studied in future work (A. D. Mackenzie, \textit{in prep.}).
Hereafter we refer to this source as NGC\,3044 ULX1 for simplicity, as it is the
brightest ULX candidate in NGC\,3044, but its catalogued 4XMM-DR10 and 2SXPS
IAU names are 4XMM\,J095343.8+013416 and 2SXPS\,J095343.7$+$013417,
respectively. Throughout this analysis, we assume a distance of $D = 20.6$\,Mpc to
NGC\,3044 (\citealt{Tully16}).
\begin{figure}
\begin{center}
\hspace*{-0.2cm}
\rotatebox{0}{
{\includegraphics[width=235pt]{./figs/image_ngc3044_xmm.eps}}
}
\end{center}
\vspace*{-0.3cm}
\caption{
{\it XMM-Newton}\ image of NGC\,3044 from OBSID 0782650101. As with Figure \ref{fig_image},
the D25 extent of NGC\,3044 is shown with the blue dashed ellipse, and the
nuclear position with a white cross. The positions of the brightest ULX candidate in
NGC\,3044, which we refer to as ULX1, and the nearby supernova 1983E are
indicated.}
\label{fig_ngc3044_img}
\end{figure}
\subsection{Observations and Data Reduction}
\label{sec_ngc3044_obs}
NGC\,3044 has been observed on four occasions by {\it XMM-Newton}, and on five occasions by
{\it Swift}. A log of these observations is given in Table \ref{tab_ngc3044_obs}. We primarily
focus on the {\it XMM-Newton}\ observations here, but also process the {\it Swift}\ observations to
provide further information on the long-term variability.
The {\it XMM-Newton}\ data for each observation are reduced following standard procedures using
the {\it XMM-Newton}\ Science Analysis System (\textsc{sas}\ v19.1.0). All of the {\it XMM-Newton}\ observations were
taken in full frame mode. Raw observation files for the {EPIC-pn}\ and {EPIC-MOS}\ detectors
are cleaned using \textsc{epchain}\ and \textsc{emchain}, respectively. In order to facilitate pulsation
searches, the cleaned {EPIC-pn}\ event files are corrected to the solar barycentre using the
DE200 solar ephemeris, as this has the best time resolution of the {\it XMM-Newton}\ detectors
(73.4\,ms in full frame mode). Source products are extracted from the cleaned event
files with \textsc{xmmselect}. Given the relative proximity of supernova SN1983E (separated by
$\sim$35$''$; see Figure \ref{fig_ngc3044_img}), we use circular source regions of
radius 15--20$''$, with the larger region size used for the higher flux observations (see
below). Background is estimated from a larger region of blank sky on the same detector
as ULX1. All of the observations suffer from periods of enhanced background to some
degree, and for each observation we determine the background threshold that
maximises the source S/N over the full {\it XMM-Newton}\ band considered in our more detailed
analysis (0.3--10.0\,keV) following the method outlined in \cite{Picon04}. Only single
and double patterned events are considered for {EPIC-pn}\ ({\small PATTERN}\,$\leq$\,4)
and single to quadruple patterned events for {EPIC-MOS}\ ({\small PATTERN}\,$\leq$\,12),
as recommended, and all of the necessary instrumental response files were generated
using \textsc{arfgen}\ and \textsc{rmfgen}. After performing the reduction separately for the two
{EPIC-MOS}\ units, we combine their individual spectra using \textsc{addascaspec}.
\begin{table}
\caption{Details of the X-ray observations of NGC\,3044 ULX1 considered in this work.}
\begin{center}
\begin{tabular}{c c c c}
\hline
\hline
\\[-0.25cm]
Mission & OBSID & Start Date & Good \\
& & & Exposure (ks)\tmark[a] \\
\\[-0.3cm]
\hline
\hline
\\[-0.15cm]
{\it XMM-Newton}\ & 0070940101 & 2001-11-24 & 6/8 \\
\\[-0.25cm]
{\it XMM-Newton}\ & 0070940401 & 2002-05-10 & 9/21 \\
\\[-0.25cm]
{\it XMM-Newton}\ & 0720252401 & 2013-05-06 & 9/11\tmark[b] \\
\\[-0.25cm]
{\it Swift}\ & 00092188001 & 2015-04-17 & 1 \\
\\[-0.25cm]
{\it Swift}\ & 00092188002 & 2015-04-19 & 1 \\
\\[-0.25cm]
{\it Swift}\ & 00092188003 & 2015-06-23 & 2 \\
\\[-0.25cm]
{\it Swift}\ & 00092188004 & 2015-06-24 & 2 \\
\\[-0.25cm]
{\it Swift}\ & 00092188005 & 2015-06-25 & 2 \\
\\[-0.25cm]
{\it XMM-Newton}\ & 0782650101 & 2016-12-07 & 80/93 \\
\\[-0.2cm]
\hline
\hline
\\[-0.15cm]
\end{tabular}
\\
\flushleft
$^{a}$ {\it XMM-Newton}\ exposures are listed for the {EPIC-pn}/MOS detectors, after
filtering for background flaring (see Section \ref{sec_ngc3044_obs}). \\
$^{b}$ ULX1 falls on a dead chip for the MOS1 detector in this observation.
\vspace*{0.3cm}
\label{tab_ngc3044_obs}
\end{center}
\end{table}
\subsection{Spectral Analysis}
\label{sec_spec}
Based on the 4XMM-DR10 data, the first two {\it XMM-Newton}\ observations (2001 and 2002) both
caught NGC\,3044 ULX1 in a lower flux state, while the latter two (2013 and 2016)
caught the source in a higher flux state. We therefore combine the {\it XMM-Newton}\ spectra from
these pairs of observations using \textsc{addascaspec}\ to provide the highest S/N data possible
for these two flux regimes. These spectra are shown in Figure \ref{fig_ngc3044_spec}.
We initially begin by fitting the high-flux data with a simple absorbed powerlaw model.
We use \hbox{\small XSPEC}\ for our spectral analysis (\citealt{xspec}), and allow for both the
Galactic absorption column of $N_{\rm{H,Gal}} = 2.33 \times 10^{20}$\,\hbox{$\rm\thinspace cm^{-2}$}\
(\citealt{NH2016}) and further absorption at the redshift of NGC\,3044 ($z = 0.00430$)
that is free to vary in all our models. Both absorbers are modelled using \textsc{tbabs}, and we
adopt solar abundances from \cite{tbabs} and absorption cross-sections from
\cite{Verner96}. We also allow for cross-normalisation constants to float between the
data from the pn and MOS detectors to account for residual calibration differences;
these factors are always within a few per cent of unity. Finally, the higher flux data are
grouped to a minimum of 25 counts per bin to facilitate the use of {$\chi^{2}$}\ minimisation.
The absorbed powerlaw model returns a fairly steep continuum, with $N_{\rm{H,high}}
= (2.1 \pm 0.2) \times 10^{21}$\,\hbox{$\rm\thinspace cm^{-2}$}\ and $\Gamma_{\rm{high}} = 2.41 \pm 0.07$
(uncertainties on the spectral parameters are quoted at the 90\% level). Unsurprisingly,
the lower flux data have a much lower S/N (in addition to the lower flux, these data have
a much lower combined exposure). We therefore group these data to just 1 count per
bin, and fit them with the same model using the Cash statistic (\citealt{cstat}). Here we
find $N_{\rm{H,low}} = 1.1^{+0.8}_{-0.7} \times 10^{21}$\,\hbox{$\rm\thinspace cm^{-2}$}\ and
$\Gamma_{\rm{low}} = 2.3 \pm 0.4$. Within the limitations of the available data, there is
therefore little evidence for spectral variability, although the parameter constraints are
not particularly tight for the lower flux data.
Although the absorbed powerlaw model captures the overall shape of the spectrum in
the 0.3-10.0\,keV band fairly well, the high-flux data have sufficient S/N that systematic
residuals to this simple model can be seen (see Figure \ref{fig_ngc3044_fits}), implying
that a more complex continuum model is required. Indeed, the quality of fit provided by
the absorbed powerlaw model for the high flux data is {$\chi^{2}$}\ = 353 for 285 degrees of
freedom (DoF), which is not formally an acceptable fit. This residual structure is fairly
typical for extreme ULXs when fit with a single powerlaw model ({\it e.g.}\
\citealt{Stobbart06, Gladstone09, Mukherjee15}), and indicates the need for distinct
continuum components above and below $\sim$1--2\,keV. We therefore fit the
high-flux data with a few more complex models often used to describe ULX spectra.
First, we fit a model consisting of a lower energy accretion disc component, and a
higher energy powerlaw continuum. We use the \textsc{diskbb}\ model (\citealt{diskbb}) for
the former, which implicitly assumes a thin disc profile (\citealt{Shakura73}), such that
the model broadly represents the classic disc--corona geometry seen in sub-Eddington
X-ray binaries. This provides a significant improvement to the simpler powerlaw fit, with
{$\chi^{2}$}/DoF = 284/283. The best-fit parameters are given in Table
\ref{tab_ngc3044_param}.
\begin{figure}
\begin{center}
\hspace*{-0.2cm}
\rotatebox{0}{
{\includegraphics[width=235pt]{./figs/ngc3044ulx1_eeuf_2states.eps}}
}
\end{center}
\vspace*{-0.3cm}
\caption{
The {\it XMM-Newton}\ spectra of NGC\,3044 ULX1 from the high- and low-flux states seen in
the available data, unfolded through a model that is constant with energy. The {EPIC-pn}\
data are shown in black and green, and the {EPIC-MOS}\ data are shown in red and blue.
The data have been rebinned for visual clarity.}
\label{fig_ngc3044_spec}
\end{figure}
\begin{figure}
\begin{center}
\hspace*{-0.2cm}
\rotatebox{0}{
{\includegraphics[width=235pt]{./figs/ngc3044ulx1_highflux_model_fits.eps}}
}
\end{center}
\vspace*{-0.2cm}
\caption{
\textit{Top panel:} The relative contributions of the best-fit \textsc{diskbb+diskpbb}
model to the high-flux {\it XMM-Newton}\ data for NGC\,3044 ULX1. The total model is shown in
black, the \textsc{diskbb} component in blue and the \textsc{diskpbb} component
in red, respectively. \textit{Lower panels:} The data/model ratio for a simple absorbed
powerlaw continuum model, a \textsc{diskbb}+\textsc{powerlaw} continuum and the
\textsc{diskbb+diskpbb} models, respectively. For the ratio panels, the colours have
the same meanings as in Figure \ref{fig_ngc3044_spec}, and the data have again
been rebinned for visual purposes.}
\label{fig_ngc3044_fits}
\end{figure}
\begin{table*}
\caption{Key parameters obtained for the various continuum model fits to the high-flux data available for NGC\,3044 ULX1}
\begin{center}
\begin{tabular}{c c c c c c}
\hline
\hline
\\[-0.2cm]
Model & Parameter & & Continuum Model \\
\\[-0.35cm]
Component & & \textsc{powerlaw} & \textsc{diskbb}+\textsc{powerlaw} & \textsc{diskbb}+\textsc{diskpbb}\ \\
\\[-0.25cm]
\hline
\hline
\\[-0.1cm]
\textsc{tbabs}\ & $N_{\rm{H}}$ [$10^{21}~\rm{cm}^{-2}$] & $2.1 \pm 0.2$ & $3.5^{+0.7}_{-0.6}$ & $3.0 \pm 0.6$ \\
\\[-0.15cm]
\textsc{diskbb}\ & $T_{\rm{in}}$ [keV] & -- & $0.16 \pm 0.02$ & $0.19 \pm 0.03$ \\
\\[-0.25cm]
& Norm & -- & $37^{+80}_{-25}$ & $13^{+26}_{-7}$ \\
\\[-0.15cm]
\textsc{powerlaw} & $\Gamma$ & $2.41 \pm 0.07$ & $2.2 \pm 0.1$ & -- \\
\\[-0.25cm]
& Norm [$10^{-5}$] & $6.1 \pm 0.3$ & $5.0^{+0.6}_{-0.5}$ & -- \\
\\[-0.15cm]
\textsc{diskpbb}\ & $T_{\rm{in}}$ [keV] & -- & -- & $1.7^{+0.7}_{-0.4}$ \\
\\[-0.25cm]
& $p$ & -- & -- & $0.56^{+0.15}_{-0.05}$ \\
\\[-0.25cm]
& Norm [$10^{-4}$] & -- & -- & $3.6^{+5.7}_{-0.5}$ \\
\\[-0.2cm]
\hline
\\[-0.2cm]
$\chi^{2}$/DoF & & 353/285 & 284/283 & 273/282 \\
\\[-0.25cm]
\hline
\hline
\\[-0.15cm]
\end{tabular}
\label{tab_ngc3044_param}
\end{center}
\end{table*}
There is still a mild hint of curvature in the spectrum at higher energies though ($E >
2$\,keV; Figure \ref{fig_ngc3044_fits}). This is seen in the majority of high S/N ULX
spectra, initially implied by {\it XMM-Newton}\ ({\it e.g.}\ \citealt{Gladstone09, Walton4517}) and then
unambiguously confirmed by the higher energy coverage provided by {\it NuSTAR}\ ({\it e.g.}\
\citealt{Bachetti13, Walton14hoIX, Rana15}). These broadband observations find that
ULX spectra are primarily described by two thermal components below 10 keV. We
therefore also fit a second model that is often considered for ULXs, combining two
accretion disc components.\footnote{The best {\it NuSTAR}\ data available for ULXs
shows that a third continuum component is typically required above $\sim$10\,keV
to model the broadband spectra ({\it e.g.}\ \citealt{Walton18ulxBB}). However, without
coverage of these energies we cannot say anything about the presence of this
component here, and so restrict ourselves to a simpler two-component model for
the data below 10\,keV.} Specifically, for the higher energy emission we replace the
powerlaw component with a \textsc{diskpbb}\ model (\citealt{diskpbb}). This allows for the
radial temperature index ($p$) to be an additional free parameter, such that the
model can mimic a thick, advection-dominated super-Eddington accretion disc
(which would be expected to have $p < 0.75$ instead of the $p = 0.75$ appropriate
for thin accretion discs; \citealt{Abram88}). The \textsc{diskbb}+\textsc{diskpbb}\ does provide a
moderate additional improvement in fit over the \textsc{diskbb}\ and powerlaw combination,
with {$\chi^{2}$}/DoF = 273/282 ({\it i.e.~\/} an improvement of $\Delta\chi^{2} = 11$ for one extra
free parameter). As our best-fit model, we show the relative contributions of the
different model components in Figure \ref{fig_ngc3044_fits}, and the parameter
constraints are also given in Table \ref{tab_ngc3044_param}.
The best-fit spectral form for NGC\,3044 ULX1 differs somewhat from that used to
compute the fluxes in 4XMM-DR10, so we also re-calculate the observed 0.3--10\,keV
fluxes for the individual {\it XMM-Newton}\ observations using the spectral models for the high-
and low-flux states discussed above. To further examine the long-term behaviour of
the source we also consider the {\it Swift}\ data at this stage. These observations can
themselves be split into two main groups, taken in April and June 2015. We process
the combined data from these two sets of observations, compute the average count
rates for each of the two groups using the same 20$''$ regions as for the {\it XMM-Newton}\ data
(correcting appropriately for the point spread function), and convert these to fluxes
using the spectral shape implied by the simple powerlaw fits to the {\it XMM-Newton}\ data. The
long term lightcurve combining the {\it XMM-Newton}\ and {\it Swift}\ data is shown in Figure
\ref{fig_ngc3044_lc}. The coverage is admittedly sparse, but the {\it Swift}\ fluxes are
consistent with the more recent {\it XMM-Newton}\ measurements, and so it appears as though
the higher flux state persisted throughout 2013--2016. We also still find the peak
luminosity of the source to be $L_{\rm{X,peak}} \sim 10^{40}$\,\hbox{erg~s$^{-1}$}, confirming the
extreme luminosity implied by our analysis of 4XMM-DR10.
\begin{figure}
\begin{center}
\hspace*{-0.2cm}
\rotatebox{0}{
{\includegraphics[width=235pt]{./figs/ngc3044ulx1_longterm_lightcurve.eps}}
}
\end{center}
\vspace*{-0.3cm}
\caption{
Longtime lightcurve for NGC\,3044 ULX1 based on the available X-ray data. {\it XMM-Newton}\
data are shown in black, and {\it Swift}\ in red. Note that the sets of {\it Swift}\ observations
taken in April and in June have been combined together here (see Section
\ref{sec_spec}).}
\label{fig_ngc3044_lc}
\end{figure}
\subsection{Timing Analysis}
The longest of the available {\it XMM-Newton}\ observations of NGC\,3044 ULX1, OBSID
0782650101, returns a total of $\sim$5000 net source counts with the {EPIC-pn}\
detector, roughly comparable to the quality of data used to detect X-ray pulsations
for some of the known ULX pulsars ({\it e.g.}\ \citealt{Israel17, Rodriguez20}). We
therefore also perform a search for pulsations on this dataset. We focus on the data
from the {EPIC-pn}\ detector here as this has both the highest count rate and the best
timing resolution of the EPIC detectors (73.4\,ms in the full-frame mode used for
this observation).
For this analysis we use the pulsar timing tools included in the \textsc{hendrics}\ package
(\citealt{bachettiHENDRICSHighENergy2018}). Since the pulse period can evolve
rapidly in ULX pulsars, either due to the secular spin-up driven by the extreme
accretion (\citealt{Fuerst16p13, Israel17, Carpano18, Vasilopoulos18}) or orbital
motion of the neutron star (\citealt{Bachetti14nat, Israel17, Fuerst18, Fuerst21}), we
perform an `accelerated' pulsation search, which considers both the frequency of
the pulsations ($f$) and its first derivative ($\dot{f}$) when searching for any
signals. Specifically, we use the \textsc{HENzsearch} script, which performs the
$Z^2_n$ search originally outlined in \cite{Buccheri83}, and allow for $n = 2$
harmonics in our search ({\it i.e.~\/} we use the $Z^2_2$ statistic). Based on the properties
of the known ULX pulsars, we focus on the frequency range of 0.01--6.75\,Hz, and
the ``fast'' search option utilized allows for $\dot{f}$ values in the range
$\pm$10$^{-8}$\,Hz~s$^{-1}$. Unfortunately we did not find any promising pulsation
candidates in this observation.
In the absence of a robust pulsation detection, we estimate an upper limit on the
pulsed fraction any undetected signal could have following the method used in
\cite{Walton21}. In short, we use the \textsc{HENzn2vspf} script, which simulates
datasets using the same GTIs and total number of events as the real data, then
uses rejection sampling to modulate the events with an increasingly strong pulsed
signal (assuming a sinusoidal pulse profile, which is appropriate for ULX pulsars),
and finally calculates the $Z^2_2$ for each pulsed fraction to see how strongly
such pulsations would have been detected. We simulate 100 datasets in order to
determine the pulsed fraction at which the $Z^2_2$ statistic reaches $\sim$40.
This threshold roughly corresponds to a 3$\sigma$ detection, and thus indicates
the equivalent upper limit on the pulsed fraction that could still be present in the
real data. We find an upper limit on the pulsed fraction of $\sim$22\% when
considering the full {\it XMM-Newton}\ bandpass.
\subsection{The Nature of NGC\,3044 ULX1}
NGC\,3044 ULX1 is a new extreme ULX discovered in our analysis that already has
high S/N data available in the archive. Although it has always been in the ULX regime
whenever observed with our current X-ray facilities (considering {\it XMM-Newton}\ and {\it Swift}\ in
combination, we have observations from 6 different epochs), sometime between
2002 and 2013 it seemed to jump up by a factor of $\sim$3--4 in luminosity from
$L_{\rm{X}} \sim 3 \times 10^{39}$\,\hbox{erg~s$^{-1}$}\ to $L_{\rm{X}} \sim 10^{40}$\,\hbox{erg~s$^{-1}$},
where it seems to have remained since (see Figure \ref{fig_ngc3044_lc}).
The 0.3--10.0\,keV X-ray spectrum observed during this high-flux period is very
similar to other extreme ULXs: the flux below 10\,keV is dominated by two continuum
components that primarily contribute above and below $\sim$1--2\,keV and,
although it is not a strong statistical detection, there is a hint of spectral curvature in
the higher energy component. We note in particular that, although there is no higher
energy coverage available here, the spectrum of NGC\,3044 ULX1 is highly
reminiscent of that seen from Holmberg II X-1 -- another extreme ULX with
$L_{\rm{X,peak}} \sim 10^{40}$\,\hbox{erg~s$^{-1}$}\ -- during the broadband observations
performed with {\it XMM-Newton}, {\it Suzaku}\ and {\it NuSTAR}\ in 2013 (\citealt{Walton15hoII}). As
noted previously, this was part of a series of broadband observations of ULXs ({\it e.g.}\
\citealt{Bachetti13, Walton14hoIX, Rana15, Mukherjee15}) that robustly confirmed
earlier indications from {\it XMM-Newton}\ ({\it e.g.}\ \citealt{Stobbart06, Gladstone09, Walton4517})
that the high-energy spectra of ULXs are distinct from those seen from
sub-Eddington black holes. While the spectra from these systems are typically
dominated by Comptonisation in an optically-thin `corona' above $\sim$2\,keV
({\it e.g.}\ \citealt{Haardt91}), the spectra of ULXs instead seem to be dominated by two
thermal components below 10\,keV, before falling away steeply at higher energies.
Indeed, the best-fitting model for the high-flux {\it XMM-Newton}\ spectra from NGC\,3044
ULX1 consists of two thermal accretion disc components.
The distinct broadband spectra of ULXs, along with the detection of X-ray
pulsations (\citealt{Bachetti14nat, Fuerst16p13, Israel17, Israel17p13, Carpano18,
Sathyaprakash19, Rodriguez20}) and extreme outflows (\citealt{Pinto16nat, Pinto17,
Pinto20, Walton16ufo, Kosec18}) from a growing number of these systems have,
together, helped clearly establish that the majority of the ULX population is
dominated by super-Eddington accretors. In this context, the two continuum
components seen in ULXs below 10\,keV likely represent the complex thermal
emisson from a hot, super-Eddington accretion disc (and potentially its associated
outflow; {\it e.g.}\ \citealt{Poutanen07, Middleton15}). Given its similarity to other, better
studied ULXs that are now well accepted to be super-Eddington accretors,
NGC\,3044 ULX1 is therefore likely another super-Eddington system. As discussed
above, these sources are of particular interest, as they may provide a local
observational window into the conditions required to rapidly grow SMBHs in the
early universe ({\it e.g.}\ \citealt{Banados18nat}).
We have searched for X-ray pulsations from NGC\,3044 ULX1, which would
unambiguously identify the accretor as another neutron star and confirm its nature
as a highly super-Eddington system. We focused on the 2016 data, which by far
have the best S/N among the available observations, and searched for pulsations
over the 0.01--6.75\,Hz frequency range based on the properties of the known
ULX pulsars, but unfortunately we did not find a robust detection of any such
variations. However, even though the available data has quite high S/N, we can only
place an upper limit of $\sim$20\% on the pulsed fraction of any pulsations present
during this observation. Although pulsations of the strength seen in NGC\,300 ULX1
can therefore be firmly excluded (pulsed fraction of $\sim$60\% below 10\,keV;
\citealt{Carpano18}), other known ULX pulsars exhibit pulsed fractions that are lower
than this limit in the {\it XMM-Newton}\ bandpass ({\it e.g.}\ \citealt{Sathyaprakash19, Rodriguez20}).
Furthermore, even in ULXs that are known to be pulsars the pulsations can be
transient, and are not always observed even when the data should have sufficient
S/N to see them ({\it e.g.}\ \citealt{Israel17, Bachetti20}). As such, even though we have
not seen any clear evidence for X-ray pulsations from NGC\,3044 ULX1, we cannot
exclude the possibility that this is another neutron star ULX.
Although the comparison with other ULXs is quite compelling, obtaining higher
energy data would be of particular use in order to more robustly confirm NGC\,3044
ULX1 as another super-Eddington accretor. In the known ULX pulsars the pulsed
fraction is seen to increase with energy, perhaps because non-pulsed components
from the accretion flow make a more significant contribution below $\sim$10\,keV
({\it e.g.}\ \citealt{Walton18ulxBB}). Higher energy coverage would therefore help to
mitigate against these issues in terms of further pulsation searches, and would also
allow us to extend the continuum spectroscopy above 10\,keV, and further confirm
that the broadband spectrum of NGC\,3044 ULX1 is similar to other ULXs.
Unfortunately, given both the fairly low peak flux from NGC\,3044 ULX1 ($\sim$2
$\times$ 10$^{-13}$\,\hbox{$\rm\thinspace erg~cm^{-2}~s^{-1}$}\ in the 0.3--10\,keV band; see Figure
\ref{fig_ngc3044_lc}) and the fairly close proximity of SN1983E (see Figure
\ref{fig_ngc3044_img}), meaningful observations of NGC\,3044 ULX1 with {\it NuSTAR}\
would likely be very challenging. This may, instead, be a suitable target for a facility
like the \textit{High Energy X-ray Probe} (\textit{HEX-P}; \citealt{HEXP_tmp}), which
would have both superior sensitivity and imaging capabilities to {\it NuSTAR}.
\section{Summary and Outlook}
\label{sec_conc}
We have compiled a new catalogue of ULX candidates, combining the latest data
releases from each of the {\it XMM-Newton}, {\it Swift}\ and {\it Chandra}\ observatories (the 4XMM-DR10,
2SXPS and CSC2 source catalogues, respectively). Our new catalogue contains
1843\ sources residing in 951\ different host galaxies, making it the largest ULX
catalogue compiled to date. Of these, 689\ sources are catalogued as ULX
candidates for the first time. Our sample also contains 71\ HLX candidates, of which
48\ are new catalogue entries. We have made significant efforts to clean the
catalogue of known non-ULX contaminants (e.g. foreground stars, background AGN,
supernovae), and estimate that the remaining contribution of unknown contaminants
is $\sim$20\%. Our primary motivation here is to unearth new sources of interest for
detailed follow-up studies, and among this new catalogue we have already found one
new extreme ULX candidate with high S/N data in the archive: NGC\,3044 ULX1. This
shows a factor of at least $\sim$4 variability on long timescales, based on the
available {\it XMM-Newton}\ and {\it Swift}\ data, with a peak luminosity of $L_{\rm{X,peak}} \sim
10^{40}$\,\hbox{erg~s$^{-1}$}\ to date. The {\it XMM-Newton}\ spectrum of the source while at this peak flux is
reminiscent of other extreme ULXs (and Holmberg II X-1 in particular), and is best-fit
by a model combining two thermal accretion disc components. This likely indicates
that NGC\,3044 ULX1 is another member of the ULX population accreting at
super-Eddington rates.
We anticipate this new catalogue will be a valuable resource for planning further
observational campaigns, both with our current X-ray imaging facilities ({\it XMM-Newton},
{\it Chandra}, {\it Swift}, {\it NuSTAR}) and with upcoming missions such as {\it XRISM}\ and, in particular,
{\it Athena}. Our new catalogue should also help to facilitate further studies of ULXs at
longer wavelengths, particularly in the era of the new optical, NIR and radio facilities
due to come online (the thirty-metre class ground-based observatories, \textit{JWST},
the SKA). Such work will be vital for determining the contribution of ULX pulsars to the
broader ULX population, their accretion physics, the prevalence of extreme outflows
among the ULX population and the impact of the winds launched by super-Eddington
accretors, and for the hunt for the first dynamically confirmed black hole ULX. Further
iterations of the {\it XMM-Newton}, {\it Swift}\ and {\it Chandra}\ serendipitous surveys, combined with the
upcoming results from {\it eROSITA}\ (\citealt{EROSITA}), will also allow us to continue
expanding this ULX sample in the future.
\section*{ACKNOWLEDGEMENTS}
DJW acknowledges support from the Science and Technology Facilities Council (STFC)
via an Ernest Rutherford Fellowship (ST/N004027/1).
TPR also acknowledges support from STFC via consolidated grant ST/000244/1.
SM acknowledges financial support from the Spanish Ministry MCIU under project
RTI2018-096686-B-C21 (MCIU/AEI/FEDER/UE), cofunded by FEDER funds and from
the Agencia Estatal de Investigaci\'{o}n, Unidad de Excelencia Mar\`{i}a de Maeztu, ref.
MDM-2017-0765.
This research has made use of data obtained with {\it XMM-Newton}, an ESA science mission with
instruments and contributions directly funded by ESA Member States, as well as
public data from the {\it Swift}\ data archive.
This work has also made use of data obtained from the {\it Chandra}\ Source Catalog,
provided by the {\it Chandra}\ X-ray Center (CXC) as part of the {\it Chandra}\ Data Archive,
as well as public data from the Swift data archive.
This paper made use of the Whole Sky Database (WSDB) created by Sergey Koposov
and maintained at the Institute of Astronomy, Cambridge by Sergey Koposov, Vasily
Belokurov and Wyn Evans with financial support from STFC and the European
Research Council (ERC), as well as the Q3C software (\citealt{Q3C}).
This research has also made use of the
NASA/IPAC Extragalactic Database (NED), which is funded by the National Aeronautics
and Space Administration and operated by the California Institute of Technology, as
well as the SIMBAD database, operated at CDS, Strasbourg, France, and we further
acknowledge usage of the HyperLEDA database.
\section*{Data Availability}
All of the raw data underlying this article are publicly available from ESA's {\it XMM-Newton}\ Science
Archive\footnote{https://www.cosmos.esa.int/web/xmm-newton/xsa}, NASA's HEASARC
database\footnote{https://heasarc.gsfc.nasa.gov/} and NASA's {\it Chandra}\ Data
Archive \footnote{https://cxc.harvard.edu/cda/}. The primary X-ray catalogues
(4XMM\footnote{http://xmmssc.irap.omp.eu/Catalogue/4XMM-DR10/4XMM\_DR10.html},
2SXPS\footnote{https://www.swift.ac.uk/2SXPS/}
CSC2\footnote{https://cxc.harvard.edu/csc/}),
galaxy catalogues
(HyperLEDA\footnote{http://leda.univ-lyon1.fr/},
CNG\footnote{https://www.sao.ru/lv/lvgdb/} and
Cosmicflows\footnote{http://edd.ifa.hawaii.edu/})
and general catalogues
(NED\footnote{https://ned.ipac.caltech.edu/},
SIMBAD\footnote{http://simbad.u-strasbg.fr/simbad/})
used in this work are also all publicly available via the links provided. The final catalogues
of ULX candidates produced here will also be made publicly available via the VizieR
archive\footnote{https://vizier.u-strasbg.fr/viz-bin/VizieR} after the publication of this
work.
\bibliographystyle{/Users/dwalton/papers/mnras}
|
\section{Introduction}
The fundamental motivation for this article is the following question:
\medskip
\par \noindent
{\it
Are there germs of meromorphic functions, in the real or in the complex plane, with a given indeterminacy structure and
prescribed zeroes and poles?
}
\medskip
\noindent
By the indeterminacy structure of a meromorphic function we mean the information given by the dicritical components of its resolution by a sequence of quadratic
blow-ups, i.e. the irreducible components of the exceptional divisor not contained in a level set of the lifted function.
We prescribe zeroes and poles by providing a finite set $\sep$ of analytic branches, i.e. irreducible analytic curves at the origin.
In the complex case, we can give a positive answer to this question once we complete $\sep$ into a finite set
of branches satisfying some conditions of axiomatic nature, so that it forms a set of ``separatrices'' (Theorem \ref{teo-structure-meromporphic}).
In the real case, the answer is always positive, whoever $\sep$ is (Theorem \ref{teo-structure-meromporphic-real}).
We handle the above question in the broader context of the local theory of singular holomorphic foliations in the complex plane, by addressing the specific problem of building the so-called logarithmic models.
They consist in the construction of germs of singular foliations of logarithmic type ---
those defined by closed meromorphic $1$-forms with simple poles --- with a prescribed set of geometric data:
dicritical components (non-invariant components of the exceptional
divisor in the reduction of singularities),
separatrices (invariant complex analytic branches), and Camacho-Sad indices. The latter are local residue-type invariants,
associated with pairs
singularity/separatrix, that play a significant role in the local topological characterization of the foliation and also impose
combinatorial restrictions along its reduction of singularities \cite{camacho1982,linsneto1986,suwa1995}.
These data may originate, for instance, from a germ of singular complex analytic foliation $\F$ of \emph{generalized curve} type \cite{camacho1984}, which means
that, if $\pi$ is a reduction of singularities for $\F$ by a sequence of
quadratic blow-ups, then the lifted foliation
$\pi^{*}\F$ has no saddle-node singularities, i.e. simple singularities with one zero eigenvalue (see Examples \ref{ex-foliation-structure} and \ref{ex-foliation-separatrices}).
In this case, a logarithmic model for $\F$ is
a germ of logarithmic foliation $\LL$ that, roughly speaking, has
the same reduction of singularities of $\F$, with dicritical components positioned in the same places, having the same
isolated separatrices (those whose lift by $\pi$ touch invariant components of the exceptional divisor) and the same Camacho-Sad indices.
All of this is considered up to the existence of some extra points, placed in the smooth part of dicritical components,
where $\LL$ has a holomorphic first integral and demands additional blow-ups in order to complete its reduction of
singularities. These are called
\emph{escape points}. They give rise to some isolated separatrices for $\LL$, called accordingly \emph{escape separatrices}.
The logarithmic model is said to be \emph{strict} if there are no escape points, thus $\F$ and $\LL$ have the
same reduction of singularities and, hence, coincident sets of isolated separatrices.
Logarithmic models of foliations in dimension two are studied in the articles \cite{corral2003},
\cite{canocorral2006} and \cite{corral2012}, where the cases considered are, respectively,
complex non-dicritical, complex dicritical and real non-dicritical.
Several ideas and arguments we present here owe much to these three papers.
Our purpose in this article is to make a construction of general nature that encompasses the three aforementioned cases and, further,
deals with the remaining real dicritical case.
Our approach is somewhat different from that of \cite{corral2003} and
\cite{canocorral2006}, bearing some resemblance to the philosophy of \cite{corral2012}.
The core of our method is the consideration of blocks of information, with an axiomatic structure inspired in the properties of foliations of generalized curve type,
having, in principle, no connection with a concrete foliation. They are built upon
the various steps of a sequence of quadratic blow-ups $\pi$, that simulates the reduction of singularities of some hypothetical foliation.
The first of these blocks is formed by the exceptional divisors of $\pi$ and of its intermediate factors, having some distinguished irreducible components
that play the role of dicritical components of the reduction divisor of a singular foliation.
The second is a finite set $\sep_{0}$ of complex
analytic branches at $(\C^{2},0)$, along with their strict transforms by all steps of $\pi$,
that impersonate separatrices of a foliation and are thus named this way.
These two blocks are called, respectively, \emph{dicritical structure} (Definition \ref{def-dic-structure}) and \emph{configuration of separatrices}
(Definition \ref{def-conf-separatrices}). Together
they form a \emph{dicritical duplet}.
Then, we have a set of local complex invariants assigned to separatrices and non-dicritical components,
which emulate Camacho-Sad indices, called \emph{system of indices} (Definition \ref{def-system-indices}).
Its attachment to a dicritical duplet forms a \emph{dicritical triplet}.
Our goal is then to assign, to separatrices and non-dicritical components, invariants in $\C^{*}$ called \emph{residues}, to be assembled in
a fourth block called \emph{system of residues} (Definition \ref{def-system-residues}), in
a consistent way with the information provided by a dicritical triplet (Definition \ref{def-consistent}), in order to form a
\emph{dicritical quadruplet}. Simultaneously, we wish that a logarithmic $1$-form with poles in $\sep_{0}$ and the given residues defines a germ
of singular foliation $\LL$ that is a \emph{logarithmic model}
for all blocks of data considered (Definition \ref{def-Q-logarithmic}).
In a nutshell, up to the existence of escape points and escape separatrices, this means that the reduction of singularities of $\LL$, its isolated separatrices, Camacho-Sad indices and residues are precisely those provided
by the dicritical quadruplet.
The existence of consistent data of residues and of a logarithmic $1$-form with these properties is given by Theorem \ref{teo-quadruplet},
which will be the main source for all other results in this paper.
Its proof relies mostly on arguments of \cite{canocorral2006}, properly adapted to our objects.
The axiomatic construction of our dicritical multiplets is carried out in Section
\ref{section-dicritical-structure}
and in Section \ref{section-system-indices}, while the proof of Theorem \ref{teo-quadruplet} is
done in Section \ref{section-existence-quadruplet}.
Next, in Section \ref{section-moromorphic-functions}, we state and prove Theorem \ref{teo-structure-meromporphic}, which gives an answer to the complex version of the question that opens this article.
We start with a dicritical duplet containing the prescribed data of indeterminacy and branches of zeroes and poles.
Then, using combinatorial tools of \cite{camacho1988},
we produce a system of rational negative indices at the final level --- the one corresponding to the whole sequence of blow-ups
$\pi$.
This condition, necessary for them to be actual
Camacho-Sad indices of reduced models of a meromorphic function,
also happens to be sufficient. Indeed,
Theorem \ref{teo-quadruplet} provides a system of rational residues, so that the wished meromorphic function arises in a straightforward way.
Then, in Section \ref{section-real-models}
and in Section \ref{section-escape}, we turn our attention to the real case, proving the existence of real logarithmic models in Theorem \ref{teo-real-logmodel}.
Our point of departure is a sort of real dicritical triplet containing only real information (sequences of quadratic blow-ups, analytic branches and indices, all of them real),
with a weakened axiomatic structure (Definition \ref{def-real-quasi-dicritical-triplet}). Its complexification can be completed into a dicritical triplet,
symmetric with respect to the involution induced
by the complex conjugation.
The application of Theorem \ref{teo-quadruplet} then provides a system of real residues and a logarithmic $1$-form,
both symmetric with respect to the conjugation. The restriction of this complex logarithmic $1$-form to the real trace provides the
logarithmic model we seek.
It turns out that, in the real case, by adding some dicritical separatrices with appropriate residues, real escape points can be eliminated
(Proposition \ref{prop-withour-real-escape}). This, combined with Theorem \ref{teo-real-logmodel}, can be used in order to produce strict real logarithmic models,
as we do in Theorem \ref{teo-strict-logmodel-real}, where they are obtained for
germs of real analytic foliations of real generalized curve type, i.e. whose real reductions of singularities lead only to non-degenerate final models. Also, by the application of
the ideas converging to Theorem \ref{teo-real-logmodel}, we obtain
Theorem \ref{teo-structure-meromporphic-real}, which gives an affirmative answer to the real version of our opening question.
Finally, in the last part of the article, Section \ref{section-sectorial}, we study sectorial decompositions of germs of real analytic vector fields.
A classical result by I. Bendixson \cite{bendixson1901} asserts that a real analytic vector field, with an isolated
singularity at $0 \in \R^{2}$, of non-monodromic type, induces a decomposition of a small neighborhood of the origin in
a finite number of sectors.
The term non-monodromic refers to the fact that there exists at least one characteristic orbit,
i.e. one accumulating to the origin
with a well defined tangent at the limit point.
Then, each sector is limited by a pair of characteristic orbits and bears a classification into hyperbolic, parabolic or elliptic,
according to the topological behavior of the orbits inside of it (see Section \ref{section-sectorial}
and also \cite{ilyashenko2008,roussarie2020} for a detailed description).
The essential information concerning a sectorial decomposition --- a finite number of real analytic semicurves, with the mentioned accumulation properties, defining sectors classified into hyperbolic, parabolic or elliptic --- would then define
a \emph{sectorial model}.
A second guiding question, also a motivation for the development
of real logarithmic models described above,
is posed in the following terms:
\medskip
\par \noindent
{\it
Given a sectorial model at $(\R^{2},0)$, is there a germ of real analytic vector field
that realizes it?
}
\medskip
\par
We answer this question in a particular case.
Within the family
of non-monodromic vector fields, we consider those
whose orbits accumulating to $0 \in \R^{2}$ are contained in the trace of
real analytic branches and call them $\ell$-analytic (Definition \ref{def-l-analytic}).
With them, we associate more refined sectorial models, also called
$\ell$-analytic, that aggregate infinitesimal information of
dicritical components, which in turn account for the existence of parabolic and elliptic sectors
(Definition \ref{def-sectorial-modeled}).
An almost immediate application of
the existence of real logarithmic models gives
Theorem \ref{teo-sectorial-modeled}, which asserts that any such a model is realized by a germ of $\ell$-analytic vector field, and
Theorem \ref{teo-sectorial-modeled-meromorphic}, which assures that it can be actually realized by a germ of real meromorphic function.
The authors are grateful to Felipe Cano for
asking us the above question on the existence of meromorphic functions.
\section{Dicritical structure and configuration of separatrices}
\label{section-dicritical-structure}
The problems we deal with concern the construction of singular holomorphic foliations with prescribed dicritical components in their reductions of singularities
and prescribed sets of invariant analytic curves. As we commented in the introduction, we approach this by means of abstract structures
that handle these information in an axiomatic manner. In this section we introduce two of them:
dicritical structure and configuration of separatrices.
\subsection{Infinitesimal classes}
We start by establishing an intrinsic way to identify an irreducible component of the exceptional divisor of a sequence of quadratic blow-ups.
Let $\pi:(\tilde{M},\D) \to (\C^{2},0)$ be a sequence of quadratic blow-ups. The set
$\D = \pi^{-1}(0)$, the \emph{exceptional divisor} of $\pi$, is a normal crossings divisor with finitely many irreducible components --- to which
we refer simply as \emph{components} --- biholomorphic to projective lines.
We call the regular points of $\D$ \emph{trace points} and its singular points \emph{corners}.
We denote by $\br_{0}$ the family of all complex analytic branches --- i.e. germs of irreducible complex analytic curves --- at $(\C^{2},0)$.
More generally, if $M$ is a complex surface, represent by
$\br_{p}$ the family of complex analytic branches at $p \in M$ and,
if $\CC \subset M$ is a complex analytic curve, by $\br_{p}(\CC) \subset \br_{p}$ the set of irreducible local components
of $\CC$ at $p$.
If $B \in \br_{0}$, we say that the strict transform $\pi^{*}B$ is \emph{transversal} to $\D = \pi^{-1}(0)$ if $\pi^{*}B$ is smooth and touches $\D$ transversally at a trace point.
In this case, we denote by
$D^{\pi}(B)$ the unique component of $\D$ touched by $\pi^{*}B$.
Let $D \subset \D$ be a component of
the exceptional divisor of the above $\pi$. We associate with $D$ an
\emph{infinitesimal class} $\kappa(D)$, which is,
essentially, the family of all sequences of blow-ups that ``generate $D$'': sequences of
blow-ups or blow-downs over $(\tilde{M}, \D)$, the latter not collapsing $D$, considered up to isomorphism.
For the sake of formality, we give a more precise definition.
Let us consider, along with $(\pi,D)$, all pairs $(\pi',D')$, where $\pi': (\tilde{M}', \D') \to (\C^{2},0)$ is a sequence of blow-ups and
$D' \subset \D'$ is a component, such that:
\begin{itemize}
\item either $\pi$ factors $\pi'$, that is, there exists a sequence of blow-ups
$\varsigma$ such that $\pi' = \pi \circ \varsigma$, and $D' = \varsigma^{*}D$;
\item or $\pi'$ factors $\pi$, that is, there exists a sequence of blow-ups
$\varsigma'$ such that $\pi = \pi' \circ \varsigma'$, and $D = \varsigma'^{*}D'$.
\end{itemize}
Next, extend this construction putting, in place of $\pi$ and $D$, any sequence of blow-ups isomorphic to $\pi$ and the isomorphic image of $D$.
The family of all these mappings is the \emph{infinitesimal class} $\kappa(D)$ of $D \subset \D$. Any of its elements is a
a \emph{realization} of $\kappa(D)$.
Recall that two sequences of blow-ups
$\pi^{(1)}: (\tilde{M}^{(1)} , \D^{(1)}) \to (\C^{2},0)$ and $\pi^{(2)}: (\tilde{M}^{(2)} , \D^{(2)}) \to (\C^{2},0)$
are \emph{isomorphic} if there is a germ of biholomorphism $\Phi: (\tilde{M}^{(1)} , \D^{(1)} ) \to (\tilde{M}^{(2)}, \D^{(2)})$ such
that $\pi^{(1)} = \pi^{(2)} \circ \Phi$.
A realization of $\kappa(D)$ is \emph{minimal} if it is factored by no other realization of $D$. A minimal realization is unique up isomorphism.
If $\D^{*}$ is a union of components of $\D$, then we define the infinitesimal class
$\kappa(\D^{*})$ as the family of sequences of blow-ups that simultaneously realize all infinitesimal classes $\kappa(D)$, for $D \subset \D^{*}$, and also separate these components.
The latter means that, if $\pi':(\tilde{M}',\D') \to (\C^{2},0)$ is a realization of $\kappa(\D^{*})$, then
no two components of the divisor $\D'= \pi'^{-1}(0)$ corresponding to components in $\D^{*}$ intersect.
We can also define, in an evident manner, a \emph{minimal realization} for $\kappa(\D^{*})$, which is unique up
to isomorphism.
In the same line of the above definitions, if $\sep \subset \br_{0}$ is a finite set, we define its \emph{equisingularity class} $\varepsilon(\sep)$
as the family of all sequences of blow-ups $\pi: (\tilde{M}, \D) \to (\C^{2},0)$ such
that $\pi$ desingularizes $\sep$. By this we mean that the strict transforms $\pi^{*}B$, for $B\in \sep$, are disjoint and
transversal to $\D$. Each such a sequence $\pi$ is a \emph{realization} of $\varepsilon(\sep)$, and we can also talk about
\emph{minimal realizations}, all of them isomorphic.
\subsection{Dicritical structures}
Let $\pi:(\tilde{M},\D) \to (\C^{2},0)$ be a composition of quadratic blow-ups as above. We write
$\pi = \sigma_{1} \circ \cdots \circ \sigma_{n}$ its factorization into individual blow-ups
$\sigma_{j} = (\tilde{M}_{j}, \D_{j}) \to (\tilde{M}_{j-1}, \D_{j-1})$, for $j=1,\ldots,n$, where $\D_{j} = (\sigma_{1} \circ \cdots \circ \sigma_{j})^{-1}(0)$, $\sigma_{j}$ is a punctual blow-up at $q_{j-1} \in \D_{j-1}$, with the convention that $q_{0} = 0$, $(\tilde{M}_{0},\D_{0}) = (\C^{2},0)$ and
$(\tilde{M}_{n},\D_{n}) = (\tilde{M},\D)$.
All objects and invariants pertaining to $\tilde{M}_{j}$ are said to be at \emph{level} $j$, with $j=0$ and $j=n$ being, respectively, the \emph{initial} and \emph{final levels}.
For each $j=1,\ldots,n$, write the factorization $\pi = \pi_{j} \circ \varsigma_{j} $ into
maps
$ \varsigma_{j}: (\tilde{M}_{n}, \D_{n}) \to (\tilde{M}_{j}, \D_{j})$
and
$\pi_{j}: (\tilde{M}_{j}, \D_{j}) \to (\tilde{M}_{0}, \D_{0})$,
where
$\varsigma_{j} = \sigma_{j+1} \circ \cdots \circ \sigma_{n}$ and
$\pi_{j} = \sigma_{1} \circ \cdots \circ \sigma_{j}$. This is depicted in the diagram:
\begin{equation}
\label{eq-sequence-blowups}
\rlap{$\underbrace{\phantom{ (\tilde{M}_{n}, \D_{n}) \stackrel{\sigma_{n}}{\longrightarrow} \cdots \stackrel{\sigma_{j+1}}{\longrightarrow}
(\tilde{M}_{j} , \D_{j})}}_{\varsigma_{j}}$}
(\tilde{M}_{n}, \D_{n}) \stackrel{\sigma_{n}}{\longrightarrow} \cdots \stackrel{\sigma_{j+1}}{\longrightarrow}
\overbrace{ (\tilde{M}_{j} , \D_{j})
\stackrel{\sigma_{j}}{\longrightarrow} (\tilde{M}_{j-1}, \D_{j-1})
\stackrel{\sigma_{j-1}}{\longrightarrow} \cdots \stackrel{\sigma_{1}}{\longrightarrow} (\tilde{M}_{0}, \D_{0})
}^{\pi_{j}} \simeq (\C^{2},0).
\end{equation}
\begin{ddef}
\label{def-dic-structure}
{\rm
A \emph{dicritical structure} at $(\C^{2},0)$ with underlying sequence of blow-ups
$\pi:(\tilde{M},\D) \to (\C^{2},0)$ is the set of data $\Delta$ given, for each $j=1,\cdots,n$, by a
decomposition $\D_{j} = \dic_{j} \cup \ndic_{j}$, where $\dic_{j}$ and $\ndic_{j}$
are unions of components of the exceptional divisor $\D_{j} = \pi_{j}^{-1}(0)$ satisfying:
\begin{enumerate}[label=(D.\arabic*)]
\item $\dic_{j}$ and $\ndic_{j}$ have no common components for $j=1,\cdots, n$;
\item $D$ is in $\dic_{j-1}$ if and only if the strict transform $\sigma_{j}^{*}D$ is in $\dic_{j}$,
for $j=2,\cdots, n$;
\item two components in $\dic_{n}$ do not intersect.
\end{enumerate}
}\end{ddef}
Components in $\dic_{j}$ are called \emph{dicritical} or \emph{non-invariant},
whereas the ones in $\ndic_{j}$ are called \emph{non-dicritical} or \emph{invariant}.
If $D = \sigma_{j}^{-1}(q_{j-1}) \subset \dic_{j}$, we will also call the blow-up $\sigma_{j}$ dicritical, the same happening in the non-dicritical case.
We will say that $n$ is the height of $\Delta$ and we denote $n = h(\Delta)$.
Note that, if we choose a point $p \in \D_{j}$, for $j=1,\ldots,n$, the dicritical structure $\Delta$ can be localized at $p$, by considering the decomposition into dicritical and non-dicritical components induced
by the sequence of blow-ups $\varsigma_{j}$ over $p$. We denote this localized dicritical structure by $\Delta_{p}$.
Evidently, if $\varsigma_{j}$ is trivial over $p$, then this dicritical structure is also trivial.
Let $\Delta$ and $\Delta'$ be two dicritical structures with underlying sequences of blow-ups
$\pi = \sigma_{1} \circ \cdots \circ \sigma_{n}$ and
$\pi' = \sigma_{1}' \circ \cdots \circ \sigma_{n'}'$.
We say that $\Delta'$ \emph{dominates} $\Delta$, and we denote $\Delta' \geq \Delta$, if:
\begin{itemize}
\item
$n' = h(\Delta') \geq h(\Delta) = n$ and $\sigma_{j}' = \sigma_{j}$ for $j=1,\ldots,n$;
\item the decomponsition $\D_{j} = \dic_{j} \cup \ndic_{j}$ is the same for $\Delta$ and $\Delta'$, for $j=1,\ldots,n$;
\item if $n'>n$, then $\varsigma_{n}' = \sigma_{n+1}' \circ \cdots \circ \sigma_{n'}'$ is a composition of non-dicritical blow-ups which is non-trivial
only over finitely many trace points of components in $\dic_{n}$.
\end{itemize}
\begin{example}
\label{ex-foliation-structure}
{\rm
Let $\F$ be a germ of singular holomorphic foliation at $(\C^{2},0)$, induced by a germ of holomorphic
$1$-form $\omega$ with isolated singularity at the origin. We take $\pi:(\tilde{M},\D) \to (\C^{2},0)$, a \emph{minimal reduction of singularities} for $\F$
(see \cite{seidenberg1968},
\cite{camacho1984}).
This means that we obtain a strict transform foliation $\tilde{\F} = \pi^{*}\F$ around the
exceptional divisor $\D = \pi^{-1}(0)$ with the following properties:
\begin{itemize}
\item $\tilde{\F}$ has a finite number of
singularities over $\D$, which are all \emph{simple}, meaning that, at each of them,
$\tilde{\F}$ is defined by a germ of holomorphic vector field with non-nilpotent linear
part with eigenvalues with ratio outside $\Q_{+}$;
\item $\tilde{\F}$ has no singularities and no points of tangency in the non-invariant components
of $\D$;
\item two non-invariant components of $\D$ do not intersect;
\item $\pi$ is minimal with these properties.
\end{itemize}
If we write the decomposition of $\pi$ as in \eqref{eq-sequence-blowups}, at each step
we have a foliation $\tilde{\F}_{j}$, the strict transform of $\F$ by $\pi_{j}$
(with the convention that $\tilde{\F}_{n} = \tilde{\F}$), which is
a germ of singular holomorphic foliation around $\D_{j}$.
We define a decomposition $\D_{j}= \dic_{j} \cup \ndic_{j}$ by setting
$D \subset \dic_{j}$ if and only if $D$ is non-invariant by $\tilde{\F}_{j}$.
This establishes a dicritical structure that we denote by $\Delta(\F)$ or by $\Delta(\omega)$.
}\end{example}
\subsection{Configuration of separatrices}
Considering a dicritical structure $\Delta$ at $(\C^{2},0)$, as defined in the previous subsection, we
write the decomposition
\begin{equation}
\label{eq-connectedcomponent}
\ndic_{n} = \cup_{k=1}^{\ell} \cl{A}_{n,k}^{\iota}
\end{equation}
into disjoint
topologically connected components.
These components, which appear in the following definition, also play an important role in some
of the arguments in this text.
For the next definition, recall that the \emph{valence} of a component $D \subset \D_{n}$, denoted $\val(D)$, is the number of intersections
of $D$ with other components of $\D_{n}$
\begin{ddef}
\label{def-conf-separatrices}
{\rm Let $\Delta$ be a dicritical structure at $(\C^{2},0)$.
A \emph{configuration of separatrices} framed on
$\Delta$
is the collection of information $\Sigma$ given
by a finite set $\sep_{0} \subset \br_{0}$, having a decomposition into
disjoint subsets $\sep_{0} = \sepdic_{0} \cup \sepndic_{0}$, together with its propagation along
the sequence of blow-ups $\pi$,
$\sep_{j} = \sepdic_{j} \cup \sepndic_{j}$, where $\sepdic_{j} = \{\pi_{j}^{*}S; S \in \sepdic_{0}\}$ and
$\sepndic_{j} = \{\pi_{j}^{*}S; S \in \sepndic_{0}\}$,
satisfying the following conditions:
\begin{enumerate}[label=(S.\arabic*)]
\item $\pi$ is a realization of $\varepsilon(\sep_{0})$ and $S \in \sepdic_{0}$ if and only if $D^{\pi}(S) \subset \dic_{n}$;
\item for every $\cl{A}_{n,k}^{\iota}$, there is $S \in \sepndic_{0}$ such
that $D^{\pi}(S) \subset \cl{A}_{n,k}^{\iota}$;
\item if $D \subset \dic_{n}$ is such that $\val(D) = 1$, then there exists $S \in \sepdic_{0}$
such that $D^{\pi}(S) = D$;
\item $\pi$ is a minimal sequence of blow-ups that simultaneously realizes the infinitesimal class $\kappa(\dic_{n})$ and
the equisingularity class $\varepsilon(\sep_{0})$.
\end{enumerate}
}\end{ddef}
Taking into account Definitions \ref{def-dic-structure} and \ref{def-conf-separatrices},
we say that the pair $\DD = (\Delta,\Sigma)$ is a
\emph{dicritical duplet} at $(\C^{2},0)$.
Branches in $\sepdic_{j}$ and in $\sepndic_{j}$ are called, respectively,
\emph{dicritical} and \emph{isolated separatrices}.
The \emph{support} of $\DD$ is defined, at each level
$j=1,\ldots,n$, as the analytic curve
$\supp_{j} = \sep_{j} \cup \ndic_{j}$, seen as germ around
$\D_{j}$. Define, at level $j$, the singular set of $\DD$ as the singular set of its support,
$\sing(\supp_{j})$.
At the initial level, we have $\supp_{0} = \sep_{0}$.
Note that $\sepdic_{0}$
may be empty. On the other hand, $\sepndic_{0}$ is empty if and only if $h(\Delta) = 1$ and
$\D_{1} = \dic_{1}$. This is the \emph{radial case}, which is trivial for our purposes and
will not be taken as our starting point. Nevertheless, in some inductive constructions,
the radial case may appear at intermediate levels of the dicritical structure,
as a result of
the localization of our objects.
Anyhow, if we wish to include the initial radial case, we should modify condition (S.3) and also ask that, if $D \in \dic_{n}$ is such that
$\val(D) =0$ --- which implies that $n = h(\Delta) =1$ --- then there are at least two separatrices in $\sepdic_{0}$.
Given two dicritical duplets $\DD = (\Delta,\Sigma)$ and $\DD' = (\Delta',\Sigma')$ at $(\C^{2},0)$,
we say that $\DD'$ \emph{dominates} $\DD$, and denote
$\DD' \geq \DD$,
if:
\begin{itemize}
\item $\Delta' \geq \Delta$;
\item $\sep_{0}' \supset \sep_{0}$, where $\sep_{0}' $ denotes the set of separatrices of
$\Sigma'$ at level $0$, and $\sep_{0}' \setminus \sep_{0}$ contains only isolated separatrices for $\Sigma'$;
\item separatrices in $\sep_{0}$ are qualified in the same way, as dicritical or isolated, for both $\Sigma$ and $\Sigma'$;
\item if $S \in \sep_{0}' \setminus \sep_{0}$,
then $D = D^{\pi}(S) \subset \dic_{n}$ and
$\pi^{*}S \cap D$ does not belong to $\supp_{n}$, the support of $\DD$ at level $n$.
\end{itemize}
We make some comments on these conditions.
For $S \in \sep_{0}' \setminus \sep_{0}$, let $q = \pi^{*}S \cap D$, where $D = D^{\pi}(S)$.
Also denote by $\varsigma_{n}' = \sigma_{n+1}' \circ \cdots \circ \sigma_{n'}'$ the blow-ups of $\pi'$ after level $n = h(\Delta)$, where
$\pi'$ is the sequence of blow-ups subjacent to $\Delta'$ and $n' = h(\Delta')$.
Since $S$ is isolated for $\Sigma'$ and $D$ is dicritical for $\Delta'$, we have that
$\varsigma_{n}'$ is non-trivial over $q$.
We call $q$ a \emph{escape point} and $S$ a
\emph{escape separatrix}.
A detailed discussion of these objects will be carried out later in
Section \ref{section-escape}. We also refer the reader to Section 5 of \cite{canocorral2006}.
As we did for a dicritical structure $\Delta$,
we can localize a configuration of separatrices $\Sigma$, framed on $\Delta$, at a point
$p \in \D_{j}$, for some $j=1,\ldots,n$, obtaining a configuration of separatrices $\Sigma_{p}$
framed on the dicritical structure $\Delta_{p}$.
The separatrices at level $0$ for $\Sigma_{p}$ will be precisely the branches in $\br_{p}(\supp_{j})$.
Branches in $\br_{p}(\sep_{j})$ will
inherit their classification as dicritical or isolated. On the other hand, a local component
of $D \subset \ndic_{j}$ at $p$ will be dicritical if and only if $\varsigma_{j}^{*}D$ touches
$\varsigma_{j}^{-1}(p)$ at a component of $\dic_{n}$ contained in $\varsigma_{j}^{-1}(p)$.
In particular, if $\varsigma_{j}$ is trivial over $p$, these local components will be isolated separatrices.
The fact that $\Sigma_{p}$ thus defined is framed on $\Delta_{p}$ is easy to check and will
be left to the reader. We will say that $\DD_{p} = (\Delta_{p},\Sigma_{p})$ is the \emph{localization}
of $\DD = (\Delta,\Sigma)$ at the point $p$.
\begin{example}
\label{ex-foliation-separatrices}
{\rm We revisit Example \ref{ex-foliation-structure}.
As we have seen, there is a dicritical structure $\Delta(\F)$ associated with
the minimal reduction of singularities of a germ of singular holomorphic foliation $\F$ at $(\C^{2},0)$.
Recall that a \emph{separatrix} for $\F$ is an analytic branch at $(\C^{2},0)$ invariant by $\F$.
Denote their set by $\text{Sep}_{0}(\F)$.
As proven in \cite{camacho1982} (or in \cite{camacho1988}), $\text{Sep}_{0}(\F) \neq \emptyset$.
From this point on, we suppose that $\F$ is of \emph{generalized curve} type (see \cite{camacho1984}),
meaning that there are no saddle-node singularities --- i.e. simple singularities with one zero eigenvalue --- in the final models of its reduction of singularities.
For a generalized curve type foliation, a desingularization for $\text{Sep}_{0}(\F)$ is
a reduction of singularities for $\F$ \cite[Th. 2]{camacho1984}. A separatrix $S \in \text{Sep}_{0}(\F)$ is classified as dicritical or isolated,
following the classification of the component $D = D^{\pi}(S)$ in the reduction divisor. Denote their sets by, respectively $\Dic_{0}(\F)$ and $\Iso_{0}(\F)$.
Isolated separatrices are finite in number. However, there are infinitely many dicritical separatrices, provided
there is some dicritical component in $\D_{n}$. In order to define a configuration of separatrices framed
on $\Delta(\F)$ we proceed as follows:
\begin{itemize}
\item $\sepndic_{0}(\F)= \Iso_{0}(\F)$ is the set of all isolated separatrices;
\item $\sepdic_{0}(\F)$ includes finitely many separatrices in $\Dic_{0}(\F)$, with at least
one separatrix attached to each component of valence one in $\dic_{n}$.
\end{itemize}
The configuration of separatrices $\Sigma(\F)$ constructed upon $\sep_{0}(\F) = \sepdic_{0}(\F) \cup \sepndic_{0}(\F)$
is framed on $\Delta(\F)$. Indeed, (S.1) and (S.3) are true by construction, while
(S.2) follows, for instance, from \cite[Prop. 4]{mol2002}. Finally, (S.4) is a consequence of
the minimality property of the reduction of singularities $\pi$ along with the above mentioned fact that the desingularization of the separatrices
and of the foliation are equivalent in the generalized curve type context.
We the have a dicritical duplet by
$\DD(\F) = (\Delta(\F),\Sigma(\F))$, also denoted by $\DD(\omega) = (\Delta(\omega),\Sigma(\omega))$ when $\F$ is defined by the germ of holomorphic $1$-form $\omega$.
This duplet is evidently non-unique provided the reduction of singularities of $\F$ contains a dicritical blow-up.
}\end{example}
\section{Systems of indices and systems of residues}
\label{section-system-indices}
In the preceding section we introduced the object $\DD = (\Delta,\Sigma)$, called
dicritical duplet. It joins two nested blocks of abstract information related to singular holomorphic foliations
at $(\C^{2},0)$: a dicritical configuration $\Delta$, which works as a chart of dicritical components, and a configuration of separatrices $\Sigma$, which is a list of mandatory invariant branches.
In this section,
we will introduce other two abstract blocks, systems of indices and systems of residues, that assemble information corresponding to Camacho-Sad indices and
residues of logarithmic $1$-forms.
These four blocks altogether will mould the construction of logarithmic foliations.
For the following definition, suppose that
the underlying sequence of blow-ups of $\Delta$ is $\pi:(\tilde{M},\D) \to (\C^{2},0)$ and that
$n = h(\Delta)$.
Recall that the support of $\DD$ at level $j$ is
$\supp_{j} = \sep_j \cup \ndic_{j}$,
for $j=0,\ldots,n$.
\begin{ddef}
\label{def-system-indices}
{\rm
A \emph{system of indices} associated with the dicritical duplet $\DD = (\Delta,\Sigma)$,
or $\DD$-system of indices for short, is the set of data $\Upsilon$
obtained by
assigning, for each $j=0,\ldots,n$, to every $p \in \supp_{j}$ and each local branch $S \in \br_{p}(\supp_{j})$, a number $I_{p}(S) \in \C$ in a way that the following conditions are respected:
\begin{enumerate}[label=(I.\arabic*)]
\item Suppose that $\sigma_{j}$ is a blow-up at $q\in \D_{j-1}$,
for some $1\leq j \leq n$. Let $p \in \supp_{j-1}$, $S \in \br_{p}(\supp_{j-1})$, $\tilde{S}= \sigma_{j}^{*}S$ and $\tilde{p} = \D_{j} \cap \tilde{S}$.
Then the following \emph{transformation law} is satisfied:
\[
I_{\tilde{p}}(\tilde{S}) =
\begin{cases} I_{p}(S) - \nu_{p}(S)^{2} & \text{if}\ \ p = q \\
I_{p}(S) & \text{if}\ \ p \neq q .
\end{cases}
\]
\item If $p$ is a regular point of $\supp_{j}$
and $S$ is the only branch of $\supp_{j}$ at $p$, then $I_{p}(S)= 0$ (\emph{regular point condition}).
\item For $1 \leq j \leq n$ and every component $D \subset \ndic_{j}$, the \emph{Camacho-Sad formula} is valid:
\[
\sum_{p \in D} I_{p}(D) = D \cdot D,
\]
where in the right side stands the self-intersection number of $D$ in $\tilde{M}_{j}$, which coincides with $c(D)$,
the first Chern class of the normal bundle of $D$ in $\tilde{M}_{j}$.
\item If $p \in \supp_{n}$ is a singular point and $S_{1}, S_{2}$ are the two smooth branches of
$\supp_{n}$ at $p$ (one of them necessarily in $\ndic_{n}$) then we have
the \emph{simple singularity rule}:
\[ I_{p}(S_{1}), I_{p}(S_{2}) \in \C^{*} \setminus \Q_{+}, \ \ \ \text{and} \ \ \
I_{p}(S_{2})= 1/I_{p}(S_{1}) .
\]
\end{enumerate}
}\end{ddef}
In this situation, we say that $\TT = (\DD,\Upsilon) = (\Delta,\Sigma,\Upsilon)$ is a \emph{dicritical triplet}
at $(\C^{2},0)$.
\begin{example}
\label{ex-foliation-indices}
{\rm We resume Examples \ref{ex-foliation-structure} and \ref{ex-foliation-separatrices}.
Let $\DD(\F) = (\Delta(\F),\Sigma(\F))$ be a dicritical duplet of a germ of singular holomorphic foliation $\F$ at $(\C^{2},0)$ of generalized curve type.
We obtain a system of indices $\Upsilon(\F)$ by setting, for each
$p \in \supp_{j}$ and each $S \in \br_{p}(\supp_{j})$, $I_{p}(S)$ as the \emph{Camacho-Sad index} of $\tilde{\F}_{j} = \pi_{j}^{*} \F$ at $p$
(\cite{camacho1982,linsneto1986,suwa1995} and also \cite{brunella1997,suwa1998}).
Conditions (I.1) to (I.4) are known properties of this index.
In particular, condition (I.3) follows from the theorem asserting that the sum of Camacho-Sad indices along a compact analytic curve invariant by a foliation in
an ambient complex surface is the curve's self intersection.
We say that $\TT(\F) = (\DD(\F),\Upsilon(\F)) = (\Delta(\F),\Sigma(\F),\Upsilon(\F))$ is a dicritical triplet associated with $\F$.
}\end{example}
A system of indices is determined in a unique way by its prescription at the final level,
as shown next.
\begin{prop}
Let $\DD = (\Delta,\Sigma)$ be a dicritical duplet at $(\C^{2},0)$. If $n = h(\Delta)$, then any set of data given by complex numbers $I_{p}(S) \in \C^{*}$, for every
$p \in \supp_{n}$ and every $S \in \br_{p}(\supp_{n})$, satisfying (I.2), (I.3) and (I.4) of Definition \ref{def-system-indices} for $j=n$, extends in a unique way to a
$\DD$-system of indices $\Upsilon$.
\end{prop}
\begin{proof}
The transformation law (I.1) ought to be used in descending the indices assigned
at the final level and, thus, it will be automatically fulfilled.
Besides, since (I.4) concerns only level $n$, it is enough to verify (I.2) and (I.3) for lower levels.
The proof goes by a reverse induction argument. The initialization, at level $n$, works by hypothesis.
Suppose then that $0 \leq j< n$ and that conditions (I.1) to (I.4) are valid at all levels above $j$.
We first prove that (I.2) is true at level $j$.
Let $p$ be a smooth point of $\supp_{j}$.
If $p \in \ndic_{j}$, then $p$ is a trace point of some component $D \subset \ndic_{j}$ and there are no components of $\sep_{j}$ at $p$.
By the minimality condition (S.4), $\varsigma_{j}$ would be non-trivial at $p$ if and
only if $\varsigma_{j}^{-1}(p)$ contained some dicritical component, say $D' \subset \dic_{n}$. But
we would have that either $D'$ disconnects $\ndic_{n}$ or $\val(D') = 1$. In both cases, by (S.2) and (S.3),
we would find some element of $\sep_{j}$ at $p$, which is impossible.
Hence, $\varsigma_{j}$ is a local isomorphism over $p$. Thus, using (I.2) at level $n$ and successively
applying (I.1), we conclude that $I_{p}(D) = 0$.
Suppose now that $p \not\in \ndic_{j}$. Then, $p$ is a trace point of a component $D \subset \dic_{j}$
and there is a unique $S \in \sep_{j}$ at $p$, which is smooth. Suppose first that $S \in \sepdic_{j}$. If $\varsigma_{j}$ were non-trivial over $p$, then, by (D.3), $\varsigma_{j}^{-1}(p)$ would contain some non-dicritical component, which in turn would imply the existence of a least one element of $\sepndic_{j}$ at $p$, which is not allowed.
Thus $\varsigma_{j}$ is trivial over $p$ and, as in the previous paragraph,
we find that $I_{p}(S) = 0$.
On the other hand, if $S \in \sepndic_{j}$, then $S$ is tangent to $D$, $\varsigma_{j}$ is non-trivial over $p$ and,
as a consequence of (S.2), $\varsigma_{j}^{-1}(p)$ contains only non-dicritical components.
By the minimality property (S.4), it is easy to see that
$\varsigma_{j}^{-1}(p)$ is a linear chain of components $D_{1} \cup \cdots \cup D_{r} \subset \ndic_{n}$
such that $D_{i} \cap D_{i+1} \neq \emptyset$ for $i=1,\ldots, r-1$,
$\tilde{S} = \varsigma_{j}^{*}S$ touches $D_{1}$ at a point $\tilde{p}$ and no other components
of $\sepndic_{n}$ touches $\varsigma_{j}^{-1}(p)$. Besides, $D_{1} \cdot D_{1} = -1$ and
$D_{i} \cdot D_{i} = -2$ for $i=2,\ldots,r$. By the application, at level $n$, of (I.3) and (I.4) along this linear chain,
we find that $I_{\tilde{p}}(\tilde{S}) = -r$. Finally, the successive application of (I.1) gives
that $I_{p}(S)= 0$. This proves (I.2) at level $j$.
In order to prove (I.3) at level $j$, suppose
that $D \subset \ndic_{j}$ is an irreducible component.
Suppose that $\sigma_{j+1}$ is a blow-up at $p \in \D_{j}$.
If $p \not\in D$,
the Camacho-Sad formula (I.3) will evidently hold for $D$, since it holds for
$\tilde{D} = \sigma_{j+1}^{*}D$ and
all invariants involved are preserved by $\sigma_{j+1}$.
Now, in the case $p \in D$, we have, on the one hand,
$\tilde{D} \cdot \tilde{D} = D \cdot D - 1$.
On the other hand,
if $\tilde{p} = \tilde{D} \cap \sigma_{j+1}^{-1}(p)$, then the transformation law (I.1) gives
$I_{\tilde{p}}(\tilde{D}) = I_{p}(D)-1$. The Camacho-Sad formula then holds for $D$, since the
indices of $D$ at points other than $p$ are preserved by $\sigma_{j+1}$. This proves (I.3) at
level $j$, concluding the proof of the proposition.
\end{proof}
Recall that a meromorphic $1$-form $\eta$
--- and also the germ of singular holomorphic foliation $\LL$ that it
defines ---
is \emph{logarithmic} if it can be written as
\begin{equation}
\label{eq-log-form}
\eta = \sum_{S \in \sep} \lambda_{S} \frac{d f_{S}}{f_{S}} + \alpha ,
\end{equation}
where $\alpha$ is some closed germ of holomorphic $1$-form,
$\sep \subset \br_{0}$ is a finite set and, for every $S \in \sep$,
$f_{S} \in \mathcal{O}_{0}$ is a local irreducible equation for $S$
and $\lambda_{S} \in \C^{*}$. Note that, since $\eta$ is closed, all branches in $\sep$ are separatrices of the foliation $\LL$.
We define a dicritical duplet $\DD(\eta) = (\Delta(\eta),\sigma(\eta))$ intrinsically associated with the
logarithmic $1$-form $\eta$. The dicritical structure is
$\Delta(\eta) = \Delta(\LL)$, the one defined by the reduction of singularities of the foliation $\LL$, as in Example \ref{ex-foliation-structure}.
As for the configuration of separatrices,
in $\Sigma(\eta)$ we include all branches in $\sep$ and also all isolated separatrices of $\LL$.
For $\Sigma(\eta)$ to be a configuration of separatrices, it remains to assure that condition (S.3) is true, in other words, that
there is $S \in \sep$ such that $D^{\pi}(S) = D$ for every
component $D \subset \dic_{n}$ with $\val(D) = 1$, where $n = h(\Delta(\eta))$ and
$\pi$ is the sequence of blow-ups beneath $\Delta(\eta)$.
Nevertheless, this happens in the cases considered in this text, more specifically, if we assume that $\eta$ is a faithful logarithmic $1$-form (see Definition \ref{def-faithful} and Section \ref{section-escape}).
Thus, henceforth we can consider that $\DD(\eta) = (\Delta(\eta),\Sigma(\eta))$ is a dicritical duplet.
We obtain a dicritical triplet $\TT(\eta) = (\DD(\eta), \Upsilon(\eta)) = (\Delta(\eta),\sigma(\eta),\Upsilon(\eta))$ by prescribing as indices those provided
by the Camacho-Sad indices of $\LL$.
In an attempt to handle residues of logarithmic $1$-forms, we add another element to our construction:
\begin{ddef}
\label{def-system-residues}
{\rm Let $\DD = (\Delta,\Sigma)$ be a dicritical duplet at $(\C^{2},0)$.
A \emph{system of residues} for $\DD = (\Delta,\Sigma)$, or simply
a \emph{$\DD$-system of residues}, is the set of data $\Lambda$
obtained by assigning
numbers $\lambda_{S} \in \C^{*}$ to each irreducible component $S \subset \supp_{j}$ ($S$ is either a
component of $\ndic_{j}$ or a germ of separatrix of $\sep_{j}$), for $j=0,\ldots,n$,
obeying the following rules:
\begin{enumerate}[label=(R.\arabic*)]
\item
Let $S \subset \supp_{j}$ and $\sigma_{j+1}$ be the factor of $\pi$ corresponding to a blow-up at a point
$q \in \supp_{j}$, for some $j<n$. If $\tilde{S} = \sigma_{j+1}^{*}S$, then $\lambda_{\tilde{S}} = \lambda_{S}$
(\emph{invariance under blow-ups}).
\item If $\sigma_{j+1}$ is a non-dicritical blow-up at $q \in \supp_{j}$, for some $0 \leq j < n$,
and $D = \sigma_{j+1}^{-1}(q)$, then
\[\lambda_{D} = \sum_{S \subset \supp_{j}} \nu_{q}(S) \lambda_{S} \neq 0 .\]
\item If $\sigma_{j+1}$ is a dicritical blow-up at $q \in \supp_{j}$, for some $0 \leq j < n$, then
\[\sum_{S \subset \supp_{j}} \nu_{q}(S) \lambda_{S} = 0 \ \ \ \text{(\emph{main resonance})}.\]
\end{enumerate}
}\end{ddef}
The sum in items (R.2) and (R.3) has finitely many non-zero terms, only those corresponding to components $S$
containing $q$.
We call it \emph{weighted balance of residues}.
It is worth pointing out that the assignment of residues at level $0$
determines, by the successive application of rules (R.1) and (R.2),
residues for all components of the support at all levels. Evidently, rules (R.2) and (R.3)
impose compatibility conditions on a set of residues at level 0 that actually engender a $\DD$-system of residues.
Two $\DD$-systems of residues $\Lambda$ and $\Lambda'$
are \emph{equivalent} if there exists a common ratio $\alpha \in \C^{*}$
for corresponding residues. We denote, in this case, $\Lambda = \alpha \Lambda'$.
If $\Lambda$ is a $\DD$-system of residues, then the principal part of the logarithmic $1$-form shown in equation \eqref{eq-log-form}
for $\sep = \sep_{0}$ is denoted by $\eta_{\Lambda}$. If $\Lambda$ and $\Lambda'$ are equivalent $\DD$-systems of residues,
with a proportionality ratio $\alpha \in \C^{*}$ as above, then $\eta_{\Lambda} = \alpha \, \eta_{\Lambda'}$, the two $1$-forms defining
the same germ of singular holomorphic foliation at $(\C^{2},0)$.
We establish a bond between
systems of indices and
systems of residues with the following:
\begin{ddef}
\label{def-consistent}
{\rm Let $\Upsilon$ and $\Lambda$ be
$\DD$-systems of indices and residues,
where
$\DD = (\Delta,\Sigma)$ is a dicritical duplet at $(\C^{2},0)$ of height $n = h(\Delta)$.
We say that $\Upsilon$ and $\Lambda$ are \emph{consistent}
if, for every $p \in \sing(\supp_{n})$,
\[ I_{p}(S_{1}) = - \frac{\lambda_{S_{2}}}{\lambda_{S_{1}}} \qquad \text{and} \qquad
I_{p}(S_{2}) = - \frac{\lambda_{S_{1}}}{\lambda_{S_{2}}} ,\]
where $S_{1}, S_{2}$ are the two local components
of $\br_{p}(\supp_{n})$.
In this situation, we say that $\QQ = (\Delta,\Sigma,\Upsilon,\Lambda)$ forms a \emph{dicritical quadruplet}.
}\end{ddef}
Note that, by (I.4), the two equalities in the definition are equivalent. Note also that there are no
conditions on the separatrices in
$\sepdic_{n}$, and this gives some degree of freedom in their assignment of residues.
This remark will be particularly useful in Section \ref{section-escape}.
Another important point is that, if $\Lambda$ and $\Lambda'$ are equivalent
$\DD$-systems of residues, then $\Lambda$ is consistent with a $\DD$-system of indices $\Upsilon$ if and only if $\Lambda'$ is.
It is apparent that, given $p \in \D_{j}$, both a system of indices $\Upsilon$ and a system of
residues $\Lambda$ can be localized at $p$, keeping their consistency relation, if it is the case.
Thus, denoting these localizations by $\Upsilon_{p}$ and $\Lambda_{p}$, we can
consider the localizations $\TT_{p}$ and $\QQ_{p}$ at $p$ of, respectively, a dicritical
triplet $\TT = (\Delta,\Sigma, \Upsilon)$ and
a dicritical
quadruplet $\QQ = (\Delta,\Sigma, \Upsilon,\Lambda)$.
It is always possible to assign residues at the final level
in a consistent way with a system of indices. More precisely:
\begin{prop}
\label{prop-residues-leveln}
Suppose that $\TT = (\Delta,\Sigma,\Upsilon)$ is a dicritical triplet at $(\C^{2},0)$. Then, it is possible to assign
numbers $\lambda_{S} \in \C^{*}$ for the irreducible components $S \subset \supp_{n}$,
where $n = h(\Delta)$, in such a way that the consistency condition of Definition
\ref{def-consistent} is satisfied.
\end{prop}
\begin{proof}
We can work separately in each
topologically connected component defined in
\eqref{eq-connectedcomponent} and we fix some $\A=\A_{n,k}^{\iota}$.
Our task is
to assign residues to each component of $\supp_{n}$ intersecting $\A$.
To that end, after fixing an arbitrary component
$D \subset \A$, define a partial order in
the components of $\A$ in the following inductive way:
$D$ is the maximal element, its immediate predecessors are the components $D'$ such that
$D \cap D' \neq \emptyset$ and so forth.
This is possible since the dual graph of $\A$ is a tree
(see the paragraph preceding Proposition \ref{prop-rational-index}).
Choose an arbitrary value $\lambda_{D} \in \C^{*}$.
Then, by the consistency condition in Definition \ref{def-consistent},
$\lambda_{D}$ determines all residues $\lambda_{S}$ for components $S \subset \supp_{n}$ intersecting
$D$: we must set
$\lambda_{S} = -\lambda_{D} I_{p}(D)$, where $p = S \cap D$. In particular, we find all residues $\lambda_{D'}$
for components $D'$ preceding $D$. Again, for each such $D'$,
all residues of components of $\supp_{n}$ intersecting $D'$ are calculated by the consistency condition.
This clearly provides an inductive process that will end once we reach the minimal elements with respect to this order.
\end{proof}
Proposition \ref{prop-residues-leveln} will provide the initial step, in a reverse induction process,
of the construction of a $\DD$-system of residues $\Lambda$ consistent with a $\DD$-system of indices $\Upsilon$.
The general step of the induction is the \emph{amalgamation procedure}, described in the next section.
For it,
we need
the proposition below.
All elements of its proof are in \cite[Prop. 3]{corral2012} and \cite[Lem. 15]{canocorral2006}.
We include a proof here adapted to our setting.
\begin{prop}
\label{prop-residues-indicies}
Let $\QQ= (\Delta,\Sigma,\Upsilon,\Lambda)$ be a dicritical quadruplet at $(\C^{2},0)$.
If $S \in \br_{0}(\sep_{0})$,
then
\[ \lambda_{S} I_{0}(S) = - \sum_{S' \in \br_{0}(\sep_{0}) \setminus S} \lambda_{S'} (S',S)_{0},\]
where $( \, \cdot \, , \cdot \, )_{0}$ denotes the
intersection number at $0 \in \C^{2}$.
\end{prop}
\begin{proof}
We will use the following induction assertion at level $j$, for $0 \leq j \leq n$:
if $p \in \supp_{j}$ and $S$ is a local component of $\supp_{j}$ at $p$, then
\[ \lambda_{S} I_{p}(S) = - \sum_{S' \in \br_{p}(\supp_{j}) \setminus S} \lambda_{S'} (S',S)_{p}.\]
We are evidently employing here data from the localized quadruplet $\QQ_{p} = (\Delta_{p},\Sigma_{p},\Upsilon_{p},\Lambda_{p})$.
At level $n$, the assertion is true: it is trivial for each regular point of $\supp_{n}$, and, for singular points, it is the very
definition of consistency. Let $1 \leq j \leq n$ and
suppose that the result is true, as stated, for all points in $\supp_{j}$.
We will prove that it is also true for points at level $j-1$. Suppose that $p \in \supp_{j-1}$. If the blow-up $\sigma_{j}$
is trivial over $p$, there is nothing to prove and,
thus, we are reduced to the case where $\sigma_{j}$ is a blow-up at $p$.
Let us denote $\tilde{S}$ and
$\tilde{S'}$ the strict transforms by $\sigma_{j}$ of the components $S$ and $S'$, $D = \sigma_{j}^{-1}(p)$
and
$\tilde{p}= \tilde{S} \cap D$. We have
\begin{eqnarray*}
\sum_{S' \in \br_{p}(\supp_{j-1}) \setminus S} \lambda_{S'} (S',S)_{p} & = & \sum_{S' \in \br_{p}(\supp_{j-1}) \setminus S} \lambda_{S'} \left(\nu_{p}(S') \nu_{p}(S)
+ (\tilde{S}',\tilde{S})_{\tilde{p}} \right) \\
& = & \nu_{p}(S) \sum_{S' \in \br_{p}(\supp_{j-1}) \setminus S} \lambda_{S'} \nu_{p}(S') +
\sum_{S' \in \br_{p}(\supp_{j-1}) \setminus S} \lambda_{S'} (\tilde{S}',\tilde{S})_{\tilde{p}} \\
& = & -\lambda_{S} \nu_{p}(S)^{2} +
\nu_{p}(S) \underbrace{ \sum_{S' \in \br_{p}(\supp_{j-1})} \lambda_{S'} \nu_{p}(S')}_{\displaystyle(*)} +
\underbrace{\sum_{S' \in \br_{p}(\supp_{j-1}) \setminus S} \lambda_{\tilde{S}'} (\tilde{S}',\tilde{S})_{\tilde{p}}}_{\displaystyle(**)}
\end{eqnarray*}
If $\sigma_{j}$ is a dicritical blow-up, then the summation $(*)$ vanishes by (R.3),
whereas $(**)$ can be calculated using the
the induction assertion at level $j$ for $\tilde{p}$,
yielding, by finally applying (R.1) and (I.1),
\[
\sum_{S' \in \br_{p}(\supp_{j-1}) \setminus S} \lambda_{S'} (S',S)_{p} =
-\lambda_{S} \nu_{p}(S)^{2} - \lambda_{\tilde{S}} I_{\tilde{p}}(\tilde{S})
= -\lambda_{S}( \nu_{p}(S)^{2} + I_{\tilde{p}}(\tilde{S}) ) = -\lambda_{S} I_{p}(S),
\]
proving thereby the result in this case.
On the other hand, if $\sigma_{j}$ is a non-dicritical blow-up, the summation $(*)$ equals
$\lambda_{D}$ by (R.2) and, by applying the induction assertion for $\tilde{p}$
followed by (R.1) and (I.1), we get
\begin{eqnarray*}
\sum_{S' \in \br_{p}(\supp_{j-1}) \setminus S} \lambda_{S'} (S',S)_{p} & = &
-\lambda_{S} \nu_{p}(S)^{2} +
\lambda_{D}(D,\tilde{S})_{\tilde{p}} +
\sum_{S' \in \br_{p}(\supp_{j-1}) \setminus S} \lambda_{\tilde{S}'} (\tilde{S}',\tilde{S})_{\tilde{p}} \\
& = &
-\lambda_{S} \nu_{p}(S)^{2} - \lambda_{\tilde{S}} I_{\tilde{p}}(\tilde{S})
= -\lambda_{S}( \nu_{p}(S)^{2} + I_{\tilde{p}}(\tilde{S}) ) = -\lambda_{S} I_{p}(S).
\end{eqnarray*}
This concludes the proof of the proposition.
\end{proof}
\section{Existence of dicritical quadruplets and logarithmic $1$-forms}
\label{section-existence-quadruplet}
We start by recalling a definition from \cite{canocorral2006} that will be crucial to the development of the results of this section. Let $\eta$ be a germ of logarithmic $1$-form at $(\C^{2},0)$,
written as
\begin{equation}
\label{eq-log-form1}
\eta = \lambda_{1} \frac{d f_{1}}{f_{1}} + \cdots + \lambda_{r} \frac{d f_{r}}{f_{r}} + \alpha,
\end{equation}
where $\alpha$ is a germ of closed holomorphic $1$-form, $r \geq 2$, and, for $i=1,\ldots,r$, $f_{i} \in \cl{O}_{0}$ is irreducible with order $\nu_{i} = \nu_{0}(f_{i})$
and
$\lambda_{i} \in \C^{*}$. Denote by $\LL$ the singular holomorphic foliation defined by $\eta$.
\begin{ddef}
\label{def-faithful}
{\rm The $1$-form
$\eta$ is \emph{$1$-faithful} if
$$\nu_{0}( f\eta) = m-1,$$
where $f = f_{1} \cdots f_{r}$ and $m = \nu_{0}(f) = \nu_{1}+ \cdots + \nu_{r}$.
We plainly say that $\eta$ is \emph{faithful} if its strict transforms are $1$-faithful along the reduction of singularities of $\LL$.
}\end{ddef}
In the definition, the closed holomorphic $1$-form $\alpha$ in \eqref{eq-log-form1} is inert, since
$\nu_{0}( f\alpha) \geq m$. Also, the definition does not depend on the choice of irreducible equations
$f_{1}, \ldots, f_{r}$ for the components of the polar set of $\eta$. This gives, in particular, the following remark, that will
be used a couple of times in the proof of Theorem \ref{teo-quadruplet} below:
if $\eta$ and $\eta'$ are germs of logarithmic $1$-forms at $(\C^{2},0)$ differing from each other
by a germ of closed holomorphic $1$-form, then $\eta$ is faithful
if and only if $\eta'$ is. In Section \ref{section-escape} we make a brief discussion on the
notion of faithfulness. Notably, the following result, from \cite[Lem. 7]{canocorral2006}, will be
explained: if
$\lambda_{0} = \sum_{i=1}^{r} \nu_{i} \lambda_{i}$ is the weighted balance of residues of $\eta$,
then
\begin{itemize}
\item $\lambda_{0} \neq 0$ implies that $\eta$ is non-dicritical at $0 \in \C^{2}$ and $1$-faithful;
\item $\lambda_{0} = 0$ and $\eta$ $1$-faithful implies that $\eta$ is dicritical $0 \in \C^{2}$ (\emph{main resonant} case).
\end{itemize}
In other words, for a faithful logarithmic $1$-form, the dicritical or non-dicritical character of
a blow-up in its reduction of singularities is determined by, respectively, the vanishing or non-vanishing
of the weighted balance of residues.
Another important point is that, if $\eta$ is faithful logarithmic and
$\pi :(\tilde{M} ,\D) \to (\C^{2},0)$ is its reduction
of singularities, then all non-dicritical components of $\D$ are poles of $\pi^{*} \eta$, while the dicritical components are neither poles nor contained in the
its set of zeroes. This can be seen below, in Section \ref{section-escape}, from formula \eqref{eq-sigma*eta} and by the discussion following formula \eqref{eq-tilde-omega}.
Then, the existence of at least one separatrix attached to each dicritical component $D \subset \D$ of valence one, a necessary condition in order that $\DD(\eta)$ be a dicritical duplet, follows from the Residue Theorem applied to the restriction of $\pi^{*} \eta$ to $D$.
Bearing in mind these observations, we make the following definition:
\begin{ddef}
\label{def-Q-logarithmic}
{\rm Let $\QQ =(\Delta,\Sigma,\Upsilon,\Lambda)$ be a dicritical quadruplet at $(\C^{2},0)$.
We say that a germ of closed meromorphic $1$-form $\eta$ is \emph{$\QQ$-logarithmic} if:
\begin{enumerate}
\item there is a writing
\[ \eta = \sum_{S \in \sep_{0}} \lambda_{S} \frac{d f_{S}}{f_{S}} + \alpha ,\]
where $\sep_{0}$ is the level $0$ of $\Sigma$, $f_{S} \in \mathcal{O}_{0}$ are local equations for the components
$S \in \sep_{0}$, $\lambda_{S}$ are the residues assigned by $\Lambda$ and
$\alpha$ is a germ of closed holomorphic $1$-form;
\item $\DD(\eta) \geq \DD$, where $\DD(\eta) = (\Delta(\eta),\Sigma(\eta))$ and $\DD = (\Delta,\Sigma)$;
\item $\eta$ is faithful.
\end{enumerate}
We say that $\eta$ is \emph{strictly $\QQ$-logarithmic} if $\DD(\eta) = \DD$.
}\end{ddef}
Above, we also say that $\eta$ is a \emph{logarithmic model} for $\QQ$.
Definition \ref{def-Q-logarithmic} deserves a few comments.
First, the poles of $\eta$ are prescribed by $\sep_{0}$, the level $0$ of the configuration of separatrices $\Sigma$, and their
residues are those defined by the system of residues $\Lambda$.
Second, the dominance condition $\DD(\eta) \geq \DD$ says, aside from the fact that the reduction of singularities
of the foliation $\LL$ is longer than the underlying sequence of blow-ups of $\Delta$, that the separatrices of $\LL$ that
are not poles of $\eta$ are escape separatrices with respect to $\DD =(\Delta,\Sigma)$.
Finally, the axiomatics defining systems of residues and systems of indices, along with the notion of consistency in
Definition \ref{def-consistent}, gives that the Camacho-Sad indices of $\LL$ with respect to the separatrices in $\sep_{0}$
are precisely those given by $\Upsilon$.
We should draw attention to the fact that asking $\eta$ to be faithful is stronger than asking that common dicritical and non-dicritical components of $\Delta(\eta)$ and $\Delta$
correspond, which is part of the information given by condition (2).
Actually, we are demanding that whenever the weighted balance of residues $\lambda_{0}$ vanishes, the corresponding blow-up is dicritical, but,
additionally, that it is dicritical in a ``1-faithful'' way.
The following theorem is a cornerstone
to all subsequent results of this article.
The remainder of this section is devoted to its proof.
\begin{maintheorem}
\label{teo-quadruplet}
Let $\TT =(\Delta,\Sigma,\Upsilon)$ be a dicritical triplet. Then there exists:
\begin{enumerate}
\item a system of
residues $\Lambda$ such that $\QQ = (\TT,\Lambda) = (\Delta,\Sigma,\Upsilon,\Lambda)$ is a dicritical quadruplet;
\item a $\QQ$-logarithmic $1$-form $\eta$.
\end{enumerate}
\end{maintheorem}
\begin{proof}
The two assertions above are proved simultaneously. We essentially adapt the arguments in \cite{canocorral2006}
to the combinatorics involved in the definition of a dicritical quadruplet.
The idea is to descend the dicritical structure, starting at the final level $n = h(\Delta)$.
For a fixed level $j$, we will furnish the following objects:
\begin{enumerate}[label=(\roman*)]
\item a \emph{partial system of residues} $\Lambda^{(j)}$, that is to say,
residue data for all components of the support for levels $j$ and higher,
satisfying (R.1), (R.2) and (R.3),
consistent with the indices of $\Upsilon$ at level $n$.
\item for each $p \in \supp_{j}$, a $\QQ^{(j)}_{p}$-logarithmic $1$-form $\eta^{(j)}_{p}$, where
$\QQ^{(j)}_{p} = (\TT_{p},\Lambda^{(j)}_{p})$ is the dicritical quadruplet at $(\tilde{M}_{j},p)$ defined by the localizations
of $\TT$ and of $\Lambda^{(j)}$ at $p$.
\end{enumerate}
At level $n$, residues of $\Lambda^{(n)}$ are provided by Proposition \ref{prop-residues-leveln}.
As for the $\QQ^{(n)}_{p}$-logarithmic $1$-forms,
at a singular point $p \in \supp_{n}$, we take local coordinates $(x,y)$ at $p$ such that
$x=0$ and $y=0$ are equations of the two components of $\supp_{n}$ at $p$, having residues $\lambda_{1}$ and $\lambda_{2}$, and
then we set $\eta^{(n)}_{p} = \lambda_{1} dx/x + \lambda_{2} dy/y$.
At a regular point $p \in \supp_{n}$, in local coordinates $(x,y)$ at $p$ such that
the component of $\supp_{n}$ at $p$ has equation $x=0$, we set $\eta^{(n)}_{p} = \lambda dx/x$,
where $\lambda \in \C^{*}$ is the corresponding residue.
Suppose that, for a fixed $j$ with $1 \leq j \leq n$, residues and logarithmic 1-forms have been provided as described
in items (i) and (ii) above. We will amalgamate data from level $j$ in order to construct, at level $j-1$, a
partial system of residues $\Lambda^{(j-1)}$, as well as a $\QQ^{(j-1)}_{p}$-logarithmic $1$-form
for every $p \in \supp_{j-1}$, where $\QQ^{(j-1)}_{p} = (\TT_{p},\Lambda^{(j-1)}_{p})$.
Actually, we only need to do this for the point
$p \in \supp_{j-1}$ that is the center of the blow-up $\sigma_{j}$.
Denote by $D = \sigma_{j}^{-1}(p)$ and let $p_{1},\ldots,p_{\ell}$ be
the points of intersection of $D$ with $\overline{\supp_{j} \setminus D}$.
In our inductive construction, in addition to the partial system of residues $\Lambda^{(j)}$,
for each $k=1,\ldots,\ell$, we have
a germ of $\QQ^{(j)}_{p_{k}}$-logarithmic
$1$-form $\eta_{p_{k}}^{(j)}$ at $p_{k}$, where
$\QQ^{(j)}_{p_{k}} = (\TT_{p_{k}},\Lambda^{(j)}_{p_{k}})$.
The non-dicritical and dicritical cases are then treated separately.
\smallskip \smallskip
\par \noindent
\underline{Case 1}: $\sigma_{j}$ is a non-dicritical blow-up.
The amalgamation of data from level $j$ is quite instantaneous here and will be a consequence
of Proposition \ref{prop-residues-indicies}.
Indeed, for each $p_{k} \in D$, we have
\begin{equation}
\label{eq-lambdaD}
\lambda_{D} I_{p_{k}}(D) = - \sum_{\tilde{S} \in \br_{p_{k}}(\supp_{j}) \setminus D} \lambda_{\tilde{S}} (\tilde{S},D)_{p_{k}},
\end{equation}
where
residues are those of $\Lambda^{(j)}$.
Applying the Camacho-Sad formula (I.3) followed by equation \eqref{eq-lambdaD}, we get
\begin{eqnarray}
\label{eq-amalgamation1}
\lambda_{D} &=& \lambda_{D}\left(- \sum_{k=1}^{\ell} I_{p_{k}}(D)\right) \ = \ \sum_{k=1}^{\ell}(- \lambda_{D} I_{p_{k}}(D)) \\
\nonumber &=&
\sum_{k=1}^{\ell} \left( \sum_{\tilde{S} \in \br_{p_{k}}(\supp_{j}) \setminus D} \lambda_{\tilde{S}} (\tilde{S},D)_{p_{k}} \right) \\
\nonumber &=& \sum_{S \in \br_{p}(\supp_{j-1})} \nu_{p}(S) \lambda_{\tilde{S}},
\end{eqnarray}
where $\tilde{S} = \sigma_{j}^{*}S$ is the strict transform of $S \in \br_{p}(\supp_{j-1})$, which
belongs to some $\br_{p_{k}}(\supp_{j})$.
Equation \eqref{eq-amalgamation1} allows us to
define $\lambda_{S} = \lambda_{\tilde{S}}$, in such a way that condition (R.2) is satisfied by the blow-up
$\sigma_{j}$.
Therefore, we can define residues at level $j-1$ by simply importing then from level $j$:
for every component $S \subset \supp_{j-1}$, set $\lambda_{S} = \lambda_{\tilde{S}}$, where $\tilde{S} = \sigma_{j}^{*}S$.
In this way, we have a partial system of residues $\Lambda^{(j-1)}$.
In our downwards construction, write the
$\QQ^{(j)}_{p_{k}}$-logarithmic
$1$-form at $p_{k} \in D = \sigma_{j}^{-1}(p)$ as
\[ \eta_{p_{k}}^{(j)} = \lambda_{D} \frac{d f_{D}}{f_{D}} +
\sum_{\tilde{S} \in \br_{p_{k}}(\supp_{j}) \setminus D} \lambda_{\tilde{S}} \frac{d f_{\tilde{S}}}{ f_{\tilde{S}}} + \alpha_{p_{k}},\]
where
$f_{\tilde{S}} \in \cl{O}_{p_{k}}$ are local equations for $\tilde{S} \in \br_{p_{k}}(\supp_{j}) \setminus D$,
$f_{D} \in \cl{O}_{p_{k}}$ is a local equation for $D$ and $\alpha_{p_{k}}$ is a germ of closed holomorphic $1$-form at $p_{k}$.
For each $S \in \br_{p}(\supp_{j-1})$, take $f_{S} \in \cl{O}_{p}$ an equation of $S$ such
that $f_{S} \circ \sigma_{j} = f_{D}^{\nu_{p}(S)} f_{\tilde{S}}$, where $\tilde{S} = \sigma_{j}^{*}S$
and $f_{D}$ is the local equation of $D$ at $p_{k} = \tilde{S} \cap D$.
We then define
\[ \eta_{p}^{(j-1)} =
\sum_{S \in \br_{p}(\supp_{j-1})} \lambda_{S} \frac{d f_{S}}{f_{S}}.\]
Setting $\QQ_{p}^{(j-1)} = (\TT_{p}, \Lambda^{(j-1)}_{p})$, then $\eta_{p}^{(j-1)}$ is $\QQ_{p}^{(j-1)}$-logarithmic.
Indeed, by equation \eqref{eq-amalgamation1},
the residue of $ \sigma_{j}^{*} \eta_{p}^{(j-1)} $ with respect to
$D = \sigma_{j}^{-1}(p)$
is
precisely $\lambda_{D} \neq 0$, which simultaneously implies that $\eta_{p}^{(j-1)}$ is $1$-faithful and
that $D$ is invariant by $ \sigma_{j}^{*} \eta_{p}^{(j-1)}$ \cite[Lem. 7]{canocorral2006}. On the other hand, faithfulness of $\eta_{p}^{(j-1)}$
at higher levels is assured by the inductive construction, since, at each point $p_{k} \in D$,
the logarithmic $1$-forms $\sigma_{j}^{*} \eta_{p}^{(j-1)}$ and $\eta_{p_{k}}^{(j)}$
differ from each other by a germ of
closed holomorphic $1$-form.
\smallskip \smallskip
\par \noindent
\underline{Case 2}: $\sigma_{j}$ is a dicritical blow-up.
We consider again $p_{1},\ldots,p_{\ell}$, which now are precisely
the points of $\supp_{j}$ on $D = \sigma_{j}^{-1}(p)$. We start by proving
the following:
\begin{assertion} $\ell \geq 2$.
\end{assertion}
\begin{proof} Indeed, if $j=n$, so that $D \subset \dic_{n}$, the result is clear:
recalling that, by condition (D.3), $\val(D)$ is actually the number of intersections
of $D$ with $\ndic_{n}$, the result
is obvious if $\val(D) \geq 2$ and, if $\val(D) = 1$, it is a consequence of condition (S.3).
Suppose now that $1 \leq j < n$ and $D \subset \dic_{j}$.
If $\varsigma_{j}$ is trivial over $D$, then the result follows from the case $j=n$.
When this is not the case, we
remark that, for each point $q \in D$ over which $\varsigma_{j}$ is non-trivial, we find,
as a consequence of (D.3) and (S.2),
at least one separatrix in $\sepndic_{j}$ at $q$.
Besides, each component of $ \D_{j}$, other than $D$, intersecting
$D$ at a point $q'$ over which $\varsigma_{j}$ is trivial is necessarily in $\ndic_{j}$ by (D.3).
We are then reduced to the situation where there is only one such point $q \in D$ and none of those points
$q' \in D$. But, in this case, denoting $\tilde{D} = \varsigma_{j}^{*}D$, then $\val(\tilde{D}) = 1$ in $\D_{n}$. By (S.3), we must have
a separatrix of $\sepdic_{j}$ touching $D$ at a point other than $q$, proving thereby the assertion.
\end{proof}
Now, at $p_{k} \in D = \sigma_{j}^{-1}(p)$, the given $\QQ_{p_{k}}^{(j)}$-logarithmic $1$-form
is written as
\[ \eta_{p_{k}}^{(j)} =
\sum_{\tilde{S} \in \br_{p_{k}}(\supp_{j}) } \lambda_{\tilde{S}} \frac{d f_{\tilde{S}}}{ f_{\tilde{S}}} + \alpha_{p_{k}},\]
where $f_{\tilde{S}} \in \cl{O}_{p_{k}}$ is a local equation for $\tilde{S}$
and $\alpha_{p_{k}}$ is a germ of closed holomorphic $1$-form at $p_{k}$.
For each $k=1,\ldots,\ell $,
denote by $\br_{p}^{p_{k}}(\supp_{j-1}) \subset \br_{p}(\supp_{j-1}) $ the subset of all branches $S$ such that $\tilde{S}= \sigma_{j}^{*}S \in \br_{p_{k}}(\supp_{j})$.
Also define
\[\rho_{p_{k}} = \sum_{S \in \br_{p}^{p_{k}}(\supp_{j-1}) } \nu_{p}(S) \lambda_{\tilde{S}}.\]
Another important point is the following:
\begin{assertion}
\label{assertion-r2}
$\rho_{p_{k}} \neq 0$ for every $k=1,\ldots,\ell$.
\end{assertion}
\begin{proof}
Suppose, to the contrary, that $\rho_{p_{k}} =0$ for some $k$.
Consider the germ of logarithmic $1$-form at $p \in \supp_{j-1}$ defined as
\[ \zeta^{(j-1)}_{p} =
\sum_{S \in \br_{p}^{p_{k}}(\supp_{j-1}) } \lambda_{\tilde{S}} \frac{d f_{S}}{ f_{S}},\]
where $f_{S} \in \cl{O}_{p}$ is the equation of $S$ corresponding to the equation $f_{\tilde{S}} \in \cl{O}_{p_{k}}$ of $\tilde{S} = \sigma_{j}^{*}S$.
Since $\zeta^{(j-1)}_{p}$ has a main resonance at $p$ and all its poles share the same tangent cone, by \cite[Prop. 10]{canocorral2006},
$\sigma_{j}$ is a non-dicritical blow-up for $\zeta^{(j-1)}_{p}$.
Denote the germ of $\sigma_{j}^{*}\zeta^{(j-1)}_{p}$ at $p_{k}$ by $\zeta^{(j)}_{p_{k}}$.
Since $\rho_{p_{k}} =0$, the logarithmic $1$-forms $\zeta^{(j)}_{p_{k}}$ and $\eta_{p_{k}}^{(j)}$ differ from each other by the germ of
closed holomorphic $1$-form $\alpha_{p_{k}}$. Thereafter, $\zeta^{(j)}_{p_{k}}$ is also $\QQ_{p_{k}}^{(j)}$-logarithmic.
We thus have that $D_{p_{k}}$, the germ of $D$ at $p_{k}$, is not a pole for
$\zeta_{p_{k}}^{(j)}$, although it is invariant.
For this to occur, $D_{p_{k}}$ should be a escape separatrix for the logarithmic $1$-form $\zeta^{(j)}_{p_{k}}$
and, as a consequence, $\varsigma_{j}^{*} D_{p_{k}}$ should touch $\varsigma_{j}^{-1}(p_{k}) \subset \D_{n}$ at a dicritical component.
However, this would mean that two components in $\dic_{n}$ meet each other, which is not allowed
by (D.3). This contradiction proves the assertion.
\end{proof}
In order to amalgamate data at level $j-1$,
we chose $c_{1},\ldots,c_{\ell} \in \C^{*}$
such that
\begin{equation}
\label{eq-linear}
c_{1}\rho_{p_{1}} + \cdots + c_{\ell}\rho_{p_{\ell}} = 0 ,
\end{equation}
which is possible by the two assertions just proven.
Next, we modify the data of $\Lambda^{(j)}$ in the following way: at each $p_{k} \in D$,
for each branch $\tilde{S} \in \br_{p_{k}}(\supp_{j})$ we replace $\lambda_{\tilde{S}}$ by
$c_{k} \lambda_{\tilde{S}}$ and perform the same correction of residues at all points in the connected component of $\ndic_{j}$ containing $p_{k}$. Once accomplished this adjustment at level $j$, we accordingly modify residues
at higher levels, obtaining new partial system of residues, that we denote by $\tilde{\Lambda}^{(j)}$.
Next, we define a partial system of residues $\Lambda^{(j-1)}$ by
using information from $\tilde{\Lambda}^{(j)}$ at levels $j$ and higher and also descending
it to level $j-1$:
for a component
$S \subset \supp_{j-1}$ , we take $\lambda_{S}$ as the residue of $\tilde{S}= \sigma_{j}^{*}S$
in $\tilde{\Lambda}^{(j)}$. We clearly have a partial system of residues. In particular,
condition (R.3) for
$\sigma_{j}$ is a consequence of \eqref{eq-linear}.
As for the $\QQ_{p}^{(j-1)}$-logarithmic $1$-form, where $\QQ_{p}^{(j-1)} = (\TT_{p}, \Lambda^{(j-1)}_{p})$,
in the same way as in the non-dicritical case,
we define
\[ \eta^{(j-1)}_{p} =
\sum_{S \in \br_{p}(\supp_{j-1}) } \lambda_{\tilde{S}} \frac{d f_{S}}{ f_{S}}.\]
It is a consequence of equation \eqref{eq-linear} that the germ of $\sigma_{j}^{*}\eta^{(j-1)}_{p}$ at $p_{k}$ differs
from $c_{k}\eta^{(j)}_{p_{k}}$ by a germ of closed holomorphic $1$-form.
This implies that $\sigma_{j}^{*}\eta^{(j-1)}_{p}$ is faithful and, thus,
$\QQ_{p_{k}}^{(j)}$-logarithmic.
On the other hand, $\eta^{(j-1)}_{p}$ is $1$-faithful as a consequence of Proposition
\ref{prop-faithfull-amalgamation}, to be proven below
in Section \ref{section-escape}.
In particular, it is dicritical for the blow-up $\sigma_{j}$.
We can thus conclude that $\eta^{(j-1)}_{p}$ is faithful, and, therefore, that it is
$\QQ_{p}^{(j-1)}$-logarithmic.
\end{proof}
\begin{example}
\label{ex-foliation-quadruplet}
{\rm In Examples \ref{ex-foliation-structure}, \ref{ex-foliation-separatrices} and \ref{ex-foliation-indices},
we built
$\TT(\F) = (\Delta(\F),\Sigma(\F),\Upsilon(\F))$, a dicritical triplet associated with a singular holomorphic foliation $\F$ at $(\C^{2},0)$.
As a consequence of Theorem \ref{teo-quadruplet}, we can complete $\TT(\F)$ into a
dicritical quadruplet
$\QQ(\F) = (\TT(\F),\Lambda(\F)) = (\Delta(\F),\Sigma(\F),\Upsilon(\F),\Lambda(\F))$ and
find a $\QQ(\F)$-logarithmic $1$-form $\eta_{\F}$. We say that $\eta_{\F}$ is a logarithmic model for $\F$.
The existence of logarithmic models for germs of singular holomorphic foliations in the plane has been established in
\cite{corral2003} and \cite{canocorral2006}.
}\end{example}
\section{Complex meromorphic functions}
\label{section-moromorphic-functions}
Let $h$ be a germ of meromorphic function at $(\C^{2},0)$.
Let $\pi:(\tilde{M},\D) \to (\C^{2},0)$
be the reduction of singularities of the foliation $\HH$ defined by the levels of $h$.
In particular, $\pi$ raises the
indeterminacy of $h$: there exists a germ of holomorphic map $\tilde{h}: \tilde{M} \to \Pe_{\C}^{1}$
in a neighborhood of $\D$ such that, outside $\D$, $\tilde{h} = h \circ \pi$.
We associate with $h$ the dicritical duplet of
the logarithmic $1$-form $\eta_{h} = d h / h$: $\DD(h) = \DD(\eta_{h})$, where
$\DD(h) = (\Delta(h),\Sigma(h))$.
Actually, we have to assume that (S.3) is satisfied, as it will be the case below.
The typical fibers of $h$ have irreducible components whose equisingularity types
are determined by its \emph{indeterminacy structure}, i.e. the dicritical components of $\dic_{n}$.
On the other hand, the isolated separatrices of $\sepndic_{0}$
are irreducible components of non-typical fibers.
\begin{example}
{\rm
Let $h= x^{p}/y^{q}$, with $p,q \in \Z_{+}$ relatively prime.
The combinatorics involved in
$\Delta(h)$ is the same of that of Euclid's algorithm of the pair $(p,q)$. The sole
dicritical blow-up will be the last one, $\sigma_{n}$.
In other words, the dicritical structure depends on the residues of the
$1$-form $\eta_{h} = dh/h = p dx/x - q dy/y$.
If $p,q > 1$, then
$S_{1}: x=0$ and $S_{2}: y=0$ are isolated separatrices.
On the other hand, if, for instance, $p> 1$ and $q=1$, then
$S_{1}$ is isolated and $S_{2}$ is dicritical.
}\end{example}
The central result of this section concerns the construction of meromorphic functions
with prescribed indeterminacy structure, zeros and poles.
The indeterminacy structure of a meromorphic function $h$ can be specified by an
infinitesimal class $\kappa$ of a union of components of some blow-up divisor, meant to identify the dicritical components of
the reduction of singularities of the foliation $\HH$ given by the levels of $h$.
We can then state:
\begin{maintheorem}
\label{teo-structure-meromporphic}
Let $\kappa$ be an infinitesimal class and $\sep$ be a finite set of branches of analytic curves at $(\C^{2},0)$.
Then there exits a germ of meromorphic function $h$ at $(\C^{2},0)$ whose indeterminacy structure is given by $\kappa$ and such that
the set of all branches of $h=0$ and $h=\infty$ contains $\sep$.
\end{maintheorem}
\begin{proof} Let $\pi:(\tilde{M},\D) \to (\C^{2},0)$ be a sequence of blow-ups that is a minimal realization for the infinitesimal class $\kappa$ and for the
equisingularity class $\varepsilon(\sep)$.
Build a dicritical structure $\Delta$ over $\pi$
with dicritical components determined by $\kappa$.
Next, complete $\sep$ into a finite subset $\sep_{0} \subset \br_{0}$ so that conditions (S.2) and (S.3) are respected,
taking care that also (S.4) is satisfied. Then $\sep_{0}$, with the decomposition given by $\Delta$, is the level 0 of a
configuration of separatrices $\Sigma$, and thus we have a dicritical duplet $\DD$ at $(\C^{2},0)$.
We obtain the theorem's statement by proving the following: there
exists a germ of meromorphic function $h$ such that
$\DD(h) \geq \DD$.
This result will follow from Theorem
\ref{teo-quadruplet} once we prove that we can associate a system of
indices $\Upsilon$ having only negative rational indices at the singular points of the support of $\DD$ at the final level. We prove this next,
in Proposition \ref{prop-rational-index}, and assume it for now. Hence, applying Theorem \ref{teo-quadruplet}, we can find
a system of residues $\Lambda$ that completes a dicritical quadruplet $\QQ= (\Delta,\Sigma,\Upsilon,\Lambda)$ as well as
a $\QQ$-logarithmic $1$-form $\eta$. An attentive look at the induction in its proof shows that $\Lambda$ can be obtained
having only rational residues.
Indeed, its initialization works by the application of
Proposition \ref{prop-residues-leveln}, where residues are obtained in $\Q^{*}$ once we depart from $\lambda_{D} \in \Q^{*}$
for the maximal element $D$ of each connected component $\A = \A_{n,k}^{\iota}$.
It is also easy to see that the amalgamation of data belonging to the general step $j$ can be done in such a
way that all residues are generated in $\Q^{*}$.
Notably, in the dicritical case, in equation \eqref{eq-linear},
if all coefficients $\rho_{p_{k}}$ are in $\Q^{*}$, then we can get
a solution vector with coordinates in $\Q^{*}$.
With the writing of Definition \ref{def-Q-logarithmic}, after
replacing some $f_{S}$ by
$u f_{S}$, where $u \in \OO_{0}$ is a unit such that $ \lambda_{S} du / u = \alpha$, and after
multiplying by the
least common multiple of the denominators of the residues,
we can write
\[ \eta = \sum_{S \in \sep_{0}} n_{S} \frac{d f_{S}}{f_{S}},\]
where $ n_{S} \in \Z^{*}$.
By setting
\[ h = \prod_{S \in \sep_{0}} f_{S}^{n_{S}} ,\]
we have that $\eta = \eta_{h} = d h /h$, proving the theorem.
\end{proof}
In order to complete the proof of Theorem \ref{teo-structure-meromporphic}, we still have
to show the possibility of
assigning negative rational indices at the final level. For this purpose,
we shall explore the combinatorics dictated by rules (I.2), (I.3) and (I.4).
Let us then
associate with $\D = \pi^{-1}(0)$ a weighted graph $\gr(\D)$, in which
each irreducible component $D$ corresponds to a vertex, still denoted by $D$, whose weight is
$c(D) = D \cdot D$, with edges marking intersections
of components. The graph $\gr(\D)$ is a tree whose intersection matrix is negative definite by \cite{laufer1971}.
More generally, if $\A \subset \D$ is a connected union of irreducible components of $\D$,
then $\gr(\A)$ is a subtree with negative definite intersection matrix.
Indeed, up to renumbering the vertices of $\gr(\D)$, the intersection matrix of $\gr(\A)$ can be seen as a diagonal
block of that of $\gr(\D)$.
We complete the proof of Theorem \ref{teo-structure-meromporphic} with the following:
\begin{prop}
\label{prop-rational-index}
Let $\DD = (\Delta,\Sigma)$ be a dicritical duplet at $(\C^{2},0)$ of height $n = h(\Delta)$. Then there exists a
system of indices $\Upsilon$ associated with $\DD$ such that $I_{p}(S) \in \Q_{-}$
for every singular point $p \in \supp_{n}$ and every $S \in \br_{p}(\supp_{n})$.
\end{prop}
\begin{proof}
Since we can work separately with each connected component of $\ndic$,
we fix some $\A = \A_{n,k}^{\iota} \subset \ndic$ and associate its weighted graph $\gr(\A)$.
Observe that a small perturbation in the weights of $\gr(\A)$ does not change the negative definiteness of its
intersection matrix. More precisely, there is $\epsilon >0$ such that if $\widetilde{\gr}(\A)$ is the weighted graph obtained from $\gr(\A)$
by replacing each weight $c(D)$ by some $\tilde{c}(D)$ satisfying
$|\tilde{c}(D) - c(D)| < \epsilon$, then the intersection matrix of $\widetilde{\gr}(\A)$ is negative definite.
Choose
$S \in \sepndic_{0}$ such
that $\pi^{*}S$ intersects $\A$ at a point $p \in D_{0} = D^{\pi}(S)$.
This $S$ exists by (S.2).
Let us order the vertices of $\gr(\A)$ in such a way that $D_{0}$ is a maximal element. Following this order, denote by $s$ be the size of
the tree $\gr(\A)$ and stratify its vertices in levels $j=1,\ldots,s$, where level $s$
corresponds to the maximal vertex $D_{0}$, level $s-1$ to its immediate predecessors and so on.
Next, for each component $D \subset \A$ and for each trace singularity $q \in D$, with $q \neq p$, choose for $I_{q}(D)$ a sufficiently small value in $\Q_{-}$
in such a way that their sum over $D$ satisfies $\left|\sum I_{q}(D) \right| < \epsilon$.
Define $\tilde{c}(D)= c(D) - \sum I_{q}(D)$. In this way,
$\widetilde{\gr}(\A)$ has a negative definite intersection matrix.
The idea now is to use \cite[Prop. 3.3]{camacho1988} in a context where the sum of indices over a component $D$ is $\tilde{c}(D)$,
in order to show that all indices, for all corners and for the point $p$, are in $\Q_{-}$.
The argument goes by induction, with the initialization and the general step treated with the same argument.
Let $1 \leq j \leq s$ and suppose that all corners in components at levels lower than $j$ have indices in $\Q_{-}$ (an empty condition,
if $j=1$).
If $D$ is a component at level $j$, denote by $\widetilde{\gr}(\A_{D})$ the subtree of $\widetilde{\gr}(\A)$ formed by $D$ and all its predecessors.
$\widetilde{\gr}(\A_{D})$ also has a negative definite intersection matrix. If $j<s$,
let $D'$ be the successor of $D$ and denote $p' = D \cap D'$. If $j=s$, set $D'= S$ and $p'=p$.
Denote by $q_{1},\ldots,q_{r}$ the corners of $D$ with its predecessors, if there are any. Then the induction hypothesis and (I.4) gives that $I_{q_{i}}(D) \in \Q_{-}$
for $i=1,\ldots,r$.
The sum of indices over $D$ equaling $\tilde{c}(D)$ gives that $I_{p'}(D) \in \Q$.
Finally, regarding
$\widetilde{\gr}(\A_{D})$ as a combinatorial portrait of an invariant divisor with only one trace singularity at $p'$,
with sums of indices given by $\tilde{c}$ in place of $c$, an application
of \cite[Prop. 3.3]{camacho1988} gives that $I_{p'}(D) \not\in \R_{\geq 0}$. Thus, $I_{p'}(D) \in \Q_{-}$ and also
$I_{p'}(D') \in \Q_{-}$.
The proposition is proved once
induction reaches level $j=s$.
\end{proof}
\begin{remark}
{\rm In Theorem
\ref{teo-structure-meromporphic} the set of branches of zeroes and poles of $h$ contains $\sep$ as a proper subset, in general. As it is clear in the beginning
of its proof, we have to include in $\sep$ some branches in order to obtain a set where
conditions (S.2) and (S.3) are valid. The necessity of condition (S.3) is apparent: if $D \subset \D$ is a dicritical
component of valence one, then the lift $\tilde{h}$ of $h$ to $\tilde{M}$, restricted $D \simeq \Pe^{1}_{\C}$, must have a zero and a pole.
One of them being given by the component of $\overline{\D \setminus D}$ intersecting $D$, the other must be given by a
separatrix touching $D$.
}\end{remark}
\section{Real logarithmic models}
\label{section-real-models}
\subsection{Symmetric dicritical quadruplets}
Let $J:(x,y) \in \C^{2} \mapsto (\bar{x},\bar{y}) \in \C^{2}$ be the canonical
anti-holomorphic involution defined by the complex conjugation. The \emph{real trace} of $\C^{2}$, the fixed set
of $J$, is identified with $\R^{2}$.
Points are said to be \emph{real} when they belong to the real trace, and \emph{non-real} when they do not.
For $f \in \OO_{0}$, let $f^{\vee} \in \OO_{0}$
be such that
$f^\vee(x,y) = \xbar{f(J(x,y))}$
for every $(x,y)$ near $0 \in \C^{2}$.
We say that $f \in \OO_{0}$ is \emph{real} if $f = f^{\vee}$. This is equivalent to $f$ having
a Taylor series expansion at $0 \in \C^{2}$ with only real coefficients or to $f$ assuming only real values over the real trace.
For a branch $S \in \br_{0}$ having $f \in \OO_{0}$ as equation, we denote by $S^{\vee} \in \br_{0}$
the branch with equation $f^{\vee}$, whose trace is the $J$-image of the trace of $S$.
We say that a branch $S \in \br_{0}$ is \emph{real} if $S=S^{\vee}$. This happens if and only if
$S$ has a real $f \in \OO_{0}$ as equation.
A germ of holomorphic $1$-form $\omega = P dx + Q dy$ at $(\C^{2},0)$ is \emph{symmetric} if $P,Q \in \OO_{0}$ are real.
In this case, the germ of holomorphic foliation $\F$ defined by $\omega$, also said to be symmetric, is invariant by $J$:
if $U$ is a $J$-invariant neighborhood of $0 \in \C^{2}$ where $\omega$ is realized, and $p \in U$ is a regular point,
denoting by $L_{p}$ the leaf through $p$, then $J(L_{p})$ is the leaf through $p^{\vee} = J(p)$.
In particular, if $p \in \R^{2}$, then
$L_{p} = J(L_{p})$. Thus, $L_{p}|_{\R^{2}}$ is a local curve in $\R^{2}$ (of topological dimension one), which is
the local leaf of the foliation $\F_{\R}$ defined by $\omega_{\R}$, the restriction of $\omega$ to the real trace.
Note also that, if $\F$ is symmetric, $S$ is a separatrix at a real point $p$ if and only if $S^{\vee}$ is.
If $\sigma: (\tilde{\C^{2}},D) \to (\C^{2},0)$ is a punctual blow-up, then there is a unique
continuous involution $J':(\tilde{\C^{2}},D) \to (\tilde{\C^{2}},D)$
such that $\sigma(J'(p)) = J(\sigma(p))$ for every $p \in \tilde{\C}^{2}$.
Its fixed set, $\tilde{\R}^{2}$, is the \emph{real trace} of $\tilde{\C^{2}}$.
Evidently, the involution $J$ will be lifted by successive blow-ups provided their centers lie in the real trace.
More generally, we say that a
sequence $\pi:(\tilde{M},\D) \to (\C^{2},0)$ of quadratic blow-ups is \emph{symmetric}
if the anti-holomorphic involution $J$
has a continuous lift $\tilde{J}$ to $\tilde{M}$.
Loosely speaking, this happens if, among the individual blow-ups factoring $\pi$,
those with non-real centers are even in number and paired by conjugation.
A dicritical structure $\Delta$ of height $n = h(\Delta)$ at $(\C^{2},0)$ is \emph{symmetric} if its underlying sequence of
blow-ups is symmetric and, for every component $D \subset \D_{n}$, we have that $D \subset \dic_{n}$ if and only if
$D^{\vee} = \tilde{J}(D) \subset \dic_{n}$.
A configuration of separatrices $\Sigma$ at $(\C^{2},0)$ is \emph{symmetric} if it is framed on a symmetric dicritical
structure $\Delta$ and if $\sep_{0}$ is invariant by $J$. Clearly, this also gives that
$\sep_{n}$ is invariant by $\tilde{J}$ and, as a consequence,
it holds that $S \in \sepdic_{0}$ if and only if $S^{\vee} \in \sepdic_{0}$.
By (S.4), if $p \in \D_{n}$ is a real point and $S \in \sep_{n}$ is a separatrix through $p$, then $S$ is real.
In particular, if $p \in \supp_{n}$ is a real singular point, then both branches of $\supp_{n}$ at $p$ are real.
We say that $\DD = (\Delta,\Sigma)$ is a \emph{symmetric dicritical duplet}.
A system of indices $\Upsilon$ associated with a symmetric dicritical duplet $\DD$ is \emph{symmetric} if
whenever $p \in \supp_{n}$ and $S \in \br_{p}(\supp_{n})$, then $I_{p^{\vee}}(S^{\vee})=\xbar{I_{p}(S)}$.
In particular, if $p \in \supp_{n}$ is real singular point, then $I_{p}(S) \in \R^{*} \setminus \Q_{+}$ for each branch $S \in \br_{p}(\supp_{n})$.
We say that $\TT =(\Delta,\Sigma,\Upsilon)$ is a \emph{symmetric dicritical triplet}.
Finally, a $\emph{symmetric}$ system of residues $\Lambda$ for the symmetric dicritical duplet $\DD =(\Delta,\Sigma)$ is one for which
$\lambda_{S^{\vee}} = \xbar{\lambda_{S}}$ whenever $S \in \sep_{0}$.
Thus, $\lambda_{S} \in \R^{*}$ whenever $S$ is a real separatrix.
The application of (R.1), (R.2) and (R.3) gives that $\lambda_{S^{\vee}} = \xbar{\lambda_{S}}$
for every component $S \subset \supp_{n}$.
If $\Upsilon$ and $\Lambda$ are consistent symmetric $\DD$-systems of indices and residues, then we say
that $\QQ = (\Delta,\Sigma,\Upsilon,\Lambda)$ is a \emph{symmetric dicritical quadruplet}.
We say that a logarithmic $1$-form as in \eqref{eq-log-form} is \emph{symmetric} if its set of poles $\sep \subset \br_{0}$ is symmetric, as well as the corresponding residues and the holomorphic $1$-form
$\alpha$.
We can restate Theorem \ref{teo-quadruplet} in the context of symmetric multiplets. The very same proof presented in Section \ref{section-existence-quadruplet} works here
and we leave to the reader
its step-by-step verification.
\begin{teo}
\label{teo-quadruplet-real}
Let $\TT =(\Delta,\Sigma,\Upsilon)$ be a symmetric dicritical triplet. Then there exists:
\begin{enumerate}
\item a symmetric system of
residues $\Lambda$ such that $\QQ = (\TT,\Lambda) = (\Delta,\Sigma,\Upsilon,\Lambda)$ is a symmetric dicritical quadruplet;
\item a $\QQ$-logarithmic symmetric $1$-form $\eta$.
\end{enumerate}
\end{teo}
In the next subsection, we will handle abstract multiplets formed by real objects that
will model the desingularization, separatrices and indices of germs of real analytic vector fields
at $(\R^{2},0)$.
The complexification of these real objects will eventually give rise to symmetric multiplets.
In order to achieve this, we first have to consider the following:
\begin{ddef}
\label{def-quasi-dicritical-triplet}
{\rm
A \emph{quasi dicritical triplet} at $(\C^{2},0)$, or \emph{q-dicritical triplet} for short, is the object
$\TT^{\aleph} = (\Delta^{\aleph} ,\Sigma^{\aleph} ,\Upsilon^{\aleph} )$ composed by:
\begin{itemize}
\item a dicritical structure $\Delta^{\aleph}$;
\item a configuration of separatrices $\Sigma^{\aleph}$ with conditions (S.2) and (S.3) removed;
\item a system of indices $\Upsilon^{\aleph}$ with condition (I.3) deleted.
\end{itemize}
}\end{ddef}
The q-dicritical triplet $\TT^{\aleph}$ is \emph{symmetric} if all its elements are invariant by conjugation.
In this case, we say that a symmetric dicritical triplet $\TT = (\Delta,\Sigma,\Upsilon )$ \emph{extends} $\TT^{\aleph}$ if:
\begin{itemize}
\item The sequence of blow-ups subjacent to $\TT$ is obtained from that of $\TT^{\aleph}$ by additional blow-ups at
non-real points; $\Delta$ and $\Delta^{\aleph}$ provide the same classification as dicritical or non-dicritical for the common
components.
\item If $\sep_{0}^{\aleph}$ and $\sep_{0}$ denote the set of separatrices at level $0$ of $\Sigma^{\aleph}$ and $\Sigma$, respectively, then
$\sep_{0}^{\aleph} \subset \sep_{0}$ and $\sep_{0} \setminus \sep_{0}^{\aleph}$ contains only non-real separatrices.
\item For separatrices in $\sep_{0}^{\aleph}$, indices assigned by $\Upsilon$ and $\Upsilon^{\aleph}$ are the same.
\end{itemize}
We can always extend a symmetric q-dicritical triplet, as shown in the following result.
\begin{prop}
\label{prop-quasi-triplet}
Let $\TT^{\aleph} = (\Delta^{\aleph} ,\Sigma^{\aleph} ,\Upsilon^{\aleph} )$ be a symmetric q-dicritical triplet at $(\C^{2},0)$.
Then there exists a symmetric dicritical triplet $\TT = (\Delta,\Sigma,\Upsilon)$ that extends $\TT^{\aleph}$.
All indices in $\Upsilon$ can be obtained in $\R$.
\end{prop}
\begin{proof} We take $\Delta = \Delta^{\aleph}$. Our task is to add non-real separatrices to $\Sigma^{\aleph}$ and assign
them indices
in such a way that rules (S.2), (S.3) and (I.3) are observed.
Let $\A = \A_{n,k}^{\iota}$ be a connected component of $\ndic_{n}$, the invariant part of the final level of $\Delta$. Let $D \subset \A$ be a component.
Define $c^{\aleph}(D) = \sum I_{p}(D)$, where $p$ runs over the singular points of $D$ with respect to $\TT^{\aleph}$, i.e. its corners and
points where separatrices in $\sep_{n}^{\aleph}$ meet $D$. If $c^{\aleph}(D) = c(D)$, we do nothing.
If
$c(D) - c^{\aleph}(D) \in \R \setminus \Q_{\geq 0}$,
we choose a non-real point $p \in D$, take $S$ any smooth branch at $p$ transversal to $D$, add it to $\sepndic_{n}$,
with index $I_{p}(S) = 2/(c(D) - c^{\aleph}(D))$ getting, as a consequence,
$I_{p}(D) = (c(D) - c^{\aleph}(D))/2$. We also include $S^{\vee}$ in $\sepndic_{n}$, with $I_{p^{\vee}}(S^{\vee}) = I_{p}(S)$.
Now, if $c(D) - c^{\aleph}(D) \in \Q_{+}$, then we choose
$b_{1},b_{2} \in \R \setminus \Q$ such that $b_{1} + b_{2} = (c(D) - c^{\aleph}(D))/2$
and two distinct, non-conjugate, non-real points $p_{1}, p_{2} \in D$.
Take $S_{1}$ and $S_{2}$ two smooth branches at, respectively, $p_{1}$ and $p_{2}$, transversal to $D$,
and add them to $\sepndic_{n}$,
with indices $I_{p_{1}}(S_{1}) = 1/b_{1}$ and $I_{p_{2}}(S_{2}) = 1/b_{2}$.
Then, $I_{p_{1}}(D) + I_{p_{2}}(D) = b_{1} + b_{2} = (c(D) - c^{\aleph}(D))/2$.
Next, also include
$S_{1}^{\vee}$ and $S_{2}^{\vee}$ to $\sepndic_{n}$
with indices $I_{p_{1}^{\vee}}(S_{1}^{\vee}) = 1/b_{1}$ and $I_{p_{2}^{\vee}}(S_{2}^{\vee}) = 1/b_{2}$.
The object
$\TT = (\Delta,\Sigma,\Upsilon)$ constructed so far satisfies (I.3) for all components $D \in \A$.
It does not comply with (S.2) if $\sep_{n}^{\aleph}$ has no separatrices touching $\A$ and if
$c^{\aleph}(D) = c(D)$ for every $D \subset \A$. If this is so, fix $D \subset \A$, pick two distinct, non-conjugate, non-real points $p_{1}, p_{2} \in D$, then choose
$b_{1},b_{2} \in \R \setminus \Q$ such that $b_{1} + b_{2} = 0$
and repeat the construction that closes the previous paragraph.
The above procedure, applied to all connected components $\A_{n,k}^{\iota} \subset \ndic_{n}$, results in the validity of (I.3) and (S.2).
To finish the proof, if for some $D \subset \dic_{n}$ with $\val(D) = 1$ condition (S.3) is not verified by a separatrix of $\sep_{n}^{\aleph}$, we add
to $\sepdic_{n}$ a pair of non-real symmetric separatrices, touching $D$ transversally at pair of different trace points, and give them both zero as indices, doing the corresponding symmetric intervention in $D^{\vee}$ if $D$ is non-real.
\end{proof}
\subsection{Real analytic 1-forms}
Suppose that $(x,y)$ are now
coordinates in $\R^{2}$. Let $C_{0}^{\omega}$ denote the ring of germs of
real analytic functions with real values at $(\R^{2},0)$.
Let $\omega_{\R} = P(x,y) dx + Q(x,y)dy$ be a germ of real analytic $1$-form at $(\R^{2},0)$, where
$P,Q \in C_{0}^{\omega}$
are relatively prime non-unities.
Then, $\omega_{\R}$ defines, near $0 \in \R^{2}$, a real analytic foliation $\F_{\R}$ with isolated singularity at the origin.
We denote by $\omega$ the complexification of $\omega_{\R}$, which amounts to regarding $(x,y)$ as coordinates of $\C^{2}$ and
$P,Q$ as elements of $\OO_{0}$. Then, according to the definition in the previous section, $\omega$ is a symmetric $1$-form,
with an isolated singularity at $0 \in \C^{2}$, defining a singular holomorphic foliation
$\F$ invariant by the involution $J$. Clearly, viewing $\R^{2}$ as the real trace of $\C^{2}$, we have that
$\F_{\R} = \F|_{\R^{2}}$.
The foliation $\F_{\R}$ has a \emph{reduction of singularities} by real quadratic blow-ups, say $\pi_{\R}:(\tilde{M}_{\R},\D_{\R}) \to (\R^{2},0)$.
This means that, for the strict transform foliation $\tilde{\F}_{\R} = \pi_{\R}^{*}\F_{\R}$, we can obtain the real equivalent of the four items listed in Example \ref{ex-foliation-structure},
with the additional condition that real eigenvalues are associated with simple singularities.
In order to obtain $\pi_{\R}$, we can consider, for instance,
the minimal reduction of singularities for $\F$, say $\pi:(\tilde{M},\D) \to (\C^{2},0)$, and
take $\tilde{M}_{\R}$ as the real trace of
$\tilde{M}$, determined by the lift of $J$, getting also $\D_{\R} = \D \cap \tilde{M}_{\R}$
and $\pi_{\R} = \pi|_{\tilde{M}_{\R}}$.
The following definition
associates with a sequence of real quadratic blow-ups an abstract triplet composed
only by real objects.
\begin{ddef}
\label{def-real-quasi-dicritical-triplet}
{\rm
A \emph{real quasi dicritical triplet}, or \emph{real q-dicritical triplet}, framed on a sequence of real quadratic blow-ups $\pi_{\R}$ is the object
$\TT^{\aleph}_{\R} = (\Delta^{\aleph}_{\R} ,\Sigma^{\aleph}_{\R} ,\Upsilon^{\aleph}_{\R})$ whose components are:
\begin{itemize}
\item $\Delta^{\aleph}_{\R}$ is the structure of real divisors produced by the sequence of blow-ups $\pi_{\R}$;
\item $\Sigma^{\aleph}_{\R}$ is formed by real analytic branches;
\item $\Upsilon^{\aleph}_{\R}$ includes only real
indices.
\end{itemize}
They satisfy the set of axioms that defines a dicritical triplet, except for (S.2), (S.3) and (I.3).
}\end{ddef}
Evidently,
the complexification of the entries of $\TT^{\aleph}_{\R}$ gives readily
a symmetric q-dicritical triplet $\TT^{\aleph}$, framed on the sequence of
complex quadratic blow-ups $\pi$, the complexification of $\pi_{\R}$.
Following the definition in \cite{risler2001}, $\F_{\R}$ is of \emph{real generalized curve} type if $\tilde{\F}_{\R}$
has no real algebraic saddle-node singularities, i.e. simple singularities with one zero eigenvalue. This is equivalent to asking that $\tilde{\F}$ has no saddle-node singularities over the real trace of $\tilde{M}$.
With a foliation of this kind, we can associate a
real q-dicritical triplet
$\TT^{\aleph}_{\R}(\F_{\R})= (\Delta^{\aleph}_{\R}(\F_{\R}),\Sigma^{\aleph}_{\R}(\F_{\R}),\Upsilon^{\aleph}_{\R}(\F_{\R}))$,
including in $\Sigma^{\aleph}_{\R}(\F_{\R})$ all isolated separatrices and a finite number of dicritical separatrices.
It gives rise, by complexification, to a symmetric q-dicritical triplet
$\TT^{\aleph}(\F_{\R})= (\Delta^{\aleph}(\F_{\R}),\Sigma^{\aleph}(\F_{\R}),\Upsilon^{\aleph}(\F_{\R}))$.
We say that a germ of $1$-form $\eta_{\R}$ at $(\R^{2},0)$ is \emph{real logarithmic} if it has a writing as in \eqref{eq-log-form1}, where each $f_{i} \in C_{0}^{\omega}$ is irreducible, the residues
$\lambda_{i}$ are real and $\alpha$ is a germ of real analytic $1$-form.
Note that the zero set of some of the $f_{i}$ may degenerate to the origin. When $f_{i} = 0$ is one dimensional, we say that it defines a \emph{real pole}
$S_{i}$ of $\eta_{\R}$.
The notion of $1$-faithful logarithmic $1$-form, set in Definition \ref{def-faithful}, extends unequivocally to the real context.
We then say that a real logarithmic $1$-form is real faithful if it is
$1$-faithful along the reduction of singularities of $\LL_{\R}$,
the germ of real analytic foliation at $(\R^{2},0)$ defined by $\eta_{\R}$.
Following the same steps of the construction of a dicritical triplet for a complex logarithmic $1$-form, preceding Definition \ref{def-system-residues},
we can build a real q-dicritical triplet
$\TT^{\aleph}_{\R}(\eta_{\R}) = (\Delta^{\aleph}_{\R}(\eta_{\R}) ,\Sigma^{\aleph}_{\R}(\eta_{\R}),\Upsilon^{\aleph}_{\R}(\eta_{\R}))$
for a real logarithmic $\eta_{\R}$.
Briefly, $\Delta^{\aleph}_{\R}(\eta_{\R})$ is based on the real reduction of singularities of $\LL_{\R}$, $\Sigma^{\aleph}_{\R}(\eta_{\R})$ includes
all real poles of $\eta_{\R}$ along with all isolated separatrices of $\LL_{\R}$, and $\Upsilon^{\aleph}_{\R}(\eta_{\R})$ collects all Camacho-Sad indices
of $\LL_{\R}$.
Next, the notion of dominance for dicritical duplets can be transposed in a straightforward way to real q-dicritical duplets, and, with it,
the concepts of escape point and escape separatrix. Keeping our notation, the writing
$\DD^{\aleph \, '}_{\R} \geq \DD^{\aleph}_{\R}$
means that $\DD^{\aleph \, '}_{\R} = (\Delta^{\aleph \, '}_{\R} ,\Sigma^{\aleph \, '}_{\R}) $ dominates $\DD^{\aleph}_{\R} = (\Delta^{\aleph}_{\R} ,\Sigma^{\aleph}_{\R})$.
\begin{ddef}
\label{def-real-log-model}
{\rm
Let $\TT^{\aleph}_{\R} = (\Delta^{\aleph}_{\R} ,\Sigma^{\aleph}_{\R} ,\Upsilon^{\aleph}_{\R})$ be a real q-dicritical
triplet at $(\R^{2},0)$.
A germ of real logarithmic $1$-form
$\eta_{\R}$ is a \emph{real logarithmic model} for $\TT^{\aleph}_{\R}$ if:
\begin{enumerate}[label=(\arabic*)]
\item the set of real poles of $\eta_{\R}$ is $\sep^{\aleph}_{0}$, the set of separatrices of
$\Sigma^{\aleph}_{\R}$ at level $0$;
\item $\DD^{\aleph}_{\R}(\eta_{\R}) \geq \DD^{\aleph}_{\R}$;
\item $\eta_{\R}$ is real faithful;
\item indices for separatrices in $\sep^{\aleph}_{0}$ are the same, for $\Upsilon^{\aleph}_{\R}$ and for $\Upsilon^{\aleph}_{\R}(\eta_{\R})$.
\end{enumerate}
}\end{ddef}
The real logarithmic model is \emph{strict} if $\DD^{\aleph}_{\R}(\eta_{\R}) = \DD^{\aleph}_{\R}$.
The following result generalizes, to the dicritical case, the main theorem in \cite{corral2012}:
\begin{maintheorem}
\label{teo-real-logmodel}
Let $\TT^{\aleph}_{\R} = (\Delta^{\aleph}_{\R} ,\Sigma^{\aleph}_{\R} ,\Upsilon^{\aleph}_{\R})$ be a real q-dicritical
triplet at $(\R^{2},0)$. Then there exists a germ of real logarithmic $1$-form $\eta_{\R}$ that is a
real logarithmic model for $\TT^{\aleph}_{\R}$.
\end{maintheorem}
\begin{proof}
We complexify $\TT^{\aleph}_{\R}$, obtaining
the symmetric q-dicritical triplet $\TT^{\aleph}$.
By Proposition \ref{prop-quasi-triplet}, $\TT^{\aleph}$ can be completed into a
symmetric dicritical triplet $\TT$ with real indices.
As a consequence of Theorem \ref{teo-quadruplet-real},
there are a symmetric system of
residues $\Lambda$, with real residues, such that $\QQ = (\TT,\Lambda)$ is a symmetric dicritical quadruplet,
and a symmetric $\QQ$-logarithmic $1$-form $\eta$, whose residues are real. If $\eta_{\R}$ is the restriction of $\eta$ to the real trace of $\C^{2}$,
then $\eta_{\R}$ is a real logarithmic model for $\TT^{\aleph}_{\R}$.
\end{proof}
In Definition \ref{def-real-log-model}, when $\TT^{\aleph}_{\R} = \TT^{\aleph}_{\R}(\F_{\R})$ for some germ of singular real analytic foliation $\F_{\R}$ of real generalized curve type,
we say that $\eta_{\R}$ is a \emph{real logarithmic model} for $\F_{\R}$. The logarithmic model is \emph{strict} if
$\eta_{\R}$ is a strict
real logarithmic model for some real q-dicritical triplet $\TT^{\aleph}_{\R}(\F_{\R})$ associated with $\F_{\R}$.
For germs of real analytic foliations, we can assure the existence of strict real logarithmic models:
\begin{maintheorem}
\label{teo-strict-logmodel-real}
Let $\F_{\R}$ be a
germ of singular real analytic foliation at $(\R^{2},0)$ of real generalized curve type.
Then there exists a strict
real logarithmic model for $\F_{\R}$.
\end{maintheorem}
\begin{proof} The existence of a logarithmic model is a consequence of Theorem \ref{teo-real-logmodel}.
The possibility of obtaining a
strict real logarithmic model is justified in the next section, by Proposition \ref{prop-withour-real-escape}
and the comments following it.
\end{proof}
As we did for complex sequences of blow-ups in Section \ref{section-dicritical-structure}, we can
define \emph{real infinitesimal classes} for sequences of real quadratic blow-ups. In this case, they are denoted by
$\kappa_{\R}$.
We can then apply the machinery of this section to the construction of real meromorphic functions, yielding the following real version of
Theorem \ref{teo-structure-meromporphic}:
\begin{maintheorem}
\label{teo-structure-meromporphic-real}
Let $\kappa_{\R}$ be a real infinitesimal class and $\sep$ be a finite set of branches of real analytic curves at $(\R^{2},0)$.
Then there exits a germ of real meromorphic function $h$ at $(\R^{2},0)$ whose indeterminacy structure is given by $\kappa_{\R}$ and such that
the set of all branches of $h=0$ and $h=\infty$ equals $\sep$.
\end{maintheorem}
\begin{proof} Information from $\kappa_{\R}$ and $\sep$ define a real q-dicritical duplet
$\DD^{\aleph}_{\R} = (\Delta^{\aleph}_{\R} ,\Sigma^{\aleph}_{\R})$. We complexify it
and add a symmetric set of complex separatrices in order that conditions (S.2) and (S.3) are accomplished, obtaining a dicritical duplet
$\DD = (\Delta,\Sigma)$.
Now we proceed as in Section \ref{section-moromorphic-functions}.
By Proposition \ref{prop-rational-index}, we can produce a system of indices $\Upsilon$ associated with $\DD$ having indices
in $\Q_{-}$, and we do it in a symmetric way with respect to the involution $J$.
As in Theorem \ref{teo-structure-meromporphic}, having as reference the symmetric dicritical triplet
$\TT = (\Delta,\Sigma,\Upsilon)$, we build a germ of meromorphic function $h$ at $(\C^{2},0)$, which is also symmetric with respect to $J$.
The restriction of $h$ to the real trace $\R^{2}$ is the desired germ of real meromorphic function.
\end{proof}
\section{Escape set and escape separatrices}
\label{section-escape}
In the construction of logarithmic models, as byproducts, some separatrices
not originally modelled by the dicritical quadruplet appear.
If $\QQ$ is a dicritical quadruplet and $\eta$ is a $\QQ$-logarithmic $1$-form, it turns out that the reduction of
singularities of the singular foliation $\LL$ defined by $\eta$ may be longer than the underlying sequence of blow-ups of $\QQ$. The additional blow-ups
start at centers called escape points, which are dicritical points for $\QQ$, outside its support, where
the strict transform of $\eta$ is closed holomorphic, having thus a holomorphic first integral.
Escape points give rise to some isolated separatrices for $\LL$, accordingly called
escape separatrices.
In this section we shall present a description of these objects.
In the real case, escape points can be eliminated in
a process called logarithmic modification, by which additional dicritical separatrices are included,
without affecting the data conveyed by the original dicritical quadruplet.
\subsection{Escape function and faithful $1$-forms}
Let $\eta$ be a logarithmic $1-$form at $(\C^{2},0)$, written as in \eqref{eq-log-form}. We can suppose that
the holomorphic closed part $\alpha$ has been incorporated into one of the equations of the polar set, in such a way that
\[ \eta = \lambda_{1} \frac{d f_{1}}{f_{1}} + \cdots + \lambda_{r} \frac{d f_{r}}{f_{r}},\]
where $r \geq 2$ and, for $i=1,\ldots,r$, $f_{i} \in \cl{O}_{0}$ is irreducible with order $ \nu_{i} = \nu_{0}(f_{i})$
and $\lambda_{i} \in \C^{*}$.
Set
$\lambda_{0} = \sum_{i=1}^{r} \nu_{i} \lambda_{i}$, the weighted balance of residues of $\eta$.
Denote by $\LL$
the germ of holomorphic foliation at $(\C^{2},0)$ defined by $\eta$.
This foliation is also induced by the holomorphic $1$-form $\omega = f \eta$, where $f = f_{1} \cdots f_{r}$,
which may have a one dimensional singular set.
For the quadratic blow-up $\sigma: (\tilde{\C}^{2},D) \to (\C^{2},0)$, we consider
coordinates $(x,t) \in \C^{2}$ such that $\sigma(x,t) = (x,xt)$, in which $D = \sigma^{-1}(0)$ has $x=0$ as an equation.
Then
the \emph{divided blow-up} of $\omega$ (even in the non-isolated singularity case) is
\begin{equation}
\label{eq-divided-blowup}
\text{(i) non-dicritical:} \ \ \tilde{\omega} = \sigma^{*}\omega / x^{m_{0}} \qquad \text{or} \qquad
\text{(ii) dicritical:} \ \ \tilde{\omega} = \sigma^{*}\omega / x^{m_{0}+1},
\end{equation}
where $m_{0} = \nu_{0}(\omega)= \nu_{0}(f \eta)$ (see, for instance, \cite{camacho1984}).
The $1$-form $\tilde{\omega}$, which induces $\tilde{\LL} = \sigma^{*} \LL$ in the coordinates $(x,t)$, is holomorphic and does not contain $D$ in its singular set.
Recall, from Definition \ref{def-faithful}, that
$\eta$ is $1$-faithful if and only if $m_{0} = \nu_{0}(\omega) = \nu_{0}(f \eta) = m-1$, where $m = \nu_{0}(f) = \nu_{1} + \cdots + \nu_{r}$.
This is equivalent to the fact that, in
\[ \omega = f \eta = \sum_{i=1}^{r} \lambda_{i} f_{1} \cdots \widehat{f_{i}} \cdots f_{r} d f_{i} , \]
the initial term is precisely that obtained by operating with the initial terms of the equations $f_{i}$.
In other words, it is expressed exclusively in terms of the residues $\lambda_{i}$ and of the equations of the tangent cones of the functions $f_{i}$.
By a linear change of coordinates,
we can assume that none of the functions $f_{i}$ contains the $y$-axis in its tangent
cone. We write, after possibly multiplying by non-zero constants,
\[ f_{i}(x,y) = (y -\alpha_{i}x)^{\nu_{i}} + g_{i}(x,y), \]
where
$\alpha_{i} \in \C$ and $g_{i}(x,y) \in \OO_{0}$ is such that $\nu_{0}(g_{i}) \geq \nu_{i}+1$.
We have
\[f_{i}(x,xt) = x^{\nu_{i}}\tilde{f}_{i}(x,t)= x^{\nu_{i}} \left((t- \alpha_{i})^{\nu_{i}} + x \tilde{g}_{i}(x,t) \right),\]
where $\tilde{g}_{i}(x,t) = g_{i}(x,xt)/x^{\nu_{i}+1}$.
It follows readily that
\begin{equation}
\label{eq-sigma*eta}
\tilde{\eta} = \sigma^{*} \eta = \lambda_{0} \frac{dx}{x} +
\lambda_{1} \frac{d \tilde{f}_{1}}{\tilde{f}_{1}} + \cdots + \lambda_{r} \frac{d \tilde{f}_{r}}{\tilde{f}_{r}}.
\end{equation}
If we set $\tilde{f} = \tilde{f}_{1} \cdots \tilde{f}_{r}$, we have
\begin{equation}
\label{eq-sigma*omega}
\sigma^{*} \omega = \sigma^{*} ( f \eta )= x^{m} \tilde{f} \tilde{\eta}.
\end{equation}
If $\lambda_{0} \neq 0$, by cancelling the poles in
\eqref{eq-sigma*eta}, we get
\[\tilde{\omega} = x \tilde{f} \tilde{\eta} = \lambda_{0} \tilde{f} dx +
x \sum_{i=1}^{r} \lambda_{i} \tilde{f}_{1} \cdots \widehat{ \tilde{f}_{i}} \cdots \tilde{f}_{r} d \tilde{f}_{i},\]
where,
as usual, ``\, $\widehat{\ \ }$\, '' indicates the absence of the corresponding factor.
It is clear that $D = \sigma^{-1}(0)$ is invariant by
$\tilde{\omega}$, that is, the blow-up $\sigma$ is non-dicritical.
Comparing with \eqref{eq-divided-blowup} and \eqref{eq-sigma*omega},
we conclude that $m_{0} = m-1$, and, thus, $\eta$ is $1$-faithful.
On the other hand, if $\lambda_{0} = 0$, then
\begin{equation}
\label{eq-tilde-omega}
\tilde{\omega} = \tilde{f} \tilde{\eta}=
\sum_{i=1}^{r} \lambda_{i} \tilde{f}_{1} \cdots \widehat{ \tilde{f}_{i}} \cdots \tilde{f}_{r} d \tilde{f}_{i}.
\end{equation}
If we suppose $\eta$ to be $1$-faithful then, by definition, $m-1 = m_{0} = \nu_{0}(f \eta)$.
Looking again at \eqref{eq-divided-blowup} and \eqref{eq-sigma*omega}, the equality
$m = m_{0} + 1$
implies that $\sigma$ is dicritical. That is, $D$ is non-invariant,
which is equivalent to the non-vanishing of
the term independent of $x$ in the coefficient of $d t$ in $\tilde{\omega}$.
Let us thus assume $\eta$ to be $1$-faithful and main resonant, i.e. $\lambda_{0}=0$.
Our goal is to track the tangency points between $\tilde{\LL} = \sigma^{*} \LL$ and $D$.
The term independent of $x$ in the coefficient of $d t$ in $\tilde{\omega}$ is
\begin{multline}
\label{eq-tang-dt}
\ \ \ \sum_{i=1}^{r} \lambda_{i} \nu_{i} (t-\alpha_{1})^{\nu_{1}}
\cdots (t-\alpha_{i})^{\nu_{i}-1} \cdots (t-\alpha_{r})^{\nu_{r}} \\
= (t-\alpha_{1})^{\nu_{1}-1} \cdots
(t-\alpha_{r})^{\nu_{r}-1}
\sum_{i=1}^{r} \lambda_{i} \nu_{i} (t-\alpha_{1})
\cdots \widehat{(t-\alpha_{i})} \cdots (t-\alpha_{r}) .
\end{multline}
Let $a_{1}, \ldots, a_{\ell}$ denote the distinct points of the set $\{\alpha_{1},\ldots,\alpha_{r}\}$.
For $k=1,\ldots, \ell$, let $I_{k} = \{1 \leq i \leq r;\ \alpha_{i} = a_{k}\}$.
For each $k$, set
\[m_{k} = \sum_{i \in I_{k}} \nu_{i} \qquad \text{and} \qquad \rho_{k} = \sum_{i \in I_{k}} \lambda_{i} \nu_{i} .\]
With this notation, we have that \eqref{eq-tang-dt} is the polynomial
\begin{equation}
\label{eq-polynomial-pdot}
P_{\eta}(t) =
\sum_{k=1}^{\ell} \rho_{k} (t-a_{1}) \cdots \widehat{(t-a_{k})} \cdots (t-a_{\ell})
\end{equation}
times $(t-a_{1})^{m_{1}-1} \cdots
(t-a_{\ell})^{m_{\ell}-1}$.
The preceding discussion allows us to register the following conclusion:
\begin{fact}
\label{fact-faithfull-P-eta}
A logarithmic $1$-form $\eta$, main resonant at $0 \in \C^{2}$, is $1$-faithful if and only if $P_{\eta} \neq 0$.
\end{fact}
The points of tangency between $\tilde{\LL}$ and $D$ out of the set $\{a_{1},\ldots,a_{\ell}\}$
are the roots of $P_{\eta}(t)$.
At each one of them, $\tilde{\eta}$ is closed and holomorphic and, thus, $\tilde{\LL}$ has a holomorphic first integral.
The roots of $P_{\eta}(t)$ determine points
of two kinds, depending on $\tilde{\LL}$ being transversal to $D$ or not.
At point $p \in D$ of the first kind, the leaf through $p$ is contained in the singular set of $\tilde{\eta}$.
However, if $p \in D$ is of the second kind, the reduction of singularities of $\LL$ will be non trivial over $p$, formed by
non-dicritical blow-ups, which give
rise to some isolated separatrices for $\LL$. It is a escape point, with respect to the branches in the polar set of $\eta$
or with respect to any dicritical quadruplet $\QQ = (\Delta,\Sigma,\Upsilon,\Lambda)$ for which
$\eta$ is $\QQ$-logarithmic.
Observe that, in this latter case, $\eta$ is faithful by hypothesis and,
from Assertion \ref{assertion-r2} in the proof of
Theorem \ref{teo-quadruplet}, we have that, in \eqref{eq-polynomial-pdot}, $\rho_{k} \neq 0$, for $k=1,\ldots, \ell$.
The vanishing of the weighted balance of residues, that is $\sum_{i=1}^{r} \nu_{i} \lambda_{i} = \sum_{k=1}^{\ell} \rho_{k} = 0$,
gives that $P_{\eta}(t)$ has degree at most $\ell -2$. The coefficient of the term of degree $\ell -2$ is,
up to changing sign,
\begin{equation}
\label{eq-tk-2}
- \sum_{k=1}^{\ell} \left( \sum_{1 \leq i \leq \ell, \, i \neq k } \rho_{i} \right)a_{k}
= \sum_{k=1}^{\ell} \rho_{k}a_{k}.
\end{equation}
Equivalent calculations in the coordinates $(u,y)$ of the blow-up, where $x = uy$, lead to the polynomial
\[\hat{P}_{\eta}(u) = - \sum_{k=1}^{\ell} \rho_{k} a_{k} (1-a_{1}u)
\cdots \widehat{(1-a_{k}u)} \cdots (1-a_{\ell}u),\]
whose constant term is $ -\sum_{k=1}^{\ell} \rho_{k}a_{k}$.
Thus, $P_{\eta}(t)$ has degree $\ell -2$ if and only if $\hat{P}_{\eta}(0) \neq 0$, which means
that $\tilde{\LL}$ and $D$ are transversal at $(u,y) = (0,0)$.
Thus, by choosing appropriately the blow-up coordinates, we can always assume that $P_{\eta}(t)$ has degree $\ell -2$ and, thus,
the search for escape points of $\eta$ on $D$ can be entirely done in the coordinates $(x,t)$. In particular, if $\ell = 2$, $P_{\eta}$ is constant
and there are no escape points.
Let us define
\begin{equation}
\label{eq-rational-eta}
R_{\eta}(t) = \sum_{k=1}^{\ell} \frac{ \rho_{k}}{t-a_{k}} = \frac{P_{\eta}(t)}{Q_{\eta}(t)} ,
\end{equation}
where $Q_{\eta}(t) = (t-a_{1}) \cdots (t-a_{\ell})$. We name $R_{\eta}$ \emph{escape function}. The affine zeroes of this rational function
are the points of tangency considered, among which we find the escape points of $\eta$ on $D$.
As a consequence of Fact \ref{fact-faithfull-P-eta}, we have:
\begin{fact}
\label{fact-faithfull-R-eta}
A logarithmic $1$-form $\eta$, main resonant at $0 \in \C^{2}$, is $1$-faithful if and only if $R_{\eta} \neq 0$.
\end{fact}
The above is the key to prove the following result, which was used in the dicritical amalgamation in the proof of
Theorem \ref{teo-quadruplet}:
\begin{prop}
\label{prop-faithfull-amalgamation}
Let
$\eta = \eta_{1} + \cdots + \eta_{\ell}$, where each
$\eta_{j}$ is a germ of logarithmic $1$-form at $(\C^2,0)$ and $\ell \geq 2$.
Suppose that:
\begin{itemize}
\item the tangent cone of the polar set of each $\eta_{k}$ is a singleton, say $p_{k} \in D = \sigma^{-1}(0)$;
\item $p_{1}, \ldots, p_{\ell} \in D$ are distinct points;
\item if $\rho_{k}$ denotes the weighted balance of residues of $\eta_{k}$, then
$\rho_{1} + \cdots + \rho_{\ell} = 0$;
\item $\rho_{k} \neq 0$ for some $k$.
\end{itemize}
Then, $\eta$ is $1$-faithfull.
\end{prop}
\begin{proof} We take advantage of the above notation and write the blow-up $\sigma$ at $0 \in \C^{2}$ in coordinates $(x,t)$ such that $\sigma(x,t) = (x,tx)$, in which
the points
$p_{k}$ have coordinates $(x,t) = (0,a_{k})$, for $k=1,\ldots,\ell$.
On account of Fact \ref{fact-faithfull-R-eta}, it is enough to see that $R_{\eta} \neq 0$.
This is however evident from formula \eqref{eq-rational-eta}, since some $\rho_{k}$ is non-zero.
\end{proof}
\subsection{Logarithmic modifications}
We remain in the setting of the previous subsection,
keeping the notation.
Let $a_{1}', \ldots, a_{s}' \in \C$ be distinct numbers, none of them in $\{a_{1},\ldots,a_{\ell}\}$.
Let $\lambda_{1}', \ldots, \lambda_{s}' \in \C^{*}$. A \emph{$1$-logarithmic modification} of $\eta$ with
parameters $\tau = \{(a_{1}',\lambda_{1}'), \ldots, (a_{s}',\lambda_{s}')\}$ is a logarithmic $1$-form of the kind
\[ \eta^{\tau} = \eta + \lambda_{1}' df_{1}' /f_{1}' + \cdots + \lambda_{s}' df_{s}' /f_{s}',\]
where, for $i=1,\ldots,s$, the function $f_{i}' \in \OO_{0}$ is an equation of a smooth branch at $(\C^{2},0)$ whose
tangent cone is $t = a_{i}'$. If $\lambda_{1}' + \cdots + \lambda_{s}'= 0$
then the polynomial $P_{\eta^{\tau}}$ has degree at most $\ell + s - 2$.
In analogy with \eqref{eq-tk-2},
the coefficient of its term of degree $\ell + s - 2$ is, up to sign,
\begin{equation}
\label{eq-tk3}
\sum_{k=1}^{\ell} \rho_{k}a_{k} + \sum_{i=1}^{s} \lambda_{i}'a_{i}'.
\end{equation}
Supposing that $P_{\eta}$ has degree $\ell - 2$,
then, for a generic choice of $(a_{1}',\ldots,a_{s}',\lambda_{1}',\ldots,\lambda_{s}')$ in the hyperplane of equation
$\lambda_{1}' + \cdots + \lambda_{s}' = 0$ in $\C^{2s}$,
the degree of $P_{\eta^{\tau}}$ is precisely $\ell + s - 2$, so that the escape points of $\eta^{\tau}$
are found among the affine roots of $R_{\eta^{\tau}}$. In this case, we say that both
$\tau$ and the corresponding $\eta^{\tau}$ are \emph{balanced}.
Now, let $\QQ = (\Delta,\Sigma,\Upsilon,\Lambda)$ be a dicritical quadruplet at $(\C^{2},0)$, main resonant at
$0 \in \C^{2}$, and $\pi$ be its underlying sequence of blow-ups. Suppose that
$\{a_{1},\ldots,a_{\ell}\}$ lists all points of the tangent cone of $\sep_{0}$
--- which coincides with the intersection of the support $\supp_{1}$ with $D = \sigma^{-1}(0)$, where $\sigma$ is the initial
blow up of $\pi$ ---
as values of $t$ of
coordinates $(x,t)$ for $\sigma$.
As in the above paragraph,
let $\tau$ be a balanced set of parameters, accompanied by a choice of smooth branches
$S_{i}' \in \br_{0}$ of equations $f_{i}'=0$, for $i=1,\ldots,s$.
We define the \emph{1-logarithmic modification} of $\QQ$ with parameters
$\tau$ as the dicritical quadruplet
$\QQ^{\tau} = (\Delta^{\tau},\Sigma^{\tau},\Upsilon^{\tau},\Lambda^{\tau})$ obtained from $\QQ$ in the following manner:
\begin{itemize}
\item $\Delta^{\tau} = \Delta$;
\item $\Sigma^{\tau}$ is built upon the decomposition at level $0$ obtained by attaching to $\sepdic_{0}$ the
curves $S_{i}'$, for $i=1,\ldots,s$.
\item $\Upsilon^{\tau}$ preserves the indices of $\Upsilon$ and assigns to the new elements of $\sepdic_{0}$ indices $I_{0}(S_{i}') = 1$.
\item $\Lambda^{\tau}$ preserves the residues of $\Lambda$ and assigns residues
$\lambda(S_{i}') = \lambda_{i}'$.
\end{itemize}
In this framework, we have:
\begin{prop}
\label{prop-log-modification}
Suppose that $\eta$ is a $\QQ$-logarithmic $1$-form for some dicritical quadruplet $\QQ = (\Delta,\Sigma,\Upsilon,\Lambda)$
at $(\C^{2},0)$, main resonant at
$0 \in \C^{2}$. Let $\eta^{\tau}$ be a balanced $1$-logarithmic
modification of $\eta$ with parameters $\tau$.
Then $\eta^{\tau}$ is $\QQ^{\tau}$-logarithmic.
\end{prop}
\begin{proof} First, the initial blow-up $\sigma$ is dicritical for $\eta^{\tau}$. Indeed,
$\eta^{\tau}$ is main resonant and $1$-faithful at $0 \in \C^{2}$, the latter being a consequence of $R_{\eta^{\tau}} \neq 0$.
Next, we examine the points of the support of
the dicritical duplet $\DD^{\tau} = (\Delta^{\tau},\Sigma^{\tau})$ over $D = \sigma^{-1}(0)$. The points $p_{k}$, corresponding to
$t= a_{k}$, for $k=1,\ldots,\ell$,
are simultaneously in the supports of $\DD^{\tau}$ and $\DD = (\Delta,\Sigma)$. At each of these points,
$\sigma^{*}\eta^{\tau}$ differs from $\sigma^{*} \eta$
by a closed holomorphic $1$-form. Hence also $\sigma^{*}\eta^{\tau}$ is $\QQ_{p_{k}}$-logarithmic, which is the same of being $\QQ_{p_{k}}^{\tau}$-logarithmic,
since $\QQ_{p_{k}}^{\tau}$ and $\QQ_{p_{k}}$ coincide. The other points of the support of $\DD^{\tau}$ are $p_{1}',\ldots,p_{s}' $, corresponding to
$t= a_{1}',\ldots,t=a_{s}'$. The fact that $\sigma^{*}\eta^{\tau}$ at each $p_{i}'$ is $\QQ_{p_{i}'}^{\tau}$-logarithmic is obvious.
\end{proof}
Suppose that $\QQ = (\Delta,\Sigma,\Upsilon,\Lambda)$ has height $n = h(\Delta)$. Let
$d$ be the number of dicritical blow-ups of $\Delta$ and
denote by $q_{j_{k}} \in \D_{j_{k}}$ their centers, where $0 \leq j_{1} < \cdots < j_{d} < n$.
Starting with $q= q_{j_{d}}$, we can consider the localization $\QQ_{q}$ and, for a balanced set of parameters $\tau_{q}$, perform
the corresponding $1$-logarithmic modification, obtaining a dicritical quadruplet
$\QQ_{q}^{\tau_{q}}$ at $(\tilde{M}_{j_{d}},q)$ = $(\tilde{M}_{j_{d}},q_{j_{d}})$. The incorporation of new dicritical separatrices and their residues
at levels of $\QQ$ lower than $j_{d}$ gives rise to a new dicritical quadruplet
$\QQ^{\tau_{q}} = (\Delta^{\tau_{q}} ,\Sigma^{\tau_{q}},\Upsilon^{\tau_{q}},\Lambda^{\tau_{q}})$ at $(\C^{2},0)$.
Since $\tau_{q}$ is balanced and all new
dicritical separatrices introduced are equisingular, their weighted balances of residues, at levels lower than $j_{d}$, will always vanish.
Therefore, at any level, these new separatrices give no contribution to the weighted balances of residues prescribed by $\QQ$.
Descending the dicritical structure, we carry out successive $1$-logarithmic modifications at each $q_{j_{k}}$ with
balanced set of parameters $\tau_{j_{k}}$.
The outcome is a dicritical quadruplet
$\QQ^{\textsc{t}} = (\Delta^{\textsc{t}},\Sigma^{\textsc{t}},\Upsilon^{\textsc{t}},\Lambda^{\textsc{t}})$, where
the superscript $\textsc{T}$ makes reference to the set of information given by the pairs $(q_{j_{k}},\tau_{j_{k}})$,
for $k=1,\ldots,d$, along with the new separatrices introduced.
Now, if $\eta$ is a $\QQ$-logarithmic $1$-form at $(\C^{2},0)$, the successive $1$-logarithmic modifications just described
can be applied to $\eta$. The result is a germ logarithmic $1$-form $\eta^{\textsc{t}}$, which, by a careful
application of Proposition \ref{prop-log-modification}, can be shown to be $\QQ^{\textsc{t}}$-logarithmic.
It is noteworthy that a $1$-logarithmic modification at a point $q_{j_{k}}$ modifies
the escape function on $D_{j_{k}+1}=\sigma_{j_{k} +1}^{-1}(q_{j_{k}})$ but does not change escape functions on dicritical components
at higher or lower levels.
\subsection{Symmetric logarithmic modifications}
Suppose now that $\eta$ is a symmetric logarithmic $1$-form at $(\C^{2},0)$, defining a germ of singular holomorphic foliation $\LL$.
A set of parameters $\tau = \{(a_{1}',\lambda_{1}'), \ldots, (a_{s}',\lambda_{s}')\}$ is \emph{symmetric} if its invariant by conjugation.
In this case
we say that $\tau$ parametrizes a
\emph{symmetric $1$-logarithmic modification} of $\eta$ provided the set of equations
$f_{1}', \ldots, f_{s}' \in \OO_{0}$ is also invariant by the involution $J$.
In particular, if $a_{i}' \in \R$, then also $\lambda_{i}' \in \R$ and $f_{i}' \in \OO_{0}$
defines a real curve. As a consequence, $\eta^{\tau}$ is also a symmetric logarithmic $1$-form.
Remark that, if $\eta$ is main resonant and symmetric, then the set of parameters involved in formula \eqref{eq-rational-eta} is symmetric under conjugation. Thus,
$P_{\eta}$ can be obtained with real coefficients (the same for $P_{\eta^{\tau}}$, for a balanced symmetric $1$-logarithmic modification $\eta^{\tau}$).
In particular, the number of real roots of $P_{\eta}$ and its degree have the same parity.
Balanced symmetric 1-logarithmic modifications can be used to eliminate real escape points of symmetric logarithmic $1$-forms, as
described in the next proposition.
\begin{prop}
\label{prof-no-real-roots}
Suppose that $\eta$ is a main resonant $1$-faithful symmetric logarithmic 1-form at $(\C^{2},0)$.
Then there is a balanced symmetric set of real parameters $\tau$ such that
the corresponding symmetric
$1$-logarithmic modification $\eta^{\tau}$ has no escape points in the real trace of the divisor of the quadratic blow-up
at $0 \in \C^{2}$.
\end{prop}
\begin{proof}
As above, we fix $(x,t)$ coordinates for the blow-up $\sigma$ at $0 \in \C^{2}$ so that
the affine zeroes of the escape function $R_{\eta}(t)$ give all tangency points between $\tilde{\eta} = \sigma^{*}\eta$ and $D=\sigma^{-1}(0)$.
We can suppose that $P_{\eta}(t)$ has even degree. If this is not so,
we make a balanced symmetric 1-logarithmic modification with parameters $\{(a_{1}',\lambda_{1}'), (a_{2}',\lambda_{2}'), (a_{3}',\lambda_{3}')\}$, where $a_{i}',\lambda_{i}' \in \R^{*}$ for $i=1,2,3$, with $\lambda_{1}',\lambda_{2}',\lambda_{3}'$ sufficiently small.
Let $t_{0}$ be a real zero of $R_{\eta}(t)$, which can be supposed to be $0 \in \C$ after a real translation in the coordinate $t$.
We write, locally,
\[ R_{\eta}(t) = \frac{P_{\eta}(t)}{Q_{\eta}(t)}=
t^{n}g(t),\]
for some $n>0$, where $g(t)$ is a holomorphic function in
a neighborhood of $0 \in \C$ such that $\alpha = g(0) \in \R^{*}$.
Choose $r>0$ sufficiently small such that
$ |\alpha|/2 < |g(t)| < 2|\alpha|$
over $\overline{ \mathbb{D}}_{r}$.
This implies that
$g(t)$ has no zeroes over $\overline{\mathbb{D}}_{r}$ and,
for $t \in \R$ with $|t|<r$, $g(t)$ and $\alpha = g(0)$ have
the same sign.
Note also that $|t^{n} g(t)| > r^{n}|\alpha|/2$ over $\partial \overline{ \mathbb{D}}_{r}$.
\smallskip \smallskip
\par \noindent
\underline{Case 1}: $n$ even.
We perform a balanced 1-logarithmic modification with parameters $\tau_{0} = \{(a,\lambda),(-a,-\lambda)\}$, where the sufficiently small $a,\lambda \in \R^{*}$, with $a>0$, are to be specified later, getting
a logarithmic $1$-form $\eta^{\tau_{0}}$ for which
\[R_{\eta^{\tau_{0}}}(t) = t^{n} g(t) + \frac{\lambda}{t-a} + \frac{-\lambda}{t+a} =
\frac{t^{n}(t^{2}-a^{2}) g(t) + 2a\lambda}{t^{2}-a^{2}}.\]
Choose $|a| < r/2$, so that $|\lambda/(t \pm a)| < 2|\lambda|/r$
over $\partial \overline{ \mathbb{D}}_{r}$.
If we also choose $\lambda \in \R^{*}$ such that $|\lambda| < r^{n+1} |\alpha|/8$, we assure that
\[ \left|\frac{\lambda}{t-a} + \frac{-\lambda}{t+a} \right| <
\frac{4 |\lambda|}{r} <
\frac{ r^{n} |\alpha|}{2} < |t^{n} g(t)| \]
over $\partial \overline{\mathbb{D}}_{r}$. Hence, applying Rouch\'es's theorem,
$ t^{n} g(t) + \lambda/(t-a) - \lambda/(t+a)$ has exactly $n+2$ zeroes in $\mathbb{D}_{r}$.
They are roots of the function $t^{n}(t^{2}-a^{2}) g(t) + 2a\lambda$.
If $t_{0} \in \R$ is one of these roots, then
\begin{equation}
\label{eq-real-root0}
t_{0}^{n}(t_{0}^{2}-a^{2}) = - \frac{2a\lambda}{g(t_{0})}.
\end{equation}
Observe that
$t^{n}(t^{2}-a^{2}) \geq -2 K a^{n+2}$ for every $t \in \R$,
where $K = \frac{1}{n+2}(\frac{n}{n+2})^{\frac{n}{2}}$.
Choose first a sufficiently small $a \in \R^{*}$, with $a>0$, and then pick $\lambda \in \R^{*}$, also small, but satisfying $|\lambda| > 2 |\alpha|K a^{n+1} $.
We then have
\[\frac{2a|\lambda|}{|g(t_{0})|} > ( 2 a ) 2 |\alpha| K a^{n+1} \frac{1}{2|\alpha|} = 2 K a^{n+2}.\]
Taking $\lambda$ with the same sign as $\alpha = g(0)$, we have
\[-\frac{2a\lambda}{g(t_{0})} < - 2 K a^{n+2} \leq t^{n}(t^{2}-a^{2}) \ \ \ \text{for every}\ t \in \R.\]
Comparing with \eqref{eq-real-root0}, we conclude that the real root $t_{0}$ cannot exist.
\smallskip \smallskip
\par \noindent
\underline{Case 2}: $n$ odd.
Now we do a non-balanced 1-logarithmic modification with parameter $\tau_{0} = \{(a,\lambda)\}$, with sufficiently small $a,\lambda \in \R^{*}$. The
logarithmic $1$-form $\eta^{\tau_{0}}$ obtained is such that
\[R_{\eta^{\tau_{0}}}(t) =
t^{n} g(t) + \frac{\lambda}{t-a} =
\frac{t^{n}(t-a) g(t) + \lambda}{t-a}.\]
If $|a| < r/2$, then $|\lambda/(t-a)| < 2|\lambda|/r$ over $\partial \overline{ \mathbb{D}}_{r}$.
Choosing $\lambda \in \R^{*}$ such that $|\lambda| < r^{n+1} |\alpha|/4$, we assure that
$|\lambda/(t-a)| < r^{n} |\alpha|/2 < |t^{n} g(t)|$ over $\partial \overline{\mathbb{D}}_{r}$. Hence, applying again Rouch\'es's theorem,
$ t^{n} g(t) + \lambda/(t-a)$ has exactly $n+1$ zeroes in $\mathbb{D}_{r}$.
These zeroes are roots of the function $t^{n}(t-a) g(t) + \lambda$.
If $t_{0} \in \R$ is one of these roots, then
\begin{equation}
\label{eq-real-root1}
t_{0}^{n}(t_{0}-a) = - \frac{\lambda}{g(t_{0})}.
\end{equation}
We have that
$t^{n}(t-a) \geq - K a^{n+1}$ for every $t \in \R$,
where $K = \frac{1}{n+1} ( \frac{n}{n+1} )^{n}$.
Choose first a sufficiently small $a \in \R^{*}$ and then pick $\lambda \in \R^{*}$, also small, but satisfying $|\lambda| > 2 |\alpha|K a^{n+1} $.
We then have
\[\frac{|\lambda|}{|g(t_{0})|} > 2 |\alpha|K a^{n+1} \frac{1}{2|\alpha|} = K a^{n+1}.\]
Taking $\lambda$ with the same sign as $\alpha = g(0)$, we have
\[-\frac{\lambda}{g(t_{0})} < - K a^{n+1} \leq t^{n}(t-a) \ \ \ \text{for every}\ t \in \R.\]
Comparing with \eqref{eq-real-root1}, this precludes the existence of such a real root.
\smallskip \smallskip
\par
We complete the proof by performing successive individual 1-logarithmic modifications as above, whenever we find
a real root for the escape function.
Once a root has been worked out, generating new roots in a small disc around it, subsequent 1-logarithmic
modifications
will not perturb any of these new roots back to the real axis,
provided all parameters are sufficiently small.
Finally, since $P_{\eta}(t)$ has real coefficients and even degree, there is an even number of real roots with $n$ odd. Thus, if any such a root exists,
there are at least two of them, allowing us to choose the residues $\lambda$ in a way that
the overall 1-logarithmic modification is balanced.
\end{proof}
Remark that in the proof of Proposition \ref{prof-no-real-roots}, all residues can be obtained in $\Q^{*}$. This is particularly important
if our goal is to use logarithmic models in order to produce real meromorphic functions, as in Theorem
\ref{teo-sectorial-modeled-meromorphic} below.
The iterated application of Proposition \ref{prof-no-real-roots} in a symmetric dicritical structure
gives promptly the following result, which was used in Theorem \ref{teo-strict-logmodel-real} in order to
obtain a strict real logarithmic model.
\begin{prop}
\label{prop-withour-real-escape}
Let $\QQ = (\Delta,\Sigma,\Upsilon,\Lambda)$ be a symmetric dicritical quadruplet at $(\C^{2},0)$
and $\eta$ be a symmetric $\QQ$-logarithmic $1$-form. Then there is a finite set of balanced symmetric logarithmic modifications, with parameters $\textsc{T}$,
such that $\eta^{\textsc{t}}$ is a symmetric $\QQ^{\textsc{t}}$-logarithmic $1$-form without real escape points.
Besides, all individual $1$-logarithmic modifications can be taken with real parameters, with residues in $\Q^{*}$.
\end{prop}
Proposition \ref{prop-withour-real-escape} completes the proof of Theorem \ref{teo-strict-logmodel-real}
on the existence of strict logarithmic models for germs of singular real analytic foliations.
Indeed, having a germ of real analytic foliation $\F_{\R}$ of real generalized curve type, we provide a q-dicritical triplet
$\TT^{\aleph}_{\R} = \TT^{\aleph}_{\R}(\F_{\R})$ and Theorem \ref{teo-real-logmodel} gives it a logarithmic model $\eta_{\R}$.
In this process, $\TT^{\aleph}_{\R} $ is complexified, then completed into a symmetric
dicritical triplet $\TT$ with real indices, which turns into
a symmetric dicritical quadruplet $\QQ = (\TT,\Lambda)$ with real residues,
to finally obtain a $\QQ$-logarithmic $1$-form $\eta$, whose decomplexification is
$\eta_{\R}$.
If $\eta_{\R}$ fails to be a strict logarithmic model for $\F_{\R}$, that is to say, if escape points exist,
we apply Proposition \ref{prop-withour-real-escape} to $\eta$.
In this case, the set of parameters $\textsc{T}$ will be chosen in such a way that all separatrices involved are actual real dicritical separatrices of $\F$, the
germ of singular holomorphic foliation defined by $\eta$. The corresponding logarithmic modifications then produce a new q-dicritical triplet $(\TT^{\aleph}_{\R})^{\textsc{t}}$, the real trace of $\TT^{\textsc{t}}$,
which in practice only incorporates the data of the real separatrices introduced,
being also a q-dicritical triplet associated with $\F_{\R}$.
Thus, $\eta_{\R}^{\textsc{t}}$, the decomplexification of $\eta^{\textsc{t}}$, is a strict logarithmic model for $\F_{\R}$.
\section{Bendixson's sectorial decomposition}
\label{section-sectorial}
The existence of a sectorial decomposition for a planar real analytic vector field with isolated singularity
appeared in the seminal paper of I. Bendixson \cite{bendixson1901}, where, following the ideas of H. Poincar\'e, he developed a qualitative study of
the orbits of a planar vector field.
Bendixson described and proved the finiteness of such a decomposition
in a sufficiently small neighborhood of the singularity, say $0 \in \R^{2}$,
when some orbit approaches the singularity with a ``determined tangent''.
In modern terminology, this is a \emph{characteristic orbit} and the situation considered is called \emph{non-monodromic}.
In this case, it turns out that all orbits approaching the origin have tangents, defined as limits of secants.
Finitely many of these orbits will be the boundary of sectors,
which have a topological classification as \emph{hyperbolic}, \emph{parabolic} or \emph{elliptic}.
In a hyperbolic sector no trajectory has the origin in its limit set,
in a parabolic one all trajectories have the origin either in the $\alpha$
or in the $\omega$-limit set and, finally,
in an elliptic sector all trajectories have the origin both in their $\alpha$ and $\omega$-limit sets.
We refer the reader to \cite{ilyashenko2008} for a proof of the existence of a sectorial decomposition based on the technique of reduction of singularities,
which unveils other analytic aspects that will be relevant to the forthcoming discussion.
Let $\pi_{\R}: (\tilde{M}_{\R}, \D_{\R}) \to (\R^{2},0)$ be a sequence of real quadratic blow-ups and let
$\rho_{\R}: (\tilde{N}_{\R}, \E_{\R}) \to (\R^{2},0)$ be the corresponding sequence of
trigonometric blow-ups, where $\tilde{N}_{\R}$ is a real analytic surface with boundaries and corners (see definitions in \cite{ilyashenko2008}).
Sequences of trigonometric blow-ups to be considered henceforth are of this kind.
There is a canonical analytic map $\psi:(\tilde{N}_{\R}, \E_{\R}) \to (\tilde{M}_{\R}, \D_{\R})$
satisfying $\pi_{\R} \circ \psi = \rho_{\R}$, obtained by
the consideration of the double covering $\mathbb{S}^{1} \to \mathbb{P}^{1}_{\R}$
for each single blow-up.
We associate with each component $D \subset \D_{\R}$ its real \emph{infinitesimal class} $\kappa_{D}$.
If $D$ has valence $m$, then the subset of trace points of $D$
has $m$ connected components, whose
pre-images by $\psi$ are $2m$ connected components of the regular
part of $\E_{\R}$. They determine $2m$ distinct \emph{infinitesimal semiclasses} and all these
semiclasses are said to be
\emph{correlate} to the infinitesimal class $\kappa_D$ and also \emph{correlate} to each other, and the
same is said of the corresponding irreducible components of $\E_{\R}$ and $\D_{\R}$.
The infinitesimal semiclass of $E \subset \E_{\R}$ will be denoted by $\hat{\kappa}_E$.
Let $E_{x} \subset \E$ be the component corresponding to the positive $x$-semiaxis.
If $E, E' \subset \E$ are two distinct components, the notation $\hat{\kappa}_{E} \prec \hat{\kappa}_{E'}$
means that either $E = E_{x}$ or $E$ lies between $E_{x}$ and $E'$, considering the counterclockwise
orientation around $0 \in \R^{2}$.
Let $U$ be a small neighborhood of $0 \in \R^{2}$.
An analytic \emph{semicurve} in $U \setminus \{0\}$ is defined by a real analytic parametrization of the kind $\gamma:[t_{0},\infty) \to U \setminus \{0\}$ or $\gamma:(-\infty,t_{0}] \to U \setminus \{0\}$,
for some $t_{0} \in \R$.
We shall only consider analytic semicurves $\gamma$ in $U \setminus \{0\}$ whose limit set is either $\ell(\gamma) = \{0\}$ or
$\ell(\gamma) = \emptyset$.
Let $X_{\R}$ be a germ of real analytic vector field, with isolated singularity at $(\R^{2},0)$, of non-monodromic type.
Any integral semicurve $\gamma$ of $X_{\R}$ such that $\ell(\gamma)=0$ has the
\emph{iterated tangents} property, meaning that $\gamma$, as well as all its lifts by successive blow-ups, have tangents at their limit points
(we refer to \cite[Sec. 2.2]{sanz2000} for a proof).
In a small neighborhood of $0 \in \R^{2}$, an integral semicurve does not intercept a real analytic branch (unless the semicurve is contained in the branch) and two integral semicurves do not intercept.
Thus, a pair $\gamma,\gamma'$ of integral solutions of $X_{\R}$, both accumulating to the origin, can be counterclockwise ordered
having the positive $x$-semiaxis $\gamma_{x}$ as reference:
$\gamma \prec \gamma'$ means that either $\gamma = \gamma_{x}$ or $\gamma$ lies between $\gamma_{x}$ and $\gamma'$ in a sufficiently small neighborhood of $0 \in \R^{2}$.
We also say that $\gamma$ has type $\hat{\kappa}_{E}$, for some $E \subset \E$, if the $\rho_{\R}$-lifiting of $\gamma$ touches
$E$ in a trace point. In this case, if $E',E'' \subset \E$ are such that $\hat{\kappa}_{E'} \prec \hat{\kappa}_{E} \prec \hat{\kappa}_{E''}$, we denote
$\hat{\kappa}_{E'} \prec \gamma \prec \hat{\kappa}_{E''}$.
A \emph{sectorial model}
corresponds to the prescription,
in a sufficiently small neighborhood $U$ of $0 \in \R^{2}$, of the following information:
\begin{itemize}
\item a finite number of real analytic semicurves $\gamma_{1},\ldots,\gamma_{n}$ in $U \setminus \{0\}$, with limit set $0 \in \R^{2}$, not intercepting each other and having the property of iterated tangents,
such that $\gamma_{1} \prec \ldots \prec \gamma_{n} $;
\item a classification of the region $\chi = \chi(\gamma_{i},\gamma_{i+1})$, named \emph{sector}, formed by points of $U$ between $\gamma_{i}$ and $\gamma_{i+1}$, where
$i=1,\ldots,n$ and $\gamma_{n+1} = \gamma_{1}$,
as hyperbolic, parabolic or elliptic,
with no parabolic sector neighboring other parabolic sectors.
\end{itemize}
Evidently, in the first of these items, we could also ask other necessary conditions for the semicurves $\gamma_{i}$ to be integral curves of
a non-monodromic real analytic vector field.
Taking into account the machinery developed throughout this text, our proposal is to produce examples of real analytic
vector fields whose sectorial decomposition fits a prescribed sectorial model. Our techniques allow us to do this
in a variety of situations, having as output a vector field
of logarithmic nature,
i.e. whose dual $1$-form is a multiple of a real logarithmic $1$-form by a unity in $C^{\omega}_{0}$.
We will exemplify this procedure within a specific category of
non-monodromic real analytic vector fields that we call \emph{$\ell$-analytic}, delimited in the following paragraph.
Denote the set of irreducible germs of real analytic curves at $(\R^{2},0)$ by $\br^{\R}_{0}$. If $S \in \br^{\R}_{0}$, then $S \setminus \{0\}$ splits into
two \emph{semibranches}, $\gamma$ and $\gamma^{*}$, said to be \emph{adjoint} to each other.
Evidently, a semibranch $\gamma$ is an integral curve of a germ of real analytic vector field $X_{\R}$ if and only if $\gamma^{*}$ is.
An analytic semicurve $\gamma$ in $U \setminus \{0\}$ is analytic in the limit,
or \emph{$\ell$-analytic} for short, if either
$\ell(\gamma) = \emptyset$ or
$\ell(\gamma) = \{0\}$ and
there is $S \in \br^{\R}_{0}$ such that $\gamma(t) \in S$ whenever $|t|$ is sufficiently large. In an abuse of language,
we say that $\gamma$ itself is an analytic semibranch. An analytic curve $\Gamma:(-\infty,\infty) \to U \setminus \{0\}$ is $\ell$-analytic
if it can be partitioned in two analytic semi-curves, both $\ell$-analytic.
\begin{ddef}
\label{def-l-analytic}
{\rm We say that a germ of real analytic vector field $X_{\R}$ is \emph{$\ell$-analytic} if
all its orbits in $U \setminus 0$ are $\ell$-analytic, for some small neighborhood $U$ of the origin.
}\end{ddef}
Clearly,
a vector field whose underlying singular foliation is defined by the levels of a real meromorphic function
is $\ell$-analytic.
Simple
saddle-type vector fields
--- whose linear part has eigenvalues $\lambda_{1}, \lambda_{2}$ satisfying $\lambda_{2}/ \lambda_{1} \in \R_{-}$ ---
are $\ell$-analytic. On the other hand, simple
node-type vector fields
--- with eigenvalues
satisfying $\lambda_{2} / \lambda_{1} \in \R_{+} \setminus \Q_{+}$ ---
are not $\ell$-analytic.
A simple vector field with one zero eigenvalue --- an \emph{algebraic saddle-node} --- is $\ell$-analytic if and only if it is a topological saddle.
In particular, its weak separatrix --- the one corresponding to the zero eigenvalue --- is convergent.
Indeed, otherwise, by the center manifold theorem, there would exist an orbit asymptotic to the non-convergent weak separatrix,
which would not be $\ell$-analytic. A non-monodromic germ of vector field is $\ell$-analytic if and only if it is of center type.
The concept of $\ell$-analytic vector field is evidently invariant under quadratic blow-ups.
Thus, if $\pi_{\R}: (\tilde{M}_{\R}, \D_{\R}) \to (\R^{2},0)$ is a reduction of singularities of $\F_{\R}$ by a sequence of real quadratic blow-ups,
then $X_{\R}$ is $\ell$-analytic if and only if all simple singularities of $\tilde{\F}_{\R}= \pi_{\R}^{*}\F_{\R}$ over $\D_{\R}$ are topological saddles.
\begin{figure}[t]
\begin{center}
\includegraphics[width=5.5cm, angle=-55]{eliptico.eps}
\includegraphics[width=5.5cm, angle=-55]{parabolico.eps}
\includegraphics[width=5.5cm, angle=-55]{hiperbolico.eps}
\put(-390,30){$\underline{\tau = 0}$}
\put(-342,0){\scriptsize {$\xi'$}}
\put(-298,-62){\scriptsize {$\xi$}}
\put(-350,-75){\scriptsize {$p$}}
\put(-375,-45){\scriptsize {$p'$}}
\put(-343,-90){\scriptsize {$E$}}
\put(-387,-33){\scriptsize {$E'$}}
\put(-355,-55){\scriptsize {$\mathcal{A}$}}
\put(-313,-25){\scriptsize {$\chi(\xi,\xi')$}}
\put(-315,-37){\scriptsize {(elliptic)}}
\put(-290,-85){\scriptsize {$\chi_{E}$}}
\put(-375,17){\scriptsize {$\chi_{E'}$}}
\put(-252,30){$\underline{\tau = 1}$}
\put(-180,-26){\scriptsize {$\gamma$}}
\put(-212,-75){\scriptsize {$p$}}
\put(-237,-45){\scriptsize {$p'$}}
\put(-205,-90){\scriptsize {$E$}}
\put(-250,-33){\scriptsize {$E'$}}
\put(-220,-55){\scriptsize {$\mathcal{A}$}}
\put(-147,-90){\scriptsize {$\chi_{E}$}}
\put(-227,15){\scriptsize {$\chi_{E'}$}}
\put(-114,30){$\underline{\tau = 2}$}
\put(-64,-3){\scriptsize {$\gamma'$}}
\put(-26,-56){\scriptsize {$\gamma$}}
\put(-71,-75){\scriptsize {$p$}}
\put(-96,-45){\scriptsize {$p'$}}
\put(-64,-90){\scriptsize {$E$}}
\put(-110,-33){\scriptsize {$E'$}}
\put(-80,-55){\scriptsize {$\mathcal{A}$}}
\put(-32,-25){\scriptsize {$\chi(\gamma,\gamma')$}}
\put(-34,-37){\scriptsize {(hyperbolic)}}
\put(-9,-90){\scriptsize {$\chi_{E}$}}
\put(-89,15){\scriptsize {$\chi_{E'}$}}
\end{center}
\end{figure}
We will try to present a more refined description of the sectorial decomposition of an $\ell$-analytic vector field $X_{\R}$.
Let $\rho_{\R}: (\tilde{N}_{\R}, \E) \to (\R^{2},0)$ be the reduction of singularities of the associated singular foliation $\F_{\R}$ by a sequence of
trigonometric blow-ups and denote $\tilde{\F}_{\R}^{\rho}= \rho_{\R}^{*}\F_{\R}$.
Let $E,E' \subset \E$ be consecutive dicritical components with $\hat{\kappa}_{E} \prec \hat{\kappa}_{E'}$.
Let $\A$
denote
the union of invariant components of $\E$ between
$E$ and $E'$, which is non-empty by the properties of the reduction of singularities.
Let $\tau$ denote the number of separatrices --- i.e. invariant semibranches --- of $\tilde{\F}_{\R}^{\rho}$ touching $\A$.
The figure gives a portrait for the cases $\tau=0,1$ and $2$.
Define $p = \A \cap E$ and $p' = \A \cap E'$.
The dicritical components $E$ and $E'$
determine
two pieces of parabolic sectors, $\chi_{E}$ and $\chi_{E'}$ --- here we say ``piece'' since they may be proper subsectors of parabolic sectors.
We attach to them the infinitesimal semiclasses $\hat{\kappa}_{E}$ and $\hat{\kappa}_{E'}$, on the grounds that their interior orbits
are of types, respectively, $E$ and $E'$.
Suppose that $\tau = 0$.
Then, leaves of $\tilde{\F}_{\R}^{\rho}$ touching $E $ in points near $p$
reach $E'$ in points near $p'$. Choosing $\xi$ and $\xi'$
orbits whose $\rho_{\R}$-lifts
touch, respectively, $E$ in a point sufficiently near $p$ and
$E'$ in a point sufficiently near $p'$, then $\chi(\xi,\xi')$ is an elliptic
sector. The semibranch $\xi$ will be the upper boundary of $\chi_{E}$ and
$\xi'$ will be the lower boundary of $\chi_{E'}$.
On the other hand,
if $\tau > 0$, then the first of the separatrices reaching $\A$, say $\gamma$,
will be the upper boundary of $\chi_{E}$ and the last one, say $\gamma'$,
will be the lower boundary of $\chi_{E'}$. If $\tau =1$, then $\gamma = \gamma'$ and $\chi_{E}$
neighbor $\chi_{E'}$. Hence, these pieces of parabolic sectors are merged, as part of the formation of a maximal parabolic sector.
On the other hand, if $\tau > 1$, there are $\tau-1$ hyperbolic sectors between $\chi_{E}$ and $\chi_{E'}$, each
one bounded by a consecutive pair of separatrices of $\A$.
In other words, hyperbolic sectors are bounded by isolated invariant semibranches of $\F_{\R}$.
Observe that elliptic sectors always neighbor parabolic sectors and
their boundaries are always dicritical invariant semibranches of $\F_{\R}$.
By possibly reducing $U$ and changing the choice of some of these semibranches, we can also suppose
that a semibranch $\gamma$ is the boundary
of a sector or of a piece of sector if and only if $\gamma^{*}$ is.
Finally, it is also clear that if $E \subset \E$ is a dicritical component defining a piece of parabolic sector, then all components of $\E$ correlate to $E$ also define pieces of parabolic sectors.
The above description matches the model we describe next.
An \emph{infinitesimal sentence} is an object $\W = W_{1} W_{2} \cdots W_{r}$,
formed by \emph{words} $W_{i}$ of one of two \emph{types}: either $W_{i}=\gamma_{i}$
is a semibranch
or $W_{i}= \hat{\kappa}_{i}$ is an infinitesimal semiclass.
The following syntax must be obeyed:
\begin{itemize}
\item $W_{1} \prec W_{2} \prec \cdots \prec W_{r}$, where ``$\prec$'' denotes the counterclockwise order
of semibranches and semiclasses described above;
\item if $W_{i}= \gamma_{i}$ is a semibranch, then the adjoint $\gamma_{i}^{*}$ is also a word of $\W$;
\item if $W_{i}= \hat{\kappa}_{i}$ is an infinitesimal semiclass, then all semiclasses correlate to $\hat{\kappa}_{i}$
are words of $\W$.
\end{itemize}
For convenience, consider $W_{0} = W_{r}$ and $W_{r+1} = W_{1}$. We then establish the following:
\begin{ddef}
\label{def-sectorial-modeled}
{\rm
Let $\W$ be an infinitesimal sentence as above.
An \emph{$\ell$-analytic sectorial model}
of pattern $\W$ is the prescription of a sectorial decomposition of a small neighborhood of $0 \in \R^{2}$
complying with the following conditions:
\begin{itemize}
\item if $W_{i} = \gamma_{i}$ is a semibranch, then it is the boundary of a hyperbolic sector or of a piece of parabolic sector;
\item if $W_{i} = \hat{\kappa}_{i}$ is an infinitesimal semiclass, then it defines a piece of parabolic sector;
if $W_{i-1}$ or $W_{i+1}$ is a semibranch, say $\gamma$, then $\gamma$ is in its boundary;
\item if $W_{i} = \gamma$ and $W_{i+1} = \gamma'$ are semibranches, then $\chi(\gamma,\gamma')$ is a hyperbolic sector;
\item if $W_{i} = \hat{\kappa}$ and $W_{i+1} = \hat{\kappa}'$ are infinitesimal semiclasses, then there is an elliptic sector $\chi(\xi,\xi')$,
bounded by semibranches $\xi$
and $\xi'$ of types, respectively,
$\hat{\kappa}$ and $\hat{\kappa}'$;
these semibranches are also boundaries of pieces of parabolic sectors;
\item neighboring pieces of parabolic sectors are merged into a single parabolic sector;
\end{itemize}
}\end{ddef}
From the definition, we infer that parabolic sectors correspond, in the sentence $\W$, to maximal sequences of words of alternating types.
Aiming at giving them a more accurate description, we could say, for instance, that each such a sequence determines an \emph{infinitesimal multitype}
for the corresponding parabolic sector.
From our discussion, every $\ell$-analytic vector field has a sectorial decomposition of the above type.
Reciprocally, we have:
\begin{maintheorem}
\label{teo-sectorial-modeled}
There are $\ell$-analytic vector fields fitting any preassigned $\ell$-analytic sectorial model.
\end{maintheorem}
\begin{proof}
Let $\W$ be the infinitesimal sentence subjacent to the given $\ell$-analytic sectorial model.
Let $\sep \subset \br_{0}^{\R}$ be the set of branches whose semibranches are words of $\W$
and $\varepsilon = \varepsilon(\sep)$ be its real equisingularity class.
Let $\kappa$ be the real infinitesimal class that aggregates all semiclasses that are words of $\W$.
Let $\pi_{\R}$ be a sequence of real quadratic blow-ups that is a minimal simultaneous realization for $\kappa$ and $\varepsilon$.
Consider a real q-dicritical triplet $\TT^{\aleph}_{\R}$ built upon $\pi_{\R}$, with the dicritical structure defined by $\kappa$ and separatrices given
by $\sep$. As for the indices, for each singular point $p$ at the final level, choose, for
the two local branches of the support at $p$, negative indices respecting (I.4).
By Theorem \ref{teo-real-logmodel}, there is a real logarithmic 1-form $\eta_{\R}$, with real residues,
which is a real logarithmic model for $\TT^{\aleph}_{\R}$.
Using Proposition \ref{prop-withour-real-escape}, after a finite number of logarithmic modifications, we obtain a real logarithmic $1$-form $\eta^{\textsc{t}}$ without
escape points.
We cancel the poles of $\eta^{\textsc{t}}$ and take the dual vector field.
This vector field has the preassigned $\ell$-analytic sectorial model
\end{proof}
We can also obtain a version of the above theorem concerning the existence of real meromorphic functions:
\begin{maintheorem}
\label{teo-sectorial-modeled-meromorphic}
There are real analytic meromorphic functions fitting any preassigned $\ell$-analytic sectorial model.
\end{maintheorem}
\begin{proof}
We follow the steps of the proof of Theorem \ref{teo-sectorial-modeled}, associating with the given sectorial model a dicritical class $\kappa$ and an equisingularity class
$\varepsilon$.
Then, as in Theorem \ref{teo-structure-meromporphic}, we produce a real meromorphic function upon these data,
using, in this process, logarithmic modifications with residues in $\Q^{*}$ in order to get rid of all real escape points, as proposed in
Proposition \ref{prop-withour-real-escape}.
\end{proof}
\section{Examples}
As an illustration of methods and concepts introduced in the text, we present below a pair of examples.
In order to lighten notation, we denote by the same symbol a curve and its strict transform by blow-up maps.
Ambiguity will be avoided by making precise the level we are working at.
Also, all pictures presented correspond to the real trace.
\begin{example}{\rm
Consider, at $(\C^{2},0)$, the branches
\[ S_{1}: f_{1} = x-y^{2} = 0 \quad \text{and} \quad S_{2}: f_{2} = x+y^{2} = 0.\]
Let $\pi = \sigma_{1} \circ \sigma_{2}$ be the sequence of blow-ups desingularizing $\sep_{0}^{\aleph} = \{S_{1},S_{2}\}$.
We will produce a symmetric logarithmic model, with rational residues, in which $\sep_{0}^{\aleph}$ are separatrices at level 0 of a q-dicritical duplet, corresponding to
a dicritical structure determined by $\pi$, where $\sigma_{1}$ is a non-dicritical blow-up and $\sigma_{2}$
is a dicritical one.
\begin{figure}[h]
\begin{center}
\includegraphics[width=12cm]{Fig1.eps}
\put(-260,73){\scriptsize {$S_{1}$}}
\put(-355,73){\scriptsize {$S_{2}$}}
\put(-263,20){\scriptsize {$S_{3}$}}
\put(-352,20){\scriptsize {$S_{4}$}}
\put(-230,43){\scriptsize {$\sigma_{1}$}}
\put(-110,43){\scriptsize {$\sigma_{2}$}}
\put(-170,13){\scriptsize {$D_{1}$}}
\put(-170,50){\scriptsize {$p$}}
\put(-43,-3){\scriptsize {$D_{1}$}}
\put(-43,73){\scriptsize {$D_{2}$}}
\put(-145,73){\scriptsize {$S_{1}$}}
\put(-198,73){\scriptsize {$S_{2}$}}
\put(-130,23){\scriptsize {$S_{3}$}}
\put(-213,23){\scriptsize {$S_{4}$}}
\put(-2,81){\scriptsize {$S_{1}$}}
\put(-88,81){\scriptsize {$S_{2}$}}
\put(-03,5){\scriptsize {$S_{3}$}}
\put(-86,5){\scriptsize {$S_{4}$}}
\end{center}
\end{figure}
Denote by $D_{1}$ and $D_{2}$ the components of the exceptional divisor introduced by, respectively, $\sigma_{1}$ and $\sigma_{2}$.
In order that (S.2) is satisfied by $D_{1}$ at the final level, we introduce the pair non-real conjugate branches (represented in the picture by dotted lines)
\[ S_{3}: f_{3} = y- \ic x =0 \quad \text{and} \quad S_{3}: f_{4} = y+ \ic x =0 ,\]
so that $\sep_{0} = \{S_{1},S_{2},S_{3},S_{4}\}$ are separatrices at level 0 of a dicritical duplet.
We shall now work at level 1. Let $p \in D_{1}$ be the center of $\sigma_{2}$. If we set $\lambda_{D_{1}} = 1$,
in order to have a zero weighted balance of residues at $p$, we can put, for instance
$\lambda_{S_{1}}= \lambda_{S_{2}}= -1/2$.
Let $q_{3}$ and $q_{4}$ denote the points of intersection of $S_{3}$ and $S_{4}$ with $D_{1}$.
Since $i_{p}(D_{1}) = 1$ and $c(D_{1})=-1$, in order to have (I.3) satisfied on $D_{1}$, we can put $i_{q_{3}}(D_{1}) = i_{q_{4}}(D_{1}) = -1$. Since $\sigma_{2}$ is trivial
over $q_{3}$ and $q_{4}$, we have, by (I.4), $i_{q_{3}}(S_{3}) = i_{q_{4}}(S_{4}) = -1$.
By the consistency condition of Definition \ref{def-consistent}, we then find $\lambda_{S_{3}}= \lambda_{S_{4}}= 1$.
The residues obtained are such that
\[ \eta= - \frac{1}{2} \frac{df_{1}}{f_{1}} - \frac{1}{2} \frac{df_{2}}{f_{2}} +
\frac{df_{3}}{f_{3}} + \frac{df_{4}}{f_{4}}
= \frac{1}{2} \frac{dh}{h} , \quad
\text{where} \quad h = \frac{f_{3}^{2}f_{4}^{2}}{f_{1}f_{2}} ,
\]
is a logarithmic model for the data described above.
Let us write $\sigma_{1}$ in coordinates $(u,y)$ such that $x=uy$ and, then,
the blow-up $\sigma_{2}$ at $p$ in coordinates $(u,t)$ such that $y= tu$.
In these coordinates, at level 2, we have $S_{1}: t-1=0$, $S_{2}: t+1=0$
and $D_{1}: t=0$.
The escape function corresponding to $\eta_{1}= \sigma_{1}^{*} \eta$ at $p$ is
\[R_{\eta_{1}} = \frac{1}{t} + \frac{-1/2}{t-1}+ \frac{-1/2}{t+1} = \frac{-1}{t(t+1)(t-1)} = \frac{P_{\eta_{1}}}{Q_{\eta_{1}}}. \]
Since $\deg(P_{\eta_{1}}) < 1$, $\eta_{2} = \sigma_{2}^{*} \eta_{1} = \pi^{*} \eta$ has a point of tangency with $D_{2}$ at the point corresponding to $t= \infty$.
The level $-1$ of $h$ has equation
\[f_{3}^{2}f_{4}^{2}+f_{1}f_{2} = x^{2}(x^{2} + 2y^2 + 1) .\]
That is, the $y$-axis is a fiber of multiplicity two, thereby contained in $\sing(\eta)$. This one-dimensional
component accounts for the point of tangency detected. It is not a escape separatrix, since it does not force the introduction of additional blow-ups in the
desingularization process.
\begin{figure}[h]
\begin{center}
\includegraphics[height=4cm]{Fig2.eps}
\put(-220,104){\scriptsize $S_{1}$}
\put(-310,104){\scriptsize $S_{2}$}
\put(-222,11){\scriptsize $S_{2}$}
\put(-308,11){\scriptsize $S_{1}$}
\put(-253,97){\scriptsize $E_{1}$}
\put(-253,17){\scriptsize $E_{2}$}
\end{center}
\end{figure}
The real figure of this example corresponds to an $\ell$-analytic sectorial model patterned on the infinitesimal
sentence
$\cl{W}=W_{1}W_{2}$, whose words are the infinitesimal semiclasses $W_{1} = \hat{\kappa}_{E_{1}}$ and
$W_{2} = \hat{\kappa}_{E_{2}}$. The semiclasses $\hat{\kappa}_{E_{i}}$, $i=1,2$, yield two parabolic sector, while the transitions
$\hat{\kappa}_{E_{i}} \leftrightarrow \hat{\kappa}_{E_{j}}$, $i \neq j$, two elliptic sectors.
}\end{example}
\begin{example}
{\rm Take, for $i=1,2,3,4$, the germs of cuspids at $(\C^{2},0)$ defined by the equations
\[ S_{i}: f_{i}(x,y) = x^3 - (y-a_{i}x)^2 =0, \]
where $a_{1} = 1$, $a_{2} = -1$, $a_{3} = 4$, $a_{4} = -4$.
We consider $\sep_{0} = \{S_{1},S_{2},S_{3},S_{4}\}$ as the initial level of a configuration of separatrices framed on a
dicritical structure determined by the sequence of blow-ups that desingularizes $\sep_{0}$, in which only the first blow-up
is dicritical.
We produce a logarithmic model with these data having rational residues.
\begin{figure}[h]
\begin{center}
\includegraphics[width=8cm]{Fig4.eps}
\put(-155,68){\scriptsize {$S_{1}$}}
\put(-177,82){\scriptsize {$S_{3}$}}
\put(-155,17){\scriptsize {$S_{2}$}}
\put(-177,3){\scriptsize {$S_{4}$}}
\put(-5,65){\scriptsize {$S_{1}$}}
\put(-25,82){\scriptsize {$S_{3}$}}
\put(-5,33){\scriptsize {$S_{2}$}}
\put(-23,14){\scriptsize {$S_{4}$}}
\put(-94,49){\scriptsize {$\sigma_{1}$}}
\put(-58,33){\scriptsize {$D_{1}$}}
\end{center}
\end{figure}
In order to zero the weighted balance of residues
at $0 \in \C^{2}$, we can take, for instance, $\lambda_{1} = \lambda_{3} = 1/10$ and $\lambda_{2} = \lambda_{4} = -1/10$. We then set
\[ \eta = \sum_{i=1}^{4} \lambda_{i} \frac{df_{i}}{f_{i}} = \frac{1}{10} \frac{dh}{h}, \quad \text{where}\ \quad h = \frac{f_{1} f_{3}}{f_{2} f_{4}}.\]
Denoting by $\sigma_{1}: (\tilde{\C^{2}},D_{1}) \to (\C^{2},0)$ the dicritical blow-up at $0 \in \C^{2}$ and considering coordinates
$(x,t)$ such that $y=tx$, at level 1, each $S_{i}$ has $x - (t-a_{i})^2 =0$ as equation.
Hence, in these coordinates, we get the escape function
\[ R_{\eta} = \frac{1}{10} \left( \frac{1}{t-1} + \frac{-1}{t+1} + \frac{1}{t-4} + \frac{-1}{t+4} \right) = \frac{t^2 -4}{(t^{2}-1)(t^{2}-16)}, \]
which has two affine zeroes, $t = 2$ and $t = -2$.
The curves of equations
\[4(t-2)^4-9(t-2)^3-8(t-2)^2x-18(t-2)^2+9(t-2)x+4x^2-18x = 0\]
and
\[-4(t+2)^4-9(t+2)^3+8(t+2)^2x+18(t+2)^2+9(t+2)x-4x^2+18x = 0\]
are invariant by $\eta_{1} = \sigma^{*} \eta$ and are tangent to the divisor $D_{1}$ at, respectively, $t=2$ and $t=-2$
(they correspond to the level curves $h= 1/81$ and $h=81$).
Their local branches at these points are, thus, escape separatrices.
Now, let us consider a set of parameters of the form $\tau = \{(2+a,\lambda),(-2-a,-\lambda)$, where $a,\lambda \in \R^{*}$ are small, to be determined
following the guidelines of the proof of Proposition \ref{prof-no-real-roots}. We consider the balanced symmetric logarithmic modification
with parameters $\tau$ given, for instance, by
\[\eta^{\tau} = \eta + \lambda \frac{d g_{1}}{g_{1}} - \lambda \frac{d g_{2}}{g_{2}}, \]
where $g_{1} = y - (2+a)x$ and $g_{2} = y + (2+a)x$.
Its escape function is
\[ R_{\eta^{\tau}} = R_{\eta} + \frac{\lambda}{t-(2+a)} + \frac{-\lambda}{t+(2+a)}= \frac{t^2 -4}{(t^{2}-1)(t^{2}-16)} + \frac{2 \lambda(2+a)}{t^{2}-(2+a)^{2}}
= \frac{ P_{\eta^{\tau}}}{ Q_{\eta^{\tau}}} ,\]
whose numerator is the biquadratic polynomial
\[ P_{\eta^{\tau}}(t) = (1 +2\lambda (2+a))t^4 +(-(2+a)^2 - 4 -34(2+a)\lambda)t^2 + (4(2+a)^2 + + 32 (2+a)\lambda).\]
The discriminant of the corresponding quadratic polynomial obtained by setting $t^{2} = z$ is
\[ \Delta = (2+a)^4 + 36 (2+a)^3 \lambda + 900 (2+a)^2 \lambda^2 -8(2+a)^2 -144(2+a)\lambda + 16 .\]
Taking, for instance, $a=1/10$ and $\lambda = -1/10$ we get $\Delta = -237215/10000 <0$.
This means that $P_{\eta^{\tau}}(t)$ has no real roots and, thus, $\eta^{\tau}$ has no real escape points.
Using these values, we have
$g_{1} = y - (2+1/10)x$ and $g_{2} = y + (2+1/10)x$, so that
\[\eta^{\tau} = \eta - \frac{1}{10} \frac{d g_{1}}{g_{1}} + \frac{1}{10} \frac{d g_{2}}{g_{2}} =
\frac{1}{10} \frac{dh_{1}}{h_{1}}, \quad \text{where} \quad h_{1} = h \frac{g_{2}}{g_{1}} = h = \frac{f_{1} f_{3}g_{2}}{f_{2} f_{4}g_{1}}.\]
We finish by inviting the reader to draw the figure corresponding to the sectorial decomposition of this example and also to try to devise other examples by himself.
}\end{example}
\bibliographystyle{plain}
|
\section{Introduction}
Open-domain question-answering (ODQA) has evolved into a core problem in Natural Language Processing (NLP), receiving growing interest from the research community \cite{raffel2020exploring, roberts2020much}. Despite the notoriously difficult challenge to correctly answer open-domain questions on arbitrary topics, recent advances of pre-trained language models (such as BERT \citep{devlin-etal-2019-bert}, BART \citep{lewis2020bart} and T5 \citep{raffel2020exploring}) have stimulated new research into additional, task-dependent pre-training steps. Specifically, recent publications show that in-domain pre-training regimes can improve models for several downstream tasks \citep{gururangan2020don}.
For open-domain question-answering, newly proposed pre-training tasks such as the Inverse Cloze Task (ICT) \citep{lee2019latent}, Body First Selection (BFS), Wiki Link Prediction (WLP) \citep{Chang2020Pre-training} and Question Answering Infused Pre-training (QUIP) \citep{jia2021question} show consistent improvements over baselines. However, most of these approaches still rely on either unlabeled text, or synthetically generated question-answer (QA) pairs. In this paper, we explore a second, somewhat orthogonal dimension to these lines of work, examining if a web-scale collection of natural QA pairs can support ODQA through in-domain pre-training.
Per definition, an ODQA system should be able to answer any question from an arbitrary domain. We believe that to approach this ability with in-domain pre-training, a suitable dataset should address the following 5 challenges: (1) Size; ODQA requires knowledge of a wide variety of topics. The underlying dataset used for in-domain pre-training hence needs to cover this abundance of domains, requiring a web-scale dataset. (2) Naturalness; While synthetic corpora can potentially capture a wide variety of language phenomena, to understand and generate truly natural language in all facets, synthetic datasets are not sufficient. (3) Quality; Given the requirement for a diverse, large-scale dataset, high data quality in terms of cleanliness and sensibility becomes a major challenge. Given that web-scale data sources require highly automated approaches operating on noisy data, assuring data quality is non-trivial. (4) Diversity; Besides size, another challenge for any ODQA in-domain pre-training dataset is the generality of the corpus. The dataset needs to support answering many diverse questions to allow models to learn general concepts. (5) Evaluation Fairness;
A web-scale question-answering dataset potentially overlaps with existing benchmark corpora, leading to inflated performance measures and impeding the evaluation fairness \citep{lewis2021question}.
To overcome these challenges, we propose a new large-scale dataset for open-domain question-answering called the \textbf{C}ommon \textbf{C}rawl \textbf{Q}uestion \textbf{A}nswering (CCQA) dataset.
Similar to popular datasets, such as C4 \citep{raffel2020exploring}, CCNet \citep{wenzek2020ccnet}, CC-100 \citep{conneau2020unsupervised}, HTLM \citep{aghajanyan2022htlm}, and CM3 \cite{aghajanyan2022cm3} we generate a large-scale, diverse and high-quality question-answering dataset from Common Crawl.
More specifically, Common Crawl allows us to obtain a large number of truly natural question-answer pairs, asked and answered by real humans on the web, rather than inferred through computational methods. Using the abundantly available \textit{schema.org} question annotation\footnote{\url{https://schema.org/Question}}, we generate question-answer pairs from explicit annotations, instead of heuristic rules, leading to high-quality data points.
In a large set of evaluations, we show that in-domain pre-training on our CCQA dataset achieves promising results across different settings, models and benchmarks. Using the rich information available on the web, we augment our dataset with additional data attributes beyond just question-answer pairs, such as votes, multiple (competing) answers, question summaries and intra-textual HTML markup, which can be used for a variety of tasks beyond question-answering in future work.
Furthermore, we evaluate the diversity and evaluation fairness of our dataset by computing topic distributions and train-test overlaps with benchmark datasets, providing additional rationale regarding the quality of our data and experiments. To summarize, our main contributions in this paper are as follows:
\begin{itemize}
\item We generate the first truly large-scale, natural question-answering dataset, containing around $130$ million unfiltered question-answer pairs ($55$M unique pairs), including about $60$ million English data points ($24$M unique pairs).
\item We present key dataset statistics, confirming the high quality of our question-answer pairs, the wide range of diverse topics and a low overlap with existing benchmarks.
\item We show the effectiveness of the dataset for in-domain pre-training by evaluating the performance of the unfiltered English subset on two question-answering tasks, three different settings, four models and five diverse benchmarks.
\end{itemize}
\begin{figure}
\centering
\includegraphics[width=.85\linewidth]{data_gen_overview_6.png}
\caption{Dataset generation overview from the initial raw HTML file (top) to general purpose, webpage aggregated question-answer pairs (bottom). M = Additional question/answer metadata. Red boxes = Non-question-answer related webpage components.}
\label{fig:dataset}
\end{figure}
\section{Related Work}
This work is inspired by a range of previous approaches using Common Crawl web-data, such as the Colossal Clean Crawled Corpus (C4) for language model pre-training \citep{raffel2020exploring}, the word/sentence representation generation corpus CCNet \citep{wenzek2020ccnet}, the CC-100 dataset for translation \citep{conneau2020unsupervised} and the markup-style language modelling HTLM corpus for zero-shot summarization \citep{aghajanyan2022htlm}. Despite all previously mentioned applications directly relying on large-scale web data from Common Crawl, their scope and application vary significantly. Compared to previously proposed datasets based on Common Crawl, we are the first to extract well-structured question-answer pairs with additional meta-data, making our corpus a valuable resource for ODQA research, and a multitude of related tasks, such as question summarization, answer rating, and answer ranking.
Further web-based datasets outside the Common Crawl domain are the TriviaQA \cite{joshi2017triviaqa} and ELI5 corpora \cite{fan2019eli5}, extracting small-scale question-answer datasets from Trivia websites and Reddit threads respectively. The large-scale GooAQ dataset \cite{khashabi2021gooaq} is similarly based on web data, however exploits the Google auto-complete feature and related answer boxes to generate semi-synthetic question-answer pairs. As a large-scale, completely synthetic dataset, the PAQ corpus \cite{lewis2021paq} automatically generates a large set of \textit{Probably Asked Questions} from Wikipedia articles. In contrast to these previously proposed datasets, our CCQA corpus presents a large-scale, natural and diverse question-answering resource in the same order of magnitude as the largest synthetic datasets.
Besides the generation of the CCQA dataset, we evaluate its potential as an in-domain pre-training corpus for open-domain question-answering. Our work is aligned with previous in-domain pre-training approaches, which have shown to improve a variety of downstream tasks \citep{gururangan2020don}. Similar to in-domain pre-training, multiple domain-dependent pre-training tasks have been proposed for open-domain question-answering. For example, \citet{lee2019latent} propose the Inverse Cloze Task (ICT), \citet{Chang2020Pre-training} introduce Body First Selection (BFS) and Wiki Link Prediction (WLP) and \citet{jia2021question} describe a novel Question Answering Infused Pre-training (QUIP) task. Along similar lines, \citet{aghajanyan2021muppet} propose pre-finetuning, an alternative to in-domain pre-training, using around $50$ domain-dependent datasets, showing that their MUPPET approach generalizes well
to many tasks. \citet{khashabi2020unifiedqa} introduce a similar concept for question-answering in their UnifiedQA framework. While we propose a somewhat orthogonal dimension to most of these works, they nevertheless present us with strong intuition regarding the effectiveness of domain-dependent pre-training.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{format_4.png}
\caption{JSON data structure following the schema.org annotation. Fasttext language labels \citep{joulin2016fasttext, joulin2017bag} added for language distinction.}
\label{fig:format}
\end{figure}
\section{The Common Crawl Question Answering (CCQA) Dataset}
\subsection{Dataset Collection}
\label{data_collection}
Our Common Crawl Question Answering (CCQA) dataset contains around $130$ million question-answer pairs ($55$M unique), extracted from 13 Common Crawl snapshots between May 2020 and May 2021\footnote{\url{https://commoncrawl.org/}}. A high-level overview of the dataset generation process is depicted in Figure~\ref{fig:dataset}. Starting from a set of raw HTML webpages, we make use of the schema.org definition to find relevant tags, such as the question, answer, author and votes (for the full set of tags see Figure~\ref{fig:format}). Using the explicit schema.org annotation (commonly used for search-engine optimization), instead of simple heuristics (e.g. question marks), we optimize the resulting corpus for high-quality data points. Specifically, due to the added efforts for website creators to define schema.org conforming meta-data, we believe that annotated question-answer pairs are likely to be relevant to the general public, mostly exclude rhetorical and contextual questions, and as a result constitute high quality QA data, despite the noisy nature of webpages.
During the dataset processing steps, we remove all HTML elements that do not contain valid schema.org markers (red in Figure~\ref{fig:dataset}) and subsequently clean every question on the webpage to only conserve markup related to the textual content of schema.org tags\footnote{Set of textual tags taken from \url{developer.mozilla.org/en-US/docs/Web/HTML/Element}}.
We further remove any unrelated markup attributes (e.g., CSS and JavaScript classes), before converting the content into a well-defined JSON object, shown in Figure~\ref{fig:format} and further described in section \ref{sec:format}.
Using the 13 consecutive Common Crawl snapshots, we generate an initial dataset of $130$ million question-answer pairs, naturally containing two types of potential duplicates: (1) Same-URL duplicates; where a webpage is updated between any two Common Crawl snapshots and (2) Content duplicates; where webpages from any Common Crawl snapshot contain same questions with potentially similar answers.
Here, we use the original, uncleaned version of the dataset, presenting a practical performance lower-bound, while leaving the exploration of additional filtering steps for future work\footnote{We provide de-duplication scripts for same-URL duplicates due to snapshot overlap at \url{https://github.com/facebookresearch/CCQA}.}.
Our dataset generation procedure is further outlined in Algorithm \ref{alg:algo_1}, found in Appendix \ref{app:algo}. For qualitative examples of our generated dataset format, we refer readers to Appendix~\ref{quali}.
\subsection{Dataset Format}
\label{sec:format}
The structured output of the dataset collection (shown in Figure~\ref{fig:format}), contains a three-level nested structure: (1) Every top-level data point represents a webpage in Common Crawl, encapsulating questions and answers found on the page, together with relevant metadata.
(2) On the second level, every question is represented as a tuple containing the question name (a short summary of the question) and question text (the main question). Questions also contain additional metadata as shown in Figure~\ref{fig:format}.
(3) Every question can contain an arbitrary number of associated answers and answer attempts, located on the third and final level of the nested structure. An answer thereby contains a mandatory accepted/suggested label, the answer text as well as optional metadata.
With this nested structure of our CCQA dataset, we allow users to verify question-answer pairs and their metadata on the original webpage, utilize additional parts of the web-document and allow future research to tackle question-answering related tasks, such as answer selection, answer rating or answer ranking.
\subsection{Dataset Dimensions}
\label{data_dimension}
To gain better insights into the massive amount of data, we present a mix of automatically obtained dataset dimensions, a small-scale human pilot study, and a set of key dataset distributions.
Regarding the small-scale human pilot study, we analyze a random subset of $400$ individual question-answer pairs and evaluate their sensibility and answerability. We define \textit{question sensibility} as to whether the annotator understands the questions itself, while
\textit{question answerability} refers to whether the question provides enough context for a perfect question-answering system to correctly answer the question. Furthermore, \emph{QA-sensibility} denotes if the question-answer pair makes sense\footnote{We do not check the answer for factual correctness but merely evaluate if it \emph{could} be the answer for the given question.}. We refer interested readers to Table~\ref{tab:sens_ans_detailed} in Appendix~\ref{app:sens_ans} for further explanations on sensibility/answerability.
\begin{table}[t]
\centering
\setlength{\tabcolsep}{0.28em}
\scalebox{0.88}{
\begin{tabular}{ccccc}
\toprule
Q-Sens$^H$ & Q-Ans$^H$ & QA-Sens$^H$ & Markup & Q-Summ \\
\midrule
96.5\% & 86\% & 82.25\% & 47.5\% & 11.7\% \\
\toprule
No A & Avg \#A$^*$ & Mean Q & Mean A & Lang Tags \\
\midrule
5.9\% & 1.41 & 43 & 57 & 77.9\% \\
\bottomrule
\end{tabular}
}
\caption{Key CCQA dataset dimensions. Q=Question, A=Answer, QA=\allowbreak Question-answer pair, Sens=\allowbreak Sensibility, Ans=\allowbreak Answerability, Lang=\allowbreak Language, Summ=\allowbreak Summarization, Mean=Average number of words, $^H$Human pilot study, $^*$Excluding questions without answers.}
\label{tab:data_dimensions}
\end{table}
As shown in Table~\ref{tab:data_dimensions}, our CCQA corpus contains nearly exclusively sensible questions, with the vast majority of them also answerable and sensible as a pair. To complement our small-scale human annotation, we further explore key dataset dimension, including the fraction of samples with advanced markup, questions containing both, question name and question text (as defined by the schema.org annotation), the number of questions without gold-answers, average question and answer length and the number of webpages with a valid language label, all indicating that the schema.org annotation highly correlates with carefully curated webpages.
Besides the key corpus-level statistics, we take a closer look at important dataset distributions in Table \ref{tab:data_distributions}. Specifically, we present the top 5 domains at the top of Table~\ref{tab:data_distributions}, showing the largest number of webpages originating from the \textit{stackexchange} domain, accounting for about $8\%$ of data points. Regarding the topical distribution of our dataset, we use the DMOZ/Curlie taxonomy, automatically extracting hierarchical topic information\footnote{\url{https://www.curlie.org}}. We randomly sample $1,000$ question webpages and show the top 5 topics in the second row of Table \ref{tab:data_distributions}. A more detailed topic distribution, also considering second-level assignments, can be found in Table~\ref{tab:data_distribution_detailed} in Appendix~\ref{app:topic_dist}.
Regarding the question-word distribution in our CCQA dataset, we observe that the majority of $36\%$ of question words are \textit{what} questions, followed by \textit{how}, \textit{when}, \textit{which} and \textit{where}. A full list of all 8 questions words and their relative appearance in our corpus can be found in Table \ref{tab:q_words_detailed} in Appendix \ref{app:q_dist}.
Lastly, expanding on the number of non-trivial markup tags presented in Table \ref{tab:data_dimensions}, we explore the frequency of HTML markup tags in our dataset in the last row in Table~\ref{tab:data_distributions}. For a list of the top-25 tags found in our corpus, we point interested readers to Table~\ref{tab:data_markup_detailed} in Appendix~\ref{app:html_dist}.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{high_level_closed_book.png}
\caption{High-level overview of the closed-book CCQA in-domain pre-training step (yellow) as part of the training pipeline for BART and T5. Language model pre-training shown in green. Task-dependent fine-tuning presented in red. Evaluation in blue. (1) Baseline pre-training/fine-tuning pipeline, (2) In-domain pre-training/fine-tuning pipeline, (3) zero-shot baseline setting and (4) zero-shot in-domain pre-training setting.}
\label{fig:high_level_closed_book}
\end{figure}
\begin{table*}[t]
\centering
\setlength{\tabcolsep}{0.2em}
\scalebox{0.9}{
\begin{tabular}{lrrrrr}
\toprule
Metric & \multicolumn{5}{c}{Top 5 Appearances in CCQA} \\
\midrule
Domains & stackexchange (07.78\%) & hotels (03.46\%) & viamichelin (02.51\%) & ccm (01.86\%) & vrbo (01.74\%) \\
Topics & Regional (38.90\%) & Society (21.10\%) & Business (08.30\%) & Sports (07.00\%) & Rec (06.20\%) \\
Q-words & What (36.20\%) & How (29.80\%) & When (09.68\%) & Which (09.64\%) & Where (06.04\%) \\
Markup & p (28.48\%) & a (14.89\%) & br (14.86\%) & li (10.04\%) & span (05.77\%) \\
\bottomrule
\end{tabular}}
\caption{CCQA dataset distribution for top 5 domains, topics according to the DMOZ/Curlie annotation, question words (Q-words, only computed on the English subset) and most common markup tags. \% for q-words and markup tags presents portion of all q-word/markup appearances. ccm=commentcamarche, Rec=Recreational.}
\label{tab:data_distributions}
\end{table*}
\section{Evaluation}
In this section, we showcase the value of our CCQA dataset with experiments on the closed-book question-answering (section \ref{cbqa}) and passage retrieval for open-book QA (section \ref{pass_ref}) tasks.
\subsection{Closed-Book Question-Answering}
\label{cbqa}
\subsubsection{Task}
The closed-book question-answering task challenges systems to answer questions without to use of additional information sources, such as knowledge bases or evidence documents. As a result, models are solely relying on the question text and the information stored inside the model weights during training. Here, we evaluate our new CCQA dataset as an in-domain pre-training corpus for this highly challenging task by converting the JSON representation into plain question-answer pairs, removing markup tags and additional metadata.
\subsubsection{Models \& Training}
Using the question-answer pairs from the CCQA dataset, we in-domain pre-train large language models for question-answering. We start with vanilla BART and T5 transformer models, shown on the left side (green) in Figure \ref{fig:high_level_closed_book}. We then further in-domain pre-train the models using a denoising or sequence-to-sequence (seq2seq) setup (yellow box in Figure~\ref{fig:high_level_closed_book}). For the denoising task, we follow the vanilla BART approach \cite{lewis2020bart}, using a concatenation of \textit{Q:$\mathbin\Vert$<question>$\mathbin\Vert$A:$\mathbin\Vert$<answer>} as the model input. For the seq2seq task, we train the model to predict the gold answer given a question as input.
With the additional in-domain pre-training step, a variety of training-flows emerge, shown as numbered circles in Figure~\ref{fig:high_level_closed_book}: \\
(1) Using a vanilla pre-trained language model to fine-tune on the benchmark dataset. \\
(2) Using the CCQA dataset for in-domain pre-training and subsequently fine-tune on the benchmark dataset. \\
(3) Using a pre-trained language model to directly infer answers on the benchmark dataset (zero-shot). \\
(4) Using the CCQA in-domain pre-trained model to directly infer answers on the benchmark dataset in zero-shot fashion.
\subsubsection{Datasets}
\label{datasets}
We evaluate the performance of our CCQA corpus as an in-domain pre-training dataset on 5 common benchmarks, based on 4 publicly available datasets in the closed-book setting:
\textbf{TriviaQA} (TQA) is a short-form, factoid-style question-answering dataset \cite{joshi2017triviaqa}. For the closed-book task, we ignore the available contexts and focus exclusively on question-answer pairs. Since the official test-split of the dataset is not publicly available, we use the official validation set as our test split and randomly sample a validation set from the training data, as commonly done in previous work \citep{roberts2020much}.
\textbf{Natural Questions} (NQ) \citep{kwiatkowski-etal-2019-natural} represents a popular corpus for question-answering research. Despite most recent work focusing on the \textbf{short-form} answers (NQ-Short), the NQ corpus also provides additional \textbf{long-form} answers (NQ-Long) for a large subset of questions. In this work, we use both, short, factoid answers and long-form responses.
\textbf{ELI5}, introduced by \citet{fan2019eli5}, constitutes the first large-scale long-form dataset for open-ended question-answering. We again do not take available evidence documents into account, but focus on the question-answer pairs only.
\textbf{GooAQ} \citep{khashabi2021gooaq} contains semi-automatically extracted question-answer pairs from the Google question auto-complete feature.
\begin{table*}[t]
\centering
\begin{threeparttable}
\setlength{\tabcolsep}{0.3em}
\centering
\scalebox{0.95}{
\begin{tabular}{lrrrrr}
\toprule
\multicolumn{6}{c}{\bf Zero-Shot}\\
\midrule
\multirow{2}{*}{Model} & \multicolumn{1}{r}{TQA} & \multicolumn{1}{r}{NQ-Short} & \multicolumn{1}{r}{NQ-Long} & \multicolumn{1}{r}{ELI5} & \multicolumn{1}{r}{GooAQ} \\
& \multicolumn{1}{r}{AR} & \multicolumn{1}{r}{AR} & \multicolumn{1}{r}{R-L} & \multicolumn{1}{r}{R-L} & \multicolumn{1}{r}{R-L}\\
\midrule
\multicolumn{6}{c}{BART Large}\\
\midrule
Rand. Init. & 0.04 & 0.11 & 0.10 & 0.26 & 0.16 \\
Vanilla & $^\dagger$4.91 & $^\dagger$1.93 & 10.39 & 11.88 & 14.67 \\
Vanilla$^a$ & & & & & \\
CCQA & \bf $^\dagger$5.14 & \bf $^\dagger$2.16 & \bf 12.18 & \bf $^\dagger$15.21 & \bf $^\dagger$17.5 \\
CCQA-d & 4.80 & 2.13 & 10.33 & 11.91 & 14.88 \\
\midrule \addlinespace
\multicolumn{6}{c}{T5 Small}\\
\midrule
Rand. Init. & 0.05 & 0.11 & 1.13 & 1.49 & 0.80 \\
Vanilla & $^\dagger$5.06 & $^\dagger$1.74 & 9.16 & 7.55 & $^\dagger$8.92 \\
Vanilla$^b$ & & & & & \\
CCQA & \bf $^\dagger$5.13 & \bf $^\dagger$1.86 & \bf $^\dagger$13.63 & \bf $^\dagger$15.28 & \bf $^\dagger$15.46 \\
\midrule \addlinespace
\multicolumn{6}{c}{T5 Base}\\
\midrule
Rand. Init. & 0.04 & 0.11 & 0.00 & 0.00 & 0.00 \\
Vanilla & $^\dagger$5.49 & $^\dagger$2.02 & $^\dagger$14.39 & 12.27 & $^\dagger$14.99 \\
Vanilla$^c$ & & & & & \\
CCQA & \bf $^\dagger$7.15 & \bf $^\dagger$3.19 & \bf $^\dagger$15.08 & \bf $^\dagger$15.69 & \bf $^\dagger$15.85 \\
\bottomrule
\end{tabular}}
\centering
\scalebox{0.95}{
\begin{tabular}{rrrrr}
\toprule
\multicolumn{5}{c}{\bf Fine-Tuned}\\
\midrule
\multicolumn{1}{r}{TQA} & \multicolumn{1}{r}{NQ-Short} & \multicolumn{1}{r}{NQ-Long} & \multicolumn{1}{r}{ELI5} & \multicolumn{1}{r}{GooAQ} \\
\multicolumn{1}{r}{EM} & \multicolumn{1}{r}{EM} & \multicolumn{1}{r}{R-L} & \multicolumn{1}{r}{R-L} & \multicolumn{1}{r}{R-L}\\
\midrule
\multicolumn{5}{c}{BART Large}\\
\midrule
0.71 & 0.75 & 16.04 & 14.37 & 16.21 \\
\bf 28.67 & 23.79 & 23.47 & 16.96 & 35.67 \\
& 26.50 & & & \\
25.82 & 22.91 & 21.25 & 17.23 & 32.53 \\
27.84 & \bf 23.96 & \bf 24.56 & \bf 17.27 & \bf 35.92 \\
\midrule \addlinespace
\multicolumn{5}{c}{T5 Small}\\
\midrule
0.44 & 0.54 & 10.86 & 13.06 & 8.71 \\
\bf 21.02 & \bf 21.16 & \bf 22.09 & 16.28 & 24.70 \\
& & 19.00 & 23.00 \\
17.55 & 19.50 & 22.05 & \bf 16.33 & \bf 25.35 \\
\midrule \addlinespace
\multicolumn{5}{c}{T5 Base}\\
\midrule
0.32 & 0.38 & 13.58 & 12.72 & 7.93 \\
\bf 26.25 & \bf 23.04 & \bf 25.36 & 16.58 & \bf 29.36 \\
23.63 & 25.94 & & & \\
22.69 & 22.32 & 24.73 & \bf 16.64 & 29.09 \\
\bottomrule
\end{tabular}
}
\begin{tablenotes}[para,raggedright]
\hbox{Results from
\item[a] \citet{lewis2021question}
\item[b] \citet{khashabi2021gooaq}
\item[c] \citet{roberts2020much}}
\end{tablenotes}
\caption{Closed-book zero-shot and fine-tuned results. Best performance of fairly computed results per sub-table \textbf{bold}. $^\dagger$Zero-shot model outperforms fully fine-tuned randomly initialized transformer of same architecture. \textit{-d} extension indicates denoising CCQA pre-training task. \textit{AR}=Answer-level recall, \textit{EM}=Exact Match, \textit{RL}=Rouge-L.
}
\label{tab:cb_results}
\end{threeparttable}
\end{table*}
\subsubsection{Metrics}
\label{metrics}
For datasets with short-form answers, we use the Exact Match (EM) metric for fine-tuned systems, in line with previous work by \citet{roberts2020much} and \citet{lewis2021paq}. While the EM metric works well for systems that are aware of the task-specific format, it punishes potentially correct answers with additional context, which we believe is overly harsh in zero-shot settings, where the specific output format is not known (e.g., training-flows (3) and (4)). Therefore, we propose using the Answer-level Recall (AR) metric for our zero-shot experiments, while limiting the answer length with the \textit{max-length} and \textit{length-penalty} inference parameters. As such, the AR metric requires the correct answer to be a continuous sub-sequence of the predicted tokens, while allowing for additional context. Since AR operates on token-level, the prediction of super/sub-words, e.g., \textit{\textbf{fun}damental} instead of \textit{\textbf{fun}}, is considered incorrect.
For long-form question-answer datasets, we choose the Rouge-L (RL) score as our evaluation metric, which has shown strong correlation with Rouge-1 and Rouge-2 scores, and is commonly used in previous work \cite{khashabi2021gooaq}.
\subsubsection{Hyper-Parameters}
We use the default parameters of the BART \citep{lewis2020bart} and T5 \citep{raffel2020exploring} models for in-domain pre-training and fine-tuning whenever possible. Regarding the in-domain pre-training on our CCQA dataset, we limit training to $800k$ steps using a batch-size of $1{,}024$. During our fine-tuning runs, we limit the number of updates to $20k$ steps with a batch-size of $256$ samples, with exception of the GooAQ dataset, which we fine-tune for $100k$ steps due to it's large size. We select the best model during our in-domain pre-training runs based on the perplexity measure, and pick the top fine-tuned model according to the final evaluation metric. We do not perform any hyper-parameter search during in-domain pre-training and fine-tuning.
For the inference step, our hyper-parameter setting is closely related to commonly used summarization parameters. We use a beam-size of 4, max-length of 140, and length-penalty of 2.0. For the fine-tuned short-form task, we choose a max-length of 30, following \citet{xiong2021answering} and a length-penalty of 1.0. All model evaluations are based on Huggingface Transformers\footnote{Experiments are executed on Nvidia V100 32GB GPUs.} \cite{wolf2019huggingface}.
\begin{figure*}[t]
\centering
\begin{tikzpicture}
\begin{axis}[
width=0.38*\textwidth,
height=5cm,
legend pos=south east,
legend style={font=\small},
ymajorgrids=true,
xmajorgrids=true,
bar width=1.5mm,
ymin=5.8, ymax=25.5,
xlabel={\bf NQ-Long},
symbolic x coords={$0$, $10^1$, $10^2$, $10^3$, $10^4$, $10^5$, full},
]
\addplot coordinates {
($0$, 9.16) ($10^1$, 15.06) ($10^2$, 15.4) ($10^3$, 16.12) ($10^4$, 17.32) ($10^5$, 21.49) (full, 22.09)
};
\addlegendentry{Vanilla}
\node at (axis cs:$0$,7.36){\tiny\textcolor{blue}{9.16}};
\node at (axis cs:$10^1$,13.06){\tiny\textcolor{blue}{15.06}};
\node at (axis cs:$10^2$,13.4){\tiny\textcolor{blue}{15.4}};
\node at (axis cs:$10^3$,14.12){\tiny\textcolor{blue}{16.12}};
\node at (axis cs:$10^4$,15.12){\tiny\textcolor{blue}{17.32}};
\node at (axis cs:$10^5$,19.49){\tiny\textcolor{blue}{21.49}};
\node at (axis cs:full,20.09){\tiny\textcolor{blue}{22.09}};
\addplot coordinates {
($0$, 13.63) ($10^1$, 15.31) ($10^2$, 16.06) ($10^3$, 16.47) ($10^4$, 16.76) ($10^5$, 21.39) (full, 22.05)
};
\addlegendentry{CCQA}
\node at (axis cs:$0$,15.63){\tiny\textcolor{red}{13.63}};
\node at (axis cs:$10^1$,17.31){\tiny\textcolor{red}{15.31}};
\node at (axis cs:$10^2$,18.06){\tiny\textcolor{red}{16.06}};
\node at (axis cs:$10^3$,18.47){\tiny\textcolor{red}{16.47}};
\node at (axis cs:$10^4$,18.99){\tiny\textcolor{red}{16.76}};
\node at (axis cs:$10^5$,23.39){\tiny\textcolor{red}{21.39}};
\node at (axis cs:full,24.05){\tiny\textcolor{red}{22.05}};
\end{axis}
\end{tikzpicture}
\begin{tikzpicture}
\begin{axis}[
width=0.38*\textwidth,
height=5cm,
legend style={font=\small},
legend pos=south east,
ymajorgrids=true,
xmajorgrids=true,
bar width=1.5mm,
ymin=5, ymax=29,
xlabel={\bf GooAQ},
symbolic x coords={$0$, $10^1$, $10^2$, $10^3$, $10^4$, $10^5$, full},
]
\addplot coordinates {
($0$, 8.92) ($10^1$, 16.07) ($10^2$, 16.63) ($10^3$, 18.13) ($10^4$, 19.15) ($10^5$, 24.77) (full, 24.7)
};
\addlegendentry{Vanilla}
\node at (axis cs:$0$,6.92){\tiny\textcolor{blue}{8.92}};
\node at (axis cs:$10^1$,14.07){\tiny\textcolor{blue}{16.07}};
\node at (axis cs:$10^2$,14.63){\tiny\textcolor{blue}{16.63}};
\node at (axis cs:$10^3$,16.13){\tiny\textcolor{blue}{18.13}};
\node at (axis cs:$10^4$,17.15){\tiny\textcolor{blue}{19.15}};
\node at (axis cs:$10^5$,22.77){\tiny\textcolor{blue}{24.77}};
\node at (axis cs:full,22.7){\tiny\textcolor{blue}{24.7}};
\addplot coordinates {
($0$, 15.46) ($10^1$, 17.12) ($10^2$, 17.33) ($10^3$, 18.44) ($10^4$, 19.11) ($10^5$, 24.83) (full, 25.35)
};
\addlegendentry{CCQA}
\node at (axis cs:$0$,17.76){\tiny\textcolor{red}{15.46}};
\node at (axis cs:$10^1$,19.12){\tiny\textcolor{red}{17.12}};
\node at (axis cs:$10^2$,19.33){\tiny\textcolor{red}{17.33}};
\node at (axis cs:$10^3$,20.44){\tiny\textcolor{red}{18.44}};
\node at (axis cs:$10^4$,21.11){\tiny\textcolor{red}{19.11}};
\node at (axis cs:$10^5$,26.83){\tiny\textcolor{red}{24.83}};
\node at (axis cs:full,27.35){\tiny\textcolor{red}{25.35}};
\end{axis}
\end{tikzpicture}
\begin{tikzpicture}
\begin{axis}[
width=0.38*\textwidth,
height=5cm,
legend pos=south east,
legend style={font=\small},
ymajorgrids=true,
xmajorgrids=true,
bar width=1.5mm,
ymin=5.5, ymax=18.5,
xlabel={\bf ELI5},
symbolic x coords={$0$, $10^1$, $10^2$, $10^3$, $10^4$, $10^5$, full},
]
\addplot coordinates {
($0$, 7.55) ($10^1$, 13.99) ($10^2$, 15.16) ($10^3$, 15.55) ($10^4$, 16.11) ($10^5$, 16.21) (full, 16.28)
};
\addlegendentry{Vanilla}
\node at (axis cs:$0$,6.55){\tiny\textcolor{blue}{7.55}};
\node at (axis cs:$10^1$,13.1){\tiny\textcolor{blue}{13.99}};
\node at (axis cs:$10^2$,14.2){\tiny\textcolor{blue}{15.16}};
\node at (axis cs:$10^3$,14.6){\tiny\textcolor{blue}{15.55}};
\node at (axis cs:$10^4$,15.1){\tiny\textcolor{blue}{16.11}};
\node at (axis cs:$10^5$,15.2){\tiny\textcolor{blue}{16.21}};
\node at (axis cs:full,15.3){\tiny\textcolor{blue}{16.28}};
\addplot coordinates {
($0$, 15.28) ($10^1$, 15.5) ($10^2$, 15.80) ($10^3$, 15.94) ($10^4$, 16.12) ($10^5$, 16.24) (full, 16.33)
};
\addlegendentry{CCQA}
\node at (axis cs:$0$,16.28){\tiny\textcolor{red}{15.28}};
\node at (axis cs:$10^1$,16.5){\tiny\textcolor{red}{15.50}};
\node at (axis cs:$10^2$,16.8){\tiny\textcolor{red}{15.80}};
\node at (axis cs:$10^3$,16.94){\tiny\textcolor{red}{15.94}};
\node at (axis cs:$10^4$,17.12){\tiny\textcolor{red}{16.12}};
\node at (axis cs:$10^5$,17.24){\tiny\textcolor{red}{16.24}};
\node at (axis cs:full,17.33){\tiny\textcolor{red}{16.33}};
\end{axis}
\end{tikzpicture}
\caption{Low resource experiments comparing the Rouge-L score of vanilla T5 Small with our CCQA pre-trained models on NQ-long (left), GooAQ (center) and ELI5 (right).}
\label{fig:low_resource}
\end{figure*}
\subsubsection{Results}
\label{closed_book}
Our main results for the closed-book question-answering task are presented in Table \ref{tab:cb_results}, showing the zero-shot and fine-tuned performance of the BART Large (top), T5 Small (center) and T5 Base (bottom) models for each of the 5 evaluation datasets. Even though we present a wide variety of benchmark results, from short-form factoid questions to long-form answers, the CCQA seq2seq pre-trained model consistently outperforms all other models on the zero-shot question-answering task. Even more importantly, the additional in-domain pre-training step achieves better zero-shot performance than fully fine-tuned, randomly initialized transformer models (as extensively used prior to 2018) in almost all settings. Specifically, our model outperforms the randomly initialized transformers on all benchmarks for T5 Small and T5 Base, as well as on 4 out of 5 datasets using BART Large.
Comparing the fully fine-tuned setting across models and datasets it becomes clear that, although oftentimes performing comparably, our CCQA seq2seq pre-trained model underperforms the vanilla models in most cases. Seq2seq in-domain pre-training on CCQA only reaches superior performance on the ELI5 dataset for all models, as well as on the GooAQ dataset for T5 Small.
Showing that seq2seq pre-training on CCQA is effective in zero-shot scenarios, however only partially improves over baselines in the fine-tuned setting, we investigate: (1) Additional experiments using the CCQA dataset for denoising-style pre-training (\textit{-d} in Table \ref{tab:cb_results}) and (2) Evaluate additional low-resource scenarios, shown in Figure \ref{fig:low_resource}.
For our denoising-style in-domain pre-training experiment, we keep the available markup information, in line with HTLM \cite{aghajanyan2022htlm}.
As shown in Table~\ref{tab:cb_results}, the in-domain CCQA denoising objective outperforms the vanilla BART Large model on 4 out of 5 benchmarks in the fine-tuned setting. We believe that this result, alongside the zero-shot performance of the seq2seq CCQA model, clearly shows the usefulness and generality of our CCQA corpus for closed-book open-domain question-answering.
Taking a closer look at low-resource scenarios, we evaluate the vanilla T5 Small model against our in-domain pre-trained approach using 5 proper subsets of the NQ-Long, GooAQ and ELI5 benchmark datasets, drawn at random. As presented in Figure \ref{fig:low_resource}, our CCQA model mostly outperforms the vanilla T5 Small model in low-resource scenarios with up to $10{,}000$ data points. While the performance of our CCQA model is consistently better on the ELI5 test-set, the vanilla baselines outperform our models fastest on the NQ-Long corpus. Additional low-resource experiments on T5 Base are shown in Table \ref{fig:app_low_resource}, in Appendix \ref{app_full_res}.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{high_level_retrieval.png}
\caption{High-level overview of the CCQA passage retrieval in-domain pre-training step (yellow) as part of the training pipeline for DPR. Language model pre-training shown in green. Task-dependent fine-tuning presented in red. Evaluation in blue. (1) Baseline pre-training/fine-tuning pipeline, (2) In-domain pre-training/fine-tuning pipeline.}
\label{fig:high_level_retrieval}
\vspace{-3mm}
\end{figure}
\subsection{Passage Retrieval}
\label{pass_ref}
\subsubsection{Task}
For the passage retrieval task, an important component of most open-book QA systems (e.g., \citet{NEURIPS2020_6b493230, izacard2021leveraging}), models aim to extract a set of evidence passages from a large collection of documents through conditional ranking.
To align our corpus with the passage retrieval task, we aggregate every question into a single data point, consisting of the question itself, alongside all available answers as either positive or negative contexts. If available, answer votes are used as a proxy to determine positive and negative (sometimes called ``hard-negative") contexts. Following the practice in \citet{fan2019eli5}, we assign every answer with at least 2 more upvotes than downvotes as a positive context and all other answer as negative. If answer votes are not available, we use the accepted/suggested label (shown in Figure~\ref{fig:format}) as an indicator for positive and negative contexts. In the absence of either criterion, we use all available answers as positive contexts.
\subsubsection{Models \& Training}
For passage retrieval, we choose the Dense Passage Retriever (DPR) \cite{karpukhin2020dense}, used in a variety of popular end-to-end open-book QA models, such as RAG \citep{NEURIPS2020_6b493230} and FiD \citep{izacard2021leveraging}. As shown in Figure~\ref{fig:high_level_retrieval}, we start with the vanilla DPR model based on BERT \cite{devlin-etal-2019-bert} and in-domain pre-train using questions and positive/negative passages from the CCQA dataset (yellow box in Figure~\ref{fig:high_level_retrieval}), similar to \citet{ouguz2021domain}. In line with the training-flows of the closed-book models, we train DPR using either the vanilla setup (pre-training $\rightarrow$ fine-tuning) or the in-domain pre-training approach (pre-training $\rightarrow$ in-domain pre-training $\rightarrow$ fine-tuning), shown as circles (1) and (2) in Figure~\ref{fig:high_level_retrieval}, respectively.
\subsubsection{Datasets \& Metrics}
Following the original DPR paper \citep{karpukhin2020dense}, we evaluate the passage retrieval task on the NQ-Short and TQA datasets presented in section \ref{datasets}, using the top-20 and top-100 retrieval accuracy (Acc@20/Acc@100) measures.
\begin{table}[t]
\centering
\setlength{\tabcolsep}{0.18em}
\scalebox{0.9}{
\begin{tabular}{lrrrr}
\toprule
\multirow{2}{*}{Model} & \multicolumn{2}{c}{TQA} & \multicolumn{2}{c}{NQ-Short} \\
& \multicolumn{1}{c}{Acc@20} & \multicolumn{1}{c}{Acc@100} & \multicolumn{1}{c}{Acc@20} & \multicolumn{1}{c}{Acc@100} \\
\midrule
DPR & 79.4 & 85.0 & 78.4 & 85.4 \\
DPR v2 & 79.5 & 85.3 & 78.3 & 85.6 \\
CCQA DPR & \textbf{80.0} & \textbf{85.6} & \textbf{79.1} & \textbf{86.3} \\
\bottomrule
\end{tabular}}
\caption{Fine-tuned Dense Passage Retriever (DPR) accuracy measure on the TQA and NQ-Short datasets. DPR represents the original DPR model \cite{karpukhin2020dense}, DPR v2 \citep{ouguz2021domain} indicates the updated codebase. CCQA DPR uses our CCQA pre-trained DPR model for retrieval fine-tuning.}
\label{tab:ob_results}
\end{table}
\begin{table}
\centering
\setlength{\tabcolsep}{0.28em}
\scalebox{0.9}{
\begin{tabular}{lrrrrr}
\toprule
Bench. (test) & TQA & NQ-S & NQ-L & ELI5 & GooAQ \\
\midrule
Bench. (train) & 11.9 & 4.9 & 5.2 & 3.0 & 26.9 \\
CCQA (train) & 0.4 & 1.9 & 2.3 & 0.5 & 26.9 \\
\bottomrule
\end{tabular}}
\caption{8-gram question overlap (in \%) between training sets and benchmark test-sets (inspired by \citet{radford2019language}). \textit{Bench (train)} refers to the overlap between the respective training- and test-portion of the benchmark datasets, \textit{CCQA (train)} identified overlaps between our dataset and the test-splits. False positive rate upper-bound by $\frac{1}{10^8}$. All inputs are normalized and lower-cased. NQ-S=\allowbreak NQ-Short, NQ-L=\allowbreak NQ-Long.}
\label{tab:bloom}
\vspace{-3mm}
\end{table}
\subsubsection{Hyper-Parameters}
We use the default DPR hyper-parameters whenever possible \citep{karpukhin2020dense}. For in-domain pre-training, we limit training to $800k$ steps using a batch-size of $1{,}536$ samples. During fine-tuning, we restrict the number of updates to $20k$ steps with a batch-size of $128$.
The best checkpoint is selected based on the Mean Reciprocal Rank (MRR) measure, following \citet{ouguz2021domain}. We do not perform any hyper-parameter search.
\subsubsection{Results}
\label{open_book}
For the passage retrieval experiments, we compare our CCQA in-domain pre-trained DPR model against the vanilla DPR model published in \citet{karpukhin2020dense}, as well as the recently enhanced version \citep{ouguz2021domain}. Table \ref{tab:ob_results} contains our empirical results, showing consistent improvements of our CCQA DPR model over the vanilla baselines. More specifically, the in-domain CCQA pre-training step increases the top-20 and top-100 accuracy score on the TQA benchmark dataset by over half a point, while the performance gap on NQ-Short shows consistent improvement of over $0.7\%$.
\subsection{Evaluation Fairness: Dataset Overlap}
\label{overlap}
With modern pre-training approaches using increasingly large datasets, accidental overlaps between pre-training corpora and benchmark datasets become more and more common \citep{lewis2021question}. To analyze this threat to the integrity of our dataset and empirical analysis, we follow \citet{radford2019language} and evaluate the 8-gram question overlap of our CCQA training portion with the test-split of benchmark datasets using bloom filters. Table \ref{tab:bloom} shows a consistently smaller question overlap between CCQA and the benchmark test set, compared to the benchmark training split itself.
\section{Conclusion and Future Work}
In this work, we presented our new web-scale CCQA dataset for in-domain model pre-training. Orthogonal to recent efforts on improving task-specific pre-training objectives, we show our dataset generation process, followed by detailed insights into key corpus dimensions of this new, large-scale, natural, and diverse question-answering dataset. In a set of empirical evaluations, we confirm the initial intuition that the corpus presents a valuable resource for open-domain question-answering research. In our zero-shot, low-resource and fine-tuned experiments for open- and closed-book QA tasks, we show promising results across multiple model architectures.
With around $130$ million question-answer pairs ($55$M unique) as well as additional meta-data, our CCQA dataset presents a versatile source of information, which has a large variety of applications in future work (e.g., question summarization, answer rating, answer ranking and many more).
\section{Ethical Considerations}
We now discuss the three major ethical considerations impacting this paper:
\paragraph{Hate-speech, Harmful Gender and Racial Biases:}
With general web-data potentially containing hate-speech and harmful gender and racial biases, we believe that our extracted dataset based on the schema.org annotations is less impacted by these issues, with the schema.org annotation representing a good proxy for high-quality, professionally curated websites. As a result, we believe that the severity of this issue is significantly reduced. Furthermore, in our human evaluation, we find no signs of the above mentioned biases. We leave computational approaches to determine dataset biases for future work (e.g., the Word Embedding Association Test \citep{Caliskan2017SemanticsDA} and Sentence Encoder Association Test \citep{may2019measuring}).
\paragraph{Data Availability:}
We do not directly provide the CCQA dataset, but enable third parties to generate the corpus through our published dataset generation scripts available at \url{https://github.com/facebookresearch/CCQA}.
\paragraph{Hallucinations and Factual Errors:}
As shown in the evaluation section, our model is able to generate reasonable answers for factoid and long-form questions. The inferred answers are fluent and human-like, but may also contain hallucinations and factual errors, especially for the challenging closed-book question-answering task. Without a guarantee of the predicted answers being factually correct, they can potentially spread misinformation if not properly corrected.
\subsubsection*{Acknowledgments}
We want to thank the anonymous reviewers for their valuable feedback and acknowledge our colleagues and peers for their help throughout this project, specifically Anchit Gupta, Gagan Aneja and Patrick Lewis for their valuable input on important decisions.
\section{Introduction}
Open-domain question-answering (ODQA) has evolved into a core problem in Natural Language Processing (NLP), receiving growing interest from the research community \cite{raffel2020exploring, roberts2020much}. Despite the notoriously difficult challenge to correctly answer open-domain questions on arbitrary topics, recent advances of pre-trained language models (such as BERT \citep{devlin-etal-2019-bert}, BART \citep{lewis2020bart} and T5 \citep{raffel2020exploring}) have stimulated new research into additional, task-dependent pre-training steps. Specifically, recent publications show that in-domain pre-training regimes can improve models for several downstream tasks \citep{gururangan2020don}.
For open-domain question-answering, newly proposed pre-training tasks such as the Inverse Cloze Task (ICT) \citep{lee2019latent}, Body First Selection (BFS), Wiki Link Prediction (WLP) \citep{Chang2020Pre-training} and Question Answering Infused Pre-training (QUIP) \citep{jia2021question} show consistent improvements over baselines. However, most of these approaches still rely on either unlabeled text, or synthetically generated question-answer (QA) pairs. In this paper, we explore a second, somewhat orthogonal dimension to these lines of work, examining if a web-scale collection of natural QA pairs can support ODQA through in-domain pre-training.
Per definition, an ODQA system should be able to answer any question from an arbitrary domain. We believe that to approach this ability with in-domain pre-training, a suitable dataset should address the following 5 challenges: (1) Size; ODQA requires knowledge of a wide variety of topics. The underlying dataset used for in-domain pre-training hence needs to cover this abundance of domains, requiring a web-scale dataset. (2) Naturalness; While synthetic corpora can potentially capture a wide variety of language phenomena, to understand and generate truly natural language in all facets, synthetic datasets are not sufficient. (3) Quality; Given the requirement for a diverse, large-scale dataset, high data quality in terms of cleanliness and sensibility becomes a major challenge. Given that web-scale data sources require highly automated approaches operating on noisy data, assuring data quality is non-trivial. (4) Diversity; Besides size, another challenge for any ODQA in-domain pre-training dataset is the generality of the corpus. The dataset needs to support answering many diverse questions to allow models to learn general concepts. (5) Evaluation Fairness;
A web-scale question-answering dataset potentially overlaps with existing benchmark corpora, leading to inflated performance measures and impeding the evaluation fairness \citep{lewis2021question}.
To overcome these challenges, we propose a new large-scale dataset for open-domain question-answering called the \textbf{C}ommon \textbf{C}rawl \textbf{Q}uestion \textbf{A}nswering (CCQA) dataset.
Similar to popular datasets, such as C4 \citep{raffel2020exploring}, CCNet \citep{wenzek2020ccnet}, CC-100 \citep{conneau2020unsupervised}, HTLM \citep{aghajanyan2022htlm}, and CM3 \cite{aghajanyan2022cm3} we generate a large-scale, diverse and high-quality question-answering dataset from Common Crawl.
More specifically, Common Crawl allows us to obtain a large number of truly natural question-answer pairs, asked and answered by real humans on the web, rather than inferred through computational methods. Using the abundantly available \textit{schema.org} question annotation\footnote{\url{https://schema.org/Question}}, we generate question-answer pairs from explicit annotations, instead of heuristic rules, leading to high-quality data points.
In a large set of evaluations, we show that in-domain pre-training on our CCQA dataset achieves promising results across different settings, models and benchmarks. Using the rich information available on the web, we augment our dataset with additional data attributes beyond just question-answer pairs, such as votes, multiple (competing) answers, question summaries and intra-textual HTML markup, which can be used for a variety of tasks beyond question-answering in future work.
Furthermore, we evaluate the diversity and evaluation fairness of our dataset by computing topic distributions and train-test overlaps with benchmark datasets, providing additional rationale regarding the quality of our data and experiments. To summarize, our main contributions in this paper are as follows:
\begin{itemize}
\item We generate the first truly large-scale, natural question-answering dataset, containing around $130$ million unfiltered question-answer pairs ($55$M unique pairs), including about $60$ million English data points ($24$M unique pairs).
\item We present key dataset statistics, confirming the high quality of our question-answer pairs, the wide range of diverse topics and a low overlap with existing benchmarks.
\item We show the effectiveness of the dataset for in-domain pre-training by evaluating the performance of the unfiltered English subset on two question-answering tasks, three different settings, four models and five diverse benchmarks.
\end{itemize}
\begin{figure}
\centering
\includegraphics[width=.85\linewidth]{data_gen_overview_6.png}
\caption{Dataset generation overview from the initial raw HTML file (top) to general purpose, webpage aggregated question-answer pairs (bottom). M = Additional question/answer metadata. Red boxes = Non-question-answer related webpage components.}
\label{fig:dataset}
\end{figure}
\section{Related Work}
This work is inspired by a range of previous approaches using Common Crawl web-data, such as the Colossal Clean Crawled Corpus (C4) for language model pre-training \citep{raffel2020exploring}, the word/sentence representation generation corpus CCNet \citep{wenzek2020ccnet}, the CC-100 dataset for translation \citep{conneau2020unsupervised} and the markup-style language modelling HTLM corpus for zero-shot summarization \citep{aghajanyan2022htlm}. Despite all previously mentioned applications directly relying on large-scale web data from Common Crawl, their scope and application vary significantly. Compared to previously proposed datasets based on Common Crawl, we are the first to extract well-structured question-answer pairs with additional meta-data, making our corpus a valuable resource for ODQA research, and a multitude of related tasks, such as question summarization, answer rating, and answer ranking.
Further web-based datasets outside the Common Crawl domain are the TriviaQA \cite{joshi2017triviaqa} and ELI5 corpora \cite{fan2019eli5}, extracting small-scale question-answer datasets from Trivia websites and Reddit threads respectively. The large-scale GooAQ dataset \cite{khashabi2021gooaq} is similarly based on web data, however exploits the Google auto-complete feature and related answer boxes to generate semi-synthetic question-answer pairs. As a large-scale, completely synthetic dataset, the PAQ corpus \cite{lewis2021paq} automatically generates a large set of \textit{Probably Asked Questions} from Wikipedia articles. In contrast to these previously proposed datasets, our CCQA corpus presents a large-scale, natural and diverse question-answering resource in the same order of magnitude as the largest synthetic datasets.
Besides the generation of the CCQA dataset, we evaluate its potential as an in-domain pre-training corpus for open-domain question-answering. Our work is aligned with previous in-domain pre-training approaches, which have shown to improve a variety of downstream tasks \citep{gururangan2020don}. Similar to in-domain pre-training, multiple domain-dependent pre-training tasks have been proposed for open-domain question-answering. For example, \citet{lee2019latent} propose the Inverse Cloze Task (ICT), \citet{Chang2020Pre-training} introduce Body First Selection (BFS) and Wiki Link Prediction (WLP) and \citet{jia2021question} describe a novel Question Answering Infused Pre-training (QUIP) task. Along similar lines, \citet{aghajanyan2021muppet} propose pre-finetuning, an alternative to in-domain pre-training, using around $50$ domain-dependent datasets, showing that their MUPPET approach generalizes well
to many tasks. \citet{khashabi2020unifiedqa} introduce a similar concept for question-answering in their UnifiedQA framework. While we propose a somewhat orthogonal dimension to most of these works, they nevertheless present us with strong intuition regarding the effectiveness of domain-dependent pre-training.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{format_4.png}
\caption{JSON data structure following the schema.org annotation. Fasttext language labels \citep{joulin2016fasttext, joulin2017bag} added for language distinction.}
\label{fig:format}
\end{figure}
\section{The Common Crawl Question Answering (CCQA) Dataset}
\subsection{Dataset Collection}
\label{data_collection}
Our Common Crawl Question Answering (CCQA) dataset contains around $130$ million question-answer pairs ($55$M unique), extracted from 13 Common Crawl snapshots between May 2020 and May 2021\footnote{\url{https://commoncrawl.org/}}. A high-level overview of the dataset generation process is depicted in Figure~\ref{fig:dataset}. Starting from a set of raw HTML webpages, we make use of the schema.org definition to find relevant tags, such as the question, answer, author and votes (for the full set of tags see Figure~\ref{fig:format}). Using the explicit schema.org annotation (commonly used for search-engine optimization), instead of simple heuristics (e.g. question marks), we optimize the resulting corpus for high-quality data points. Specifically, due to the added efforts for website creators to define schema.org conforming meta-data, we believe that annotated question-answer pairs are likely to be relevant to the general public, mostly exclude rhetorical and contextual questions, and as a result constitute high quality QA data, despite the noisy nature of webpages.
During the dataset processing steps, we remove all HTML elements that do not contain valid schema.org markers (red in Figure~\ref{fig:dataset}) and subsequently clean every question on the webpage to only conserve markup related to the textual content of schema.org tags\footnote{Set of textual tags taken from \url{developer.mozilla.org/en-US/docs/Web/HTML/Element}}.
We further remove any unrelated markup attributes (e.g., CSS and JavaScript classes), before converting the content into a well-defined JSON object, shown in Figure~\ref{fig:format} and further described in section \ref{sec:format}.
Using the 13 consecutive Common Crawl snapshots, we generate an initial dataset of $130$ million question-answer pairs, naturally containing two types of potential duplicates: (1) Same-URL duplicates; where a webpage is updated between any two Common Crawl snapshots and (2) Content duplicates; where webpages from any Common Crawl snapshot contain same questions with potentially similar answers.
Here, we use the original, uncleaned version of the dataset, presenting a practical performance lower-bound, while leaving the exploration of additional filtering steps for future work\footnote{We provide de-duplication scripts for same-URL duplicates due to snapshot overlap at \url{https://github.com/facebookresearch/CCQA}.}.
Our dataset generation procedure is further outlined in Algorithm \ref{alg:algo_1}, found in Appendix \ref{app:algo}. For qualitative examples of our generated dataset format, we refer readers to Appendix~\ref{quali}.
\subsection{Dataset Format}
\label{sec:format}
The structured output of the dataset collection (shown in Figure~\ref{fig:format}), contains a three-level nested structure: (1) Every top-level data point represents a webpage in Common Crawl, encapsulating questions and answers found on the page, together with relevant metadata.
(2) On the second level, every question is represented as a tuple containing the question name (a short summary of the question) and question text (the main question). Questions also contain additional metadata as shown in Figure~\ref{fig:format}.
(3) Every question can contain an arbitrary number of associated answers and answer attempts, located on the third and final level of the nested structure. An answer thereby contains a mandatory accepted/suggested label, the answer text as well as optional metadata.
With this nested structure of our CCQA dataset, we allow users to verify question-answer pairs and their metadata on the original webpage, utilize additional parts of the web-document and allow future research to tackle question-answering related tasks, such as answer selection, answer rating or answer ranking.
\subsection{Dataset Dimensions}
\label{data_dimension}
To gain better insights into the massive amount of data, we present a mix of automatically obtained dataset dimensions, a small-scale human pilot study, and a set of key dataset distributions.
Regarding the small-scale human pilot study, we analyze a random subset of $400$ individual question-answer pairs and evaluate their sensibility and answerability. We define \textit{question sensibility} as to whether the annotator understands the questions itself, while
\textit{question answerability} refers to whether the question provides enough context for a perfect question-answering system to correctly answer the question. Furthermore, \emph{QA-sensibility} denotes if the question-answer pair makes sense\footnote{We do not check the answer for factual correctness but merely evaluate if it \emph{could} be the answer for the given question.}. We refer interested readers to Table~\ref{tab:sens_ans_detailed} in Appendix~\ref{app:sens_ans} for further explanations on sensibility/answerability.
\begin{table}[t]
\centering
\setlength{\tabcolsep}{0.28em}
\scalebox{0.88}{
\begin{tabular}{ccccc}
\toprule
Q-Sens$^H$ & Q-Ans$^H$ & QA-Sens$^H$ & Markup & Q-Summ \\
\midrule
96.5\% & 86\% & 82.25\% & 47.5\% & 11.7\% \\
\toprule
No A & Avg \#A$^*$ & Mean Q & Mean A & Lang Tags \\
\midrule
5.9\% & 1.41 & 43 & 57 & 77.9\% \\
\bottomrule
\end{tabular}
}
\caption{Key CCQA dataset dimensions. Q=Question, A=Answer, QA=\allowbreak Question-answer pair, Sens=\allowbreak Sensibility, Ans=\allowbreak Answerability, Lang=\allowbreak Language, Summ=\allowbreak Summarization, Mean=Average number of words, $^H$Human pilot study, $^*$Excluding questions without answers.}
\label{tab:data_dimensions}
\end{table}
As shown in Table~\ref{tab:data_dimensions}, our CCQA corpus contains nearly exclusively sensible questions, with the vast majority of them also answerable and sensible as a pair. To complement our small-scale human annotation, we further explore key dataset dimension, including the fraction of samples with advanced markup, questions containing both, question name and question text (as defined by the schema.org annotation), the number of questions without gold-answers, average question and answer length and the number of webpages with a valid language label, all indicating that the schema.org annotation highly correlates with carefully curated webpages.
Besides the key corpus-level statistics, we take a closer look at important dataset distributions in Table \ref{tab:data_distributions}. Specifically, we present the top 5 domains at the top of Table~\ref{tab:data_distributions}, showing the largest number of webpages originating from the \textit{stackexchange} domain, accounting for about $8\%$ of data points. Regarding the topical distribution of our dataset, we use the DMOZ/Curlie taxonomy, automatically extracting hierarchical topic information\footnote{\url{https://www.curlie.org}}. We randomly sample $1,000$ question webpages and show the top 5 topics in the second row of Table \ref{tab:data_distributions}. A more detailed topic distribution, also considering second-level assignments, can be found in Table~\ref{tab:data_distribution_detailed} in Appendix~\ref{app:topic_dist}.
Regarding the question-word distribution in our CCQA dataset, we observe that the majority of $36\%$ of question words are \textit{what} questions, followed by \textit{how}, \textit{when}, \textit{which} and \textit{where}. A full list of all 8 questions words and their relative appearance in our corpus can be found in Table \ref{tab:q_words_detailed} in Appendix \ref{app:q_dist}.
Lastly, expanding on the number of non-trivial markup tags presented in Table \ref{tab:data_dimensions}, we explore the frequency of HTML markup tags in our dataset in the last row in Table~\ref{tab:data_distributions}. For a list of the top-25 tags found in our corpus, we point interested readers to Table~\ref{tab:data_markup_detailed} in Appendix~\ref{app:html_dist}.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{high_level_closed_book.png}
\caption{High-level overview of the closed-book CCQA in-domain pre-training step (yellow) as part of the training pipeline for BART and T5. Language model pre-training shown in green. Task-dependent fine-tuning presented in red. Evaluation in blue. (1) Baseline pre-training/fine-tuning pipeline, (2) In-domain pre-training/fine-tuning pipeline, (3) zero-shot baseline setting and (4) zero-shot in-domain pre-training setting.}
\label{fig:high_level_closed_book}
\end{figure}
\begin{table*}[t]
\centering
\setlength{\tabcolsep}{0.2em}
\scalebox{0.9}{
\begin{tabular}{lrrrrr}
\toprule
Metric & \multicolumn{5}{c}{Top 5 Appearances in CCQA} \\
\midrule
Domains & stackexchange (07.78\%) & hotels (03.46\%) & viamichelin (02.51\%) & ccm (01.86\%) & vrbo (01.74\%) \\
Topics & Regional (38.90\%) & Society (21.10\%) & Business (08.30\%) & Sports (07.00\%) & Rec (06.20\%) \\
Q-words & What (36.20\%) & How (29.80\%) & When (09.68\%) & Which (09.64\%) & Where (06.04\%) \\
Markup & p (28.48\%) & a (14.89\%) & br (14.86\%) & li (10.04\%) & span (05.77\%) \\
\bottomrule
\end{tabular}}
\caption{CCQA dataset distribution for top 5 domains, topics according to the DMOZ/Curlie annotation, question words (Q-words, only computed on the English subset) and most common markup tags. \% for q-words and markup tags presents portion of all q-word/markup appearances. ccm=commentcamarche, Rec=Recreational.}
\label{tab:data_distributions}
\end{table*}
\section{Evaluation}
In this section, we showcase the value of our CCQA dataset with experiments on the closed-book question-answering (section \ref{cbqa}) and passage retrieval for open-book QA (section \ref{pass_ref}) tasks.
\subsection{Closed-Book Question-Answering}
\label{cbqa}
\subsubsection{Task}
The closed-book question-answering task challenges systems to answer questions without to use of additional information sources, such as knowledge bases or evidence documents. As a result, models are solely relying on the question text and the information stored inside the model weights during training. Here, we evaluate our new CCQA dataset as an in-domain pre-training corpus for this highly challenging task by converting the JSON representation into plain question-answer pairs, removing markup tags and additional metadata.
\subsubsection{Models \& Training}
Using the question-answer pairs from the CCQA dataset, we in-domain pre-train large language models for question-answering. We start with vanilla BART and T5 transformer models, shown on the left side (green) in Figure \ref{fig:high_level_closed_book}. We then further in-domain pre-train the models using a denoising or sequence-to-sequence (seq2seq) setup (yellow box in Figure~\ref{fig:high_level_closed_book}). For the denoising task, we follow the vanilla BART approach \cite{lewis2020bart}, using a concatenation of \textit{Q:$\mathbin\Vert$<question>$\mathbin\Vert$A:$\mathbin\Vert$<answer>} as the model input. For the seq2seq task, we train the model to predict the gold answer given a question as input.
With the additional in-domain pre-training step, a variety of training-flows emerge, shown as numbered circles in Figure~\ref{fig:high_level_closed_book}: \\
(1) Using a vanilla pre-trained language model to fine-tune on the benchmark dataset. \\
(2) Using the CCQA dataset for in-domain pre-training and subsequently fine-tune on the benchmark dataset. \\
(3) Using a pre-trained language model to directly infer answers on the benchmark dataset (zero-shot). \\
(4) Using the CCQA in-domain pre-trained model to directly infer answers on the benchmark dataset in zero-shot fashion.
\subsubsection{Datasets}
\label{datasets}
We evaluate the performance of our CCQA corpus as an in-domain pre-training dataset on 5 common benchmarks, based on 4 publicly available datasets in the closed-book setting:
\textbf{TriviaQA} (TQA) is a short-form, factoid-style question-answering dataset \cite{joshi2017triviaqa}. For the closed-book task, we ignore the available contexts and focus exclusively on question-answer pairs. Since the official test-split of the dataset is not publicly available, we use the official validation set as our test split and randomly sample a validation set from the training data, as commonly done in previous work \citep{roberts2020much}.
\textbf{Natural Questions} (NQ) \citep{kwiatkowski-etal-2019-natural} represents a popular corpus for question-answering research. Despite most recent work focusing on the \textbf{short-form} answers (NQ-Short), the NQ corpus also provides additional \textbf{long-form} answers (NQ-Long) for a large subset of questions. In this work, we use both, short, factoid answers and long-form responses.
\textbf{ELI5}, introduced by \citet{fan2019eli5}, constitutes the first large-scale long-form dataset for open-ended question-answering. We again do not take available evidence documents into account, but focus on the question-answer pairs only.
\textbf{GooAQ} \citep{khashabi2021gooaq} contains semi-automatically extracted question-answer pairs from the Google question auto-complete feature.
\begin{table*}[t]
\centering
\begin{threeparttable}
\setlength{\tabcolsep}{0.3em}
\centering
\scalebox{0.95}{
\begin{tabular}{lrrrrr}
\toprule
\multicolumn{6}{c}{\bf Zero-Shot}\\
\midrule
\multirow{2}{*}{Model} & \multicolumn{1}{r}{TQA} & \multicolumn{1}{r}{NQ-Short} & \multicolumn{1}{r}{NQ-Long} & \multicolumn{1}{r}{ELI5} & \multicolumn{1}{r}{GooAQ} \\
& \multicolumn{1}{r}{AR} & \multicolumn{1}{r}{AR} & \multicolumn{1}{r}{R-L} & \multicolumn{1}{r}{R-L} & \multicolumn{1}{r}{R-L}\\
\midrule
\multicolumn{6}{c}{BART Large}\\
\midrule
Rand. Init. & 0.04 & 0.11 & 0.10 & 0.26 & 0.16 \\
Vanilla & $^\dagger$4.91 & $^\dagger$1.93 & 10.39 & 11.88 & 14.67 \\
Vanilla$^a$ & & & & & \\
CCQA & \bf $^\dagger$5.14 & \bf $^\dagger$2.16 & \bf 12.18 & \bf $^\dagger$15.21 & \bf $^\dagger$17.5 \\
CCQA-d & 4.80 & 2.13 & 10.33 & 11.91 & 14.88 \\
\midrule \addlinespace
\multicolumn{6}{c}{T5 Small}\\
\midrule
Rand. Init. & 0.05 & 0.11 & 1.13 & 1.49 & 0.80 \\
Vanilla & $^\dagger$5.06 & $^\dagger$1.74 & 9.16 & 7.55 & $^\dagger$8.92 \\
Vanilla$^b$ & & & & & \\
CCQA & \bf $^\dagger$5.13 & \bf $^\dagger$1.86 & \bf $^\dagger$13.63 & \bf $^\dagger$15.28 & \bf $^\dagger$15.46 \\
\midrule \addlinespace
\multicolumn{6}{c}{T5 Base}\\
\midrule
Rand. Init. & 0.04 & 0.11 & 0.00 & 0.00 & 0.00 \\
Vanilla & $^\dagger$5.49 & $^\dagger$2.02 & $^\dagger$14.39 & 12.27 & $^\dagger$14.99 \\
Vanilla$^c$ & & & & & \\
CCQA & \bf $^\dagger$7.15 & \bf $^\dagger$3.19 & \bf $^\dagger$15.08 & \bf $^\dagger$15.69 & \bf $^\dagger$15.85 \\
\bottomrule
\end{tabular}}
\centering
\scalebox{0.95}{
\begin{tabular}{rrrrr}
\toprule
\multicolumn{5}{c}{\bf Fine-Tuned}\\
\midrule
\multicolumn{1}{r}{TQA} & \multicolumn{1}{r}{NQ-Short} & \multicolumn{1}{r}{NQ-Long} & \multicolumn{1}{r}{ELI5} & \multicolumn{1}{r}{GooAQ} \\
\multicolumn{1}{r}{EM} & \multicolumn{1}{r}{EM} & \multicolumn{1}{r}{R-L} & \multicolumn{1}{r}{R-L} & \multicolumn{1}{r}{R-L}\\
\midrule
\multicolumn{5}{c}{BART Large}\\
\midrule
0.71 & 0.75 & 16.04 & 14.37 & 16.21 \\
\bf 28.67 & 23.79 & 23.47 & 16.96 & 35.67 \\
& 26.50 & & & \\
25.82 & 22.91 & 21.25 & 17.23 & 32.53 \\
27.84 & \bf 23.96 & \bf 24.56 & \bf 17.27 & \bf 35.92 \\
\midrule \addlinespace
\multicolumn{5}{c}{T5 Small}\\
\midrule
0.44 & 0.54 & 10.86 & 13.06 & 8.71 \\
\bf 21.02 & \bf 21.16 & \bf 22.09 & 16.28 & 24.70 \\
& & 19.00 & 23.00 \\
17.55 & 19.50 & 22.05 & \bf 16.33 & \bf 25.35 \\
\midrule \addlinespace
\multicolumn{5}{c}{T5 Base}\\
\midrule
0.32 & 0.38 & 13.58 & 12.72 & 7.93 \\
\bf 26.25 & \bf 23.04 & \bf 25.36 & 16.58 & \bf 29.36 \\
23.63 & 25.94 & & & \\
22.69 & 22.32 & 24.73 & \bf 16.64 & 29.09 \\
\bottomrule
\end{tabular}
}
\begin{tablenotes}[para,raggedright]
\hbox{Results from
\item[a] \citet{lewis2021question}
\item[b] \citet{khashabi2021gooaq}
\item[c] \citet{roberts2020much}}
\end{tablenotes}
\caption{Closed-book zero-shot and fine-tuned results. Best performance of fairly computed results per sub-table \textbf{bold}. $^\dagger$Zero-shot model outperforms fully fine-tuned randomly initialized transformer of same architecture. \textit{-d} extension indicates denoising CCQA pre-training task. \textit{AR}=Answer-level recall, \textit{EM}=Exact Match, \textit{RL}=Rouge-L.
}
\label{tab:cb_results}
\end{threeparttable}
\end{table*}
\subsubsection{Metrics}
\label{metrics}
For datasets with short-form answers, we use the Exact Match (EM) metric for fine-tuned systems, in line with previous work by \citet{roberts2020much} and \citet{lewis2021paq}. While the EM metric works well for systems that are aware of the task-specific format, it punishes potentially correct answers with additional context, which we believe is overly harsh in zero-shot settings, where the specific output format is not known (e.g., training-flows (3) and (4)). Therefore, we propose using the Answer-level Recall (AR) metric for our zero-shot experiments, while limiting the answer length with the \textit{max-length} and \textit{length-penalty} inference parameters. As such, the AR metric requires the correct answer to be a continuous sub-sequence of the predicted tokens, while allowing for additional context. Since AR operates on token-level, the prediction of super/sub-words, e.g., \textit{\textbf{fun}damental} instead of \textit{\textbf{fun}}, is considered incorrect.
For long-form question-answer datasets, we choose the Rouge-L (RL) score as our evaluation metric, which has shown strong correlation with Rouge-1 and Rouge-2 scores, and is commonly used in previous work \cite{khashabi2021gooaq}.
\subsubsection{Hyper-Parameters}
We use the default parameters of the BART \citep{lewis2020bart} and T5 \citep{raffel2020exploring} models for in-domain pre-training and fine-tuning whenever possible. Regarding the in-domain pre-training on our CCQA dataset, we limit training to $800k$ steps using a batch-size of $1{,}024$. During our fine-tuning runs, we limit the number of updates to $20k$ steps with a batch-size of $256$ samples, with exception of the GooAQ dataset, which we fine-tune for $100k$ steps due to it's large size. We select the best model during our in-domain pre-training runs based on the perplexity measure, and pick the top fine-tuned model according to the final evaluation metric. We do not perform any hyper-parameter search during in-domain pre-training and fine-tuning.
For the inference step, our hyper-parameter setting is closely related to commonly used summarization parameters. We use a beam-size of 4, max-length of 140, and length-penalty of 2.0. For the fine-tuned short-form task, we choose a max-length of 30, following \citet{xiong2021answering} and a length-penalty of 1.0. All model evaluations are based on Huggingface Transformers\footnote{Experiments are executed on Nvidia V100 32GB GPUs.} \cite{wolf2019huggingface}.
\begin{figure*}[t]
\centering
\begin{tikzpicture}
\begin{axis}[
width=0.38*\textwidth,
height=5cm,
legend pos=south east,
legend style={font=\small},
ymajorgrids=true,
xmajorgrids=true,
bar width=1.5mm,
ymin=5.8, ymax=25.5,
xlabel={\bf NQ-Long},
symbolic x coords={$0$, $10^1$, $10^2$, $10^3$, $10^4$, $10^5$, full},
]
\addplot coordinates {
($0$, 9.16) ($10^1$, 15.06) ($10^2$, 15.4) ($10^3$, 16.12) ($10^4$, 17.32) ($10^5$, 21.49) (full, 22.09)
};
\addlegendentry{Vanilla}
\node at (axis cs:$0$,7.36){\tiny\textcolor{blue}{9.16}};
\node at (axis cs:$10^1$,13.06){\tiny\textcolor{blue}{15.06}};
\node at (axis cs:$10^2$,13.4){\tiny\textcolor{blue}{15.4}};
\node at (axis cs:$10^3$,14.12){\tiny\textcolor{blue}{16.12}};
\node at (axis cs:$10^4$,15.12){\tiny\textcolor{blue}{17.32}};
\node at (axis cs:$10^5$,19.49){\tiny\textcolor{blue}{21.49}};
\node at (axis cs:full,20.09){\tiny\textcolor{blue}{22.09}};
\addplot coordinates {
($0$, 13.63) ($10^1$, 15.31) ($10^2$, 16.06) ($10^3$, 16.47) ($10^4$, 16.76) ($10^5$, 21.39) (full, 22.05)
};
\addlegendentry{CCQA}
\node at (axis cs:$0$,15.63){\tiny\textcolor{red}{13.63}};
\node at (axis cs:$10^1$,17.31){\tiny\textcolor{red}{15.31}};
\node at (axis cs:$10^2$,18.06){\tiny\textcolor{red}{16.06}};
\node at (axis cs:$10^3$,18.47){\tiny\textcolor{red}{16.47}};
\node at (axis cs:$10^4$,18.99){\tiny\textcolor{red}{16.76}};
\node at (axis cs:$10^5$,23.39){\tiny\textcolor{red}{21.39}};
\node at (axis cs:full,24.05){\tiny\textcolor{red}{22.05}};
\end{axis}
\end{tikzpicture}
\begin{tikzpicture}
\begin{axis}[
width=0.38*\textwidth,
height=5cm,
legend style={font=\small},
legend pos=south east,
ymajorgrids=true,
xmajorgrids=true,
bar width=1.5mm,
ymin=5, ymax=29,
xlabel={\bf GooAQ},
symbolic x coords={$0$, $10^1$, $10^2$, $10^3$, $10^4$, $10^5$, full},
]
\addplot coordinates {
($0$, 8.92) ($10^1$, 16.07) ($10^2$, 16.63) ($10^3$, 18.13) ($10^4$, 19.15) ($10^5$, 24.77) (full, 24.7)
};
\addlegendentry{Vanilla}
\node at (axis cs:$0$,6.92){\tiny\textcolor{blue}{8.92}};
\node at (axis cs:$10^1$,14.07){\tiny\textcolor{blue}{16.07}};
\node at (axis cs:$10^2$,14.63){\tiny\textcolor{blue}{16.63}};
\node at (axis cs:$10^3$,16.13){\tiny\textcolor{blue}{18.13}};
\node at (axis cs:$10^4$,17.15){\tiny\textcolor{blue}{19.15}};
\node at (axis cs:$10^5$,22.77){\tiny\textcolor{blue}{24.77}};
\node at (axis cs:full,22.7){\tiny\textcolor{blue}{24.7}};
\addplot coordinates {
($0$, 15.46) ($10^1$, 17.12) ($10^2$, 17.33) ($10^3$, 18.44) ($10^4$, 19.11) ($10^5$, 24.83) (full, 25.35)
};
\addlegendentry{CCQA}
\node at (axis cs:$0$,17.76){\tiny\textcolor{red}{15.46}};
\node at (axis cs:$10^1$,19.12){\tiny\textcolor{red}{17.12}};
\node at (axis cs:$10^2$,19.33){\tiny\textcolor{red}{17.33}};
\node at (axis cs:$10^3$,20.44){\tiny\textcolor{red}{18.44}};
\node at (axis cs:$10^4$,21.11){\tiny\textcolor{red}{19.11}};
\node at (axis cs:$10^5$,26.83){\tiny\textcolor{red}{24.83}};
\node at (axis cs:full,27.35){\tiny\textcolor{red}{25.35}};
\end{axis}
\end{tikzpicture}
\begin{tikzpicture}
\begin{axis}[
width=0.38*\textwidth,
height=5cm,
legend pos=south east,
legend style={font=\small},
ymajorgrids=true,
xmajorgrids=true,
bar width=1.5mm,
ymin=5.5, ymax=18.5,
xlabel={\bf ELI5},
symbolic x coords={$0$, $10^1$, $10^2$, $10^3$, $10^4$, $10^5$, full},
]
\addplot coordinates {
($0$, 7.55) ($10^1$, 13.99) ($10^2$, 15.16) ($10^3$, 15.55) ($10^4$, 16.11) ($10^5$, 16.21) (full, 16.28)
};
\addlegendentry{Vanilla}
\node at (axis cs:$0$,6.55){\tiny\textcolor{blue}{7.55}};
\node at (axis cs:$10^1$,13.1){\tiny\textcolor{blue}{13.99}};
\node at (axis cs:$10^2$,14.2){\tiny\textcolor{blue}{15.16}};
\node at (axis cs:$10^3$,14.6){\tiny\textcolor{blue}{15.55}};
\node at (axis cs:$10^4$,15.1){\tiny\textcolor{blue}{16.11}};
\node at (axis cs:$10^5$,15.2){\tiny\textcolor{blue}{16.21}};
\node at (axis cs:full,15.3){\tiny\textcolor{blue}{16.28}};
\addplot coordinates {
($0$, 15.28) ($10^1$, 15.5) ($10^2$, 15.80) ($10^3$, 15.94) ($10^4$, 16.12) ($10^5$, 16.24) (full, 16.33)
};
\addlegendentry{CCQA}
\node at (axis cs:$0$,16.28){\tiny\textcolor{red}{15.28}};
\node at (axis cs:$10^1$,16.5){\tiny\textcolor{red}{15.50}};
\node at (axis cs:$10^2$,16.8){\tiny\textcolor{red}{15.80}};
\node at (axis cs:$10^3$,16.94){\tiny\textcolor{red}{15.94}};
\node at (axis cs:$10^4$,17.12){\tiny\textcolor{red}{16.12}};
\node at (axis cs:$10^5$,17.24){\tiny\textcolor{red}{16.24}};
\node at (axis cs:full,17.33){\tiny\textcolor{red}{16.33}};
\end{axis}
\end{tikzpicture}
\caption{Low resource experiments comparing the Rouge-L score of vanilla T5 Small with our CCQA pre-trained models on NQ-long (left), GooAQ (center) and ELI5 (right).}
\label{fig:low_resource}
\end{figure*}
\subsubsection{Results}
\label{closed_book}
Our main results for the closed-book question-answering task are presented in Table \ref{tab:cb_results}, showing the zero-shot and fine-tuned performance of the BART Large (top), T5 Small (center) and T5 Base (bottom) models for each of the 5 evaluation datasets. Even though we present a wide variety of benchmark results, from short-form factoid questions to long-form answers, the CCQA seq2seq pre-trained model consistently outperforms all other models on the zero-shot question-answering task. Even more importantly, the additional in-domain pre-training step achieves better zero-shot performance than fully fine-tuned, randomly initialized transformer models (as extensively used prior to 2018) in almost all settings. Specifically, our model outperforms the randomly initialized transformers on all benchmarks for T5 Small and T5 Base, as well as on 4 out of 5 datasets using BART Large.
Comparing the fully fine-tuned setting across models and datasets it becomes clear that, although oftentimes performing comparably, our CCQA seq2seq pre-trained model underperforms the vanilla models in most cases. Seq2seq in-domain pre-training on CCQA only reaches superior performance on the ELI5 dataset for all models, as well as on the GooAQ dataset for T5 Small.
Showing that seq2seq pre-training on CCQA is effective in zero-shot scenarios, however only partially improves over baselines in the fine-tuned setting, we investigate: (1) Additional experiments using the CCQA dataset for denoising-style pre-training (\textit{-d} in Table \ref{tab:cb_results}) and (2) Evaluate additional low-resource scenarios, shown in Figure \ref{fig:low_resource}.
For our denoising-style in-domain pre-training experiment, we keep the available markup information, in line with HTLM \cite{aghajanyan2022htlm}.
As shown in Table~\ref{tab:cb_results}, the in-domain CCQA denoising objective outperforms the vanilla BART Large model on 4 out of 5 benchmarks in the fine-tuned setting. We believe that this result, alongside the zero-shot performance of the seq2seq CCQA model, clearly shows the usefulness and generality of our CCQA corpus for closed-book open-domain question-answering.
Taking a closer look at low-resource scenarios, we evaluate the vanilla T5 Small model against our in-domain pre-trained approach using 5 proper subsets of the NQ-Long, GooAQ and ELI5 benchmark datasets, drawn at random. As presented in Figure \ref{fig:low_resource}, our CCQA model mostly outperforms the vanilla T5 Small model in low-resource scenarios with up to $10{,}000$ data points. While the performance of our CCQA model is consistently better on the ELI5 test-set, the vanilla baselines outperform our models fastest on the NQ-Long corpus. Additional low-resource experiments on T5 Base are shown in Table \ref{fig:app_low_resource}, in Appendix \ref{app_full_res}.
\begin{figure}[t]
\centering
\includegraphics[width=\linewidth]{high_level_retrieval.png}
\caption{High-level overview of the CCQA passage retrieval in-domain pre-training step (yellow) as part of the training pipeline for DPR. Language model pre-training shown in green. Task-dependent fine-tuning presented in red. Evaluation in blue. (1) Baseline pre-training/fine-tuning pipeline, (2) In-domain pre-training/fine-tuning pipeline.}
\label{fig:high_level_retrieval}
\vspace{-3mm}
\end{figure}
\subsection{Passage Retrieval}
\label{pass_ref}
\subsubsection{Task}
For the passage retrieval task, an important component of most open-book QA systems (e.g., \citet{NEURIPS2020_6b493230, izacard2021leveraging}), models aim to extract a set of evidence passages from a large collection of documents through conditional ranking.
To align our corpus with the passage retrieval task, we aggregate every question into a single data point, consisting of the question itself, alongside all available answers as either positive or negative contexts. If available, answer votes are used as a proxy to determine positive and negative (sometimes called ``hard-negative") contexts. Following the practice in \citet{fan2019eli5}, we assign every answer with at least 2 more upvotes than downvotes as a positive context and all other answer as negative. If answer votes are not available, we use the accepted/suggested label (shown in Figure~\ref{fig:format}) as an indicator for positive and negative contexts. In the absence of either criterion, we use all available answers as positive contexts.
\subsubsection{Models \& Training}
For passage retrieval, we choose the Dense Passage Retriever (DPR) \cite{karpukhin2020dense}, used in a variety of popular end-to-end open-book QA models, such as RAG \citep{NEURIPS2020_6b493230} and FiD \citep{izacard2021leveraging}. As shown in Figure~\ref{fig:high_level_retrieval}, we start with the vanilla DPR model based on BERT \cite{devlin-etal-2019-bert} and in-domain pre-train using questions and positive/negative passages from the CCQA dataset (yellow box in Figure~\ref{fig:high_level_retrieval}), similar to \citet{ouguz2021domain}. In line with the training-flows of the closed-book models, we train DPR using either the vanilla setup (pre-training $\rightarrow$ fine-tuning) or the in-domain pre-training approach (pre-training $\rightarrow$ in-domain pre-training $\rightarrow$ fine-tuning), shown as circles (1) and (2) in Figure~\ref{fig:high_level_retrieval}, respectively.
\subsubsection{Datasets \& Metrics}
Following the original DPR paper \citep{karpukhin2020dense}, we evaluate the passage retrieval task on the NQ-Short and TQA datasets presented in section \ref{datasets}, using the top-20 and top-100 retrieval accuracy (Acc@20/Acc@100) measures.
\begin{table}[t]
\centering
\setlength{\tabcolsep}{0.18em}
\scalebox{0.9}{
\begin{tabular}{lrrrr}
\toprule
\multirow{2}{*}{Model} & \multicolumn{2}{c}{TQA} & \multicolumn{2}{c}{NQ-Short} \\
& \multicolumn{1}{c}{Acc@20} & \multicolumn{1}{c}{Acc@100} & \multicolumn{1}{c}{Acc@20} & \multicolumn{1}{c}{Acc@100} \\
\midrule
DPR & 79.4 & 85.0 & 78.4 & 85.4 \\
DPR v2 & 79.5 & 85.3 & 78.3 & 85.6 \\
CCQA DPR & \textbf{80.0} & \textbf{85.6} & \textbf{79.1} & \textbf{86.3} \\
\bottomrule
\end{tabular}}
\caption{Fine-tuned Dense Passage Retriever (DPR) accuracy measure on the TQA and NQ-Short datasets. DPR represents the original DPR model \cite{karpukhin2020dense}, DPR v2 \citep{ouguz2021domain} indicates the updated codebase. CCQA DPR uses our CCQA pre-trained DPR model for retrieval fine-tuning.}
\label{tab:ob_results}
\end{table}
\begin{table}
\centering
\setlength{\tabcolsep}{0.28em}
\scalebox{0.9}{
\begin{tabular}{lrrrrr}
\toprule
Bench. (test) & TQA & NQ-S & NQ-L & ELI5 & GooAQ \\
\midrule
Bench. (train) & 11.9 & 4.9 & 5.2 & 3.0 & 26.9 \\
CCQA (train) & 0.4 & 1.9 & 2.3 & 0.5 & 26.9 \\
\bottomrule
\end{tabular}}
\caption{8-gram question overlap (in \%) between training sets and benchmark test-sets (inspired by \citet{radford2019language}). \textit{Bench (train)} refers to the overlap between the respective training- and test-portion of the benchmark datasets, \textit{CCQA (train)} identified overlaps between our dataset and the test-splits. False positive rate upper-bound by $\frac{1}{10^8}$. All inputs are normalized and lower-cased. NQ-S=\allowbreak NQ-Short, NQ-L=\allowbreak NQ-Long.}
\label{tab:bloom}
\vspace{-3mm}
\end{table}
\subsubsection{Hyper-Parameters}
We use the default DPR hyper-parameters whenever possible \citep{karpukhin2020dense}. For in-domain pre-training, we limit training to $800k$ steps using a batch-size of $1{,}536$ samples. During fine-tuning, we restrict the number of updates to $20k$ steps with a batch-size of $128$.
The best checkpoint is selected based on the Mean Reciprocal Rank (MRR) measure, following \citet{ouguz2021domain}. We do not perform any hyper-parameter search.
\subsubsection{Results}
\label{open_book}
For the passage retrieval experiments, we compare our CCQA in-domain pre-trained DPR model against the vanilla DPR model published in \citet{karpukhin2020dense}, as well as the recently enhanced version \citep{ouguz2021domain}. Table \ref{tab:ob_results} contains our empirical results, showing consistent improvements of our CCQA DPR model over the vanilla baselines. More specifically, the in-domain CCQA pre-training step increases the top-20 and top-100 accuracy score on the TQA benchmark dataset by over half a point, while the performance gap on NQ-Short shows consistent improvement of over $0.7\%$.
\subsection{Evaluation Fairness: Dataset Overlap}
\label{overlap}
With modern pre-training approaches using increasingly large datasets, accidental overlaps between pre-training corpora and benchmark datasets become more and more common \citep{lewis2021question}. To analyze this threat to the integrity of our dataset and empirical analysis, we follow \citet{radford2019language} and evaluate the 8-gram question overlap of our CCQA training portion with the test-split of benchmark datasets using bloom filters. Table \ref{tab:bloom} shows a consistently smaller question overlap between CCQA and the benchmark test set, compared to the benchmark training split itself.
\section{Conclusion and Future Work}
In this work, we presented our new web-scale CCQA dataset for in-domain model pre-training. Orthogonal to recent efforts on improving task-specific pre-training objectives, we show our dataset generation process, followed by detailed insights into key corpus dimensions of this new, large-scale, natural, and diverse question-answering dataset. In a set of empirical evaluations, we confirm the initial intuition that the corpus presents a valuable resource for open-domain question-answering research. In our zero-shot, low-resource and fine-tuned experiments for open- and closed-book QA tasks, we show promising results across multiple model architectures.
With around $130$ million question-answer pairs ($55$M unique) as well as additional meta-data, our CCQA dataset presents a versatile source of information, which has a large variety of applications in future work (e.g., question summarization, answer rating, answer ranking and many more).
\section{Ethical Considerations}
We now discuss the three major ethical considerations impacting this paper:
\paragraph{Hate-speech, Harmful Gender and Racial Biases:}
With general web-data potentially containing hate-speech and harmful gender and racial biases, we believe that our extracted dataset based on the schema.org annotations is less impacted by these issues, with the schema.org annotation representing a good proxy for high-quality, professionally curated websites. As a result, we believe that the severity of this issue is significantly reduced. Furthermore, in our human evaluation, we find no signs of the above mentioned biases. We leave computational approaches to determine dataset biases for future work (e.g., the Word Embedding Association Test \citep{Caliskan2017SemanticsDA} and Sentence Encoder Association Test \citep{may2019measuring}).
\paragraph{Data Availability:}
We do not directly provide the CCQA dataset, but enable third parties to generate the corpus through our published dataset generation scripts available at \url{https://github.com/facebookresearch/CCQA}.
\paragraph{Hallucinations and Factual Errors:}
As shown in the evaluation section, our model is able to generate reasonable answers for factoid and long-form questions. The inferred answers are fluent and human-like, but may also contain hallucinations and factual errors, especially for the challenging closed-book question-answering task. Without a guarantee of the predicted answers being factually correct, they can potentially spread misinformation if not properly corrected.
\subsubsection*{Acknowledgments}
We want to thank the anonymous reviewers for their valuable feedback and acknowledge our colleagues and peers for their help throughout this project, specifically Anchit Gupta, Gagan Aneja and Patrick Lewis for their valuable input on important decisions.
|
\section{Introduction and main results}\label{section:introduction}
In order to explain the motivation for our work and to state our main results, we set up the following terminology regarding \emph{higher rank lattices}.
\begin{terminology}
Let $G$ be any connected semisimple real Lie group with finite center, no nontrivial compact factor and real rank $\operatorname{rk}_{\mathbf{R}}(G) \geq 2$. Let $\Gamma < G$ be any \emph{irreducible lattice}, meaning that $\Gamma$ is a discrete subgroup of $G$ with finite covolume such that $N \cdot \Gamma$ is a dense subgroup of $G$ for every noncentral closed normal subgroup $N \lhd G$. In what follows, if all the above conditions are satisfied, then we simply say that $\Gamma < G$ is a \emph{higher rank lattice}.
\end{terminology}
The following examples of higher rank lattices are particular cases of general results due to Borel--Harish-Chandra \cite{BHC61}.
\begin{Examples}
For every $d \geq 2$, the special linear group $\operatorname{SL}_d(\mathbf{R})$ is a connected simple real Lie group with finite center $\mathscr Z(\operatorname{SL}_d(\mathbf{R})) = \{\pm 1_d\}$ and real rank $\operatorname{rk}_{\mathbf{R}}(\operatorname{SL}_d(\mathbf{R})) = d - 1$.
\begin{enumerate}
\item For every $d \geq 3$, $\operatorname{SL}_d(\mathbf{Z}) < \operatorname{SL}_d(\mathbf{R})$ is a higher rank lattice.
\item For every $d \geq 2$ and every square free integer $q \in \mathbf{N}$,
\begin{equation*}
\Gamma \coloneqq \{ (g, g^\sigma) \mid g \in \operatorname{SL}_d(\mathbf{Z}[\sqrt{q}]) \} < \operatorname{SL}_d(\mathbf{R}) \times \operatorname{SL}_d(\mathbf{R}) \coloneqq G
\end{equation*}
is a higher rank lattice, where $\sigma$ is the order $2$ automorphism of $\mathbf{Q}(\sqrt{q})$.
\end{enumerate}
\end{Examples}
The main inspiration for our work is Margulis'\! celebrated \emph{normal subgroup theorem} which states that for any higher rank lattice $\Gamma < G$, any normal subgroup $N \lhd \Gamma$ is either finite and contained in $\mathscr Z(\Gamma)$ or $N$ has finite index in $\Gamma$ (see \cite[Theorem IV.4.9]{Ma91}). Margulis'\! remarkable strategy to prove the normal subgroup theorem consists of two ``halves'': the \emph{amenability half} and the \emph{property} (T) \emph{half}. Indeed, assuming that $N \lhd \Gamma$ is a noncentral normal subgroup, to prove that the quotient group $\Gamma/ N$ is finite, Margulis showed that $\Gamma/N$ is both amenable and has property (T). The proof of the amenability half relies on Margulis'\! \emph{factor theorem} which states that any measurable $\Gamma$-factor of the homogeneous space $G/P$, where $P < G$ is a minimal parabolic subgroup, is measurably isomorphic to a $G$-factor whence of the form $G/Q$, where $P < Q < G$ is an intermediate parabolic subgroup (see \cite[Theorem IV.2.11]{Ma91}). Margulis'\! strategy has been used to prove a normal subgroup theorem for various classes of irreducible lattices in product groups (see \cite{BM00, Sh99, BS04, CS13}) and to understand the structure of point stabilizers of ergodic probability measure preserving actions of higher rank lattices (see \cite{SZ92, CP12}). More recently, Margulis'\! strategy has been adapted to the noncommutative setting to study \emph{characters} of higher rank lattices (see \cite{CP13, Pe14}).
In that respect, for any countable discrete group $\Lambda$, we denote by $\mathscr P(\Lambda)$ the space of positive definite functions $\varphi : \Lambda \to \mathbf{C}$ normalized so that $\varphi(e) = 1$. Then $\mathscr P(\Lambda) \subset \ell^\infty(\Lambda)$ is a weak-$\ast$ compact convex subset. Thanks to the Gelfand--Naimark--Segal (GNS) construction, to any positive definite function $\varphi \in \mathscr P(\Lambda)$ corresponds a triple $(\pi_\varphi, \mathscr H_\varphi, \xi_\varphi)$, where $\pi_\varphi : \Lambda \to \mathscr U(\mathscr H_\varphi)$ is a unitary representation and $\xi_\varphi \in \mathscr H_\varphi$ is a unit vector such that the linear span of $\pi_\varphi(\Lambda)\xi_\varphi$ is dense in $\mathscr H_\varphi$ and
\begin{equation*}
\forall \gamma \in \Lambda, \quad \varphi (\gamma) = \langle \pi_\varphi(\gamma)\xi_\varphi, \xi_\varphi\rangle.
\end{equation*}
We consider the conjugation action $\Lambda \curvearrowright \mathscr P(\Lambda)$ defined by
\begin{equation*}
\forall \gamma, g \in \Lambda, \forall \varphi \in \mathscr P(\Lambda), \quad (\gamma \varphi)(g) \coloneqq \varphi(\gamma^{-1} g \gamma).
\end{equation*}
A fixed point $\varphi \in \mathscr P(\Lambda)$ for the conjugation action is called a \emph{character}. We denote by $\operatorname{Char}(\Lambda) \subset \mathscr P(\Lambda)$ the weak-$\ast$ compact convex subset of all characters. Any countable discrete group $\Lambda$ always admits at least two characters: the \emph{trivial} character $1_\Lambda$ and the \emph{regular} character $\delta_e$. The GNS representation of the regular character $\delta_e$ coincides with the left regular representation $\lambda : \Lambda \to \mathscr U(\ell^2(\Lambda))$. An important source of characters comes from ergodic theory. Indeed, for any probability measure preserving action $\Lambda \curvearrowright (X, \nu)$ on a standard probability space, the function $\varphi : \Lambda \to \mathbf{C} : \gamma \mapsto \nu(\operatorname{Fix}(\gamma))$ defines a character. The action $\Lambda \curvearrowright (X, \nu)$ is (essentially) free if and only if the above character $\varphi$ is equal to $\delta_e$.
For any unitary representation $\pi : \Lambda \to \mathscr U(\mathscr H_\pi)$, we consider the unital $\operatorname{C}^*$-algebra
$$\operatorname{C}^*_\pi(\Lambda) \coloneqq \operatorname{C}^*(\{\pi(\gamma) \mid \gamma \in \Lambda \}) \subset \mathrm B(\mathscr H_\pi)$$ endowed with the conjugation action $\operatorname{Ad}(\pi) : \Lambda \curvearrowright \operatorname{C}^*_\pi(\Lambda)$. We then regard the state space $\mathfrak S(\operatorname{C}^*_\pi(\Lambda))$ as a $\Lambda$-invariant weak-$\ast$ compact convex subset of $\mathscr P(\Lambda)$ via the mapping $\mathfrak S(\operatorname{C}^*_\pi(\Lambda)) \hookrightarrow \mathscr P(\Lambda) : \psi \mapsto \psi \circ \pi$. When $\pi = \lambda$ is the left regular representation, $\operatorname{C}^*_\lambda(\Lambda)$ is the \emph{reduced} group $\operatorname{C}^*$-algebra which is endowed with the canonical faithful trace $\tau_\Lambda$ defined by $\tau_\Lambda : \operatorname{C}^*_\lambda(\Lambda) \to \mathbf{C} : a \mapsto \langle a \delta_e, \delta_e\rangle$.
Given unitary representations $\pi_i : \Lambda \to \mathscr U(\mathscr H_i)$, $i = 1, 2$, we say that $\pi_2$ is \emph{weakly contained} in $\pi_1$ if the map $\pi_1(\Lambda) \to \pi_2(\Lambda) : \pi_1(\gamma) \mapsto \pi_2(\gamma)$ is well defined and extends to a $\ast$-homomorphism $\operatorname{C}^*_{\pi_1}(\Lambda) \to \operatorname{C}^*_{\pi_2}(\Lambda)$. Following \cite{Be89}, we say that a unitary representation $\pi : \Lambda \to \mathscr U(\mathscr H_\pi)$ is \emph{amenable} if the trivial representation $1_\Lambda$ is weakly contained in $\pi \otimes \overline \pi$. If $\pi$ contains a finite dimensional subrepresentation, then $\pi$ is amenable. If $\Lambda$ has property (T), then conversely any amenable representation $\pi : \Lambda \to \mathscr U(\mathscr H_\pi)$ contains a finite dimensional subrepresentation.
We now present in a unified way the main results we obtained in \cite{BH19} (joint work with R.\ Boutonnet) and \cite{BBHP20} (joint work with U.\ Bader, R.\ Boutonnet and J.\ Peterson). Our first main result deals with the \emph{existence of characters}. It is a fixed point theorem for the affine action of higher rank lattices on their space of positive definite functions.
\begin{letterthm}[\cite{BH19, BBHP20}]\label{thm:main}
Let $\Gamma < G$ be any higher rank lattice. Then any nonempty $\Gamma$-invariant weak-$\ast$ compact convex subset $\mathscr C \subset \mathscr P(\Gamma)$ contains a character.
\end{letterthm}
Our second main result deals with the \emph{classification of characters} of higher rank lattices. Bekka \cite{Be06} obtained the first character rigidity results in the case $\Gamma = \operatorname{SL}_d(\mathbf{Z})$ for $d \geq 3$. More recently, using a different approach based on Margulis'\! strategy discussed above, Peterson \cite{Pe14} obtained character rigidity results for arbitrary higher rank lattices (see also \cite{CP13} for the case of irreducible lattices in certain product groups). The operator algebraic framework we developed in \cite{BH19, BBHP20} enables us to obtain a new and more conceptual proof of Peterson's character rigidity results \cite{Pe14}.
\begin{letterthm}[Peterson, \cite{Pe14}]\label{thm:characters}
Let $\Gamma < G$ be any higher rank lattice. Then any character $\varphi \in \operatorname{Char}(\Gamma)$ is either supported on $\mathscr Z(\Gamma)$ or its GNS representation $\pi_\varphi$ is amenable.
In case $G$ has a simple factor with property \emph{(T)}, any character $\varphi \in \operatorname{Char}(\Gamma)$ is either supported on $\mathscr Z(\Gamma)$ or its GNS representation $\pi_\varphi$ contains a finite dimensional subrepresentation.
\end{letterthm}
Theorem \ref{thm:characters} generalizes Margulis'\! normal subgroup theorem \cite{Ma91} and Stuck--Zimmer's stabilizer rigidity theorem \cite{SZ92}. Also, Theorem \ref{thm:characters} solved a conjecture formulated by Connes (see \cite{Jo00}). For other recent results regarding classification of characters, we refer the reader to \cite{CP13, PT13, Be19, BF20, LL20}.
Combining Theorems \ref{thm:main} and \ref{thm:characters}, we obtain new results regarding the simplicity and the unique trace property for the $\operatorname{C}^*$-algebra $\operatorname{C}^*_\pi(\Gamma)$ associated with an arbitrary nonamenable (resp.\! weakly mixing) unitary representation $\pi : \Gamma \to \mathscr U(\mathscr H_\pi)$. In particular, Corollary \ref{cor:C*} provides a far reaching generalization of the results obtained by Bekka--Cowling--de la Harpe \cite{BCH94} for the reduced $\operatorname{C}^*$-algebra $\operatorname{C}^*_\lambda(\Gamma)$.
\begin{lettercor}[\cite{BH19, BBHP20}]\label{cor:C*}
Let $\Gamma < G$ be any higher rank lattice. Let $\pi : \Gamma \to \mathscr U(\mathscr H_\pi)$ be any unitary representation. Then $\operatorname{C}^*_\pi(\Gamma)$ admits a trace.
Assume moreover that $G$ has trivial center. If $\pi$ is not amenable, then $\lambda$ is weakly contained in $\pi$ and the unique $\ast$-homomorphism $\Theta : \operatorname{C}^*_\pi(\Gamma) \to \operatorname{C}^*_\lambda(\Gamma) : \pi(\gamma) \mapsto \lambda(\gamma)$ satisfies the following properties:
\begin{enumerate}
\item $\tau_\Gamma \circ \Theta$ is the unique trace on $\operatorname{C}^*_\pi(\Gamma)$.
\item $\ker(\Theta)$ is the unique proper maximal ideal of $\operatorname{C}^*_\pi(\Gamma)$.
\end{enumerate}
In case $G$ has property \emph{(T)}, the above properties hold as soon as $\pi$ does not contain any nonzero finite dimensional subrepresentation.
\end{lettercor}
In case $\Lambda$ is a countable discrete group and $\pi = \lambda$, Hartman--Kalantar \cite{HK17} initiated the study of the noncommutative dynamical system $\Lambda \curvearrowright \operatorname{C}^*_\lambda(\Lambda)$ and obtained a new characterization of the simplicity and the unique trace property for $\operatorname{C}^*_\lambda(\Lambda)$ (see also \cite{KK14, BKKO14}).
As a byproduct of our operator algebraic methods, we also obtain a topological analogue of Stuck--Zimmer's stabilizer rigidity theorem \cite{SZ92}. In particular, our next result gives a positive answer to a recent problem raised by Glasner--Weiss \cite{GW14}.
\begin{letterthm}[\cite{BH19, BBHP20}]\label{thm:dynamics}
Let $\Gamma < G$ be any higher rank lattice and assume that $G$ has trivial center. Let $\Gamma \curvearrowright X$ be any minimal action on a compact space. Then at least one of the following assertions hold:
\begin{enumerate}
\item There exists a $\Gamma$-invariant Borel probability measure on $X$.
\item The action $\Gamma \curvearrowright X$ is topologically free.
\end{enumerate}
In case $G$ has a simple factor with property \emph{(T)}, either $X$ is finite or the action $\Gamma \curvearrowright X$ is topologically free.
\end{letterthm}
As we will explain, all the main results stated above are consequences of a dynamical dichotomy theorem for $\Gamma$-equivariant faithful normal unital completely positive (ucp) maps $\Phi : M \to \operatorname{ L}^\infty(G/P)$, where $M$ is an arbitrary von Neumann algebra endowed with an ergodic action $\Gamma \curvearrowright M$ (see Theorem \ref{thm:NCNZ} and Theorem \ref{thm:dynamical} below). This dynamical dichotomy theorem is one of the key novelties of our operator algebraic framework. Both its statement and its proof rely on von Neumann algebra theory and depend heavily on whether the connected semisimple real Lie group $G$ is simple or not.
In our first joint work \cite{BH19}, we dealt with the case where $G$ is simple and $\operatorname{rk}_{\mathbf{R}}(G) \geq 2$ (and more generally where all its simple factors $G_i$ satisfy $\operatorname{rk}_{\mathbf{R}}(G_i) \geq 2$). In that case, we obtained the following noncommutative analogue of Nevo--Zimmer's structure theorem \cite{NZ97, NZ00}. We denote by $P < G$ a minimal parabolic subgroup and whenever $P < Q < G$ is an intermediate parabolic subgroup, we denote by $p_Q : G/P \to G/Q : gP \mapsto gQ$ the canonical factor map and by $p_Q^\ast : \operatorname{ L}^\infty(G/Q) \to \operatorname{ L}^\infty(G/P) : f \mapsto f \circ p_Q$ the corresponding unital normal embedding.
\begin{letterthm}[\cite{BH19}]\label{thm:NCNZ}
Let $\Gamma < G$ be any higher rank lattice and assume that $G$ is simple. Let $M$ be any von Neumann algebra, $\Gamma \curvearrowright M$ any ergodic action and $\Phi : M \to \operatorname{ L}^\infty(G/P)$ any $\Gamma$-equivariant faithful normal ucp map. Then the following dichotomy holds:
\begin{itemize}
\item Either $\Phi(M) = \mathbf{C} 1$.
\item Or there exist a proper parabolic subgroup $P < Q < G$ and a $\Gamma$-equivariant unital normal embedding $\iota : \operatorname{ L}^\infty(G/Q) \hookrightarrow M$ such that $\Phi \circ \iota = p_Q^\ast$.
\end{itemize}
\end{letterthm}
Theorem \ref{thm:NCNZ} extends the work of Nevo--Zimmer in two ways. Firstly, we deal with arbitrary von Neumann algebras $M$ (instead of measure spaces $(X, \nu)$) and secondly, we deal with $\Gamma$-actions $\Gamma \curvearrowright M$ (instead of $G$-actions $G \curvearrowright (X, \nu)$). We refer to Section \ref{section:structures} for the correspondence between equivariant ucp maps and stationary states. The remarkable feature of Theorem \ref{thm:NCNZ} is that when $\Phi : M \to \operatorname{ L}^\infty(G/P)$ is not invariant, there is a nontrivial $\Gamma$-invariant \emph{commutative} von Neumann subalgebra $M_0 \subset M$ such that $M_0 \cong \operatorname{ L}^\infty(G/Q)$. This allows us to exploit the dynamical properties of the ergodic action $\Gamma \curvearrowright G/Q$ and the fact that every noncentral element $\gamma \in \Gamma \setminus \mathscr Z(\Gamma)$ acts (essentially) freely on $G/Q$. In particular in \cite{BH19}, we used Theorem \ref{thm:NCNZ} to derive all the main results stated above.
In our second joint work \cite{BBHP20}, we dealt with the case where $G$ is not simple. We point out that when $G$ has a rank one simple factor (e.g.\ $G = \operatorname{SL}_2(\mathbf{R}) \times \operatorname{SL}_2(\mathbf{R})$), a Nevo--Zimmer structure theorem does not hold and the method used in \cite{BH19} to prove the main results does not apply. Instead, we proceeded as follows \cite{BBHP20}. Firstly, for \emph{any} higher rank lattice $\Gamma < G$, we formulated a general dynamical dichotomy theorem \emph{invariant vs.\ singular} for $\Gamma$-equivariant faithful normal ucp maps $\Phi : M \to \operatorname{ L}^\infty(G/P)$ (see Theorem \ref{thm:dynamical} below) and we showed that all the main results stated above can be derived from this general dichotomy theorem. In case $G$ is simple, the dynamical dichotomy Theorem \ref{thm:dynamical} is a straightforward consequence of Theorem \ref{thm:NCNZ}. Secondly, to prove the dynamical dichotomy Theorem \ref{thm:dynamical} in the case where $G$ is not simple, we developed a new method based on the product structure in $G$. In that respect, the tools developed in \cite{BH19} and \cite{BBHP20} are complementary.
For any nonsingular action $\Lambda \curvearrowright (X, \nu)$ on a standard probability space, we denote by $\operatorname{ L}(\Lambda \curvearrowright X)$ the corresponding \emph{group measure space} von Neumann algebra (see Section \ref{section:preliminaries}). For higher rank lattices $\Gamma < G$ where $G$ is simple with trivial center, we present yet another application of Theorem \ref{thm:NCNZ} that will appear in \cite{BBH21} (joint work with U.\ Bader and R.\ Boutonnet). The next result can be regarded as a noncommutative analogue of Margulis'\! factor theorem.
\begin{lettercor}[\cite{BBH21}]\label{thm:NCFT}
Let $\Gamma < G$ be any higher rank lattice and assume that $G$ is simple with trivial center. Let $\operatorname{ L}(\Gamma) \subset M \subset \operatorname{ L}(\Gamma \curvearrowright G/P)$ be any intermediate von Neumann subalgebra. Then there is a unique intermediate parabolic subgroup $P < Q < G$ such that
\begin{equation*}
M = \operatorname{ L}(\Gamma \curvearrowright G/Q).
\end{equation*}
\end{lettercor}
In Section \ref{section:preliminaries}, we give some preliminary background on Poisson boundaries, semisimple Lie groups and operator algebras. In Section \ref{section:structures}, we introduce the notion of boundary structures for von Neumann algebras and state the dynamical dichotomy theorem. We also outline the main steps of the proof of Theorem \ref{thm:NCNZ}. In Section \ref{section:main}, we sketch the proofs of Theorems \ref{thm:main}, \ref{thm:characters} and Corollary \ref{cor:C*} based on the dynamical dichotomy theorem. We also discuss open problems related to our main results. In Section \ref{section:NCFT}, we discuss Corollary \ref{thm:NCFT} and its relevance for Connes'\! rigidity conjecture.
\begin{Remark}
In this survey article, we only consider higher rank lattices in connected semisimple \emph{real} Lie groups to simplify the exposition and to focus on the main ideas. However, we point out that all our main results do hold for higher rank lattices in semisimple algebraic groups defined over arbitrary local fields. We refer the reader to \cite{BBHP20, BBH21} for more general statements and further details.
\end{Remark}
\subsection*{Acknowledgments}
I would like to express my deepest gratitude to my co-authors for our exciting joint works. I also thank Adrian Ioana for his useful remarks.
\section{Preliminaries}\label{section:preliminaries}
\subsection{Poisson boundaries}\label{subsection:poisson}
Let $H$ be any locally compact second countable (lcsc) group. We say that a Borel probability measure $\mu \in \operatorname{Prob}(H)$ is \emph{admissible} if the following conditions are satisfied:
\begin{enumerate}
\item $\mu$ is absolutely continuous with respect to the Haar measure;
\item $\operatorname{supp}(\mu)$ generates $H$ as a semigroup;
\item $\operatorname{supp}(\mu)$ contains a neighborhood of the identity element $e \in H$.
\end{enumerate}
We say that a bounded measurable function $F : H \to \mathbf{C}$ is (right) $\mu$-\emph{harmonic} if
\begin{equation*}
\forall g \in H, \quad F(g) = \int_H F(gh) \, \mathrm{d}\mu(h).
\end{equation*}
Any $\mu$-harmonic function is continuous. We denote by $\operatorname{Har}^\infty(H, \mu) \subset \operatorname{C}_b(H)$ the space of all (right) $\mu$-harmonic functions. The left translation action $\lambda : H \curvearrowright \operatorname{C}_b(H)$ leaves the subspace $\operatorname{Har}^\infty(H, \mu)$ globally invariant.
Let $(X, \nu)$ be any standard probability space endowed with a measurable action $H \curvearrowright X$. We say that $(X, \nu)$ is a $(H, \mu)$-\emph{space} if $\nu$ is $\mu$-\emph{stationary}, that is, $\mu \ast \nu = \nu$. For any $(H, \mu)$-space $(X, \nu)$, define the \emph{Poisson transform} $\Psi_\mu : \operatorname{ L}^\infty(X, \nu) \to \operatorname{Har}^\infty(H, \mu)$ by the formula
\begin{equation*}
\forall f \in \operatorname{ L}^\infty(X, \nu), \forall g \in H, \quad \Psi_\mu(f)(g) = \int_X f(g x) \, \mathrm{d}\nu(x).
\end{equation*}
The mapping $\Psi_\mu : \operatorname{ L}^\infty(X, \nu) \to \operatorname{Har}^\infty(H, \mu)$ is $H$-equivariant, unital, positive and contractive.
\begin{theorem}[Furstenberg, \cite{Fu62b}]
There is a unique $(H, \mu)$-space $(B, \nu_B)$ for which the Poisson transform $\Psi_\mu : \operatorname{ L}^\infty(B, \nu_B) \to \operatorname{Har}^\infty(H, \mu)$ is bijective.
\end{theorem}
The $(H, \mu)$-space $(B, \nu_B)$ is called the $(H, \mu)$-\emph{Poisson boundary}. For a construction of the $(H, \mu)$-space $(B, \nu_B)$, we also refer to \cite{BS04, Fu00}. The $(H, \mu)$-space $(B, \nu_B)$ enjoys remarkable ergodic theoretic properties. In that respect, let $(E, \|\cdot\| )$ be any separable continuous isometric Banach $H$-module and $\mathscr C \subset E^*$ any nonempty $H$-invariant weak-$\ast$ compact convex subset. Denote by $\operatorname{Bar} : \operatorname{Prob}(\mathscr C) \to \mathscr C$ the $H$-equivariant continuous barycenter map. A point $c \in \mathscr C$ is $\mu$-\emph{stationary} if $\operatorname{Bar}({\iota_c}_\ast\mu) = c$ where $\iota_c : H \to \mathscr C : g \mapsto gc$ is the orbit map associated with $c \in \mathscr C$. By Markov--Kakutani's fixed point theorem, the subset $\mathscr C_\mu \subset \mathscr C$ of all $\mu$-stationary points in $\mathscr C$ is not empty.
The following theorem due to Furstenberg provides the existence (and uniqueness) of boundary maps (see also \cite[Section 2]{BS04}).
\begin{theorem}[Furstenberg, \cite{Fu62b}]
Let $c \in \mathscr C_\mu$ be any $\mu$-stationary point. Then there exists an (essentially) unique $H$-equivariant measurable map $\beta : B \to \mathscr C$ such that
\begin{equation*}
\operatorname{Bar}({\beta}_\ast \nu_B ) = c.
\end{equation*}
We say that $\beta : B \to \mathscr C$ is the $H$-equivariant \emph{boundary map} associated with $c \in \mathscr C_\mu$.
\end{theorem}
\subsection{Semisimple Lie groups}
Let $G$ be any connected semisimple real Lie group with finite center and no nontrivial compact factors. Fix an Iwasawa decomposition $G = KAV$, where $K < G$ is a maximal compact subgroup, $A < G$ is a Cartan subgroup and $V < G$ is a unipotent subgroup. Denote by $L \coloneqq \mathscr Z_G(A)$ the centralizer of $A$ in $G$ and set $P \coloneqq L V$. Then $P < G$ is a minimal parabolic subgroup. Since $K \curvearrowright G/P$ is transitive, $G/P$ is a compact homogeneous space and there exists a unique $K$-invariant Borel probability measure $\nu_P \in \operatorname{Prob}(G/P)$. The measure class of $\nu_P$ coincides with the unique $G$-invariant measure class on $G/P$.
\begin{example}
Assume that $G = \operatorname{SL}_d(\mathbf{R})$ for $d \geq 2$. Then we may take $K = \operatorname{SO}_d(\mathbf{R})$, $A < G$ the subgroup of diagonal matrices and $V < G$ the subgroup of strict upper triangular matrices. In that case, $P = AV < G$ is the subgroup of upper triangular matrices. The homogeneous space $G/P$ is the \emph{full flag variety} which consists in all flags $\{0\} \subset W_1 \subset \cdots \subset W_d = \mathbf{R}^d$, where $W_i \subset \mathbf{R}^d$ is a vector subspace such that $\dim_{\mathbf{R}}(W_i) = i$ for every $1 \leq i \leq d$.
\end{example}
Observe for any left $K$-invariant Borel probability measure $\mu_G \in \operatorname{Prob}(G)$, the probability measure $\mu_G \ast \nu_P$ is $K$-invariant on $G/P$ and so $\mu_G \ast \nu_P = \nu_P$, that is, $(G/P, \nu_P)$ is a $(G, \mu_G)$-space. Furstenberg \cite{Fu62a} proved the following fundamental result describing the Poisson boundary of semisimple Lie groups.
\begin{theorem}[Furstenberg, \cite{Fu62a}]\label{thm:poisson-G}
Let $\mu_G \in \operatorname{Prob}(G)$ be any $K$-invariant admissible Borel probability measure. Then $(G/P, \nu_P)$ is the $(G, \mu_G)$-Poisson boundary.
\end{theorem}
For lattices $\Gamma < G$ in connected semisimple \emph{real} Lie groups as above, Furstenberg \cite{Fu67} also showed that $(G/P, \nu_P)$ can be regarded as the $(\Gamma, \mu_\Gamma)$-Poisson boundary with respect to a well chosen probability measure $\mu_\Gamma \in \operatorname{Prob}(\Gamma)$ (see also \cite{Fu00} and the references therein).
\begin{theorem}[Furstenberg, \cite{Fu67}]\label{thm:poisson-Gamma}
Let $\Gamma < G$ be any lattice. Then there exists a probability measure $\mu_\Gamma \in \operatorname{Prob}(\Gamma)$ with full support such that $(G/P, \nu_P)$ is the $(\Gamma, \mu_\Gamma)$-Poisson boundary.
\end{theorem}
We call a probability measure $\mu_\Gamma \in \operatorname{Prob} (\Gamma)$ as in Theorem \ref{thm:poisson-Gamma} a \emph{Furstenberg measure}. Combining Theorems \ref{thm:poisson-G} and \ref{thm:poisson-Gamma}, we have
\begin{equation*}
\operatorname{Har}^\infty(G, \mu_G) \quad \underset{G\text{-equiv.}}{\cong} \quad \operatorname{ L}^\infty(G/P, \nu_P) \quad \underset{\Gamma\text{-equiv.}}{\cong} \quad \operatorname{Har}^\infty(\Gamma, \mu_\Gamma).
\end{equation*}
A combination of Theorem \ref{thm:poisson-Gamma} and \cite{GM89} implies that for any intermediate parabolic subgroup $P < Q < G$, the map
\begin{equation}\label{eq:uniqueness}
\delta \circ p_Q : G/P \to \operatorname{Prob}(G/Q) : gP \mapsto \delta_{gQ}
\end{equation}
is the (essentially) unique $\Gamma$-equivariant measurable mapping $\zeta : G/P \to \operatorname{Prob}(G/Q)$.
\subsection{Operator algebras}
A $\operatorname{C}^*$-\emph{algebra} $A$ is a Banach $\ast$-algebra endowed with a complete norm $\|\cdot\|$ that satisfies the $\operatorname{C}^*$-identity: $\|a^* a\| = \|a\|^2$, for every $a \in A$.
Any $\operatorname{C}^*$-algebra $A$ admits a faithful isometric $\ast$-representation on a Hilbert space $\pi : A \to \mathrm B(\mathscr H)$. After identifying $A$ with $\pi(A)$, we may regard $A \subset \mathrm B(\mathscr H)$ as a concrete $\operatorname{C}^*$-algebra. Unless stated otherwise, all $\operatorname{C}^*$-algebras and all linear mappings between $\operatorname{C}^*$-algebras are always assumed to be unital.
We denote by $\mathfrak S(A)$ the \emph{state space} of $A$. Then $\mathfrak S(A) \subset \operatorname{Ball}(A^*)$ is a weak-$\ast$ compact convex subset. We say that an action $\sigma : H \curvearrowright A$ is \emph{continuous} if the action map $H \times A \to A : (g, a) \mapsto \sigma_g(a)$ is continuous. We then simply say that $A$ is a $H$-$\operatorname{C}^*$-algebra. The continuous action $H \curvearrowright A$ induces a weak-$\ast$ continuous affine action $H \curvearrowright \mathfrak S(A)$. We may apply the results from Subsection \ref{subsection:poisson} to the $H$-invariant weak-$\ast$ compact convex set $\mathscr C = \mathfrak S(A)$. When $\mu \in \operatorname{Prob}(H)$ is an admissible Borel probability measure, we denote by $\mathfrak S_\mu(A) \subset \mathfrak S(A)$ the nonempty weak-$\ast$ compact convex subset of all $\mu$-stationary states.
\begin{examples} We will consider the following examples of $\operatorname{C}^*$-algebras.
\begin{enumerate}
\item For any compact metrizable space $X$, the space $\operatorname{C}(X)$ of all continuous functions on $X$ endowed with the uniform norm $\|\cdot\|_\infty$ is a commutative $\operatorname{C}^*$-algebra. Any commutative $\operatorname{C}^*$-algebra arises this way. We identify the set $\operatorname{Prob}(X)$ of Borel probability measures on $X$ with the state space $\mathfrak S(\operatorname{C}(X))$ via the continuous mapping $\operatorname{Prob}(X) \to \mathfrak S(\operatorname{C}(X)) : \nu \mapsto \int_X \cdot \; \mathrm{d}\nu$. Any continuous action by homeomorphisms $H \curvearrowright X$ naturally gives rise to a continuous action $H \curvearrowright \operatorname{C}(X)$ in the above sense.
\item For any countable discrete group $\Lambda$ and any unitary representation $\pi : \Lambda \to \mathscr U(\mathscr H_\pi)$, define the $\operatorname{C}^*$-algebra
\begin{equation*}
\operatorname{C}^*_\pi(\Lambda) \coloneqq \operatorname{C}^*(\left\{ \pi(\gamma) \mid \gamma \in \Lambda\right\}) \subset \mathrm B(\mathscr H_\pi)
\end{equation*}
and consider the conjugation action $\operatorname{Ad}(\pi) : \Lambda \curvearrowright \operatorname{C}^*_\pi(\Lambda)$. The state space $\mathfrak S(\operatorname{C}^*_\pi(\Lambda))$ is a $\Lambda$-invariant weak-$\ast$ compact convex subset of $\mathscr P(\Lambda)$ via the mapping $\mathfrak S(\operatorname{C}^*_\pi(\Lambda)) \hookrightarrow \mathscr P(\Lambda) : \psi \mapsto \psi \circ \pi$. If $\pi = \lambda$ is the left regular representation, then $\operatorname{C}^*_\lambda(\Lambda)$ is the \emph{reduced} group $\operatorname{C}^*$-algebra. Moreover, the state $\tau_\Lambda : \operatorname{C}^*_\lambda(\Lambda) \to \mathbf{C}: a \mapsto \langle a \delta_e, \delta_e\rangle$ is a faithful trace.
\end{enumerate}
\end{examples}
A \emph{von Neumann algebra} (or $\operatorname{W}^*$-\emph{algebra}) $M$ is a unital $\operatorname{C}^*$-algebra which admits a faithful unital $\ast$-representation $\pi : M \to \mathrm B(\mathscr H)$ such that $\pi(M) \subset \mathrm B(\mathscr H)$ is closed with respect to the weak (equivalently strong) operator topology. After identifying $M$ with $\pi(M)$, we may regard $M \subset \mathrm B(\mathscr H)$ as a concrete von Neumann algebra. By von Neumann's bicommutant theorem, a unital $\ast$-subalgebra $M \subset \mathrm B(\mathscr H)$ is a von Neumann algebra if and only if $M$ is equal to its own bicommutant $M^{\prime\prime}$, that is, $M = M^{\prime\prime}$. There is a unique Banach space predual $M_\ast$ such that $M = (M_\ast)^*$. The ultraweak topology on $M$ coincides with the weak-$\ast$ topology arising from the identification $M = (M_\ast)^*$. A linear mapping between von Neumann algebras is \emph{normal} if it is continuous with respect to the ultraweak topology. We say that an action $\sigma : H \curvearrowright M$ is \emph{continuous} if the corresponding action map $H \times M_\ast \to M_\ast : (g, \varphi) \mapsto \varphi \circ \sigma_g^{-1}$ is continuous (see e.g.\ \cite[Proposition X.1.2]{Ta03a}). We then simply say that $M$ is a $H$-von Neumann algebra. The action $H \curvearrowright M$ is \emph{ergodic} if the fixed point von Neumann subalgebra $M^H = \left\{ x \in M \mid \forall g \in H, \sigma_h(x) = x \right \}$ is trivial.
\begin{examples}\label{ex:vN} We will consider the following examples of von Neumann algebras.
\begin{enumerate}
\item For any standard probability space $(X, \nu)$, the space $ \operatorname{ L}^\infty(X, \nu)$ of all $\nu$-equivalence classes of (essentially) bounded measurable functions endowed with the (essential) uniform norm $\|\cdot\|_\infty$ is a commutative von Neumann algebra. Any commutative von Neumann algebra arises this way. Any nonsingular action $H \curvearrowright (X, \nu)$ naturally gives rise to a continuous action $H \curvearrowright \operatorname{ L}^\infty(X, \nu)$ in the above sense. When no confusion is possible, we simply write $\operatorname{ L}^\infty(X) = \operatorname{ L}^\infty(X, \nu)$.
\item For any countable discrete group $\Lambda$ and any nonsingular action $\Lambda \curvearrowright (X, \nu)$ on a standard probability space, define the \emph{group measure space von Neumann algebra}
\begin{equation*}
\operatorname{ L}(\Lambda \curvearrowright X) \coloneqq \{ f \otimes 1, \pi(\gamma) \mid f \in \operatorname{ L}^\infty(X), \gamma \in \Lambda\}^{\prime\prime} \subset \mathrm B(\operatorname{ L}^2(X, \nu) \otimes \ell^2(\Lambda))
\end{equation*}
where $\pi : \Lambda \to \mathscr U(\operatorname{ L}^2(X, \nu) \otimes \ell^2(\Lambda))$ is the unitary representation defined by
\begin{equation*}
\forall \xi \in \operatorname{ L}^2(X, \nu), \forall \gamma, h \in \Lambda, \quad \pi(\gamma)(\xi \otimes \delta_h) = \sqrt{\frac{\mathrm{d}(\nu\circ \gamma^{-1})}{\mathrm{d}\nu}} \, \xi \circ \gamma^{-1} \otimes \delta_{\gamma h}.
\end{equation*}
When $(X, \nu)$ is a singleton, the von Neumann algebra $\operatorname{ L}(\Lambda \curvearrowright X)$ coincides with the \emph{group von Neumann algebra} $\operatorname{ L}(\Lambda)$. When the action $\Lambda \curvearrowright (X, \nu)$ is (essentially) free and ergodic, the von Neumann algebra $\operatorname{ L}(\Lambda \curvearrowright X)$ is a factor whose type coincides with the type of the action (see e.g.\ \cite[Theorem XIII.1.7]{Ta03b}).
\end{enumerate}
\end{examples}
A von Neumann algebra $M \subset \mathrm B(\mathscr H)$ is \emph{amenable} if there exists a norm one projection $\operatorname{ E} : \mathrm B(\mathscr H) \to M$. By Connes'\! fundamental result \cite{Co75}, $M$ is amenable if and only if $M$ is \emph{approximately finite dimensional}, that is, there exists an increasing net of finite dimensional subalgebras $M_i \subset M$ such that $\bigvee_{i \in I} M_i = M$.
\section{Dynamical dichotomy for boundary structures}\label{section:structures}
\subsection{Boundary structures}
For any $\operatorname{C}^*$-algebra $A\subset \mathrm B(\mathscr H)$ and any $n \geq 1$, $\mathrm M_n(A) \coloneqq \mathrm M_n(\mathbf{C}) \otimes A \subset \mathrm B(\mathscr H^{\oplus n})$ is naturally a $\operatorname{C}^*$-algebra. Let $A, B$ be any $\operatorname{C}^*$-algebras. A linear map $\Phi : A \to B$ is said to be \emph{unital completely positive} (ucp) if $\Phi$ is unital and if for every $n \geq 1$, the linear map $\Phi^{(n)} : \mathrm M_n(A) \to \mathrm M_n(B) : [a_{ij}]_{ij} \mapsto [\Phi(a_{ij})]_{ij}$ is positive. Any unital $\ast$-homomorphism $\pi : A \to B$ is a ucp map. When $A$ or $B$ is commutative, any unital positive linear map $\Phi : A \to B$ is automatically ucp (see e.g.\ \cite[Theorems 3.9 and 3.11]{Pa02}).
\begin{definition}[\cite{BBHP20}]
Let $\Gamma < G$ be any higher rank lattice and $M$ any $\Gamma$-von Neumann algebra with separable predual. A $\Gamma$-\emph{boundary structure} $\Phi : M \to \operatorname{ L}^\infty(G/P)$ is a $\Gamma$-equivariant faithful normal ucp map. We say that $\Phi$ is \emph{invariant} if $\Phi(M) = \mathbf{C} 1$.
\end{definition}
We will simply say that $\Phi : M \to \operatorname{ L}^\infty(G/P)$ is a \emph{boundary structure} instead of a $\Gamma$-boundary structure when it is understood that $M$ is a $\Gamma$-von Neumann algebra. In this survey, we only deal with higher rank lattices in connected semisimple \emph{real} Lie groups. In that setting, the notion of boundary structure is equivalent to the notion of stationary state. Indeed, fix a Furstenberg measure $\mu_\Gamma \in \operatorname{Prob}(\Gamma)$ so that $(G/P, \nu_P)$ is the $(\Gamma, \mu_\Gamma)$-Poisson boundary (see Theorem \ref{thm:poisson-Gamma}).
\begin{itemize}
\item If $\Phi : M \to \operatorname{ L}^\infty(G/P)$ is a boundary structure, then $\varphi \coloneqq \nu_P \circ \Phi \in M_\ast$ is a faithful normal $\mu_\Gamma$-stationary state on $M$. Moreover, if $\Phi$ is invariant, then $\varphi$ is $\Gamma$-invariant.
\item Conversely, let $\varphi \in M_\ast$ be any faithful normal $\mu_\Gamma$-stationary state on $M$. Define the $\Gamma$-equivariant faithful normal ucp map
$$\Phi : M \to \operatorname{Har}^\infty(\Gamma, \mu_\Gamma) : x \mapsto (\gamma \mapsto \varphi(\gamma^{-1} x)).$$
Since $\operatorname{Har}^\infty(\Gamma, \mu_\Gamma) \cong \operatorname{ L}^\infty(G/P, \nu_P)$ as $\Gamma$-operator systems, we may further regard $\Phi : M \to \operatorname{ L}^\infty(G/P)$ as a boundary structure such that $\varphi = \nu_P \circ \Phi$. If $\varphi$ is $\Gamma$-invariant, then $\Phi$ is invariant.
\end{itemize}
\begin{remark}
The notion of boundary structure was developed in \cite{BBHP20} to replace the notion of stationary state used in \cite{BH19} in order to deal with higher rank lattices in semisimple algebraic groups defined over \emph{arbitrary} local fields.
\end{remark}
It is useful to restrict boundary structures to separable $\operatorname{C}^*$-subalgebras. Let $M$ be any $\Gamma$-von Neumann algebra with separable predual. A globally $\Gamma$-invariant separable ultraweakly dense $\operatorname{C}^*$-subalgebra $A \subset M$ is called a \emph{separable model} for the action $\Gamma \curvearrowright M$. If $\Phi : M \to \operatorname{ L}^\infty(G/P)$ is a boundary structure, then $(\nu_P \circ \Phi)|_A \in \mathfrak S_{\mu_\Gamma}(A)$ and the restriction $\Phi|_A : A \to \operatorname{ L}^\infty(G/P)$ gives arise to the $\Gamma$-equivariant boundary map $\beta : G/P \to \mathfrak S(A) : b \mapsto \beta_b$ such that $\operatorname{Bar}(\beta_\ast \nu_P) = (\nu_P \circ \Phi)|_A$, where
\begin{equation*}
\forall a \in A, \quad \Phi(a)(b) = \beta_b(a).
\end{equation*}
We present several examples of boundary structures.
\begin{example}[Boundary structure arising from unitary representations]\label{example:boundary-rep}
Let $\pi : \Gamma \to \mathscr U(\mathscr H_\pi)$ be any unitary representation and set $A \coloneqq \operatorname{C}^*_\pi(\Gamma)$. Choose an extremal $\mu_\Gamma$-stationary state $\varphi \in \mathfrak S_{\mu_\Gamma}(A)$ and consider the GNS triple $(\pi_\varphi, \mathscr H_\varphi, \xi_\varphi)$. Denote by $\beta : G/P \to \mathfrak S(A) : b \mapsto \beta_b$ the $\Gamma$-equivariant boundary map associated with $\varphi \in \mathfrak S_{\mu_\Gamma}(A)$. By duality, we may consider the $\Gamma$-equivariant ucp map $\Phi : A \to \operatorname{ L}^\infty(G/P) : a \mapsto( b \mapsto \beta_b(a))$ which satisfies $\nu_P \circ \Phi = \varphi$. Set $M \coloneqq \pi_\varphi(A)^{\prime\prime} = (\pi_\varphi \circ \pi)(\Gamma)^{\prime\prime}$. By extremality, the conjugation action $\operatorname{Ad} (\pi_\varphi \circ \pi) : \Gamma \curvearrowright M$ is ergodic. Moreover, the $\Gamma$-equivariant ucp map
\begin{equation*}
\pi_\varphi(A) \to \operatorname{ L}^\infty(G/P) : \pi_\varphi(a) \mapsto \Phi(a)
\end{equation*}
is well defined and extends to a boundary structure $\Phi : M \to \operatorname{ L}^\infty(G/P)$. We refer to \cite[Proof of Theorem A]{BH19} for further details.
\end{example}
\begin{example}[Boundary structure arising from characters]\label{example:boundary-characters} Let $\varphi : \Gamma \to \mathbf{C}$ be any extremal character. Simply denote by $(\pi, \mathscr H, \xi)$ the GNS triple associated with $\varphi \in \operatorname{Char}(\Gamma)$. Denote by $J : \mathscr H \to \mathscr H : \pi(\gamma)\xi \mapsto \pi(\gamma)^*\xi$ the canonical conjugation. Following \cite{Pe14}, define the \emph{noncommutative Poisson boundary} $\mathscr B$ as the von Neumann algebra of all $\nu_P$-equivalence classes of (essentially) bounded measurable functions $f : G/P \to \mathrm B(\mathscr H)$ satisfying $f(\gamma b) = \operatorname{Ad}(J \pi(\gamma) J)(f(b))$ for every $\gamma \in \Gamma$ and almost every $b \in G/P$. Observe that $\mathbf{C} 1 \otimes \pi(\Gamma)^{\prime\prime} \subset \mathscr B$. Since $P$ is amenable, $\mathscr B$ is an amenable von Neumann algebra. By extremality, the conjugation action $\operatorname{Ad}(\pi) : \Gamma \curvearrowright \mathscr B$ is ergodic. Moreover,
\begin{equation*}
\Phi : \mathscr B \to \operatorname{ L}^\infty(G/P) : f \mapsto (b \mapsto \langle f(b)\xi, \xi\rangle)
\end{equation*}
is a boundary structure. When $\varphi = \delta_e$ is the regular character, the noncommutative Poisson boundary $\mathscr B$ coincides with the group measure space von Neumann algebra $\operatorname{ L}(\Gamma \curvearrowright G/P)$ and the boundary structure $\Phi : \operatorname{ L}(\Gamma \curvearrowright G/P) \to \operatorname{ L}^\infty(G/P)$ is the canonical $\Gamma$-equivariant conditional expectation. We refer to \cite[Proof of Theorem C]{BH19} for further details.
\end{example}
\begin{example}[Boundary structure arising from topological dynamics]\label{example:boundary-top}
Let $\Gamma \curvearrowright X$ be any minimal action on a compact metrizable space. Choose an extremal $\mu_\Gamma$-stationary Borel probability measure $\nu \in \operatorname{Prob}_{\mu_\Gamma}(X)$. By minimality, we have $\operatorname{supp}(\nu) = X$. Denote by $\beta : G/P \to \operatorname{Prob}(X) : b \mapsto \beta_b$ the $\Gamma$-equivariant boundary map associated with $\nu \in \operatorname{Prob}_{\mu_\Gamma}(X)$. By duality, we may consider the $\Gamma$-equivariant ucp map $\Phi : \operatorname{C}(X) \to \operatorname{ L}^\infty(G/P) : f \mapsto( b \mapsto \beta_b(f))$ which satisfies $\nu_P \circ \Phi = \nu$. By extremality, the nonsingular action $\Gamma \curvearrowright (X, \nu)$ is ergodic. Moreover, $\Phi : \operatorname{C}(X) \to \operatorname{ L}^\infty(G/P)$
extends to a boundary structure $\Phi : \operatorname{ L}^\infty(X, \nu) \to \operatorname{ L}^\infty(G/P)$.
\end{example}
The notion of boundary structure is well adapted to induction. Indeed, let $\Phi : M \to \operatorname{ L}^\infty(G/P)$ be any $\Gamma$-boundary structure. Denote by $\widehat M \coloneqq \operatorname{Ind}_\Gamma^G(M) \cong \operatorname{ L}^\infty(G/\Gamma) \mathbin{\overline{\otimes}} M$ the induced $G$-von Neumann algebra. Since $G/P$ is a $G$-space, we have $\operatorname{Ind}_\Gamma^G(\operatorname{ L}^\infty(G/P)) \cong \operatorname{ L}^\infty(G/\Gamma) \mathbin{\overline{\otimes}} \operatorname{ L}^\infty(G/P)$, where $G \curvearrowright G/\Gamma \times G/P$ acts diagonally. Denote by $\nu_{\Gamma} \in \operatorname{Prob}(G/\Gamma)$ the unique $G$-invariant Borel probability measure. Then the map $\widehat \Phi \coloneqq \nu_\Gamma \otimes \Phi : \widehat M \to \operatorname{ L}^\infty(G/P)$ is a $G$-equivariant faithful normal ucp map. We then refer to $\widehat \Phi$ as the \emph{induced} $G$-\emph{boundary structure}. Note that $\Phi$ is invariant if and only if $\widehat \Phi$ is invariant.
This framework provides a more conceptual approach to the \emph{stationary induction} considered in \cite[Section 4]{BH19}. Let $\mu_G \in \operatorname{Prob}(G)$ be any $K$-invariant admissible Borel probability measure. Let $\varphi $ be any faithful normal $\mu_\Gamma$-stationary state on $M$ and define the corresponding $\Gamma$-boundary structure $\Phi : M \to \operatorname{ L}^\infty(G/P)$ such that $\nu_P \circ \Phi = \varphi$. Consider the induced $G$-boundary structure $\widehat \Phi : \widehat M \to \operatorname{ L}^\infty(G/P)$. Then $\widehat \varphi \coloneqq \nu_P \circ \widehat \Phi$ is a faithful normal $\mu_G$-stationary state on $\widehat M$. Moreover, $\varphi$ is $\Gamma$-invariant if and only if $\widehat \varphi$ is $G$-invariant.
\subsection{The dynamical dichotomy theorem for boundary structures}
Let $A$ be any separable $\operatorname{C}^*$-algebra. We say that $\phi, \psi \in \mathfrak S(A)$ are \emph{pairwise singular} and write $\phi \perp \psi$ if there exists a sequence $(a_k)_k$ in $A$ such that $0 \leq a_k \leq 1$ for every $k \in \mathbf{N}$ and for which $\lim_k \phi(a_k) = 0 = \lim_k \psi(1 - a_k)$. This notion naturally extends the notion of pairwise singularity of Borel probability measures on metrizable compact spaces. Observe that for any unital $\operatorname{C}^*$-subalgebra $B \subset A$ and any states $\phi, \psi \in \mathfrak S(A)$, if $\phi|_B \perp \psi|_B$, then $\phi \perp \psi$. We introduce the following terminology.
\begin{definition}[\cite{BBHP20}]
Let $\Gamma < G$ be any higher rank lattice. Let $M$ be any $\Gamma$-von Neumann algebra with separable predual and $\Phi : M \to \operatorname{ L}^\infty(G/P)$ any boundary structure. We say that $\Phi$ is \emph{singular} if there exists a separable model $A \subset M$ for the action $\Gamma \curvearrowright M$ such that the corresponding $\Gamma$-equivariant boundary map $\beta : G/P \to \mathfrak S(A) : b \mapsto \beta_b$ satisfies the following property:
\begin{equation}\label{eq:singularity}
\text{For every } \gamma \in \Gamma \setminus \mathscr Z(\Gamma), \text{ for almost every } b \in G/P, \quad \beta_{\gamma b} \perp \beta_b.
\end{equation}
\end{definition}
The notion of singularity for boundary structures is quite robust. If $\Phi : M \to \operatorname{ L}^\infty(G/P)$ is singular, then for {\em every} separable model $A \subset M$, the corresponding $\Gamma$-equivariant boundary map $\beta : G/P \to \mathfrak S(A) : b \mapsto \beta_b$ satisfies \eqref{eq:singularity} (see \cite[Proposition 4.10]{BBHP20}). This implies the following useful fact. If $M_0 \subset M$ is a $\Gamma$-invariant von Neumann subalgebra and if the restriction $\Phi |_{M_0} : M_0 \to \operatorname{ L}^\infty(G/P)$ is singular, then $\Phi$ is singular as well.
In case the action $\Gamma \curvearrowright M$ is given by conjugation, singular boundary structures enjoy the following useful vanishing property.
\begin{proposition}[\cite{BBHP20}]\label{prop:dynamical-singular}
Let $M$ be any von Neumann algebra with separable predual and $\pi : \Gamma \to \mathscr U(M)$ any unitary representation. Consider the conjugation action $\operatorname{Ad}(\pi) : \Gamma \curvearrowright M$. Let $\Phi : M \to \operatorname{ L}^\infty(G/P)$ be any singular boundary structure. Then for every $\gamma \in \Gamma \setminus \mathscr Z(\Gamma)$, we have $\Phi(\pi(\gamma)) = 0$.
\end{proposition}
\begin{proof}
The proof is similar to \cite[Lemma 2.2]{HK17}. We may choose a separable model $A \subset M$ for the conjugation action $\operatorname{Ad}(\pi) : \Gamma \curvearrowright M$ such that $\pi(\Gamma) \subset A$. Denote by $\beta : G/P \to \mathfrak S(A)$ the $\Gamma$-equivariant boundary map arising from $\Phi|_A$. Let $\gamma \in \Gamma \setminus \mathscr Z(\Gamma)$ be any element. Choose a conull measurable subset $Y \subset G/P$ such that for every $b \in Y$, we have $\beta_{\gamma b} = \gamma \beta_b= \beta_b \circ \operatorname{Ad}(\pi(\gamma)^*)$ and $\beta_{\gamma b} \perp \beta_b$. Let $b \in Y$ be any point and choose a sequence $(a_k)_k$ in $A$ such that $0 \leq a_k \leq 1$ for every $k \in \mathbf{N}$ and for which $\lim_k \beta_{\gamma b}(a_k) = 0 = \lim_k \beta_b(1 - a_k)$. Then Cauchy--Schwarz inequality implies that
\begin{align*}
|\beta_b((1 - a_k)\pi(\gamma) )| &= |\beta_b((1 - a_k)^{1/2} \cdot (1 - a_k)^{1/2}\pi(\gamma) )| \\
&\leq \beta_b(1 - a_k)^{1/2} \to 0.
\end{align*}
Likewise, Cauchy--Schwarz inequality implies that
\begin{align*}
|\beta_{b}(a_k\pi(\gamma) )| &= |\beta_{\gamma b}(\pi(\gamma) a_k^{1/2} \cdot a_k^{1/2})| \\
&\leq \beta_{\gamma b}(a_k)^{1/2} \to 0.
\end{align*}
Then $\beta_b(\pi(\gamma)) = \beta_b((1 - a_k)\pi(\gamma) ) + \beta_b(a_k\pi(\gamma) ) \to 0$ and so $\beta_b(\pi(\gamma)) = 0$. Since this holds true for every $b \in Y$, it follows that $\Phi(\pi(\gamma)) = 0$.
\end{proof}
As we mentioned in the Introduction, the following dynamical dichotomy theorem \emph{invariant vs.\ singular} for boundary structures is the key novelty in our operator algebraic framework.
\begin{theorem}[\cite{BH19, BBHP20}]\label{thm:dynamical}
Let $M$ be any ergodic $\Gamma$-von Neumann algebra with separable predual and $\Phi : M \to \operatorname{ L}^\infty(G/P)$ any boundary structure. Then $\Phi$ is either invariant or singular.
\end{theorem}
The proof of Theorem \ref{thm:dynamical} depends heavily upon whether the ambient connected semisimple real Lie group $G$ is simple or not.
\textbf{In case $G$ is simple}, let us explain why Theorem \ref{thm:NCNZ} implies Theorem \ref{thm:dynamical}. Let $\Phi : M \to \operatorname{ L}^\infty(G/P)$ be any non-invariant boundary structure. By Theorem \ref{thm:NCNZ}, there exist a proper parabolic subgroup $P < Q < G$ and a $\Gamma$-equivariant unital normal embedding $\iota : \operatorname{ L}^\infty(G/Q) \hookrightarrow M$ such that $\Phi \circ \iota = p_Q^\ast$. Set $M_0 \coloneqq \iota(\operatorname{ L}^\infty(G/Q)) \subset M$. Then $A \coloneqq \iota(\operatorname{C}(G/Q)) \subset M_0$ is a separable model for the action $\Gamma \curvearrowright M_0$ and the $\Gamma$-equivariant boundary map corresponding to $\Phi|_A$ is exactly $\delta \circ p_Q : G/P \to \operatorname{Prob}(G/Q) : gP \mapsto \delta_{gQ}$. Since any element $\gamma \in \Gamma \setminus \mathscr Z(\Gamma)$ acts (essentially) freely on $G/Q$ (see e.g.\! \cite[Lemma 6.2]{BH19}), it follows that the restriction $\Phi|_{M_0} : M_0 \to \operatorname{ L}^\infty(G/P)$ is singular. Thus, $\Phi$ is singular.
\textbf{In case $G$ is not simple}, we have to use a different approach. Following \cite{BBHP20}, we outline the main steps of the proof of Theorem \ref{thm:dynamical} in the particular case where $G = G_1 \times G_2$ with $G_1$ and $G_2$ noncompact connected simple Lie groups with finite center. This particular case already contains all the main conceptual difficulties. Let $i = 1, 2$. Denote by $p_i : G \to G_i$ the canonical factor map. Denote by $P_i < G_i$ a minimal parabolic subgroup and set $P = P_1 \times P_2 <G$. By Theorem \ref{thm:poisson-G}, $(G_i, \nu_{P_i})$ is the $(G_i, \mu_i)$-Poisson boundary with respect to appropriate Borel probability measures $\mu_i \in \operatorname{Prob}(G_i)$ and $\nu_{P_i} \in \operatorname{Prob}(G_i/P_i)$. Let $\Phi : M \to \operatorname{ L}^\infty(G/P)$ be any non-invariant boundary structure. Our goal is to show that $\Phi$ is singular.
\textbf{Step 1: Induction.} Denote by $\widehat \Phi : \widehat M \to \operatorname{ L}^\infty(G/P)$ the induced $G$-boundary structure. Since $\Phi$ is not invariant, $\widehat \Phi$ is not invariant either, that is, $\widehat \Phi(\widehat M) \neq \mathbf{C} 1$.
\textbf{Step 2: Reduction to the von Neumann algebra of $G_1$-continuous elements.} Exploiting the product structure $G = G_1 \times G_2$ and up to permuting the indices, we show that the restriction of $\widehat \Phi$ to the $G_2$-fixed point von Neumann subalgebra $\widehat M^{G_2} \subset \widehat M$ still satisfies $\Phi(\widehat M^{G_2}) \neq \mathbf{C} 1$ and moreover $\widehat \Phi(\widehat M^{G_2}) \subset \operatorname{ L}^\infty(G_1/P_1)$. Exploiting that $\Gamma < G_1 \times G_2$ is irreducible and that $(G_i, \nu_{P_i})$ is the $(G_i, \mu_i)$-Poisson boundary, $i = 1, 2$, we show that the $G_1$-von Neumann algebra $\widehat M^{G_2}$ is $\Gamma$-isomorphic to the $\Gamma$-von Neumann subalgebra $M_1 \subset M$ of all elements $x \in M$ for which the action map $\Gamma \to M : \gamma \mapsto \sigma_\gamma(x)$ extends continuously to $G_1$. We say that $M_1 \subset M$ is the von Neumann subalgebra of $G_1$-\emph{continuous elements}. Moreover, under the identification $\widehat M^{G_2} = M_1$, we naturally have the identification $\widehat \Phi|_{\widehat M^{G_2}} = \Phi|_{M_1}$. Then $M_1$ is a $G_1$-ergodic von Neumann algebra and $\Phi|_{M_1} : M_1 \to \operatorname{ L}^\infty(G_1/P_1)$ is a $G_1$-boundary structure such that $\Phi(M_1) \neq \mathbf{C} 1$. Since $\Phi|_{M_1}$ is the restriction to $M_1$ of the $\Gamma$-boundary structure $\Phi$, it suffices to show that $\Phi|_{M_1}$ is singular.
\textbf{Step 3: Singularity of $\Phi|_{M_1}$.} We may choose a separable model $A_1 \subset M_1$ for the continuous action $G_1 \curvearrowright M_1$. Since $G_1 \curvearrowright G_1/P_1$ is transitive, $\Phi|_{A_1} : A_1 \to \operatorname{ L}^\infty(G_1/P_1)$ gives rise to a $G_1$-equivariant continuous boundary map $\beta : G_1/P_1 \to \mathfrak S(A_1) : b \mapsto \beta_b$. Since the action $G_1 \curvearrowright M_1$ is ergodic, the $P_1$-invariant state $\psi \coloneqq \beta_{P_1} \in \mathfrak S(A_1)$ is extremal among $P_1$-invariant states. We then show that for every $g \in G_1$, either $g\psi \perp \psi$ or $g \psi = \psi$. Since $\psi$ is not $G_1$-invariant on $A_1$, the stabilizer $Q_1 = \operatorname{Stab}_{G_1}(\psi)$ is a proper parabolic subgroup such that $P_1 < Q_1$. Since any element $g \in G_1 \setminus \mathscr Z(G_1)$ acts (essentially) freely on $G_1/Q_1$ and since $p_1(\Gamma \setminus \mathscr Z(\Gamma)) \subset G_1 \setminus \mathscr Z(G_1)$, it follows that the restriction $\Phi|_{M_1}$ is singular. Thus, $\Phi$ is singular.
\subsection{Outline of the proof of Theorem \ref{thm:NCNZ}}
Following \cite{BH19}, we outline the main steps of the proof of Theorem \ref{thm:NCNZ}. We may assume that $G$ is a connected simple real Lie group with trivial center and real rank $\operatorname{rk}_{\mathbf{R}}(G) \geq 2$. Recall that $G$ admits an Iwasawa decomposition $G = KAV$ where $K < G$ is a maximal compact subgroup, $A < G$ is a Cartan subgroup and $V < G$ is a unipotent subgroup. Set $P = LV$ where $L = \mathscr Z_G(A)$ and note that $P < G$ is a minimal parabolic subgroup. Likewise, write $\overline P = L \overline V$ for the opposite minimal parabolic subgroup. Note that $G = \langle P, \overline V\rangle$ and the map $\overline V \to G/P : \overline v \mapsto \overline v P$ defines a measurable isomorphism. Let $\Phi : M \to \operatorname{ L}^\infty(G/P)$ be any non-invariant boundary structure. Our goal is to show that there exist a proper parabolic subgroup $P < Q < G$ and a $\Gamma$-equivariant unital normal embedding $\iota : \operatorname{ L}^\infty(G/Q) \hookrightarrow M$.
\textbf{Step 1: Induction.} Exactly as in the proof of Theorem \ref{thm:dynamical}, denote by $\widehat \Phi : \widehat M \to \operatorname{ L}^\infty(G/P)$ the induced $G$-boundary structure which satisfies $\widehat \Phi(\widehat M) \neq \mathbf{C} 1$. We may choose a separable model $A \subset \widehat M$ for the continuous action $G \curvearrowright \widehat M$. Since $G \curvearrowright G/P$ is transitive, $\widehat \Phi|_{A} : A \to \operatorname{ L}^\infty(G/P)$ gives rise to a $G$-equivariant continuous boundary map $\beta : G/P \to \mathfrak S(A) : b \mapsto \beta_b$. Denote by $\psi \coloneqq \beta_P \in \mathfrak S(A)$ the corresponding $P$-invariant state, consider the GNS representation $\pi_\psi :A \to \mathrm B( \mathscr H_\psi)$ and set $N \coloneqq \pi_\psi(A)^{\prime\prime}$. Then $N$ is a $P$-von Neumann algebra and we may consider the induced $G$-von Neumann algebra $\operatorname{Ind}_P^G(N)$. Moreover, we may regard $\widehat M \subset \operatorname{Ind}_P^G(N)$ as a $G$-von Neumann subalgebra. We point out that the normal state $\psi \in N_\ast$ need not be faithful on $N$. Since we only give a sketch of the proof, we will assume that $\psi$ is faithful on $N$. We refer to \cite[Theorem 5.1]{BH19} for the general proof.
\textbf{Step 2: Construction of a well behaved von Neumann subalgebra.} In this step, we build upon Nevo--Zimmer's proof of \cite[Theorem 1]{NZ00}. Since $\widehat \Phi(\widehat M) \neq \mathbf{C} 1$, the $P$-invariant state $\psi \in \mathfrak S(A)$ is not $G$-invariant whence not $\overline V$-invariant. Using the real rank assumption $\operatorname{rk}_{\mathbf{R}}(G) \geq 2$, there is a strict intermediate parabolic subgroup $P < P_0 < G$ with Levi decomposition $P_0 = L_0 V_0$, where $L_0 = \mathscr Z_G(A_0)$, $A_0 < A$, and $V_0 < V$, such that $\psi \in \mathfrak S(A)$ is not $\overline V_0$-invariant. Choose a nontrivial element $s \in A_0$ so that $s$ (resp.\! $s^{-1}$) acts by conjugation as a contracting automorphism of $V_0$ (resp.\! $\overline V_0$). By Mautner phenomenon, any $s$-fixed element in $N$ is necessarily $V_0$-fixed. Since the subgroup $\langle s, V_0\rangle$ is normal in $P$, it follows that $N^s \subset N$ is a $P$-invariant von Neumann subalgebra. Using these assumptions, we show that $\mathscr M_0 = \widehat M \cap \operatorname{Ind}_P^G(N^s)$ is a $G$-von Neumann subalgebra such that $\widehat \Phi(\mathscr M_0) \neq \mathbf{C} 1$.
\textbf{Step 3: From noncommutative to commutative.} We reached the point where we can no longer rely on Nevo--Zimmer's argument \cite{NZ00}. Indeed, $\mathscr M_0$ is not commutative and so we cannot use the Gauss map trick from \cite[Section 3]{NZ00} to conclude. We adopt the following new strategy. Using the induction in two steps, we may write $\operatorname{Ind}_P^G (N^s) = \operatorname{Ind}_{P_0}^G (\operatorname{Ind}_P^{P_0}(N^s)) \cong \operatorname{ L}^\infty(\overline V_0, \operatorname{Ind}_P^{P_0}(N^s))$. Regarding $\mathscr M_0 \subset \operatorname{ L}^\infty(\overline V_0, \operatorname{Ind}_P^{P_0}(N^s))$ as a $G$-von Neumann subalgebra, denote by $\mathscr N_0 \subset \operatorname{Ind}_P^{P_0}(N^s)$ the von Neumann subalgebra generated by the essential values of all the elements $f \in \mathscr A_0$, where $\mathscr A_0 \subset \mathscr M_0$ is an ultraweakly dense separable $\operatorname{C}^*$-subalgebra. On the one hand, by construction, we have $\mathscr M_0 \subset \operatorname{ L}^\infty(\overline V_0, \mathscr N_0) \cong \operatorname{ L}^\infty(\overline V_0) \mathbin{\overline{\otimes}} \mathscr N_0$. On the other hand, exploiting that $s^{-1}$ acts by conjugation as a contracting automorphism of $\overline V_0$, we show that $\mathbf{C} 1 \mathbin{\overline{\otimes}} \mathscr N_0 \subset \mathscr M_0$. Therefore, we have the inclusions $\mathbf{C} 1 \mathbin{\overline{\otimes}} \mathscr N_0 \subset \mathscr M_0 \subset \operatorname{ L}^\infty(\overline V_0) \mathbin{\overline{\otimes}} \mathscr N_0$. By considering the centers, we also have $\mathbf{C} 1 \mathbin{\overline{\otimes}} \mathscr Z(\mathscr N_0) \subset \mathscr Z( \mathscr M_0 ) \subset \operatorname{ L}^\infty(\overline V_0) \mathbin{\overline{\otimes}} \mathscr Z(\mathscr N_0)$.
Since $G \curvearrowright \mathscr M_0$ is faithful and since $s$ acts identically on $\mathbf{C} 1 \mathbin{\overline{\otimes}} \mathscr N_0$, we have $\bfC1 \mathbin{\overline{\otimes}} \mathscr N_0 \subsetneq \mathscr M_0$. Exploiting Ge--Kadison's splitting theorem \cite{GK95}, we show that $\mathbf{C} 1 \mathbin{\overline{\otimes}} \mathscr Z(\mathscr N_0) \neq \mathscr Z( \mathscr M_0 )$. This further implies that $\Phi(\mathscr Z(\mathscr M_0)) \neq \mathbf{C} 1$. We may now apply Nevo--Zimmer's result \cite[Theorem 1]{NZ00} to the commutative $G$-von Neumann algebra $\mathscr Z(\mathscr M_0)$ to obtain a proper parabolic subgroup $P < Q < G$ and a $G$-equivariant unital normal embedding $\iota : \operatorname{ L}^\infty(G/Q) \hookrightarrow \mathscr Z(\mathscr M_0) \subset \widehat M$.
\textbf{Step 4: Back to the lattice.} We use the simple argument given in \cite{BBH21}. Since the action $G \curvearrowright \operatorname{C}(G/Q)$ is $\|\cdot\|$-continuous, by $G$-equivariance, it follows that $\iota(\operatorname{C}(G/Q))$ is contained in the $\operatorname{C}^*$-subalgebra $\operatorname{C}^*_b(G, M)^\Gamma \subset \widehat M$ of all $\Gamma$-equivariant bounded continuous functions $f : G \to M$. Consider the evaluation $\ast$-homomorphism $\mathsf{ev} : \operatorname{C}^*_b(G, M)^\Gamma \to M : f \mapsto f(e)$. Then $\mathsf{ev} \circ \iota : \operatorname{C}(G/Q) \to M$ is a $\Gamma$-equivariant $\ast$-homomorphism which extends uniquely to a $\Gamma$-equivariant unital normal embedding $\operatorname{ L}^\infty(G/Q) \hookrightarrow M$ thanks to \eqref{eq:uniqueness}.
\section{Proofs of the main results}\label{section:main}
In this section, following \cite{BH19, BBHP20}, we explain how to use the dynamical dichotomy Theorem \ref{thm:dynamical} to prove the main results stated in the Introduction. We fix a higher rank lattice $\Gamma < G$ and a Furstenberg measure $\mu_\Gamma \in \operatorname{Prob}(\Gamma)$ so that $(G/P, \nu_P)$ is the $(\Gamma, \mu_\Gamma)$-Poisson boundary (see Theorem \ref{thm:poisson-Gamma}). Since the proofs of Theorems \ref{thm:main} and \ref{thm:dynamics} are similar, we only give the proofs of Theorems \ref{thm:main}, \ref{thm:characters} and Corollary \ref{cor:C*}.
\begin{proof}[Proof of Theorem \ref{thm:main}]
Denote by $\pi : \Gamma \to \mathscr U(\mathscr H_\pi)$ the \emph{universal} unitary representation, meaning that $\pi$ is equal to the orthogonal direct sum of all cyclic unitary representations of $\Gamma$. Then $A \coloneqq \operatorname{C}^*_\pi(\Gamma)$ coincides with the full $\operatorname{C}^*$-algebra $\operatorname{C}^*(\Gamma)$ and we may use the identification $\mathfrak S(A) = \mathscr P(\Gamma)$. Let $\mathscr C \subset \mathfrak S(A)$ be any nonempty $\Gamma$-invariant weak-$\ast$ compact convex subset. We claim that any $\mu_\Gamma$-stationary state $\varphi \in \mathscr C$ is $\Gamma$-invariant. More generally, we claim that any $\mu_\Gamma$-stationary state on $A$ is $\Gamma$-invariant. By Krein--Milman theorem, it suffices to show that any extremal $\mu_\Gamma$-stationary $\varphi \in \mathfrak S(A)$ is $\Gamma$-invariant. Letting $M = \pi_\varphi(A)^{\prime\prime}$, consider the boundary structure $\Phi : M \to \operatorname{ L}^\infty(G/P)$ such that $\nu_P \circ \Phi = \varphi$ as in Example \ref{example:boundary-rep}. By Theorem \ref{thm:dynamical}, $\Phi$ is either invariant or singular. If $\Phi$ is invariant, then $\varphi \in \operatorname{Char}(\Gamma)$. If $\Phi$ is singular, then Proposition \ref{prop:dynamical-singular} implies that for every $\gamma \in \Gamma \setminus \mathscr Z(\Gamma)$, we have $\Phi(\pi(\gamma)) = 0$ and so $\varphi(\gamma) = 0$. Then $\varphi$ is supported on $\mathscr Z(\Gamma)$ and so $\varphi \in \operatorname{Char}(\Gamma)$.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:characters}]
Let $\varphi \in \operatorname{Char}(\Gamma)$ be any character. We may assume that $\varphi$ is an extremal character. Denote by $\mathscr B$ the noncommutative Poisson boundary and consider the boundary structure $\Phi : \mathscr B \to \operatorname{ L}^\infty(G/P)$ as in Example \ref{example:boundary-characters}. By Theorem \ref{thm:dynamical}, $\Phi$ is either invariant or singular. If $\Phi$ is invariant, then for every $f \in \mathscr B$, the function $\Phi(f) : G/P \to \mathbf{C} : b \mapsto \langle f(b)\xi, \xi\rangle$ is (essentially) constant. Since $\bfC1 \otimes \pi_\varphi(\Gamma)^{\prime\prime} \subset \mathscr B$ and since the linear span of $\pi_\varphi(\Gamma)\xi_\varphi$ is dense in $ \mathscr H_\varphi$, it easily follows that every $f \in \mathscr B$ is (essentially) constant as a function $G/P \to \mathrm B(\mathscr H_\varphi)$. This further implies that $\mathscr B = \mathbf{C} 1 \otimes \pi_\varphi(\Gamma)^{\prime\prime}$ and so $\pi_\varphi(\Gamma)^{\prime\prime}$ is amenable. This further implies that $\pi_\varphi$ is amenable. If $\Phi$ is singular, then for every $\gamma \in \Gamma \setminus \mathscr Z(\Gamma)$ we have $\Phi(\pi(\gamma)) = 0$ and so $\varphi (\gamma) = 0$. Thus, $\varphi$ is supported on $\mathscr Z(\Gamma)$.
If $G$ has property (T), then $\Gamma$ has property (T). If $\varphi$ is not supported on $\mathscr Z(\Gamma)$, then $\pi_\varphi$ is amenable and so $\pi_\varphi$ necessarily contains a finite dimensional subrepresentation. In the more general case where $G$ has a simple factor with property (T), we refer to the proof of \cite[Proposition 7.5]{BBHP20}.
\end{proof}
\begin{proof}[Proof of Corollary \ref{cor:C*}]
Let $\pi : \Gamma \to \mathscr U(\mathscr H_\pi)$ be any unitary representation and set $A \coloneqq \operatorname{C}^*_\pi(\Gamma)$. Regarding $\mathfrak S(A) \subset \mathscr P(\Gamma)$ as a $\Gamma$-invariant weak-$\ast$ compact convex subset, Theorem \ref{thm:main} implies that the $\operatorname{C}^*$-algebra $A$ admits a trace.
Assume moreover that $G$ has trivial center and that $\pi$ is not amenable. Let $\varphi \in \mathfrak S(A)$ be any trace. Regarding $\varphi \in \operatorname{Char}(\Gamma)$ as a character, the GNS representation $\pi_\varphi$ is weakly contained in $\pi$ and so $\pi_\varphi$ is not amenable. Theorem \ref{thm:characters} implies that $\varphi = \delta_e$. Then $\pi_\varphi = \lambda$ is the left regular representation and $\lambda$ is weakly contained in $\pi$. Denote by $\Theta : \operatorname{C}^*_\pi(\Gamma) \to \operatorname{C}^*_\lambda(\Gamma)$ the unique $\ast$-homomorphism such that $\Theta(\pi(\gamma)) = \lambda(\gamma)$ for every $\gamma \in \Gamma$. Then $\tau_\Gamma \circ \Theta$ is the unique trace on $A = \operatorname{C}^*_\pi(\Gamma)$. Let $J \lhd A$ be any proper ideal and define $\rho : A \to A/J$. Then the unitary representation $\rho \circ \pi$ is weakly contained in $\pi$ and so $\rho \circ \pi$ is not amenable. The previous reasoning implies that $\lambda$ is weakly contained in $\rho \circ \pi$ and so there is a $\ast$-homomorphism $\overline \Theta : A/J \to \operatorname{C}^*_\lambda(\Gamma)$ such that $\Theta = \overline \Theta \circ \rho$. Then $J = \ker(\rho) \subset \ker(\Theta)$. Therefore, $\ker(\Theta) \lhd \operatorname{C}^*_\pi(\Gamma)$ is the unique maximal proper ideal.
\end{proof}
We now discuss open problems in relation with our main results. Let $(E, \|\cdot\|)$ be any separable Banach $\Gamma$-module and $\mathscr C \subset E^*$ any nonempty $\Gamma$-invariant weak-$\ast$ compact convex subset. Let $\mu \in \operatorname{Prob}(\Gamma)$ be any probability measure. By analogy with \cite{Fu98}, we say that the affine action $\Gamma \curvearrowright \mathscr C$ is $\mu$-\emph{stiff} if any $\mu$-stationary point is invariant. The proof of Theorem \ref{thm:main} shows that for any Furstenberg measure $\mu_\Gamma \in \operatorname{Prob}(\Gamma)$, the affine action $\Gamma \curvearrowright \mathscr P(\Gamma)$ is $\mu_\Gamma$-stiff. It is natural to ask whether the stiffness property holds for more general probability measures $\mu \in \operatorname{Prob}(\Gamma)$.
\begin{problem}\label{prob:stationary}
Let $\mu \in \operatorname{Prob}(\Gamma)$ be \emph{any} probability measure such that $\langle \operatorname{supp}(\mu)\rangle = \Gamma$. Is the action $\Gamma \curvearrowright \mathscr P(\Gamma)$ $\mu$-stiff?
\end{problem}
Problem \ref{prob:stationary} requires a new strategy as we can no longer identify the $(\Gamma, \mu)$-Poisson boundary with $(G/P, \nu_P)$. We note that in case $\Gamma$ has trivial center, it is showed in \cite{HK17} that for any probability measure $\mu \in \operatorname{Prob}(\Gamma)$ such that $\langle \operatorname{supp}(\mu)\rangle = \Gamma$, the affine action $\Gamma \curvearrowright \mathfrak S(\operatorname{C}^*_\lambda(\Gamma))$ is $\mu$-stiff and the canonical trace $\tau_\Gamma$ is the only $\mu$-stationary state on $\operatorname{C}^*_\lambda(\Gamma)$. Problem \ref{prob:stationary} is particularly relevant in case $\Gamma \coloneqq \operatorname{SL}_d(\mathbf{Z}) < \operatorname{SL}_d(\mathbf{R}) \coloneqq G$ for $d \geq 3$. To draw a parallel with homogeneous dynamics, we point out that the affine action $\operatorname{SL}_d(\mathbf{Z}) \curvearrowright \operatorname{Prob}(\mathbf{T}^d)$ is $\mu$-stiff for every $d \geq 2$ and every probability measure $\mu \in \operatorname{Prob}(\operatorname{SL}_d(\mathbf{Z}))$ such that $\langle \operatorname{supp}(\mu)\rangle = \operatorname{SL}_d(\mathbf{Z})$ \cite{BFLM09}. We refer the reader to \cite{BFLM09} and \cite{BQ09} for more general results regarding measure rigidity.
We say that a countable discrete group $\Lambda$ is \emph{character rigid} if for any extremal character $\varphi \in \operatorname{Char}(\Lambda)$, either $\varphi$ is supported on $\mathscr Z(\Lambda)$ or the GNS representation $\pi_\varphi$ is finite dimensional. Assuming that $G$ has a simple factor with property (T), Theorem \ref{thm:characters} says that $\Gamma$ is character rigid. It is showed in \cite{PT13} that $\operatorname{SL}_d(\mathbf{Z}[S^{-1}])$ is character rigid for every $d \geq 2$ and every nonempty set of primes $S$. In view of Margulis'\! normal subgroup theorem which holds for arbitrary higher rank lattices, the next problem is of fundamental importance.
\begin{problem}\label{prob:NCSZ}
Let $\Gamma < G$ be \emph{any} higher rank lattice. Is $\Gamma$ character rigid?
\end{problem}
Problem \ref{prob:NCSZ} is also discussed in \cite[Question 2.1]{Ge18} for characters arising from ergodic probability measure preserving actions $\Gamma \curvearrowright (X, \nu)$, in connection with Stuck--Zimmer's results \cite{SZ92}. To answer positively Problem \ref{prob:NCSZ}, it would suffice to prove for every extremal character $\varphi \in \operatorname{Char}(\Gamma)$ that is not supported on $\mathscr Z(\Gamma)$, the tracial GNS factor $\pi_\varphi(\Gamma)^{\prime\prime}$ has property (T) in the sense of Connes--Jones \cite{CJ83}. This would correspond to the \emph{property} (T) \emph{half} in Margulis'\! normal subgroup theorem.
\section{Noncommutative factor theorem and Connes'\ rigidity conjecture}\label{section:NCFT}
Connes \cite{Co80} obtained the first rigidity result in von Neumann algebras by showing that for any icc (infinite conjugacy classes) countable discrete group $\Lambda$ with property (T), the type $\mathrm{II}_1$ factor $M = \operatorname{ L}(\Lambda)$ has countable outer automorphism group $\operatorname{Out}(M)$ and countable fundamental group $\mathscr F(M)$\footnote{The \emph{fundamental group} $\mathscr F(M)$ of a type $\mathrm{II}_1$ factor $M$ is defined as the subgroup of $\mathbf{R}^*_+$ that consists of all $\frac{\tau(p)}{\tau(q)}$, where $p, q \in M$ are projections for which $pMp \cong q M q$.}. This result prompted Connes to state the following bold conjecture (see \cite[Problem V.B.1]{Co94}).
\begin{connes}
Let $\Lambda_1$ and $\Lambda_2$ be any icc countable discrete groups with property (T) such that $\operatorname{ L}(\Lambda_1) \cong \operatorname{ L}(\Lambda_2)$. Show that $\Lambda_1 \cong \Lambda_2$.
\end{connes}
We say that a discrete group $\Lambda$ is W$^*$-\emph{superrigid} if whenever $\Upsilon$ is another discrete group such that $\operatorname{ L}(\Lambda) \cong \operatorname{ L}(\Upsilon)$, we have $\Lambda \cong \Upsilon$. Using \cite{CJ83}, Connes'\! rigidity conjecture asks whether every icc countable discrete group $\Lambda$ with property (T) is W$^*$-superrigid.
A first deep result towards Connes'\! rigidity conjecture was obtained by Cowling--Haagerup \cite{CH88} where they showed that for every $n \geq 2$ and every lattice $\Lambda$ in the rank one connected simple Lie group $\operatorname{Sp}(n, 1)/\{\pm 1\}$, the type $\mathrm {II}_1$ factor $\operatorname{ L}(\Lambda)$ retains the integer $n \geq 2$. For the last two decades, Popa's \emph{deformation/rigidity} theory \cite{Po06} has led to tremendous progress regarding classification and rigidity results for group (resp.\ group measure space) von Neumann algebras. In particular, the first examples of W$^*$-superrigid groups were obtained by Ioana--Popa--Vaes \cite{IPV10}. The examples constructed in \cite{IPV10} are generalized wreath products groups and so they do not have property (T). It is still an open problem to find an example of a W$^*$-superrigid countable discrete group $\Lambda$ with property (T). For other recent results regarding classification and rigidity results for von Neumann algebras, we refer the reader to the surveys \cite{Va10, Io12, Io18}.
Connes'\! rigidity conjecture is particularly relevant for the class of higher rank lattices. In this context, celebrated strong rigidity results by Mostow and Margulis (see \cite{Mo73} and \cite[Chapter VI]{Ma91}) show that whenever $\Gamma_i < G_i$ is a higher rank lattice, where $G_i$ has trivial center, $i = 1, 2$, if $\Gamma_1 \cong \Gamma_2$, then $G_1 \cong G_2$. In view of the strong rigidity results by Mostow and Margulis, we state the following version of Connes'\! rigidity conjecture for higher rank lattices.
\begin{conjecture}
For every $i = 1, 2$, let $G_i$ be any connected simple real Lie group with trivial center and real rank $\operatorname{rk}_{\mathbf{R}}(G_i) \geq 2$ and $\Gamma_i < G_i$ any lattice. If $\operatorname{ L}(\Gamma_1) \cong \operatorname{ L}(\Gamma_2)$, then $G_1 \cong G_2$.
\end{conjecture}
Popa's \emph{deformation/rigidity} theory cannot be used to tackle the above conjecture because higher rank lattices are somehow ``too rigid''. As suggested by Connes himself (see the discussion in \cite[Section 4]{Jo00}), it is natural to try and develop a strategy building upon the works of Furstenberg, Margulis and Zimmer. In what follows, we assume that $G$ is a connected simple real Lie group with trivial center and real rank $\operatorname{rk}_{\mathbf{R}}(G) \geq 2$. We fix a minimal parabolic subgroup $P < G$. Let $\Gamma < G$ be any lattice (e.g.\ $\Gamma \coloneqq \operatorname{PSL}_d(\mathbf{Z}) < \operatorname{PSL}_d(\mathbf{R}) \coloneqq G$ for $d \geq 3$). Then $\Gamma$ is icc and the group von Neumann algebra $\operatorname{ L}(\Gamma)$ is a type $\mathrm{II}_1$ factor. Moreover, the nonsingular action $\Gamma \curvearrowright (G/P, \nu_P)$ is (essentially) free and ergodic and the corresponding von Neumann factor $\operatorname{ L}(\Gamma \curvearrowright G/P)$ is amenable and of type $\mathrm{III}_1$ (see e.g.\ \cite[Proposition 4.7]{BN11}). We now give the proof of Corollary \ref{thm:NCFT} by combining Theorem \ref{thm:NCNZ} with Suzuki's results \cite{Su18}.
\begin{proof}[Proof of Corollary \ref{thm:NCFT}]
Denote by $\operatorname{ E} : \operatorname{ L}(\Gamma \curvearrowright G/P) \to \operatorname{ L}^\infty(G/P)$ the canonical $\Gamma$-equivariant conditional expectation. Let $\operatorname{ L}(\Gamma) \subset M \subset \operatorname{ L}(\Gamma \curvearrowright G/P)$ be any intermediate von Neumann subalgebra and consider the boundary structure $\Phi = \operatorname{ E}|_M : M \to \operatorname{ L}^\infty(G/P)$. By Theorem \ref{thm:NCNZ}, there are two cases to consider.
Firstly, assume that $\Phi$ is invariant. Following Examples \ref{ex:vN}(2), we simply denote by $u_\gamma \in \operatorname{ L}(\Gamma) \subset \operatorname{ L}(\Gamma \curvearrowright G/P)$ the canonical unitaries implementing the action $\Gamma \curvearrowright G/P$. For every $x \in M$, write $x = \sum_{\gamma \in \Gamma} x_\gamma u_\gamma$ for its Fourier expansion, where $x_\gamma = \operatorname{ E}(x u_\gamma^*)$ for every $\gamma \in \Gamma$. Since $\operatorname{ E}|_M = \Phi$ is invariant and since $\operatorname{ L}(\Gamma) \subset M$, it follows that $x_\gamma \in \mathbf{C} 1$ for every $\gamma \in \Gamma$ and every $x \in M$. This implies that $M = \operatorname{ L}(\Gamma)$ (see e.g.\ \cite[Lemma 6.8]{AHHM18}).
Secondly, assume that $\Phi$ is not invariant. Then there exist a proper parabolic subgroup $P < Q < G$ and a $\Gamma$-equivariant unital normal embedding $\iota : \operatorname{ L}^\infty(G/Q) \hookrightarrow M$ such that $\operatorname{ E} \circ \iota = p_Q^\ast$. This further implies that $\operatorname{ L}(\Gamma \curvearrowright G/Q) = \operatorname{ L}(\Gamma) \vee \operatorname{ L}^\infty(G/Q) \subset M$. Since the nonsingular action $\Gamma \curvearrowright (G/Q, \nu_Q)$ is (essentially) free (see e.g.\ \cite[Lemma 6.2]{BH19}), a combination of \cite[Theorem 3.6]{Su18} and \cite[Theorem IV.2.11]{Ma91} implies that there exists a parabolic subgroup $P < R < Q$ such that $M = \operatorname{ L}(\Gamma \curvearrowright G/R)$.
\end{proof}
It is well known that there are exactly $2^{\operatorname{rk}_{\mathbf{R}}(G)}$ intermediate parabolic subgroups $P < Q < G$. Thus, Corollary \ref{thm:NCFT} implies that there are exactly $2^{\operatorname{rk}_{\mathbf{R}}(G)}$ intermediate von Neumann subalgebras $\operatorname{ L}(\Gamma) \subset M \subset \operatorname{ L}(\Gamma \curvearrowright G/P)$. In particular, the inclusion $\operatorname{ L}(\Gamma) \subset \operatorname{ L}(\Gamma \curvearrowright G/P)$ retains the real rank $\operatorname{rk}_{\mathbf{R}}(G)$. We believe this observation could be useful to tackle Connes' rigidity conjecture and to show that the group von Neumann algebra $\operatorname{ L}(\Gamma)$ retains the real rank $\operatorname{rk}_{\mathbf{R}}(G)$.
|
\section{Introduction}
\par Both high sensitivity observations and simulations show that the matter content in the Universe is not randomly distributed but resides in a complex structure called the `cosmic web' \citep{2015JCAP...01..036J,2015A&A...580A.119V,2014MNRAS.445.3706V,2006astro.ph..8019L,2005Natur.435..629S,1996Natur.380..603B}. Billions of stars, quasars, galaxies, and intergalactic gas make up the pattern of thin walls or filaments inside the Large-Scale Structure (LSS). These filaments are on the scales of a few tens to hundreds of Mpc and are surrounded by large voids. Furthermore, dense and massive galaxy clusters are born at the intersection of filaments. Hence, it is essential to study galaxy clusters in the `cosmic web' environment to understand how the Universe evolved and how matter flows within filaments and interacts with galaxy clusters.
\par To date, mapping of the `cosmic web' was generally done using optical redshift surveys; finding densities in galaxy distributions was the only available tool to trace large scale filaments \citep{2014MNRAS.438.3465T}. Thankfully, technical enhancements in the observations of superclusters at other wavelengths such as radio band, have enhanced our knowledge of the cosmic web, its origin, evolution, physical condition, and properties \citep{2018MNRAS.479..776V, galaxies5010016, 2012MNRAS.423.2325A}. Further, there are many other critical questions that depend on the future high quality and sensitive multiwavelength observations to directly measure baryon density and its distribution in the cosmic web, as well as the total energy budget to drive the material inside filaments, magnetic field strength and overall growth of the structure.
\par In the last two decades, galaxy clusters have been extensively studied in the radio bands. Low-frequency radio observations (at 1 GHz and below) have revealed the existence of Mpc size diffuse radio sources associated with the cluster formation process \citep[and references therein]{2012A&ARv..20...54F}. These extended diffuse radio sources further split into two classes -- radio halos and relics \citep[and references therein]{2019SSRv..215...16V}. These two classes of sources are similar in their radio properties, aside from their positions and morphologies. Giant radio halos are spherical and can generally be found at the cluster centre or core, while radio relics are elongated in one direction and located at the cluster peripheries. Formation mechanisms of these radio sources are an open debate; however, observations hint that these radio sources are generated due to the turbulence (radio halos) and shock structures (radio relics) activated by the cluster merging process. {Recently, low-frequency observations have also revealed another class of diffuse radio source, know as `radio-ridge' found between two galaxy clusters in low dense regions \citep{2019Sci...364..981G,2019A&A...630A..77B}. The formation mechanism of this new class of radio source is not yet fully understood \citep{2020PhRvL.124e1101B}}.
\par In the present era, with the ability to conduct high sensitivity and angular resolution radio observations with next-generation telescopes, such as the MeerKAT \citep{2018ApJ...856..180C,2009IEEEP..97.1522J}, Australian Square Kilometer Array Pathfinder (ASKAP; \cite{2016PASA...33...42M}), and LOw Frequency ARray (LOFAR; \cite{2013A&A...556A...2V}), the time has arrived to utilise these telescopes to observe superclusters and filaments in order to detect shocks, particle acceleration and magnetic fields beyond cluster volume \citep{2020PhRvL.124e1101B,2019Sci...364..981G,2019A&A...630A..77B}. We have observed the {\it Saraswati} supercluster \citep{2017ApJ...844...25B} with the MeerKAT telescope to study the supercluster environment in radio frequency regime, characterise cluster merger shock and/or accretion shock related to relics, and mapping of diffuse radio sources with high dynamic range imaging.
\par With the advancement in radio interferometric observations, there is an increase in demand to produce the most sensitive radio maps with the highest dynamic range \citep{2011A&A...527A.108S}. In order to do so, one has to deal with many observational and data calibration errors, which place constraints on achieving such sensitivity. These errors include antenna pointing inaccuracy, time and frequency-dependent primary beam variation across fields, stability of the ionosphere over which observation is conducted, unknown radio frequency interference (RFI) sources, thermal noise and gain variation of receivers, to name a few. These time and frequency varying errors can also be related to one or more source-specific directions, known as direction-dependent errors. Sometimes strong point sources in the primary beam can limit the sensitivity required for observations. This problem is more severe when studying diffuse radio sources in the vicinity of a strong radio source(s) which hampered the observation by generating artefacts. Primary calibration processes (even self-calibration) are not adequate enough to fully characterise the properties of these strong radio sources, hence these poorly deconvolved sources challenge us in the imaging of faint and diffuse radio sources. One common solution to this issue is known as ``source peeling'' \citep{2004SPIE.5489..817N}. The general principle here is to obtain source-specific calibration solutions towards the strong source (which we need to be peeled) location or, in other words, apply direction-dependent calibration towards the particular source direction and solve the complex gain of that source and then subtracted it from the visibility. A number of software packages implement variations on this approach \citep[e.g.][]{2018MNRAS.478.2399K,Rioja_2018,2016ApJS..223....2V,2014ASInC..13..469I,2010A&A...524A..61N,2007ASPC..376..127M}, with the aim of producing the best scientific images.
\subsection{{\it Saraswati} supercluster}
\par The {\it Saraswati} supercluster was recently discovered in the famous Stripe 82 region \citep{2011AJ....142....3H} of the Sloan Digital Sky Survey (SDSS; \citealt{2000AJ....120.1579Y}), at redshift $z$ $\sim$ 0.3 \citep{2017ApJ...844...25B}. The total mass and size of this supercluster are $\sim$ 2 $\times$ 10$^{16}$ $M_\odot$ and $\sim$ 200 Mpc, respectively, making it one of the largest observed structures in the Universe. This massive supercluster hosts 43 galaxy clusters at a mean redshift of $z$ $\sim$ 0.28 and has an average density contrast of $\delta = 1.62$. Furthermore, it is surrounded by a complex network of galaxy filaments and large voids. There are a total of 24 voids of the size $\sim$ 40-170 Mpc identified in the {vicinity of the} supercluster. {The {clean spectroscopic} redshifts are derived from the LEGACY \citep{2009ApJS..182..543A}, BOSS \citep{2013AJ....145...10D} and SOUTHERN\footnote{https://www.sdss.org/dr12/algorithms/legacy_special_target/\#southern} (LBS) programs of the SDSS-III DR12 database \citep{2015ApJS..219...12A}}. A total of 3016 galaxies have been found within {and around} the {\it Saraswati} supercluster region \citep{2017ApJ...844...25B}.
\par The {\it Saraswati} supercluster's dense central (gravitationally bound) region extends to a radius of 20 Mpc and encompasses a mass of at least 4 $\times$ 10$^{15}$ $M_\odot$ (20\% of total supercluster mass) comprised of five massive galaxy clusters, including Abell 2631. These five massive clusters form the dense, bound core of the {\it Saraswati} supercluster. Abell 2631 is the most massive cluster located in the core of the {\it Saraswati} supercluster, and it is also extremely rich (richness class 3) and hot {($\sim$ 9.60 keV, \citealt{2009ApJS..182...12C})}. The complex merging galaxy cluster ZwCl 2341.1+0000, which is situated at $\sim$ 45 Mpc from the core, is the second massive cluster and is located in the trail of galaxies which makes up a southern filamentary network that is connected to the main central region. In the NRAO VLA Sky Survey (NVSS) data \citep{1998AJ....115.1693C}, it was found that ZwCl 2341.1+0000 hosts double giant diffuse radio relics connected by faint `radio-bridge'. The overall structure is indicative of the infall and merger dynamics of several galaxy groups/clusters during the initial phase of the cluster formation process \citep{2002NewA....7..249B}. Later on, in the Giant Meterwave Radio Telescope (GMRT) low-frequency observation of ZwCl 2341.1+000, double radio relics were detected, but there was no radio bridge or halo found between double relics (\citealt[][hereafter V09]{2009A&A...506.1083V}). However, there was marginalised faint radio emission detected between these two relics, along the filament network, in the Very Large Array (VLA) L-band observation (\citealt[][hereafter G10]{2010A&A...511L...5G}). However, due to lack of sensitivity in the VLA data, the nature of this faint emission could not be determined and could thus not be studied further. The estimated virial radii of both clusters are $\sim$ 1.93 Mpc and $\sim$ 1.62 Mpc, respectively, corresponding to the angular sizes of 8$'$ and 6.5$'$.
In this paper, we present our MeerKAT and GMRT (archival data) analysis results of the MOSS. Firstly, we focus on the direction-dependent calibration technique for MeerKAT data for improving our images. We also show the point source subtraction method used to analyse diffuse radio sources. Both these techniques make use of new in-house software.
\par As this section has provided a clear introduction to the research, the rest of the paper will be structured as follows: \S 2 gives details of the radio observation of the {\it Saraswati} supercluster; \S 3 describes the radio data reduction procedures; \S 4 presents our results; \S 5 gives the discussion and conclusion. In this paper, we have assumed $H_{0}$ = 70 km\,s$^{-1}$\,Mpc$^{-1}$, $\Omega_{\mathrm{M}}$ = 0.3 and $\Omega_{\Lambda}$ = 0.7. At redshift $z$ of 0.27, 1$''$ = 4.14 kpc, and luminosity distance $D_{\mathrm{L}}$ = 1376 Mpc.
\section{Radio observations}
\subsection{MeerKAT L-band observation}
\par MeerKAT is the precursor of the Square Kilometer Array (SKA) located in the Karoo desert of South Africa \citep{2018ApJ...856..180C,2009IEEEP..97.1522J}. We observed the central part of the {\it Saraswati} supercluster with the MeerKAT, which includes two pointings centred on each of the A2631 and ZwCl2341.1+0000 galaxy clusters. The separation between the two pointing is 1.5$^{\circ}$. MeerKAT observational details are provided in Table \ref{MK_observation}.
\begin{table}
\caption{MeerKAT observations of the {\it Saraswati} supercluster.}
\label{MK_observation}
\begin{tabular}{lll}
\hline
\hline
&Observation date: & 2019-06-15 \\
&Observation time: & 00:41:27.8 \\
&Number of pointings: & 2 \\
&Phase centre of A2631 & 23h37m40.6s\\
& & +0d16m36.0s\\
&Phase centre of ZwCl2341.1+0000 & 23h43m39.7s\\
& & +0d19m51.0s\\
&Number of antennas: & 60\\
&Total observation time:& 14hrs \\
&Central frequency: & 1283 MHz\\
&Total bandwidth: & 900 MHz \\
&Channel width: & 208 kHz \\
&Total number of channels: & 4016 \\
&Dump time: & 8s\\
&Cross products: & XX,XY,YX,YY\\
&Bandpass and flux calibrator: & J1939--6342 \\
&Gain calibrator: & J2357--1125\\
\hline
\end{tabular}
\end{table}
\subsection{GMRT 325 MHz observation}
\par ZwCl 2341.1+0000 was observed with the legacy GMRT on 24 June 2011 and 20 July 2011 at 325 MHz. To the best of our knowledge, this archival data set was not published before. The pointing of antennas were chosen at RA: 23h43m45.012482s, DEC: +00d18m00.08115s. The total observation time was 22 hours. This observation was performed using 256 channels, 30 MHz bandwidth and four cross products. To fix the absolute amplitude scale and bandpass calibration, 3C286 and 3C48 calibrators were observed, along with the J2340+135 (RA: 23h40m33.232373s; DEC: +13d33m00.98106s) phase calibrator.
\section{Data analysis}
\par We reduced both the MeerKAT and GMRT data sets using the \textbf{C}ontainerized \textbf{A}utomated \textbf{R}adio \textbf{A}stronomy \textbf{Cal}ibration (CARACal\footnote[1]{https://caracal.readthedocs.io/en/latest/}) pipeline \citep{2020arXiv200602955J}. The use of this pipeline is also mentioned in our previous work \cite{2020MNRAS.499..404P}. Briefly, the CARACal pipeline is written using the Stimela\footnote[2]{https://github.com/ratt-ru/Stimela} pipelining framework, which is python-based and uses containers (Docker and Singularity) specially developed for independent radio interferometry scripting, which allows users to use general radio astronomy software. This permits the combination of the appropriate radio astronomy software packages into a single pipeline. CARACal is an open-source and flexible pipeline that comes with pre-defined input settings for the reduction of data from MeerKAT and other telescopes like GMRT and JVLA in a variety of configuration files, which can be modified by the user as per their individual needs. We also made use of the Stimela package to carry out tasks such as excising discrete radio point sources (from {\it{uv}} domain) blended with extended emission and direction-dependent calibration. More details of both these methods are given in \S \ref{ddcal} and \S \ref{ptsub}.
\subsection{MeerKAT data analysis}
\par {We used standard bandpass and gain calibration tasks of Common Astronomy Software Application (CASA) for cross-calibration inside the CARACal pipeline}, with several rounds of automatic RFI flagging (before and after calibration). We then used the AOFlagger \citep{offringa-2012-morph-rfi-algorithm} software with a custom strategy optimised for MeerKAT data to remove the RFI. We also applied the mask of known RFI-affected channels to the data. In order to fix the absolute scale of the flux calibrator, we used the local sky model of the calibrator derived using the Reynolds scale \citep{ReyATNF.352..508R}. The absolute flux density of the primary or flux calibrator J1939--6342 is $\sim$ 14.90 Jy at the rest frequency of 1.4 GHz. After bootstrapping the flux density scale from the flux calibrator, the flux density of the secondary calibrator J2357--1125 is 2.07 $\pm$ 0.01 at 1.28 GHz (at 1.4 GHz, its value is $\sim$ 1.8 Jy and after scaling to 1.28 GHz with a spectral index value of -0.7, its expected value is $\sim$ {1.9 Jy}). Once the data were calibrated, we then split the two target sources from the multi-source measurement set. To reduce the data size and imaging time, we averaged every five channels in the target data. We then imaged ZwCl 2341.1+0000 and A2631 fields separately in the self-calibration rounds. In the pipeline, for direction-independent imaging and self-calibration tasks, we used the WSClean \citep{2014MNRAS.444..606O} and CubiCal \citep{2018MNRAS.478.2399K} software. We performed three phase-only self-calibration rounds to improve the antenna-based phase solutions. In the deconvolution procedure, we employed wide-band and multiscale cleaning options to better image the diffuse emission. We performed joint deconvolution on five sub-bands and generated a multi-frequency synthesis (MFS) cube (the bandwidth of each sub-band image is 170 MHz and has 164 channels). In joint deconvolution, we used the 2$^{nd}$ order polynomial (\texttt{nterms = 2}) in the spectral fitting. We made images with different Briggs robust weighting values to generate high- and low-resolution images. Further, we used these self-calibrated visibilities of ZwCl 2341.1+0000 and A2631 fields for direction-dependent calibration.
\subsection{GMRT data analysis}
\par For the 325 MHz GMRT archival data reduction, we also used the CARACal pipeline. We used the default settings to perform the initial flux and gain calibrations. For RFI removal, we used the \texttt{tfcrop} and \texttt{rflag} options of CASA, accessible inside the pipeline. For the flux calibrator, we used the Perley-Butler 2010 (hereafter PB10) scale. This gave the flux density of the flux calibrator 3C286 to be $\sim$ 24.83 Jy. The phase calibrator for this observation was 2340+135 (RA:23h40m33.22s, Dec:+13d33m00.92s). Similar to the MeerKAT data, we performed the joint deconvolution on five sub-bands (each sub-band of 6 MHz bandwidth with 51 channels) with multiscale cleaning and generated an MFS image using the 2$^{nd}$ order polynomial fitting. We performed three rounds of phase and two rounds of amplitude and phase self-calibration. The goal of the present data analysis is to test the CARACal pipeline for the GMRT antennas and tune the pipeline for uGMRT wide-band data analysis for future use.
\subsection{MeerKAT and GMRT flux calibration accuracy}
\label{cal_ac}
In order to obtain the calibration accuracy, we compared the flux densities of unresolved and discrete point radio sources between MeerKAT and NVSS observations of the ZwCl2341.1+0000 and A2631 fields. In both images, we manually selected isolated bright point sources (of flux density $>$ 1 mJy) to avoid any source blending or overlapping. We made a list of these point sources and used the CASA's \texttt{imfit} task to fit the Gaussian and measure the flux densities from both MeerKAT and NVSS images. In this analysis, we used the MeerKAT primary beam \citep{2021MNRAS.502.2970A} corrected images. We scaled the 1283 MHz MeerKAT flux density values to 1400 MHz with a spectral index value of -0.7. We have shown our flux densities comparison result in Figure \ref{NVSS_MeerKAT}. We found that the overall (mean) flux density accuracy between MeerKAT and NVSS is $\sim$ 10\% and median flux density accuracy $\sim$ 8\%.. For GMRT observations, this error is also estimated to be 10\% \citep{2004ApJ...612..974C}.
\begin{figure}
\includegraphics[scale=0.4]{NVSS_vs_MeerKAT_combine.png}
\caption{Total flux density comparison between MeerKAT and NVSS for ZwCl2341.1+0000 and A2631 fields. The solid black line shows the best-fit relationship between two observations and dashed red line shows the one-to-one relationship. Each side histogram shows the flux density distribution for MeerKAT and NVSS, with the blue line indicating the median value.}
\label{NVSS_MeerKAT}
\end{figure}
\subsection{Flux density error estimation and calibration scales}
\label{cal_scale}
\par In the procedure of uncertainty estimation on the total flux densities, two main factors should be evaluated. The first factor is the data calibration accuracy. From our estimation (\S\ref{cal_ac}), this calibration error was measured to be $\sim$ 10$\%$. The second factor is the noise present in the data. For extended radio sources, this error is characterised by the rms in the image multiplied with the square root of the ratio of the solid angle (of the extended source) to that of the synthesised beam, which is the number of beams across the extended source. In addition to this, we also included uncertainty due to point source subtraction from {\it uv} data \citep{2019SSRv..215...16V}. These three uncertainties factors are unrelated; therefore they are added in quadrature to estimate the absolute error on the flux densities of the extended sources, as shown below:
\begin{equation}
\Delta{S} = [(\sigma_{\mathrm{amp}}S)^2 + (\sigma_{\mathrm{rms}}\sqrt{n_{\mathrm{beams,ext}}})^2) + \sigma_{sub}^{2}]^{1/2},
\label{cal_err}
\end{equation}
\par where $S$ is the flux density, $\sigma_{\mathrm{amp}}$ is the flux calibration uncertainty (10\%), $\sigma_{\mathrm{rms}}$ is the image rms noise, and $n_{\mathrm{beams,ext}}$ is the number of beams in the full extent of the extended source, $\sigma_{sub}^{2}$ = $\sum_{i} (\sigma_{\mathrm{rms}}\sqrt{n_{\mathrm{beams,pt}}})^2$, where $n_{\mathrm{beams,pt}}$ is the number of beams in the subtracted point sources.
The GMRT (325 MHz) and MeerKAT flux densities are on the PB10 and \cite{ReyATNF.352..508R} scales, respectively. The agreement between these two scales is within 2\% (private communication with MeerKAT commissioning team).
For spectral index measurements (\S \ref{spxplot}), we took measured flux densities at 241 and 610 MHz from the V09. The authors used the \cite{1977A&A....61...99B} standard; hence we scaled the 325 MHz flux densities by the ratio between the \cite{1977A&A....61...99B} to PB10 standards.
\subsection{Direction-dependent calibration}
\label{ddcal}
\par Radio interferometric wide-band and wide-field of view observations are often dynamic range limited. A number of instrumental and observational errors affect the quality of such radio observations. These errors typically manifest themselves as artefacts associated with bright and strong sources in the observed field, particularly off-axis sources. These artefacts can extend throughout the map (being, in a sense, modulated by the extensive sidelobes of the radio interferometer point spread function (PSF)) and therefore hamper the detection of faint and diffuse emission and imaging of extended radio source(s) elsewhere in the map (including at the centre). There are several techniques available to minimise this problem by employing some form of direction-dependent calibration \citep[e.g.][]{2019RNAAS...3..110W,2011MNRAS.414.1656K,2011A&A...527A.107S, 2009A&A...501.1185I}. One of the first such methods proposed was ``source peeling'' which deals with strong sources by solving for the gain towards them one by one, followed by the source subtraction from the visibility ({\it uv}) data. A generalization of this is the ``differential gains'' technique \citep{2011A&A...527A.107S,2011A&A...527A.108S}, which can be seen as a form of simultaneous peeling, calibrating and correcting towards multiple bright sources simultaneously. In this work, we employed the CubiCal software \citep{2018MNRAS.478.2399K} in conjunction with the DDFacet imaging software \citep{2018A&A...611A..87T} to peel bright and strong sources. CubiCal solves the following Radio Interferometer Measurement Equation (RIME)
\begin{equation}
D_{p,q} = G_{p}(\sum_{i=1}^{n}dE^{(i)}_{p}P^{(i)}_{p}S_{p,q}^{(i)}P_{q}^{^{(i)}H}dE_{q}^{^{(i)}H}+S_{p,q}^{\mathrm{(di)}})G_{q}^{H}
\end{equation}
where $G$ terms are the direction-independent errors affecting the entire field of view. In our case, we constrain these to phase-only solutions (the equivalent of traditional phase selfcal). The $dE$ terms represent direction-dependent effects associated with $i=1...n$ bright sources. These use full complex 2x2 Jones matrix solutions \citep{2015MNRAS.449.2668S}. The $P$ terms implemented antenna-specific primary beam models (time and frequency varying). The $S^{(i)}$ terms denote \emph{intrinsic} (and direction-dependent) sky models for the individual sources being peeled, while $S^{\mathrm{(di)}}$ represents the ``direction-independent'' model for the rest of the emission in the field, derived from clean model components.
The sky model is generated using the DDFacet imaging software, and source-specific models (direction-dependent) are predicted using on-the-fly targeted faceting from subsets of the sky model. CubiCal can solve for both of these ($G$ and $dE$) errors simultaneously, while taking into account most of the sky flux. This limits the flux suppression of underlying radio sources (compact and extended sources).
Firstly, self-calibrated visibilities (i.e. with a $G$ solution only) are used to generate an initial sky model with DDFacet. We then designate regions (using the SAOImage ds9 viewer) corresponding to the individual bright sources that require peeling. CubiCal then uses these regions to split the sky model predict into $S^{(i)}$ and $S^{\mathrm{(di)}}$ components, and performs simultaneous $G$ and $dE$ solutions as per the RIME above. The bright sources (using their best-fitting $dE$ solutions) are subtracted from the visibilities, the residuals are corrected for $G$, and then re-imaged again using DDFacet. This procedure results in images with the bright sources (and associated artefacts) mostly removed, with improved image quality.
\subsection{Point source subtraction}
\label{ptsub}
\par In radio interferometry observations generally, unresolved and discrete point sources are blended with extended diffuse radio emission. Sometimes, the brightness of diffuse emission is obscured due to the presence of very bright point sources, and hence it is not easy to image the extended sources. In order to estimate the flux densities of extended sources, we need to subtract these point sources blended with the diffuse emission. Removing point sources is also an important task for generating a high-resolution spectral index map. There are various techniques available for excising point sources. A standard method is to use the box cleaning option in various clean algorithms. These boxes can be used for (affecting) strong point sources while cleaning and generating the model, which will later be subtracted from calibrated data. The other widely used method is to image the data with uniform weighting (and exclude the inner interferometry baselines that correspond to the angular size of the extended source that needs to be image) to generate the high-resolution map. This high-resolution data should contain only point sources and resolve out any extended emission. This will create models (of only point sources) which can then be subtracted from the calibrated data before re-imaging the residual visibility at lower resolution. Both methods, however, depend on the different angular scales of the underlying radio sources and are performed iteratively. They are, therefore, time-consuming when the data volume is large. Hence, it is necessary to develop a method that can subtract the point sources efficiently in a standard imaging procedure without more iterations. To do this, we used the Crystalball software\footnote[3]{https://github.com/paoloserra/crystalball} which can be employed in a Stimela script (or pipeline) to perform the point source subtraction. In this procedure, a user has to first make a region file that lists the coordinates of point sources which they want to remove. This region file can be manually generated via the SAOImage ds9 viewer. The user then runs WSClean with the \texttt{-save-source-list} option (in the final self-calibration imaging). This will generate a clean components file\footnote[4]{https://sourceforge.net/p/wsclean/wiki/ComponentList/}. Crystalball takes these two files, point source regions and clean component list, and generates another column in a given Measurement Set. This can be subtracted from the visibility and the residuals re-imaged. In our case, we made a source list from a high-resolution map of the ZwCl2341.1+0000 (uniform weighting and excluding inner MeerKAT baselines of 1 km resulting in 5$''$ beam size). In this high-resolution map, there are 1800 point sources that have been detected in the PyBDSF \citep{2015ascl.soft02007M} generated catalogue. We selected only compact sources i) detected in the high-resolution image (uniform weighting), ii) blended in double radio relics and nearby strong sources (Figure \ref{dd_labled_marked}) and iii) situated in between relics region. This technique of point source subtraction from visibilities is very efficient in the removal of many point sources simultaneously. The advantage of this method is that the Crystalball software does a direct Fourier transform (DFT) on the clean component list generated by the WSClean. This is more accurate than using model images because it can predict a smooth spectral polynomial at native channel resolution. The other advantage of this method is that all the required software are part of the Stimela container; hence one can implement all mentioned tasks in a pipeline.
\section{Results}
\par We have shown our MeerKAT results from Figure \ref{Xray_radio_img} to Figure \ref{radio_pt_img} as well as from Figure \ref{dd_labled_marked} to Figure \ref{A2631_dde_compt_src}. In this section, our main focus is to study the diffuse and extended radio sources after direction-dependent calibration. There are double radio relics that have already been discovered in the ZwCl 2341.1+0000 \citep{2002NewA....7..249B}. We show the MeerKAT and GMRT direction-independent images of double radio relics in Figure \ref{Xray_radio_img}. To compare and match the resolutions, we tapered the MeerKAT image (\texttt{robust = 0}) with 10$''$ to match the beam size of the GMRT image. As previously shown in the GMRT and VLA observations (G10,V09), there are two double radio relics; one in the north-west and another in the south-east directions clearly visible in the MeerKAT L-band and 325 MHz GMRT data. Similar to the previous observations, these two relics are situated perpendicular to the merger axis (north-south). Furthermore, they are located outside the hot and dense X-ray intra-cluster medium (ICM) region at the periphery of the merger. Both relics have a regular symmetric structure, slightly deviating from the usual elongated arc-like structure attributed to relics. We also noticed that the southern relic is larger (496 $\times$ 1241 kpc$^{2}$) in size and more elongated compared to the northern relic (270 $\times$ 426 kpc$^{2}$). The size of the southern relic is similar to the 610 MHz observation as measured by V09, but the linear size of the northern relic in the MeerKAT observation is larger than the 610 MHz measurements.
In the MeerKAT data, there is a strong and unresolved point source (RA:23:43:16.569, Dec:+0:26:20.392) located at a distance of $\sim$ 7$'$ ($\sim$ 1.7 Mpc) from the north relic (NR), $\sim$ 15$'$ ($\sim$ 3.7 Mpc) from the south relic (SR) and $\sim$ 9$'$ from the phase-centre (pointing centre) of the ZwCl 2341.1+0000 field, marked in the green box region in Figure \ref{dd_calibration} (left). The total flux density of this radio source is $\sim$ 87 mJy, and it causes artefacts and hence poor image quality around the central part of the cluster where two relics are situated. In Figure \ref{dd_calibration} (left) shows the CARACal image (final self-calibrated) generated with WSClean. As seen in this Figure, clearly the artefacts are affecting both north and south radio relics. Then in addition to this strong source, we also chose four different strong point sources or directions to apply the direction-dependent calibration. We used DDFacet to generate the facet based initial sky model required to calibrate the data in different directions. Figure \ref{dd_calibration} (middle) shows the DDFacet generated image which is still uncalibrated for distinct directions. Finally, Figure \ref{dd_calibration} (right) shows the direction-dependent calibrated image. We have seen clear improvement around the relic region before and after the direction-dependent calibration. We found a 25$\%$ rms noise decrease in the region between two relics after the direction-dependent calibration. All these maps are made at \texttt{robust = 0} value. This results in a FWHM of 8$''$ $\times$ 6$''$. Figure \ref{zwcl_dd_calibration} shows the full field of the ZwCl 2341.1+0000 before (left image) and after (right image) the direction-dependent calibration. In Figure \ref{zwcl_dde_compt_src}, we have shown four different point sources towards calibration applied and peeled the sources in order to improve the image quality. Similar to this, we also applied direction-dependent calibration to the A2631 field and results are shown in Figure \ref{A2631_dd_calibration} and Figure \ref{A2631_dde_compt_src}. In both of these fields, for every peeled source, we estimated local rms, minimum clean component value and sum of the total negative components around 8$'$ region. We have given these values in Table \ref{DD_stat}.
In the ZwCl2341.1+0000 radio map, we noticed discrete and blended point sources as shown in Figure \ref{dd_labled_marked}. Some of them are also reported in V09 which we labelled in Figure \ref{dd_labled_marked}(left). Figure \ref{dd_labled_marked}(right) shows the image of only discrete point sources, generated using the Crystalball (as mentioned in \S \ref{ptsub}). We subtracted these point sources from the MeerKAT (direction-dependent calibrated) and GMRT visibilities using our point source subtraction technique. We listed the properties of subtracted point sources in Table \ref{pt_list}. We have shown results in Figure \ref{radio_pt_img}(a) and (b) for MeerKAT and GMRT, respectively. We calculated flux densities of both radio relics and reported these values in Table \ref{relic_prop}. We also compared flux densities of compact radio sources before and after direction-dependent calibration and showed the result in Figure ~\ref{dd_calibration_compare}. We found an agreement within 2$\%$; if we include only bright point sources (of $>$ 1 mJy), then the agreement is within 0.5\%.
\par {In our forthcoming papers, we will give (1) a detailed analysis of these relics and the MeerKAT sub-band spectral index map (Kincaid et al. in prep.) and (2) radio source identifications and catalogue, spectral indices, and overall radio galaxy populations and their properties (for clusters - ZwCl 2341.1+0000 and A2631) in the core region of the Saraswati Supercluster (Parekh et al. in prep.)}
\begin{center}
\begin{figure*}
\centering
\begin{subfigure}[t]{0.45\textwidth}
\includegraphics[width=1\textwidth]{Radio_X-ray.png}
\caption{}
\label{rfidtest_xaxis1}
\end{subfigure}
\begin{subfigure}[t]{0.45\textwidth}
\includegraphics[width=1\textwidth]{MeerKAT_GMRT.png}
\caption{}
\label{rfidtest_yaxis2}
\end{subfigure}
\caption{(a) ZwCl 2341+0000 {\it Chandra} smoothed X-ray color image. The green contours show the MeerKAT observation. (b) GMRT 325 MHz radio contours on MeerKAT radio color image. In both images, the contour levels are [1,2,4,8,16,32,64,128] $\times$ 3$\sigma$. Negative contours are shown by the magenta color. For the MeerKAT radio image, the beam size is 12$''$ $\times$ 10$''$ and 1$\sigma$ = 20 $\mu$Jy beam$^{-1}$. In the GMRT radio image, the beam size is 11$''$ $\times$ 7$''$ and 1$\sigma$ = 30 $\mu$Jy beam$^{-1}$.}
\label{Xray_radio_img}
\end{figure*}
\end{center}
\begin{center}
\begin{figure*}
\centering
\includegraphics[width=1\textwidth]{ds9+new.png}
\caption{Results of direction-dependant calibration. The left image (only shown central part of ZwCl 2341.1+0000) is the self-calibrated direction-independent image generated from the CARACal. The middle image is the self-calibrated image generated using the DDFacet. The right image is the direction-dependent image (after calibrating towards five directions using the CubiCal). In the left image, we marked the bright and strong radio source in the green box, which affects the region of relics. All images have the same resolution of 8$''$ $\times$ 6$''$ and colour scale. We calculated the local rms, minimum pixel and sum of total negative component values inside the 8$'$ box size around the marked strong source before and after direction-dependent calibration. These values are listed in Table \ref{DD_stat}.}
\label{dd_calibration}
\end{figure*}
\end{center}
\begin{center}
\begin{figure*}
\centering
\begin{subfigure}[t]{0.55\textwidth}
\includegraphics[width=1\textwidth]{meerkat_pt_sub_dde_mask.png}
\caption{}
\label{rfidtest_xaxis1}
\end{subfigure}
\hspace{-2cm}
\begin{subfigure}[t]{0.55\textwidth}
\includegraphics[width=1\textwidth]{GMRT_ptsub.png}
\caption{}
\label{rfidtest_yaxis2}
\end{subfigure}
\caption{Discrete point sources around radio relics have been subtracted for (a) MeerKAT (direction-dependent image) and (b) GMRT (direction-independent image). For the MeerKAT radio image, the beam size is 8$''$ $\times$ 6$''$ and in the GMRT radio image, the beam size is 11$''$ $\times$ 7$''$.}
\label{radio_pt_img}
\end{figure*}
\end{center}
\begin{table}
\caption{Flux density measurements of diffuse radio sources.}
\label{relic_prop}
\begin{tabular}{ccccccccc}
\hline
\hline
Frequency (MHz) & 343 & 1283 \\
\hline
Synthesized beam ($''$ $\times$ $''$,$^{\circ}$) & 11 $\times$ 8, 79 & 8 $\times$ 6, -3 \\
rms ($\mu$Jy beam$^{-1}$) & 30 & 14 \\
\hline
North relic (mJy) & 18.0 $\pm$ 1.85 & 5.0 $\pm$ 0.51 \\
South relic (mJy) & 64.0 $\pm$ 6.41 & 17.0 $\pm$ 1.72\\
\hline
Synthesized beam ($''$ $\times$ $''$,$^{\circ}$) & & 50 $\times$ 50, 0 \\
rms ($\mu$Jy beam$^{-1}$) & & 77 \\
\hline
Halo (mJy) & - & 3.3 $\pm$ 0.50 \\
\hline
\end{tabular}
\end{table}
\vspace{-1.5cm}
\section{Is there diffuse halo present in ZwCl2341.1+0000?}
\par In the high-resolution (\texttt{robust = 0}) MeerKAT observation of ZwCl2341.1+0000, we could not detect any large radio emission between two radio relics. We applied {\it{uv}}-tapering to the MeerKAT data to match the resolution with the NVSS where \cite{2002NewA....7..249B} have been known to show the possibility of an extended radio bridge between relics. We generated images with different tapering using the direction-dependent calibrated- and point source subtracted data. At 50$''$ resolution, as shown in Figure ~\ref{mk_halo_img}, we detected a radio source between double radio relics. This suggests that the specific radio emission is resolved out in the high-resolution images. Based on its position, it could be a candidate radio halo, however, the nature of this radio source is not clear. The radio emission visible in the MeerKAT data is irregular and complex. This radio emission is patchy and faint (surface brightness of $\sim$ 0.13 mJy beam$^{-1}$) and extended around the central region of ZwCl2341.1+0000, connected with both radio relics. The detection level of this diffuse emission is only 2$\sigma$ in the present data. The rms of the image is 77 $\mu$Jy beam$^{-1}$ which is a factor of two better than previous detections in the VLA (L-band, D-configuration) as reported in G10, and five times better than the NVSS. Their beam sizes, however, were $\sim$ 80$''$ and 45$''$, respectively. We have not detected a halo emission in the 80$''$ beam size of the MeerKAT data (of which rms $\sim$ 0.2 mJy beam$^{-1}$). We found the size of the possible candidate halo (at 2$\sigma$ level) is $\sim$ 221$''$ $\times$ 286$''$ (914 $\times$ 1183 kpc$^{2}$) which is comparable to typical radio halo sizes. The flux density within 2$\sigma$ level is reported in Table \ref{relic_prop}. It is difficult to study its morphology and other radio properties in the current data. More data is needed to confirm this detection.
\begin{center}
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{MK_halo.png}
\caption{MeerKAT image of ZwCl 2341.1+0000 with 50$''$ beam size and rms sensitivity $1\sigma = 77 \mu$Jy beam$^{-1}$ (after direction-dependent calibration). The Contour levels are $[1,3,5,7,9,11] \times 2 \sigma$. The first contour is drawn at 2$\sigma$.}
\label{mk_halo_img}
\end{figure}
\end{center}
\section{Spectral index plot of relics and P1.4 GHz values}
\label{spxplot}
\par We have shown a spectral index plot of both radio relics in Figure ~\ref{1d_spx}. We derived the spectra between 241 to 1283 MHz. We took 241 and 610 MHz measurements from V09 and 323 and 1283 MHz from this work. For the SR, we combined flux densities of regions RS-1, RS-2, and RS-3 as labelled by V09. We matched the different calibration scales (\S \ref{cal_scale}) in order to estimate flux densities. {The minimum separation between antenna pairs of MeerKAT is 29 m, hence MeerKAT is insensitive to image a structure larger than $\sim$ 27$'$ at 1283 MHz. Similarly, GMRT is sensitive only up to $\sim$ 32$'$and 17$'$ at 325 and 610 MHz bands, respectively. This should be considered while comparing flux densities of extended sources using different arrays. In our case, the sizes of both relics are less than these largest detectable structures of different data sets.} We found a straight power-law fit for the SR of spectral index value $\alpha_{\mathrm{south}}$ = -1.01 $\pm$ 0.08. For the NR, we found a break at 610 MHz, and the single power-law model ({red solid line in Figure~\ref{1d_spx}, of $\alpha$ = -0.98 $\pm$ 0.10}) does not fit the data. We found a flat spectrum between 241 to 610 MHz, which became steep thereafter. We measured two slopes, $\alpha_{\mathrm{1,north}}$ = -0.51 $\pm$ 0.10 and $\alpha_{\mathrm{2,north}}$ = -1.38. The value of $\alpha_{\mathrm{1,north}}$ is similar to the value given by V09 ($\alpha_{\mathrm{north}}$ = -0.49 $\pm$ 0.18).
\par For double radio relics and the candidate halo, we estimated its K-corrected (rest-frame) radio power ($P_{1.4\mathrm{GHz}}$) using the following:
\begin{equation}
P_{1.4\mathrm{GHz}} = \frac{4 \pi D_{\mathrm{L}}^{2}S_{v}}{(1+z)^{1+\alpha}},
\end{equation}
\par where $D_{\mathrm{L}}$ is the source luminosity distance at $z$ = 0.27, and $S_{v}$ is the source flux density. For the north radio relic, $\alpha$ $\sim$ -1.3 (between 610 and 1283 MHz) gives radio power at 1.4 GHz to be $P_{1.4\mathrm{GHz}}$ = (1.08 $\pm$ 0.11) $\times$ 10$^{24}$ W Hz$^{-1}$. For the south radio relic, $\alpha$ $\sim$ -1.0 gives radio power to be $P_{1.4\mathrm{GHz}}$ = (3.53 $\pm$ 0.35) $\times$ 10$^{24}$ W Hz$^{-1}$. For radio halo, $P_{1.4\mathrm{GHz}}$ = (7.17 $\pm$ 1.08) $\times$ 10$^{23}$ W Hz$^{-1}$. We used $\alpha$ = -1.3 to scale the halo flux density to 1.4 GHz.
\begin{center}
\begin{figure}
\centering
\includegraphics[width=0.4\textwidth]{Spectral_plot_1D_new.png}
\caption{1D spectral index plot for north relic (NR) and south relic (SR). {The red solid line shows the best-fit line between 241 and 1283 MHz for the NR}.}
\label{1d_spx}
\end{figure}
\end{center}
\section{Discussion and conclusion}
\par Using the MeerKAT telescope, we observed the centrally surrounded two massive galaxy clusters, A2631 and ZwCl 2341.1+0000, of the {\it Saraswati} supercluster. In this work, we aimed to test the direction-dependent calibrations on wide-field of view and broadband MeerKAT data and show the improvement in diffuse radio source imaging. Furthermore, we also studied particular of the ZwCl2341.1+0000 merging galaxy cluster. The cluster is elongated and displays the disturbances in X-ray emission. For more details about the dynamics and merging geometry of the cluster, we refer to \cite{2014MNRAS.443.2463O}. Previously, in low frequency and high-resolution GMRT data, two radio relics have been detected (V09). NVSS and low-resolution VLA data also suggested the presence of relics and extended diffuse radio source in the cosmic web environments in overlap with a chain of galaxies.
\subsection{Direction-dependent effects}
\par The peeling techniques we have employed to improve the image qualities of our MeerKAT data are discussed in \S \ref{ddcal}. In our direction-dependent analysis, we found that within 35$'$ from the phase centre, strong, poorly deconvolved sources can be peeled quite well, with considerable attenuation of the associated artefacts.
Figure \ref{zwcl_dde_compt_src} shows the peeled sources, ordered in increasing distance from centre. Our first peeled source (shown in Figure \ref{dd_calibration}) is 8$'$ off-axis, and was dramatically improved by the direction-dependent calibration. The other four sources (Figure \ref{zwcl_dde_compt_src}(a), (b), (c) and (d)) are at distances of 12$'$, 26$'$, 31$'$ and 38$'$, respectively. As one can see, sources within $<$ 35$'$ can be improved. For the other field of A2631, the three selected strong sources (Figure \ref{A2631_dde_compt_src}(a), (b) and (c)) are at distances of 23$'$, 30$'$ and 51$'$, respectively, from the phase-centre. As one can see, the level of artefacts around the first two sources are improved after the direction-dependent calibrations, whilst residual artefacts are still visible (albeit at a low level) for the third source. As a quantitative measure, we calculate the local rms, minimum pixel and sum of total negative pixel values in an $8'\times 8'$ box centred on each peeled source (Table \ref{DD_stat}). The rationale for looking at negative pixels is that imaging artefacts due to calibration errors manifest themselves in a positive-negative pattern (fundamentally, because PSF sidelobes are positive-negative), while physical emission is all positive. The distribution of the negative pixels is therefore indicative of the artefact level. As we can see in Table \ref{DD_stat} (sources are given in order of increasing distance from centre), there are clear decrements in the (absolute) sum of negative values after direction-dependent calibration. Local rms is also improved after the direction-dependent calibration, and (absolute) minimum pixel values are also reduced after the direction-dependent calibration, although this does not seem to be a significant indicator.
\par The success of the peeling process will also depend on source location in the given image as well as source structure. Since MeerKAT provides wide-bandwidth observations, it is important to account for the variation of the primary beam across the broad frequency range. The size of the primary beam and its half-power points vary radially with the frequencies within the MeerKAT band. Furthermore, if antenna beam patterns are not modelled well across the wide bandwidth, then accuracy in the estimate of the flux densities of off-axis sources (away from the phase-centre) decrease \citep{10.1093/mnras/stz702}. Hence, the calibration process computes incorrect gains for these sources. This makes it difficult to improve residual artefacts around strong sources situated at the edges of image. With the advancement in MeerKAT future beam modeling, we can hope to further alleviate the artefacts associated with such edge sources.
\subsection{Diffuse radio sources in ZwCl2341.1+0000}
\par The candidate radio halo (or bridge) in ZwCl2341.1+0000 was first detected in the NVSS survey which provides a beam size of $\sim$ 45$''$ and was not detected in high-resolution GMRT observation, even at low frequencies. In low-resolution VLA L-band observation of ZwCl2341.1+0000 (G10), it was not possible to separate three diffuse components of the north and south radio relics and candidate radio halo lying between them. Our MeerKAT L-band observation provides the highest resolution and sensitivity compared to the previous 1.4 GHz observations. We could not detect the candidate radio halo in a high-resolution (8$''$ $\times$ 6$''$) MeerKAT image. However, we do see a tentative detection of it at a 2 sigma significance level after tapering the MeerKAT image to a 50$''$ beam size. The candidate halo emission is resolved out in high-resolution data and could be detected with compact array configurations only where extended radio emissions have higher surface brightness. The radio halo emission in the MeerKAT observation is very irregular and patchy. The rms statistics and detection level are better in the direction-dependent tapered image than the direction-independent image. The tapering of the {\it uv} data increases the rms noise; hence we could not taper to larger beam sizes to improve the surface brightness of the halo emission. Based on the cluster X-ray mass and halo size scaling relationship \citep{2013ApJ...777..141C,2007MNRAS.378.1565C}, the expected linear size of the halo is $\sim$ 756 kpc, however, the largest linear size of the halo that we could measure was $\sim$ 1183 kpc. This is comparable to the size estimated by G10. The flux density measurement of the halo in MeerKAT observation is $\sim$ 3 times lower than that reported by G10. This could be due to the difference in beam sizes and array configurations of MeerKAT and VLA. The expected power of the radio halo based on the scaling relation \citep{2013ApJ...777..141C} is $\sim$ 8.5 $\times$ 10$^{23}$ W Hz$^{-1}$, which is close to our measurement (7.17 $\pm$ 1.08) $\times$ 10$^{23}$ W Hz$^{-1}$, in MeerKAT data. {It has been shown that the radio power of giant radio halos is correlated with the X-ray masses of their host clusters \citep{osti_22518716,2000ApJ...544..686L}. Based on this, the radio power of the candidate halo of ZwCl 2341.1+0000 is consistent with the $P_{1.4GHz}$ vs Mass relation of giant halo clusters}. However, the nature of this candidate radio halo emission is still unclear and it could also be related to the cosmic web network of the {\it Saraswati} supercluster. Its formation mechanism could also be related to the LSS formation instead of cluster merger. More data is required to characterise the properties of this puzzling candidate radio halo.
\par G10 have reported the total flux density of all three components as $\sim$ 28.5 mJy. In our observation, we found total flux densities of the double radio relics and halo are = 25.3 $\pm$ 1.86 mJy, which is comparable. G10 measured the total linear size of all three diffuse components as $\sim$ 2.2 Mpc, while we estimated the total linear size as $\sim$ 2.4 Mpc. We also noticed that the NR (in MeerKAT observation) has a larger linear size than compared to the 610 MHz data (V09), but we also note that these observations have different sensitivities and background rms values. This could suggest that our sensitive MeerKAT observations have detected more diffuse emissions in the northern radio relic (towards the north direction) than previous observations. { Based on radio relics cluster $P_{1.4GHz}$ vs Mass relationship \citep{2017MNRAS.472..940K}, the radio power of both NR and SR are consistent with this relation}. Furthermore, G10 reported the spectral index value between 610 and 1400 MHz for north and south relics as $\sim$ -1.2. We fitted a straight line with spectral index, $\alpha_{1283\mathrm{}}^{610\mathrm{}}$ $\sim$ -1.2 for the SR, which is comparable with G10. The spectral index value for SR, $\alpha_{1283\mathrm{}}^{241\mathrm{}}$ = -1.01 $\pm$ 0.08. {Based on radio relic spectral indices compilation \citep{2021MNRAS.506..396W}, we found an average spectral index value of known radio relics is $\sim$ -1.2. The spectral index value of the SR is comparable to this average value}. The NR is faint compared to the SR, and its spectrum appears complex (Figure \ref{1d_spx}). There is a spectral break at 610 MHz with a steeper spectrum ($\alpha_{1400\mathrm{}}^{610\mathrm{}}$ $\sim$ -1.38) towards the high frequency and a flattening towards lower frequencies ($\alpha_{610\mathrm{}}^{241\mathrm{}}$ = -0.51 $\pm$ 0.10). {This spectral break, if confirmed with further sensitive low frequency data, then is very unusual and inconsistent with the diffusive shock acceleration (DSA) mechanism \citep{1983RPPh...46..973D}, which advocates the single power-law distributions for relic spectra. In the last decade, several radio relics have been observed over $>$ 1 GHz frequency bands \citep{2015A&A...575A..45T,2014MNRAS.441L..41S,2014ApJ...794...24O,2013A&A...555A.110S}. It was found that the A2256 also shows uncommon spectral break at $>$ 1.4 GHz \citep{2015A&A...575A..45T}. This indicates the halting of injection of electrons in the relic system, which eventually leads to cutoff at high frequencies of the integrated spectrum. The alternative explanation could be for this spectral break is clusters merging process is happening recently (young merger) where relics have just formed, and the electrons have not yet had time to lose their energy to steepen the spectrum. We will study this spectral break in more detail in our forthcoming paper (Kincaid et al. in prep.).} We emphasize that all different instruments (MeerKAT, VLA and GMRT) have different {\it uv} coverage, antenna configurations, observation conditions and sensitivities; hence our spectral indices comparisons with other observations are affected by these constraints.
\par In this work, we have shown the direction-dependent calibration technique for galaxy cluster fields. It both improves the image quality and allow us to estimate the flux densities of diffuse radio sources after eliminating artefacts caused by strong sources. This method is important in the wide-bandwidth and wide-field of view observations of superclusters or galaxy clusters in order to detect faint and diffuse radio emissions in the presence of strong radio source(s). This technique can be implemented in a pipeline and can be applied to other data. In our next paper, we will study the merger shock properties in the supercluster environment and broadband spectral index studies of diffuse radio sources.
\\
\\
\noindent{\it Acknowledgements.} The financial assistance of the South African Radio Astronomy Observatory (SARAO) towards this research is hereby acknowledged. The MeerKAT telescope is operated by the South African Radio Astronomy Observatory, which is a facility of the National Research Foundation, an agency of the Department of Science and Innovation. This work is based upon research supported by the South African Research Chairs Initiative of the Department of Science and Technology and National Research Foundation. V. Parekh acknowledge Hiroki Akamatsu, Bill Cotton and Annalisa Bonafede for fruitful discussion. (Part of) the data published here have been reduced using the CARACal pipeline, partially supported by ERC Starting grant number 679629 ``FORNAX", MAECI Grant Number ZA18GR02, DST-NRF Grant Number 113121 as part of the ISARP Joint Research Scheme, and BMBF project 05A17PC2 for MeerKAT. Information about CARACal can be obtained online under the URL: https://caracal.readthedocs.io.
RK acknowledges the support of the Department of Atomic Energy, Government of India
under project no. 12-R\&D-TFR-5.02-0700. SS acknowledges the support of the European Regional Development Fund and the Mobilitas Pluss postdoctoral research grant MOBJD660.
\\
\noindent{\it Data Availability.} The data underlying this article are subject to an embargo. Once the embargo expires the data will be available [https://archive.sarao.ac.za/; https://www.sarao.ac.za/wp-content/uploads/2019/12/MeerKAT-Telescope-and-Data-Access-Guidelines.pdf].
|
\section{Introduction}
The structure of the set of rational proper maps between complex unit balls is a very classical topic in Several Complex Variables. Among the many unsolved problems in this topic, there is the well-known \textit{gap phenomenon}, which will be recalled now. Fix an integer $n\geq 2$. For each $k\in\mathbb N^+$ such that $k(k+1)/2<n$, define the closed interval $\mathcal I_k:=[kn+1,(k+1)n-\frac{k(k+1)}{2}-1].$ The classical theorem of Faran~\cite{faran} amounts to saying that when $N\in\mathcal I_1=[n+1,2n-2]$, any local holomorphic map sending an open piece of $\partial\mathbb B^n$ to $\partial\mathbb B^N$ actually maps $\partial\mathbb B^n$ to a linear section $\partial\mathbb B^n\subset\partial\mathbb B^N$. In other words, there are no ``new'' maps when $N$ increases from $n$ to $2n-2$.
Then, it was discovered by Huang-Ji-Xu~\cite{hjx} that the same phenomenon holds for $N\in\mathcal I_2=[2n+1,3n-4]$ and later by Huang-Ji-Yin~\cite{HJY} for $N\in\mathcal I_3=[3n+1,4n-7]$. The \textit{Gap Conjecture}, formulated in~\cite{HJY2}, states that the gap phenomenon holds whenever $N\in\mathcal I_k$.
In this article we are going to establish the existence of similar gaps for all levels \textit{at once} and also to demonstrate the gap phenomenon actually holds for \textit{all} generalized balls (whose definition will be recalled below).
\begin{theorem}\label{intro thm 2}
Let $k,n\in\mathbb N^+$ such that $n>k(k+1)$. For the local proper holomorphic maps between generalized balls, the gap phenomenon holds over the intervals
$$
\mathcal J_k:=[kn+k, (k+1)n-(k^2+1)].
$$
\end{theorem}
For Theorem~\ref{intro thm 2}, what we are going to prove is that when the target dimension is within $\mathcal J_k$, then the image must lie in a hyperplane (Theorem~\ref{gap thm}). Thus, in the case of ordinary unit balls, it can be directly interpreted as there are no new proper holomorphic maps within the gaps, as described above. However, we will see that formulating precisely the gap phenomenon for all generalized balls as ``no new maps within the gaps" needs a bit more work and this will be done in Section~\ref{proper map section}. There the theorem above will be stated more precisely as Theorem~\ref{gap thm 2}.
Note that although the interval $\mathcal J_k$ in our theorem is smaller than the $\mathcal I_k$ in the original Gap Conjecture, this is to be expected since our theorem holds for \textit{all generalized balls}. As a matter of fact, the lower bound for $\mathcal J_k$ is sharp in the present context, as will be demonstrated after Theorem~\ref{gap thm 2}. On the other hand, we do not know at this point whether the upper bound for $\mathcal J_k$ is sharp for generalized balls.
Our proof for Theorem~\ref{intro thm 2} consists of two main ingredients: the orthogonality preserved by the relevant proper maps; and a hyperplane restriction theorem for holomorphic mappings. Regarding the study of orthogonality, it originated from an earlier work~\cite{GN} of the authors. There we proposed a coordinate free approach to the rigidity problems related to real hyperquadrics on the projective space and generalized a number of well-known rigidity theorems by using rather simple arguments. The reader is referred to~\cite{GN} for the detail of this approach using orthogonality, although we will briefly recall the basics wherever needed in this article.
To state our hyperplane restriction theorem, we first bring out the fact that every positive integer $A$ can be written as certain sums of binomial coefficients. Fix any $n\in\mathbb N^+$, there exist unique positive integers $a_n>a_{n-1}>\cdots>a_\delta$, where $\delta\geq 1$ and $a_j\geq j$ for every $j$, such that
$A=\binom{a_n}{n}+\cdots+\binom{a_\delta}{\delta}$. This is called the \textit{$n$-th Macaulay's representation} of $A$ and its existence and uniqueness can be proved by a greedy algorithm. These representations originally appeared in Macaulay's work of homogeneous ideals in polynomial rings~\cite{Ma}. Using the $n$-th Macaulay representation of $A$, we define the operation $A^{-<n>}:=\binom{a_n-1}{n-1}+\cdots+\binom{a_\delta-1}{\delta-1}$. In what follows, ``span" means the projective linear span.
\begin{theorem}\label{hyperplane thm}
Let $f:U\subset\mathbb P^n\rightarrow\mathbb P^M$ be a local holomorphic map such that $\dim(\mathrm{span}(f(U))\geq N$ for some positive integer $N\leq M$. Then, for a general hyperplane $H$ such that $H\cap U\neq\varnothing$, $\dim(\mathrm{span}(f(H\cap U))\geq N^{-<n>}$.
\end{theorem}
The equality in the theorem can hold, for example, when $f$ is a rational map whose components are all the linearly independent monic monomials of a fixed degree. Our theorem is obtained from combining Green's hyperplane restriction theorem (Theorem~\ref{green}) with a new combinatorial identity (Lemma~\ref{binom}). It holds for any local holomorphic maps between projective spaces and we believe that it will find applications elsewhere.
We now briefly explain the idea behind our proof for the gap phenomenon. Suppose $f$ is a local proper map from an $n$-dimensional generalized ball to an $N$-dimensional generalized ball such that the image of $f$ is not contained in any hyperplane. We first use our hyperplane restriction theorem repeatedly to obtain dimension estimates for the linear spans of the images of the linear subspaces in all dimensions. Since we know that $f$ is orthogonal, for any pair of orthogonal subspaces in the source projective space, their images under $f$ will span two orthogonal subspaces in the target projective space. If $N$ falls within any of the ``gaps'', i.e. the intervals $\mathcal J_k$ introduced earlier, by some amount of arithmetic we can show that the previously obtained dimension estimates will imply that there are two orthogonal subspaces in the target space whose sum of dimensions is at least $N$, which is impossible.
\section{Macaulay representation and a hyperplane restriction theorem for holomorphic mappings}
Every positive integer $A$ can be written as certain sums of binomial coefficients. Fix any $n\in\mathbb N^+$, there exist unique positive integers $a_n>a_{n-1}>\cdots>a_\delta$, where $\delta\geq 1$ and $a_j\geq j$ for every $j$, such that
$A=\binom{a_n}{n}+\cdots+\binom{a_\delta}{\delta}$. This is called the $n$-th Macaulay's representation of $A$. These representations naturally appeared in the works of Macaulay~\cite{Ma} and Green~\cite{Gr} on homogeneous ideals in polynomial rings. There are several operations pertaining to the Macaulay's representations, as follows.
Let $A=\binom{a_n}{n}+\cdots+\binom{a_\delta}{\delta}$ be the $n$-th Macaulay's representation of $A$, define
$$
A_{<n>}=\binom{a_n-1}{n}+\cdots+\binom{a_\delta-1}{\delta};\,\,\,\,\,\,\,\,
A^{-<n>}=\binom{a_n-1}{n-1}+\cdots+\binom{a_\delta-1}{\delta-1}.
$$
Here, we employ the convention that $\binom{a}{b}=0$ whenever $a<b$ or $b=0$. The seemingly peculiar choices of notations for these operations are due the fact that $A_{<n>}$ was used by Green and $A^{<n>}:=\binom{a_n+1}{n+1}+\cdots+\binom{a_\delta+1}{\delta+1}$ was used by Macaulay and $A^{-<n>}$ is in some sense the opposite of $A^{<n>}$.
We will need the following Green's hyperplane restriction theorem, which has already been used in~\cite{GLV} to study CR mappings between real hyperquadrics.
\begin{theorem}[\cite{Gr}]\label{green}
Let $W$ be a complex vector subspace of $H^0(\mathcal O_{\mathbb P^n}(d))$ of codimension $c$.
Let $W_H\subset H^0(\mathcal O_H(d))$ be the restriction of $W$ to a general hyperplane $H$ and $c_H$ be its codimension. Then, $c_H\leq c_{<d>}$.
\end{theorem}
We begin by stating a key combinatorial lemma related to these operations, which connects Green's hyperplane theorem and our Theorem~\ref{hyperplane thm}. To streamline the presentation, we defer its proof to the Section~\ref{last section}.
\begin{lemma}\label{binom}
Suppose $m,k\geq 1$ and $A,B\geq 0$. If $A+B=\binom{m+k}{k}-1$, then $$A^{-<m>}+B_{<k>}=\binom{m+k-1}{k}-1.$$
Here, we adopt the convention that $0^{-<m>}=0_{<k>}=0$.
\end{lemma}
We can now prove our hyperplane restriction theorem.
\begin{proof}[Proof of Theorem~\ref{hyperplane thm}]
We first prove the theorem for rational maps. Let $f:\mathbb P^n\dasharrow\mathbb P^M$ be a rational map such that $\dim(\mathrm{span}(f(\mathbb P^n))\geq N$. Let $f=[f_0,\ldots,f_M]$, where $f_j\in\mathbb C[z_0,\ldots,z_n]$, $0\leq j\leq M$, are homogeneous polynomials of degree $d$ without non-constant common factors.
Let $\mathcal H_{d,n}\subset\mathbb C[z_0,\ldots,z_n]$ be the vector subspace of homogeneous polynomials of degree $d$ and $W\subset\mathcal H_{d,n}$ be the subspace spanned by $f_0,\ldots,f_M$.
Thus, we have
$$
N+1\leq\dim(W)\leq\dim(\mathcal H_{d,n})=\binom{n+d}{d}.
$$
For a hyperplane $H\subset\mathbb P^n$, choose a set of homogeneous coordinates on $H$ and let $W_H\subset\mathcal H_{d,n-1}$ be the subspace spanned by restrictions of $f_0,\ldots,f_M$ on $H$. By Green's Theorem~\ref{green}, for a general hyperplane $H$, we have
$$
\dim(\mathcal H_{d,n-1})-\dim(W_H)\leq \left(\dim(\mathcal H_{d,n})-\dim(W)\right)_{<d>}
\leq \left(\dim(\mathcal H_{d,n})-N-1\right)_{<d>},
$$
in which the last inequality follows from the fact that $c_{<d>}\leq c'_{<d>}$ if $c\leq c'$. Thus,
\begin{equation}\label{dim wh}
\dim(W_H)
\geq\binom{n+d-1}{d}-\left(\binom{n+d}{d}-N-1\right)_{<d>}
\end{equation}
Now let $B=\binom{n+d}{d}-N-1$. Then $B\geq 0$ and
$$
N+B=\binom{n+d}{d}-1.
$$
By Lemma~\ref{binom}, we have from~$(1)$
\begin{eqnarray*}
\dim(W_H
&\geq&\binom{n+d-1}{d}-B_{<d>}\\
&=&\binom{n+d-1}{d}-\left(\binom{n+d-1}{d}-1-N^{-<n>}\right)\\
&=&N^{-<n>}+1.
\end{eqnarray*}
Thus $\dim(\mathrm{span}(f(H))\geq N^{-<n>}$ and we have proved the theorem for rational maps.
For the general case, if $f:U\subset\mathbb P^n\rightarrow\mathbb P^M$ is a local holomorphic map such that $\dim(\mathrm{span}(f(U))\geq N$, then for a sufficiently large $k$, the $k$-th order jet of $f$ can be represented by a rational map $f^\flat:\mathbb P^n\dasharrow\mathbb P^m$ (e.g. a truncated Taylor polynomial of $f$ at a point in $U$ after homogenization) such that $\dim(\mathrm{span}(f^\flat(\mathbb P^n))\geq N$. Since the restriction of $f^\flat$ to a hyperplane $H\subset\mathbb P^n$ represents the $k$-th order jet of $f|_{U\cap H}$, we see from the proven case of rational maps that for a general hyperplane $H$, we have
$
\dim(\mathrm{span}(f(H\cap U))\geq \dim(\mathrm{span}(f^\flat(H))\geq N^{-<n>}.
$
\end{proof}
\noindent\textbf{Remark.} The equality in Theorem~\ref{hyperplane thm} can hold since the equality can hold in Green's theorem~\cite{Gr}. One can also see directly that the equality holds when $F$ is the rational map whose components are all the linearly independent monic monomials of a fixed degree in $\mathbb C[z_0,\ldots,z_n]$.
Sometimes it is convenient to use the following counterpart of Theorem~\ref{hyperplane thm} and we will also elaborate a couple of special cases which are very useful for the study of CR mappings between real hyperquadrics (e.g. see~\cite{GN}).
\begin{theorem}\label{faran type}
Let $g:U\subset\mathbb P^m\rightarrow\mathbb P^{m'}$ be a local holomorphic map and $\ell\in\mathbb N^+$ such that $\ell\leq m-1$. If $g$ maps $\ell$-planes to $\ell'$-planes, then it maps $(\ell+1)$-planes to $((\ell'+1)^{<\ell>}-1)$-planes. In particular,
$(i)$ if $\ell'\leq\ell-1$, then the image of $g$ is contained in an $\ell'$-plane;
$(ii)$ if $\ell\leq \ell'\leq 2\ell-1$, then $g$ maps $(\ell+k)$-planes to $(\ell'+k)$-planes for $k\geq 0$;
\end{theorem}
\begin{proof}
Suppose on the contrary the image of a general $(\ell+1)$-plane under $g$ is not contained in any $((\ell'+1)^{<\ell>}-1)$-plane. Since $((\ell'+1)^{<\ell>})^{-<\ell+1>}=\ell'+1$, Theorem~\ref{hyperplane thm} implies that the image of a general $\ell$-plane is not contained in any $\ell'$-plane.
If $\ell'\leq\ell-1$, then $\ell'+1=\binom{\ell}{\ell}+\binom{\ell-1}{\ell-1}+\cdots+\binom{\delta}{\delta}$ for some $\delta\geq 1$, so $(\ell'+1)^{<\ell>}-1=\ell'$. Therefore we deduce inductively that the image of $g$ is contained in an $\ell'$-plane.
If $\ell\leq \ell'\leq 2\ell-1$, then $\ell'+1=\binom{\ell+1}{\ell}+\binom{\ell-1}{\ell-1}+\binom{\ell-2}{\ell-2}+\cdots+\binom{\delta}{\delta}$ for some $\delta\geq 1$. Thus,
$(\ell'+1)^{<\ell>}-1=\ell'+1$ and so $g$ maps $(\ell+1)$-planes to $(\ell'+1)$-planes. Moreover, as $\ell+1\leq\ell'+1< 2(\ell+1)-1$, we can proceed inductively and the desired result follows.
\end{proof}
\noindent\textbf{Remark.} One can apply Theorem~\ref{faran type} repeatedly to get the following simple formula. Under the same hypotheses, if the $\ell$-th Macaulay's representation of $\ell'+1$ is $\binom{\lambda_\ell}{\ell}+\cdots+\binom{\lambda_\delta}{\delta}$, then for any $k\in\mathbb N^+$, $g$ maps every $(\ell+k)$-plane to some linear subspace of dimension
$\binom{\lambda_\ell+k}{\ell+k}+\cdots+\binom{\lambda_\delta+k}{\delta+k}-1$.
\section{Gap phenomenon for local orthogonal maps}\label{gap}
We now recall some basics of \textit{local orthogonal maps} and the reader can see~\cite{GN} for more detail.
Let $r,s,t\in\mathbb N$ such that $r+s+t>0$. Denote by $\mathbb C^{r,s,t}$ the Euclidean space $\mathbb C^{r+s+t}$ equipped with the standard Hermitian bilinear form $\langle z, w\rangle_{r,s,t}$ of signature $(r,s,t)$, i.e.
$$
\langle z, w\rangle_{r,s,t}
=z_1\bar w_1+\cdots+z_r\bar w_r-z_{r+1}\bar w_{r+1}-\cdots - z_{r+s}\bar w_{r+s},
$$
where $z=(z_1,\ldots,z_{r+s+t})$ and $w=(w_1,\ldots,w_{r+s+t})$.
Define the indefinite norm $\|z\|^2_{r,s,t}=\langle z, z\rangle_{r,s,t}$ and call any $z\in \mathbb C^{r,s,t}$ a \textit{positive point} if $\|z\|^2_{r,s,t}>0$; a \textit{negative point} if $\|z\|^2_{r,s,t}<0$ and a \textit{null point} if $\|z\|^2_{r,s,t}=0$. If $\langle z, w\rangle_{r,s,t}=0$, we say that $z$ is orthogonal to $w$ and write $z\perp w$. In addition, the \textit{orthogonal complement} of $z$ is defined as
$$z^{\perp}=\{w\in \mathbb C^{r,s,t} \mid \langle z, w\rangle_{r,s,t}=0\}.$$
We denote by $\mathbb P^{r,s,t}:=\mathbb P\mathbb C^{r,s,t}$ the projectivization. We write $\mathbb C^{r,s}$ and $\mathbb P^{r,s}$ instead of $\mathbb C^{r,s,0}$ and $\mathbb P^{r,s,0}$.
On $\mathbb P^{r,s,t}$, even though the inner product is no longer defined, the notions of positive points, negative points and null points still make sense. Likewise, and more importantly, the orthogonality remains well defined. On $\mathbb P^{r,s,t}$, the set of positive points $\mathbb B^{r,s,t}\subset\mathbb P^{r,s,t}$ is called a \textit{generalized ball} since $\mathbb B^{1,s}$ is just the ordinary $s$-dimensional complex unit ball $\mathbb B^s$ embedded in $\mathbb P^s$. In addition, the boundary $\partial \mathbb B^{r,s,t}$ of $\mathbb B^{r,s,t}$ is simply the set of null points on $\mathbb P^{r,s,t}$.
\begin{definition}[\cite{GN}]\label{orthogonal def}
Let $U \subset \mathbb P^{r,s,t}$ be a connected open set containing a null point. We call a holomorphic map $f: U\rightarrow\mathbb P^{r',s',t'}$ \textbf{orthogonal} if $f(p)\perp f(q)$ for any $p, q\in U$ such that $p\perp q$.
\end{definition}
\begin{proposition}\label{contradict}
Let $f:U\subset\mathbb P^{r,s,t}\rightarrow\mathbb P^{r',s'}$ be a local orthogonal map. Then, for every linear subspace $E\subset\mathbb P^{r,s,t}$ such that $E\cap U\neq\varnothing$ and $E^\perp\cap U\neq\varnothing$,
$$
\dim(\mathrm{span}(f(E\cap U))+\dim(\mathrm{span}(f(E^\perp\cap U))\leq \dim(\mathbb P^{r',s'})-1.
$$
\end{proposition}
\begin{proof}
By orthogonality, $f(E^\perp\cap U)\subset (f(E\cap U))^\perp$ and since the Hermitian form on $\mathbb C^{r',s'}$ is non-degenerate, we have
$$
\dim(\mathrm{span}(f(E^\perp\cap U))+1\leq\dim(\mathbb C^{r',s'})-(\dim(\mathrm{span}(f(E\cap U))+1),
$$ and the desired result follows.
\end{proof}
Let $n,N\in\mathbb N$ such that $n+1\leq N<\binom{n+2}{2}=\binom{n+2}{n}$. By considering the $n$-th Macaulay representation of $N$, we deduce that $N$ is of the following form:
$$
N=N(n;a,b):=\binom{n+1}{n}+\cdots+\binom{n-a+1}{n-a}+ b
$$
for some integers $a,b\geq 0$ such that $b\leq n-a-1$. In fact, the $n$-th Macaulay's representation of $N(n;a,b)$ is
$$
N(n;a,0)=\binom{n+1}{n}+\cdots+\binom{n-a+1}{n-a}
$$
and for $b\geq 1$,
$$
N(n;a,b)=\binom{n+1}{n}+\cdots+\binom{n-a+1}{n-a}+ \binom{n-a-1}{n-a-1}+\cdots+\binom{n-a-b}{n-a-b}.
$$
\begin{lemma}\label{nab}
$\displaystyle
N(n;a,b)^{-<n>}=\left\{
\begin{matrix}[lcl]
N(n-1;a,b)&\mathrm{if}& n-a-b\geq 2;\\
N(n-1;a,b-1)&\mathrm{if}& n-a-b=1\,\,\mathrm{and}\,\,b\geq 1.
\end{matrix}
\right.
$
\end{lemma}
\begin{proof}
It is an immediate consequence of the $n$-th Macaulay representation of $N(n;a,b)$
described above.
\end{proof}
\begin{proposition}\label{dim prop}
Let $a,b,n$ be non-negative integers. Let $g:U\subset\mathbb P^n\rightarrow\mathbb P^{N(n;a,b)}$ be a local holomorphic map whose image is not contained in a proper linear subspace. Let $D_m=\dim(\mathrm{span}(g(M\cap U)))$ for a general $m$-dimensional linear subspace $M$ intersecting $U$. Then,
$$
D_m\geq\left\{
\begin{matrix}[lcl]
N(m;a,b) &\,\,\mathrm{if}& a+b+1\leq m\leq n-1;\\
N(m;a,m-a-1) &\,\,\mathrm{if}& a+1\leq m\leq a+b
\end{matrix}
\right.
$$
\end{proposition}
\begin{proof}
We will apply Theorem~\ref{hyperplane thm} and Lemma~\ref{nab} repeatedly with $m$ descending from $n-1$.
If $n-a-b\geq 2$, from the first line of Lemma~\ref{nab}, we get that $D_m\geq N(m;a,b)$ for $m=n-1,n-2,\ldots,a+b+1$. The proof is complete here if $b=0$. For $b\geq 1$, when we reach $m=a+b+1$, we have $N(m;a,b)=N(m;a,m-a-1)$. Thus, we deduce from the second line of Lemma~\ref{nab} that for $m=a+b,\dots, a+1$, we always have $D_m\geq N(m;a,m-a-1)$.
If $n-a-b=1$, the condition for the first inequality never holds and so we just need to prove the second inequality. The argument is exactly the same as that in the previous paragraph by using the second line of Lemma~\ref{nab}.
\end{proof}
\begin{theorem}\label{gap thm}
Let $f$ be a local orthogonal map from $\mathbb P^{r,s}$ to $\mathbb P^{r',s'}$ and $n:=\dim(\mathbb P^{r,s})$, $n':=\dim(\mathbb P^{r',s'})$. If there exists a non-negative integer $a$ such that
$$(a+1)(n+1)\leq n'\leq (a+2)n-(a^2+2a+2),$$
then the image of $f$ lies in a hyperplane of $\mathbb P^{r',s'}$
\end{theorem}
\begin{proof}
We will prove by contradiction. Suppose in the inequality in the hypotheses is satisfied for some $a$ and the image of $f$ is not contained in any hyperplane in $\mathbb P^{r',s'}$.
Since
$$
(a+1)(n+1)=\binom{n+1}{n}+\cdots+\binom{n-a+1}{n-a}+\dfrac{(a+1)a}{2}
$$
and
$$
(a+2)n-(a^2+2a+2)=\binom{n+1}{n}+\cdots+\binom{n-a+1}{n-a}+\left(n-\dfrac{a^2+5a+6}{2}\right),
$$
it follows that
$$
n'=\binom{n+1}{n}+\cdots+\binom{n-a+1}{n-a}+b=N(n;a,b)
$$
for some $b$ satisfying
\begin{equation}\label{ineq 1}
\dfrac{(a+1)a}{2}\leq b\leq n-\dfrac{a^2+5a+6}{2}.
\end{equation}
Let $n_1:=\left[\dfrac{n-1}{2}\right]$ and
$
n_2:=\left\{
\begin{matrix}n_1&\mathrm{ if\,\,} n \mathrm{\,\, is\,\, odd;}\\
n_1+1&\mathrm{ if\,\,} n \mathrm{\,\, is\,\, even.}
\end{matrix}\right.
$
Then $n_1+n_2+1=n$. Since $\langle\cdot,\cdot\rangle_{r,s}$ is non-degenerate, for an $n_1$-dimensional linear subspace in $\mathbb P^{r,s}$, its orthogonal complement is of dimension $n_2$ and conversely, any $n_2$-dimensional linear subspace is the orthogonal complement of some $n_1$-dimensional linear subspace. Let $D_m$ be the dimension of the linear span of the image under $f$ of a general $m$-dimensional linear subspace intersecting the domain of definition of $f$. We are going to use Proposition~\ref{contradict} to reach a contradiction by showing that $D_{n_1}+D_{n_2}\geq n'$.
By~(\ref{ineq 1}), we have $n\geq a^2+3a+3$, thus
$$
n_1\geq \dfrac{n-2}{2}\geq\dfrac{a^2+3a+1}{2}\geq a+\dfrac{1}{2},
$$
from which we always have $n_1\geq a+1$ since $n_1$ is an integer.
We now consider the following two cases separately:
$$
\mathbf{ Case\,\, I:\,\,\,} b\leq n_1-a-1
\,\,\,\,\,\,\,\,\,\,\mathrm{and}\,\,\,\,\,\,\,\,\,\,
\mathbf{ Case\,\, II:\,\,\,}n_1-a\leq b.
$$
In \textbf{Case I}, since $$n-1\geq n_2\geq n_1\geq a+b+1,$$ so by Proposition~\ref{dim prop} and (\ref{ineq 1}),
\begin{eqnarray*}
D_{n_1}+D_{n_2}&\geq& N(n_1;a,b)+N(n_2;a,b)\\
&=&\dfrac{a+1}{2}(2n_1+2-a)+b+\dfrac{a+1}{2}(2n_2+2-a)+b\\
&=&(a+1)(n+1-a)+2b\\
&=&(a+1)(n+1-\dfrac{a}{2})+b+\left(b-\dfrac{(a+1)a}{2}\right)\\
&\geq&(a+1)(n+1-\dfrac{a}{2})+b\\
&=&N(n;a,b)=n',
\end{eqnarray*}
which contradicts Proposition~\ref{contradict}.
In \textbf{Case II}, we have $a+1\leq n_1\leq a+b$ and thus by Proposition~\ref{dim prop},
$$
D_{n_1}\geq N(n_1;a;n_1-a-1)=(a+2)n_1-\dfrac{a^2+a}{2}
$$
and
$$
D_{n_2}\geq\left\{
\begin{matrix}
N(n_2;a;b) &\mathrm{\,\,if}& n_1=a+b &\textrm{and}& n_2=n_1+1;\\
N(n_2;a;n_2-a-1) &\mathrm{\,\,if}& n_1<a+b&\textrm{or}& n_2=n_1.
\end{matrix}
\right.
$$
Therefore,
{\small
$$
D_{n_1}+D_{n_2}\geq\left\{
\begin{matrix}
(a+2)(n_1+n_2+1)-(a^2+a)+b-n_2-1 &\mathrm{\,\,if}& n_1=a+b &\textrm{and}& n_2=n_1+1;\\
(a+2)(n_1+n_2)-(a^2+a) &\mathrm{\,\,if}& n_1<a+b&\textrm{or}& n_2=n_1,
\end{matrix}
\right.
$$
}
which simplifies to
$$
D_{n_1}+D_{n_2}\geq\left\{
\begin{matrix}
(a+2)n-(a^2+2a+2)&\mathrm{\,\,if}& n_1=a+b &\textrm{and}& n_2=n_1+1;\\
(a+2)n-(a^2+2a+2)&\mathrm{\,\,if}& n_1<a+b&\textrm{or}& n_2=n_1.
\end{matrix}
\right.
$$
Thus, we always have $D_{n_1}+D_{n_2}\geq n'$, which again contradicts Proposition~\ref{contradict}.
\end{proof}
\section{Proper maps between generalized balls}\label{proper map section}
We will now translate the results of the previous section to results for local proper holomorphic maps and formulate precisely the gap phenomenon for all generalized balls.
Let $V$ be a complex vector space equipped with a Hermitian inner product $H_V$ (possibly degenerate or indefinite) of signature $(r;s;t)$, where $\dim(V)=r+s+t$. Let $\mathbb PV$ be its projectivization. Similar to $\mathbb P^{r,s,t}$, the notion of positivity, negativity, nullity and orthogonality can be defined on $\mathbb PV$. In addition, any linear isometry $F:\mathbb C^{r,s,t}\rightarrow V$ induces a biholomorphic map $\tilde F:\mathbb P^{r,s,t}\rightarrow\mathbb PV$ preserving all these notions. Sufficient for our purpose, we can simply identify any such projective space $\mathbb PV$ with $\mathbb P^{r,s,t}$ through any such biholomorphism and we write $\mathbb PV\cong\mathbb P^{r,s,t}$ for such identification.
Now let $H$ be a complex linear subspace in $\mathbb C^{r,s,t}$ and the restriction of $\langle\cdot,\cdot\rangle_{r,s,t}$ on $H$ has the signature $(a; b; c)$. Obviously, we have $0\le a\le r$, $0\le b\le s$, $0\le c\le \min\{r-a,s-b\}+t$ and $a+b+c=\dim(H)$. Then $\mathbb PH \cong \mathbb P^{a,b,c}$. We call $\mathbb P H$ an\textit{ $(a,b,c)$-subspace} of $\mathbb P^{r,s,t}$. We will often denote an $(a,b,c)$-subspace by $H^{a,b,c}$.
We use the following definition for local proper holomorphic maps between generalized balls:
\begin{definition}
A local holomorphic map $f:U\subset\mathbb P^{r,s,t}\rightarrow\mathbb P^{r',s',t'}$, defined on a connected open set $U$ such that $U\cap\partial \mathbb B^{r,s,t}\neq\varnothing$, is called a \textbf{local proper holomorphic map} from $\mathbb B^{r,s,t}$ to $\mathbb B^{r',s',t'}$ if $f(U\cap \mathbb B^{r,s,t})\subset \mathbb B^{r',s',t'}$ and $f(U\cap\partial \mathbb B^{r,s,t})\subset\partial \mathbb B^{r',s',t'}$.
\end{definition}
\begin{proposition}\label{equiv2}
By shrinking the domain of definition if necessary, a local proper holomorphic map from $\mathbb B^{r,s,t}$ to $\mathbb B^{r',s',t'}$ is a local orthogonal map from $\mathbb P^{r,s,t}$ to $\mathbb P^{r',s',t'}$.
\end{proposition}
\begin{proof}
The proof is essentially the same as the one used to prove that such a local proper holomorphic map respects the Segre varieties associated to the boundaries of the generalized balls. Note that in this case a Segre variety is just the orthogonal complement of a single point. (Readers may see~\cite{GN}, Proposition 2.5 therein, for the detail.)
\end{proof}
The reason why it is a bit more complicated to formulate the gap phenomenon for all generalized balls is that by adding ``canceling components'' one can trivially modify any given proper map $f$ between generalized balls to get a new map $f^\sharp$, whose target is a bigger generalized ball. The image of $f^\sharp$ lies in a hyperplane of the target generalized ball, but is not equivalent to $f$ under automorphisms. This is in contrast to the case where the target is an ordinary unit ball. To incorporate this, we will need a couple of definitions:
\begin{definition}[\cite{GN}] In Definition~\ref{orthogonal def}, a local orthogonal map from $\mathbb P^{r,s,t}$ to $\mathbb P^{r',s',t'}$ is called \textbf{null} if its image lies entirely in a null space of $\mathbb P^{r',s',t'}$.
\end{definition}
\noindent\textbf{Remark.} If a local proper holomorphic map from $\mathbb B^{r,s,t}$ to $\mathbb B^{r',s',t'}$ is null, then it follows that its image is contained in $\partial\mathbb B^{r',s',t'}$.
For any $(r',s')$-subspace $H^{r',s'}\subset\mathbb P^{R,S}$, by considering the corresponding orthogonal decomposition of $\mathbb C^{R,S}$, it is not difficult to see that there are two canonical projections $\pi$ and $\pi^\perp$ (as rational maps) from $\mathbb P^{R,S}$ to $H^{r',s'}$ and $(H^{r',s'})^\perp$ respectively.
\begin{definition}\label{null prolong}
Let $f$ be a local proper holomorphic map from $\mathbb B^{r,s}$ to $\mathbb B^{R,S}$. If there exists an $(r',s')$-subspace $H^{r',s'}\subset\mathbb P^{R,S}$, such that either
$(i)$ the image of $f$ is contained in $H^{r',s'}$; or
$(ii)$ $\pi\circ f$ is a local proper holomorphic map from $\mathbb B^{r,s}$ to $\mathbb B^{R,S}\cap H^{r',s'}\cong \mathbb B^{r',s'}$ and $\pi^\perp\circ f$ is null,
then we say that $f$ is a {\it \textbf{null prolongation}} of $\pi\circ f$.
\end{definition}
\noindent\textbf{Example.} If $f=[f_1,\ldots,f_{r'+s'}]$ is a rational proper holomorphic map from $\mathbb B^{r,s}$ to $\mathbb B^{r',s'}$, where each $f_j\in\mathbb C[z_1,\ldots,z_{r+s}]$ is a degree-$d$ homogeneous polynomial, then for any homogeneous $\psi,\phi\in \mathbb C[z_1,\ldots,z_{r+s}]$ with $\deg(\phi)=\deg(\psi)+d$, the map $$F:=[\psi f_1,\ldots,\psi f_{r'},\phi,\psi f_{r'+1},\ldots,\psi f_{r'+s'},\phi],$$ which is locally proper from $\mathbb B^{r,s}$ to $\mathbb B^{r'+1,s'+1}$, is a null prolongation of $f$.
On the other hand, a similar construction is not possible if we restrict ourselves to unit balls. Indeed, it is easy to see that for maps between unit balls, only $(i)$ can happen in Definition~\ref{null prolong}, i.e. a null prolongation can only be a local proper holomorphic map whose image lies in a smaller dimensional unit ball.
\begin{theorem}\label{gap thm 2}
Let $f$ be a local proper holomorphic map from $\mathbb B^{r,s}$ to $\mathbb B^{R,S}$ and $n:=\dim(\mathbb B^{r,s})$ and $N:=\dim(\mathbb B^{R,S})$. If there exists a positive integer $k$ such that
$$kn+k\leq N\leq (k+1)n-(k^2+1),$$
then there exists an $(r',s')$-subspace $H^{r',s'}\subset\mathbb P^{R,S}$,
with $\dim(H^{r',s'})=kn+k-1$, such that $f$ is a null prolongation of some local proper holomorphic map from $\mathbb B^{r,s}$ to $\mathbb B^{R,S}\cap H^{r',s'}\cong \mathbb B^{r',s'}$.
\end{theorem}
\begin{proof}[Proof and remarks.]
In Theorem~\ref{gap thm}, by substituting $k:=a+1$, we see that whenever the hypotheses are satisfied, the image of $f$ is contained in a hyperplane $H\subset\mathbb P^{r,s}$. Write $H\cong\mathbb P^{r_1,s_1,t_1}$. Let $H^{r_1,s_1}\subset H$ be any $(r_1,s_1)$-subspace and $\pi:H\dasharrow H^{r_1,s_1}$, $\pi^\perp:H\dasharrow (H^{r_1,s_1})^\perp$ be the canonical projections. (Note that $(H^{r_1,s_1})^\perp\cong\mathbb P^{0,0,t_1}$ is a null space in $\mathbb P^{R,S}$.) Then, it follows that $f$ is a null prolongation of $f_1:=\pi\circ f$, where the latter is locally proper from $\mathbb B^{r,s}$ to $\mathbb B^{R,S}\cap H^{r_1,s_1}\cong\mathbb B^{r_1,s_1}$. The desired result then follows if we repeat the argument for a finite number of times, because if $f_1$ is a null prolongation of another local proper holomorphic map $f_2$ from $\mathbb B^{r,s}$ to $\mathbb B^{r_1,s_1}\cap H^{r_2,s_2}\cong\mathbb B^{r_2,s_2}$ for some $(r_2,s_2)$-subspace $H^{r_2,s_2}$, then $f$ is also a null prolongation of $f_2$.
For proper holomorphic maps from $\mathbb B^{1,n}\cong \mathbb B^n$ to $\mathbb B^{1,N}\cong\mathbb B^N$, Faran's result~\cite{faran} is essentially the statement that the conclusion of the theorem holds for $n+1\leq N\leq 2n-2$. Moreover, the same conclusion has been shown by Huang-Ji-Xu~\cite{hjx} to hold for $2n+1\leq N\leq 3n-4$ and by Huang-Ji-Yin~\cite{HJY} for $3n+1\leq N\leq 4n-7$.
The lower bound $kn+k$ of our gap is actually optimal. This can be seen by considering the expansion of $\left(\sum^k_{j=1}|z_j|^2-\sum^{n+1}_{j=k+1}|z_j|^2\right)\left(\sum^k_{j=1}|z_j|^4\right)$. The expansion is a sum of (plus or minus) norm squares of $kn+k$ linearly independent cubic monomials. Using these monomials as components we get a rational proper map from $\mathbb B^{k,n+1-k}$ to $\mathbb B^{k^2,k(n-k+1)}$ whose image does not lie in any hyperplane. Note that $\dim(\mathbb B^{k,n+1-k})=n$ and $\dim(\mathbb B^{k^2,k(n-k+1)})=kn+k-1$.
\end{proof}
\section{Proof of Lemma~\ref{binom}}\label{last section}
We will prove by induction and first show that the lemma is true for $m=1$ or $k=1$.
Suppose $k=1$ and $A+B=\binom{m+1}{1}-1=m$. If $A=0$ and $B=m$, then $A^{-<m>}+B_{<1>}=0+(m-1)=m-1=\binom{m}{1}-1$. If $1\leq A\leq m$, then
$
A=\binom{m}{m}+\binom{m-1}{m-1}+\cdots+\binom{m-A+1}{m-A+1}.
$
Hence, $A^{-<m>}+B_{<1>}=A+B-1=\binom{m}{1}-1$.
Suppose $m=1$ and $A+B=\binom{1+k}{k}-1$. Then $B\leq k$ and hence $B_{<k>}=0$. Thus, $A^{-<1>}+B_{<k>}=0+0=\binom{k}{k}-1$.
Suppose now $A+B=\binom{m+k}{k}-1=\binom{m+k}{m}-1=\binom{m+k-1}{m}+\binom{m+k-1}{k}-1$. Since $A,B$ are integers, we have either $A\geq\binom{m+k-1}{m}$ or $B\geq\binom{m+k-1}{k}$.
If $A\geq \binom{m+k-1}{m}$, then the $m$-th Macaulay's representation of $A$ is of the form
$A=\binom{m+k-1}{m}+\binom{a_{m-1}}{m-1}+\cdots+\binom{a_\delta}{\delta}$. Since $(A-\binom{m+k-1}{m})+B=\binom{m+k-1}{k}-1$, by the induction hypothesis we get
{\footnotesize
\begin{eqnarray*}
\left(A-\binom{m+k-1}{m}\right)^{-<m-1>}+B_{<k>}&=&\binom{m+k-2}{k}-1\\
\Rightarrow\,\,\,\,\, \binom{a_{m-1}-1}{m-2}+\cdots+\binom{a_\delta-1}{\delta-1}+B_{<k>}&=&\binom{m+k-2}{k}-1\\
\Rightarrow\,\,\,\,\, \binom{m+k-2}{m-1}+\binom{a_{m-1}-1}{m-2}+\cdots+\binom{a_\delta-1}{\delta-1}+B_{<k>}&=&\binom{m+k-2}{m-1}+\binom{m+k-2}{k}-1\\
\Rightarrow\,\,\,\,\, A^{-<m>}+B_{<k>}&=&\binom{m+k-1}{k}-1
\end{eqnarray*}
}
If $B\geq \binom{m+k-1}{k}$, then the $k$-th Macaulay's representation of $B$ is of the form
$B=\binom{m+k-1}{k}+\binom{b_{k-1}}{k-1}+\cdots+\binom{b_\epsilon}{\epsilon}$. Since $A+(B-\binom{m+k-1}{k})=\binom{m+k-1}{k-1}-1$, by the induction hypothesis we get
{\footnotesize
\begin{eqnarray*}
A^{-<m>}+\left(B-\binom{m+k-1}{k}\right)_{<k-1>}&=&\binom{m+k-2}{k-1}-1\\
\Rightarrow\,\,\,\,\, A^{-<m>}+\binom{b_{k-1}-1}{k-1}+\cdots+\binom{b_\epsilon-1}{\epsilon}&=&\binom{m+k-2}{k-1}-1\\
\Rightarrow\,\,\,\,\, \binom{m+k-2}{k}+A^{-<m>}+\binom{b_{k-1}-1}{k-1}+\cdots+\binom{b_\epsilon-1}{\epsilon}&=&\binom{m+k-2}{k}+\binom{m+k-2}{k-1}-1\\
\Rightarrow\,\,\,\,\, A^{-<m>}+B_{<k>}&=&\binom{m+k-1}{k}-1
\end{eqnarray*}
}
\noindent{\bf Acknowledgements.} The authors would like to thank Prof. Xiaojun Huang, Ming Xiao, Wanke Yin and Yuan Yuan for their comments on the first version. The first author was partially supported by Institute of Marine Equipment, Shanghai Jiao Tong University and
Shanghai Science and Technology Plan projects (No. 21JC1401900). The second author was partially supported by Science and Technology Commission of Shanghai Municipality (STCSM) (No. 13dz2260400).
|
\section{Introduction}
\label{sec:introduction}
\input{sections/introduction}
\section{Related Work}
\label{sec:related_work}
\input{sections/related_work.tex}
\section{Framework for Identifying Shortcuts}
\label{sec:method}
\input{sections/method.tex}
\section{Experiments}
\label{sec:experiment}
\input{sections/experiments.tex}
\section{Conclusion}
\input{sections/conclusion.tex}
\subsection{Tasks and Datasets}
\noindent \textbf{Task 1: Sentiment classification.} For the task of sentiment classification, we use several datasets in our experiments. To find shortcuts in Stanford Sentiment Treebank (SST-2) \cite{socher-etal-2013-recursive} dataset, we first train a model on SST-2 training set which consists of $67,349$ sentences. We then evaluate the model on SST-2 training set\footnote{We use training set of SST-2 because the test set has a very limited number of examples.} and Yelp~\cite{asghar2016yelp} test set and obtain attention scores. For cross-dataset analysis, we compare the important tokens extracted from SST-2 and Yelp.
Similarly, we train another model on $80,000$ amazon kitchen reviews \cite{amazon_data}, and apply it on the kitchen review dev set and the amazon electronics dev set, both having $10,000$ reviews.
\noindent \textbf{Task 2: Occupation classification.} Following \citet{pruthi-etal-2020-learning}, we use the biographies \cite{De_Arteaga_2019} to predict whether the occupation is a surgeon or physician (non-surgeon). The training data consists of $17,629$ biographies and the dev set contains $2,519$ samples.
\noindent \textbf{Models.}
We use the attention scores over BERT \cite{devlin-etal-2019-bert} based classification models as they have achieved the state-of-art performance. Note that our proposed framework can also be easily extended to models with different architectures. BERT-based models have the advantage that we can directly use the attention scores as explanations of model decisions. For models with other architectures, we can use explanation techniques such as LIME~\cite{LIME} or Path Integrated Gradient approaches~\cite{pig} to provide explanations.
\noindent \textbf{Evaluation.}
Evaluating identified shortcuts in machine learning or deep leaning based models can be difficult.
We do not have ground-truth labels for the shortcuts identified through our framework, and whether a token is a shortcut or not can be subjective even with human annotators, and it can further depend on the context.
Faced with these challenges, we carefully designed a task and adopted Amazon Mechanical Turk for evaluation. We post the identified shortcuts after each step in our framework, along with several sample sentences containing the token, as additional context, to the human annotator. We ask the question ``does the word determine the sentiment in the sentence'' and ask the annotator to provide a ``yes''/``no'' answer\footnote{In the instruction, we further specify ``select yes'' if the highlighted word is a determining factor for the sentiment label, and we provide a few example sentences along with their shortcuts as references. The exact template is shown in the Appendix.} to the question based on the answer that holds true for the majority of the provided sentences (we also experimented with adding an option of ``unsure'' but found most annotators do not choose that option). Each identified shortcut is verified by 3 annotators.
\begin{table*}[t!]
\begin{center}
\small
\begin{tabular}{c|c|c|c|c|c|c}
\toprule
Methods & \makecell{SST-2 $\rightarrow$\\ Kitchen} &
\makecell{SST-2 $\rightarrow$\\ Electronics}& \makecell{ Kitchen $\rightarrow$\\ SST-2} & \makecell{Kitchen $\rightarrow$\\ Electronics} & \makecell{Electronics $\rightarrow$\\ SST-2}&
\makecell{Electronics $\rightarrow$\\ Kitchen}\\
\midrule
No Mitigation &87.43 & \textbf{84.30} & 71.45 & 98.22 &73.05 & 98.79\\
Test RM &87.50 & 83.96 & 71.56 & 98.07 & 72.94 & 98.77\\
Train RM & 87.72 & 84.13 & \textbf{72.82} &\textbf{98.60} & \textbf{74.08} & 98.79\\
Train \& Test RM & \textbf{87.76} & 83.74 & \textbf{72.82} & \textbf{98.62} & \textbf{74.08} & \textbf{98.80}\\
\bottomrule
\end{tabular}
\end{center}
\vspace{-0.1in}
\caption{Domain generalization results on SST-2 and Amazon Kitchen/Electronics datasets. RM means shortcuts removed, Train/Test corresponds to shortcuts removal during training and test time, respectively.}
\label{tab:transfer}
\end{table*}
\subsection{Experimental Results}
We summarized the top important tokens after each step in our framework (Table~\ref{tab:important_words}). We also report the precision score (the percentage of tokens) out of the top $50$ tokens identified as true shortcuts by human annotators in Table~\ref{tab:precision_exp}.
Across all datasets, we see that the precision score increases after each step, which demonstrates that our proposed framework can consistently improve shortcut identification more precisely. Specifically, after the first step, the precision score of shortcuts is low\footnote{Some cells have ``-'' importance score due to no shortcut is identified by human annotators in the top-$K$ identified tokens.} because most of the top extracted tokens are important tokens only (thus many of them are genuine). After the second step (cross-dataset analysis) and the third step (knowledge-aware perturbation), we see a significant increase of the shortcuts among the top-$K$ extracted tokens. Table~\ref{tab:qualitative} shows examples of perturbing shortcut tokens leading to model predictions changes.
\textbf{Agreement analysis over annotations.} Since this annotation task is non-trivial and sometimes subjective, we further compute the intraclass correlation score~\cite{bartko1966intraclass} for the Amazon Mechanical Turk annotations. Our collected annotations reaches an intraclass correlation score of $0.72$, showing a good agreement among annotators. Another agreement we analyze is showing annotators $5$ sample sentences compared to showing them all sentences, to avoid sample bias. We ask annotators to annotate a batch of $25$ tokens with all sentences containing the corresponding token shown to them. The agreement reaches $84.0\%$, indicating that showing $5$ sample sentences does not significantly affect annotator's decision on the target token. More details of Amazon Mechanical Turk interface can be found in the Appendix.
\begin{table}[t]
\begin{center}
\small
\begin{tabular}{c|c|c}
\toprule
Top $10$ extracted tokens & Precision & Recall \\
\midrule
\makecell{\colorbox{yellow!50}{ms.}, \colorbox{yellow!50}{mrs.}, \colorbox{yellow!50}{she}, \colorbox{yellow!50}{her}, \\ \colorbox{yellow!50}{he}, reviews, been, \\favorite, \colorbox{yellow!50}{his}, practices} & 0.60 & 0.50\\
\bottomrule
\end{tabular}
\end{center}
\vspace{-0.1in}
\caption{Identified shortcuts (highlighted tokens are overlapped with the pre-specified impermissible tokens from \citet{pruthi-etal-2020-learning}) in occupation classification.}
\label{tab:occupation}
\end{table}
\subsection{A Case Study: Occupation Classification}
\citet{pruthi-etal-2020-learning} derived an occupation dataset to study the gender bias in NLP classification tasks. The task is framed as a binary classification task to distinguish between ``surgeons'' and ``physicians''. These two occupations are chosen because they share similar words in their biographies and a majority of surgeons are male. The dataset is further tuned -- downsample minority classes (female surgeons and male physicians) by a factor of ten to encourage the model to rely on gendered words to make predictions. \citet{pruthi-etal-2020-learning} also provides a pre-specified list of impermissible tokens~\footnote{he, she, her, his, him, himself, herself, mr, ms, mr., mrs., ms. We removed ``hers'' and ``mrs'' from the original list since they do not appear in dev data.} that a robust model should assign low attention scores to. We instead treat this list of tokens as shortcuts and analyze the efficacy of our proposed framework on identifying these tokens. These impermissible tokens can be regarded as shortcuts because they only reflect the gender of the person, thus by definition should not affect the decision of a occupation classification model.
Table~\ref{tab:occupation} presents the result on identifying the list of impermissible tokens. Among the top ten tokens selected by our method, $6$ of them are shortcuts.
Furthermore, $9$ out of $12$ impermissible tokens are captured in the top $50$ tokens selected by our method.
This further demonstrates that our method can effectively find shortcuts in this occupation classification task, in a more automated way compared to existing approaches that rely on pre-defined lists.
\begin{table}[t!]
\begin{center}
\small
\begin{tabular}{c|c|c|c}
\toprule
Dataset & Methods & C1 & C2\\
\midrule
\multirow{4}{*}{\makecell{Amazon \\ Kitchen}} & No Mitigation & 99.15 & 0.0\\
& Test RM & 99.21 & 0.18\\
& Train RM & 99.15 & 0.24\\
& Train \& Test RM & 99.15 & 0.24\\
\bottomrule
\end{tabular}
\end{center}
\vspace{-0.1in}
\caption{Accuracy on challenging datasets. C1: test subset that has shortcuts; C2: test subset that has shortcuts and are wrongly predicted by the original model.}
\label{tab:challenging}
\end{table}
\begin{table}[t!]
\begin{center}
\small
\begin{tabular}{c|cccc}
\toprule
& Male & Female & $\Delta$ & Overall \\
\midrule
No Mitigation & 94.02 & 99.50 & 5.48 & 97.46 \\
Test RM & 92.28 & 96.36 & 4.08 & 94.84 \\
Train RM & 93.26 & 92.40 & 0.86 & 92.66 \\
Train \& Test RM & 94.46 & 99.06 & 4.60 & 97.34\\
\bottomrule
\end{tabular}
\end{center}
\vspace{-0.1in}
\caption{Accuracy and performance gap of male and female groups in Occupation Classification task.}
\label{tab:bias}
\vspace{-0.1in}
\end{table}
\begin{table}[t]
\begin{center}
\small
\begin{tabular}{c|c|c|c|c}
\toprule
$\lambda$ & 4 & 6 & 8 & 10 \\
\midrule
4 & 1.00 & 0.78 & 0.62 & 0.56\\
6 & 0.78 & 1.00 & 0.84 & 0.76\\
8 & 0.62 & 0.84 & 1.00 & 0.92\\
10 & 0.56 & 0.76 & 0.92 & 1.00\\
\bottomrule
\end{tabular}
\end{center}
\vspace{-0.1in}
\caption{Overlap of top $50$ tokens when changing $\lambda$.
}
\label{tab:ablation}
\end{table}
\begin{table}[htb]
\begin{center}
\small
\begin{tabular}{c|c|c|c|c}
\toprule
\multirow{2}{*}{Dataset} & \multirow{2}{*}{Method} & \multicolumn{1}{c|}{@10} & \multicolumn{1}{c|}{@20} & \multicolumn{1}{c}{@50} \\
& & Prec. & Prec. & Prec. \\
\midrule
\multirow{2}{*}{SST-2} & Attention & 0.40 & 0.35 & 0.32 \\
& Integrated Gradient & 0.30 & 0.3 & 0.34 \\
\midrule
\multirow{2}{*}{Yelp} & Attention & 0.60 & 0.50 & 0.56 \\
& Integrated Gradient & 0.50 & 0.55 & 0.60 \\
\bottomrule
\end{tabular}
\end{center}
\vspace{-0.1in}
\caption{Ablation study on using Integrated Gradient to extract important tokens.}
\label{tab:ablation_ig}
\vspace{-0.1in}
\end{table}
\subsection{Mitigating Shortcuts}
We also study mitigating shortcuts by masking out the identified shortcuts. Specifically, we use shortcut tokens identified by human annotators and mask them out in training set and re-train the model (Train RM), during test time directly (Test RM), and both (Train \& Test RM) as described in Sec~\ref{sec:mitigate}. We evaluate these three approaches in multiple settings: 1) domain generalization; 2) challenging datasets; 3) gender bias. As shown in Table~\ref{tab:transfer}, masking out shortcuts, especially in training data, can improve model's generalization to out-of-distribution data.
Note in this setting, different from existing domain transfer work \cite{pan_transfer_learning_survey}, we do not assume access to labeled data in the target domain during training, instead we use our proposed approach to identify potential shortcuts that can generalize to unseen target domains.
As a result, we also observe model's performance improvement on challenging datasets (Table~\ref{tab:challenging}). Table~\ref{tab:bias} demonstrates that mitigating shortcuts helps to reduce the performance gap ($\Delta$) between male and female groups, resulting in a fairer model.
Note the original performance might degrade slightly due to models learning different but more robust feature representations, consistent with findings in existing work \cite{tsipras2018robustness}.
\paragraph{Ablation Study}
We conduct an ablation study of changing the hyper-parameter $\lambda$ in the first step of extracting important tokens. As shown in Table~\ref{tab:ablation}, our method is not very sensitive to the choice of $\lambda$. In Table~\ref{tab:ablation_ig}, we show that Attention scores and Integrated Gradient can both serve as a reasonable method for extracting important tokens in our first step, suggesting the flexibility of our framework.
\subsection{Identify Tokens Key to Model's Decision}
The first step of the framework aims to identify the top-$K$ most important tokens that affect model's decision making process.
We look at the importance at the token-level.\footnote{In this paper, we mostly focus on unigrams. Our method can also be easily extended to multi-gram, text span or other type of features by summing the attention scores over spans. For a vocabulary of wordpieces as used in BERT, we concatenate wordpieces with a prefix of ``\#\#'' to form unigrams and sum the attention scores.}
In general, depending on how the tokens are being used in model's decision process, they can be roughly divided into three categories: ``genuine'', ``spurious'', and others (e.g., tokens that are not useful for a model's prediction).
\textit{Genuine} tokens are tokens that causally affect a task's label \cite{pmlr-v119-srivastava20a, zhao_spurious_2020}, and thus the correlations between those tokens and the labels are what we expect the model to capture and to more heavily rely on.
On the other hand, \textit{spurious} tokens, or \textit{shortcuts} as commonly denoted in prior work \cite{Geirhos_2020, minderer2020automatic}, are features that correlate with task labels but are not genuine, and thus might fail to transfer to challenging test conditions \cite{Geirhos_2020} or out-of-distribution data; spurious tokens do not causally affect task labels \cite{pmlr-v119-srivastava20a, zhao_spurious_2020}.
In this step, we will extract both genuine tokens and shortcut tokens because they are both likely to affect a model's prediction.
We rely on interpretability techniques to collect information on whether a certain input token is important to model's decision making. In this paper, we use the attention score in BERT-based models as an explanation of model predictions \cite{clark-etal-2019-bert, kovaleva-etal-2019-revealing}, due to its simplicity and fast computation. Recent work~\cite{hiddencut} also reveals that attention scores outperform other explanation techniques in regularizing redundant information.
Other techniques \cite{LIME, pig, chen-etal-2020-generating, jacovi2021contrastive} can also be used in this step.
As an example, given a sentence ``\emph{Spielberg is a good director.}'', assuming ``\emph{good}'' is a genuine token and ``\emph{Spielberg}'' is a shortcut token, we expect that in a BERT-based sentiment classification model, the attention scores for ``\emph{good}'' and ``\emph{Spielberg}'' are higher and thus will be extracted as important tokens. On the other hand, for ``\emph{is}'', ``\emph{a}'' and ``\emph{director}'' the attention scores would be lower as they are relatively less useful to the model decision.
We now describe this step using sentiment classification task as an example (more details can be found in Algorithm~\ref{alg:step1}). Let $f$ be a well trained sentiment classification model. Given a corpus $\mathcal{D}$, for each input sentence $s_i$, $i=1, \ldots, n$ for a total of $n$ sentences in the corpus, we apply $f$ on it to obtain the output probability $p^{pos}_i$ and $p^{neg}_i$ for positive and negative label respectively.
We then extract attention scores $\{a^1_{i}, a^2_{i}, \dots, a^m_{i}\}$ for tokens $\{t^1_{i}, t^2_{i}, \dots, t^m_{i}\}$ in sentence $s_i$, where $m$ is the length of the sentence. In BERT-based classification models, the embedding of [CLS] token in the final layer is fed to a classification layer. We thus extract the attention scores of each token $t$ used for computing the embedding of the [CLS] token and average them across different heads.
If $p^{pos}_i > p^{neg}_i$, we obtain the updated attention score $\Tilde{a}^j_i = a^j_i * p^{pos}_i$, otherwise $\Tilde{a}^j_i = -a^j_i * p^{neg}_i$.
For each token $t$ in the vocabulary $\mathcal{V}$, we compute the average attention score: $\Bar{a}_t = \frac{1}{mn} \cdot \Sigma_{i=1}^n\Sigma_{j=1}^{m}[\Tilde{a}^j_i \cdot \mathbbm{1}(t^j_i=t)]$,
where we aggregate the attention scores $\Tilde{a}^j_i$ for token $t$, across all $n$ sentences in the corpus.
We then normalize the attention scores across the vocabulary to obtain the importance score for each token $t$: $I_t = \Bar{a}_t / \Sigma_{t\in \mathcal{V}}\Bar{a}_{t}$. This can lead to very small $I_t$ for certain tokens, thus we take the log of all importance scores to avoid underflow, $I'_t = \log(I_t)$.
So far, we have computed the importance score for each token. However, we observe that some tokens appearing only very a few times could accidentally have very high importance scores. Thus, we propose to penalize the tokens with low frequencies:
$\hat{I}_t = I'_t - \lambda/\log(1+c_t),$
where $c_t$ is the frequency of token $t$ and $\lambda$ is a temperature parameter to adjust the degree that we want to penalize over the frequency.
\begin{algorithm}[t]
\SetAlgoLined
\SetKwInOut{Input}{Input}\SetKwInOut{Output}{Output}
\Input{Sentiment classification model: $f$\\
Text corpus: $\mathcal{D}$}
// Obtain attention scores for tokens in each input sentence $s_i\in\mathcal{D}$:\\
\For{$i = 1$ to n}{
$p_i^{pos}$, $p_i^{neg}$, $\{a_i^1, a_i^2, ..., a_i^m\}$ = $f(s_i)$\;
\For{$j = 1$ to m}{
if $p^{pos}_i > p^{neg}_i$:
$\Tilde{a}^j_i = a^j_i \cdot p^{pos}_i$\;
else:
$\Tilde{a}^j_i = -a^j_i \cdot p^{neg}_i$\;
}
}
// Use $\{\Tilde{a}_i^j\}$ to compute an importance score for each token $t$ in the vocabulary $\mathcal{V}$:\\
$Importance = dict()$\\
\For{$i = 1$ to n}{
\For{$j = 1$ to m}{
$Importance[t_i^j]$.append($\Tilde{a}_i^j$)\;
}
}
// Normalize the importance score and penalize low-frequency tokens:\\
\For{$t$ in $\mathcal{V}$}{
$\Bar{a}_t = average(Importance[t])$\;
$I_t = \Bar{a}_t / \Sigma_{t\in \mathcal{V}}\Bar{a}_{t}$\;
$I'_t = \log(I_t)$\;
$\hat{I}_t = I'_t - \lambda / \log(1 + frequency[t])$\;
}
\Output{A list of tokens sorted according to their importance scores:\\
\{$t_1, t_2, ..., t_{|\mathcal{V}|}$\}, \\
where $\hat{I}_{t_i} \geq \hat{I}_{t_2} \geq ... \geq \hat{I}_{t_{|\mathcal{V}|}}$
}
\caption{Important Token Extraction.}
\label{alg:step1}
\end{algorithm}
\begin{table*}[t]
\begin{center}
\small
\setlength\tabcolsep{2.5pt}
\begin{tabular}{l}
\toprule
\textit{Shortcut token}: \textbf{bread}\\
\makecell[l]{\textit{Original:} I bought this in the hopes it would keep \textbf{bread} I made fresh. However, after a few times of usings the I found out \\that moister w still getting in bread would become stale or moldy ...(\textcolor{red}{Neg})}\\
\makecell[l]{\textit{Perturbed:} I bought this in the hopes it would keep \textbf{loaf} I made fresh. However, after a few times of usings the I found out \\that moister w still getting in bread would become stale or moldy ... (\textcolor{blue}{Pos})}\\
\midrule
\textit{Shortcut token}: \textbf{iPhone} \\
\makecell[l]{\textit{Original:} I lost my original TV remote, and found this one thinking it was the same one. ... Now this one is merely a back \\up. Also, I have the Samsung remote app on my \textbf{iPhone}, which also works just as good as these remotes. (\textcolor{blue}{Pos})}\\
\makecell[l]{\textit{Perturbed:} I lost my original TV remote, and found this one thinking it was the same one. ... Now this one is merely a back \\up. Also, I have the Samsung remote app on my \textbf{ipod}, which also works just as good as these remotes. (\textcolor{red}{Neg})}\\
\bottomrule
\end{tabular}
\end{center}
\vspace{-0.1in}
\caption{Examples of shortcut tokens with significant performance drop during knowledge-aware perturbation.}
\label{tab:qualitative}
\vspace{-0.1in}
\end{table*}
\subsection{Cross-Dataset Stability Analysis}
\label{sec:cross-dataset}
As mentioned before, the tokens that are important to a model's prediction could be either genuine or spurious, thus in this step, we want to categorize the extracted tokens into these two categories and maintain a list of tokens that are more likely to be ``spurious''.
In many real-world NLP tasks, if we have access to datasets from different sources or domains, then we can perform a \textit{cross-dataset} analysis to more effectively identify ``spurious'' tokens.
The reasoning is that ``spurious'' tokens tend to be important for a model's decision making on one dataset but are less likely to transfer or generalize to other datasets, e.g. ``\emph{Spielberg}'' could be an important token for movie reviews but is not likely to be useful on other review datasets (e.g., for restaurants or hotels). On the other hand, genuine tokens are more likely to be important across multiple datasets, for example, tokens like ``\emph{good}'', ``\emph{bad}'', ``\emph{great}'', ``\emph{terrible}'' should remain useful across various sentiment classification datasets. Thus, in this step, we try to distinguish ``genuine'' tokens from ``spurious'' tokens from the top extracted important tokens after the first step. Our idea is to compare tokens' importance ranking and find the ones that have very different ranks across datasets.
To this end, we conduct a cross-dataset stability analysis. Specifically, we apply the same model $f$ on two datasets A and B, and obtain two importance ranking lists. Since importance scores may have different ranges on the two datasets, we normalize all importance scores to adjust the value to be in the range of $[0, 1]$:
$$\Tilde{I}^A_t = \frac{\hat{I}^A_t - \min(\{\hat{I}^A_t|t\in\mathcal{V}\})}{\max(\{\hat{I}^A_t|t\in\mathcal{V}\}) - \min(\{\hat{I}^A_t|t\in\mathcal{V}\})}$$
$$\Tilde{I}^B_t = \frac{\hat{I}^B_t - \min(\{\hat{I}^B_t|t\in\mathcal{V}\})}{\max(\{\hat{I}^B_t|t\in\mathcal{V}\}) - \min(\{\hat{I}^B_t|t\in\mathcal{V}\})}$$
where $\Tilde{I}^A_t$ and $\Tilde{I}^B_t$ are normalized importance scores on dataset A and B respectively. We then subtract $\Tilde{I}^B_t$ from $\Tilde{I}^A_t$ and re-rank all tokens according to their differences. Tokens with largest differences are the ones with high importance scores in dataset A but low importance scores in dataset B, thus they are more likely to be ``shortcut'' tokens in dataset A. Similarly, we can also extract tokens with largest differences from dataset B by subtract $\Tilde{I}^A_t$ from $\Tilde{I}^B_t$.
\subsection{Knowledge-aware Perturbation}
The cross-dataset analysis is an efficient way to remove important tokens that are ``genuine'' across multiple datasets, after which we can obtain a list with tokens that are more likely to be ``spurious''. However, on this list, domain-specific genuine tokens can still be ranked very high, e.g., ``\emph{ambitious}'' from a movie review dataset and ``\emph{delicious}'' from a restaurant review dataset. This is because domain-specific genuine tokens have similar characteristics as shortcuts, they are effective for a model's decision making on a certain dataset but could appear very rarely (and thus could be deemed as not important) on another dataset.
Hence, in this section, we further propose a slightly more expensive and a more fine-grained approach to verify whether a token is indeed ``spurious'', through \textit{knowledge-aware perturbation}.
For each potential shortcut token, we extract $N$ synonyms by leveraging the word embeddings curated for synonym extraction~\cite{counter-fitting}, plus WordNet~\cite{wordnet} and DBpedia~\cite{10.5555/1785162.1785216}. More specifically, for each top token $t$ in the list generated by the previous step, we first search counter-fitting word vectors to find synonyms with cosine similarity larger than a threshold\footnote{We set it as $0.5$ following the set up in~\cite{textfooler}.} $\tau$. Additionally we search in WordNet and DBpedia to obtain a maximum of $N$ synonyms for each token $t$. Then we extract a subset $S_t$ from $\mathcal{D}$, which consists of sentences containing $t$. We perturb all sentences in $S_t$ by replacing $t$ with its synonyms. The resulted perturbed set $S'_t$ is $N$ times of the original set $S_t$. We apply model $f$ on $S_t$ and $S'_t$ and obtain accuracy $acc_t$ and $acc'_t$. Since we only perturb $S_t$ with $t$'s synonyms, the semantic meaning of perturbed sentences should stay close to the original sentences. Thus, if $t$ is a genuine token, $acc'_t$ is expected to be close to $acc_t$. On the other hand, if $t$ is a shortcut, model prediction can be different even the semantic meaning of the sentence does not change a lot (see examples in Table~\ref{tab:qualitative}). Thus, we assume tokens with larger differences between $acc_t$ and $acc'_t$ are more likely to be shortcuts and tokens with smaller differences are more likely to be domain specific ``genuine'' words. From the potential shortcut token list computed in Sec~\ref{sec:cross-dataset}, we remove tokens with performance difference smaller than $\mathbbm{\delta}$ to further filter domain specific ``geniue'' tokens .
\begin{table*}[t]
\begin{center}
\small
\setlength\tabcolsep{2.5pt}
\begin{tabular}{c|c}
\toprule
Dataset & Top \textbf{important} tokens extracted from each dataset \\
\midrule
SST-2 & terrific, impeccable, exhilarating, refreshingly, irresistible, heartfelt, thought-provoking, ...\\
Yelp & Awesome, Definitely, is, Excellent, Very, Great, Good, Best, attentive, worth, definitely, Highly, ...\\
Amazon Kitchen & utensils, thermometer, Cuisinart, Definitely, Pyrex, Bought, utensil, Arrived, Recommend, ...\\
\midrule
Dataset & Top \textbf{shortcuts} extracted from each dataset (verified by human annotators) \\
\midrule
SST-2 & recycled, seal, sitcom, longest, fallen, qualities, rises, impact, translate, emphasizes, ...\\
Yelp & ambiance, tastes, bartenders, patio, burgers, staff, watering, donuts, cannot, pancakes, regulars, ...\\
Amazon Kitchen & utensils, Cuisinart, Rachael, Pyrex, utensil, Breville, Zojirushi, Corelle, Oxo, dehumidifier, ...\\
\bottomrule
\end{tabular}
\end{center}
\vspace{-0.15in}
\caption{Top \textit{important} tokens and top \textit{shortcut} tokens identified by our proposed framework and further verified by human annotators. Many shortcuts reflect the characteristics of the datasets, e.g. ``captures'' from a movie review dataset, ``burgers'' from a restaurant review dataset and brand names from an Amazon kitchen review dataset. }
\label{tab:important_words}
\end{table*}
\begin{table*}[htb]
\begin{center}
\small
\begin{tabular}{c|c|cc|cc|cc}
\toprule
\multirow{2}{*}{Dataset} & \multirow{2}{*}{Method} & \multicolumn{2}{c|}{@10} & \multicolumn{2}{c|}{@20} & \multicolumn{2}{c}{@50} \\
& & Prec. & Impor. & Prec. & Impor. & Prec. & Impor. \\
\midrule
\multirow{3}{*}{SST-2} & 1 & 0.00 & - & 0.05 & 0.97 & 0.02 & 0.96\\
& 2 & 0.10 & 0.95 & 0.05 & 0.94 & 0.04 & 0.93\\
& 3 & 0.40 & 0.90 & 0.35 & 0.87 & 0.32 & 0.85\\
\midrule
\multirow{3}{*}{Yelp} & 1 & 0.10 & 0.96 & 0.05 & 0.95 & 0.18 & 0.95\\
& 2 & 0.40 & 0.89 & 0.25 & 0.89 & 0.30 & 0.88\\
& 3 & 0.60 & 0.89 & 0.50 & 0.87 & 0.56 & 0.87\\
\midrule
\multirow{3}{*}{\makecell{Amazon Kitchen}} & 1 & 0.70 & 0.98 & 0.80 & 0.96 & 0.78 & 0.95\\
& 2 &1.00 & 0.97 & 1.00 & 0.95 & 1.00 & 0.95\\
& 3 &1.00 & 0.97 & 1.00 & 0.95 & 1.00 & 0.95\\
\bottomrule
\end{tabular}
\end{center}
\vspace{-0.15in}
\caption{We report the precision as well as the averaged importance score $\Tilde{I}$ of identified ``shortcuts'' after each step based on our framework. The identified ``shortcut'' is a true shortcut or not is verified by 3 independent human annotators (Amazon Turkers). We can see that the precision increases after each step in our framework, demonstrating the utility of cross-dataset analysis (step 2) and knowledge-aware perturbation (step 3). }
\label{tab:precision_exp}
\vspace{-0.1in}
\end{table*}
\subsection{Mitigation via Identified Shortcuts}
\label{sec:mitigate}
In this section, we describe how the identified shortcuts can be further utilized to improve robustness in NLP models.
More specifically, we propose \textit{targeted} approaches to mitigate the identified shortcuts including three variants:
(1) a \textit{training-time} mitigation approach: we mask out the identified shortcuts during training time and re-train the model;
(2) an \textit{inference-time} mitigation approach: we mask out the identified shortcuts during inference time only, in this way we save the extra cost of re-training a model;
(3) we combine both approach (1) and (2).
In the experiment section, we will demonstrate the effect of each approach over a set of benchmark datasets. We found that by masking out shortcuts in datasets, models generalize better to challenging datasets, out-of-distribution data, and also become more fair.
|
\section{Introduction}
A lithium (Li) ion battery is a complicated electrochemical system and its performance is dependent on a multitude of material properties, among which the solid-state diffusion coefficient $D_s$ of Li$^+$ is one of the key parameters, since the mass transport in these particles is the rate-limiting processes for thin electrodes, and the corresponding resistances constitute a major part of the total battery overpotential. The value of $D_s$ is, however, not straight-forwardly obtained, which is a major obstacle for more precise battery modelling. While the diffusion coefficient can be theoretically estimated from density functional theory (DFT) \cite{zheng2011diffusion,dathar2011calculations} or Molecular Dynamics (MD) simulations \cite{islam2000ionic,islam2014lithium}, it is more frequently experimentally determined by electrochemical approaches, including the galvanostatic intermittent titration technique (GITT), potentiostatic intermittent titration technique (PITT), electrochemical impedance spectroscopy (EIS) and cyclic voltammetry (CV) \cite{santos2018revisiting}. Of these, GITT is arguably the most popular choice due to its comparative robustness, while still being time-consuming.
GITT was original proposed by Weppner and Huggins \cite{weppner1977determination} for a Li$_3$Sb electrode, and then later generalized for porous insertion electrode materials used in lithium-ion batteries \cite{chen2020development,nickol2020gitt,dees2009analysis}. In the GITT test, a current pulse is applied when the cell is in an equilibrium state and the voltage response follows a linear relationship with the square root of time. The diffusion coefficient can thus be calculated from the voltage response, assuming that the particle size is known. Normally the current pulses are applied at different states of charge (SOC) as the solid-state diffusion coefficient is a SOC dependent parameter. Since an equilibrium initial condition is required before the current pulse, a long relaxation time is needed in between each SOC level, which makes the GITT method very time consuming. In recent years, however, the alternative intermittent current interruption (ICI) technique has been proposed \cite{lacey2017influence} as a faster alternative to GITT \cite{chien2021fast}. Using ICI, the voltage response during a current interruption is monitored, which also follows a linear relationship with the square root of time. Since the ICI method does not require the cell to be in an equilibrium state, the most time consuming part in the GITT method, i.e. the relaxation period, can thus be omitted.
Employing the same experimental object and setup, different methods can sometimes provide fairly similar results for the diffusion coefficient $D_s$. For example, Chien et al. obtained diffusion coefficient values with the GITT and ICI methods with an excellent agreement with each other \cite{chien2021fast}, and a similar comparison has been made between the GITT and EIS methods displaying a good match \cite{deng2020consistent}. Nevertheless, the reported $D_s$ values from different literature of one specific material often vary within several orders of magnitude and this fact makes it challenging to decide a proper value to use in battery modelling. The reason for this large spread of the reported $D_s$ values can be explained by the differences in the electrode design, electrolyte composition, current magnitude, fitting method, resting period, etc. Moreover, the validity of GITT, as well as other techniques employed for this purpose, is being questioned and discussed as the underlying assumptions are not completely valid when they are applied for porous insertion lithium-ion electrodes \cite{liu2006analytical,garcia2018quantitative}. Dees et al. has shown that the $D_s$ value estimated with the GITT theory is generally lower that the value used in their DC electrochemical model \cite{dees2009analysis}.
It is not feasible to strictly validate the electrochemical approaches for estimating the solid-state Li$^+$ diffusion coefficient experimentally, since it is difficult to have a diffusion coefficient value that can be trusted as the truth. However, if these electrochemical approaches are valid for real batteries, they shall as well be valid for continuum models of Li-ion batteries, as a model is a simplification of the real world but at the same time keeps all the assumptions used in the electroanalytical methods. The philosophy in this work is that if experiments with these electrochemical methods are simulated with a valid battery model, one should be able to extract the same solid-state diffusion coefficient as what is put in the model, as a proof of the method validity. This is essential so that these electrochemical methods can be trusted to parameterize accurate models.
Therefore, a theoretical validation is demonstrated in this work for the GITT and ICI method when applying them for a porous electrode. The classical Doyle-Fuller-Newman (DFN) model \cite{doyle1993modeling,fuller1994simulation} is one of the most commonly employed continuum models for simulating lithium-ion batteries and it is considered to be a fairly faithful representation of a real battery system. In this work, a DFN model previously implemented in MATLAB \cite{geng2021bridging} with parameters from \cite{chen2020development} for LiNi$_{0.8}$Mn$_{0.1}$Co$_{0.1}$O$_2$ is used for black-box testing. The input is current sequences designed according to the GITT and ICI methods and the output is the battery terminal voltage which is used to estimate the solid-state diffusion coefficient. The advantage is that after the black-box testing, different physical phenomenon in the model can be studied in details to demonstrate the origins for any deviations between the estimated $D_s$ value and the true value used in the model. The main purpose of the study is thus to reveal how the porous electrode structure will affect the solid-state diffusion coefficient estimation when employing the GITT and ICI methods, which was originally proposed for a more uniform electrode material. This work also demonstrates the impact on the $D_s$ estimation accuracy with different fitting methods, fitting lengths, current magnitude, electrode thickness as well as electrolyte properties.
\section{Conclusions}
A theoretical validation of the GITT and ICI methods was demonstrated with a porous lithium-ion NMC811 electrode employing a DFN model. Although the theoretical derivation of the GITT and ICI methods are based on multiple assumptions which are most likely not entirely valid for this type of lithium-ion battery electrode, both methods can still provide a very good estimation of the $D_s$ value, with the maximum deviation being an underestimation of 66 \%. The estimation is more accurate when the voltage profile has a more linear shape, since the voltage response can then be better described with the analytical solution of Fick's law. For the GITT method, a longer fitting time can sometime provide a better estimation, while for the ICI method it is preferred to keep the fitting length short. With the parameter set used in this work, the solid-state diffusion coefficient is underestimated in most of the case. A higher current magnitude, a thinner electrode and a higher electrolyte diffusion coefficient can lead to a higher estimation for the $D_s$ value, but can also render an overestimation. For the purpose of identifying the solid-state diffusion coefficient, the ICI method has a significant advantage of being time-effective. It should be acknowledged that GITT can be used to study also other properties of an electrode material, including the open circuit voltage and hysteresis. However, the ICI technique is much more convenient and powerful and it is suitable to be incorporated as part of a reference performance test for aging diagnostics.
\section*{Acknowledgment}
The authors would like to thank Energimyndigheten (P42789-1) for the financing of this work. Y.-C.C. and DB acknowledges STandUP for Energy.
\input{Nomenclature}
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\bibliographystyle{IEEEtran}
\section{Results and discussions}
The simulated voltage profiles during the GITT and ICI tests are shown in Fig.~\ref{fig:GITT_ICI_full_soc_voltage}, which resemble the experimental results reported in literature \cite{chien2021fast,dees2009analysis}. With the symmetric cell setup described in the previous section, the delithiation and lithiation processes are simulated simultaneously. In the GITT test, the voltage step response is clearly visible under a C/10 current pulse and in total 35 current pulses are applied through the entire SOC range. With the simulation protocol, i.e. 15 minutes pulse with 1.5 hours relaxation, the time required to complete a GITT test in a real experiment will be 63 hours. On the other hand, 53 current interruptions take place in the ICI test, with the corresponding experiment time being merely 9 hours. The ICI method thereby allows more identification occasions for the diffusion coefficient, with less than 15 \% of the time it takes to complete a GITT measurement.
\begin{figure}[!ht]
\centering
\includegraphics{figures/GITT_ICI_full_soc_voltage.pdf}
\caption{Simulated voltage profiles of an NMC811 electrode during lithiation and delithiation processes with GITT and ICI methods.}
\label{fig:GITT_ICI_full_soc_voltage}
\end{figure}
\begin{figure}
\centering
\begin{subfigure}[b]{0.49\textwidth}
\centering
\includegraphics{figures/GITT_ICI_full_soc_D_s_new.pdf}
\caption{}
\label{fig:GITT_ICI_full_soc_D_s}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.49\textwidth}
\centering
\includegraphics{figures/GITT_ICI_full_soc_dEdx_new.pdf}
\caption{}
\label{fig:GITT_ICI_full_soc_dEdx}
\end{subfigure}
\caption{(a) The derived diffusion coefficient $D_s$ from the GITT and ICI simulations. (b) Estimated $dE/dx$ from the GITT and ICI simulations.}
\end{figure}
The derived diffusion coefficient $D_s$ is presented in Fig.~\ref{fig:GITT_ICI_full_soc_D_s}. In the GITT method, the voltage response between $t$ = 1 s and $t$ = 21 s (the current is applied at $t$ = 0 s) is used to for the fitting in \eqref{eq:tau<1 fitting}, and the voltage response between $t$ = 1 s and $t$ = 5 s (the current is interrupted at $t$ = 0 s) is used in the ICI method to fit with the expression in \eqref{eq:ICI_fitting}. The diffusion coefficient used in the model is 1.48$\times$10$^{-15}$ m$^2$/s \cite{chen2020development} and it is plotted as the 'true value' in Fig.~\ref{fig:GITT_ICI_full_soc_D_s}. Although in reality the diffusion coefficient of lithium ions is SOC dependent for most electrode materials \cite{persson2010lithium}, the parameter value used in this work is a constant for simplicity and for making a clear demonstration.
It can be seen that the two methods can estimate the diffusion coefficient $D_s$ successfully, with only small deviations from the value used in the model. Since the true value of $D_s$ in the model, as well as other electrochemical parameters (see Table 1), are constant, ideally the estimated $D_s$ value from the GITT and ICI methods should be a constant as well. However, as it is shown in Fig.~\ref{fig:GITT_ICI_full_soc_D_s}, the estimated $D_s$ from the GITT and ICI methods has a significant dependence on the SOC. It can be observed that the trend of the estimated $D_s$ in Fig.~\ref{fig:GITT_ICI_full_soc_D_s} has a good correlation with the $dE/dx$ profile in Fig.~\ref{fig:GITT_ICI_full_soc_dEdx}, and this is a hint that the deviations are related with the violation of one assumption in the mathematical derivation: \textit{The voltage profile is linear during a short interval, i.e. $dE/dc$ is a constant}. The true value of the $dE/dx$ term is calculated from the open circuit voltage profile in \eqref{eq:ocv}. Although the SOC interval during a current pulse is relatively small (1.785 \% SOC change in a GITT current pulse), the minor variation in the $dE/dx$ term cannot be ignored and its impact is directly reflected in the $D_s$ estimation. This will be demonstrated in the next part. One more observation in Fig.~\ref{fig:GITT_ICI_full_soc_dEdx} is that the GITT method can provide a better estimation of the $dE/dx$ profile. This is expected, since in the GITT method, the $dE/dx$ term is estimated between two equilibrium states and it is not affected by the current direction, taking into account that the open circuit voltage hysteresis is not included in the model. However, the ICI method obtains $dE/dx$ during a galvanostatic charge and discharge, and therefore the estimation is slightly shifted depending on the current direction. As can be noted in Fig.~\ref{fig:GITT_ICI_full_soc_D_s} and Fig.~\ref{fig:GITT_ICI_full_soc_dEdx}, in the beginning of the lithiation and delithiation processes in the ICI test, both the $dE/dx$ term and the estimated $D_s$ value have a larger mismatch from the true value, but the agreement improves with the time. Despite some minor shifts in the $dE/dx$ estimation, the $D_s$ value obtained from the ICI method has a very good agreement with those calculated using the GITT method.
To demonstrate how the $dE/dx$ term affects the $D_s$ estimation, two cases at different SOC levels are studied in detail. In the first case ($x$ = 0.35), $dE/dx$ is changing rapid during the current pulse, while in the second case ($x$ = 0.8) $dE/dx$ is comparatively constant, which in turn fulfills the assumptions of the GITT derivation to a higher degree. Moreover, results with different fitting approaches and different fitting lengths are demonstrated as well. For the GITT method, the full expression fitting in \eqref{eq:full_expression_fitting} and with the $\tau \ll$ 1 approximation fitting in \eqref{eq:tau<1 fitting} are applied with a fitting length varying from 20 s to 900 s. When it comes to the ICI method, the expression in \eqref{eq:ICI_fitting} is used to find the diffusion coefficient, and the fitting length is from 1 s to 4 s. The voltage response in the first second is always excluded in the fitting to avoid the transient behaviours in other electrochemical processes in the system. The voltage response and its fitting are presented in Fig.~\ref{fig:dEdx_change_case1_case2} and the estimated diffusion coefficient are compared with the true value used in the DFN model.
\begin{figure}
\centering
\includegraphics{figures/Fitting_result_case_1.pdf}
\caption{The voltage response of LiNi$_{0.8}$Mn$_{0.1}$Co$_{0.1}$O$_2$ and its fitting to obtain its diffusion coefficient $D_s$ in the GITT and ICI methods. The voltage response in black is the numerical solution simulated with a DFN model, and voltage response in colors are the analytical solutions with the fitted $D_s$, described in equations \eqref{eq:full_expression_fitting} as dashed line for the GITT method, \eqref{eq:tau<1 fitting} as the solid line for the GITT method and \eqref{eq:ICI_fitting} as the solid line for the ICI method, respectively.}
\label{fig:dEdx_change_case1_case2}
\end{figure}
It can be observed that the analytical solutions of Fick's law (colored lines) cannot capture the voltage response accurately when $x$ = 0.35 in Fig.~\ref{fig:dEdx_change_case1_case2}a, due to the relatively large variation of the $dE/dx$ term at this SOC level, as indicated in Fig.~\ref{fig:GITT_ICI_full_soc_dEdx}. As a consequence, the estimated $D_s$ when $x$ = 0.35 has a larger deviation from the true value of 1.48$\times$10$^{-15}$ m$^2$/s, for both the full expression fitting (dashed lines) and $\tau \ll$ 1 fitting (solid lines). It is seen in Fig.~\ref{fig:dEdx_change_case1_case2}a that this estimation can be improved with a longer fitting length. On the other hand, the analytical solutions (colored lines) highly resemble the numerical solution with a DFN model (black lines) when $x$ = 0.8 in Fig.~\ref{fig:dEdx_change_case1_case2}b, since the voltage profile is more linear during the current pulse, i.e. the $dE/dx$ term is close to a constant. Under these conditions, the estimated $D_s$ is closer to the true value, especially using the full expression fitting. The validity of the $\tau \ll$ 1 approximation was discussed in conjunction with Fig.~\ref{fig:Fickslaw_solution}, with the upper time axis calculated with the $r_s$ and $D_s$ values used in the model. The approximation is considered to be valid when $\tau <$ 0.0032, i.e. $t<$ 58.92 s. Therefore, the estimation discrepancy increases when the fitting length is increased from 20 s to 50 s and 100 s in Fig.~\ref{fig:dEdx_change_case1_case2}b, as this approximation is losing its validity. When the fitting length is increased even longer, the overpotential contributions from the charge transfer reaction and electrolyte mass transportation plays a role in the fitting, causing an opposite effect to the invalidity of the $\tau \ll$ 1 approximation, and improves the estimation in this case. Based on Fig.~\ref{fig:dEdx_change_case1_case2}a and Fig.~\ref{fig:dEdx_change_case1_case2}b, it can be concluded that the diffusion coefficient $D_s$ can be estimated more accurately when the voltage profile is more linear. The same conclusion can be made with the ICI simulation results in Fig.~\ref{fig:dEdx_change_case1_case2}c and Fig.~\ref{fig:dEdx_change_case1_case2}d, that the estimated $D_s$ with the ICI method is closer to the true value at $x$ = 0.8 with a constant $dE/dx$. Moreover, it can be observed that in the ICI method, it is preferable to keep the fitting length short to better fulfill the $\tau \ll$ 1 approximation. Despite the fact that the diffusion coefficient $D_s$ is slightly underestimated in most of the presented cases, this is considered as a fairly accurate estimation, with the maximum error only being 66\%.
\begin{figure}
\centering
\begin{subfigure}[b]{0.49\textwidth}
\centering
\includegraphics{figures/Current_distribution.pdf}
\caption{}
\label{fig:Current_distribution}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.49\textwidth}
\centering
\includegraphics{figures/Overpotential.pdf}
\caption{}
\label{fig:Overpotential}
\end{subfigure}
\caption{During a current pulse at $x$ = 0.8, (a) the current distribution in the electrode during one GITT pulse and (b) Overpotential at the current collector side.}
\end{figure}
However, even when the voltage profile is linear at $x$ = 0.8, it is still not possible to obtain exactly the true value of $D_s$ irrespective of the fitting method and fitting length used. This can be explained with the violations of the other two assumptions in the derivation listed in the theory section: \textit{The current density at the particle surface is constant} and \textit{the over-potential caused by other electrochemical processes can be ignored}. To address these two issues, the current distribution in the electrode during one GITT pulse and the overpotential development are studied in detail in Fig.~6. In this analysis, the local current density and the overpotential at the particle closest to the current collector are studied, as it directly determines the measured terminal voltage. The current distribution during a current pulse in the GITT method at $x$ = 0.8 is shown in Fig.~\ref{fig:Current_distribution}. When the current is applied at $t$ = 0 s, the current density is higher at the separator side, since the electronic conductivity is much higher than the ionic conductivity and the current tends to flow via the least resistive path. With time passing, the current density becomes more even through the porous electrode. The actual local current $j$ at the current collector side is presented in Fig.~\ref{fig:Current_distribution}, and as can be seen, it is increasing with time. For the GITT fitting, the current is assumed to have an even distribution and the local current density $j$ is estimated to be $j = I/(S_a L)$ = 0.17 A/m$^2$ during a C/10 current pulse (plotted as the black line). It can be observed that in the beginning of the pulse, the actual current density at the current collector side is lower than 0.17 A/m$^2$, meaning that 0.17 A/m$^2$ is an overestimation of the local current density. This is one of the reasons for the disagreement between the estimated $D_s$ value and the true value.
Besides the local current density, another factor causing this discrepancy is the overpotential caused by other electrochemical processes. As shown in Fig.~\ref{fig:Overpotential}, the measurable voltage response $E$ is the sum of the stoichiometry based voltage and the overpotential. In a DFN model, the overpotential is a consequence of three processes: the charge transfer reaction, electrolyte mass transport and the diffusion in the solid state. The charge transfer and solid diffusion overpotential are presented for the particle closest to the current collector, and the electrolyte mass transport overpotential is the difference between the electrolyte potential at the current collector point and electrolyte overpotential at the middle of the separator (the reference potential). Although it is only the solid diffusion overpotential that is the desired to use for the $D_s$ estimation, in reality it is not possible to separate it from the other two overpotentials. It can be observed in Fig.~\ref{fig:Overpotential} that both the charge transfer and electrolyte overpotential show a transient behaviour in the beginning of the pulse, which shall be avoided in the fitting. This is why the voltage response in the first second has been excluded in the previous demonstrations. These two overpotentials stabilize with time, and a higher electrolyte diffusion coefficient $D_l$ and a faster reaction rate will speed up the stabilization. In the later part of the current pulse, these two overpotentials constitute an offset in the voltage response while the solid diffusion overpotential is still developing with time. This offset will still impact the fitting result as it introduces errors in the $dE/d\sqrt{t}$ fitting.
To further demonstrate how the violations of the last three assumptions in the theory section can affect the result, four setups are studied, from the most simplified scenario to a more realistic system. Four black-box models are configured according to Table~\ref{tab:case_setup} and the simulated voltage responses are plotted as solid lines in Fig.~\ref{fig:Case_1_2_3_4}. The simulated voltage response are then fitted with the analytical solutions, both with the full expression in (10), plotted as dashed lines, and the $\tau \ll 1$ approximation in (12), not shown in the figure. The first setup is an ideal scenario, where the voltage profile of the electrode material is perfectly linear, with a constant local current density and excluding the rest of the system. With this setup, the analytical solution of Fick's law can precisely represent the voltage response, as shown in Fig.~\ref{fig:Case_1_2_3_4}, and the true $D_s$ value can be accurately estimated without any deviations using the full expression fitting \eqref{eq:full_expression_fitting}. However, the $\tau \ll$ 1 fitting method leads to a 7.9 \% underestimation, since the approximation is conditionally valid. In the next setup, the voltage profile is a function of SOC, described in \eqref{eq:ocv} at $x$ = 0.35. The instant observation is that the analytical solution does not match the numerical solution anymore and the estimated $D_s$ values differ more than 40 \% compared to the true value. To complicate the system further, an uneven current distribution in the electrode is included in Setup 3. With this more realistic current distribution, the local current density at the current collector side is lower than the average value, as demonstrated in Fig.~\ref{fig:Current_distribution}, and the voltage response in Setup 3 is therefore slightly lower than what is simulated with Setup 2. In the final setup, the overpotential caused by other electrochemical processes are included in the terminal voltage. This setup has the same condition as in a DFN model and the simulated voltage profile is identical to what was presented in Fig.~\ref{fig:dEdx_change_case1_case2}a. Based on the fitting results in Table.~\ref{tab:case_setup}, it can be noted that more violations of these assumptions do not necessarily lead to a worse estimation. The exact impact depends on the combination of the parameters in the system, as well as the current magnitude, which will be demonstrated in the next part.
\begin{table}[!ht]
\centering
\caption{Four setups to investigate the impact of violations of assumptions.}
\begin{tabular}{| c | c |c|c | c| c|} \hline
& $dE/dx$ & Local current & \makecell{Overpotential caused \\ by other processes} & \makecell{Fitted $D_s$ with \eqref{eq:tau<1 fitting} \\ between 1-20 s} & \makecell{Fitted $D_s$ with \eqref{eq:full_expression_fitting} \\ between 1-900 s}\\ \hline
Setup 1 & Constant & Constant & Excluded & 1.36$\times$10$^{-15}$ m$^2$/s (-7.9\%) & 1.48$\times$10$^{-15}$ m$^2$/s (0\%) \\ \hline
Setup 2 & Varying & Constant & Excluded & 0.85$\times$10$^{-15}$ m$^2$/s (-42.6\%) & 2.15$\times$10$^{-15}$ m$^2$/s (+45.3\%) \\ \hline
Setup 3 & Varying & Varying & Excluded & 0.98$\times$10$^{-15}$ m$^2$/s (-33.6\%) & 2.53$\times$10$^{-15}$ m$^2$/s (+70.6\%) \\ \hline
Setup 4 (DFN model) & Varying & Varying & Included & 0.57$\times$10$^{-15}$ m$^2$/s (-61.7\%) & 1.73$\times$10$^{-15}$ m$^2$/s (+16.6\%) \\ \hline
\end{tabular}
\label{tab:case_setup}
\end{table}
\begin{figure}[!ht]
\centering
\includegraphics{figures/Case_1_2_3_4.pdf}
\caption{The voltage response during a current pulse simulated with a black-box configured according to Table.~\ref{tab:case_setup} (solid lines) and its fitting with analytical expressions with (10) (dashed lines).}
\label{fig:Case_1_2_3_4}
\end{figure}
\begin{figure}[!ht]
\centering
\begin{subfigure}[b]{0.49\textwidth}
\centering
\includegraphics{figures/D_s_sweep_current_thickness.pdf}
\caption{}
\label{fig:D_s_sweep_current_thickness}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.49\textwidth}
\centering
\includegraphics{figures/D_s_sweep_current_thickness_higher_Dl.pdf}
\caption{}
\label{fig:D_s_sweep_current_thickness_higher_Dl}
\end{subfigure}
\caption{The difference between the estimated $D_s$ value and the true value with a variation in the electrode thickness and current magnitude with (a) the parameters listed in Table \ref{tab:parametersNMC811} (b) the diffusion coefficient in the electrolyte $D_l$ is increased by 10 times.}
\end{figure}
In the studies above, the estimation accuracy of $D_s$ is affected by the fitting approach and fitting length. Besides, the battery design, current magnitude and the rest of the system can have a large impact on the results as well. To illustrate such effects, the result of the estimated $D_s$ with a variation in the electrode thickness and current magnitude is presented in Fig.~\ref{fig:D_s_sweep_current_thickness}. The scenario investigated is that a current pulse is applied at $t$ = 0 s at $x$ = 0.8, and the voltage response between $t$ = 1 s and $t$ = 21 s is used for fitting with the $\tau \ll$ 1 approximation in \eqref{eq:tau<1 fitting}. The same study is then performed when the diffusion coefficient in the electrolyte $D_l$ in Table \ref{tab:parametersNMC811} is increased by 10 times; the results of this approach is shown in Fig.~\ref{fig:D_s_sweep_current_thickness_higher_Dl}. For both sets of parameters, the estimated value of $D_s$ is higher when employing a thinner electrode and a higher current magnitude. With a 40 $\mu$m thick electrode and C/10 current, the estimated $D_s$ is 32\% lower than the true value in Fig.~\ref{fig:D_s_sweep_current_thickness}, while the $D_s$ estimation becomes only 5\% lower than the true value in Fig.~\ref{fig:D_s_sweep_current_thickness}, meaning that an increased electrolyte diffusion coefficient $D_l$ can result in a higher estimation of the $D_s$ value. This does not necessarily mean that this can improve the estimation, since it can lead to an overestimation under some conditions. When the GITT method is used for estimating lithium-ion battery aging, a decreasing trend in the $D_s$ can be observed with the increase of the cycling numbers \cite{chien2021fast}. The result in Fig.~8 implies that apart from the active material degradation being the main reason, the decrease in $D_s$ could also be partly due to the degradation of the electrolyte and a decreased $D_l$.
\section*{Supplementary material}
\begin{figure}[!ht]
\centering
\includegraphics{figures/D_s_sweep_current_thickness_analysis.pdf}
\caption{The overpotential and current distribution during a 0.1 C current pulse at $x$ = 0.8. $I$ = 2.49 A/m$^2$ for $L$ = 37.8 $\mu$m, $I$ = 7.47 A/m$^2$ for $L$ = 113.4 $\mu$m.}
\label{fig:D_s_sweep_current_thickness_analysis}
\end{figure}
\section{Simulation setup}
The mathematical model is a DFN model implemented in MATLAB as described in \cite{geng2021bridging} and the parameters are from \cite{chen2020development}, listed in Table \ref{tab:parametersNMC811}. Two values are here adjusted from the original values reported in \cite{chen2020development}; one is the electronic conductivity of the electrode material and the other is the reaction rate constant in the charge transfer reaction. The electronic conductivity of LiNi$_{0.8}$Mn$_{0.1}$Co$_{0.1}$O$_2$ (NMC811) is reported to be 0.18 S/m in \cite{chen2020development}. However, conductive carbon additives are often added in cell manufacturing to improve the conductivity of the electrode. Therefore, a higher value 10 S/m is used in this work to be closer to reality. The other parameter, the reaction rate constant in the charge transfer reaction, is reduced to increase the charge transfer resistance and thus improve the stability of the system. Instability issues appear under a certain parameter combination with solvers used in \cite{geng2021bridging} but can be overcome by slightly adjusting the parameters. This adjustment is considered to not affect the conclusions in this study for the solid-state diffusion coefficient identification.
A symmetric cell setup is employed: NMC811 ($x_0=x_{min}$) - separator - NMC811 ($x_0=x_{max}$). When the cell is charged, one electrode NMC811 ($x_0=x_{min}$) is being lithiated while the other electrode NMC811 ($x_0=x_{max}$) is being delithiated. This setup has the advantage that the lithiation and delithiation processes of one electrode material can be simulated within one simulation, thereby saving computational time. The electrolyte potential in the middle of the separator is used as the reference potential (ground) to mimic a three-electrode cell. The DFN model is a pseudo two-dimensional (P2D) model where one dimension is in the cross-section direction of the cell to address the current distribution in the electrodes, and the electrolyte concentration distribution through the electrodes and the separator, while another dimension is in the radius direction of the particles for the solid-state diffusion processes. Both dimensions are evenly meshed, with 10 mesh elements in each electrode and the separator and 100 mesh elements in each particle. The time step is 0.1 s during the region used for fitting, and 5 s for the rest to shorten the simulation time.
\input{Table_parameters}
The applied current is 0.1 C = 0.498 mA/cm$^2$ for both the GITT and ICI simulations. In the GITT simulation, the current pulse is 900 s with 1.5 h of relaxation in between. In reality, lithium-ion cells take longer time to reach equilibrium again after the pulse, but 1.5 h is sufficient in these simulations. For the ICI procedure, the current is interrupted for 5 s every 5 min.
Three identification approaches are used to identify the diffusion coefficient $D_s$
\begin{itemize}
\item GITT full expression fitting according to \eqref{eq:full_expression_fitting}
\item GITT $\tau \ll 1$ approximation fitting according to \eqref{eq:tau<1 fitting}
\item ICI fitting according to \eqref{eq:ICI_fitting}
\end{itemize}
In \eqref{eq:tau<1 fitting} and \eqref{eq:ICI_fitting}, $dE/d \sqrt(t)$ is fitted within a selected region, as shown in Fig.\ref{fig:voltage_profile_illustration}. The starting point of the fitting region is one second after the current is switched in order to avoid the transient period of other electrochemical processes. In reality, this can be adjusted to a later point in time to exclude the effect of double layer capacitance. Different fitting lengths are used in the data processing. With the full expression fitting in \eqref{eq:full_expression_fitting}, the fitting length is unlimited. However, with the $\tau = D_st/r_s^2 \ll 1$ fitting in \eqref{eq:tau<1 fitting} and \eqref{eq:ICI_fitting}, the fitting length should not be too long to diverge from the conditions of approximation.
\section{Theory}
Following the DFN model, the mass transport process in the electrode particles is modelled by Fick's law in a spherical coordinate system
\begin{equation}
\frac{\partial c_s}{\partial t} = D_s (\frac{\partial ^2 c_s}{\partial r^2} + \frac{2}{r} \frac{\partial c_s}{\partial r})
\label{eq:Ficks_law}
\end{equation}
where $c_s$ is the lithium-ion concentration in the solid active material, $t$ is the time, $D_s$ is the solid-state diffusion coefficient of lithium ions, and $r$ is the distance from the particle center. When the cell is in an equilibrium state
\begin{equation}
c_s|_{t=0} = c_0
\end{equation}
If a constant current $j$ is applied to the surface of a particle at $t$ = 0 s, then the lithium ion flux at the surface is
\begin{equation}
\frac{\partial c_s}{\partial r}|_{r = r_s} = -j/(FD_s)
\end{equation}
where $F$ is the Faraday constant. The symmetric shape of the particle fulfills the condition
\begin{equation}
\frac{\partial c_s}{\partial r}|_{r=0} = 0
\end{equation}
The analytical solution of (1) with the boundary conditions in (2)-(4) is
\begin{equation}
c_{s,surf} = c_0 + \frac{j r_s}{F D_s} f(\frac{D_s t}{r_s^2})
\end{equation}
\begin{equation}
f(\frac{D_s t}{r_s^2}) = f(\tau) = 3 \tau +0.2 -2 \sum_{n=1}^{\infty} \frac{1}{\lambda _n^2} exp(-\lambda _n^2 \tau)
\end{equation}
where $\lambda _n$ is the positive roots of $\lambda = \tan (\lambda)$ \cite{nickol2020gitt}. The expression in (6) can be approximated to be
\begin{equation}
f(\tau) =
\begin{cases}
\frac{2}{\sqrt{\pi}} \sqrt{\tau} & \tau \ll 1\\
3 \tau +0.2 & \tau \gg 1
\end{cases}
\end{equation}
According to the mathematical description above, the diffusion coefficient $D_s$ could be found from the step response of $c_s$. In most cases, the concentration is very difficult to measure directly, therefore the terminal voltage $E$ - the experimentally easily measurable parameter - is used instead. The step response (5) is therefore rewritten as
\begin{equation}
E = E_0 + \frac{dE}{dc} \frac{j r_s}{F D_s} f(\frac{D_s t}{r_s^2})
\end{equation}
In (8), $dE/dc$ is often assumed to be a constant within a small SOC interval, i.e. the voltage profile is linear
\begin{equation}
\frac{dE}{dc} = \frac{\Delta E}{\Delta c} = \frac{\Delta E_{GITT}}{It_{pulse} /(FL\epsilon_s)}
\label{eq:dEdc}
\end{equation}
where $I$ is the current density on the electrode, $t_{pulse}$ is the time length of the current pulse, $L$ is the electrode length, $\epsilon_s$ is the volume fraction of the solid, and $\Delta E_{GITT}$ is the voltage difference between two equilibrium states before and after the current pulse. Furthermore, the local current density on particle $j$ in (8) is $j = I/(S_a L)$, where $S_a = 3 \epsilon_s/r_s$ is the specific surface area. Combining (8) and (9) gives
\begin{equation}
E = E_0 + \frac{ \Delta E_{GITT}}{t_{pulse}} \frac{r_s^2}{3D_s} f(\frac{D_s t}{r_s^2})
\label{eq:full_expression_fitting}
\end{equation}
When $\tau = D_st/r_s^2 \ll 1$, \eqref{eq:full_expression_fitting} can be approximated as
\begin{equation}
E = E_0 + \frac{2}{3}\frac{\Delta E_{GITT}}{t_{pulse}} \frac{r_s \sqrt{t}}{\sqrt{\pi D_s}}
\end{equation}
\noindent and thus the diffusion constant can be found to be
\begin{equation}
D_s = \frac{4}{9 \pi} (\frac{r_s}{dE/d\sqrt{t}} \frac{\Delta E_{GITT}}{t_{pulse}})^2
\label{eq:tau<1 fitting}
\end{equation}
The approach in \eqref{eq:tau<1 fitting} is one of the most commonly used fitting methods employed for GITT, and typically only a chosen region of the voltage step response is used to fit $dE/d\sqrt{t}$ to fulfill the approximation condition $\tau \ll 1$. The validity of this approximation is demonstrated in Fig.~\ref{fig:Fickslaw_solution}. It can be observed that the system response follows a linear relationship with $\sqrt{\tau}$ in the beginning and this approximation is considered to be valid when $\tau <$ 0.0032 \cite{chien2021fast}. The time axis is shown for two sets of parameters reported for LiNi$_{0.8}$Mn$_{0.1}$Co$_{0.1}$O$_2$: $r_s$ = 5.22 $\mu$m, $D_s = 1.48 \times 10^{-15}$ m$^2$/s \cite{chen2020development} and $r_s$ = 3.8 $\mu$m, $D_s = 5 \times 10^{-13}$m$^2$/s \cite{sturm2019modeling}. In the first case, $\tau <$ 0.0032 corresponds to $t = \tau r_s^2/D_s<$ 58.92 s and therefore the voltage response within this time length can be used to fit $dE/d\sqrt{t}$ in \eqref{eq:tau<1 fitting} for the diffusion coefficient identification. In contrast, the condition $\tau <$ 0.0032 corresponds to $t <$ 0.09 s in the latter case, meaning that the $\tau \ll 1$ approximation fails to describe the system response almost instantly. For such material properties, the expression in \eqref{eq:full_expression_fitting} can be used instead, which is not limited by the time constant of the system \cite{delacourt2011measurement}. In this work, both of the two methods will be examined. Moreover, it should be acknowledged that there exist other less common variants to estimate the $\Delta E/\Delta c$ in (8), resulting in similar expressions as those described above \cite{nickol2020gitt,wei2015kinetics,shen2013least}, but those variants will not be further discussed here.
\begin{figure}[!ht]
\centering
\includegraphics{figures/Fickslaw_solution.pdf}
\caption{The function in (6) is plotted as a solid line (n = 1272 for $\lambda _n$), while its approximations in (7) are plotted as dashed lines.}
\label{fig:Fickslaw_solution}
\end{figure}
In the GITT method, the solution of voltage response in (8) is solved with the initial condition where the cell is in an equilibrium state, meaning that a sufficient relaxation time period needs to take place between the current pulses. This generally very long time consumption has been considered to be the main disadvantage of GITT when compared with other methods. Alternatively, the more efficient ICI technique \cite{lacey2017influence} can be considered as a possible replacement. In ICI, the step response when the current is switched off is utilized. When the current is stopped at $t = t_1$, the concentration on the surface is
\begin{equation}
c_{s,surf} = c_0 + \frac{j r_s}{F D_s} f(\frac{D_s t}{r_s^2})-H(t)\frac{j r_s}{F D_s} f(\frac{D_s (t-t_1)}{r_s^2})
\end{equation}
\begin{equation}
H (t) =
\begin{cases}
0 & t<t_1\\
1 & t\geq t_1
\end{cases}
\end{equation}
where $t$ is counted from the time when the current is switched on \cite{chien2021fast}. Similar as in the GITT method, the initial concentration response when the current stops also follows a linear relationship with $\sqrt{t}$ according to the approximation in (7). In principle, the diffusion coefficient $D_s$ could be derived in a similar manner as in \eqref{eq:tau<1 fitting}, resulting in the expression
\begin{equation}
D_s = \frac{4}{9 \pi} (\frac{r_s}{dE/d\sqrt{t}} \frac{\Delta E_{ICI}}{\Delta t})^2
\label{eq:ICI_fitting}
\end{equation}
where $dE/d \sqrt{t}$ is fitted during the zero current period and $\Delta E_{ICI}/ \Delta t$ is estimated from a pseudo open circuit voltage, which is obtained by subtracting the IR drop from the voltage curve, as shown in Fig.~\ref{fig:voltage_profile_illustration}. Moreover, the ICI method does not require an equilibrium initial condition as in the GITT method, and the cell can instead be in a dynamic state under a constant current before the current interruption. This feature renders the ICI method considerably less time consuming.
\begin{figure}
\centering
\includegraphics{figures/voltage_profile_illustration_new_ICI.pdf}
\caption{The voltage response during one current pulse, comparing the GITT (above) and ICI (below) approaches.}
\label{fig:voltage_profile_illustration}
\end{figure}
It should be acknowledged that the derivation above is based the following assumptions:
\begin{itemize}
\item Fick's law is applied.
\item The particle has a spherical shape and the diffusion is symmetric.
\item The voltage profile is linear during a short interval, i.e. $dE/dc$ is a constant.
\item The current density at the particle surface is constant.
\item The over-potential caused by other electrochemical processes can be ignored.
\end{itemize}
The first two assumptions are the same as in the general DFN model, and their impact is thus not examined. The impact of the latter three assumptions will be studied hereafter to verify the validity of the GITT and ICI methods when they are applied for a porous lithium ion electrode.
|
\section{Introduction}
The discrepancies among the theoretical and experimental values of $R_{D^{(*)}}$ and $R_{K^{(*)}}$ observables might lead to a pathway to the discovery of beyond the standard model (BSM) physics. They are defined as follows,
\begin{align}
R_{D^{(*)}} = \dfrac{{\mathcal {B}}(B\rightarrow D^{(*)}\tau\bar\nu)}{{\mathcal {B}}(B\rightarrow D^{(*)}\hat{\ell}\bar\nu)} \quad \mbox{and}\quad R_{K^{(*)}} = \dfrac{{\mathcal {B}}(B\rightarrow K^{(*)}\mu^+\mu^-)}{{\mathcal {B}}(B\rightarrow K^{(*)}e^+e^-)}\ .\label{eq:anomalies}
\end{align}
The recent measurement of $R_{K^{(*)}}$ are smaller than the theoretical prediction by $3.1\sigma$ \cite{Bordone:2016gaq, Hiller:2003js, Blanke:2018sro, Calibbi:2015kma, Calibbi:2017qbu, Crivellin:2018yvo}. Whereas the measurements of the $R_{D^{(*)}}$ observables show a combined excess of $3.1\sigma$ \cite{HFLAV:2019otj} than the Standard Model (SM) value.
From the many models proposed to explain the anomalies, we focus on the $U_1\equiv(\mathbf{3},\mathbf{1},2/3)$ leptoquark (LQ) model that can explain both the anomalies simultaneously. LQs are colored triplet, scalar or vector bosons. They carry both lepton and baryon numbers and decay to the leptons and quarks. LQs appear in many models like, $SU(4)$ Pati-Salam models, Technicolor models, RPV SUSY models, etc. Their coupling to SM leptons and quarks makes them an ideal candidate to explain the $B$-anomalies.
\section{The $\rm{U}_1$ Lagrangian}
The Lagrangian for the interactions between $\rm{U_1}$ and the SM quarks and leptons is given as,
\begin{equation}
\label{eq:GenLagU1}
\mathcal{L} \supset
x^{LL}_{1~ij}~\bar{Q}^{i}\gamma_{\mu}U^{\mu}_{1}P_{L}L^{j} + x_{1~ij}^{RR}~\bar{d}^i_{R}\gamma_{\mu}U^{\mu}_{1}P_{R}\ell^j_{R} + \textrm{h.c.}
\end{equation}
Here, $Q_i$ and $L_j$ are the SM left-handed quark and lepton doublets, respectively. The $d_{R}^{i}$ and $\ell_{R}^{j}$ are the down-type right-handed quarks and leptons, respectively. The indices $i,j=\{1,\ 2,\ 3\}$ stand for the quark and lepton generations. Thus, the $x_{1~ij}^{LL}$ and $x_{1~ij}^{RR}$ are $3\times 3$ matrices in the flavour space. Here, we consider only real values of the couplings in the matrices and set all couplings but the ones involved in the $b\to c\tau\bar{\nu}$ (i.e., $\lambda^L_{23}$, $\lambda^{L/R}_{33}$) and $b\to s\mu^+\mu^-$ (i.e., $\lambda^{L/R}_{22}$, $\lambda^{L/R}_{32}$) decays to be zero.
\begin{equation}
x^{LL}_1 =
\begin{pmatrix}
0 & 0 & 0 \\
0 & {\lambda^L_{22}} & \lambda^L_{23} \\
0 & {\lambda^L_{32}} & \lambda^L_{33}
\end{pmatrix};~~
x^{RR}_1 =
\begin{pmatrix}
0 & 0 & 0 \\
0 & {\lambda^R_{32}} & 0 \\
0 & {\lambda^R_{32}} & \lambda^R_{33}
\end{pmatrix}. \label{eq:rkcouplings}
\end{equation}
\subsection{The $R_{D^{(*)}}$ \& $R_{K^{(*)}}$ scenarios}
To explain the $R_{D^{(*)}}$ scenarios, the new physics needs to contribute to the $b\to c\tau\bar{\nu}$ transition. Hence, in our case, the $\rm{U}_1$ needs to couple with both $c$-$\nu$ and $b$-$\tau$ pairs. To obtain the required couplings, we first consider some minimal single-coupling scenarios and then move on to more complex multi-coupling scenarios. In the first single-coupling scenario, we
consider only $\lambda^L_{23}$ to be non-zero in the following manner,
\begin{equation}
\mathcal{L} \supset \lambda_{23}^{L}[\bar{c}_L\gamma_{\mu}\nu_{L} + \bar{s}_L \gamma_{\mu}\tau_{L})] U_{1}^{\mu}
= \lambda_{23}^{L}[\bar{c}_L\gamma_{\mu}\nu_{L} + (V^*_{cd}\bar{d}_L + V^*_{cs}\bar{s}_L + V^*_{cb}\bar{b}_L) \gamma_{\mu}\tau_{L})] U_{1}^{\mu},
\end{equation}
i.e., the $c$-$\nu$ coupling is generated directly but the $b$-$\tau$ coupling is an effective one, generated by the mixing among the down-type quarks. In ref.~\cite{Bhaskar:2021pml}, this scenario is called RD1A.
Similarly, if we put all but $\lambda^L_{33}$ to be zero, i.e., we generate the $b$-$\tau$ coupling directly and $c$-$\nu$ coupling via quark mixing, we get the scenario RD1B with the following lagrangian,
\begin{equation}
\mathcal{L} \supset \lambda_{33}^{L}[\bar{t}_L \gamma_{\mu}\nu_{L} + \bar{b}_L\gamma_{\mu}\tau_{L}] U_{1}^{\mu}
= \lambda_{33}^{L}[(V_{ub}\bar{u}_L + V_{cb}\bar{c}_L + V_{tb}\bar{t}_L) \gamma_{\mu}\nu_{L}) + \bar{b}_L\gamma_{\mu}\tau_{L}] U_{1}^{\mu}.
\end{equation}
The possible single and multi-coupling scenarios have been summarized in Table~\ref{tab:RDobs}.
There we also show some of the possible scenarios one can construct to accommodate the $R_{K^{(*)}}$ anomalies. These scenarios may seem similar from an effective-field-theory point of view, but their LHC signatures are quite different. Further details on these scenarios can be found in ref.~\cite{Bhaskar:2021pml}.
\begin{table}[t!]
\caption{Summary of the single and multi-coupling scenarios to explain the $R_{D^{(*)}}$ and $R_{K^{(*)}}$ anomalies (from ref.~\cite{Bhaskar:2021pml}).}
\begin{center}
{\linespread{1.3}\footnotesize
\begin{tabular*}{\textwidth}{l@{\extracolsep{\fill}}ccclcccc}
\hline
$R_{D^{(*)}}$ scenarios & $\lambda_{c\nu}^L$ & $\lambda_{b\tau}^L$ & $\lambda_{b\tau}^R$ &$R_{K^{(*)}}$ scenarios&$\lambda_{s\mu}^L$ & $\lambda_{b\mu}^L$ & $\lambda_{s\mu}^R$ & $\lambda_{b\mu}^R$ \\
\hline\hline
RD1A & $\lambda_{23}^L$ & $V_{cb}^*\lambda_{23}^L$ & $-$ & RK1A & $V_{cs}^*\lambda_{22}^L$ & $V_{cb}^*\lambda_{22}^L$ & $-$ & $-$ \\
RD1B & $V_{cb}\lambda_{33}^L$ & $\lambda_{33}^L$ & $-$ & RK1B & $V_{ts}^*\lambda_{32}^L$ & $V_{tb}^*\lambda_{32}^L$ & $-$ & $-$ \\
& & & & RK1C & $-$ & $-$ & $V_{cs}\lambda_{22}^R$ & $V_{cb}\lambda_{22}^R$ \\
& & & & RK1D & $-$ & $-$ & $V_{ts}\lambda_{32}^R$ & $V_{tb}\lambda_{32}^R$ \\
\hline
RD2A & $V_{cs}\lambda_{23}^{L} + V_{cb}\lambda_{33}^{L}$ & $\lambda_{33}^{L}$ & $-$ & RK2A & $\lambda_{22}^L$ & $\lambda_{32}^L$ & $-$ & $-$ \\
RD2B & $V_{cs}\lambda_{23}^{L}$ & $-$ & $\lambda_{33}^{R}$ & RK2B & $\lambda_{22}^L$ & $-$ & $-$ & $\lambda_{32}^R$ \\
& & & & RK2C & $-$ & $\lambda_{32}^L$ & $\lambda_{22}^R$ & $-$ \\
& & & & RK2D & $-$ & $-$ & $\lambda_{22}^{R}$ & $\lambda_{32}^{R}$ \\\hline
RD3 & $V_{cb}\lambda^L_{33} + V_{cs} \lambda^L_{23}$ & $\lambda^L_{33}$ & $\lambda_{33}^{R}$ & RK4 & $\lambda_{22}^{L} $ & $\lambda_{32}^{L}$ & $\lambda_{22}^{R}$ & $\lambda_{32}^{R}$ \\
\hline
\end{tabular*}
}
\label{tab:RDobs}
\end{center}
\end{table}
\section{Production at the LHC and data recast}
We generate Monte Carlo events for all LQ production modes in the $R_{D^{(*)}}$ and $R_{K^{(*)}}$ scenarios. While recasting the LHC dilepton data~\cite{ATLAS:2020zms,CMS:2021ctt} we systematically combine the events that can contribute to these searches~\cite{Mandal:2012rx,Mandal:2015vfa,Mandal:2016csb,Mandal:2018kau,Aydemir:2019ynb}. The $U_1$ can be produced at the LHC via two types of processes---resonant productions (like pair and the single productions) and non-resonant productions (like $t$-channel LQ exchange). In the RD1A (RD1B) scenario, the pair production mode, which is mostly QCD mediated, gives the following final states,
\begin{equation}
\label{eq:pairRD1A}
pp\to U_1 U_1 \rightarrow s\tau \, s\tau (b\tau \, b\tau)\equiv \tau\tau + 2j.
\end{equation}
The relevant single-production channels in the $R_{D^{(*)}}$ ($R_{K^{(*)}}$) scenarios can be written as
\begin{equation}
\label{eq:singRD1A}
pp\to U_1\tau(\mu) + U_1\tau(\mu) j \rightarrow \tau\tau j(\mu\mu j) + \tau\tau jj(\mu\mu jj) \equiv \tau\tau(\mu\mu) + n\,j.
\end{equation}
The single productions depend on the $U_1$-$\ell$-$q$ couplings ($\lambda$) and play a significant role in the limits.
The $\lambda$-dependent non-resonant process where a $\rm{U}_1$ is exchanged in the $t$-channel can lead to $\tau\tau$ ($\mu\mu$) final states.
Moreover, this process interferes with the SM background processes like $pp\to Z\gamma^*\to\tau\tau (\mu\mu)$.
Since there is destructive interference between the BSM process and the SM background, we use a $\chi^2$ test instead of a simple recast to recast the latest $\tau\tau$ ($\mu\mu$) search data from ATLAS~\cite{ATLAS:2020zms} (CMS~\cite{CMS:2021ctt}) to obtain bounds. The $\chi^2$ can be expressed as follows~\cite{Bhaskar:2021pml},
\begin{equation}
\chi^2 = \sum_i^{\rm bins} \left(\frac{ \left[N_{p}+N^{incl}_s+N_t-N_{\times}\right]^i + N_{BG}^i - N_{D}^i }{\sqrt{\left(\Delta N^i_{Stat}\right)^2+\left(\Delta N^i_{Syst}\right)^2}}\right)^2
\label{eqnchisq}
\end{equation}
where $N_{p}$, $N^{incl}_s$, $N_t$, $N_{\times}$, $N_{BG}$, and $N_{D}$ are the number of events from pair production, single production, pure-BSM ($t$-channel $U_1$ exchange), the BSM-SM interference, the SM background, and the data, respectively and $N^i_{Stat}=\sqrt{N_D^i}$ and $N^i_{Syst}=0.1\, N_D^i$ are the statistical and systematic uncertainties. For a range of $U_1$ masses, we use the total transverse mass distribution from the $\tau\tau$ data and the invariant mass distribution from the $\mu\mu$ data to find the best fit parameters (i.e., couplings) and the $2\sigma$ exclusion limits on them.
\section{Results}
\begin{figure*}[!t]
\centering
\captionsetup[subfigure]{labelformat=empty}
\subfloat[\quad\quad\quad(a)]{\includegraphics[width=0.4\textwidth]{LM23L_combined}\label{fig:RD1alla}}\quad\quad
\subfloat[\quad\quad\quad(b)]{\includegraphics[width=0.4\textwidth]{1500_RD2A_combflalim}\label{fig:RD1allb}}
\caption{(a) The region (in purple) excluded by the LHC $\tau\tau$ data and the region (in yellow) favoured by the flavour observables in a $\lambda^L_{23}$-$M_{{U}_1}$ plane. (b) For a $1.5$TeV ${U}_1$, the blue color region is allowed by the LHC $\tau\tau$ data (from ref.~\cite{Bhaskar:2021pml}).}
\label{fig:RD1all}
\end{figure*}
\begin{figure*}[!t]
\centering
\captionsetup[subfigure]{labelformat=empty}
\subfloat[\quad\quad\quad(a)]{\includegraphics[width=0.4\textwidth]{RK_LHC_LM22L}\label{fig:RK12alla}}\quad\quad
\subfloat[\quad\quad\quad(b)]{\includegraphics[width=0.4\textwidth]{LM22L_32L_2000}\label{fig:RK12allb}}
\caption{a) Excluded region (purple) by the LHC and the $R_{K^{(*)}}$-favoured region (yellow). (b) For a $2.0$ TeV $U_1$, the regions allowed by the LHC (in blue) and relevant to the $R_{K^{(*)}}$ observables (in yellow)~\cite{Bhaskar:2021pml}.}
\label{fig:RK12all}
\end{figure*}
\begin{figure*}[!t]
\captionsetup[subfigure]{labelformat=empty}
\centering\includegraphics[width=0.4\textwidth]{plot_RK2A_1500}\label{fig:RK2Aa}
\caption{Region in the $\lambda^L_{22}$-$\lambda^L_{32}$ plane for a $1.5$ TeV $U_1$ allowed by the LHC data (blue) and simultaneously explaining the $R_{D^{(*)}}$ and $R_{K^{(*)}}$ anomalies (yellow)~\cite{Bhaskar:2021pml}.}
\label{fig:RK1500}
\end{figure*}
In Figure~\ref{fig:RD1alla}, we show the exclusion limits on $\lambda^L_{23}$ in purple ($\tau\tau$) and the $R_{D^{(*)}}$-favoured region (FlavD, which also satisfies other flavour bounds like $F_L(D^*)$, $P_\tau(D^*)$ etc.~\cite{Bhaskar:2021pml}) in yellow. Exclusion region (in magenta) come from direct searches of LQ. We find the single-coupling $R_{D^{(*)}}$ scenarios are not favoured by the LHC data. However, we see an improvement in a two-coupling $R_{D^{(*)}}$ scenario. We show the region allowed by the LHC data in blue for the RD2A scenario for a $1.5$ TeV $U_1$ in Figure~\ref{fig:RD1allb}. There is a considerable overlap between this region and the one required by the relevant flavour observables (in yellow and green).
In Figure \ref{fig:RK12alla}, we plot the limits on the single coupling scenario RK1A (where only $\lambda^L_{22}\neq0$). Here we see the LHC $\mu\mu$ data (in purple) is not much restrictive for the $R_{K^{(*)}}$-favoured region (FlavK). FlavK (in yellow) also consists of other flavour observables relevant to this particular scenario. However, a recast of the direct search for scalar LQ~ \cite{ATLAS:2020dsk} (magenta dotted line) effectively rules out a $1.5$ TeV LQ in the one- or two-coupling $R_{K^{(*)}}$-scenarios (Table \ref{tab:RDobs}). We show the region allowed by the LHC data for a heavier LQ in the RK2A scenario in Figure~\ref{fig:RK12allb}.
However, a $1.5$ TeV $U_1$ can simultaneously revolve both the anomalies. In Figure~\ref{fig:RK1500}, we show such a possibility where we have considered four new couplings to be nonzero (two relevant for the $R_{D^{(*)}}$-observable and the other two, for $R_{K^{(*)}}$).\footnote{See ref.~\cite{Mandal:2018kau} for a similar analysis with the scalar LQ $S_1$ and refs.~\cite{Chandak:2019iwj,Bhaskar:2021gsy} for the discovery prospects of LQs that couples with third generation quarks at the HL-LHC.}
\section{Conclusions}
We obtained precise bounds on the parameter space of the $U_1$ leptoquark in all possible simple $R_{D^{(*)}}$- and $R_{K^{(*)}}$- anomalies-motivated scenarios from the latest $\tau\tau$ and $\mu\mu$ resonance search data at the LHC. The current dilepton data from the LHC put stringent bounds on the $U_1$ model parameters needed to explain the $R_{D^{(*)}}$ and the $R_{K^{(*)}}$ anomalies. The obtained bounds are precise since we considered the contribution of all production modes (including the signal-background interference contribution) in the signal while recasting. In the low-mass region, the resonant modes play a significant role in obtaining the limits. We found a $1.5$ TeV LQ can explain both $R_{D^{(*)}}$ and $R_{K^{(*)}}$ anomalies simultaneously while satisfying all relevant limits.
\acknowledgments
A.B. and S.M. acknowledge support from the Science and
Engineering Research Board (SERB), DST, India, under Grant No. ECR/2017/000517. D.D. acknowledges the DST, Government of India for the INSPIRE Faculty Fellowship (Grant No. IFA16-PH170). T.M. is supported by the intramural grant from IISER-TVM. C.N. is supported by the DST-Inspire Fellowship.
|
\section{Introduction}
\label{sec:introduction}
Racing requires each vehicle to make sub-second decision in a fast changing environment and operate at its physical limits \cite{liniger2015optimization}, when any safety infraction could lead to catastrophic failure. Thus, autonomous racing is a particularly challenging proving ground for autonomous agents to optimize performance, while adhering to safety constraints. In the reinforcement learning (RL) literature, it is common to define safety as satisfying safety specifications \cite{ray2019benchmarking} under the constrained Markov decision process (CMDP) framework \cite{altman1999constrained}, which extends the Markov decision process (MDP) by incorporating constraints on expected cumulative costs.
Due to the low sensor cost and high information content, camera-based perception is gaining increasing popularity in autonomous vehicles \cite{strobel2020accurate}. While end-to-end autonomous driving on visual input is an extensively-researched topic for urban driving, largely thanks to the release of the CARLA simulator \cite{dosovitskiy2017carla}, it is less so for high-speed racing, which may be partly attributed to the lack of open-source, high-fidelity simulation environments. The recent release of \texttt{Learn-to-Race}~(\texttt{L2R}) \cite{herman2021learn} changes that and lowers the barrier of entry for autonomous racing research.
In this work, we study the problem of constrained RL for autonomous racing, using the vehicle's ego-camera view and speed as input. Due to the nature of the task, the autonomous agent needs to be able to 1) identify and avoid unsafe scenarios and 2) make fast safety verification given the perception data. In Figure \ref{fig:safe_examples}, we show examples of ego-camera views and the corresponding safety value $Q_S(x, u)$, estimated by our proposed safety critic, and the distance to road boundary $l(x)$. While it is straightforward to determine whether a state is safe based on the vehicle pose, which is illustrated for reference, the distinction from ego-camera views is much more subtle. Also evident from the examples is that safety does not necessarily corresponds to distance to road boundary.
Regarding the requirement for fast decision-making, in our experiments, \texttt{L2R}~operates under the setting where the simulator executes the agent's command upon receiving it, and does not wait for the agent to complete its computation. Thus, high latency can adversely impact agent performance, where, as discussed in prior art \cite{strobel2020accurate}, perception stacks in autonomous race-cars account for nearly 60\% of total latency.
\begin{figure*}[t!]
\vspace{-0.5cm}
\sbox\twosubbox{%
\resizebox{\dimexpr\textwidth-1em}{!}{%
\includegraphics[height=3cm]{figures/fm_examples.pdf}
\hspace{0.5cm}
\includegraphics[height=3cm]{figures/framework_new.pdf}
}
}
\setlength{\twosubht}{\ht\twosubbox}
\centering
\subcaptionbox{Examples of Safe vs. Unsafe States\footnotemark \label{fig:safe_examples}}{%
\includegraphics[height=0.99\twosubht]{figures/fm_examples.pdf}}
\hspace{0.5cm}
\subcaptionbox{SAGE~ Architecture \label{fig:spar:architecture}}{%
\includegraphics[height=0.99\twosubht]{figures/framework_new.pdf}}
\caption[]{SAGE~Overview. (a) The safety critic, $Q_S$, verifies the safety of a state-action pair by checking if $Q_S(x, u)\geq\epsilon$. Some examples of safe vs. unsafe states are provided, using safety margin $\epsilon=3$, $u=\mathbf{0}$, and speed = 10m/s. (While vehicle pose is NOT available to the safety critic, we illustrate them here for reference.) (b) SAGE~consists of two policies, which are in charge of safety and performance, independently. The safety controller intervenes when the current state-action pair is deemed unsafe by the safety critic.
}
\vspace{-0.1cm}
\label{fig:framework}
\end{figure*}
Given these considerations, we propose to incorporate Hamilton-Jacobi (HJ) reachability theory, a safety verification method for general non-linear systems, into the CMDP framework. HJ reachability not only provides a control-theoretical approach to learn about safety, but also enables low-latency safety verification. As a reachable set takes into consideration all possible trajectories over a specified time horizon, safety verification via HJ reachability only requires evaluating the safety value of the current state.
Furthermore, safety verification under HJ Reachability theory does not depend on the performance policy. Thus, we can bypass the challenges involved with solving a constrained optimization problem with a neural policy, and decompose the problem of learning under safety constraints into (a) optimizing for performance, and (b) updating the safety value function. Given this intuition, we learn two policies that independently manage safety and performance (Figure \ref{fig:spar:architecture}): the performance policy focuses exclusively on optimizing performance, while the safety critic verifies if the current state is safe and intervenes when necessary. We refer to our approach as \textbf{S}afe \textbf{A}utonomous racin\textbf{G} on \textbf{E}go-vision (SAGE).
\footnotetext{Unsafe here refers to $Q_S(x, u)<\epsilon$. }
Aside from our proposed method, our key contributions are as follows. Due to the scalability issue \cite{bansal2017hamilton}, existing works on HJ reachability exclusively study problems defined on physical states, e.g., vehicle pose, instead of high-dimensional sensory inputs, such as RGB images. We demonstrate that the HJ safety value function can be learned directly on visual context via neural approximation, the highest-dimensional problem studied by HJ safety analysis to-date, thereby expanding the applications of HJ reachability to high-dimensional systems where dynamics models may not be available.
Secondly, we compare the HJ Bellman update rule \cite{fisac2019bridging} to alternatives for learning a safety critic \cite{srinivasan2020learning, bharadhwaj2020conservative} on two classical control benchmarks, where safe
states are known, analytically. Given the same off-policy samples, the HJ Bellman update rule is more accurate and sample efficient.
Finally, we evaluate our methods on \texttt{Safety Gym} \cite{ray2019benchmarking} and \texttt{Learn-to-Race}~(\texttt{L2R}) \cite{herman2021learn}, a recently-released, high-fidelity autonomous racing environment, which challenges the agent to make safety-critical decisions in a complex and fast-changing environment. While SAGE~is by no means free from failure, it has significantly fewer constraint violations compared to other constrained RL baselines in \texttt{Safety Gym}. We also report new state-of-the-art results on the \texttt{L2R}~benchmark task, and show that incorporating a learnable safety critic grounded in control theory boosts performance especially during the initial learning phase.
\begin{comment}
\begin{itemize}
\item Motivation: Autonomous agents should respect safety constraints.
\blue{(Reminder: Can either pitch as a l2r application or formulate in a more general way)} \bc{Disambiguate safety}
\item Model-free vs Model-based; Drive home the point that model-free methods which depend on on-policy samples to estimate whether an agent is going to violate constraints, is doomed to fail during the initial learning phase. ``Safety is not determined by how much a system fails on average, but by whether it fails at all."
\item Motivate the benefits of having a model; Briefly mention other related approaches; Describe HJI and its strength;
\item However, there are challenges with HJI: 1) computational cost grows exponentially with the state space, and 2) the need of a model.
\item
\end{itemize}
\end{comment}
\begin{comment}
\paragraph{Notes} Compared to urban driving, the racing car must make safety-critical, real-time decisions based on multimodal inputs in a fast-changing environment, with consideration of competing agents in multi-agent racing. Further, a racing car operates at its physical limit and is significantly less stable compare to a street car, where small mistakes can steer the car out-of-boundary.
Current algorithmic solutions are split between:
More classical approaches: demand privileged information, require significant parameter-tuning, are limited in performance capacity
Approximate driving methods: provide no guarantees of safety and are prone to overfitting on the training scenarios.
Without constraints, autonomous agents are not guaranteed to adhere to safe behaviors---neither during learning, nor during task execution.
We explore ways, e.g., HJ Reachability, to embed logical constraints in autonomous agents, while still allowing them to explore sample-efficiently and safely
Approaches based on HJ reachability theory require offline calculation of the safe set, as a costly partial derivative of the value function (given privileged information about the track, vehicular dynamics, etc.), where the expense in solving the PDE scales exponentially with the dimension of the state space
Addressing challenges in safe policy learning and optimization, we propose the use of safety constraint sets for autonomous racing, inspired by Hamilton-Jacobi (HJ) Reachability theory
We will further show that an arbitrary learning policy that is coupled with this safe controller is able to learn performant driving behavior, both safely and sample-efficiently.
Current algorithmic solutions are split between more classical approaches that demand privileged information, require significant parameter-tuning, and are limited in their performance capacity; versus approximate driving methods that provide no guarantees of safety and are prone to overfitting on the training scenarios.
To address these challenges in safe policy optimization, we propose the use of safety constraint sets for autonomous racing, inspired by the theoretical foundations of Hamilton-Jacobi (HJ) Reachability. Here, we define a safety controller that intervenes whenever an agent approaches bad states, and we show that even an agent that generates actions at random is guaranteed to stay on the drivable area; we further show that an arbitrary learning policy that is coupled with this safe controller is able to learn performant driving behavior, both safely and sample-efficiently.
\end{comment}
\section{Related Work}
\label{sec:related_work}
\para{Autonomous racing.} One approach for autonomous racing is via model predictive control \cite{liniger2015optimization, rosolia2017autonomous, kabzan2019learning}, which solves an optimization problem with a model of the system dynamics. Aside from the challenges in modeling the complex dynamics, a significant drawback of such approach is the dependence on extensive sensor installation for localization and state estimation \cite{cai2021vision}. Another approach is to use a modular pipeline \cite{kabzan2019learning, strobel2020accurate}, starting from perception on raw sensory inputs, to localization and object-detection, and finally to planning and control. While this approach is most commonly used in practice, disadvantages of the approach include over-complexity and error propagation \cite{yurtsever2020survey, francis2021core}. Recently, there is a lot of interest in using RL-based approaches for autonomous racing. In \cite{fuchs2020superhuman, chisari2021learning}, RL agents were trained using low-dimensional features as inputs. In \cite{chen2015deepdriving, drews2017aggressive}, intermediate features were extracted from perception pipelines to determine control actions. In \cite{cai2021vision, weiss2020deepracing}, RL agents were trained end-to-end on visual inputs by imitating expert demonstration; in \cite{cai2021vision}, a data-driven model of the environment was further utilized to train the agent by unrolling future trajectories.
In comparison to racing, there is significantly more literature on end-to-end autonomous driving for urban scenarios \cite{codevilla2018end, ohn2020learning, codevilla2019exploring, chen2020learning, zhang2021learning, prakash2021multi, zhang2021end, zhang2021learning}. It is beyond our scope to cover this large research field, and we refer interested readers to survey papers, such as \cite{yurtsever2020survey, grigorescu2020survey}, for more information. While we focus on high-speed racing its unique challenges, we believe the discussion here for safety analysis on ego-vision is also relevant to urban driving.
\begin{comment}
\begin{itemize}
\item Optimization-based: ; Downside: Requires a model of the vehicle, Requires additional sensors for
\item Modular Vision Pipeline: ; Downside: error propagation and over-complexity \cite{yurtsever2020survey}
\item RL-based: selected states \citet{fuchs2020superhuman, chisari2021learning}; \cite{chen2015deepdriving, drews2017aggressive} extract intermediate representation from perception with which determines the control actions;
\item RL on vision \cite{cai2021vision}
\item Autonomous driving: classic imitative modeling approach to urban driving scenarios ; Survey \cite{yurtsever2020survey, grigorescu2020survey}
\end{itemize}
\end{comment}
\para{Constrained reinforcement learning.} There is growing interest in enforcing some notion of safety in RL algorithms, e.g., satisfying safety constraints, avoiding worst-case outcomes, or being robust to environmental stochasticity \cite{garcia2015comprehensive}.
We focus on the notion of safety as satisfying constraints. CMDP \cite{altman1999constrained} is a widely-used framework for studying RL under constraints, where the agent maximizes cumulative rewards, subject to limits on cumulative costs characterizing constraint violations. Solving a CMDP problem is challenging, because the policy needs to be optimized over the set of feasible states; this requires off-policy evaluation of the constraint functions, to determine whether a policy is feasible \cite{achiam2017constrained}. As a result, safety grows with experience, but requires diverse state-action pairs, including unsafe ones \cite{srinivasan2020learning}.
Furthermore, one needs to solve a constrained optimization problem with a non-convex neural policy. This may be implemented with techniques inspired by convex optimization, such as primal-dual updates \cite{bharadhwaj2020conservative} and projection \cite{yang2020projection}, or by upper bounding the expected cost at each policy iteration \cite{achiam2017constrained}.
Most relevant to our work is \cite{bharadhwaj2020conservative, srinivasan2020learning, thananjeyan2021recovery}, which also uses a safety critic to verify if a state is safe; we compare our control-theoretic learning rule with theirs in Section \ref{sec:toy_exp}
\para{Guaranteed safe control.}
Guaranteeing the safety of general continuous nonlinear systems is challenging, but there are several approaches that have been successful. These methods typically rely on knowledge of the environment dynamics.
Control barrier functions (CBFs) provide a measure of safety with gradients that inform the acceptable safe actions \cite{ames2019control}.
For specific forms of dynamics, e.g., control-affine \cite{cheng2019end}, and unlimited actuation bounds, this approach can be scalable to higher-dimensional systems and can be paired with an efficient online quadratic program for computing the instantaneous control \cite{cheng2019end}. Unfortunately, finding a valid control barrier function for a general system is a nontrivial task. Lyapunov-based methods \cite{chow2018lyapunov, chow2019lyapunov} suffer from the same limitation of requiring hand-crafted functions.
HJ reachability is a technique that uses continuous-time dynamic programming to directly compute a value function that captures the optimal safe control for a general nonlinear system \cite{bansal2017hamilton, fisac2018general}. This method can provide hard safety guarantees for systems, subject to bounded uncertainties and disturbances. There are two major drawbacks to HJ reachability. The first is that the technique suffers from the curse of dimensionality and scales exponentially with number of states in the system. Because of this, the technique can only be used directly on systems of up to 4-5 dimensions \cite{bansal2017hamilton}. When using specific dynamics formulations and/or restricted controllers, this upper limit can be extended \cite{chen2018decomposition, kousik2020bridging}. Second, because of this computational cost, the value function is typically computed offline based on assumed system dynamics and bounds on uncertainties. This can lead the safety analysis to be invalid or overly conservative.
There are many attempts in injecting some form of control theory into RL algorithms. In comparison to works that assume specific problem structure \cite{cheng2019end, dean2019safely} or existence of a nominal model \cite{cheng2019end, bastani2021safe}, our proposed approach is applicable to general nonlinear systems and does not require a model. But, we do assume access to a distance metric defined on the state space.
Our primary inspiration is recent work by \cite{fisac2019bridging} that connects HJ reachability with RL and introduces a HJ Bellman update, which can be applied to deep Q-learning for safety analysis. This method loses hard safety guarantees due to the neural approximation, but enables scalable learning of safety value function. However, an agent trained using the method in \cite{fisac2019bridging} will focus exclusively on safety. Thus, we extend the method by formulating it within the CMDP framework, thereby enabling performance-driven learning.
\section{Preliminaries}
\label{sec:preliminaries}
\vspace{-0.2cm}
\para{Constrained MDPs.} The problem of RL with safety constraints is often formulated as a CMDP. On top of the MDP tuple $(\mathcal{X}, \mathcal{U}, R, \mathcal{F})$, where $\mathcal{X}$ is the state space, $\mathcal{U}$ is the action space, $\mathcal{F}: \mathcal{X} \times \mathcal{U}\xrightarrow{} \mathcal{X}$ characterizes the system dynamics, and $R:\mathcal{X} \times \mathcal{U} \xrightarrow{} \mathbb{R} $ is the reward function, CMDP includes an additional set of cost functions, $\{C_1,\dots,C_m\}$, where each $C_i:\mathcal{X} \times \mathcal{U} \xrightarrow{} \mathbb{R}$ maps state-action transitions to costs characterizing constraint violations.
The objective of RL is to find a policy $\pi: \mathcal{X}\xrightarrow{} \mathcal{P}(\mathcal{U})$ that maximizes the expected cumulative rewards, $V^\pi_R(x) =\mathbb{E}_{x_k, u_k \sim \pi} \left[\sum_{k=0}^\infty \gamma^k R(x_k, u_k)|x_0 =x\right]$, where $\gamma\in [0, 1)$ is a temporal discount factor. Similarly, the expected cumulative costs are defined as $V^\pi_{C_i}(x) = \mathbb{E}_{x_k, u_k \sim \pi} \left[\sum_{k=0}^\infty \gamma^k C_i(x_k, u_k)|x_0 =x\right]$;
CMDP requires the policy to be feasible by imposing a limit for the costs, i.e., $V_{C_i}(\pi)\leq \chi_i, \forall i$. Putting everything together, the RL problem in a CMDP is:
\begin{align}
\label{eq:cmdp}
\pi^{*} = \arg\max_{\pi} \;\; V^\pi_R(x)
\quad \textrm{s.t.} \;\;\ V^{\pi}_{C_i}(x)\leq \chi_i \;\;\forall i.
\end{align}
\para{HJ Reachability.} To generate the safety constraint, one can apply HJ reachability to a general nonlinear system model, denoted as $\dot{x} = f(x,u)$. Here $x\in \mathbb{R}^n$ is the state, $u$ is the control contained within a compact set $\mathcal{U}$. $f$ is assumed uniformly continuous and bounded, and Lipschitz in x for all u. For discrete-time approximations, the time step $\Delta t >0$ is used.
We denote all allowable states as $\mathcal{K}$, for which there exists a terminal reward $l(x)$, such that $ x\in \mathcal{K} \iff l(x)\geq 0$. An $l(x)$ that satisfies this condition is the signed distance to the boundary of $\mathcal{K}$. Taking autonomous driving as an example, $\mathcal{K}$ is the drivable area and $l(x)$ is the distance to road boundary or obstacle. This set $\mathcal{K}$ is the complement of the failure set that must be avoided. The goal of this HJ reachability problem is to compute a safety value function that maps a state to its safety value, with respect to $l(x)$, over time. This is done by capturing the minimum reward achieved over time by the system applying an optimal control policy:
\begin{equation}
V_S(x,T) = \sup_{u(\cdot)}\min_{t\in [0, T]} l(\xi^{u}_{x,T}(t)),
\end{equation}
where $\xi$ is the state trajectory, $T<0$ is the initial time, and $0$ is the final time. To solve for this safety value function, a form of continuous dynamic programming is applied backwards in time, from $t = 0$ to $t = T$, using the Hamilton-Jacobi-Isaacs Variational Inequality (HJI-VI):
\begin{multline}
\min \left\{ \frac{\partial V_S}{\partial t}+ \max_{u \in \mathcal{U}} \langle f(x, u), \nabla V_S(x) \rangle, l(x)-V_S(x, t) \right\}=0,
\\V_S(x, 0) = l(x).
\label{eq:HJIVI}
\end{multline}
The super-zero level set of this function is called the reachable tube, and describes all states from which the system can remain outside of the failure set for the time horizon. For the infinite-time, if the limit exists, we define the converged value function as $V_S(x) = \lim_{T \rightarrow -\infty}V_S(x,T)$. While the HJI-VI is difficult to solve, once solved, safety verification only requires evaluating the safety value of the current state.
Once the safety value function is computed, the optimal safe control can be found online by solving the Hamiltonian: $\pi_S^*(x) = \arg \max_{u \in \mathcal{U}} \langle f(x, u), \nabla V_S(x) \rangle$. This safe control is typically applied in a least-restrictive way, wherein the safety controller becomes active only when the system approaches the boundary of the reachable tube, i.e., $u\sim \pi$ if $V_S(x,T)\geq 0$ and $\pi_S^*$ otherwise.
The newly introduced discounted safety Bellman equation \cite{fisac2019bridging} modifies the HJI-VI in \eqref{eq:HJIVI} in a time-discounted formulation for discrete time:
\begin{multline}
V_S(x) = (1-\gamma)l(x) + \gamma \min\left\{ l(x), \max_{u\in \mathcal{U}} V_S(x+f(x,u)\Delta t) \right\},\\
V_S(x, 0) = l(x).
\end{multline}
This formulation induces a contraction mapping, which enables convergence of the value function when applied to dynamic programming schemes, commonly used in RL.
\section{Safe Autonomous racinG on Ego-vision}
\label{sec:approach}
In this section, we describe our framework for safety-aware autonomous racing. We are inspired by guaranteed-safe methods, such as HJ reachability, which provides a systematic way to verify safety. Thus, we formulate our problem as a combination of constrained RL and HJ reachability theory, adopting a control-theoretic approach to learn safety.
Building upon prior work on neural approximation of HJ Reachability \cite{fisac2019bridging}, we demonstrate that it is possible to directly update the safety value function on high-dimensional sensory input, thereby expanding the scope of applications to problems previously inaccessible. We highlight the notable aspects of our framework:
\parai{i) HJ reachability provides a control-theoretic and low-latency way to verify safety.} By incorporating HJ Reachability theory in the CMDP framework, we have a control-theoretic update rule to learn about safety and can verify safety by evaluating the safety value of the current state. Another positive outcome of the formulation is that the original constrained problem is decomposed into two unconstrained optimization problems, making our formulation more amenable to gradient-based learning.
\parai{ii) Scales to high-dimensional visual context.} Compared to standard HJ Reachability methods, whose computational complexity scales exponentially with the state dimension, we updated the safety value directly on vision embedding, with neural approximation. This is the highest-dimensional problem studied studied via HJ reachability to-date.
\begin{algorithm}[t!]
\SetAlgoLined
\textbf{Initialize: } Performance actor $\pi$ and critic $Q$; \\
\textbf{Initialize: } Safety actor $\pi_S$ and critic $Q_S$; \\
\For{i = 0, $\dots$, \# Episodes}
{$x $ = env.reset()\\
\While{not terminal}{
{u $\sim \pi(x)$;\\
\emph{// The safe actor intervenes when the current state-action is deemed unsafe by the safety critic.}\\
\If{$Q_{S}(x, u)< \epsilon$}{
$u \sim \pi_{S}(x)$
}
$x', r $ = env.step($u$)\\
\emph{// Update performance actor-critic and safety actor-critic. See Appendix \ref{sec:algorithm} for details.}\\
}
}
}
\caption{SAGE-Environment Interaction}
\label{alg:spar}
\end{algorithm}
\begin{figure*}[t!]
\vspace{-1cm}
\sbox\twosubbox{%
\resizebox{\dimexpr\textwidth-1em}{!}{%
\includegraphics[height=3cm]{figures/DoubleIntegrator.pdf}
\includegraphics[height=3cm]{figures/DubinsCar.png}
\includegraphics[height=3cm]{figures/CompALL.pdf}
}
}
\setlength{\twosubht}{\ht\twosubbox}
\centering
\subcaptionbox{Double Integrator\label{fig:DoubleIntegrator}}{%
\includegraphics[height=0.9\twosubht]{figures/DoubleIntegrator.pdf}}
\subcaptionbox{Dubins' Car\label{fig:DubinsCar}}{%
\includegraphics[height=0.95\twosubht]{figures/DubinsCar.png}}
\subcaptionbox{Performance comparison of learning rules (averaged over 5 random seeds)\label{fig:toy_accuracy}}{%
\includegraphics[height=\twosubht]{figures/CompALL.pdf}
}
\caption{We use two classical control benchmarks, \textit{double integrator} and \textit{Dubins' car}, to evaluate the performance of different learning rules for safety analysis. (a) shows the safety value function of the double integrator and the black line delineates $V_S(x)=0$, within which the particle can remain within the allowable range of $x\in [-1, 1]$. (b) shows the iso-surface of the safety value function at 0, i.e., $V_S(x)=0$, for Dubins' Car, within which the car can reach a unit circle at the origin. The performance comparison is summarized in (c).
}
\vspace{-0.5cm}
\label{fig:classical_control_bench}
\end{figure*}
\para{Problem formulation.} We inject HJ reachability theory into the CMDP framework. Starting with Eqn. \ref{eq:cmdp}, we can interpret the negative of a cost as a reward for safety and, without loss of generality, reverse the direction of the inequality constraint. Recall that the super-zero level set of the safety value function, i.e., $\{x|V_s(x)\geq0\}$, designates all states from which the system can remain within the set of allowable states, $\mathcal{K}$, over infinite time horizon.
Thus, the safety value function derived from HJ Reachability can be plugged into CMDP (Eqn. \ref{eq:problem_formulation}): %
\begin{equation}
\label{eq:problem_formulation}
\pi^{*} = \arg\max_{\pi} \;\; V^{\pi}_R(x),
\quad \textrm{s.t.} \;\;\ V_{\text{\textit{S}}}(x) \geq \epsilon,
\end{equation}
where $\epsilon\geq0$ is introduced as a safety margin. A key difference from the original CMDP formulation (Eqn. \ref{eq:cmdp}) is that constraint satisfaction, $V_S(x)\geq \epsilon$, no longer depends on the policy, $\pi$. Thus, we can bypass the challenges of solving CMDPs (Section \ref{sec:related_work}) and decompose learning under safety constraints into optimizing for performance and updating safety value estimation. While a number of works have similar dual-policy architecture \cite{cheng2019end, bastani2021safe, thananjeyan2021recovery}, ours design is informed by HJ Reachability theory. Another difference is that HJ Reachability considers safety as absolute, and there is no mechanism to allow for some level of safety infraction, and thus $\chi_i$ in Eqn. \ref{eq:cmdp} is not longer present.
\para{Update of Safety Critic.} We apply HJ Bellman update, in place of standard Bellman backup, to learn the safety value function. The learning rule proposed by \cite{fisac2019bridging} is defined on discrete action space, which we modify for continuous action space (Eqn. \ref{eqn:safey_update}). While the safety actor is sub-optimal during learning, the resulting HJ Bellman target is an under-estimation of the safety value, as $Q_S(x', u')\leq \max_{u'\in \mathcal{U}} Q_S(x', u')$. Note that $Q_S(x, u)$ is updated model-free using state-action transitions, and only additionally requires $l(x)$. We assume $l(x)$ can be acquired from the vehicle's sensing capability \cite{bmw_2021} or estimated from perception \cite{chen2015deepdriving}.
\begin{equation}
\label{eqn:safey_update}
\begin{split}
Q_S(x, u) &= (1-\gamma)l(x) + \gamma\min \{l(x), Q_S(x', u')\},\\
u'&\sim \pi_S(x').
\end{split}
\end{equation}
\para{SAGE.} We propose \textsc{SAGE}, which consists of a performance policy and a safety policy. The safety backup controller is applied in a least restrictive way, only intervening when the RL agent is about to enter into an unsafe state, i.e., $u\sim \pi$, if $Q_S(x, u)\geq \epsilon$ and $u\sim \pi_S$ otherwise. The performance policy may be implemented with any RL algorithm. Since we expect the majority of samples to be from the performance policy, it is more appropriate to update the safety actor critic with an off-policy algorithm. In this work, we base our implementation of the safety actor critic on soft-actor critic (SAC) \parencite{haarnoja2018soft}. The safety critic is updated with Eqn. \ref{eqn:safey_update}, and the safety actor is updated via policy gradient through the safety critic, i.e., $\nabla_u Q_S(x, u)$. Algorithm \ref{alg:spar} provides an overview for \texttt{SAGE} and a detailed version is presented in Appendix \ref{sec:algorithm}.
\section{Experiments}
\label{sec:experiments}
We evaluate \texttt{SAGE} on three sets of benchmarks, of increasing difficulty. While the our intended application is autonomous racing, the first two set of benchmarks can be considered as some abstraction of vehicles with the objective of avoiding obstacles and/or moving towards goals. Firstly, we evaluate on two classical control tasks, where the safe vs. unsafe states are known analytically, and we compare the HJ Bellman update used in \texttt{SAGE} to alternatives for learning safety critics in the literature. Secondly, we compare \texttt{SAGE} to constrained RL baselines in Safety Gym. Finally, we challenge \texttt{SAGE} in \texttt{Learn-to-Race}~and conduct ablation to better understand how different components of \texttt{SAGE} contribute to its performance.
\subsection{Experiment: Classical Control Benchmarks}\label{sec:toy_exp}
\begin{figure*}[!]
\centering
\includegraphics[width=0.8\linewidth]{figures/gym_plots_full.pdf}
\caption{Performance of SAGE~with comparison to baselines in the CarGoal1-v0 (top row) and PointGoal1-v0 (bottom row) benchmarks (averaged over 5 random seeds). In Goal tasks, agents must navigate to observed goal locations (indicated by the green regions), while avoiding obstacles (e.g., vases in cyan, and hazards in blue).}
\label{fig:gym_results}
\end{figure*}
As mentioned earlier, safety critics have been trained in other works \cite{bharadhwaj2020conservative, srinivasan2020learning} with different learning rules. The objective here is to compare the HJ Bellman update with alternatives. Thus, we focus on safety analysis with off-policy samples, and
evaluate on two classical control benchmarks \textit{Double Integrator} \cite{fisac2019bridging} and \textit{Dubins' Car} \cite{bansal2017hamilton}, where the safe / lively\footnote{Liveness refers the ability to reach the specified goal \cite{hsu2021safety}.} states (Figure \ref{fig:DoubleIntegrator} and \ref{fig:DubinsCar}) and the optimal safety controller are known analytically. Double Integrator characterizes a particle moving on the x-axis, with velocity $v$. By controlling the acceleration, the objective is to keep the particle on a bounded range on x-axis. Dubins' Car is a simplified car model, where the car moves at a constant speed. By controlling the turning rate, the goal is to reach a unit circle regardless of the heading. More information on the two tasks are provided in Appendix \ref{sec:toy_model}.
In this experiment, we generate state-action pairs with a random policy, and evaluate the safety value function with respect to the optimal safety controller, $\pi^*_S$. In both \textit{Safety Q-functions for RL} (SQRL) \cite{srinivasan2020learning} and the \textit{Conservative Safety Critic} (CSC) \cite{bharadhwaj2020conservative}, the safety value function is defined as the expected cumulative cost, i.e., $Q^\pi_C(x, u) \coloneqq \mathbb{E}_{x_k, u_k \sim \pi} \left[\sum_{k=0}^\infty \gamma^k C(x_k)|x_0=x, u_0=u\right]$, where $C(x_k)=1$ if a failure occurs at $x_k$ and 0 otherwise. In this case, both the environment and optimal safety policy are deterministic. Thus, by definition, $Q^{\pi_S^*}_C(x, \pi_S^*(x))$ should be 0 if $x$ is a safe state. SQRL uses the standard Bellman backup to propagate the failure signal. On top of that, CSC uses conservative Q-learning (CQL) \cite{kumar2020conservative} to correct for difference between the behavior policy, i.e., the random policy, and the evaluation policy, i.e., the optimal safety policy, and overestimates $Q_C$ to err on the side of caution.
Since the safe vs. unsafe states are known for these benchmark tasks, we can directly compare the performance of these safety critics learned with different learning rules (Figure \ref{fig:toy_accuracy}). While the theoretical cut-off for safe vs. unsafe states is 0, the performance of SQRL is sensitive to the choice of the cut-off; thus, we report AUROC instead. For both CQL and SQRL, we do a grid search around the hyperparameters used in the original papers and report the best results. The implementation details and additional results are included in Appendix \ref{sec:toy_comp}. Directly applying Bellman update for safety analysis, as in SQRL, performs reasonably well on \textit{Double Integrator}, but not on the more challenging \textit{Dubins' Car}. In our experiment, CQL consistently under-performs SQRL. In comparison, HJ Bellman update has AUROC close to 1 on both tasks and has very small variance over different runs. It is worth-noting that the result with the HJ Bellman update is achieved without explicitly addressing the distribution mismatch \cite{voloshin2019empirical}, which typically challenges off-policy evaluation problems.
This experiment only compares the efficacy of the different learning rules for safety critic given the same off-policy samples, and does not intend to compare other aspects of SQRL and CSC.
One caveat is that SQRL and CQL uses a binary signal for failures, while HJ Bellman update has access to the distance, $l(x)$. On one hand, HJ Bellman update does assume more information. On the other hand, it may be more practical to learn safety from distance measurements then experiencing failures. Applied to autonomous driving, this translates to learning to avoid obstacle from distance measurements that are becoming prevalent on cars with assisted driving capabilities \cite{bmw_2021}, in comparison to experiencing collisions.
\begin{figure*}[t]
\sbox\twosubbox{%
\resizebox{\dimexpr\textwidth-1em}{!}{%
\includegraphics[height=3cm]{figures/racetrack_overhead_view.png}
\includegraphics[height=3cm]{figures/racetrack_speedtrap_narrow.png}
\includegraphics[height=3cm]{figures/racetrack_agent_egoview.png}
}
}
\setlength{\twosubht}{\ht\twosubbox}
\centering
\subcaptionbox{Aerial\label{fig:racetrack:aerial}}{%
\includegraphics[height=0.95\twosubht]{figures/racetrack_overhead_view.png}}
\subcaptionbox{Third-person\label{fig:racetrack:third_person}}{%
\includegraphics[height=0.95\twosubht]{figures/racetrack_speedtrap_narrow.png}}
\subcaptionbox{Ego-view\label{fig:racetrack:ego_view}}{%
\includegraphics[height=0.95\twosubht]{figures/racetrack_agent_egoview.png}
}
\caption{We use the \texttt{Learn-to-Race}~(\texttt{L2R}) framework \cite{herman2021learn} for evaluation; this environment provides simulated racing tracks that are modeled after real-world counterparts, such as the famed Thruxton Circuit in the UK (\texttt{Track01:Thruxton}, (a)). Here, learning-based agents can be trained and evaluated according to challenging metrics and realistic vehicle and environmental dynamics, making \texttt{L2R}~a compelling target for safe reinforcement learning. Each track features challenging components for autonomous agents, such as sharp turns (shown in (b)), where SAGE only uses ego-camera views (shown in (c)) and speed.
}
\vspace{-0.5cm}
\label{fig:racetrack}
\end{figure*}
\subsection{Experiment: \texttt{Safety Gym}} \label{exp:gym}
We additionally evaluate our proposed approach, SAGE, in Safety Gym \parencite{ray2019benchmarking}. Specifically, we evaluate on the standard CarGoal1-v0 and PointGoal1-v0 benchmarks, where the agent navigates to a goal while avoiding hazards. We compare SAGE~against baselines including: Constrained Policy optimization (CPO) \parencite{achiam2017constrained}, an unconstrained RL algorithm (Proximal Policy optimization (PPO) \parencite{schulman2017proximal}), and its Lagrangian variant (PPO-Lagrangian). By default, distance measurements from LiDAR are available to all baselines in these benchmarks, and thus SAGE~has direct access to $l(x)$. Episodic Performance and Cost curves are shown in Figure \ref{fig:gym_results} and implementation details are included in Appendix \ref{app:sec:sgym_details}.
PPO-SAGE has significantly fewer constraint violations, compared to other baselines, and the number of violations decreases over time.
While CPO and PPO-Lagrangian take into account that a certain number of violations are permissible, there is no such mechanism in SAGE, as HJ Reachability theory defines safety in an absolute sense. While the inability to allow for some level of safety infractions, unfortunately, compromises performance, SAGE learns mature obstacle-avoidance behaviors, compared to some policies, which may ignore traps in favor of fast navigation to goal locations. Violations that do occur in SAGE result from neural approximation error, and the number of violations decreases over time as the safety actor-critic gains experience, despite the randomized and constantly-changing episodic layouts.
\vspace{-0.2cm}
\subsection{Experiment: \texttt{Learn-to-Race}}
\para{Task Overview.} In this paper, we evaluate our approach using the Arrival Autonomous Racing Simulator, through the newly-introduced and OpenAI-gym compliant \texttt{Learn-to-Race}~(\texttt{L2R}) task and evaluation framework \parencite{herman2021learn}. \texttt{L2R}~provides multiple simulated racing tracks, modeled after real-world counterparts, such as Thruxton Circuit in the UK (\texttt{Track01:Thruxton}; see Figure \ref{fig:racetrack}).
\texttt{L2R}~can provide access to RGB images from any specified location, semantic segmentation, and vehicle states (e.g., pose, velocity). In each episode, an agent is spawned on the selected track. At each time-step, it uses its observations to determine normalized steering angle and acceleration. All learning-based agents receive the reward specified by \texttt{L2R}, which is formulated as a weighted sum of reward for driving fast and penalty for leaving the drivable area; the main objective is to complete laps in as little time as possible. Additional metrics are defined to evaluate driving quality.
\para{Implementation Details.} To characterize the performance of our approach, we report results on the Average Speed and the Episode Completion Percentage (\textit{ECP}) metrics \parencite{herman2021learn} as proxies for agent performance and safety, respectively. We report other metrics defined by \texttt{L2R}~in Appendix \ref{sec:add_results}.
We use \texttt{Track01:Thruxton} in \texttt{L2R}~(Fig. \ref{fig:racetrack}) for all stages of agent interaction with the environment. During training, the agent is spawned at random locations along the race track and uses a stochastic policy. During evaluation, the agent is spawned at a fixed location and uses a deterministic policy. The episode terminates when the agent successfully finishes a lap, leaves the drivable area, collides with obstacles, or does not progress for a number of time-steps. For each agent, we report averaged results across 5 random seeds, evaluated every 5000 steps over an episode (one lap). We use SAC as the performance policy, and all agents only have access to ego-camera view (Figure \ref{fig:racetrack:ego_view}) and speed, unless specified otherwise. The implementation, including network architecture and hyperparameters, are detailed in Appendix \ref{sec:implementation_details}.
\begin{figure*}[t]
\vspace{-0.0cm}
\centering
\includegraphics[width=0.9\linewidth]{figures/main_result.pdf}
\caption{\textbf{Left:} Episode percent completion and \textbf{Right:} speed evaluated every 5000 steps over an episode (a single lap) and averaged over 5 random seeds. Results reported based on \texttt{Track01:Thruxton} in \texttt{L2R}.}
\label{fig:main_results}
\end{figure*}
\para{Ablation Study.}
To demonstrate the benefit of utilizing domain knowledge in the form of a nominal model, we use a kinematic bike model \cite{kong2015kinematic} to calculate the safety value and derive the corresponding safety controller, detailed in Appendix in \ref{sec:nominal_ac}. We refer to this as the \textit{static actor-critic}. In all our experiments, only the static actor-critic has access to vehicle pose, specifically location and heading. We evaluate the performance of this static actor-critic by coupling a random agent with it (\texttt{SafeRandom}). We test \texttt{SafeRandom} on a series of safety margins to account for unmodelled dynamics; the performance averaged over 10 random seeds is summarized in Figure \ref{fig:SafeRandom}. For instance, $\epsilon\geq4.2$ achieves 80+\% ECP, in comparison to 0.5\% ECP by \texttt{Random} agent.
We examine the effect of having a safety controller, by comparing \texttt{SAC} with an instance of itself that is coupled with a static safety actor-critic (\texttt{SafeSAC}). We set the safety margin $\epsilon$ to be 4.2, based on empirical results from \texttt{SafeRandom}. We also compare the performance of using the static safety actor-critic (\texttt{SafeSAC}) and a dynamically-updating one (\texttt{SAGE}). Since the \texttt{SAGE} agent is expected to have a better characterization of the safety value, the agent no longer depends on a large safety margin to remain safe and thus \texttt{SAGE} uses a safety margin of 3.0m, which accounts for the vehicle dimensions\footnote{The HJ reachable tube is computed with respect to the back axle of the vehicle and does not account for the physical dimension of the vehicle. Using the car length as the safety margin is a rough engineering estimate.}. We also report results of \texttt{SafeSAC} with the same safety margin in in Appendix \ref{sec:add_results}
\begin{comment}
\begin{table*}[t]
\small
\captionsetup{margin=0.5cm}
\caption{\small\texttt{Learn-to-Race}~task \parencite{herman2021learn} results on \texttt{Track01} (Thruxton Circuit), with respect to the task metrics, Episode Completion Percentage (\textbf{ECP}) and Average Adjusted Track Speed (\textbf{AATS}). Arrows ($\uparrow\downarrow$) indicate directions of better performance.}
\label{tab:main_results}
\centering
\begin{tabular}{l ccc}
\toprule
\textbf{Agent} & \textbf{ECP} (\%) ($\uparrow$) & \textbf{AATS} (km/h) ($\uparrow$) \\
\midrule
\texttt{Random} & $0.5\pm 0.3$ & $11.9\pm 3.8$ \\ [1pt]
\texttt{MPC} & $100.0\pm 0.0$ & $45.1 \pm 0.0$ \\ [1pt]
\texttt{Human} & $100.0\pm 0.0$ & $171.2 \pm 3.4$ \\ [1pt]
\hline
\texttt{SAC} & $61.6\pm 38.6$ & $48.0\pm 30.9$ \\ [1pt]
\texttt{SafeRandom ($\delta$ = 4.2)} & $83.1 \pm 24.5$ & $11.2\pm 0.9$ \\ [1pt]
\texttt{SafeSAC ($\delta$ = 4.2)} & $49.1\pm 41.7$ & $33.8\pm 26.2$ \\ [1pt]
\texttt{SAGE ($\delta$ = 3.0)} & $\textbf{79.9}\pm23.2$ & \textbf{53.3}$\pm3.8$ \\ [1pt]
\bottomrule
\end{tabular}
\end{table*}
\end{comment}
\para{Results.}
\label{sec:result}
The performance comparison between different agents is summarized in Figure \ref{fig:main_results}.
In interpreting the results, note that a single lap in \texttt{Track01:Thruxton} is 3.8km, whereas CARLA, the \textit{de facto} environment for urban driving research, has in total 4.3km drivable roads in the original benchmark \cite{codevilla2019exploring}. Thus, successfully completing an episode, i.e., a lap, is quite challenging.
\parai{The static safety actor-critic significantly boosts initial safety performance.} With the help of the static safety actor-critic, the \texttt{SafeSAC} can complete close to 80\% of a lap, in comparison to slightly more than 5\% with \texttt{SAC}. This, again, showcases the benefit of injecting domain knowledge in the form of a nominal model. However, there are two notable limitations with the static safety controller. Firstly, it is extremely conservative, hard-braking whenever the vehicle is less safe. As a result, the \texttt{SafeSAC} agent has an initial speed of less than 10km/h. Secondly, as the \texttt{SAC} learns to avoid activating the safety controller and drive faster, the static safety controller is no longer able to recover the vehicle from marginally safe states. In fact, by applying the `optimal' safety action from Eqn. \ref{eq:safety_control}, i.e., maximum brake and steer towards centerline, the vehicle will lose traction and spin out of control. As a result, the ECP actually decreases over time for \texttt{SafeSAC}.
\parai{SAGE~learns safety directly from vision context and can recover from marginally safe states more smoothly.} Having a safety actor-critic that is dedicated to learning about safety significantly boosted the initial safety performance of \texttt{SAGE}, in comparison to the \texttt{SAC} agent, even though both the performance and the safety actor-critics are randomly initialized. Moreover, this shows that the safety value function can be learned from scratch on vision embedding. In practice, we envision the safety actor-critic to be warm-started with the nominal model or observational data, and fine-tuned by interactions with the environment. Furthermore, the learnable safety actor-critic can recover from marginally safe states smoothly, avoiding the two undesirable behaviors from the static actor-critic. A qualitative comparison of such behaviors is available at the \href{https://sites.google.com/view/safeautonomousracing/home}{\underline{anonymized paper website}}.
While \texttt{SAGE} outperforms other baselines, there is still significant performance gap with human, as the speed record at Thruxton Circuit is 237 km/h (average speed).
\section{Conclusion}
\label{sec:conclusion}
In this paper, we propose \texttt{SAGE} for end-to-end autonomous racing, which can learn to identify unsafe states from ego-camera views and recover from unsafe states, despite the complex dynamics with unstable regimes. We demonstrate on two classical control benchmarks that the HJ Bellman update is more effective than alternatives for learning the safety critic. Compared to constrained RL baselines in the Safety Gym, we show that \texttt{SAGE} has significantly fewer constraint violations. We report the new state-of-the-art result on \texttt{Learn-to-Race}, and we demonstrate that the safety value can be learned directly on visual context, thereby expanding HJ reachability to broader applications.
Throughout our experiments, we find it is highly effective to inject domain knowledge, in the form of nominal model or control-theoretic learning rule. In our experiments, the safety actor-critics were randomly initialized. But, in practice, we expect it to be pre-trained with a nominal model and/or observational data, prior to interacting with the environment. While neural approximation enables us to scale HJ reachability to high-dimensional visual inputs, we unfortunately lose the hard guarantees on safety. An important next-step is to characterize neural approximation error and find ways to retain the safety guarantees with function approximators.
\clearpage
\begin{comment}
\subsubsection*{Acknowledgments}
We thank the reviewers for their detailed reviews, and we thank Jaime Fisac and Mo Chen for useful comments and discussions. This work was supported, in part, by a doctoral research fellowship from Bosch Research and computing resources from Honda Research and CMU's Parallel Data Lab.
\end{comment}
{\small
\bibliographystyle{ieee_fullname}
\section{Classical Control Benchmarks}
The objective of this section is to compare the learning rule proposed by \cite{fisac2019bridging}, i.e., $Q(x, u) = (1-\gamma)l(x) + \gamma\min \{l(x), \max_{u'\in \mathcal{U}}Q(x', u'\}$
with alternatives for learning safety value function. We evaluate it on two classical control benchmarks, \textit{Double Integrator} and \textit{Dubins' Car}, as described in Section \ref{sec:toy_model}, where the analytical solution to safe states and optimal safe actions are known. Thus, we implement the learning rule here as Eqn. \ref{eq:lr_known_safety}. This is slightly different from the general case of Eqn. \ref{eqn:safey_update}, where the optimal safety policy is unknown.
\begin{multline}\label{eq:lr_known_safety}
Q(x, u) = (1-\gamma)l(x) + \gamma\min \{l(x), Q(x', u'\},\\
\text{where}\; u' = \pi^*_S.
\end{multline}
\subsection{Model Dynamics} \label{sec:toy_model}
\para{Double Integrator.} The double integrator models a particle moving along the x-axis at velocity $v$. The control input is the acceleration $a$. The goal in this case is keep the particle within a fixed boundary, in this case $x\in [-1, 1]$, subject to $a\in [-1, 1]$.
\begin{equation}
\begin{cases}
\dot{x}=v\\
\dot{v}=a
\end{cases}
\end{equation}
By solving the Hamiltonian, i.e., $\pi_S^*(x) = \arg \max_{u \in \mathcal{U}} \langle f(x, u), \nabla V_S(x) \rangle$, we can get the optimal safe control as:
\begin{equation}\label{eq:opt_DI}
a^*=
\begin{cases}
\overline{a} & \text{if}\;\; \partial V_S/\partial v \geq 0\\
\underline{a} & \text{otherwise}
\end{cases}
\end{equation}
\para{Dubins' Car.} The Dubins' car models a vehicle moving at constant speed, in this case $v=1$. Similar to the kinematic vehicle model, $x, y, \phi$ describes the position and heading of the vehicle, and control input is the turning rate $u\in[-1, 1]$. The goal is to reach a unit circle centred at the origin.
\begin{equation}
\begin{cases}
\dot{x}=v\cos(\phi)\\
\dot{y}=v\sin(\phi)\\
\dot{\phi}=u
\end{cases}
\end{equation}
Note that Dubins' Car is a reach task, i.e. reaching a specified goal, instead of an avoid task, i.e. avoiding specified obstacles. The reach task can be simply implemented by setting $\pi_S^*(x) = \arg \min_{u \in \mathcal{U}} \langle f(x, u), \nabla V_S(x) \rangle$ \citep{bansal2017hamilton}. In other words, the optimal safe action for a given state is the one that minimises the distance to the goal. The corresponding optimal safe control is
\begin{equation}
u^*=
\begin{cases}
\underline{u} & \text{if}\;\; \partial V_S/\partial \theta \geq 0\\
\overline{u} & \text{otherwise}
\end{cases}
\end{equation}
The ground truth safety value function for these two benchmarks are shown in Figure \ref{fig:DoubleIntegrator} and \ref{fig:DubinsCar}.
\para{Implementation \& Evaluation.} We use a neural network with hidden layers of size [16, 16] for the double integrator and [64, 64, 32] for Dubins' car. We use ADAM \cite{kingma2014adam} as the optimiser with a learning rate of 0.001, batch size of 64. We update the safety value function over 25K steps for Double Integrator and 50K steps for Dubin's Car, and report classification accuracy every 1000 steps averaged over 5 random seeds. While the safety value is defined over continuous state space, we evaluate the performance over a discrete mesh on the state space. By definition, the safety value at a given state $x$ is $Q(x, u^*)$, where $u^*=\pi^*_S(x)$.
\para{Qualitative Results.} Qualitative comparison between the ground truth value and that learned via HJ Bellman update is shown in Figure \ref{fig:V_est_DI} and \ref{fig:V_est_Dubins}. As we can see, the neural approximation largely recovers the ground truth value, except for minute difference.
\begin{figure}[h!]
\centering
\includegraphics[width=\linewidth]{figures/suppl/DI_V_hat.pdf}
\caption{A comparison between the ground truth safety value and that learned via HJ Bellman update for double integrator; The black line delineates $V_S(x)=0$.}
\label{fig:V_est_DI}
\end{figure}
\begin{figure}[h!]
\centering
\begin{subfigure}[b]{0.22\textwidth}
\centering
\includegraphics[clip, trim= 5cm 2cm 5cm 2cm, width=\linewidth]{figures/suppl/DB_view1.pdf}
\caption{}
\label{fig:DC}
\end{subfigure}
\quad
\begin{subfigure}[b]{0.22\textwidth}
\centering
\includegraphics[clip, trim= 5cm 2cm 5cm 2cm, width=\linewidth]{figures/suppl/DB_view2.pdf}
\caption{}
\label{fig:Dubins}
\end{subfigure}
\caption{A comparison between isosurface of the ground truth safety value (blue) and that learned via HJ Bellman update (green) for Dubins' car }
\label{fig:V_est_Dubins}
\end{figure}
While we do not need to learn the safety actor in this case, we further demonstrate that $\nabla_u Q_S(x, u)$ can indeed by used to update the safety actor. In Figure \ref{fig:nabla_Q}, we compare the ground truth $\partial{V}/\partial{v}$, with which one can determine the optimal safe action with Eqn. \ref{eq:opt_DI}, and the gradient through the safety critic, i.e., $\nabla_u Q_S(x, u)$. We can see that $\nabla_u Q_S(x, u)$, consistently point towards the correct optimal safe action within the safe set, i.e., the area delineated by the black line. The safety value outside the safe set is not learned, as the region is outside the support of data when episodes terminate upon failures.
\begin{figure}[h!]
\centering
\includegraphics[width=0.75\linewidth]{figures/suppl/nabla_Q.pdf}
\caption{The gradients through the safety critic, i.e., $\nabla_u Q_S(x, u)$, consistently point towards the correct optimal safe action, as indicated by $\partial{V}/\partial{v}$ (Eqn. \ref{eq:opt_DI}), within the safe set (the area delineated by the black line) for double integrator.}
\label{fig:nabla_Q}
\end{figure}
\begin{comment}
\subsection{Comparison of Implementation Details} \label{sec:convergence}
On a high level, we are interested in how common RL implementation techniques affect convergence. The specific questions we attempt to answer are:
\begin{itemize}
\item \textit{Is a slow-moving target network necessary for convergence?} It is common practice in RL algorithms to keep a copy of slow-moving target network $Q'$ for stability \citep{fujimoto2018addressing}, as the circular dependency between value estimate and policy results in accumulation of residual error and divergent updates. The target network is commonly updated with $Q' \xleftarrow{} \tau Q + (1-\tau)Q'$, where $\tau$ is a small number in (0, 1]. We examine how different value of $\tau$ affects convergence.
\item \textit{Is the Clipped Double Q-learning technique conducive to learning the safety value?} The Clipped Double Q-learning was popularised by TD3 \citep{fujimoto2018addressing} and also adopted in SAC \citep{pmlr-v80-haarnoja18b}. The technique addresses the overestimation of the value function by keeping two value networks, and computing the Bellman backup with the smaller estimate of the two. We examine if the same technique is conducive to learning the safety value.
\item \textit{Is the baseline technique helpful?} Deducting an action-independent baseline from the value function is commonly used for variance reduction in policy gradient algorithms. We expect that by choosing $l(x)$ as the baseline, the safety critic can focus on learning whether an action increases or decreases distance to obstacle / goal from the current state.
\end{itemize}
\parai{A slow-moving target network is not necessary for convergence.} We compare the performance at $\tau =[0.01, 0.1, 1]$. As we can see in Figure \ref{fig:by_tau}, the convergence were barely affected by the choice of $\tau$, and the safety value converged without problem using $\tau=1$, which is equivalent to not keeping a target network at all. We hypothesise that is because the actions come from the behaviour policy, and thus are independent of the safety critic. This removes the circular dependency between value estimation and policy and is consistent with the observation in \cite{fujimoto2018addressing}. In practice, we also expect the majority of the actions to be from the performance policy.
Since $\tau=0.1$ does slightly better on \textit{Dubins' Car}, we use $\tau=0.1$ for the result of the experiment.
\begin{figure}[h!]
\centering
\includegraphics[width = \linewidth]{figures/suppl/Toy_Accuracy.pdf}
\caption{Performance comparison under $\tau=0.01, 0.1, 1$}
\label{fig:by_tau}
\end{figure}
\parai{Clipped Double Q-learning is unnecessary.} In Figure \ref{fig:toy_double}, we provide a comparison on vanilla DQN vs. clipped double Q-learning, which uses the minimum of the two estimates from the Q-networks to calculate the target value. Aside from using the minimum of the two estimates as in the regular case, we also implement a version that uses the maximum of the estimates. In neither case, the clipped double Q-learning technique consistently improved performance. We hypothesise that the HJ Bellman backup already clips $Q'$ with $l(x)$, and thus the overestimation error is no longer a concern.
\begin{figure}[h!]
\centering
\includegraphics[width = \linewidth]{figures/suppl/Toy_AccuracyByDouble.pdf}
\caption{Performance comparison with clipped double Q-learning}
\label{fig:toy_double}
\end{figure}
\parai{Baseline technique does not have a significant impact on the accuracy of the safety critic.} We consider updating the safety critic with the baselined version of safety Q-value, i.e. $\Psi_S(x, u) = Q_S(x, u)-l(x)$. The corresponding update rule is given by Eqn. \ref{eqn:baselined-Q}. Contrary to our expectation, using the baseline technique doesn't have a significant impact on the accuracy of the safety critic (Figure \ref{fig:toy_baseline}).
\begin{equation}\label{eqn:baselined-Q}
\Psi_S(x, u) = \gamma\min \{0, \Psi_S(x', u') + l(x')-l(x)\}
\end{equation}
\begin{figure}[h!]
\centering
\includegraphics[width = \linewidth]{figures/suppl/Toy_AccuracyByBaseline.pdf}
\caption{Performance comparison with and without the baseline}
\label{fig:toy_baseline}
\end{figure}
To summarise our observations,
\begin{itemize}
\item The target network for safety critic may be updated more rapidly as we expect the majority of actions to come from the performance policy, removing the circular dependency between value estimation and policy.
\item The Clipped Double Q-learning is not conducive to learning the safety value function, as the HJ Bellman backup clips $Q'$ with $l(x)$, therefore overestimate error does not appear to be a concern.
\item Contrary to our expectation, deducting an action-independent baseline from the value function does not have a significant impact on learning the safety critic.
\end{itemize}
\end{comment}
\subsection{Learning Rule Comparison for Safety Critic }\label{sec:toy_comp}
Firstly, we describes the approaches pertaining to learning the safety critic in \cite{srinivasan2020learning, bharadhwaj2020conservative}.
In both Safety Q-functions for RL (SQRL) \citep{srinivasan2020learning} and Conservative Safety Critic (CSC) \citep{bharadhwaj2020conservative}, the safety critic is defined as the expected cumulative cost, i.e. $Q^\pi_C(x, u) \coloneqq \mathbb{E}_{x_k, u_k \sim \pi} \left[\sum_{k=0}^\infty \gamma^k C(x_k)|x_0=x, u_0=u\right]$, where $C(x_k)=1$ if a failure occurs at $x_k$ and 0 otherwise. Both papers endowed the safety critic, $Q^\pi_C(x, u)$ with a probabilistic interpretation, i.e. the expected probability of failure.
\para{SQRL.} The safety critic is trained by propagating the failure signal using the standard Bellman backup, as in Eqn. \ref{eq:SQRL}, where $\mathcal{D}$ denotes the replay memory, $\gamma_S$ is a time-discount parameter, and $\bar{Q}^\pi_C$ is the delayed target network. This approach for learning the safety critic is also adopted in \cite{thananjeyan2021recovery}.
\begin{equation}\label{eq:SQRL}
Q_C^\pi \leftarrow {E}_{x, u, x'\sim \mathcal{D}}\left[C(x) + \gamma_S(1-C(x)) \mathbb{E}_{u'\sim \pi} \bar{Q}^\pi_C(x', u')\right]
\end{equation}
\para{CSC.} On top of using Bellman backup to propagate the failure signals, CSC uses conservative Q-learning (CQL) \citep{kumar2020conservative} to correct for the distribution mismatch between the behaviour policy and the evaluation policy, and overestimate $Q_C$ to err on the side of caution. The resulting objective is given in Eqn. \ref{eq:CQL}, where $\mathcal{B}^\pi Q(x, u)=C(x)+\gamma\mathbb{E}_{x'\sim P(x'|x, u), u'\sim \pi(x')} \bar{Q}(x', u')$ is the Bellman operator and $\alpha$ is a hyperparamter that controls the extent of conservativeness. If $\alpha =0$, the objective is the same as that of SQRL.
\begin{equation}\label{eq:CQL}
\begin{split}
\mathcal{L}&= \frac{1}{2} \mathbb{E}_{s,a\sim \mathcal{D}}\left[Q_C(x, u) -\mathcal{B}^\pi \bar{Q}_C(x, u))\right] \\&-\alpha \left[ \mathbb{E}_{x\sim\mathcal{D}, u\sim \pi(x)}Q_C(x, u)-\mathbb{E}_{x, u\sim \mathcal{D}}Q_C(x, u)\right]
\end{split}
\end{equation}
Note that CSC reversed the sign in front of $\alpha$ compared to the original implementation in CQL so as to over-estimate $Q_C$. This learning objective does not guarantee point-wise conservativeness, but conservativeness in expectation, i.e. $\mathbb{E}_\pi \hat{Q}_c^\pi(x, u)\geq \mathbb{E}_\pi {Q}_c^\pi(x, u)$. $\tau$ is the rate for polyak averaging of the target network, i.e. $Q' \xleftarrow{} \tau Q + (1-\tau)Q'$.
\para{Implementation Details.} In \cite{srinivasan2020learning}, the authors used a learning rate of $3\times10^{-4}$, $\gamma_S=0.7$. Using the same learning rate, we did grid search over $\gamma_S=[0.7, 0.9]$ and $\tau =[0.1, 0.01]$. We observed that $\gamma_S=0.9$ had better performance, and thus selected $\gamma_S = 0.9$ and $\tau=0.1$.
In \cite{bharadhwaj2020conservative}, the authors used a learning rate of $2\times10^{-4}$, $\gamma_S=0.99$, and selected $\alpha = 0.5$ from from 0.05, 0.5, and 5. Using the same learning rate and $\gamma_S$, we did grid search over $\alpha=[0.01, 0.05, 0.5, 5]$ and $\tau =[0.1, 0.01]$. We selected $\alpha = 0.01$ and $\tau=0.1$.
\para{Results.} The main results are summarized in Figure \ref{fig:toy_accuracy}. In Figure \ref{fig:V_DI_comp}, we show a qualitative comparison between the ground truth safety value and that learned via different learning rules. In interpreting the results, note that both the environment and optimal safety policy are deterministic. Thus, $Q^{\pi_S^*}_C(x, \pi_S^*(x))$ should be 0 if $x$ is a safe state, following the definition. Due to the difference in definition, i.e., $Q_S(x)\geq 0$ is safe for HJ safety value and $Q_C(x)\leq 0$ is safe in SQRL and CQL, we plot $1-Q_C$ such that in Figure \ref{fig:V_DI_comp} the larger value consistently indicates safety and the cut-off for safe vs. unsafe is 0.
SQRL largely captures the correct safe states, though the classification performance is highly dependent on picking an appropriate threshold. CQL does underestimate the level of safety (and overestimate $Q_C$) as intended, and the pattern of underestimation appear to corresponds to $\nabla_x Q_S$ (refer to Figure \ref{fig:nabla_Q}).
\begin{figure}[h!]
\centering
\includegraphics[width=\linewidth]{figures/suppl/V_DI_comp.pdf}
\caption{Comparison between the group truth safety value and the safety critics from different learning rules for double integrator}
\label{fig:V_DI_comp}
\end{figure}
\begin{comment}
We also compare the false positive (FP) and false negative (FN) rate of the safety critic learned by HJ Bellman update and by CQL. With the goal of identifying unsafe states, CQL should have low FN rate by overestimating $Q_C$. As shown in Figure \ref{fig:toy_error}, CQL indeed has lower FN rate compared to HJ Bellman update, but at the expense of significantly larger FP rate. A secondary benefit of learning the safety critic with HJ Bellman update is that it adheres
\begin{figure}[h!]
\centering
\includegraphics[width = \linewidth]{figures/suppl/Toy_error.pdf}
\caption{A breakdown of false negative (FN) and false negative (FP) rate}
\label{fig:toy_error}
\end{figure}
\end{comment}
\newpage
\section{SAGE~Algorithm}\label{sec:algorithm}
SAGE~relies on a dual policy structure, the rationale of which is explained in Section \ref{sec:approach}. This pairing of a safety policy and a performance policy is important, as we are able to decompose the problem of learning under safety constraints into optimizing for performance and updating the safety value function, separately.
We optimize the performance policy using SAC, but it may be switched for any other comparable RL algorithms. The safety policy is used least-restrictively, that is only intervene when the RL agent is about to enter into an unsafe state and thus allowing the performance policy maximum freedom in exploring safely. Instead of using the optimal safe policy from solving Hamiltonian, the safe policy is updated via gradients through the safety critic, same as other actor-critic algorithms.
\begin{algorithm}[h!]
\SetAlgoLined
\textbf{Initialize: } performance critic $Q_{\phi}$ and actor $\pi_\theta$; \\
\textbf{Initialize: } safety critic $Q_{\phi_S}$, and actor $\pi_{\theta_S}$; target networks $\phi'_S\leftarrow \phi_S$;\\
\textbf{Initialize: } replay buffer $\mathcal{D}$;\\
\For{i = 0, $\dots$, \# Episodes}{
$x $ = env.reset()\\
\While{not terminal}{
{u $\sim \pi_{\theta}(x)$;\\
\emph{// The safe actor intervenes when the current state-action is deemed unsafe by the safety critic.}\\
\If{$Q_{\phi_S}(x, u)< \epsilon$}{
$u \sim \pi_{\theta_S}(x)$
}
$x', r $ = env.step($u$)\\
$\mathcal{D}$.store($x, a, x', r$)\\
$x=x'$\\
{Update performance critic $Q_{\phi}$ and actor $\pi_\theta$ with preferred RL algorithm;\\
\emph{// Update the safety critic:}\\
Sample N transitions $(x, u, x')$ from $\mathcal{D}$;\\
\emph{// Calculate the target value with the discounted Bellman safety update }
$$y = (1-\gamma)l(x)+ \gamma \min\{l(x), Q_{\phi'_S}(x', u')\},$$ $$\text{where}\;u'\sim \pi_{\theta_S}$$
$$\mathcal{L}_{\phi_S} = N^{-1}\sum (Q_{\phi_S}(x, u)-y)^2$$
$$\phi_S \leftarrow\phi_S- \alpha\nabla_{\phi_S} \mathcal{L}_{\phi_S}$$
\emph{// Update the safety actor with deterministic policy gradient:}
$$\theta_S \leftarrow\theta_S + \alpha N^{-1}\sum \nabla_u Q(x, u) \nabla_{\theta_S} \pi_{\theta_S}(x)$$
\emph{// Update the target networks:}
$$\phi'_S\leftarrow \tau\phi_S+(1-\tau)\phi'_S $$
}
}
}}
\caption{\texttt{SAGE}: Safe Autonomous Racing via Approximate Reachability on Ego-vision}
\label{suppl:alg:spar}
\end{algorithm}
\section{Details: Safety Gym Experiment}
\label{app:sec:sgym_details}
Following the default CarGoal1-v0 and PointGoal1-v0 benchmarks in Safety Gym, all agents were given LiDARs observation with respect to hazard, goal, and vase, with avoiding hazards as the safety constraints. Both environments were initialised with a total of 8 hazards and 1 vase. Agent's are endowed with accelerometer, velocimeter, gyro, and magnetometer sensors; their LiDAR configurations included 16 bins, with max distance of 3.
The baselines we considered, i.e., CPO, PPO and PPO-Lagrangian follows the default implementation that comes with Safety Gym. PPO-SAGE~wraps the proposed safety actor critic around the PPO base agent. Despite PPO being an on-policy algorithm, the SAGE~safety critic was implemented with off-policy updates, using prioritised memory replay based on the TD-error of predicting safety value. Since $l(x)$ is small in this environment, we scaled cost by a factor of 100. For the safety actor-critic, We used $\gamma_S$ annealing from 0.85 to 1 following \cite{fisac2019bridging}, $\tau = 0.005$, critic learning rate of 0.001, actor learning rate of 0.0003, and $\alpha =0.2$ (regularisation on policy entropy). We used a safety margin $\epsilon = 0.25$, mainly to account for the dimension of the hazards (radius = 0.2).
For each model, on each Safety Gym benchmark, results were reported as the average across 5 instances. All experiments in Safety Gym were run on an Intel(R) Core(TM) i9-9920X CPU @ 3.50GHz -- with 1 CPU, 12 physical cores per CPU, and a total of 24 logical CPU units.
\section{Static Safety Actor-Critic Derived from Kinematic Bike Model}\label{sec:nominal_ac}
\begin{figure*}[h!]
\vspace{-0.0cm}
\centering
\begin{subfigure}[b]{0.25\textwidth}
\centering
\includegraphics[width=0.8\textwidth]{figures/bike_model.pdf}
\caption{nominal model}
\label{fig:bike_model}
\end{subfigure}
\begin{subfigure}[b]{0.69\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/HJ_example.pdf}
\caption{$V_s(x, y, v, \phi)$ computed via the nominal model, where v=12m/s}
\label{fig:HJ_contour}
\end{subfigure}
\caption{(a) We compute the safety value function, via a kinematic vehicle model. (b) We illustrate different views of the 4D state space, given fixed velocity and three different yaw angles, indicated by the blue arrows.}
\label{fig:HJ_warmstart}
\end{figure*}
To demonstrate the benefit of utilizing domain knowledge in the form of a nominal model and to compare with the learnable safety actor-critic in \texttt{SAGE}, we use the kinematic vehicle model \cite{kong2015kinematic} (see Figure \ref{fig:bike_model}), which is a significant simplification of a realistic race car model \cite{kabzan2019learning}, to compute the safety value and corresponding `optimal'\footnote{only with respect to the nominal model} safety controller. The dynamics and `optimal' safety control is given in Eqn. \ref{eq:bike_model} and \ref{eq:safety_control}, where the state is $\mathbf{x}=[x, y, v, \phi]$, and the action is $\mathbf{u}=[a, \delta]$. $x, y, v, \phi$ are the vehicle's location, speed, and yaw angle. $a$ is the acceleration, and $\delta$ is the steering angle. The actions are bounded, i.e., $a\in [\underline{a}, \overline{a}]$ and $\delta \in [\underline{\delta}, \overline{\delta}]$. $L=3$m is the car length.
\begin{equation}\label{eq:bike_model}
f(\mathbf{x}, \mathbf{u}) = \begin{cases}
\dot{x}=v\cos(\phi)\\
\dot{y}=v\sin(\phi)\\
\dot{v}=a\\
\dot{\phi}=v\tan{\delta}/L
\end{cases}
\end{equation}
\begin{equation}\label{eq:safety_control}
\begin{split}
a^*&=
\begin{cases}
\underline{a} & \text{if}\;\; {\partial V_S}/{\partial v}\leq 0\\
\overline{a} & \text{else}
\end{cases}, \\
\delta^*&=
\begin{cases}
\overline{\delta} & \text{if}\;\; {\partial V_S}/{\partial \phi} \geq 0\\
\underline{\delta} & \text{else}
\end{cases}
\end{split}
\end{equation}
Intuitively, the `optimal' safety policy brakes and steers towards the center of the track as much as possible. We also derive the `optimal' safety policy here. The optimal safety control is derived by solving the Hamiltonian as given in Eqn. \ref{eq:hamiltonian}. By definition, $\nabla V_S(\mathbf{x}) = [{\partial V_S}/{\partial x }, {\partial V_S}/{\partial y}, {\partial V_S}/{\partial v}, {\partial V_S}/{\partial \phi}].$
\begin{subequations}\label{eq:derivation}
\begin{align}
\pi_S^*(\mathbf{x}) &= \arg \max_{\mathbf{u} \in \mathcal{U}} \langle f(\mathbf{x}, \mathbf{u}), \nabla V_S(\mathbf{x}) \rangle \label{eq:hamiltonian}\\
&= \arg \max_{[a, \delta] \in \mathcal{U}} [v\cos(\phi)\frac{\partial V_S}{\partial x } + v\sin(\phi) \frac{\partial V_S}{\partial y } \nonumber\\
&+ a \frac{\partial V_S}{\partial v} + v\tan{\delta}/L \frac{\partial V_S}{\partial \phi}]\\
&= \arg \max_{[a, \delta] \in \mathcal{U}} [a \frac{\partial V_S}{\partial v}+ v\tan{\delta}/L \frac{\partial V_S}{\partial \phi}]\label{eq:derv_end}
\end{align}
\end{subequations}
From Eqn \ref{eq:derv_end}, it is clear that the actions given by Eqn. \ref{eq:safety_control} maximize the Hamiltonian.
We parametrized the racetrack as a cubic spline and computed $l(\mathbf{x})$ by projection onto the spline.
Setting $V_S(\mathbf{x}, 0) = l(\mathbf{x})$, we calculated the backward reachable tube using the code from \cite{optimized_dp}. For efficient computation, we divided the racetrack into overlapping segments and computed the safety value segment-wise.
Fig. \ref{fig:HJ_contour} illustrates resulting safety value function at slices of state space, as the agent enters into a sharp turn. It is clear that the safety value at each location can be quite different from the initialization, $l(\mathbf{x})$.
\section{Details: Learn-to-Race Experiment}\label{sec:implementation_details}
\para{Vision Encoder.} We condition the optimization of the performance policy as well as the safety value updates on pretrained embedding of vehicle's visual scene context. The perception module maps ego-images from the on-board RGB camera to feature embedding of reduced dimension. To learn this mapping, we use a standard variational autoencoding (VAE) \cite{kingma2013auto} paradigm, with a convolutional encoder.
We use an image reconstruction objective with binary cross-entropy loss, Adam optimizer \cite{kingma2014adam}, and a latent vector dimension of 32. We train the VAE encoder to reconstruct ego-images, sampled from the vehicle's front camera during random agent execution; examples are provided in Figure \ref{fig:img_reconstruction}. We further refine the encoder by training the VAE module to reconstruct projected road boundaries, illustrated in Figure \ref{fig:road_boundaries}, with inputs in the left column and the reconstructed outputs in the right column.
\begin{figure}[h!]
\centering
\includegraphics[width=\linewidth]{figures/suppl/real-image-reconstruction.pdf}
\caption{VAE image reconstruction, with real images in the left column and reconstructed images in the right column.}
\label{fig:img_reconstruction}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=\linewidth]{figures/suppl/projected-boundary-reconstruction.pdf}
\caption{VAE reconstruction of projected road boundary images, with real images in the left column and reconstructed images in the right column.}
\label{fig:road_boundaries}
\end{figure}
\para{Neural Architecture.} As illustrated in Figure \ref{fig:enc_arch}, the vision encoder takes an image as input to produce a latent vector, which is concatenated with speed and action embedding and passed to the performance and safety actor-critics. The specific implementation of layers are summarized in Table \ref{suppl:tab:hyperparameters}.
\begin{figure}[h!]
\centering
\includegraphics[width=\linewidth]{figures/suppl/encoder_arch.pdf}
\caption{SAGE neural architecture overview.}
\label{fig:enc_arch}
\end{figure}
Specifically, we use a squashed Gaussian policy (Eqn. \ref{eq:squashed_gaussian}) for both performance and safety actors, following \cite{haarnoja2018soft}.
\begin{equation}\label{eq:squashed_gaussian}
u = \tanh(\mu(x)+\sigma(x)\odot \xi),\quad \xi\sim \mathcal{N}(0, \mathbf{I})
\end{equation}
\para{Agent training details.} During training, the agent is spawned at random locations along the race track and uses a stochastic policy. During evaluation, the agent is spawned at a fixed location and uses a deterministic policy. The episode terminates when the agent successfully finishes a lap, leaves the drivable area, collides with obstacles, or does not progress for a number of steps. For each agent, we report averaged results across 5 random seeds, evaluated every 5000 steps over an episode (one lap). In total, we train each agent over 250,000 steps, and evaluate it over 50 episodes.
During its interaction with the environment, the agent receives a $192 \times 144$ ego-camera view and its speed at each time-step. The agent encodes the RGB image frame and its speed to a 40-dimensional feature representation, subsequently used as input to both actor-critic networks. We initialise the replay buffer with 2000 random transitions, following \cite{SpinningUp2018}. After 2000 steps, we perform a policy update at each time step. For the SafeSAC agent, we only save state-action transitions from the performance actor to the replay buffer. For the SAGE~agent, we save all state-action transitions.
\begin{table*}[h!]
\caption{Network Architecture}
\label{suppl:tab:hyperparameters}
\begin{center}
\scriptsize
\resizebox{0.9\textwidth}{!}{
\begin{tabular}{l c l c l c l}
\toprule
Operation & & Input (dim.) & & Output (dim.) & & Parameters\\
\midrule
\multicolumn{7}{c}{\textsc{Visual Encoder}} \\
\hline
\textit{Conv2d} & & $(N, \text{chan}, 42, 144)$, chan : 3$\rightarrow$32 & & \texttt{conv1} & & k$\coloneqq$(4,4), s$\coloneqq$2, p$\coloneqq$1, activation$\coloneqq$ReLU \\
\textit{Conv2d} & & \texttt{conv1}, chan : 32$\rightarrow$64 & & \texttt{conv2} & & k$\coloneqq$(4,4), s$\coloneqq$2, p$\coloneqq$1, activation$\coloneqq$ReLU \\
\textit{Conv2d} & & \texttt{conv2}, chan : 64$\rightarrow$128 & & \texttt{conv3} & & k$\coloneqq$(4,4), s$\coloneqq$2, p$\coloneqq$1, activation$\coloneqq$ReLU \\
\textit{Conv2d} & & \texttt{conv3}, chan : 128$\rightarrow$256 & & \texttt{conv4} & & k$\coloneqq$(4,4), s$\coloneqq$2, p$\coloneqq$1, activation$\coloneqq$ReLU \\
\textit{Flatten} & & --- & & --- & & --- \\
\midrule
\multicolumn{7}{c}{\textsc{Visual Encoder Bottleneck Representation}} \\
\textit{Linear} (mu) & & $N \times \text{h\_dim}$ & & $N \times 32$ & & --- \\
\textit{Linear} (sigma) & & $N \times \text{h\_dim}$ & & $N \times 32$ & & --- \\
\midrule
\multicolumn{7}{c}{\textsc{Visual Decoder} (only for pre-training Visual Encoder)} \\
\hline
\textit{Unflatten} & & --- & & --- & & --- \\
\textit{ConvTranspose2d} & & \texttt{encoder.conv4}: encoder.conv4.chan: 256 $\rightarrow$128 & & \texttt{convtranspose1} & & k$\coloneqq$(4,4), s$\coloneqq$2, p$\coloneqq$1, activation$\coloneqq$ReLU \\
\textit{ConvTranspose2d} & & \texttt{convtranspose1}, chan : 128 $\rightarrow$64 & & \texttt{convtranspose2} & & k$\coloneqq$(4,4), s$\coloneqq$2, p$\coloneqq$1, activation$\coloneqq$ReLU \\
\textit{ConvTranspose2d} & & \texttt{convtranspose2}, chan : 64 $\rightarrow$32 & & \texttt{convtranspose3} & & k$\coloneqq$(4,4), s$\coloneqq$2, p$\coloneqq$1, activation$\coloneqq$ReLU \\
\textit{ConvTranspose2d} & & \texttt{convtranspose3}, chan : 32 $\rightarrow$3 & & \texttt{convtranspose4} & & k$\coloneqq$(4,4), s$\coloneqq$2, p$\coloneqq$1, activation$\coloneqq$Sigmoid \\
\midrule
\multicolumn{7}{c}{\textsc{Safety Actor-Critic}} \\
\hline
\texttt{actor\_network} & & --- & & --- & & --- \\
\texttt{q\_function1} & & --- & & --- & & --- \\
\texttt{q\_function2} & & --- & & --- & & --- \\
\midrule
\multicolumn{7}{c}{\textsc{Performance Actor-Critic}} \\
\hline
\texttt{actor\_network} & & --- & & --- & & --- \\
\texttt{q\_function1} & & --- & & --- & & --- \\
\texttt{q\_function2} & & --- & & --- & & --- \\
\midrule
\multicolumn{7}{c}{\textsc{Actor Network (Policy): SquashedGaussianMLPActor}} \\
\hline
\textit{Linear} & & $N \times 32$ & & $N \times 64$ & & activation$\coloneqq$ReLU \\
\textit{Linear} & & $N \times 64$ & & $N \times 64$ & & activation$\coloneqq$ReLU \\
\textit{Linear} & & $N \times 64$ & & $N \times 32$ & & activation$\coloneqq$ReLU \\
\textit{Linear} (projection: mu\_layer) & & $N \times 32$ & & $N \times 3$ & & --- \\
\textit{Linear} (projection: log\_std\_layer) & & $N \times 32$ & & $N \times 3$ & & --- \\
\midrule
\multicolumn{7}{c}{\textsc{Q function}} \\
\hline
\texttt{speed\_encoder} & & --- & & --- & & --- \\
\texttt{regressor} & & --- & & --- & & --- \\
\midrule
\multicolumn{7}{c}{\textsc{Speed Encoder}} \\
\hline
\textit{Linear} & & $N \times 1$ & & $N \times 8$ & & activation$\coloneqq$ReLU \\
\textit{Linear} & & $N \times 8$ & & $N \times 8$ & & activation$\coloneqq$Identity \\
\midrule
\multicolumn{7}{c}{\textsc{Regressor}} \\
\hline
\textit{Linear} & & $N \times 42$ & & $N \times 32$ & & activation$\coloneqq$ReLU \\
\textit{Linear} & & $N \times 32$ & & $N \times 64$ & & activation$\coloneqq$ReLU \\
\textit{Linear} & & $N \times 64$ & & $N \times 64$ & & activation$\coloneqq$ReLU \\
\textit{Linear} & & $N \times 64$ & & $N \times 32$ & & activation$\coloneqq$ReLU \\
\textit{Linear} & & $N \times 32$ & & $N \times 32$ & & activation$\coloneqq$ReLU \\
\textit{Linear} & & $N \times 32$ & & $N \times 1$ & & activation$\coloneqq$Identity \\
\bottomrule
\end{tabular}
}
\end{center}
\label{tab:parameters}
\end{table*}
\begin{table*}[h!]
\small
\captionsetup{margin=0.3cm}
\caption{\small\texttt{Learn-to-Race}~task \parencite{herman2021learn} results on \texttt{Track01} (Thruxton Circuit), for learning-\textit{free} agents, with respect to the task metrics: Episode Completion Percentage (\textbf{ECP}), Episode Duration (\textbf{ED}), Average Adjusted Track Speed (\textbf{AATS}), Average Displacement Error (\textbf{ADE}), Trajectory Admissibility (\textbf{TrA}), Trajectory Efficiency (\textbf{TrE}), and Movement Smoothness (\textbf{MS}). Arrows ($\uparrow\downarrow$) indicate directions of better performance, across agents. \textbf{Bold} results in tables \ref{tab:base_ltr_task_lf} and \ref{tab:base_ltr_task_lb} are generally best, however, asterisks (*) indicate metrics which may be misleading, for incomplete racing episodes.}
\label{tab:base_ltr_task_lf}
\resizebox{\textwidth}{!}{
\centering
\begin{tabular}{lccccccc}
\toprule
\textbf{Agent} & \textbf{ECP} ($\uparrow$) & \textbf{ED}* ($\downarrow$) & \textbf{AATS} ($\uparrow$) & \textbf{ADE} ($\downarrow$) & \textbf{TrA} ($\uparrow$) & \textbf{TrE} ($\uparrow$) & \textbf{MS} ($\uparrow$) \\
\midrule
\texttt{HUMAN} & $\textbf{100.0}\pm 0.0$ & $78.6\pm 5.2$ & $\textbf{79.29}\pm 4.7$ & $2.4\pm 0.1$ & $0.93\pm 0.01$ & $\textbf{1.00}\pm 0.02$ & $\textbf{11.7}\pm 0.1$ \\ [1pt]
\hline
\texttt{Random} & $0.50\pm 0.30$ & $\textbf{4.67}\pm 3.2$ & $11.90\pm 3.80$ & $1.5\pm 0.60$ & $0.81\pm 0.04$ & $0.33\pm 0.38^{*}$
& $6.7\pm 1.1$ \\ [1pt]
\texttt{MPC} & $100.0\pm 0.0$ & $301.40\pm 10.10$ & $45.10\pm 0.0$ & $\textbf{0.90}\pm 0.10$ & $\textbf{0.98}\pm 0.01$ & $0.85\pm 0.03$ & $10.4\pm 0.60$ \\ [1pt]
\bottomrule
\end{tabular}
}
\end{table*}
\begin{table*}[h]
\small
\captionsetup{margin=0.3cm}
\caption{\small\texttt{Learn-to-Race}~task \parencite{herman2021learn} results on \texttt{Track01} (Thruxton Circuit), for learning-\textit{based} agents.}
\label{tab:base_ltr_task_lb}
\resizebox{\textwidth}{!}{
\centering
\begin{tabular}{lccccccc}
\toprule
\textbf{Agent} & \textbf{ECP} ($\uparrow$) & \textbf{ED}* ($\downarrow$) & \textbf{AATS} ($\uparrow$) & \textbf{ADE} ($\downarrow$) & \textbf{TrA} ($\uparrow$) & \textbf{TrE} ($\uparrow$) & \textbf{MS} ($\uparrow$) \\
\midrule
\texttt{SAC} & $61.61\pm 38.57$ & $272.75\pm256.51$ & $47.99\pm 30.9$ & $1.54\pm1.07$ & $\textbf{0.94}\pm0.02$ & $\textbf{0.28}\pm0.12$ & $11.84\pm2.12$ \\ [1pt]
\texttt{SafeRandom} (ours), $\delta=3.0$ & $36.46\pm23.71$ & $654.37\pm447.05$ & $8.44\pm 1.37$ & $3.93\pm0.21$ & $0.81\pm0.10$ & $0.00\pm0.00$ & $13.21\pm1.88$ \\ [1pt]
\texttt{SafeRandom} (ours), $\delta=4.2$ & $63.63\pm39.46$ & $761.80\pm494.65$ & $11.68\pm 1.07$ & $2.74\pm0.16$ & $0.90\pm0.07$ & $0.02\pm0.01$ & $\textbf{13.63}\pm2.01$ \\ [1pt]
\texttt{SafeSAC} (ours), $\delta=3.0$ & $25.70\pm11.31$ & $66.90\pm23.22$ & $49.67\pm3.34$ & $1.35\pm0.05$ & $0.86\pm0.06$ & $0.14\pm0.05$ & $8.46\pm2.35$ \\ [1pt]
\texttt{SafeSAC} (ours), $\delta=4.2$ & $49.05\pm41.66$ & $617.52\pm842.49$ & $33.83\pm26.21$ & $1.80\pm0.63$ & $0.91\pm0.12$ & $0.07\pm0.11$ & $10.03\pm2.75$ \\ [1pt]
\texttt{SAGE} (ours) & $\textbf{79.94}\pm23.20$ & \textbf{59.19}$\pm29.99$ &\textbf{53.28}$\pm3.76$ & \textbf{0.99}$\pm0.17$ & $0.91\pm0.03$ & $0.22\pm0.03$ & $9.27\pm1.68$ \\ [1pt]
\bottomrule
\end{tabular}
}
\end{table*}
\para{Implementation Details.} For all experiments, we implemented the models using the \texttt{PyTorch} 1.8.0. We optimised both the performance and safety actor-critic with Adam \cite{kingma2014adam}, with a learning rate of 0.003. We used $\gamma=0.99$ for the performance critic, and annealed $\gamma_S$ from 0.85 to 1 for the safety critic following \cite{fisac2019bridging}. We used $\tau=0.005$ for the performance critic, and $\tau=0.05$ for the safety critic. For both the performance and safety actor, we include the policy entropy term with $\alpha=0.2$. We used a batch size of 256, and a replay buffer size of 250,000.
\para{Computing hardware.} For rendering the simulator and performing local agent verification and analysis, we used a single GPU machine, with the following CPU specifications: Intel(R) Core(TM) i5-4690K CPU @ 3.50GHz; 1 CPU, 4 physical cores per CPU, total of 4 logical CPU units. The machine includes a single GeoForce GTX TITAN X GPU, with 12.2GB GPU memory. For generating multi-instance experimental results, we used a cluster of three multi-GPU machines with the following CPU specifications: 2x Intel(R) Xeon(R) Gold 5218R CPU @ 2.10GHz; 80 total CPU cores using a Cascade Lake architecture; memory of 512 GiB DDR4 3200 MHz, 16x32 GiB DIMMs. Each machine includes 8x NVIDIA GeForce RTX 2080 Ti GPUs, each with 11GB GDDR6 of GPU memory. Experiments were orchestrated on the these machines using Kubernetes, an open-source container deployment and management system.
All experiments were conducted using version 0.7.0.182276 of the Arrival Racing Simulator. The simulator and \texttt{Learn-to-Race}~framework \parencite{herman2021learn} are available for academic-use, here: \url{https://learn-to-race.org}.
\section{Additional Results}\label{sec:add_results}
\para{Performance of the SafeRandom agent.} Recall that the \texttt{SafeRandom} agent takes random actions and uses the safety value function precomputed from the nominal model. The optimal safety controller intervene whenever the safety value of the current state falls belong the safety margin. The safety margin is necessary because 1) the nominal model is a significant over-simplification of vehicle dynamics, and 2) the HJ Reachability computation does not take into consideration of the physical dimension of the vehicle.
The performance of the \texttt{SafeRandom} agent at different safety margin is summarised in Figure \ref{fig:SafeRandom}. For safety margin $\epsilon\geq 4.2$, the \texttt{SafeRandom} agent can finish 80+\% of the lap, and thus we use $\epsilon=4.2$ as the safety margin for the \texttt{SafeSAC} agent. On the other hand, the performance decrease drastically when the safety margin is reduced to 3.
\begin{figure}[h!]
\centering
\includegraphics[width = \linewidth ]{figures/suppl/SafeRandom.pdf}
\caption{Performance of the \texttt{SafeRandom} agent at different safety margin (averaged over 10 random seeds)}
\label{fig:SafeRandom}
\end{figure}
\para{SafeSAC \& SAGE~performance with same safe margin.} While we choose the safety margin $\epsilon$ based on performance of the \texttt{SafeRandom} agent over a range of margins and our best engineering judgement, some may wonder if the superior performance of SAGE~over SafeSAC may be attributed to the use of different safety margins. Thus, we also show here the performance of a SafeSAC agent with the same safety margin as SAGE~in Figure \ref{fig:SafeSAC-3}. Given the smaller safety margin, the ECP is low initially, which is inline with the observation from \texttt{SafeRandom}. Furthermore, the ECP barely improves over time. As the performance agent learns to drive faster, it is increasingly difficulty for the static actor-critic to catch the vehicle in marginally safe states.
\begin{figure}[h!]
\centering
\includegraphics[width = \linewidth ]{figures/suppl/SafeSAC-3.pdf}
\caption{Performance of \texttt{SafeSAC} ($\epsilon=3$) with comparison to \texttt{SAGE} }
\label{fig:SafeSAC-3}
\end{figure}
\para{Learn-to-Race benchmark results.} In tables \ref{tab:base_ltr_task_lf} and \ref{tab:base_ltr_task_lb}, we follow \parencite{herman2021learn} in reporting on all of their driving quality metrics, for the \texttt{Learn-to-Race}~benchmark: Episode Completion Percentage (ECP), Episode Duration (ED), Average Adjusted Track Speed (AATS), Average Displacement Error (ADE), Trajectory Admissibility (TrA), Trajectory Efficiency (TrE), and Movement Smoothness (MS).
We highlight the fact that such metrics as TrA, TrE, and MS are most meaningful for agents that \textit{also} have high ECP results. Taking TrA, for example, safe policies score higher ECP values but may spend more time in inadmissible positions (as defined by the task, i.e., with at least one wheel touching the edge of the drivable area), compared to policies without a safety backup controller that may quickly terminate episodes by driving out-of-bounds (thus spending less time in the inadmissible positions). On the other hand, policies that have low completion percentages also have low ED scores, due to more frequent failures and subsequent environment resets.
We observe new state-of-the-art performance received by our approach, across the driving quality metrics, in the \texttt{Learn-to-Race}~benchmark.
\begin{comment}
\begin{table*}[t]
\small
\captionsetup{margin=0.5cm}
\caption{\small\texttt{Learn-to-Race}~task \parencite{herman2021learn} results on \texttt{Track01} (Thruxton Circuit), with respect to the task metrics, Episode Completion Percentage (\textbf{ECP}) and Average Adjusted Track Speed (\textbf{AATS}). Arrows ($\uparrow\downarrow$) indicate directions of better performance.}
\label{tab:main_results}
\centering
\begin{tabular}{l ccc}
\toprule
\textbf{Agent} & \textbf{ECP} (\%) ($\uparrow$) & \textbf{AATS} (km/h) ($\uparrow$) \\
\midrule
\texttt{Random} & $0.5\pm 0.3$ & $11.9\pm 3.8$ \\ [1pt]
\texttt{MPC} & $100.0\pm 0.0$ & $45.1 \pm 0.0$ \\ [1pt]
\texttt{Human} & $100.0\pm 0.0$ & $171.2 \pm 3.4$ \\ [1pt]
\hline
\texttt{SAC} & $61.6\pm 38.6$ & $48.0\pm 30.9$ \\ [1pt]
\texttt{SafeRandom ($\delta$ = 4.2)} & $83.1 \pm 24.5$ & $11.2\pm 0.9$ \\ [1pt]
\texttt{SafeSAC ($\delta$ = 4.2)} & $49.1\pm 41.7$ & $33.8\pm 26.2$ \\ [1pt]
\texttt{SAGE~($\delta$ = 3.0)} & $\textbf{79.9}\pm23.2$ & \\ [1pt]
\bottomrule
\end{tabular}
\end{table*}
\end{comment}
|
\section{Introduction}
The diffusion processes can be classified according to the relation between the mean squared displacement (MSD) defined as $\langle x^2(t)\rangle=\int_{-\infty}^{\infty} x^2P(x,t)dx$ and the time $t$. Usually, the MSD appears as $\langle x^2(t)\rangle\sim t^\alpha$; the diffusion process is called normal diffusion if $\alpha=1$, otherwise it is termed as anomalous diffusion \cite{BG1990,SLLS2004,MJCB2014}. A representative example of normal diffusion is Brownian motion \cite{NelsonE}, which is discovered by irregular motion of the pollen and other suspended particles in water. Anomalous diffusion is ubiquitous in a wide range of fields, such as diffusion processes in physics \cite{Norregaard2017}, in finance \cite{s3}, in ecology \cite{s4}, and in biology \cite{Franosch2013}, etc. Specifically, an anomalous diffusion is called subdiffusion \cite{SM1975,SW2009,JLOM2013,WSTK2011,JMJM2012,WST2010,TBKK2013} if $0<\alpha<1$ and superdiffusion \cite{CGE2000,BS2002,RJLM2015} if $\alpha>1$. Random walk process, including continuous time random walk (CTRW) and L\'{e}vy walk, is one of the most popular models to describe anomalous diffusion. There are two series of independent identically distributed (i.i.d.) random variables in CTRW model, which respectively are waiting time~$\tau$~following the distribution $\phi(\tau)$ and jump length~$l$~following the probability density function (PDF) $\lambda(l)$ \cite{s16,s11,M1969}. The corresponding subdiffusion or superdiffusion process can be modeled with infinite $\langle \tau\rangle$ and finite $\langle l^2\rangle$ or finite $\langle \tau\rangle$ and infinite $\langle l^2\rangle$, respectively. L\'evy flight \cite{F1994a,F1994b,s16} is a classical model which displays superdiffusion when the jump length density behaves as $\lambda(l)\sim1/|l|^{1+\mu}$, $0<\mu<2$. L\'{e}vy walk is another important model of random walks which is space and time coupled through finite propagation speed \cite{Zaburdaev,BFK2000}. The traditional L\'{e}vy walk is the one with constant value of speed $v_0$ and the corresponding walking length of which for each finished step is $v_0 \tau$, where $\tau$ is the walking time for each step of movement following the density $\phi(\tau)$.
The ordinary diffusion process mentioned above is implicitly with the assumption that the medium is static, which implies the distance between two stationary particles does not change with time. In reality, almost all the media the particles move in are non-static. The expected resolution of the studied dynamics and the amplitude of the displacement of the media decide whether or not the non-static behaviors can be ignored. The expansion and contraction are two typical features of the media, which
can be observed in many different areas, such as biology \cite{Simpson2015,Simpson2015b,Crampin1999,Crampin2002}, cosmology \cite{Berezinsky2006,Berezinsky2007,Aloisio2009}, fluids \cite{Haba2014}. In biology, the formation of pigmentation patterns depends on the growing concomitant tissues and organs. High energy cosmic rays is also an example of expansion from the perspective of cosmology \cite{Kotera2008}. Moreover, in \cite{Crampin1999,Berezinsky2006,Yates2014,Averbukh2014,Simpson2015,Simpson2015b,Yuste2016,F}, the authors show that the expansion of medium brings about enormous difference on the behavior of the diffusion particles, which motivates people to pay attention to the movement of the particles in non-static medium.
A growing number of scholars concentrate on analyzing the transport characters of diffusion process in non-static medium. The relevant Fokker-Planck equation which describes diffusion in non-static medium has been derived via generalized Chapman-Kolmogorov equation in \cite{Yuste2016}. The crossover effects has been discovered in a non-static medium with a power-law scale factor. Moreover, in \cite{F}, the authors introduce a comoving coordinate which can be thought as a reference frame where the medium looks to be static. Further in \cite{F,Sokolov2012}, the Fokker-Planck equation in comoving coordiante as well as the long-time asymptotic behavior of comoving MSDs have been derived for CTRW model in a non-static medium. Additionally, based on the relation between physical and comoving coordinates, the relevant physical statistics of CTRW in a non-static medium are obtained. Moreover, by introducing the conformal time, the dynamical behavior of L\'{e}vy flight in a non-static medium has also been discussed in \cite{F}, which concludes that in a non-static medium the diffusion coefficient of fractional diffusion equation for subdiffusive CTRW is time-dependent.
In this paper we discuss the dynamical behavior of L\'{e}vy walk in non-static medium. We first build the model and derive the governing equations. Then using Hermite polynomials expansion established in \cite{Pengbo} solves the equation to calculate the statistical observables.
The method of Hermite orthogonal polynomials expansion is good at calculating MSD even for some cases that are unsolvable by ordinary method of integral transform, such as L\'evy walks moving under external potential \cite{Pengbo1,Tian2021}. In addition, for the Langevin equation describing the diffusion processes under the action of external potential \cite{SIM2003,coffey,Metzle1999,JMF1999,CKGM2003}, Hermite polynomials expansion is also a doable method
\cite{Barkai1998,Metzle1999,Henry2010,Langlands2010}.
This paper is organized as follows. In Sec. \ref{sec 2}, we introduce the L\'{e}vy walk model in a non-static medium and construct the master equation in the frame of comoving coordinate. In Sec. \ref{sec 3}, the corresponding MSDs are calculated in comoving coordinate by the approach of Hermite polynomials expansion, and the asymptotic behaviors of MSDs in physical space are obtained as well. In Sec. \ref{sec 4}, the kurtosis in physical coordinate for the case of localization with exponential scale factor is calculated and the evolutions are analyzed in detail; the stationary distributions for some cases are also discussed through numerical simulations. In Sec. \ref{sec 5}, we obtain the mean first passage time by numerical simulations. The paper is concluded with some discussions in Sec. \ref{sec 6}.
\section{L\'evy walk in a non-static media}\label{sec 2}
The most representative L\'evy walk process is isotropic and with constant velocity $v_0$ \cite{Zaburdaev,First},
and then the length that L\'evy walk particle moves for each step of renewal is $v_0 \tau$, where $\tau$ indicating the walking duration of each step follows the density $\phi(\tau)$. If L\'evy walk particle stays at position $x$ after finishing some steps of moving, then the next step will arrive at $x \pm v_0 \tau$. It can be seen that the spatiotemporal coupling is
the most striking feature of L\'evy walk. Another significant difference between L\'evy walk and CTRW is that the former one has no rest after each step of walking, while the latter one must rest after each instant jumping, leading to much more difference when considering non-static media.
In this paper, we focus on one-dimension system with uniform expansion or contraction, and the physical coordinate (denoted as $y$) and comoving coordinate (denoted as $x$) are correlated with each other in the following way
\begin{equation}\label{1.2}
y(t)=a(t)x,
\end{equation}
where $a(t_0)=1$ for the initial time $t_0$. The function $a(t)$ is called as scale factor in the context of
cosmology \cite{cosmological,Introducing}. Since the initial condition of $a(t_0)=1$, the physical coordinate and comoving coordinate are same at time $t_0$. According to the property of scale factor $a(t)$, we can classify the medium; specifically, we say a medium is expanding (or contracting) if $\dot{a}(t)>0$ (or $\dot{a}(t)<0$). One can see that the distance between two stationary particles always changes with time, since the medium the particles are embedded always expands or contracts.
The CTRW model in a non-static medium has been studied in \cite{F}. Let us take the $n$th step of CTRW process for instance. For ordinary CTRW model, the process will wait at $y_{n-1}$ for a period of random time then instantly jumps $\Delta y_n$ relative to $y_{n-1}$ drawn from the PDF $\lambda(\Delta y_n)$. However, the process will move along with the non-static medium, even though it is waiting for next jump as shown in Fig. \ref{model}(a). Assuming the CTRW particle arrives at $y_{n-1}$ at time $t_{n-1}$, then due to non-static medium, it comes to $y'_{n-1}$ at time $t_n$, and further has a jump with the length of $\Delta y_n$ relative to $y'_{n-1}$. Since the scale factors also change with time, we must borrow comoving coordinates to determine $y'_{n-1}$, which is $x_{n-1}^+=x_n^-$ (`+' and `-' respectively mean right and left limit); further from the relation between physical coordinate and comoving coordinate \eqref{1.2}, it holds
\begin{equation*}
y'_{n-1}= \frac{a(t_n)}{a(t_{n-1})} y_{n-1}.
\end{equation*}
We denote $\Delta y'_n=y'_{n-1}-y_{n-1}$.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.45]{CTRW_expanding.eps}
\includegraphics[scale=0.45]{lw_expanding.eps}
\caption{CTRW and L\'evy walk in expanding media. The comoving and physical coordinates are $x$ and $y$, respectively. Figure (a) represents CTRW process in expanding medium, where $\Delta y_n$ is a random variable drawn from the PDF $\lambda(\Delta y_n)$, $\Delta y'_n$ represents the displacement caused by expanding medium from time $t_{n-1}$ to $t_n$, and $\Delta x_n=\Delta y_n/a(t_n)$ denotes the jumping length in comoving space. Figure (b) illustrates L\'evy walk process in expanding medium with $\Delta y_n = v_0 (t_n-t_{n-1})$, and the other notations share similar meanings with those in (a).}
\label{model}
\end{figure}
In this paper, we mainly consider the dynamic behavior of L\'evy walk in a non-static medium. As shown in Fig. \ref{model}(b), since L\'evy walk process has no rest, the trajectory behaves much different. Consider the time $\xi \in (t_{n-1},t_n]$. Since L\'evy walk does not wait, its movement in physical coordinate consists of two parts: one is movement caused by non-static medium, $y'_\xi = (a(\xi)/a(t_{n-1})) y_{n-1}$; the other one is walking length relative to $y'_\xi$, that is, $\pm v_0 (\xi-t_{n-1})$. Whereas in CTRW model, if we also consider the time $\xi$ between the $(n-1)$th and $n$th jumping events, then the movement is only influenced by non-static medium. This may cause a major difference between CTRW and L\'evy walk models, in fact, assuming that $t=\xi$ is the final observation time, then at time $t$ the L\'evy walk particle in non-static medium locates at
\begin{equation}\label{model_yt}
y_{t}= \frac{a(t)}{a(t_{n})} y_{n} \pm v_0 (t-t_n).
\end{equation}
Obviously, the static media whose scale factor $a(t)=1$ is the case considered in ordinary L\'evy walk problem \cite{Zaburdaev,First}, which indicates the distance between two stationary particles does not change with time.
As discussed above, the walking length $\Delta y$ for L\'evy walk in physical coordinate is determined by the walking time $\tau$ drawn from the PDF $\phi(\tau)$ or the survival time; with a little abuse of notation we still denote it as $\tau$ following the survival probability
\begin{equation}\label{survival_prob_t}
\psi(\tau)=\int_\tau^\infty \phi(\tau') d \tau',
\end{equation}
therefore $\Delta y$ can be given by the following conditional density
\begin{equation}\label{1.3}
\lambda(\Delta y \vert \tau)=\frac{1}{2}\delta(\Delta y-v_0\tau)+\frac{1}{2}\delta(\Delta y+v_0\tau),
\end{equation}
where $\delta(\cdot)$ represents the Dirac $\delta$-function.
It is not easy to directly consider the process in physical coordinate since the relation of \eqref{model_yt} is hard to be conveniently utilized. Therefore in the following we will change the process into the one with comoving coordinate.
In order to transfer the L\'evy walk process from physical coordinate to comoving coordinate, first of all we need to identify the walking length $\Delta x$ for each step of moving. It can be accomplished by the relation between $\Delta y$ and $\Delta x$ at time $t$ \cite{F}, i.e, $\Delta y=a(t) \Delta x$; specially we choose $t$ to be the renewal time when the walker finishes a few steps of walking (such as $n$ steps) or the overall observation time, and take $\tau$ to be the walking time of current step (the $n$th step) or the survival time so that $(t-\tau)$ should be the time when current step starts. Denote the conditional density of $\Delta x$ as $\lambda(\Delta x \vert t; \tau)$, which means the particle in comoving coordinate walks $\Delta x$ for a period of $\tau$ at time $t$.
According to the relation between $\Delta y$ and $\Delta x$ and combining with \eqref{1.3}, one has
\begin{equation}\label{1.4}
\begin{split}
\lambda(\Delta &x|t;\tau) = a(t) \lambda\big(a(t) \Delta x \vert \tau\big) \\
& =\frac{1}{2}a(t)\left( \delta \left(a(t) \Delta x-v_0 \tau\right)+\delta \left(a(t) \Delta x+v_0 \tau\right)\right)\\
&=\frac{1}{2} \left[ \delta\left( \Delta x -\frac{v_0\tau}{a(t)} \right)+ \delta\left( \Delta x +\frac{v_0\tau}{a(t)} \right)\right].
\end{split}
\end{equation}
We use $W(x,t)$ to denote the PDF of finding a particle at position $x$ in comoving coordinate at time $t$. Assume the related PDF in physical space to be $P(y,t)$. According to the relation between comoving and physical coordinates \eqref{1.2}, one gets
\begin{equation}\label{1.5}
P(y,t)= \frac{1}{a(t)} W\left(\frac{y}{a(t)},t\right).
\end{equation}
The PDF $P(y,t)$ as well as some important statistics such as the first two moments of L\'evy walk in non-static medium can be calculated according to \eqref{1.5}. In the following we aim to derive the transport equations of L\'{e}vy walk in comoving coordinate. As one can see in the comoving coordinate, if the particle locates at $x$ at time $t$ which is the renewal time, and the walking time for this step $\tau$ following the PDF $\phi(\tau)$, then the previous renewal event happens at time $t-\tau$ at position $x - \Delta x$ with $\Delta x$ given by $\lambda(\Delta x \vert t; \tau)$. If we denote the corresponding PDF of finding particle just arriving at position $x$ at renewal time $t$ as $q(x,t)$, accordingly we have
\begin{equation}\label{1}
\begin{split}
q(x,t) & \! =\!\int_{-\infty}^{\infty}d\Delta x\int_{0}^{t} q\Big(x-\Delta x,t-\tau\Big)\phi(\tau)\lambda(\Delta x|t;\tau)d\tau\\
& +P_0(x)\delta(t),
\end{split}
\end{equation}
where the second term on r.h.s. represents the initial distribution.
Then the PDF in the comoving coordinate denoted as $W(x,t)$ can be given though \eqref{1}, which is
\begin{equation}\label{1.}
W(x,t)\!=\!\int_{-\infty}^{\infty}d\Delta x\int_{0}^{t} q\Big(x-\Delta x,t-\tau\Big)\psi(\tau)\lambda(\Delta x|t;\tau)d\tau,
\end{equation}
where $\psi(\tau)$ is defined in \eqref{survival_prob_t}. By taking Laplace transform, defined as $\hat{g}(s)=\mathscr{L}_t \{g(t)\}(s)=\int_{0}^{\infty} e^{-s t}g(t)dt$, on survival probability $\psi(\tau)$ there exists
\begin{equation}\label{survival_prob}
\hat{\psi}(s)=\frac{1-\hat{\phi}(s)}{s}.
\end{equation}
Substituting the expression of $\lambda(\Delta x|t;\tau)$ given by \eqref{1.4} into \eqref{1} and \eqref{1.}, respectively, we have
\begin{equation}\label{1..6}
\begin{split}
q(x,&t)=\frac{1}{2}\int_{0}^{t}q\bigg(x-\frac{v_0\tau}{a(t)},t-\tau\bigg)\phi(\tau)d\tau \\
&+\frac{1}{2} \int_{0}^{t}q\bigg(x+\frac{v_0\tau}{a(t)},t-\tau\bigg)\phi(\tau)d\tau +P_0(x)\delta(t)
\end{split}
\end{equation}
and
\begin{equation}\label{1..7}
\begin{split}
W(x,t)&=\frac{1}{2} \int_{0}^{t}q\bigg(x-\frac{v_0\tau}{a(t)},t-\tau\bigg)\psi(\tau)d\tau \\
&+\frac{1}{2} \int_{0}^{t}q\bigg(x+\frac{v_0\tau}{a(t)},t-\tau\!\bigg)\psi(\tau)d\tau.
\end{split}
\end{equation}
The normalization of $W(x,t)$ can be immediately verified by Fourier transform, defined as $\tilde{f}(k)=\mathscr{F}_{x}\{f(x)\}(k)=\int_{-\infty}^{\infty}e^{-i k x}f(x)dx$. In fact by taking Fourier-Laplace transform on \eqref{1..6} and \eqref{1..7}, and letting $k=0$, there exists
\begin{equation*}
\begin{split}
\hat{\tilde{q}}(0,s) &=\frac{1}{1-\hat{\phi}(s)},\\
\widehat{\widetilde{W}}(0,s) &= \hat{\psi}(s) \hat{\tilde{q}}(0,s)=1/s,
\end{split}
\end{equation*}
indicating normalization of PDF $W(x,t)$ after inverse Laplace transform.
In the following, to theoretically obtain the behaviors of moments in comoving and physical coordinates, we utilize Hermite orthogonal polynomials to approach PDF \cite{Pengbo1,Pengbo,Tian2021}.
\section{Hermite polynomial approximation to L\'{e}vy walk in non-static medium}\label{sec 3}
The Hermite polynomials form an orthogonal basis of the Hilbert space with the inner product $\langle f,g \rangle=\int_{-\infty}^{\infty} f(x)\bar{g}(x)e^{-x^2} dx$ \cite{hermit_intro}. In this section, we utilize the Hermite polynomials to approach the PDF of L\'{e}vy walk in comoving coordinate. According to the complete orthogonal system, we assume that in Hilbert space $q(x,t)$ and $W(x,t)$ can be, respectively, represented as
\begin{align}
q(x,t)&=\sum_{n=0}^{\infty} H_n(x) T_n(t) e^{-x^2},\label{1.9}\\
W(x,t)&=\sum_{n=0}^{\infty} H_n(x) R_n(t) e^{-x^2},\label{1.10}
\end{align}
where $H_n(x), n=0,1,\cdots,$ represent the Hermite polynomials, $\{T_n(t)\}$ and $\{R_n(t)\}$ are a series of functions with respect to $t$ to be determined.
We take the initial distribution of the particle as a Dirac $\delta$-function, i.e., $P_0(x)=\delta(x)$. Substituting the assumed forms of $q(x,t)$ and $W(x,t)$ in \eqref{1.9} and \eqref{1.10} into \eqref{1..6} and \eqref{1..7}, respectively, in Laplace space the iteration relation of $\widehat{T}_m(s)$ and the relation between $\widehat{R}_m(s)$ and $\widehat{T}_m(s)$ are given as
\begin{widetext}
\begin{equation}\label{1.11}
\begin{split}
\sqrt{\pi} 2^m m!\widehat{T}_m(s)=& \frac{1}{2}\sum_{k=0}^{m}\frac{2^k \sqrt{\pi} m! }{(m-k)!}\left((2 v_0)^{m-k}+(-2 v_0)^{m-k}\right)\mathscr{L}_{t}\left\{ a(t)^{k-m} \mathscr{L}^{-1}_{s}\left[(-1)^{m-k}\hat{\phi}^{(m-k)}(s)\widehat{T}_k(s)\right](t)\right\}(s)\\
&+H_m(0),
\end{split}
\end{equation}
where $\hat{\phi}^{(m-k)}(s)=\frac{d^{m-k}}{ds^{m-k}} \phi(s)$, and
\begin{equation}\label{1.12}
\sqrt{\pi} 2^m m!\widehat{R}_m(s)= \frac{1}{2}\sum_{k=0}^{m}\frac{2^k \sqrt{\pi} m! }{(m-k)!}\left((2 v_0)^{m-k}+(-2 v_0)^{m-k}\right)\mathscr{L}_{t}\left\{ a(t)^{k-m} \mathscr{L}^{-1}_{s}\left[(-1)^{m-k}\hat{\psi}^{(m-k)}(s)\widehat{T}_k(s)\right](t)\right\}(s).
\end{equation}
\end{widetext}
The detailed derivations are shown in Appendix \ref{App_B}.
The long time asymptotic behaviors of the first two moments in both coordinates attract our interest. Firstly, we concentrate on the first two moments for the process in comoving coordinate. From \cite{Zaburdaev}, there exists
\begin{equation}\label{mmt}
\langle \hat{x}^m(s)\rangle=(i)^m \left. \frac{\partial^m}{\partial k^m} \widehat{\widetilde{W}}(k,s)\right|_{k=0}.
\end{equation}
Further, the specific form of $W(x,t)$ can be obtained by combining \eqref{1.10} with \eqref{a1}, which is
\begin{equation}\label{1.13}
W(x,t)=\sum_{n=0}^{\infty}(-1)^n \frac{d^n}{d x^n} e^{-x^2} R_n(t).
\end{equation}
Applying the Fourier transform $x\to k$ and Laplace transform $t\to s$ on \eqref{1.13} lead to
\begin{equation}\label{1.14}
\widehat{\widetilde{W}}(k,s)=\sum_{n=0}^{\infty} \sqrt{\pi}(-i k)^n e^{-\frac{k^2}{4}} \widehat{R}_n(s).
\end{equation}
The first two comoving moments of the process in Laplace space can be further represented as
\begin{equation}\label{firstmoment}
\langle \hat{x}(s)\rangle =i \left.\frac{\partial}{\partial k} \widehat{\widetilde{W}}(k,s)\right|_{k=0}=\sqrt{\pi} \widehat{R}_1(s)
\end{equation}
and
\begin{equation}\label{msd}
\begin{split}
\langle\hat{x}^2(s)\rangle & =- \frac{\partial^2}{\partial k^2} \widehat{\widetilde{W}}(k,s)|_{k=0}\\
&=\frac{\sqrt{\pi}}{2} \widehat{R}_0(s)+2\sqrt{\pi} \widehat{R}_2(s).
\end{split}
\end{equation}
In order to calculate the first two moments of the process in comoving coordinate, the values of $\widehat{R}_0(s), \widehat{R}_1(s)$, and $\widehat{R}_2(s)$ should be calculated first. Taking $m=0,1,2$ in \eqref{1.11} and \eqref{1.12} results in
\begin{equation}\label{1.16}
\begin{split}
&\widehat{T}_0(s)=\frac{1}{\sqrt{\pi}\big(1-\hat{\phi}(s)\big)}, \\
& \widehat{R}_0(s)=\frac{1}{s\sqrt{\pi}},\\
&\widehat{T}_1(s)=\widehat{R}_1(s)=0.
\end{split}
\end{equation}
Additionally,
\begin{equation}\label{1.17}
\begin{split}
& \sqrt{\pi}2^3\left(1-\hat{\phi}(s)\right) \widehat{T}_2(s)\\
& =\sqrt{\pi} (2 v_0)^2\mathscr{L}_{t}\left\{\frac{1}{a^2(t)}\mathscr{L}^{-1}_{s}\left[\hat{\phi}''(s)\widehat{T}_0(s)\right](t)\right\}(s)-2,
\end{split}
\end{equation}
\begin{equation}\label{1.18}
\begin{split}
& \sqrt{\pi}2^3 \widehat{R}_2(s)- \sqrt{\pi}2^3\hat{\psi}(s)\widehat{T}_2(s)\\
& =\sqrt{\pi} (2 v_0)^2\mathscr{L}_{t}\left\{\frac{1}{a^2(t)}\mathscr{L}^{-1}_{s}\left[\hat{\psi}''(s)\widehat{T}_0(s)\right](t)\right\}(s).\\
\end{split}
\end{equation}
Moreover, the first moment in comoving coordinate $\langle x(t)\rangle=0$ since $\widehat{R}_1(s)=0$, which indicates the process in comoving coordinate is symmetric. It should be noted that the normalization of the PDF $W(x,t)$ can also be verified through Hermite polynomials expansion and the result of $R_0=1/\sqrt{\pi}$, the detailed calculations of which can be found in \cite{Pengbo,Pengbo1}.
Next we change the results from comoving coordinate to the physical coordinate. In fact from \eqref{1.5}, the following relations of the first two moments in different coordinates can be easily obtained,
\begin{equation}\label{1..21}
\langle y(t)\rangle=a(t)\langle x(t)\rangle
\end{equation}
and
\begin{equation}\label{1.21}
\langle y^2(t)\rangle=a^2(t)\langle x^2(t)\rangle.
\end{equation}
Obviously the average moment $\langle y(t)\rangle=0$ since $\langle x(t)\rangle=0$, which means the expansion or contraction of medium does not change the symmetry of the process. It can be easily noticed that the PDF $P(y,t)$ is also normalized from the normalization of $W(x,t)$.
In the following, by considering some representative walking time PDFs $\phi(\tau)$ and scale factors $a(t)$, we discuss the long-time asymptotic behaviors of the MSDs in both comoving and physical coordinates.
\section{Dynamics of L\'evy walk in non-static medium} \label{sec 4}
\subsection{Exponentially distributed walking time}
In this subsection the walking time PDF is assumed to be exponential distribution, i.e., $\phi(\tau)=\lambda e^{-\lambda\tau}$, with $\lambda>0$. For different scale factors $a(t)$, we are going to analyze the asymptotic behavior of the statistical properties of the process in comoving and physical coordinates.
\subsubsection{Exponential scale factor}
We focus on the case that the non-static medium is described by the exponential scale factor
\begin{equation}\label{1.19}
a(t)=\exp(H t)
\end{equation}
with the ``Hubble constant'' $H=a'(t)/a(t)$ \cite{cosmological, Introducing}. The case $H>0$ corresponds to an expanding medium while the case $H<0$ represents a contracting one.
Combining \eqref{msd} with \eqref{1.16}, \eqref{1.17}, as well as \eqref{1.18}, the MSDs in comoving coordinate can be got
\begin{equation}\label{1.20}
\langle x^2(t)\rangle\sim
\left\{
\begin{split}
&\frac{\lambda v_0^2}{H (2 H+\lambda)^2}, \,\quad\quad\quad \mbox{if $H>0$}, \\
& \frac{(2 H-\lambda)v_0^2}{H \lambda^2}e^{-2 H t}, \quad \mbox{if $H<0$}.
\end{split}
\right.
\end{equation}
The results in \eqref{1.20} are supported by Fig. \ref{expexp}(a) and (b). It can be concluded from \eqref{1.20} that for $H>0$ the comoving variance tends to be a constant for sufficient long time which relies on the value of $H$ and $\lambda$ as well as the velocity $v_0$ of L\'{e}vy walks. In contrast, the MSD in comoving coordinate displays exponential growth for $H<0$. The similar conclusions for subdiffusion CTRW in non-static medium can be found in \cite{F}.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.28]{expaexpHpositive}
\includegraphics[scale=0.28]{expaexpHnegative}
\includegraphics[scale=0.28]{yexpaexpHpositive}
\includegraphics[scale=0.28]{yexpaexpHnegative}
\caption{Numerical simulations of the MSDs of L\'{e}vy walk in uniform expansion medium with exponential scale factor by sampling over $10^4$ realizations. The walking time PDF of L\'{e}vy walk behaves as exponential distribution $\phi(\tau)=\lambda e^{-\lambda\tau}$ with $\lambda=1$. The parameters are $v_0=1$ and $x_0=y_0=0$. For (a) and (c), we take $H=0.1$; for (b) and (d), $H=-0.1$. The solid lines in (a), (b) and (c), (d) respectively are the theoretical results shown in \eqref{1.20} and \eqref{1.22}.}
\label{expexp}
\end{figure}
Now we turn to analyzing MSDs in physical coordinate. Combining \eqref{1.20} with the relation \eqref{1.21}, we have
\begin{equation}\label{1.22}
\langle y^2(t)\rangle\sim
\left\{
\begin{split}
& \frac{\lambda v_0^2}{H (2 H+\lambda)^2}e^{2 H t}, \quad \mbox{if $H>0$}, \\
& \frac{(2 H-\lambda)v_0^2}{H \lambda^2}, \quad \quad \quad ~ \mbox{if $H<0$},
\end{split}
\right.
\end{equation}
which have been verified by Fig. \ref{expexp}(c) and (d).
For $H>0$, the variance in physical space exponentially increases with respect to the time $t$. However, the MSD keeps a constant for sufficiently long time for $H<0$, which implies a stationary propagator function in physical space can be reached for long time limit. Compared with the MSD of L\'{e}vy walk in static medium which behaves as $\langle y^2(t)\rangle \sim t$, one can conclude that the dominant term of the diffusive processes is the displacement induced by exponential scale factor of the medium.
Next we numerically simulate the stationary distribution $P^{st}(x)$ in comoving or physical space.
For L\'{e}vy walk with $\phi(\tau)=\lambda e^{-\lambda\tau}$ in comoving coordinate and a positive $H$, the role of parameter $\lambda$ can be concluded from Fig. \ref{pstm}(a), specifically the increase of $\lambda$ changes $P^{st}(x)$ from bimodal distribution to unimodal one. While Fig. \ref{pstm}(b) indicates that the change of velocity $v_0$ does not make the stationary distribution become from bimodal to unimodal or vice versa; increasing $v_0$ can only flatten $P^{st}(x)$. Similarly, it can be concluded from Fig. \ref{pstm}(c) that the variation of Hubble constant $H>0$ can neither change the unimodality of stationary distribution, and the decrease of $H$ also plays the role of flattening $P^{st}(x)$. Interestingly, the stationary distribution in physically coordinated $P^{st}(y)$ for L\'evy walk process with $H<0$ has different properties. Specifically, according to Fig. \ref{pstym}(a) we conclude that the increasing absolute value of Hubble constant $|H|$ changes $P^{st}(y)$ from unimodal distribution to bimodal one. The increase of $v_0$ can only flatten the stationary distribution concluded from Fig. \ref{pstym}(b), and the unimodal or bimodal property cannot be affected by variation of $v_0$. The role of $\lambda$ turns out to be completely different from the one in comoving coordinate, specifically from Fig. \ref{pstym}(c) the stationary distribution is flattened as the parameter $\lambda$ decreases whereas the unimodal or bimodal property keeps the same with the change of $\lambda$.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.19]{pstlambdachangevHfix}
\includegraphics[scale=0.19]{pstvchangeHlambdafix}
\includegraphics[scale=0.19]{pstHchangevlambdafix}
\caption{Stationary PDFs of L\'{e}vy walk with $\phi(\tau)=\lambda e^{-\lambda\tau}$ and exponentially expanding scale factor in comoving coordinates by sampling over $2\times10^4$ realizations. Here we assume $x_0=0$. For panel (a), we take $H=0.1$, $v_0=1$; for panel (b), we use $H=0.1$, $\lambda=1$; for panel (c), $\lambda=1$ and $v_0=1$.}
\label{pstm}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.19]{pstyHchangevlambdafix}
\includegraphics[scale=0.19]{pstyvchangeHlambdafix1}
\includegraphics[scale=0.19]{pstylambdachangevHfix}
\caption{Stationary PDFs of L\'{e}vy walk with $\phi(\tau)=\lambda e^{-\lambda\tau}$ and exponentially contracting scale factor in physical coordinates by sampling over $2\times10^4$ realizations. Here we assume $y_0=0$. For panel (a), we take $v_0=1$ and $\lambda=1$; for panel (b), we use $H=-0.1$ and $\lambda=1$; for panel (c), $H=-0.1$ and $v_0=1$;.}
\label{pstym}
\end{figure}
Another quantity to characterize the property of PDF for stochastic process is kurtosis $K$, which measures the tails of the PDF. For a symmetric process with zero spatial average, the kurtosis can be defined as \cite{Pengbo1,Tian2021}
\begin{equation}\label{K}
K=\frac{\langle y^4(t)\rangle}{\langle y^2(t)\rangle^2}.
\end{equation}
It is usually to choose $3$ as a benchmark obtained from normal distribution, and then make a comparison.
Specifically for $K>3$ (or $K<3$), we call the corresponding PDF to be leptokurtic (or platykurtic). In this part, we only consider the kurtosis for the stationary density in physical coordinate, therefore we choose $H<0$. According to the definition of kurtosis, the fourth moment $\langle y^4(t)\rangle$ is required additionally. Similarly we first consider the corresponding statistical quantity in comoving coordinate $\langle x^4(t)\rangle$, and then convert it to the one in physical coordinate. It can be obtained from \eqref{mmt} and \eqref{1.14} that
\begin{equation}\label{fourthmomentxs}
\langle \hat{x}^4(s)\rangle \!=\frac{3}{4}\sqrt{\pi} \widehat{R}_0(s)+6\sqrt{\pi} \widehat{R}_2(s)+24\sqrt{\pi} \widehat{R}_4(s),
\end{equation}
in which $\widehat{R}_4(s)$ can be explicitly calculated through \eqref{1.11} and \eqref{1.12}. Then by applying inverse Laplace transform on \eqref{fourthmomentxs}, there exists
\begin{equation}\label{fourthmomentx}
\langle x^4(t)\rangle \!\sim\!\frac{3e^{-4 H t}\!(32 H^4\!-\!40 H^3 \lambda \!+\!16 H^2\lambda^2\!-\!6 H \lambda^3\!+\!\lambda^4)\! v_0^4 }{H^2 (2 H-\lambda)^2 \lambda^4}.
\end{equation}
Finally according to the relation of \eqref{1.5}, the asymptotic form of fourth moment in physical space is
\begin{equation}\label{fourthmomenty}
\begin{split}
\langle y^4(t)\rangle & = a^4(t) \langle x^4(t)\rangle\\
& \sim\frac{3(32 H^4-40 H^3 \lambda+16 H^2 \lambda^2-6 H \lambda^3+\lambda^4) v_0^4 }{H^2 (2 H-\lambda)^2 \lambda^4}.
\end{split}
\end{equation}
Combining with the MSD in \eqref{1.22} with $H<0$, for sufficiently long time there exists
\begin{equation}\label{Kvalue}
K\sim\frac{3(32 H^4-40 H^3 \lambda+16 H^2 \lambda^2-6 H \lambda^3+\lambda^4)}{(2 H-\lambda)^4},
\end{equation}
which is verified by Fig. \ref{K}. Therefore, the asymptotic behavior of kurtosis is determined by Hubble constant and the inverse average of exponentially distributed walking time. At first for any fixed $H<0$ we consider $K$ as a function of $\lambda$. As shown in Fig. \ref{K}(a),
$\lim_{\lambda\to 0} K=6$ indicating that the corresponding stationary density is leptokurtic for small $\lambda$. Then kurtosis decreases as $\lambda$ increases until $K$ reaches the minimum $K_{\min}=27/4-3\sqrt{2}\approx 2.51$ at $\lambda=-2(1+\sqrt{2})H$, and the minimum has no relevance to $H$, so that the stationary distribution enters the platykurtic category. With the further increase of $\lambda$, $K$ increases to finally reach the limit $3$ since $\lim_{\lambda\to\infty} K=3$. On the other hand, for fixed $\lambda>0$, we consider $K$ as a function of $H$, then the limit can be easily obtained, $\lim_{H\to 0}K=3$, which has been verified in Fig. \ref{K}(b). Further when $|H|$ increases, $K$ decreases to the same minimum $K_{\min}$ at $H=-\lambda(\sqrt{2}-1)/2$, indicating that the stationary density is platykurtic. Afterwards, $K$ keeps increasing to the limit $\lim_{H\to -\infty} K=6$, therefore the stationary density changes from platykurtic category to leptokurtosis one.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.28]{K}
\includegraphics[scale=0.28]{kH}
\caption{Numerical simulations of kurtosis of L\'{e}vy walk in a non-static medium with exponential contraction by sampling over $3\times10^4$ realizations. The walking time PDF of L\'{e}vy walk behaves as exponential distribution $\phi(\tau)=\lambda e^{-\lambda\tau}$. The initial speed of L\'{e}vy walk is assumed to be $v_0=1$. For (a), we fix $H=-0.1$ and for (b) we take $\lambda=1$. The solid lines are the theoretical results \eqref{Kvalue}.}
\label{K}
\end{figure}
\subsubsection{Power-law scale factor}
Power-law scale factor also has many different types of applications in cosmology, which has the form
\begin{equation}\label{1.23}
a(t)=\left(\frac{t+t_0}{t_0}\right)^\beta,
\end{equation}
where $t$ is the time elapsed after the initial time $t_0$ and the positive value of $\beta$ corresponds to an expanding medium while the negative value of $\beta$ describes a contracting one.
The asymptotic behaviors of the MSDs can be obtained by using \eqref{msd}, \eqref{1.16}, \eqref{1.17}, and \eqref{1.18}, which behave as
\begin{equation}\label{1.24}
\langle x^2(t)\rangle\sim
\left\{
\begin{split}
&\frac{2 t_0^{2\beta}v_0^2}{\lambda(1-2\beta)}t^{1-2\beta}, ~~ \mbox{if $\beta<\frac{1}{2}$, }\\
& C_{1}+\frac{2 t_0 v_0^2 }{\lambda}\ln t, \quad \text{if}~\beta=\frac{1}{2},\\
& C_2,\quad\quad\quad\quad\quad\quad ~ \text{if}~\beta>\frac{1}{2},
\end{split}
\right.
\end{equation}
where the constants
$C_{1}=2 t_0 v_0^2\big(-1+e^{\lambda t_0}(\lambda t_0-1)\Gamma(0,\lambda t_0)-\ln t_0\big)/\lambda$ and $C_2=2 t_0 v_0^2 \big(2-2\beta+e^{\lambda t_0}(2\beta-1)(-2+2\beta+\lambda t_0)E_{2\beta}(\lambda t_0)\big)/[\lambda(2\beta-1)]$ with exponentially integral function
$E_n(x)=\int_{1}^{\infty}e^{-x t}/t^n dt$ and $\Gamma(a,b)$ being the incomplete gamma function. Therefore, the asymptotic long-time behavior of the MSD in comoving coordinate displays a power-law growth with exponent $1-2\beta$ when $\beta<\frac{1}{2}$, while when $\beta=1/2$ the MSD asymptotically behaves like a logarithmic growth. Finally when $\beta>1/2$, the MSD asymptotically becomes a constant after sufficiently long time in comoving coordinate. The results of \eqref{1.24} are confirmed by the numerical simulations given in Fig. \ref{exppowlaw}(a), (b), and (c).
Correspondingly, the asymptotic behaviors of the MSDs can be obtained in physical coordinate by \eqref{1.21} for different categories of $\beta$ compared to the critical value $1/2$,
\begin{equation}\label{1.25}
\langle y^2(t)\rangle\sim
\left\{
\begin{split}
& \frac{2 v_0^2}{\lambda(1-2\beta)}t, ~~\quad\quad\quad \mbox{if $\beta<\frac{1}{2}$}, \\
& C_1 t +\frac{2 t_0 v_0^2}{\lambda} t \ln t, \quad~ \mbox{if $\beta=\frac{1}{2}$}, \\
& C_2 t^{2\beta}, \quad\quad\quad\quad\quad\quad \mbox{if $\beta>\frac{1}{2}$}.
\end{split}
\right.
\end{equation}
Intuitively, the $\beta$ controls the effects of expanding (for $\beta>0$) or contracting (for $\beta<0$) medium. As a result, for weakly expanding medium with $0<\beta<1/2$ or contracting medium, the corresponding L\'evy walk in non-static medium behaves asymptotically as a normal diffusion, which turns out to be completely different from the result of exponential scale factor shown in \eqref{1.22}, and also implies that the power-law expanding medium is much weaker than the exponential one. Besides the result of normal diffusion also indicates that the movement of the particle mainly depends on the intrinsic stochastic motion in this category of $\beta$. When the expanding effect of the medium is stronger with $\beta>\frac{1}{2}$, then the process in physical coordinate behaves like superdiffusion, which indicates expanding medium takes in charge.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.28]{expapowlawsmall}
\includegraphics[scale=0.28]{expapowerlawmiddle}
\includegraphics[scale=0.28]{expapowlawlarge}
\includegraphics[scale=0.28]{yexpapowlawsmall}
\includegraphics[scale=0.28]{yexpapowerlawmiddle}
\includegraphics[scale=0.28]{yexpapowlawlarge}
\caption{Numerical simulations of the MSDs of L\'{e}vy walk in a uniformly power-law expansion medium by sampling over $10^4$ realizations. The walking time PDF of L\'{e}vy walk behaves as exponential distribution $\phi(\tau)=\lambda e^{-\lambda\tau}$ with $\lambda=1$. The parameter $v_0=1$. For (a) and (d), we take $t_0=1$, $\beta=0.1$ (square), and $\beta=-0.1$ (star); for (b) and (e), we use $\beta=0.5$, $t_0=1$ (square), and $t_0=10$ (star); for (c) and (f), we take $t_0=1$ and $\beta=1$. The solid lines in (a), (b), (c) are the theoretical results shown in \eqref{1.24}, and the solid lines in (d), (e), (f) are the theoretical results given in \eqref{1.25}.}
\label{exppowlaw}
\end{figure}
\subsection{Power-law distributed walking time}
In the following, we consider the MSDs of L\'evy walk process in non-static medium whose walking time follows Pareto distribution
\begin{equation}\label{1.26}
\phi(\tau)=\frac{1}{\tau_0}\frac{\alpha}{(1+\tau/\tau_0)^{1+\alpha}}
\end{equation}
with $\tau_0>0$ and $\alpha>0$. The corresponding asymptotic form for $\alpha\neq 1,2$ in Laplace space after neglecting higher order of small $s$ behaves like \cite{Zaburdaev}
\begin{equation}\label{waiting_time_appr}
\hat{\phi}(s)\sim1-\frac{\tau_0}{\alpha-1}s-\tau_0^\alpha \Gamma(1-\alpha) s^\alpha+\frac{\tau_0^2 s^2}{(\alpha-1)(\alpha-2)}.
\end{equation}
Similarly, the asymptotic behaviors of the MSDs in both coordinates with respect to different types of scale factors $a(t)$ are considered in the following.
\subsubsection{Exponential scale factor}
The iteration relation of ${\widehat{T}_m(s)}$ as well as the relation between $\widehat{R}_m(s)$ and $\widehat{T}_m(s)$ for $m=0, 1, 2$ can be obtained by substituting $a(t)$ \eqref{1.19}, $\phi(\tau)$ \eqref{1.26} into \eqref{1.17}, \eqref{1.18}, respectively. Further, combining with \eqref{msd}, we have the following long time asymptotic behavior for $H>0$ in comoving coordinate
\begin{equation}\label{1.27}
\begin{split}
\langle x^2(t)\rangle &\sim \frac{\alpha v_0^2(1-\alpha-2 H \tau_0)}{4 H^2\big(1-\tau_0^\alpha2^\alpha\alpha e^{2H\tau_0}H^\alpha\Gamma(-\alpha,2 H \tau_0)\big)} \\
& +\frac{\alpha v_0^2 e^{2 H \tau_0}(-1+\alpha)\alpha E_{1+\alpha}(2 H \tau_0)}{4 H^2\big(1-\tau_0^\alpha2^\alpha\alpha e^{2H\tau_0}H^\alpha\Gamma(-\alpha,2 H \tau_0)\big)}\\
& +\frac{\alpha v_0^2 e^{2 H \tau_0}(4\alpha H\tau_0+4H^2\tau_0^2)E_{1+\alpha}(2 H \tau_0)}{4 H^2\big(1-\tau_0^\alpha2^\alpha\alpha e^{2H\tau_0}H^\alpha\Gamma(-\alpha,2 H \tau_0)\big)},
\end{split}
\end{equation}
which is a constant.
On the other hand, for $H<0$ there exists
\begin{equation}\label{1.28}
\langle x^2(t)\rangle\sim
\left\{
\begin{split}
\frac{1}{2}(\alpha-2)(\alpha-1)v_0^2 e^{-2 H t} t^2, & \quad \mbox{if $0<\alpha<1$}, \\
\frac{(\alpha-1)\tau_0^{\alpha-1}v_0^2 }{(3-\alpha)} e^{-2 H t}t^{3-\alpha}, &\quad \mbox{if $1<\alpha<2$}.
\end{split}
\right.
\end{equation}
For exponential expansion medium $H > 0$, the comoving variance tends to be a constant when the time is large enough, which is similar to the previous case of exponentially distributed walking time with exponential expansion. On the contrast, the MSDs display exponential growth multiplied by power-law one for exponential contraction medium $H <0$.
Furthermore, the MSDs in physical space can be figured out for sufficiently long time $t$. For $H>0$ we get
\begin{equation}\label{1.29}
\begin{split}
\langle y^2(t)\rangle&\sim \frac{\alpha v_0^2(1-\alpha-2 H \tau_0)e^{2 H t}}{4 H^2\big(1-\tau_0^\alpha2^\alpha\alpha e^{2H\tau_0}H^\alpha\Gamma(-\alpha,2 H \tau_0)\big)} \\
& +\frac{\alpha v_0^2e^{2 H (t+\tau_0)}(-1+\alpha)\alpha E_{1+\alpha}(2 H \tau_0)}{4 H^2\big(1-\tau_0^\alpha2^\alpha\alpha e^{2H\tau_0}H^\alpha\Gamma(-\alpha,2 H \tau_0)\big)}\\
& +\frac{\alpha v_0^2e^{2 H (t+\tau_0)}(4\alpha H\tau_0+4H^2\tau_0^2)E_{1+\alpha}(2 H \tau_0)}{4 H^2\big(1-\tau_0^\alpha2^\alpha\alpha e^{2H\tau_0}H^\alpha\Gamma(-\alpha,2 H \tau_0)\big)}.
\end{split}
\end{equation}
And for $H<0$ we have
\begin{equation}\label{1.30}
\langle y^2(t)\rangle\sim \left\{
\begin{split}
& \frac{1}{2}(2-\alpha)(1-\alpha)v_0^2 t^2, ~\quad \mbox{if $0<\alpha<1$}, \\
& \frac{(\alpha-1)\tau_0^{\alpha-1}v_0^2}{(3-\alpha)}t^{3-\alpha}, \quad~~ \mbox{if $1<\alpha<2$}.
\end{split}
\right.
\end{equation}
A difference between the cases of positive and negative $H$ is that the MSD in physical coordinate asymptotically behaves like exponential growth for $\alpha\in (0,1)\cup (1,2)$ for $H>0$, while for the negative case the type of diffusion relies on the category of $\alpha$, specifically when $\alpha\in (0,1)$ ballistic diffusion can be observed and for $\alpha\in(1,2)$ the process behaves like superdiffusion. A significant difference for the case of $H<0$ can also be found by comparing the result of \eqref{1.30} with the localization one of \eqref{1.22}. Intuitively, a superdiffusive type of L\'evy walk in physical coordinate can keep the diffusion exponent, which is a kind of stable property mentioned in \cite{XD2018}. Further it can be found in \cite{Pengbo} that the MSD of ordinary L\'evy walk can be obtained through multiplying $\langle y^2(t)\rangle$ in \eqref{1.30} with $2/(2-\alpha)$ for $\alpha\in(0,1)\cup(1,2)$. Therefore one can conclude that the contracting exponential medium cannot localize superdiffusive or ballistic types of L\'evy walks, and it can only make the diffusion constant become slower comparing to the ordinary case instead of changing the diffusion exponent. The results of \eqref{1.29} and \eqref{1.30} are verified in Fig. \ref{powlawexp}.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.28]{powaexpHpositive}
\includegraphics[scale=0.28]{powaexpHnegativeb}
\includegraphics[scale=0.28]{ypowaexpHpositive}
\includegraphics[scale=0.28]{ypowaexpHnegativeb}
\caption{Numerical simulations of the MSDs of L\'{e}vy walk in a uniformly exponential expansion medium by sampling over $10^4$ realizations. The walking time PDF of L\'{e}vy walk behaves as Pareto distribution with $\tau_0=1$. The parameter $v_0=1$. For (a) and (c), we take $H=10^{-1}$ and the value of $\alpha$ respectively is $\alpha=0.5$ (circles) and $\alpha=1.5$ (stars). The value of H in (b) and (d) is $H=-10^{-1}$ respectively with $\alpha=0.5$ (circles) and $\alpha=1.5$ (stars). The solid lines in (a), (b), (c), (d) respectively are the theoretical results shown in \eqref{1.27} and \eqref{1.28} as well as \eqref{1.29} and \eqref{1.30}.}
\label{powlawexp}
\end{figure}
\subsubsection{Power-law scale factor}
The asymptotic behaviors of the MSDs are studied when the scale factor $a(t)$ is considered to be power-law function \eqref{1.23}. Similarly, after some calculations one can obtain the asymptotic behaviors of the MSDs in comoving coordinates for different categories of $\alpha\in (0,1)\cup(1,2)$ and $\beta$. Specifically, for $0<\alpha<1$, there exists
\begin{equation}\label{1.31}
\langle x^2(t)\rangle\sim
\left\{
\begin{split}
& \frac{(1-\alpha)\left(2+(\alpha-2)\beta\right)t_0^{2\beta}v_0^2}{2(1-\beta)} t^{2-2\beta}, \quad \mbox{if $\beta<1$}, \\
& \alpha(1-\alpha)t_0^2v_0^2\big(\gamma+\ln t\big),\quad \mbox{if $\beta=1$}, \\
& A_1 t^0, \quad\quad\quad\quad\quad\quad\quad\quad ~~ \mbox{if $\beta>1$};
\end{split}
\right.
\end{equation}
on the other hand for $1<\alpha<2$,
\begin{equation}\label{1.32}
\langle x^2(t)\rangle\sim
\left\{
\begin{split}
& C_3 t^{3-\alpha-2\beta}, \quad \mbox{if $\beta<\frac{3-\alpha}{2}$ }, \\
& \frac{\alpha(\alpha-1)t_0^{3-\alpha}\tau_0^{\alpha-1}v_0^2}{2-\alpha} \big(\gamma+\ln(t)\big), \quad\mbox{if $\beta=\frac{3-\alpha}{2}$},\\
& A_2 t^0,\quad \mbox{if $\beta>\frac{3-\alpha}{2}$},
\end{split}
\right.
\end{equation}
where $C_3= \frac{2(\alpha-1)\left(3-\alpha+(\alpha-2)\beta\right)t_0^{2\beta}v_0^2 \tau_0^{\alpha-1}}{(6-5\alpha+\alpha^2)(3-\alpha-2\beta)}$ and $\gamma \approx 0.577216$ represents Euler–Mascheroni constant. Besides, we do not explicitly get the values of the constants $A_1$ and $A_2$ in \eqref{1.31} and \eqref{1.32} but the power-law distributed walking time can only lead to the result of localization.
Similarly, from the relation between comoving and physical coordinates \eqref{1.21}, the MSDs in physical coordinates can be obtained from \eqref{1.31} and \eqref{1.32}. Specifically, for $0<\alpha<1$,
\begin{equation}\label{1.311}
\langle y^2(t)\rangle\sim
\left\{
\begin{split}
& \frac{(1-\alpha)\left(2+(\alpha-2)\beta\right)v_0^2}{2(1-\beta)} t^{2}, \quad \mbox{if $\beta<1$ }, \\
& \alpha(1-\alpha)v_0^2\big(\gamma+\ln(t)\big)t^{2}, \quad \mbox{if $\beta=1$},\\
& A_1\left(\frac{t}{t_0}\right)^{2\beta}, \quad\quad\quad\quad\quad\quad~ \mbox{if $\beta>1$};
\end{split}
\right.
\end{equation}
and for $1<\alpha<2$,
\begin{equation}\label{1.322}
\langle y^2(t)\rangle\sim
\left\{
\begin{split}
& C_3 t^{3-\alpha}, \quad \mbox{if $\beta<\frac{3-\alpha}{2}$ }, \\
& \frac{\alpha(\alpha-1)\tau_0^{\alpha-1}v_0^2\big(\gamma+\ln(t)\big)}{2-\alpha}t^{3-\alpha},\quad \mbox{if $\beta=\frac{3-\alpha}{2}$},\\
& A_2\left(\frac{t}{t_0}\right)^{2\beta}, \quad \mbox{if $\beta>\frac{3-\alpha}{2}$}.
\end{split}
\right.
\end{equation}
The MSDs obtained in both coordinates are verified in Fig. \ref{powlawpowlaw}. As we can observe from \eqref{1.311} and \eqref{1.322}, there exists a critical value of $\beta$ denoted as $\beta_c$ such that when $\beta<\beta_c$ the L\'evy walk in physical coordinate can keep the ballistic diffusion or superdiffusion, whereas for $\beta>\beta_c$ the MSDs in physical coordinates asymptotically behave like $\langle y^2(t)\rangle\sim t^{2\beta}$ for $\alpha\in(0,1) \cup (1,2)$. Therefore, when $\beta>\beta_c$ the expanding medium will play a key role to accelerate the L\'evy walk process in physical coordinate, and when $\beta<\beta_c$ the effect of non-static medium will become weaker so that the L\'evy walk can keep the diffusion exponent. The specific value of $\beta_c$ relies on the categories of $\alpha$; if $\alpha\in (0,1)$ then $\beta_c=1$ and if $\alpha\in (1,2)$ then $\beta_c=(3-\alpha)/2$. When $\beta=\beta_c$, a logarithmic correction of $t$ is needed.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.28]{powapowerlawgammasmall}
\includegraphics[scale=0.28]{powapowerlawgammasmallnegative}
\includegraphics[scale=0.28]{powapowerlawgammalog}
\includegraphics[scale=0.28]{powapowerlawgammaconstant}
\includegraphics[scale=0.28]{ypowapowerlawgammasmall}
\includegraphics[scale=0.28]{ypowapowerlawgammasmallnegative}
\includegraphics[scale=0.28]{ypowapowerlawgammalog}
\includegraphics[scale=0.28]{ypowapowerlawgammaconstant}
\caption{Numerical simulations of the MSDs of L\'{e}vy walk in a uniformly power-law expansion medium by sampling over $10^4$ realizations. The walking time PDF of L\'{e}vy walk and $v_0$ are the same as the ones in Fig. \ref{powlawexp}. And $\alpha=0.5$ (circles) and $\alpha=1.5$ (stars). For $\alpha=0.5$, we respectively take $\beta=0.6$, $\beta=-0.1$, $\beta=1$, and $\beta=1.5$ in (a) (e), (b) (f), (c) (g), and (d) (h); for $\alpha=1.5$, we respectively use $\beta=0.5$, $\beta=-0.1$, $\beta=0.75$, and $\beta=1$ in (a) (e), (b) (f), (c) (g), and (d) (h). The solid lines in (a), (b) and (c), (d) are the theoretical results shown in \eqref{1.31} and \eqref{1.32} and the solid lines in (e), (f) and (g), (h) are the theoretical results shown in \eqref{1.311} and \eqref{1.322}.}
\label{powlawpowlaw}
\end{figure}
\section{Mean First passage time}\label{sec 5}
In this section, we mainly consider the asymptotic behaviors of mean first passage time $\langle t_f\rangle$, which has wide applications in many different areas. The first passage time $t_f$ is a random time when a stochastic process reaches the spatial boundary for the first time. Since the non-static medium causes the physical space always to expand or contract, the spatial boundary also changes correspondingly. Therefore it is reasonable to consider L\'evy walk process moving in a given domain of comoving space. Specifically, we consider the one dimensional L\'evy walk process moving in an interval $[-L,L]$ of comoving space with $L>0$ and absorbing boundary $\pm L$. Additionally, we choose the scale factor $a(t)$ to be power-law function \eqref{1.23}. According to the numerical simulations with respect to different types of PDFs $\phi(\tau)$ of walking time, we present the asymptotic behaviors of $\langle t_f\rangle$ as a function of $L$
When the walking time PDF of L\'{e}vy walk is exponential distribution, i.e., $\phi(\tau)=\lambda e^{-\lambda\tau}$, the asymptotic behavior of the mean first passage time for large $L$ is given by numerical simulation in Fig. \ref{mfptexp}, which matches
\begin{equation}\label{mfpt1}
\langle t_f\rangle\sim L^{-\frac{\lambda}{10}+8\beta+\frac{13}{10}}.
\end{equation}
For L\'{e}vy walk with power-law walking time $\phi(\tau)=\frac{1}{\tau_0}\frac{\alpha}{(1+\tau/\tau_0)^{1+\alpha}}$, when $0<\alpha<1$, through numerical simulation we find
\begin{equation}\label{mfpt2}
\langle t_f\rangle\sim L^{\frac{\alpha}{2}+\frac{3}{2}\beta+\frac{4}{5}}.
\end{equation}
Similarly, when $1<\alpha<2$, by numerical simulation we have
\begin{equation}\label{mfpt3}
\langle t_f\rangle\sim L^{10 \alpha+15 \beta-\frac{t_0}{2}-17}.
\end{equation}
The above results are verified by Fig. \ref{mfptpower}.
\begin{figure}[htbp]
\centering
\includegraphics[width=8cm]{MFPTexpapowerlaw}
\caption{Numerical simulations of MFPT of L\'{e}vy walk in comoving coordinate where the physical space is doing power-law expansion. The results are obtained by sampling over $10^4$ realizations. The walking time PDF of L\'{e}vy walk behaves as exponential distribution $\phi(\tau)=\lambda e^{-\lambda\tau}$. The parameters are $v_0=1$ and $x_0=0$. For circles, the parameters are $\lambda=1$, $\beta=0.1$, and $t_0=1$; for stars, $\lambda=2$, $\beta=0.1$, and $t_0=1$; for squares, $\lambda=1$, $\beta=0.1$, and $t_0=2$; for plus, $\lambda=1$, $\beta=0.2$, and $t_0=1$. The solid lines represent the result of $\langle t_f\rangle \sim L^{-\frac{\lambda}{10}+8\beta+\frac{13}{10}}$.}
\label{mfptexp}
\end{figure}
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.28]{MFPTpowersmallapowerlawt0gamma}
\includegraphics[scale=0.28]{MFPTpowerlargeapowerlawt0gamma}
\caption{Numerical simulations of MFPT of L\'{e}vy walk in comoving coordinate where the physical space is doing the power-law expansion. The results are obtained by sampling over $10^4$ realizations. The walking time PDF of L\'{e}vy walk behaves as power-law distribution $\phi(\tau)=\frac{1}{\tau_0}\frac{\alpha}{(1+\tau/\tau_0)^{1+\alpha}}$. The parameters are $v_0=1$ and $x_0=0$. The panel (a) is with $0<\alpha<1$. For circles, the parameters are $\alpha=0.5$, $\beta=0.1$, and $t_0=1$; for star, $\alpha=0.7$, $\beta=0.1$, and $t_0=1$; for squares, $\alpha=0.5$, $\beta=0.1$, and $t_0=2$; for plus, $\alpha=0.5$, $\beta=0.2$, and $t_0=1$. The panel (b) is with $1<\alpha<2$. For circles, the parameters are $\alpha=1.5$, $\beta=0.5$, and $t_0=1$; for stars, $\alpha=1.6$, $\beta=0.5$, and $t_0=1$; for squares, $\alpha=1.5$, $\beta=0.5$, and $t_0=2$. The solid lines respectively represent the results of $\langle t_f\rangle \sim L^{\frac{\alpha}{2}+\frac{3}{2}\beta+\frac{4}{5}}$ and $\langle t_f\rangle \sim L^{10 \alpha+15 \beta-\frac{t_0}{2}-17}$.}
\label{mfptpower}
\end{figure}
\section{Conclusion}\label{sec 6}
In this paper, we establish the model of L\'evy walk in non-static medium and study the stochastic dynamics based on the model.
We first build the governing equation for the PDF of the position of the particle in comoving coordinate. Further, we utilize Hermite orthogonal polynomials to approach the solution to governing equation, which can also lead to the form of the MSD in comoving coordinate. According to the time dependent scale factor, we build up the relation of the MSD in comoving coordinate with the one in physical coordinate. Next we give some representative examples of scale factors $a(t)$ and walking time density $\phi(\tau)$ to obtain the statistical properties in both coordinates through analytical analyses and numerical simulations.
For L\'evy walk with exponentially distributed walking time, we first consider the case of the scale factor $a(t)$ to be an exponential function. When the Hubble constant $H>0$ indicating an expanding medium, the MSD in comoving coordinate tends to be a constant, leading to exponential growth in physical coordinate. On the contrary for the case of contracting medium with $H<0$, we reach a different conclusion, specifically, the MSDs in comoving and physical coordinates show exponential growth and a constant, respectively. The results also indicate that the non-static medium with exponential scale factor has significant influence on the L\'evy walk whose walking time obeys exponential PDF with average $1/\lambda$. Additionally we simulate the stationary distributions for the processes with $H>0$ in comoving coordinate and $H<0$ in the physical coordinate, it turns out that a transition from unimodal distribution to the bimodal one can be observed with the decreasing of $\lambda$ for the former case or increasing of absolute value of the Hubble constant $|H|$ for the latter case. Next, the kurtosis $K$ which is also an important quantity for stochastic process has been theoretically studied in physical coordinate for $H<0$. According to the simulation, we find $K$ is a function of $H$ and $\lambda$ for sufficiently long time, and the evolutions of $K$ for each variable are analyzed. Then we consider the case that the scale factor is a power-law function; by calculating and analyzing the MSD $\langle y^2(t)\rangle$ in physical coordinate we find that for $\beta<1/2$ there is $\langle y^2(t)\rangle\sim t$ indicating the domination of L\'evy walk process comparing to non-static medium, whereas when $\beta>1/2$ it is $\langle y^2(t)\rangle \sim t^{2\beta}$ indicating that the expending medium dominates the process.
Next for L\'{e}vy walk with a Pareto distributed walking time with $\alpha\in(0,1)\cup (1,2)$, we also consider exponential scale factor first. When $H>0$ so that the medium expands, the constant MSD in comoving coordinate implies exponential growth of the MSD in physical coordinate; and for contracting medium with $H<0$, the MSD in physical coordinate keeps the same diffusion exponential as the ordinary L\'evy walk in static medium, and the contracting medium in this case can only shrink the value of diffusion constant. Therefore for the exponential scale factor, one can conclude that the expanding medium with $H>0$ can significantly affect L\'evy walk process whereas the contracting medium has negligible influence on the diffusion. Finally for the case of power-law scale factor, a critical value $\beta_c$ is found; when $ \beta<\beta_c$ the MSD in physical coordinate keeps the same diffusion exponent as the ordinary case in static medium. For different region of $\alpha$, the critical value is also different, specifically, $\beta_c=1$ and $(3-\alpha)/2$ for $0<\alpha<1$ and $1<\alpha<2$, respectively. Besides, the asymptotic behaviors of mean first passage time in comoving coordinate are also discussed through numerical simulations.
\section*{Acknowledgements}
This work was supported by the National Natural Science Foundation of China under Grant No. 12071195, and the AI and Big Data Funds under Grant No. 2019620005000775.
\begin{appendix}
\section{A brief introduction of Hermite polynomials}\label{Appen_A}
Hermite polynomials are a set of orthogonal polynomials defined on $(-\infty,\infty)$ with weight function $e^{-x^2}$ \cite{hermit_intro}. One way of standardizing the Hermite polynomials is to explicitly give them as
\begin{equation}\label{a1}
H_n(x)=(-1)^n e^{x^2} \frac{d^n}{d x^n} e^{-x^2}.
\end{equation}
Furthermore, its orthogonality can be expressed as
\begin{equation}\label{a2}
\int_{-\infty}^{\infty} H_n(x)H_m(x)e^{-x^2}dx=\sqrt{\pi}2^n n! \delta_{n,m},
\end{equation}
where $\delta_{n,m}$ is the Kronecker delta function.
By Taylor's expansion, there exists
\begin{equation}\label{a3}
H_n(x+y)=\sum_{k=0}^{n}\binom{n}{k} H_k(y) (2 x)^{n-k};
\end{equation}
and the following holds
\begin{equation}\label{a4}
H_n(\gamma x)=\sum_{j=0}^{\lfloor\frac{n}{2}\rfloor} \gamma^{n-2 j} (\gamma^2-1)^j \binom{n}{2 j} \frac{(2 j)!}{j!}H_{n-2 j}(x),
\end{equation}
where $\lfloor\frac{n}{2}\rfloor$ is the biggest integer smaller than $\frac{n}{2}$. The Hermite polynomials evaluated at zero,
called Hermite number, are
\begin{equation}\label{a5}
H_n(0)=
\begin{cases}
0, &\mbox{if $n$ is odd}, \\
(-1)^{\frac{n}{2}} 2^{\frac{n}{2}} (n-1)!!, & \mbox{if $n$ is even}.
\end{cases}
\end{equation}
In particular,
\begin{equation}\label{a6}
H_0(x)=1,\quad H_1(x)=2 x,\quad {\rm and}~~ H_2(x)=4 x^2-2.
\end{equation}
\section{Derivations of \eqref{1.11} and \eqref{1.12}}\label{App_B}
Inserting \eqref{1.9} into \eqref{1..6} leads to
\begin{widetext}
\begin{equation}\label{B1}
\begin{split}
\sum_{n=0}^{\infty} H_n(x)T_n(t) e^{-x^2}&=\frac{1}{2}\int_{0}^{t}\sum_{n=0}^{\infty} H_n\bigg(x-\frac{v_0 \tau}{a(t)} \bigg)
T_n(t-\tau) \exp\Bigg[-\bigg(x-\frac{v_0 \tau}{a(t)}\bigg)^2\Bigg] \phi(\tau) d\tau\\
& +\frac{1}{2}\int_{0}^{t}\sum_{n=0}^{\infty} H_n\left(x+\frac{v_0 \tau}{a(t)}\right)
T_n(t-\tau) \exp\left[-\left(x+\frac{v_0 \tau}{a(t)}\right)^2\right] \phi(\tau) d\tau
+P_0(x)\delta(t).
\end{split}
\end{equation}
Multiplying $H_m(x)$, $m=0,1,2,\cdots,$ on both side of \eqref{B1} and integrating $x$ over $(-\infty,+\infty)$, according to the orthogonal properties of Hermite polynomials \eqref{a2}, the left hand of \eqref{B1} becomes
\begin{equation}\label{B2}
\int_{-\infty}^{+\infty}\sum_{n=0}^{\infty} H_n(x)T_n(t)H_m(x) e^{-x^2} dx=\sqrt{\pi}2^m m! T_m(t).
\end{equation}
We next pay attention to the first term of the right hands of \eqref{B1} in detail. Multiplying $H_m(x)$, $m=0,1,2,\cdots,$ and then integrating $x$ over $(-\infty,+\infty)$ lead to
\begin{equation}\label{B3}
\begin{split}
&\frac{1}{2}\int_{0}^{t} \phi(\tau) T_n(t-\tau) \int_{-\infty}^{+\infty} H_m(x)\sum_{n=0}^{\infty} H_n\left(x-\frac{v_0 \tau}{a(t)}\right)
\exp\left[-\left(x-\frac{v_0 \tau}{a(t)}\right)^2\right] dx d\tau \\
=& \frac{1}{2}\int_{0}^{t} \phi(\tau) T_n(t-\tau) \int_{-\infty}^{+\infty} H_m\left(y+\frac{v_0 \tau}{a(t)}\right)\sum_{n=0}^{\infty} H_n(y)
\exp(-y^2) dy d\tau \\
=&\frac{1}{2}\int_{0}^{t} \phi(\tau) T_n(t-\tau) \int_{-\infty}^{+\infty}\sum_{k=0}^{m}\binom{m}{k} H_k(y) \left(\frac{2v_0 \tau}{a(t)}\right)^{m-k}\sum_{n=0}^{\infty} H_n(y)
\exp(-y^2) dy d\tau \\
=&\frac{1}{2}\sum_{k=0}^{m}\binom{m}{k} \int_{0}^{t} \phi(\tau) T_n(t-\tau)\left(\frac{2v_0 \tau}{a(t)}\right)^{m-k} \int_{-\infty}^{+\infty} H_k(y)\sum_{n=0}^{\infty} H_n(y)
\exp(-y^2) dy d\tau \\
=&\frac{1}{2}\sum_{k=0}^{m}\frac{m!}{k!(m-k)!} \int_{0}^{t} \sqrt{\pi}2^k k! \left(\frac{2v_0 \tau}{a(t)}\right)^{m-k}T_k(t-\tau) \phi(\tau) d\tau,
\end{split}
\end{equation}
which is obtained by utilizing the properties of Hermite polynomials \eqref{a2} and \eqref{a3} in Appendix \ref{Appen_A}. The second term can be treated in the similar way and it reduces to $\frac{1}{2}\sum_{k=0}^{m}\frac{m!}{k!(m-k)!} \int_{0}^{t} \sqrt{\pi}2^k k! \left(\frac{-2v_0 \tau}{a(t)}\right)^{m-k}T_k(t-\tau) \phi(\tau) d\tau$. Since the initial distribution of particles is Dirac-delta function, i.e., $P_0(x)=\delta(x)$, the third term becomes $\int_{-\infty}^{+\infty}H_m(x)\delta(x)\delta(t)dx=\delta(t)H_m(0)$.
Finally the iteration relation of ${T_m(t)}$ can be obtained
\begin{equation}\label{B4}
\sqrt{\pi} 2^m m!T_m(t)= \frac{1}{2}\sum_{k=0}^{m}\frac{m!}{k!(m-k)!}\int_{0}^{t}\sqrt{\pi}2^k k!\left[\left(\frac{2 v_0 \tau}{a(t)}\right)^{m-k}+\left(\frac{-2 v_0 \tau}{a(t)}\right)^{m-k}\right]T_k(t-\tau)\phi(\tau)d\tau+\delta(t)H_m(0).
\end{equation}
Similarly, by substituting \eqref{1.10} into \eqref{1..7}, we find the relation between $R_m(t)$ and $T_m(t)$ as
\begin{equation}\label{B5}
\sqrt{\pi} 2^m m!R_m(t)= \frac{1}{2}\sum_{k=0}^{m}\frac{m!}{k!(m-k)!}\int_{0}^{t}\sqrt{\pi}2^k k!\left[\left(\frac{2 v_0 \tau}{a(t)}\right)^{m-k}+\left(\frac{-2 v_0 \tau}{a(t)}\right)^{m-k}\right]T_k(t-\tau)\psi(\tau)d\tau .
\end{equation}
\end{widetext}
Equations \eqref{1.11} and \eqref{1.12} can be got in Laplace space by taking Laplace transform on \eqref{B4} and \eqref{B5} with the help of convolution theorem.
\end{appendix}
|
\section*{Abstract}
Many problems in classification involve huge numbers of irrelevant features. Model selection reveals the crucial features, reduces the dimensionality of feature space, and improves model interpretation. In the support vector machine literature, model selection is achieved by $\ell_1$ penalties. These convex relaxations seriously bias parameter estimates toward 0 and tend to admit too many irrelevant features. The current paper presents an alternative that replaces penalties by sparse-set constraints. Penalties still appear, but serve a different purpose. The proximal distance principle takes a loss function $L(\boldsymbol{\beta})$ and adds the penalty $\frac{\rho}{2}\mathop{\rm dist}\nolimits(\boldsymbol{\beta}, S_k)^2$ capturing the squared Euclidean distance of the parameter vector $\boldsymbol{\beta}$ to the sparsity set $S_k$ where at most $k$ components of $\boldsymbol{\beta}$ are nonzero. If $\boldsymbol{\beta}_\rho$ represents the minimum of the objective $f_\rho(\boldsymbol{\beta})=L(\boldsymbol{\beta})+\frac{\rho}{2}\mathop{\rm dist}\nolimits(\boldsymbol{\beta}, S_k)^2$, then $\boldsymbol{\beta}_\rho$ tends to the constrained minimum of $L(\boldsymbol{\beta})$ over $S_k$ as $\rho$ tends to $\infty$. We derive two closely related algorithms to carry out this strategy. Our simulated and real examples vividly demonstrate how the algorithms achieve much better sparsity without loss of classification power.
\bigskip
\noindent
\textbf{keywords:} sparsity; discriminant analysis; unsupervised learning; Julia
\newpage
\section{Introduction}
Support vector machines (SVMs) are powerful pattern recognition tools \cite{cortes1995supportvector} with a wide range of applications across machine learning and statistics. Success stories in supervised learning include optical character recognition \cite{decoste2002training}, image segmentation \cite{barghout2015spatialtaxon}, text categorization \cite{joachims1998text,pradhan2004shallow}, protein structure prediction \cite{dunbrack2006sequence}, and early detection and classification of human cancers \cite{sewak2007svm}. It is possible to extend SVM to online algorithms for resource-limited computing environments \cite{cauwenberghs2000incremental,laskov2006incremental} and beyond classification to unsupervised learning \cite{ben-hur2002support}.
Many variations on supervised SVMs have been proposed. For example, the classic soft-margin classifier features a $\ell_{2}$ penalty term on model coefficients, $\sum_{j} \beta_{j}^{2}$, that regularizes classification. To induce model selection, this ridge penalty has since been amended to an $\ell_{1}$ penalty \cite{zhu20031norm}. Other loss functions, such as quadratic and Huber hinge errors, have been proposed as alternatives to the typical hinge loss $u_+=\max\{0,u\}$ to promote better prediction and robustness to outliers \cite{groenen2008svmmaj}.
The aim of this work is to present a flexible framework that attacks sparsity directly and generates attractive search algorithms. We focus on primal problems under the squared-hinge loss with extensions to kernel machines for nonlinear data. Our major contribution is to replace sparsity inducing penalties with sparsity constraint sets. This goal is achieved by estimating parameters by an iterative proximal distance algorithm \cite{lange2016mma}. Our proximal distance algorithm comes in two closely related flavors. Both exhibit comparable, and often superior, predictive accuracy to existing SVM approaches. Although our proximal distance algorithms are sometimes slower than competing methods, the output of both is easier to interpret and better at revealing the sparse signals hidden in high-dimensional data.
\section{Motivation}
In binary classification we adopt the squared hinge-loss $\max\{0,1-u\}^{2}$ in the criterion
\begin{equation}
\label{eq:basic-primal-model}
L(\boldsymbol{\beta} \mid \boldsymbol{y}, \boldsymbol{X}) = \frac{1}{2n} \sum_{i=1}^{n} \max\{0, 1-y_{i} \boldsymbol{x}_{i}^{\top} \boldsymbol{\beta} \}^{2}.
\end{equation}
Here the $n$ observations $(y_{i}, \boldsymbol{x}_{i})$ consist of a class label $y_{i} \in \{-1,1\}$ and a feature vector $\boldsymbol{x}_{i} \in \mathbb{R}^{p+1}$.
The parameter vector $\boldsymbol{\beta} \in \mathbb{R}^{p+1}$
defines the hyperplane separating the two classes. The last component $\beta_{p+1}$ of $\boldsymbol{\beta}$ represents the model intercept; the last component $x_{i, p+1}$ of each $\boldsymbol{x}_i$ is accordingly 1. The form of the loss allows for a small margin of error in classification.
Rather than directly minimize $L(\boldsymbol{\beta} \mid \boldsymbol{y}, \boldsymbol{X})$ defined by equation (\ref{eq:basic-primal-model}), we turn to the MM principle \cite{lange2016mma,lange2000optimization} and invoke the quadratic majorization
\[
\max\{0, 1-u\}^{2} \le \begin{cases}
(u_{m} - u)^{2}, & u_{m} \ge 1 \\
(1 - u)^{2}, & u_{m} < 1
\end{cases}
\]
at iteration $m$ suggested by Groenen et al \cite{groenen2008svmmaj}. Note the two sides of the majorization agree when $u=u_m$. Given that all $y_{i}^{2} = 1$, the term by term application of the majorization creates the overall quadratic surrogate
\begin{equation}
\label{eq:basic-surrogate}
g(\boldsymbol{\beta} \mid \boldsymbol{\beta}_{m})
= \frac{1}{2n} \| \boldsymbol{z}_{m} - \boldsymbol{X} \boldsymbol{\beta} \|^{2},
\qquad
\boldsymbol{z}_{mi} = \begin{cases}
\boldsymbol{x}_{i}^{\top} \boldsymbol{\beta}_{m} & y_{i} \boldsymbol{x}_{i}^{\top} \boldsymbol{\beta}_{m} \ge 1 \\
y_{i} & y_{i} \boldsymbol{x}_{i}^{\top} \boldsymbol{\beta}_{m} < 1
\end{cases}.
\end{equation}
This maneuver reduces the original minimization problem to sequence of easier minimization problems that can be solved by iteratively-reweighted least squares (IRLS); specifically, $\boldsymbol{\beta}_{m+1} = \mathop{\rm argmin}\nolimits_{\boldsymbol{\beta}}~\frac{1}{2n} \| \boldsymbol{z}_{m} - \boldsymbol{X} \boldsymbol{\beta} \|^{2}$. The MM principle implies that
every iteration decreases the objective function (\ref{objective_function}). Nguyen and McLachlan also apply the MM principle to support vector machines and make the connection to IRLS, albeit with a different majorization \cite{nguyen2017iterativelyreweighted}. The surrogate (\ref{eq:basic-surrogate}) is nicer because all weights in IRLS remain 1.
\section{Sparse SVMs via Distance Penalization}
The convex surrogate (\ref{eq:basic-surrogate}) does not admit a unique solution in general unless its Hessian $\boldsymbol{X}^{\top} \boldsymbol{X}$ is positive definite. The standard approach in the literature augments the (squared) hinge-loss (\ref{eq:basic-primal-model}) with the ridge penalty $\lambda \sum_{j=1}^{p} \beta_{j}^{2}$, omitting the intercept. Putting $\lambda > 0$ (a) induces an upper bound of $\lambda^{-1}$ on the margin separating distinct classes in the original problem, (b) enforces strict convexity, and (c) guarantees a unique minimum of both the objective and surrogate. Other regularizations are tractable and achieve the same goal \cite{groenen2008svmmaj}.
Here we focus on inducing feature selection in soft-margin classifiers. In brief, we impose a set constraint on the loss (\ref{eq:basic-primal-model}) that directly controls sparsity in $\boldsymbol{\beta}$. The set $S_{k} \subset \mathbb{R}^{p+1}$ determined by the hyperparameter $k$ limits $\boldsymbol{\beta}$ to have at most at most $k$ nonzero values among its first $p$ components. Fortunately, projection onto $S_{k}$ is straightforward, and the hyperparameter choices $k \in \{0,1,2,\ldots,p\}$ are finite in number. The projection operator $P_{S_k}(\boldsymbol{\beta})$ sets to zero all but the largest $k$ entries in magnitude of $\beta_1$ through $\beta_p$. This goal can be achieved efficiently by a partial sort of these entries. The intercept is ignored by projection, and the value of $k$ may be chosen by cross-validation.
Following the penalty method of constrained optimization \cite{beltrami1970algorithmica,courant1943variationala}, we minimize the continuously differentiable but unconstrained objective
\begin{equation}
f_\rho(\boldsymbol{\beta}) = L(\boldsymbol{\beta} \mid \boldsymbol{y}, \boldsymbol{X}) + \frac{\rho}{2(p-k+1)} \mathop{\rm dist}\nolimits(\boldsymbol{\beta}, S_{k})^{2} \label{objective_function}
\end{equation}
for a large value of the annealing parameter $\rho \ge 0$. The distance penalty enforces near sparsity. In the limit as $\rho$ tends to $\infty$, the solution vector $\boldsymbol{\beta}_\rho$ tends to the solution of the constrained problem. We attach the prefactor $(p-k+1)^{-1}$ to account for the number of nonzero entries in the difference $\boldsymbol{\beta} - P_{S_{k}}(\boldsymbol{\beta})$. This choice facilitates comparisons between different sparsity levels $k$. Combining our previous majorization (\ref{eq:basic-surrogate}) with distance majorization
\[
\mathop{\rm dist}\nolimits(\boldsymbol{u}, S) \le \|\boldsymbol{u} - P_{S}(\boldsymbol{u}_{m})\|^{2},
\]
yields the sum of squares surrogate
\begin{equation}
\label{eq:generic-surrogate}
g_{\rho}(\boldsymbol{\beta} \mid \boldsymbol{\beta}_{m})
=
\frac{1}{2}
\left\|
\begin{bmatrix}
a \boldsymbol{z}_{m} \\
b \boldsymbol{p}_{m}
\end{bmatrix}
-
\begin{bmatrix}
a \boldsymbol{X} \\
b \boldsymbol{D}
\end{bmatrix} \boldsymbol{\beta}
\right\|_{2}^{2}
\end{equation}
with the choices
\[
a = \frac{1}{\sqrt{n}}, \quad
b = \sqrt{ \frac{\rho}{p-k+1} }, \quad
\boldsymbol{p}_{m} = P_{S_{k}}(\boldsymbol{\beta}_{m}), \quad
\boldsymbol{D} = \boldsymbol{I}_{p \times p}.
\]
In this form it is clear that our model formulation contains the $\ell_{2}$ penalty as a special case; simply substitute $\boldsymbol{p}_{m} = \boldsymbol{0}$ and $b = \sqrt{\lambda}$. Moreover, the MM principle restores differentiability and permits the use of exotic regularization penalties with computable projection operators. Although convexity is lost, sparsity constraints permit identification of features driving a classifier's decision boundary. Our previous experience \cite{keys2019proximala, landeros2020extensions,xu2017generalizeda} supports the value of the proximal distance principle in building parsimonious models with non-convex set constraints.
While soft-margin classifiers are known to perform decently even on data that are not linearly separable, they do not generalize well to inherently nonlinear data. Instead, one can invoke the``kernel trick'' via the representation $\boldsymbol{\beta} = \sum_{i} y_{i} \alpha_{i} \phi(\boldsymbol{x}_{i})$ with possibly infinite dimensional feature map $\phi(\boldsymbol{x})$. Substituting this representation into equation (\ref{eq:basic-primal-model}) yields the loss
\[
L(\boldsymbol{\alpha} \mid \boldsymbol{y}, \boldsymbol{X}) = \frac{1}{2n} \sum_{i=1}^{n} \max\Big\{0, 1
-y_{i} \sum_{j=1}^{n} y_{j} \alpha_{j} K_{ij}\Big\}^{2}.
\]
Applying our previous majorization yields the same surrogate (\ref{eq:basic-surrogate}) but with $\boldsymbol{X}$ replaced by the matrix $\boldsymbol{K} \boldsymbol{Y}$ and $\boldsymbol{\beta}$ replaced by the vector $\boldsymbol{\alpha} \in \mathbb{R}^{n}$. Here $K_{ij} = \langle \phi(\boldsymbol{x}_{i}), \phi(\boldsymbol{x}_{j}) \rangle$ defines a positive definite kernel and $\boldsymbol{Y} = \mathrm{Diag}(\boldsymbol{y})$. Unfortunately, feature selection is generally impossible under this representation. Sparsity constraints instead become a mechanism for controlling the number of support vectors (representative samples). We focus our attention on the $L(\boldsymbol{\beta})$ model rather than $L(\boldsymbol{\alpha})$ model, with the understanding that one can readily pass to the kernel version as needed.
\section{Algorithms}
The gradient of the surrogate (\ref{eq:basic-surrogate}) reads
\begin{equation}
\label{eq:surrogate-gradient}
\nabla g_{\rho}(\boldsymbol{\beta} \mid \boldsymbol{\beta}_{m}) =
(a^{2} \boldsymbol{X}^{\top} \boldsymbol{X} + b^{2} \boldsymbol{I}) \boldsymbol{\beta}
- a^{2} \boldsymbol{X}^{\top} \boldsymbol{z}_{m} - b^{2} \boldsymbol{p}_{m}.
\end{equation}
Taking the thin SVD $\boldsymbol{X} = \boldsymbol{U} \boldsymbol{\Sigma} \boldsymbol{V}^{\top}$ of $\boldsymbol{X}$ enables one to directly solve the stationary equation $\nabla g_{\rho}(\boldsymbol{\beta} \mid \boldsymbol{\beta}_{m}) = \boldsymbol{0}$.
In the appendix we derive the iteration map
\begin{eqnarray*}
\boldsymbol{\beta}_{m+1}
&=& \boldsymbol{p}_{m} + a^{2} \boldsymbol{V} (a^{2} \boldsymbol{\Sigma}^{2} + b^{2} \boldsymbol{I})^{-1} \left(
\boldsymbol{\Sigma} \boldsymbol{U}^{\top} \boldsymbol{z}_{m} - \boldsymbol{\Sigma}^{2} \boldsymbol{V}^{\top} \boldsymbol{p}_{m}
\right) \\
&=& \boldsymbol{p}_{m} + \sum_{j=1}^{r}
\frac{ a^{2} }{ a^{2} s_{j}^{2} + b^{2} } \left[
s_{j} \boldsymbol{u}_{j}^{\top} \boldsymbol{z}_{m} - s_{j}^{2} \boldsymbol{v}_{j}^{\top} \boldsymbol{p}_{m}
\right] \boldsymbol{v}_{j}.
\end{eqnarray*}
The second line of this MM update can be implemented as described in Algorithm \ref{alg:MM}. Each update step requires one copy to initialize $\boldsymbol{\beta}_{m+1}$ as $\boldsymbol{p}_m$ followed by two dot products (sizes $n$ and $p+1$) and one BLAS \texttt{AXPY} per singular value. The time complexity of each iteration therefore scales as $\mathcal{O}[2nr + (2r+1)(p+1)]$. Alternatively, one may implement the updates based on the first line above using BLAS Level 2 operations at the cost of extra memory allocations.
Let us highlight a few of the merits of the MM algorithm. The Gram matrix $\boldsymbol{X}^\top\boldsymbol{X}$ is never formed. The row and column dimensions of $\boldsymbol{X}$ are decoupled so one can expect favorable performance for many high-dimensional problems. Fortunately, the expensive SVD of $\boldsymbol{X}$ need only be extracted once and applies across all values of $\rho$ and all sparsity levels $k$. This allows one to solve each IRLS subproblem quickly. As already noted, the MM algorithm leads to a steady decrease in the objective $f_\rho(\boldsymbol{\beta})$ for $\rho$ fixed. The shrewd reader may observe that the initial value $\boldsymbol{\beta}_{0}$, the choice of annealing schedule sending $\rho \to \infty$, and the non-uniqueness of the projection operator may impact the the quality of the ultimate solution. The second of these three concerns is the most worrisome. Non-uniqueness almost never occurs, and gentle annealing usually weaves its way to a good solution. On the other hand, too gentle annealing wastes time.
\begin{algorithm}[htp]
\small
\caption{MM}
\label{alg:MM}
\begin{algorithmic}[1]
\Require $\boldsymbol{U}, \boldsymbol{V}, \boldsymbol{s}$ such that $\boldsymbol{X} = \boldsymbol{U} \mathrm{Diag}(\boldsymbol{s}) \boldsymbol{V}^{\top}$
\State Set $a^{2} \gets 1/n$ and $b^{2} \gets \rho / (p-k+1)$.
\While {not converged}
\State Set $z_{mi} \gets
\boldsymbol{x}_{i}^{\top} \boldsymbol{\beta}_{m} \: \text{if} \: y_{i} \boldsymbol{x}_{i}^{\top} \boldsymbol{\beta}_{m} \ge 1 \: \text{and} \: y_i \: \text{otherwise}$
for $i = 1,2,\ldots,n$.
\State Project $\boldsymbol{p}_{m} \gets P_{S_{k}}( \boldsymbol{\beta}_{m} )$.
\State Copy $\boldsymbol{\beta}_{m+1} \gets \boldsymbol{p}_{m}$.
\For {$j = 1,2,\ldots,r$}
\State Evaluate $uz = \boldsymbol{u}_{j}^{\top} \boldsymbol{z}_{m}$ and $vp = \boldsymbol{v}_{j}^{\top} \boldsymbol{p}_{m}$.
\State Compute $c_{1} = \frac{ a^{2} s_{j} }{ a^{2} s_{j}^{2} + b^{2} }$,
$c_{2} = \frac{ a^{2} s_j^{2} }{ a^{2} s_{j}^{2} + b^{2} }$, and
$c_{3} = c_{1} uz - c_{2} vp$.
\State Accumulate $\boldsymbol{\beta}_{m+1} \gets \boldsymbol{\beta}_{m+1} + c_{3} \boldsymbol{v}_{j}$.
\EndFor
\EndWhile
\end{algorithmic}
\end{algorithm}
Alternatively, the tangency condition $\nabla g_{\rho}(\boldsymbol{\beta}_{m} \mid \boldsymbol{\beta}_{m}) = \nabla f_\rho(\boldsymbol{\beta}_m)$ suggests implementing gradient descent. Minimizing the surrogate $g_{\rho}(\boldsymbol{\beta} \mid\boldsymbol{\beta}_{m})$ in the direction $-\nabla f_\rho(\boldsymbol{\beta}_m)$ leads to the update
\begin{eqnarray*}
\boldsymbol{\beta}_{m+1} & = & \boldsymbol{\beta}_{m} - t_{m}
\nabla f(\boldsymbol{\beta}_m) \\
t_{m} & = & \frac{\|\nabla f_\rho(\boldsymbol{\beta}_m)\|^{2} }{
a^{2} \|\boldsymbol{X} \nabla f_\rho(\boldsymbol{\beta}_m)\|^{2} + b^{2} \|\nabla f_\rho(\boldsymbol{\beta}_m)\|^2}
\end{eqnarray*}
derived in the appendix. The fact that the surrogate is quadratic and strictly convex in $\boldsymbol{\beta}$ implies that $\boldsymbol{\beta}_{m+1}$ furnishes its exact minimum along the steepest descent direction. Once again, minimizing the surrogate forces a decrease in $f_\rho(\boldsymbol{\beta})$. In practice, it is helpful to add a small perturbation to the denominator of $t_{m}$ to guard against the indeterminate form $0 / 0$. The convergence properties of both the MM and SD algorithms are shared by more general proximal distance algorithms and sketched in the references \cite{keys2019proximala,lange2021nonconvex}.
Algorithm \ref{alg:SD} summarizes the SD algorithm in pseudocode.
Our steepest descent algorithm has several advantages. First, the updates of Algorithm \ref{alg:SD} are simpler than those of its sibling Algorithm \ref{alg:MM}. Second, changing the hyperparameters $\rho$ and $k$ imposes no additional overhead. Third, gradient descent makes little memory demand beyond storing input data and model coefficients. Fourth, the quadratic surrogate relies on an exact line search that implicitly incorporates curvature information from the objective. Fifth, the method is simple to implement and can run on both CPUs and GPUs.
\begin{algorithm}[htp]
\small
\caption{SD}
\label{alg:SD}
\begin{algorithmic}[1]
\State Set $a^{2} \gets 1/n$ and $b^{2} \gets \rho / (p-k+1)$.
\While {not converged}
\State Set $v_{i} = - a^{2} y_{i} \max\{0, 1-y_{i} \boldsymbol{x}_{i}^{\top} \boldsymbol{\beta}_{m}\}$ for $i = 1,2,\ldots,n$.
\State Project $\boldsymbol{p}_{m} \gets P_{S_{k}}( \boldsymbol{\beta}_{m} )$.
\State Evaluate $\nabla f_\rho(\boldsymbol{\beta}_{m}) \gets b^{2} [ \boldsymbol{\beta}_{m} - \boldsymbol{p}_{m} ]
+ \boldsymbol{X}^{\top} \boldsymbol{v}$.
\State Compute $t_{m}$ as described.
\State Iterate $\boldsymbol{\beta}_{m+1} \gets \boldsymbol{\beta}_{m} - t_m \nabla f_\rho(\boldsymbol{\beta}_{m})$.
\EndWhile
\end{algorithmic}
\end{algorithm}
We conclude this section by pointing out that both proximal distance algorithms benefit from Nesterov acceleration \cite{nesterov1983method}.
Starting with $j=1$, each update $\boldsymbol{\beta}_{m+1}$ can be potentially be replaced according to the rule
\[
\boldsymbol{\beta}_{m+1} \mapsto
\boldsymbol{\beta}_{m+1} + \frac{j-1}{j+r-1} [\boldsymbol{\beta}_{m+1} - \boldsymbol{\beta}_{m}],
\]
where $r=3$ is the preferred choice \cite{su2014differential}. Because Nesterov
acceleration does not preserve the descent property, we omit it in the early iterations of the MM and SD algorithms. This actions lets the algorithms hit a neighborhood of the optimal point before we aggressively accelerate.
\section{Numerical Experiments}
Here we apply our algorithms to supervised classification involving our own synthetic examples and datasets from the UCI Machine Learning Repository \cite{dua2019uci}.
The chosen datasets, which address both overdetermined and underdetermined problems, are listed in Table \ref{tab:1}.
\begin{table}[htp]
\centering
\scriptsize
\input{tables/Table1.tex}
\caption{
\label{tab:1}
Summary of datasets for numerical experiments.
Data are standardized to have mean 0 and variance 1 unless otherwise stated. The symbol $\ast$ indicates data are normalized using the min-max transformation; $\dagger$ indicates data are not transformed.
}
\end{table}
\noindent Let us briefly elaborate on our two synthetic examples:
\begin{itemize}
\item[(a)] \texttt{synthetic}: Targets are simulated as $y_i = \mathop{\rm sgn}\nolimits{\boldsymbol{x}_i^\top \boldsymbol{\beta}_{0}}$ with $\boldsymbol{x}_i$ a sample from $\mathcal{N}(\boldsymbol{0}, \boldsymbol{\Sigma})$. There are two causal predictors $\beta_{2} = -\beta_{1} = 10$; otherwise $\beta_{j} = 0$.
The covariance structure of $\boldsymbol{x}_i$ is chosen so that only the two informative features $(x_{1}, x_{2})$ are strongly (positively) correlated. All remaining pairs $(x_{i}, x_{j})$ are uncorrelated.
\item[(b)] \texttt{spiral}: The data consist of three noisy spirals with 600, 300, and 100 samples in each class and variances 0.01, 0.04, and 0.09, respectively.
These data are highly nonlinear, so we evaluate our algorithms using a Gaussian kernel version of SVM classification.
\end{itemize}
We use the standard decision rule $\boldsymbol{x} \mapsto \mathop{\rm sgn}\nolimits (\boldsymbol{x}^{\top} \boldsymbol{\beta}$) for binary classification. In examples with $c > 2$ classes, we train $\binom{c}{2}$ SVMs under the one-versus-one paradigm \cite{hsu2002comparison}. Briefly, if classes are labeled $\{1, 2, \ldots, c\}$, the decision rule is $\boldsymbol{x} \mapsto \arg\max v_{j}$, where $v_{j}$ is the number of votes cast for class $j$ when class $j$ is compared to every other class $i$.
Algorithms \ref{alg:MM} and \ref{alg:SD} are referred to as MM and SD, respectively. Our implementations of MM and SD use Nesterov acceleration with restarts. In both cases, the algorithm is run with fixed $\rho$ until the following convergence criterion is met
\[
\|\nabla f_{\rho}(\boldsymbol{\beta}_{m+1})\|^{2} < \epsilon
\quad \text{or} \quad
m+1 > m_{\text{inner}}.
\]
where $f_\rho(\boldsymbol{\beta})$ is the penalized objective (\ref{objective_function}).
We fix $\epsilon = 10^{-6}$ and allow a maximum $m_{\text{inner}} =10^{4}$ or $10^{5}$ iterations across all examples. This rule ensures that an algorithm makes sufficient progress towards minimizing $f_\rho(\boldsymbol{\beta})$. In practice, we observe that both algorithms require far fewer than $10^5$ iterations on a given subproblem. Starting from $\rho=1$, we update $\rho \to M \times \rho$ for some multiplier $M>1$ every $i_{\text{outer}}=100$ outer iterations; thus, the maximum value of $\rho$ is $M^{100}$. Taking $d_{i} = (n-k+1)^{-1} \mathop{\rm dist}\nolimits(\boldsymbol{\beta}_{i}, S_{k})^{2}$ as the distance at outer iteration $i$, each algorithm halts before all $100$ outer whenever
\[
d_{i} \le 10^{-6}
\quad \text{or} \quad
|d_{i} - d_{i-1}| < 10^{-6} (1 + d_{i-1}).
\]
This ensures solutions are close to the chosen constraint set. These heuristic stopping rules strike a balance between delivering accurate solutions and avoiding expensive iterations that make little progress in minimization. Taking all
100 outer iterations indicates a poor annealing schedule or slow convergence. The general procedure is summarized in Algorithm \ref{alg:pdi}.
\begin{algorithm}[htp]
\scriptsize
\caption{Proximal Distance Iteration}
\label{alg:pdi}
\begin{algorithmic}[1]
\State Set tolerance $\epsilon$, multiplier $a$, and sparsity level $s = 1-k/n$.
\State Set outer iterations $i_{\mathrm{outer}}$ and inner iterations $m_{\mathrm{inner}}$.
\State Initialize $\rho$, $\boldsymbol{\beta}$, and $d_{0} \gets (n-k+1)^{-1}\mathop{\rm dist}\nolimits(\boldsymbol{\beta}, S_{k})^{2}$.
\For {$i \gets 1,i_{\mathrm{outer}}$}
\For {$m \gets 1,m_{\mathrm{inner}}$}
\State Solve the subproblem $\boldsymbol{\beta}_{m+1} \gets \mathop{\rm argmin}\nolimits_{\boldsymbol{\beta}} g_{\rho}(\boldsymbol{\beta} \mid \boldsymbol{\beta}_{m})$
\If {$\|\nabla f_{\rho}(\boldsymbol{\beta}_{m+1})\|^{2} < \epsilon$}
\State Break.
\EndIf
\EndFor
\State Update $\boldsymbol{\beta} \gets \boldsymbol{\beta}_{m}$.
\State Set $d_{i} \gets (n-k+1)^{-1}\mathop{\rm dist}\nolimits(\boldsymbol{\beta}, S_{k})^{2}$.
\If {$d_{i} < 10^{-6}$ OR $|d_{i} - d_{i-1}| < 10^{-6}[1 + d_{i-1}]$}
\State Break.
\Else
\State $\rho \gets a \times \rho$
\EndIf
\EndFor
\State Project final estimate $\boldsymbol{\beta} \gets P_{S_{k}}(\boldsymbol{\beta})$.
\end{algorithmic}
\end{algorithm}
\subsection{Implementation}
We implement our algorithms and numerical experiments in the Julia programming language \cite{Julia-2017}. Our software package calls DataDeps.jl to automatically download data sets and run any preprocessing \cite{DataDeps.jl-2018}. All of our experiments were carried out on a Manjaro Linux desktop machine equipped with an Intel i9-10900KF @ 4.9 GHz (using 8 cores) and 32 GB RAM @ 3600 MHz. Intel's BLAS implementation, MKL, was used for all linear algebra subroutines.
\subsection{Sensitivity to Initial Conditions}
Sparsity sets $S_{k}$ are closed but not convex, so imposing sparsity constraints on the original minimization problem $\min_{\boldsymbol{\beta}} L(\boldsymbol{\beta})$ destroys theoretical guarantees afforded by convexity. Thus, solutions generated by both MM and SD are potentially sensitive to the initial guess $\boldsymbol{\beta}_{0}$. To address this concern, we evaluate our proximal distance algorithms on the basis of (a) the number of iterations taken, (b) the final value of the penalized objective $f_{\rho}(\boldsymbol{\beta})$, (c) the squared gradient norm $\|\nabla f_{\rho}(\boldsymbol{\beta})\|^{2}$, (d) distance to the constraint set, (e) the number of support vectors, and (f) prediction accuracy on a test set. For each data set, we trained each algorithm 100 times, randomly sampling components of $\boldsymbol{\beta}_{0}$ from the standard normal distribution. The sparsity level was fixed such that solutions were at least 50\% sparse. Results are summarized in Figure \ref{fig:1}.
\begin{figure}[htp]
\centering
\includegraphics[width=\textwidth]{Fig1.png}
\caption{
Sensitivity of algorithms to initial conditions across data sets.
\label{fig:1}
}
\end{figure}
While there is variation across all four evaluation metrics, we do not observe substantial outliers.
The most variable metric is test accuracy, while the remaining metrics are relatively stable.
Interestingly, there is no clear winner between MM and SD. In the remaining experiments we initialize $\boldsymbol{\beta}$ based on univariate regression with $y_{i} \sim x_{ij} \beta_{j} + \beta_{p+1}$; that is
\begin{equation}
\label{eq:initheuristic}
\beta_{j} = \frac{ \sum_{i=1}^{n} (x_{ij} - \bar{x}_{j}) (y_{i} - \bar{y}) }{ \sum_{i=1}^{n} (x_{ij} - \bar{x}_{j})^{2} },
\quad
j = 1,2,\ldots,p;
\qquad
\beta_{p+1} = \bar{y}.
\end{equation}
\subsection{Prediction and Model Selection}
Our motivation for incorporating sparsity in the model (\ref{eq:basic-primal-model}) is to build feature selection into SVMs directly. Thus, we are interested in studying both predictive accuracy and estimation capability. Specifically, we claim that both MM and SD are capable of recovering a sparse model $\boldsymbol{\beta}$
with reasonable training data $(y_{i}, \boldsymbol{x}_{i})$. These advances improve SVM's predictive power and our understanding of the critical features of the data.
To address these issues, we simulate $(\boldsymbol{y}, \boldsymbol{X}, \boldsymbol{\beta})$ as follows:
\begin{itemize}
\item[(a)] The number of true causal features is fixed at $k_{0} = 50$.
\item[(b)] Feature vectors $\boldsymbol{x}_{i}$ are drawn independently from the multivariate normal distribution $\mathcal{N}(\boldsymbol{0}, \boldsymbol{I})$.
\item[(c)] Model parameters corresponding to a causal features, $\beta_{j}$, are selected uniformly from the union $[-10,-2] \cup [2, 10]$. The remaining components are set to 0.
\item[(d)] Labels $y_{i}$ are simulated under the binary decision rule $\mathop{\rm sgn}\nolimits(\boldsymbol{x}_i^\top\boldsymbol{\beta})$.
\item[(e)] The numbers of samples $n$ and features $p$ are varied to cover different extremes of high-dimensional data.
\end{itemize}
In each problem instance, we determine predictive accuracy from the proportion of correctly classified samples in a test set and assess a binary classifier's ability to select informative features using false discovery and omission rates (FDR and FOR). For completeness, the formulas
\begin{eqnarray*}
FDR &= \frac{(1-SPC) \times (1-q)}{(1-SPC) \times (1-q) + SEN \times q} \\
FOR &= \frac{(1-SEN) \times q}{(1-SEN) \times q + SPC \times (1-q)}
\end{eqnarray*}
yield the desired quantities, where sensitivity and specificity (SEN and SPC, respectively) are determined by comparing the fitted coefficients $\hat{\boldsymbol{\beta}}$ against the true coefficients $\boldsymbol{\beta}$.
Here $q = k_{0} / n$, the proportion of causal features, adjusts both FDR and FOR to allow comparisons under different ground truths for sparsity. Note that we do not use warm starts in solving for a specific sparsity level.
Figure \ref{fig:2} summarizes our results.
\begin{figure}[htp]
\centering
\includegraphics[width=0.5\textwidth]{Fig2A.png}%
\includegraphics[width=0.5\textwidth]{Fig2B.png}
\caption{
Predictive and estimation capabilities of algorithms MM (blue) and SD (orange) across high-dimensional scenarios.
The black vertical line in the left panels highlights the true sparsity level, $100 \times (1-k_{0} / n)$.
In the right-hand panels, we report representative results based on the true sparsity level to demonstrate scaling behavior in an ideal regime.
\label{fig:2}
}
\end{figure}
Both MM and SD have similar performance characteristics as feature or sample dimensions increase. In general, our algorithms minimize FOR and FDR near the true sparsity level, and both error metrics increase if one insists on too simple of a model. FOR is consistently low, demonstrating that the methods are good at eliminating uninformative features, especially as the number of training samples increases.
However, FDR may be inflated in the underdetermined regime $(n=500, p=1000)$ if data are too noisy. We attribute this result to the large number of uncorrelated features $\boldsymbol{x}_{i}$ simulated. Another potential issue is that $\boldsymbol{\beta}$ is initialized with the heuristic (\ref{eq:initheuristic}) rather than reusing previous solutions as warm starts. We consider warm starts in the next experiments.
\subsection{Cross-Validation}
Without prior knowledge of the true number of causal features, one must tune the sparsity level of the solutions. Here we study compatibility of $10$-fold cross-validation of our algorithms with warm starts.
Specifically, in each fold we initialize $\boldsymbol{\beta}$ with the heuristic (\ref{eq:initheuristic}) and then solve (\ref{eq:basic-primal-model}) without any sparsity constraints. Then solution sparsity $s = 1-k/n$ is slowly increased from fully dense (0\%) to fully sparse (100\%) by solving the distance penalized version of the basic soft-margin model. Let us briefly clarify our terminology:
\begin{itemize}
\item[(a)] A training set (Tr) is used to fit model coefficients in each fold.
\item[(b)] A validation set (V) is used to tune the hyperparameter $k$ (or equivalently, $s$) based on prediction accuracy within each fold.
\item[(c)] A test set (T) is used to evaluate performance of selected models. This set contains samples never encountered in the training or test sets.
\end{itemize}
We report selected evaluation metrics in the following tables, aggregated as averages over cross-validation folds. Table \ref{tab:2} illustrates performance of MM on the \texttt{synthetic} example in detail, and Table \ref{tab:3} reports the highlights across all 8 examples.
\begin{table}[htp]
\tiny
\centering
\input{tables/Table2.tex}
\caption{
Results for 10-fold cross-validation on \texttt{synthetic} data set using MM at various sparsity levels $s$.
The highlighted row corresponds to results for the true sparsity level.
\label{tab:2}
}
\end{table}
Accuracy across training, validation, and test sets peak around the true sparsity level $s=99.7\%$ in the \texttt{synthetic} example. Table \ref{tab:2} shows that MM can quickly explore different sparsity levels and that distance-to-sparsity sets is a compelling form of regularization for SVMs. In Table \ref{tab:3} we observe that the combination of our proximal distance algorithms with cross-validation can select parsimonious, predictive models across a wide range of examples. The MM and SD algorithms share similar performance characteristics, but the latter tends to deliver solutions faster due to its cheap computational cost per iteration.
However, the two methods diverge in fitting quality solutions on higher dimensional datasets, namely the \texttt{optdigits} ($5620 \times 64$), \texttt{letter-recognition} ($20000 \times 16$), and \texttt{TCGA-PANCAN-HiSeq} ($801 \times 10000$).
Specifically, MM and SD may select models with differing sparsity levels under cross-validation.
The difference stems from slow progress in driving the distance penalty downhill under the SD algorithm as reported in Figure \ref{fig:3}.
Indeed, the \texttt{TCGA-PANCAN-HiSeq} example demonstrates that SD may fail to fit training data under our chosen stopping rules which are designed to limit iterations in the face of slow convergence.
Poor adherence to a specified constraint set $S_{k}$ ultimately leads SD astray in model selection.
\begin{table}[htp]
\tiny
\centering
\input{tables/Table3.tex}
\caption{
Summary of 10-fold cross-validation results. Column V is the average validation set accuracy used to select a model in cross-validation. The selected model is then evaluated on a test set T. Total iteration count and total time are based on solution path traversal from $s = 0\%$ to $s = 100\%$.
\label{tab:3}
}
\end{table}
\begin{figure}[htp]
\centering
\includegraphics[width=0.48\linewidth]{Fig3A.png}%
\includegraphics[width=0.48\linewidth]{Fig3B.png}
\includegraphics[width=0.48\linewidth]{Fig3C.png}%
\includegraphics[width=0.48\linewidth]{Fig3D.png}
\includegraphics[width=0.48\linewidth]{Fig3E.png}%
\includegraphics[width=0.48\linewidth]{Fig3F.png}
\caption{
Convergence metrics (left) and evaluation metrics (right) for Algorithms MM and SD on the \texttt{TCGA-PANCAN-HiSeq} example.
\label{fig:3}
}
\end{figure}
\subsection{Comparison to LIBSVM}
We compare our algorithms to algorithms implemented in the Julia version LIBSVM.jl of LIBSVM \cite{chang2011libsvm} and LIBLINEAR \cite{fan2008liblinear}. Specifically we compare MM and SD to:
\begin{itemize}
\item[(a)] L2R, an option from LIBLINEAR for the standard $\ell_{2}$ regularized version of criterion (\ref{eq:basic-primal-model}).
\item[(b)] L1R, an option from LIBLINEAR for the $\ell_{1}$ regularized version of (\ref{eq:basic-primal-model}).
\item[(c)] SVC, an option from LIBSVM chosen specifically for its nonlinear example \texttt{spiral}. SVC is run using a Gaussian kernel and assumes the standard hinge-loss model.
\end{itemize}
Table \ref{tab:4} reports test set accuracy after 10-fold cross-validation to select the sparsity hyperparameter $(s)$ in our methods and the classifier margin ($C$) in the LIBSVM methods. Both LIBLINEAR and LIBSVM classifiers rely on the OVO paradigm for multiclass assignment. The final column in Table \ref{tab:4} reports the observed sparsity in the final model. If the classification task involves $c>2$ classes, then we average observed sparsity levels of the non-intercept terms in $\boldsymbol{\beta}$ over each of the $\binom{c}{2}$ SVMs. Our SVM training algorithms are comparable to existing approaches, albeit slower in the examples \texttt{TCGA-PANCAN-HiSeq} and \texttt{spiral}. To our credit, the \texttt{synthetic} example underscores the superiority of sparsity constraints over shrinkage-based penalties.
\begin{table}[htp]
\tiny
\centering
\input{tables/Table4.tex}
\caption{
Comparison of our algorithms (MM and SD) against existing approaches from LIBSVM (L2R, L1R, and SVC). Optimal values for each metric are highlighted with bold text.
\label{tab:4}
}
\end{table}
\newpage
\section{Discussion}
We have demonstrated the benefits of conceptually simple proximal distance algorithms for binary and multiclass classificaiton problems on both linear SVMs and nonlinear kernel SVMs. The proximal distance principle makes it possible to attack parsimony directly through squared distance penalties. This direct approach (a) restores differentiability via quadratic surrogate functions, (b) avoids the shrinkage inherent in lasso-based algorithms, (c) identifies sparser models with good predictive power, and (d) substitutes a discrete interpretable sparsity level for the continuous uninterpretable hyperparameters of competing methods. To our surprise, the more expensive Algorithm MM scales better on high-dimensional data due to its ability to quickly drive solutions close to a desired sparsity set.
While we are pleased with our results, particularly for binary classification tasks, much is left to be desired for multiclass problems. Relying on multiple SVMs to handle multiclass problems introduces $\binom{c}{2}$ subproblems for $c$ true classes. Furthermore, different decision boundaries in the OVO paradigm may be driven by different features, obscuring the universal features that discriminate between classes.
Hence, it is natural to investigate multiclass methods beyond hyperplane separation. Our previous research on multivertex discriminant analysis (MVDA) \cite{lange2008mm} explored a multiclass model that represents classes geometrically as vertices of a regular simplex embedded in Euclidean space rather than binary choices from $\{-1,1\}$. MVDA takes advantage of $\epsilon$-insensitive norms and generalizes to nonlinear classification via the kernel trick \cite{wu2010multicategory}. We plan to revisit MVDA and incorporate sparsity based on the proximal distance principle and possibly Huber hinge errors \cite{burg2016gensvm}. Given the length of the current paper and the many unresolved challenges ahead,
this goal is best left to a future paper.
\bibliographystyle{abbrv}
|
\section{\@startsection{section}{1}%
\z@{.7\linespacing\@plus\linespacing}{.5\linespacing}%
{\normalfont\bfseries\scshape\centering}}
\def\subsection{\@startsection{subsection}{2}%
\z@{.5\linespacing\@plus\linespacing}{.5\linespacing}%
{\normalfont\bfseries\scshape}}
\def\subsubsection{\@startsection{subsubsection}{3}%
\z@{.5\linespacing\@plus\linespacing}{-.5em
{\normalfont\bfseries\itshape}}
\definecolor{darkgreen}{rgb}{0,0.4,0}
\definecolor{BrickRed}{rgb}{0.65,0.08,0}
\hypersetup{colorlinks=true,linkcolor=blue,citecolor=red,filecolor=BrickRed,urlcolor=darkgreen}
\linespread{1.05}
\newcommand{\begin{equation}}{\begin{equation}}
\newcommand{\end{equation}}{\end{equation}}
\newcommand{generating function}{generating function}
\newcommand{generating functions}{generating functions}
\newcommand{formal power series}{formal power series}
\newcommand{{\sc Maple}}{{\sc Maple}}
\newcommand{U_1}{U_1}
\newcommand{U_3}{U_3}
\newcommand{U_4}{U_4}
\newcommand{U_0}{U_0}
\newcommand{U_1}{U_1}
\newcommand{U_2}{U_2}
\newcommand{N}{N}
\newcommand{c}{c}
\newcommand{r}{r}
\newcommand{\tilde z}{\tilde z}
\newcommand{Sym}{Sym}
\newcommand{R}{R}
\newcommand{r}{r}
\DeclareMathOperator{\Rat}{Rat}
\DeclareMathOperator{\numsmall}{num}
\DeclareMathOperator{\den}{den}
\newcommand{C}{C}
\newcommand{c}{c}
\newcommand{\overrightarrow{g}}{\overrightarrow{g}}
\DeclareMathOperator\OS{OS}
\newcommand{\ns}{{\mathbb N}}
\newcommand{\zs}{{\mathbb Z}}
\newcommand{\qs}{{\mathbb Q}}
\newcommand{\cs}{{\mathbb C}}
\newcommand{\rs}{{\mathbb R}}
\newcommand{\mathcal{O}}{\mathcal{O}}
\newcommand{o}{o}
\newcommand{\V}{\mathbb{V}}
\newcommand{\mathcal{A}}{\mathcal{A}}
\newcommand{\mathcal{B}}{\mathcal{B}}
\newcommand{\mathcal{C}}{\mathcal{C}}
\newcommand{\mathcal{Q}}{\mathcal{Q}}
\newcommand{\tilde{\mathcal{C}}}{\tilde{\mathcal{C}}}
\newcommand{\mathcal{D}}{\mathcal{D}}
\newcommand{\mathcal{E}}{\mathcal{E}}
\newcommand{\mathcal{F}}{\mathcal{F}}
\newcommand{\mathcal{H}}{\mathcal{H}}
\newcommand{\mathcal{I}}{\mathcal{I}}
\newcommand{\mathcal{K}}{\mathcal{K}}
\newcommand{\mathcal{L}}{\mathcal{L}}
\newcommand{\mathcal{N}}{\mathcal{N}}
\newcommand{\mathcal{R}}{\mathcal{R}}
\newcommand{\mathcal{S}}{\mathcal{S}}
\newcommand{\mathcal{T}}{\mathcal{T}}
\newcommand{\mathcal{U}}{\mathcal{U}}
\newcommand{\mathbb{C}}{\mathbb{C}}
\newcommand{\mathbb{N}}{\mathbb{N}}
\newcommand{\mathbb{Q}}{\mathbb{Q}}
\newcommand{{\mathbb Z}}{{\mathbb Z}}
\newcommand{{\mathbb R}}{{\mathbb R}}
\newcommand{\widetilde Q}{\widetilde Q}
\newcommand{\widetilde C}{\widetilde C}
\newcommand{\KK}{\mathbb{K}}
\newcommand{\varepsilon}{\varepsilon}
\newcommand{ \hfill $_{\blacksquare} $ }{ \hfill $_{\blacksquare} $ }
\newcommand{ \hfill ${\lozenge} $}{ \hfill ${\lozenge} $}
\newtheorem{theo}{Theorem}[section]
\newtheorem{lemma}[theo]{Lemma}
\newtheorem{prop}[theo]{Proposition}
\newtheorem{coro}[theo]{Corollary}
\newtheorem{definition}[theo]{Definition}
\newtheorem{conjecture}[theo]{Conjecture}
\def$\hfill{\Box}${$\hfill{\Box}$}
\theoremstyle{remark}
\newtheorem*{remark}{Remark}
\newtheorem{example}[theo]{Remark}
\newcommand{{\rm sign} }{{\rm sign} }
\newcommand{{\rm id} }{{\rm id} }
\newcommand{\mathbb{A}}{\mathbb{A}}
\newcommand{\mathbb{K}}{\mathbb{K}}
\newcommand{\mathcal S}{\mathcal S}
\newcommand{S}{S}
\newcommand{\bar{x}}{\bar{x}}
\newcommand{\bar{y}}{\bar{y}}
\newcommand{\bar{u}}{\bar{u}}
\newcommand{\bar{v}}{\bar{v}}
\newcommand{\bar 1}{\bar 1}
\newcommand{\uroot}{\zeta}
\newcommand{\bzeta}{\bar\zeta}
\newcommand{a}{a}
\newcommand{u}{u}
\newcommand{v}{v}
\newcommand{w}{w}
\newcommand{B_1}{B_1}
\newcommand{B_2}{B_2}
\newcommand{ \tilde z}{ \tilde z}
\newcommand{\tilde{S}}{\tilde{S}}
\newcommand{\tilde w}{\tilde w}
\newcommand{\hat{S}}{\hat{S}}
\newcommand{\hat{R}}{\hat{R}}
\DeclareMathOperator\Pol{Pol}
\DeclareMathOperator\pol{pol}
\DeclareMathOperator\Alg{Alg}
\def\emm#1,{{\em #1}}
\newcommand*\circled[1]{\tikz[baseline=(char.base)]{
\node[shape=circle,draw,inner sep=1pt] (char) {\footnotesize #1};}}
\newcommand{\mathrel{\rotatebox[origin=c]{-45}{$\hookrightarrow$}}}{\mathrel{\rotatebox[origin=c]{-45}{$\hookrightarrow$}}}
\newcommand{\mathrel{\rotatebox[origin=c]{45}{$\hookrightarrow$}}}{\mathrel{\rotatebox[origin=c]{45}{$\hookrightarrow$}}}
\newcommand{\textcolor{blue}{\hookNEarrow}}{\textcolor{blue}{\mathrel{\rotatebox[origin=c]{45}{$\hookrightarrow$}}}}
\newcommand{\textcolor{blue}{\hookrightarrow}}{\textcolor{blue}{\hookrightarrow}}
\newcommand{\OEIS}[1]{\href{http://oeis.org/#1}{OEIS~#1}}
\newcommand{\OEISs}[1]{\href{http://oeis.org/#1}{#1}}
\begin{document}
\author{Mireille Bousquet-M\'elou \and Michael Wallner}
\thanks{MBM was partially supported by the ANR projects DeRerumNatura (ANR-19-CE40-0018) and Combiné (ANR-19-CE48-0011). MW was supported by an Erwin Schr\"odinger Fellowship and a Stand-Alone Project of the Austrian Science Fund (FWF): J~4162 and P~34142.}
\date{\today}
\title{Walks avoiding a quadrant and the reflection principle}
\begin{abstract}
We continue the enumeration of plane lattice walks
with small steps avoiding the negative quadrant, initiated by the first author in 2016.
We solve in detail a new case, namely the king model where all eight nearest neighbour steps are allowed. The associated generating function\ is proved to be the sum of a simple, explicit D-finite series (related to the number
of walks confined to the first quadrant), and an algebraic one. This was already the case for the two models solved
by the first author in 2016.
The principle of the approach is also the same,
but challenging theoretical and computational difficulties arise as we now handle algebraic series of larger degree.
We expect a similar algebraicity phenomenon to hold for the seven \emm Weyl,\ step sets, which are those for which walks confined to the first quadrant can be counted using the reflection principle. With this paper, this is now proved for three of them. For the remaining four, we predict the D-finite part of the solution, and in three of the four cases, give evidence for the algebraicity of the remaining part.
\end{abstract}
\maketitle
\section{Introduction}
\label{sec:intro}
Over the last two decades, the enumeration of walks in the non-negative quadrant
\[
\mathcal{Q}:= \{ (i,j) : i \geq 0 \text{ and } j \geq 0 \}
\]
has attracted a lot of attention and established its own scientific community with {close to a hundred research papers; see, e.g., \cite{bomi10} and citing papers.
One of its {attractive} features is the diversity of the used tools, such as
algebra on formal power series~\cite{bomi10,Mishna-jcta}, bijective approaches~\cite{Bern07,chyzak-yeats},
computer algebra\cite{BoKa08,KaKoZe08},
complex analysis\cite{raschel-unified,BeBMRa-17},
probability theory\cite{BoRaSa14,DenisovWachtel15}, and difference
Galois theory\cite{DHRS-17}. Most of the attention has focused on walks \emm with small steps,, that is, taking their steps in a fixed subset $\mathcal S$ of $\{-1,0, 1\}^2\setminus{(0,0)}$. For each such step set~$\mathcal S$ (often called a \emm model, henceforth), one considers a trivariate generating function\ $Q(x,y;t)$ defined by
\begin{equation}\label{Q-def}
Q(x,y;t)= \sum_{n \ge 0}\sum_{i,j \in \mathcal{Q}} q_{i,j}(n) x^i y^j t^n,
\end{equation}
where $q_{i,j}(n)$ is the number of quadrant walks with steps in $\mathcal S$, starting from $(0,0)$, ending at $(i,j)$, and having in total $n$ steps. For each $\mathcal S$, one now knows whether and where this series fits in the following classical hierarchy of series:
\[
\hbox{ rational } \subset \hbox{ algebraic } \subset \hbox{ D-finite }
\subset \hbox{ D-algebraic}.
\]
Recall that a series (say $Q(x,y;t)$ in our case) is \emm rational, if it is the ratio of two polynomials, \emm algebraic, if it satisfies a polynomial equation (with coefficients that are polynomials in the variables), \emm D-finite, if it satisfies three \emm linear, differential equations (one in each variable), again with polynomial coefficients, and finally \emm D-algebraic,
if it satisfies three \emm polynomial, differential equations. It has been known since the $1980$s~\cite{gessel-factorization} that the generating function\ of walks confined to a half-plane is algebraic. This explains why $Q(x,y;t)$ is algebraic in some cases, for instance when $\mathcal S=\{\rightarrow, \uparrow, \leftarrow\}$: indeed, confining walks to the first quadrant is then equivalent to confining them to the right half-plane $i\ge 0$. It was shown in~\cite{bomi10} that exactly $79$ (essentially distinct)
quadrant problems with small steps are \emm not, equivalent to any half-plane problem.
One central result in the classification of these $79$ models is that $Q(x,y;t)$ is D-finite if and only if a certain group, which is easy to construct from the step set $\mathcal S$, is finite~\cite{bomi10,BoRaSa14,KuRa12,BoKa08,Mishna-Rechni,melczer-mishna-sing}.
Since any strictly convex closed cone can be deformed into the first quadrant, the enumeration of walks confined to $\mathcal{Q}$ captures all such counting problems (provided we consider all possible step sets $\mathcal S$, not only small steps).
Similarly, any non-convex closed cone in two dimensions can be deformed into the three-quadrant plane
\begin{equation}\label{Ccone-def}
\mathcal{C} := \{ (i,j) : i \geq 0 \text{ or } j \geq 0 \},
\end{equation}
and in 2016, the first author initiated the enumeration of lattice paths confined to $\mathcal{C}$~\cite{Bousquet2016}.
Therein, the two most natural models of walks were studied:
\emm simple walks, with steps in $\{\rightarrow, \uparrow, \leftarrow, \downarrow \}$,
and \emm diagonal walks, with steps in $\{\nearrow, \nwarrow, \swarrow, \searrow \}$.
In both cases, the generating function
\begin{equation}\label{C-def}
C(x,y;t)=\sum_{n\ge 0 } \sum_{i,j\in \mathcal{C}}c_{i,j}(n) x^i y^j t^n
\end{equation}
defined analogously to $Q(x,y;t)$ (see~\eqref{Q-def})
was proved to differ from the series
\begin{equation}\label{Q-combin}
\frac{1}{3}\left( Q\left(x,y;t\right) - \frac{1}{x^2}\,Q\!\left(\frac{1}{x},y;t\right) - \frac{1}{y^2}\,Q\!\left(x,\frac{1}{y};t\right) \right)
\end{equation}
by an \emm algebraic, one. In both cases, the underlying group is finite, hence $Q(x,y;t)$ is D-finite and $C(x,y;t)$ is D-finite as well.
\begin{figure}[hb]
\centering
\scalebox{0.99}{\input{kingwalk3qp}}
\caption{A king walk in the three-quadrant plane $\mathcal{C}$. The associated generating function is D-finite and transcendental (i.e., non-algebraic).}
\label{fig:king}
\end{figure}
It became then natural to explore more three-quadrant problems, in
particular to understand whether the D-finiteness of $C(x,y;t)$ was
again related to the finiteness of the associated group -- at least
for the $74$ three-quadrant problems that are not equivalent to a
half-plane problem; see Section~\ref{sec:interestingmodels}. Using an asymptotic argument, Mustapha quickly proved that the~$51$ three-quadrant problems associated with an infinite group have, like their quadrant counterparts, a non-D-finite solution~\cite{Mustapha2019Walks}. Regarding exact solutions, Raschel and Trotignon obtained in~\cite{RaschelTrotignon2018Avoiding} sophisticated integral expressions for eight step sets. Four of them have a finite group (namely $\{\rightarrow, \uparrow, \leftarrow, \downarrow\}$, $\{\nearrow, \leftarrow, \downarrow \}$, $\{\rightarrow, \uparrow, \swarrow \}$, and $\{\rightarrow, \nearrow, \uparrow, \leftarrow, \swarrow, \downarrow \}$), and these expressions imply that they are D-finite (at least in $x$ and $y$).
In fact, the latter three are now known to be
algebraic~\cite{mbm-tq-kreweras}. The other four have an infinite group
and have been further studied by Dreyfus and Trotignon: one of them is
D-algebraic, the other three are
not~\cite{dreyfus-trotignon}. Furthermore, the remarkable results of
Budd~\cite{Budd2020Winding} and Elvey
Price~\cite{ElveyPrice2020Winding} on the winding number of various
families of plane walks provide explicit D-finite expressions for
several generating functions\ of three-quadrant walks starting and ending close to
the origin, in particular for step sets
$\{\rightarrow,\nearrow, \leftarrow, \swarrow\}$
and
$\{\rightarrow, \nwarrow, \leftarrow, \searrow \}$
in Budd's paper, as well as
$\{\uparrow, \leftarrow, \searrow \}$
and
$\{\rightarrow, \uparrow, \nwarrow, \leftarrow, \downarrow, \searrow\}$
in Elvey Price's\footnote{In both papers, when steps $\nwarrow$ or $\searrow$ are allowed, one includes in the enumeration walks using jumps from $(-1,0)$ to $(0,-1)$ and vice-versa. Such jumps are forbidden in this paper, but we show in the last section that allowing them in king walks does not significantly modify the form of our results.}.
\medskip
\noindent{\bf Main results.}
In this paper we enrich the collection of completely solved cases with
the {\emph{king walks}}, in which all eight nearest neighbour steps $\rightarrow, \nearrow, \uparrow, \nwarrow, \leftarrow, \swarrow,
\downarrow, \searrow $ are allowed; see Figure~\ref{fig:king}. This is again a finite group model, and the series $Q(x,y;t)$ is a well-understood D-finite series~\cite{bomi10}.
Here we determine $C(x,y;t)$, and show that the algebraicity phenomenon of~\cite{Bousquet2016} persists: the series $C(x,y;t)$ differs from the linear combination~\eqref{Q-combin} by an algebraic series, this time of degree $216$.
For the simple and diagonal walks of~\cite{Bousquet2016} this
algebraic series was of degree $72$ ``only''. The generating function \ $C_{i,j}$ of
walks ending at a prescribed position $(i,j)$ differs from a series
of the form $\pm Q_{k,\ell}/3$
by an algebraic series of degree at most $24$ (while
this degree was bounded by $8$ in the two models
of~\cite{Bousquet2016}).
Moreover, we explain why we expect a similar property to hold (with variations on the linear combination~\eqref{Q-combin} of series~$Q(\cdot, \cdot)$)
for the seven models of Table~\ref{tab:weyl}. These are precisely the models for which the quadrant
problem can be solved using the reflection principle~\cite{gessel-zeilberger}, and for this reason we call them \emm Weyl models,. We predict the relevant linear combination of series~$Q$, and we give evidence of the algebraicity phenomenon for the three rightmost models of Table~\ref{tab:weyl}.
However, we also expect the effective solution of these models to be challenging in computational terms, because the relevant algebraic series will most likely have very large degrees.
\begingroup
\setlength{\tabcolsep}{4pt}
\begin{table}
\centering
\begin{tabular}{l|c|c|c|c}
model $\mathcal S$&
\begingroup
\setlength{\tabcolsep}{1pt}
\begin{tabular}{cccc}
\includegraphics[height=0.8cm]{10101010}&\includegraphics[height=0.8cm]{01010101} &\includegraphics[height=0.8cm]{11111111}&\includegraphics[height=0.8cm]{11011101}\\
\small{simple}& \small{diagonal}& \small{king} &\small{diabolo }
\end{tabular}
\endgroup
&
\begin{tabular}{cc}
\includegraphics[height=0.8cm]{10010010}&\includegraphics[height=0.8cm]{10111011} \\
\small{tandem} &\small{double-tandem}
\\
\end{tabular}
&
\begin{tabular}{c}
\includegraphics[height=0.8cm]{00110011} \\
\small{Gouyou-Beauchamps}
\end{tabular}
\\
\hline
group $G$ & \begin{minipage}{3cm}
$(x,y)$,$(\bar x, y)$,\\
$(\bar x, \bar y)$,$(x, \bar y)$
\end{minipage} &
\begin{minipage}{3cm} $(x,y), (\bar x y,y)$,\\
$(\bar x y, \bar x), (\bar y, \bar x)$,\\
$(\bar y, x\bar y ), (x,x\bar y )$
\\
\end{minipage} & \begin{minipage}{3cm}
\mbox{}\\$(x, y), (\bar x y, y),$\\ $(\bar x y, {\bar x}^2 y), (\bar x, {\bar x}^2y)$,\\
$ (\bar x, \bar y), (x\bar y, \bar y),$\\$(x\bar y, x^2\bar y), (x,x^2\bar y )$\\
\end{minipage}\\
new steps &
\includegraphics[height=0.8cm]{10101010}\hskip 3mm\includegraphics[height=0.8cm]{01010101} \hskip 2mm\includegraphics[height=0.8cm]{11111111}\hskip 3mm\includegraphics[height=0.8cm]{11011101}& \includegraphics[height=0.9cm]{A2-steps}& \includegraphics[height=0.8cm]{10101010}\\
\begin{minipage}{2cm}
\vskip -36mm Weyl\\ chambers
\end{minipage}
&\scalebox{0.7}{\input{A1-carre.pdf_t}} & \scalebox{0.7}{\input{A2.pdf_t}} & \scalebox{0.7}{\input{B2.pdf_t}}
\end{tabular}
\vskip 3mm
\caption{The seven Weyl models, with their usual names and groups (defined in Section~\ref{sec:group}). The next rows show how to deform the steps and the plane so that walks in the first quadrant correspond to walks in a Weyl chamber, and walks avoiding the negative quadrant to walks avoiding a Weyl chamber.}
\label{tab:weyl}
\end{table}
\endgroup
\medskip\noindent
{\bf Outline of the paper.}
We begin in Section~\ref{sec:general} with generalities on the enumeration of walks with small steps confined to the three-quadrant cone $\mathcal{C}$, and on the related functional equations. We describe the $74$ (essentially distinct) models of interest -- those that are not equivalent to a half-plane model -- and define the group associated with a model. In Section~\ref{sec:OS} we state and justify our conjecture on the form of $C(x,y;t)$ in the seven Weyl cases. The next three sections are devoted to the solution of the king model in three quadrants: in Section~\ref{sec:KingResults} we state our results in details, and we prove them in Sections~\ref{sec:KingEquation} and~\ref{sec:quadratic}. In Section~\ref{sec:combi}, we give combinatorial proofs, based on the reflection principle, of some identities obtained so far via functional equations, and generalize them to all seven Weyl models. In Section~\ref{sec:final} we conclude with a few comments, in particular about what happens to the king model when one allows steps between $(0,-1)$ and $(-1,0)$.
This paper is the full version of an extended abstract~\cite{mbm-wallner-AofA} that was published in the proceedings of the \emph{Analysis of Algorithms} Conference in 2020.
\section{Enumeration in the three-quadrant plane: basic tools}
\label{sec:general}
Let us begin with some definitions and notation on formal power series.
Let $\mathbb{A}$ be a commutative ring and $x$ an indeterminate. We denote by
$\mathbb{A}[x]$ (resp.~$\mathbb{A}[[x]]$) the ring of polynomials (resp.~formal power series) in $x$
with coefficients in $\mathbb{A}$. If $\mathbb{A}$ is a field, then $\mathbb{A}(x)$ denotes the field
of rational functions in $x$, and $\mathbb{A}((x))$ the
field of Laurent series in $x$, that is, series of the form
$ \sum_{n \ge n_0} a_n x^n$,
with $n_0\in \zs$ and $a_n\in \mathbb{A}$. This notation is generalized to polynomials, fractions, and series in several indeterminates.
The coefficient of $x^n$ in a series $F(x)$ is denoted by
$[x^n]F(x)$. We denote partial derivatives with indices: for instance, for a series $F$ involving the indeterminate $x$, we write $F_x$ for $\partial F/\partial x$.
We denote {with} bars the reciprocals of variables: that is, $\bar{x}=1/x$,
so that $\mathbb{A}[x,\bar{x}]$ is the ring of Laurent polynomials in $x$ with
coefficients in $\mathbb{A}$.
We will often handle series of $\qs(x)((t))$, and consider $\qs(x)$ as a subring of $\qs((x))$ (that is, we expand rational functions in $x$ around $x=0$). For $F(x;t) \in \qs((x))((t))$, of the form
\[
F(x;t)= \sum_{n\ge n_0} t^n \sum_{i \ge i_0(n)} a(n, i) x^i,
\]
the \emm non-negative part of $F$ in, $x$ is the
following series in $t$ and $x$:
\[
[x^{\geq}]F(x;t)= \sum_{n\ge n_0} t^n \sum_{i \ge \max(0, i_0(n))} a(n, i) x^i.
\]
We define analogously the \emm positive part, of $F$, denoted by $[x^>]F$. The \emm negative part, of $F$ is $[x^<]F:=F-[x^\ge ]F$. Observe that, if $F(x;t) \in \qs(x)((t))$, this convention makes the roles of $x$ and $1/x$ non-symmetric: the negative part of $F(x)$ is not always obtained by inverting $x$ in the positive part of $F(\bar{x})$. For instance, if we take $F(x)=1/(1+x)$, its negative part is $0$, but $F(\bar{x})=1/(1+\bar{x})=x/(1+x)$ has a non-trivial positive part. However, if $F(x;t) \in \qs[x, \bar{x}]((t))$, then the expected symmetry holds.
If $\mathbb{A}$ is a field, a power series $F(x) \in \mathbb{A}[[x]]$
is \emm algebraic, (over $\mathbb{A}(x)$) if it satisfies a
non-trivial polynomial equation $P(x, F(x))=0$ with coefficients in
$\mathbb{A}$. Otherwise it is \emm transcendental,. It is \emm differentially finite, (or \emm D-finite,) if it satisfies a non-trivial linear
differential equation with coefficients in $\mathbb{A}(x)$. For
multivariate series, D-finiteness requires the
existence of a differential equation \emm in each variable,. {We
refer to~\cite{Li88,lipshitz-df} for general results on D-finite series.}
We usually omit the dependency in $t$ of our series, writing for instance $C(x,y)$ for $C(x,y;t)$.
For a series $F(x,y) \in \qs[x, \bar{x}, y, \bar{y}][[t]]$ and two integers
$i$ and $j$, we denote by $F_{i,j}$ the coefficient of $x^i y^j$ in $F(x,y)$.
This is a series in $\qs[[t]]$. We also denote
\begin{equation} \label{Fhv}
F_{-,0}(\bar{x}) = \sum_{i<0}F_{i,0}\, x^i \quad\text{and} \quad
F_{0,-}(\bar{y}) = \sum_{j<0}F_{0,j}\, y^j .
\end{equation}
These two series lie in $\bar{x}\qs[\bar{x}][[t]]$ and $\bar{y}\qs[\bar{y}][[t]]$, respectively.
\subsection{A functional equation}
We fix a subset $\mathcal S$ of $\{-1,0, 1\}^2\setminus\{(0,0)\}$ and we consider
walks with steps in $\mathcal S$ that start from
$(0,0)$ and remain in the cone $\mathcal{C}$ defined by~\eqref{Ccone-def}.
By this, we mean that not only must every vertex of the walk
lie in $\mathcal{C}$, but also every edge: a walk containing a step from
$(-1,0)$ to $(0,-1)$ (or vice versa) does not lie in $\mathcal{C}$. We often say
for short that our walks \emm avoid the negative quadrant,. The \emm step polynomial, of $\mathcal S$ is defined by
\begin{equation}\label{eq:steppolynomial}
S(x,y)=\sum_{(i,j)\in \mathcal S} x^i y^j%
= \bar{y} H_-(x)+ H_0(x) +y H_+(x)
= \bar{x} V_-(y) + V_0(y) +x V_+(y),
\end{equation}
for some Laurent polynomials $H_-, H_0, H_+$ and $V_-, V_0, V_+$
recording horizontal and vertical displacements, respectively.
We denote by $C(x,y) \equiv C(x,y;t)$ the
generating function~\eqref{C-def} of walks confined to $\mathcal{C}$. In the expression~\eqref{C-def}, $c_{i,j}(n)$ is the number of walks of length $n$ that go from
$(0,0)$ to $(i,j)$ and
are confined to $\mathcal{C}$.
Constructing walks confined to $\mathcal{C}$ step by step gives the following functional equation:
\[
C(x,y)= 1 +tS(x,y) C(x,y) -t\bar{y} H_-(x) C_{-,0}(\bar{x}) -t\bar{x} V_-(y)C_{0,-}(\bar{y}) -t\bar{x}\bar{y}
C_{0,0} \mathbbm 1_{(-1,-1)\in \mathcal S},
\]
where we have used the notation~\eqref{Fhv}. Note that the series $C_{-,0}(\bar{x})$ and $C_{0,-}(\bar{y})$ count walks ending on the horizontal and vertical boundaries of $\mathcal{C}$ (but not at $(0,0)$).
On the right-hand side, the term $1$ accounts for the
empty walk, the next term describes the extension of a walk in $\mathcal{C}$
by one step of $\mathcal S$, and each of the other three terms corresponds to
a ``bad'' move, either starting from the negative $x$-axis, or from
the negative $y$-axis, or from $(0,0)$. Equivalently,
\begin{equation}\label{eqfunc-gen}
K(x,y)C(x,y)= 1 -t\bar{y} H_-(x) C_{-,0}(\bar{x}) -t\bar{x} V_-(y)C_{0,-}(\bar{y}) -t\bar{x}\bar{y}
C_{0,0} \mathbbm 1_{(-1,-1)\in \mathcal S},
\end{equation}
where $K(x,y):=1-tS(x,y)$ is the \emm kernel, of the equation.
As recalled in the introduction, the enumeration of walks confined to the non-negative quadrant $\mathcal{Q}$ has been studied intensively over the last $20$ years.
The associated generating function\ $Q(x,y)\equiv Q(x,y;t)\in \qs[x,y][[t]]$ defined in~\eqref{Q-def} satisfies a similarly looking equation~\cite[Lem.~4]{bomi10}:
\begin{equation}\label{eqfunc-gen-qu}
K(x,y)Q(x,y)= 1 -t\bar{y} H_-(x) Q(x,0) -t\bar{x} V_-(y)Q(0,y) +t\bar{x}\bar{y}
Q(0,0) \mathbbm 1_{(-1,-1)\in \mathcal S}.
\end{equation}
\begin{remark} In two recent references dealing with the winding number of plane lattice walks~\cite{Budd2020Winding,ElveyPrice2020Winding}, it seems more natural to count walks in which all vertices lie in $\mathcal{C}$, but not necessarily all edges: this means that there may be steps form $(-1,0)$ to $(0,-1)$, and vice-versa. Counting these walks would add two terms to the right-hand side of~\eqref{eqfunc-gen}, namely
\begin{equation}\label{ajout}
t \bar{y} C_{-1,0} \mathbbm 1_{(1,-1)\in \mathcal S} + t \bar{x} C_{0,-1} \mathbbm 1_{(-1,1)\in \mathcal S}.
\end{equation}
We discuss in the final section of the paper
the enumeration of king walks in $\mathcal{C}$ when these two steps are allowed. The results are qualitatively the same as when they are forbidden.
\end{remark}
\subsection{Interesting step sets}
\label{sec:interestingmodels}
As in the quadrant case~\cite{bomi10}, we can decrease the number of
step sets that are worth being considered thanks to a few simple observations (\emm a priori,, there are $2^8$ of them):
\begin{itemize}
\item Since the cone $\mathcal{C}$ (as well as the quarter plane $\mathcal{Q}$) is $x/y$-symmetric, the counting problems defined by $\mathcal S$ and by its mirror image $\overline \mathcal S:=\{(j,i): (i,j) \in \mathcal S\}$ are equivalent; the associated generating functions\ are related by $\overline C(x,y)=C(y,x)$.
%
\item If all steps of $\mathcal S$ are contained in the right half-plane $\{(i,j): i\ge 0\}$, then \emm all, walks with steps in $\mathcal S$ lie in $\mathcal{C}$, and the series $C(x,y)=1/(1-tS(x,y))$ is simply rational. The series $Q(x,y)$ is known to be algebraic in this case~\cite{banderier-flajolet,bousquet-petkovsek-recurrences,Duchon98,gessel-factorization}.
%
\item If all steps of $\mathcal S$ are contained in the left half-plane $\{(i,j): i\le 0\}$, then confining a walk to $\mathcal{C}$ is equivalent to confining it to the upper half-plane: the associated generating function\ is then algebraic, and so is $Q(x,y)$.
%
\item If all steps of $\mathcal S$ lie (weakly) above the first diagonal
($i=j$), then confining a walk to~$\mathcal{C}$ is again equivalent to confining it to the upper half-plane: the associated generating function\ is then algebraic, and so is $Q(x,y)$.
\item If all steps of $\mathcal S$ lie (weakly) above the second diagonal ($i+j=0$), then all walks with steps in $\mathcal S$ lie in $\mathcal{C}$, and $C(x,y)=1/(1-tS(x,y))$ is simply rational. In this case however, the series $Q(x,y)$ is not at all trivial~\cite{bomi10,Mishna-Rechni}. Such step sets are sometimes called \emm singular, in the framework of quadrant walks.
%
\item
Finally, if all steps of $\mathcal S$ lie (weakly) below the second diagonal, then a walk confined to~$\mathcal{C}$ moves for a while along the second diagonal, and then either stops there or leaves it into the NW or SE quadrant using
a South, South-West, or West step. It cannot leave the chosen quadrant anymore and behaves therein like a half-plane walk.
By polishing this observation, one can prove that $C(x,y)$ is algebraic (while $Q(x,y)=1$).
\end{itemize}
By combining these arguments, one finds that there are
$74$ essentially distinct models of walks avoiding the negative quadrant that are worth studying: the $79$ models considered for quadrant walks (see~\cite[Tables~1--4]{bomi10}) except the $5$ ``singular'' models for which all steps of~$\mathcal S$ lie weakly above the diagonal $i+j=0$.
\subsection{The group of the model}
\label{sec:group}
One important tool in the systematic approach to quadrant walks
is a certain group $G$ of birational transformations associated with the step set $\mathcal S$. It was introduced in~\cite{bomi10}, and is an algebraic variant of a group introduced much earlier in the study of \emm random, walks in the quadrant~\cite{fayolle-livre,flatto-hahn,malyshev}.
We assume from now on that $\mathcal S$ contains positive and negative steps in the horizontal and vertical directions (otherwise the problem degenerates, as explained above). We define two bi-rational transformations $\phi$ and $\psi$, acting
on pairs $(u,v)$ of coordinates (which will be, typically, rational functions of $x$ and $y$):
\[
\phi: (u,v)\mapsto \left(\bar{u}\, \frac {V_{-}(v)}{V_{+}(v)}, v\right)
\qquad \hbox{and} \qquad
\psi: (u,v)\mapsto \left(u, \bar{v}\, \frac {H_{-}(u)}{H_{+}(u)}\right),
\]
where $H_-$, $H_+$, $V_-$, and $V_+$ are defined by~\eqref{eq:steppolynomial}. Each transformation fixes one coordinate, and transforms the other \emph{so as to leave the step polynomial $S(u,v)$, defined by~\eqref{eq:steppolynomial}, unchanged.}
Note that $\phi$ and $\psi$ are both involutions. The group $G$ is
the group generated by these two transformations. It is isomorphic to
a dihedral group of order $2n$, with $n\in\ns\cup
\{\infty\}$. The \emm length, of $g\in G$, denoted $\ell(g)$,
is the smallest $\ell$ such that $g$ can be written as a product of
$\ell$ generators $\phi$ and $\psi$. The \emm sign, of $g\in G$,
denoted $\varepsilon_g$, is defined by $\varepsilon_g = (-1)^{\ell(g)}$.
Note that for any $g \in G$, we have $S(g(x,y))=S(x,y)$.
Among the $74$ interesting models identified in the previous subsection, exactly $23$ have a finite group; see~\cite{bomi10}. For the remaining $51$ models, an asymptotic argument implies that the series $C_{0,0}$ that counts walks ending at $(0,0)$ is not D-finite, which implies that $C(x,y;t)$ is not D-finite; see~\cite{Mustapha2019Walks}. Among the $23$ models with a finite group,
\begin{itemize}
\item $16$ have a vertical symmetry (say) and a group of order $4$,
\item $5$ have a group of order $6$, and
\item $2$ have a group of order $8$.
\end{itemize}
These models and groups are listed in~\cite[Tables~1--3]{bomi10}. Another classification of these groups distinguishes the $7+4$ models with a \emm monomial, group (meaning that for every $g\in G$, the pair $g(x,y)$ consists of two Laurent monomials in $x$ and $y$), shown in Tables~\ref{tab:weyl} and~\ref{tab:zero} of this paper, from the $12$ non-monomial ones (Table~\ref{tab:non-monomial}).
\begin{table}[htb]
\centering
\begin{tabular}{c|c|c|c|c|c|c}
$\mathcal S$ & \includegraphics[height=0.8cm]{01001001} \includegraphics[height=0.8cm]{01101011} &\includegraphics[height=0.8cm]{11001001} \includegraphics[height=0.8cm]{11101011} & \includegraphics[height=0.8cm]{11010101} \includegraphics[height=0.8cm]{11110111}
& \includegraphics[height=0.8cm]{10011100} \includegraphics[height=0.8cm]{10111110}&\includegraphics[height=0.8cm]{01011101} \includegraphics[height=0.8cm]{01111111}& \includegraphics[height=0.8cm]{10010100} \includegraphics[height=0.8cm]{10110110}\\
\hline
$G$ & \begin{minipage}{20mm}
$(x,y),(\bar x,y)$,\\ $(\bar x, \bar y \frac{1}{x+\bar x})$,\\ $(x, \bar y \frac{1}{x+\bar x}) $
\end{minipage}&
\begin{minipage}{20mm}
$(x,y),(\bar x,y)$,\\
$(\bar x,\bar y\,\frac {1}{x+1+\bar x})$,\\
$(x,\bar y\,\frac {1}{x+1+\bar x}) $
\end{minipage}&
\begin{minipage}{20mm}
$(x,y),(\bar x, y)$,\\
$(\bar x, \bar y\, \frac {x+\bar x}{x+1+\bar x})$,\\
$(x, \bar y\, \frac {x+\bar x}{x+1+\bar x})$
\end{minipage}&
\begin{minipage}{28mm}
$(x,y),(\bar x,y)$,\\
$(\bar x,\bar y\,(x+1+\bar x))$,\\
$(x,\bar y\,(x+1+\bar x))$\end{minipage}&
\begin{minipage}{20mm}$(x,y), (\bar x,y)$,\\
$(\bar x,\bar y\, \frac{x+1+\bar x}{x+\bar x})$,\\
$(x,\bar y\, \frac{x+1+\bar x}{x+\bar x})$\end{minipage}&
\begin{minipage}{22mm}$(x,y),(\bar{x},y)$,\\
$(\bar x,\bar{y}(x+\bar{x}))$,\\
$(x,\bar{y}(x+\bar{x}))$
\end{minipage}
\end{tabular}
\vskip 4mm
\caption{The $12$ models with a finite non-monomial group.}
\label{tab:non-monomial}
\end{table}
\section{A conjectural form of \texorpdfstring{$C(x,y)$}{C(x,y)} for Weyl models}
\label{sec:OS}
It is shown in~\cite{bomi10} that $19$ of the $23$ quadrant problems with a finite group can be solved in a uniform manner by considering the \emm orbit sum, of $xy$,
where for any series $F(x,y)$ in $\qs(x,y)[[t]]$, we define the \emm orbit sum, of $F$ by
\[
\OS (F(x,y)):= \sum_{g\in G}\varepsilon_g\, g(F(x,y)),
\]
with $g(F(x,y)):=F(g(x,y))$. To explain the relevance of orbit sums
observe that the quadrant equation~\eqref{eqfunc-gen-qu}, once multiplied by $xy$, reads
\[
K(x,y) xy Q(x,y)= xy-R(x)-S(y),
\]
where the series $R(x)$ does \emm not, involve $y$ and the series $S(y)$ does not involve $x$. Recall moreover that for every $g \in G$ we have $K(g(x,y))=1-tS(g(x,y))=1-tS(x,y)=K(x,y)$. By linearity this gives
\begin{align}
K(x,y) \sum_{g\in G} \varepsilon_g \, g(xyQ(x,y))&= \OS (K(x,y)xyQ(x,y)) \nonumber\\
&= \OS(xy) -\sum _{g\in G} \varepsilon_g \,g(R(x))-\sum _{g\in G} \varepsilon_g \,g(S(y)) \nonumber\\
& = \OS(xy), \label{OS-Q}
\end{align}
since $\varepsilon_{g \circ \psi}=\varepsilon_{g \circ \phi}=- \varepsilon_g$ while $g(R(x)) = (g \circ \psi)(R(x))$ and $g(S(y)) = (g \circ \phi)(S(y))$.
Analogously, for walks confined to $\mathcal{C}$, the form of the functional equation~\eqref{eqfunc-gen} implies that:
\begin{equation}\label{OS-C}
K(x,y) \sum_{g\in G} \varepsilon_g\, g(xyC(x,y))= \OS (xy).
\end{equation}
\begin{remark}
If we decide to allow steps between $(-1,0)$ and $(0,-1)$ in the walks that we count, thus adding the terms~\eqref{ajout} to the right-hand side of the functional equation~\eqref{eqfunc-gen}, the orbit sum of $xy C(x,y)$ is still $\OS(xy)/K(x,y)$.
\end{remark}
\subsection{Vanishing orbit sums and algebraicity}
\begin{table}[b]
\centering
\begin{tabular}{l|c|c|}
model & \begin{tabular}{ccc}
\includegraphics[height=0.8cm]{01001010}&\includegraphics[height=0.8cm]{10100100} &\includegraphics[height=0.8cm]{11101110}\\
\small{Kreweras}& \small{reverse Kreweras}& \small{double-Kreweras}
\end{tabular}&
\begin{tabular}{c}
\includegraphics[height=0.8cm]{01100110}\\ \small{Gessel}
\end{tabular}
\\
\hline
group &\begin{minipage}{3cm}
$(x,y), (\bar x \bar y,y)$,\\
$(\bar x \bar y, x)$, $(y,x)$,\\
$(y, \bar x \bar y), (x,\bar x\bar y)$
\end{minipage} &
\begin{minipage}{3cm}
\mbox{}\\$(x, y), (\bar x \bar y, y),$\\ $(\bar x \bar y, x^2y), (\bar x, x^2y)$,\\
$ (\bar x, \bar y), (xy, \bar y),$\\ $(xy, {\bar x}^2 \bar y), (x, \bar y {\bar x}^2)$\\
\end{minipage} \\
new steps & \includegraphics[height=0.9cm]{kreweras-steps} & \includegraphics[height=0.8cm]{10101010}\\
\begin{minipage}{2cm}
\vskip -36mm quadrant\\ walks
\end{minipage}
& \includegraphics[scale=0.7]{kreweras} & \includegraphics[scale=0.7]{gessel}
\end{tabular}
\vskip 3mm
\caption{The four models for which $\OS(xy)=0$, with their names and groups. After deformation, the first quadrant becomes a union of two/three
Weyl chambers. The three-quadrant plane corresponds to the complement of this region.}
\label{tab:zero}
\end{table}
As was first observed in~\cite{bomi10}, the orbit sum $\OS(xy)$ is zero for exactly four of the $23$ models with a finite group. The four models are shown in Table~\ref{tab:zero}. For each of them, one has:
\[
\sum_{g\in G} \varepsilon_g \, g(xyQ(x,y))=0=\sum_{g\in G} \varepsilon_g \, g(xyC(x,y)).
\]
That is, the orbit sums of $xyQ(x,y)$ and $xyC(x,y)$ vanish.
It is known that these four models are precisely those for which $Q(x,y)$ is algebraic.
One can derive the algebraicity from the fact that $\OS(xy)=0$~\cite{BeBMRa-16,BeBMRa-17,BostanKurkovaRaschel2016}. These derivations strongly suggest that, more generally, for any finite group model and any point $(a,b) \in \mathcal{Q}$ such that $\OS(x^{a+1} y^{b+1})=0$, the generating function\ for walks in $\mathcal{Q}$ starting from $(a,b)$ is algebraic. This is proved in some cases beyond the case $(a,b)=(0,0)$; see the discussion in~\cite[Sec.~7.2]{BeBMRa-17}. Note that the corresponding generating function\ $\widetilde Q(x,y)$ is defined by
\[
K(x,y)\widetilde Q(x,y)= x^a y^b -t\bar{y} H_-(x) \widetilde Q(x,0) -t\bar{x} V_-(y)\widetilde Q(0,y) +t\bar{x}\bar{y}
\widetilde Q(0,0) \mathbbm 1_{(-1,-1)\in \mathcal S}.
\]
By linearity, it is thus expected that for any polynomial $I(x,y) \in \qs[x,y]$ satisfying $\OS(xyI)=0$, the series $\widetilde Q(x,y) \in \qs[x,y][[t]]$ defined by
\[
K(x,y)\widetilde Q(x,y)= I(x,y) -t\bar{y} H_-(x) \widetilde Q(x,0) -t\bar{x} V_-(y)\widetilde Q(0,y) +t\bar{x}\bar{y}
\widetilde Q(0,0) \mathbbm 1_{(-1,-1)\in \mathcal S}
\]
is algebraic.
Could a similar algebraicity phenomenon hold for three-quadrant problems? That is, could it be that, for any finite group model, and any Laurent polynomial $I(x,y) \in \qs[x,\bar{x},y,\bar{y}]$ having its support in $\mathcal{C}$ and satisfying $\OS(xyI)=0$, the series $\widetilde C(x,y)$ defined by
\begin{equation}\label{eq-tC}
K(x,y)\widetilde C(x,y)= I(x,y) -t\bar{y} H_-(x) \widetilde C_{-,0}(\bar{x}) -t\bar{x} V_-(y)\widetilde C_{0,-}(\bar{y}) -t\bar{x}\bar{y}
\widetilde C_{0,0} \mathbbm 1_{(-1,-1)\in \mathcal S}
\end{equation}
is algebraic? (Again, if $I(x,y)$ is reduced to a single monomial $x^ay^b$, then $\widetilde C$ counts walks in~$\mathcal{C}$ starting from $(a,b)$.) Several results and guesses support this belief:
\begin{enumerate}
\item It was conjectured in\cite{Bousquet2016} that for the four models of Table~\ref{tab:zero}, for which $\OS(xy)=0$, the series $C(x,y)$ that counts walks in $\mathcal{C}$ starting from $(0,0)$ is algebraic. This was mostly based on a guessed polynomial equation satisfied by $C_{0,0}$. The algebraicity of $C(x,y)$ (and $C_{0,0}$) is now proved for the first three of these models~\cite{mbm-tq-kreweras}, with explicit algebraic expressions. Moreover, Theorem~23 in~\cite{Budd2020Winding}, taken with $\alpha=0$, $\beta_-=-\pi/2$, and $\beta_+=3\pi/4$, proves that $C_{0,0}$ is algebraic as well for the fourth model.
\item For the simple square lattice model $\mathcal S=\{\rightarrow,\uparrow, \leftarrow, \downarrow\}$, the orbit of $(x,y)$ under the action of $G$ is $\{(x,y), (\bar{x}, y), (x, \bar{y}),(\bar{x},\bar{y})\}$. The orbit sum $\OS(xy)$ is non-zero, and the series $C(x,y)$ is not algebraic. But it was proved that for walks starting at $(-1,0)$, that is, for $I(x,y)=\bar{x}$, the generating function\ of walks confined to $\mathcal{C}$ \emm is, algebraic~\cite[Thm.~6]{Bousquet2016}, and one observes that $\OS(xyI)=\OS(y)=0$. Moreover, the generating function\ of walks in $\mathcal{C}$ starting from $(-1,b)$ and ending at $(0,0)$ is also algebraic~\cite[Cor.~2]{Bousquet2016},
and for $I(x,y)= \bar{x} y^b$ it also holds that $\OS(xyI)=0$.
\item\label{item3} Still for the square lattice model, the heart of the derivation of $C(x,y)$ in\cite{Bousquet2016} is to prove that the series $\widetilde C(x,y)$ defined by~\eqref{eq-tC} with $I(x,y)= (2+\bar{x}^2+\bar{y}^2)/3$ is algebraic. Observe that $\OS(xyI)=0$.
\item Similar results hold for the diagonal model $\mathcal S=\{\nearrow, \nwarrow, \swarrow, \searrow \}$, which has the same group as the square lattice model. More precisely, for $I(x,y)=(2+\bar{x}^2+\bar{y}^2)/3$ the series $\widetilde C(x,y)$ is algebraic~\cite[Thm.~4]{Bousquet2016}, and moreover walks confined to $\mathcal{C}$ starting at $(-1,b)$ and ending at $(0,0)$ have an algebraic generating function~\cite[Cor.~5]{Bousquet2016}.
\item Finally, it is conjectured by Raschel and
Trotignon~\cite[p.~9]{RaschelTrotignon2018Avoiding} that for any
finite group model, walks in $\mathcal{C}$ starting from $(-1,b)$ (or
$(b,-1)$) have an algebraic generating function. This series
satisfies~\eqref{eq-tC} with again $I(x,y)= \bar{x} y^b$, and thus $\OS(xyI)=0$.
\end{enumerate}
We could add to this list more algebraicity results for walks with
a fixed endpoint confined to certain cones;
see~\cite[Thm.~23]{Budd2020Winding}
and~\cite[Cor.~4]{ElveyPrice2020Winding}. In these two papers
steps between $(-1,0)$ and $(0,-1)$ are allowed (when they belong
to $\mathcal{C}$), and thus the series under consideration obey slightly
different functional equations. We also refer
to~\cite{buchacher-kauers-trotignon} for more exotic series with
zero orbit sums that are \emm not, algebraic.
\subsection{Vanishing orbit sums for three-quadrant walks}
As mentioned in item~\eqref{item3} above, the heart of the derivation of $C(x,y)$ for the simple square lattice is the solution of~\eqref{eq-tC} with $I(x,y)= (2+\bar{x}^2+\bar{y}^2)/3$. Recall the associated series $xy\widetilde C$ has orbit sum zero. What is then the connection between $\widetilde C(x,y)$ and the series $C(x,y)$ we are interested~in?
It is not hard to construct, for any finite group model, a series $\widetilde C(x,y)$ related to $C(x,y)$ such that $xy\widetilde C$ has orbit sum zero. In sight of the functional equations~\eqref{OS-Q} and~\eqref{OS-C}, an obvious choice is $\widetilde C(x,y):=C(x,y)-Q(x,y)$. However, we would also like $\widetilde C(x,y)$ to be characterized by a functional equation resembling~\eqref{eq-tC},
in which every unknown series is explicitly described as a sub-series of $\widetilde C(x,y)$. But this is not the case for the above choice of $\widetilde C$. Indeed, if for instance we take $\mathcal S=\{\rightarrow, \uparrow, \leftarrow, \downarrow\}$, we have
\begin{equation}\label{eqQ-square}
K(x,y) Q(x,y) = 1-t\bar{y} Q(x,0) -t\bar{x} Q(0,y) ,
\end{equation}
\begin{equation}\label{eqC-square}
K(x,y) C(x,y) = 1-t\bar{y} C_{-,0}(\bar{x}) -t\bar{x} C_{0,-}(\bar{y}) ,
\end{equation}
and, if we take $\widetilde C(x,y):=C(x,y)-Q(x,y)$, we obtain, by extracting terms of the forms $x^iy^0$ and $x^0y^i$ with $i<0$:
\[
C_{-,0}(\bar{x})= \widetilde C_{-,0}(\bar{x}), \qquad C_{0,-}(\bar{y}) =\widetilde C_{0,-}(\bar{y}),
\]
so that
\[
K(x,y) \widetilde C(x,y)= -t\bar{y} \widetilde C_{-,0}(\bar{x}) -t\bar{x} \widetilde C_{0,-}(\bar{y}) +t\bar{y} Q(x,0) +t\bar{x} Q(0,y) ,
\]
which does not look very encouraging. However, we can get more leeway as follows. Observe that for any model $\mathcal S$ with a finite group $G$, any $h \in G$, and any series $F\in \qs(x,y)[[t]]$, we have $\OS (h(F ))=\varepsilon_h \OS ( F)$. Therefore, for any $h\in G$, the series $ \varepsilon_h h(xyQ(x,y))$ has the same orbit sum as $xy Q(x,y)$ and $xy C(x,y)$. Consequently, for any collection of real numbers $\lambda_h, h \in G$ such that $\sum_h \lambda_h=1$, the series
\begin{equation}\label{A-def-gen}
\widetilde C(x,y) := C(x,y) - \bar{x}\bar{y} \sum _{h \in G} \varepsilon_h \lambda_h h( xyQ(x,y))
\end{equation}
is such that $xy \widetilde C$ has vanishing orbit sum. Can we choose the $\lambda_h$ such that $\widetilde C(x,y)$ is defined by an equation not involving $Q$? Returning to the square lattice case, if we choose
\[
\widetilde C(x,y):=C(x,y) - \frac{\bar{x}\bar{y}}3\big( xy Q(x,y)- \bar{x} y Q(\bar{x},y) -x\bar{y} Q(x,\bar{y})\big),
\]
we obtain
\[
C_{-,0}(\bar{x})= \widetilde C_{-,0}(\bar{x}) - \frac {\bar{x}^2} 3\, Q(\bar{x},0), \quad
C_{0,-}(\bar{y})= \widetilde C_{0,-}(\bar{y}) - \frac {\bar{y}^2} 3\, Q(0,\bar{y}), \quad
\]
and the combination of~\eqref{eqQ-square} (written for $(x,y)$, $(\bar{x},y)$, and $(x,\bar{y}))$ and~\eqref{eqC-square} results in the following simple equation:
\[
K(x,y) \widetilde C(x,y)= \frac{2+\bar{x}^2+\bar{y}^2}3 -t\bar{y} \widetilde C_{-,0}(\bar{x}) -t\bar{x} \widetilde C_{0,-}(\bar{y}),
\]
which involves no specialization of $Q$. The following proposition tells us that a similar choice exists for any of the Weyl models of Table~\ref{tab:weyl}. In fact, it is easy to check that this choice is always unique.
\begin{prop}\label{prop:A-def-gen}
Let $\mathcal S$ be one of the Weyl models shown in Table~\ref{tab:weyl}. Let $2d$ be the order of the associated group $G$. Then $d=2, 3$ or $4$. Let $\omega:=\phi\psi\phi \cdots$ (with $d$ generators) be the only element of length $\ell(\omega)=d$ in $G$.
Define
\begin{align}
A(x,y) &= C(x,y) - \frac{\bar{x} \bar{y}} {2d-1} \sum_{h\in G\setminus\{\omega\}} \varepsilon_h\, h(xyQ(x,y)) \label{Aexpr-1}\\
&= C(x,y) -\frac{\bar{x} \bar{y}} {2d-1} \left( \frac{\OS(xy)}{K(x,y)} - \varepsilon_\omega\, \omega(xyQ(x,y))\right).\label{Aexpr-2}
\end{align}
Then $xyA(x,y)$ has orbit sum zero, and is characterized by the following equation:
\begin{align}
\label{A-func-eq}
\begin{aligned}
K(x,y) A(x,y)&= 1 -\frac {\bar{x}\bar{y}}{2d-1} \left(
\OS(xy)-(-1)^d\bar{x}\bar{y}\right)\\
& \quad -t\bar{y} H_-(x) A_{-,0}(\bar{x}) -t\bar{x} V_-(y)A_{0,-}(\bar{y}) -t\bar{x}\bar{y} A_{0,0} \mathbbm 1_{(-1,-1)\in \mathcal S}.
\end{aligned}
\end{align}
\end{prop}
\begin{proof}
First, the equivalence between the two expressions of $A(x,y)$ comes from~\eqref{OS-Q}. Then, the orbit sum of $xyA$ vanishes because, as noticed above, $xyC$ and each $\varepsilon_h \, h(xyQ(x,y))$ have the same orbit sum.
Now we want to write an equation for $A(x,y)$, using the defining equations of $C$ and $Q$. Let us first express $C_{-,0}(\bar{x})$ in terms of $A$ and $Q$, by extracting from~\eqref{Aexpr-1} terms of the form $x^iy^0$, with $i<0$. By examination of the three possible groups, detailed in Table~\ref{tab:weyl}, we see that only one element $h$ contributes, namely $\omega^-=\phi\psi\cdots$ with $d-1$ generators. More explicitly,
\begin{equation}\label{Cm0}
C_{-,0}(\bar{x})=A_{-,0}(\bar{x})+\frac{ (-1)^{d-1}}{2d-1}
\begin{cases}
\bar{x}^d Q(\bar{x},0) & \text{if } d=2 \text{ or } 4,\\
\bar{x}^d Q(0,\bar{x}) & \text{if } d=3.
\end{cases}
\end{equation}
Analogously, when we extract from~\eqref{Aexpr-1} terms of the form $x^0 y^j$, with $j<0$, the only group element that contributes is $\omega^+=\phi\psi\cdots $ with $(d+1)$ generators, and
\begin{equation}\label{C0m}
C_{0,-}(\bar{y})= A_{0,-}(\bar{y})+\frac {(-1)^{d+1}}{2d-1}
\begin{cases}
\bar{y}^m Q(0,\bar{y}) & \text{if } d=2 \text{ or } 4,\\
\bar{y}^3 Q(\bar{y},0) & \text{if } d=3,
\end{cases}
\end{equation}
where $m=2$ for $d=2$ and $m=3$ for $d=4$. Finally, the only element $h$ that contributes to the coefficient of $x^0y^0$ in $C(x,y)$ is the identity, and
\begin{equation}\label{C00}
C_{0,0}= A_{0,0} + \frac 1 {2d-1} Q(0,0).
\end{equation}
We now start from~\eqref{Aexpr-2} to write an equation defining $A(x,y)$.
By examining again the three possible groups we see that $\omega(x,y)=(\bar{u},\bar{v})$ with
\begin{equation}\label{uv}
(u,v)=
\begin{cases}
(x, y) & \text{if } d=2 \text{ or } 4,\\
(y,x) &\text{if } d=3.
\end{cases}
\end{equation}
Let us finally denote $\delta= \mathbbm 1_{(-1,-1)\in \mathcal S}$. Then
\begin{align*}
K(x,y) A(x,y) &= K(x,y)C(x,y) - \frac{\bar{x} \bar{y}}{2d-1}\left( {\OS(xy)} -(-1)^d K(x,y) \omega(xy Q(x,y))\right) \\
&= 1 -t\bar{y} H_-(x) C_{-,0}(\bar{x}) -t\bar{x} V_-(y)C_{0,-}(\bar{y}) -t\delta \bar{x}\bar{y}
C_{0,0}
- \frac{\bar{x} \bar{y}}{2d-1} {\OS(xy)}\\
& \quad + \frac{(-1)^d \bar{x}^2 \bar{y} ^2}{2d-1} \Big(
1 -tv H_-(\bar{u}) Q(\bar{u},0) -tu V_-(\bar{v})Q(0,\bar{v}) +t\delta uv
Q(0,0)\Big).
\end{align*}
Here, we have used~\eqref{eqfunc-gen} to express $K(x,y)C(x,y)$, and~\eqref{eqfunc-gen-qu} to express $K(x,y) \omega(Q(x,y))=K(\bar{u}, \bar{v}) Q(\bar{u}, \bar{v})$, with $(u,v)$ given by~\eqref{uv}.
We now express $C_{-,0}(\bar{x})$, $C_{0,-}(\bar{y})$, and $C_{0,0}$ thanks to~\eqref{Cm0}--\eqref{C00},
and examine separately the cases $d=2,4$ and $d=3$.
If $d=2$ or $4$, we have
\begin{align*}
K(x,y) A(x,y) &= 1 -t\bar{y} H_-(x) A_{-,0}(\bar{x}) -t\bar{x} V_-(y)A_{0,-}(\bar{y}) -t\delta \bar{x}\bar{y} A_{0,0} - \frac{\bar{x} \bar{y} ({\OS(xy)-\bar{x}\bar{y}})}{2d-1} \\
& \quad - t\delta \frac{\bar{x} \bar{y}}{2d-1} \big( Q(0,0)-\bar{x}\bar{y} x y Q(0,0) \big) \\
& \quad + \frac {t\bar{y}}{2d-1} Q(\bar{x}, 0)\left( \bar{x}^d H_-(x)-\bar{x}^2 H_-(\bar{x})\right)
\\
& \quad + \frac{t\bar{x}}{2d-1} Q(0, \bar{y}) \Big( \bar{y}^m V_-(y)-\bar{y}^2 V_-(\bar{y})\Big),
\end{align*}
and the announced equation follows by observing that for each of the 5 models under consideration (shown in the first and third columns of Table~\ref{tab:weyl}),
\[
\bar{x}^{d-2} H_-(x)= H_-(\bar{x}) \quad \text{and} \quad \bar{y}^{m-2} V_-(y)= V_-(\bar{y}).
\]
For the two models such that $d=3$ (central column in Table~\ref{tab:weyl}), we have $\delta=0$, and
\begin{align*}
K(x,y) A(x,y) &= 1 -t\bar{y} H_-(x) A_{-,0}(\bar{x}) -t\bar{x} V_-(y)A_{0,-}(\bar{y}) - \frac{\bar{x} \bar{y}({\OS(xy)+ \bar{x}\bar{y}})}{2d-1} \\
&\ \quad - \frac {t\bar{y}}{2d-1} Q(0,\bar{x})\left( \bar{x}^d H_-(x)-\bar{x}^2 V_-(\bar{x})\right)
\\
&\ \quad - \frac{t\bar{x}}{2d-1} Q(\bar{y},0) \Big( \bar{y}^3 V_-(y)-\bar{y}^2 H_-(\bar{y})\Big),
\end{align*}
and now the announced equation follows from the fact that
\[
\bar{x} H_-(x)=V_-(\bar{x}). \qedhere
\]
\end{proof}
\begin{remark}
Let us explain why we only consider the seven Weyl models in Proposition~\ref{prop:A-def-gen}.
\begin{itemize}
\item For the four models of Table~\ref{tab:zero}, we have seen that $\OS(xy)=0$. Hence $xy C$ has orbit sum zero, and we can simply take $A(x,y)=C(x,y)$.
%
\item The remaining $12$ models (Table~\ref{tab:non-monomial}) have a vertical symmetry, and a group of order~$4$. The orbit of $(x,y)$ consists of the pairs $(x,y),(\bar{x},y)=\phi(x,y)$, and two pairs $(x, \bar{y} r(x))=\psi(x,y)$ and $(\bar{x},\bar{y} r(x))=\phi\psi(x,y)$ where $r(x)=r(\bar{x})$ is a rational function in $x$, \emm not reduced to a monomial,. Let us consider a series $\widetilde C(x,y)$ of the form~\eqref{A-def-gen}. If at least one of the coefficients $\lambda_{\psi}$ or $\lambda_{\phi\psi}$ is non-zero, then it is not clear how to express $C_{0,-}(\bar{y})$ in terms of $Q$ and $A$, as we did in~\eqref{C0m}. If these two coefficients are taken to be $0$, then
\[
\widetilde C(x,y)=C(x,y)-\lambda Q(x,y) +(1-\lambda) \bar{x}^2 Q(\bar{x},y),
\]
and the reader can check that the equation for $\widetilde C$ still involves specializations of $Q$, for any choice of $\lambda$.
\end{itemize}
\end{remark}
Let us thus return to the seven Weyl models. The reason for the construction of $A(x,y)$ (and of the notation change $\widetilde C \rightarrow A$) is the following, partially proved, conjecture.
\begin{conjecture}\label{conj:Weyl}
For any of the seven Weyl models of Table~\ref{tab:weyl}, the series $A(x,y)$ defined in Proposition~\ref{prop:A-def-gen} is algebraic. In particular, $C(x,y)$ is D-finite.
For any of the four models of Table~\ref{tab:zero}, the series $C(x,y)$ is algebraic (as $Q(x,y)$ itself).
\end{conjecture}
Let us recall that the conjecture is proved for the first two models of Table~\ref{tab:weyl} in\cite{Bousquet2016}, and in this paper for the third (king steps). The second part of the conjecture is proved for the first three models of Table~\ref{tab:zero} in\cite{mbm-tq-kreweras} (D-finiteness was established earlier in\cite{RaschelTrotignon2018Avoiding}).
This leaves us with five models for which the conjecture is open: four Weyl models, and the (conjecturally algebraic) Gessel model. Based on the solved cases, we believe that algebraicity should hold in a strong sense, and in particular, that each series $A_{i,j}$ (for Weyl models) or $C_{i,j}$ (for the models of Table~\ref{tab:zero}) should be algebraic. For the four remaining Weyl models, we have tried to guess (using the \texttt{gfun} package~\cite{gfun} in {\sc Maple}) a polynomial equation for the series $A_{-1,0}$, which coincides with the generating function\ $C_{-1,0}$ of walks ending at $(-1,0)$.
We could not guess anything for the diabolo model (using the counting sequence up to length $n=4000$), but we discovered equations of degree $24$ for each of the next three.
The degree of $C_{-1,0}$ is $4$ (resp.\ $8$, $24$) for the three solved Weyl models. For Gessel's model, $C_{0,0}$ was conjectured to be algebraic of degree $24$ in~\cite{Bousquet2016}, and algebraicity was proved since then in~\cite[Thm.~23]{Budd2020Winding} (taken with $\alpha=0$, $\beta_-=-\pi/2$, and $\beta_+=3\pi/4$).
\section{The king walks: statement of the results}
\label{sec:KingResults}
We now fix the step set to be $\mathcal S=\{-1,0,1\}^2\setminus\{(0,0)\}$. We still denote by $Q(x,y)$ the generating function\ of walks confined to the first quadrant, and by $C(x,y)$ the generating function\ of walks avoiding the negative quadrant. The orbit of $(x,y)$ under the action of $G$ is $\{(x,y), (\bar{x}, y), (x, \bar{y}), (\bar{x}, \bar{y})\}$. Recall from~\cite{bomi10} that $Q(x,y)$ can be expressed in terms of a simple rational function:
\[
xyQ(x,y)= [x^> y^>] \frac{\OS(xy)}{K(x,y)}
= [x^> y^>] \frac{(x-\bar{x})(y-\bar{y})}{1 - t( x + xy + y + \bar{x} y + \bar{x} + \bar{x} \bar{y} + \bar{y} + x \bar{y} )}.
\]
From Proposition~\ref{prop:A-def-gen} we get an expression of $C(x,y)$ of the form
\[
C(x,y)= A(x,y) + \frac{1}{3}\big( Q(x,y) - \bar{x}^2\,Q\!\left(\bar{x},y\right) - \bar{y}^2\,Q\!\left(x,\bar{y}\right) \big),
\]
where, as announced, $A(x,y)$ is algebraic. More precisely, we write
\begin{equation}\label{A-PM}
A(x,y)=P(x,y)+\bar{x} M(\bar{x}, y)+\bar{y} M(\bar{y}, x),
\end{equation}
where $P(x,y)$ and $M(x,y)$ belong to $\qs[x,y][[t]]$, and prove that $P$ and $M$ are algebraic.
\begin{theo} [{\bf The GF of king walks}]
\label{theo:king}
The generating function\ of king walks starting from $(0,0)$,
confined to $\mathcal{C}$, and ending in the first quadrant (resp.~at a negative
abscissa) is
\begin{equation}\label{sol-king}
\frac 1 3 Q(x,y) + P(x,y), \qquad \left(\hbox{resp.} -\frac 1 3
\bar{x}^2 Q(\bar{x},y) + \bar{x} M(\bar{x},y)\right),
\end{equation}
where $P(x,y)$ and $M(x,y)$ are algebraic of degree $216$ over $\qs(x,y,t)$.
The generating function of walks ending at a negative ordinate follows using the $x/y$-symmetry of the step set.
The series $P$ can be expressed in terms of $M$ by:
\begin{equation}\label{Psol-king}
P(x,y)=\bar{x} \big( M(x,y)-M(0,y)\big) +\bar{y} \big(
M(y,x)-M(0,x)\big),
\end{equation}
and $M$ is defined by the following equation:
\begin{align}
\begin{aligned}
K(x,y) &\left(2M(x,y)-M(0,y)\right) =
\frac{2x}{3}
-2t \bar{y}(x+1+\bar{x})M(x,0)
+t \bar{y}(y+1+\bar{y})M(y,0) \\
&\qquad
+t(x-\bar{x})(y+1+\bar{y})M(0,y)
-t\left(1 + \bar{y}^2 - 2\bar{x}\bar{y} \right)%
M_{0,0}
-t\bar{y}
M_{1,0},
\end{aligned}
\label{eqMMM-king}
\end{align}
where $K(x,y) = 1 - t( x + xy + y + \bar{x} y + \bar{x} + \bar{x} \bar{y} + \bar{y} + x \bar{y} )$.
The specializations $M(x,0)$ and $M(0,y)$ are algebraic
of degree $72$ over $\qs(x,t)$ and $\qs(y,t)$, respectively, and $M_{0,0}$ and
$M_{1,0}$ have degree $24$ over $\qs(t)$.
\end{theo}
We give in the following two subsections
a complete algebraic description of all the series
needed to reconstruct $P(x,y)$ and $M(x,y)$ from~\eqref{Psol-king}
and~\eqref{eqMMM-king}, namely, the univariate series $M_{0,0}$ and
$M_{1,0}$ (Section~\ref{sec:univariate})
and the bivariate series $M(x,0)$ and $M(0,y)$ (Section~\ref{sec:bivariate}).
A combinatorial proof of~\eqref{Psol-king} is given in Section~\ref{sec:combi}, together with a generalization to other starting points and other Weyl models.
\subsection{Univariate series}
\label{sec:univariate}
We define in three steps an extension of $\mathbb{Q}(t)$ of degree $24$, schematized by
\[
\mathbb{Q}(t) \stackrel{4}{\hookrightarrow} \mathbb{Q}(t,u) \stackrel{3}{\hookrightarrow} \mathbb{Q}(t,v) \stackrel{2}{\hookrightarrow} \mathbb{Q}(t,w),
\]
where $u, v, w \in \qs[[t]]$ and the numbers
give the degrees of the successive extensions.
First, let $u=t+t^2+\mathcal{O}(t^3)$ be the only power series in $t$ satisfying the quartic equation
\begin{equation}\label{u-def}
(1-3u)^3 (1+u)t^2+(1 + 18u^2
-27u^4)t-u=0.
\end{equation}
Equivalently,
\begin{equation}\label{u-def-alt}
\frac {u}{(1+u)(1-3u)^3}= \frac {t(1+t)}{1-8t}.
\end{equation}
Second, let $v = t+3t^2+\mathcal{O}(t^3)$ be the only series with constant term zero satisfying the cubic equation
\begin{equation}\label{v-def}
(1+3v-v^3)u -v(v^2+v+1)=0.
\end{equation}
Clearly, it holds that $u\in \qs(v)$, and hence we have $\qs(t,u,v)=\qs(t,v)$. The minimal equation of $v$ over $\mathbb{Q}(t)$ can be written as follows:
\begin{equation}\label{algv}
{\frac {v \left( {v}^{2}+v+1 \right) \left( {v}^{3}-3\,v-1 \right) ^{
3}}{ \left( {v}^{2}+4\,v+1 \right) \left( 4\,{v}^{3}+3\,{v}^{2}-1
\right) ^{3}}}= \frac {t(1+t)}{1-8t}.
\end{equation}
Third, define
\begin{equation}\label{w-def}
w = \sqrt{1 + 4v - 4v^3 - 4v^4} = 1+2t+4t^2+\mathcal{O}(t^3).
\end{equation}
One can check that $w$ has degree $24$ over $\qs(t)$. Hence the extension $\qs(t,w)$ contains $v$.
\smallskip
We can now make the series $M_{0,0}$ and $M_{1,0}$ occurring in~\eqref{eqMMM-king} explicit. Note that by~\eqref{sol-king}, the series $M_{0,0}$ coincides with the series $C_{-1,0}$ that counts
walks in $\mathcal{C}$ ending at $(-1,0)$. It is
\begin{equation}\label{Cm10-expr}
M_{0,0} =C_{-1,0}= \frac{1}{2t} \left( \frac{w(1+2v)}{1+4v-2v^3}-1\right)
= t+2t^2+17t^3+80t^4+536t^5+\mathcal{O}(t^6).
\end{equation}
Analogously, we have
\[
C_{0,0}= \frac 1 3 Q_{0,0} +P_{0,0} \qquad \text{and} \qquad C_{-2,0}= -\frac 1 3 Q_{0,0} +M_{1,0},
\]
where $ P_{0,0} = 2M_{1,0}$ (by~\eqref{Psol-king}) and
\begin{align}\label{Mm10-expr}
M_{1,0}&=
\frac{1}{6t^2}\left( 1+2t + \frac{(1-2t)(1+2v)(16v^6+24v^5+7v^4-24v^3-30v^2-10v-1)}{w(v^4+8v^3+6v^2+2v+1)(1+4v-2v^3)} \right).
\end{align}
More generally, we have the following counterpart of~\cite[Cor.~2 and
Cor.~5]{Bousquet2016}.
\begin{prop}[\bf Walks ending at a prescribed position]
\label{cor:coeffs}
Let $w$ be the above defined series in $t$.
For $j\ge 0$, the series $C_{-1,j}$ belongs to $ \qs(t,w)$, and
is thus algebraic.
More generally, for $i\ge 1$ and $j\ge 0$, the series $C_{-i,j}$ is
D-finite of the form
$$
- \frac 1 3 Q_{i-2,j} + \Rat(t,w)
$$
for some rational function $\Rat$. It is transcendental as soon as $i\ge 2$.
For $i\ge 0$ and $j\ge 0$, the series $C_{i,j}$ is D-finite and transcendental of the form
$$
\frac 1 3 Q_{i,j} + \Rat(t,w).
$$
\end{prop}
Another series of interest is $C(1,1)$, which counts all walks in $\mathcal{C}$, regardless of their endpoint. It reads
\[
C(1,1)=A(1,1) - \frac 1 3 Q(1,1),
\]
and we prove that $A(1,1)$ has degree $24$ over $\mathbb{Q}(t)$. A rational
expression for $A(1,1)$ in terms of $v$ and $w$ is given in
Proposition~\ref{prop:univariate_series}. However, $Q(1, 1)$ is
transcendental~\cite{Bostan-etal-2017-qp} hence $C(1, 1)$ is transcendental too.
We also obtain detailed asymptotic results, which refine general results of Denisov--Wachtel~\cite{DenisovWachtel15} and Mustapha~\cite{Mustapha2019Walks}
(who only obtained estimates up to a multiplicative constant).
\begin{coro}
\label{coro:asy}
The number $c(n)$ of $n$-step king walks confined to $\mathcal{C}$ and
ending anywhere, and
the number $c_{0,0}(n)$ of such walks in $\mathcal{C}$ ending at the origin satisfy for $n \to \infty$:
\begin{align*}
c(n) &= \left(\frac{2^{32} K}{3^7}\right)^{\! \! 1/6} \frac{1}{\Gamma(2/3)} \frac{8^n}{n^{1/3}}
- \frac{8}{9 \pi} \frac{8^n}{n} + \mathcal{O}\left(\frac{8^n}{n^{4/3}}\right)
,
\\
c_{0,0}(n) &= \left(\frac{2^{29} K}{3^7} \right)^{\! \! 1/3} \,\frac{\Gamma(2/3)}{\pi} \frac{8^n}{n^{5/3}}
- \left( \frac{2^{62} L}{3^{31}} \right)^{1/6} \frac{1}{\Gamma(2/3) n^{7/3}} + \mathcal{O}\left(\frac{8^n}{n^{8/3}}\right)
,
\end{align*}
where
$K$ and $L$ are the unique real roots of
\[101^6 K^3 - 601275603 \, K^2 + 92811 \, K - 1\]
and
\begin{align*}
101^{18} L^3 &- 342130847546623941461342020714770 \, L^2
\\&
+ 25258724190403343220341683641 \, L - 5078^6.
\end{align*}
\end{coro}
\subsection{Bivariate series}
\label{sec:bivariate}
It remains to describe the series $M(x,0)$ and $M(0,x)$ involved in~\eqref{eqMMM-king}. Both are cubic over $\mathbb{Q}(t,w,x)$, and we express them explicitly in terms of a parametrizing series $U_1$ that satisfies a reasonably compact cubic equation over $\mathbb{Q}(t,v,x)$. Details are given in Proposition~\ref{prop:RhatShat}. We also refer to Figure~\ref{fig:algSR} on page~\pageref{fig:algSR} for the structure of all series involved in the paper.
In Table~\ref{tab:degrees}, we compare the degrees of several relevant algebraic series in the king's model and the simple and diagonal models solved in~\cite{Bousquet2016}\footnote{The details on the series $M(1,0)$, $M(0,1)$, $M(1,1)$, and $A(1,1)$ are not stated in~\cite{Bousquet2016}, but they can be found in the {\sc Maple}\ sessions accompanying this paper on the author's webpage.}. This
gives a hint of the technical difficulties that arise in the solution of the king's model.
\renewcommand{\arraystretch}{1.5}
\begin{table}[hb]
\centering
\begin{tabular}{c|ccccccc}
\toprule
Series & $M(x,y)$ & $M(x,0)$ & $M(0,y)$ & $M(1,1)$& $M(0,1)$ &$M_{0,0}$&
\begin{minipage}{23mm}
$ M(1,0),M_{1,0}$, \\ $A(1,1)$, $A_{0,0}$ \end{minipage}
\\
\hline
Simple/Diag. & 72 & 24 & 12 & 16 & 8 & 4/-- &8\\ \hline
King & 216 & 72 & 72 & 48 & 48 & 24& 24 \\
\bottomrule
\end{tabular}
\vskip 3mm
\caption{A comparison of the degrees of various algebraic series for the simple (or diagonal) model~\cite{Bousquet2016} and for the king's model (this paper). The series $M_{0,0}$ is zero for the diagonal model.}
\label{tab:degrees}
\end{table}
\renewcommand{\arraystretch}{1.0}
\section{The king walks: an equation with only one catalytic variable}
\label{sec:KingEquation}
Our starting point is the functional equation~\eqref{eqfunc-gen}, specialized to
\begin{align*}
S(x,y) &= (x +1+\bar{x})(y+1+\bar{y})-1 \\
&=x + xy + y + \bar{x} y + \bar{x} + \bar{x} \bar{y} + \bar{y} + x \bar{y}.
\end{align*}
We use the $x/y$-symmetry of $S(x,y)$, which induces a
bijection between walks ending on the negative $x$- and $y$-axis,
and implies that
\[
C_{-,0}(\bar{x})= C_{0,-}(\bar{x})=: C_-(\bar{x}).
\]
This gives
\begin{align}
\label{eq:kernelC1}
K(x,y) C(x,y) &= 1 - t \bar{y}(x+1+\bar{x}) C_{-}(\bar{x}) - t \bar{x}(y+1+\bar{y}) C_{-}(\bar{y}) - t \bar{x} \bar{y} C_{0,0},
\end{align}
where as usual, the kernel is
$
K(x,y) = 1 - tS(x,y).
$
Multiplying by $xy$ gives
\begin{align*}
xy K(x,y) C(x,y) &= xy - t (x^2+x+1) C_{-}(\bar{x}) - t (y^2+y+1) C_{-}(\bar{y}) - t C_{0,0}.
\end{align*}
As observed before, the generating function\ $Q(x,y)$ of quadrant walks satisfies similarly:
\[
xy K(x,y) Q(x,y) = xy - t (x^2+x+1) Q(x,0) - t (y^2+y+1) Q(0,y)+ t Q_{0,0}.
\]
The subsequent solution follows the same steps as for the simple walk and the diagonal walk in~\cite{Bousquet2016}.
But in practise, the king model turns out to be much heavier, and raises
serious computational difficulties.
In what follows, we
focus on the points of the derivation that differ
from~\cite{Bousquet2016}. We have performed all computations with the
computer algebra system {\sc Maple}. The corresponding sessions are available on the authors' webpages.
\subsection{A series \texorpdfstring{$A(x,y)$}{A(x,y)} with orbit sum zero}
As discussed in Section~\ref{sec:OS}, and summarized in Proposition~\ref{prop:A-def-gen}, it makes sense to introduce a new series $A(x,y)$ defined by
\begin{equation}\label{AC-def-king}
C(x,y)= A(x,y) + \frac 1 3 \left( Q(x,y) - \bar{x}^2 Q(\bar{x},y) - \bar{y}^2 Q(x,\bar{y})\right).
\end{equation}
Note that any monomial $x^iy^jt^n$ that occurs in $A(x,y)$ is such that $(i,j)\in \mathcal{C}$. Then $xy A(x,y)$ has orbit sum zero, meaning that
\begin{equation}\label{eq:orbitA}
xy A(x,y) - \bar{x} y A(\bar{x},y) + \bar{x} \bar{y} A(\bar{x},\bar{y}) - x \bar{y} A(x,\bar{y}) = 0.
\end{equation}
Moreover, $A(x,y)$ is defined by the functional equation~\eqref{A-func-eq}, which reads:
\begin{align}
\label{eq:kernelA1}
K(x,y) A(x,y) &=
\frac{2+\bar{x}^2+\bar{y}^2}{3}
-t\bar{y}(x+1+\bar{x})A_{-}(\bar{x})
-t\bar{x}(y+1+\bar{y})A_{-}(\bar{y})
-t\bar{x} \bar{y} A_{0,0}.
\end{align}
We now focus on the determination of $A(x,y)$, which should be algebraic according to our final Theorem~\ref{theo:king}. The next step is to split the series $A(x,y)$ into three parts, which involve polynomials in $x$ and $y$ instead of Laurent polynomials.
\subsection{Reduction to a quadrant-like problem for \texorpdfstring{$M(x,y)$}{M(x,y)}}
\label{sec:red_quadrant:king}
We now separate in $A(x,y)$ the contributions of the three quadrants, again using the $x/y$-symmetry of the step set:
\begin{align}
\label{eq:Asep}
A(x,y) = P(x,y) + \bar{x} M(\bar{x},y) + \bar{y} M(\bar{y},x),
\end{align}
where $P(x,y)$ and $M(x,y)$ lie in $\mathbb{Q}[x,y][[t]]$.
Note that this identity defines $P$ and $M$ uniquely in terms of $A$.
Replacing $A$ by this expression, and extracting the positive part in $x$ and $y$ from the orbit equation~\eqref{eq:orbitA} relates the series $P$ and $M$ by
\begin{align}\label{PM-king}
xyP(x,y) &= y \left(M(x,y)-M(0,y)\right) + x \left(M(y,x) - M(0,x)\right),
\end{align}
which is Equation~\eqref{Psol-king} in Theorem~\ref{theo:king}, and also the same as \cite[Eq.~(22)]{Bousquet2016}.
For a combinatorial proof of this equation see Section~\ref{sec:combi}.
We could now follow the lines of proof of~\cite[Sec.~2.3]{Bousquet2016} to obtain the functional equation~\eqref{eqMMM-king} for $M(x,y)$.
However, we prefer to describe a slightly different -- and more combinatorial -- way to derive this equation.
Clearly, $A(x,y)$ counts walks confined to $\mathcal{C}$, starting either from $(0,0)$, $(-2,0)$, or $(0,-2)$, with a weight $2/3$ in the first case and $1/3$ in each of the other two cases. In sight of the splitting~\eqref{eq:Asep} of $A(x,y)$, the series $P(x,y)$ counts such walks ending in the first quadrant, and $\bar{x} M(\bar{x},y)$ those ending at a negative abscissa. By combining these two observations and constructing these walks step by step, we can write directly a pair of equations for $P$ and $M$:
\begin{align*}
K(x,y)P(x,y) &= \frac 2 3 -t\bar{y}(x+1+\bar{x}) P(x,0) -t\bar{x}(y+1+\bar{y})P(0,y)+t\bar{x} \bar{y} P_{0,0}\\
&\quad+ t(x+1+\bar{x}) M(0,x)
-t\bar{x} M_{0,0}+ t(y+1+\bar{y})M(0,y)-t \bar{y} M_{0,0},
\end{align*}
\begin{align}
\label{eqM-king}
\begin{aligned}
K(x,y) \bar{x} M(\bar{x},y) &= \frac 1 3 \bar{x}^2 -t\bar{y}(x+1+\bar{x}) \bar{x} M(\bar{x},0)
-t(y+1+\bar{y})M(0,y)+t \bar{y} M_{0,0}\\
&\quad+ t\bar{x}(y+1+\bar{y})P(0,y)-t\bar{x} \bar{y} P_{0,0}.
\end{aligned}
\end{align}
In the first equation for instance, the term $t(y+1+\bar{y})M(0,y)-t \bar{y} M_{0,0}$ counts walks that come from the NW quadrant and enter the non-negative quadrant through the $y$-axis. We will in fact ignore the first equation and replace it by the link~\eqref{PM-king} between $P$ and $M$. Extracting the coefficient of $x^1$ in~\eqref{PM-king} gives
\[
P(0,y)= M_x(0,y)+ \bar{y}\left( M(y,0)-M_{0,0}\right).
\]
Extracting now the coefficient of $y^0$ gives
\[
P_{0,0}=2 M_x(0,0) =2M_{1,0}.
\]
We plug these two identities into~\eqref{eqM-king}: upon replacing $x$ by $\bar{x}$ and then dividing by~$x$, we find:
\begin{align}
\label{eqM-prelim-king}
\begin{aligned}
K(x,y) M(x,y) &= \frac{1}{3} x -t\bar{y}(x+1+\bar{x}) M(x,0)
-t\bar{x}(y+1+\bar{y})M(0,y)+t \bar{x}\bar{y} M_{0,0}\\
&\quad+ t(y+1+\bar{y})
\left(M_x(0,y)+ \bar{y}\left( M(y,0)-M_{0,0}\right)\right)
-2t \bar{y} M_{1,0}.
\end{aligned}
\end{align}
This is not yet~\eqref{eqMMM-king}, as there is one more series involved here, namely $M_x(0,y)$. However, by extracting the coefficient of $x^0$ in the above equation, we find one more relation:
\[
t\bar{y}(y+1+\bar{y}) M(y,0)+(ty+t\bar{y}-1) M(0,y)+2t(y+1+\bar{y})M_x(0,y)-t\bar{y}(\bar{y}+2+y)M_{0,0}-3t\bar{y} M_{1,0} =0 .
\]
Combined with~\eqref{eqM-prelim-king}, this now gives~\eqref{eqMMM-king}.
\subsection{Cancelling the kernel: an equation between bivariate series}
Next we will cancel the kernel $K$. As a polynomial in $y$,
the kernel admits only one root that is a formal power series in $t$:
\[
Y(x) = \frac{1 - t(x+\bar{x}) - \sqrt{ (1-t(x+\bar{x}))^2-4t^2(x+1+\bar{x})^2 }}{ 2t(x+1+\bar{x})}
= (x+1+\bar{x})t + \mathcal{O}(t^2).
\]
Note that $Y(x)=Y(\bar{x})$.
We specialize~\eqref{eqMMM-king} to the pairs $(x,Y(x)), (\bar{x},Y(x)), (Y(x),x)$, and $(Y(x), \bar{x})$ (the left-hand side vanishes for each specialization since $K(x,y)=K(y,x)$, yet this symmetry is not part of the group of the model), and we eliminate $M(0,Y)$, $M(Y,0)$, and $M(\bar{x},0)$ from the four resulting equations. We obtain:
\begin{align}
\label{eq:3Ms}
\begin{aligned}
(x+1+\bar{x})\left(Y(x)-\frac{1}{Y(x)}\right) & \left( xM(0,x) - 2\bar{x} M(0,\bar{x}) \right)
+ 3(x+1+\bar{x})M(x,0)
\\ & \qquad
-\frac{2\bar{x} Y(x)}{t} + 3M_{1,0}
+ (2Y(x) -x - \bar{x}) M_{0,0} = 0.
\end{aligned}
\end{align}
We have now eliminated the trivariate series $M(x,y)$. We are left with three bivariate series, namely $M(0,x)$, $M(0, \bar{x})$, and $M(x,0)$.
In the next section we eliminate the term $M(x,0)$, so as to end with two specializations of the series $M(0,x)$.
\subsection{An equation between \texorpdfstring{${M(0,x)}$}{M(0,x)} and \texorpdfstring{${M(0,\bar{x})}$}{M(0,1/x)}}
Let us denote the discriminant occurring in $Y(x)$ by
\begin{equation}\label{Delta-def}
\Delta(x) := (1-t(x+\bar{x}))^2-4t^2(x+1+\bar{x})^2
= (1-t(3(x+\bar{x})+2))(1+t(x+\bar{x}+2)),
\end{equation}
and introduce the notation
\begin{align}
\label{eq:RSdef}
\begin{aligned}
R(x) &:= t^2 M(x,0) = \frac{xt^2}{3} + \left(1+\frac{x^2}{3}\right)t^3 + \mathcal{O}(t^4), \\
S(x) &:= txM(0,x) = x(1+x)t^2 + 2x(1+x+x^2)t^3 + \mathcal{O}(t^4).
\end{aligned}
\end{align}
Note that $t^2M_{0,0}=R_0=tS_1$ and $t^2M_{1,0}=R_1$. Then~\eqref{eq:3Ms} reads
\begin{multline}\label{eq:3Ms2}
\sqrt{\Delta(x)} \left( S(x) - 2S(\bar{x}) + \frac{xR_0-t}{t(1+x+x^2)} \right) =
\\ 3(x+1+\bar{x})R(x) + 3R_1
+ \frac{1 - t(x+\bar{x})}{t(1+x+x^2)}(xR_0-t) - (x+\bar{x})R_0
=: \hat{R}(x),
\end{multline}
where we defined $\hat{R}(x)$ as a shorthand for the right-hand side.
Observe that introducing
\begin{align}
\label{eq:STcardanosubs}
\hat{S}(x): = S(x) - \frac{3R_0/t
-2x-\bar{x}}{3(x+1+\bar{x})},
\end{align}
allows us to rewrite the above equation as
\begin{equation}\label{Rh-Sh}
\sqrt{\Delta(x)} \left( \hat{S}(x) - 2\hat{S}(\bar{x}) \right) = \hat{R}(x).
\end{equation}
Before we go into the details of the next steps, let us describe their principle. We consider both sides of~\eqref{eq:3Ms2} as power series in $t$ whose coefficients are Laurent series in $x$. We square Equation~\eqref{eq:3Ms2} and extract the negative part in
$x$, as defined at the beginning of Section~\ref{sec:general}. On the right-hand side, the terms involving
$R(x)$ (mostly) disappear as this series involves only non-negative powers of~$x$. On the left-hand side, the terms involving only $S(x)$ mostly disappear as well. There remain terms involving only $S(\bar{x})$, as well as the negative part of
$\Delta(x)S(x)S(\bar{x})$. In other words, the result is an expression
for the negative part of
$\Delta(x)S(x)S(\bar{x})$ in terms of $S(\bar{x})$ and univariate series. Using the symmetry of $\Delta(x)$ in~$x$ and~$\bar{x}$, we
will then express the \emm positive, part of
$\Delta(x)S(x)S(\bar{x})$ in terms of $S(x)$ and univariate series. We will thus reconstruct an expression of $\Delta(x)S(x)S(\bar{x})$ that
does not involve $R(x)$, as in~\cite[Sec.~2.5]{Bousquet2016}.
In order to make the above programme effective, we need the following lemma, which tells us how to extract the non-negative part of certain series as those that we meet when we square~\eqref{eq:3Ms2}.
\begin{lemma}
\label{lem:pospole}
Let $\zeta=e^{2i\pi/3}$ and $\bar \zeta=e^{-2i\pi/3}$ be the two primitive cubic roots of unity. Let $F(x) \in \mathbb{C}[x]((t))$.
Then,
\[
[x^{\geq}] \frac{F(\bar{x})}{1+x+x^2} =\frac 1 {1-\zeta}\, \frac{F(\zeta)}{1-\zeta x}
+\frac 1 {1-\bzeta}\,\frac{F(\bzeta)}{1-\bzeta x}
\]
and
\begin{multline*}
[x^{\geq}] \frac{F(\bar{x})}{(1+x+x^2)^2} = \frac 2 3
\left(\frac 1 {1-\zeta}\, \frac{F(\zeta)}{1-\zeta x} +\frac 1 {1-\bzeta}\,\frac{F(\bzeta)}{1-\bzeta x} \right)
\\ +\frac 1 {(1-\zeta)^2} \left( \frac{\zeta F'(\zeta)} {1-\zeta x}+ \frac{F(\zeta)}{(1-\zeta x)^2} \right)
+\frac 1 {(1-\bzeta)^2} \left( \frac{\bzeta F'(\bzeta)} {1-\bzeta x} + \frac{F(\bzeta)}{(1-\bzeta x)^2} \right).
\end{multline*}
In fact, the first formula holds for $F(x) \in \bar{x}\mathbb{C}[x]((t))$, and the second for $F(x) \in \bar{x}^3\mathbb{C}[x]((t))$.
\end{lemma}
\begin{proof}
By linearity, it suffices to prove the lemma when $F(x)= x^k$, for $k\ge -1$ in the first part, $k\ge -3$ in the second part. A key ingredient are the following partial fraction expansions:
\[
\frac 1 {1+x+x^2} = \frac 1 {(1-\zeta)(1-\zeta x)}
+ \frac 1 {(1-\bzeta)(1-\bzeta x)},
\]
\[
\frac 1 {(1+x+x^2)^2} =\frac 2 3 \left( \frac 1 {(1-\zeta)(1-\zeta x)}
+ \frac 1 {(1-\bzeta)(1-\bzeta x)}\right)
+ \frac 1 {(1-\zeta)^2(1-\zeta x)^2}
+ \frac 1 {(1-\bzeta)^2(1-\bzeta x)^2} .
\]
Then we work out each piece separately, first focussing on the case $k\ge 0$. For instance,
\[
[x^\ge] \frac{\bar{x}^k}{1-\zeta x}= \bar{x}^k \sum_{n\ge k} \zeta^n x^n= \frac{\zeta^k}{1-\zeta x}
\]
and
\[
[x^\ge] \frac{\bar{x}^k}{(1-\zeta x)^2}= \bar{x}^k \sum_{n\ge k} (n+1)\zeta^n x^n= \zeta^k \sum_{n\ge 0} (k+n+1) \zeta^n x^n= \frac{k\zeta^k}{1-\zeta x} +\frac{\zeta^k}{(1-\zeta x)^2} .
\]
To complete the proof, we check that the first (resp.\ second) identity of the lemma holds as well if $F(x)=\bar{x}^\ell$ for $\ell=1$ (resp.\ $\ell =1,2,3$).
\end{proof}
By expanding~\eqref{eq:3Ms2} at $x=\zeta$ and $x=\bzeta$, we derive the values of $S(x)$ at these two points, which will be useful in sight of the above lemma:
\begin{align}
\label{eq:S1}
&S(\zeta) = S(\bzeta) = - \frac{R_0 + 3 R_1}{1+t} = -t^2 -11t^4 -30t^5 + \mathcal{O}(t^6).
\end{align}
Now, as already observed, the right-hand side $\hat{R}(x)$ of~\eqref{eq:3Ms2} is mostly positive in $x$, meaning that the valuation in $x$ of the coefficient of $t^n$ is bounded from below, uniformly in $n$. We now square both sides. The negative part of (the square of) the right-hand side is easily obtained by an expansion around $x=0$, and found to be
\[
\left( 2 R_0+t \right) ^{2}{\bar{x}}^{2}+2 \bar{x} \left( 2 R_0+t
\right) \left( 2 R_0+6 R_1-1-t \right).
\]
In the square of the left-hand side of~\eqref{eq:3Ms2} some terms are also mostly positive -- in fact all terms that do not involve $S(\bar{x})$. Their negative parts can be extracted as above by an expansion around $x=0$. Some other terms, like $\Delta(x) S(\bar{x})^2$, are mostly negative, and we subtract their non-negative parts, obtained via an expansion at $x=\infty$ (which is legitimate due to their Laurent polynomial coefficients in $x$). And finally there are two tricky terms:
\[
\Delta(x) S(x) S(\bar{x}) \quad \hbox{ and } \quad \Delta(x) \frac{S(\bar{x})(xR_0/t-1)}{1+x+x^2},
\]
which require some care. We leave the first term untouched, since what we want to determine is precisely its negative part. The numerator of the second term is a series in $t$ with coefficients in $\mathbb{Q}[x,\bar{x}]$. We expand it at infinity, using $S_0=0$ and $S_1=R_0/t$, and obtain,
\[
\Delta(x) S(\bar{x})(xR_0/t-1) = -3 R_0^2 \,x^2+ F(\bar{x}),
\]
for a series $F(x) \in \bar{x}\qs[x]((t))$. Then we divide this by $(1+x+x^2)$. The term $x^2/(1+x+x^2)$ has no negative part, and we apply Lemma~\ref{lem:pospole} to express the negative part of $F(\bar{x})/(1+x+x^2)$. After having treated all terms, we reach an identity of the form
\[
[x^<] \big(\Delta(x)S(\bar{x})S(x) \big)=
\Delta(x) S(\bar{x})^2 - \Delta(x) \frac{S(\bar{x})(xR_0-t)}{t(1+x+x^2)}
+\frac{\Pol(R_0,R_1,t,x)}{t x^2(1+x+x^2)}
\]
for some polynomial $\Pol$ with rational coefficients. We can now replace $x$ by $\bar{x}$ to obtain an expression of the positive part of
$\Delta(x)S(\bar{x})S(x)$ (which is a series in $\qs[x, \bar{x}][[t]]$). We finally denote by $P_0$ the coefficient of $x^0$ in $\Delta(x) S(x)S(\bar{x})$,
and obtain an expression of $\Delta(x) S(x)S(\bar{x})$ in terms of $S(x)$, $S(\bar{x})$, $P_0$, $R_0$, and $R_1$, which can be written as:
\begin{align}
\label{eq:2Ms}
\begin{aligned}
\Delta(x) \left( S(x)^2 + S(\bar{x})^2 - S(x) S(\bar{x}) + \frac{S(x)(xt-R_0) + S(\bar{x}) ( \bar{x} t - R_0 )}{t(x+1+\bar{x})} \right)
= \\
(R_0 + 3R_1) \left(
(2R_0+t) \left( x+\bar{x}+ \frac {1+t}{t(x+1+\bar{x})} \right)-1-t
\right) \\
- (1+4t)(x+\bar{x})R_0 + (t^2 + tR_0 + R_0^2) (x^2+\bar{x}^2) - P_0.
\end{aligned}
\end{align}
As in~\cite{Bousquet2016} the numerator of the right-hand side as a polynomial in $x$ is not divisible by $\Delta(x)$, nor by any of its factors.
Observe that~\eqref{eq:2Ms} can also be written in terms of the series $\hat{S}$ defined by~\eqref{eq:STcardanosubs}, and then takes the following form:
\begin{align}
\label{eq:2MsSt}
\Delta(x) \left( \hat{S}(x)^2 - \hat{S}(x) \hat{S}(\bar{x}) + \hat{S}(\bar{x})^2 \right) = \frac{{\Pol}(P_0,R_0, R_1, t,x)}{x^4 t^2 (x+1+\bar{x})^2},
\end{align}
where ${\Pol}$ is another polynomial with rational coefficients.
This simpler form in terms of $\hat{S}$ will guide us in the following final step, in which we eliminate $S(\bar{x})$ and obtain an equation in which $S(x)$ is the only bivariate series.
\subsection{An equation for \texorpdfstring{${M(0,x)}$}{M(0,x)} only}
\label{sec:cat}
We would like to extract the positive part of~\eqref{eq:2Ms}, but we are stopped by the mixed term $S(x) S(\bar{x})$. However,
from the structure visible in~\eqref{eq:2MsSt}, we observe that a multiplication by $\hat{S}(x) + \hat{S}(\bar{x})$
eliminates this mixed term, leaving us with the following cubic
equation in $\hat{S}$:
\[
\Delta(x) \left( \hat{S}(x)^3+ \hat{S}(\bar{x})^3 \right) =
\left( \hat{S}(x)+ \hat{S}(\bar{x})\right)\frac{\Pol(P_0,R_0, R_1, t,x)}{x^4 t^2 (x+1+\bar{x})^2}.
\]
We then rewrite this in terms of $S$ rather than $\hat{S}$, and extract the non-negative part in $x$, using the same tools as in the previous subsection. We refer for full details to the accompanying {\sc Maple}\ worksheet. The terms that are mostly positive or mostly negative in $x$ do not raise any difficulties. The two tricky terms are those that involve $S(\bar{x})^2/(1+x+x^2)$ and $S(\bar{x})/(1+x+x^2)^2$. Their non-negative parts are extracted using Lemma~\ref{lem:pospole}. When processing the latter term, three additional univariate series occur, namely $S_2$, $S'(\zeta)$, and $S'(\bzeta)$. We find it more convenient to work with the real and imaginary parts of $\zeta S'(\zeta)$, and to define series $B_1$ and $B_2$ by
\begin{align} \label{eq:B1B2def}
\begin{aligned}
(1+t)^2 \zeta S'(\zeta) &= B_1 + i\sqrt{3}B_2, \\
(1+t)^2 \bzeta S'(\bzeta) &= B_1 - i\sqrt{3}B_2.
\end{aligned}
\end{align}
We use several times $S_1=R_0/t$ and the expressions~\eqref{eq:S1} of $S(\zeta)$ and $S(\bzeta)$. At the end we obtain a polynomial identity between $S(x)$, $P_0$, $R_0$, $R_1$, $S_2$, $B_1$, $B_2$, $x$, and $t$.
We can reduce to four the number of univariate series involved in this equation as follows. First, we expand the equation around $x=0$ at first order: this gives an expression of $P_0$ in terms of the $5$ other univariate series. We replace $P_0$ by this expression in the functional equation, and now expand at first order around $x=\zeta$: this gives
\[
3t^2 S_2=- 3tR_0-3t^2-2B_1.
\]
In the end we get a cubic equation in $S(x)$:
\begin{align}
\label{eq:Pol1}
\Pol(S(x), R_0, R_1, B_1, B_2, t, x) = 0,
\end{align}
for a polynomial $\Pol(x_0, x_1, x_2, x_3, x_4 ,t, x)$ with rational coefficients. In the terminology of~\cite{BMJ06}, this is an equation with only one \emm catalytic variable,, namely $x$, as opposed to the original functional equation for $M(x,y)$ that had two catalytic variables, $x$ and $y$.
We can describe the above polynomial $\Pol$ in a reasonably compact form thanks to some of its properties: first, when we introduce the series $\hat{S}(x)$ defined by~\eqref{eq:STcardanosubs}, there is no quadratic term (in $\hat{S}(x)$).
Then, the coefficients of the resulting equation are (almost) symmetric in $x$ and $\bar{x}$, and they become symmetric if we introduce the series $\hat{S}(x)/(x-\bar{x})$. Now we can write the equation in terms of a new variable $y:=x+1+\bar{x}$.
Then we observe one more property, namely that the coefficients are (almost) invariant when we replace $y$ by $\bar{y} (1+1/t)$. We refer to our {\sc Maple}\ worksheet for details. If we denote
\begin{equation}\label{z-def}
z= t(x+1+\bar{x} ) + \frac{1+t}{x+1+\bar{x} }
\end{equation}
and
\[
\tilde{S}(x)= \frac{x+1+\bar{x} }{x-\bar{x}}\, \hat{S}(x)= \frac 1{x-\bar{x}} \left( (x+1+\bar{x} ) S(x)- \frac{R_0} t + \frac {2x} 3 + \frac\bar{x} 3\right),
\]
then Equation~\eqref{eq:Pol1} reads:
{
\newcommand{\shrink}[1]{\!#1\!}
\newcommand{\shrink{+}}{\shrink{+}}
\newcommand{\shrink{-}}{\shrink{-}}
\begin{align}
\label{eqSc}
\begin{aligned}
0 &= 27 {t}^{2} \left( 2 t\shrink{+} z\shrink{+} 1 \right) \left( 10 t\shrink{-} 3 z\shrink{+} 1 \right) {\tilde S(x)}^{3} \\
& +\big( \left( 216 {t}^{2} \shrink{-} 27 {z}^{2} \shrink{+} 54 t \right) R_0^{2} \shrink{+} 27 t \left( 6 R_1 t\shrink{-} 6 R_1 z\shrink{+} 6 {t}^{2} \shrink{+} 2 zt \shrink{-} {z}^{2} \shrink{+} 2 B_2\shrink{+} t\shrink{+} z \right) R_0
\\
& \quad - 9 {t}^{2} \left( 27 R_1^{2} \shrink{-} 27 R_1 t\shrink{+} 9 R_1 z\shrink{+} 5 {t}^{2} \shrink{-} 2 zt\shrink{+} 3 B_1\shrink{-} 3 B_2\shrink{-} 9 R_1\shrink{+} 6 t\shrink{-} 2 z\shrink{+} 1 \right) \big) \tilde S(x)\\
& +\left( 72 {t}^{2} \shrink{-} 9 {z}^{2} \shrink{+} 18 B_1\shrink{+} 18 t \right) R_0^{2} \shrink{+} 9 t \left( 6 R_1 t\shrink{-} 6 R_1 z\shrink{+} 6 {t}^{2} \shrink{+} 2 zt \shrink{-} {z}^{2} \shrink{+} 2 B_1\shrink{+} 2 B_2\shrink{+} t\shrink{+} z \right) R_0 \\
& \quad -{t}^{2} \left( 81 R_1^{2} \shrink{-} 81 R_1 t\shrink{+} 27 R_1 z\shrink{-} 5 {t}^{2} \shrink{-} 10 zt\shrink{+} 3 {z}^{2} \shrink{-} 9 B_1\shrink{-} 9 B_2\shrink{-} 27 R_1\shrink{+} 6 t\shrink{-} 4 z\shrink{+} 2 \right).
\end{aligned}
\end{align}
}
\section{The king walks: algebraicity}
\label{sec:quadratic}
In~\cite{BMJ06}, a general method to solve equations in one catalytic variable was developed, proving in particular that their solutions are systematically algebraic (provided the equation is \emm proper, in a certain natural sense). In Section~\ref{sec:system} we first use the results
of~\cite{BMJ06} to obtain a system of four polynomial equations relating the series $R_0, R_1, B_1$, and $B_2$. Combined with a few initial terms, this system characterizes these four series.
Unfortunately, it turns out to be too big for us to
obtain individual equations for each of the four series, be it by bare hand elimination or using Gr\"obner bases: we did obtain polynomial equations for $R_0$ and $R_1$, of degree $24$ in each case, but not for the other two series.
Instead, as detailed in Section~\ref{sec:guess}, we have resorted to a guess-and-check approach,
consisting in \emph{guessing} such equations (of degree $12$ or $24$,
depending on the series), and then \emph{checking} that they satisfy the
system.
\subsection{A polynomial system relating \texorpdfstring{$R_0$, $R_1$, $B_1$, and $B_2$}{R0, R1, B1, and B2}}
\label{sec:system}
We start from the cubic equation~\eqref{eq:Pol1}. The approach of~\cite{BMJ06} instructs us to consider the series $X$ (in $t$, or in a fractional power of $t$)), satisfying
\begin{equation}\label{dx0}
\Pol_{x_0}(S(X), R_0, R_1, B_1, B_2, t, X) = 0,
\end{equation}
where $\Pol_{x_0}$ stands for the derivative of $\Pol$ with respect to its
first variable.
The number of such series $X$ and their first terms depend only
on the first terms of the series $S(x)$, $R_0$, $R_1$, $B_1$, and $B_2$; see~\cite[Thm.~2]{BMJ06}. We find that $6$ such series exist:
\begin{align*}
X_1(t) &= i+2t^2+4t^3+(36-2i)t^4+ \mathcal{O}(t^5), \\
X_2(t) &=
-i+2t^2+4t^3+(36+2i)t^4+ \mathcal{O}(t^5), \\
X_3(t) &= \sqrt{t}+t+\frac{3}{2}t^{3/2} + 3t^2 + \frac{51}{8}t^{5/2} + 14t^3 + \mathcal{O}(t^{7/2}), \\
X_4(t) &= -\sqrt{t}+t-\frac{3}{2}t^{3/2} + 3t^2 -
\frac{51}{8}t^{5/2} + 14t^3 + \mathcal{O}(t^{7/2}), \\
X_5(t) &= i \sqrt{t}-it^{3/2}+2i t^{5/2} + t^3 - 4i t^{7/2} + 2t^4 + \mathcal{O}(t^{9/2}),\\
X_6(t) &
-i \sqrt{t}+it^{3/2}-2i t^{5/2} + t^3 + 4i t^{7/2} + 2t^4 + \mathcal{O}(t^{9/2}).
\end{align*}
Note that the coefficients of $X_1$ and $X_2$ (resp.~$X_5$ and $X_6$)
are conjugates of one another. As discussed in~\cite{BMJ06}, each of these
series $X$ also satisfies
\begin{equation}\label{dx}
\Pol_{x}(S(X), R_0, R_1, B_1, B_2, t, X) = 0,
\end{equation}
where $\Pol_x$ is the derivative with respect to the last variable of $\Pol$, and of course
\begin{equation}\label{original}
\Pol(S(X), R_0, R_1, B_1, B_2, t, X) = 0.
\end{equation}
Using this, we can easily identify two of the series $X_i$: indeed,
eliminating $B_1$ and $B_2$ between the three equations~\eqref{dx0},~\eqref{dx}, and~\eqref{original} gives a polynomial equation between $S(X)$, $R_0$, $R_1$, $t$, and $X$, which factors. Remarkably, its simplest non-trivial factor only
involves $t$ and $X$, and reads
\vspace{-1mm}
\begin{equation}\label{X34}
X^2-t(1+X)^2(1+X^2).
\end{equation}
By looking at the first terms of the $X_i$'s and at the other
factors, one concludes that the above equation holds for $X_3$ and $X_4$, which are thus explicit. The other four series $X_i$ satisfy another equation in $S(X)$, $X$, $R_0$, $R_1$, and $t$, which we will not use.
Let $D(x_1,\ldots,x_4,t,x)$ be the discriminant of $\operatorname{Pol}(x_0,\ldots,x_4,t,x)$ with respect to $x_0$. According to~\cite[Thm.~14]{BMJ06},
each $X_i$ is a {\emph{double root}} of $D(R_0,R_1, B_1,B_2,t,x)$, seen as a polynomial in $x$.
Hence this polynomial, which involves four unknown series $R_0, R_1, B_1,B_2$, has (at least)
$6$ double roots. This seems more information than we need.
However, we shall see that there is some redundancy in the $6$
series $X_i$, which comes from the properties of $\Pol$
that we used at the end of Section~\ref{sec:cat} to write it in a compact form.
We first observe that $D$ factors as
\[
D(R_0,R_1, B_1,B_2,t,x)= 27x^2t^2(1+x+x^2)^2 \Delta(x)
D_1(R_0,R_1, B_1,B_2,t,x),
\]
where $\Delta(x)$ is defined by~\eqref{Delta-def}, and $D_1$ has
degree $24$ in $x$. It is easily checked that none of the $X_i$'s are
roots of the prefactors, so they are double roots of $D_1$. But
we observe that $\bar{x}^{12}D_1$ is symmetric in $x$ and $\bar{x}$. That is,
\[
D_1(R_0,R_1, B_1,B_2,t,x)= x^{12}D_2(R_0,R_1, B_1,B_2,t,x+1+\bar{x}),
\]
for some polynomial $D_2(x_1, \ldots, x_4,t,y)\equiv D_2(y)$ of degree 12 in
$y$. Since each $X_i$ is a double root of $D_1$, each series
$Y_i:=X_i+1+1/X_i$, for $1\le i\le 6$, is a double root of $D_2$. The series $Y_i$, for $2\le i \le 6$, are easily seen from their first
terms to be distinct, but the first terms of $Y_1$ and $Y_2$
suspiciously agree: one suspects (and rightly so), that $X_2=1/X_1$,
and carefully concludes that $D_2$ has (at least) $5$ double roots in
$y$. Moreover, since $X_3$ and $X_4$ satisfy~\eqref{X34}, the
corresponding series $Y_3$ and $Y_4$ are the roots of $ 1+t=tY_i^2$,
that is, $Y_{3,4}=\pm \sqrt{1+1/t}$. The other roots start as follows:
\[
Y_2= 1+4t^2+8t^3 + \mathcal{O}(t^4),\qquad
Y_{5,6}= \mp \frac i{\sqrt t} + 1+ t^2 \pm i t^{5/2} + \mathcal{O}(t^3).
\]
This is not yet the end of the story: indeed, $D_2$ appears to be
almost symmetric in $y$ and $1/y$.
More precisely, we observe that
\[
D_2(R_0,R_1, B_1,B_2,y)= y^6 D_3\left(R_0,R_1, B_1,B_2,
ty+\frac{t+1}y\right),
\]
for some polynomial $D_3(R_0,R_1, B_1,B_2,t,z)\equiv D_3(z)$ of degree $6$ in $z$. It follows that each series $Z_i:= tY_i+(1+t)/Y_i$, for $2\le i \le 6$, is a root of $D_3(z)$, and even a double root, unless $tY_i^2=1+t$, which precisely occurs for $i=3,4$. One finds $Z_{3,4}= \pm 2\sqrt{t(1+t)}$,
\[
Z_2= 1+ 2t -4 t^2 +\mathcal{O}(t^3),\qquad
Z_{5,6}= 2t+2t^3 +\mathcal{O}(t^4).
\]
Since $Z_5$ and $Z_6$ seem indistinguishable, we conclude that $D_3(z)$ has (at least) two double roots $Z_2$ and $Z_5$, and a factor $(z^2-4t(1+t))$ coming from the simple roots at $Z_3$ and $Z_4$.
We
\pagebreak
\noindent can thus write
\[
D_3(z)= \sum_{i=0}^6 d_i z^i= \left({z}^{2}-4\,t(1+t) \right) (\alpha z^2 +\beta z +\gamma)^2,
\]
where the $d_i$ are explicit in terms of $R_0$, $R_1$, $B_1$, and $B_2$. We can determine $\alpha$, $\beta$, and $\gamma$ in terms of the $d_i$ by matching the three monomials of highest degree, and this gives:
\[
D_3(z)= \sum_{i=0}^6 d_i z^i= {\frac { \left({z}^{2}-4\,t(1+t) \right) \left( 8\,{z}^{2}d_{6}^{2}+4\,zd_{{5}}d_{{
6}}+16\,{t}^{2}d_{6}^{2}+16\,td_{6}^{2}+4\,d_{{4}}d_{{6}}-d_{5}^{2} \right) ^{2}}{64\,d_{6}^{3}}}.
\]
Extracting from this identity the coefficients of $z^0, \ldots, z^3$ gives four polynomial
relations between the coefficients~$d_i$, resulting in four polynomial
relations between the four series $R_0$, $R_1$, $B_1$, and $B_2$.
We give below the degrees and number of terms in each of them.
\begin{table}[ht]
\centering
\begin{tabular}{l@{\hskip 8mm}ccccc|c}
\toprule
Degree in &$R_0$& $R_1$& $B_1$& $B_2$ & $t$ &Number of terms\\
\midrule
Eq.~$1$ & $5$ & $3$ & $1$ & $1$ & $7$ & $72$ \\
Eq.~$2$ & $6$ & $4$ & $2$ & $2$ & $7$ & $132$ \\
Eq.~$3$ & $5$ & $5$ & $2$ & $2$ & $9$ & $192$ \\
Eq.~$4$ & $6$ & $6$ & $3$ & $3$ & $10$ & $276$ \\
\bottomrule
\end{tabular}
\vskip 4mm
\caption{Properties of the four polynomial equations defining the four main unknown series $R_0$, $R_1$, $B_1$, and $B_2$.}
\label{tab:system}
\end{table}
\vspace{-3mm}
We will now check that the solution of this system is unique if we add the conditions $R_0=\mathcal{O}(t^3)$, $R_1=\mathcal{O}(t^2)$, $B_1=\mathcal{O}(t^2)$, $B_2=\mathcal{O}(t^2)$, which are directly deduced from the definitions of $R(x)$, $B_1$, and $B_2$ in~\eqref{eq:RSdef} and~\eqref{eq:B1B2def}. We write accordingly $R_0=t^3 \tilde R_0$, $R_1=t^2 \tilde R_1$, $B_1=t^2 \tilde B_1$, $B_2=t^2 \tilde B_2$ in the system, divide each equation by a power of $t$ so that it becomes non-trivial at $t=0$ (and, as it happens, linear in each series at this point). We finally form linear combinations of these four equations so that the system, evaluated at $t=0$, is triangular. We refer again to our {\sc Maple} \ sessions for details.
As explained at the beginning of this subsection, we have been able to derive directly from this system polynomial equations (of degree $24$) for $R_0$ and $R_1$ by successive eliminations, but not for the other two series. At the end we resorted to a guess-and-check approach.
\subsection{Guess-and-check}
\label{sec:guess}
The functional equation~\eqref{eq:kernelA1} defining $A(x,y)$ encodes a simple recurrence for the numbers $a_{i,j}(n)$ that count (weighted) walks of length $n$ by the positions of their endpoints $(i,j) \in \mathcal{C}$:
\begin{align*}
a_{i,j}(n+1) &= a_{i-1,j-1}(n) + a_{i-1,j}(n) + a_{i-1,j+1}(n) + a_{i,j-1}(n) \\ & \quad + a_{i,j+1}(n) + a_{i+1,j-1}(n)
+ a_{i+1,j}(n) + a_{i+1,j+1}(n) ,
\end{align*}
with $a_{i,j}(n)=0$ for $(i,j) \not\in \mathcal{C}$ and initial conditions $a_{0,0}(0)=2/3$, $a_{-2,0}(0)=a_{0,-2}(0)=1/3$, and $a_{i,j}(0)=0$ otherwise.
We implemented this recurrence in the programming language~$C$ using modular arithmetic and the Chinese remainder theorem to compute
these numbers up to $n=2000$ (this effectively bounds $i$ and $j$ to $2000$ as well, since $a_{i,j}(n)=0$ if $i>n$ or $j>n$).
For this purpose, we used approximately $100$ primes of size $\approx 2^{64}$, and we actually computed $3A(x,y)$ rather than $A(x,y)$, as it has integer coefficients.
The series $R_0$, $R_1$, $B_1$, and $B_2$ are related to $A(x,y)$ as follows. First, observe that by~\eqref{eq:RSdef} it holds that $R_0= t^2 A_{-1,0}$ and $R_1=t^2 A_{-2,0}$.
Second, for $B_1$ and $B_2$ defined in~\eqref{eq:B1B2def}, we also start from~\eqref{eq:RSdef}, which implies that
$S'(\zeta) = t M(0,\zeta) + t\zeta M_{y}(0,\zeta)$ where $M(0,y) = \sum_{j \geq 0} A_{-1,j} y^j$.
In order to compute $M(0,\zeta)$ we used $\zeta^2 = -1-\zeta$, with $\zeta=(-1+i\sqrt 3)/2$, which implies that $6M(0,\zeta) = \alpha_1 + i\sqrt{3}\alpha_2$ with $\alpha_1,\alpha_2 \in {\mathbb Z}[[t]]$. Hence, the initial coefficients of the series $\alpha_1$ and $\alpha_2$ may be computed using modular arithmetic. The same holds for $\zeta M_y(0,\zeta)$, which then allows to reconstruct the coefficients of $B_1$ and $B_2$. Then we were able to guess polynomial equations satisfied by
$R_0$, $R_1$, $B_1$, and $B_2$ using the \texttt{gfun} package in {{\sc Maple}}~\cite{gfun}.
We refer for full details to the accompanying {\sc Maple}\ worksheet.
Of course, the equations obtained for $R_0$ and $R_1$ coincide with those that we derived from the system of the previous subsection.
Details on the corresponding equations are shown in Table~\ref{tab:guessed}. We note that the degree $24$ equation for $B_2$ is in fact a degree~$12$ equation for $B_2^2$.
\begin{table}[ht!]
\centering
\begin{tabular}{cccc}
\toprule
Generating function & Degree in $GF$ & Degree in $t$ & Number of terms \\
\hline
$R_0$ & $24$ & $36$ & $323$ \\
$R_1$ & $24$ & $36$ & $623$
\\
$B_1$ & $12$ & $24$ & $229$ \\
$B_2$ & $24$ & $60$ & $477$ \\
\bottomrule
\end{tabular}
\vskip 4mm
\caption{Properties of the guessed polynomial equations for the four main unknown series $R_0$, $R_1$, $B_1$, $B_2$.}
\label{tab:guessed}
\end{table}
\vspace{-2mm}
We now have to check that the guessed series satisfy the system obtained in the previous subsection. This turns out to be much easier once the algebraic structure of these series is elucidated. We explain in Appendix~\ref{app:subextensions} how this can be done. We believe that this can be of interest to readers handling algebraic series of large degree. After this step, one obtains expressions for $R_0$, $R_1$, $B_1$, and $B_2$ in terms of the series $v$ and $w$ of Section~\ref{sec:univariate}. We have not tried a direct check of the system based on the four guessed equations of Table~\ref{tab:guessed}.
\begin{prop}\label{prop:explicit-univariate}
Let $u,v, w \in \qs[[t]]$ be the series defined in
Section~\ref{sec:univariate} by~\eqref{u-def},~\eqref{v-def},
and~\eqref{w-def}, respectively. Then the four series that occur in
the equation in one catalytic variable defining $S(x)$ are:
\allowdisplaybreaks
\newcommand{\\[1.5mm]}{\\[1.5mm]}
\begin{align*}
R_0 &= \frac{t}{2} \left( \frac{w(1+2v)}{1+4v-2v^3}-1\right), \\[1.5mm]
R_1 &= \frac{1}{6}\left( 1+2t + \frac{(1-2t)(1+2v)(16v^6+24v^5+7v^4-24v^3-30v^2-10v-1)}{w(v^4+8v^3+6v^2+2v+1)(1+4v-2v^3)} \right), \\[1.5mm]
B_1 &= \frac{3v^2 (1-8t) (1+4v+v^2)(v^2-1)(1+2v)}{2(1-3v^2-4v^3)^3(1+4v-2v^3)},\\[1.5mm]
B_2 &= \frac{(1+2v)(1-2t)}{2w(v^4+8v^3+6v^2+2v+1)^2(2v^3-4v-1)}
\left(4tv^{12}+68tv^{11}+16(22t+1)v^{10} \right. \\
&\quad \left. +12(67t+2)v^9+5(192t-5)v^8+8(61t-10)v^7
-(286t+41)v^6 -2(394t-33)v^5 \right. \\
&\quad \left.
-(738t-113)v^4 -4(97t-17)v^3 -(126t-19)v^2
-2(12t-1)v -2t \right).
\end{align*}
\end{prop}
\begin{proof}
It suffices to check that the four series above satisfy the initial conditions $R_0=\mathcal{O}(t^3)$, $R_1=\mathcal{O}(t^2)$, $B_1=\mathcal{O}(t^2)$, $B_2=\mathcal{O}(t^2)$, and the system of 4 polynomial equations established in Section~\ref{sec:system}, the properties of which are summarized in Table~\ref{tab:system}. The first point is straightforward. Then we take each equation of the system in turn, replace the four unknown series by the above expressions, take the numerator of the resulting equation (which is a polynomial in $t$, $v$, and~$w$), and reduce it first modulo Equation~\eqref{w-def} defining $w$ over $\qs(v)$. In each case, we note that the remainder does not involve $w$, an encouraging sign. Then we reduce further modulo Equation~\eqref{algv} defining $v$ over $\qs(t)$. In each case, we find zero, so that the system holds for the above values of $R_0$, $R_1$, $B_1$, and $B_2$. This completes the proof.
\end{proof}
Note that this proves in particular the announced expressions~\eqref{Cm10-expr} and~\eqref{Mm10-expr} for the series $M_{0,0}=R_0/t^2$ and $M_{1,0}=R_1/t^2$; see~\eqref{eq:RSdef}. We claim that at this stage, we have proved the algebraicity of the series $P(x,y)$ and $M(x,y)$. Recall that by definition, walks in $\mathcal{C}$ ending in the first quadrant (resp.\ at negative abscissa) have generating functions
\[
\frac 1 3 Q(x,y) + P(x,y), \qquad \left(\hbox{resp.} -\frac 1 3
\bar{x}^2 Q(\bar{x},y) + \bar{x} M(\bar{x},y)
\right).
\]
\begin{coro}
The series $P(x,y)$ and $M(x,y)$ are algebraic over $\mathbb{Q}(t,x,y)$.
\end{coro}
\begin{proof}
We work our way backwards starting from the 4 univariate algebraic series of Proposition~\ref{prop:explicit-univariate}. Since $S(x)=tx M(0,x)$ satisfies a cubic equation $\Pol(S(x), R_0, R_1, B_1, B_2, t, x) = 0$,
where the polynomial $\Pol$
has non-zero leading coefficient in its first variable, $S(x)$ and $M(0,x)$ are algebraic of degree at most $72$. We will see that this bound is tight. It then follows from~\eqref{eq:3Ms2} that $R(x)=t^2M(x,0)$ is algebraic as well. We now return to~\eqref{eqMMM-king}, which expresses $M(x,y)$: since $M_{0,0}=R_0/t^2$ and $M_{1,0}=R_1/t^2$, we conclude that $M(x,y)$ is algebraic. We finally use the relation~\eqref{Psol-king} between $P(x,y)$ and $M(x,y)$ to conclude that $P(x,y)$ is algebraic.
\end{proof}
In the next subsection, we determine the degree of all algebraic series of interest, and give closed form expressions for $S(x)$ and $R(x)$ in terms of the already defined series $v$ and $w$, and a ``simple'' cubic extension of $\mathbb{Q}(t,v,x)$.
\subsection{Back to \texorpdfstring{${S(x)}$}{S(x)} and \texorpdfstring{${R(x)}$}{R(x)}}
\label{sec:bivariate-sol}
In this subsection, we prove that $S(x)$ and $R(x)$ belong to the same cubic extension of $\mathbb{Q}(t,w,x)$, and describe this extension in (reasonably) compact terms.
We give two descriptions of this extension by rational parametrizations (in fact, a third one hides in Appendix~\ref{app:P0}). Remarkably, they define cubic extensions of $\mathbb{Q}(t,v,x)$ rather than $\mathbb{Q}(t,w,x)$.
The first one is in terms of the variable $\tilde y:= t(x+\bar{x}+1)/(1-2t)$ and involves $v$ but not $t$. The second one, however, involves the original variable $x$, and now $t$ and $v$.
More precisely, let $U_1\equiv U_1(x)$ be the unique series of the form $U_1=xt^2+\mathcal{O}(t^3)$ satisfying
\begin{equation}\label{P1-param}
\tilde y = \frac\kappa {U_1} \frac{N(U_1)}{N(r_1/U_1)},
\end{equation}
where
\[
N(U)=
U+{v}^{2}w^2 -{\frac {
{v}^{4} w^2 \left( v^2-1 \right) \left( {v}^{2}+v+1
\right) }{U}},
\]
with $w^2=1+4v-4v^3-4v^4$ as before, and
\begin{align}
\kappa&={\frac { \left( {v}^{3}-3\,v-1 \right) ^{2} \left( {v}^{2}+v+1
\right) {v}^{2}}{{v}^{4}+8\,{v}^{3}+6\,{v}^{2}+2\,v+1}}, \notag \\
\quad
r_1&=-{v}^{3}w^2 \left( {v}^{2}+v+1 \right) \left( {v}^{3}-3\,v-1 \right). \label{kappa-r1}
\end{align}
Let $U_2\equiv U_2(x)$ be the unique series $U_2=\bar{x}+\mathcal{O}(t)$ that satisfies
\begin{equation}\label{xP2}
x=U_2\, \frac{M(U_2)}{M(1/U_2)},
\end{equation}
where $M(U)= 1/U+\alpha+\beta U$, with
\[
\alpha= v -{\frac { \left( {v}^{3}-3\,v-1 \right) \beta}{{v}^{2}+v+1}}
\]
and
\[
\beta = \frac{\left( {v}^{2}+v+1 \right)
\left( \left( 2\,{v}^{5}+15\,{v}^{4}+20\,{v}^{3}
+16\,{v}^{2}+6\,v+1 \right) t+ v \left( {v}
^{3}-3\,v-1 \right)\right)
}
{t \left( {v}^{4}+8\,{v}^{3}+6\,{v}^{2}+2\,v+1 \right) \left( 2\,{v}^{3
}+3\,{v}^{2}+6\,v+1 \right)
}.
\]
The series $U_1$ and $U_2$ generate the same cubic extension of $\mathbb{Q}(t,v,x)$. In particular,
\begin{equation}\label{P12}
\frac 1 U_1 = a\left( U_2+ \frac 1 U_2 + {\frac {{v}^{2}+4\,v+1}{{v}^{2}+v+1}}\right)
\end{equation}
with
\[
a={\frac {\beta\,t \left( {v}^{4}+8\,{v}^{3}+6\,{v}^{2}+2\,v+1 \right)
}{ {v}^{3}(1-2t) \left( v^2-1 \right) \left( {v}^{2}+v+1 \right)
\left( {v}^{3}-3\,v-1 \right) }}.
\]
One can also express $U_2$ as an element of $\mathbb{Q}(t,v,x,U_1)$ by combining~\eqref{xP2} and~\eqref{P12}. Finally, one can check that $U_1$ and $U_2$ have degree $36=12\times 3$ over $\mathbb{Q}(t,x)$. Therefore, we have $\mathbb{Q}(t,v,x,U_1)=\mathbb{Q}(t, x, U_1)=\mathbb{Q}(t,x,U_2)$.
\begin{prop}\label{prop:RhatShat}
Let $v$ and $w$ be the series of $\mathbb{Q}[[t]]$ defined by~\eqref{v-def} and~\eqref{w-def}. Let $U_1(x)$ and $U_2(x)$ be defined above. The series $R(x)=t^2M(x,0)$ and $S(x)=txM(0,x)$ are algebraic of degree $72$ over
$\qs(x,t)$ and belong to $\qs(t,x,w,U_1)=\qs(t,x,w,U_2)$.
More precisely, the series
\begin{equation}\label{St-def}
\tilde{S}(x)= \frac 1{x-\bar{x}} \left( (x+1+\bar{x} ) S(x)- \frac{R_0} t + \frac {2x} 3 + \frac\bar{x} 3\right)
\end{equation}
and
\begin{equation}\label{eq:Rhat}
\hat{R}(x)= 3(x+1+\bar{x})R(x) + 3R_1
+ \frac{1 - t\bar{x}(x+\bar{x})(x+1)^2}{t(x+1+\bar{x})}(R_0-t\bar{x}) + t(1+\bar{x}^2)
\end{equation}
belong respectively to $\qs(t,x,U_1 (x))$
and $w\,\qs(t,x,U_1 (x))$. In particular,
\[
\tilde S(x)+\frac 1 3 =-
\frac{{v}^{2}w^2 (1+2v)\left( {v}^{2}+4\,v+1 \right) ^{2}}{2v^3-4v-1}
\frac{1 }{D(U_1) D(r_1/U_1)}
\]
where $r_1$ is given by~\eqref{kappa-r1} and
\[
D(U)=(v+1) U + v w^2(v^2-1) + (v-1) \frac{r_1} U.
\]
Recall that $R_0$ and $R_1$ lie in $\mathbb{Q}(t,w)$, and are given by Proposition~\ref{prop:explicit-univariate}.
\end{prop}
\begin{proof}
We return to the cubic equation that defines $S(x)$, written in the form~\eqref{eqSc} in terms of~$z$ and $\tilde{S}(x)$ and we replace $R_0$, $R_1$, $B_1$, $B_2$ by their expressions in terms of $t$, $v$, and $w$.
Then we observe that only even powers of $w$ occur: hence, using the defining equation~\eqref{w-def} of $w$, we obtain a cubic equation for $\tilde S(x)$ involving only $t$, $v$, and of course the variable $z$ defined by~\eqref{z-def}. This equation has degree
$2$ in~$z$. We lower the degree in $t$ to $1$ using the minimal polynomial~\eqref{algv} of $v$. Now the coefficient of $z^2$ does not involve $t$, the coefficient of $z^1$ is a multiple of $(1-2t)$, and the coefficient of $z^0$ is a multiple of $(1-8t)$. But observe that the minimal equation of $v$ can also be written as
\[
\frac{1-8t}{(1-2t)^2}= {\frac { \left( {v}^{2}+4\,v+1 \right) \left( 4\,{v}^{3}+3\,{v}^{2}-1
\right) ^{3}}{ \left( 4\,{v}^{4}+4\,{v}^{3}-4\,v-1 \right) \left( {v
}^{4}+8\,{v}^{3}+6\,{v}^{2}+2\,v+1 \right) ^{2}}}.
\]
This gives a cubic equation for $\tilde S(x)$, with coefficients in $\mathbb{Q}(v, \tilde z)$ where
\begin{equation}\label{zt-def}
\tilde z= \frac z{1-2t}= \frac{yt}{1-2t} + \frac{1+t}{y(1-2t)},
\end{equation}
where as before $y=x+1+\bar{x}$. It is remarkable that this equation does \emph{not} involve $t$. Its genus (in $ \tilde z$ and $\tilde S$) is found to be zero and thus this equation admits a rational parametrization. We give one in Appendix~\ref{app:P0} (see~\eqref{zt-P0}), in terms of a series denoted by $U_0(x)$, for which we have
\[
\tilde S(x)+\frac 1 3 =
- {\frac {{v}^{2} \left( v^2-1 \right) \left( 2 v+1 \right) \left( {v}^{2}+4 v+1 \right) ^{2}}{ \left( 2 {v}^{3}-4 v-
1 \right)
\left(w^2{U_0}^{2}+{v}^{2}
\left( v^2-1 \right) \left( 2 v+1 \right) \left( 2 {v}^{3}+3 {v}^{2}+6 v+1 \right)\right)
}}.
\]
But it may be better to parametrize our extensions in terms of $x$ than $\tilde z$. Let us first get back to $y=x+1+\bar{x}$, or rather to $\tilde y = {yt}/({1-2t})$, and observe that $\tilde z$ can be written as
\begin{equation}\label{zt-yt}
\tilde z= \tilde y + \frac {t(1+t)}{(1-2t)^2}\frac 1 {\tilde y}= \tilde y + \frac q{\tilde y}
\end{equation}
where
\[
q= -{\frac {v \left( {v}^{2}+v+1 \right) \left( {v}^{3}-3\,v-1 \right) ^{3}}
{
w^2 \left( {v}^{4}+8\,{v
}^{3}+6\,{v}^{2}+2\,v+1 \right) ^{2}}}
.
\]
This means that $\tilde S(x)$ also satisfies a cubic equation with coefficients
in $\mathbb{Q}(\tilde y, v)$, again not involving $t$.
This equation is also found to have genus $0$ (in $\tilde y$ and $\tilde S$) and can be parametrized rationally by introducing the series $U_1$ defined by~\eqref{P1-param}. Indeed, if, in the equation relating $\tilde y $ and $\tilde S$, we replace $\tilde y$ by its expression in terms of $U_1$, the equation factors into a linear term in $\tilde S$, and a quadratic one. Provided we choose the correct determination of $U_1$, given by $U_1=xt^2+\mathcal{O}(t^3)$, then the term that vanishes is the linear one, and this gives the expression of $\tilde S$ stated in the proposition.
Observe that replacing $U_1$ by $r_1/U_1$ in~\eqref{P1-param}
replaces $\tilde y $ by $q/\tilde y$ (because
$\kappa^2=q r_1$), and thus leaves~$\tilde z $ unchanged; see~\eqref{zt-yt}. Analogously, the series $U_0(x)$ that parametrizes the equation in $\tilde z$ and $\tilde S$ (see Appendix~\ref{app:P0}) is invariant by this transformation, and reads
\[
U_0=\frac{1-v^2}{w^2}\left(U_1 +v^2w^2 +\frac{r_1}{U_1}\right).
\]
One can actually go even further, as the equation that relates the original variable $x$ and the series $U_1$ (now with coefficients in $\mathbb{Q}(t,v)$) also has genus zero. It can be parametrized by introducing the series $U_2$ defined by~\eqref{xP2}. Indeed, if we replace, in the equation relating~$x$ and~$U_1$, the variable $x$ by its expression in terms of $U_2$, we observe again a factorization, which leads to~\eqref{P12} once the correct determination of $U_2$ is chosen.
One readily checks that $U_1$ and $U_2$ (and $U_0$ as well) have degree $36$ over $\mathbb{Q}(t,x)$.
Thus $\tilde S(x)$ belongs to $\mathbb{Q}(t,x,U_1)=\mathbb{Q}(t,x,U_2)$, while $S(x)$, which involves $R_0$ and hence~$w$, belongs to $\mathbb{Q}(t,w,x,U_1)=\mathbb{Q}(t,w,x,U_2)$ and has degree at most $72$. To prove that this bound is tight, one can eliminate $w$ and $v$ in the equation defining $S(x)$. It is enough to do it for $x=2$, for instance, as we find that $S(2)$ has degree $72$.
\medskip
We now wish to determine the series $R(x)=t^2M(x,0)$, which is expressed in terms of $S(x)$ and $S(\bar{x})$ in~\eqref{Rh-Sh}. Equivalently,
\[
\hat R (x)= \frac{(x-\bar{x})\sqrt{\Delta(x)}}{x+1+\bar{x} } \left( \tilde{S}(x) +2 \tilde{S}(\bar{x})\right).
\]
We could of course eliminate $\tilde{S}(x)$ and $\tilde{S}(\bar{x})$ to determine a polynomial equation satisfied by $\hat{R}(x)$ over $\qs(t,x,v)$, but there is an algebraic structure in the above equation, which will
save us these calculations. Let us denote $\Pol(s)=s^3+ps+q$ the monic minimal polynomial of $\tilde{S}(x)$ over $\mathbb{Q}(\tilde z,v)$. One of its root is of course $s_1=\tilde{S}(x)$, another one is $s_2=\tilde{S}(\bar{x})$ (because $\tilde z$ is invariant under $x\mapsto \bar{x}$) and the third one is $s_3=-\tilde{S}(x)-\tilde{S}(\bar{x})$ (because there is no quadratic term in $\Pol$). Hence $ \tilde{S}(x) +2 \tilde{S}(\bar{x})=s_2-s_3$. It is not hard to see that, if we denote by $\delta(\tilde z)=-4p^3-27q^2$ the discriminant of $\Pol(s)$, and choose its square root so that
\[
\sqrt{\delta(\tilde z)} = (s_1-s_2)(s_1-s_3)(s_2-s_3),\]
then
\[
\sqrt{\delta(\tilde z)} (s_2-s_3)=-6p s_1^2 + 9q s_1 -4p^2.
\]
Hence
\begin{equation}\label{Rh-expr}
\hat R (x)=\frac{(x-\bar{x})}{x+1+\bar{x} }\sqrt{\frac{\Delta(x)}{\delta(\tilde z)}}
\left( 9q\tilde{S}(x) -6p\tilde{S}(x)^2-4p^2\right),
\end{equation}
for some $p, q \in \mathbb{Q}(\tilde z,v)$. Hence the proof of the proposition will be complete if we prove that ${\Delta(x)}/(w^2{\delta(\tilde z)})$ is a square in $\mathbb{Q}(t,v,x)$.
After several reductions, described in our {\sc Maple}\ session, we obtain
\begin{equation}\label{ratio:delta}
\sqrt{\frac{\Delta(x)}{\delta(\tilde z)}}=
\frac{w\Delta(x)^2 (v^4+8v^3+6v^2+2v+1)^2(2v^3-4v-1)^3(x-\bar{x})^3}
{y^2(ty^2-t-1)(v^2+4v+1)^2(v^2-1)(2v+1)(1-2t)^2P(\tilde z)}
\end{equation}
where we denote as before $y=x+1+\bar{x}$ and
{
\newcommand{\shrink}[1]{\hspace{-0.1em}#1\hspace{-0.1em}}
\newcommand{\shrink{+}}{\shrink{+}}
\newcommand{\shrink{-}}{\shrink{-}}
\begin{align*}
P(\tilde z) &=
- w^2 {\tilde z}^{2} \left( {v}
^{4}\myp8 {v}^{3}\myp6 {v}^{2}\myp2 v\myp1 \right) ^{2}\\
&\quad - ( v\mym1 )
\left( 8 {v}^{7}\myp16 {v}^{6}\myp40 {v}^{5}\myp72 {v}^{4}\myp85 {v}^{3}\myp53
{v}^{2}\myp13 v\myp1 \right) \left( {v}^{4}\myp8 {v}^{3}\myp6 {v}^
{2}\myp2 v\myp1 \right)\tilde z\\
&\quad + 2 v \left( 2 {v}^{11}\myp2 {v}^{10}\myp12 {v}^{9}\myp18
{v}^{8}\myp23 {v}^{7}\myp22 {v}^{6}\myp5 {v}^{5}\mym29 {v}^{4}\mym57 {v}^{3}\shrink{-}
40 {v}^{2}\mym11 v\mym1 \right) .
\end{align*}
}%
From this point on, we can combine~\eqref{Rh-expr} and~\eqref{ratio:delta} with the various parametrizations (by $U_0$, $U_1$, or $U_2$) introduced above to write closed-form expressions for $\hat{R}(x)$. We give one in Appendix~\ref{app:P0} in terms of $U_0$; see~\eqref{Rhat-P0}.
The degree of $R(x)$ is clearly $72$ at most. We determine it at $x=2$ by elimination of $U_0$, $w$, and $v$, and find it to be $72$; hence the bound is tight.
\end{proof}
\begin{figure}[!ht]
\hskip -25mm \scalebox{1}{\input{extensions.pdf_t}}
\caption{Structure of the various fields involved in the solution of king walks in $\mathcal{C}$. We have indicated the degrees, and where the main series lie.}
\label{fig:algSR}
\end{figure}
\begin{proof}
[End of the proof of Theorem~\ref{theo:king}.]
We have proved all statements of this theorem, except for the
degrees of the trivariate generating functions \ $M(x,y)$, $P(x,y)$, and $A(x,y)$. It
is clear from~\eqref{eqMMM-king},~\eqref{Psol-king}, and~\eqref{A-PM}
that they belong to $\KK(w, U_1(x), U_1(y))$,
where $\KK=\mathbb{Q}(t,x,y)$
and hence that they have degree at most $72\times 3=216$ over
$\KK$. We check (by specializing $x$, $y$, and even $t$ to real
values where all series converge, like $x=3$, $y=2$, and $t=1/100$) that there is no unexpected degree reduction.
\end{proof}
We get the final picture of the algebraic extensions shown in Figure~\ref{fig:algSR}.
\begin{comment}
\begin{figure}[!ht]
\centering
\scalebox{1.1}{
\begin{tabular}{ccccc}
&& $\KK(v,w,\tilde{S}(x))$ && \\
& $\stackrel{3}{\mathrel{\rotatebox[origin=c]{45}{$\hookrightarrow$}}}$ && $\stackrel{3}{\mathrel{\rotatebox[origin=c]{-45}{$\hookrightarrow$}}}$ &\\
$\KK = \mathbb{Q}(t,x,y) \stackrel{24}{\hookrightarrow}\KK(v,w)$ &&&& $\KK(v,w,\tilde{S}(x),\tilde{S}(y))$\\
& $\stackrel{3}{\mathrel{\rotatebox[origin=c]{-45}{$\hookrightarrow$}}}$ && $\stackrel{3}{\mathrel{\rotatebox[origin=c]{45}{$\hookrightarrow$}}}$ &\\
&& $\KK(v,w,\tilde{S}(y))$ &&
\end{tabular}
}
\caption{Define $\KK = \mathbb{Q}(t,x,y)$. We have $R(x) \in \KK(v,w,\tilde{S}(x))$ and $M(x,y) \in \KK(v,w,\tilde{S}(x),\tilde{S}(y))$. Thus, $M(x,y)$ and $A(x,y)$ are of degree $216$ over $\mathbb{Q}(t,x,y)$.}
\label{fig:algSR}
\end{figure}
\end{comment}
\subsection{Some interesting univariate series}
\label{sec:proofcoro}
In this subsection we examine various univariate series of interest, like those that are involved in the enumeration of all walks in $\mathcal{C}$, or of walks ending on the boundaries of $\mathcal{C}$. We also prove the results of Proposition~\ref{cor:coeffs} dealing with walks ending at a specific point, and the asymptotic results of Corollary~\ref{coro:asy}.
\begin{prop}\label{prop:univariate_series}
The series $R(1)=t^2M(1,0)$ is algebraic of degree $24$ over $\mathbb{Q}(t)$ and belongs to $\mathbb{Q}(t,w)$. More precisely,
\[
R(1)+\frac t 3 =-\frac{\numsmall} {3w(2v^3-4v-1)(v^4+8v^3+6v^2+2v+1)(2v^3+3v^2+6v+1)},
\]
where
\begin{align*}
\numsmall&=v \left( v+1 \right) \left( 2 {v}^{3}+4 {v}^{2}+5 v+1 \right)
\left( 4 {v}^{6}+3 {v}^{5}-8 {v}^{4}-6 {v}^{3}+12 {v}^{2}+11 v+
2 \right)\\
&\quad +\left( 96 {v}^{10}+272 {v}^{9}+446 {v}^{8}+384 {v}^{7}+3 {v}^{6
}-464 {v}^{5}-553 {v}^{4}-298 {v}^{3}-87 {v}^{2}-14 v-1 \right) t
.
\end{align*}
The series $S(1)=tM(0,1)$ is algebraic of degree $48$ over $\mathbb{Q}(t)$ and belongs to a quadratic extension of $\mathbb{Q}(t,w)$. More precisely,
\begin{equation}\label{S-T}
S(1)+\frac 1 2 =w\tilde w,
\end{equation}
where $\tilde w=1/2+ \mathcal{O}(t)$ has degree $2$ over $\mathbb{Q}(t,v)$, and satisfies~\eqref{T-def} (in Appendix~\ref{sec:appendixC}).
\\
The series $M(1,1)$ and $P(1,1)$ are algebraic of degree $48$ and belong to $\mathbb{Q}(t,w,\tilde w)$. \\
The series $A(1,1)$ and $A_{0,0}$ are algebraic of degree $24$ and belong to $\mathbb{Q}(t,w)$. More precisely,
\[
A(1,1)+\frac 1 {3t}= -\frac{w \times \numsmall'}{3t(1-2t)(4v^3+3v^2-1)^2(2v^3-4v-1)(2v^3+3v^2+6v+1)},\]
with
\begin{align*}
\numsmall'&=2(4v^3+3v^2+4v+1)(4v^3+3v^2-1)^2t\\
&\quad +(v+1)(16v^9+72v^8+94v^7+86v^6+3v^5+61v^4+68v^3+24v^2+7v+1) ,
\end{align*}
while
\[
A_{0,0}=P_{0,0} = \frac{2 R_1}{t^2},
\]
where $R_1$ is given in Proposition~\ref{prop:explicit-univariate}.
\end{prop}
\begin{proof}
We begin with the series $S(1)$: we set $x=1$ in the cubic equation~\eqref{eq:Pol1} satisfied by $S(x)$, and observe that the equation factors. The factor that vanishes is quadratic in $S(1)$. (The fact that $S(1)$ is quadratic can also be seen from~\eqref{eq:2Ms}.) Then we replace $R_0, R_1, B_1, B_2$ by their expressions from Proposition~\ref{prop:explicit-univariate}. We then reduce the degree of $t$ and $w$ in this equation by taking remainders (in $t$ and $w$) modulo~\eqref{algv} and~\eqref{w-def}. The coefficient of $w$ in this equation has a factor $(1+2S(1))$, which suggests to write~\eqref{S-T}. Now $\tilde w$ is quadratic over $\mathbb{Q}(t,v)$, but is found not to belong to $\mathbb{Q}(t,w)$. Its minimal equation over $\mathbb{Q}(t,v)$ can be written as~\eqref{T-def}.
Now in order to determine $R(1)$, we set $x=1$ in the square of~\eqref{eq:3Ms2}, and perform similar reductions as for $S(1)$. For $M(1,1)$, we use the defining equation of $M(x,y)$ (see~\eqref{eqMMM-king}), of course at $x=y=1$, and obtain
\[
(1-8t) M(1,1)=\frac 1 3-\frac {R_1+3R(1)}{2t} +(1-8t) \frac{S(1)}{2t},
\]
from which the properties stated in the proposition easily follow. We then combine the above expression of $M(1,1)$ with~\eqref{Psol-king} to obtain
\[
(1-8t) P(1,1)=\frac 2 3-\frac {R_1+3R(1)}{t} -(1-8t) \frac{S(1)}{t}.
\]
Since $A(x,y)$ is given by~\eqref{eq:Asep}, we then find
\[
(1-8t) A(1,1)=\frac 4 3-2\frac {R_1+3R(1)}{t}.
\]
We observe that the series $S(1)$ is not involved in this expression, and therefore $A(1,1)$ has degree $24$ only. Finally, we obtain from~\eqref{Psol-king} that $A_{0,0}=P_{0,0}= 2 M_{1,0}= 2 R_1/t^2$, which thus also has degree $24$.
\end{proof}
Let us now prove Proposition~\ref{cor:coeffs}, which deals with walks ending at a specific point.
\begin{proof}[Proof of Proposition~\ref{cor:coeffs}.]
According to~\eqref{sol-king} and~\eqref{Psol-king}, it suffices to prove that all series $M_{i,j}$ belong to $\mathbb{Q}(t,w)$.
Let us first prove this when $i=0$ or $j=0$, that is, for the coefficients of the series $S(x)=txM(0,x)$ and $R(x)=t^2M(x,0)$. For $S(x)$, we write $S(x)=xT(x)$, and observe that the cubic equation~\eqref{eq:Pol1} satisfied by $S(x)$, with coefficients in $\qs(t,x,R_0, R_1, B_1, B_2)$, reads
\[
3 t (R_0^2+R_0 t+t^2) (T(x)-R_0/t)= x\, \widetilde \Pol(t,x,T(x),R_0, R_1, B_1, B_2),
\]
for some polynomial $\widetilde \Pol$. This implies that $T_0=S_1=R_0/t$, as we already know from the definitions of $R(x)$ and $S(x)$, and then, by induction on $i$, that the series $S_i$ belong to $\mathbb{Q}(t,w)$ (because the series $R_0, R_1, B_1, B_2$ do). It then follows that the coefficients of $R(x)$ also belong to this field, using~\eqref{eq:Rhat},~\eqref{Rh-expr}, and~\eqref{ratio:delta}.
We finally return to the equation~\eqref{eqMMM-king} that defines $M(x,y)$. It reads
$K(x,y)M(x,y)= F(t,x,y)$, where $F(t,x,y)$ is a Laurent series in $x$ and $y$, having coefficients in $\mathbb{Q}(t,w)$ as we have just proved. We extract the coefficient of $x^i y^j$ in this equation, for $i,j\ge 0$, and thus obtain a linear expression $tM_{i+1, j+1}$ in terms of series $M_{k, \ell}$, where $k\le i+1$ and $\ell \le j+1$, one equality being strict, and series of $\mathbb{Q}(t,w)$. We then conclude by induction on $i+j$.
The fact that $C_{i,j}$ is transcendental (except for $i=-1$ or $j=-1$), follows from the fact that $Q_{i,j}$ is transcendental for $i,j\ge 0$, because its coefficients grow like $8^n n^{-3}$, which contradicts algebraicity.
\end{proof}
We finally prove the asymptotic results of Corollary~\ref{coro:asy}.
\begin{proof}[Proof of Corollary~\ref{coro:asy}.] We apply here the principles of the \emm singularity analysis, of algebraic series~\cite[Sec.~VII.7]{flajolet-sedgewick}. The series $u$ defined by~\eqref{u-def} is found to have radius of convergence $1/8$, and a unique singularity of minimal modulus, at $t=1/8$.
Moreover, as $t$ approaches $1/8$ from below, $u$ has the following Puiseux expansion:
\[
u= \frac 1 3-\frac 2 9\,{6}^{1/3}(1-8t)^{1/3}+\frac 1 {27}{6}^{2/3}(1-8t)^{2/3}
+\frac 1{27}(1-8t) + \mathcal{O}\left((1-8t)^{4/3}\right).
\]
Then the series $v$ defined in~\eqref{v-def}, seen as a series in $u$, has a radius of convergence larger than $u_c:=1/3$, and is thus analytic at $u_c$. At this point it attains the value $v_c\approx 0.455\ldots$,
which is the only real root of $4v^3+3v^2-1$. As $t$ approaches $1/8$ from below, one finds
\[
v= v_c-\frac 1 3\,v_c \left( 1+2\,v_c \right) {6}^{1/3}(1-8t)^{1/3}+{\frac { \left( 8\,{v_c}^{2}+11\,v_c+2 \right) }{54}}(1-8t) + \mathcal{O}\left((1-8t)^{4/3}\right).
\]
Finally, the series $w$, seen as a series in $v$, is analytic at $v_c$, where it is equal to $w_c:=\sqrt{3v_c^2+12v_c+3}/2$. As $t$ approaches $1/8$ from below, one finds
\[
w= w_c-\frac 2 9{6}^{2/3} \,v_c\,w_c\left( 1+2\,v_c
\right)(1-8t) ^{2/3}+ \mathcal{O}\left((1-8t)\right).
\]
More terms of the singular expansions of these three series are available in our {\sc Maple}\ session. We plug these expansions in the expressions of $A(1,1)$ and $A_{0,0}$ given in the previous proposition and obtain
\begin{equation}\label{A11sing}
A(1,1)= - \frac{2^5 6^{1/3} w_c\left(28 v_c^2+61 v_c-86\right)}{3^3 101 (1-8t)^{2/3}} + cst+ \mathcal{O}\left((1-8t)^{1/3}\right),
\end{equation}
\begin{multline*}
A_{0,0}=cst
-{\frac {2^9\,{6}^{2/3} { w_c}\, \left( 6716\,{{ v_c}}^{2}+2165\,{
v_c}-1582 \right) \left(1 -8\,t \right) ^{2/3}}{3^4 101^2}}
%
+ cst \left(1 -8\,t \right)
\\+\frac {2^8\,{6}^{1/3} { w_c}\, \left( 344660\,{{ v_c}}^{2}+688535\,{ v_c }-718546 \right) \left(1 -8\,t \right) ^{4/3}}{
3^5 101^3}
+\mathcal{O}\left( \left(1 -8\,t \right) ^{5/3}\right),
\end{multline*}
where each symbol $cst$ stands for a real constant that may vary from place to place, but has no implication on the asymptotic behaviour of the coefficients of our series. The series we are really interested in are
\[
C(1,1)= A(1,1)-\frac 1 3 Q(1,1)
\]
and
\[
C_{0,0}=A_{0,0} +\frac 1 3 Q_{0,0}.
\]
Recall from~\cite[Thm.~VI.1]{flajolet-sedgewick} that for $\alpha \not \in \{-1, -2, \ldots\}$, it holds that
\[
[t^n] (1-8t)^{-\alpha-1} = \frac {8^n n^\alpha} {\Gamma(\alpha+1)} + \mathcal{O} (8^nn^{\alpha-1} ).
\]
In particular, the $n$th coefficient in $A(1,1)$ grows like $8^n n^{-1/3}$, while the estimate corresponding to the remainder is in $8^n n^{-4/3}$. Moreover, it is proved in~\cite{Bostan-etal-2017-qp,MelczerMishna2016} that
\[
[ t^n] Q(1,1)= \frac 8{3\pi} \frac{8^n}{n} + \mathcal{O}\left(\frac{8^n}{n^2}\right),
\]
so that $Q(1,1)$ contributes to the second order term in the asymptotic behaviour of the number $c(n)$ of $n$-step walks in $\mathcal{C}$. We then compute the minimal polynomial over $\mathbb{Q}$ of the constant occurring in the first term of~\eqref{A11sing}, and put the two contributions together to obtain the first part of the corollary.
Now consider the series $C_{0,0}$. Since the coefficient of $t^n $ in $Q_{0,0}$ grows like $8^n/n^3$ (see~\cite{Bostan-etal-2017-qp,DenisovWachtel15}), the first two terms in the expansion of $c_{0,0}(n)$ come from the above expansion of $A_{0,0}$, and this yields the second part of the corollary.
\end{proof}
\section{Combinatorial proofs of some identities on square lattice walks}
\label{sec:combi}
\newcommand{\Walkse}[3]{#1^{#2}_{#3}}
\newcommand{\Qse}[2]{\Walkse{Q}{#1}{#2}}
\newcommand{\Cse}[2]{\Walkse{C}{#1}{#2}}
\newcommand{\Ase}[2]{\Walkse{A}{#1}{#2}}
\newcommand{\Hse}[2]{\Walkse{H}{#1}{#2}}
As already observed in~\cite[Sec.~7.1]{Bousquet2016} for the simple and diagonal models, the first two equations of Theorem~\ref{theo:king}, combined with the $x/y$-symmetry of our step set, imply that for $i,j \ge 0$,
\[
C_{i,j}=Q_{i,j}+C_{-i-2, j}+ C_{i,-j-2}.
\]
As suggested in\cite{Bousquet2016}, this can be proved using the reflection principle. This is what we do in this section. Further, we establish identities of this type for more general starting points and endpoints, and all Weyl models of Table~\ref{tab:weyl}. We begin in Section~\ref{sec:combi4} with the four models having a group of order $4$, and develop in Section~\ref{sec:combi-general} a general setting.
\subsection{A group of order \texorpdfstring{$4$}{4}: simple, diagonal, king, and diabolo walks}
\label{sec:combi4}
As shown in Table~\ref{tab:weyl} there are four step sets associated with
the Weyl group $A_1\times A_1$, of order $4$.
Mimicking the action of this group on $\rs^2$, we decompose the three-quarter plane $\mathcal{C}$ into three disjoint parts:
\begin{align*}
\mathcal{Q} &= \{ (i,j) : i \geq 0 \text{ and } j \geq 0 \} &&\text{(the first quadrant)}, \\
\mathcal{L} &= \{ (i,j) : i \leq -1 \text{ and } j \geq 0 \} &&\text{(the left quadrant)}, \\
\mathcal{B} &= \{ (i,j) : i \geq 0 \text{ and } j \leq -1 \} &&\text{(the bottom quadrant)}.
\end{align*}
As before, let $C_{i,j}$ (resp.\ $Q_{i,j}$) be the number of walks confined to $\mathcal{C}$ (resp.\ $\mathcal{Q}$) ending at $(i,j)$.
More generally, for any starting point $(a,b)$ we write $\Qse{a,b}{i,j}$ (resp.~$C^{a,b}_{i,j}$) for the length generating function\ of walks confined to $\mathcal{Q}$ (resp.~$\mathcal{C}$), starting from $(a,b)$ and ending at $(i,j)$.
A step set~$\mathcal S$ is called \emph{vertically symmetric} (or \emm v-symmetric,) if for all $(i,j) \in \mathcal S$ one has $(-i,j) \in \mathcal S$;
it is called \emph{horizontally symmetric} (or \emm h-symmetric,) if for all $(i,j) \in \mathcal S$ one has $(i,-j) \in \mathcal S$.
The four models that we consider in this subsection are the only v- and h-symmetric models among all small step models.
\begin{figure}[htb]
\centering
\includegraphics[height=4.2cm]{bijection1}
\qquad
\includegraphics[height=4.2cm]{bijection2}
\qquad
\includegraphics[height=4.2cm]{bijection3}
\caption{By the reflection principle, walks in the three-quarter plane $\mathcal{C}$ from $(0,0)$ to $(i,j)$ with $i,j \geq 0$ are in bijection with the union of three sets of walks: walks in $\mathcal{C}$ ending at $(-i-2,j)$, walks in $\mathcal{C}$ ending at $(i,-j-2)$, and walks staying completely in the first quadrant~$\mathcal{Q}$, ending at $(i,j)$. For more such identities see Proposition~\ref{prop:bijection}.}
\label{fig:bijection}
\end{figure}
\begin{prop}
\label{prop:bijection}
Let $\mathcal S$ be one of the four v- and h-symmetric small step models, and let $(a,b)$ be a starting point in $\mathcal{C}$.
For $(i,j) \in \mathcal{Q}$ we have
\begin{align*}
\Cse{a,b}{i,j} &= \Cse{a,b}{-i-2,j} + \Cse{a,b}{i,-j-2} +
\begin{cases}
\Qse{a,b}{i,j} & \text{ if } a,b \geq 0, \\
0 & \text{ if } a = -1 \text{ or } b=-1, \\
- \Qse{-a-2,b}{i,j} & \text{ if } a < -1, \\
- \Qse{a,-b-2}{i,j} & \text{ if } b < -1.
\end{cases}
\end{align*}
Furthermore, there exists an explicit bijection proving each of these identities.
\end{prop}
\begin{proof}
The proof idea is to suitably reflect the walks along the lines $x=-1$ and $y=-1$ which directly results in bijections for the claimed identities.
We fix an endpoint $(i,j) \in \mathcal{Q}$.
First, for a starting point $(a,b) \in \mathcal{Q}$ we partition the walks confined to $\mathcal{C}$ into three classes as shown in Figure~\ref{fig:bijection}: a walk either always stays in the first quadrant and is therefore counted by $\mathcal{Q}_{i,j} ^{a,b}$,
or it leaves the first quadrant.
In the latter case it either touches the line $x=-1$ or $y=-1$.
We cut the walk at the last point $(k,\ell)$ where this happens and reflect the second part of the walk, going from $(k,\ell)$ to $(i,j)$, along this line. As $\mathcal{S}$ is v- and h-symmetric we get a walk in $\mathcal{C}$ with steps in $\mathcal{S}$ ending either at $(-i-2,j)$ or $(i,-j-2)$.
The reverse bijection is analogous. One key point here is that a walk from $(a,b)$ to $(-i-2,j)$ (say) will necessarily touch the line $x=-1$,
and will touch it after any visit to the line $y=-1$.
Second, if the starting point $(a,b)$ is on the line $x=-1$ or $y=-1$ then the same argument applies, with $Q_{i,j}^{a,b}=0$ because no path can be entirely in the first quadrant $\mathcal{Q}$.
Third, if $a<-1$, the path starts left of the line $x=-1$, and thus cannot be contained in the first quadrant either. Moreover, a difficulty arises when defining the reverse construction: a walk starting from $(a,b)$ and ending at $(-i-2,j)$ may not touch the line $x=-1$, and thus cannot be reflected along this line (there is no such problem with walks ending at $(i,-j-2)$). But these walks are in essence walks in a quadrant: reflecting them along the line $x=-1$ gives walks from $(-a-2,b)$ to $(i,j)$ confined to the first quadrant $\mathcal{Q}$.
Fourth, for $b<-1$ the reasoning is analogous.
\end{proof}
The above proposition implies in particular the three formulas given in \cite[Sec.~7.1]{Bousquet2016}:
for $i,j \geq 0$ we have for any v- and h-symmetric step set and the three starting points $(0,0)$, $(-1,0)$, and $(-2,0)$:
\begin{align*}
\Cse{0,0}{i,j} &= \Cse{0,0}{-i-2,j} + \Cse{0,0}{i,-j-2} + Q_{i,j}^{{0,0}}, \\
\Cse{-1,0}{i,j} &= \Cse{-1,0}{-i-2,j} + \Cse{-1,0}{i,-j-2}, \\
\Cse{-2,0}{i,j} &= \Cse{-2,0}{-i-2,j} + \Cse{-2,0}{i,-j-2} -Q_{i,j}^{{0,0}}.
\end{align*}
Let us reformulate Proposition~\ref{prop:bijection} in terms of trivariate (rather than univariate) generating functions. For $(a,b) \in \mathcal{C}$, let $C^{a,b}(x,y)$ denote the generating function \ of walks in $\mathcal{C}$ that start from $(a,b)$:
\begin{equation}\label{Cab-def}
C^{a,b}(x,y)= \sum_{(i,j)\in \mathcal{C}} \Cse{a,b}{i,j} x^i y^j.
\end{equation}
We also define (uniquely) series $P^{a,b}(x,y)$, $L^{a,b}(x,y)$, and $B^{a,b}(x,y)$ in $\qs[x,y][[t]]$ by
\begin{equation}\label{Cab-split}
C^{a,b}(x,y)= P^{a,b}(x,y)+ \bar{x} L^{a,b}(\bar{x},y)+ \bar{y} B^{a,b}(x,\bar{y}).
\end{equation}
Then Proposition~\ref{prop:bijection} can be reformulated as follows.
%
\begin{prop} \label{prop:bijection-3v}
Let $\mathcal{S}$ be one of the four v- and h-symmetric small step models. For $(a,b)\in~\mathcal{C}$, the above defined series are related by
\begin{align*}
P^{a,b}(x,y) &= \bar{x} \left( L^{a,b}(x,y)-L^{a,b}(0,y)\right) +\bar{y} \left( B^{a,b}(x,y)-B^{a,b}(x,0)\right) \\
&+
\begin{cases}
Q^{a,b}(x,y) & \text{if } a,b \ge 0, \\
0 &\text{if } a=-1 \text{ or } b=-1, \\
- Q^{-a-2,b}(x,y) & \text{if } a<-1, \\
- Q^{a,-b-2}(x,y) & \text{if }b<-1.
\end{cases}
\end{align*}
\end{prop}
\begin{proof}
We multiply the identities of Proposition~\ref{prop:bijection} by $x^i y^j$ and sum over all $i,j\ge 0$.
\end{proof}
Now we will use these results to generalize Equation~\eqref{Psol-king} to the four models under consideration.
First, we define the generating function $A(x,y)$ as in
Proposition~\ref{prop:A-def-gen}, or equivalently by~\eqref{AC-def-king}.
It satisfies the following functional equation:
\begin{equation}\label{eq:diaboloAfunceq}
K(x,y)A(x,y)= \frac{2+\bar{x}^2+\bar{y}^2}3
-t\bar{y} H_-(x) A_{-,0}(\bar{x}) -t\bar{x} V_-(y)A_{0,-}(\bar{y}) -t\bar{x}\bar{y} A_{0,0} \mathbbm 1_{(-1,-1)\in \mathcal S}.
\end{equation}
Hence $A(x,y) = \sum_{(i,j) \in \mathcal{C}} A_{i,j} x^i y^j$ can be interpreted as the generating function\ of walks starting from $(0,0)$, $(-2,0)$, or $(0,-2)$ with weights $2/3$, $1/3$, and $1/3$, respectively.
In particular, if we now define the series $P(x,y), L(x,y), B(x,y) \in \qs[x,y][[t]]$ by
\begin{equation}\label{A-split-diab}
A(x,y)= P(x,y) + \bar{x} L(\bar{x},y) + \bar{y} B(x, \bar{y}),
\end{equation}
(observe that $B(x, \bar{y})= L(\bar{y},x)$ for an $x/y$-symmetric model), we have
\begin{equation}\label{linear}
P(x,y) = \frac 1 3 \left(2 P^{0,0}(x,y) + P^{-2,0}(x,y) + P^{0,-2}(x,y)\right),
\end{equation}
and analogously for the series $L$ and $B$.
Then Proposition~\ref{prop:bijection-3v} implies the following generalization of Equation~\eqref{Psol-king}.
\begin{coro}
\label{coro:PLBgroup4}
In the case of simple, diagonal, king, or diabolo walks, the power series $P(x,y)$, $L(x,y)$, and $B(x,y)$ defined in~\eqref{A-split-diab} obey the following identity
\[
P(x,y)=\bar{x} \big( L(x,y)-L(0,y)\big) +\bar{y} \big( B(x,y)-B(x,0) \big).
\]
\end{coro}
\begin{proof}
Applying Proposition~\ref{prop:bijection-3v} for $(a,b) = (0,0)$, $(-2,0)$, and $(0,-2)$ to~\eqref{linear} makes all contributions of the series $Q$ vanish and shows the claim.
\end{proof}
\begin{remark}
Let us define the series $A(x,y)$ as in~\eqref{eq:diaboloAfunceq}, but with weights $w_0$,
$w_x$, and $w_y$ for walks starting from $(0,0)$, $(-2,0)$, and
$(0,-2)$ respectively (rather than $2/3$, $1/3$, $1/3$). This only changes the initial term
in~\eqref{eq:diaboloAfunceq}, and Corollary~\ref{coro:PLBgroup4} still holds, provided that $w_0=w_x+w_y$.
\end{remark}
\medskip
In the next subsection, we give a higher level explanation of what happens here, more in the spirit of Gessel's and Zeilberger's proof of the reflection principle in~\cite{gessel-zeilberger}, and thus obtain statements that are valid for all Weyl models.
\subsection{A general result for Weyl models}
\label{sec:combi-general}
\newcommand\sdot[1]{\smash{\overset{\mathbin{\vcenter{\hbox{\raisebox{-2mm}{\hspace{0.2mm}\huge.}}}}}{#1}}}
We now consider one of the seven Weyl models $\mathcal S$ of Table~\ref{tab:weyl}, with a group $G$ of order $2d$, $d\in\{2, 3, 4\}$. Recall the definition of this group from Section~\ref{sec:group}, and the definition of the length $\ell(g)$ and sign $\varepsilon_g$ of $g \in G$. This group acts on steps, seen as elements of the vector space~$\zs^2$: for $g\in G$, the corresponding element $\vec g$ sends $(i,j)$ to $(k,\ell)$ if $g(x^i y^j)=x^k y^\ell$ (recall that we have defined $g(F(x,y)):=F(g(x,y))$ for any rational function $F(x,y)$). By construction of $G$, the set of steps $\mathcal S$ is invariant under this action of $G$. The group $G$ also acts on points of the plane, that is, on the \emm affine space, $\zs^2$, by $\sdot g (a,b)=(c,d)$ where $x^cy^d=\bar{x}\bar{y} \, g(x^{a+1}y^{b+1})$ (the shift by $xy$ is a bit unfortunate, and would be avoided by considering the positive quadrant $\{(a,b): a >0, b>0\}$ rather than the non-negative quadrant $\mathcal{Q}$).
For $g \in G$, we denote $\mathcal{Q}_g= \sdot g (\mathcal{Q})$. The $2d$ domains $\mathcal{Q}_g$, for $g \in G$, are disjoint; see Figure~\ref{fig:reflection}. For $(a,b)\in \mathcal{Q}$, the orbit of $(a,b)$ under the affine action of $G$ consists of $2d$ distinct points of the plane. In particular, the orbit sum $\OS(x^{a+1}y^{b+1})= \sum_g \varepsilon_g g(x^{a+1}y^{b+1})$ is non-zero. The union of the $2d$ domains $\mathcal{Q}_g$ does not cover the whole plane. For the points $(a,b)$ that are not in this union, the orbit of $(a,b)$ under the affine action of $G$ has cardinality less that $2d$, and in fact $\OS(x^{a+1}y^{b+1})= 0$. The complement of $\cup_g \mathcal{Q}_g$ is the union of $d$ lines (also called \emm walls, to match the terminology of~\cite{gessel-zeilberger}), defined, for each $g\in G$ such that $\varepsilon_g=-1$ (i.e., $\ell(g)$ is odd), by $W_g=\{(a,b)\in \zs^2 : \sdot g(a,b)=(a,b)\}$. The lines are dashed in our figures, and correspond to the reflection axes once the steps are straightened (as in Table~\ref{tab:weyl}). For instance, in all cases we have $W_\phi=\{(a,b):a=-1\}$ and $W_\psi=\{(a,b):b=-1\}$. Any two of the lines $W_g$ intersect at the point $(-1,-1)$. An important property is that a walk that is not entirely contained in a domain $\mathcal{Q}_g$ must touch one of these lines.
\begin{figure}[htb]
\centering
\includegraphics[width=0.35\textwidth]{Reflection_Simple}
\hskip -6mm \includegraphics[width=0.35\textwidth]{Reflection_Doubletandem}
\hskip -6mm \includegraphics[width=0.355\textwidth]{Reflection_Gouyou}
\caption{The $2d$ domains $\mathcal{Q}_g$ for $g \in G$, where $G$ has order $2d=4,6, 8$. They are separated by the $d$ walls $W_g$, for $g\in G$ such that $\varepsilon_g=-1$.}
\label{fig:reflection}
\end{figure}
We adopt the same notation $C^{a,b}_{i,j}$ and $C^{a,b}(x,y)$ as in the previous subsection. The generalization of Proposition~\ref{prop:bijection} reads as follows.
\begin{prop}\label{prop:bij3v-general}
Let $\mathcal{S}$ be one of the Weyl models of Table~\ref{tab:weyl}. Let $2d$ be the order of the associated group $G$. Let $\omega=\phi\psi\psi\cdots$ (with $d$ generators) be the only element of length $d$ in~$G$. For any starting point $(a,b) \in \mathcal{C}$ and any endpoint $(i,j)\in \mathcal{Q}$, we have:
\[
\sum_{g\in G\setminus\{\omega\}} \varepsilon_g\, C^{a,b}_{\sdot g(i,j)} =
\begin{cases}
0 & \text{if } (a,b) \not \in \bigcup_g \mathcal{Q}_g, \\
\varepsilon_h \, Q_{i,j} ^{\sdot h(a,b)} & \text{if } \sdot h(a,b) \in \mathcal{Q} \text{ for } h \in G.
\end{cases}
\]
\end{prop}
\begin{proof}
Recall that for $(i,j)\in \mathcal{Q}$, the $2d$ endpoints $\sdot g(i,j)$ are distinct. Hence the left-hand side of the above identity counts walks in $\mathcal{C}$, starting from $(a,b)$ and ending at one of the $2d-1$ points in the (affine) orbit of $(i,j)$ that are not in the negative quadrant, with a sign that depends on the domain $\mathcal{Q}_g$ where the walk ends. Observe that for each walk the parameters $(a,b)$, $(i,j)$, and $g$ are uniquely determined.
We will define a (partial) sign-reversing involution $\iota$ on these walks. The idea is sketched in Figure~\ref{fig:tandem-reflection-walks} for a group of order $6$, that is, for tandem or double tandem walks.
Let $w$ be such a walk. If it does not intersect any of the walls, then $\iota(w)$ is undefined. In this case, the starting point $(a,b)$ of $w$ must be in one (and exactly one) of the domains $\mathcal{Q}_g$, say in $\mathcal{Q}_{h^{-1}}$ (so that $\sdot h(a,b)\in \mathcal{Q}$). Then the endpoint of $w$ must be in $\mathcal{Q}_{h^{-1}}$ as well, and applying $\sdot h$ to the walk $w$ (seen as a sequence of vertices) sends $w$ to a walk joining $\sdot h(a,b)$ to $ (i,j)$ in $\mathcal{Q}$. Hence the signed number of walks that do not intersect any wall is given by the right-hand side of the identity.
Now assume that $w$ intersects one of the $d$ walls, and write $w=(w_0, \ldots, w_n)$ where the $w_i$'s are points of $\mathcal{C}$. Consider the largest $m$ such that $w_m$ is on one of the walls $W_h$. Note that the group element $h$ is uniquely defined, because the walls only intersect at $(-1,-1)$, which is not in~$\mathcal{C}$. Moreover, we have $m<n$ because the final point $\sdot g(i,j)$ is not on a wall. More generally, all points $w_{m+1}, \ldots, w_n=\sdot g(i,j)$ lie in $\mathcal{Q}_g$. Now, form the walk $\iota(w):=(w_0, \ldots, w_m=\sdot h (w_{m}), \sdot h (w_{m+1}),\ldots, \sdot h (w_{n}))$. Note that the points $\sdot h (w_{m+1}),\ldots, \sdot h (w_{n})$ lie in the domain $\mathcal{Q}_{hg}$. The new walk has still steps in $\mathcal S$, because $\mathcal{S}$ is invariant under the (vectorial) action of $G$.
Let us prove that it lies in the three-quadrant cone $\mathcal{C}$. This holds obviously for the first $m$ steps. If this were not true for the rest of the walk, then either the step $(w_m, \sdot h (w_{m+1}))$ would be one of the two forbidden steps joining $(-1,0)$ to $(0,-1)$ (but this is impossible because $\sdot h (w_{m+1})$ is not on a wall), or all points $\sdot h (w_{m+1}),\ldots, \sdot h (w_{n})$ would be in the domain $\mathcal{Q}_\omega$. But this is not possible either since $w_m=\sdot h (w_{m})$ would then have both coordinates negative.
Since $\iota(w)$ ends at $\sdot h\circ \sdot g(i,j)$, its sign is $-\varepsilon_g$ (because $h$ has odd sign). Its last visit to a wall is clearly $w_m \in W_h$, so $\iota \circ \iota(w)=w$ and we have indeed constructed a sign reversing involution of walks that visit at least one wall. This concludes the proof.
\end{proof}
\begin{figure}
\makebox[\textwidth][c]{
\includegraphics[width=0.27\textwidth]{Doubletandem_reflection_walk1}
\hskip -1.5mm
\includegraphics[width=0.27\textwidth]{Doubletandem_reflection_walk2}
\hskip -1.5mm
\includegraphics[width=0.27\textwidth]{Doubletandem_reflection_walk3}
\hskip -1.5mm
\includegraphics[width=0.27\textwidth]{Doubletandem_reflection_walk4}}
\caption{The involution of Proposition~~\ref{prop:bij3v-general} for walks starting at $(0,0)$ and a group of order $6$. The above eight walks capture all possible values of the pair $(g,h)$, where $\mathcal{Q}_g$ is the domain in which the walk ends and $W_h$ the last visited wall. The figure also shows the action of $h$ on the steps.}
\label{fig:tandem-reflection-walks}
\end{figure}
\pagebreak
Let us now reformulate the above proposition in terms of trivariate generating functions, as in Proposition~\ref{prop:bijection-3v}. Given $(a,b)\in \mathcal{C}$ and $g\in G$, the generating function\ of walks in $\mathcal{C}$ starting from $(a,b)$ and ending in $\mathcal{Q}_g$ reads
\begin{align*}
\sum_{(k, \ell) \in \mathcal{Q}_g} C^{a,b}_{k, \ell} x^k y^\ell
& =
\sum_{(i, j) \in \mathcal{Q}} C^{a,b}_{\sdot g (i,j)}\bar{x} \bar{y} \, g\!\left(x^{i+1} y^{j+1} \right)
\\
& = \bar{x} \bar{y} \, g(xy) P_g^{a,b}(g(x,y)),
\end{align*}
where $P_g^{a,b}(x,y) = \sum_{(i, j) \in \mathcal{Q}} C^{a,b}_{\sdot g (i,j)} x^i y^j$ is a series in $\qs[x,y][[t]]$. For instance, when $G$ has order $4$, it follows from~\eqref{Cab-split} that
\begin{align*}
P_{\rm id} ^{a,b}(x,y) &= P^{a,b}(x,y),\\
P_\phi^{a,b}(x,y)&=\bar{x}\left(L^{a,b}(x,y)- L^{a,b}(0,y)\right), \\
P^{a,b}_\Psi(x,y)&= \bar{y} \left( B^{a,b}(x,y)-B^{a,b}(x,0)\right).
\end{align*}
The generalization of Proposition~\ref{prop:bijection-3v} reads as follows.
\begin{prop} \label{prop:bijection-3v-gen} Let $\mathcal{S}$ be one of the Weyl models of Table~\ref{tab:weyl}. Let $2d$ be the order of the associated group $G$. Let $\omega=\phi\psi\psi\cdots$ (with $d$ generators) be the only element of length $d$ in~$G$. For any starting point $(a,b) \in \mathcal{C}$, we have:
\[
\sum_{g \in G \setminus\{\omega\}} \varepsilon_g P_g^{a,b}(x,y)=
\begin{cases}
0 & \text{if } (a,b) \not \in \bigcup_g \mathcal{Q}_g, \\
\varepsilon _h Q^{\sdot h(a,b)}(x,y) & \text{if } \sdot h(a,b) \in \mathcal{Q} \text{ for } h \in G.
\end{cases}
\]
\end{prop}
\begin{proof}
Multiply the identity of Proposition~\ref{prop:bij3v-general} by $x^i y^j$ and sum over $i, j \ge 0$.
\end{proof}
Our final result deals with the series $A(x,y)$ defined in Proposition~\ref{prop:A-def-gen}. Observe that the orbit sum of $xy$ can be written in terms of the affine orbit of $(0,0)$:
\[
\bar{x} \bar{y} \OS(xy) =\sum_{\substack{h \in G \\[1mm] (c,d)=\sdot h (0,0)}}
\varepsilon_h \, x^c y^d.
\]
Hence, from the functional equation~\eqref{A-func-eq} satisfied by $A(x,y)$, we conclude that this series counts weighted walks in $\mathcal{C}$, and more precisely, that
\[
A(x,y)= \frac{2d-2}{2d-1}\, C^{0,0}(x,y) - \frac 1 {2d-1} \sum_{h \in G\setminus\{{\rm id} , \omega\}} \varepsilon_h \, C^{\sdot h(0,0)}(x,y).
\]
For $g\in G$, let us denote by $\bar{x} \bar{y} \, g(xy) P_g(g(x,y))$ the contribution in $A(x,y)$ of (weighted) walks ending in $\mathcal{Q}_g$. As before, this notation is designed so that $P_g(x,y) \in \qs[x,y][[t]]$. Then
\begin{equation}\label{P-Pab}
P_g(x,y)= \frac{2d-2}{2d-1}\, P^{0,0}_g(x,y) - \frac 1 {2d-1} \sum_{h \in G\setminus\{{\rm id} , \omega\}} \varepsilon_h \, P^{\sdot h(0,0)}_g(x,y).
\end{equation}
We can now state the generalization of Corollary~\ref{coro:PLBgroup4}.
\begin{coro}
The above defined series $P_g(x,y)$ are related by
\[
\sum_{g \in G \setminus\{\omega\}} \varepsilon_g P_g(x,y)=0.
\]
\end{coro}
\begin{proof}
We first use the identity~\eqref{P-Pab}, and then Proposition~\ref{prop:bijection-3v-gen}. This gives:
\begin{align*}
\sum_{g \in G \setminus\{\omega\}} \varepsilon_g P_g(x,y)
&= \frac{2d-2}{2d-1}\, \sum_{g \in G \setminus\{\omega\}} \varepsilon_g P^{0,0}_g(x,y) - \frac 1 {2d-1} \sum_{h \in G\setminus\{{\rm id} , \omega\}} \varepsilon_h
\sum_{g \in G \setminus\{\omega\}} \varepsilon_g P^{\sdot h(0,0)}_g(x,y)
\\
&= \frac{2d-2}{2d-1}\, Q^{0,0}(x,y) - \frac 1 {2d-1} \sum_{h \in G\setminus\{{\rm id} , \omega\}} \varepsilon_h \varepsilon_{h^{-1}} Q^{0,0}(x,y)
\\
& =0,
\end{align*}
since $\varepsilon_h \varepsilon_{h^{-1}} =1$ and $G$ has order $2d$.
\end{proof}
\begin{remark}
It is easy to see that the results of this section hold as well if we allow steps between the points $(0,-1)$ and $(-1,0)$, as in~\cite{Budd2020Winding,ElveyPrice2020Winding}.
\end{remark}
\section{Final comments}
\label{sec:final}
The first question raised by this paper is whether all seven models of Table~\ref{tab:weyl} actually obey the pattern described in Conjecture~\ref{conj:Weyl}. Does $C(x,y)$ differ from the linear combination of series $Q(\cdot, \cdot)$ given in Proposition~\ref{prop:A-def-gen} by an algebraic series? This is now proved for three of these seven models.
In terms of techniques, one can of course try to extend the approach of this paper to the other four Weyl models. Another idea would be to try to use the technique, based on \emm invariants,, that has been used recently~\cite{mbm-tq-kreweras} to solve the first three (algebraic) models of Table~\ref{tab:zero}. In fact, it is shown in~\cite{mbm-tq-kreweras} that this approach also works for the simple and diagonal models. Can it be adapted to the four unsolved Weyl cases?
to Gessel's model (number four in Table~\ref{tab:zero})?
Next to these $7+4=11$ models, there remain $12$ models with finite and non-monomial group, as shown in Table~\ref{tab:non-monomial}.
The non-monomial group action when applied to power series, prevents the efficient extraction of (positive/negative) parts.
For this reason the methods of this paper become even more complicated, and probably new approaches have to be developed.
Another question is whether the generating function\ for walks in other cones -- possibly larger than $2\pi$, as in~\cite{Budd2020Winding,ElveyPrice2020Winding} -- may satisfy a similar algebraicity phenomenon; that is, decompose into a simple D-finite series with the same orbit sum and an algebraic one.
We conclude with a sketch of the solution of the king model in which we
allow moves from $(0,-1)$ to $(-1,0)$ and back, as in~\cite{Budd2020Winding,ElveyPrice2020Winding}.
\subsection*{Allowing steps between \texorpdfstring{$(-1,0)$}{(-1,0)} and \texorpdfstring{$(0,-1)$}{(0,-1)} in king walks}
As already mentioned in this paper,
in two recent references dealing with the winding number of plane lattice walks~\cite{Budd2020Winding,ElveyPrice2020Winding}, it seems more natural to count walks in which all vertices lie in $\mathcal{C}$, but not necessarily all edges: that is, one allows steps form $(-1,0)$ to $(0,-1)$, and vice versa. It is natural to ask whether this choice leads to simpler series. This is why we have re-run our {\sc Maple}\ sessions on this variant of the king model. The first steps of the derivation, until the determination of the series $R_0$, $R_1$, $B_1$, and $B_2$ (as in Proposition~\ref{prop:explicit-univariate}) appear to be a bit simpler, but this stops being the case as soon as we return to the series $R(x)$ and $S(x)$. Let us give a few details.
First, the only changes in our basic functional equations are a term $t(\bar{x}+\bar{y}) C_{-1,0}$ in the right-hand side of~\eqref{eq:kernelC1}, and a term $t\bar{x} M_{0,0}$ in the right-hand side of~\eqref{eqM-king}. The new series $\hat{S}(x)$ is obtained from~\eqref{eq:STcardanosubs} by deleting the term in $R_0$. We still denote $\tilde{S}(x)=(x+\bar{x}+1)\hat{S}(x)/(x-\bar{x})$, and then the equation in one catalytic variable that we have to solve reads, with the same notation as in~\eqref{eqSc}:
\begin{align*}
0 &= 27 \left( 2 t+z+1 \right) \left( 10 t-3 z+1 \right) {\tilde{S}(x)}^{3}+
\big( \left( 54-243 R_1+54 t \right) R_0 -243 R_1^{2}+243 R_1 t \\
& \quad -81 zR_1 -45 {t}^{2}+18 zt-27 B_1+27 B_2+81 R_1-54 t+18 z-9 \big) \tilde{S}(x)\\
&\quad -81 {R_0}^{2}+ \left( 81 t+27-27 z-162 R_1 \right) R_0-81 R_1^{2}+81 R_1 t-27 zR_1\\
&\quad +5 {t}^{2}+10 zt-3 {z}^{2}+9 B_1+9 B_2+27 R_1-6 t+4 z-2=0.
\end{align*}
The system defining $R_0$, $R_1$, $B_1$, and $B_2$ is also a bit more compact, and in fact we can derive polynomial equations for each individual series without having to guess them first. They are now all of degree $24$ (while $B_1$ had degree $12$ in the first setting), and are found to belong to $\mathbb{Q}(t,w)$. For instance, we now have
\[
R_0= \frac{v(1-2t) }{v^4+8v^3+6v^2+2v+1}\left( 1+2v+ \frac {2v^3-4v-1}{2w}\right).
\]
Then we get back to $S(x)$, and that is where things become after all a bit more complicated than in the first setting. For instance, $\tilde{S}(x)$ has now degree $72$ rather than $36$. It can be written as $\tilde{S}_0(x)+w \tilde{S}_1(x)$, where both series $\tilde{S}_i(x)$, now of degree $36$, belong to $\mathbb{Q}(v,U_0(x))$ and hence to $\mathbb{Q}(v,U_1(x))$, where $U_0$ and $U_1$ are series defined in Appendix~\ref{app:P0} and Section~\ref{sec:bivariate}, respectively. Finally, both series $R(x)$ and $S(x)$ are found to belong to $\mathbb{Q}(t,w, x, U_1(x))$ and have degree $72$ over $\qs(t,x)$.
\section*{Acknowledgements}
We are extremely thankful to Mark van Hoeij, who helped us a lot in finding a simple description of our algebraic series of high degree. Our warm thanks also go to Bruno Salvy for his help with several {\sc Maple}\ problems that we met.
|
\section{Introduction}
In this paper, $G$ is a simple graph without isolated vertices, with vertex
set $V=V(G)$ and edge set $E=E(G)$. The \emph{order} $|V|$ of $G$ is denoted
by $n=n(G)$. For a vertex $v\in V$, the \emph{open neighborhood} of $v$ is
the set $N(v)=\{u\in V\mid uv\in E\},$ the \emph{closed neighborhood} of $v$
is the set $N[v]=N(v)\cup \{v\}$, and the \emph{degree} of $v$ is $\deg
_{G}(v)=|N(v)|$.
Let $u$ and $v$ be two vertices in $G.$ A $uv$\textit{-path} is a path with
endvertices $u$ and $v$, and the \textit{distance} between $u$ and $v$ is
the length of a shortest $uv$-path. The \emph{diameter} of $G$, denoted by
\mathrm{diam}(G)$, is the maximum distance between vertices of $G$. We write
$P_{n}$ and $C_{n}$ for the \emph{path} and \emph{cycle} of order $n$,
respectively. {Let }$A$ and $B$ {are} two disjoint subgraphs
(not necessarily induced) }of a graph $G.$ If there is an edge
e $ having one endvertex in $A$ and the other one in $B,$ then $A+B+e$ will
denote the graph formed by $A$ and $B$ for which we add only the edge $e.$
We also denote by $G-A$ the subgraph of $G$ induced by $V(G)-V(A).
For a set $D,$ let $B(D)$ be the set of vertices in $V\setminus D$ that have
a neighbor in $D$. The \textit{differential} of a set $D$ is defined in~\cit
{M} as $\partial (D)=|B(D)|-|D|$, and the maximum value of $\partial (D)$
for any subset $D$ of $V$ is the \textit{differential} of $G$, denoted
\partial (G)$. Differential of graphs has been studied extensively in
several papers, in particular \cite{B0,B3,B4,B5,B6,B1,B2,kkcs}. In 2017,
Bermudo \cite{B0} proved that for any graph $G$ with order $n\geq 15$,
minimum degree two and without any induced tailed $5$-cycle graph of seven
vertices or tailed $5$-cycle graph of seven vertices together with a
particular edge, it is satisfied $\partial (G)\geq \frac{5n}{17}$. Moreover,
he posed the following conjecture.
\begin{conj}[\protect\cite{B0}]
\label{conj}Let $G$ be a graph of order $n\ge 6k + 9$, minimum degree
\delta\ge 2$, which does not contain any induced $\{C_5,C_8,\ldots,C_{3k+2
\} $-cycles. Then $\partial(G)\ge \frac{(2k+3)n}{6k+11}.$
\end{conj}
A \emph{Roman dominating function} (RDF) on a graph $G$ is a function
f:V(G)\rightarrow \{0,1,2\}$ such that every vertex $u\in V(G)$ with $f(u)=0$
has a neighbor $v$ with $f(v)=2$. The \textit{weight} of an RDF $f$ is the
value $f(V(G))=\sum_{u\in V(G)}f(u),$ and the \emph{Roman domination number}
$\gamma _{R}(G)$ is the minimum weight of an RDF on $G$. The Roman
domination number of graphs was introduced in 2004 by Cockayne et al.\ in
\cite{CDHH} and is now well-studied in graph theory. The literature on Roman
domination and its variations has been surveyed and detailed in two book
chapters and three surveys \cite{Ch1, Ch2, Ch3, Ch4, Ch5}.
{In \cite{cham}, it has been shown that, if $G$ is a graph of
order $n\geq 9$ and minimum degree $\delta \geq 2$, then $\gamma _{R}(G)\leq
\frac{8n}{11}$. It was also shown in \cite{B0} that $\gamma _{R}(G)\leq
\frac{12n}{17}$ for any graph }$G$ {with order $n\geq 15$, minimum degree
two and without any induced tailed 5-cycle graph of seven vertices or tailed
5-cycle graph of seven vertices together with a particular edge. }
In this paper, we improve the aforementioned known results by showing that
if $G$ is a graph satisfying the statement of Conjecture \ref{conj}, then
\gamma _{R}(G)\leq \frac{(4k+8)n}{6k+11}.$ U{sing the Gallai-type result
involving the differential and the Roman domination number of graphs
established by Bermudo, Fernau and Sigarreta \cite{B6} who proved that
\gamma _{R}(G)+\partial (G)=n$, our bound leads to }$\partial (G)\geq \frac
(2k+3)n}{6k+11}$ which settles Conjecture {\ref{conj}. }
We close this section by recalling the exact values of the Roman domination
number of paths and cycles given in \cite{CDHH}, namely $\gamma
_{R}(P_{n})=\gamma _{R}(C_{n})=\lceil \frac{2n}{3}\rceil .$
\section{Some useful lemmas}
We gather in this section some results that will be useful to us thereafter.
For technical reasons, we will often consider three Roman dominating
functions $f_{1},f_{2}$, and $f_{3}$ on a graph $G$, where we use
\overrightarrow{f}$ to denote the $3$-tuple $(f_{1},f_{2},f_{3})$, and
\overrightarrow{f}(v)$ for $(f_{1}(v),f_{2}(v),f_{3}(v))$ for a vertex $v$.
A vertex $v$ is said to be $\overrightarrow{f}$\textit{-strong} if
f_{j}(v)=2$ for some $j\in \{1,2,3\}$. Moreover, the weight of
\overrightarrow{f}$ is $\omega (\overrightarrow{f})=\sum_{j=1}^{3}\omega
(f_{j})$. Clearly, $\omega (f_{j})\leq \omega (\overrightarrow{f})/3$ for
some $j\in \{1,2,3\}$. \ Also, if $H$ is an induced subgraph of $G$ and $f$
an RDF on $G$, then we denote the restriction of $f$ on $H$ by $f|_{V(H)}$
and let $f(V(H))=\omega (f,H).$
For integers $m$ and {$\ell$} such that $m\geq 3$ and $\ell \geq 1$, let
C_{m,\ell }$ be the graph obtained from a cycle {$C_{m}=x_{1}x_{2}\ldots
x_{m}x_{1}$} and a path $P=y_{1}y_{2}\ldots y_{\ell }$ by adding the edge
x_{1}y_{1},$ with {$y_{i}\notin V(C_{m})$} for all possible $i$. The graph
C_{m,\ell }$ will be called a \textit{tailed }$m$\textit{-cycle graph} of
order $m+\ell $. {We call an \textit{ear }of a cycle }$C$ in a graph $G,$ {to a
path }$P$ {in }$G-C$ {whose endvertices are adjacent to some vertices in }
C. $
\begin{lemma}
\label{ear1}Let $G$ be a graph, $u,v\in V(G)$ and $\overrightarrow{f
=(f_{1},f_{2},f_{3})$ be a 3-tuple of RDFs of $G$ such that $u$ and $v$ are
\overrightarrow{f}$-strong. If $H$ is a graph obtained from $G$ by adding a
path $Q=y_{1}\ldots y_{\ell }$ and the edges $uy_{1},vy_{\ell }$, then
\overrightarrow{f}$ can be extended to a 3-tuple of RDFs $\overrightarrow{g}$
of $H$ such that {$\omega (\overrightarrow{g},Q)\leq 2\ell$} and each vertex in
V(Q)-\{y_{1},y_{\ell }\}$ is $\overrightarrow{g}$-strong.
\end{lemma}
\noindent \textbf{Proof.} By assumption, $f_{i}(u)=2$ and $f_{j}(v)=2$ for
some $i,j\in \{1,2,3\}$. Let us consider the following two cases.
\smallskip \noindent \textbf{Case 1.} $i\neq j$.\newline
Assume, without loss of generality, that $i=1$ and $j=2$. Consider the
following situations.
\smallskip \textbf{Subcase 1.1.} $\ell \equiv 0\pmod 3$.\newline
Define the functions $g_{1},g_{2}$ and $g_{3}$ on $V(H)$ as follows:
g_{1}(z)=f_{1}(z)$ for all $z\in V(G)$, $g_{1}(y_{3i+3})=2$ for $0\leq i\leq
\frac{\ell }{3}-1$, and $g_{1}(z)=0$ otherwise; $g_{2}(z)=f_{2}(z)$ for all
z\in V(G)$, $g_{2}(y_{3i+1})=2$ for $0\leq i\leq \frac{\ell }{3}-1$, and
g_{2}(z)=0$ otherwise; and $g_{3}(z)=f_{3}(z)$ for all $z\in V(G)$,
g_{3}(y_{3i+2})=2$ for $0\leq i\leq \frac{\ell }{3}-1$, and $g_{3}(z)=0$
otherwise.
\smallskip \textbf{Subcase 1.2.} $\ell \equiv 1\pmod 3$.\newline
Define the functions $g_{1},g_{2}$ and $g_{3}$ on $V(H)$ as follows:
g_{1}(z)=f_{1}(z)$ for all $z\in V(G)$, $g_{1}(y_{3i+3})=2$ for $0\leq i\leq
\frac{\ell -4}{3}$, and $g_{1}(z)=0$ otherwise; $g_{2}(z)=f_{2}(z)$ for all
z\in V(G)$, $g_{2}(y_{3i+2})=2$ for $0\leq i\leq \frac{\ell -4}{3}$, and
g_{2}(z)=0$ otherwise; and $g_{3}(z)=f_{3}(z)$ for $z\in V(G)$,
g_{3}(y_{3i+1})=2$ for $0\leq i\leq \frac{\ell -1}{3}$, and $g_{3}(z)=0$
otherwise.
\smallskip \textbf{Subcase 1.3.} $\ell \equiv 2\pmod 3$.\newline
Define the functions $g_{1},g_{2}$ and $g_{3}$ on $V(H)$ as follows:
g_{1}(z)=f_{1}(z)$ for all $z\in V(G)$, $g_{1}(y_{3i+3})=2$ for $0\leq i\leq
\frac{\ell -5}{3}$, $g_{1}(y_{\ell })=1$ and $g_{1}(z)=0$ otherwise;
g_{2}(z)=f_{2}(z)$ for all $z\in V(G)$, $g_{2}(y_{3i+2})=2$ for $0\leq i\leq
\frac{\ell -5}{3}$, $g_{2}(y_{\ell -1})=1$ and $g_{2}(z)=0$ otherwise; and
g_{3}(z)=f_{3}(z)$ for all $z\in V(G)$, $g_{3}(y_{3i+1})=2$ for $0\leq i\leq
\frac{\ell -2}{3}$, and $g_{3}(z)=0$ otherwise.
\bigskip
In either subcase, $g_{1},g_{2},g_{3}$ are RDFs of $H$ and thus
g=(g_{1},g_{2},g_{3})$ is a 3-tuple of RDFs of $H.$ In addition, $\omega
(\overrightarrow{g},Q)\leq 2\ell$ and each vertex of $V(Q)-\{y_{1},y_{\ell }\}$ is
\overrightarrow{g}$-strong.
\smallskip \noindent \textbf{Case 2.} $i=j$.\newline
Assume, without loss of generality, that $i=j=1$. Consider again the
following situations.
\smallskip \textbf{Subcase 2.1.} $\ell \equiv 0\pmod 3$.\newline
Define the functions $g_{1},g_{2}$ and $g_{3}$ on $V(H)$ as follows:
g_{1}(z)=f_{1}(z)$ for all $z\in V(G)$, $g_{1}(y_{2})=1$, $g_{1}(y_{3i+4})=2$
for $0\leq i\leq \frac{\ell -6}{3}$, and $g_{1}(z)=0$ otherwise;
g_{2}(z)=f_{2}(z)$ for all $z\in V(G)$, $g_{2}(y_{3i+2})=2$ for $0\leq i\leq
\frac{\ell }{3}-1$, and $g_{2}(z)=0$ otherwise; $g_{3}(z)=f_{3}(z)$ for all
z\in V(G)$, $g_{3}(y_{1})=1$, $g_{3}(y_{3i+3})=2$ for $0\leq i\leq \frac
\ell }{3}-1$, and $g_{3}(z)=0$ otherwise.
\smallskip \textbf{Subcase 2.2.} $\ell \equiv 1\pmod 3$.\newline
Define the functions $g_{1},g_{2}$ and $g_{3}$ on $V(H)$ as follows:
g_{1}(z)=f_{1}(z)$ for $z\in V(G)$, $g_{1}(y_{2})=g_{1}(y_{\ell -1})=1$,
g_{1}(y_{3i+4})=2$ for $0\leq i\leq \frac{\ell -7}{3}$, and $g_{1}(z)=0$
otherwise; $g_{2}(z)=f_{2}(z)$ for all $z\in V(G)$, $g_{2}(y_{1})=1$,
g_{2}(y_{3i+3})=2$ for $0\leq i\leq \frac{\ell -4}{3}$, and $g_{2}(z)=0$
otherwise; $g_{3}(z)=f_{3}(z)$ for all $z\in V(G)$, $g_{3}(y_{\ell })=1$,
g_{3}(y_{3i+2})=2$ for $0\leq i\leq \frac{\ell -4}{3}$, and $g_{3}(z)=0$
otherwise.
\smallskip \textbf{Subcase 2.3.} $\ell \equiv 2\pmod 3$.\newline
Define the functions $g_{1},g_{2}$ and $g_{3}$ on $V(H)$ as follows:
g_{1}(z)=f_{1}(z)$ for all $z\in V(G)$, $g_{1}(y_{3i+3})=2$ for $0\leq i\leq
\frac{\ell -5}{3}$ and $g_{1}(z)=0$ otherwise; $g_{2}(z)=f_{2}(z)$ for all
z\in V(G)$, $g_{2}(y_{3i+1})=2$ for $0\leq i\leq \frac{\ell -2}{3}$, and
g_{2}(z)=0$ otherwise; $g_{3}(z)=f_{3}(z)$ for all $z\in V(G)$,
g_{3}(y_{3i+2})=2$ for $0\leq i\leq \frac{\ell -2}{3}$, and $g_{3}(z)=0$
otherwise.
\bigskip
In either subcase, $g_{1},g_{2},g_{3}$ are RDFs of $H$ and thus
g=(g_{1},g_{2},g_{3})$ is a 3-tuple of RDFs of $H.$ Moreover, $\omega
(\overrightarrow{g},Q)\leq 2\ell$ and each vertex of $V(Q)-\{y_{1},y_{\ell }\}$ is
\overrightarrow{g}$-strong.$\hfill \Box $
\begin{lemma}
\label{tailedcycle-3p+1}Let $G$ be a graph, $u\in V(G)$ and $\overrightarrow
f}=(f_{1},f_{2},f_{3})$ a 3-tuple of RDFs of $G$ such that $u$ is
\overrightarrow{f}$-strong.
\begin{itemize}
\item[1.] If $H$ is obtained from $G$ by adding a cycle $C_{3p+1}=x_{1}x_{2
\ldots x_{3p+1}x_{1}\;(p\geq 1)$ and the edge $ux_{1}$, then
\overrightarrow{f}$ can be extended to a 3-tuple $\overrightarrow{g}$ of
RDFs of $H$ such that {$\omega (\overrightarrow{g},C_{3p+1})\leq 2(3p+1)$} and each vertex
in $V(C_{3p+1})-\{x_{3p+1}\}$ is $\overrightarrow{g}$-strong.
\item[2.] If $H$ is obtained from $G$ by adding a tailed cycle $C_{3p+1,\ell
}\;(p\geq 1)$ and the edge $uy_{\ell }$, then $\overrightarrow{f}$ can be
extended to a 3-tuple $\overrightarrow{g}$ of RDFs of $H$ such that $\omega
(\overrightarrow{g},C_{3p+1,\ell })\leq 2(3p+1+\ell )$ and each vertex of
V(C_{3p+1,\ell })-\{x_{3p+1}\}$ is $\overrightarrow{g}$-strong.
\item[3.] If $H$ is obtained from $G$ by adding a cycle $C_{3p+2}=x_{1}x_{2
\ldots x_{3p+2}x_{1}\;(p\geq 1)$ and the edge $ux_{1}$, then
\overrightarrow{f}$ can be extended to a 3-tuple $\overrightarrow{g}$ of
RDFs of $H$ such that $\omega (\overrightarrow{g},C_{3p+2})\leq 2(3p+2)+1$ and each
vertex of $C_{3p+2}$, is $\overrightarrow{g}$-strong.
\item[4.] If $H$ is obtained from $G$ by adding a tailed cycle $C_{3p+2,\ell
}\;(p\geq 1)$ and the edge $uy_{\ell }$, then $\overrightarrow{f}$ can be
extended to a 3-tuple $\overrightarrow{g}$ of RDFs of $H$ such that $\omega
(\overrightarrow{g},C_{3p+2,\ell })\leq 2(3p+2+\ell )+1$ and each vertex of
C_{3p+2,\ell }$, is $\overrightarrow{g}$-strong.
\end{itemize}
\end{lemma}
\noindent \textbf{Proof.} Since $u$ is $\overrightarrow{f}$-strong, let us
assume, without loss of generality, that $f_{1}(u)=2$.
1) Define the functions $g_{1},g_{2}$ and $g_{3}$ on $V(H)$ as follows:
g_{1}(z)=f_{1}(z)$ for all $z\in V(G)$, $g_{1}(x_{3i+3})=2$ for $0\leq i\leq
p-1$, and $g_{1}(z)=0$ otherwise; $g_{2}(z)=f_{2}(z)$ for all $z\in V(G)$,
g_{2}(x_{3p})=1$, $g_{2}(x_{3i+1})=2$ for $0\leq i\leq p-1$, and $g_{2}(z)=0$
otherwise; $g_{3}(z)=f_{3}(z)$ for all $z\in V(G)$, $g_{3}(x_{3p+1})=1$,
g_{3}(x_{3i+2})=2$ for $0\leq i\leq p-1$, and $g_{3}(z)=0$ otherwise.
Clearly, $g_{1},g_{2},g_{3}$ are RDFs of $H$ and thus $g=(g_{1},g_{2},g_{3})$
is a 3-tuple of RDFs of $H.$ In addition, $\omega (\overrightarrow{g},C_{3p+1})\leq
2(3p+1)$ and each vertex of $V(C_{3p+1})-\{x_{3p+1}\}$ is
\overrightarrow{g}$-strong.
2) Consider the following cases.
\smallskip \noindent \textbf{Case 1.} $\ell \equiv 0\pmod 3$.\newline
Define the functions $g_{1},g_{2}$ and $g_{3}$ on $V(H)$ as follows:
g_{1}(z)=f_{1}(z)$ for all $z\in V(G)$, $g_{1}(x_{3i+3})=2$ for $0\leq i\leq
p-1$, $g_{1}(y_{3i+1})=2$ for $0\leq i\leq \frac{\ell }{3}-1$, and
g_{1}(z)=0$ otherwise; $g_{2}(z)=f_{2}(z)$ for all $z\in V(G)$,
g_{2}(x_{3p})=1$, $g_{2}(x_{3i+1})=2$ for $0\leq i\leq p-1$,
g_{2}(y_{3i+3})=2$ for $0\leq i\leq \frac{\ell }{3}-1$, and $g_{2}(z)=0$
otherwise; $g_{3}(z)=f_{3}(z)$ for all $z\in V(G)$, $g_{3}(x_{3p+1})=1$,
g_{3}(x_{3i+2})=2$ for $0\leq i\leq p-1$, $g_{3}(y_{3i+2})=2$ for $0\leq
i\leq \frac{\ell }{3}-1$, and $g_{3}(z)=0$ otherwise.
\smallskip \noindent \textbf{Case 2.} $\ell \equiv 1\pmod 3$.\newline
Define the functions $g_{1},g_{2}$ and $g_{3}$ on $V(H)$ as follows:
g_{1}(z)=f_{1}(z)$ for all $z\in V(G)$, $g_{1}(x_{3p+1})=1$,
g_{1}(x_{3i+2})=2$ for $0\leq i\leq p-1$, $g_{1}(y_{3i+2})=2$ for $0\leq
i\leq \frac{\ell -4}{3}$, and $g_{1}(z)=0$ otherwise; $g_{2}(z)=f_{2}(z)$
for all $z\in V(G)$, $g_{1}(x_{3p})=1$, $g_{2}(x_{3i+1})=2$ for $0\leq i\leq
p-1$, $g_{2}(y_{3i+3})=2$ for $0\leq i\leq \frac{\ell -4}{3}$, and
g_{2}(z)=0$ otherwise; $g_{3}(z)=f_{3}(z)$ for all $z\in V(G)$,
g_{3}(x_{3i+3})=2$ for $0\leq i\leq p-1$, $g_{3}(y_{3i+1})=2$ for $0\leq
i\leq \frac{\ell -1}{3}$, and $g_{3}(z)=0$ otherwise.
\smallskip \noindent \textbf{Case 3.} $\ell \equiv 2\pmod 3$.\newline
Define the functions $g_{1},g_{2}$ and $g_{3}$ on $V(H)$ as follows:
g_{1}(z)=f_{1}(z)$ for all $z\in V(G)$, $g_{1}(x_{3p})=1$,
g_{1}(x_{3i+1})=2 $ for $0\leq i\leq p-1$, $g_{1}(y_{3i+3})=2$ for $0\leq
i\leq \frac{\ell -5}{3}$, and $g_{1}(z)=0$ otherwise; $g_{2}(z)=f_{2}(z)$
for all $z\in V(G)$, $g_{2}(x_{3p+1})=1,$ $g_{2}(x_{3i+2})=2$ for $0\leq
i\leq p-1$, $g_{2}(y_{3i+2})=2$ for $0\leq i\leq \frac{\ell -2}{3}$, and
g_{2}(z)=0$ otherwise; $g_{3}(z)=f_{3}(z)$ for all $z\in V(G)$,
g_{3}(x_{3i+3})=2$ for $0\leq i\leq p-1$, $g_{3}(y_{3i+1})=2$ for $0\leq
i\leq \frac{\ell -2}{3}$, and $g_{3}(z)=0$ otherwise.
\bigskip
In either case, $g_{1},g_{2},g_{3}$ are RDFs of $H$ and thus
\overrightarrow{g}=(g_{1},g_{2},g_{3})$ is a 3-tuple of RDFs of $H.$ In
addition, $\omega (\overrightarrow{g},C_{3p+1,\ell })\leq 2(3p+1+\ell) $
and each vertex of $V(C_{3p+1,\ell })-\{x_{3p+1}\}$ is $\overrightarrow{g}
-strong.
The proofs of the remaining items are similar and therefore omitted. $\hfill
\Box $
\begin{lemma}
\label{cyclepath}
\begin{enumerate}
\item \emph{\ Let $C=v_{1}v_{2}\ldots v_{t}v_{1}$ be a cycle on $t\geq 4$
vertices with $t\equiv 1\pmod 3$. Then $C$ has a 3-tuple of RDFs
\overrightarrow{f}=(f_{1},f_{2},f_{3})$ such that $\omega (\overrightarrow{f
)\leq 2t+1$ and all vertices of $C$ but $v_{t}$ are $\overrightarrow{f}
-strong. }
\item \emph{Let $C=v_{1}v_{2}\ldots v_{t}v_{1}$ be a cycle on $t\geq 3$
vertices with $t\equiv 0\pmod 3$. Then $C$ has a 3-tuple of RDFs
\overrightarrow{f}=(f_{1},f_{2},f_{3})$ such that $\omega (\overrightarrow{f
)\leq 2t$ and all vertices of $C$ are $\overrightarrow{f}$-strong.}
\item \emph{Let $C_{m,\ell }$ be a tailed $m$-cycle with $m\equiv 1\pmod
3$ and $V(C_{m,\ell })=\{x_{1},x_{2},\ldots ,x_{m},y_{1},y_{2},\ldots
,y_{\ell }\},$ where the $x_{i}$'s induce a cycle {$C_{m}$ and the }$y_{i}
's induce a path $P_{\ell }$. Then $G$ has a 3-tuple of RDFs
\overrightarrow{f}=(f_{1},f_{2},f_{3})$ such that $\omega (\overrightarrow{f
)\leq 2(m+\ell )+1$ and all vertices of $C_{m,\ell }$ but $x_{m}$ are
\overrightarrow{f}$-strong.}
\end{enumerate}
\end{lemma}
\emph{\ }
\noindent \textbf{Proof. }1) Define the functions $f_{1},f_{2}$ and $f_{3}$
on $V(C)$ as follows: $f_{1}(v_{t-1})=1,f_{1}(v_{3i+1})=2$ for $0\leq i\leq
\frac{t-4}{3}$ and $f_{1}(x)=0$ otherwise; $f_{2}(v_{t})=1,$
f_{2}(v_{3i+2})=2$ for $0\leq i\leq \frac{t-4}{3}$ and $f_{2}(x)=0$
otherwise; $f_{3}(v_{1})=1,f_{3}(v_{3i+3})=2$ for $0\leq i\leq \frac{t-4}{3}$
and $f_{3}(x)=0$ otherwise. Clearly, $f_{1},f_{2},f_{3}$ are RDFs of $C.$
Hence $\overrightarrow{f}=(f_{1},f_{2},f_{3})$ is a 3-tuple of RDFs of $C$,
with $\omega (\overrightarrow{f})\leq 2t+1$ and all vertices of $C$ except
v_{t}$ are $\overrightarrow{f}$-strong.
2) Define the functions $f_{1},f_{2}$ and $f_{3}$ on $V(C)$ as follows:
f_{1}(v_{3i+1})=2$ for $0\leq i\leq \frac{t-3}{3}$ and $f_{1}(x)=0$
otherwise; $f_{2}(v_{3i+2})=2$ for $0\leq i\leq \frac{t-3}{3}$ and
f_{2}(x)=0$ otherwise; $f_{3}(v_{3i+3})=2$ for $0\leq i\leq \frac{t-3}{3}$
and $f_{3}(x)=0$ otherwise. Clearly, $f_{1},f_{2},f_{3}$ are RDFs of $C$.
Hence $\overrightarrow{f}=(f_{1},f_{2},f_{3})$ is a 3-tuple of RDFs of $C$
with $\omega (\overrightarrow{f})\leq 2t$ and each vertex of $C$ are
\overrightarrow{f}$-strong.
3) Define the functions $f_{1},f_{2}$ and $f_{3}$ on $V(C_{m,\ell }).$ For
vertices on $C_{m}$ as follows: $f_{1}(x_{m-1})=1,f_{1}(x_{3i+1})=2$ for
0\leq i\leq \frac{m-4}{3}$ and $f_{1}(x)=0$ otherwise; $f_{2}(x_{m})=1,$
f_{2}(x_{3i+2})=2$ for $0\leq i\leq \frac{m-4}{3}$ and $f_{2}(x)=0$
otherwise; $f_{3}(x_{3i+3})=2$ for $0\leq i\leq \frac{m-4}{3}$ and
f_{3}(x)=0$ otherwise. Moreover, the $f_{i}$'s are defined for the vertices
on $P_{\ell }$ according to $\ell $ as follows.
If $\ell \equiv 0\pmod 3$, then $f_{1}(y_{3i+3})=2$ for $0\leq i\leq \frac
\ell -3}{3}$ and $f_{1}(x)=0$ otherwise; $f_{2}(y_{3i+2})=2$ for $0\leq
i\leq \frac{\ell -3}{3}$ and $f_{2}(x)=0$ otherwise; $f_{3}(y_{\ell })=1$,
f_{3}(y_{3i+1})=2$ for $0\leq i\leq \frac{\ell -3}{3}$ and $f_{3}(x)=0$
otherwise.
If $\ell \equiv 1\pmod 3$, then $f_{1}(y_{3i+3})=2$ for $0\leq i\leq \frac
\ell -4}{3}$ and $f_{1}(x)=0$ otherwise; $f_{2}(y_{\ell })=1,$
f_{2}(y_{3i+2})=2$ for $0\leq i\leq \frac{\ell -4}{3}$ and $f_{2}(x)=0$
otherwise; $f_{3}(y_{3i+1})=2$ for $0\leq i\leq \frac{\ell -1}{3}$ and
f_{3}(x)=0$ otherwise.
If $\ell \equiv 2\pmod 3$, then $f_1(y_{\ell })=1,$ $f_{1}(y_{3i+3})=2$ for
0\leq i\leq \frac{\ell -5}{3}$ and $f_{1}(x)=0$ otherwise;
f_{2}(y_{3i+2})=2 $ for $0\leq i\leq \frac{\ell -2}{3}$ and $f_{2}(x)=0$
otherwise; $f_{3}(y_{3i+1})=2$ for $0\leq i\leq \frac{\ell -2}{3}$ and
f_{3}(x)=0$ otherwise.
Clearly, in either case $f_{1},f_{2},f_{3}$ are RDFs of {$C_{m,\ell}$} and thus
\overrightarrow{f}=(f_{1},f_{2},f_{3})$ is a 3-tuple of RDFs of {$C_{m,\ell}.$} Also,
\omega (\overrightarrow{f})\leq 2(m+\ell )+1$ and all vertices of $C_{m,\ell
}$ but $x_{m}$ are $\overrightarrow{f}$-strong.$\hfill \Box $
\begin{lemma}
\label{MainLem}Let $C_{i}=x_{1}^{i}x_{2}^{i}\ldots x_{n_{i}}^{i}x_{1}^{i}\;
be a cycle of order $n_{i},$ for $i\in \{1,2,\ldots ,s\}.$
\begin{enumerate}
\item If $n_{1}\equiv 2\pmod 3$ and $G$ is a graph obtained from $C_{1}$ and
$C_{2}$ by identifying the vertices $x_{1}^{1}$ and $x_{1}^{2}$, then $G$
has a 3-tuple of RDFs $\overrightarrow{f}$ such that $\omega
\overrightarrow{f})\leq 2n(G)+1$, and all vertices of $V(G)
\{x_{2}^{2},x_{n_{2}}^{2}\}$ are $\overrightarrow{f}$-strong.
\item If $n_{1}\equiv 2\pmod 3$, $n_{2}\equiv 1\pmod 3$ and $G$ is obtained
from $C_{1}$ and $C_{2}$ by adding either the edge $x_{1}^{1}x_{1}^{2}$ or a
path $z_{1}\ldots z_{3k}\;(k\geq 1)$ and the edges $x_{1}^{1}z_{1},$
x_{1}^{2}z_{3k}$, then $G$ has a 3-tuple of RDFs $\overrightarrow{f
=(f_1,f_2,f_3)$ such that $\omega (\overrightarrow{f})\leq 2n(G)+1$ and each
vertex of $G$ but $x_{n_{2}}^{2}$ is $\overrightarrow{f}$-strong.
\item If $n_{1}\equiv 2\pmod 3$, $n_{2}\equiv 1\pmod 3$ and $G$ is obtained
from $C_{1}$ and $C_{2}$ by adding for $k\geq 1$, a path $z_{1}\ldots
z_{3k+1}$ and the edges $x_{1}^{1}z_{1},$ $x_{1}^{2}z_{3k+1}$, then $G$ has
a 3-tuple of RDFs $\overrightarrow{f}$ such that $\omega (\overrightarrow{f
)\leq 2n(G)+1$ and all vertices of $G$ but $x_{n_{2}}^{2}$ are
\overrightarrow{f}$-strong.
\item If $n_{1}\equiv 2\pmod 3$, $n_{2}\equiv 1\pmod 3$ and $G$ is obtained
from $C_{1}$ and $C_{2}$ by adding for $k\geq 1$ a path $z_{1}\ldots
z_{3k+2}\;$and the edges $x_{1}^{1}z_{1},$ $x_{1}^{2}z_{3k+2}$, then $G$ has
a 3-tuple of RDFs $\overrightarrow{f}$ such that $\omega (\overrightarrow{f
)\leq 2n(G)+1$ and all vertices of $G$ but $x_{n_{2}}^{2}$ are
\overrightarrow{f}$-strong.
\item If $n_{i}\equiv 2\pmod 3$ for $i\in \{1,2\}$ and $G$ is obtained from
C_{1}$ and $C_{2}$ by adding either the edge $x_{1}^{1}x_{1}^{2}$ or a path
z_{1}\ldots z_{3k}\;(k\geq 1)$ and the edges $x_{1}^{1}z_{1},$
x_{1}^{2}z_{3k}$, then $G$ has a 3-tuple of RDFs $\overrightarrow{f}$ such
that $\omega (\overrightarrow{f})\leq 2n(G)+2$ and all vertices of $G$ are
\overrightarrow{f}$-strong.
\item If $n_{i}\equiv 2\pmod 3$ for $i\in \{1,2\}$ and $G$ is obtained fro
\textbf{\ }$C_{1}$ and $C_{2}$ by adding a path $z_{1}\ldots
z_{3k+1}\;(k\geq 1)$ and the edges $x_{1}^{1}z_{1},$ $x_{1}^{2}z_{3k+1}$,
then $G$ has a 3-tuple of RDFs $\overrightarrow{f}$ such that $\omega
\overrightarrow{f})\leq 2n(G)+2$ and all vertices of $G$ are
\overrightarrow{f}$-strong.
\item If $n_{i}\equiv 2\pmod 3$ for $i\in \{1,2\}$ and $G$ is obtained from
C_{1}$ and $C_{2}$ by adding a path $z_{1}\ldots z_{3k+2}\;(k\geq 1)$ and
the edges $x_{1}^{1}z_{1},$ $x_{1}^{2}z_{3k+2}$, then $G$ has a 3-tuple of
RDFs $\overrightarrow{f}$ such that $\omega (\overrightarrow{f})\leq 2n(G)+2$
and all vertices of $G$ are $\overrightarrow{f}$-strong.
\item If $n_{1}\equiv 2\pmod 3$, $n_{2}\equiv 0\pmod 3$ and $G$ is obtained
from $C_{1}$ and $C_{2}$ by adding either the edge $x_{1}^{1}x_{1}^{2}$ or a
path $z_{1}\ldots z_{3k}\;(k\geq 1)$ and the edges $x_{1}^{1}z_{1},$
x_{1}^{2}z_{3k}$, then $G$ has a 3-tuple of RDFs $\overrightarrow{f}$ such
that $\omega (\overrightarrow{f})\leq 2n(G)+1$ and all vertices of $G$ are
\overrightarrow{f}$-strong.
\item If $n_{1}\equiv 2\pmod 3$, $n_{2}\equiv 0\pmod 3$ and $G$ is obtained
from $C_{1}$ and $C_{2}$ by adding a path $z_{1}\ldots z_{3k+1}\;(k\geq 1)$
and the edges $x_{1}^{1}z_{1},$ $x_{1}^{2}z_{3k+1}$, then $G$ has a 3-tuple
of RDFs $\overrightarrow{f}$ such that $\omega (\overrightarrow{f})\leq
2n(G)+1$ and all vertices of $G$ are $\overrightarrow{f}$-strong.
\item If $n_{1}\equiv 2\pmod 3$, $n_{2}\equiv 0\pmod 3$ and $G$ is obtained
from $C_{1}$ and $C_{2}$ by adding a path $z_{1}\ldots z_{3k+2}\;(k\geq 1)$
and the edges $x_{1}^{1}z_{1},$ $x_{1}^{2}z_{3k+2}$, then $G$ has a 3-tuple
of RDFs $\overrightarrow{f}$ such that $\omega (\overrightarrow{f})\leq
2n(G)+1$ and all vertices of $G$ are $\overrightarrow{f}$-strong.
\item {If $s\geq 3$, $n_{i}\equiv 2\pmod 3$ for each $i$ and $G$ is obtained
from $C_{1},\ldots ,C_{s}$ by adding a new vertex $x$ and the edges
xx_{1}^{1},\ldots ,xx_{1}^{s}$, then $G$ has a 3-tuple of RDFs
\overrightarrow{f}$ such that $\omega (f,G)\leq 2n(G)-s+4$ and the vertex $x$
is $\overrightarrow{f}$-strong.}
\end{enumerate}
\end{lemma}
\noindent\textbf{Proof.}
\begin{enumerate}
\item Define the functions $f_{1},f_{2}$ and $f_{3}$ on $V(G)$ as follows$.$
For vertices on $C_{1}$: $f_{1}(x_{3i+1}^{1})=f_{2}(x_{3i+2}^{1})=2$ for
each $0\leq i\leq \frac{n_{1}-2}{3}$ and $f_{i}(x)=0$ otherwise,
for $i=1,2$, and $f_{3}(x_{3i+3}^{1})=2$ for each $0\leq i\leq
\frac{n_{1}-5}{3}$, $f_{3}(x_{1}^{1})=2$, and $f_{3}(x)=0$ otherwise.} Now
for vertices on $C_{2}$ but $x_{1}^{2}$:
If $n_{2}\equiv 0\pmod 3$, then let $f_{1}(x_{3i+4}^{2})=2$ for each $0\leq
i\leq \frac{n_{2}-6}{3}$ and $f_{1}(x)=0$ otherwise; $f_{2}(x_{3i+3}^{2})=2$
for each $0\leq i\leq \frac{n_{2}-3}{3}$ and $f_{2}(x)=0$ otherwise;
f_{3}(x_{3i+2}^{2})=2$ for each $0\leq i\leq \frac{n_{2}-3}{3}$ and
f_{3}(x)=0$ otherwise.
If $n_{2}\equiv 1\pmod 3$, then let $f_{1}(x_{3i+4}^{2})=2$ for each $0\leq
i\leq \frac{n_{2}-7}{3}$, $f_{1}(x_{n_{2}-1}^{2})=1$ and $f_{1}(x)=0$
otherwise; $f_{2}(x_{3i+3}^{2})=2$ for each $0\leq i\leq \frac{n_{2}-4}{3}$
and $f_{2}(x)=0$ otherwise; {$f_{3}(x_{n_{2}}^{2})=1$,}
f_{3}(x_{3i+2}^{2})=2$ for each $0\leq i\leq \frac{n_{2}-4}{3}$ and
f_{3}(x)=0$ otherwise.
If $n_{2}\equiv 2\pmod 3$, then let $f_{1}(x_{3i+4}^{2})=2$ for each $0\leq
i\leq \frac{n_{2}-5}{3}$ and $f_{1}(x)=0$ otherwise; $f_{2}(x_{3i+3}^{2})=2$
for each $0\leq i\leq \frac{n_{2}-5}{3}$, $f_{2}(x_{n_{2}}^{2})=1$ and
f_{2}(x)=0$ otherwise;
$f_{3}(x_{3}^{2})=f_{3}(x_{n_{2}-1}^{2})=1$, $f_{3}(x_{3i+5}^{2})=2$ for each
$0\leq i\leq \frac{n_{2}-8}{3}$ and $f_{3}(x)=0$ otherwise.
In either case, $f_{1},f_{2}$ and $f_{3}$ are RDFs of $G.$ Hence $\omega
\overrightarrow{f})\leq 2n(G)+1$, and all vertices of $V(G)
\{x_{2}^{2},x_{n_{2}}^{2}\}$ are $\overrightarrow{f}$-strong. \textbf{\ }
\item Define the functions $f_{1},f_{2}$ and $f_{3}$ on $V(G)$ as follows:
f_{1}(x_{3i+1}^{1})=2$ for each $0\leq i\leq \frac{n_{1}-2}{3}$,
f_{1}(z_{3i+3})=2$ for $0\leq i\leq k-1$, $f_{1}(x_{3i+3}^{2})=2$ for each
0\leq i\leq \frac{n_{2}-4}{3}$ and $f_{1}(x)=0$ otherwise;
f_{2}(x_{3i+2}^{1})=2$ for each $0\leq i\leq \frac{n_{1}-2}{3}$,
f_{2}(z_{3i+2})=2$ for $0\leq i\leq k-1$, $f_{2}(x_{3i+2}^{2})=2$ for each
0\leq i\leq \frac{n_{2}-4}{3}$, $f_{2}(x_{n_{2}}^{2})=1$ and $f_{2}(x)=0$
otherwise; $f_{3}(x_{n_{1}}^{1})=1$, $f_{3}(x_{3i+3}^{1})=2$ for each $0\leq
i\leq \frac{n_{1}-5}{3}$, $f_{3}(z_{3i+1})=2$ for $0\leq i\leq k-1$,
f_{3}(x_{3i+1}^{2})=2$ for each $0\leq i\leq \frac{n_{2}-4}{3}$,
f_{3}(x_{n_{2}-1}^{2})=2$ and $f_{3}(x)=0$ otherwise.
Clearly $f_{1},f_{2}$ and $f_{3}$ are RDFs of $G,$ and thus $\overrightarrow
f}=(f_{1},f_{2},f_{3})$ is a 3-tuple of RDFs of $G.$ Moreover, $\omega
\overrightarrow{f})\leq 2n(G)+1$ and each vertex of $G$ but $x_{n_{2}}^{2}$
is $\overrightarrow{f}$-strong.
\item Define the functions $f_{1},f_{2}$ and $f_{3}$ on $V(G)$ as follows:
f_{1}(x_{3i+1}^{1})=2$ for each $0\leq i\leq \frac{n_{1}-2}{3}$,
f_{1}(z_{3i+3})=2$ for $0\leq i\leq k-1$, $f_{1}(x_{3i+2}^{2})=2$ for each
0\leq i\leq \frac{n_{2}-4}{3}$, $f_{1}(x_{n_{2}}^{2})=1$ and $f_{1}(x)=0$
otherwise; $f_{2}(x_{3i+2}^{1})=2$ for each $0\leq i\leq \frac{n_{1}-2}{3}$,
$f_{2}(z_{3i+2})=2$ for $0\leq i\leq k-1$, $f_{2}(x_{3i+1}^{2})=2$ for each
0\leq i\leq \frac{n_{2}-4}{3}$, $f_{2}(x_{n_{2}-1}^{2})=1$ and $f_{2}(x)=0$
otherwise; $f_{3}(x_{n_{1}}^{1})=1$, $f_{3}(x_{3i+3}^{1})=2$ for each $0\leq
i\leq \frac{n_{1}-5}{3}$, $f_{3}(z_{3i+1})=2$ for $0\leq i\leq k$,
f_{3}(x_{3i+3}^{2})=2$ for each $0\leq i\leq \frac{n_{2}-4}{3}$ and
f_{3}(x)=0$ otherwise.
Clearly $f_{1},f_{2}$ and $f_{3}$ are RDFs of $G$ and thus $\overrightarrow{
}=(f_{1},f_{2},f_{3})$ is a 3-tuple of RDFs of $G.$ Also, $\omega
\overrightarrow{f})\leq 2n(G)+1$ and all vertices of $G$ but $x_{n_{2}}^{2}$
are $\overrightarrow{f}$-strong.
\item Define the functions $f_{1},f_{2}$ and $f_{3}$ on $V(G)$ as follows:
f_{1}(x_{3i+1}^{1})=2$ for each $0\leq i\leq \frac{n_{1}-2}{3}$,
f_{1}(z_{3i+3})=2$ for $0\leq i\leq k-1$, $f_{1}(x_{3i+1}^{2})=2$ for each
0\leq i\leq \frac{n_{2}-4}{3}$, $f_{3}(x_{n_{2}-1}^{2})=1$ and $f_{1}(x)=0$
otherwise; $f_{2}(x_{3i+2}^{1})=2$ for each $0\leq i\leq \frac{n_{1}-2}{3}$,
$f_{2}(z_{3i+2})=2$ for $0\leq i\leq k$, $f_{2}(x_{3i+3}^{2})=2$ for each
0\leq i\leq \frac{n_{2}-4}{3}$ and $f_{2}(x)=0$ otherwise;
f_{3}(x_{n_{1}}^{1})=1$, $f_{3}(x_{3i+3}^{1})=2$ for each $0\leq i\leq \frac
n_{1}-5}{3}$, $f_{3}(z_{3i+1})=2$ for $0\leq i\leq k$,
f_{3}(x_{3i+2}^{2})=2 $ for each $0\leq i\leq \frac{n_{2}-4}{3}$,
f_{3}(x_{n_{2}}^{2})=1$ and $f_{3}(x)=0$ otherwise.
Clearly $f_{1},f_{2}$ and $f_{3}$ are RDFs of $G$ and thus $\overrightarrow{
}=(f_{1},f_{2},f_{3})$ is a 3-tuple of RDFs of $G.$ Also, $\omega
\overrightarrow{f})\leq 2n(G)+1$ and each vertex of $G$ but $x_{n_{2}}^{2}$
is $\overrightarrow{f}$-strong.
\item Define the functions $f_{1},f_{2}$ and $f_{3}$ on $V(G)$ as follows:
f_{1}(x_{3i+1}^{1})=2$ for each $0\leq i\leq \frac{n_{1}-2}{3}$,
f_{1}(z_{3i+3})=2$ for $0\leq i\leq k-1$, $f_{1}(x_{3i+3}^{2})=2$ for each
0\leq i\leq \frac{n_{2}-5}{3}$, $f_{1}(x_{n_{2}}^{2})=1$ and $f_{1}(x)=0$
otherwise; $f_{2}(x_{3i+2}^{1})=2$ for each $0\leq i\leq \frac{n_{1}-2}{3}$,
$f_{2}(z_{3i+2})=2$ for $0\leq i\leq k-1$, $f_{2}(x_{3i+2}^{2})=2$ for each
0\leq i\leq \frac{n_{2}-2}{3}$ and $f_{2}(x)=0$ otherwise;
f_{3}(x_{n_{1}}^{1})=1$, $f_{3}(x_{3i+3}^{1})=2$ for each $0\leq i\leq \frac
n_{1}-5}{3}$, $f_{3}(z_{3i+1})=2$ for $0\leq i\leq k-1$,
f_{3}(x_{3i+1}^{2})=2$ for each $0\leq i\leq \frac{n_{2}-2}{3}$ and
f_{3}(x)=0$ otherwise.
Clearly $f_{1},f_{2}$ and $f_{3}$ are RDFs of $G$ and thus $\overrightarrow{
}=(f_{1},f_{2},f_{3})$ is a 3-tuple of RDFs of $G.$ Further, $\omega
\overrightarrow{f})\leq 2n(G)+2$ and all vertices of $G$ are
\overrightarrow{f}$-strong.
\item The proof is similar to that of item (5).
\item The proof is similar to that of item (5).
\item Define the functions $f_{1},f_{2}$ and $f_{3}$ on $V(G)$ as follows:
f_{1}(x_{3i+1}^{1})=2$ for each $0\leq i\leq \frac{n_{1}-2}{3}$,
f_{1}(z_{3i+3})=2$ for $0\leq i\leq k-1$, $f_{1}(x_{3i+3}^{2})=2$ for each
0\leq i\leq \frac{n_{2}-3}{3}$ and $f_{1}(x)=0$ otherwise;
f_{2}(x_{3i+2}^{1})=2$ for each $0\leq i\leq \frac{n_{1}-2}{3}$,
f_{2}(z_{3i+2})=2$ for $0\leq i\leq k-1$, $f_{2}(x_{3i+2}^{2})=2$ for each
0\leq i\leq \frac{n_{2}-3}{3}$ and $f_{2}(x)=0$ otherwise;
f_{3}(x_{n_{1}}^{1})=1$, $f_{3}(x_{3i+3}^{1})=2$ for each $0\leq i\leq \frac
n_{1}-5}{3}$, $f_{3}(z_{3i+1})=2$ for $0\leq i\leq k-1$,
f_{3}(x_{3i+1}^{2})=2$ for each $0\leq i\leq \frac{n_{2}-3}{3}$ and
f_{3}(x)=0$ otherwise.
Then $f_{1},f_{2}$ and $f_{3}$ are RDFs of $G$ and thus $\overrightarrow{f
=(f_{1},f_{2},f_{3})$ is a 3-tuple of RDFs of $G$ with the desired property.
\item The proof is similar to that of item (8).
\item The proof is similar to that of item (8).
\item {Define the function $f_{1}$ by $f_{1}(x)=2$, $f_{1}(x_{n_{j}}^{j})=1$
for each $1\leq j\leq s$, $f_{1}(x_{i}^{j})=2$ for each $j$ and each
i\equiv 0\pmod 3$, and $f_{1}(y)=0$ otherwise, and set $\overrightarrow{f
=(f_{1},f_{1},f_{1})$. }Clearly $f_{1}$ is an RDF of $G$ and thus
\overrightarrow{f}=(f_{1},f_{2},f_{3})$ is a 3-tuple of RDFs of {$G$ such
that $\omega (\overrightarrow{f})\leq 2n(G)-s+4$ and the vertex $x$ is
\overrightarrow{f}$-strong as desired.} $\hfill \Box $
\end{enumerate}
\begin{lemma}
\label{ear2}Let $H$ be a graph obtained from a cycle $C_{3p+2}=x_{1}x_{2
\ldots x_{3p+2}x_{1}$ and a path $Q=y_{1}\ldots y_{\ell }$ where $\ell
\equiv 1\;\mathrm{or}\;2\pmod 3$ by adding the edge $y_{1}x_{1}$ and joining
$y_{\ell }$ to some vertices in $V(C_{3p+2})-\{x_{1}\}$ with the condition
that:
\begin{description}
\item[(a)] if $\ell\equiv 1\pmod 3$ and $y_{\ell}x_{j}\in E(H)$, then
j\not\equiv 2\pmod 3$,
\item[(b)] if $\ell\equiv 2\pmod 3$ and $y_{\ell}x_{j}\in E(H)$, then
$j\equiv 2\pmod 3$.}
\end{description}
Then there exists a 3-tuple $\overrightarrow{g}=(g_{1},g_{2},g_{3})$ of RDF
of $H$ such that $\omega (\overrightarrow{g},H)\leq 2n(H)+1$ and each vertex
of $H$ but $y_{1},y_{\ell }$ is $\overrightarrow{g}$-strong.
\end{lemma}
\noindent \textbf{Proof.} First let $\ell \equiv 1\pmod 3$ and $y_{\ell
}x_{j}\in E(H).$ Define the functions $g_{1},g_{2}$ and $g_{3}$ on $V(H)$ as
follows, depending on whether $j\equiv 0\pmod 3$ or $j\equiv 1\pmod 3.$
If $j\equiv 0\pmod 3,$ then let $g_{1}(x_{3i+1})=2$ for $0\leq i\leq p$,
g_{1}(y_{3i+3})=2$ for $0\leq i\leq \frac{\ell -4}{3}$, and $g_{1}(z)=0$
otherwise; $g_{2}(x_{3i+2})=2$ for $0\leq i\leq p$, $g_{2}(y_{\ell })=1$,
g_{2}(y_{3i+2})=2$ for $0\leq i\leq \frac{\ell -4}{3}$, and $g_{2}(z)=0$
otherwise; $g_{3}(x_{1})=2,$ $g_{3}(x_{3i+3})=2$ for $0\leq i\leq p-1$,
g_{3}(y_{3i+2})=2$ for $0\leq i\leq \frac{\ell -4}{3}$, and $g_{1}(z)=0$
otherwise.
If $j\equiv 1\pmod 3$, then let $g_{1}(x_{3i+1})=2$ for $0\leq i\leq p$,
g_{1}(y_{3i+3})=2$ for $0\leq i\leq \frac{\ell -4}{3}$, and $g_{1}(z)=0$
otherwise; $g_{2}(x_{3i+2})=2$ for $0\leq i\leq p$, $g_{2}(y_{\ell })=1$,
g_{2}(y_{3i+2})=2$ for $0\leq i\leq \frac{\ell -4}{3}$, and $g_{2}(z)=0$
otherwise; $g_{3}(x_{1})=2,$ $g_{3}(x_{3i+3})=2$ for $0\leq i\leq p-1$,
g_{3}(y_{3i+3})=2$ for $0\leq i\leq \frac{\ell -4}{3}$, and $g_{1}(z)=0$
otherwise.
Second, let $\ell \equiv 2\pmod 3$ and $y_{\ell }x_{j}\in E(H),$ where
$j\equiv 2\pmod 3$.} Define the functions $g_{1},g_{2}$ and
g_{3}$ on $V(H)$ as follows: $g_{1}(x_{p})=1,$ $g_{1}(x_{3i+3})=2$ for
0\leq i\leq p-1$, $g_{1}(y_{3i+1})=2$ for $0\leq i\leq \frac{\ell -2}{3}$,
and $g_{1}(z)=0$ otherwise; $g_{2}(x_{3i+2})=2$ for $0\leq i\leq p$,
g_{2}(y_{3i+2})=2$ for $0\leq i\leq \frac{\ell -5}{3}$, $g_2(y_{\ell-1})=1$,
and $g_{2}(z)=0$ otherwise; $g_{3}(x_{3i+1})=2$ for $0\leq i\leq p$,
g_{3}(y_{3i+3})=2$ for $0\leq i\leq \frac{\ell -5}{3}$, $g_3(y_{\ell})=1$,
and $g_{1}(z)=0$ otherwise.
Clearly, $g_{1},g_{2},g_{3}$ are RDFs of $H$ and thus $\overrightarrow{g
=(g_{1},g_{2},g_{3})$ is a 3-tuple of RDFs of $H$ with the desired property.
$\hfill \Box $
\section{Partial answer to Conjecture \protect\ref{conj}}
In this section, we give a positive answer to Conjecture \ref{conj} for some
particular graphs. We start with the following two lemmas.
\begin{lemma}
\label{induced}\emph{Let }$k\geq 1$ \emph{be an integer and let $G$ be a
connected graph with $\delta \geq 2$, which does not contain {\ neither }any
induced $\{C_{5},C_{8},\ldots ,C_{3k+2}\}$-cycles {\ nor }any cycle of
length $\equiv 0\pmod 3$. Let $C$ be a cycle of $G$ with length $\ell (C)$
\equiv 2\pmod 3$. Then }
\begin{enumerate}
\item \emph{if $C$ is induced in $G$, then $\ell (C)\geq 3k+5$, and }
\item \emph{if $C$ is not induced in $G$, then $\ell (C)\geq 6k+8$. }
\end{enumerate}
\end{lemma}
\noindent \textbf{Proof.} Item (1) is immediate since $G$ does not contain
any induced $\{C_{5},C_{8},\ldots ,C_{3k+2}\}$-cycles and $\ell (C)\equiv
\pmod 3$. To prove item (2), let $C=v_{1}v_{2}\ldots v_{3p+2}v_{1}$ be a
cycle which is not induced in $G$. Hence $C$ has a chord, say without loss
of generality, $v_{1}v_{i}\in E(G)$. Consider the two paths
P=v_{i+1}v_{i+2}\ldots v_{3p+2}v_{1}$ and $Q=v_{2}v_{3}\ldots v_{i}$. Let
n(P)$ and $n(Q)$ denote the order of $P$ and $Q,$ respectively. Clearly
n(P)+n(Q)=3p+2$. Now, if $n(P)\equiv 0\pmod 3$, then $v_{1}v_{2}v_{3}\ldots
v_{i}v_{1}$ is a cycle of length $\equiv 0\pmod 3,$ contradicting the fact
that $G$ has no cycle of such length. Hence $n(P)\not\equiv 0\pmod 3$, and
likewise $n(Q)\not\equiv 0\pmod 3$. Moreover, since $n(P)+n(Q)=3p+2,$ we
deduce that $n(P)\not\equiv 2\pmod 3$ and $n(Q)\not\equiv 2\pmod 3$. Hence
n(P)\equiv 1\pmod 3$ and $n(Q)\equiv 1\pmod 3$. Consider the cycles
C_{1}=v_{1}v_{2}v_{3}\ldots v_{i}v_{1}$ and $C_{2}=v_{1}v_{i}v_{i+1}\ldots
v_{3p+2}v_{1}$. Then $\ell (C_{1})\equiv 2\pmod 3$ and $\ell (C_{2})\equiv
\pmod 3$. If $C_{1}$ and $C_{2}$ are induced in $G$, then by item (1) we
have $i\geq 3k+5$ and $3p+4-i\geq 3k+5$ and thus $\ell (C)=3p+2\geq
3k+5+i-2\geq 3k+5+3k+5-2=6k+8$. Hence we assume that $C_{1}$ is not induced
in $G$. By repeating the above process we can see that the subgraph
G[V(C_{1})]$ has an induced cycle of length $\equiv 2\pmod 3$ and so
|V(C_{1})|=i\geq 3k+5$. If $C_{2}$ is an induced cycle, then by item (1) we
have $3p+4-i\geq 3k+5$ and so $\ell (C)=3p+2\geq 3k+5+i-2\geq
3k+5+3k+5-2=6k+8$. Now if $C_{2}$ is not an induced cycle, then a similar
argument as above shows that $G[V(C_{2})]$ has an induced cycle of length
\equiv 2\pmod 3$ yielding also $\ell (C)\geq 6k+8$.$\hfill \Box $
\begin{lemma}
\label{12}Let $G$ be a connected\textbf{\ }graph with minimum degree $\delta
\geq 2$ and let $G_{1}$ and $G_{2}$ be two non-null subgraphs of $G$ such
that $V(G)=V(G_{1})\cup V(G_{2})$. Then one of the following holds:
\begin{enumerate}
\item $G_{1}$ has a path $P=v_{1}\ldots v_{t}$ such that both $v_{1}$ and
v_{t}$ have neighbors in $G_{2}$ and $N_{G}(v_{1})\cup N_{G}(v_{t})\subseteq
V(G_{2})\cup V(P)$.
\item $G_{1}$ has a cycle $C=v_{1}v_{2}\ldots v_{t}v_{1}$ such that $v_{1}$
has neighbors in $G_{2}$ and $N_{G}(v_{t})\subseteq V(G_{2})\cup V(C)$.
\item $G_{1}$ contains a tailed $m$-cycle, say $C_{m,\ell }$, such that
y_{\ell }$ is adjacent to some vertex in $G_{2}$ and $N_{G}(x_{2})\cup
N_{G}(x_{m})\subseteq V(G_{2})\cup V(C_{m,\ell })$.
\end{enumerate}
\end{lemma}
\noindent \textbf{Proof.} Let $\mathcal{P}$ be the family of all longest
paths (not necessarily induced) in $G_{1}$ such that at least one of their
end-points has a neighbor in $G_{2}$ and let $Q=\{v\in V(G)\mid \mathrm
there\;is\;a\;path}$ $v_{1},\ldots ,v_{t}(=v)\in \mathcal{P}\;\mathrm
such\;that}\;v_{1}\;\mathrm{has}$ $\mathrm{a}$ $\mathrm{neighbor}$ $\mathrm
in}\;G_{2}\}$. Choose a vertex $v\in Q$ such that the length of its
corresponding path $P=v_{1},\ldots ,v_{t}(=v)\in \mathcal{P}$ is as long as
possible.
First let $v$ be adjacent to some vertex in $G_{2}$. By the definition of
set $Q,$ we have $v_{1}\in Q$, and from the choice of $v$ we deduce that
N_{G}(v_{1})\cup N_{G}(v_{t})\subseteq V(G_{2})\cup V(P).$ Hence item (1)
holds. Suppose now $v$ has no neighbor in $G_{2}$. It follows from the
choice of $v$ and the fact $\delta \geq 2$ that $v$ has at least two
neighbors in $V(P)$. Let $j$ be the smallest index such that $vv_{j}$ $\in
E(G).$ Now, if $j=1,$ that is $v$ is adjacent to $v_{1}$, then clearly
N_{G}(v_{t})\subseteq V(P)$ and thus item (2) holds. Hence we assume that
j\neq 1.$ Then $v_{1}\ldots v_{j}...v_{t-1}vv_{j}$\textbf{\ }is a tailed
cycle contained in $G_{1}.$ Observe that the path with endvertices $v_{1}$
and $v_{j+1}$ starting from $v_{1}$ to $v_{j}$ and then passing through
v_{t}$ to $v_{j+1}$ is also a longest path with same length as $P.$ Since
v=v_{t}$ has no neighbor in $G_{2},$ we may assume by analogy that $v_{j+1}$
has no neighbor in $G_{2}$ and thus all its neighbors are on\textbf{\ }$P$
which forms a tailed cycle and thus item (3) holds$.$\textbf{\ }$\hfill \Box
$
\begin{theorem}
\label{Th1}\emph{\emph{Let} $k\geq 1$ be an integer and let $G$ be a
connected graph of order $n\geq 6k+9$ and minimum degree at least 2 such
that $G$ has no cycle with length $\equiv 0\;\mathrm{or}\;2\pmod 3$. Then
\gamma _{R}(G)\leq \frac{(4k+8)n}{6k+11}$.}
\end{theorem}
\noindent \textbf{Proof.} Let $Q=z_{1}z_{2}\ldots z_{r}$ be a longest path
in $G$. If $V(G)=V(Q)$, then we have $\gamma _{R}(G)\leq \frac{2n+1}{3}
\frac{(4k+8)n}{6k+11}$. Hence, we assume that $V(Q)\subsetneqq V(G)$. By the
choice of $Q$ we have $N_{G}(z_{1})\cup N_{G}(z_{r})\subseteq V(Q)$. Since
\delta (G)\geq 2$, $z_{1}$ is adjacent to some $z_{j}$ with $j\equiv 1\pmod
,$ because $G$ has no cycle with length $\equiv 0\;\mathrm{or}\;2\pmod 3$.
Let $G_{2}^{0}$ be the graph obtained from the path $Q$ to which we add the
edge $z_{1}z_{j}$ and let $G_{1}^{0}$ be the graph induced by
V(G)-V(G_{2}^{0}).$ Observe that $G_{2}^{0}$ is a tailed $j$-cycle
C_{j,r-j}.$ By Lemma \ref{cyclepath}-(3), $G_{2}^{0}$ has a 3-tuple of RDFs
\overrightarrow{f}=(f_{1},f_{2},f_{3})$ such that $\omega (\overrightarrow{f
)\leq 2r+1$ and all vertices of $C_{j,r-j}$ but $z_{1}$ are $\overrightarrow
f}$-strong. According to Lemma \ref{12}, we consider the following three
possibilities.
\begin{description}
\item[(a)] $G_{1}^{0}$ has a path $P=v_{1}\ldots v_{t}$ such that
v_{1},v_{t}$ are adjacent to some vertices in $V(G_{2}^{0})-\{z_{1},z_{t}\}
, say $u,v$ (possibly $u=v$) and $N_{G}(v_{1})\cup N_{G}(v_{t})\subseteq
V(G_{2}^{0})\cup V(P)$.\newline
Let $G_{2}^{1}$ be the graph obtained from $G_{2}^{0}$ and the path $P$ by
adding the edges $v_{1}u,v_{t}v$. By Lemma \ref{ear1}, $\overrightarrow{f}$
can be extended to a 3-triple of RDFs $\overrightarrow{g}$ of $G_{2}^{1}$
such that $\omega (\overrightarrow{g},P)\leq 2n(P)$ and each vertex in
V(P)-\{v_{1},v_{t}\}$ is $\overrightarrow{g}$-strong. Note that $\omega
\overrightarrow{g})=\omega (\overrightarrow{f})+\omega (\overrightarrow{g
,P)\leq 2n(G_{2}^{1})+1.$
\item[(b)] $G_{1}^{0}$ has a cycle $C=v_{1},\ldots ,v_{t}v_{1}$ such that
v_{1}$ is adjacent to a vertex in $G_{2}^{0}$, say $u$, and
N_{G}(v_{t})\subseteq V(G_{2}^{0})\cup V(C)$.\newline
Since $G$ has no cycle of length $\equiv 0\;\mathrm{or}\;2\pmod 3$, we have
t\equiv 1\pmod 3$. Let $G_{2}^{1}$ be the graph obtained from $G_{2}^{0}$
and the cycle $C$ by adding the edge $v_{1}u$. By Lemma \re
{tailedcycle-3p+1}, $\overrightarrow{f}$ can be extended to a 3-triple of
RDFs $\overrightarrow{g}$ of $G_{2}^{1}$ such that $\omega (g,C)\leq 2n(C)$
and each vertex in $V(C)-\{v_{t}\}$ is $\overrightarrow{g}$-strong. In
addition, it is clear that $\omega (\overrightarrow{g})\leq \omega
\overrightarrow{f})+\omega (\overrightarrow{g},C)\leq 2n(G_{2}^{1})+1$.
\item[(c)] $G_{1}^{0}$ contains a tailed $m$-cycle $C_{m,\ell }$, such that
y_{\ell }$ is adjacent to some vertex in $G_{2}^{0}$, say $u$, and
N_{G}(x_{2})\cup N_{G}(x_{m})\subseteq V(G_{2}^{0})\cup V(C_{m,\ell })$
\newline
As above in (b), since $G$ has no cycle of length $\equiv 0\;\mathrm{or}\;
\pmod 3$, we have $m\equiv 1\pmod 3$. Let $G_{2}^{1}$ be the graph obtained
from $G_{2}^{0}$ and the tailed $m$-cycle $C_{m,\ell }$ by adding the edge
uy_{\ell }$. By Lemma \ref{tailedcycle-3p+1}, $\overrightarrow{f}$ can be
extended to a 3-triple of RDFs $\overrightarrow{g}$ of $G_{2}^{1}$ such that
$\omega (\overrightarrow{g},C_{m,\ell })\leq 2n(C_{m,\ell })$ and each
vertex of $C_{m,\ell }$ but $x_{m}$\ is $\overrightarrow{g}$-strong.
Therefore, we also have $\omega (\overrightarrow{g})\leq 2n(G_{2}^{1})+1.$
\end{description}
Now, let $G_{1}^{1}=G-G_{2}^{1}$. By repeating the above process, we obtain
a $3$-tuple of RDFs $G$ that is $\overrightarrow{h}=(h_{1},h_{2},h_{3})$
such that $\omega (\overrightarrow{h})\leq 2n(G)+1\leq \frac{3n(4k+8)}{6k+8}
. Therefore, $\omega (h_{j})\leq \frac{n(4k+8)}{6k+8}$ for some $j\in
\{1,2,3\},$ and this completes the proof.$\hfill \Box $
\begin{theorem}
\label{Th2}\emph{\emph{Let} $k\geq 1$ be an integer and let $G$ be a
connected graph of order $n\geq 6k+9$ with minimum degree at least 2 and
having a cycle $C$ with length $\equiv 0\pmod
3$ such that any other cycle of $G$ with length $\equiv 0\;\mathrm{or}\;
\pmod 3$ has at least a common vertex with $C$. Then $\gamma _{R}(G)\leq
\frac{(4k+8)n}{6k+11}$.}
\end{theorem}
\noindent \textbf{Proof.} Assume that the vertices of the cycle $C$ with
length $\equiv 0\pmod
3$ are labelled by $z_{1}z_{2}\ldots z_{r}z_{1}.$ If $V(G)=V(C)$, then
clearly $\gamma _{R}(G)\leq \frac{2n}{3}<\frac{(4k+8)n}{6k+11}$. Hence, we
assume that $V(C)\subsetneqq V(G)$. Let $G_{2}^{0}=C$ and let $G_{1}^{0}$ be
the graph induced by $V(G)-V(G_{2}^{0}).$ By Lemma \ref{cyclepath},
G_{2}^{0}$ has a 3-tuple of RDFs $\overrightarrow{f}=(f_{1},f_{2},f_{3})$
such that $\omega (\overrightarrow{f})\leq 2r$ and all vertices of $C$ are
\overrightarrow{f}$-strong. Now, according to Lemma \ref{12}, we consider
the following three possibilities.
\begin{description}
\item[(a)] $G_{1}^{0}$ has a path $P=v_{1}\ldots v_{t}$ such that
v_{1},v_{t}$ have neighbors in $V(G_{2})$, say $u,v$ (possibly $u=v$), and
N_{G}(v_{1})\cup N_{G}(v_{t})\subseteq V(G_{2}^{0})\cup V(P)$.\newline
Let $G_{2}^{1}$ be the graph obtained from $G_{2}^{0}$ and the path $P$ by
adding the edges $v_{1}u$ and $v_{t}v$. By Lemma \ref{ear1},
\overrightarrow{f}$ can be extended to a 3-triple of RDFs $\overrightarrow{g}
$ of $G_{2}^{1}$ such that $\omega (g,P)\leq 2n(P)=2t$ and each vertex in
V(P)-\{v_{1},v_{t}\}$ is $\overrightarrow{g}$-strong. In this case, we have
\omega (\overrightarrow{g})=\omega (\overrightarrow{f})+\omega
\overrightarrow{g},P)\leq 2n(G_{2}^{1}).$
\item[(b)] $G_{1}^{0}$ has a cycle $C^{\prime }=v_{1},\ldots ,v_{t}v_{1}$
such that $v_{1}$ is adjacent to a vertex in $G_{2}^{0}$, say $u$, and
N_{G}(v_{t})\subseteq V(G_{2}^{0})\cup V(C^{\prime })$.\newline
By assumption, we have $t\equiv 1\pmod 3$. Let $G_{2}^{1}$ be\ the graph
obtained from $G_{2}^{0}$ and the cycle $C^{\prime }$ by adding the edge
v_{1}u$. By Lemma \ref{tailedcycle-3p+1}, $\overrightarrow{f}$ can be
extended to a 3-triple of RDFs $\overrightarrow{g}$ of $G_{2}^{1}$ such that
$\omega (\overrightarrow{g},C^{\prime })\leq 2n(C^{\prime })=2t$ and each
vertex in $V(C^{\prime })-\{v_{t}\}$ is $\overrightarrow{g}$-strong.
Moreover, we also obtain $\omega (\overrightarrow{g})\leq 2n(G_{2}^{1})$.
\item[(c)] $G_{1}^{0}$ contains a tailed $m$-cycle $C_{m,\ell }$, such that
y_{\ell }$ is adjacent to some vertex in $G_{2}^{0}$, say $u$, and
N_{G}(x_{2})\cup N_{G}(x_{m})\subseteq V(G_{2}^{0})\cup V(C_{m,\ell })$
\newline
As above, $m\equiv 1\pmod 3$. Let $G_{2}^{1}$ be the graph obtained from
G_{2}^{0}$ and the tailed $m$-cycle $C_{m,\ell }$ by adding the edge
uy_{\ell }$. By Lemmas \ref{tailedcycle-3p+1}, $\overrightarrow{f}$ can be
extended to a 3-triple of RDFs $\overrightarrow{g}$ of $G_{2}^{1}$ such that
such that $\omega (\overrightarrow{g},C_{m,\ell })\leq 2n(C_{m,\ell })$ and
each vertex of $C_{m,\ell }$ but $x_{m}$\ is $\overrightarrow{g}$-strong. In
addition, we have $\omega (\overrightarrow{g})\leq 2n(G_{2}^{1})$.
\end{description}
Let $G_{1}^{1}=G-G_{2}^{1}$. By repeating the above process, we obtain a
3-tuple of RDFs $G$ that is $\overrightarrow{h}=(h_{1},h_{2},h_{3})$ such
that $\omega (\overrightarrow{h})\leq 2n(G)+1\leq \frac{3n(4k+8)}{6k+8}$.
Therefore $\omega (f_{j})\leq \frac{n(4k+8)}{6k+8}$ for some $j\in
\{1,2,3\}, $ and this completes the proof.$\hfill \Box $
\begin{theorem}
\label{Th3}\emph{Let $k\geq 1$ be an integer and let $G$ be a connected
graph of order $n\geq 6k+9\;$and minimum degree at least 2 which does not
contain {neither }any induced $\{C_{5},C_{8},\ldots ,C_{3k+2}\}$-cycles {nor
}any cycle of length $\equiv 0\pmod 3$, and every two distinct cycles of
length $\equiv 2\pmod 3$ (if any) have at least a common vertex. If $G$ has
a cycle $C$ with length $\equiv 2\pmod 3$, then $\gamma _{R}(G)\leq \frac
(4k+8)n}{6k+11}$.}
\end{theorem}
\noindent \textbf{Proof.} Let $C=z_{1}z_{2}\ldots z_{p}z_{1}$ be a cycle of
length $\equiv 2\pmod 3$ in $G$ {chosen first not induced, if it exists,
otherwise it is of course induced. }If $V(G)=V(C)$, then we have $\gamma
_{R}(G)\leq \frac{2n+2}{3}<\frac{(4k+8)n}{6k+11}$. Hence, we can assume that
$V(C)\subsetneqq V(G)$.
First assume there is either a cycle $C^{\prime }=x_{1}x_{2}\ldots
x_{m}x_{1} $ such that $x_{1}$ is adjacent to a vertex {of }$C$, say $z_{1}
, and $N_{G}(x_{m})\subseteq V(C)\cup V(C^{\prime })$, or a tailed $m$-cycle
$C_{m,\ell }$ in $G$ such that $y_{\ell }$ is adjacent to a vertex {of} $C$,
say $z_{1}$, and $N_{G}(x_{m})\subseteq V(C)\cup V(C_{m,\ell })$. By
assumption $m\equiv 1\pmod 3$. Let $G_{2}^{0}=C+C^{\prime }+x_{1}z_{1}$ or
G_{2}^{0}=C+C_{m.\ell }+y_{\ell }z_{1}$ {(depending on which situation
occurs, the first or the second one), }and let $G_{1}^{0}=G-G_{2}^{0}$. By
Lemma \ref{MainLem}, $G_{2}^{0}$ has a 3-tuple of RDFs $\overrightarrow{f
=(f_{1},f_{2},f_{3})$ such that $\omega (\overrightarrow{f})\leq
2n(G_{2}^{0})+1$ and all vertices of $G_{2}^{0}$ but $x_{m}$ are
\overrightarrow{f}$-strong. Considering our assumption and Lemma \ref{12},
one of the following situations holds. }
\begin{description}
\item[(a)] $G_{1}^{0}$ has a path $P=v_{1},\ldots ,v_{t}$ such that $v_{1}$
and $v_{t}$ are adjacent to some vertices in $V(G_{2}^{0})$, say $u,v$
(possibly $u=v$) and $N_{G}(v_{1})\cup N_{G}(v_{t})\subseteq
V(G_{2}^{0})\cup V(P)$. {We note that }$x_{m}\notin \{u,v\}$, {since }
C^{\prime }$ or $C_{m,\ell }$ {has been chosen so that }$x_{m}$ satisfies
N_{G}(x_{m})\subseteq V(C)\cup V(C_{m,\ell })$. {Hence }$u$ and $v$ are
\overrightarrow{f}$-strong. \newline
Let $G_{2}^{1}$ be {the graph }obtained from $G_{2}^{0}$ and the path $P$ by
adding the edges $v_{1}u,v_{t}v$. By Lemma \ref{ear1}, $\overrightarrow{f}$
can be extended to a 3-triple $\overrightarrow{g}$ such that $\omega
\overrightarrow{g})\leq 2n(G_{2}^{1})+1$ and all vertices of
V(P)-\{v_{1},v_{t}\}$ are $\overrightarrow{g}$-strong.
\item[(b)] $G_{1}^{0}$ has a cycle $C^{\prime \prime }=v_{1},\ldots
,v_{t}v_{1}$ such that $v_{1}$ is adjacent to a vertex in $G_{2}^{0}$, say
u $, and $N_{G}(v_{t})\subseteq V(G_{2}^{0})\cup V(C^{\prime \prime })$. {A
same argument as in item (a) shows that }$u\neq x_{m}$, {and thus }$u$ is
\overrightarrow{f}$-strong. \newline
By assumption, we have $t\equiv 1\pmod 3$. Let $G_{2}^{1}$ be {\ the graph
obtained from $G_{2}^{0}$ and the cycle $C^{\prime \prime }$ by adding the
edge $v_{1}u$. By Lemma \ref{tailedcycle-3p+1}, $\overrightarrow{f}$ can be
extended to a 3-triple $\overrightarrow{g}$ such that $\omega
\overrightarrow{g})\leq 2n(G_{2}^{1})+1$ and each vertex $V(C^{\prime \prime
})-\{v_{t}\}$ is $\overrightarrow{g}$-strong.
\item[(c)] $G_{1}^{0}$ contains a tailed $m^{\prime }$-cycle $C_{m^{\prime
},\ell ^{\prime }}$, such that $y_{\ell ^{\prime }}$ is adjacent to some
vertex in $G_{2}^{0}$, say $u$, and $N_{G}(x_{2})\cup N_{G}(x_{m})\subseteq
V(G_{2}^{0})\cup V(C_{m^{\prime },\ell ^{\prime }})$. {Note that }$u\neq
x_{m}$ and $u$ is $\overrightarrow{f}$-strong. \newline
As above $m^{\prime }\equiv 1\pmod 3$. Let $G_{2}^{1}$ be {the graph
obtained from $G_{2}^{0}$ and the tailed cycle $C_{m^{\prime },\ell ^{\prime
}}$ by adding the edge $uy_{\ell ^{\prime }}$. By Lemma \re
{tailedcycle-3p+1}, $\overrightarrow{f}$ can be extended to a 3-triple
\overrightarrow{g}$ such that $\omega (\overrightarrow{g})\leq
2n(G_{2}^{1})+1$ and all vertices of $V(C_{m^{\prime },\ell ^{\prime
}})-\{x_{m^{\prime }}\}$ are $\overrightarrow{g}$-strong.
\end{description}
Let $G_{1}^{1}=G-G_{2}^{1}$. By repeating {the }above process, we obtain a
3-tuple of RDFs $\overrightarrow{g^{\prime }}=(g_{1}^{\prime },g_{2}^{\prime
},g_{3}^{\prime })$ such that $\omega (\overrightarrow{g^{\prime }})\leq
2n(G)+1\leq \frac{3n(4k+8)}{6k+8}$. It follows that $\omega (g_{j}^{\prime
})\leq \frac{n(4k+8)}{6k+8}$ for some $j\in \{1,2,3\}$ as desired.
{Next we can }assume that there is {neither a }cycle $C^{\prime
}=(x_{1}x_{2}\ldots x_{m}x_{1})$ such that $x_{1}$ is adjacent to a vertex
in $C$ and $N_{G}(x_{m})\subseteq V(C)\cup V(C^{\prime })$, {nor a }tailed
m $-cycle $C_{m,\ell }$ with $m\equiv 1\pmod 3$ in $G$ such that $y_{\ell }$
is adjacent to a vertex in $C$ and $N_{G}(x_{m})\subseteq V(C)\cup
V(C_{m,\ell })$. Let $H_{2}^{0}=C$ and $H_{1}^{0}=G-H_{2}^{0}$. It follows
from Lemma \ref{12} and the assumptions that $H_{1}^{0}$ has a path
P=v_{1},\ldots ,v_{t}$ such that $v_{1},v_{t}$ are adjacent to some vertices
in $V(H_{2}^{0})$, say $z_{1},z_{j}$ (possibly $j=1$) and $N_{G}(v_{1})\cup
N_{G}(v_{t})\subseteq V(H_{2}^{0})\cup V(P)$. We consider the following
cases.
\smallskip \noindent \textbf{Case 1.} $j=1$.\newline
Let $G_{2}^{1}$ be {the graph }obtained from $H_{2}^{0}$ and the path $P$ by
adding the edges $v_{1}z_{1},v_{t}z_{1}$ and let $G_{1}^{1}=G-G_{2}^{1}$. By
Lemma \ref{MainLem}{-(1)}, $G_{2}^{1}$ has a triple $\overrightarrow{g}$ of
RDFs such that $\omega (\overrightarrow{f})\leq 2n(G_{2}^{1})+1$ and all
vertices of $G_{2}^{1}$ but $v_{1},v_{t}$ are $\overrightarrow{g}$-strong.
If $V(G)=V(G_{2}^{1})$ {(and hence }$G_{1}^{1}$ is empty), then the result
follows. {Hence, a}ssume that $V(G)\neq V(G_{2}^{1})$. By the assumptions
and Lemma \ref{12}, we deduce that $G_{1}^{1}$ has a path $P^{\prime
}=v_{1}^{\prime },\ldots ,v_{t^{\prime }}^{\prime }$ such that
v_{1}^{\prime },v_{t^{\prime }}^{\prime }$ are adjacent to some vertices in
V(G_{2}^{1})$, say $u,v$ (possibly $u=v$) and $N_{G}(v_{1}^{\prime })\cup
N_{G}(v_{t^{\prime }}^{\prime })\subseteq V(G_{2}^{1})\cup V(P^{\prime })$.
Let $G_{2}^{2}$ be obtained from $G_{2}^{1}$ and the path $P^{\prime }$ by
adding the edges $v_{1}^{\prime }u,v_{t^{\prime }}^{\prime }v$ and let
G_{1}^{2}=G-G_{2}^{2}$. {Note that }$v_{1},v_{t}\notin \{u,v\}$ {and thus }
u,v$ are $\overrightarrow{g}$-strong. By Lemma \ref{ear1}, $\overrightarrow{
}$ can be extended to a 3-triple $\overrightarrow{g^{\prime }}$ such that
\omega (\overrightarrow{f_{2}})\leq 2n(G_{2}^{2})+1$ and all vertices of
V(P^{\prime })-\{v_{1}^{\prime },v_{t^{\prime }}^{\prime }\}$ are
\overrightarrow{g^{\prime }}$-strong. By repeating {the }above process, we
obtain a 3-tuple of RDFs $\overrightarrow{g^{\ast }}=(g_{1}^{\ast
},g_{2}^{\ast },g_{3}^{\ast })$ such that $\omega (\overrightarrow{g^{\ast }
)\leq 2n(G)+1\leq \frac{3n(4k+8)}{6k+8}$. It follows that $\omega
(g_{r}^{\ast })\leq \frac{n(4k+8)}{6k+8}$ for some $r\in \{1,2,3\}$ as
desired.
\smallskip \noindent \textbf{Case 2.} $j\neq 1$.\newline
We distinguish the following {three }subcases.
\smallskip \textbf{Subcase 2.1.} $t\equiv 1\pmod 3$.\newline
Since $G$ has no cycle of length $\equiv 0\pmod 3$, we have $j\not\equiv
\pmod 3$. Let $G_{2}^{1}$ be {the graph }obtained from $H_{2}^{0} $ and the
path $P$ by adding the edges $v_{1}z_{1},v_{t}z_{j}$ and let
G_{1}^{1}=G-G_{2}^{1}$. By Lemma \ref{ear2}, $G_{2}^{1}$ has a triple
\overrightarrow{f}$ of RDFs such that $\omega (\overrightarrow{f})\leq
2n(G_{2}^{1})+1$ and all vertices of $G_{2}^{1}$ but $v_{1},v_{t}$ are
\overrightarrow{f_{1}}$-strong. As in Case 1, we can obtain a 3-tuple of
RDFs $\overrightarrow{g^{\ast }}=(g_{1}^{\ast },g_{2}^{\ast },g_{3}^{\ast })$
such that $\omega (\overrightarrow{g^{\ast }})\leq 2n(G)+1\leq \frac{3n(4k+8
}{6k+8}$ {yielding the desired result}.
\smallskip \textbf{Subcase 2.2.} $t\equiv 2\pmod 3$.\newline
{Observe that if }$j\equiv 0\pmod 3$, {then }
z_{1}v_{1}...v_{t}z_{j}z_{j+1}...z_{p}z_{1}$ would be a cycle of length
\equiv 0\pmod 3$, a contradiction, {and if $j\equiv 1\pmod 3$,
then }$z_{1}v_{1}...v_{t}z_{j}z_{j-1}...z_{2}z_{1}$ would be a cycle of
length $\equiv 0\pmod 3$, a contradiction again. Hence $j\equiv 2\pmod 3$}.
Now, as }in Subcase 2.1, we can get the result.
Considering Subcases 2.1 and 2.2, we may assume all ears of $C$ in
G_{1}^{0} $ have length $\equiv 0\pmod 3$.
\smallskip \textbf{Subcase 2.3.} $t\equiv 0\pmod 3$.\newline
Considering the cycles generated by $C+P+v_{1}z_{1}+v_{t}z_{j}$ and that
fact that $G$ has no cycle of length $\equiv 0\pmod 3$, we {\ deduce that }
j\equiv 2\pmod 3$. Let $C_{1}=(z_{1}z_{2}\ldots z_{p}z_{1})$,
C_{2}=(z_{1}v_{1}v_{2}\ldots v_{t}z_{j}z_{j-1}\ldots z_{2}z_{1})$ and
C_{3}=(z_{1}v_{1}v_{2}\ldots v_{t}z_{j}z_{j+1}\ldots z_{p}z_{1})$. Clearly
the cycles $C_{1},C_{2}$, $C_{3}$ are {all }of length $\equiv 2\pmod 3$.
{Assume first that }$C_{1}$ is not an induced cycle in $G$. Then by Lemma
\ref{induced} and considering the ear we have $n(C_{1}\cup C_{2})\geq 6k+11
. Let $G_{2}^{0}=C_{1}\cup C_{2}$ and $G_{1}^{0}=G-G_{2}^{0} $. It is not
hard to see that $G_{2}^{0}$ has a 3-tuple $\overrightarrow{f}$ of RDFs such
that $\omega (\overrightarrow{f})\leq 2n(G_{2}^{0})+2$ and all vertices of
G_{2}^{0}$ but $v_{1},v_{t}$ are $\overrightarrow{f}$-strong. If
V(G)=V(G_{2}^{0})$, then the result follows. {Hence }assume that $V(G)\neq
V(G_{2}^{0})$. By the assumptions and Lemma \ref{12}, we deduce that
G_{1}^{0}$ has a path $P^{\prime }=v_{1}^{\prime },\ldots ,v_{t^{\prime
}}^{\prime }$ such that $v_{1}^{\prime },v_{t^{\prime }}^{\prime }$ are
adjacent to some vertices in $V(G_{2}^{0})$, say $u,v$ (possibly $u=v$) and
N_{G}(v_{1}^{\prime })\cup N_{G}(v_{t^{\prime }}^{\prime })\subseteq
V(G_{2}^{0})\cup V(P^{\prime })$. {Note that }$u,v\notin \{v_{1},v_{t}\}$
since }$N_{G}(v_{1})\cup N_{G}(v_{t})\subseteq V(H_{2}^{0})\cup V(P)$. Thus
{u,v}$ are $\overrightarrow{f}$-strong. {Now, let }$G_{2}^{1}$ be {the graph
}obtained from $G_{2}^{1}$ and the path $P^{\prime }$ by adding the edges
v_{1}^{\prime }u,v_{t^{\prime }}^{\prime }v$ and let $G_{1}^{1}=G-G_{2}^{1}
. By Lemma \ref{ear1}, $\overrightarrow{f} $ can be extended to a 3-triple
\overrightarrow{g}$ such that $\omega (\overrightarrow{g})\leq
2n(G_{2}^{2})+1$ and all vertices\ of $P^{\prime }$ but $v_{1}^{\prime
},v_{t^{\prime }}^{\prime }$ are $\overrightarrow{f_{2}}$-strong. By
repeating above process, we obtain a 3-tuple of RDFs $\overrightarrow
g^{\ast }}=(g_{1}^{\ast },g_{2}^{\ast },g_{3}^{\ast })$ such that $\omega
\overrightarrow{g^{\ast }})\leq {2n(G)+2}\leq \frac{3n(4k+8)}
6k+8}$. It follows that $\omega (\overrightarrow{g_{j}^{\ast }})\leq \frac
n(4k+8)}{6k+8}$ for some $j\in \{1,2,3\}$ as desired.
Assume now that $C_{1}$ is {an induced cycle}. By the choice of $C,$ we may
assume that $G$ has no cycle of length $\equiv 2\pmod 3$ which is not
induced. {Hence the cycle }$C_{2}$ {is also induced. }Let
G_{2}^{0}=C_{1}\cup C_{2}$ and $G_{2}^{1}=G-G_{2}^{0}$. {There are the
following two possibilities. }
\begin{itemize}
\item $V(G)=V(G_{2}^{0})$.\newline
Suppose $n(C_{1})=3t_{1}+2$ and $t=3t_{2}$. Using the fact that $n\geq 6k+9,$
we obtain
\begin{equation*}
\begin{array}{lll}
n & = & n(C_{1})+t \\
& = & 3t_{1}+2+3t_{2} \\
& \geq & 3(2k+3
\end{array
\end{equation*
implying that $t_{1}+t_{2}\geq 2k+3-2/3$. {Since }$t_{1}+t_{2}$ {is integer,
we deduce that }$t_{1}+t_{2}\geq 2k+3,$ and thus $n\geq 6k+11$. {Now, it }is
easy to see that $\gamma _{R}(G)\leq \frac{2n+2}{3}\leq \frac{(4k+8)n}{6k+11}
$.
\item $V(G_{2}^{0})\subsetneqq V(G)$.\newline
Clearly $G_{2}^{0}$ has a triple $\overrightarrow{f^{0}}$ of RDFs such that
\omega (\overrightarrow{f^{0}})\leq n(G_{2}^{0})+2$ and all vertices of
G_{2}^{0}$ but $v_{1},v_{t}$ are $\overrightarrow{f^{0}}$-strong. By the
assumptions and Lemma \ref{12}, we deduce that $G_{1}^{0}$ has a path
P_{1}=v_{1}^{1},\ldots ,v_{q_{1}}^{1}$ such that $v_{1}^{1},v_{q_{1}}^{1}$
are adjacent to some vertices in $V(G_{2}^{0})$, say $u,v$ (possibly $u=v$)
and $N_{G}(v_{1}^{1})\cup N_{G}(v_{q_{1}}^{1})\subseteq V(G_{2}^{1})\cup
V(P_{1})$. {Recall that }$u,v\in \{v_{1},v_{t}\}$ {\ and thus they are }
\overrightarrow{f^{0}}$-strong. {Moreover, }since every cycle of $G$
intersects $C_{1}$, we have $V(C_{1})\cap \{u,v\}\neq \emptyset $. {Hence
vertices }$u,v$ {may belong to }$C_{1}$, $C_{2}$ or $C_{3}$. {Now, }seeing
Case 1 and Subcase 2.1 and 2.2, we may assume that $q_{1}\equiv 0\pmod 3$.
Let $q_{1}=3q_{1}^{\prime }$ and let $G_{2}^{1}$ be {the graph }obtained
from $G_{2}^{0}$ and the path $P_{1}$ by adding the edges
v_{1}^{1}u,v_{q_{1}}^{1}v$ and let $G_{1}^{1}=G-G_{2}^{1}$. By Lemma \re
{ear1}, $\overrightarrow{f^{0}}$ can be extended to a 3-triple
\overrightarrow{f^{1}}$ {of }$G_{2}^{1}$ such that {$\omega (\overrightarrow
f^{1}})\leq 2n(G_{2}^{1})+2$} and all vertices {\ of }$P_{1}$ but
v_{1}^{1},v_{q_{1}}^{1}$ are $\overrightarrow{f^{1}}$-strong. If
V(G)=V(G_{2}^{1})$, then $n=3t_{1}+3t_{2}+3q_{1}^{\prime }+2$. As above we
can see that $n\geq 6k+11,$ {implying that }$\gamma _{R}(G)\leq \frac{2n+2}{
}\leq \frac{(4k+8)n}{6k+11}$. {Hence assume that }$V(G_{2}^{1})\subsetneqq
V(G)$. By the assumptions and Lemma \ref{12}, we deduce that $G_{1}^{1}$ has
a path $P_{2}=v_{1}^{2},\ldots ,v_{q_{2}}^{2}$ such that
v_{1}^{2},v_{q_{2}}^{2}$ are adjacent to some vertices in $V(G_{2}^{1})$,
say $u^{\prime },v^{\prime }$ (possibly $u^{\prime }=v^{\prime }$) and
N_{G}(v_{1}^{2})\cup N_{G}(v_{q_{2}}^{2})\subseteq V(G_{2}^{1})\cup V(P_{2})
. Since every cycle of $G$ intersects $C_{1}$, we have $V(C_{1})\cap
\{u^{\prime },v^{\prime }\}\neq \emptyset $. On the other hand, we note that
$u^{\prime },v^{\prime }$ lies on a cycle of length $\equiv 2\pmod 3$.
Seeing Case 1 and Subcase 2.1 and 2.2, we may assume that $q_{2}\equiv
\pmod 3$. Let $q_{2}=3q_{2}^{\prime }$ and let $G_{2}^{2}$ be {the graph
obtained from $G_{2}^{1}$ and the path $P_{2}$ by adding the edges
v_{1}^{2}u,v_{q_{2}}^{2}v$ and let $G_{1}^{2}=G-G_{2}^{2}$. By Lemma \re
{ear1}, $\overrightarrow{f^{1}}$ can be extended to a 3-triple
\overrightarrow{f^{2}}$ such that $\omega (\overrightarrow{f^{2}})\leq
2n(G_{2}^{2})+2$ and all {of }$P_{2}$ but $v_{1}^{2},v_{q_{2}}^{2}$ are
\overrightarrow{f^{2}}$-strong. If $V(G)=V(G_{2}^{2})$, then
n=3t_{1}+3t_{2}+3q_{1}^{\prime }+3q_{2}^{\prime }+2$. As above we can see
that $n\geq 6k+11,$ {implying that }$\gamma _{R}(G)\leq \frac{2n+2}{3}\leq
\frac{(4k+8)n}{6k+11}$. {\ Hence suppose that }$V(G_{2}^{k})\subsetneqq V(G)
. ~ By repeating {the }above process, we obtain a subgraph $G_{2}^{k}$ with
n(G_{2}^{k})\geq 6k+11$ and having a 3-tuple $\overrightarrow{f^{k}}$ of
RDFs such that $\omega (\overrightarrow{f^{k}})\leq {
2n(G_{k}^{2})+2}$ and all vertices of $G_{2}^{k}$ with a neighbor outside of
$G_{2}^{k}$ are $\overrightarrow{f^{k}}$-strong. If $V(G)=V(G_{2}^{k})$,
then the result follows immediately. {Otherwise, let
G_{1}^{k}=G-V(G_{2}^{k})$}. Now using Lemma \ref{ear1} we can obtain extend
G_{2}^{k}$ to a subgraph $G_{2}^{k+1}$ by adding an ear in $G_{1}^{k}$ and
extend $\overrightarrow{f^{k}}$ to a 3-tuple $\overrightarrow{f^{k+1}}$ of
RDFs such that $\omega (\overrightarrow{f^{k+1}})\leq 2n(G_{k+1}^{2})+2$ and
all vertices of $G_{2}^{k+1}$ with a neighbor outside of $G_{2}^{k+1}$ are
\overrightarrow{f^{k+1}}$-strong. By repeating this process we obtain a
3-tuple $g$ of RDFs of $G$ such that $\omega (\overrightarrow{g})\leq
2n(G)+2 $ and this leads to the result as above. $\hfill \Box $
\end{itemize}
\section{Some more lemmas}
Let $\mathcal{F}_{i}$ be the family of all cycles of length $\equiv i\pmod 3$
with $i\in \{0,1,2\}$. Let $\mathcal{F}_{0,2}$ be the family of all
connected }graphs obtained from a cycle $C$ of $\mathcal{F}_{0}$ and a cycle
$C^{\prime }$ of $\mathcal{F}_{2}$ by {joining a vertex }$x$ {of }$C$ {\and
a vertex }$y$ {of }$C^{\prime }$ {by either an edge }$xy${\ or by a
nontrivial path that we add so that one of the envertices of the path is
attached to }$x$ and the other one to $y$; $\mathcal{F}_{2,2}$ be the family
of all {connected} graphs obtained from two cycles in $\mathcal{F}_{2}$ by
adding an edge between them; and let $\mathcal{F}_{3}$ be the family of all
graphs $G$ obtained from a graph $G^{\prime }$ in $\mathcal{F}_{0,2}$ and a
graph $G^{\prime \prime }$ in $\mathcal{F}_{2,2}$ by adding {either an edge
or a path joining a vertex of }$G^{\prime }$ to a vertex of $G^{\prime
\prime }$ {so that all vertices of the path become of degree two in }$G.$
\bigskip\
Let $\mathcal{B}_{r,s}\;(r+s\geq 2)$ be the family of {connected }graphs
obtained from $r$ tailed cycles $C_{n_{1},\ell _{1}},\ldots ,C_{n_{r},\ell
_{r}}$ and $s$ cycles $C_{m_{1}},\ldots ,C_{m_{s}}$, where $n_{i}\equiv
\pmod 3$ and $m_{j}\equiv 2\pmod 3$ for each $i,j$, by adding a new vertex
z $ ({which we call }special vertex) and {\ joining by edges }$z$ to the
unique leaf }of each graph $C_{n_{i},\ell _{i}}$ and to one vertex of each
cycle $C_{m_{j}}$. {Moreover, each of the }$s$ {cycles will be called a
\textit{near cycle} of }$z.${\ }Let $\mathcal{E}=\cup _{r,s\geq 0;r+s\geq 2
\mathcal{B}_{r,s}$.
\begin{lemma}
\label{Lam3}\emph{Let $G$ be a connected graph {with at least two disjoint
cycles and let }$\mathcal{F}$ be a family of pairwise disjoint cycles of
length $\equiv 0,2\pmod 3$ in $G$ with $|\mathcal{F}|\geq 2$. Then $G$ has
two disjoint subgraphs $G_{1}$ (possibly null) and $G_{2}$ such that
V(G)=V(G_{1})\cup V(G_{2})$, $G_{1}$ has no cycle of $\mathcal{F}$ and each
component of $G_{2}$ is in $\mathcal{F}_{0}\cup \mathcal{F}_{0,2}\cup
\mathcal{F}_{2,2}\cup \mathcal{E}.$}
\end{lemma}
\noindent \textbf{Proof.} The proof is by induction on the number of cycles
in $\mathcal{F}$. First let $|\mathcal{F}|=2$ with $\mathcal{F
=\{C_{1},C_{2}\}$. {Since }$G$ {is connected, let }$P $ be a shortest path
joining a vertex of }$C_{1}$ {to a vertex of $C_{2}$. If both $C_{1},C_{2}$
have length $\equiv 0\pmod 3$, then let $G_{2}=C_{1}\cup C_{2}$ and if one
of the {two }cycles has length $\equiv 2\pmod 3$, then let
G_{2}=C_{1}+C_{2}+P$. Assume that $G_{1}=G-V(G_{1})$. Clearly $G_{1}$ has no
cycle of $\mathcal{F}$ and each component of $G_{2}$ belongs to $\mathcal{F
_{0}\cup \mathcal{F}_{0,2}\cup \mathcal{F}_{2,2}\cup \mathcal{E}$,
establishing the base case.} }
\noindent Next let $|\mathcal{F}|=3$ and $\mathcal{F}=\{C_{1},C_{2},C_{3}\}
. Assume that $C_{i}=x_{1}^{i}x_{2}^{i}\ldots x_{n_{i}}^{i}x_{1}^{i}$ for
i\in \{1,2,3\}$. If {each cycle of }$\mathcal{F}$ has length $\equiv
\pmod 3
$, then let $G_{2}=C_{1}\cup C_{2}\cup C_{3}$ and $G_{1}=G-G_{2}$. {Clearly
the result holds. Hence assume that }one of the {three }cycles has length
\equiv 2\pmod 3$, say $C_{1}$. Let $P$ be a shortest path {joining a vertex
of }$C_{1}$ {to a vertex in }$C_{2}$ or $C_{3}$. Assume, without loss of
generality, that $P$ {joins }$C_{1}$ and $C_{2}$, where
P=(x_{1}^{1}=)z_{0}z_{1}...z_{k}(=x_{1}^{2}).$ If $C_{3}$ has length $\equiv
0\pmod 3,$ {then by setting }$G_{2}=\left( C_{1}\cup C_{2}\cup P\right) \cup
C_{3}$ and $G_{1}=G-G_{2},$ {it is clear that the result holds. Hence we
assume that $C_{3}$ has length $\equiv 2\pmod 3.$ {Now, let }
Q=(x_{1}^{3}=)y_{0}y_{1}...y_{s}$ be a shortest path {joining a vertex of }
C_{3}$ {to a vertex }$y_{s}$ {belonging to }$V(C_{1})\cup V(C_{2})\cup V(P)
. Assume that $y_{s}\in V(P)-\{x_{1}^{1},x_{1}^{2}\}$, say $y_{s}=z_{m}.$ If
$C_{2}$ has length $\equiv 0\pmod 3,$ then {by setting }$G_{2}=\left(
C_{1}\cup C_{3}\cup P^{\prime }\right) \cup C_{2},$ with $P^{\prime
}=(x_{1}^{1}=)z_{0}z_{1}...z_{m},y_{s-1},\ldots,y_{0},$ and $G_{1}=G-G_{2},$
we get the desired result. Hence we assume that $C_{2}$ has length $\equiv
\pmod 3.$ In this case, the result holds by letting $G_{2}=C_{1}\cup
C_{2}\cup C_{3}\cup P\cup Q$ and $G_{1}=G-G_{2}.$ Finally, assume, without
loss of generality, that\ $y_{s}\in C_{2},$ say $y_{s}=x_{j}^{2}$ (possibly
j=1$). Let $G_{2}=C_{1}Px_{1}^{2}\ldots x_{j}^{2}QC_{3}$ and $G_{1}=G-G_{2}
. {Note that }$G_{2}$belongs to {$\mathcal{B}_{0,2}\cup \mathcal{B}_{1,1},$}
{and clearly the desired result holds. } }
\noindent Assume now that $|\mathcal{F}|\geq 4$. If all cycles in $\mathcal{
}$ have length $\equiv 0\pmod 3$, then the subgraphs $G_{2}=\cup _{i=1}^{
\mathcal{F}|}C_{i}$ and $G_{1}=G-G_{2}$ satisfy the conditions {\ and the
result holds. }Hence we assume that one of the cycles in $\mathcal{F} $,
say }$C_{0},${\ }has length $\equiv 2\pmod 3$. Let $\mathcal{F}^{\prime }
\mathcal{F}-\{C_{0}\}$ and let $G^{1}=G-V(C_{0})$. We consider two cases.
\smallskip \noindent \textbf{Case 1.} $G^{1}$ is connected.\newline
Then $\mathcal{F}^{\prime }$ is a family of disjoint cycles {of length }
\equiv 0,2\pmod 3$ in $G^{1}$ {with }\emph{$|\mathcal{F}^{\prime }|\geq 2$}.
By the induction hypothesis, $G^{1}$ has two disjoint subgraphs
G_{1}^{\prime }$ (possibly null) and $G_{2}^{\prime }$ such that
V(G^{1})=V(G_{1}^{\prime })\cup V(G_{2}^{\prime })$, $G_{1}^{\prime }$ has
no cycle of $\mathcal{F}^{\prime }$, where each component of $G_{2}^{\prime
} $ is in $\mathcal{F}_{0}\cup \mathcal{F}_{0,2}\cup \mathcal{F}_{2,2}\cup
\mathcal{E}.$ Let $H_{1},\ldots ,H_{p}$ be the components of $G_{2}^{\prime
} $. {Suppose without loss of generality that $P:=(x_{1}^{0}=)v_{0}v_{1
\ldots v_{t}$ is a shortest path between $V(C_{0})$ and $V(G_{2}^{\prime })$
in $G$ where $v_{t}\in V(G_{2}^{\prime })$.} Without loss of generality,
assume that $v_{t}\in V(H_{1}).$ If $H_{1}\in \mathcal{F}_{0}$, then let
H_{1}^{\prime }=H_{1}+P+C_{0}$ and clearly {the }two subgraphs
G_{2}=H_{1}^{\prime }\cup H_{2}\cup \ldots H_{p}$ and $G_{1}=G-V(G_{2})$
satisfy the conditions {and result follows.
For the next, we can assume that $H_{1}$ contains at least two cycles. We
distinguish the following. }
\smallskip \textbf{Subcase 1.1.} $H_{1}\in \mathcal{F}_{2,2}$.\newline
Suppose $H_{1}$ is obtained from two cycles $C_{1}=x_{1}^{1}\ldots
x_{m_{1}}^{1}x_{1}^{1}$ and $C_{2}=x_{1}^{2}\ldots x_{m_{2}}^{2}x_{1}^{2}$
by adding a path $Q=(x_{1}^{1}=)z_{0}z_{1}\ldots z_{t}(=x_{1}^{2})$. {We
further assume, without loss of generality, that }$v_{t}=x_{j}^{2}\in
V(C_{2})$. {Let }$H_{1}^{\prime }$ {be the graph obtained from }$C_{0}\cup
C_{1}$ {to which we add the path }$Px_{j-1}^{2}\ldots x_{2}^{2}Q$, {in other
words, }$H_{1}^{\prime }$ {is obtained from }$C_{0}+P+H_{1} $ {by removing
vertices }$x_{j+1}^{2},...,x_{m_{2}}^{2}.$ {Note that }$H_{1}^{\prime }$
belongs to either $\mathcal{F}_{0,2}$ or {$\mathcal{B}_{0,2}\cup\mathcal{B
_{1,1} .$ } Now let $G_{2}^{\prime \prime }=H_{1}^{\prime }\cup H_{2}\cup
\ldots H_{p}$ and $G_{1}^{\prime \prime }=G-G_{2}^{\prime \prime }. $ Then
the subgraphs $G_{2}^{\prime \prime }$ and $G_{1}^{\prime \prime }$ satisfy
the conditions {and result follows}. }
\smallskip \textbf{Subcase 1.2.} $H_{1}\in \mathcal{F}_{0,2}$
\newline
Using an argument similar to that described in the case $|\mathcal{F}|=3$,
we can obtain two subgraphs $G_{2}^{\prime \prime }$ and $G_{1}^{\prime
\prime }$ satisfying the conditions {and yielding the desired result}.
\smallskip \textbf{Subcase 1.3.} $H_{1}\in \mathcal{B}_{r,s}$ where $r+s\geq
2$.\newline
Let $z^{\ast }$ {be the special vertex of }$H_{1}.$ {If }$v_{t}=z^{\ast }$,
then $H_{1}^{\prime }=H_{1}+P+C_{0}$ is a subgraph belonging to $\mathcal{E}$
and {thus }the subgraphs $G_{2}=H_{1}^{\prime }\cup H_{2}\cup \ldots H_{p}$
and $G_{1}=G-V(G_{2})$ satisfy the conditions and the result follows. Hence
we assume that $v_{t}\neq z^{\ast }.$ First let $r+s=2.$ Then $H_{1}$ is
obtained from two cycles $C_{1}=x_{1}^{1}\ldots x_{m_{1}}^{1}x_{1}^{1}$ and
C_{2}=x_{1}^{2}\ldots x_{m_{2}}^{2}x_{1}^{2}$ by adding a path
Q=(x_{1}^{1}=)z_{0}z_{1}\ldots z_{t}(=x_{1}^{2}),$where $t\geq 2.$ If
v_{t}\in \{z_{1},\ldots ,z_{t-1}\}$, then let $H_{1}^{\prime
}=H_{1}+C_{0}+P. $ {Clearly, }$H_{1}^{\prime }\in \mathcal{B}_{r,s}$ where
r+s=3$, and thus the subgraphs $G_{2}=H_{1}^{\prime }\cup H_{2}\cup \ldots
H_{p}$ and $G_{1}=G-V(G_{2})$ satisfy the conditions {and the result yields.
Now, suppose, without loss of generality, that }$v_{t}=x_{j}^{2}\in
V(C_{2}). $ {Let }$H_{1}$ be obtained $C_{0}\cup C_{1}$ {to which we add the
path }$Px_{j-1}^{2}\ldots x_{2}^{2}Q,$ {and set }$G_{2}^{\prime \prime
}=H_{1}^{\prime }\cup H_{2}\cup \ldots H_{p}$ and $G_{1}^{\prime \prime
}=G-G_{2}^{\prime \prime }.$ {Clearly, }$G_{1}^{\prime \prime }$ and
G_{2}^{\prime \prime }$ {satisfy the condition and the desired result
follows.}
{Now let } $r+s\geq 3.$ {Assume that }$v_{t}$ {belongs to one of the }${s+r}$
{cycles of }$H_{1},$ {say} $C^{\prime }.$ {\ Let }$H_{1}^{\prime }=C^{\prime
}+C_{0}+P$ and $H_{1}^{\prime \prime }$ be the graph obtained from $H_{1}$
by deleting {the }vertices of $V(C^{\prime })$ and {the path \ (if any)
joining }$z^{\ast }$ to $V(C^{\prime })$ in $H_{1}$. {\ Note that }
H_{1}^{\prime }$ belongs to either {$\mathcal{F}_{2,2}$} or {
\mathcal{B}_{0,2}\cup \mathcal{B}_{1,1}.$} {Now the subgraphs }
G_{2}=H_{1}^{\prime }\cup H_{1}^{\prime \prime }\cup H_{2}\cup \ldots H_{p}$
and $G_{1}=G-V(G_{2})$ satisfy the conditions and {the desired result holds.
Assume that }$v_{t}$ belongs to a path on a tailed cycle $C_{m,\ell }$ of
H_{1},$ and let $P^{\prime }$ be the subpath between $v_{t}$ and the cycle
C^{\prime }$ of $C_{m,\ell }.$ {Let }$H_{1}^{\prime }=C^{\prime
}+C_{0}+P+P^{\prime }$ and $H_{1}^{\prime \prime }$ be the graph obtained
from $H_{1}$ by deleting {the }vertices of $V(C_{m,\ell }).$ {Note that }
H_{1}^{\prime }$ belongs to either {$\mathcal{F}_{2,2}$} or
\mathcal{B}_{0,2}.$ {Now the subgraphs }$G_{2}=H_{1}^{\prime }\cup
H_{1}^{\prime \prime }\cup H_{2}\cup \ldots H_{p}$ and $G_{1}=G-V(G_{2})$
satisfy the conditions and {the desired result holds.}
\smallskip \noindent \textbf{Case 2.} $G^{1}$ is disconnected.\newline
Let $M_{1},\ldots ,M_{t}$ be the components of $G^{1}$. Assume first that a
component $M_{i}$ contains all cycles of $\mathcal{F}-\{C_{0}\},$ say $M_{1}
. Let $K$ be the subgraph of $G$ induced by $V(C_{0})\cup V(M_{1})$. Clearly
$K$ is connected. Using an argument similar to that described in Case 1 on
K-C_{0}$, we get the result.
Henceforth, we may assume that no $M_{i}$ contains all cycles of $\mathcal{F
-\{C_{0}\}$ for each $i$. Now, assume that a component $M_{i}$ contains at
least two cycles of $\mathcal{F}$, say $M_{1}.$ Let $G^{2}=G-V(M_{1})$.
Clearly $G^{2}$ is connected. Let $\mathcal{F}_{1}=\{C\mid C\in \mathcal{F}\
\mathrm{and}\;V(C)\subseteq V(M_{1})\}$ and $\mathcal{F}_{2}=\mathcal{F}
\mathcal{F}_{1}$. By the induction hypothesis, $M_{1}$ has two subgraphs
K^{1},K^{2}$ such that $K^{1}$ does not contain any cycle of $\mathcal{F
_{1} $ and each component of $K^{2}$ belongs to $\mathcal{F}_{0}\cup
\mathcal{F}_{0,2}\cup \mathcal{F}_{2,2}\cup \mathcal{E}.$ Moreover, $G^{2}$
has two subgraphs $K_{1}^{\prime }$ and $K_{2}^{\prime }$ such that
K_{1}^{\prime }$ does not contain any cycle of $\mathcal{F}_{2}$ and each
component of $K_{2}^{\prime } $ belongs to $\mathcal{F}_{0}\cup \mathcal{F
_{0,2}\cup \mathcal{F}_{2,2}\cup \mathcal{E}.$\textbf{\ }Now the two
subgraphs $G_{1}=K^{1}\cup K_{1}^{\prime }$ and $G_{2}=K^{2}\cup
K_{2}^{\prime }$ satisfies the conditions yielding the desired result.
{From now on, we can assume that }each $M_{i}$ contains at most one cycle of
$\mathcal{F}$. {Suppose that only the }$s$ first $M_{i}$ contains exactly
one cycle $C_{i}$ of $\mathcal{F}.$ Let {$C_{i}=x_{1}^{i}x_{2}^{i}\ldots
x_{n_{i}}^{i}x_{1}^{i}$ for $0\leq i\leq s.$ In addition, let
P_{i}:=(x_{1}^{i}=)w_{0}^{i}\ldots w_{\ell _{i}}^{i}$ be a shortest
nontrivial path (possibly of order two) between $V(C_{i})$ and $V(C_{0})$ in
$G$ for each $1\leq i\leq s,$ where $w_{\ell _{i}}^{i}\in V(C_{0})$. }If all
cycles $C_{1},\ldots ,C_{s}$ have length $\equiv 0\pmod 3$, then the
subgraphs $G_{2}=(C_{1}+P_{1}+C_{0})\cup C_{2}\cup \ldots \cup C_{s}$ and
G_{1}=G-G_{2}$ satisfy the conditions {and the result follows}. {Hence, we
assume that some }cycle $C_{i}\;(i\geq 1)$ has length $\equiv 2\pmod 3$.
Note that the }paths $P_{i}$'s {minus their end-vertices belonging to }
V(C_{0})$ are disjoint. {If some }$C_{i}$ has length $\equiv 0\pmod 3,$ say
C_{1}$, {then let }$L=C_{0}\cup (\cup _{i=2}^{s}C_{i})\cup (\cup
_{i=2}^{s}P_{i})$. By the induction hypothesis, $L$ has two subgraphs
L_{1},L_{2}$ such that $L_{1}$ has no cycle of $\mathcal{F}-\{C_{1}\}$ and
each component of $L_{2}$ belongs to $\mathcal{F}_{0}\cup \mathcal{F
_{0,2}\cup \mathcal{F}_{2,2}\cup \mathcal{E}$. Now $G_{2}=L_{2}\cup C_{1}$
and $G_{1}=G-G_{2}$ satisfy the conditions and the result holds. Hence we
can assume that all cycles $C_{1},\ldots ,C_{s}$ have length $\equiv
\pmod 3
$. Let {$L=C_{0}\cup (\cup _{i=1}^{s}C_{i})\cup (\cup _{i=1}^{s}P_{i})$}.
Let $x_{i_{1}}^{0},\ldots ,x_{i_{t}}^{0}$ be the vertices of $C_{0}$ with
degree at least three and assume, without loss of generality, that
i_{1}<i_{2}<\ldots <i_{t}$.} Consider the following situations.
\smallskip \textbf{Subcase 2.1.} $t=2$.\newline
{If $\deg (x_{i_{1}}^{0}),\deg (x_{i_{2}}^{0})\geq 4$, then let $G_{2}$ be
the graph obtained from $L$ by deleting all vertices of $V(C_{0})
\{x_{i_{1}}^{0},x_{i_{2}}^{0}\}$. Otherwise, let $G_{2}$ be the graph
obtained from $L$ by deleting either the edge }$x_{i_{1}}^{0}x_{i_{2}}^{0}$
(if any) or all the vertices $x_{i_{1}+1},\ldots ,x_{i_{2}-1}$. Then the
subgraphs $G_{2}$ and $G_{1}=G-G_{2}$ satisfies the conditions and the
result holds. }
\smallskip \textbf{Subcase 2.2.} $t=3$.\newline
{If $\deg (x_{i_{1}}^{0}),\deg (x_{i_{2}}^{0}),\deg (x_{i_{3}}^{0})\geq 4$,
then let $G_{2}$ be the graph obtained from $L$ by deleting all vertices of
V(C_{0})-\{x_{i_{1}}^{0},x_{i_{2}}^{0},x_{i_{3}}^{0}\}$. If $\deg
(x_{i_{1}}^{0})=\deg (x_{i_{2}}^{0})=\deg (x_{i_{3}}^{0})=3$, then let
G_{2} $ be the graph obtained from $L$ by deleting either the edge }
x_{i_{2}}^{0}x_{i_{3}}^{0}$ (if any) or all the vertices {of
\{x_{i_{2}+1}^{0},x_{i_{2}+2}^{0},\ldots ,x_{i_{3}-1}^{0}\}$. If, without
loss of generality, $\deg (x_{i_{1}}^{0})=3$ and $\deg (x_{i_{3}}^{0})\geq 4
. Let $G_{2}$ be the graph obtained from $L$ by deleting either the edge
x_{i_{2}}^{0}x_{i_{3}}^{0}$ (if any) or {all vertices between of }
x_{i_{2}}^{0}$ and $x_{i_{3}}^{0}$ {as well all vertices between
x_{i_{3}}^{0}$ and $x_{i_{1}}^{0} $ {starting from }$x_{i_{3}+1}^{0}.$ {In
either situation, the }subgraphs $G_{2}$ and $G_{1}=G-G_{2}$ satisfies the
conditions {and the result follows}. }}
\smallskip \textbf{Subcase 2.3.} $t\geq 4$.\newline
{If $\deg (x_{i_{1}}^{0}),\deg (x_{i_{2}}^{0}),\deg (x_{i_{3}}^{0})\geq 4$,
then let $G_{2}$ be the graph obtained from $L$ by deleting all vertices of
V(C_{0})-\{x_{i_{1}}^{0},x_{i_{2}}^{0},\ldots ,x_{i_{t}}^{0}\}$. If $\deg
(x_{i_{1}}^{0})=\deg (x_{i_{2}}^{0})=\ldots =\deg (x_{i_{t}}^{0})=3$, then
let $G_{2}$ be the graph obtained from $L$ by deleting all vertices of
\bigcup_{j=1}^{\lfloor t/2\rfloor
}\{x_{i_{2j}+1}^{0},x_{i_{2j}+2}^{0},\ldots ,x_{i_{2j+1}-1}^{0}\}$. Assume
without loss of generality that $\deg (x_{i_{1}}^{0})=3$ and $\deg
(x_{i_{t}}^{0})\geq 4$. Let $L^{1}$ be the component of $L
\{x_{i_{1}-1}^{0}x_{i_{1}}^{0},x_{i_{2}}^{0}x_{i_{2}+1}^{0}\}$ containing
x_{i_{1}}^{0}$, and let $L^{2}$ be the component of $L
\{x_{i_{3}-1}^{0}x_{i_{3}}^{0}\}$ containing $x_{i_{3}}^{0}$ if $\deg
(x_{i_{3}}^{0})\geq 4$, and be the component of $L
\{x_{i_{3}}^{0}x_{i_{3}-1}^{0},x_{i_{4}}^{0}x_{i_{4}+1}^{0}\}$ containing
x_{i_{3}}^{0}$ if $\deg (x_{i_{3}}^{0})=3$. Repeating this process we obtain
a sequence $L^{1},\ldots ,L^{p}$ of subgraphs $L$ which contains all cycles
of $L$ but $C_{0}$. Now the subgraphs $G_{2}=\cup _{i=1}^{p}L^{i}$ and
G_{1}=G-G_{2}$ satisfies the conditions and the result follows.}$\hfill \Box
$
\begin{lemma}
\label{mainn}Let $G$ be a connected graph with at least two disjoint cycles
of length $\equiv 0,2\pmod 3$, and let $\mathcal{F}$ be the family of all
cycles of $G$ with length $\equiv 0\;\mathrm{or}\;2\pmod 3$. Then there
exists a maximal subfamily $\mathcal{T}$ of pairwise disjoint cycles of
\mathcal{F}$ with $|\mathcal{T}|\geq 2$ and two disjoint subgraphs $G_{1}$
(possibly null) and $G_{2}$ of $G$ such that $V(G)=V(G_{1})\cup V(G_{2})$,
G_{1}$ has no cycle of $\mathcal{F}$ and each component of $G_{2}$ belongs
to $\mathcal{F}_{0}\cup \mathcal{F}_{0,2}\cup \mathcal{F}_{2,2}\cup \mathcal
E}$.
\end{lemma}
\noindent \textbf{Proof.} {By Lemma \ref{Lam3}, for any maximal subfamily }
\mathcal{T}$ of pairwise disjoint cycles of $\mathcal{F}$ with $|\mathcal{T
|\geq 2$, $G$ has two disjoint subgraphs $G_{1}^{\mathcal{T}}$ {and }$G_{2}^
\mathcal{T}}$ such that $V(G)=V(G_{1}^{\mathcal{T}})\cup V(G_{2}^{\mathcal{T
})$, $G_{1}^{\mathcal{T}}$ has no cycle of $\mathcal{T}$ {and }each
component of $G_{2}^{\mathcal{T}}$ is in $\mathcal{F}_{0}\cup \mathcal{F
_{0,2}\cup \mathcal{F}_{2,2}\cup \mathcal{E}. $ {Now, let }$c_{\mathcal{T}}$
{denote the }number of cycles of $G_{2}^{\mathcal{T}}$, {and let }$s_
\mathcal{T}}$ be the sum of the lengths of paths between two cycles in the
components of $G_{2}^{\mathcal{T}}$ {that belong to }$\mathcal{F}_{0,2}\cup
\mathcal{F}_{2,2}\cup (\cup _{r,s\geq 0;r+s=2}B_{r,s}).$ {\ Moreover, let }
\begin{equation*}
c_{\mathcal{F}}=\max \{c_{\mathcal{T}}\mid \mathcal{T}\;\mathrm
is\;a\;maximal\;subfamily\;of\;pairwise\;disjoint\;cycles\;of}\;\mathcal{F}\
\mathrm{with}\;|\mathcal{T}|\geq 2\}.
\end{equation*
Choose a triple $(\mathcal{T},G_{1}^{\mathcal{T}},G_{2}^{\mathcal{T}})$ such
that: (i) $c_{\mathcal{F}}=c_{\mathcal{T}}$; (ii) subject to (i): $s_
\mathcal{T}}$ is maximized. {Notice that }$G_{2}^{\mathcal{T}}$ {may not
contain all cycles of }$\mathcal{T}$. We claim that {the }two disjoint
subgraphs $G_{1}^{\mathcal{T}}$ and $G_{2}^{\mathcal{T}}$ {chosen in this
way yield the desired result. }
{It is clear that it suffices to show }that $G_{1}^{\mathcal{T}}$ has no
cycle of $\mathcal{F}$. {Hence, suppose }to the contrary that $G_{1}^
\mathcal{T}}$ contains at least one cycle of $\mathcal{F}$. Let $G_{2}^{1}$\
be obtained from $G_{2}^{\mathcal{T}}$ by adding { a maximum set
of pairwise of cycles of $\mathcal{F}$ } with length $\equiv 0\pmod 3$ {\
belonging to }$G_{1}^{\mathcal{T}}$ and let $G_{1}^{1}=G-G_{2}^{1}$. Note
that if $G_{1}^{\mathcal{T}}$ {contains no }cycle of $\mathcal{F} $ with
length $\equiv 0\pmod 3$, then $G_{2}^{1}=G_{2}^{\mathcal{T}}$. {Now, l}et
\mathcal{T}_{1}^{1}$ be the family of all cycles of $G_{2}^{1}$ {that belong
}to $\mathcal{F}$ and {et }$\mathcal{T}^{1}$ be a maximal subfamily of
\mathcal{F}$ such that $\mathcal{T}_{1}^{1}\subseteq \mathcal{T}^{1}$. If
G_{1}^{1}$ does not contain any cycle of $\mathcal{F}$, then the family
\mathcal{T}^{1}$ and the subgraphs $G_{1}^{1}$ and $G_{2}^{1}$ satisfy the
conditions {which }leads to a contradiction because of $c_{\mathcal{T
^{1}}>c_{\mathcal{T}}$. {Hence we }assume that $G_{1}^{1}$ contains at least
one cycle of $\mathcal{F}$.
Let first $H^{1},\ldots H^{r}$ be the components of $G_{1}^{1}$ which
contains at least two disjoint cycles of $\mathcal{F}$ (if any), and let
\mathcal{F}^{i}$ be a maximal subfamily of pairwise disjoint cycles of
\mathcal{F}$ {that are }in $H^{i}$ with $|\mathcal{F}^{i}|\geq 2$, for each
1\leq i\leq r$. By Lemma \ref{Lam3}, $H^{i}$ has two subgraphs
H_{1}^{i},H_{2}^{i}$ such that $H_{1}^{i}$ has no cycles of $\mathcal{F}^{i}$
and each component of $H_{2}^{i}$ is in $\mathcal{F}_{0}\cup \mathcal{F
_{0,2}\cup \mathcal{F}_{2,2}\cup \mathcal{E}$. Let $G_{2}^{2}=G_{2}^{1}\cup
(\cup _{i=1}^{r}H_{2}^{i})$, if $r\geq 1$ and $G_{1}^{2}=G-G_{2}^{2}$. Let
\mathcal{T}_{2}^{1}$ be the family of all cycles of $G_{2}^{2}$ {\ that
belong }to $\mathcal{F}$ and let $\mathcal{T}^{2}$ be a maximal subfamily of
$\mathcal{F}$ such that $\mathcal{T}_{1}^{2}\subseteq \mathcal{T}^{2}$. If
G_{1}^{2}$ does not contain any cycle of $\mathcal{F}$, then the family
\mathcal{T}^{2}$ and the subgraphs $G_{1}^{2}$ and $G_{2}^{2}$ satisfy the
conditions {which }leads to a contradiction because $c_{\mathcal{T}^{2}}>c_
\mathcal{T}}$. {Hence, we }assume that $G_{1}^{2}$ contains at least one
cycle of $\mathcal{F}$. If $G_{1}^{2}$ has a component with at least two
disjoint cycles of $\mathcal{F}$, then we proceed as above.
Henceforth, we {can }assume that {each }component of $G_{1}^{2}$ has at most
one cycle of $\mathcal{F}$. Let $C_{0}$ be a cycle of $G_{1}^{2}$ {belonging
}to $\mathcal{F}$. Clearly, $C_{0}=x_{1}^{0}x_{2}^{0}\ldots
x_{m_{0}}^{0}x_{1}^{0}$ is connected to a component of $G_{2}^{2}$ by {some
path (possibly an edge). Let $P=(x_{1}^{0}=)v_{0}v_{1}\ldots v_{t}$ be a
shortest path between $V(C_{0})$ and $V(G_{2}^{2})$. Then $v_{t}$ belongs to
{a component of }$G_{2}^{2}$, {say }$H_{1}.$ If $H_{1}\in \mathcal{F}_{0}$,
then let $H_{1}^{\prime }=H_{1}+P+C_{0}$ and let
G_{2}^{3}=(G_{2}^{2}-H_{1})\cup H_{1}^{\prime }$. Hence, assume that $H_{1}$
contains at least two cycles. We distinguish the following cases.
\smallskip \textbf{Case 1.} $H_{1}\in \mathcal{F}_{2,2} $
\newline
Suppose $H_{1}$ is obtained from two cycles $C_{1}=x_{1}^{1}\ldots
x_{m_{1}}^{1}x_{1}^{1}$ and $C_{2}=x_{1}^{2}\ldots x_{m_{2}}^{2}x_{1}^{2}$
by adding an edge $x_{1}^{1}x_{1}^{2}$. We further assume, without loss of
generality, that $v_{t}=x_{j}^{2}\in V(C_{2})$. {Let }$H_{1}^{\prime }$ be
the graph obtained from $C_{0}\cup C_{1}$ to which we add the path
Px_{j-1}^{2}\ldots x_{2}^{2}Q$, {in other words, $H_{1}^{\prime }$ {is
obtained from }$C_{0}+P+H_{1}$ by removing vertices
x_{j+1}^{2},...,x_{m_{2}}^{2}.$ Note that }$H_{1}^{\prime }$ belongs to
either $\mathcal{B}_{0,2}\cup \mathcal{B}_{1,1}.$ Now let
G_{2}^{3}=(G_{2}^{2}-H_{1})\cup H_{1}^{\prime} $ which we will discuss
further below.
{ \smallskip \textbf{Case 2.} $H_{1}\in \mathcal{F}_{0,2} $
\newline
Suppose $H_{1}$ is obtained from two cycles $C_{1}=x_{1}^{1}\ldots
x_{m_{1}}^{1}x_{1}^{1}$ and $C_{2}=x_{1}^{2}\ldots x_{m_{2}}^{2}x_{1}^{2}$
by adding a path $Q=(x_{1}^{1}=)z_{0}z_{1}\ldots z_{k}(=x_{1}^{2})$. Suppose
without loss of generality that $m_1\equiv 0\pmod 3$ and $m_2\equiv \pmod 3
. If $v_t=z_j$ for some $j$, then let $H_1^{\prime}$ be obtained from
C_{0}\cup C_{2}$ to which we add the path $Pz_j\ldots z_{k}$, and let
G_{2}^{3}=(G_{2}^{2}-H_{1})\cup (H_{1}^{\prime} \cup C_1)$. Suppose that
v_t\in V(C_1)\cup V(C_2)$. {We further assume, without loss of generality,
that }$v_{t}=x_{j}^{2}\in V(C_{2})$. {Let }$H_{1}^{\prime }$ {be the graph
obtained from }$C_{0}\cup C_{1}$ {to which we add the path }
Px_{j-1}^{2}\ldots x_{2}^{2}Q$, {in other words, }$H_{1}^{\prime }$ {is
obtained from }$C_{0}+P+H_{1}$ {by removing vertices }
x_{j+1}^{2},...,x_{m_{2}}^{2}.$ {Note that }$H_{1}^{\prime }$ belongs to
either $\mathcal{F}_{0,2}$ or $\mathcal{B}_{0,2}\cup \mathcal{B}_{1,1}.$ Now
let $G_{2}^{3}=(G_{2}^{2}-H_{1})\cup H_{1}^{\prime } $ which we will discuss
further below. }
\smallskip \textbf{Case 3.} $H_{1}\in \mathcal{B}_{r,s}$ where $r+s\geq 2$
\newline
Let $z^{\ast }$ {be the special vertex of }$H_{1}.$ {If }$v_{t}=z^{\ast }$,
then $H_{1}^{\prime }=H_{1}+P+C_{0}$ is a subgraph belonging to $\mathcal{E}
. In this case, let $G_{2}^{3}=(G_{2}^{2}-H_{1})\cup H_{1}^{\prime }$ which
we will discuss further below. Hence we assume that $v_{t}\neq z^{\ast }.$
First let $r+s=2.$ Then $H_{1}$ is obtained from two cycles
C_{1}=x_{1}^{1}\ldots x_{m_{1}}^{1}x_{1}^{1}$ and $C_{2}=x_{1}^{2}\ldots
x_{m_{2}}^{2}x_{1}^{2}$ by adding a path $Q=(x_{1}^{1}=)z_{0}z_{1}\ldots
z_{t}(=x_{1}^{2}),$where $t\geq 2.$ If $v_{t}\in \{z_{1},\ldots ,z_{t-1}\}$,
then let $H_{1}^{\prime }=H_{1}+C_{0}+P$ and $G_{2}^{3}=(G_{2}^{2}-H_{1}
\cup H_{1}^{\prime }$ . {Now, suppose, without loss of generality, that }
v_{t}=x_{j}^{2}\in V(C_{2}). $ {Let }$H_{1}$ be obtained {from }$C_{0}\cup
C_{1}$ {to which we add the path }$Px_{j-1}^{2}\ldots x_{2}^{2}Q.$ S{et }
G_{2}^{3}=(G_{2}^{2}-H_{1})\cup H_{1}^{\prime }$ {which we will discuss
further below}.
Now let $r+s\geq 3.$ {Assume that }$v_{t}$ {belongs to one of the }${s+r}$
cycles of $H_{1},$ {say} $C^{\prime }.$ Let $H_{1}^{\prime }=C^{\prime
}+C_{0}+P$ and $H_{1}^{\prime \prime }$ be the graph obtained from $H_{1}$
by deleting the vertices of $V(C^{\prime })$ and the path (if any) joining
z^{\ast }$ to $V(C^{\prime })$ in $H_{1}$. Note that $H_{1}^{\prime }$
belongs to either $\mathcal{F}_{0,2}$ or $\mathcal{B}_{0,2}\cup \mathcal{B
_{1,1}.$ Now let $G_{2}^{3}=(G_{2}^{2}-H_{1})\cup H_{1}^{\prime }$ . Assume
that $v_{t}$ belongs to a path on a tailed cycle $C_{m,\ell }$ of $H_{1},$
and let $P^{\prime }$ be the subpath between $v_{t}$ and the cycle
C^{\prime }$ of $C_{m,\ell }.$ {Let }$H_{1}^{\prime }=C^{\prime
}+C_{0}+P+P^{\prime }$ and $H_{1}^{\prime \prime }$ be the graph obtained
from $H_{1}$ by deleting {the }vertices of $V(C_{m,\ell }).$ {Note that }
H_{1}^{\prime }$ {belongs to either }$\mathcal{F}_{0,2}$ or $\mathcal{B
_{0,2}.$ Suppose $G_{2}^{3}=(G_{2}^{2}-H_{1})\cup H_{1}^{\prime }\cup
H_{1}^{\prime \prime }$. Obviously either the number of cycles of $G_{2}^{3}$
is greater than the number of cycles of $G_{2}^{\mathcal{T}}$ or the sum of
lengths of paths between two cycles of $G_{2}^{3}$ that belong to $\mathcal{
}_{0,2}\cup \mathcal{F}_{2,2}\cup (\cup _{r,s\geq 0;r+s=2}B_{r,s})$ is
greater than the corresponding sum of $G_{2}^{\mathcal{T}}$. Let
G_{1}^{3}=G-G_{2}^{3}.$ Let $\mathcal{T}_{3}^{1}$ be the family of all
cycles of $G_{2}^{3}$ which belongs to $\mathcal{F}$ and let $\mathcal{T
^{3} $ be a maximal subfamily of $\mathcal{F}$ such that $\mathcal{T
_{1}^{3}\subseteq \mathcal{T}^{3}$. If $G_{1}^{3}$ does not contain any
cycle of $\mathcal{F}$, then the family $\mathcal{T}^{3}$ and the subgraphs
G_{1}^{3}$ and $G_{2}^{3}$ satisfy the conditions which leads to a
contradiction because {of either }$c_{\mathcal{T}^{3}}>c_{\mathcal{T}}$ or
s_{\mathcal{T}^{3}}>s_{\mathcal{T}}$. Hence we assume that $G_{1}^{3}$
contains at least one cycle of $\mathcal{F}$. We repeat the above precess.
Since $G$ is finite, this process will stop and we obtain a maximal
subfamily $\mathcal{T}^{\prime }$ of pairwise disjoint cycles of $\mathcal{F}
$ with $|\mathcal{T}^{\prime }|\geq 2 $ and two disjoint subgraphs $G_{1}$
(possibly null), $G_{2}$ of $G$ such that $V(G)=V(G_{1})\cup V(G_{2})$,
G_{1}$ has no cycle of $\mathcal{F}$ and each component of $G_{2}$ belongs
to $\mathcal{F}_{0}\cup \mathcal{F}_{0,2}\cup \mathcal{F}_{2,2}\cup \mathcal
E}$. $\hfill \Box $
\begin{lemma}
\label{mainnn}\emph{Let $G$ be a connected graph with at least two disjoint
cycles of length $\equiv 0,2\pmod 3$, and let $\mathcal{F}$ be the family of
all cycles of $G$ with length $\equiv 0\;\mathrm{or}\;2\pmod 3$. Then there
exists a maximal subfamily $\mathcal{T}$ of pairwise disjoint cycles of
\mathcal{F}$ with $|\mathcal{T}|\geq 2$ and two disjoint subgraphs $G_{1}$
(possibly null), $G_{2}$ of $G$ such that $V(G)=V(G_{1})\cup V(G_{2})$,
G_{1}$ has no cycle of $\mathcal{F}$ and each component of $G_{2}$ belongs
to $\mathcal{F}_{0}\cup \mathcal{F}_{0,2}\cup \mathcal{F}_{2,2}\cup \mathcal
F}_{3}\cup \mathcal{E}$.}
\end{lemma}
\noindent \textbf{Proof.} Let $(\mathcal{T},G_{1},G_{2})$ {be the triple
satisfying the conditions of Lemma \ref{mainn}. Hence }$G_{1}$ has no cycle
of $\mathcal{F}$ {and each component of }$G_{2}$ {belongs to }$\mathcal{F
_{0}\cup \mathcal{F}_{0,2}\cup \mathcal{F}_{2,2}\cup \mathcal{E}$. {If there
are no two components }$H_{1}\in \mathcal{F}_{0,2},$ $H_{2}\in \mathcal{F
_{2,2}$ of $G_{2}$ {joined by a path }$P$ {in }$G$ {with all its vertices,
except the end-vertices, belong to }$V(G_{1})$, then $G_{1}$ and $G_{2}$ are
{the }desired subgraphs. {Hence we assume }that there {are two components }
H_{1}\in \mathcal{F}_{0,2}$ and $H_{2}\in \mathcal{F}_{2,2}$ of $G_{2}$
joined by a path }$P$ {in }$G$ {with all its vertices, except the
end-vertices, belong to }$V(G_{1}).$ Let $G_{2}^{\prime }$ be {\ the graph
obtained from $G_{2}$ by adding the path $P$ and let $G_{1}^{\prime
}=G-G_{2}^{\prime }$. Clearly $G_{1}^{\prime }$ and $G_{2}^{\prime }$
satisfy the conditions {and the result follows}. We can repeat this process
until we get two }subgraphs $G_{1}^{\ast }$ (possibly null) and $G_{2}^{\ast
}$ such that $V(G)=V(G_{1}^{\ast })\cup V(G_{2}^{\ast })$, $G_{1}^{\ast }$
has no cycle of $\mathcal{F}$, each component of $G_{2}^{\ast }$ is in
\mathcal{F}_{0}\cup \mathcal{F}_{0,2}\cup \mathcal{F}_{2,2}\cup \mathcal{F
_{3}\cup \mathcal{E}$ and {such that no path in }$G$ {like to the one
described above joins two components }$H^{\prime }\in \mathcal{F}_{0,2}$ and
$H^{\prime \prime }\in \mathcal{F}_{2,2}$ of $G_{2}^{\ast }.$ $\hfill \Box $
\bigskip
{From now on, }a graph in $\mathcal{(}\cup _{r+s\geq 2;s\leq 2}\mathcal{B
_{r,s})\cup \mathcal{F}_{3}\cup \mathcal{F}_{0,2}\cup \mathcal{F}_{0}$ {will
be called }\emph{strong}. {Also, the }special vertex of each graph in
\mathcal{B}_{r,s}$ {will be }called a \textit{strong vertex}.
\begin{lemma}
\label{strong}\emph{Let $k\geq 1$ be an integer and let $G$ be a graph of
order $n$ and minimum degree $\delta \geq 2$, which does not contain any
induced $\{C_{5},C_{8},\ldots ,C_{3k+2}\}$-cycles. If $G$ {is strong, }then
G$ has a 3-tuple $\overrightarrow{f}$ of {RDFs} such that $\omega
\overrightarrow{f})\leq \frac{(4k+8)3n}{6k+11}$ and all vertices of $G$ are
\overrightarrow{f}$-strong. }
\end{lemma}
\noindent \textbf{Proof.} Let $G\in (\cup _{r+s\geq 2;s\leq 2}\mathcal{B
_{r,s})\cup \mathcal{F}_{3}\cup \mathcal{F}_{0,2}\cup \mathcal{F}_{0}.$
Assume first that $G\in \mathcal{F}_{0}$. Then $\gamma _{R}(G)=\frac{2n}{3}
\frac{(4k+8)n}{6k+11}$. Let $G=x_{1}x_{2}\ldots x_{3t}x_{1}$ and define for
j\in \{1,2,3\}$ the functions $f_{j}$ on $V(G)$ as follows:
f_{j}(x_{3i+j})=2$ for $0\leq i\leq t-1$ and $f_{j}(x)=0$ otherwise. Clearly
$f_{j}$ is an $\gamma _{R}(G)$-function for each $j\in \{0,1,2\}$ and the
triple $\overrightarrow{f}=(f_{0},f_{1},f_{2})$ satisfies the desired result.
Assume now that $G\in \mathcal{F}_{0,2}$. {Since }$G$ has no induced
\{C_{5},C_{8},\ldots ,C_{3k+2}\}$-cycles, {we deduce that cycle of lentgh }
\equiv 2\pmod 3$ in $\mathcal{F}_{0,2}${\ has order at least }$3k+5$, and
thus $G$ has order at least $3k+8.$ {Now b}y Lemma \ref{MainLem}, $G$ has a
3-tuple $\overrightarrow{f}$ of RDFs such that $\omega (\overrightarrow{f
)\leq 2n+1$ and all vertices of $G$ are $\overrightarrow{f}$-strong. {A
simple calculation shows that }$\omega (\overrightarrow{f})\leq \frac
(4k+8)3n}{6k+11}$.
Next assume that\textbf{\ }$G\in \mathcal{F}_{3}$. {By definition, }$G$ is
obtained from a graph $G_{1}\in \mathcal{F}_{0,2}$ and a graph in $G_{2}\in
\mathcal{F}_{2,2}$ by adding either an edge $vw$ {or a path }$Q$ {joining a
vertex of }$G_{1}$ to a vertex of $G_{2}$ {so that all vertices of }$Q${\
become of degree two in }$G.$ Let $G_{1}$ be obtained from two cycles
C_{1}=x_{1}^{1}x_{2}^{1}\ldots x_{n_{1}}^{1}x_{1}^{1}\in \mathcal{F}_{0}$
and $C_{2}=x_{1}^{2}x_{2}^{2}\ldots x_{n_{2}}^{2}x_{1}^{2}\in \mathcal{F
_{2} $ by adding {\ either the edge }$x_{1}^{1}x_{1}^{2}$ {or a }path $P$
between $x_{1}^{1}$ and $x_{1}^{2}.$ {By Lemma \ref{MainLem} (items 8,9,10),
$G_{1}$ has a a 3-tuple $\overrightarrow{f}=(f_{1},f_{2},f_{3})$ of RDFss
such that $\omega (\overrightarrow{f},G_{1})\leq 2n(G_{1})+1$ and all
vertices of $G_{1} $ are $\overrightarrow{f}$-strong. Moreover, }let $G_{2}$
be obtained from two cycles $C_{3}=x_{1}^{3}x_{2}^{3}\ldots
x_{n_{1}}^{3}x_{1}^{3}\in \mathcal{F}_{2}$ and $C_{4}=x_{1}^{4}x_{2}^{4
\ldots x_{n_{4}}^{4}x_{1}^{4}\in \mathcal{F}_{2}$ by adding the edge
x_{1}^{3}x_{1}^{4}$. Without loss of generality, we assume that the added
edge $uv$ or the path $Q$ is between $V(C_{3})$ and $V(G_{1})$. {By
sequentially applying Lemmas \ref{tailedcycle-3p+1} (items 3,4) (once on }
uv $ or $Q$ and $C_{3},$ {and then on the resulting graph with }$C_{4}),$ {
\overrightarrow{f}$ can be extended to a triple $\overrightarrow{g}$ of RDFs
of $G$ such that $\omega (\overrightarrow{g},G_{2})\leq 2n(G_{2})+2$ and
each vertex of $G_{2}$ is $\overrightarrow{g}$-strong. Since }$G$ has no
induced $\{C_{5},C_{8},\ldots ,C_{3k+2}\}$-cycles, {we deduce that order
each cycle of lentgh }$\equiv 2\pmod 3$ in $G$ is at least $3k+5.$ {Using
the fact that }$G$ has three cycles {of length }$\equiv 2\pmod 3$ and one
cycle {of lentgh }$\equiv 0\pmod 3,$ we have {$n(G)\geq 9k+18.$ Therefore
\omega (\overrightarrow{g})\leq 2n(G)+3\leq \frac{(4k+8)3n(G)}{6k+11}.$ }
Using a similar argument we can show that for any graph $G\in \mathcal{\cup
_{r+s\geq 2;s\leq 2}B_{r,s}$ the result is {also }true. $\hfill \Box $
\begin{lemma}
\label{non-strong1}\emph{Let $k\geq 1$ be an integer and let $G$ be a graph
of order $n$, minimum degree $\delta \geq 2$, which does not contain any
induced $\{C_{5},C_{8},\ldots ,C_{3k+2}\}$-cycles. If $G\in \mathcal{B
_{r,s} $ with $s\geq 3$, then $G$ has a 3-tuple $\overrightarrow{f}$ of RDFs
such that $\omega (\overrightarrow{f})\leq \frac{(4k+8)3n}{6k+11}$ and the
special vertex as well as all vertices on tailed cycles of $G$ are
\overrightarrow{f}$-strong. }
\end{lemma}
\noindent \textbf{Proof.} Suppose $G$ be obtained from $r\geq 0$ graphs
C_{n_{1},\ell _{1}},\ldots ,C_{n_{r},\ell _{r}}$ and $s\geq 3$ cycles
C_{m_{1}},\ldots ,C_{m_{s}}$, where $n_{i}\equiv 2\pmod 3$ and $m_{j}\equiv
\pmod 3$ for each $i,j$, by adding a new vertex $z$ (special vertex)
attached to {endvertices of the }$C_{n_{i},\ell _{i}}$'s and to one vertex
of each cycle $C_{m_{j}}$. {We first note that each of the }$r+s\geq 3$
cycles has order at least }$3k+5$, {and thus each tailed cycle contains at
least }${(3k+5)+1}$ vertices. {Hence }${n(G)\geq (3k+5)(s+r)+r+1.}$ {\ Now,
if }$r=0$, then the result follows from Lemmas \ref{MainLem}-(11) {and the
previous fact}. {Hence assume }that $r\geq 1$. Let $H$ be obtained from $G$
by deleting all vertices of $C_{n_{i},\ell _{i}}$'s. By Lemma \ref{MainLem}
(item 11), $H$ has a triple $\overrightarrow{f}$ such that $\omega
\overrightarrow{f})\leq 2n(H)-s+4$ and $z$ is $\overrightarrow{f}$-strong.
Since }${n(H)\geq (3k+5)s+1,}$ we deduce that $\omega (\overrightarrow{f
)\leq \frac{(4k+8)3n(H)}{6k+11}$. {Now, by applying repeatedly }Lemma \re
{tailedcycle-3p+1}-(4) {on }$C_{n_{1},\ell _{1}},\ldots ,C_{n_{r},\ell _{r}}
, we can extend $\overrightarrow{f}$ to a triple $\overrightarrow{g}$ of $G$
such that $\omega (g,\cup _{i=1}^{r}C_{n_{1},\ell _{1}})\leq
\sum_{i=1}^{r}(2n(C_{n_{1},\ell _{1}})+1)$ and all newly added vertices are
\overrightarrow{g}$-strong. {Therefore, }$\omega (\overrightarrow{g})\leq
2n+r+4-s.$ {Now by the previous fact on the order and the calculation, we
can see that }$2n+r+4-s\leq \frac{(4k+8)3n}{6k+11}$, {which proves the
result.}$\hfill \Box ${\ }
\begin{lemma}
\label{non-strong2} Let $k\geq 1$ be an integer and let $G\in \mathcal{F
_{2,2}$ be a graph of order $n$, minimum degree $\delta \geq 2$, which does
not contain any induced $\{C_{5},C_{8},\ldots ,C_{3k+2}\}$-cycles. Then
\begin{enumerate}
\item $G$ has a $3$-tuple $f$ of RDFs such that $\overrightarrow{f}\leq \
2n(G) + 1 \leq \frac{(4k+8)3n(G)}{(6k+11)}.$
\item If H is a graph obtained from $G$ and a cycle
C_{3p+1}=x_{1}...x_{3p+1}x_1$ by adding an edge between them, then H has a
3-tuple $f$ of RDFs such that $\omega(\overrightarrow{f}) \leq \frac
(4k+8)3n(G)}{(6k+11)}$ and all vertices of H but $x_{3p+1}$ are
\overrightarrow{f}$-strong.
\item If $H$ is a graph obtained from $G$ and a tailed cycle $C_{3p+1,\ell }$
with vertex set $x_1,\ldots ,x_{3p+1}$, $y_{1},\ldots ,y_{\ell }$ by joining
$y_{\ell }$ to a vertex of G, then H has a 3-tuple $f $ of RDFs such that
\omega(\overrightarrow{f}) \leq \frac{\emph{(4k+8)3n(G)}}{(6k+11)}$ and all
vertices of $H$ but $x_{3p+1}$ are $\overrightarrow{f}$-strong.
\end{enumerate}
\end{lemma}
\noindent \textbf{Proof.} (1) is easy to show and so we prove only (2) and
(3). Let $G\in \mathcal{F}_{2,2}$ be formed from two cycles $C_{1}$ and
C_{2}$ by adding an edge between them, and let $H$ be obtained from $G$ and
the cycle $C_{3p+1}$ (resp. tailed cycle $C_{3p+1,\ell }$) by adding an edge
$xy$ (resp. $xy_{\ell }$), where without loss of generality $x\in V(C_{2})$.
Let $K$ be the graph obtained from $H$ by deleting all vertices of $V(C_{1})
. By Lemma \ref{MainLem} (items 2,3 and 4), $K$ has a 3-tuple
\overrightarrow{g}$ of RDFs of $K$ such that $\omega (\overrightarrow{g
)\leq 2n(K)+1$ and all vertices of $K$ except $x_{3p+1}$ are
\overrightarrow{g}$-strong. Now by Lemma \ref{tailedcycle-3p+1}, we can
extend $\overrightarrow{g}$ to a 3-tuple $\overrightarrow{f}$ of RDFs of $H$
such that $\omega (\overrightarrow{f})\leq 2n(H)+2$ and all vertices of $H$
except $x_{3p+1}$ are $\overrightarrow{g}$-strong. By assumption we have
n(H)\geq 6k+14$ and {thus one can check that }$\omega (\overrightarrow{f
)\leq \frac{(4k+8)3n(H)}{6k+11}$. $\hfill\Box $
\section{\protect\smallskip Proof of Conjecture \protect\ref{conj}}
Now we are ready to state our main result.{\ }
\begin{theorem}
\label{Theorem} \emph{Let $G$ be a graph of order $n\ge 6k + 9$, minimum
degree $\delta\ge 2$, which does not contain any induced
\{C_5,C_8,\ldots,C_{3k+2}\}$-cycles. Then $\gamma_R(G)\le \frac{(4k+8)n}
6k+11}.$}
\end{theorem}
\noindent \textbf{Proof.} Let $\mathcal{F}$ be the family of all cycles of
G $ with length $\equiv 0\;\mathrm{or}\;2\pmod 3$. If $|\mathcal{F}|=0$,
then the result follows from Theorem \ref{Th1} and if $|\mathcal{F}|\geq 1$
and $\mathcal{F}$ contains a cycle which intersect any cycle of $\mathcal{F}
, then the result follows from Theorems \ref{Th2} and \ref{Th3}. Henceforth,
we assume that each cycle of $\mathcal{F}$ belongs to a maximal subfamily
\mathcal{T}$ of pairwise disjoint cycles of $\mathcal{F}$ with $|\mathcal{T
|\geq 2$. Let $(G_{1}^{1},G_{2}^{1}),\ldots ,(G_{1}^{m},G_{2}^{m})$ be all
pairs of subgraph such that $V(G)=V(G_{1}^{i})\cup V(G_{2}^{i})$, $G_{1}^{i}$
has no cycle of $\mathcal{F}$ and each component of $G_{2}^{i}$ belongs to
\mathcal{F}_{0}\cup \mathcal{F}_{0,2}\cup \mathcal{F}_{2,2}\cup \mathcal{F
_{3}\cup \mathcal{E}$. Let{\ }$s_{(G_{1}^{i},G_{2}^{i})}$ be the sum of the
lengths of paths between two cycles in the components of $G_{2}^{\mathcal{T
} $ {that belong to }$\mathcal{F}_{0,2}\cup \mathcal{F}_{2,2}\cup (\cup
_{r,s\geq 0;r+s=2}B_{r,s}).$ {Among all pairs }$(G_{1}^{i},G_{2}^{i}),$ let
(G_{1},G_{2})$ {be one chosen so that: }
\begin{description}
\item[(C$_{1}$)] the {number of }strong components of $G_{2}$ is maximized.
\item[(C$_{2}$)] subject to Condition (C$_{1}$): the number of cycles of
G_{2}$ belonging to $\mathcal{F}$ is maximized.
\item[(C$_{3}$)] subject to Conditions (C$_{1}$) and (C$_{2}$): the number
of components of $G_{2}$ in $\mathcal{F}_{2,2}$ is minimized.
\item[(C$_{4}$)] subject to Conditions (C$_{1}$), (C$_{2}$) and (C$_{3}$):
s_{(G_{1},G_{2})}$ is maximized.
\end{description}
We proceed with some {further claims that are needed for our proof. }
\smallskip \noindent \textbf{Claim 1.} Let $M$ be a component of $G_{2}$
such that $M\in \mathcal{F}_{2,2}$. Then there is no path $v_{0}v_{1}\ldots
v_{t+1}\;(t\geq 1)$ in $G$ such that $v_{0},v_{t+1}\in M$, $v_{1},\ldots
,v_{t}\in V(G_{1})$ and $v_{0}$ and $v_{t+1}$ belong to different cycles of
M$.\newline
\noindent \textbf{Proof of Claim 1.} Suppose, to the contrary, that there is
path $P=v_{0}v_{1}\ldots v_{t+1}\;(t\geq 1)$ in $G$ such that
v_{0},v_{t+1}\in M$, $v_{1},\ldots ,v_{t}\in V(G_{1})$ and $v_{0}$ and
v_{t+1}$ belong to different cycles of $M$. {Let }$e^{\ast }$ {be the edge
joining the two cycles of }$M${\ and let }$M^{\prime }$ be obtained from $M$
by deleting $e^{\ast }$ and adding path $P$. {Set }$G_{2}^{\prime
}=(G_{2}-M)\cup M^{\prime }$ and $G_{1}^{\prime }=G-G_{2}$. Clearly
V(G)=V(G_{1}^{\prime })\cup V(G_{2}^{\prime })$, $G_{1}$ has no cycle of
\mathcal{F}$ and each component of $G_{2}^{\prime }$ {is in }$\mathcal{F
_{0}\cup \mathcal{F}_{0,2}\cup \mathcal{F}_{2,2}\cup \mathcal{F}_{3}\cup
\mathcal{E}$. But $G_{2}^{\prime }$ has one more strong component than
G_{2},$ {contradicting our choice of }$(G_{1},G_{2}).$ $\ \ \ \ \ \ \ \ \ \
\ \hspace{3cm}\Box $ \ \ \
\bigskip\
\smallskip \noindent \textbf{Claim 2.} {For any two components }$M_{1}$ and
M_{2}$ of $G_{2}$ {belonging to }$\mathcal{F}_{2,2}$, {there }is no path
v_{0}v_{1}\ldots v_{t+1}\;(t\geq 1)$ in $G$ such that $v_{0}\in M_{1}$,
v_{t+1}\in M_{2}$ and $v_{1},\ldots ,v_{t}\in V(G_{1})$.\newline
\noindent \textbf{Proof of Claim 2.} Suppose, to the contrary, that {for
two components $M_{1},M_{2}$ of $G_{2}$ {\ belonging to }$\mathcal{F}_{2,2},$
there is a path $v_{0}v_{1}\ldots v_{t+1}\;(t\geq 1)$ in $G_{1}$ such that
v_{0}\in M_{1}$ and $v_{t+1}\in M_{2}$. Suppose {that } $M_{1}$ is obtained
from two cycles $C_{1}=u_{1}^{1}\ldots u_{m_{1}}^{1}u_{1}^{1}$ and
C_{2}=u_{1}^{2}\ldots u_{m_{2}}^{2}u_{1}^{2}$ by adding {the edge }
u_{1}^{1}u_{1}^{2}$, and let $M_{2}$ {be }obtained from two cycles
C_{3}=u_{1}^{3}\ldots u_{m_{3}}^{3}u_{1}^{3}$ and $C_{4}=u_{1}^{4}\ldots
u_{m_{4}}^{4}u_{1}^{4}$ by adding the edge $u_{1}^{3}u_{1}^{4}$. {Moreover,
assume, without loss of generality, that }$v_{0}=u_{j}^{2}\in V(C_{2})$
where $j\geq m_{2}/2$ ({by relabeling the vertices if necessary) }and
v_{t+1}=u_{b}^{3}\in V(C_{3})$ where $b\geq m_{3}/2$ ({by relabeling the
vertices if necessary)}. {Now, let }$M$ be the subgraph obtained from $C_{1}$
and $C_{4} $ by adding the path $u_{1}^{1}u_{1}^{2}u_{2}^{2}\ldots
u_{j}^{2}v_{1}v_{2}\ldots v_{t}u_{b}^{3}u_{b-1}^{3}\ldots u_{1}^{3}u_{1}^{4}
. {Set }$G_{2}^{\prime }=(G_{2}-(M_{1}\cup M_{2}))\cup M$ and $G_{1}^{\prime
}=G-G_{2}^{\prime }$. If $G_{1}^{\prime }$ has no cycle of $\mathcal{F}$,
then by considering the pair $(G_{1}^{\prime },G_{2}^{\prime }) $ we get
one more strong component in }$G_{2}^{\prime }$ {than in }$G_{2},$
contradicting our choice of }$(G_{1},G_{2}).$ {Hence we }assume that
G_{1}^{\prime }$ has some cycles of $\mathcal{F}$.
First let $G_{1}^{\prime }$ has exactly one cycle $C$ of $\mathcal{F}$. If
C $ has length $\equiv 0\pmod 3$, then as above we get a contradiction by
considering the subgraphs $G_{2}^{\prime \prime }=G_{2}^{\prime }\cup C$ and
$G_{1}^{\prime \prime }=G-G_{2}^{\prime \prime }$. Hence suppose $C$ has
length $\equiv 2\pmod 3$. Since $G_{1}$ has no cycle of $\mathcal{F}$, we
may assume that $C$ contains one of the vertices $u_{j+1}^{2},\ldots
,u_{m_{2}}^{2}$. {Let }$\ell \in \{j+1,\ldots ,m_{2}\}$ be the smallest
index {such that }$u_{\ell }^{2}\in V(C)$. {Let }$M^{\prime }=(M\cup
C)+u_{j}^{2}u_{j+1}^{2}\ldots u_{\ell }^{2}$. Clearly $M^{\prime }$ is
strong {because it belongs to }$\mathcal{B}_{r,s},$ {with }$r+s\geq 3$ and
s\leq 2$. {By considering the subgraphs }$G_{2}^{\prime \prime
}=(G_{2}-(M_{1}\cup M_{2}))\cup M^{\prime }$ and $G_{1}^{\prime \prime
}=G-G_{2}^{\prime \prime },$ {the pair }$(G_{1}^{\prime \prime
},G_{2}^{\prime \prime })$ {leads to a contradiction on the choice of }
(G_{1},G_{2}).$
Now let $G_{1}^{\prime }$ has at least two disjoint cycles $C$ and
C^{\prime }$ of $\mathcal{F}$. Using an argument similar to that described
in the proof of Lemma \ref{mainn}, we can obtain a pair $(G_{1}^{\prime
\prime },G_{2}^{\prime \prime })$ such that $G_{1}^{\prime \prime }$ has no
cycle of $\mathcal{F}$ and each component of $G_{2}^{\prime \prime }$
belongs to $\mathcal{F}_{0}\cup \mathcal{F}_{0,2}\cup \mathcal{F}_{2,2}\cup
\mathcal{F}_{3}\cup \mathcal{E}$, {where either }$G_{2}^{\prime \prime }$
has more strong components than }$G_{2}$ or the number of cycles of
G_{2}^{\prime \prime }$ {belonging to }$\mathcal{F}$ is greater than the\
number of cycles of $G_{2}$ {\ belonging to}$\mathcal{F}$ or
s_{(G_{1}^{\prime \prime },G_{2}^{\prime \prime })}>s_{(G_{1},G_{2})}.$ {In
either case, we obtain a }contradiction. $\hfill\Box $
\bigskip
Recall that a component of $\mathcal{B}_{r,s}$ is not strong when $s\geq 3.$
\smallskip \noindent \textbf{Claim 3.} Let $M_{1}$ and $M_{2}$ be two
non-strong components of $G_{2}$ such that $M_{1}\in \mathcal{F}_{2,2}$ and
M_{2}\in \mathcal{B}_{r,s}$. Then there is no path $v_{0}v_{1}\ldots
v_{t+1}\;(t\geq 1)$ in $G$ such that $v_{1},\ldots ,v_{t}\in V(G_{1})$,
v_{0}\in M_{1}$, $v_{t+1}\in M_{2}$ and $v_{t+1}$ is\ not {the }special
vertex of $M_{2}$.\newline
\noindent \textbf{Proof of Claim 3.} Suppose, to the contrary, {\ that
there is a path $v_{0}v_{1}\ldots v_{t+1}\;(t\geq 1)$ in $G$ such that
v_{1},\ldots ,v_{t}\in V(G_{1})$, $v_{0}\in M_{1}$, $v_{t+1}\in M_{2}$ and
v_{t+1}$ is not\textbf{\ }special vertex of $M_{2}$. Suppose $M_{1}$ is
obtained from two cycles $C^{1}=u_{1}^{1}\ldots u_{m_{1}}^{1}u_{1}^{1}$ and
C^{2}=u_{1}^{2}\ldots u_{m_{2}}^{2}u_{1}^{2}$ by adding the edge
u_{1}^{1}u_{1}^{2}$ and let $M_{2}$ obtained from $r\geq 0$ tailed-cycle
C_{n_{1},\ell _{1}}$, $\ldots ,C_{n_{r},\ell _{r}}$ and $s\geq 3$ cycles
C_{m_{1}},\ldots ,C_{m_{s}}$, where $n_{i}\equiv 2\pmod 3$ and $m_{j}\equiv
\pmod
3$ for each $i,j$, by adding a new vertex $z$ (special vertex) and attaching
$z$ to the leaf of each tailed cycle $C_{n_{i},\ell _{i}}$ and to one vertex
of each cycle $C_{m_{j}}$. Without loss of generality, that we may assume
that $v_{1}$ is adjacent to the vertex $u_{j}^{2}\in V(C_{2})$ where $j\geq
m_{2}/2$ ({by relabeling the vertices if necessary)}.
First let $v_{t+1}$ belongs to a cycle $C_{m_{i}}=w_{1}^{i}w_{2}^{i}\ldots
w_{m_{i}}^{i}w_{1}^{i}$ for some $i.$ {Without loss of generality, let }$i=1$
and {$v_{t+1}=w_{q}^{1}$.} Let $M_{2}^{\prime }$ be obtained
from $M_{2}$ by deleting the vertices of $C_{m_{1}}$, and $M_{1}^{\prime
}=(C^{1}\cup C_{m_{1}})\cup u_{1}^{1}u_{1}^{2}u_{2}^{2}\ldots
u_{j}^{2}v_{1}\ldots v_{t}w_{q}^{1}$. {In this case, consider the subgraphs
$G_{2}^{1}=(G_{2}-(M_{1}\cup M_{2}))\cup (M_{1}^{\prime }\cup M_{2}^{\prime
})$ and $G_{1}^{1}=G-G_{2}^{1}$ {{which we will be discussing later}. }
Now {assume that }$v_{t+1}$ {belongs to a tailed cycle }$C_{n_{i},\ell _{i}}$
for some $i$, say $i=1$. Let $C=w_{1}^{1}w_{2}^{1}\ldots
w_{n_{1}}^{1}w_{1}^{1}$ be the cycle of $C_{n_{1},\ell _{1}}$ and
P=y_{1}^{1}\ldots y_{\ell _{1}}^{1}$ be the tail of $C_{n_{1},\ell _{1}}$
such that $w_{1}^{1}y_{1}^{1}\in E(G)$. {Consider the two situations
depending on whether }$v_{t+1}$ {is on the cycle or the tail. If }
v_{t+1}\in V(C)$, say $v_{t+1}=w_{q}^{1}$, {then let }$M_{2}^{\prime }$ be
obtained from $M_{2}$ by deleting the vertices of $C_{n_{1},\ell _{1}}$, and
$M_{1}^{\prime }=(C^{1}\cup C)\cup u_{1}^{1}u_{1}^{2}u_{2}^{2}\ldots
u_{j}^{2}v_{1}\ldots v_{t}w_{q}^{i}$. {In this case, consider the subgraphs
$G_{2}^{1}=(G_{2}-(M_{1}\cup M_{2}))\cup (M_{1}^{\prime }\cup M_{2}^{\prime
})$ and $G_{1}^{1}=G-G_{2}^{1}.$ {If }$v_{t+1}\in V(P),$ say
v_{t+1}=y_{q}^{1}$, {then let }$M_{2}^{\prime }$ be obtained from $M_{2}$ by
deleting the vertices of $C_{n_{1},\ell _{1}}$, and $M_{1}^{\prime
}=(C^{1}\cup C)\cup u_{1}^{1}u_{1}^{2}u_{2}^{2}\ldots u_{j}^{2}v_{1}\ldots
v_{t}y_{q}^{1}y_{q-1}^{1}\ldots y_{1}^{1}w_{1}^{1}$. {In this case, consider
the subgraphs }$G_{2}^{1}=(G_{2}-(M_{1}\cup M_{2}))\cup (M_{1}^{\prime }\cup
M_{2}^{\prime })$ and $G_{1}^{1}=G-G_{2}^{1}$.
{Observe that in any situation, either }the number of cycles of $G_{2}^{1}$
belonging to }$\mathcal{F}$ is greater than {the one of }$G_{2}$ {that are
in }$\mathcal{F}$ or $s_{(G_{1}^{1},G_{2}^{1})}>s_{(G_{1},G_{2})}$. {Now, if
}$G_{1}^{1}$ has no cycle of $\mathcal{F}$, then the pair
(G_{1}^{1},G_{2}^{1})$ leads to a contradiction. Otherwise, by repeating
above process we can obtain a pair $(G_{1}^{\prime },G_{2}^{\prime })$ such
that $G_{1}^{\prime}$ has no cycle of $\mathcal{F}$ and each component of
G_{2}^{\prime}$ belongs to $\mathcal{F}_{0}\cup \mathcal{F}_{0,2}\cup
\mathcal{F}_{2,2}\cup \mathcal{F}_{3}\cup \mathcal{E}$, {where either the
number of }strong components of $G_{2}^{\prime }$ is greater than the {one
of }$G_{2}$ or the number of cycles of $G_{2}^{\prime }$ {that are in }
\mathcal{F}$ is greater than the number of cycles of $G_{2}$ {belonging to }
\mathcal{F}$ or $s_{(G_{1}^{\prime },G_{2}^{\prime })}>s_{(G_{1},G_{2})}.$
In either case, we have a contradiction and the desired claim follows. }
\hfill\Box $
\bigskip
\smallskip \noindent \textbf{Claim 4.} {If }$M\in \mathcal{B}_{r,s}\;${is a
non-strong component of }$G_{2}$ {\ with a special vertex }$z,$ then there
is no path $v_{0}v_{1}\ldots v_{t}v_{t+1}$ $(t\geq 1)$ in $G$ such that
v_{1},\ldots ,v_{t}\in V(G_{1})$, $v_{0},v_{t+1}\in V(M)-\{z\}$ and
v_{0},v_{t+1}$ belong to different {near cycles of }$z.$ \newline
\noindent \textbf{Proof of Claim 4.} Let $M$ be obtained from $r\geq 0$
tailed-cycle $C_{n_{1},\ell _{1}}$, $\ldots ,C_{n_{r},\ell _{r}}$ and $s\geq
3$ cycles $C_{m_{1}},\ldots ,C_{m_{s}}$, where $n_{i}\equiv 2\pmod 3$ and
m_{j}\equiv 2\pmod
3$ for each $i,j$, by adding a new vertex $z$ (special vertex) and attaching
$z$ to the leaf of each tailed cycle $C_{n_{i},\ell _{i}}$ and to one vertex
of each cycle $C_{m_{j}}$. {Moreover, let }$C_{m_{i}}=z_{1}^{i}z_{2}^{i
\ldots z_{m_{i}}^{i}z_{1}^{i}$ for {\ each }$i\in \{1,...,s\}$ and let
V(C_{n_{i},\ell _{i}})=\{x_{1}^{i},\ldots ,x_{n_{i}}^{i},y_{1}^{i},\ldots
,y_{\ell _{i}}^{i}\},$ where $x_{1}^{i},\ldots ,x_{n_{i}}^{i}$ {induce in
order the cycle }of $C_{n_{i},\ell _{i}}$ and $y_{1}^{i},\ldots ,y_{\ell
_{i}}^{i}$ {induce in order the }tail of $C_{n_{i},\ell _{i}}$.
Suppose, to the contrary, that there is a path $P=v_{0}v_{1}\ldots
v_{t}v_{t+1}$ in $G$ such that $v_{1},\ldots ,v_{t}\in V(G_{1}),$
v_{0},v_{t+1}\in V(M)-\{z\}$ and $v_{0},v_{t+1}$ belong to different {near
cycles of }$z.$
First let $r+s=3$. Then $r=0$ and $s=3$. Assume, without loss of generality,
that $v_{0}=u_{k}^{1}$ and $v_{t+1}=u_{j}^{2}$ where $j\leq m_{1}/2$ and
k\leq m_{2}/2$. Let $M^{\prime }$ be obtained from $C_{1},C_{3}$ by adding
the path $Pu_{j-1}^{2}\ldots u_{1}^{2}zu_{1}^{3}$. Note that if
v_{t+1}=u_{1}^{2}$, then the added path will be {simply }$Pzu_{1}^{3}$.
Consider the subgraph }$G_{2}^{\prime }=(G_{2}-M)\cup M^{\prime }$. If
G-G_{2}^{\prime }$ has no cycle of $\mathcal{F}$, then the pair
(G-G_{2}^{\prime },G_{2}^{\prime })$ {provides a number of strong components
in }$G_{2}^{\prime }$ {greater than the one of }$G_{2},$ {contradicting our
choice of the pair }$(G_{1},G_{2}).$ {Assume now that }$G-G_{2}^{\prime }$
has exactly\textbf{\ }one cycle $C$ of $\mathcal{F}$. Then $V(C)$ meets at
least a vertex of $\{u_{j+1}^{2},\ldots ,u_{m_{2}}^{2}\} $ and let $p$ be
the largest integer that $u_{p}^{2}\in V(C) $. Let $M^{\prime \prime }$ be
obtained from $M^{\prime }\cup C$ by adding the path $u_{1}^{2}\ldots
u_{p}^{2}$. {Consider the subgraph }$G_{2}^{\prime \prime }=(G_{2}-M^{\prime
})\cup M^{\prime \prime } $. Then, {as above, }the pair $(G-G_{2}^{\prime
\prime },G_{2}^{\prime \prime })$ leads to a contradiction. {Hence we can
assume that }$G-G_{2}^{\prime }$ has at least two disjoint cycles of
\mathcal{F}$. Clearly each of these cycles meets at least a vertex of
\{u_{j+1}^{2},\ldots ,u_{m_{2}}^{2}\}$. Consider the subgraph $G^{\prime }$
of $G-G_{2}^{\prime }$ induced by the vertices of these cycles and the
vertices of $\{u_{j+1}^{2},\ldots ,u_{m_{2}}^{2}\}$, {and let }$K_{1}$ and
K_{2}$ be two disjoint subgraphs of {$G^{\prime }$} satisfying
the conditions of Lemma \ref{mainn}. Then the pair $(G_{2}^{\prime }\cup
K_{2},G-(G_{2}^{\prime }\cup K_{2}))$ leads to a contradiction because the
number of strong components of $G_{2}^{\prime }\cup K_{2}$ is greater than
the number of strong components of $G_{2}$.
Now let $r+s\geq 4$, {and assume that }$P$ connects two cycles $C$ and
C^{\prime }$ of $M$ {that are at distance one from }$z.$ Let $M^{\prime }$
be obtained from $M$ by deleting the vertices of $C\cup C^{\prime }$, and
let $M^{\prime \prime }=C+C^{\prime }+P$. {Now, if we consider the subgraphs
}$G_{2}^{\prime \prime }=(G_{2}-M)\cup (M^{\prime }\cup M^{\prime \prime })$
and $G_{1}^{\prime \prime }=G-G_{2}^{\prime \prime }$, {then one can see, as
above, that the pair }$(G_{1}^{\prime \prime },G_{2}^{\prime \prime })$
leads to a contradiction. $\hfill\Box $
\bigskip
\smallskip \noindent \textbf{Claim 5.} Let $M_{1},M_{2}\in \mathcal{E}$ be
two non-strong components of $G_{2}$ and let $z_{i}$ be the special vertex
of $M_{i}$. Then there is no path $P=v_{0}v_{1}\ldots v_{t}v_{t+1}\;(t\geq
1) $ in $G$ such that $v_{1},\ldots v_{t}\in V(G_{1})$, $v_{0}$ belongs to a
{near cycle }$C_{1}$ {of }$z_{1}$ and $v_{t+1}$ belongs to {near cycle }
C_{2}$ {of }$z_{2}$.\newline
\noindent \textbf{Proof of Claim 5.} {Suppose to the contrary that such a
path }$P$ exists. Let $M_{i}^{\prime }$ be obtained from $M_{i}$ by deleting
the vertices of $V(C_{i})$ for {each }$i\in \{1,2\} $ and {let }
M=(C_{1}\cup C_{2})+P$. {\ Consider the subgraphs }$G_{2}^{\prime
}=(G_{2}-(M_{1}\cup M_{2}))\cup (M_{1}^{\prime }\cup M_{2}^{\prime }\cup M)$
and $G_{1}^{\prime }=G-G_{2}^{\prime }$. {Since }$G_{2}^{\prime }${{\ {has
more }}}strong components than $G_{2},$ {the pair }$(G_{1}^{\prime
},G_{2}^{\prime })$ contradicts the choice of the {pair }$(G_{1},G_{2}).$
\hfill\Box $
\bigskip
{Now, let }$K_{0}$ be the subgraph of $G_{2}$ that consists of all
non-strong components of $G_{2}$ and {let }$H_{0}$ be the subgraph of $G_{2}$
that consists of all strong components of $G_{2}$. By Lemma \ref{strong},
each component $M$ of $H_{0}$ has a $3$-tuple $\overrightarrow{f_{M}}$ of
RDFs of $M$ such that $\omega (\overrightarrow{f_{M}})\leq \frac{(4k+8)3n(M
}{6k+11}$ and all vertices of $M$ are $\overrightarrow{f_{M}}$-strong.
Therefore, by combining }these $3$-tuples we obtain a 3-tuple
\overrightarrow{f_{0}}$ of RDFs of $H_{0}$ such that $\omega
\overrightarrow{f_{0}})\leq \frac{(4k+8)3n(H_{0})}{6k+11}$ and all vertices
of $H_{0}$ are $\overrightarrow{f_{0}}$-strong. {Moreover, }by Lemmas \re
{non-strong1} and \ref{non-strong2}, each component $M$ of $K_{0}$ has a
3-tuple $\overrightarrow{g_{M}}$ of RDFs of $M$ such that $\omega
\overrightarrow{g_{M}})\leq \frac{(4k+8)3n(M)}{6k+11},$ and {if further }
M\in \mathcal{B}_{r,s}\;(s\geq 3)$, then its special vertex {as well as all
all vertices }on tailed cycle are $\overrightarrow{g_{M}}$-strong.
Therefore, by combining }these $3$-tuples we obtain a $3$-tuple
\overrightarrow{g_{0}}$ of RDFs of $K_{0}$ such that $\omega
\overrightarrow{g_{0}})\leq \frac{(4k+8)3n(K_{0})}{6k+11}$ and all vertices
of $K_{0}$ are $\overrightarrow{g_{0}}$-strong {except vertices on near
cycles of some }special vertex or vertices on the component in $\mathcal{F
_{2,2}$.
If {there is a path }$P_{1}=v_{0}v_{1}\ldots v_{t}v_{t+1}$ in $G$ such that
v_{1},\ldots ,v_{t}\in V(G_{1})$, $N(v_{1})\cup N(v_{t})\subseteq
V(H_{0}\cup K_{0})\cup V(P_{1})$ and both $v_{0},v_{t+1}$ belong to a
component $M\in \mathcal{F}_{2,2}$ of $K_{0}$, then we deduce from Claim 1
that both of $v_{0},v_{t+1}$ belong to {same }cycle of $M$. {Let }$M^{\prime
}=M+P$. It follows from Lemma \ref{MainLem}-(5) {and Lemmas }\ref{ear1} and
\ref{induced} that $M^{\prime }$ has a 3-tuple $\overrightarrow{f^{\prime }}$
of RDFs of $M^{\prime }$ such that $\omega (\overrightarrow{f^{\prime }
)\leq 2n(M^{\prime })+2\leq \frac{(4k+8)3n(M^{\prime })}{6k+11}$ and all
vertices {of }$M^{\prime }$ are $\overrightarrow{f^{\prime }}$-strong
except }$v_{1}$ and $v_{t}$. {In this case, let }$K_{0}^{1}=K_{0}-M$,
\overrightarrow{g_{0}^{1}}=\overrightarrow{g_{0}}|_{K_{0}^{1}}$ {(the
restriction of }$\overrightarrow{g_{0}}$ on $K_{0}^{1})$ and
H_{0}^{1}=H_{0}\cup (M+P)=H_{0}\cup M^{\prime }.$ {Let }$\overrightarrow
f_{0}^{1}}$ be a 3-tuple {\ of RDFs }obtained by combining the 3-tuples
f_{0}$ and $\overrightarrow{f^{\prime }}$. Clearly all vertices of
H_{0}^{1} $ which have {\ a }neighbor outside $H_{0}^{1}\cup K_{0}^{1}$ are
\overrightarrow{f_{0}^{1}}$-strong. By repeating this process we obtain two
\ sequences }of subgraphs $K_{0}\supseteq K_{0}^{1}\supseteq \ldots
\supseteq K_{0}^{d}$ and $H_{0}\subseteq H_{0}^{1}\subseteq \ldots \subseteq
H_{0}^{d}$ so that: \textrm{(i)} {there is no path }$P=w_{0}w_{1}\ldots
w_{r}w_{r+1}$ in $G$ with $w_{1},\ldots ,w_{r}\in V(G_{1})-(\cup
_{i=1}^{d}V(P_{i}))$, $N(w_{1})\cup N(w_{r})\subseteq V(H_{0}^{d}\cup
K_{0}^{d})\cup (\cup _{i=1}^{d}V(P_{i}))$ and both $v_{0},v_{t+1}$ belong to
a component $M\in \mathcal{F}_{2,2}$ of $K_{0}^{d}$, and \textrm{(ii)}
H_{0}^{d}$ has a 3-tuple $\overrightarrow{f_{0}^{d}}$ such that all {its
vertices which have{a} neighbor outside $V(H_{0}^{d}\cup K_{0}^{d})$ are
\overrightarrow{f_{0}^{d}}$-strong. Let $H_{1}=H_{0}^{d}$, $K_{1}=K_{0}^{d}
, $\overrightarrow{g_{1}}=\overrightarrow{g_{0}}|_{K_{0}^{d}}$ and
\overrightarrow{f_{1}}=f_{0}^{d}$. {Observe that $\omega
\overrightarrow{f_{1}})\leq \frac{(4k+8)3n(H_{1})}{6k+11}$ and $\omega
\overrightarrow{g_{1}})\leq \frac{(4k+8)3n(K_{1})}{6k+11}$. If
V(G)=V(H_{1}\cup K_{1})$, then by combining 3-tuple $\overrightarrow{f_{1}}$
of $H_{1}$ and 3-tuple $\overrightarrow{g_{1}}$ of $K_{1}$, we get a 3-tuple
$\overrightarrow{h}$ of $G$ such that $\omega (\overrightarrow{h})\leq \frac
(4k+8)3n(G)}{6k+11}$ which will prove the theorem.} {Hence assume }that
V(G)\neq V(H_{1}\cup K_{1})$, {and let }$G_{2}^{1}=H_{1}\cup K_{1}$ and
G_{1}^{1}=G-G_{2}^{1}$.
If {there is a path }$P_{1}=v_{0}v_{1}\ldots ,v_{t}v_{t+1}$ in $G$ such that
$v_{0},v_{t+1}\in V(H_{1})$, $v_{1},\ldots ,v_{t}\in V(G_{1}^{1})$ and
N(v_{1})\cup N(v_{t})\subseteq V(H_{1})\cup V(P_{1})$, then let
H_{1}^{1}=H_{1}+P$. By Lemma \ref{ear1}, we can extend $\overrightarrow{f_{1
}$ to a 3-tuple $\overrightarrow{f_{1}^{1}}$ of RDFs of $H_{1}^{1}$ such
that $\omega (\overrightarrow{f_{1}^{1}})\leq \frac{(4k+8)3n(H_{1}^{1})}
6k+11}$, {where }all vertices of $H_{1}^{1}$ but $v_{1}$ {and }$v_{t}$ are
\overrightarrow{f_{1}^{1}}$-strong. {Now, if there is a path }
P_{2}=z_{0}z_{1}\ldots ,z_{m}z_{m+1}\;(m\geq 1)$ in $G$ such that
z_{0},z_{m+1}\in V(H_{1}^{1})$, $z_{1},\ldots ,z_{m}\in
V(G_{1}^{1})-V(P_{1}) $, $N(z_{1})\cup N(z_{m})\subseteq V(H_{1}^{1})\cup
V(P_{2})$, then let $H_{1}^{2}=H_{1}^{1}+P_{2}$. By Lemma \ref{ear1}, we can
extend $\overrightarrow{f_{1}^{1}}$ to a 3-tuple $\overrightarrow{f_{1}^{2}}$
of RDFs of $H_{1}^{2}$ such that $\omega (\overrightarrow{f_{1}^{2}})\leq
\frac{(4k+8)3n(H_{1}^{2})}{6k+11}$ and all new vertices but $z_{1},z_{m}$
are $\overrightarrow{f_{1}^{2}}$-strong. By repeating this process we obtain
a {sequence }of subgraphs $H_{1}\subseteq H_{1}^{1}\subseteq \ldots
\subseteq H_{1}^{q}$ so that {there is no path }$P=w_{0}w_{1}\ldots
,w_{r}w_{r+1}$ in $G$ such that $w_{0},w_{r+1}\in V(H_{1}^{q})$,
w_{1},\ldots ,w_{r}\in V(G_{1}^{1})-(\cup _{i=1}^{q}V(P_{i}))$ and
N(w_{1})\cup N(w_{r})\subseteq V(H_{1}^{q})\cup V(P)$. Moreover, $H_{1}^{q}$
has a 3-tuple $\overrightarrow{f_{1}^{q}}$ of RDFs of $H_{1}^{q}$ such that
\omega (\overrightarrow{f_{1}^{q}})\leq \frac{(4k+8)3n(H_{1}^{q})}{6k+11}$
and all vertices are $\overrightarrow{f_{1}^{q}}$-strong unless the vertices
which have no neighbors outside of $H_{1}^{q}\cup K_{1}$. If
V(G)=V(H_{1}^{q}\cup K_{1})$, {then as above, by combining
3-tuple $\overrightarrow{f_{1}^{q}}$ and 3-tuple $\overrightarrow{g_{2}}
\overrightarrow{g_{1}}$, the result follows.} {Hence }assume that $V(G)\neq
V(H_{1}^{q}\cup K_{1})$, {{{and let }}}$H_{2}=H_{1}^{q}$,
\overrightarrow{f_{2}}=\overrightarrow{f_{1}^{q}}$, $K_{2}=K_{1}$,
\overrightarrow{g_{2}}=\overrightarrow{g_{1}}$, $G_{2}^{2}=H_{2}\cup K_{2}$
and $G_{1}^{2}=G-G_{2}^{2}$. {In the following we will use Lemma
\ref{12} by applying its three items, one by one (in any order), starting
with the subgraph }$G_{1}^{2}$ {and obtaining each time (when the
item occurs) a sequence of subgraphs. The last subgraph of the sequence will
be used for the next item. }
\smallskip \noindent \textbf{Case 1.} {$G_{1}^{2}$} contains a
tailed $m$-cycle $C_{m,\ell }\;({m\equiv 1\pmod 3})$, with
vertex set $\{x_{1},\ldots ,x_{m},y_{1},\ldots ,y_{\ell }\}$, such that
y_{\ell }$ is adjacent to some vertex $x$ of $G_{2}^{2}$ and
N_{G}(x_{m})\subseteq {V(G_{2}^{2})}\cup V(C_{m,\ell })$.\newline
First {assume that }$x\in V(H_{2})$. Then $x$ is $\overrightarrow{f_{2}}
-strong. {Let }$H_{2}^{1}$ be obtained from $H_{2}$ by adding the tailed
cycle $C_{m,\ell }$ and the edge $xy_{\ell }$. By Lemma \re
{tailedcycle-3p+1}, $\overrightarrow{f_{2}}$ can be extended to a 3-tuple
\overrightarrow{f_{2}^{1}}$ or RDFs of $H_{2}^{1}$ such that $\omega
\overrightarrow{f_{2}^{1}})\leq \frac{(4k+8)3n(H_{2}^{1})}{6k+11}$ and all
new vertices but $x_{m}$ are $\overrightarrow{f_{2}^{1}}$-strong. Set also
K_{2}^{1}=K_{2}$ and $\overrightarrow{g_{2}^{1}}=\overrightarrow{g_{1}}$.
\noindent Now {assume that }$x$ belongs to a component $M$ of $K_{2}$ such
that $M\in \mathcal{F}_{2,2}$. {Let }$M^{\prime }$ be obtained from $M$ and
C_{m,\ell }$ by adding the edge $xy_{\ell }$. By Lemma \ref{MainLem}{-(5)
and Lemma }\ref{tailedcycle-3p+1}, one can see that $M^{\prime }$ has a
3-tuple $f_{M^{\prime }}$ of RDFs such that $\omega (\overrightarrow
f_{M^{\prime }}})\leq 2n(M^{\prime })+2\leq \frac{(4k+8)3n(M^{\prime })}
6k+11}$ and all of its vertices but $x_{m}$ are $\overrightarrow
f_{M^{\prime }}}$-strong. {Let }$H_{2}^{1}=H_{2}\cup M^{\prime }$,
K_{2}^{1}=K_{2}-M$, $g_{1}^{2}$ {be }the restriction of $g$ on $K_{2}^{1}$
and 3-tuple $\overrightarrow{f_{2}^{1}}$ is obtained from combining
\overrightarrow{f_{M^{\prime }}}$ and $\overrightarrow{f_{2}}$. Note that
all vertices of $H_{2}^{1}$ which have neighbor in {
G_{1}^{2}-V(C_{m,\ell})$} are $\overrightarrow{f_{2}^{1}}$-strong.
\noindent Next assume that $x$ belongs to a component $M$ of $K_{2}$ such
that $M\in \mathcal{B}_{r,s}\;({s\geq 3})$ and $x$ is
\overrightarrow{g}$-strong. {{{Let }}}$M^{\prime }$ {{
be }}}obtained {{{from }}}$M$ and $C_{m,\ell }$ by
adding the edge $xy_{\ell }$ and let $H_{2}^{1}=H_{2}$ and
K_{2}^{1}=(K_{2}-M)\cup M^{\prime }$. By Lemma \ref{tailedcycle-3p+1},
\overrightarrow{g_{2}}$ can be extended to a 3-tuple $g_{2}^{1}$ of RDFs of
K_{2}^{1}$ such that $\omega (\overrightarrow{g_{2}^{1}})\leq \frac
(4k+8)3n(K_{2}^{1})}{6k+11}$ and all newly added vertices but $x_{m}$ are
\overrightarrow{g_{2}^{1}}$-strong.
Finally, {assume that }$x$ belongs to a component $M$ of $K_{2}$ such that
M\in \mathcal{B}_{r,s}\;({s\geq 3})$ and $x$ is not
\overrightarrow{g}$-strong. Then $x$ belongs to a {near }cycle $C$ {from the
}special vertex of $M$. Let $M^{\prime }$ be obtained from $M$ by deleting
the vertices of $C$ and {let }$M^{\prime \prime }$ be obtained from $C$ and
C_{m,\ell }$ by adding the edge $xy_{\ell }$. {In this case, let }
H_{2}^{1}=H_{2}\cup M^{\prime \prime }$, $K_{2}^{1}=K_{2}-V(C)$ and
g_{1}^{2}$ is the restriction of $\overrightarrow{g_{2}}$ on $K_{2}^{1}$. By
Lemma \ref{MainLem} (items 2,3,4), $f_{2}$ can be extended to a 3-tuple of
RDFs $f_{2}^{1}$ of $H_{2}^{1}$ such that $\omega (\overrightarrow{f_{2}^{1}
)\leq \frac{(4k+8)3n(H_{2}^{1})}{6k+11}$ and all newly added vertices but
x_{m}$ are $\overrightarrow{f_{2}^{1}}$-strong.
By repeatedly applying the above argument we obtain two\ {{{
sequences }}}of subgraphs $H_{2}\supseteq H_{2}^{1}\supseteq \ldots
\supseteq H_{2}^{r_{1}}$ and {$K_{2},K_{2}^{1},\ldots
,K_{2}^{s_{1}}$} such that there is no tailed cycle $C_{m,\ell }\;(
m\equiv 1\pmod 3})$ in $G-(H_{2}^{r_{1}}\cup K_{2}^{s_{1}})$
whose end-vertex is adjacent to a vertex of $H_{2}^{r_{1}}\cup K_{2}^{s_{1}}
. Let $H_{3}=H_{2}^{r_{1}}$, $K_{3}=K_{2}^{s_{1}}$, $f_{3}$ be a 3-tuple of
RDFs of $H_{3}$ such that all vertices of $H_{3}$ which have a neighbor
outside $H_{3}\cup K_{3}$ are $\overrightarrow{f_{3}}$-strong, and $g_{3}$
be a 3-tuple of RDFs of $K_{3}$ such that all newly added vertices of $K_{3}$
which have no neighbor outside $H_{3}\cup K_{3}$ are $\overrightarrow{g_{3}}
-strong. {Let }$G_{2}^{3}=H_{3}\cup K_{3}$ and $G_{1}^{3}=G-G_{2}^{3}$.
\smallskip \noindent \textbf{Case 2.} $G_{1}^{3}$ contains a cycle
C_{m}=x_{1}x_{2}\ldots x_{m}x_{1}\;({m\equiv 1\pmod 3})$ such
that $N_{G}(x_{m})\subseteq V(G_{2}^{3})\cup V(C_{m})$ and there is an edge
x_{1}y$ with $y\in V(G_{3}^{2})$ Applying an argument similar to that
described in Case 1, we obtained subgraphs $H_{4}$ and $K_{4}$ such that
H_{3}\subseteq H_{4}$, and a 3-tuple $f_{4}$ of RDFs of $H_{4}$ so that all
vertices of $H_{4}$ {having a }neighbor outside $H_{4}\cup K_{4}$ are
\overrightarrow{f_{4}}$-strong, and a 3-tuple $\overrightarrow{g_{4}}$ of
RDFs of $K_{4}$ so that all newly added vertices of $K_{3}$ {that have a
neighbor outside $H_{4}\cup K_{4}$ are $\overrightarrow{g_{4}}$-strong.
Assume that $G_{2}^{4}=H_{4}\cup K_{4}$ and $G_{1}^{4}=G-G_{2}^{4}$.
Let $\overrightarrow{h}$ be a 3-tuple defined on $G_{2}^{4}$ {\ obtained by
combining }$\overrightarrow{f_{4}}$ and $\overrightarrow{g_{4}} $.
\smallskip \noindent \textbf{Case 3.} $G_{1}^{4}$ has a path
P=v_{0}v_{1},\ldots ,v_{t}v_{t+1}\;(t\geq 1)$ such that $v_{0},v_{t+1}\in
G_{2}^{4}$, $v_{1},\ldots ,v_{t}\in V(G_{1}^{4})$ and $N_{G}(v_{1})\cup
N_{G}(v_{t+1})\subseteq V(G_{2}^{4})\cup V(P)$.\newline
By Claims 1,2,3,4 {and }5, at least one of the vertices $v_{0},v_{t+1}$ is
\overrightarrow{h}$-strong. First {assume that each of }$v_{0}$ and $v_{t+1}$
is $\overrightarrow{h}$-strong. {Let }$G_{2}^{5}=G_{2}^{4}+P$ and
G_{1}^{5}=G-G_{2}^{5}$. By Lemma \ref{ear1}, we can extend $\overrightarrow{
}$ to a 3-tuple $\overrightarrow{h^{1}}$ of RDFs of $G_{2}^{5}$ such that
\omega (\overrightarrow{h^{1}})\leq \frac{(4k+8)3n(G_{2}^{5})}{6k+11}$ and
all vertices of $G_{2}^{5}$ which have neighbor in $G_{1}^{5}$ are
\overrightarrow{h^{1}}$-strong. {Assume now, without loss of generality,
that }$v_{0}$ is $\overrightarrow{h}$-strong and $v_{t+1}$ is not
\overrightarrow{h}$-strong. It follows that $v_{t+1}$ is on {a near }cycle
from a special vertex of a component of $K_{4}$ or is in a component of
K_{4}$ {that belongs }to $\mathcal{F}_{2,2}$. If $v_{t+1}$ is on a {near
cycle $C$ from a special vertex, then let $G_{2}^{5}=G_{2}^{4}+P$. By Lemma
\ref{tailedcycle-3p+1}, $h|_{G_{2}^{4}-C}$ can be extended to a 3-tuple
\overrightarrow{h^{1}}$ of RDFs of $G_{2}^{5}$ such that $\omega
\overrightarrow{h^{1}})\leq \frac{(4k+8)3n(G_{2}^{5})}{6k+11}$ and all
vertices of $G_{2}^{5}$ which have neighbors in $G_{1}^{5}-V(P)$ are
\overrightarrow{h^{1}}$-strong. If $v_{t_{1}}$ is in a component $M$ of
K_{4}$ {belonging }to $\mathcal{F}_{2,2}$, then let $G_{2}^{5}=G_{2}^{4}+P$.
By applying Lemma \ref{tailedcycle-3p+1} twice, $h|_{G_{2}^{4}-M}$ can be
extended to a 3-tuple $\overrightarrow{h^{1}}$ of RDFs of $G_{2}^{5}$ such
that $\omega (\overrightarrow{h^{1}})\leq \frac{(4k+8)3n(G_{2}^{5})}{6k+11}$
and all vertices of $G_{2}^{5}$ which have neighbors in $G_{1}^{5}-V(P)$ are
$\overrightarrow{h^{1}}$-strong.
By repeating this process we obtain a 3-tuple $\overrightarrow{h}$ of RDFs
G $ such that $\omega (\overrightarrow{h})\leq \frac{(4k+8)3n(G)}{6k+11}$,
\ implying that }$\gamma _{R}(G)\leq \frac{(4k+8)n(G)}{6k+11}$ as desired.
\ \ \ \hfill \Box $
\bigskip
{{{Now, the next result settling }}}Conjecture \ref{conj} is an
immediate consequence of Theorem \ref{Theorem} and the Gallai-type result
\gamma _{R}(G)+\partial (G)=n$ which is valid for every graph $G$ of order
n.$
\begin{corollary}
\emph{Let $G$ be a graph of order $n\ge 6k + 9$, minimum degree $\delta\ge 2
, which does not contain any induced $\{C_5,C_8,\ldots,C_{3k+2}\} $-cycles.
Then $\partial(G)\ge \frac{(2k+3)n}{6k+11}.$}
\end{corollary}
\section{Acknowledgment}
This work was supported by the National Key R \& D Program of China (Grant No. 2019YFA0706402)
and the Natural Science Foundation of Guangdong Province under grant 2018A0303130115.
|
\section{Introduction}
Applying ideas from Brill-Noether theory, we
construct effective divisors in the moduli space of \mbox{$k$-canonical} divisors on algebraic curves.
To describe the ambient moduli space in question, start with
the $k$-th Hodge bundle $\mathbb{E}^k_g$. This is the vector bundle of stable $k$-differentials over the moduli space $\overline{\mathcal{M}}_g$ of stable curves of genus $g$.
It is defined as $\mathbb{E}^k_g:=\pi_* \left( \omega_\pi^{\otimes k}\right)$, where
$\pi\colon \mathcal{C}_g\rightarrow \overline{\mathcal{M}}_g$ is the universal curve with relative dualizing sheaf $\omega_\pi$.
The projectivization $\P\mathbb{E}^k_g$ of $\mathbb{E}^k_g$ compactifies the moduli space of $k$-canonical divisors on smooth algebraic curves.
Various divisor classes on $\P\mathbb{E}^k_g$ have recently been considered in the literature, including the class of the closure of the divisorial stratum in $\P\mathbb{E}^k_g$ consisting of $k$-differentials with a double zero, and the first Chern class of Prym-Tyurin vector bundles
\cite{korotkin2011tau, korotkin2013tau, sauvagetcohomology, korotkin2019tau}.
Here our focus is on divisors consisting of $k$-differentials which vanish at a Brill-Noether special point. To start, consider the effective divisor
\[
\mathbb{W}^k_g := \left\{(C, \mu) \in \P\E^k_{g} \; \big| \; C \text{ is smooth and } \mu \text{ vanishes at a Weierstrass point} \right\}\!.
\]
We compute the class of $\mathbb{W}^k_g$ in terms of standard generators of the Picard group of $\P\E^k_{g}$ with rational coefficients, generalizing the case $k=1$ treated in \cite{gheorghita2018effective}:
\begin{thm}
\label{thm:kW}
For $k\geq 1$ and $g\geq 2$, one has
\begin{multline*}
\left[\overline{\mathbb{W}}^k_g\right] = -g(g^2-1)\,\eta + k(6g^2+4g+2)\,\lambda - k{g+1 \choose 2}\,\delta_0 \\
- \sum_{i=1}^{\lfloor g/2 \rfloor} k(g+3)i(g-i)\,\delta_i \,\,\in\,\,\textup{Pic}_\mathbb{Q}\left(\P\E^k_{g}\right).
\end{multline*}
\end{thm}
Here, $\eta := c_1\big(\mathscr{O}_{\P\mathbb{E}^k_g}(-1)\big)$; $\lambda$ is the pull-back from $\overline{\mathcal{M}}_g$ of the first Chern class of the Hodge bundle; $\delta_0$ is the pull-back from $\overline{\mathcal{M}}_g$ of the class of the locus of curves whose general element has a non-disconnecting node;
and $\delta_i$ is the pull-back from $\overline{\mathcal{M}}_g$ of the class of the locus of curves whose general element has a disconnecting node and components of genera $i$ and $g-i$, for $1\leq i\leq \lfloor g/2 \rfloor$.
Additionally, we consider divisors in $\P\E^k_{g}$ obtained by imposing vanishing at an arbitrary Brill-Noether special point.
We briefly review the setup and refer the reader to \S\ref{sec:BNdiv} for the required background.
For $g\geq 2$ and a sequence
\[
\bm{a}:0\leq a_0< \dots <a_r\leq d,
\]
define $\rho(g,r,d,\bm{a}):= g-(r+1)(g-d+r)-\sum_{i=0}^r (a_i-i)$. When $\rho(g,r,d,\bm{a})=-1$,
a general curve of genus $g$ contains only finitely many points $P$
where a linear series $\ell$ of type $\mathfrak{g}^r_d$ has vanishing sequence $\bm{a}^\ell(P)\geq \bm{a}$.
In this case, the locus $\mathcal{M}_{g,d}^{\bm{a}}$ of pointed curves $(C,P)$ which admit a linear series $\ell$ of type $\mathfrak{g}^r_d$ with $\bm{a}^\ell(P)\geq \bm{a}$ is a proper subvariety with a unique irreducible divisorial component inside the moduli space $\mathcal{M}_{g,1}$ of pointed genus-$g$ curves \cite{EH-1}. For $d=2g-2$, $r=g-1$, and $\bm{a}=(0,1,2,\dots,g-2,g)$, one obtains the divisor of curves with a marked Weierstrass point. The class of the divisorial component of $\overline{\mathcal{M}}_{g,d}^{\bm{a}}$ lies in the cone spanned by the pullback of the Brill-Noether divisor class $\mathcal{BN}_g$ from $\overline{\mathcal{M}}_g$ and the class $\mathcal{W}_g$ of the divisor of marked Weierstrass points \cite{EH-1}. Specifically, the divisorial component of $\overline{\mathcal{M}}_{g,d}^{\bm{a}}$ has class equal to $\mu_{g,d,\bm{a}} \, \mathcal{BN}_g + \nu_{g,d,\bm{a}} \, \mathcal{W}_g$, for some nonnegative rational coefficients $\mu_{g,d,\bm{a}}$ and $\nu_{g,d,\bm{a}}$ computed in \cite{FT}. These classes have been used in various settings, including the study of the birational geometry of moduli spaces of pointed curves \cite{MR1953519, MR2530855}, and the proof of the non-varying property of sums of Lyapunov exponents for certain strata of abelian and quadratic differentials in low genus \cite{MR3033521, chen2014quadratic}.
The loci $\mathcal{M}_{g,d}^{\bm{a}}$ induce natural subvarieties of $\P\mathbb{E}^k_g$.
Namely, given $g\geq 2$ and a sequence $\bm{a}:0\leq a_0< \dots <a_r\leq d$,
define the locus ${\H}^{\bm{a}}_{g,d}$ in $\P\E^k_{g}$ as
\[
{\H}^{\bm{a}}_{g,d}:=\left\{ (C, \mu)\in \P\E^k_{g} \,\Bigg|\,
\begin{array}{l}
\mbox{$C$ is smooth and $\mu$ vanishes at some point $P$}\\[0.2cm]
\mbox{such that }\bm{a}^\ell(P) \geq \bm{a} \mbox{ for some $\ell \in G^r_d(C)$}
\end{array}
\right\}.
\]
When $\rho(g,r,d,\bm{a})=-1$, the locus ${\H}^{\bm{a}}_{g,d}$ has a divisorial component.
For $d=2g-2$, $r=g-1$, and $\bm{a}=(0,1,2,\dots,g-2,g)$, the locus ${\H}^{\bm{a}}_{g,d}$ specializes to the above divisor $\mathbb{W}^k_g$.
We show that the class of the closure of the divisorial component of ${\H}^{\bm{a}}_{g,d}$, denoted by $\left[ \overline{\H}^{\bm{a}}_{g,d} \right] \in \textup{Pic}_\mathbb{Q}\left(\P\E^k_{g}\right)$, lies in the cone spanned by the pullback of the Brill-Noether divisor class $\mathcal{BN}_g$ from $\overline{\mathcal{M}}_g$ and the divisor class $\left[\overline{\mathbb{W}}^k_g\right]$ from Theorem \ref{thm:kW}, analogously to the result from \cite{EH-1} for the divisorial components of $\mathcal{M}_{g,d}^{\bm{a}}\subset{\mathcal{M}}_{g,1}$. Specifically, we have:
\begin{cor}
\label{cor:kBN}
For $k\geq 1$, $g \geq 2$, and
$\bm{a}: 0 \leq a_0 < \dots < a_r \leq d$ such that $\rho(g,r,d,\bm{a}) =-1$,
one has
\[
\left[ \overline{\H}^{\bm{a}}_{g,d} \right] = 2k(g-1)\,\mu_{g,d,\bm{a}} \, \mathcal{BN}_g + \nu_{g,d,\bm{a}} \left[\overline{\mathbb{W}}^k_g\right] \,\,\in\,\,\textup{Pic}_\mathbb{Q}\left(\P\E^k_{g}\right).
\]
\end{cor}
Next, we show how Theorem \ref{thm:kW} for $g=2$ complements a result from \cite{korotkin2019tau}.
The closure of the divisorial stratum in $\P\mathbb{E}^k_g$ of $k$-differentials with a double zero is computed for all $(g, k) \neq (2,2)$ in \cite{korotkin2011tau, korotkin2013tau, sauvagetcohomology, korotkin2019tau}.
For $g=k=2$, the formula from \cite{korotkin2019tau} specializes to a weighted sum of two components. Namely,
one has
\[
\left[\overline{\H}^2_2(2, 1, 1)\right]+ 2\left[\overline{\H}^2_2(2,2)\right] = 72 \lambda -10\eta -6\delta_0-6\delta_1
\]
from \cite[Def.~1.3, Thm 1.12]{korotkin2019tau}, where ${\H}^2_2(2, 1, 1)\subset \P\mathbb{E}^2_2$ is the divisorial stratum of quadratic differentials vanishing at a Weierstrass point, and ${\H}^2_2(2,2)\subset \P\mathbb{E}^2_2$ is the divisorial stratum consisting of squares of holomorphic differentials.
Since one has $\mathbb{W}^2_2 = {\H}_2^2(2,1,1)$, we deduce:
\begin{cor}
The closure of the divisorial stratum ${\H}^2_2(2,2)\subset \P\mathbb{E}^2_2$ parametrizing squares of holomorphic differentials has class
\[
\left[\overline{\H}^2_2(2,2)\right] = -2\eta + 12\lambda - \delta_0 \,\, \in\,\, \mathrm{Pic}\left(\P\mathbb{E}^2_{2}\right).
\]
\end{cor}
We present two independent proofs of Theorem \ref{thm:kW} and Corollary \ref{cor:kBN}. For the first proof,
the key idea is to regard the locus ${\H}^{\bm{a}}_{g,d}$ as the projection of a locus in the space $\P\mathbb{E}^k_{g,1}$ obtained by pulling-back $\P\mathbb{E}^k_g$ via $\overline{\mathcal{M}}_{g,1}\rightarrow \overline{\mathcal{M}}_g$. Let
\begin{equation}
\label{eq:incdiv}
\mathbb{H}^k_{g, 1} \subset \P\mathbb{E}^k_{g,1}
\end{equation}
be the \textit{incidence divisor} consisting of smooth pointed curves together with the class of a stable $k$-differential vanishing at the marked point.
In Proposition \ref{prop:BNirreducible}, we show that the divisorial component of $\overline{\H}^{\bm{a}}_{g,d}$ is the push-forward of the intersection of $\overline{\mathbb{H}}^k_{g, 1}$ and the pull-back of the divisorial component of $\overline{\mathcal{M}}_{g,d}^{\bm{a}}\subset\overline{\mathcal{M}}_{g,1}$, and thus deduce the divisorial class of $\overline{\H}^{\bm{a}}_{g,d}$ in \S\ref{sec:proofThmskWkBN}.
An alternative approach is pursued in \S\ref{sec:testfamilies}, where the class of the divisorial component of $\overline{\H}^{\bm{a}}_{g,d}$ is computed by intersecting with various test families.
This proof has the advantage of showing explicit restrictions of the locus $\overline{\H}^{\bm{a}}_{g,d}$, including restrictions to pencils of curves on K3 surfaces following \cite{MR852158, cukierman1993curves, MR2123229} and pencils of Du Val curves following \cite{arbarello2016explicit}.
Finally, we obtain the following result on the incidence divisor.
Using a dense collection of Teichm\"uller curves of abelian and quadratic differentials
having negative intersection with the incidence divisor, we show:
\begin{thm}
\label{thm:Z1extremal}
For $k\in \{1,2\}$, the class of $\overline{\mathbb{H}}^k_{g, 1}$ is rigid and extremal in $\overline{\mathrm{Eff}}^1\left( \P\mathbb{E}^k_{g,1}\right)$.
\end{thm}
To review the notation:
A class $E$ in the cone $\overline{\mathrm{Eff}}^1( X)$ of pseudo-effective divisor classes on a projective variety $X$ is called \textit{extremal} if $E = E_1 +E_2$ for $E_1$ and $E_2$ in $\overline{\mathrm{Eff}}^1( X)$ implies that both $E_1$ and $E_2$ are proportional to~$E$. An effective cycle class $E$ is called \textit{rigid} if any effective cycle with class $mE$ is supported on the support of $E$.
It is natural to ask whether the class of $\overline{\mathbb{H}}^k_{g, 1}$ is rigid and extremal also for $k\geq 3$. More generally, one can consider higher codimensional incidence varieties consisting of smooth $n$-pointed curves together with the class of a stable $k$-differential vanishing at all marked points. We have recently started the study of the classes of such varieties in \cite{gt}. It is natural to ask whether the rigidity and extremality of the incidence divisor for $k\in \{1,2\}$ extend to its higher codimensional counterparts.
\subsection*{Acknowledgements}
We would like to thank Dawei Chen for helpful conversations on $k$-differentials and the incidence variety compactification.
\section{Background on pointed Brill-Noether divisors}
\label{sec:BNdiv}
Here we review the background on pointed Brill-Noether theory required in \S\S\ref{sec:proofThmskWkBN}-\ref{sec:testfamilies} following \cite{MR910206, EH-1, FT}.
For a smooth algebraic curve $C$, a linear series of type $\mathfrak{g}^r_d$ on $C$ is a pair $(L,V)$ where $L\in \mathrm{Pic}^d(C)$ and $V\subseteq H^0(C,L)$ is a subspace of dimension $r+1$.
The variety $G^r_d(C)$ parametrizes linear series of type $\mathfrak{g}^r_d$ on $C$.
For $\ell = (L,V)$ in $G^r_d(C)$, the \textit{vanishing sequence} of $\ell$ at a point $P$ in $C$
\[
\bm{a}^\ell(P): 0 \leq a_0 < \dots < a_r \leq d
\]
is defined as the increasing sequence of vanishing orders of sections in $V$ at~$P$.
For a sequence $\bm{a}: 0 \leq a_0 < \dots < a_r \leq d$, the \textit{adjusted Brill-Noether number} is defined as
\[
\rho\left(g,r,d,\bm{a}\right) :=g - (r+1)(g-d+r) - \sum_{i=0}^r \left(a_i - i\right).
\]
The pointed version of the Brill-Noether Theorem \cite{MR910206} states that a general pointed curve $(C,P)$ of genus $g>0$ admits a linear series $\ell \in G^r_d(C)$ with vanishing sequence $\bm{a}^\ell(P) = \bm{a}$ if and only if
\[\sum_{i=0}^r(a_i - i + g - d + r)_+ \leq g,\]
where $(n)_+ := \max\{n, 0\}$ for $n \in \mathbb{Z}$.
This implies \mbox{$\rho(g,r,d,\bm{a}^\ell(P))\geq 0$} for a general $(C,P)$ and any $\ell\in G^r_d(C)$. We refer to \cite{FT2} for explicit examples of smooth pointed curves satisfying the pointed Brill-Noether Theorem.
When $g \in \{0,1\}$, one has $\rho(g,r,d,\bm{a}^\ell(P)) \geq 0$ for any $(C,P)$ and any $\ell \in G^r_d(C)$. However, when $g \geq 2$ and
$\bm{a}: 0 \leq a_0 < \dots < a_r \leq d$ such that $\rho(g,r,d,\bm{a}) <0$, the locus
\[
\mathcal{M}^{\bm{a}}_{g,d}:=\left\{ (C,P)\in \mathcal{M}_{g,1} \,|\, \bm{a}^\ell(P) \geq \bm{a} \mbox{ for some } \ell\in G^r_d(C)\right\}
\]
is a proper subvariety of $\mathcal{M}_{g,1}$: when $\rho(g,r,d,\bm{a})=-1$, it contains a unique divisorial component, while all components have higher codimension if $\rho(g,r,d,\bm{a})<-1$ \cite{EH-1}.
For example, when $d=2g-2$, \mbox{$r=g-1$,} and $\bm{a}=(0,1,2,\dots,g-2,g)$, the locus $\mathcal{M}^{\bm{a}}_{g,d}$ is the irreducible divisor of curves with a marked Weierstrass point.
In general, $\mathcal{M}^{\bm{a}}_{g,d}$ may not be irreducible, see \cite[\S 2]{EH-1}.
Assume $\rho(g,r,d,\bm{a}) =-1$, and denote by $\left[\overline{\mathcal{M}}^{\bm{a}}_{g,d}\right]\in \mathrm{Pic}\left( \overline{\mathcal{M}}_{g,1}\right)$ the class of the closure of the divisorial component of $\mathcal{M}^{\bm{a}}_{g,d}$.
After \cite{EH-1}, one has
\begin{equation}
\label{eq:muBNnuW}
\left[\overline{\mathcal{M}}^{\bm{a}}_{g,d}\right] = \mu_{g,d,\bm{a}} \, \mathcal{BN}_g + \nu_{g,d,\bm{a}} \, \mathcal{W}_g \,\,\in\,\, \mathrm{Pic}\left( \overline{\mathcal{M}}_{g,1}\right),
\end{equation}
for some $\mu_{g,d,\bm{a}},\nu_{g,d,\bm{a}} \in \mathbb{Q}_{\geq 0}$, where $\mathcal{W}_g$ is the class of the divisor of curves with a marked Weierstrass point \cite{Cuk}:
\begin{equation}
\label{eq:W}
\mathcal{W}_g := {g+1 \choose 2}\psi - \lambda - \sum_{i=1}^{g-1} {g-i+1 \choose 2}\delta_i \,\,\in\,\, \mathrm{Pic}\left( \overline{\mathcal{M}}_{g,1}\right),
\end{equation}
and $\mathcal{B}\mathcal{N}_g$ is the pullback of the Brill-Noether divisor class from $\overline{\mathcal{M}}_g$ \cite{MR910206}:
\begin{equation}
\label{eq:BN}
\mathcal{BN}_g := (g+3)\lambda - \frac{g+1}{6}\delta_0 - \sum_{i=1}^{g-1} i(g-i)\delta_i \,\,\in\,\, \mathrm{Pic}\left( \overline{\mathcal{M}}_{g,1}\right).
\end{equation}
Explicit formulae for $\mu_{g,d,\bm{a}} $ and $\nu_{g,d,\bm{a}}$ were computed in \cite{FT}, and make an appearance in our computations. These values are expressed in terms of
the number
\begin{equation}
\label{eq:ngdadef}
n_{g,d,\bm{a}}:= \#\left\{ (P, \ell) \in C \times G^r_d(C) \,| \, \bm{a}^\ell(P) \geq \bm{a}\right\},
\end{equation}
where $C$ is a general curve of genus $g \geq 2$. Let $\delta^i_j$ be the Kronecker delta. After \cite{FT}, one has
\begin{equation}
\label{eq:ngda}
\begin{split}
n_{g,d,\bm{a}} = g! \sum_{0 \leq k_1 < k_2 \leq r}& \left( (a_{k_2} - a_{k_1})^2 -1\right) \\
&\times \frac{\prod_{0 \leq i < j \leq r} \left(a_j - \delta^{k_1}_j - \delta^{k_2}_j - a_i + \delta^{k_1}_i + \delta^{k_2}_i \right)}{\prod_{i=0}^r \left( g-d+r+a_i-\delta^{k_1}_i-\delta^{k_2}_i \right)!}
\end{split}
\end{equation}
and consequently, one has
\begin{align}
\label{eq:munu}
\begin{split}
\mu_{g,d,\bm{a}} &= -\frac{n_{g,d,\bm{a}}}{2(g^2-1)} + \frac{1}{4 \binom{g-1}{2}}\sum_{i=0}^r n_{g-1,d,\bm{a}^i},\\
\nu_{g,d,\bm{a}} &= \frac{n_{g,d,\bm{a}}}{g(g^2-1)}
\end{split}
\end{align}
where $\bm{a}^i:=\left(a_0+1-\delta_0^i, \dots, a_r + 1 - \delta_r^i \right)$.
\section{The incidence divisors and pointed Brill-Noether divisors}
\label{sec:proofThmskWkBN}
Here we prove Theorem \ref{thm:kW} and Corollary \ref{cor:kBN}.
The argument involves the incidence divisor ${\mathbb{H}}^k_{g, 1}\subset \P\mathbb{E}^k_{g,1}$ from \eqref{eq:incdiv}.
The closure $\overline{\mathbb{H}}^k_{g, 1}$ in $\P\mathbb{E}^k_{g,1}$ is described by the incidence variety compactification introduced in \cite{bcggm1, bcggm}.
The class of $\overline{\mathbb{H}}^k_{g, 1}$ is given by:
\begin{lemma}[{[\citefirst[\S1.6]{sauvagetcohomology}[\S4]{korotkin2019tau}}]
\label{lem:Z1class}
One has $\overline{\mathbb{H}}^k_{g, 1} \equiv k \,\psi -\eta$ in $\mathrm{Pic}\left( \P\mathbb{E}^k_{g,1}\right)$ for $k\geq 1$ and $g\geq 2$.
\end{lemma}
\noindent Here $\psi$ is the first Chern class of the cotangent line bundle at the marked point.
To prove Theorem \ref{thm:kW} and Corollary \ref{cor:kBN},
consider the forgetful morphisms:
\[
\begin{tikzcd}
& \P\mathbb{E}^k_{g,1} \arrow{dr}{\pi} \arrow{dl}[swap]{\varphi} \\
\overline{\mathcal{M}}_{g,1} && \P\mathbb{E}^k_g.
\end{tikzcd}
\]
The key step in the proof is the study of
the intersection in $\P\mathbb{E}^k_{g,1}$ of $\overline{\mathbb{H}}^k_{g, 1}$ and the pullbacks via $\varphi$ of pointed Brill-Noether divisors in $\overline{\mathcal{M}}_{g,1}$.
Specifically, we have:
\begin{proposition}
\label{prop:BNirreducible}
For $k\geq 1$, $g\geq 2$ and $\bm{a}: 0 \leq a_0 < \dots < a_r \leq d$ such that $\rho(g,r,d,\bm{a}) =-1$, one has
\[
\pi_*\left(\left[\overline{\mathbb{H}}^k_{g, 1}\right] \cdot \varphi^*\left[\overline{\mathcal{M}}^{\bm{a}}_{g,d}\right]\right) = \left[\overline{\H}^{\bm{a}}_{g,d}\right] \,\, \in\,\,\mathrm{Pic}\left( \P\mathbb{E}^k_{g} \right).
\]
\end{proposition}
Before proving Proposition \ref{prop:BNirreducible}, we show how this implies Theorem \ref{thm:kW} and Corollary \ref{cor:kBN}.
Recall the Weierstrass divisor class $\mathcal{W}_g$ from \eqref{eq:W}.
\begin{proof}[Proof of Theorem \ref{thm:kW}]
For $d=2g-2$, $r=g-1$, and \mbox{$\bm{a}=(0,1,\dots,g-2,g)$,} one has the specializations
$\overline{\H}^{\bm{a}}_{g,d} \equiv \overline{\mathbb{W}}^k_g $ in $\mathrm{Pic}\left(\P\mathbb{E}^k_{g} \right)$ and
$\overline{\mathcal{M}}^{\bm{a}}_{g,d} \equiv \mathcal{W}_g$ in $\mathrm{Pic}\left( \overline{\mathcal{M}}_{g,1} \right)$. Thus Proposition \ref{prop:BNirreducible} implies $\overline{\mathbb{W}}^k_g \equiv\pi_* \left( \left[\overline{\mathbb{H}}^k_{g, 1}\right]\cdot \varphi^*\,\mathcal{W}_g \right)$ in $\mathrm{Pic}\left( \P\mathbb{E}^k_{g} \right)$.
To compute this, consider first the intersection
\begin{multline*}
\left[\overline{\mathbb{H}}^k_{g, 1}\right]\cdot \varphi^*\,\mathcal{W}_g \\
= \left(k\,\psi -\eta\right) \left( {g+1 \choose 2} \psi - \lambda -\sum_{i=1}^{g-1}{g-i+1 \choose 2}\delta_i \right) \,\,\in\,\, A^2\left(\P\mathbb{E}^k_{g,1}\right),
\end{multline*}
where we used Lemma \ref{lem:Z1class}.
The push-forward via $\pi\colon \P\mathbb{E}^k_{g,1} \rightarrow\P\mathbb{E}^k_{g}$ is
\begin{multline*}
\pi_* \left(\left[\overline{\mathbb{H}}^k_{g, 1}\right]\cdot \varphi^*\,\mathcal{W}_g \right) = -g(g^2-1)\,\eta + k {g+1 \choose 2} \kappa_1 -k(2g-2)\,\lambda \\
-\sum_{i=1}^{\lfloor g/2 \rfloor} k\left( (2i-1) {g-i+1 \choose 2} +(2g-2i-1) {i+1 \choose 2} \right)\delta_i
\end{multline*}
in $\mathrm{Pic}\left( \P\mathbb{E}^k_{g} \right)$.
Here, we used
\begin{align*}
\kappa_1:=\pi_* \left(\psi^2\right), \qquad \pi_* \left(\psi\lambda\right)=(2g-2)\lambda, \qquad \pi_* \left(\psi\eta\right)=(2g-2)\eta, \\
\pi_* \left(\psi\delta_i\right)=(2i-1)\delta_i \,\,\,\mbox{and}\,\,\, \pi_* \left(\psi\delta_{g-i}\right)=(2g-2i-1)\delta_i \,\,\,\mbox{for $1\leq i \leq \lfloor g/2 \rfloor$.}
\end{align*}
Mumford's formula $\kappa_1=12 \lambda -\sum_{i=0}^{\lfloor g/2 \rfloor} \delta_i$ and simplifying yield
\begin{align*}
\pi_* \left(\left[ \overline{\mathbb{H}}^k_{g, 1}\right]\cdot \varphi^*\,\mathcal{W}_g \right) =& -g(g^2-1)\,\eta + k(6g^2+4g+2)\,\lambda - k{g+1 \choose 2}\delta_0 \\
&-\sum_{i=1}^{\lfloor g/2 \rfloor} k(g+3)i(g-i)\,\delta_i.
\end{align*}
The statement follows.
\end{proof}
\begin{proof}[Proof of Corollary \ref{cor:kBN}]
From Proposition \ref{prop:BNirreducible}, Lemma \ref{lem:Z1class}, and \eqref{eq:muBNnuW}, one has
\begin{align*}
\left[\overline{\H}^{\bm{a}}_{g,d}\right]
&=\pi_*\left( \left[\overline{\mathbb{H}}^k_{g, 1} \right]\cdot \varphi^*\left[ \overline{\mathcal{M}}^{\bm{a}}_{g,d} \right]\right) \\
&= \pi_*\big(\left(k\,\psi-\eta\right)\left(\mu_{d,g,\bm{a}}\, \mathcal{BN}_g+\nu_{d,g,\bm{a}}\,\mathcal{W}_g \right) \big)\\
&= k(2g-2)\,\mu_{d,g,\bm{a}}\, \mathcal{BN}_g+\nu_{d,g,\bm{a}}\,\pi_* \left(\left[\overline{\mathbb{H}}^k_{g, 1}\right]\cdot \varphi^*\,\mathcal{W}_g \right)
\end{align*}
in $\mathrm{Pic}\left( \P\mathbb{E}^k_{g} \right)$.
Again by Proposition \ref{prop:BNirreducible}, one has $\pi_* \left( \left[\overline{\mathbb{H}}^k_{g, 1}\right]\cdot \varphi^*\,\mathcal{W}_g \right)\equiv \overline{\mathbb{W}}^k_g$ in $\mathrm{Pic}\left( \P\mathbb{E}^k_{g} \right)$, hence the statement.
\end{proof}
Expanding the formula in Corollary \ref{cor:kBN}, we deduce:
\begin{corollary}
\label{cor:expcoeffBNdiv}
For $k\geq 1$, $g \geq 2$, and
$\bm{a}: 0 \leq a_0 < \dots < a_r \leq d$ such that $\rho(g,r,d,\bm{a}) =-1$,
one has
\[
\left[\overline{\H}^{\bm{a}}_{g,d}\right] = c_\eta\,\eta + c_\lambda\,\lambda - \sum_{i=0}^{\lfloor g/2 \rfloor}c_i\,\delta_i \,\in\,\mathrm{Pic}\left(\P\mathbb{E}^k_{g} \right),
\]
where
\begin{align*}
c_\eta &= -g(g^2-1)\,\nu_{g,d,\bm{a}}, \\
c_\lambda &= 2(g-1)(g+3)k\,\mu_{g,d,\bm{a}} + 2(3g^2 + 2g +1)k\,\nu_{g,d,\bm{a}}, \\
c_0 &= \frac{g^2-1}{3}k\,\mu_{g,d,\bm{a}} + \frac{g(g+1)}{2}k\,\nu_{g,d,\bm{a}}, \\
c_i &= 2i(g-i)(g-1)k\,\mu_{g,d,\bm{a}} + i(g-i)(g+3)k\,\nu_{g,d,\bm{a}} \quad \mbox{for $i\geq 1$,}
\end{align*}
and $\mu_{g,d,\bm{a}}$ and $\nu_{g,d,\bm{a}}$ are given by \eqref{eq:munu}.
\end{corollary}
We now turn to the proof of Proposition \ref{prop:BNirreducible}.
\begin{proof}[Proof of Proposition \ref{prop:BNirreducible}]
By definition, the locus ${\H}^{\bm{a}}_{g,d}$ is the only component of $\pi\left(\overline{\mathbb{H}}^k_{g, 1} \cap \varphi^{-1}\left(\overline{\mathcal{M}}^{\bm{a}}_{g,d}\right)\right)$ in the restriction of $\P\mathbb{E}^k_{g}$ over the locus of \textit{smooth} curves $\mathcal{M}_g$. Also, $\pi\colon {\mathbb{H}}^k_{g, 1} \cap {\mathcal{M}}^{\bm{a}}_{g,d}\rightarrow {\H}^{\bm{a}}_{g,d}$ is generically of degree one. This implies the statement over $\mathcal{M}_g$.
Using the exact sequence
\[
\mathrm{Pic}\left(\P\mathbb{E}^k_{g}\big|_{\overline{\mathcal{M}}_g\setminus \mathcal{M}_g} \right) \rightarrow \mathrm{Pic}\left(\P\mathbb{E}^k_{g} \right) \rightarrow \mathrm{Pic}\left(\P\mathbb{E}^k_{g}\big|_{\mathcal{M}_g} \right) \rightarrow 0,
\]
we deduce that one has
\begin{equation}
\label{eq:E}
\pi_*\left(\left[\overline{\mathbb{H}}^k_{g, 1}\right] \cdot \varphi^* \left[\overline{\mathcal{M}}^{\bm{a}}_{g,d}\right]\right) = \left[\overline{\H}^{\bm{a}}_{g,d}\right] +E \qquad \in\mathrm{Pic}\left( \P\mathbb{E}^k_{g} \right)
\end{equation}
for some cycle $E$ which is either zero, or supported only over the divisor $\overline{\mathcal{M}}_g\setminus \mathcal{M}_g$ of singular curves.
Write
\[
\P\mathbb{E}^k_{g}\big|_{\overline{\mathcal{M}}_g\setminus \mathcal{M}_g}=\Delta_0\cup \cdots\cup\Delta_{\lfloor g/2 \rfloor},
\]
where $\Delta_0$ is the divisor whose general element consists of an irreducible nodal curve together with the class of a general stable $k$-differential, and $\Delta_i$ is the divisor
whose general element consists of a reducible nodal curve with two components of genera $i$ and $g-i$ together with the class of a general stable $k$-differential, for $1 \leq i \leq \lfloor g/2 \rfloor$.
To show that $E=0$,
it is enough to argue that the preimage under $\pi$ of a \textit{general} element of $\Delta_i$, for $0 \leq i \leq \lfloor g/2 \rfloor$, is disjoint from $\overline{\mathbb{H}}^k_{g, 1} \cap \varphi^{-1}\left(\overline{\mathcal{M}}^{\bm{a}}_{g,d}\right)$.
First, consider the case $i > 0$. Let $(X, \mu)$ be a general element of $\Delta_i$, i.e., $X$ is a nodal curve obtained by identifying the marked points of two \textit{general} pointed curves $(C_1, Q_1)$ and $(C_2, Q_2)$ of genus $i$ and $g-i$, respectively, and $\mu$ is the class of a \textit{general} stable $k$-differential on $X$.
There are finitely many points on $X$ which are limits of a Brill-Noether special point $P$ on a nearby smooth curve $C$ such that $\bm{a}^{\ell}(P) \geq \bm{a}$ for some $\ell\in G^r_d(C)$. Such limit points are away from the node of $X$. That is, the stable pointed curve obtained when a marked point in $X$ collides with the node, thus creating a rational bridge containing the marked point, is not a limit of Brill-Noether special pointed curves \cite[Theorem (1.1)]{MR910206}.
In particular, the codimension-two locus in $\pi^{-1}\left( \Delta_i\right)\subset \P\mathbb{E}^k_{g,1}$ whose general element consists of a curve with a pointed rational bridge and general stable $k$-differential is not a component of the intersection $\overline{\mathbb{H}}^k_{g, 1} \cap \varphi^{-1}\left( \overline{\mathcal{M}}^{\bm{a}}_{g,d}\right)$.
Moreover, since $\mu$ is general, the zeros of $\mu$ avoid the limits in $X$ of Brill-Noether special points.
Indeed, $\mu$ is a general element of
\begin{equation*}
\P \left(H^0\left(\omega^{\otimes k}_{C_1}\left(kQ_1\right)\right) \times H^0\left(\omega^{\otimes k}_{C_2}\left(kQ_2\right)\right)\right)
\end{equation*}
satisfying the \textit{$k$-residue condition}.
That is, $\mu$ is the class of a pair of general $k$-differentials $\mu_1$ and $\mu_2$ on the two components $C_1$ and $C_2$ with poles of order $k$ at $Q_1$ and $Q_2$, respectively, and with appropriate scaling such that
\[
\mathrm{Res}^k_{Q_1}(\mu_1) = (-1)^k \,\mathrm{Res}^k_{Q_2}(\mu_2).
\]
Here, since $\mu_i$ has a pole of order $k$ at $Q_i$, the $k$-residue $\mathrm{Res}^k_{Q_i}(\mu_i)$ is computed as the coefficient of $t_i^{-k}(dt_i)^k$ in the Laurent series expansion of $\mu_i$ at $Q_i$, for a formal coordinate $t_i$ at $Q_i$, with $i=1,2$.
Hence, the zeros of $\mu$, equal to the union of the zeros of $\mu_1$ and $\mu_2$, are generically away from the limits in $X$ of Brill-Noether special points on nearby smooth curves.
Finally, consider the case $i=0$.
Let $(X,\mu)$ be a general element of $\Delta_0$ in $\P\mathbb{E}^k_{g}$, i.e., $X$ is an irreducible nodal curve and $\mu$ is the class of a general stable $k$-differential on $X$.
Pulling back via the normalization map $\widetilde{X}\rightarrow X$, a general stable $k$-differential on $X$ is identified with a general $k$-differential on $\widetilde{X}$ having poles of
order $k$ at the two preimages of the node.
First, consider the stable pointed curve $(X_0, P_0)$ obtained when a marked point in $X$ collides with the node, thus creating a rational component containing the marked point $P_0$ and meeting the rest of the curve in two points.
We claim that the codimension-two locus in $\pi^{-1}\left( \Delta_0\right)\subset \P\mathbb{E}^k_{g,1}$ whose general element consists of such a pointed curve $(X_0, P_0)$ together with
a general stable $k$-differential $\mu_0$ is not a component of the intersection $\overline{\mathbb{H}}^k_{g, 1} \cap \varphi^{-1}\left( \overline{\mathcal{M}}^{\bm{a}}_{g,d}\right)$.
Indeed, the restriction of $\mu_0$ to each component of $X_0$ has
poles of order $k$ at the two nodes. It follows that the restriction of $\mu_0$ to the rational component has no zeros for degree reasons, hence $\mu_0$ does not vanish at ~$P_0$. This implies that $(X_0, P_0, \mu_0)$ does not lie in $\overline{\mathbb{H}}^k_{g, 1}$, hence the claim.
It remains to consider marked points in $X$ that are away from the node.
For this, one argues similarly to the case $i>0$, as the zeros of $\mu$ generically avoid the finitely many limits in $X$ of Brill-Noether special points on nearby smooth curves.
It follows that $E=0$ in \eqref{eq:E}, hence the statement.
\end{proof}
\section{Test curves}
\label{sec:testfamilies}
We consider here the intersections of the loci $\overline{\H}^{\bm{a}}_{g,d}$ with various curves in $\P\mathbb{E}^k_{g}$. These will serve a double purpose: We provide explicit examples of restrictions of the loci $\overline{\H}^{\bm{a}}_{g,d}$ to families of curves with $k$-differentials, and the resulting computations provide a second independent proof of Theorem~\ref{thm:kW} and Corollary~\ref{cor:expcoeffBNdiv}.
The classes $\eta, \lambda$, and $\delta_i$ for $i=0,\dots, \lfloor g/2 \rfloor$ freely generate $\mathrm{Pic}\left(\P\mathbb{E}^k_g \right)$ for $g\geq 3$, while for $g=2$, $\mathrm{Pic}\left(\P\mathbb{E}^k_2 \right)$ is generated by $\eta, \lambda,\delta_0,\delta_1$ modulo the relation $10\lambda = \delta_0 + 2\delta_1$.
Hence, given $g\geq 2$ and $\bm{a}: 0 \leq a_0 < \dots < a_r \leq d$ such that $\rho(g, r, d, \bm{a}) =-1$,
we can write
\begin{equation}
\label{eq:Hcoeff}
\left[ \overline{\H}^{\bm{a}}_{g,d} \right] = c_\eta \,\eta + c_\lambda \,\lambda - \sum_{i=0}^{\lfloor g/2 \rfloor} c_i \,\delta_i \,\,\in\,\, \mathrm{Pic}_{\mathbb{Q}}\left(\P\mathbb{E}^k_{g}\right),\quad \mbox{for some $c_\eta, c_\lambda, c_i\in \mathbb{Q}$}.
\end{equation}
All the coefficients $c_\eta, c_\lambda, c_i$, with $0\leq i\leq \lfloor g/2 \rfloor$, are given by Theorem \ref{thm:kW} and Corollary \ref{cor:expcoeffBNdiv}. The intersections with the following families allow one to independently recover these coefficients.
\subsection{A pencil of $k$-canonical divisors on a general curve}
\label{sec:ceta}
Let $C$ be a general genus $g$ curve $k$-canonically embedded in $\P^{N}$, where $N=g-1$ if $k=1$, or $N=(g-1)(2k-1) -1$ if $k\geq 2$. Let $\Lambda \cong \P^{N-3}$ be a fixed general subspace in $\P^{N}$, and consider the one-dimensional family of hyperplanes in $\P^N$ containing $\Lambda$.
As every such hyperplane cuts a $k$-canonical divisor on $C$,
this family gives rise to a pencil of $k$-canonical divisors on $C$, hence a pencil $A$ in $\P\mathbb{E}^k_g$.
One has
\[
A \cdot \eta = -1, \quad A \cdot \lambda = 0, \quad A \cdot \delta_i = 0, \quad \mbox{for all $i$.}
\]
The pencil $A$ intersects $\overline{\H}^{\bm{a}}_{g,d}$ transversely along the Brill-Noether special points $P$ in $C$ such that $\bm{a}^\ell(P) \geq \bm{a}$ for some $\ell\in G^r_d(C)$, hence one has
\[
A \cdot \left[\overline{\H}^{\bm{a}}_{g,d}\right] = n_{g, d, \bm{a}}
\]
where $n_{g,d,\bm{a}}$ is the number defined in \eqref{eq:ngdadef}. Combining with \eqref{eq:munu}, this gives
\begin{equation}
\label{eq:ceta}
c_\eta = -g(g^2-1)\,\nu_{g,d,\bm{a}}.
\end{equation}
\subsection{Curves on K3 surfaces}
\label{sec:curvesonK3}
Next, we consider a Lefschetz pencil of curves of genus $g\geq 3$ lying on a general K3 surface $S$ of degree $2g-2$ in $\mathbb{P}^g$.
Let \mbox{$\mathcal{X}\rightarrow S$} be the blow-up at the $2g-2$ base-points of the pencil, and let \mbox{$\pi\colon\mathcal{X}\rightarrow \mathbb{P}^1$} be the corresponding family of curves.
Fix $k$ general sections in $S$ and let $\Sigma_1, \dots, \Sigma_k$ be their proper transform in $\mathcal{X}$. Since $S$ has canonical sections, the cycle $\Sigma_1+\cdots+\Sigma_k$ on $\mathcal{X}$ restricts to a $k$-canonical divisor on each fiber of $\pi$, hence this gives rise to a pencil $\tau\colon\mathbb{P}^1\rightarrow \P\mathbb{E}^k_{g}$.
The intersections with the generators are
\[
\tau^* \eta = k, \qquad
\tau^* \lambda = g+1, \qquad
\tau^* \delta_0 = 6g+18, \qquad
\tau^* \delta_i = 0, \qquad \mbox{for $i>0$.}
\]
The intersections with $\lambda$ and the $\delta_i$ are classical \cite{cukierman1993curves, MR2123229}. The degree of $\eta$ can be computed by intersecting the relation $\omega_{\mathcal{X}/\P^1}^{\otimes k} = \pi^*\eta \otimes \mathcal{O}_\mathcal{X}(\Sigma_1 + \cdots + \Sigma_k)$ valid on $\mathcal{X}$
with the class of one of the $2g-2$ exceptional divisors (this is as in \cite{chen2018positivity}, see also \cite[Example 3.2]{gheorghita2018effective} for a similar computation).
To compute the intersection of the pencil $\tau$ with the divisor class $\left[\overline{\H}^{\bm{a}}_{g,d}\right]$, we start from the locus of corresponding Brill-Noether special points in $\mathcal{X}$.
After identifying $\overline{\mathcal{M}}_{g,1} \rightarrow \overline{\mathcal{M}}_{g}$ with the universal curve over $\overline{\mathcal{M}}_{g}$,
one has a moduli map $\mathcal{X}\rightarrow \overline{\mathcal{M}}_{g,1}$. The pull-back of the divisor class $\left[\overline{\mathcal{M}}_{g,d}^{\bm{a}}\right]$ on $\overline{\mathcal{M}}_{g,1}$ from \eqref{eq:muBNnuW} via such map is
\begin{align}
\label{eq:Xagd}
\begin{split}
\left[\mathcal{X}^{\bm{a}}_{g,d}\right]:= &\nu_{g,d,\bm{a}}\,{g+1 \choose 2}\, c_1\left(\omega_{\mathcal{X}/\mathbb{P}^1}\right) \\
& +(\mu_{g,d,\bm{a}}\,(g+3) -\nu_{g,d,\bm{a}})\, \lambda - \mu_{g,d,\bm{a}}\, \frac{g+1}{6}\, \delta_0.
\end{split}
\end{align}
It follows that
\[
\tau^* \left[\overline{\H}^{\bm{a}}_{g,d}\right] = \left[\mathcal{X}^{\bm{a}}_{g,d}\right]\cdot \left([\Sigma_1]+ \cdots+[\Sigma_k]\right).
\]
By the adjunction formula, one has $\Sigma_j \equiv K_f \equiv f+ K_{\mathcal{X}}\equiv f+\sum_{i=1}^{2g-2}E_i$, for each $j=1,\dots,k$,
where $f$ is a fiber of $\pi$, and $E_i$ are the exceptional curves on~$\mathcal{X}$.
This gives
\begin{align}
\label{eq:K3Hagd}
\tau^* \left[\overline{\H}^{\bm{a}}_{g,d}\right] = \left[\mathcal{X}^{\bm{a}}_{g,d}\right]\cdot k\left[ f+\sum_{i=1}^{2g-2}E_i\right] =2k(g+1)(g-1)^2\,\nu_{g,d,\bm{a}}.
\end{align}
Here we used that the nonzero intersections are given by
\begin{align*}
c_1\left(\omega_{\mathcal{X}/\mathbb{P}^1}\right) \cdot f = 2g-2, \quad
c_1\left(\omega_{\mathcal{X}/\mathbb{P}^1}\right) \cdot E_i=1, \quad
\lambda\cdot E_i = \tau^* \lambda= g+1, \\
\delta_0\cdot E_i = \tau^* \delta_0= 6g+18, \mbox{ for each $i=1,\dots,2g-2$}.
\end{align*}
Note how \eqref{eq:K3Hagd} is independent of $\mu_{g,d,\bm{a}}$, as the pencil has zero intersection with the Brill-Noether class $\mathcal{BN}_g$ \cite{MR852158}.
It follows that the intersection with \eqref{eq:Hcoeff} gives
\begin{equation}
\label{eq:K3}
2k(g+1)(g-1)^2\,\nu_{g,d,\bm{a}} = k\,c_\eta +(g+1) \, c_\lambda - (6g+18)\,c_0.
\end{equation}
\subsection{A Du Val pencil}
We consider here a pencil of Du Val curves of genus $g\geq 2$, as introduced in \cite{arbarello2016explicit}.
We briefly review the definition. Let $S'$ be the blow-up of $\P^2$ at nine general points $P_1,\dots, P_9$ (the generality condition is explicitly given in \cite{arbarello2016explicit}), and consider the linear system on $S'$
\[
L_g:=\left|3g \ell -g\left( E_ 1 + \cdots + E_ 8\right) -(g-1) E_9 \right|
\]
where $\ell$ is the proper transform of a line in $\P^2$, and $E_1,\dots, E_9$ are the exceptional curves in $S'$.
There exists a unique curve $J'\equiv 3\ell -(E_ 1 + \cdots + E_ 9)$ in $S'$ corresponding to the unique smooth plane cubic curve passing through the points $P_1,\dots,P_9$,
and for each $C'\in L_g$, one has $C'\cdot J'=1$. It follows that $C'\cap J'=:\{P_{10}\}$ is independent of $C'$, and thus $P_{10}$ is a base-point of~$L_g$.
Let $S\rightarrow S'$ be the blow-up at $P_{10}$. Denote still by $E_1,\dots, E_9$ the inverse image in $S$ of the exceptional curves in $S'$, and let $E_{10}$ be the exceptional curve in $S$ over $P_{10}$. Let $J$ and $C$ be the proper transforms of $J'$ and $C'\in L_g$. One has
\begin{align*}
-K_S &\equiv J \equiv 3\ell -(E_ 1 + \cdots + E_ {10}),\\
C &\equiv 3g \ell -g\left( E_ 1 + \cdots + E_ 8\right) -(g-1) E_9 - E_{10}.
\end{align*}
The linear system $|C|$ is base-point free and gives rise to a map \mbox{$S\rightarrow \overline{S}\subset \P^g$} contracting $J$ into an elliptic singularity. The surface $\overline{S}$ has canonical sections and deforms in $\P^g$ to a nonsingular K3 surface of degree $2g-2$ \cite{arbarello2015hyperplane}.
Consider a Lefschetz pencil in $|C|$. Let $\mathcal{X}\rightarrow S$ be the blow-up at the $C^2=2g-2$ base-points of the pencil, and let $\pi\colon \mathcal{X}\rightarrow \P^1$ be the corresponding family of curves.
As in \S\ref{sec:curvesonK3}, consider the proper transforms $\Sigma_1, \dots, \Sigma_k$ in $\mathcal{X}$ of $k$ general sections in $S$. Since $\overline{S}$ has canonical sections, the cycle $\Sigma_1+\cdots+\Sigma_k$ cuts $k$-canonical divisors on the fibers of $\pi$, hence this gives rise to a pencil $\tau\colon\mathbb{P}^1\rightarrow \P\mathbb{E}^k_{g}$.
The intersections with the generators are
\begin{align*}
\tau^* \eta &= k, &
\tau^* \lambda &= g, &
\tau^* \delta_0 &= 6(g+1), \\
&&\tau^* \delta_1 &= 1, &
\tau^* \delta_i &= 0, \qquad\mbox{for $i>1$.}
\end{align*}
The intersections with $\lambda$ and the $\delta_i$ are computed in \cite{arbarello2016explicit}, and the intersection with $\eta$ is computed as in \S\ref{sec:curvesonK3}.
We note that there is a unique fiber $J+D$ of $\pi$ contributing to $\tau^* \delta_1$ and consisting of the elliptic curve $J$ and a $g-1$ curve
\[
D\equiv 3(g-1) \ell -(g-1)\left( E_ 1 + \cdots + E_ 8\right) -(g-2) E_9
\]
meeting in one point.
Consider the moduli map $\gamma\colon\mathcal{X}\rightarrow \overline{\mathcal{M}}_{g,1}$. Since
$\gamma^*\delta_1 =J$ and $\gamma^*\delta_{g-1} =D$,
the pull-back via $\gamma$ of the divisor class $\left[\overline{\mathcal{M}}_{g,d}^{\bm{a}}\right]$ expressed by \eqref{eq:muBNnuW} is
\begin{align*}
\left[\mathcal{X}^{\bm{a}}_{g,d}\right]:= &\nu_{g,d,\bm{a}}\,{g+1\choose 2}\, c_1\left(\omega_{\mathcal{X}/\mathbb{P}^1}\right)
+(\mu_{g,d,\bm{a}}\,(g+3) -\nu_{g,d,\bm{a}})\, \lambda \\
&- \mu_{g,d,\bm{a}}\, \frac{g+1}{6}\, \delta_0
-\left(\mu_{g,d,\bm{a}}\,(g-1)+ \nu_{g,d,\bm{a}}\,{g \choose 2}\right)J\\
& -\left(\mu_{g,d,\bm{a}}\,(g-1) + \nu_{g,d,\bm{a}}\right)D.
\end{align*}
The intersection with the divisor class $\left[\overline{\H}^{\bm{a}}_{g,d}\right]$ is
\[
\tau^* \left[\overline{\H}^{\bm{a}}_{g,d}\right] = \left[\mathcal{X}^{\bm{a}}_{g,d}\right]\cdot \left([\Sigma_1]+ \cdots+[\Sigma_k]\right).
\]
By the adjunction formula, one has
\[
\Sigma_j \equiv K_f \equiv f+ K_{\mathcal{X}}\equiv f+K_S+\sum_{i=1}^{2g-2}F_i \equiv f-J+\sum_{i=1}^{2g-2}F_i,
\]
for each $j=1,\dots,k$,
where $f$ is a fiber of $\pi$, and $F_i$ are the exceptional curves over the base-points of the pencil in~$S$.
This gives
\begin{align}
\label{eq:DVHagd}
\tau^* \left[\overline{\H}^{\bm{a}}_{g,d}\right] = \left[\mathcal{X}^{\bm{a}}_{g,d}\right]\cdot k\left[ f-J+\sum_{i=1}^{2g-2}F_i\right] = k(2g-3)(g^2-1) \,\nu_{g,d,\bm{a}}.
\end{align}
The result is independent of $\mu_{g,d,\bm{a}}$, and indeed the pencil has zero intersection with the class $\mathcal{BN}_g$ \cite{arbarello2016explicit}.
In \eqref{eq:DVHagd}, we used that the nonzero intersections are given by
\begin{align*}
c_1\left(\omega_{\mathcal{X}/\mathbb{P}^1}\right) \cdot f = 2g-2, \quad
c_1\left(\omega_{\mathcal{X}/\mathbb{P}^1}\right) \cdot F_i=1, \quad
\lambda\cdot F_i = \tau^* \lambda= g, \\
\delta_0\cdot F_i = \tau^* \delta_0= 6(g+1),\quad
D\cdot F_i = \tau^* \delta_{1}= 1, \quad
\mbox{for $i=1,\dots,2g-2$}.
\end{align*}
For the intersection $D\cdot F_i$, note that $C\cdot D=2g-2$, hence the restrictions of all base-points of the pencil to the fiber $J+D$ of $\pi$ are in $D$.
Intersecting with \eqref{eq:Hcoeff}, we deduce the following relation on the coefficients of the class $\left[\overline{\H}^{\bm{a}}_{g,d}\right]$:
\begin{equation}
\label{eq:DV}
k(2g-3)(g^2-1) \,\nu_{g,d,\bm{a}} = k\,c_\eta +g \, c_\lambda - 6(g+1)\,c_0 - \, c_1.
\end{equation}
\subsection{A pencil of hyperelliptic curves}
\label{sec:hyppencil}
We consider here a pencil of hyperelliptic curves, following \cite[pg.~361-363]{MR1070600}. Let $S\rightarrow \P^2$ be a double cover branched over a general smooth curve of degree $2g+2$ in $\P^2$. This gives a two-dimensional family of hyperelliptic curves of genus $g$. Consider a general pencil of hyperelliptic curves in $S$.
Such a pencil has two base points \cite[pg.~361-363]{MR1070600}.
Let $\pi\colon\mathcal{X} \rightarrow \P^1$ be the corresponding family of curves obtained by blowing-up the two base points.
A choice of $k(g-1)$ lines in $\P^2$ gives a $k$-canonical divisor on each curve in the family after pulling back to $\mathcal{X}$. This gives rise to a pencil in $\P\mathbb{E}^k_g$. One has
\begin{align*}
\deg\eta &= k, & \deg\lambda &=\frac{g(g+1)}{2}, \\
\deg\delta_0& = 2(g+1)(2g+1), & \deg\delta_i&=0 &\mbox{for $i\geq 1$.}
\end{align*}
The intersections with $\lambda$ and $\delta_i$ are computed in \cite[pg.~361-363]{MR1070600}, and the intersection with $\eta$ can be computed as in \S \ref{sec:curvesonK3}.
Furthermore, let $\Sigma_i$, for $i=1,\dots, k(g-1)$, be the pullbacks to $\mathcal{X}$ of the fixed $k(g-1)$ lines in $\P^2$, let $f$ be a fiber of $\pi$, and $E_1,E_2$ be the two exceptional curves. By the adjunction formula, one has $\Sigma_i \equiv K_f \equiv f+E_1 + E_2$ on $\mathcal{X}$.
The pointed Brill-Noether divisor class $\big[\mathcal{X}^{\bm{a}}_{g,d}\big]$ on $\mathcal{X}$ is as in \eqref{eq:Xagd}.
Similarly to \S \ref{sec:curvesonK3}, the intersection of the pencil $\pi\colon\mathcal{X} \rightarrow \P^1$ with the divisor class $\left[\overline{\H}^{\bm{a}}_{g,d}\right]$ equals
\begin{align*}
\big[\mathcal{X}^{\bm{a}}_{g,d}\big] \cdot & \sum_{i=1}^{k(g-1)} [\Sigma_i ]= \big[\mathcal{X}^{\bm{a}}_{g,d}\big] \cdot k(g-1)(f+E_1 + E_2) \\
&= kg(g+1)(g-1)^2 \,\nu_{g,d,\bm{a}} - \frac{1}{3}k(g+1)(g-1)^2(g-2)\,\mu_{g,d,\bm{a}}.
\end{align*}
Here we used that the nonzero intersections are given by
\begin{align*}
c_1\left(\omega_{\mathcal{X}/\mathbb{P}^1}\right) \cdot f = 2g-2,\quad
c_1\left(\omega_{\mathcal{X}/\mathbb{P}^1}\right) \cdot E_i=1, \quad
\lambda\cdot E_i =\frac{g(g+1)}{2}, \\
\delta_0\cdot E_i = 2(g+1)(2g+1), \mbox{ for each $i=1,2$.}
\end{align*}
It follows that the coefficients of the class $\left[\overline{\H}^{\bm{a}}_{g,d}\right]$ in \eqref{eq:Hcoeff} satisfy
\begin{equation}
\begin{split}
\label{eq:hyp}
kg(g+1)(g-1)^2 \,\nu_{g,d,\bm{a}} -& \frac{1}{3}k(g+1)(g-1)^2(g-2)\,\mu_{g,d,\bm{a}} \\
&= k\,c_\eta +\frac{g(g+1)}{2}\,c_\lambda - 2(g+1)(2g+1)\,c_0.
\end{split}
\end{equation}
\subsection{Pull-back to $\overline{\mathcal{M}}_{0,g}$}
For $g\geq 4$, we define here a map
\[
\xi\colon \overline{\mathcal{M}}_{0,g} \rightarrow \P\mathbb{E}^k_g.
\]
Consider an elliptic curve $J$ with distinguished point $Q\in J$, and let $\mu_J$ be a general $k$-differential on $J$ with a pole of order $k-1$ at $Q$ and regular elsewhere.
Specifically, given a degree $d>0$, it will be enough to select $\mu_J\in H^0\left(J, \mathscr{O}_J((k-1)Q)\right)$ which does not vanish at the finitely many points $P\in J$ such that $\mathscr{O}_J(P-Q)$ is an $e$-torsion point in $\mathrm{Pic}^0(J)$ for some $e\leq d$.
Then $\xi$ is defined by mapping
a stable $g$-pointed rational curve to the stable curve of genus $g$ obtained by attaching $g$ copies of $J$ at the $g$ marked points, and assigning the stable $k$-differential which restricts as $\mu_J$ on each copy of $J$ and is zero on all rational components.
Let $\epsilon_i$ be the class of the divisor in $\overline{\mathcal{M}}_{0,g}$ whose generic element consists of two rational components, one of them containing exactly $i$ marked points, for $2\leq i \leq \lfloor g/2 \rfloor$. One has $\xi^* \eta =0$, and from \cite{MR910206}:
\[
\xi^* \lambda =0,
\quad
\xi^* \delta_1 =-\sum_{i=2}^{\lfloor g/2 \rfloor} \frac{i(g-i)}{g-1}\epsilon_i,
\quad
\xi^* \delta_i=\epsilon_i, \,\, \mbox{ for $2\leq i \leq \lfloor g/2 \rfloor$}.
\]
\begin{lemma}
For $g\geq 4$ and $\bm{a}: 0 \leq a_0 < \dots < a_r \leq d$ such that $\rho(g, r, d, \bm{a}) <0$,
the locus $\overline{\H}^{\bm{a}}_{g,d}$ in $\P\mathbb{E}^k_g$ is disjoint from the image of $\xi$.
In particular, when $\rho(g, r, d, \bm{a}) =-1$, it follows that
$\xi^*\left[\overline{\H}^{\bm{a}}_{g,d}\right]=0$ in $\mathrm{Pic}\left( \overline{\mathcal{M}}_{0,g} \right)$, hence one has
\begin{equation}
\label{eq:ci}
c_i=\frac{i(g-i)}{g-1}\,c_1, \qquad \mbox{for $2\leq i \leq \lfloor g/2 \rfloor$}.
\end{equation}
\end{lemma}
\begin{proof}
Let $(X,\mu)$ in $\P\mathbb{E}^k_g$ be an arbitrary element in the image of $\xi$.
For $(X,\mu)$ to be in $\overline{\H}^{\bm{a}}_{g,d}$, there need to be a twisted $k$-differential of type $\mu$ (in the sense of \cite{bcggm1, bcggm}) vanishing at a limit in $X$ of a Brill-Noether special point of some linear series of type $\mathfrak{g}^r_d$ on a nearby smooth curve.
Since the restriction $\mu_J$ of $\mu$ to each elliptic tail of $X$ is generically nonzero, all twisted $k$-differentials of type $\mu$
restrict as $\mu_J$ on each elliptic tail.
Points on rational components of $X$ are always Brill-Noether general \cite{EH-1}, and Brill-Noether special points of linear series of type $\mathfrak{g}^r_d$
specialize to points $P$ on elliptic tails of $X$ such that $\mathscr{O}_J(P-Q)$ is an $e$-torsion point in $\mathrm{Pic}^0(J)$ for some $e\leq d$ [\citefirst{MR846932}[Lemma 2.1]{osserman2014simple}.
By assumption, $\mu_J$ vanishes away from such points, hence
the image of the map $\xi$ is disjoint from the locus $\overline{\H}^{\bm{a}}_{g,d}$ in $\P\mathbb{E}^k_g$.
\end{proof}
\subsection{Second proof of Theorem \ref{thm:kW} and Corollary \ref{cor:expcoeffBNdiv}}
Expanding the class $\left[\overline{\H}^{\bm{a}}_{g,d}\right]$ as in \eqref{eq:Hcoeff}, the relations in \eqref{eq:ceta}, \eqref{eq:K3}, \eqref{eq:DV}, \eqref{eq:hyp}, \eqref{eq:ci}
allow one to determine all the coefficients $c_\eta, c_\lambda, c_i$, with $0\leq i\leq \lfloor g/2 \rfloor$ modulo the required relation $10\lambda = \delta_0 + 2\delta_1$ in the case $g=2$.
\hfill$\square$
\section{Extremality and rigidity of the incidence divisor}
In this section we prove Theorem \ref{thm:Z1extremal}. To deduce extremality, we use:
\begin{lemma}[{\cite[Lemma 4.1]{MR3071469}}]
\label{lem:extremality}
Let $D$ be an irreducible effective divisor in a projective variety~$X$, and let $\mathscr{C}$ be a set of irreducible effective curves contained in $D$ such that $\bigcup_{C\in \mathscr{C}} C$ is Zariski dense in $D$. If for every curve $C$ in $\mathscr{C}$ one has
\[
C \cdot (D + B) \leq 0, \qquad \mbox{for a fixed big divisor class $B$ on $X$,}
\]
then $D$ is extremal in $\overline{\mathrm{Eff}}^1(X)$.
\end{lemma}
To review the notation, a divisor class on $X$ is \textit{big} if it lies in the interior of the pseudo-effective cone $\overline{\mathrm{Eff}}^1(X)$.
We emphasize that in Lemma \ref{lem:extremality}, the curves in $\mathscr{C}$ are not required to be moving in $D$. We apply Lemma \ref{lem:extremality} below to deduce the extremality of the divisor $\overline{\mathbb{H}}^k_{g, 1}$ using a set $\mathscr{C}$ of Teichm\"uller curves in $\overline{\mathbb{H}}^k_{g, 1}$ available for $k=1,2$. While such Teichm\"uller curves are not moving in $\overline{\mathbb{H}}^k_{g, 1}$, their union is indeed dense in $\overline{\mathbb{H}}^k_{g, 1}$.
\begin{proof}[Proof of Theorem \ref{thm:Z1extremal}]
First, we show the extremality for $k=1$. Let $\mathcal{C}\subset \P\mathbb{E}^1_{g,1}$ be the closure of a Teichm\"uller curve generated by some element $(C, P, \mu)$
in $\mathbb{H}^1_{g, 1}$ such that $C$ is smooth and $\mu$ is a canonical divisor on $C$ supported at distinct points, including $P$.
By varying $(C, P, \mu)$,
the collection $\mathscr{C}$ of such Teichm\"uller curves $\mathcal{C}$
is dense in $\overline{\mathbb{H}}^1_{g, 1}$.
From \cite{MR3033521}, one has for each $\mathcal{C}\in \mathscr{C}$:
\begin{align}
\begin{split}
\label{eq:intersectionCTeich}
&\mathcal{C} \cdot \lambda = -\frac{\chi}{2}L, \qquad\qquad\qquad\qquad\qquad\,\,
\mathcal{C} \cdot \delta_0 = -\frac{3\,\chi}{2}(4L - g+1), \\
&\mathcal{C} \cdot \psi = 2\, \frac{\mathcal{C} \cdot \lambda - \mathcal{C} \cdot \delta_0/12}{g-1} = -\frac{\chi}{4}, \qquad
\mathcal{C} \cdot \delta_i = 0, \quad\mbox{ for $i \geq 1$,}
\end{split}
\end{align}
where $\chi$ is the Euler characteristic of $\mathcal{C}$ (a different sign convention is used for $\chi$ in \cite{MR3033521}), and $L$ is the sum of the top $g$ Lyapunov exponents of~$\mathcal{C}$ (we refer to \cite[\S 1.4]{eskin2014sum} for more on Lyapunov exponents).
Specifically, $\mathcal{C}$ is disjoint from $\delta_i$ for $i>0$ from \cite[Cor.~3.2]{MR3033521}.
For any Teichm\"uller curve, the intersections with $\lambda$ and $\delta_0$ are computed in terms of the area Siegel-Veech constant $c_{\rm{SV}}(\mathcal{C})$ of $\mathcal{C}$ \cite[\S 1.6]{eskin2014sum}: namely,
$\deg \delta_0 = -6\chi\, c_{\rm{SV}}(\mathcal{C})$ from \cite[Proof of Prop.~4.5]{MR3033521} and $\deg \lambda = \frac{L}{12\, c_{\rm{SV}}(\mathcal{C})} \deg \delta_0$ from \cite[Prop.~4.5]{MR3033521}.
Since $\pi(\mathcal{C})$ is generically contained in the principal stratum of $\P\mathbb{E}^1_g$ consisting of curves with canonical divisors supported at \textit{distinct} points, where $\pi\colon \P\mathbb{E}^1_{g,1}\rightarrow \P\mathbb{E}^1_g$ is the forgetful map, one has $c_{\rm{SV}}(\mathcal{C})=L-\frac{1}{4}(g-1)$ from [\citefirst[Thm.~1]{eskin2014sum}[Thm.~4.4]{MR3033521}, hence the intersections with $\lambda$ and $\delta_0$ in \eqref{eq:intersectionCTeich} follow. The intersection with $\psi$ follows from \cite[Prop.~4.8]{MR3033521} using that for the generator $(C,P,\mu)$ of $\mathcal{C}$ the support of the canonical divisor $\mu$ contains $P$ with multiplicity one.
Moreover, from [\citefirst{moller2006variations}[(15)]{MR3033521}, one has
\[
\mathcal{C} \cdot \eta = -\frac{\chi}{2}.
\]
Combining this with Lemma \ref{lem:Z1class}, we have
\[
\mathcal{C} \cdot \overline{\mathbb{H}}^1_{g, 1} = \mathcal{C} \cdot (\psi - \eta) = \frac{\chi}{4}.
\]
Since $\psi$ is ample on any nonconstant family not all of whose elements are singular \cite[Thm.~6.33]{MR1631825}, one has $\mathcal{C} \cdot \psi>0$.
We deduce $\chi < 0$ and thus $\mathcal{C} \cdot \overline{\mathbb{H}}^1_{g, 1} <0$. We conclude that $\mathcal{C}\subset \overline{\mathbb{H}}^1_{g, 1}$ for all $\mathcal{C}\in \mathscr{C}$.
To deduce the extremality of $\overline{\mathbb{H}}^1_{g, 1}$, we apply Lemma \ref{lem:extremality} for a certain big divisor class $B$ on $\P\mathbb{E}^1_{g,1}$.
Select an ample divisor class
\[
A := c_\eta \,\eta + c_\lambda \,\lambda + c_\psi \,\psi + \sum_{i=0}^{g-1} c_i\,\delta_i \,\,\in\,\, \mathrm{Pic}\left(\P\mathbb{E}^1_{g,1}\right),
\]
and define $B:=dA$ for a sufficiently small $d>0$ such that $\mathcal{C} \cdot (\overline{\mathbb{H}}^1_{g, 1} + B) \leq 0$ for all $\mathcal{C} \in \mathscr{C}$. Namely, let
\[
d := \inf_{\mathcal{C}\in \mathscr{C}} \Bigg\{ \frac{1}{2c_\eta + c_\psi -6(g-1)c_0 +2L(c_\lambda +12c_0)} \Bigg\}. \]
The expression in the brackets comes from solving for $d$ in $\mathcal{C} \cdot (\overline{\mathbb{H}}^1_{g, 1} + dA) = 0$ for a given Teichm\"uller curve, and it is positive since $\mathcal{C} \cdot \overline{\mathbb{H}}^1_{g, 1} <0$ and $\mathcal{C} \cdot A >0$. Furthermore, the value in the brackets depends only on the sum $L$ of the top $g$ Lyapunov exponents of $\mathcal{C}$. Since $0\leq L\leq g$ (as the top $g$ Lyapunov exponents are nonnegative and at most $1$ [\citefirst[\S 5]{kontsevich1997lyapunov}[\S 1.4]{eskin2014sum}), the infimum here is indeed positive, and thus $B=dA$ is big. The extremality of $\overline{\mathbb{H}}^1_{g, 1}$ follows by Lemma \ref{lem:extremality}.
\smallskip
Next, we show the extremality for $k=2$. The argument is similar to the previous case. Let $\mathcal{C}\subset \P\mathbb{E}^2_{g,1}$ be the closure of a Teichm\"uller curve generated by some $(C, P, \mu)$ in $\mathbb{H}^2_{g, 1}$ such that $C$ is smooth and $\mu$ is the class of a quadratic differential on $C$ vanishing at distinct points, including $P$, and $\mu$ is not the class of the square of an Abelian differential \cite[\S 4.1]{chen2014quadratic}. The collection $\mathscr{C}$ of such Teichm\"uller curves $\mathcal{C}$ is dense in $\overline{\mathbb{H}}^2_{g, 1}$.
From \cite[Prop.~4.2 and (2)]{chen2014quadratic}, one has
\begin{align}
\label{eq:intersectionCTeich2}
\mathcal{C} \cdot \lambda = -\frac{\chi}{36}\big(18\, c_{\rm{SV}}(\mathcal{C}) + 5(g-1)\big), \quad
\mathcal{C} \cdot \delta = -6\,\chi \, c_{\rm{SV}}(\mathcal{C}), \quad
\mathcal{C} \cdot \psi = -\frac{\chi}{3},
\end{align}
where, as before, $\chi$ and $c_{\rm{SV}}(\mathcal{C})$ are the Euler characteristic and the area Siegel-Veech constant of $\mathcal{C}$, respectively (a different sign convention is used for $\chi$ in \cite{chen2014quadratic}),
and $\delta=\delta_0+\cdots + \delta_{g-1}$ is the total boundary divisor class. From [\citefirst{moller2006variations}[Proof of Prop.~4.2]{chen2014quadratic}, one has $\mathcal{C} \cdot \eta = -\chi$. By Lemma \ref{lem:Z1class}, we conclude that
\[
\mathcal{C} \cdot \overline{\mathbb{H}}^2_{g, 1} = \mathcal{C} \cdot (2\psi-\eta) = \frac{\chi}{3}.
\]
As in the case $k=1$, one argues that $\chi<0$, hence $\mathcal{C} \cdot \overline{\mathbb{H}}^2_{g, 1} < 0$. Now, select an ample divisor class
\[
A:= c_\eta \eta + c_\lambda \lambda + c_\psi \psi + \sum_{i=0}^{ g-1 } c_i \delta_i \,\,\in\,\, \mathrm{Pic}\left(\P\mathbb{E}_{g,1}^2\right),
\]
and let $c_\delta := \max_{\, i=0}^{\, g-1} c_i$. Then $A^+= c_\eta \eta + c_\lambda \lambda + c_\psi \psi + c_\delta \delta$ is big.
Define $B:=dA^+$ for some sufficiently small $d>0$ such that $\mathcal{C} \cdot (\overline{\mathbb{H}}^2_{g, 1} +B) \leq 0$ for all $\mathcal{C} \in \mathscr{C}$. Specifically, let
\[
d:= \inf_{\mathcal{C}\in \mathscr{C}} \Bigg\{ \frac{12}{36c_\eta + 12c_\psi + 5(g-1)c_\lambda +c_{\rm{SV}}(\mathcal{C})(18c_\lambda +216c_\delta)} \Bigg\}.
\]
The expression in the brackets is the value of $d$ such that $\mathcal{C} \cdot (\overline{\mathbb{H}}^2_{g, 1} +dA^+) = 0$, and is positive, since $\mathcal{C} \cdot \overline{\mathbb{H}}^2_{g, 1} < 0$, $\mathcal{C} \cdot A > 0$, and $\mathcal{C} \cdot \delta_i \geq 0$ for all~$i$ (as the general element of $\mathcal{C}$ is a \textit{smooth} pointed curve with quadratic differential), hence $\mathcal{C} \cdot A^+ > 0$.
From \cite[Thm.~2]{eskin2014sum}, $c_{\rm{SV}}(\mathcal{C})$ is the sum of certain Lyapunov exponents minus a positive constant; in particular, $c_{\rm{SV}}(\mathcal{C})$ is bounded. It follows that $d$ is positive, and thus $B$ is big. By Lemma \ref{lem:extremality}, $\overline{\mathbb{H}}^2_{g, 1}$ is extremal.
\smallskip
Finally, we prove the rigidity of $\overline{\mathbb{H}}^k_{g, 1}$ simultaneously for $k\in \{1,2\}$. Suppose that there exists an effective divisor $D$ such that $D \equiv m\,\overline{\mathbb{H}}^k_{g, 1}$, for some $m>0$. We may assume that $D$ does not contain $\overline{\mathbb{H}}^k_{g, 1}$, since otherwise we could simply consider the divisor $D\setminus\overline{\mathbb{H}}^k_{g, 1}$ and reduce the coefficient $m$. For all Teichm\"uller curves $\mathcal{C}\in\mathscr{C}$ as above, we have that $\mathcal{C} \cdot \overline{\mathbb{H}}^k_{g, 1} < 0$, and thus $\mathcal{C} \cdot D < 0$. This means that $D$ contains the entire collection $\mathscr{C}$ of such curves. Since $\mathscr{C}$ is dense in $\overline{\mathbb{H}}^k_{g, 1}$, we deduce that $D$ contains $\overline{\mathbb{H}}^k_{g, 1}$, which contradicts the assumption.
\end{proof}
\bibliographystyle{alphanumN}
|
\section{Introduction}
The gravitational wave (GW) astronomy will provide unprecedented opportunity to test fundamental physics. Currently, more than 50 coalescing compact binaries events have already been observed during the three running stages of the LIGO/VIRGO \citet{LIGOScientific:2018mvr}, \citet{LIGO_01, LIGO_02} and other dozens more should be observed in the coming years. Crucially, the observation of the GWs from neutron star binaries GW170817 and its associated electromagnetic counterpart GRB 170817A have marked the dawn of multi-messenger cosmology \citet{LIGOScientific:2017vwq, Goldstein:2017mmi, Savchenko:2017ffs, LIGOScientific:2017zic,LIGOScientific:2017ync, DES:2016apb,Arcavi:2017xiz, Tanvir:2017pws}. Such events are dubbed as standard sirens (SS), the gravitational analogue of standard candles \citet{Holz:2005df, Palmese:2019ehe}. The wealth of science that joint detections bring can not be probed with either messenger alone. Other counterparts like neutrino emissions and polarization features associated to this GW event can shed more light on the nature of the merger \citet{ANTARES:2017bia,Shakeri:2018qal}. The importance of such a detection is because the redshift and the location of the source are obtained with more precision breaking the degeneracies in the parameter space. In particular, one appealing application for the SS observations, is the possibility of utilizing events like these to estimate cosmological parameters \citet{Schutz:1986gp}. This provides one complementary probe to constrain cosmological parameters, since the cosmological parameters are encoded in the luminosity distance provided from these events.
An independent measurement of the Hubble constant, $H_0$, using the standard siren approach results in $H_0=70^{12}_{-8}$ km s${}^{-1}$ Mpc${}^{-1}$ \citet{LIGOScientific:2017adf}. The errors for one event are still large. However, with the advent of third generation detectors like Einstein Telescope \citet{Maggiore:2019uih}, Cosmic Explorer \citet{Sathyaprakash:2019nnu} which are based on the earth and space-borne detectors like LISA \citet{LISA:2017pwj, Baker:2019nia} and TianQin \citet{TianQin:2020hid,TianQin:2015yph}, the constraining power of GWs with electromagnetic counterparts will increase significantly and they could extend these observations to large redshifts. The LISA will operate in the millihertz band with the objective to be an all-sky GW survey \citet{Baker:2019nia}. Science with LISA brings opportunities and challenges in terms of complications which arise because of its motion around the earth. Basically, LISA can be considered as two detectors. It will be launched in three identical drag-free spacecraft forming an equilateral triangle with the arm length of about $2.5\times10^6 km$ \citet{Cutler:1997ta,Cornish:2002rt}.
The implications of cosmological studies using the standard sirens have motivated focused studies on the nature of dark energy, modified gravity, dark matter, and several other fundamental questions in modern cosmology \citet{Nishizawa:2019rra,LISACosmologyWorkingGroup:2019mwx,Belgacem:2018lbp,Belgacem:2017ihm, Ezquiaga:2021ler,Cai:2021ooo,Nishizawa:2017nef,Nunes:2019bjq,Yang:2019vni,Matos:2021qne,Jiang:2021mpd,Wang:2019tto,DAgostino:2019hvh,Pan:2021tpk,Kalomenopoulos:2020klp,Yang:2021qge,Gray:2019ksv,Lagos:2020mzy,Tasinato:2021wol,Bonilla:2019mbm,deSouza:2019ype,Dalang:2019rke,Zhang:2020axa,Yang:2020wby,Baker:2020apq,Fu:2020btp,Belgacem:2019zzu,Bonilla:2021dql,Mastrogiovanni:2020gua, Nunes:2020rmr,Baral:2020mzs,Bernardo:2021vsj,Mastrogiovanni:2020mvm,Wang:2021srv,Zhang:2021kqn, Canas-Herrera:2021qxs,Canas-Herrera:2019npr}.
Even if the detected event has no electromagnetic counterpart, it is possible to use other methods to study cosmological parameters \citet{Mukherjee:2020mha,Zhu:2021aat,Garoffolo:2020vtd, Mukherjee:2020hyn,Borhanian:2020vyr,Wang:2020dkc,Feeney:2020kxk,Mastrogiovanni:2021wsd}
In this article, we will forecast bounds for a motivated modified gravity model named No Slip gravity \citet{Linder:2018jil}, a subclass of the Horndeski gravity model \citet{Horndeski}, a general scalar tensor theory with second order field equations. The main characteristics of No Slip model are given by the speed of GWs propagation which is equal to the speed of light, and equality between the effective gravitational coupling strengths to matter and light, but yet different from Newton’s constant, which is capable of generating an effective gravitational coupling. Some observational perspectives of this class were previously investigated in \citet{Brush:2018dhg, Mitra:2020vzq,Linder:2020xza,Brando:2019xbv,Nunes:2020rmr}.
Our approach is to use the LISA standard sirens to predict the bounds on the No Slip gravity baseline parameters. In following analysis, we integrate the modified luminosity distance from standard sirens to constrain the main free parameters of the model. Also, we combine our results with previous results where cosmic microwave background data were used for No Slip gravity, to get tight constraints on the model.
This manuscript is organized as follows. In section \ref{2}, we present the modified luminosity distance in our model. Section \ref{sec:data} is devoted to the LISA standard sirens where we show how we generate the mock data. In section \ref{results}, we present our main results. Finally, our final remarks are included in section \ref{5}.
\section{Propagation of Gravitational Waves in Modified Gravity}
\label{2}
One particular way by which modified theories leave their imprints on the cosmological observables is the modified propagation equations for the tensorial part of the perturbations. This leads to a notion of GW luminosity distance different from the electromagnetic luminosity distance. This provides an arena to test the modified theories in the context of GWs by the standard sirens, the GW events with the associated electromagnetic counterparts.
The standard expression of the luminosity distance in a universe with matter density fraction $\Omega_m$, radiation density fraction $\Omega_R$ and dark energy density $\rho_{de}$ is defined as
\begin{equation}
d^{em}_L(z)=\frac{1+z}{H_0}\int_0^z\,
\frac{d\tilde{z}}{E(\tilde{z})}\, ,
\end{equation}
where
\begin{equation}
\label{E(z)}
E(z)=\sqrt{\Omega_R (1+z)^4+\Omega_m (1+z)^3+\rho_{de}(z)/\rho_0 }\, ,
\end{equation}
where we have $\rho_0=3H_0^2/(8\pi G)$ \citet{Belgacem:2017ihm}.
In GR the transverse taceless part of metric perturbations, $\partial_{i}h^{ij}=h^{i}_{i}=0$, produces the gravitational waves. The propagation equation for GWs on a flat FRW background in absence of anisotropic stress tensor for the polarization A is given by
\begin{align}
h''_A +2 {\cal H} h'_A+k^2 h_A=0, \quad\, A=+,\times
\end{align}
where the prime denotes derivative with respect to the conformal time and ${\cal H}=a'/a$.
Defining
\begin{align}
h_{A}(\eta,k)=\frac{\chi_{A} (\eta,k)}{a(\eta)}\, ,
\end{align}
we find that
\begin{align}
\chi_{A}''(\eta,k)+\left(k^2-\frac{a''}{a} \right)\chi_{A}(\eta,k)=0 \,.
\end{align}
This shows that the amplitude of gravitational waves decrease as $1/a$ and in GR the luminosity distance for the gravitational waves from a binary is the same as electromagnetic waves $d_{L}^{em}=d_{L}^{gw}$.
In a generic modified theory the equation for GWs in the Fourier space reads as
\begin{align}
h''_A +2 {\cal H}[1-\delta(\eta)] h'_A+k^2 h_A=0\, .
\label{newdis}
\end{align}
We have neglected the modulations in $k^2$ term as they change the speed of GWs and we assume the vacuum field equations. Changing the speed of GWs is not favored given the stringent limit set by GW170817 $|c_{\rm gw}-c|/c< O(10^{-15})$.
To solve Eq.~\eqref{newdis} one introduces
\begin{align}
h_{A}(\eta,k)=&\frac{\chi_{A} (\eta,k)}{\bar{a}(\eta)}\,,\\
\frac{\bar{a}'}{\bar{a}}=&{\cal H} \left(1-\delta(\eta) \right).
\end{align}
This way one finds that the luminosity distance for the GWs obeys a different equation. We have \citet{Belgacem:2017ihm, Finke:2021aom, Mastrogiovanni:2020mvm,Hogg:2020ktc}
\begin{align}
d_L^{\,\rm gw}(z)=d_L^{\,\rm em}(z)\exp\left\{-\int_0^z \,\frac{dz'}{1+z'}\,\delta(z')\right\}\, .
\label{dis}
\end{align}
The modified gravity we consider which changes the propagation of GWs as we have in Eqs.~\eqref{newdis} and ~\eqref{dis} is the No Slip model. This modified theory has the advantage of not changing the propagation speed for GWs. Therefore, it is favoured by the current observations. This model exhibits simple behavior in terms of phenomenological parametrizations and has been applied to study the CMB \citet{Brush:2018dhg}.
No Slip model belongs to the well-motivated larger class of Horndeski scalar-tensor theories \citet{Horndeski}. The Horndeski is the most general scalar-tensor theory with second order equations of motion in four dimensions from which different modified theories can be obtained such as Quintessence, k-essence and $f(R)$ gravity. The Lagrangian for the theory is presented in Refs. \citet{Kobayashi:2019hrl,Gleyzes:2014rba} and other references therein. This theory can be tuned such that GWs propagate at the speed of light to conform with the limit set by GW170817. Moreover, a very important prediction of modified gravity theories is the modification of Einstein equations in a way where the coupling of gravity to matter and light could be different. In other words, the Poisson equation and the lensing equation have different dependence on the metric perturbations.
The perturbations on the FRW background in the conformal Newtonian gauge are given by
\begin{align}
ds^2=a(\eta)^2\left\lbrace -\left(1+2 \Psi \right)d\eta^2+\left(1-2\Phi \right)\delta_{ij}dx^{i}dx^j \right\rbrace ,
\end{align}
where $\Phi$ and $\Psi$ are metric perturbations. It is usually a combination, $\Phi+\Psi$, which appears in the geodesic equations for photons \citet{Bertacca:2014wga,Allahyari:2017gmq}. More explicitly, the Einstein equations take the form
\begin{align}
\nabla^{2} & \Psi = 4\pi a^{2} G_N G_{\rm matter} \rho_m\,\delta_ m \\
\nabla^{2} & (\Psi+\Phi) = 8\pi a^{2} G_N G_{\rm light}
\rho_m\,\delta_m \,,
\end{align}
where the quantity $\rho_m$ is the background matter density, and
$\delta_m=\delta\rho_m/\rho_m$ is the comoving density perturbation and $G_{\rm light}$ and $ G_{\rm matter}$ are functions of time and $G_N$ is the gravitational constant.
To quantify the effect of gravity on matter and photons the gravitational slip parameter is defined by
\begin{align}
\eta=\frac{G_{matter}}{G_{light}}.
\end{align}
In the No Slip model the gravitational interaction is gauged so that for the gravitational slip parameter we have $\eta=1$. This way, we find that the effective gravitational couplings are simply related by
\begin{align}
G_{matter}=G_{light}=\frac{m_{p}^2}{M_{*}^{2}}\,,
\end{align}
where $m_{p}$ and $M_{*}$ are Planck and time-dependent Planck mass respectively. $M_{*}$ is a function of the terms in the Horndeski Lagrangian \citet{Kase:2018aps}, \citet{Linder:2015rcz}. We find that the gravitational coupling strength is modified. Therefore, the growth of structures and lensing potentials will deviate from GR prediction.
The other condition to impose on the Horndeski model to derive No Slip model is the speed of GWs. The Hornedeski model can accommodate different propagation speeds for GWs than the speed of light. In Horndeski model the second order action for tensorial perturbations reads as
\begin{align}
S^{(2)}_{\rm tensor}=
\frac{1}{8}\int a^3 q_t\left[
\dot h_{ij}^2-\frac{c_{gw}^2}{a^2}(\partial_k h_{ij})^2
\right] dtd^3x,
\label{action2:tensor}
\end{align}
where $q_t$ is defined in Ref.~\citet{Kase:2018aps}. The speed of tensorial perturbations is given by $c_{gw}$ which is function of various terms in the Horndeski Lagrangian.
In No Slip model we impose the condition to have $c_{gw}=1$. Varying the action in Eq.~\eqref{action2:tensor} and imposing the conditions mentioned, we get
\begin{align}
\ddot{h}_A+\left(3H +\frac{\dot{q}_t}{q_t} \right) \dot{h}_A
+ \frac{k^2}{a^2} h_A=0\,.
\label{tensor}
\end{align}
This way we find that $G_{matter}=G_{light}=1/(16\pi G_N G_4)$ and $ q_t=2 G_4=M_*^2$ where $G_4$ is one of the terms in the Horndeski Lagrangian~ \citet{Kase:2018aps}.
Comparing Eqs.~\eqref{tensor} and ~\eqref{newdis}, we have $-2\delta=\frac{d \ln{q_t}}{d \ln{a}}$. We find that it is the running of the Planck mass which controls the tensor perturbations in this theory.
At this point, one could define a simple phenomenological parametrization to capture the main features of the theory. One could start with a parametrization for $\delta$ or $M_*$. As we are interested in studying the GW luminosity distance we take the first appraoch, namely $\delta$ parametrization.
Specializing to our case study here, No Slip model, we also use the phenomenological parametrization introduced by \citet{LISACosmologyWorkingGroup:2019mwx}
\begin{align}
\label{delta2}
\delta(z) = \frac{n(1-\xi)}{1-\xi+\xi(1+z)^n}\;,
\end{align}
where $\xi$ and $n$ are constants. This parametrization is applicable to models in which we expect some simple properties in low and high redshifts. More explicitly, we find that at the high redshifts $z\rightarrow \infty$, we have $\delta(z)\rightarrow 0$. Moreover, when $z\rightarrow0$, this yields $\delta(z)\rightarrow n(1-\xi)$ where for $\xi=1$ we have $\delta(z)\rightarrow 0$. Thus, we recover the standard expression for the luminosity distance at high and low redshifts. We expect this parametrization to be a viable parametrization as theory is not complicated. Please also note that in models where there are also other degrees of freedom other than one scalar field, the parametrization in Eq.~\eqref{delta2} might not capture the details of the theory. Examples where this parametrization might not work are bigravity theories \citet{Comelli:2012db,Hassan:2011zd}.
Alternatively, another parametrization starting with variation of Planck mass is given in Refs. \citet{Brush:2018dhg, Mitra:2020vzq},
\begin{align}
\left( \frac{m_{p}}{M_{*}}\right)^{-2}=1+\frac{\mu}{1+(\frac{a}{a_{t}})^{-\tau} } \,,
\label{linder}
\end{align}
where $\mu$ is the amplitude of transition, $a_t$ is the scale factor at the transition time and $\tau$ is the rapidity. This parametrization has a simple transition from unity in the past to some constant value in the future. We find that simple parametrizations of the model are adequate for No Slip model.
Here, we find it more convenient to work in $n,\xi$ parametrization, as defined in Eq.~(\ref{delta2}). For the sake of comparison, the mapping from set $(n,\xi)$ to set $(\tau,\mu)$ has been provided in Ref.~\citet{Mitra:2020vzq}
\\
\begin{align}
&\xi = \lim_{z\to \infty}\frac{M_{*}(0)}{M_{*}(z)} \,, \label{eq:Xi0param} \\
& n \simeq \frac{\alpha_{M0}}{2(\xi-1)} \,,
\label{eq:nparam}
\end{align}
where $\alpha_{M0}=-2\delta(0)$. One could check that the mapping generates the same behavior for $\delta(z)$ in two parametrizations. The parametrization in Eq.~\eqref{linder} was suggested to produce a stable and ghost free model. Therefore, this mapping ensures that parametrization in Eq.~\eqref{delta2} is a good choice.
\section{LISA Standard Sirens}
\label{sec:data}
LISA is a space-borne detector with its sensitivity peak around 1 millihertz. Among astrophysical sources LISA can reach including Galactic binaries \citet{Breivik:2017jip,Korol:2017qcx}, stellar-origin black hole binaries \citet{Sesana:2016ljz} and extreme-mass-ratio inspirals \citet{Babak:2017tow}, LISA will also observe massive black hole binaries (MBHBs) from $10^4$ to $10^7$ solar masses \citet{Klein:2015hvg}. The high signal to noise ratio (SNR) of the detected signals will allow for more precise parameter estimations. Among the most probable LISA sources with electromagnetic counterparts are MBHBs. In particular, MBHBs are supposed to merge in gas rich environments and within the LISA frequency band allowing for electromagnetic followups to determine their redshifts. The prospect of MBHBs which could have EM counterparts extends up to $z$ $\sim$7 providing a unique probe of the universe at high redshifts \citet{Klein:2015hvg}.
Our catalog is based on the model in Ref.~\citet{Tamanini:2016zlh} where they use a semi analytic model which allows tracing the galactic baryonic structures and dark matter mergers. In addition, the model integrates the black hole seeding at high redshifts and the delays between the merger
of two galaxies and that of the massive BHs residing in the galaxies.
Theoretical models and simulations can predict the redshift distribution and merger rate of MBHBs. Depending on the initial conditions for blackhole formation at high redshifts, there are two scenarios light seed and heavy seed. In the light seed scenario massive BHs are assumed to grow from the remnants of
population III (pop III) stars forming at $z\approx15-20$. In the heavy seed scenario, the massive BHs are assumed to form from the collapse of protogalactic disks \citet{Tamanini:2016zlh}.
The result of the scenarios produce three categories of population models named Pop III, Delay and No Delay \citet{Madau:2001sc,Volonteri:2007ax,Klein:2015hvg}. Among all the merger events, one chooses the events with a SNR threshold and localization error $\Delta\Omega$.
The redshift distribution of sources based on LISA sensitivity is reported in Ref.~\citet{Tamanini:2016zlh,Tamanini:2016uin}.
For each source one then calculates the associated GW luminosity distance assuming a given cosmology. We need to consider the prospective errors on $d_L$ based on the LISA sensitivity. Using the criteria in Ref. \cite{Speri:2020hwc} where they use the results of Ref.~\citet{Marsat:2020rtl} to fix the proportionality factor in the approximation $\sigma_{\rm LISA} / d_L \propto 2 d_L$ , we have
\begin{equation}
\frac{\sigma_{\rm LISA}(z)}{d_L(z)} = 0.05 \left(\frac{d_L(z)}{36.6\, {\rm Gpc}}\right) \,.
\end{equation}
However, this will give us the instrument error. We also need to consider three pieces of ingredients in our error estimation \citet{Wang:2021srv, Speri:2020hwc}. They are weak lensing
\begin{equation}
\sigma_{\rm delens}(z) = F_{\rm delens}(z) \sigma_{\rm lens}(z) \,,
\end{equation}
where we have
\begin{equation}
F_{\rm delens}(z) = 1 - \frac{0.3}{\pi / 2} \arctan\left(\frac{z}{z_*}\right) \,,
\end{equation}
\begin{equation}
\frac{\sigma_{\rm lens}(z)}{d_L(z)} = 0.066 \left( \frac{1 - (1+z)^{-0.25}}{0.25} \right)^{1.8} \,,
\end{equation}
peculiar velocity
\begin{equation}
\frac{\sigma_v(z)}{d_L(z)} = \left[ 1 + \frac{c (1+z)^2}{H(z)d_L(z)} \right] \frac{\sqrt{\langle v^2 \rangle}}{c} \,,
\end{equation}
with $\langle v^2 \rangle = 500$ km/s,
and redshift measurement error \citet{Speri:2020hwc}
\begin{equation}
\sigma_{\rm photo}(z) = 0.03 (1+z) \quad{\rm if}\; z>2 \,.
\end{equation}
The total error is given by adding in quadrature all contributions:
\begin{equation}
\langle\sigma_{d_L} \rangle^2=\langle \sigma_\mathrm{photo} \rangle^2+\langle \sigma_{lens} \rangle^2+\langle \sigma_{LISA} \rangle^2+\langle \sigma_{v} \rangle^2\;.
\end{equation}
We generate three catalogs based on these criteria for the aforementioned distributions, namely, Pop III, Delay and No Delay for a ten-year mission. We generate our mock catalog from a fiducial input baseline with $H_0=67.4$ km s${}^{-1}$ Mpc${}^{-1}$, $\Omega_m=0.315$, $\xi = 1.0328$, $n = 0.3176$. The pair values $(\xi,n)$ are chosen in order to generate a mock catalog using modified gravity corrections. The motivation for this choice meets in the sense that we do not assume the generation of the full simulated catalogs fixed on the $\Lambda$CDM baseline. Since we are interested in investigating deviations from the standard cosmology, we choose some minimal input deviation from $\Lambda$CDM cosmology to create the mock data. Then, the forecast of experiments, in this case, the LISA mission, will quantify the expected error bar on the baseline of the model. Note that our choice for the pair values $(\xi,n)$ is compatible with current constraints \citet{Brush:2018dhg}, \citet{Mitra:2020vzq}.
\begin{figure}
\begin{center}
\includegraphics[width=1\columnwidth]{LISA_II_dist.pdf}
\caption{Redshift distribution of MBHB standard sirens, namely, Pop III, Delay and No Delay.}
\label{z-dist}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=1\columnwidth]{LISAII.pdf}
\caption{LISA standard sirens at all redshifts from the three population models we consider in this work.}
\label{SH1}
\end{center}
\end{figure}
Figure \ref{z-dist} shows the redshift distributions of the mock catalogs and Figure \ref{SH1} shows the luminosity distance and the corresponding errors based on our input baseline, as well as the three categories of LISA SS sources.
In what follows, we will analyze the free parameters of the theory and present our main results.
\section{Results and Discussion}
\label{results}
In this section, we present our main results on the model parameters by performing Bayesian Monte Carlo Markov Chain analysis. We create new likelihoods in \texttt{MontePython} code \citet{monte} to analyze the SS mock data as presented above. Then, we use the Metropolis - Hastings mode in \texttt{CLASS}+\texttt{MontePython} code~\citet{class, monte, monte_V3} to derive the constraints on cosmological parameters ensuring a Gelman - Rubin convergence criterion of $R - 1 < 10^{-3}$~\citet{gelman1992}. We divide our analysis as follows: First, we will analyze our baseline for each category of populations as defined in section \ref{sec:data}, namely, Pop III, Delay and No Delay samples. Then, we will do a joint analysis using the CMB information as obtained in the context of No Slip gravity in Ref.~\citet{Brush:2018dhg}. We simulate LISA + CMB constraints on the No Slip gravity adding appropriate Gaussian priors on the parameters $H_0$ and $w_{\rm cdm}$ as derivatives in Ref.~\citet{Brush:2018dhg}.
\begin{table*}
\centering
\caption{Constraints at 68\% CL from the Pop III samples and their combination with CMB information.}
\label{tab:main_results_1}
\begin{tabular}{cccc}
\hline
Parameter & Pop III & Pop III + CMB \\
\hline
$H_0$ [km s${}^{-1}$ Mpc${}^{-1}$] & $66.71\pm 0.95$ & $ 66.92\pm 0.56$\\
$\Omega{}_{m} $ & $0.31\pm 0.11 $ & $ 0.3160\pm 0.0070 $\\
$n $ & $0.64^{+0.56}_{-0.68}$ & $0.15^{+0.69}_{-0.80} $\\
$\xi $ & $0.98\pm 0.33 $ & $ 1.016^{+0.069}_{-0.083} $\\
\end{tabular}
\end{table*}
\begin{table*}
\centering
\caption{Same as table \ref{tab:main_results_1}, but for the Delay samples.}
\label{tab:main_results_2}
\begin{tabular}{cccc}
\hline
Parameter & Delay & Delay + CMB \\
\hline
$H_0$ [km s${}^{-1}$ Mpc${}^{-1}$] & $70.8\pm 4.5 $ & $67.78\pm 0.91 $ \\
$\Omega{}_{m } $ & $0.456^{+0.049}_{-0.13} $ & $0.3098\pm 0.0092$ \\
$n $ & $0.98\pm 0.55$ & $0.42^{+0.47}_{-0.78}$ \\
$\xi$ & $1.362^{+0.081}_{-0.22}$ & $1.13^{+0.14}_{-0.34} $ \\
\end{tabular}
\end{table*}
\begin{table*}
\centering
\caption{Same as table \ref{tab:main_results_1}, but for the No Delay samples.}
\label{tab:main_results_3}
\begin{tabular}{cccc}
\hline
Parameter & No Delay & No Delay + CMB \\
\hline
$H_0$ [km s${}^{-1}$ Mpc${}^{-1}$]& $66.87^{+0.92}_{-0.63}$ & $67.21\pm 0.47 $ \\
$\Omega{}_{m } $ & $0.238\pm 0.078 $ & $0.3133\pm 0.0062 $ \\
$n $ & $0.43^{+0.37}_{-0.55} $ & $0.07^{+0.71}_{-0.54} $ \\
$\xi $ & $0.75\pm 0.31$& $1.037^{+0.076}_{-0.17}$ \\
\end{tabular}
\end{table*}
\begin{figure}
\begin{center}
\includegraphics[width=1\columnwidth]{LISA_I.pdf}
\caption{One dimensional posterior distributions and two dimensional joint contours for the parameter space $H_0$, $\Omega_m$ and $\xi$ for LISA Pop III samples and Pop III + CMB prior. The $\xi =1$ prediction recovers the GR.}
\label{LISA_I}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=1\columnwidth]{Delay_2.pdf}
\caption{Same as in Figure \ref{LISA_I}, but for the Delay samples.}
\label{LISA_II}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[width=1\columnwidth]{LISA_III.pdf}
\caption{Same as in Figure \ref{LISA_I}, but for the No Delay samples.}
\label{LISA_III}
\end{center}
\end{figure}
\begin{figure*}
\includegraphics[width=8cm]{1D_n_LISA.pdf} \,\,\,\,\,\,\,
\includegraphics[width=8cm]{1D_n_LISA_CMB.pdf}
\caption{1D marginalized posterior probability distributions for $n$, obtained for each LISA mock sample on the left panel and in combination with CMB prior on the right panel.}
\label{n}
\end{figure*}
We summarize the main results of our statistical analysis in tables \ref{tab:main_results_1}, \ref{tab:main_results_2} and \ref{tab:main_results_3} for the cases with Pop III, Delay and No Delay mock data, respectively. The tables also show the best fit for LISA + CMB prior combination. Figures \ref{LISA_I}, \ref{LISA_II} and \ref{LISA_III} show the 2D joint posterior distributions at 68\%~CL and 95\%~CL for some parameters of interest. As expected by the definition in Eq. (\ref{delta2}), we note that the parameters $n$ and $\xi$ are completely degenerate statistically. In all our analysis, we choose to keep both parameters free, although only the parameter $\xi$ would suffice to quantify the deviations of GR. Figure \ref{n} shows the one dimensional marginalized posterior probability distributions for the parameter $n$.
We find that the free parameter which quantifies deviation of GR, $\xi$, can be measured with 33\%, 21\% and 41\% accuracy from the Pop III, Delay and No Delay populations, respectively. As already pointed out in \citet{Mitra:2020vzq}, we see a significant degeneracy between the parameters $\xi$ and $n$, which on the other hand is also very degenerate with the matter density $\Omega_m$ and the Hubble constant, $H_0$. Note that for $n=0$, or equivalently $\xi=1$, the equations of the model also reduce to GR. It is interesting to note that the parameter $\xi$ is positively correlated with $\Omega_m$, and does not show significant correlation with $H_0$. As the matter density increases, the luminosity distance to the different GWs sources will decrease, but this can be balanced by increasing the effects from the modified gravity model. We find that the Hubble parameter can be simultaneously fit to 1.4\%, 6.2\% and 1.1\% accuracy from the Pop III, Delay and No Delay populations, respectively.
To test gravity through the propagation of gravitational waves is a unique probe, though the model cannot be distinguished due to strong degeneracy in the whole parameter space using only the SS events. Therefore, in this case, it will be important to combine the SS sample with other probes. Geometrical probes measured by electromagnetic waves like Supernovae Type Ia, Baryon Acoustic Oscillation distance, cosmic chronometers, among others, are insensitive on the No Slip gravity free parameters, but can provide independent constraints on the $\Omega_m$ and $H_0$, and then to help break this degeneracy on $\xi$ and $n$. On the other hand, the No Slip gravity prediction is sensitive to the amplitude of late-time matter fluctuations (weak lensing and redshift-space distortions measurements) and CMB anisotropies. As defined previously, we combine the MBHB standard sirens populations with CMB prior information in order to have a forecast of how much this combination can improve the precision on the parametric space. Taking LISA + CMB prior, we note a 20\%, 33\% and 15\% accuracy from the Pop III, Delay and No Delay populations, respectively. Moreover, $H_0$ can be simultaneously fit to 0.8\%, 1.2\% and 0.7\% accuracy from the Pop III, Delay and No Delay populations, respectively. Note that the cases Delay and No Delay populations, generate higher and lower values for $\Omega_{m}$, respectively, when compared with the value observed by CMB. On the other hand, the parameter $\xi$ can not be measured via electromagnetic waves probes as discussed earlier. Any possible improvements on $\xi$ will happen by breaking the degeneracy on the other parameters, in special $\Omega_{m}$, because $\xi$ and $\Omega_{m}$ are correlated with each other. When we add the CMB prior, the pair ($\Omega_{m}$, $H_0$) tends to (0.31, 67.4). Looking at the parametric space $\xi-\Omega_{m}$ (see Fig. \ref{LISA_II}, \ref{LISA_III}), we can notice this effect. Therefore, the compression on $\Omega_{m}$ to Planck-values, can increase the relative error on $\xi$ in some cases. We note it more significantly in the Delay case. A way to robustly improve the total constraint on $\xi$ would be to combine our analysis with weak lensing and redshift-space distortions measurements. Because in this case, all probes are very sensitive to $\xi$ effects. These steps will be left for future perspectives.
\\
We have extended the analysis in Ref.~\citet{Mitra:2020vzq}
considering mock data from the LISA perspective.
We see an overall improvement on the precision of the Hubble constant, $H_0$, and $\Omega_m$ by including the CMB information. However, we see that by including the CMB it is difficult to distinguish the model from GR as the 68\% credible intervals on $n$ and $\xi$ in tables \ref{tab:main_results_1}, \ref{tab:main_results_2} and \ref{tab:main_results_3} contain the GR prediction, namely $\xi=1$. We note the prospective strength of implementing the CMB is that the degeneracy between $\xi$ and $\Omega_m$ can be broken.
\section{Final remarks}\label{5}
We have presented a forecast analysis for the standard siren events which are expected to be observed in LISA frequency band for three categories of population models, named, Pop III, Delay and No Delay. Then, we performed a parameter estimation analysis to test modifications of GR inspired by the No Slip gravity, where the speed of GWs propagation is equal to the speed of light, and the effective gravitational coupling strengths to matter and light are equal, but yet different from Newton’s constant coupling.
In our analysis we have also included the priors from the CMB and find that perspectives towards No Slip gravity can be well tested within the expected perspective for cosmological observations with LISA. Adding the CMB information could help improve our parameter estimation and the shape of covariances in the parameter space. In combination with CMB information, we find a 15\% accuracy on the modified gravity free parameters and 0.7\% accuracy on the Hubble parameter.
In this present work we have considered one space-borne detector. A more sophisticated future study will be to include both LISA and TianQin or Taiji \citet{Ruan:2018tsw}. This could increase
the number of the GW events with electromagnetic counterparts and improve the whole parameter space of the model. In the next decades GW detectors will form a powerful detector network. This will allow to do joint detections for events. This can be utilized to do complementary studies of modified gravity theories with modified GW propagation equations. On the other hand, the modified GW propagation arises naturally from any modified gravity theory. It may be interesting to search and improve the expected study of LISA sources without electromagnetic counterpart, and use these information to test modified GW propagation, since several LISA sources are expected at large cosmological distance \citet{Mukherjee:2020mha,Borhanian:2020vyr,Wang:2020dkc,Feeney:2020kxk}. Thus, we can significantly increase our sample, number of events for doing statistics. Also, it may be interesting to search the modified GW propagation effects on the gravitational wave background signal expected in LISA band \citet{DOrazio:2018jnv,Samsing:2018isx, Kowalska}.
\section*{Acknowledgments}
The authors thank the referee for some useful comments that improved the manuscript. R.C.N. acknowledges financial support from the Funda\c{c}\~{a}o de Amparo \`{a} Pesquisa do Estado de S\~{a}o Paulo (FAPESP, S\~{a}o Paulo Research Foundation) under the project No. 2018/18036-5. DFM thanks the Research Council of Norway for their support. Some computations were performed on resources provided by UNINETT Sigma2 -- the National Infrastructure for High Performance Computing and Data Storage in Norway.
\section*{Data Availability}
The data underlying this article will be shared on reasonable request to the corresponding author.
\bibliographystyle{mnras}
|
\subsection{MPPI}\label{subsec:MPPI}
In this paper, we follow the procedures described in \cite{p:williams2017informationMPPI} to use MPPI algorithm.
The MPPI module requires system dynamics and initial state $x_0$ to sample trajectories.
The MPPI horizon $T_{\mathrm{MPPI}}$, the input sampling covariance parameter $\nu$, and the number of trajectory samples $K$ are required as algorithm parameters.
\subsection{Half-space Generation}\label{subsec:halfspacegen}
Safe half-spaces are generated by the ``HSGen'' procedure.
It takes the first $T_{\mathrm{CS}}$ states of the sequence $X^{T_{\mathrm{MPPI}}}$ and projects the position vectors $p_\ell$ onto each obstacle $\mathcal{O}_{j}$.
Then, it computes the supporting hyperplane $\mathcal{H}_{\ell, j} := \{ p \in \mathbb{R}^{2} ~|~ a_{\ell,j}^{\mathrm{T}} p - b_{\ell,j} = 0 \},$
at the point of projection such that $a_{\ell, j}^{\mathrm{T}} p_\ell - b_{\ell, j} \geq 0 \Rightarrow p_\ell \notin \mathcal{O}_{j}.$
The projection of position $p_\ell$ at time $\ell$ onto the obstacle $\mathcal{O}_j$ is denoted as $z_{\ell, j}$ and is defined as $z_{\ell,j} := s_j + h_{\ell, j} r_j$,
where $h_{\ell,j}:= (p_\ell - s_j)/\lVert p_\ell - s_j \rVert_2$ is the unit normal vector to obstacle $\mathcal{O}_j$ at the point $z_{\ell, j}$
towards $p_\ell$.
We set $ a_{\ell,j} = h_{\ell,j}$ and, $b_{\ell,j} = a_{\ell,j}^{\mathrm{T}} z_{\ell, j}$.
So, we can express $a_{\ell,j}$ and $b_{\ell,j}$ in terms of $p_l, s_j$ and $r_j$ as follows:
{\small{
\begin{align}\label{eq:halfspace}
a_{\ell,j} &= \frac{(p_\ell - s_j)}{\lVert p_\ell - s_j \rVert_2}, & b_{\ell,j} & = \frac{(p_\ell - s_j)^{\mathrm{T}} s_{j} }{\lVert p_\ell - s_j \rVert_2} + r_j.
\end{align}
}}
The halfspace generation process is repeated for each obstacle $\mathcal{O}_j$ and every time step $\ell$. The halfspace parameters are gathered in the set of tuples $\mathcal{S}_k^{\mathrm{obs}} = \{ (a_{\ell,j}, b_{\ell,j}) \}_{\ell=0,\dots,T_{MC}}^{ j=1,\dots,N_{\mathrm{obs}}} $ to be used in Constrained Covariance Steering.
The condition $p_{\ell} \notin \mathcal{O}_j$ is not necessary for the half-space generation procedure.
Even if $p_\ell \in \mathcal{O}_j$, the procedure described by equations in \eqref{eq:halfspace} generates a half-space $\mathcal{H}_{\ell, j}$ such that $a_{\ell,j}^{\mathrm{T}} p - b_{\ell,j} \leq 0$ holds for all $p \in \mathcal{O}_{j}$.
\subsection{Constrained Covariance Steering}\label{subsec:covsteer}
The goal of the Constrained Covariance Steering module is to minimize the deviation of the actual state and control sequence from the reference state and control sequences which are computed by the MPPI algorithm while satisfying the safety constraints. This problem can be formally stated as the following stochastic optimal control problem:
\begin{subequations}\label{prob:problem2}
\begin{align}
\underset{\pi \in \Pi}{\textrm{minimize}} & \quad J(\pi):= \mathbb{E} \Big[\sum_{\ell=0}^{T_{CS}-1} \delta x_{\ell}^{\mathrm{T}} Q_{\ell} \delta x_{\ell} + \delta u_{\ell}^{\mathrm{T}} R_{\ell} \delta u_{\ell} \nonumber \\
& \qquad \qquad \quad + \delta x_{T_{CS}}^{\mathrm{T}} Q_{T_{CS}} \delta x_{T_{CS}} \Big] \label{prob2:obj}\\
\textrm{subject to} & \quad x_{\ell+1} = A_\ell x_\ell + B_\ell u_\ell + w_\ell, \quad \forall \ell \in \mathcal{I}_{t} \label{prob2:dynamics}\\
& \quad u_\ell = \pi(x_{0}, \dots, x_{\ell}), \qquad\qquad\; \forall \ell \in \mathcal{I}_{\mathrm{t}}\\
& \quad \mathbb{P} \left[ a_{\ell, j}^{\mathrm{T}} p_\ell - b_{\ell,j} \geq 0 \right] \geq 1 - P_{\mathrm{fail}}, \quad \nonumber \\
& \qquad \qquad\qquad\qquad\qquad\qquad \forall \{\ell, j\} \in \mathcal{I} \label{prob2:constrsafe}
\end{align}
\end{subequations}
\noindent where $\Pi$ denotes the set of all admissible control policies, $\delta x_{\ell} = x_{\ell} - x_{\ell}^{\mathrm{MPPI}}$, $\delta u_{\ell} = u_{\ell} - u_{\ell}^{\mathrm{MPPI}}$, $\mathcal{I}_{\mathrm{t}} := \{0, \dots, T_{CS}\}$, $\mathcal{I}_{\mathrm{o}} := \{1, \dots, N_{\mathrm{obs}}\}$, $\mathcal{I} = \mathcal{I}_{\mathrm{t}} \times \mathcal{I}_\mathrm{o}$.
The stochastic optimal control problem defined in
\eqref{prob:problem2}
can be cast as a deterministic optimization problem by fixing the policy as in \eqref{eq:inputCS}
as explained in Section \ref{sec:CovSteering}.
The resulting finite-dimensional deterministic optimization problem is given by:
\begin{subequations}\label{prob:problem3}
\begin{align}
\underset{\bm{\Bar{u}}, \bm{\mathcal{H}}, \bm{\mathcal{K}} }{\textrm{minimize}} & \quad \mathcal{J}(\bm{\Bar{u}}, \bm{\mathcal{H}}, \bm{\mathcal{K}}):=\delta \bm{\Bar{x}}^{\mathrm{T}} \bm{Q} \delta \bm{\Bar{x}} + \delta \bm{\Bar{u}}^{\mathrm{T}} \bm{\mathcal{R}} \delta \bm{\Bar{u}} \nonumber\\
& \qquad \qquad + \operatorname{tr}(\bm{\mathcal{Q}} \mathrm{var}_{\bm{x}} ) + \operatorname{tr}(\bm{\mathcal{R}} \mathrm{var}_{\bm{u}}) \label{prob3:obj} \\
\textrm{subject to} & \quad a_{\ell,j}^{\mathrm{T}} \mathbf{P}_{\ell} \mu_{\bm{x}} - b_{\ell,j} \geq \nonumber \\
& \qquad \qquad \alpha \lVert \bm{\zeta}^{\mathrm{T}} \mathbf{P}_{\ell}^{\mathrm{T}} a_{\ell,j} \rVert_2, ~~ \forall \{\ell, j\} \in \mathcal{I} \label{prob3:constr1}\\
& \quad \bm{\zeta} = \left[ (\mathbf{G}_0 + \mathbf{G_u} \bm{\mathcal{H}}) ~ (\mathbf{G_w} + \mathbf{G_w} \bm{\mathcal{K}}) \right]\mathbf{R} \label{prob3:constr2}
\end{align}
\end{subequations}
\noindent where $\mathbf{R}\mathbf{R}^{\mathrm{T}} = \operatorname{bdiag}(\Sigma_0, \mathbf{W})$, $\bm{x}$ and $\bm{u}$ are defined as in Section \ref{sec:CovSteering}, $\delta \bm{\Bar{x}} = \mu_{\bm{x}} - \bm{x}^{\mathrm{MPPI}}$, $\delta \bm{\Bar{u}} = \mu_{\bm{u}} - \bm{u}^{\mathrm{MPPI}}$, $\bm{\mathcal{Q}}=\operatorname{bdiag}(Q_0, \dots, Q_{T_{CS}})$, $\bm{\mathcal{R}}=\operatorname{bdiag}(R_0, \dots, R_{T_{CS}-1})$.
$\mathbf{P}_{\ell}$ is defined such that $p_\ell = \mathbf{P}_{\ell} \bm{x}$
and $\alpha = \varphi^{-1}(1 - P_{\mathrm{fail}})$ where $\varphi$ is the cumulative density function of normal random variable with zero mean and unit variance.
Finally, $\mu_{\bm{x}}$, $\mu_{\bm{u}}$, $\mathrm{var}_{\bm{x}}$ and $\mathrm{var}_{\bm{u}}$ are defined as in \eqref{eq:meanX}-\eqref{eq:covV}. In addition, we observe that $\mathrm{var}_{\bm{x}} = \bm{\zeta} \bm{\zeta}^{\mathrm{T}}$ where $\bm{\zeta}$ is given by \eqref{prob3:constr2}.
To see the equivalence of optimization problems in \eqref{prob:problem3} and \eqref{prob:problem2},
first, observe that the objective function in \eqref{prob:problem2} can be written as $\mathbb{E}[\delta \bm{x}^{\mathrm{T}} \bm{\mathcal{Q}} \delta \bm{x} + \delta \bm{u}^{\mathrm{T}} \bm{\mathcal{R}} \delta \bm{u} ]$
where $\delta \bm{x} = \bm{x} - \bm{x}^{\mathrm{MPPI}}$, $\delta \bm{u} = \bm{u} - \bm{u}^{\mathrm{MPPI}}$.
Using the cyclic permutation property of trace operator, the linearity of expectation and the equalities $\mathrm{var}_{\delta \bm{x}} = \mathrm{var}_{\bm{x}}$ and $\mathrm{var}_{\delta \bm{u}} = \mathrm{var}_{\bm{u}}$, it follows that the objective functions in \eqref{prob2:obj} and \eqref{prob3:obj} are equivalent.
Proposition \ref{prop:chanceconstraint} along with the expressions $\mu_{p_\ell} = \mathbf{P}_{\ell} \mu_x$ and $\mathrm{var}_{p_{\ell}} = \mathbf{P}_{\ell} \bm{\zeta} \bm{\zeta}^{\mathrm{T}} \mathbf{P}_{\ell}^{\mathrm{T}}$ show that \eqref{prob3:constr1}, \eqref{prob3:constr2} $\Leftrightarrow$ \eqref{prob2:constrsafe}.
\begin{proposition}\label{prop:chanceconstraint}
Let $ p \sim \mathcal{N}(\mu_{p}, \Sigma_{p})$, where $a, \mu_p \in \mathbb{R}^n$, $b \in \mathbb{R}$, $\Sigma_p \in \mathbb{S}_n^{+}$, and $P \in (0, 1/2]$. Then, $\mathbb{P}[ a^{\mathrm{T}} p - b \geq 0] \geq 1 - P$ if and only if $a^{\mathrm{T}} \mu_p - b \geq \varphi^{-1} (1-P) \lVert \mathbf{R}^{\mathrm{T}} a \rVert_2 $ where $\varphi : \mathbb{R} \rightarrow (0, 1) $ is the cumulative density function of normally distributed random variable with zero mean and unit variance and, finally $\mathbf{R}$ is such that $\mathbf{R} \mathbf{R}^{\mathrm{T}} = \Sigma_p$.
\end{proposition}
The problem in \eqref{prob:problem3} has a convex quadratic objective function and the constraint in \eqref{prob3:constr2} is affine.
Also, the constraint in \eqref{prob3:constr1} corresponds to a second-order cone constraint since $\alpha = \varphi^{-1}(1-P_{\mathrm{fail}}) \geq 0$ for all $P_{\mathrm{fail}} \in (0, 0.5]$ \cite{b:boyd2004convex}.
Thus, problem \ref{prob:problem3} can be solved for global optimal solution $( \bm{\Bar{v}}^{\star}, \bm{\mathcal{H}}^{\star}, \bm{\mathcal{K}}^{\star})$ using off-the-shelf solvers such as \cite{mosek2010mosek}.
\subsection{Discussion}\label{subsec:discussionAlgorithm}
First, the algorithm is initialized by setting $\Bar{x}_{0} = x_0$ and $\Sigma_0 = \bm{0}$ where $\Bar{x}_0$ and $\Sigma_0$ represent the initial nominal state and initial covariance respectively.
Then, using $\Bar{x}_k$ as the initial state, MPPI generates a pair of reference state and input sequences $(X^{T_{\mathrm{MPPI}}}, U^{T_{\mathrm{MPPI}}})$.
The state sequence is used to generate a safe convex region over which the constraints \eqref{prob2:constrsafe} are satisfied based on the technique that will be described in Section \ref{subsec:halfspacegen}.
Then, we formulate a corresponding CCS problem that seeks for a control policy in the form of \eqref{eq:inputCS}
that will guarantee the satisfaction of chance constraints while minimizing the deviation from reference state and control sequences.
If the largest eigenvalue of the computed covariance $\Sigma_{k+1}$ exceeds a predetermined threshold $\sigma_{\mathrm{max}}$, then the nominal state $\Bar{x}_{k}$ is set equal to the real state $x_k$ and covariance $\Sigma_k$ is set to $\bm{0}$.
Next, the nominal state $\Bar{x}_{k+1}$ and covariance matrix $\Sigma_{k+1}$ will be updated as described in lines 9-10 in Algorithm \ref{al:CSMPPI}.
It is worth mentioning that the CCS module uses the disturbance noise covariance in its formulation. However, this information is usually unknown in real-world scenarios. But, this can easily be handled by over-approximating the noise covariance, that is, by taking $\mathbf{W}_k \succeq \mathbf{W}_k^{\mathrm{real}}$, where $\mathbf{W}_{k}^{\mathrm{real}}$ is the actual noise covariance that is acting on the system.
This allows the CCS module to find a policy that satisfies the safety constraints.
Furthermore, this condition implies that the state will stay in the safe region with probability greater than $1 - P_{\mathrm{fail}}$.
Although this approach may generate overly conservative policies, system identification techniques can be used to learn the actual noise covariance \cite{p:feng2014kalman} and hence reduce conservativeness.
The final component of our algorithm is the use of nominal dynamics, which is the noise-free version of the real dynamics.
Since the CCS module assures that chance constraints are satisfied and it uses the nominal state as the initial mean state in its formulation,
the safety margins of the nominal state will be greater than the real state.
Also, by computing the covariance in line 10 in Algorithm \ref{al:CSMPPI}, we compute the high probability region where the real state lies. Then, this covariance value is used as initial covariance in the CCS procedure in the next step to guarantee the satisfaction of the chance constraints.
\section{Introduction}\label{sec:introduction}
\input{intro}
\section{Problem Statement and Preliminaries}\label{sec:problem-statement}
\input{problem}
\section{Brief Review of MPPI and TUBE-MPPI}\label{sec:MPPI}
\input{mppi}
\section{Linear Covariance Steering Theory}\label{sec:CovSteering}
\input{covariance-steering}
\section{Main Algorithm}\label{sec:MainAlgorithm}
\input{main-algorithm}
\section{Numerical Experiments}\label{sec:Numerical Experiments}
\input{experiments}
\section{Conclusion}\label{sec:Conclusion}
\input{conclusion}
\bibliographystyle{ieeetr}
\subsection{Notation}
We denote by $\mathbb{R}^n$ the set of $n$-dimensional real vectors.
We use $\mathbb{E}\left[\cdot\right]$ and $\mathbb{P}(E)$ to denote the expectation functional and the probability of the random event $E$, respectively.
Given a vector $x$, its 2-norm is denoted by $\lVert x \rVert_2$ and
given a matrix $\mathbf{A} \in \mathbb{R}^{n \times m}$, we denote its Frobenius norm by $\lVert\mathbf{A} \rVert_{F}$
and its trace by $\operatorname{tr}(\mathbf{A})$.
We use $\bm{0}$ and $I_n$ to denote the zero matrix
and the $n \times n$ identity matrix, respectively.
We will denote the convex cone of $n\times n$ symmetric positive semi-definite (symmetric positive definite) matrices by $\mathbb{S}^{+}_n$ ($\mathbb{S}^{++}_n$).
We write $\mathrm{bdiag}(A_1,$ $\dots, A_\ell)$ to denote the block diagonal matrix formed by the matrices $A_i$, $i\in \{1,\dots, \ell\}$. $\cup_{i=1}^{N} \mathcal{O}_i$ to denote the union of sets $\mathcal{O}_i$ indexed by $i \in \{1, \dots, N\}$.
We denote by $\mu_z$ and $\mathrm{var}_z$ the mean and the variance of a random vector $z$, respectively.
We write $ x \sim \mathcal{N}(\mu, \Sigma)$ to represent the random variable $x$ is normally distributed with mean $\mu$ covariance $\Sigma$.
\subsection{Problem Statement}
The motion of the agent is described by a discrete-time stochastic linear system:
\begin{align}
x_{k+1} = A_k x_k + B_k u_k + w_k, \label{eq:linear-dynamics}
\end{align}
where $A_k \in \mathbb{R}^{4 \times 4}$, $B_k \in \mathbb{R}^{4 \times 2},$ $x_k \in \mathcal{X} \subseteq \mathbb{R}^{4}$ is the state of the agent which is decomposed as $x_k = [p_k^\mathrm{T}, v_k^\mathrm{T}]^{\mathrm{T}}$,
where $p_k = [p^x_k, p^y_k]^{\mathrm{T}} \in \mathbb{R}^{2}$ is the position and $v_k = [v^x_k, v^y_k]^{\mathrm{T}} \in \mathbb{R}^{2}$ is the velocity,
$u_k = [u_k^x, u_k^y]^{\mathrm{T}} \in \mathbb{R}^2$ is the control input and
$w_k \in \mathbb{R}^{4}$ is the disturbance.
We assume that $w_k \sim \mathcal{N}(\mathbf{0}, \mathbf{W}_k)$ and $\mathbb{E}[w_k w_l^{\mathrm{T}}] = \bm{0}$ for all $k \neq l$.
The choice of the objective function will be determined by the particular application. However, in this paper, we use the objective function utilized in the formulation of the information-theoretic MPPI \cite{p:williams2017informationMPPI}, which is defined as follows:
{\small{
\begin{align}\label{eq:objective-function}
\mathcal{L}(X^{N}, U^{N-1}) = \Phi(x_{N}) + \sum_{k=0}^{N-1} \left( q(x_k) + \lambda u_k^{\mathrm{T}} R_k u_k \right),
\end{align}
}}
\noindent where $X^N = \{ x_0, x_1, \dots, x_N \}$ is the state sequence, $U^{N-1} = \{ u_0, u_1, \dots, u_{N-1} \}$ is the control input sequence, $q: \mathbb{R}^{n} \rightarrow \mathbb{R}$ is the state-dependent term of the running cost function, $\Phi : \mathbb{R}^{2n} \rightarrow \mathbb{R}$ is the terminal cost function, $R_k \in \mathbb{S}_n^{+}$ and $\lambda \geq 0$. While the term $\lambda u_k^{\mathrm{T}} R_k u_k$ penalizes the control input, $q(x_k)$ and $\Phi(x_N)$ are task dependent and each of them can either be a smooth function as in \cite{p:williams2017informationMPPI} or a sum of indicator functions for obstacle avoidance as in \cite{p:williams2018tubeMPPI}.
We assume that the position space is populated by $N_{\mathrm{obs}}$ obstacles.
The $i$-th obstacle is parametrized by its position, $s_i \in \mathbb{R}^{n}$, and its radius, $r_i \geq 0 $,
and the region it occupies is denoted as $\mathcal{O}_i := \{ p \in \mathbb{R}^{n} | \lVert x-s_i \rVert_{2} \leq r_i\}$.
The position space is defined as $\mathcal{X} \subseteq \mathbb{R}^{n}$ and the safe region is defined as $\mathcal{X}_{\mathrm{safe}} := \mathcal{X}\backslash\mathcal{O}$ where $\mathcal{O} = \cup_{j=1}^{N_{\mathrm{obs}}} \mathcal{O}_j $. Then, the safe trajectory optimization problem can be formally stated as follows:
\begin{subequations}\label{prob:prob1}
\begin{align}
\underset{X^{N}, \; U^{N-1}}{\textrm{minimize}} & \quad \mathbb{E} [\mathcal{L}(X^{N}, U^{N-1})] \\
\textrm{subject to} & \quad x_0= \Bar{x}_0, ~~ \eqref{eq:linear-dynamics} \nonumber \\
& \quad \mathbb{P}(x_k \in \mathcal{X}_{\mathrm{safe}}) \geq 1 - \mathrm{P}_{\mathrm{fail}}, ~~ \forall k \in \mathcal{I}^{N}
\end{align}
\end{subequations}
where $\mathcal{I}^{N} = \{ 0, 1, \dots, N \}$, $\Bar{x}_{0} \in \mathbb{R}^{2n}$ and $\mathrm{P}_{\mathrm{fail}} \in (0,0.5]$ is the acceptable level of the probability of the failure.
\noindent Since the system dynamics in \eqref{eq:linear-dynamics} includes the random noise term $w_k$, the state sequence $X^{k}$ is a random process thus the problem in \ref{prob:prob1} is a stochastic trajectory optimization problem.
|
\section*{Introduction}
Great progress has been achieved recently in various physical platforms for quantum computing, most notably is the 54-qubit programmable superconducting processor \cite{Arute2019}. High fidelity two-qubit gates were also demonstrated for trapped ions \cite{Ballance2016}, neutral atoms in optical tweezers \cite{Levine2019}, and spin qubits in silicon \cite{He2019,Huang2019} and GaAs \cite{Nichol2017}. These experimental implementations are based on tunable coupling between qubits where the interaction is switched on only when the two qubit gates are implemented. In solid state quantum computers, tunable couplers typically involve more circuit elements and require their own external control for tuning the magnitude of the interaction \cite{Arute2019,Chen2014,He2019} leading to overheads in fabrication and wiring. For solving real-world problems, a quantum computer needs a large number of qubits \cite{Fowler2012}, and the complexity of the tunable couplers adds to the technological difficulties in scaling up the device. In contrast, fixed couplers do not require the extra components for controlling the magnitude of the interaction, resulting in a substantial simplification of the hardware architecture and hence a significant advantage for scaling up.
An important requirement for quantum computing with fixed coupling is the ability to cancel the unwanted evolution due to the fixed interaction on qubits where no gate is needed. In NMR quantum computing, where the qubits have fixed longitudinal couplings, this is achieved by applying a series of cleverly designed of refocusing pulses \cite{Vandersypen2001,Tsunoda2020}. For large arrays of qubits these series become increasingly complex, which is a bottleneck for scaling up \cite{Tsunoda2020}. In this paper we describe a simple method for quantum computing on qubit arrays with fixed coupling without refocusing pulses. Instead, we rely on a key observation that, by driving a specific subarray, one can implement any gate on the driven qubits, and \emph{at the same time} implement an identity operator on all undriven qubits, effectively cancelling the unwanted evolution on these undriven qubits. Any arbitrary quantum circuit can then be implemented by changing the driven subarray between the time steps. An illustration of the driving pattern and the implementation of gates is given in Fig.~\ref{fig:circuit}. Our method can be scaled up to an arbitrarily large array in a straightforward manner, opening an alternative pathway for a simplified quantum computer hardware architecture based entirely on fixed coupling.
In principle, designing the subarray could be difficult. This is because simulating a constantly interacting system of qubits is in general not possible due to the exponential wall: the cost in memory and time increases exponentially with the number of qubits. Thus, one cannot predict the unitary gate implemented by a driving pulse. In our method this problem is avoided, because the driven subarray can be chosen such that the total Hamiltonian of the system can be decomposed into a sum of commuting blocks of only a few qubits. Each block has a low dimensional Hilbert space, and thus its unitary evolution can be simulated and optimized efficiently. This decomposition exists when the qubit-qubit coupling term is longitudinal, i.e., diagonal in the computational basis, for example, the ZZ interaction.
\begin{figure*}[t]
\centering
\includegraphics[width=2\columnwidth]{fig_overall.png}
\caption{\label{fig:circuit} a) An example of implementing quantum gates on an array of qubits with fixed couplings. In any given step only a sub-array can be driven. This sub-array can be chosen to satisfy specific requirements so that the array's Hamiltonian can be decomposed into commuting few-qubit blocks. Each drive can implement a gate on the driven qubit, and through its combination with the fixed qubit-qubit couplings also implements an identity operator on the neighboring undriven qubits. In the next step a different subarray is driven for implementing gates on a different set of qubits. Here C-X on two adjacent qubits denotes the CNOT gate and $I$ the identity gate. b) Illustration of a quantum circuit in our scheme. The key feature is that any idle interval between gates can be used to realize with an active identity gate, for preventing unwanted evolution due to the fixed couplings. }
\end{figure*}
An appealing feature of our method is the robustness of the gates against uncertainty in all the physical parameters of the array. By using robust optimal control we find pulses for realizing gates with fidelities around $99.99\%$ despite a $1\%$ uncertainty in all the qubit-qubit and drive-qubit couplings, and a detuning uncertainty at 0.1\% of the qubit-qubit coupling strength. This robustness of the fidelity against uncertainties is crucial for an architecture with entirely fixed couplers because it is not possible to isolate a qubit or qubit pair for a precise measurement of the parameter values, and hence there is always a significant residual uncertainty even after the device characterisation process.
This paper is organised as follows: We first describe the key details of our method, including the driving pattern to make the Hamiltonian decomposable, the application of a universal set of gates, and the implementation of an arbitrary circuit. Next, we show how to use optimal control to make the gates robust against parameter uncertainty in the Hamiltonian. Finally, we discuss potential physical realisations of our method.
\section*{Results}
We first describe our method for implementing an arbitrary quantum circuit on qubit arrays with fixed longitudinal coupling. We consider a system of qubits coupled by fixed nearest-neighbor longitudinal interaction, i.e., an interaction that commutes with the bare qubit's Hamiltonian. For simplicity we choose the ZZ interaction, which has been realised experimentally for superconducting qubits \cite{Johnson2011,Collodo2020}. When a subset of qubits is driven by external fields, the system's Hamiltonian is
\begin{align}
\mathcal{H}(t)=&-\sum_{j}\frac{\omega_j}{2}\sigma_j^z+\sum_{j\in \mathcal{L}} d_j E_j(t) \sigma^x_j+\sum_{jk}J_{jk}\sigma_j^z\sigma_k^z, \nonumber \\
E_j(t)=& \ E^x_j(t)\cos\left(\nu_j t\right)+E^y_j(t)\sin\left(\nu_j t\right) ,\nonumber
\end{align}
where $\nu_j$ is the frequency of the drive on the $j$-th qubit, $E^x_j(t)$ and $E^y_j(t)$ the two quadratures of the field, $d_j$ the $j$-th qubit's dipole matrix element, and $\mathcal{L}$ the driven subset. Typically, the qubit's transition energy, $\omega_j$, is much larger than the interaction, $J_{jk}$, and hence $\ket{0,0,...,0}$ is the ground state of the undriven Hamiltonian and can be initialised by cooling.
In the frame rotating with the qubits' frequencies, described by the unitary transformation $U_0(t)=e^{i\sum_{j}\frac{\omega_j}{2}\sigma^z_j t}$, the Hamiltonian in the rotating wave approximation is
\begin{align}
H(t)&\approx \sum_{j\in \mathcal{L}} \frac{1}{2} \left[ \Omega_j(t) \sigma^x_j+ \Omega'_j(t)\sigma^y_{j}\right]+\sum_{jk}J_{jk}\sigma_j^z \sigma_k^z ,
\end{align}
where
\begin{align}
\Omega_j(t)=\Omega^x_j(t) \cos(\delta_j t)+\Omega^y_j(t)\sin(\delta_j t),\nonumber \\
\Omega'_j(t) = \Omega^y_j(t) \cos(\delta_j t)-\Omega^x_j(t)\sin(\delta_j t).
\end{align}
Here, $\Omega^{x,y}_j(t)\equiv d_j E^{x,y}_j(t)$ is the Rabi frequency and $\delta_j\equiv \nu_j-\omega_j$ is the detuning.
\subsection*{Hamiltonian decomposition into commuting blocks}
Computing the unitary evolution of a many-body Hamiltonian like $H(t)$ is in general intractable due to the exponential complexity of the wave function, unless one can decompose the Hamiltonian into a sum commuting few-qubit blocks, i.e., $H(t)=\sum_{l}H_l(t)$, where all the $H_l(t)$ are mutually commuting. The unitary evolution after a time duration $T$ is then $U(T)=\prod_l U_l(T)$ where $U_l(T)=\mathcal{T} e^{-i\int_{0}^{T}dt' H_l(t')}$ and $\mathcal{T}$ is the time-ordering operator. $U_l(T)$ can be efficiently computed since it involves only a few qubits. The $U_l(T)$ factors are also mutually commuting and can be seen as \textit{parallel} quantum gates applied on separate qubit blocks.
\begin{figure}[b]
\centering
\includegraphics[width=\columnwidth]{fig_2d.png}
\caption{\label{fig:2D} a) An arbitrarily large 2D array of coupled qubits in a honeycomb array. Each link represents a ZZ coupling term. A yellow (cyan) vertex represents a driven (undriven) qubit. The total Hamiltonian can be decomposed into a sum of identical four-qubit blocks that commute with each other (enclosed in the dashed circles). An undriven qubit is shared by three neighboring blocks. (b) The pattern of driving for implementing a two-qubit gate, resulting in a central row of identical six-qubit blocks (enclosed by the dashed rectangles). The rest of the array can be decomposed into the four-qubit blocks as in (a).}
\end{figure}
We find that the simplest geometry that allows the decomposition of $H(t)$ into few-qubit commuting blocks is a honeycomb array of qubits with nearest neighbor ZZ coupling, as shown in Fig.~\ref{fig:2D}. We consider an alternating driving pattern where only the subarray colored in yellow in Fig.~\ref{fig:2D}a is driven, then $H(t)=\sum_{j\in \mathcal{L}}H_j(t)$ where $\mathcal{L}$ is the driven subarray and
\begin{align}\label{eq:blockH}
H_j(t)=\frac{1}{2} \left[\Omega_{j}(t) \sigma^x_{j}+\Omega'_{j}(t) \sigma^y_{j}\right]+\sum_{k\in \text{NB}_j}J_{jk}\sigma_{j}^z \sigma_{k}^z,
\end{align}
where $\text{NB}_j$ is the set of the three nearest neighbors of qubit $j$. Each of the block Hamiltonians, $H_j$, has only four qubits and they commute with each other. This can be seen in a more transparent way by the graphical representation in Fig.~\ref{fig:2D}a. Each link in the graph represents a $ZZ$ coupling term; a yellow vertex represents a single qubit driving term, and a cyan vertex represents an undriven qubit and hence nothing in $H(t)$. All the ZZ terms commute with each other. A yellow vertex does not commute only with the three vertices connected to it, because $\sigma^x$ and $\sigma^y$ do not commute with $\sigma^z$. Thus, the total Hamiltonian can be expressed as a sum of commuting four-qubit blocks enclosed by the dash circles in Fig.~\ref{fig:2D}a. Note that each block has only one driven qubit in the center. The qubits at the intersection of two neighboring blocks must not be driven for the commutativity to hold. We will show below that it is possible to implement a single qubit gate on the driven qubit without changing the state of the undriven qubits, at the end of the gate, despite the permanent ZZ interaction in the block.
The driving pattern need to be modified slightly for implementing two-qubit gates. In a conventional device with tunable couplers the qubit-qubit interaction is turned on only when a two-qubit entangling gate is applied. In our case the ZZ coupling is always on, and in general it entangles all the qubits at all times. However, we find that it is still possible to implement a specific two-qubit entanging gate, for example the CNOT gate, between two targeted qubits by driving both. Turning on the drives on two neighboring qubits results in the pattern of Fig.~\ref{fig:2D}b where the central row is built from identical six-qubit blocks. The rest of the array can be driven in the alternating pattern as before. The reader may wonder why the six-qubit blocks are required for the entire central row when only one two-qubit gate is needed. This is necessary for applying the identity operators on all undriven qubits for cancelling the actions of the fixed ZZ coupling, which requires that any undriven qubit must have at least one neighboring driven qubit (more details below). If a link, $J_{jk}\sigma^z_j\sigma^z_k$, is not connected to any driven qubit, then it commutes with every other terms in the Hamiltonian, and its contribution to the total unitary evolution is simply the factor $e^{-iJ_{jk}\sigma^z_j\sigma^z_k T}$, which cannot be cancelled due to the absence of control.
\subsection*{Applying gates using optimal control}\label{sec:gate}
\begin{figure}[t]
\centering
\includegraphics[width=0.85\columnwidth]{fig_block.png}
\caption{\label{fig:block} a) A star graph with undriven qubits on the boundary and a driven subet of one or more qubits in the center. Each undriven qubit must be connected to at least one driven qubit. By utilizing optimal control on the driven subset it is possible to apply a unitary of the form $U_{\mathcal{C}}\otimes I_{\mathcal{B}}$ where $U_{\mathcal{C}}$ is a target unitary on the central driven subset, and $I_{\mathcal{B}}\equiv I\otimes I\otimes\dots\otimes I$ is the identity in the Hilbert space of the undriven qubits on the boundary (see text). b) Examples of applying a Hamadard gate on the four-qubit block and CNOT (CX) gate on the six-qubit block of Fig.~\ref{fig:2D}. In both cases the identity operators are applied on all undriven qubits.}
\end{figure}
We now describe how to apply targeted gates on the driven qubits while at the same time apply the identity operators on the neighboring undriven qubits. Note that the four and six qubit blocks in Fig.~\ref{fig:2D}a and \ref{fig:2D}b have the form of a star graph where only a central subset of qubits is driven, as depicted in Fig.~\ref{fig:block}a. Our method lies in the key numerical finding that, for such a star graph, it is possible to use optimal control algorithm to find pulse shapes, $\Omega_j^{x,y}(t)$ where $j\in \textit{driven subset}$, to implement a unitary operation of the type $U_{\mathcal{C}}\otimes I_{\mathcal{B}}$, where $U_{\mathcal{C}}$ is a unitary acting on the driven subset, and $I_{\mathcal{B}}$ the identity matrix acting on the undriven subset at the boundary. The net effect is that the gate $U_{\mathcal{C}}$ is applied to the driven subset while the rest remains unchanged. If the driven subset has one (two) qubit, then $U_{\mathcal{C}}$ is a single qubit (two-qubit) gate.
Obviously the qubits on the boundary are acted on by the ZZ interactions, and hence their states are changed, during the pulse, but by choosing the right shape one can use the combined effect of the central driving term and the ZZ connectors to ensure that the identity operators are applied at then end of the pulse, removing the ZZ interactions in a stroboscopic fashion. This can be partly understood by looking at the Baker–Campbell–Hausdorff expansion. For the four-qubit block with the Hamiltonian of Eq.~\eqref{eq:blockH} , for example, the unitary evolution in a small time step is
\begin{align}
e^{-i H_j(t) dt}&= e^{-i \left(H^d_j(t)+H^{int}_j\right)dt}\nonumber \\
&\approx e^{-i H_j^d(t) dt}e^{-i H^{int}_j dt}e^{-[H^d_j(t),H^{int}_j]dt^2/2},
\end{align}
where $H^d_j(t)\equiv (1/2) \left[\Omega_{j}(t) \sigma^x_{j}+\Omega'_{j}(t) \sigma^y_{j}\right]$ is the driving term, $H^{int}_j\equiv \sum_{k\in \text{NB}_j}J_{jk}\sigma_{j}^z \sigma_{k}^z$ the ZZ interaction terms, and $[H^d_j(t),H^{int}_j]$ the commutator of the two. While the first term is responsible for applying a gate on the driven qubit, the second is the unwanted evolution due to the ZZ interaction. Since $[H^d_j(t),H^{int}_j]=\sum_{k\in \text{NB}_j}iJ_{jk}\left[\Omega_j(t)\sigma_{j}^x-\Omega'_j(t)\sigma_{j}^y\right] \sigma_{k}^z$, it is obvious that the third term allows partial control of the undriven qubits, labelled by $k$, through shaping $\Omega_j^{x,y}(t)$; and we find that this is sufficient for undoing the evolution due to the ZZ interaction.
Using optimal control we are able to obtain pulses for realising the unitary operator $U_{\mathcal{C}}\otimes I_{\mathcal{B}}$ with maximum fidelity, $F=1$, up to numerical precision, where $U_{\mathcal{C}}$ is the Hadamard, $\pi/8$ and the \emph{direct} identity gate on the the single driven qubit of the four-qubit block (we use ``\emph{direct} identity gate" to refer to an identity gate applied on an driven qubit to differentiate it from the identity operators applied on the undriven qubits). The same was achieved where $U_{\mathcal{C}}$ is the CNOT gate and the direct two-qubit identity gate, $I\otimes I$, on the two driven qubits of the six-qubit block of Fig.~\ref{fig:block}b. Note that in all of these examples the identity operators are applied on the undriven qubits. These one and two qubit gates form a unversal set, i.e., a set from which any muti-qubit unitary can be approximated from with arbitrary precision, \emph{allowing the implementation of an arbitrary quantum circuit} \cite{Nielsen2000}. More details of the optimal control algorithm and pulse shapes are given later where we discuss the robustness of the these gates against parameter uncertainty in the Hamiltonian.
\subsection*{Implementing quantum circuits}
An implementation of quantum computation on the array is illustrated in Fig.~\ref{fig:circuit} where the driven subarray is varied from one step to the next to apply the target gate on the right qubits. As can be seen in Fig.~\ref{fig:circuit} the identity operators are applied on the undriven qubits at every step. Note that the commuting blocks are not fixed, but are changed constantly during the execution of a quantum circuit, depending on where the gates are applied and whether they are single or two-qubit gates.
We show in Fig.~\ref{fig:driving_pattern} a simple example of how the driven/undriven subarrays and the blocks are varied during the implementation of a simple quantum circuit. Consider the following sequence of gates on two qubits, denoted by A and B: a two-qubit gate on A and B, followed by a single qubit gate on A, and then a single qubit gate on B. Note how the driven/undriven subarrays and the blocks are changed at each step. In step 1 both qubits, A and B, are driven in a six-qubit block, in step 2 only qubit A is driven in a four-qubit block and qubit B now becomes an undriven qubit, and in step 3 qubit B is driven and qubit A undriven. At each step gates can also be implemented in parallel on the driven qubits other than A and B. This parallel processing helps reduce the number of steps in a computation. If there is no gate on a driven qubit at a given step one simply applies the direct identity gate to keep its state unchanged. The undriven qubits are always subjected to the identity operators at all steps. Following this example it is straightforward to derive the driving pattern for an arbitrary quantum circuit.
\begin{figure}[h]
\centering
\includegraphics[width=\columnwidth]{fig_driving_pattern.png}
\caption{\label{fig:driving_pattern} An illustration of how the blocks are changed/shifted during an implementation of the simple quantum circuit described in the text.}
\end{figure}
In our method the undriven subarray is crucial for the Hamiltonian decomposition into commuting blocks, but this means that there are always qubits that have no gate at a given step, leading to an overhead in the number of steps compared with conventional quantum computation with tunable couplers. The exact amount of this overhead depends on the gate configuration of a circuit. However, as shown in Fig.~\ref{fig:2D}, the number of undriven qubits is no more than half of the total qubits, and it can be shown that the overhead is at worst a factor of 2, which does not change the computational complexity of a quantum algorithm.
Now we discuss \emph{readout} for the qubit array which obviously takes a finite duration. At the end of the computation the drives are turned off and the system evolves according to the background Hamiltonian $H_0=\sum_{j,k}J_{j,k}\sigma_{j}^z \sigma_{k}^z$. The final wavefunction can be expanded in the computational basis, $\ket{\psi_f}~=~\sum_{j_1,...,j_N = 0,1}c_{j_1,...,j_N}\ket{j_1,...,j_N}$. Since $H_0$ consists of exclusively $\sigma^z$ terms, $\ket{j_1,...,j_N}$ is an eigenstate of $H_0$ and the evolution under $H_0$ introduces only addition phases in the coefficients $c_{j_1,...,j_N}$. Since the solution of a computation is usually encoded in the bit string of the final wavefunction, as is the case for the Shor's algorithm for factoring \cite{Shor1997} and HHS algorithm for solving linear systems \cite{Harrow2009}, the additional phase will not change the result.
\subsection*{Robustness against parameter uncertainty}\label{sec:robustness}
\begin{table*}[ht]
\centering
\begin{center}
\begin{tabular}{ |c|c|c|c|c|c|c|c| }
\hline
&&&&&&&\\[-1em]
$\Delta J/\bar{J}$ & $\Delta \alpha$ & $\Delta \delta/\bar{J}$ & $H\otimes I_\mathcal{B}$ & $\pi/8\otimes I_\mathcal{B}$ & $I\otimes I_\mathcal{B}$ & CX$\otimes I_\mathcal{B}$ & $I_2\otimes I_\mathcal{B}$
\\ \hline
&&&&&&&\\[-1em]
0\% & 0\% & 0\% & 10 & 10 & 10 & 10 & 10 \\ \hline
&&&&&&&\\[-1em]
0.1\% & 0.1\% &0.1\% & 5.6 (4.5) & 5.5 (4.6) & 5.5 (4.7) & 4.4 (4.1) & 4.6 (4.3) \\ \hline
&&&&&&&\\[-1em]
1\% & 1\% &0.1\% & 5.6 (2.8) & 5.5 (2.7) & 5.6 (2.7)&3.7 (2.4) & 4.2 (2.3)\\ \hline
&&&&&&&\\[-1em]
5\% & 5\% &0.1\% & 5.3 (1.4) & 5.3 (1.2) & 5.4 (1.1)&3.2 (1.1) &4.0 (0.9)\\
\hline
\end{tabular}
\caption{Maximum worst-case fidelities. The figures shown are the exponents of the infidelity, i.e., $-\log_{10}\left(1-\mathcal{F}_{\text{max}}\right)$, which is the number of nines in $\mathcal{F}_{\text{max}}$. The fidelities are calculated for the single qubit gates on the four-qubit block, and the two-qubit gates on the six-qubit block of Fig.~\ref{fig:2D}. Four levels of uncertainty in the qubit-qubit coupling strength and control amplitude are considered, while the uncertainty in the detuning is kept at $0.1\%$ of the coupling strength. The maximum amplitude of the Rabi frequencies is constrained to less than $10\bar{J}$. $I$ and $I_2$ are the direct identity gates for one and two qubits, respectively. The gate duration is $T=2\pi/\bar{J}$, divided into $M=100$ time bins. The figures inside the bracket are results obtained with non-robust optimization.}
\label{tab:gates}
\end{center}
\end{table*}
We now describe the optimal control algorithm for maximising the fidelity of the unitary $\mathcal{U_C}\otimes\mathcal{I_B}$ and how to make this unitary robust against parameter uncertainty in the Hamiltonian. We divide the pulse duration, $T$, into $M$ time bins of interval $\Delta t$. In each time bin the field amplitudes are kept constant. The set of the Rabi frequencies form the control vector, $\bm{c}=\{\Omega_{jn}^{\mu}: 1\leq n\leq M; \mu=x,y; j\in \mathcal{C} \}$, where $\mathcal{C}$ is the driven subset. The unitary evolution $U_{\mathcal{G}}(T)$ of the star graph $\mathcal{G}$ is then a function of $\bm{c}$. Each qubit-qubit coupling, $J_{jk}$, and detuning, $\delta_j$, is allowed to vary independently in the uncertainty intervals $[\bar{J}-\Delta J/2,\bar{J}+\Delta J/2]$, and $[\bar{\delta}-\Delta \delta/2,\bar{\delta}+\Delta \delta/2]$, respectively. The uncertainty in the Rabi frequencies can be caused by that in the dipole-matrix elements, or a slow drift in the drive leading to changes in the field amplitudes from one experiment to the next. This can be modelled by replacing $\Omega_j^{x,y}(t)$ in $H_\mathcal{G}$ by $\alpha_j \Omega_j^{x,y}(t) $ where $\alpha_j$ is a dimensionless parameter that varies in the interval $[1-\Delta \alpha/2,1+\Delta \alpha/2]$. Now the unitary $U_{\mathcal{G}}(T)$ also depends on $J_{jk}$, $\alpha_j$, and $\delta_j$. The qubit-qubit coupling and dipole-matrix element of a qubit have to be measured or estimated at the point of fabrication, so their values can change substantially after more qubits are added to the array due to additional interaction or experimental drift. In contrast, we find that it is possible to determine the frequency of every qubit in the completed array with typically very precise spectroscopic measurement. In the array the resonant frequency of each qubit is shifted due to the ZZ interactions, but there exists a procedure of one and two-photon absorption measurements that can be combined to cancel these shifts and obtain the bare qubit frequency, $\omega_j$ (see Method). Thus, we assume that the driving fields are tuned to resonance, $\bar{\delta}=0$, with residual detuning uncertainties much smaller than the average qubit-qubit interaction, $\Delta \delta_j/\bar{J}=0.1\%$, which is typical for superconducting qubits \cite{footnote}. This small detuning uncertainty should be achievable in most physical realisations of qubits owing to the high accuracy of spectroscopic measurements.
Denote the set of these uncertain parameters as $\bm{v}$, then the robust optimal control problem can be defined as a max-min optimization problem: We find an optimal control that maximizes the minimum fidelity over $\bm{v}$,
\begin{align}\label{eq:fidelity}
\mathcal{F}_{\text{max}}= \max_{\bm{c}} \mathcal{F}(\bm{c}), \quad \mathcal{F}(\bm{c})=\min_{v\in\mathcal{V}} F(\bm{c},\bm{v}),
\end{align}
where $\mathcal{V}$ is the hypercube containing the possible values of $\bm{v}$, and $F(\bm{c},\bm{v})$ the gate fidelity based on the trace distance
\begin{align}
F(\bm{c},\bm{v})=\left\vert\frac{1}{D}\, \text{tr}\!\left[U_{\mathcal{G}}^{\dagger}(T)\left(U_{\mathcal{C}} \otimes I_{\mathcal{B}} \right)\right] \right \vert^2,
\end{align}
where $D$ is the dimension of the Hilbert space of $\mathcal{G}$, $I_{\mathcal{B}}$ the identity matrix of the subset of undriven qubits on the boundary, and $U_{\mathcal{C}}$ the target unitary that we want to apply on the central driven subset. In numerical computation one chooses a set of sampling points, $\bm{v}_i$ , in $\mathcal{V}$, and find the minimum fidelity in this set. We found that when the uncertainties are all smaller than $5\%$ and $F(\bm{c},\bm{v})$ is larger than $99\%$ its minimum over $\bm{v}$ always lies at one of the extreme points of $\mathcal{V}$, i.e., one of the corners of the hypercube. Thus, we can redefine $\mathcal{F}(\bm{c})\equiv \min_{\bm{v}_i\in\mathcal{X}} F(\bm{c},\bm{v}_i)$ where $\mathcal{X}$ is the discreet set of the extreme points of $\mathcal{V}$. This drastically reduces the number of sampling points where $F(\bm{c},\bm{v})$ has to be computed. There are $2^{n_u}$ extreme points in a hypercube of $n_u$ uncertain parameters. For example, the 6-qubit block in Fig.~\ref{fig:2D}b has 9 uncertain parameters, five $J$s, two $\alpha$s, and two $\delta$s, so we need to compute $2^9\equiv 512$ values of $F(\bm{c},\bm{v}_i)$ for any given $\bm{c}$.
We develop a numerical computation for gradient-based robust optimal control that can handle systems with up to 12 qubits and multiple uncertain parameters. Starting with a random initial guess for $\bm{c}$, we use gradients to identify a step $\delta \bm{c}$ to maximize $\min_{i}\nabla_{\bm{c}} {F(\bm{c},\bm{v}_i)}.\delta \bm{c}$, the first order increment, so that $F(\bm{c},\bm{v}_i)$ is increased for all $\bm{v}_i$. In this way $\mathcal{F}(\bm{c})$, the minimum fidelity over $\bm{v}_i$, can be increased to a value very close to one. In practice we find that one can also raise $\mathcal{F}(\bm{c})$ by maximizing the average fidelity, $\sum_{i=1}^{n_{\mathcal{X}}} F(\bm{c},\bm{v}_i)/n_{\mathcal{X}}$, with gradient ascent, which in some cases works faster than trying to increase $F(\bm{c},\bm{v}_i)$ for all $\bm{v}_i$. For the $n$-th time bin the time evolution is calculated by the mid-point rule $e^{-i H_{\mathcal{G}}(t_n- \Delta t/2)\Delta t}+O(\Delta t^3)$ which is accurate when the time step is small. The matrix exponentiation is sped up by using the Krylov subspace method on sparse matrices; and the gradient computed from a simple and efficient second-order formula \cite{deFouquieres2011}.
For the four-qubit block of Fig.~\ref{fig:2D}a, we derive optimal pulses to realize $U_{\mathcal{C}}\otimes I_{\mathcal{B}}$ where $U_\mathcal{C}$ is the Hadamard and $\pi/8$ gates \cite{Nielsen2000}. And for the six-qubit block of Fig.~\ref{fig:2D}b we want $U_{\mathcal{C}}$ to be the CNOT gate. These three gates form a universal set where any multi-qubit unitary can be approximated from with arbitrary precision \cite{Nielsen2000}, allowing universal quantum computing on the array. For a system with exclusively fixed coupling, it is also necessary to implement the identity gate that keeps all the qubits in a block unchanged despite the permanent interaction.
In Table~\ref{tab:gates} we show the robust fidelities obtained for the universal set and the identity gates at various levels of uncertainty. The gate duration is divided into $M=100$ time bins. For $1\%$ uncertainty the fidelity is higher than 99.999\% for the single qubit gates and exceeds 99.98\% for the two-qubit gates. Even if the uncertainty is as high as 5\% five-nines fidelities are still achieved for the single qubit gates, and above 99.94\% for the two-qubit gates. This can be improved by increasing the number of initial guesses, relaxing the constraints, or raising the number of control variables. The optimal pulse shape for the Hadamard gate is shown in Fig.~\ref{fig:optm}, and the pulse shapes for the other gates in Table~\ref{tab:gates} is given in the Supplemental Materials. We choose $T=2\pi/\bar{J}$ but this specific value of the duration is not essential; the same order of magnitude is achieved for the fidelities when $T$ is changed by 10\%. In order to see the effectiveness of robust optimal control we also calculate the fidelities with non-robust optimal control: We first neglect all the uncertainties and optimize the fidelity for the ideal case where $J_j=\bar{J}$, $\alpha_j=1$ and $\delta_j=0$ for all $j$, then we use the obtained optimal control, $\bm{c}_{\text{ideal}}$, to calculate the minimum fidelity in the hypercube, $\mathcal{F}=\min_{v\in\mathcal{V}} F(\bm{c}_{\text{ideal}},\bm{v})$. The results are shown in the parentheses of Table~\ref{tab:gates}. robust optimization improves the fidelities by two to three nines when the uncertanties are significantly large ($1\%$ and $5\%$).
\begin{figure}[b]
\centering
\includegraphics[width=0.9\columnwidth]{fig_optm_bar.png}
\caption{\label{fig:optm} Optimal pulse for the Hadamard gate at $1\%$ uncertainty in Table~\ref{tab:gates}.}
\end{figure}
\section*{Discussion}\label{sec:discuss}
A promising physical realisation of our model is the superconducting qubits including the flux and transmon qubits. A direct ZZ interaction between flux qubits can be realised by coupling the qubits inductively, as demonstrated in quantum annealers \cite{Johnson2011,Tennant2021}. There is another interesting scheme based on the inductive longitudinal coupling of the flux qubits with a common bus cavity \cite{Billangeon2015,Didier2015}, which can be scaled up to 2D arrays \cite{Richer2016}. In addition, a cross-Kerr ZZ interaction has been recently demonstrated for transmon qubits using a flux-tunable coupler \cite{Collodo2020}. While flux qubits are very good two-level systems and hence our results are applicable, for transmons the leakage to higher excited states must be accounted for in the optimal control algorithm \cite{Motzoi2009}. The ZZ coupling is also the natural interaction in a nuclear magnetic resonance quantum computer \cite{Vandersypen2001}, for which sophisticated pulse-shaping is available \cite{Yang2020}, making it a good test bed for our model.
In experimental realisations, the pulse duration, $T$, is limited by the coherence time, $T_2$, of the qubits. The decoherence rate of a block of $N'$ qubits is enhanced by a factor of approximately $N'$ in the worst case, giving rise to a lifetime of $T_2'=T_2/N'$. The fidelity of a multi-qubit unitary on the block is then bounded by $F\leq 1-T/T_2'\equiv 1-N'T/T_2$. Therefore, to achieve a fidelity $F$ for the four-qubit and six-qubit blocks in Table~\ref{tab:gates} the pulse duration needs to be shorter than $T_2(1-F)/N'$ where $N'=4$ and 6, respectively.
Although a honeycomb array is the focus of this paper, the qubits can be arranged in any physical shape that has the same connectivity, for example a square array with each qubit connected to only three nearest neighbors. Moreover, there exist driving patterns that satisfy the conditions of commutitativity and robust control for other geometries such as square arrays and one-dimensional chains (see Supplemental Materials). One can also envisage a hybrid architecture where large clusters of fixed couplers are connected with tunable couplers, keeping the number of required tunable couplers low. Such a modular structure can help ease the technological difficulties in scaling quantum computers.
To conclude, we find that it is feasible to implement quantum computing with accurate operations on a 2D qubit array with exclusively fixed couplers. The quantum gates are robust against significant uncertainty in the qubit's frequency, qubit-qubit and drive-qubit coupling caused by fabrication imperfection and/or slowly fluctuating fields. Both 1D and 2D geometries are possible with a robust optimization process capable of handling the minimal cluster size required for each scheme. Our proposal shows that scalability can be accelerated with simplified hardware architecture based on fixed longitudinal coupling schemes, thus motivating further development of this coupling in different physical platforms.
\section*{Method}
\subsection*{Calculating fidelity and gradient}
We first describe how the fidelity and its gradients are calculated with the midpoint rule. The time duration is divided into $M$ equal time bins with $t_0=0$ and $t_M=T$. The field amplitudes are kept constant during each time bin. The Hamiltonian of a star graph, $\mathcal{G}$, at the midpoint of the $n-$th interval from $t_{n-1}$ to $t_{n}$ is
\begin{equation}
H_{\mathcal{G},n}=\sum_{j\in \mathcal{C}} \alpha_j \left[ \Omega_{jn} \sigma^x_j+ \Omega'_{jn}\sigma^y_{j}\right]+\sum_{j,k\in \mathcal{G}}J_{jk}\sigma_j^z \sigma_k^z ,\nonumber
\end{equation}
where $\alpha_j$ is the dimensionless factor introduced to model the uncertainty in the Rabi frequencies (see the Main Text), $\mathcal{C}$ is the driven subset in the center of the graph and $J_{jk}\neq 0$ only for nearest neighbors, and
\begin{align}
\Omega_{jn}&=\Omega^x_{jn} \cos(\delta_j (t_n-\Delta t/2))+\Omega^y_{jn}\sin(\delta_j (t_n-\Delta t/2)),\nonumber \\
\Omega'_{jn} &= \Omega^y_{jn} \cos(\delta_j (t_n-\Delta t/2))-\Omega^x_{jn}\sin(\delta_j (t_n-\Delta t/2)),\nonumber
\end{align}
where $\Omega^{x,y}_{jn}$ are the Rabi frequencies of the driving field on the $j$-th driven qubit during the interval from $t_{n-1}$ to $t_{n}$. They are the elements of the $2MN_{\mathcal{C}}\times 1$ control vector, $\bm{c}$, where $N_{\mathcal{C}}$ is the number of qubits in the driving subset. The unitary evolution from $t_{n-1}$ to $t_{n}$, $U_n=e^{-i H_{\mathcal{G},n} \Delta t}+O(\Delta t^3)$, is computed with the $\text{expm}$ function in Matlab. For an efficient calculation of the fidelity and the gradients we compute and store all the $U_n$, and then obtain the forward and backward unitary propagation operators \cite{deFouquieres2011, Motzoi2011}, defined by
\begin{align}
U^f_n&=U_nU_{n-1}\dots U_1, \nonumber \\
U^b_{n+1}&=U_MU_{M-1}\dots U_{n+1}, \nonumber
\end{align}
using the recursive relations $U^f_n=U_n U^f_{n-1}$ and $U^b_{n+1}=U^b_{n+2}U_{n+1}$. Then the fidelity is
\begin{align}
F(\bm{c},\bm{v})=\left\vert\frac{1}{2^{N_{\mathcal{G}}}}\, \text{tr}\!\left[{
U_M^{f}}^{\dagger}\left(U_{\mathcal{C}} \otimes I_{\mathcal{B}} \right)\right] \right \vert^2, \nonumber
\end{align}
where $N_{\mathcal{G}}$ is the number of qubits in the star graph.
For computing the gradients, we note that
\begin{align}
H_{\mathcal{G},n}= \sum_{\mu=x,y}\sum_{j\in \mathcal{C}} \Omega^{\mu}_{jn} K^{\mu}_{jn}+\sum_{j,k\in \mathcal{G}}J_{jk}\sigma_j^z \sigma_k^z , \nonumber
\end{align}
where
\begin{align}
K^{x}_{jn}&=g_j\left[\sigma_j^x \cos(\delta_j (t_n-\Delta t/2)) - \sigma_j^y \sin(\delta_j (t_n-\Delta t/2))\right], \nonumber \\
K^{y}_{jn}&=\alpha_j\left[\sigma_j^y \cos(\delta_j (t_n-\Delta t/2)) + \sigma_j^x \sin(\delta_j (t_n-\Delta t/2))\right]. \nonumber
\end{align}
One can show that the derivative of $U_n\equiv e^{-i H_{\mathcal{G},n}\Delta t}$ with respect to $\Omega^{\mu}_{jn}$ is \cite{deFouquieres2011}
\begin{align}
\frac{\partial U_n}{\partial \Omega^{\mu}_{jn}}=\left\{-i \Delta t K^{\mu}_{jn} +\frac{\Delta t^2}{2}\left[H_{\mathcal{G},n},K^{\mu}_{jn} \right]\right\}U_n+O(\Delta t^3), \nonumber
\end{align}
where $\left[H_{\mathcal{G},n},K^{\mu}_{j,n} \right]$ is a commutator. It follows that the derivative of $U_M^f\equiv U^b_{n+1}U_n U^f_{n-1}$ is
\begin{align}
\frac{\partial U_M^f}{\partial \Omega^{\mu}_{jn}}=U^b_{n+1}\left\{-i \Delta t K^{\mu}_{jn} +\frac{\Delta t^2}{2}\left[H_{\mathcal{G},n},K^{\mu}_{jn} \right]\right\}U^f_n, \nonumber
\end{align}
and from this it is straight forward to compute the gradient of the fidelity. The most computationally expensive part of the calculation is the matrix exponentiation for obtaining $U_n$, which is done $M$ times.
\subsection*{Robust optimisation}
We use two algorithms to raise the worst-case fidelity, $\mathcal{F}(\bm{c})$, defined in Eq.~\ref{eq:fidelity}. The first is based on sequential convex programming \cite{Allen2017}. We start with initial guesses $\bm{c}_0$ and $\bm{u}_0$ for the control vector and the upper limit (trust region) of the step, respectively. Then a step $|\delta \bm{c}|<\bm{u}_0$ is found to maximize $\min_{i}\nabla_{\bm{c}}F(\bm{c},\bm{v}_i).\delta\bm{c}$, i.e., to maximize the minimum first-order increment over $\bm{v}_i$. This ensures all the fidelities at the extreme points of the hypercube, $\mathcal{V}$, are increased. The above optimization problem can be solved by sequential convex programming (SCP). We used the YALMIP toolbox and SPDT3 package in Matlab for this purpose. If a step can be found such that $\min_{i}\nabla_{\bm{c}}F(\bm{c},\bm{v}_i).\delta\bm{c}$ is positive then we increase the upper bound $\bm{u}_0$ by 1.15, otherwise we decrease it by 2. We chose these factors as they give the fastest convergence in our numerical tests. The procedure is repeated until either the maximum iteration is reached or the step's upper bound drops below a small tolerance. The second approach is to simply maximize the average fidelity,
\begin{align}
\bar{\mathcal{F}}(\bm{c})=\sum_{i=1}^{n_{\mathcal{X}}} F(\bm{c},\bm{v}_i)/n_{\mathcal{X}},
\end{align}
using a quasi-Newton method. Obviously this does not guarantee that the worst-case fidelity over $\bm{v}$ is increased, as the mean can be increased without increasing the minimum value in the set. However, we found that in our calculations the worst-case fidelity is always improved substantially when we maximize the average fidelity. We optimize $\bar{\mathcal{F}}(\bm{c})$ using the interior-point method implemented in Matlab's fmincon function, where the Hessian is computed from the exact gradients with the BFGS approximation.
In our numerical tests the first algorithm is more sensitive on the initial guesses of the control parameters. For the two-qubit gates in Table~I the computation is very expensive and hence it is not practical to run the optimization with too many initial guesses. We find that for the same running time the second algorithm gives higher fidelities, and the results in Table~I are obtained with it.
\subsection*{Spectroscopic measurement of qubit frequency}
\begin{figure}[b]
\centering
\includegraphics[width=0.8\columnwidth]{fig_measure.png}
\caption{\label{fig:measure} The frequency of qubit 1 can be determined by one and two-photon absorption measurements on the four qubits of the numbered unit cell (see text).}
\end{figure}
In this section we propose a procedure using one and two-photon absorption measurements for determining the frequency of every qubit in the 2D honeycomb array. Initially all qubits in the array are in the ground state, $\ket{0}$. We want to measure the frequency of qubit 1 in Fig.~\ref{fig:measure}. Due to the ZZ interactions with neighboring qubits causing a dispersive shift the one-photon absorption peak of qubit 1 is at
\begin{align}\label{eq:1photon1}
\omega^p_1=\omega_1-2(J_{12}+J_{13}+J_{14}).
\end{align}
The shift is the energy cost for flipping three bonds from $\ket{00}$ to $\ket{01}$. Similarly, the one-photon absorption peaks for the other numbered qubits are
\begin{align}\label{eq:1photon2}
\omega^p_2&=\omega_2-2 J_{21}-2\sum_{\substack{k\in \text{NB}_2 \\k\neq 1}}{J_{2k}}, \nonumber \\
\omega^p_3&=\omega_3-2 J_{31}-2\sum_{\substack{k\in \text{NB}_3 \\k\neq 1}}J_{3k},\nonumber \\
\omega^p_4&=\omega_4-2 J_{41}-2\sum_{\substack{k\in \text{NB}_4 \\k\neq 1}}J_{4k},
\end{align}
where $\text{NB}_j$ is the set of nearest neighbors of qubit $j$. The last sums are the dispersive shifts caused by the ZZ intreaction with neighbors that are not qubit 1. Summing all the equations in Eqs.~\eqref{eq:1photon1} and \eqref{eq:1photon2}, and using $J_{kl}=J_{lk}$, one obtains
\begin{align}\label{eq:1photon}
\sum_{j=1}^{4}\omega^p_4=\sum_{j=1}^{4}\omega_j-4\sum_{j=2}^{4}J_{1k}-2\sum_{j=2}^4\sum_{\substack{k\in \text{NB}_j \\k\neq 1}} J_{jk}.
\end{align}
If both qubits 1 and 2 are subjected to the the same field, for example by splitting the field in a Mach–Zehnder interferometer set-up with one qubit in each arm, one can observe the two-photon transition $\ket{0}_1\ket{0}_2\rightarrow\ket{1}_1\ket{1}_2$. Now the coupling energy of the bond between qubit 1 and 2 are not changed, and the two-photon resonance happens at the frequency
\begin{align}\label{eq:2photon1}
2\omega^p_{12}&=\omega_1+\omega_2-2 (J_{13}+J_{14})-2\sum_{\substack{k\in \text{NB}_2 \\k\neq 1}}{J_{2k}},
\end{align}
Similarly, the two-photon resonance frequency for the other qubit pairs are
\begin{align}\label{eq:2photon2}
2\omega^p_{13}&=\omega_1+\omega_3-2 (J_{12}+J_{14})-2\sum_{\substack{k\in \text{NB}_3 \\k\neq 1}}{J_{3k}}, \nonumber \\
2\omega^p_{14}&=\omega_1+\omega_4-2 (J_{12}+J_{13})-2\sum_{\substack{k\in \text{NB}_4 \\k\neq 1}}{J_{4k}}.
\end{align}
Summing all the euqations in Eqs.~\eqref{eq:2photon1} and \eqref{eq:2photon2}, one obtains
\begin{align}\label{eq:2photon}
2\sum_{j=2}^{4} \omega^p_{1j}=2 \omega_1 +\sum_{j=1}^4\omega_j -4\sum_{j=2}^{4}J_{1k}-2\sum_{j=2}^4\sum_{\substack{k\in \text{NB}_j \\k\neq 1}} J_{jk}.
\end{align}
Finally, subtracting Eq.~\eqref{eq:1photon} from Eq.~\eqref{eq:2photon}, the dispersive shifts are cancelled and one arrives at
\begin{align}
\omega_1= \sum_{j=2}^{4}
\omega^p_{1j}-\frac{1}{2}\sum_{j=1}^{4}\omega^p_j, \nonumber
\end{align}
which gives the bare frequency of qubit 1 in terms of the one and two-photon resonant frequencies. In most qubit platforms spectroscopic measurements are precise and thus the uncertainty in the qubit's frequency can be small.
\subsection*{Data Availability}
The data for this work are available without restriction \cite{zenodo}.
\subsection*{Code Availability}
The Matlab code for this work are available without restriction \cite{zenodo}.
|
\section{Introduction}\label{sec.int}
Modal $\mu$-calculus, first proposed by Kozen in~\cite{Kozen82,Kozen83}, has been well known among modal logicians. Equipped with notations similar to $\lambda$-calculus~\cite{Arnold01}, modal $\mu$-calculus serves as an extremely expressive yet still decidable language (in fact EXPTIME-complete, {\em cf.}~\cite{Emerson88,Emerson99}) that is able to cover a wide range of variations of modal logic, especially for those study fields which are tightly related to theoretical computer science, such as temporal logic~\cite{Stirling01}, propositional dynamic logic~\cite{Streett81} as well as finite model theory~\cite{Gradel07}.
Rigidly speaking, semantics of modal $\mu$-calculus is commonly defined via Tarski's fixed-point theorem, {\em cf.}~\cite{Bradfield06}. However intuitively, {\em e.g.}, consider the following modal $\mu$-calculus formula $\nu Z.\phi\land[a]Z$, where $\nu$ is the dual of $\mu$. Here, what this $\nu Z.$ operator actually does is assigning the rest part of the formula, namely $\phi\land[a]Z$, to a formula variable $Z$. Hence, the subformula $\phi\land[a]Z$ is now able to talk about itself via this formula variable $Z$. Through such self-reference, our ``core'' formula $\phi$ gets recursively ``passed down'' along the binary relation $R_a$, so the original formula $\nu Z.\phi\land[a]Z$ intuitively says that $\phi$ is true along every $R_a$-path. So far so good. Then quite naturally, if we regard the binary relation $R_a$ as a \textit{horizontal} dimension in our Kripke model, we may also wonder whether this core formula $\phi$ can similarly be passed down recursively along other types of routes in the model, {\em e.g.}, on a \textit{vertical} dimension? Actually our answer is yes, and in parallel with modal $\mu$-calculus, we would like to call this novel kind of modal logic as modal $\xi$-calculus.
So after all, where lies this another \textit{vertical} dimension in modal $\xi$-calculus? Let us turn our attention to first-order modal logic, where we are able to discover another natural hierarchy, namely between the whole Kripke model $\mathcal{M}$ itself and any element $a$ that belongs to the (constant) domain of $\mathcal{M}$. We may also note that, {\em e.g.}, there exists arbitrary unary predicate $P$ so that $Pa$ forms an atomic formula of normal first-order modal logic, but surely, it will be much better if such unary predicates can be \textit{intrinsically} constructed, rather than \textit{extrinsically} designated as extra symbol like $P$. Thus, what if $a$ is a Kripke model as well, exactly the same type as $\mathcal{M}$ so that the hierarchy between $\mathcal{M}$ and $a$ is truly a well-defined \textit{vertical} dimension, and more importantly, any formula can also be evaluated in the model $a$ just as in the model $\mathcal{M}$ so that a formula itself can simply act like an unary operator! Moreover, since now $a$ is the same type of Kripke model like $\mathcal{M}$, then just as $\mathcal{M}$ possesses a domain which contains $a$, of course $a$ also possesses its own domain, and once again, any element in $a$'s domain is also the same type of Kripke model. Through such recursion, a tree-like \textit{vertical} genealogy gets established among a bunch of Kripke models, and so we call this type of Kripke model as genealogical Kripke model. Furthermore, since now any formula acts like an unary predicate, then just as modal $\mu$-calculus, once a formula is equipped with the ability to talk about itself, it can similarly get passed down recursively on the \textit{vertical} dimension, namely along a sequence like $a\in\text{\rm dom}(\mathcal{M})$, $a'\in\text{\rm dom}(a)$, $\cdots$. Thus more precisely, we call our logic as first-order modal $\xi$-calculus, abbreviated as $\text{\sf FOMC}$. For new notations that we adopt, loosely speaking, if $\phi$ is an $\text{\sf FOMC}$-formula, then adding a pair of question marks around it produces an unary predicate as $\text{\rm?`}\phi\text{\rm?}$, hence $\text{\rm?`}\phi\text{\rm?} a$ is also an $\text{\sf FOMC}$-formula with a structure parallel to that of $Pa$. Also, we introduce the $\xi X.$ operator as $\xi X.\phi$, very similar to modal $\mu$-calculus.
Importantly, here may we provide a few useful tips for readers to follow this paper more smoothly. In fact, no matter how detailedly we manage to explain our na\"{i}ve intuitions behind first-order modal $\xi$-calculus and genealogical Kripke models, instinctive inaccuracy of natural language might still present obstacles toward full comprehension. Therefore, readers are strongly recommended not only to carefully digest all the crucial definitions in Section~\ref{sec.pre}, but also to frequently take some preview on those vivid examples in Section~\ref{sec.exam}, which will certainly help to navigate readers' understanding through intuitive pictures. Generally speaking, although the formal definition may seem a little complex at first, once having completely understood, readers will certainly agree that first-order modal $\xi$-calculus is nothing but just a quite natural mathematical generalization of common modal $\mu$-calculus.
Speaking of practical application, we firstly have to determine how to interpret genealogical Kripke models with suitable realistic meaning. Probably, computer scientists feel it tempting to interpret a model $\mathcal{M}$ as a process on a modern computer, $\mathcal{M}$'s (possibly empty) domain contains all of its children processes, and every child process is again represented by a genealogical Kripke model. This is exactly how a modern computer operating system (such as Linux) commonly handles multiple concurrent processes: on the one hand, there exists a genealogical hierarchy between the parent process and the child process; on the other hand, every process is technically represented by the same data structure and thus applicable to all the same properties, namely in logic, all the same unary predicates. While a lot of other interpretations can also fit into our first-order modal $\xi$-calculus pretty well, within this paper we shall mainly focus on interpreting as computer processes to help to illustrate our logic more vividly. Meanwhile, the other \textit{horizontal} dimension, namely the binary relation $R$ in a genealogical Kripke model, is usually interpreted as temporal logic's chronological future for computer processes, and thus often assumed to be reflexive and transitive~\cite{Gabbay94}. Nevertheless, it may not always be so under every condition, and hence for the most general purpose we choose not to presume any restriction on $R$.
The rest of this paper is organized as the following. Section~\ref{sec.pre} strictly defines mathematical preliminaries. Section~\ref{sec.exam} provides several concrete examples to illustrate powerful practical usage of first-order modal $\xi$-calculus. Section~\ref{sec.bis} studies bisimulation of the logic. Section~\ref{sec.conc} concludes this paper and proposes some feasible directions for future work.
\section{Preliminaries}\label{sec.pre}
This section defines the necessary preliminaries. Should readers find any part of the content difficult to understand, referring to some examples in Section~\ref{sec.exam} as a vivid assistance is strongly suggested.
Also, paragraphs started with a $\bigstar$ mark contain important clarifications. Honestly speaking, this paper indeed includes a handful of concepts and definitions about which readers might be vulnerable to getting confused, perhaps either because the formal definition seems a bit complex (but actually quite natural and intuitive once fully understood), or because it is somewhat different from the ordinary definition (while noticing such subtle difference may prove to be crucial for correct understanding). Thus those clarifications are particularly prepared with very careful elaboration, just in case of various kinds of possible confusion, and so readers are barely encouraged to simply neglect them but had better read those clarifications as one organic, comprehensive part of this entire paper.
\begin{definition}[Formula in Language $\text{\sf FOMC}$]\label{def.for}
Given a countable set of propositional letters $\mathbf{P}$, and a countable set of constant symbols $\mathbf{C}$, well formed formula $\phi$ in Language $\text{\sf FOMC}$ is recursively defined as the following BNF:
\begin{align*}
\phi::=X\mid\top\mid p\mid\text{\rm?`}\phi\text{\rm?} x\mid\text{\rm?`}\phi\text{\rm?} c\mid\neg\phi\mid(\phi\land\phi)\mid\Box\phi\mid\forall x.\phi\mid\xi X.\phi
\end{align*}
where $p\in\mathbf{P}$, $c\in\mathbf{C}$, $x$ stands for arbitrary model variable and $X$ stands for arbitrary formula variable. In the rest of this paper, we will follow this routine to use lowercase letters for model variables and uppercase letters for formula variables.
$\perp$, $\lor$, $\to$, $\Diamond$ and $\exists$ are abbreviations defined as common.
\end{definition}
From Definition~\ref{def.for} we can tell that in Language $\text{\sf FOMC}$, there exist two distinct kinds of variables: model variables like $x,y$ and formula variables like $X,Y$. There can also exist some constant symbols like $c,d\in\mathbf{C}$, which may be called as model constants as well. Thus intuitively, as their names literally suggest, a model variable or a model constant stands for a model (precisely speaking a genealogical Kripke model, introduced later in Definition~\ref{def.model}), while a formula variable represents an $\text{\sf FOMC}$-formula. Readers will gradually become aware of all these notations' respective indispensable function in our logic.
For now let us firstly pay our attention to one issue: usually in logic, variables in a formula may be either free or bound, and so are they here, too. Roughly speaking in an $\text{\sf FOMC}$-formula, a $\forall x.$ operator binds a model variable $x$, while a $\xi X.$ operator (plus some extra condition) binds a formula variable $X$. The formal definition is as the following:
\begin{definition}[Free Variable and Bound Variable]\label{def.bound}
In any given $\text{\sf FOMC}$-formula $\phi$, a model variable $x$ either is bound by the nearest $\forall x.$ whose scope includes this $x$; otherwise it is a free model variable, namely, not within any $\forall x.$'s scope.
In any given $\text{\sf FOMC}$-formula $\phi$, a formula variable $X$ either is bound, iff it is firstly within the scope of the nearest $\text{\rm?`}\text{\rm?}$ pair, and then bound by the nearest $\xi X.$ whose scope includes this $\text{\rm?`}\text{\rm?}$ pair; otherwise it is a free formula variable.
\end{definition}
\begin{example}\label{eg.bound}
For a model variable $x$, whether free or bound is just like ordinary first order logic, {\em e.g.}, $x$ is bound in $\text{\sf FOMC}$-formula $\forall x.\text{\rm?`} p\text{\rm?} x$ and is free in $\text{\sf FOMC}$-formula $\text{\rm?`} p\text{\rm?} x$.
For a formula variables $X$, however, it must firstly be within the scope of a $\text{\rm?`}\text{\rm?}$ pair, after which could it be bound by some $\xi X.$ whose scope includes that $\text{\rm?`}\text{\rm?}$ pair. This ``extra condition'' requires very careful attention, {\em e.g.}, $X$ is bound in $\text{\sf FOMC}$-formula $\xi X.\text{\rm?`} X\text{\rm?} c$ but is free in $\text{\sf FOMC}$-formula $\xi X.X$ because it is not within the scope of any $\text{\rm?`}\text{\rm?}$ pair.
\end{example}
\phantom{nothing}
\noindent$\bigstar$ \textbf{Important Clarification} Note that from Definition~\ref{def.bound}, a bound formula variable is not entirely similar to an ordinary bound model variable, {\em e.g.}, from Example~\ref{eg.bound} we know that $\xi X.X$ does \textit{not} bind its $X$. Intuitively, such stipulation is meant to ensure that circular evaluation will not occur when we later define semantics in Definition~\ref{def.sem}.
Together with the concept of free and bound variables arises the concept of logic sentences. The general idea as such is still quite common: a sentence, also called as a closed formula, is usually defined as a formula without any free variables. However here due to somewhat tangled nature of recursion in $\text{\sf FOMC}$-formulae, only after undergoing very careful analysis can we finally settle down to the following precise conditions in the definition of $\text{\sf FOMC}$-sentences:
\begin{definition}[Sentence in Language $\text{\sf FOMC}$]\label{def.sen}
An $\text{\sf FOMC}$-formula $\phi$ is an $\text{\sf FOMC}$-sentence, iff all of the following conditions hold:
\begin{itemize}
\item $\phi$ contains neither free model variables nor free formula variables.
\item If $\xi X.\psi$ is a subformula of $\phi$, then $\xi X.\psi$ contains neither free model variables nor free formula variables.
\item If $\text{\rm?`}\psi\text{\rm?} x$ or $\text{\rm?`}\psi\text{\rm?} c$ is a subformula of $\phi$, then $\psi$ contains no free model variables.
\end{itemize}
\end{definition}
\begin{remark}\label{rem.sen}
Actually, the formal logical language studied in this paper is $\text{\sf FOMC}$-sentences in Definition~\ref{def.sen}, rather than $\text{\sf FOMC}$-formulae in Definition~\ref{def.for}. Essentially, Definition~\ref{def.for} defines $\text{\sf FOMC}$-formulae by BNF, only as nothing more than an auxiliary notion. Those extra restrictions in Definition~\ref{def.sen} will later on ensure a well-defined semantics ({\em cf.}~Theorem~\ref{th.sem}), which cannot just be naturally generalized onto all $\text{\sf FOMC}$-formulae.
\end{remark}
Next, let us turn to the semantical aspect and define genealogical Kripke models, with respect to which semantics of $\text{\sf FOMC}$-sentences will be interpreted:
\begin{definition}[Genealogical Kripke Model]\label{def.model}
A genealogical Kripke model $\mathcal{M}$ is recursively defined as a tuple $(S_\mathcal{M},R_\mathcal{M},V_\mathcal{M},N_\mathcal{M},I_\mathcal{M},T_\mathcal{M})$ where:
\begin{itemize}
\item $S_\mathcal{M}$ is a non-empty set of possible worlds.
\item $R_\mathcal{M}\subseteq S_\mathcal{M}\times S_\mathcal{M}$ is a binary relation on $S_\mathcal{M}$.
\item $V_\mathcal{M}:\mathbf{P}\to\P(S_\mathcal{M})$ is a valuation function.
\item $N_\mathcal{M}$ is a set of genealogical Kripke models.
\item $I_\mathcal{M}:S_\mathcal{M}\times\mathbf{C}\to N_\mathcal{M}$ is a partial assignment function.
\item $T_\mathcal{M}:S_\mathcal{M}\times N_\mathcal{M}\to\bigcup\limits_{\mathcal{N}\in N_\mathcal{M}}S_\mathcal{N}$ is a function such that for any possible world $s\in S_\mathcal{M}$ and any model $\mathcal{N}\in N_\mathcal{M}$, $T_\mathcal{M}(s,\mathcal{N})\in S_\mathcal{N}$.
\end{itemize}
When the model $\mathcal{M}$ is clear from the context, we can omit $\mathcal{M}$ in the subscript and hence denote the model $\mathcal{M}$ simply as a tuple $(S,R,V,N,I,T)$.
\end{definition}
In case of any misunderstanding about the above Definition~\ref{def.model}, we would like to present (a little longer) clarification as well as several technical remarks right away, while elaboration on the intuitive interpretation of Definition~\ref{def.model} will soon follow after Remark~\ref{rem.model}.
\phantom{nothing}
\noindent$\bigstar$ \textbf{Important Clarification} In Definition~\ref{def.model}, a genealogical Kripke model $\mathcal{M}$ contains some part $N_\mathcal{M}$, which is again a set of genealogical Kripke models. Hence Definition~\ref{def.model} is a recursive definition, just like that a logic language is usually defined through BNF, and so by Axiom of Regularity in set theory~\cite{Jech03}, circular inclusion is forbidden here as well, therefore, we must have $\mathcal{M}\notin N_\mathcal{M}$. Note that in fact $N$ can be $\emptyset$, which exactly forms the basic case of this inductive definition; in other words, $N$ is automatically guaranteed to reach the $\emptyset$ dead point after a finite depth of generations, thus genealogical Kripke models are always well-founded and infinite downward chains will never occur.
Just in case of any likely confusion, here may we emphasize over this issue once again. Especially, readers familiar with modal $\mu$-calculus might naturally think that it is also possible for a genealogical Kripke model here to have itself as one of its own children models, namely $\mathcal{M}\in N_\mathcal{M}$; or more generally, to contain an infinite downward chain $\cdots\mathcal{M}_3\in N_{\mathcal{M}_2}$, $\mathcal{M}_2\in N_{\mathcal{M}_1}$, $\mathcal{M}_1\in N_\mathcal{M}$. Definitely, we agree that these situations are both conceptually conceivable and mathematically manageable, nonetheless, they simply do \textit{not} appear here because a recursive definition in ZFC automatically rules them out. Such ``restriction'' is essentially a direct result of the recursive nature of Definition~\ref{def.model}, and so actually, no extra restriction at all needs to be added into the definition. Readers are suggested to draw an analogy from the recursive BNF in Definition~\ref{def.for}, so that readers can also understand in a similar way, in Definition~\ref{def.model} we have made neither fault nor negligence. Finally for better explicitness, though a formal proof is omitted (which can be found in any standard textbook on set theory, {\em e.g.},~\cite{Kunen80}), we would like to rigorously summarize our main point in the above discussion as the following theorem:
\begin{theorem}\label{th.model}
There does not exist an infinite downward chain of genealogical Kripke models $\mathcal{M}_i$, where $i\in\omega$, such that $\forall i\in\omega$, $\mathcal{M}_{i+1}\in N_{\mathcal{M}_i}$.
\end{theorem}
\begin{remark}\label{rem.model}
The following are several other minor points concerning Definition~\ref{def.model}, to which readers are suggested to pay attention as well:
\begin{itemize}
\item $\mathcal{N}$ and $N$ are different: the former denotes some model, while the latter denotes some part of a model, which is a set of (other) models.
\item On the one hand, $I$ is generally a \textit{partial} function, i.e., it is not guaranteed to be total. In fact, since the codomain $N$ may be empty, sometimes $I$ is simply impossible to be total.
\item On the other hand, $T$ is a \textit{total} function, and when its codomain is empty, it is easy to see that its domain must also be empty, so this is not a problem.
\end{itemize}
\end{remark}
Having ensured that mathematically, readers can grasp a correct understanding about Definition~\ref{def.model}, we next provide a brief intuitive explanation on what every part of a genealogical Kripke model stands for. Assume that we interpret the genealogical Kripke model $\mathcal{M}$ as a computer process, then $N_\mathcal{M}$ represents all its children processes, and at every possible state $s\in S_\mathcal{M}$ of the parent process $\mathcal{M}$, for any constant symbol $c\in\mathbf{C}$, $I_\mathcal{M}(s,c)\in N_\mathcal{M}$ (if defined) selects a specific child process; for any child process $\mathcal{N}\in N_\mathcal{M}$, $T_\mathcal{M}(s,\mathcal{N})\in S_\mathcal{N}$ designates in which state this child process $\mathcal{N}$ currently is. Therefore fixing a constant symbol $c\in\mathbf{C}$, its interpretation $I_\mathcal{M}(s,c)$ still depends on $s$ and so may vary from possible world to possible world, thus even though we decide to call $c$ as a model ``constant'', readers should keep aware that $c$ is after all a \textit{local} constant rather than a \textit{global} one. (Of course if needed, we are also able to easily introduce another notion of global model constants.)
Anyway, since as of data structure, both the parent process $\mathcal{M}$ and the child process $\mathcal{N}$ are the same type of object, namely a process, it makes perfect sense that both of them are represented by a genealogical Kripke model so that they share exactly the same mathematical form. (By the way, $\mathcal{N}$ may also have its own children processes as well, which are therefore $\mathcal{M}$'s grandchildren processes.) Here then arises a quite natural question, however: should there be any correlation between the parent process's current state $s\in S_\mathcal{M}$ and the child process's current state $T_\mathcal{M}(s,\mathcal{N})\in S_\mathcal{N}$, where $\mathcal{N}\in N_\mathcal{M}$? We think the answer depends: in some cases it is quite plausible to assume such kind of relation, while in some other cases both the parent process and the children processes can act independently. Therefore, for the most general purpose, here we choose not to put any extra restrictions upon the relation between the parent process and the children processes.
We move on to define semantics of $\text{\sf FOMC}$-sentences. To start with, we define interpretations both for any model variable $x$ and for any formula variable $X$:
\begin{definition}[Interpretation]
An interpretation $i$ is a (partial) function. For any $(k,v)\in i$, intuitively, $i$ interprets the key $k$ as the value $v$. We also denote:
\begin{align*}
i[k:=v]=\left\{\begin{array}{ll}
i\cup\{(k,v)\}, & \quad\text{if }i(k)\text{ is undefined}\\
(i\setminus\{(k,v')\})\cup\{(k,v)\}, & \quad\text{if }i(k)=v',\text{ namely }(k,v')\in i
\end{array}\right.
\end{align*}
\end{definition}
\begin{definition}[Semantics]\label{def.sem}
Given a genealogical Kripke model $\mathcal{M}$ and an $\text{\sf FOMC}$-sentence $\phi$, valuation of $\phi$ over $\mathcal{M}$ (with two interpretations $i$ and $j$, the first one for model variables while the second one for formula variables, and initially $i=j=\emptyset$) is recursively defined as $\llbracket\phi\rrbracket^\mathcal{M}_{\emptyset,\emptyset}\subseteq S_\mathcal{M}$ by the following:
\begin{align*}
\llbracket X\rrbracket^\mathcal{M}_{i,j}= & \llbracket\psi\rrbracket^\mathcal{M}_{i,j},\text{ where }j(X)=\psi\\
\llbracket\top\rrbracket^\mathcal{M}_{i,j}= & S_\mathcal{M}\\
\llbracket p\rrbracket^\mathcal{M}_{i,j}= & V_\mathcal{M}(p)\\
\llbracket\text{\rm?`}\phi\text{\rm?} x\rrbracket^\mathcal{M}_{i,j}= & \{s\in S_\mathcal{M}\mid T_\mathcal{M}(s,\mathcal{N})\in\llbracket\phi\rrbracket^\mathcal{N}_{\emptyset,j}\},\text{ where }i(x)=\mathcal{N}\\
\llbracket\text{\rm?`}\phi\text{\rm?} c\rrbracket^\mathcal{M}_{i,j}= & \{s\in S_\mathcal{M}\mid I_\mathcal{M}(s,c)\text{ is defined and }T_\mathcal{M}(s,I_\mathcal{M}(s,c))\in\llbracket\phi\rrbracket^{I_\mathcal{M}(s,c)}_{\emptyset,j}\}\\
\llbracket\neg\phi\rrbracket^\mathcal{M}_{i,j}= & S_\mathcal{M}\setminus\llbracket\phi\rrbracket^\mathcal{M}_{i,j}\\
\llbracket\phi\land\psi\rrbracket^\mathcal{M}_{i,j}= & \llbracket\phi\rrbracket^\mathcal{M}_{i,j}\cap\llbracket\psi\rrbracket^\mathcal{M}_{i,j}\\
\llbracket\Box\phi\rrbracket^\mathcal{M}_{i,j}= & \{s\in S_\mathcal{M}\mid\text{for all }t\in S_\mathcal{M}\text{ such that }sR_\mathcal{M} t,t\in\llbracket\phi\rrbracket^\mathcal{M}_{i,j}\}\\
\llbracket\forall x.\phi\rrbracket^\mathcal{M}_{i,j}= & S_\mathcal{M}\cap(\bigcap_{\mathcal{N}\in N_\mathcal{M}}\llbracket\phi\rrbracket^\mathcal{M}_{i[x:=\mathcal{N}],j})\\
\llbracket\xi X.\phi\rrbracket^\mathcal{M}_{i,j}= & \llbracket\phi\rrbracket^\mathcal{M}_{i,j[X:=\phi]}
\end{align*}
For any genealogical Kripke model $\mathcal{M}$, any possible world $s\in S_\mathcal{M}$ and any $\text{\sf FOMC}$-sentence $\phi$, we denote $\mathcal{M},s\vDash\phi$ iff $s\in\llbracket\phi\rrbracket^\mathcal{M}_{\emptyset,\emptyset}$.
\end{definition}
Before proving soundness of Definition~\ref{def.sem} in Theorem~\ref{th.sem}, we would like to elaborate further on what on earth this definition of semantics intuitively captures:
\begin{itemize}
\item For $\llbracket X\rrbracket^\mathcal{M}_{i,j}$, if $j$ interprets the formula variable $X$ as $\psi$, then valuation of $X$ just equals to valuation of $\psi$. Later on we will show that in such a case $j(X)$ is guaranteed to be defined.
\item For $\llbracket\text{\rm?`}\phi\text{\rm?} x\rrbracket^\mathcal{M}_{i,j}$, the idea is to treat $\text{\rm?`}\phi\text{\rm?}$ as a unary predicate, hence if $i$ interprets the model variable $x$ as some child model $\mathcal{N}\in N_\mathcal{M}$, then for any parent model $\mathcal{M}$'s state $s\in S_\mathcal{M}$, $\text{\rm?`}\phi\text{\rm?} x$ is true at $\mathcal{M},s$ iff $\phi$ is true at the child model $\mathcal{N}$'s current state $T_\mathcal{M}(s,\mathcal{N})$, namely $T_\mathcal{M}(s,\mathcal{N})\in\llbracket\phi\rrbracket^\mathcal{N}_{\emptyset,j}$. Note that the interpretation $j$ for formula variables gets inherited but the interpretation $i$ for model variables is reset to $\emptyset$, because between two models $\mathcal{M}$ and $\mathcal{N}$, their language $\text{\sf FOMC}$ keeps the same but their domains $N_\mathcal{M}$ and $N_\mathcal{N}$ are separate. Later on we will also show that in such a case $i(x)$ is guaranteed to be defined.
\item For $\llbracket\text{\rm?`}\phi\text{\rm?} c\rrbracket^\mathcal{M}_{i,j}$, we simply attempt to interpret the constant symbol $c$ as a child model $I_\mathcal{M}(s,c)\in N_\mathcal{M}$ and then we can follow the similar idea as above.
\item For $\llbracket\forall x.\phi\rrbracket^\mathcal{M}_{i,j}$, all possible interpretations for the model variable $x$ as some child model $\mathcal{N}\in N_\mathcal{M}$ gets considered and added into the interpretation $i$, respectively. Then the compound universal formula values as the interaction of all possibilities, also together intersected with $S_\mathcal{M}$ just in case that $N_\mathcal{M}$ may be $\emptyset$.
\item For $\llbracket\xi X.\phi\rrbracket^\mathcal{M}_{i,j}$, just like modal $\mu$-calculus, this $\xi X.$ operator here intuitively assigns the following formula $\phi$ to the formula variable $X$, thus this interpretation gets added into $j$.
\end{itemize}
\phantom{nothing}
\noindent$\bigstar$ \textbf{Important Clarification} Readers familiar with semantics of modal $\mu$-calculus may wonder, why we have not made use of Tarski's fixed-point theorem in the definition of semantics, just as modal $\mu$-calculus normally does. The reason is that from Theorem~\ref{th.model}, we already know that a genealogical Kripke model can only have a finite depth of generations, hence the fixed point can simply be found by induction through finitely many steps. We also point out that unlike modal $\mu$-calculus, here $\phi$ in $\xi X.\phi$ does not have to be positive, either. Despite these minor technical differences, after all, our first-order modal $\xi$-calculus indeed generalizes the ordinary modal $\mu$-calculus onto the \textit{vertical} dimension through keeping the very core idea totally intact. Readers who may still possess doubts about this point hitherto are extremely eagerly suggested, in any case, to carefully reread Section~\ref{sec.int} --- the very beginning introductory part --- and especially the first several paragraphs, so as to amply convince themselves with quite intuitive justification.
\begin{remark}
Another point we would like to explain about the semantics in Definition~\ref{def.sem} is, why we always designate that $\mathcal{M},s\nvDash\text{\rm?`}\phi\text{\rm?} c$ when $I_\mathcal{M}(s,c)$ is not defined. This might simply seem like our factitious choice, but actually we have some good reasons to justify it. Let us interpret genealogical Kripke models as computer processes, then $I_\mathcal{M}(s,c)$ is essentially a process pointer that either points to one of $\mathcal{M}$'s children processes $\mathcal{N}\in N_\mathcal{M}$, or is the \textbf{NULL} pointer, which can be regarded as pointing to certain fixed \textbf{NULL} process. Therefore if $I_\mathcal{M}(s,c)$ is undefined, then for any fixed $\text{\sf FOMC}$-formula $\phi$, we should anticipate either always $s\in\llbracket\text{\rm?`}\phi\text{\rm?} c\rrbracket^\mathcal{M}_{i,j}$ or always $s\notin\llbracket\text{\rm?`}\phi\text{\rm?} c\rrbracket^\mathcal{M}_{i,j}$, i.e., depending on how this default \textbf{NULL} process is designed in advance but not depending on the specific current state $s\in S_\mathcal{M}$. Here we just arbitrarily designate that $s\notin\llbracket\text{\rm?`}\phi\text{\rm?} c\rrbracket^\mathcal{M}_{i,j}$ for every $\phi$, which does not really cause an issue, because the expected behavior of $\text{\sf FOMC}$-formula $\text{\rm?`}\phi\text{\rm?} c$ when letting $s\in\llbracket\text{\rm?`}\phi\text{\rm?} c\rrbracket^\mathcal{M}_{i,j}$ can be exactly simulated by $\text{\sf FOMC}$-formula $\neg\text{\rm?`}\neg\phi\text{\rm?} c$ when letting $s\notin\llbracket\text{\rm?`}\neg\phi\text{\rm?} c\rrbracket^\mathcal{M}_{i,j}$.
\end{remark}
At last, before finishing this section of mathematical preliminaries, we have to complete the indispensable task of proving the following Theorem~\ref{th.sem}, showing that our semantics is surely well defined with respect to $\text{\sf FOMC}$-sentences.
\begin{theorem}\label{th.sem}
Semantics in Definition~\ref{def.sem} is well defined.
\end{theorem}
\begin{proof}
For any $\text{\sf FOMC}$-sentence $\phi$, we recursively evaluate $\llbracket\phi\rrbracket^\mathcal{M}_{\emptyset,\emptyset}$.
On the one hand, we claim that this evaluating process always goes on well. Namely, we have to make sure that whenever necessary, $i$ and $j$ will always provide interpretation for any model variable $x$ and any formula variable $X$, respectively. This feature is guaranteed by conditions in Definition~\ref{def.sen}, intuitively as the following:
\begin{itemize}
\item $\phi$ contains neither free model variables nor free formula variables. This clause is just the ordinary closed-formula condition.
\item
\begin{enumerate}
\item If $\xi X.\psi$ is a subformula of $\phi$, then $\xi X.\psi$ contains no free model variables, or equivalently, $\psi$ contains no free model variables. This half clause is needed because $X$ may appear in a subformula of $\psi$ as $\text{\rm?`}\cdots X\cdots\text{\rm?} x$ or $\text{\rm?`}\cdots X\cdots\text{\rm?} c$, namely, to be passed down to a child model. As explained by the intuition of semantics defined in Definition~\ref{def.sem}, only interpretation for formula variables gets inherited by the child model while interpretation for model variables is reset to $\emptyset$, thus, $\psi$ should not contain any free model variables.
\item If $\xi X.\psi$ is a subformula of $\phi$, then $\xi X.\psi$ contains no free formula variables, either. This half clause will become of use later on, i.e., in the latter on-the-other-hand part.
\end{enumerate}
\item If $\text{\rm?`}\psi\text{\rm?} x$ or $\text{\rm?`}\psi\text{\rm?} c$ is a subformula of $\phi$, then $\psi$ contains no free model variables, due to the same intuitive reason as the above clause.
\end{itemize}
Based on the above intuitive intentions of all the clauses in Definition~\ref{def.sen}, it then immediately becomes just self-evident that $i(x)$ and $j(X)$ will always be properly defined whenever in need during the whole evaluating process.
On the other hand, we claim that this evaluating process will eventually terminate instead of going on infinitely. In fact, the only possible intermediate step when the current subformula may become more complicated is rewriting $\llbracket X\rrbracket^\mathcal{N}_{i,j}$ to $\llbracket\psi\rrbracket^\mathcal{N}_{i,j}$. This indicates $j(X)=\psi$, so $\xi X.\psi$ is a subformula of $\phi$, and since $\phi$ is an $\text{\sf FOMC}$-sentence, $\xi X.\psi$ contains no free formula variables. Therefore, appearance of any formula variable $Y$ in $\psi$ must be within the scope of a $\text{\rm?`}\text{\rm?}$ pair, and thus if we continue evaluating $\llbracket\psi\rrbracket^\mathcal{N}_{i,j}$ and later on encounter another intermediate stage $\llbracket Y\rrbracket^{\mathcal{N}'}_{i',j'}$, then $\mathcal{N}'$ must be a direct descendant of $\mathcal{N}$. By Theorem~\ref{th.model}, $\mathcal{N}'\neq\mathcal{N}$ and the maximal depth of generations in model $\mathcal{M}$ is finite. Not only do we need finite steps to descend from $\mathcal{N}$ to one of its descendants $\mathcal{N}'$, but also we have finite depth of generations to descend, namely, neither circle nor infinite chain will occur and thus recursively evaluating $\llbracket\phi\rrbracket^\mathcal{M}_{\emptyset,\emptyset}$ is well founded.
\end{proof}
\section{Examples}\label{sec.exam}
We have seen that theoretically, first-order modal $\xi$-calculus is just a very natural mathematical generalization of \textit{horizontal} modal $\mu$-calculus onto the other \textit{vertical} dimension. While practically, first-order modal $\xi$-calculus can depict the recursive structure of computer processes in a genealogical Kripke model, which will be demonstrated by a series of typical examples in the following that cover some of the most interesting topics in concurrency control~\cite{Breshears09}. Starting easily, we first reveal some simple examples without the $\xi X.$ operator.
\phantom{nothing}
\noindent$\bigstar$ \textbf{Important Clarification} As having been introduced in Section~\ref{sec.int}, now the binary relation $R$ is interpreted as temporal logic and thus assumed to be reflexive and transitive~\cite{Goldblatt92}. Therefore for neatness, when drawing the graph for a model we may omit some $R$ relation arrows, but readers should always be aware that the actual relation $R$ must be understood as the reflexive and transitive closure.
\begin{example}\label{exa.de}
As the following graph shows, the parent process $\mathcal{M}$ possesses two children processes $\mathcal{N}_1$ and $\mathcal{N}_2$, with one propositional letter $r$ representing that a process is currently running and one constant symbol $c$ that points to the currently running child process. We have $\mathcal{M},s_0\vDash\Box\exists x.\text{\rm?`} r\text{\rm?} x\land\neg\exists x.\Box\text{\rm?`} r\text{\rm?} x\land\Box\text{\rm?`} r\text{\rm?} c$, which intuitively reads as: there always exists some running child process, but no single child process is running forever, and it is always true that the currently running child process is currently running. Divergence between the first two clauses is well known in first-order modal logic as \textit{de dicto} vs. \textit{de re}~\cite{Hughes96}, while the last clause simply claims something trivial under our present interpretation.
$$\xymatrix@C=10pt@R=15pt{
\mathcal{M}: & \txt{$s_0$\\$c:=\mathcal{N}_1$}\ar[rr] & & \txt{$s_1$\\$c:=\mathcal{N}_2$}\ar[rr] & & \txt{$s_2$\\$c:=\mathcal{N}_1$}\\
\mathcal{N}_1,running\ar@{--}[ur] & \mathcal{N}_2,stopped\ar@{--}[u] & \mathcal{N}_1,stopped\ar@{--}[ur] & \mathcal{N}_2,running\ar@{--}[u] & \mathcal{N}_1,running\ar@{--}[ur] & \mathcal{N}_2,dead\ar@{--}[u]\\
\mathcal{N}_1: & \txt{$stopped$\\$\neg r$}\ar@<1ex>[r] & \txt{$running$\\$r$}\ar@<1ex>[l]\ar[r] & \txt{$dead$\\$\neg r$}\\
\mathcal{N}_2: & \txt{$stopped$\\$\neg r$}\ar@<1ex>[r] & \txt{$running$\\$r$}\ar@<1ex>[l]\ar[r] & \txt{$dead$\\$\neg r$}
}$$
\end{example}
\begin{example}
As the following graph shows, $\mathcal{N}_1$ and $\mathcal{N}_2$ are two children processes that require both resources $a$ and $b$ to run. $\mathcal{N}_1$ asks for resource $a$ first while $\mathcal{N}_2$ does the opposite, hence in theory we can imagine that a dead lock easily occurs when $\mathcal{N}_1$ occupies resource $a$ and $\mathcal{N}_2$ occupies resource $b$. Nonetheless, in this actual case, $\mathcal{N}_1$ and $\mathcal{N}_2$'s common parent process $\mathcal{M}$ manages to forbid certain combinations of states so as to prevent any possible dead lock. Therefore we have $\mathcal{M},s_0\vDash\exists x.\exists y.(\text{\rm?`}\Diamond(a\land\neg b)\text{\rm?} x\land\text{\rm?`}\Diamond(\neg a\land b)\text{\rm?} y)\land\neg\Diamond\exists x.\exists y.(\text{\rm?`} a\land\neg b\text{\rm?} x\land\text{\rm?`}\neg a\land b\text{\rm?} y)$, namely, dead lock has been avoided under the parent process $\mathcal{M}$'s supervision.
$$\xymatrix@C=10pt@R=15pt{
& & \mathcal{N}_1,stopped\ar@{--}[dr] & \mathcal{N}_2,waiting\ar@{--}[d] & \mathcal{N}_1,stopped\ar@{--}[dr] & \mathcal{N}_2,running\ar@{--}[d]\\
& & & s_3\ar@(ur,l)[rr] & & s_4\ar@(dl,r)[dllll]\\
\mathcal{M}: & s_0\ar@(dr,l)[rr]\ar@(u,l)[urr] & & s_1\ar@(dr,l)[rr] & & s_2\ar@(ul,ur)[llll]\\
\mathcal{N}_1,stopped\ar@{--}[ur] & \mathcal{N}_2,stopped\ar@{--}[u] & \mathcal{N}_1,waiting\ar@{--}[ur] & \mathcal{N}_2,stopped\ar@{--}[u] & \mathcal{N}_1,running\ar@{--}[ur] & \mathcal{N}_2,stopped\ar@{--}[u]\\
\mathcal{N}_1: & \txt{$stopped$\\$\neg a,\neg b$}\ar[r] & \txt{$waiting$\\$a,\neg b$}\ar[r] & \txt{$running$\\$a,b$}\ar@(ul,ur)[ll]\\
\mathcal{N}_2: & \txt{$stopped$\\$\neg a,\neg b$}\ar[r] & \txt{$waiting$\\$\neg a,b$}\ar[r] & \txt{$running$\\$a,b$}\ar@(ul,ur)[ll]
}$$
\end{example}
\begin{remark}
From the examples above readers may have already noticed that actually, genealogical Kripke models are constructed from an external omniscient viewpoint, and we also implicitly assume that all the states of the child process are completely transparent to the parent process. Therefore, the parent process changes its state in accordance with the child process, even when nothing has really changed in the parent process itself. Nevertheless, $\text{\rm?`}\text{\rm?}$ pairs are still necessary for the parent process to query the states of its children processes, and hence as a possible future research, by putting restrictions on formulae we will then be able to limit the parent process' knowledge about its children processes.
\end{remark}
Next, we turn to a handful of more complicated examples, which clearly demonstrate that in practical applications, the introduction of the $\xi X.$ operator enables the very same core formula to be passed among different processes recursively.
\begin{example}
As the following graph shows, process $\mathcal{M}$ is not running until $s_3$ when its two children processes $\mathcal{N}_1$ and $\mathcal{N}_2$ have both finished, and $\mathcal{N}_1$ again has a child process $\mathcal{N}_3$ whose finish must be waited, too. Due to limit of space here we only draw out one possible history of execution as the root model $\mathcal{M}$. Anyway, we can see that there is always some process running, which can be formulated as $\mathcal{M},s_0\vDash\Box\xi X.(r\lor\exists x.\text{\rm?`} X\text{\rm?} x)$.
$$\xymatrix@C=20pt@R=15pt{
& \mathcal{N}_1,stopped\ar@{--}[dr] & \mathcal{N}_2,running\ar@{--}[d] & \mathcal{N}_1,dead\ar@{--}[dr] & \mathcal{N}_2,dead\ar@{--}[d]\\
\mathcal{M}: & \txt{$s_0$\\$\neg r$}\ar[r] & \txt{$s_1$\\$\neg r$}\ar[r] & \txt{$s_2$\\$\neg r$}\ar[r] & \txt{$s_3$\\$r$}\\
& \mathcal{N}_1,waiting\ar@{--}[u] & \mathcal{N}_2,stopped\ar@{--}[ul] & \mathcal{N}_1,running\ar@{--}[u] & \mathcal{N}_2,dead\ar@{--}[ul]\\
\mathcal{N}_1: & \txt{$pending$\\$\neg r$}\ar@<1ex>[r] & \txt{$waiting$\\$\neg r$}\ar@<1ex>[l]\ar[r] & \txt{$stopped$\\$\neg r$}\ar@<1ex>[r] & \txt{$running$\\$r$}\ar@<1ex>[l]\ar[r] & \txt{$dead$\\$\neg r$}\\
& \mathcal{N}_3,stopped\ar@{--}[u] & \mathcal{N}_3,running\ar@{--}[u] & \mathcal{N}_3,dead\ar@{--}[u] & \mathcal{N}_3,dead\ar@{--}[u] & \mathcal{N}_3,dead\ar@{--}[u]\\
\mathcal{N}_2: & \txt{$stopped$\\$\neg r$}\ar@<1ex>[r] & \txt{$running$\\$r$}\ar@<1ex>[l]\ar[r] & \txt{$dead$\\$\neg r$}\\
\mathcal{N}_3: & \txt{$stopped$\\$\neg r$}\ar@<1ex>[r] & \txt{$running$\\$r$}\ar@<1ex>[l]\ar[r] & \txt{$dead$\\$\neg r$}
}$$
\end{example}
\begin{example}
We finally present several advanced examples of more complex formulae and briefly explain their intuitive meanings, without providing sample graphs:
\begin{itemize}
\item $\text{\sf FOMC}$-sentence $\xi X.\forall x.\text{\rm?`} X\text{\rm?} x$ may look like nonsense at first glance, but it is actually valid everywhere. In fact, it simply reflects our assumption that there exists no circular nor infinite genealogical hierarchy, namely, a formal expression of Theorem~\ref{th.model}.
\item $\text{\sf FOMC}$-sentence $\xi X.(\Box p\land\forall x.\text{\rm?`}\Diamond X\text{\rm?} x)$ says that since the very beginning, the root process keeps to have property $p$, and furthermore as time goes on, property $p$ gradually diffuses down to the children processes layer by layer, so that eventually all the descendant processes will get infected by property $p$. Undoubtedly, such universal kind of spreading pattern --- no matter this property $p$ is a computer virus or anything else --- simply happens everyday and everywhere in our actual computers, network topology, and even human society.
\item $\text{\sf FOMC}$-sentence $\xi X.(\Diamond\forall y.\text{\rm?`}\xi Y.(p\land\forall z.\text{\rm?`} Y\text{\rm?} z)\text{\rm?} y\land\forall x.\text{\rm?`} X\text{\rm?} x)$ says that for every process, there will be a time when all its descendant processes possess property $p$, {\em e.g.}, $p$ represents being a dead zombie process so that it can be cleaned up by its parent process~\cite{Tanenbaum15}.
\end{itemize}
\end{example}
\phantom{nothing}
\noindent$\bigstar$ \textbf{Important Clarification} Until now, a variety of pragmatic examples have been vividly exhibited, which should adequately convince readers of first-order modal $\xi$-calculus' robust universality. It hence does not look likely that such a complex language with quite strong expressivity can be interpreted by any simpler logic that is familiar to us. In order to justify the above claim, a standard approach is to compare our first-order modal $\xi$-calculus with some other well-known logic and to show that the former is strictly more expressive. Such kind of work can definitely be carried out without much difficulty, however honestly speaking, considering the rather novel structure of genealogical Kripke models and the powerful expressivity along with undecidability of first-order quantifiers, we simply think the complexity of first-order modal $\xi$-calculus is so apparent that formal mathematical comparison seems quite redundant.
Therefore after all, let us just informally reason a few more words on this topic. Perhaps as one possible alternative to first-order modal $\xi$-calculus, certain sort of multi-modal logic might be suggested, {\em e.g.}, viewing the instantaneous parent-child relation between states of two processes as a new modality while also introducing another global modality for the $\forall x.$ quantifier. Such kind of multi-modal logic may capture a decent fragment of first-order modal $\xi$-calculus, nonetheless as illustrated in Example~\ref{exa.de}, the crucial \textit{de-dicto-de-re} distinction is known to be uniquely characteristic to first-order modal logic~\cite{Fitting02} and thus cannot get expressed in any multi-modal logic without first-order quantifiers.
\section{Bisimulation}\label{sec.bis}
As a quite powerful tool for studying modal logic, the concept of bisimulation plays a particularly significant role over fields related to theoretical computer science, such as process algebra~\cite{Ponse95}. In this section, we propose a bisimulation notion for first-order modal $\xi$-calculus and prove the bisimulation theorem in two directions, namely in one direction bisimulation implies logical equivalence, and in the other direction logical equivalence implies bisimulation under the Hennessy-Milner property~\cite{Blackburn01}.
\begin{definition}[Pointed Genealogical Kripke Model]
A pointed genealogical Kripke model $\mathcal{M},s$ is a genealogical Kripke model $\mathcal{M}$ with a fixed state $s\in S$.
\end{definition}
The following Definition~\ref{def.bis} provides the notion of bisimulation between two pointed genealogical Kripke models. Before such formal mathematical definition comes, however, we would like to depict an intuitive picture. Now that first-order-like Kripke models are dealt with, say two models $\mathcal{M}$ and $\mathcal{N}$ are bisimilar, then we not only have to designate which world in $\mathcal{M}$ is bisimilar to which world in $\mathcal{N}$ through a binary relation $Z\subseteq S_\mathcal{M}\times S_\mathcal{N}$, but also need to prescribe the correspondence between children models in $\mathcal{M}$'s and $\mathcal{N}$'s domains, viz. $N_\mathcal{M}$ and $N_\mathcal{N}$. Thus suppose $(u,v)\in Z$, then for this pair of counterparts, we need to know \textit{at present} which child model in $N_\mathcal{M}$ corresponds to which child model in $N_\mathcal{N}$ through a similar binary relation $f((u,v))\subseteq N_\mathcal{M}\times N_\mathcal{N}$. Nevertheless, there is no reason to prevent the valuation of $f$ from changing according to the pair $(u,v)$, hence generally speaking, with respect to some fixed $Z$, $f$ should then be a function from $Z$ to $\P(N_\mathcal{M}\times N_\mathcal{N})$. Therefore the detailed definition goes as the following:
\begin{definition}[Bisimilar Pointed Models]\label{def.bis}
Two pointed models $\mathcal{M},s$ and $\mathcal{N},t$ (with the same fixed Language $\text{\sf FOMC}$) are bisimilar, iff there exist a binary relation $Z\subseteq S_\mathcal{M}\times S_\mathcal{N}$ and a function $f:Z\to\P(N_\mathcal{M}\times N_\mathcal{N})$ such that:
\begin{itemize}
\item $(s,t)\in Z$.
\item $\forall(u,v)\in Z\forall p\in\mathbf{P}$, $u\in V_\mathcal{M}(p)$ iff $v\in V_\mathcal{N}(p)$.
\item $\forall(u,v)\in Z\forall\mathcal{M}'\in N_\mathcal{M}$, $\exists\mathcal{N}'\in N_\mathcal{N}$ such that $(\mathcal{M}',\mathcal{N}')\in f((u,v))$.
\item $\forall(u,v)\in Z\forall\mathcal{N}'\in N_\mathcal{N}$, $\exists\mathcal{M}'\in N_\mathcal{M}$ such that $(\mathcal{M}',\mathcal{N}')\in f((u,v))$.
\item $\forall(u,v)\in Z\forall(\mathcal{M}',\mathcal{N}')\in f((u,v))$, $\mathcal{M}',T_\mathcal{M}(u,\mathcal{M}')$ and $\mathcal{N}',T_\mathcal{N}(v,\mathcal{N}')$ are bisimilar.
\item $\forall(u,v)\in Z\forall c\in\mathbf{C}$, either one of the following holds:
\begin{enumerate}
\item both $I_\mathcal{M}(u,c)$ and $I_\mathcal{N}(v,c)$ are undefined;
\item both $I_\mathcal{M}(u,c)$ and $I_\mathcal{N}(v,c)$ are defined, and moreover, $I_\mathcal{M}(u,c),T_\mathcal{M}(u,I_\mathcal{M}(u,c))$ and $I_\mathcal{N}(v,c),T_\mathcal{N}(v,I_\mathcal{N}(v,c))$ are bisimilar.
\end{enumerate}
\item $\forall(u,v)\in Z\forall u'\in S_\mathcal{M}$ if $uR_\mathcal{M} u'$, then $\exists v'\in S_\mathcal{N}$, $vR_\mathcal{N} v'$, $(u',v')\in Z$, $f((u,v))\subseteq f((u',v'))$.
\item $\forall(u,v)\in Z\forall v'\in S_\mathcal{N}$ if $vR_\mathcal{N} v'$, then $\exists u'\in S_\mathcal{M}$, $uR_\mathcal{M} u'$, $(u',v')\in Z$, $f((u,v))\subseteq f((u',v'))$.
\end{itemize}
\end{definition}
\begin{remark}
Once again, Definition~\ref{def.bis} is a recursive definition, for in order to check whether $\mathcal{M},s$ and $\mathcal{N},t$ are bisimilar, we have to firstly calculate the bisimilar situations between models in $N_\mathcal{M}$ and models in $N_\mathcal{N}$. Nonetheless, just like Theorem~\ref{th.model}, this issue does not really cause any problem here either and thus Definition~\ref{def.bis} is well defined.
\end{remark}
\begin{theorem}\label{th.bis}
If $\mathcal{M},s$ and $\mathcal{N},t$ are bisimilar, then for any $\text{\sf FOMC}$-sentence $\phi$, $\mathcal{M},s\vDash\phi\iff\mathcal{N},t\vDash\phi$.
\end{theorem}
\begin{proof}
Suppose $\mathcal{M},s$ and $\mathcal{N},t$ are bisimilar with certain fixed $Z$ and $f$, so $(s,t)\in Z$. We inductively prove that, for any possible \textit{intermediate} stages $\llbracket\phi\rrbracket^\mathcal{M}_{i,j}$ and $\llbracket\phi\rrbracket^\mathcal{N}_{k,j}$ when evaluating arbitrary two $\text{\sf FOMC}$-sentences with respect to arbitrary two genealogical Kripke models,\footnote{We do not care about what the \textit{initial} two $\text{\sf FOMC}$-sentences are at the beginning of evaluating; nor do we care about what the \textit{initial} two models are or whether they are just $\mathcal{M}$ or $\mathcal{N}$.} if for every $(x,\mathcal{M}')\in i$, there exists $(x,\mathcal{N}')\in k$ such that $(\mathcal{M}',\mathcal{N}')\in f((s,t))$,\footnote{This condition is \textit{not} asymmetric as it looks like, since $i$ and $k$ are partial functions.} then $s\in\llbracket\phi\rrbracket^\mathcal{M}_{i,j}$ iff $t\in\llbracket\phi\rrbracket^\mathcal{N}_{k,j}$.\footnote{This claim is stronger than the original theorem, because suppose $\phi$ is an $\text{\sf FOMC}$-sentence, then we can simply start evaluating $\phi$ from $\mathcal{M}$ and $\mathcal{N}$ as the initial models, and since the initial stages are also intermediate stages but at the initial stages $i=k=\emptyset$, the condition trivially holds so we have $s\in\llbracket\phi\rrbracket^\mathcal{M}_{\emptyset,\emptyset}$ iff $t\in\llbracket\phi\rrbracket^\mathcal{N}_{\emptyset,\emptyset}$, namely $\mathcal{M},s\vDash\phi$ iff $\mathcal{N},t\vDash\phi$.}
As a matter of fact, two nested levels of induction are needed in this proof. The outer level of induction deals with models $\mathcal{M}$'s and $\mathcal{N}$'s maximal depths of generations, whose basic case is when $N_\mathcal{M}=N_\mathcal{N}=\emptyset$. Then, the inner level of induction deals with $\text{\sf FOMC}$-formula $\phi$'s structure. Firstly, at the basic step of the outer level of induction, for the inner level of induction, we only concentrate on those new cases beyond normal propositional modal logic:
\begin{itemize}
\item Suppose $\phi$ is in the form of $X$. Since the interpretation $j$ for formula variables is the same, and also from proof of Theorem~\ref{th.sem} we know that infinite loop of $X$ will not occur, this case then holds by induction hypothesis of the inner level.
\item Suppose $\phi$ is in the form of $\text{\rm?`}\psi\text{\rm?} x$. This case is impossible, because now $N_\mathcal{M}=N_\mathcal{N}=\emptyset$ and there is no appropriate interpretation for model variable $x$, contradicting Theorem~\ref{th.sem}.
\item Suppose $\phi$ is in the form of $\text{\rm?`}\psi\text{\rm?} c$. Also because $N_\mathcal{M}=N_\mathcal{N}=\emptyset$, now both $I_\mathcal{M}(u,c)$ and $I_\mathcal{N}(v,c)$ must be undefined and thus $\llbracket\text{\rm?`}\psi\text{\rm?} c\rrbracket^\mathcal{M}_{i,j}=\llbracket\text{\rm?`}\psi\text{\rm?} c\rrbracket^\mathcal{N}_{k,j}=\emptyset$.
\item Suppose $\phi$ is in the form of $\forall x.\psi$. Also because $N_\mathcal{M}=N_\mathcal{N}=\emptyset$, we trivially have $\llbracket\forall x.\psi\rrbracket^\mathcal{M}_{i,j}=S_\mathcal{M}$ and $\llbracket\forall x.\psi\rrbracket^\mathcal{N}_{k,j}=S_\mathcal{N}$.
\item Suppose $\phi$ is in the form of $\xi X.\psi$. This case holds by induction hypothesis of the inner level.
\end{itemize}
Next, we move on to the inductive step of the outer level of induction. For the inner level of induction, we again only mention several technically subtle cases that differ from routine proof:
\begin{itemize}
\item Suppose $\phi$ is in the form of $\text{\rm?`}\psi\text{\rm?} x$, and $(x,\mathcal{M}')\in i$, $(x,\mathcal{N}')\in k$. Since $(s,t)\in Z$ and $(\mathcal{M}',\mathcal{N}')\in f((s,t))$, from the condition in Definition~\ref{def.bis} we know that $\mathcal{M}',T_\mathcal{M}(s,\mathcal{M}')$ and $\mathcal{N}',T_\mathcal{N}(t,\mathcal{N}')$ are bisimilar, then by induction hypothesis of the outer level we know that $T_\mathcal{M}(s,\mathcal{M}')\in\llbracket\psi\rrbracket^{\mathcal{M}'}_{\emptyset,j}$ iff $T_\mathcal{N}(t,\mathcal{N}')\in\llbracket\psi\rrbracket^{\mathcal{N}'}_{\emptyset,j}$, therefore, $s\in\llbracket\phi\rrbracket^\mathcal{M}_{i,j}$ iff $t\in\llbracket\phi\rrbracket^\mathcal{N}_{k,j}$.
\item Suppose $\phi$ is in the form of $\text{\rm?`}\psi\text{\rm?} c$. Since $(s,t)\in Z$, from the condition in Definition~\ref{def.bis} we know that either both $I_\mathcal{M}(s,c)$ and $I_\mathcal{N}(t,c)$ are undefined, then we have $s\notin\llbracket\text{\rm?`}\psi\text{\rm?} c\rrbracket^\mathcal{M}_{i,j}$ and $t\notin\llbracket\text{\rm?`}\psi\text{\rm?} c\rrbracket^\mathcal{N}_{k,j}$; or both $I_\mathcal{M}(s,c)$ and $I_\mathcal{N}(t,c)$ are defined, $I_\mathcal{M}(s,c),T_\mathcal{M}(s,I_\mathcal{M}(s,c))$ and $I_\mathcal{N}(t,c),T_\mathcal{N}(t,I_\mathcal{N}(t,c))$ are bisimilar, then by induction hypothesis of the outer level we know that $T_\mathcal{M}(s,I_\mathcal{M}(s,c))\in\llbracket\psi\rrbracket^{I_\mathcal{M}(s,c)}_{\emptyset,j}$ iff $T_\mathcal{N}(t,I_\mathcal{N}(t,c))\in\llbracket\psi\rrbracket^{I_\mathcal{N}(t,c)}_{\emptyset,j}$, therefore, $s\in\llbracket\phi\rrbracket^\mathcal{M}_{i,j}$ iff $t\in\llbracket\phi\rrbracket^\mathcal{N}_{k,j}$.
\item Suppose $\phi$ is in the form of $\Box\psi$. Since $(s,t)\in Z$, from the condition in Definition~\ref{def.bis} we know that for every $u\in S_\mathcal{M}$ such that $sR_\mathcal{M} u$, there exists $v\in S_\mathcal{N}$ such that $tR_\mathcal{N} v$, $(u,v)\in Z$, $f((s,t))\subseteq f((u,v))$, and vice versa. Hence, fix arbitrary such pair of $(u,v)\in Z$, we only have to show that $u\in\llbracket\psi\rrbracket^\mathcal{M}_{i,j}$ iff $v\in\llbracket\psi\rrbracket^\mathcal{N}_{k,j}$, and because $f((s,t))\subseteq f((u,v))$, for every $(x,\mathcal{M}')\in i$ there exists $(x,\mathcal{N}')\in k$ such that $(\mathcal{M}',\mathcal{N}')\in f((u,v))$, so by induction hypothesis of the inner level the above claim holds, therefore, $s\in\llbracket\phi\rrbracket^\mathcal{M}_{i,j}$ iff $t\in\llbracket\phi\rrbracket^\mathcal{N}_{k,j}$.
\item Suppose $\phi$ is in the form of $\forall x.\psi$. Since $(s,t)\in Z$, from the condition in Definition~\ref{def.bis} we know that for every $\mathcal{M}'\in N_\mathcal{M}$ there exists $\mathcal{N}'\in N_\mathcal{N}$ such that $(\mathcal{M}',\mathcal{N}')\in f((s,t))$, and vice versa. Hence, fix arbitrary such pair of $(\mathcal{M}',\mathcal{N}')\in f((s,t))$, we only have to show that $s\in\llbracket\psi\rrbracket^\mathcal{M}_{i[x:=\mathcal{M}'],j}$ iff $t\in\llbracket\psi\rrbracket^\mathcal{N}_{k[x:=\mathcal{N}'],j}$, and because $(\mathcal{M}',\mathcal{N}')\in f((s,t))$, for every $(x,\mathcal{M}'')\in i[x:=\mathcal{M}']$ there exists $(x,\mathcal{N}'')\in k[x:=\mathcal{N}']$ such that $(\mathcal{M}'',\mathcal{N}'')\in f((s,t))$, so by induction hypothesis of the inner level the above claim holds, therefore, $s\in\llbracket\phi\rrbracket^\mathcal{M}_{i,j}$ iff $t\in\llbracket\phi\rrbracket^\mathcal{N}_{k,j}$.
\end{itemize}
\end{proof}
Similar to ordinary propositional modal logic~\cite{Hennessy85}, the reverse of Theorem~\ref{th.bis} also holds, provided that both $\mathcal{M}$ and $\mathcal{N}$ are image-finite models. To start with, however, we have to define what an image-finite model is, which is still a recursive definition and actually requires finiteness on two dimensions, namely both horizontally along the binary relation $R$ and vertically along the set of children models $N$:
\begin{definition}[Image-Finite Genealogical Kripke Model]
A genealogical Kripke model $\mathcal{M}$ is image-finite, iff all of the following conditions hold:
\begin{itemize}
\item For every $s\in S$, the set $\{t\in S\mid sRt\}$ is finite.
\item The set $N$ is finite.
\item For every $\mathcal{N}\in N$, the model $\mathcal{N}$ is image-finite.
\end{itemize}
\end{definition}
For proof of the reverse of Theorem~\ref{th.bis}, as a preparation let us firstly focus on the following Lemma~\ref{lem.hm}, which constitutes the major difficulty throughout the entire inductive proof. Just like the diagram method in model theory of first order logic~\cite{Marker02}, here we will also expand the language from $\text{\sf FOMC}$ to $\text{\sf FOMC}'$ by adding a corresponding new constant symbol for every element in the domain, i.e., every child model.
Formally, the expansion works as follows. Suppose $\mathcal{M}$ and $\mathcal{N}$ are two image-finite genealogical Kripke models, $s\in S_\mathcal{M}$, $t\in S_\mathcal{N}$, and for any $\text{\sf FOMC}$-sentence $\phi$, $\mathcal{M},s\vDash\phi\iff\mathcal{N},t\vDash\phi$. Let two sets of fresh constant symbols be $\mathbf{C}_\mathcal{M}=\{c_{\mathcal{M}'}\mid\mathcal{M}'\in N_\mathcal{M}\}$ and $\mathbf{C}_\mathcal{N}=\{c_{\mathcal{N}'}\mid\mathcal{N}'\in N_\mathcal{N}\}$, we expand the language to $\text{\sf FOMC}'$ by enlarging $\mathbf{C}'=\mathbf{C}\cup\mathbf{C}_\mathcal{M}\cup\mathbf{C}_\mathcal{N}$, but meanwhile restricting that the appearance of any $c\in\mathbf{C}_\mathcal{M}\cup\mathbf{C}_\mathcal{N}$ should not be within the scope of any $\xi X.$ or the scope of any $\text{\rm?`}\text{\rm?}$ pair. Thus, $c$ will never get passed down vertically.
Now suppose there arbitrarily exist two fixed functions $g:N_\mathcal{M}\to N_\mathcal{N}$ and $h:N_\mathcal{N}\to N_\mathcal{M}$, then using these two functions, we naturally view $\mathcal{M}$ and $\mathcal{N}$ also as genealogical Kripke models for the expanded language $\text{\sf FOMC}'$ by letting $I_\mathcal{M}'(u,c_{\mathcal{M}'})=\mathcal{M}'$, $I_\mathcal{M}'(u,c_{\mathcal{N}'})=h(\mathcal{N}')$, $I_\mathcal{N}'(v,c_{\mathcal{M}'})=g(\mathcal{M}')$, $I_\mathcal{N}'(v,c_{\mathcal{N}'})=\mathcal{N}'$ for every $u\in S_\mathcal{M}$, $v\in S_\mathcal{N}$, $c_{\mathcal{M}'}\in\mathbf{C}_\mathcal{M}$, $c_{\mathcal{N}'}\in\mathbf{C}_\mathcal{N}$.
\begin{lemma}\label{lem.hm}
Suppose $\mathcal{M}$ and $\mathcal{N}$ are two image-finite genealogical Kripke models, $s\in S_\mathcal{M}$, $t\in S_\mathcal{N}$, and for any $\text{\sf FOMC}$-sentence $\phi$, $\mathcal{M},s\vDash\phi\iff\mathcal{N},t\vDash\phi$. Do the expansion as above. Then there exist two functions $g:N_\mathcal{M}\to N_\mathcal{N}$ and $h:N_\mathcal{N}\to N_\mathcal{M}$ such that:
\begin{itemize}
\item For any natural number $n\in\omega$ and any $sR^n_\mathcal{M} u$, there exists $tR^n_\mathcal{N} v$ so that for any $\text{\sf FOMC}'$-sentence $\phi$, $\mathcal{M},u\vDash\phi\iff\mathcal{N},v\vDash\phi$.
\item For any natural number $m\in\omega$ and any $tR^m_\mathcal{N} v$, there exists $sR^m_\mathcal{M} u$ so that for any $\text{\sf FOMC}'$-sentence $\psi$, $\mathcal{M},u\vDash\psi\iff\mathcal{N},v\vDash\psi$.
\end{itemize}
\end{lemma}
\begin{proof}
Consider the nontrivial situation when $N_\mathcal{M}\neq\emptyset$ and $N_\mathcal{N}\neq\emptyset$. We prove by contradiction. Suppose not, namely, for every possible functions $g:N_\mathcal{M}\to N_\mathcal{N}$ and $h:N_\mathcal{N}\to N_\mathcal{M}$, either one of the following two cases holds:\footnote{In the following, $\bigwedge\limits_v$ denotes conjunction over all possible $v$, and the same for $u$, $g$ and $h$, all of which only have finitely many possibilities due to image-finiteness. $\phi$ and $\psi$ generally depend on parameters $v$, $u$, $g$ and $h$, but we just omit too many subscripts for neatness.}
\begin{itemize}
\item There exist some natural number $n\in\omega$ and $sR^n_\mathcal{M} u$ such that for any $tR^n_\mathcal{N} v$, there exists some $\text{\sf FOMC}'$-sentence $\phi$ so that $\mathcal{M},u\vDash\phi\nLeftrightarrow\mathcal{N},v\vDash\phi$, and without loss of generality, we may assume $\mathcal{M},u\vDash\phi$ and $\mathcal{N},v\nvDash\phi$. Since the possible number of all the different $v$ is finite, by conjunction we have $\mathcal{M},s\vDash\Diamond^n\bigwedge\limits_v\phi$ as $sR^n_\mathcal{M} u$, but $\mathcal{N},t\nvDash\Diamond^n\bigwedge\limits_v\phi$ as $tR^n_\mathcal{N} v$ for every $v$.
\item There exist some natural number $m\in\omega$ and $tR^m_\mathcal{N} v$ such that for any $sR^m_\mathcal{M} u$, there exists some $\text{\sf FOMC}'$-sentence $\psi$ so that $\mathcal{M},u\vDash\psi\nLeftrightarrow\mathcal{N},v\vDash\psi$, and without loss of generality, we may assume $\mathcal{M},u\nvDash\psi$ and $\mathcal{N},v\vDash\psi$. Since the possible number of all the different $u$ is finite, by conjunction we have $\mathcal{N},t\vDash\Diamond^m\bigwedge\limits_u\psi$ as $tR^m_\mathcal{N} v$, but $\mathcal{M},s\nvDash\Diamond^m\bigwedge\limits_u\psi$ as $sR^m_\mathcal{M} u$ for every $u$, namely, $\mathcal{M},s\vDash\neg\Diamond^m\bigwedge\limits_u\psi$ but $\mathcal{N},t\nvDash\neg\Diamond^m\bigwedge\limits_u\psi$.
\end{itemize}
To sum up, as $\mathcal{M}$ is not affected by $g$, we can actually conclude that for every possible functions $g$ and $h$, we always have $\mathcal{M},s\vDash\bigvee\limits_h(\bigwedge\limits_g\Diamond^n\bigwedge\limits_v\phi\land\bigwedge\limits_g\neg\Diamond^m\bigwedge\limits_u\psi)$. This is an $\text{\sf FOMC}'$-sentence, in the front of which we can add existential quantifiers binding every $c_{\mathcal{M}'}\in\mathbf{C}_\mathcal{M}$ as well as universal quantifiers binding every $c_{\mathcal{N}'}\in\mathbf{C}_\mathcal{N}$ so as to convert it back to an $\text{\sf FOMC}$-sentence, namely $\mathcal{M},s\vDash\exists\vec{c}_{\mathcal{M}'}.\forall\vec{c}_{\mathcal{N}'}.\bigvee\limits_h(\bigwedge\limits_g\Diamond^n\bigwedge\limits_v\phi\land\bigwedge\limits_g\neg\Diamond^m\bigwedge\limits_u\psi)$, and since $\mathcal{M},s\vDash\phi\iff\mathcal{N},t\vDash\phi$ for any $\text{\sf FOMC}$-sentence $\phi$, we obtain $\mathcal{N},t\vDash\exists\vec{c}_{\mathcal{M}'}.\forall\vec{c}_{\mathcal{N}'}.\bigvee\limits_h(\bigwedge\limits_g\Diamond^n\bigwedge\limits_v\phi\land\bigwedge\limits_g\neg\Diamond^m\bigwedge\limits_u\psi)$, and so back again to $\text{\sf FOMC}'$-sentence we can conclude that there exists some $g$ such that $\mathcal{N},t\vDash\bigvee\limits_h(\bigwedge\limits_g\Diamond^n\bigwedge\limits_v\phi\land\bigwedge\limits_g\neg\Diamond^m\bigwedge\limits_u\psi)$. However, because $\mathcal{N}$ is not affected by $h$, for each disjunctive branch here we always have $\mathcal{N},t\nvDash\bigwedge\limits_g\Diamond^n\bigwedge\limits_v\phi\land\bigwedge\limits_g\neg\Diamond^m\bigwedge\limits_u\psi$ for any $g$, thus a contradiction.
\end{proof}
At last, we can state the reverse of Theorem~\ref{th.bis}, usually called Hennessy-Milner Theorem, as the following:
\begin{theorem}[Hennessy-Milner]\label{th.hm}
If $\mathcal{M}$ and $\mathcal{N}$ are two image-finite genealogical Kripke models, $s\in S_\mathcal{M}$, $t\in S_\mathcal{N}$, and for any $\text{\sf FOMC}$-sentence $\phi$, $\mathcal{M},s\vDash\phi\iff\mathcal{N},t\vDash\phi$. Then $\mathcal{M},s$ and $\mathcal{N},t$ are bisimilar.
\end{theorem}
\begin{proof}
According to Lemma~\ref{lem.hm}, fix two functions $g:N_\mathcal{M}\to N_\mathcal{N}$ and $h:N_\mathcal{N}\to N_\mathcal{M}$, let $Z$ be the set of all ``zig-and-zag'' $(u,v)$ pairs described in the two clauses of conclusion in Lemma~\ref{lem.hm}, and let $f((u,v))=g\cup h^{-1}$ for every $(u,v)\in Z$. We still make use of inductive proof with respect to models $\mathcal{M}$'s and $\mathcal{N}$'s maximal depths of generations, and so as induction hypothesis, we also know that this theorem already holds between any pair of $\mathcal{M}',I_\mathcal{M}(u,\mathcal{M}')$ and $\mathcal{N}',I_\mathcal{N}(v,\mathcal{N}')$, where $(u,v)\in Z$, $\mathcal{M}'\in N_\mathcal{M}$ and $\mathcal{N}'\in N_\mathcal{N}$. For any $\text{\sf FOMC}$-sentence $\phi$, by Lemma~\ref{lem.hm} we have $\mathcal{M},u\vDash\text{\rm?`}\phi\text{\rm?} c_{\mathcal{M}'}\iff\mathcal{N},v\vDash\text{\rm?`}\phi\text{\rm?} c_{\mathcal{M}'}$, namely $\mathcal{M}',I_\mathcal{M}(u,\mathcal{M}')\vDash\phi\iff g(\mathcal{M}'),I_\mathcal{N}(v,g(\mathcal{M}'))\vDash\phi$, hence by induction hypothesis, $\mathcal{M}',I_\mathcal{M}(u,\mathcal{M}')$ and $g(\mathcal{M}'),I_\mathcal{N}(v,g(\mathcal{M}'))$ are bisimilar, and similarly $h(\mathcal{N}'),I_\mathcal{M}(u,h(\mathcal{N}'))$ and $\mathcal{N}',I_\mathcal{N}(v,\mathcal{N}')$ are bisimilar, too. It is then not difficult to check that such $Z$ and $f$ are indeed a bisimulation between $\mathcal{M},s$ and $\mathcal{N},t$.
\end{proof}
\section{Conclusions and Future Work}\label{sec.conc}
This paper proposes first-order modal $\xi$-calculus, a logic for expressively describing genealogical Kripke models. With illustration of several vivid examples, we have witnessed how genealogical Kripke models naturally depict hierarchic and concurrent practical phenomena --- particularly like computer processes, as well as how sentences of first-order modal $\xi$-calculus succinctly but exactly capture lots of most interesting properties of the Kripke structure. The intuitive picture is clear to understand, while mathematically speaking, the logic itself is also technically challenging, which can be perceived through our effortful management in order to precisely characterize the logic's expressivity through a bisimulation notion.
Nonetheless, this paper is still no more than a piece of primitive work, based on which we would be glad to suggest some plausible future research directions:
\begin{itemize}
\item Despite that genealogical Kripke models form a rather novel kind of shape, we can still attempt to capture part of its features through other formal languages, which will then be put onto a thorough comparison in contrast to our first-order modal $\xi$-calculus, investigating overall aspects such as expressivity, succinctness, standard translation as well as characterization theorem. It should also be worthwhile trying to combine first-order modal $\xi$-calculus with modal $\mu$-calculus into an integrated framework, and studying them all together from a united algebraic perspective.
\item Being called a sort of logical calculus, sound and complete proof systems need to be further established for our first-order modal $\xi$-calculus, {\em e.g.}, trying a tableau-kind system might be a good point to start~\cite{Agostino99}. Similar to normal modal logic, we may also reasonably anticipate that different logics dwell on different frame classes, or under different additional restrictions over genealogical Kripke models, {\em e.g.}, it sounds tempting to naturally consider various heritage correlations between the parent and the children. Moreover, different proof systems can also possess different computational properties, including decidability, complexity, model checking and so on, which should all prove quite essential in future study and practice as this logic's utility in computer science has been heavily suggested.
\item Last but not least, it seems a quite promising and exciting approach to extend the definition of genealogical Kripke models as well as the corresponding semantics, {\em e.g.}, what if a model $\mathcal{M}$ is allowed to reflexively refer to itself as one of its own children models, namely $\mathcal{M}\in N_\mathcal{M}$? As shown by Theorem~\ref{th.model}, a non-standard-non-recursive definition is necessary, and so tight relevance to non-well-founded set theory~\cite{Aczel88}, as well as process algebra which has developed very fruitful in theoretical computer science~\cite{Sangiorgi11}, could then be reasonably expected. Another plausible extension of genealogical Kripke models might be to loosen the restriction on constant domain~\cite{Fitting98}, since in reality, construction of new processes as well as destruction of dead zombie ones are just taking place constantly.
\end{itemize}
\section*{Acknowledgment}
\noindent The author owes much thank to Satoshi Tojo and Mizuhito Ogawa for helpful advice about choosing appropriate notations and writing more clearly. A handful of anonymous reviewers on previous versions of the manuscript have also provided tremendous valuable suggestions to help the author improve this paper.
|
\section{Introduction}
Inflation -- a period of cosmic acceleration in the very early and
energetic universe -- is on the brink of being tested in new detail
through the detection of primordial gravitational waves. These tensor
modes reveal the energy scale of inflation, and together with the
scalar (density) perturbations give key indications of the inflationary
physics. The tensor to scalar ratio $r$ and the scalar power spectrum
slope $n_s$ form a parameter space with different classes of models
lying in different regions of it.
A useful and intriguing approach is to explore universality classes
rather than individual models of the inflationary potential or slow
roll parameters. This has the philosophy that the scalar tilt $n_s-1$
should be a function purely of the number of e-folds of inflation $N$,
without additional scales entering.
Traditionally the ansatz has meant taking an expansion such that
$n_s-1\sim 1/N$ plus higher order terms in powers of $1/N$. Here
we keep the philosophy of $N$ being the determining factor, but allow
for $\ln N$ corrections to the leading order -- still without introducing
any other scale. We explore the effects of the next to leading order
term, in particular on the $n_s$--$r$
plane, and implications for experimental limits.
\section{The $n_s$--$r$ Relation} \label{sec:relation}
For slow roll inflation one has a hierarchy of derivatives of the
scalar field potential, or Hubble parameter, and from these one can
compute the observables of the
scalar tilt $n_s-1$ and tensor to scalar ratio $r$.
While one can work within a specific model of the potential $V(\phi)$
or expansion $H(\phi)$, many classes of inflation theory exhibit a
universality relation $n_s(N)$. One can start instead with that more
model independent approach
as the ansatz, as advocated early by
\cite{1303.3925,1309.1285,1402.2059,1411.7237,1412.0678} and many
others since then. (But see \cite{1609.04739} for
the limitations of such an ansatz.)
The tensor to scalar ratio $r$ is given by a differential
equation under the slow roll assumption,
\begin{equation}
\frac{d\ln r}{dN}-\frac{r}{8}=n_s-1\,. \label{eq:dr}
\end{equation}
The standard universality relation has
\begin{equation}
n_s-1=-\frac{\al}{N}\,,
\end{equation}
where $\al$ is of order one, and with this we can solve
Eq.~\eqref{eq:dr} for $r$.
\subsection{$\al=$ constant}
When $\al$ is constant, one has the well known solution
\begin{equation}
r=\frac{8(\al-1)}{N+cN^\al}\,,
\end{equation}
where $c$ is a constant of integration. This gives two asymptotic
branches, where $r\sim N^{-1}$ and where $r\sim N^{-\al}$.
Since under slow roll $r=16\eps$ and the slow roll parameter
\begin{equation}
\eps=-d\ln H/dN=\frac{1}{2}\left(\frac{dV/d\phi}{V}\right)^2\,,
\end{equation}
then by converting $N$ to $\phi$,
\begin{equation}
\phi=\int dN\,\sqrt{2\eps}\,,
\end{equation}
one can see that the first case corresponds to
\begin{equation}
V(\phi)\sim \phi^{2(\al-1)}\,,
\end{equation}
and the second case is
\begin{eqnarray}
V(\phi)&\sim& \phi^{(2-2\al)/(2-\al)}\qquad [\al\ne2]\\
V(\phi)&\sim& \left(1-e^{-\phi\sqrt{2/c}}\,\right)\qquad [\al=2]\,,
\end{eqnarray}
(rolling off the nearly flat plateau at large $\phi$ in the $\al=2$
case).
\subsection{Logarithmic running of $\al$}
However, $\alpha$ generally gets terms beyond the leading order
constant term. For $\al(N)$,
Eq.~\eqref{eq:dr} can be solved to give
\cite{1412.0678}
\begin{equation}
\frac{r}{8}=\left[-e^{\int (dN'/N')\,\al(N')}\,\int dN'\,e^{-\int (dN''/N'')\,\al(N'')}+c\,e^{\int (dN'/N')\,\al(N')}\right]^{-1}\,, \label{eq:ral}
\end{equation}
where $c$ is again an integration constant and the integrals are
evaluated up to a $N$ e-folds.
The point of universality is to not introduce any time dependence or
scale other than the e-fold scale $N$. Thus we expect that higher order
terms inducing a variation
of $\al(N)$ should be in some series expansion in $N$ with coefficients
of order one. The natural ansatz is $\al(N)=\al_0+\al_1/N+\al_2/N^2+\dots$.
This has been studied for many cases and arises from, for example,
hilltop inflation \cite{1309.1285}.
One could also consider $\ln N$ effects, and indeed Starobinsky
inflation \cite{staro} has $\al(N)=2-3\ln N/N$. Since $\ln N/N>1/N$ then
we expect this correction to play a larger role in altering the relation
in the $n_s$--$r$ plane than a simple $1/N$ next to leading order term.
However, there is potentially an even larger next to leading order term:
$1/\ln N>\ln N/N$. This is the term we consider in this paper:
\begin{equation}
\al(N)=\al_0+\frac{\al_1}{(\ln N)^s}\,, \label{eq:lnn}
\end{equation}
where the next to leading order term is suppressed by $(\ln N)^s$.
Using Eq.~\eqref{eq:ral} this gives, for $s\ne1$,
\begin{equation}
\frac{r}{8}=\left[-N^{\al_0}e^{[\al_1/(1-s)](\ln N)^{1-s}}\,\int dN\,N^{-\al_0} e^{-[\al_1/(1-s)](\ln N)^{1-s}}+cN^{\al_0}e^{[\al_1/(1-s)](\ln N)^{1-s}}\right]^{-1}\,.
\end{equation}
The $s=1$ case is simpler, with
\begin{equation}
\frac{r}{8}=\left[-N^{\al_0}(\ln N)^{\al_1}\,\int dN\,N^{-\al_0}(\ln N)^{-\al_1}+cN^{\al_0}(\ln N)^{\al_1}\right]^{-1}\,.
\end{equation}
For simplicity, we consider two analytic cases: Case A with $s=1$ and $\al_1=-1$, and Case B
with arbitrary $s$ and $\al_0=1$.
Case A yields
\begin{equation}
\frac{r}{8}=\left[\frac{N}{(\al_0-1)}+\frac{N}{(\al_0-1)^2\ln N}+\frac{cN^{\al_0}}{\ln N}\right]^{-1}\,.
\label{eq:rcaseA}
\end{equation}
This is interesting: the first branch gains an extra (positive) term
in the denominator, lowering $r$ for a given $\al_0$, while
the second branch has a $\ln N$ suppression, raising $r$.
Figure~\ref{fig:rns} shows the behavior in the $n_s$--$r$ plane
for the standard relation (leading order term only), and for Case A with the
next to leading order $1/\ln N$ term, i.e.\ Eq.~\eqref{eq:rcaseA}. In the region of greatest interest,
$0.96\lesssim n_s\lesssim 0.97$, the second branch dominates
for $c\sim1$ (here we took $c=1$). Indeed the next to leading order
term enhances the value of $r$ for a given $n_s$, making
inflationary gravitational waves easier to detect.
\begin{figure}[!htb]
\centering
\includegraphics[width=0.6\columnwidth]{rnsa2.ps}
\caption{
Relation between $r$ and $n_s$ for inflation models following
the standard leading order ansatz (black) or including a
next to leading order $1/\ln N$ term (blue; Case A: $s=1$, $\al_1=-1$).
Asymptotic branches due to the two terms are shown, dashed
and solid (with $c=1$) respectively, while the dotted curves give
the full sum. Curves are shown for $N=60$
and $\al_0$ runs from 3 to 1.05, from slightly off the left side to slightly off the
right side.
}
\label{fig:rns}
\end{figure}
The next to leading order term
gives a multiplicative enhancement of $r$ on the second branch
by a factor $\ln N\approx4$
(we show results for $N=60$) for fixed $\al_0$. However, it shifts $n_s$ as well, so
the vertical and horizon displacements combine to give a smaller
gain in r, by $(\ln N)/e\sim1.5$. Since the first branch has shallower slope,
the horizontal shift actually erases the expected lowering of $r$,
giving a slight enhancement. In any case, this region of relatively
high $r$ is disfavored by data, especially after the new BICEP/KECK
results, $r<0.036$ at 95\% confidence level \cite{bicep}.
To clarify the shift induced in $r$ we plot the $r$--$\al_0$ plane
for Case A in Fig.~\ref{fig:ralpha}; this gives a purer picture of $r$ but note
that a given value of $\al_0$ will have differing values of $n_s$
between the curves using leading order alone and those including
the next to leading order. The factor of four enhancement for the
second branch and the reduction by a factor $1+1/[(\al_0-1)\ln N]$
for the first branch are clear.
\begin{figure}[!htb]
\centering
\includegraphics[width=0.6\columnwidth]{ralphaa2.ps}
\caption{
As Fig.~\ref{fig:rns} for Case A but now in the $r$--$\al_0$ plane,
so $n_s$ runs along the curves. Values $\al_0\approx2$
give $n_s\approx0.96$--0.97.
}
\label{fig:ralpha}
\end{figure}
Turning to Case B, with arbitrary $s$ but $\al_0=1$, gives
\begin{equation}
\frac{r}{8}=\left[\frac{N(\ln N)^s}{\al_1}+cN\,e^{[\al_1/(1-s)](\ln N)^{1-s}}\right]^{-1}\,,
\end{equation}
or when $s=1$,
\begin{equation}
\frac{r}{8}=\left[\frac{N\,\ln N}{(\al_1-1)}+cN\,(\ln N)^{\al_1}\right]^{-1}\,. \label{eq:ral0s1}
\end{equation}
For integer $s>1$, we find $n_s$ too high to be viable, and a
low $r$, so the $s=1$ case is most relevant, with a large enough
$\al_1$ to give a viable $n_s-1=-1/N-\al_1/(N\ln N)$.
Figure~\ref{fig:rnss1a01}
shows the results for Case B ranging over $\al_1=1.05$--5. Again the second
branch is the dominant one for viably small $r$. In fact,
the constraint on $n_s$ forces smaller $r$ than in the
previous models, with $n_s=0.970$ (0.973) having
$\al_1=3.2$ (2.5) and $r=1.4\times 10^{-3}$ ($3.6\times10^{-3}$).
\begin{figure}[!htb]
\centering
\includegraphics[width=0.6\columnwidth]{rnss1a01a2.ps}
\caption{
As Fig.~\ref{fig:rns} but now with $\al_1$ free and fixing
$\al_0=1$ to get a simple solution. Curves run over
$\al_1=1.05$ (right endpoints) to 5 (left endpoints).
}
\label{fig:rnss1a01}
\end{figure}
\section{Conclusions}
The ansatz $n_s-1\sim 1/N$ is an attractive quasi model independent,
or universality class, approach to inflation. It has the virtue that no scale
enters different from order one other than the number of e-folds of inflation.
Beyond leading order terms that also depend only on $N$ have the same
property. Here we examined the largest possible next to leading order
term, $1/\ln N$.
Logarithmic terms in the expansion also appear, though in a different form,
in well known inflation theories such as Starobinsky inflation. Here we
explore what impact this larger correction can have on the power spectrum
tilt $n_s$ and tensor to scalar ratio $r$. Equation~\eqref{eq:lnn} gives
our basic ansatz, and Eqs.~\eqref{eq:rcaseA} and \eqref{eq:ral0s1} the
solutions of most interest. For viable values of $n_s$ and $r$, the
second branch is most relevant, and we find that for Case A the next
to leading order term can make inflationary gravitational waves more
detectable by increasing $r$ by a factor $\sim1.5$ for a given measured value of
$n_s$.
\acknowledgments
This work is supported in part by the Energetic Cosmos Laboratory and by the
U.S.\ Department of Energy, Office of Science, Office of High Energy
Physics, under contract no.~DE-AC02-05CH11231.
|
\section{Introduction}
Neural networks (NNs) may easily learn certain training sets, but typically they do not generalize on systematically different test sets.
Examples of systematic generalization \citep{fodor1988connectionism}
include generalization to sequences longer than those seen during training---productivity, and algorithmic combinations of previously learned rules---systematicity.
Despite recent efforts \citep{bahdanau2019closure, korrel2019transcoding, lake2019compositional, li2019compositional, russin2019compositional, csordas2021devil}, systematic generalization generally remains unsolved \citep{fodor1990connectionism, lake2017generalization, liska2018memorize, greff2020binding, hupkes2019compositionality}.
On some datasets, the best performing models are neuro-symbolic hybrids \citep{chen2020compositional, liu2020compositional} using task-specific symbolic functions.
However, their applicability to other datasets remains limited \citep{furrer2020compositional, shaw2020compositional}.
A big question is: which type of architectural inductive bias encourages the training process to select ``good'' solutions which generalize systematically?
The popular Transformers \citep{vaswani2017attention} also often fail to generalize on algorithmic tasks (e.g.~\citet{liska2018memorize, dubois2020location, chaabouni2021can, csordas2021devil, ontanon2021making}), even on tasks with intuitive solutions that can be simply expressed in terms of Transformer attention patterns.
Given an input sequence of length $N$ and a Transformer encoder of depth $T$,
solving an algorithmic task is often all about routing the relevant information
to the right node/operation at the right time in the $T$-by-$N$ grid represented by Transformer columns (illustrated in Figure \ref{fig:geometirc_order}/Left).
Effectively the task is to learn to draw an \textit{adaptive control flow} on the canvas of Transformer columns. In fact, recent work by \citet{weiss21} introduced a programming language called RASP, which is specifically designed to express solutions to sequence processing problems, and which has a direct equivalent to the operations in Transformer encoders.
However, it is shown that Transformers learn solutions expressed in RASP only through intermediate supervision of attention patterns, and sometimes, even such supervision fails.
Generally speaking, Transformers fail to find easily interpretable and/or symbolic solutions to algorithmic tasks.
We conversely hypothesize that attention-based NNs that are able to find intuitive solutions
(achieving interpretable attention patterns)
could improve systematic generalization.
Here we point out that regular Transformers lack some basic ingredients for learning such ``intuitive'' solutions to algorithmic problems.
As a remedy, we propose simple architectural modifications to help them learn data routing.
As a first step towards validating our model, we focus on
the popular length generalization task of compositional table lookup (CTL;
\citet{liska2018memorize, hupkes2018learning, dubois2020location}),
as well as two more complex tasks: a simple arithmetic task and a variant of ListOps \citep{nangia2018listops} designed
to test the compositional generalization ability of NNs.
Our novel Neural Data Router (NDR) achieves 100\% generalization accuracy (never reported before; \citet{dubois2020location}) on the CTL task,
and obtains nearly perfect accuracy on both the proposed simple arithmetic
and ListOps tasks.
We show that the attention and gating patterns of NDR
tend to be interpretable as
plausible control flows.
\section{Improving Transformers for Learning Adaptive Control Flow}
\label{sec:control_flow}
We argue that the following components are needed to build Transformers capable of learning adaptive control flow.
\textbf{First}, composing known operations in an arbitrary order requires that all operations are available at every computational step.
This can be easily achieved by sharing the weights of the layers, as is done in Universal Transformers \citep{dehghani2019universal}.
\textbf{Second}, the network should be sufficiently deep, at least as deep as the deepest data dependency in the computational graph built
from elementary operations (e.g., in the case of a parse tree, this is the depth of the tree).
Otherwise, multiple operations must be fused into a single layer and hinder natural and elegant compositions.
\textbf{Third}, inputs in some columns should be kept unchanged until it is their turn to be processed.
The regular Transformer lacks a mechanism for skipping the whole transformation step by simply copying the input to the next step/layer.
We propose a special gating function, \textit{copy gate}, to implement such a mechanism (Sec.~\ref{sec:gating}).
\textbf{Finally}, many algorithmic tasks require combining several local computations in the right order.
This typically implies that attention should not focus on all possible matches at a given time but only on the closest match.
We propose and investigate a new type of attention with
a corresponding inductive bias called \textit{geometric attention} (Sec.~\ref{sec:geo-attn}). Using both the geometric attention and copy gate, our model implements a ``neural data routing mechanism'', which can adaptively serialize the input problem.
We refer to the resulting new Transformer as Neural Data Router (NDR). In the experimental section (Sec.~\ref{sec:exp_top}), we evaluate this model on three algorithmic tasks requiring length generalization
and demonstrate its effectiveness.
\subsection{Copy Gate: Learning to Skip Operations (Vertical Flow)}
\label{sec:gating}
Each layer of the regular Transformer consists of
one self-attention and one feedforward block.
The input to each of these blocks is directly connected
to the corresponding output via a residual connection \citep{srivastava2015training, resnet}.
However, such a connection does not allow for skipping the transformation of the entire layer and simply passing the unchanged input to the next layer.
Here we propose to add an explicit gate, which we call \textit{copy gate}, to facilitate such a behavior.
We consider a $T$-layer Transformer encoder and an input sequence of length $N$.
Since each layer corresponds to one \textit{computational step},
we often refer to a layer as a step $t$.
We denote the Transformer state of column $i$ in layer $t$ as ${\bm{h}}^{(i, t)} = {\tens{H}}_{t,i} \in \mathbb{R}^{d}$ where $d$ is the state size,
and ${\tens{H}}_{t} \in \mathbb{R}^{N \times d}$ denotes
the states of all $N$ columns in layer $t$.
In the copy gate-augmented Transformer (Figure \ref{appendix:fig:copy_structure} in the appendix), each column $i$ in layer $(t+1)$ processes the input ${\tens{H}}_{t}$ similarly to regular Transformers:
\begin{align}
\label{eq:sa}
{\bm{a}}^{(i, t+1)} &= \layernorm(\mha({\bm{h}}^{(i, t)}, {\tens{H}}_{t}, {\tens{H}}_{t}) + {\bm{h}}^{(i, t)}) \\
\label{eq:ffn}
{\bm{u}}^{(i, t+1)} &= \layernorm(\ffn^\textrm{data}({\bm{a}}^{(i, t+1)}))
\end{align}
using the standard multi-head attention operation \citep{vaswani2017attention} $\mha$ with a query obtained from ${\bm{h}}^{(i, t)}$ and keys/values from ${\tens{H}}_{t}$, but the output is gated (using ${\bm{g}}^{(i, t+1)} \in \mathbb{R}^d$) as:
\begin{align}
\label{eq:gate}
{\bm{g}}^{(i, t+1)} &= \sigma(\ffn^\textrm{gate}({\bm{a}}^{(i, t+1)})) \\
\label{eq:out}
{\bm{h}}^{(i, t+1)} &= {\bm{g}}^{(i, t+1)} \odot {\bm{u}}^{(i, t+1)} + (1 - {\bm{g}}^{(i, t+1)}) \odot {\bm{h}}^{(i, t)}
\end{align}
We use the basic two-layer feedforward block \citep{vaswani2017attention} for both $\ffn^\textrm{data}$ and $\ffn^\textrm{gate}$ which transforms input ${\bm{x}} \in \mathbb{R}^{d}$ to:
\begin{equation}
\ffn ({\bm{x}}) = {\bm{W}}_2 \max({\bm{W}}_1 {\bm{x}} +{\bm{b}}_1, 0)+{\bm{b}}_2
\end{equation}
but with separate parameters and different dimensionalities:
for $\ffn^\textrm{data}$
${\bm{W}}_1^\textrm{data} \in \mathbb{R}^{d_\text{FF} \times d}$, ${\bm{W}}_2^\textrm{data} \in \mathbb{R}^{d \times d_\text{FF}}$, while for $\ffn^\textrm{gate}$ ${\bm{W}}_1^\textrm{gate}, {\bm{W}}_2^\textrm{gate} \in \mathbb{R}^{d \times d}$,
with biases ${\bm{b}}_1^\textrm{data} \in \mathbb{R}^{d_\text{FF}}$ and ${\bm{b}}_2^\textrm{data}, {\bm{b}}_1^\textrm{gate}, {\bm{b}}_2^\textrm{gate} \in \mathbb{R}^{d}$.
When the gate is closed i.e.~${\bm{g}}^{(i, t+1)} = 0$ in Eq.~\ref{eq:out},
the entire transformation is skipped and the input
is copied over to the next layer ${\bm{h}}^{(i, t+1)} = {\bm{h}}^{(i, t)}$.
Crucially, we parameterize the gate (Eq.~\ref{eq:gate}) as a function
of the output of the self-attention (Eq.~\ref{eq:sa}), such that the
decision to copy or transform the input for each column depends
on the states of all columns.
This is a crucial difference compared to previously proposed gatings
in Transformers, which are solely motivated by training stability \citep{parisotto2020stabilizing} or by a common practice from convolution-based models \citep{chaabouni2021can}.
None of the previous approaches
can implement the behavior of our copy gate (see Sec.~\ref{sec:rel} on related work). \looseness=-1
The bias of the gate ${\bm{b}}_2^\textrm{gate}$ is initialized to $-3$ \citep{hochreiter1997long}.
This ensures that no update happens initially to create a better gradient flow between layers.
It also encourages the model to skip layers unless they have an important contribution in the corresponding step.
\subsection{Geometric Attention: Learning to Attend to the Closest Match (Horizontal Flow)}
\label{sec:geo-attn}
We propose \textit{geometric attention} designed to attend to the closest matching element.
Like in regular self-attention,
given an input sequence $[{\bm{x}}^{(1)}, {\bm{x}}^{(2)}, ..., {\bm{x}}^{(N)}]$ with ${\bm{x}}^{(i)} \in \mathbb{R}^{d_{\text{in}}}$, each input is projected to
key ${\bm{k}}^{(i)} \in \mathbb{R}^{d_{\text{key}}}$, value ${\bm{v}}^{(i)} \in \mathbb{R}^{d_{\text{value}}}$, query ${\bm{q}}^{(i)} \in \mathbb{R}^{d_{\text{key}}}$ vectors, and the dot product is
computed for each key/query combination.
In our geometric attention, the dot product is followed
by a sigmoid function to obtain a score between 0 and 1:
\begin{align}
\label{eq:geo-attn-prob}
{\bm{P}}_{i,j} = \sigma( {\bm{k}}^{(j)\top}{\bm{q}}^{(i)})
\end{align}
which will be treated as a probability of the key at (source) position $j$ matching the query at (target) position $i$.
These probabilities are finally converted to the attention scores ${\bm{A}}_{i,j}$ as follows:
\begin{align}
\label{eq:geo-attn}
{\bm{A}}_{i,j} = {\bm{P}}_{i,j} \prod_{k \in \mathbb{S}_{i,j}} (1-{\bm{P}}_{i,k})
\end{align}
where $\mathbb{S}_{i,j}$ denotes the set of all (source) indices which are closer to $i$ than $j$ is to $i$, and when two indices have the same distance to $i$, we consider the one which is to the right of $i$ (i.e., greater than $i$) to be closer, i.e.,
\begin{align}
\label{eq:order}
\mathbb{S}_{i,j} &=
\begin{cases}
k \in \{1,...,N\}\setminus{\{i,j\}}: |i-k| < |i-j|, & \text{if } i < j\\
k \in \{1,...,N\}\setminus{\{i,j\}}: |i-k| \leq |i-j| , & \text{if } j < i \\
\end{cases}
\end{align}
In addition, we explicitly zero out the diagonal by setting
${\bm{A}}_{i,i}=0$ for all $i = {1,...,N}$.
The ordering of source indices is illustrated in Figure \ref{fig:geometirc_order}/Right.
The resulting scores ${\bm{A}}_{i,j}$ are the attention scores
used to compute the weighted averages of the value vectors.
By using the terms $(1-{\bm{P}}_{i,k})$ in Eq.~\ref{eq:geo-attn}, when there is a match, it downscales any other more distant matches.
Two recent works \citep{BrooksRLS21, banino2021pondernet} use such a parameterized geometric distribution in the form of Eq.~\ref{eq:geo-attn} (see Sec.~\ref{sec:rel} on related work).
The resulting attention function has a complexity of $O(N^2)$, similar to the regular self-attention used in Transformers \citep{vaswani2017attention}.
Eq.~\ref{eq:geo-attn} can be implemented in a numerically stable way in log space.
The products can then be calculated using cumulative sums, subtracting the elements for the correct indices in each position.
\begin{figure}[t!]
\begin{center}
\includegraphics[height=24mm]{figures/trafo_math.pdf}\hspace{20mm}
\includegraphics[height=24mm]{figures/geometric_order.pdf}
\caption{Left:~an ideal sequence of computations in a Transformer for an arithmetic expression.
Right: ordering (numbers in the grid) of source positions used in geometric attention (Eq.~\ref{eq:order}; $N=5$).
}
\label{fig:geometirc_order}
\end{center}
\end{figure}
\paragraph{Directional encoding.}
In practice, we augment Eq.~\ref{eq:geo-attn-prob} with
an additional \textit{directional encoding}.
In fact, the only positional information available in the geometric
attention presented above is the ordering used to define the product in Eqs.~\ref{eq:geo-attn}-\ref{eq:order}.
In practice, we found it crucial to augment the score computation of Eq.~\ref{eq:geo-attn-prob} with additional \textit{directional information},
encoded as a scalar ${\bm{D}}_{i,j} \in \mathbb{R}$
for each target/source position pair $(i, j)$:
\begin{align}
{\bm{D}}_{i,j} =
\begin{cases}
{\bm{W}}_\text{LR}{\bm{h}}^{(i)} + b_\text{LR}, & \text{if }i\leq j \\
{\bm{W}}_\text{RL}{\bm{h}}^{(i)} + b_\text{RL}, & \text{if } i > j
\end{cases}
\end{align}
where ${\bm{h}}^{(i)} \in \mathbb{R}^{d}$ denotes the input/state at position $i$ and ${\bm{W}}_\text{LR}, {\bm{W}}_\text{RL} \in \mathbb{R}^{1 \times d}$, $b_\text{LR}, b_\text{RL} \in \mathbb{R}$
are trainable parameters.
This directional information is integrated into the score computation
of Eq.~\ref{eq:geo-attn-prob} as follows (akin to how \citet{dai2019transformer} introduce the relative
positional encoding \citep{schmidhuber92ncchunker} as an extra term in the computation of attention scores):
\begin{align}
{\bm{P}}_{i, j} &= \sigma \big( \alpha \big({\bm{W}}_{q}{\bm{h}}^{(i)} + {\bm{b}}_q \big)^\top {\bm{W}}_{k, E} {\bm{h}}^{(j)} +
\beta {\bm{D}}_{i,j} + \gamma \big)
\label{eq:geom_att_prob_direction}
\end{align}
where the matrix ${\bm{W}}_{q} \in \mathbb{R}^{d_\text{head} \times d}$ maps the states to queries, ${\bm{b}}_q \in \mathbb{R}^{d_\text{head}}$ is a bias for queries, ${\bm{W}}_{k, E} \in \mathbb{R}^{d_\text{head} \times d}$ maps states to keys (we note that $d_\text{head}$ is typically the size of the key, query and value vectors for each head, $d_\text{head} = \frac{d}{n_\mathrm{heads}}$),
and $\alpha, \beta, \gamma \in \mathbb{R}$ are learned scaling coefficients and bias, initialized to $\alpha=\frac{1}{\sqrt{d_\text{head}}}, \beta=1, \gamma=0$.
Using this additional directional information, each query (position $i$) can potentially learn to restrict its attention to either the left or right side.
\section{Experiments}
\label{sec:exp_top}
We evaluate the proposed methods on three tasks:
the compositional table lookup \citep{liska2018memorize, hupkes2018learning}, a custom variant of ListOps \citep{nangia2018listops},
and a simple arithmetic task which we propose.
In all cases, the task is designed to test the compositional generalization ability of NNs:
the model has to learn to apply operations
seen during training in a longer/deeper compositional way (productivity). Further experimental details for each task can be found
in Appendix \ref{app:impl_detail}. \looseness=-1
\subsection{Compositional Table Lookup}
\label{sec:exp_ctl}
\paragraph{Task.} The compositional table lookup task \citep{liska2018memorize, hupkes2018learning, dubois2020location} is constructed based on a set of symbols and unary functions defined over these symbols.
Each example in the task is defined by one input symbol and a list
of functions to be applied sequentially, i.e., the
first function is applied to the input symbol and the resulting output
becomes the input to the second function, and so forth.
There are eight possible symbols.
Each symbol is traditionally represented by a 3-bit bitstring \citep{liska2018memorize}. However, in practice, they are simply processed as one token \citep{dubois2020location}.
The functions are bijective and randomly generated.
Each function is represented by a letter.
An example input is `\texttt{101 d a b}', which corresponds to the expression $b(a(d(101)))$; the model has to predict the correct output symbol.
We note that there exists a sequence-to-sequence variant of this task \citep{dubois2020location} where the model
has to predict all intermediate steps (thus trained with intermediate supervision).
We directly predict the final output.
An ideal model should be able to solve this task independently of the presentation order, that is, it should not matter
whether the task is encoded as `\texttt{101 d a b}' or `\texttt{b a d 101}'.
We thus study both forward (former) and backward (latter) variants of the task.
To evaluate systematic generalization, the train/valid/test
sets reflect different numbers of compositions: samples with 1-5/6-8/9-10 operations, respectively.
To best of our knowledge, no previous work has reported perfect accuracy on this task
through an NN.
We refer the readers to Sec.~\ref{sec:rel} for further details
on the previous work.\looseness=-1
\paragraph{Results.} We consider five different baselines: an LSTM \citep{hochreiter1997long}, bidirectional LSTM \citep{schuster1997bidirectional}, DNC \citep{graves2016hybrid, csordas2019improving},
Universal Transformers \citep{vaswani2017attention, dehghani2019universal}, and its relative position variants \citep{csordas2021devil}.
For Transformers, the prediction is based on the last column in the final layer (we conduct an ablation study on this choice in Appendix \ref{app:abla}).
The hyper-parameters used for each model
can be found in Table \ref{tab:hyperparams} in the appendix.
We also provide an ablation study on the number of layers needed for generalization in Appendix \ref{app:abla}, which supports our claim on the necessity for a ``sufficiently'' deep architecture.
The main results on this task are shown in Table \ref{tab:performance}.
The LSTM and DNC perform well in the forward variant, achieving perfect generalization for longer sequences, but fail on the backward variant. This is not surprising since in the forward case, input symbols are presented in the ``right'' processing order to the LSTM.
As expected, the bidirectional LSTM performs well in both presentation orders,
since one of its processing directions is always aligned with the order of computation.
However, for an arbitrary task, the order of processing is not given. For example, for ListOps (Sec.~\ref{sec:listops}), the processing should start from the deepest point in the parse tree, which is probably somewhere in the middle of the sequence.
The experiments on other tasks (Sec.~\ref{sec:simple_arithmetic} and \ref{sec:listops}) requiring arbitrary processing orders show that bidirectional LSTMs do not generalize well in such tasks.
This is not satisfactory since our goal is to create a generic architecture which can solve arbitrary problems with an arbitrary underlying input processing order.
While the Transformer seems to be a good candidate for learning
problem dependent processing orders, the baseline Transformer variants
fail to generalize in this task in both directions.
By introducing the copy gate (Sec.~\ref{sec:gating}), the relative Transformer can solve the forward task, but not the backward one. Our analysis showed that the network learns to attend to the last operation based on the relative position information.
Since the result is read from the last column, this position changes with the sequence length.
The model thus fails to generalize to such arbitrary offsets.
To address this issue, we introduce a simple mechanism to let the model choose between absolute and relative positional encodings at each position (see Appendix \ref{sec:app-absrel}).
The resulting model effectively manages to use the absolute position for the prediction and perform well in both directions.
However, such a combination of absolute/relative positional encoding might be an overly specific bias.
A more generic solution, geometric attention (Sec. \ref{sec:geo-attn}), also achieved perfect generalization and was found easier to train.
We present the corresponding visualization of our model in Sec.~\ref{sec:analysis}.
\begin{table*}[ht]
\centering
\small
\caption{Accuracy on \textbf{compositional table lookup} dataset.}
\label{tab:performance}
\begin{tabular}{l c c c c}
\toprule
& \multicolumn{2}{c}{IID} & \multicolumn{2}{c}{Longer} \\
\cmidrule(lr){2-3} \cmidrule(lr){4-5}
Model & Forward & Backward & Forward & Backward \\
\midrule
LSTM & \bf{1.00 $\pm$ 0.00} & 0.59 $\pm$ 0.03 & \bf{1.00 $\pm$ 0.00} & 0.22 $\pm$ 0.03 \\
Bidirectional LSTM & \bf{1.00 $\pm$ 0.00} & \bf{1.00 $\pm$ 0.00} & \bf{1.00 $\pm$ 0.00} & \bf{1.00 $\pm$ 0.00} \\
DNC & \bf{1.00 $\pm$ 0.00} & 0.57 $\pm$ 0.06 & \bf{1.00 $\pm$ 0.00} & 0.18 $\pm$ 0.02 \\
\midrule
Transformer & \bf{1.00 $\pm$ 0.00} & 0.82 $\pm$ 0.39 & 0.13 $\pm$ 0.01 & 0.12 $\pm$ 0.01 \\
\quad + rel & \bf{1.00 $\pm$ 0.00} & \bf{1.00 $\pm$ 0.00} & 0.23 $\pm$ 0.05 & 0.13 $\pm$ 0.01 \\
\quad + rel + gate & \bf{1.00 $\pm$ 0.00} & \bf{1.00 $\pm$ 0.00} & \bf{0.99 $\pm$ 0.01} & 0.19 $\pm$ 0.04 \\
\quad + abs/rel + gate & \bf{1.00 $\pm$ 0.00} & \bf{1.00 $\pm$ 0.00} & \bf{0.98 $\pm$ 0.02} & \bf{0.98 $\pm$ 0.03} \\
\quad + geom. att. & 0.96 $\pm$ 0.04 & 0.93 $\pm$ 0.06 & 0.16 $\pm$ 0.02 & 0.15 $\pm$ 0.02 \\
\quad + geom. att. + gate (NDR) & \bf{1.00 $\pm$ 0.00} & \bf{1.00 $\pm$ 0.00} & \bf{1.00 $\pm$ 0.00} & \bf{1.00 $\pm$ 0.00} \\
\bottomrule
\end{tabular}
\end{table*}
\subsection{Simple Arithmetic} \label{sec:simple_arithmetic}
In order to validate the success of the proposed model on
a task that involves more complex data flows and operations,
we propose the \textit{simple arithmetic} task.
\paragraph{Task.}
The task is to execute an arithmetic expression consisting of nested modulo 10 additions and multiplications.
This requires the model to process tree-structured data flows,
which is presumably more difficult than the sequential
processing required for the CTL task.
Each operation is surrounded by brackets, such that the boundaries of operations are easy to determine.
For example `\texttt{((4*7)+2)}' should evaluate to `\texttt{0}' (30 modulo 10).
The expressions are generated randomly.
The tree depth is up to 5 for the training set, 6 for the validation set,
and 7-8 for the test set.
The depth
is measured as the number of operations, ignoring the leaves, so the example above has a depth of 2.
The sequence length is limited to at most 50 tokens.
\paragraph{Results.} Table \ref{tab:performance_arithmetics} shows the results.
All considered models perform well on the IID validation data, but none except the NDR performs well on the generalization test set, which achieves near-perfect accuracy of 98\%.
We also note that the NDR learns very quickly:
while all other models require about 200\,K steps to converge,
the NDR achieves near-perfect accuracy
after 50\,K steps of training.
\begin{table*}[ht]
\centering
\small
\caption{Performance of different models on the \textbf{simple arithmetic} dataset. All models are trained for 200\,K iterations, except the NDR which we stop training at 100\,K. We also report the performance after 50\,K iterations, where it can be seen that NDR converges significantly faster than the others.}
\label{tab:performance_arithmetics}
\begin{tabular}{l c c c c}
\toprule
& IID (1..5) & \multicolumn{2}{c}{Test (7..8)} \\ \cmidrule(r){2-2} \cmidrule(r){3-4}
& 200\,K & 200\,K & 50\,K \\
\midrule
LSTM & \bf{0.99 $\pm$ 0.00} & 0.74 $\pm$ 0.02 & 0.72 $\pm$ 0.01 \\
Bidirectional LSTM & \bf{0.98 $\pm$ 0.01} & 0.82 $\pm$ 0.06 & 0.80 $\pm$ 0.04 \\
\midrule
Transformer & \bf{0.98 $\pm$ 0.01} & 0.47 $\pm$ 0.01 & 0.29 $\pm$ 0.01 \\
\quad + rel & \bf{1.00 $\pm$ 0.00} & 0.77 $\pm$ 0.04 & 0.40 $\pm$ 0.05 \\
\quad + abs/rel + gate & \bf{1.00 $\pm$ 0.01} & 0.80 $\pm$ 0.16 & 0.73 $\pm$ 0.15 \\
\quad + geom. att. + gate (NDR) & \bf{1.00 $\pm$ 0.00} & \bf{0.98 $\pm$ 0.01} & \bf{0.98 $\pm$ 0.01} \\
\bottomrule
\end{tabular}
\end{table*}
\subsection{ListOps} \label{sec:listops}
We also evaluate our model on a variant of the ListOps task \citep{nangia2018listops} which is a popular task commonly used to evaluate parsing abilities of NNs \citep{havrylov2019cooperative, ShenTHLSC19, XiongZCTFLS21, tay2021long, irie2021going}. Some special architectures such as \citet{ChowdhuryC21} can almost perfectly generalize to longer sequences on this task.
However, as far as we know,
no Transformer variant has been reported to be fully successful.
\paragraph{Task.}
The task consists of executing nested list operations written in prefix notation. All operations have a list of arguments that can be either a digit (from 0 to 9)
or recursively another operation with its own list of arguments.
The operations are min, max, median and sum.
The sum is modulo 10, and the median is followed by the floor function such that the output of any operation lies
between 0 and 9.
For example: \texttt{[MED 4 8 5 [MAX 8 4 9 ] ]} should return \texttt{6}.
There are two well-known variants:
the original one by \citet{nangia2018listops} and
the ``Long Range Arena'' variant by \citet{tay2021long}
which have different maximum numbers of arguments in each function and maximum sequence lengths.
In both variants, there is no strict control of the depth of data samples: there is simply a certain pre-defined probability that each argument in the list
is expanded into another list (which may increase the tree depth).
This is not suitable for evaluating systematic generalization
in terms of compositionality (over the problem depth).
We propose instead to generate clean train, valid, and test splits with disjoint depths: up to depth 5 for training,
depth 6 for validation and depths 7 and 8 for test.
Importantly, we make sure that a depth-$K$ sample
effectively requires computation until depth-$K$
(otherwise min, max, and med operations
could potentially find the output without executing all of its arguments).
By dissociating the splits by the depth, we can clearly identify
models which fail to generalize compositionally.
Apart from the depth specifications, all train/valid/test sets
share the same settings as follows:
the maximum sequence length is 50 (tokens),
the probability of recursively sampling another function inside a list is 30\% at each position, and the maximum number of arguments for a function is 5.
The train set consists of 1M, the valid and test sets of 1K sequences.
\paragraph{Results.}
Table \ref{tab:performance_listops} shows the results.
Like on the other
tasks, the baseline LSTM and Transformers do not generalize well on the test set consisting of deeper problems, while they achieve a near-perfect accuracy on IID data.
In contrast, our model achieves near-perfect generalization.
\begin{table*}[ht]
\centering
\small
\caption{Performance of different models on balanced \textbf{ListOps} dataset. All models are trained for 200\,K iterations, except all \texttt{+gate} variants which converge after 100\,K steps.
The numbers in the parentheses indicate the problem depths (1-5 for the IID, and 7-8 for the test set).}
\label{tab:performance_listops}
\begin{tabular}{l c c}
\toprule
& IID (1..5) & Test (7..8) \\
\midrule
LSTM & \bf{0.99 $\pm$ 0.00} & 0.71 $\pm$ 0.03 \\
Bidirectional LSTM & \bf{1.00 $\pm$ 0.00} & 0.57 $\pm$ 0.04 \\
\midrule
Transformer & 0.98 $\pm$ 0.00 & 0.74 $\pm$ 0.03 \\
\quad + rel & \bf{0.98 $\pm$ 0.01} & 0.79 $\pm$ 0.04 \\
\quad + abs/rel + gate & \bf{1.00 $\pm$ 0.01} & 0.90 $\pm$ 0.06 \\
\quad + geom. att. + gate (NDR) & \bf{1.00 $\pm$ 0.00} & \bf{0.99 $\pm$ 0.01} \\
\bottomrule
\end{tabular}
\end{table*}
\section{Analysis}
\label{sec:analysis}
In this section, we provide some visualizations of
attention and gating patterns of the NDR
and the corresponding analyses.
For more visualizations, we refer the readers to Appendix \ref{app:visual}.
\paragraph{Compositional Table Lookup.}
Figure \ref{fig:gates} shows the gating and attention patterns
of the NDR model
for an example of the backward presentation task.
As shown in Fig.~\ref{fig:gates}/Bottom, the gates of different columns open sequentially one after another when the input is available for them.
Fig.~\ref{fig:gates}/Top shows the corresponding attention maps.
Each column attends to the neighbouring one, waiting for its computation to be finished.
The behavior of the last column is different: it always attends to the second position of the sequence, which corresponds to the last operation to be performed.
\begin{figure}[ht]
\begin{center}
\includegraphics[width=130mm]{figures/att_and_gates.pdf}
\caption{
Example visualization of NDR.
For other models, see Appendix \ref{app:visual}.
Top: Attention map for different steps.
The x/y-axis corresponds to source/target positions, respectively.
Each position focuses on the column to the right, except the last one where the result is read from, which focuses on the last operation. The focus becomes clear only once the result is available.
Bottom: gate activations for different steps/layers. The gates remain closed until the data dependencies are satisfied.}
\label{fig:gates}
\end{center}
\end{figure}
\begin{figure}[ht]
\begin{center}
\includegraphics[width=130mm]{figures/listops_annot2.pdf}
\caption{
Example visualization of NDR on ListOps. The top row shows \textit{head 13} in different steps, which controls which arguments are used in which step. The bottom row shows different heads in different key steps.
Please refer to Sec.~\ref{sec:analysis} for the step-by-step description.
More visualizations are provided in the appendix: Fig.~\ref{appendix:fig:all_attention_geom_listops} shows the max of attention over all heads for all steps,
Fig.~\ref{appendix:fig:all_attention_geom_listops_h13} shows all steps of \textit{head 13},
and Fig.~\ref{appendix:fig:all_gates_geom_listops} shows the corresponding gates.}
\label{fig:listops_annot}
\end{center}
\end{figure}
\paragraph{ListOps.}
We can also identify how the NDR processes the data in ListOps.
Different attention heads play different roles.
We highlight the core observations in Figure \ref{fig:listops_annot}.
The input for this example is: \texttt{[SM [MED [MIN 1 7 4 [MAX 2 4 0 8 9 ] ] 7 ] 5 [MED 8 5 8 ] 0 7 ]}.
First of all, we find that there is a head (\textit{head 13} in Figure \ref{fig:listops_annot}, \textit{first row}) which seems to be responsible for connecting operators and their arguments: the operands/arguments of an operation attend to the operator.
In step 0 ($t=0$ in the figure), we can recognize that the operations at the deepest level,
namely \texttt{MAX} and the second \texttt{MED} have all the arguments ready (as is shown by vertical lines on the columns corresponding to \texttt{MAX} and \texttt{MED}).
The model indeed identifies that these two operations are ready to be executed and that they can be processed in parallel (these arguments-to-operation attention patterns remain
for a few steps).
We note that at this stage, the last argument of \texttt{MIN} is not ready yet (\texttt{[MIN 1 7 4 [MAX 2 4 0 8 9 ] ]}).
We can see that only arguments which are already ready (\texttt{1 7 4}) attend to the operator (see the column of \texttt{MIN}).
In step 1 ($t=1$, \textit{2nd row}), we can see that \textit{head 5} copies the expected result of \texttt{MAX}, \texttt{9} to the column of the operator (we note that this only requires one step as \texttt{9} is always the result of \texttt{MAX} when it is one of the arguments of \texttt{MAX}).
Similarly in step 2, \textit{head 7} (\textit{2nd row}) seems to copy the result of the second \texttt{MED},
\texttt{8} to the operator column.
In step 3 ($t=3$, \textit{1st row}), we recognize that the result of \texttt{MAX} is marked as an argument for \texttt{MIN}
in \textit{head 13} which is responsible for communication between
operators and their arguments.
This is shown by the new attention which appears at $t=3$ in \textit{head 13} from the source position \texttt{MAX} to the target position \texttt{MIN} (a pattern which is not visible at $t=2$).
In \textit{head 3}, $t=6$ (\textit{2nd row}), the expected result of \texttt{MIN}, which is \texttt{1},
is copied to the operator, similarly to the patterns we observed above for \texttt{MAX} and \texttt{MED}.
In \textit{head 13}, $t=6$ (\textit{1st row}), all arguments for the first \texttt{MED}
are now also recognized (the result of \texttt{MIN} which is \texttt{1}, and \texttt{7}).
Finally in $t=7$ (\textit{2nd row}), two heads, \textit{head 3} and \textit{head 5} seem to copy/gather two inputs
needed to compute the corresponding median, \texttt{1} and \texttt{7}, and store them in the column of the operator \texttt{MED}.
A complete visualization of further steps can be found in Appendix \ref{app:plot_listops}.
We noticed that some of the heads do not seem to play a key role; we focused on interpreting those which seem to participate
in the main computation.
For ListOps, we also partially find the attention patterns
described above in the baseline Transformer with relative positional encoding,
at least on some inspected examples,
which also explains its rather high accuracy.
\section{Discussion}
\paragraph{Learning adaptive serialization.}
The NDR architecture can be understood as performing adaptive serialization of the problem.
A key requirement for reusable computation is decomposing the problem into reusable building blocks, typically applied in sequential steps.
The granularity of the decomposition determines the degree of reusability: fusing operations in a single step makes the processing faster (fewer steps), but also more specialized. Learning the most granular solutions is thus preferable for generalization.
At the same time, not all processing should happen serially: branches of the computational graph that do not have common data dependencies can be processed independently in parallel, which we empirically observe
in our NDR in the ListOps example (Sec.~\ref{sec:analysis}).
This enables the architecture to get away with a number of computational steps reflecting the depth of the computational graph rather than the length of the input.
\paragraph{Bottom up approach for improving model architectures.}
Transformers have seen tremendous successes across various application domains
\citep{devlin2019bert, gpt3, dosovitskiy2021an}.
Impressive results have been reported when they are scaled up
with a large amount of data \citep{gpt3}.
On the other hand, simple tasks like those highlighted in the present work demonstrate that the Transformer architecture still struggles with basic reasoning.
Particularly in algorithmic tasks, it is often the case that a sub-optimal choice of architecture/optimization method makes the model fall back to simple memorization.
We argue that it is crucial to look at isolated problems which test
specific generalization capability.
This calls for a bottom-up approach: building on toy tasks that focus on individual aspects of generalization and using them for improving models.
\section{Related Work}
\label{sec:rel}
\paragraph{Gating inside Transformers.}
Several prior works have proposed to use some sort of gating
within Transformer architectures \citep{parisotto2020stabilizing, chaabouni2021can}.
Our proposed \textit{copy gate} is different from those
as it satisfies two important properties.
First, our copy gate allows the model to skip the \textit{entire} Transformer layer
(i.e., both the self-attention and the feedforward blocks) when
the gate is closed.
Second, the gate function is conditioned on the attention output
such that the decision of opening or closing depends
on information from all columns.
While multiple gating variants have been proposed by \citet{parisotto2020stabilizing} to stabilize Transformers for reinforcement learning, none of them can produce this behavior.
Empirically, we also tried out a few other gating variants which do not satisfy the two properties above; we found them not to improve over regular Transformers in our preliminary experiments on compositional table lookup. Recent work by
\citet{chaabouni2021can} also makes use of ``gating'' in Transformers through a gated linear unit (GLU) activation function
commonly used in convolutional NNs \citep{dauphin2017lm}.
Transformer models with such an activation function were reported to outperform RNN baselines on a systematic generalization task \citep{DessiB19}.
Unlike our copy gate or \citet{parisotto2020stabilizing}'s gating, such a gating activation does not have the ``residual'' term
(i.e.~a closed gate zeros out the input) which allows the model to skip a transformation.
In a more general context,
benefits of the GLU activation in Transformers vary across tasks \citep{irie19trafolm, shazeer2020glu}.
In language modeling, no improvement is typically obtained by
using the standard highway gate instead of the residual connection in Transformers \citep{irie2020phd},
while it yields improvements when combined with convolutional layers \citep{kimcharacter}.\looseness=-1
\paragraph{Parameterized geometric distributions.}
Two recent works \citep{BrooksRLS21,banino2021pondernet} have used a form of parameterized geometric distribution (PGD; in the form of Eq.~\ref{eq:geo-attn}).
\citet{BrooksRLS21} have used such a distribution to parameterize
the movement of a pointer on a sequence of instructions.
\citet{banino2021pondernet} have used it to implement adaptive computation time \citep{Schmidhuber:12slimnn, graves2016adaptive}.
We use the PGD to obtain a generic attention mechanism as a replacement
of the standard self-attention used in Transformers \citep{vaswani2017attention}.
\paragraph{Compositional table lookup.}
CTL task was proposed for evaluating
the compositional ability of NNs \citep{liska2018memorize}.
Previous works evaluated RNNs, RNNs with attention,
and Transformers on this task with
limited success \citep{hupkes2018learning, dubois2020location}.
\citet{dubois2020location} have proposed a special attention mechanism
to augment the recurrent architecture.
While they obtained good performance for the forward presentation order, the proposed model failed in the backward one.
In contrast, two of our approaches (Sec.~\ref{sec:exp_ctl})
achieve 100\% generalization accuracy for both orders.
\paragraph{Positional encodings.}
Many previous works have focused on improving positional encoding \citep{schmidhuber92ncchunker, vaswani2017attention}
for self-attention.
Most notably, the relative positional encoding \citep{schmidhuber92ncchunker, shaw2018selfattention, dai2019transformer} was found useful for improving
systematic generalization of Transformers \citep{csordas2021devil}.
Here we also present two new approaches related to positional encoding.
One is the gated combination of absolute and relative positional encoding (Sec.~\ref{sec:exp_ctl}; details in Appendix \ref{sec:app-absrel}).
We show that absolute positional encoding can complement relative
positional encoding.
The former enables the model to always attend
to a specific position, as is needed for the CTL task in the last step,
while the gating allows it to use relative positional encoding for other positions/steps.
Second, we introduce directional encoding to augment geometric attention.
Unlike positional encoding which can overfit to a range of positions seen during training, the direction information is found to be robust
and to be a crucial augmentation of the geometric attention. \looseness=-1
\section{Conclusion}
We proposed a new view on the internal operations of Transformer encoders as a dynamic dataflow architecture between Transformer columns.
This overcomes two shortcomings of traditional Transformers: the problem of routing and retaining data in an unaltered fashion, which we solve by an additional copy gate, and the problem of learning length-independent attention patterns, which we solve by geometric attention.
Our new model, the Neural Data Router (NDR), generalizes to compositions longer than those seen during training on the popular compositional lookup table task in both forward and backward directions.
NDR also achieves near perfect performance on simple arithmetic and ListOps tasks in settings that test systematic generalization in terms of computational depth.
In general, the gates and the attention maps collectively make the architecture more interpretable than the baselines. Future work will extend this encoder-only architecture
to a full sequence-to-sequence model and evaluate it
on other standard tasks in systematic generalization requiring generation of variable-length output sequences.
\section*{Acknowledgments}
We thank Imanol Schlag and Sjoerd van Steenkiste for helpful discussions and suggestions on an earlier version of the manuscript.
This research was partially funded by ERC Advanced grant no: 742870, project AlgoRNN,
and by Swiss National Science Foundation grant no: 200021\_192356, project NEUSYM.
We are thankful for hardware donations from NVIDIA \& IBM. The resources used for the project were partially provided by Swiss National Supercomputing Centre (CSCS) project s1023.
|
\subsection{Sufficient conditions for detectability}
\label{sec:detectability-nse}
\begin{definition}\label{def:detectability-class}
Take $\mathcal C\in\mathscr{L}(L^2)$ such that $\mathcal{V}\subseteq\mathscr{D}(\mathcal C)$. $\mathcal C$ is said to be of class $\mathscr{C}_1(h,\Omega)$ or of class $\mathscr{C}_2(h,\Omega)$ if~\eqref{eq:CPoin} or ~\eqref{eq:CPoin2} respectively holds for some $h>0$ and $C_\Omega>0$
\begin{align}
\forall\vec u\in\mathring{\mathcal{V}}:\quad \|\vec u - \mathcal C\vec u\|_{L^2}^2&\le h^2 C_\Omega \|\nabla\vec u\|_{L^2}^2 \label{eq:CPoin}\\
\forall\vec u\in \mathscr{D}(A):\quad \|\vec u - \mathcal C\vec u\|_{L^2}^2&\le h^2 C_\Omega \|\Delta\vec u\|_{L^2}^2 \label{eq:CPoin2}
\end{align}
\end{definition}
\begin{prop}\label{prop:detectability-nse-1}
Recall Def.~\ref{def:detectability}: let $\vec z,\vec u$ solve~\eqref{eq:NSEu}-\eqref{eq:NSEz}. Recall from~\cite[p.]{NSE-Turbulence-book2001} the interpolation inequality and define constants $C_{1,2}$:
\begin{align}
&\forall\vec u\in\mathscr{D}(A):\|\nabla\vec u\|^2_{L^2}\le C_{\nabla} \|\vec u\|_{L^2} \|A\vec u\|_{L^2} \label{eq:InterpIneq}\\
&C_1^2=(1+\lambda_1^{-1})/(2\pi), \quad C_2^2 = \|\vec \ell\|^2_{\mathbb{R}^2}/(32\pi^3)\label{eq:C1C2}
\end{align}
Take $\mathcal C\in\mathscr{C}_1(h,\Omega)$. Then NSE is detectable in $\mathring{\mathcal{V}}$ if for some $1<\kappa<<2$
\begin{equation}
\label{eq:h-bnd-detect}
h< \frac{\lambda_1^{\frac12}\nu\bigl(\frac{3\nu}{4} -\frac{C_2}{\lambda_1\Gamma}\bigr)}{C_1 C_\nabla C^{\frac12}_\Omega \kappa^{\frac12}\|\vec f\|_{L^2}\log^{\frac12}\bigl(1+\frac{4\pi^2 \kappa \|\vec f\|_{L^2}^2\Gamma^2 }{\nu^2\ll_1\ll_2}\bigr)}
\end{equation}
moreover, if $\mathcal C\in\mathscr{C}_2(h,\Omega)$ then NSE is detectable in $\mathring{\mathcal{V}}$ if $h$ verifies~\eqref{eq:h-bnd-detect} without $\lambda_1^{\frac12}$ in the numerator.
\end{prop}
\begin{example}\label{exmpl:detectability-nse-1}
The class of operators $\mathscr{C}_1(h,\Omega)$ generalizes the notion of so-called \textit{determining modes} for NSE which is well-known in NSE's literature (see e.g.~\cite[p.123]{NSE-Turbulence-book2001}). The case of determining modes corresponds to $\mathcal C\in\mathscr{C}_1(h,\Omega)$ of finite rank, $\mathcal C\vec u$ represents the projection of $\vec u$ onto a certain finite-dimensional subspace of $\mathcal{H}$, and in this case $\mathcal C$ is bounded in $\mathcal{H}$. As an example, consider the case of spatial averages over a partition of $\Omega$ (e.g.~\cite{emilia2,azouani2014continuous}): Let $\Omega=\cup_{j=1}^N \Omega_j$, $\Omega_j\cap\Omega_i=\varnothing$ for $i\ne j$, $\Omega_j$ -- a rectangle with sides of length $h_x^j$ and $h_y^j$ respectively, and let $\xi_j$ denote the indicator function of $\Omega_j$. An example of a uniform square partition of $\Omega$ is provided in the left panel of Fig.~\ref{fig:1}. Let \[
\mathcal C\vec u(x,y) = \sum_{j=1}^N\bigl[
\begin{smallmatrix}
(u_1,\xi_j)\\(u_2,\xi_j)
\end{smallmatrix}\bigr]\frac{\xi_j(x,y)}{(h_x^j h_y^j)^{\frac12}}, \quad \xi_j(x,y)=1, (x,y)\in\Omega_j
\] Then~\eqref{eq:CPoin} holds for $h^2=\max_j(\max\{h_x^j,h_y^j\})^2$ and $C_\Omega=(4\pi^2)^{-1}$ by Poincare inequality~\eqref{eq:Poincare}.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.49\columnwidth]{figures/Obs}
\includegraphics[width=0.49\columnwidth]{figures/State}
\caption{Example of the 1st component of $\mathcal C\vec u$ (left panel), versus $1st$ component of $\vec u$}
\label{fig:1}
\end{center}
\end{figure}
\end{example}
\begin{remark}
Fig.~\ref{fig:1} provides a good illustration of the detectability concept: the left panel represents 1st component of $\mathcal C\vec u$, the projection of the $1st$ component of $\vec u$ (right panel) onto the orthonormal set of $N$ indicator functions $\xi_j$, and $\vec u$ solves NSE~\eqref{eq:NSEu}. Comparing panels of Fig.~\ref{fig:1} it is evident that there is little or no hope to reconstruct $\vec u$ from $\mathcal C\vec u$, e.g. by interpolation, without using the structure of NSE~\eqref{eq:NSEu}. In this situation one can reconstruct the continuous state from observations, if NSE is detectable, i.e. if $h^2=\max_j(\max\{h_x^j,h_y^j\})^2$ is small enough (e.g.~\eqref{eq:h-bnd-detect}) for the diffusion term $\nu((\vec u,\vec u))$ to dominate the non-linear term $b$ in the ``unmeasured'' infinite-dimensional orthogonal complement of the span of $\xi_j$. In other words, the span of $\xi_j$ (and more generally the measured output $\mathcal C\vec u$) need to ``reveal'' a certain part of the state space where the nonlinear term $b$ is not ``dominated'' by the diffusion term $\nu((\vec u,\vec u))$.
\end{remark}
\begin{example}\label{exmpl:detectability-nse-1}
The class of operators $\mathscr{C}_2(h,\Omega)$ generalizes notion of so-called \textit{determining nodes} for NSE (see e.g.~\cite[p.131]{NSE-Turbulence-book2001}). The case of determining nodes corresponds to $\mathcal C\in\mathscr{C}_2(h,\Omega)$ of the form $\mathcal C\vec u=\sum_{j=1}^n \vec u(t,\vec x_j)\vec{\phi}_j$ for some grid nodes $\vec x_j\in\Omega$ and functions $\vec{\phi}_j\in H^1_p(\Omega)^2$. Clearly, $\mathcal C$ is not bounded in $\mathcal{H}$ as it is defined for $\vec u\in H^2_p(\Omega)^2\cap\mathcal{H}$, but it is a closed linear operator such that $\mathscr{D}(A)\subset\mathscr{D}(C)$. As an example, consider the case of pointwise evaluations over nodes $\vec x_j$ a finite element grid, here $\vec{\phi}_j$ are 2D hat basis functions. In this case $\mathcal C$ verifies~\eqref{eq:CPoin2} indeed with constants $h$ and $C_\Omega$ which can be found in either~\cite[Th 12.3.4]{Hutson1980}.
\end{example}
\subsection{Observer design}
\label{sec:observer-design}
Proposition~\ref{prop:detectability-nse-1} describes classes $\mathscr{C}_1(h,\Omega)$ and $\mathscr{C}_2(h,\Omega)$ of linear operators $\mathcal C$ which make NSE detectable, provided~\eqref{eq:h-bnd-detect} holds. However, as is, the condition~\eqref{eq:h-bnd-detect} cannot be used in practise as one needs to know the interpolation constant $C_\nabla$. This is not surprising as condition~\eqref{eq:h-bnd-detect} in the descriptions of $\mathscr{C}_1(h,\Omega)$ and $\mathscr{C}_2(h,\Omega)$ accounts for generic $\vec f$ and $\vec F$. The following result further refines condition~\eqref{eq:h-bnd-detect} by imposing a certain structure on $\vec F$: namely, it is demonstrated that the classical Luenberger-type output-feedback control $\vec F =\vec g+L\mathcal C(\vec u-\vec z)$ guarantees condition $A)$ of Def.~\ref{def:detectability}, and $\vec e=\vec u-\vec z\to0$, provided $\vec f-\vec g$ ``dissapeares'' at infinity and $h$ verifies an easy-to-check condition.
\begin{lemma}\label{l:observer-existence}
Let $\vec u$ solve~\eqref{eq:NSEu} on $\Omega=(-\frac{\ell_1}2, \frac{\ell_1}2)\times (-\frac{\ell_2}2, \frac{\ell_2}2)$ and let $\mathcal C\in\mathscr{C}_1(h,\Omega)$ or $\mathcal C\in\mathscr{C}_2(h,\Omega)$. For any $\vec g\in\mathring{\mathcal{H}}$ the following equation ($\vec e=\vec u-\vec z$, $\vec \phi\in\mathring{\mathcal{V}}$):
\begin{align}
\displaystyle\frac{d}{dt}(\vec z,\vec \phi) &+ b(\vec z,\vec z,\vec\phi) + \nu((\vec z,\vec \phi)) = (\vec g+L\mathcal C\vec e,\vec \phi)\,, \label{eq:NSE-LO}
\end{align}
has the unique solution $\vec z\in L^2(0,T,\mathscr{D}(A))\cap C(0,T,\mathring{\mathcal{V}})$.
\end{lemma}
\begin{theorem}\label{t:1}
Let $\vec u$ solve~\eqref{eq:NSEu} and $\vec z$ solve~\eqref{eq:NSE-LO} provided
\begin{equation}
\label{eq:fgdecay}
\limsup_{t\to\infty}\frac1T\int_t^{t+T}\|\vec f(s)-\vec g(s)\|^2_{L^2} ds=0\,.
\end{equation}
For $\kappa>1$ pick $t^\star>0$ and $T>0$ such that $\theta_{t^\star,T}\le \kappa\|\vec f\|^2_{L^2}\nu^{-2}$ where \[
\theta_{t,T} = \frac{2\|\vec f\|^2_{L^2}}{T\nu^3\lambda_1} + \frac{2\mathbb{e}^{(-\lambda_1 \nu)t}}{T\nu}\|\nabla\vec u_0\|^2_2 +\frac{\|\vec f\|^2_{L^2}}{\nu^2}
\] and take $\Gamma_{\mathrm{max}}\in \operatorname{Argmax}\Theta(\Gamma)$ where
\begin{align*}
\Theta(\Gamma)&=\frac{(2\pi)^{\frac32}\nu(\nu -\frac{\|\vec \ell\|_{\mathbb{R}^2}}{\sqrt{32\pi^3}\lambda_1\Gamma}) \log^{-\frac12}\bigl(1+\frac{4\pi^2 \kappa\|\vec f\|^2_{L^2}\Gamma^2 }{\nu^2\ll_1\ll_2}\bigr)}{C_\Omega (4\kappa\pi^2+\kappa\max\{\ell_1,\ell_2\}^2)^{\frac12} \|\vec f\|_{L^2}}
\end{align*}
Then $V(t)=\|\vec u(t)-\vec z(t)\|^2_{H^1}\to 0$ for $t>t^\star$ if
\begin{itemize}
\item $\mathcal C\in\mathscr{C}_1(h,\Omega)$, and $h^2<\beta \Theta(\Gamma_{\mathrm{max}})$, $L=\beta \hat L_\nabla$ for $0<\beta<1$ where
\begin{align}
\label{eq:hatL}
\hat L_\nabla \coloneqq\frac{2(\nu -\frac{\|\vec \ell\|_{\mathbb{R}^2}}{\sqrt{32\pi^3}\lambda_1\Gamma_{\mathrm{max}}})}{h^{2} C_\Omega}, \quad \vec \ell=
\bigl[\begin{smallmatrix}
\ell_1\\\ell_2
\end{smallmatrix}\bigr],
\end{align}
\item $\mathcal C\in\mathscr{C}_2(h,\Omega)$ and $h< \beta C_\Omega^{\frac12}\theta^{-1}\Theta(\Gamma_{\mathrm{max}})$, $L=\theta\hat L_\Delta$ for $\theta>1$, where
\begin{align}\label{eq:hatL1}
\hat L_\Delta\coloneqq \frac{(\kappa+\kappa\lambda_1^{-1})^\frac12 \|\vec f\|_{L^2}}{\nu\sqrt{2\pi}} \log^{\frac12}\bigl(1+\frac{4\pi^2 \kappa \|\vec f\|_{L^2}^2 \Gamma_{\mathrm{max}}^2 }{\nu^2\ll_1\ll_2}\bigr),
\end{align}
\end{itemize}
\end{theorem}
\begin{remark}
If $\|\vec f(s)-\vec g(s)\|_{\mathbb{R}^2}=0$ for $s>s^\star$ then $V\to0$ \emph{exponentially fast} for $\beta=1$ and $t>\max\{t^\star,s^\star\}$. If $\|\vec f\|_{L^2},\|\nabla \vec u_0\|_{L^2}<\bar C$ for a given $\bar C$ then $t^\star$ and $T$ can be computed explicitly for every $\kappa$.
\end{remark}
\section{Introduction}
\label{sec:introduction}
\input{introduction.tex}
\subsection{Mathematical preliminaries}
\label{sec:math-prel-1}
\input{preliminaries.tex}
\section{Problem statement}
\label{sec:problem-statement}
\input{prob_statement}
\section{Main results}
\label{sec:main-results}
\input{detectability.tex}
\input{observer.tex}
\section{Experiments}
\label{sec:experiments}
\input{experiments.tex}
\section{Conclusions}
\label{sec:conclusions}
In this work we obtained proposed {\it easy-to-verify} detectability conditions for Navier-Stokes Equation (NSE) in two spatial dimensions and unknown inputs, and designed a globally converging observer of Luenberger type for the case of continuous in time measurements. A promising research direction now is to consider the case of sampled measurements.
\bibliographystyle{plain}
\subsection{Observer design}
\label{sec:observer-design}
Proposition~\ref{prop:detectability-nse-1} describes classes $\mathscr{C}_1(h,\Omega)$ and $\mathscr{C}_2(h,\Omega)$ of linear operators $\mathcal C$ which make NSE detectable, provided~\eqref{eq:h-bnd-detect} holds. However, as is, the condition~\eqref{eq:h-bnd-detect} cannot be used in practise as one needs to know the interpolation constant $C_\nabla$. This is not surprising as condition~\eqref{eq:h-bnd-detect} in the descriptions of $\mathscr{C}_1(h,\Omega)$ and $\mathscr{C}_2(h,\Omega)$ accounts for generic $\vec f$ and $\vec F$. The following result further refines condition~\eqref{eq:h-bnd-detect} by imposing a certain structure on $\vec F$: namely, it is demonstrated that the classical Luenberger-type output-feedback control $\vec F =\vec g+L\mathcal C(\vec u-\vec z)$ guarantees condition $A)$ of Def.~\ref{def:detectability}, and $\vec e=\vec u-\vec z\to0$, provided $\vec f-\vec g$ ``dissapeares'' at infinity and $h$ verifies an easy-to-check condition.
\begin{lemma}\label{l:observer-existence}
Let $\vec u$ solve~\eqref{eq:NSEu} on $\Omega=(-\frac{\ell_1}2, \frac{\ell_1}2)\times (-\frac{\ell_2}2, \frac{\ell_2}2)$ and let $\mathcal C\in\mathscr{C}_1(h,\Omega)$ or $\mathcal C\in\mathscr{C}_2(h,\Omega)$. For any $\vec g\in\mathring{\mathcal{H}}$ the following equation ($\vec e=\vec u-\vec z$, $\vec \phi\in\mathring{\mathcal{V}}$):
\begin{align}
\displaystyle\frac{d}{dt}(\vec z,\vec \phi) &+ b(\vec z,\vec z,\vec\phi) + \nu((\vec z,\vec \phi)) = (\vec g+L\mathcal C\vec e,\vec \phi)\,, \label{eq:NSE-LO}
\end{align}
has the unique solution $\vec z\in L^2(0,T,\mathscr{D}(A))\cap C(0,T,\mathring{\mathcal{V}})$.
\end{lemma}
\begin{theorem}\label{t:1}
Let $\vec u$ solve~\eqref{eq:NSEu} and $\vec z$ solve~\eqref{eq:NSE-LO} provided
\begin{equation}
\label{eq:fgdecay}
\limsup_{t\to\infty}\frac1T\int_t^{t+T}\|\vec f(s)-\vec g(s)\|^2_{L^2} ds=0\,.
\end{equation}
For $\kappa>1$ pick\footnote{If $\|\vec f\|_{L^2},\|\nabla \vec u_0\|_{L^2}<\bar C$ for a given $\bar C$ then $t^\star$ and $T$ can be computed explicitly for every $\kappa$.} $t^\star>0$ and $T>0$ such that $\theta_{t^\star,T}\le \kappa\|\vec f\|^2_{L^2}\nu^{-2}$ where \[
\theta_{t,T} = \frac{2\|\vec f\|^2_{L^2}}{T\nu^3\lambda_1} + \frac{2\mathbb{e}^{(-\lambda_1 \nu)t}}{T\nu}\|\nabla\vec u_0\|^2_2 +\frac{\|\vec f\|^2_{L^2}}{\nu^2}
\] and take $\Gamma_{\mathrm{max}}\in \operatorname{Argmax}\Theta(\Gamma)$ where
\begin{align*}
\Theta(\Gamma)&=\frac{(2\pi)^{\frac32}\nu(\nu -\frac{\|\vec \ell\|_{\mathbb{R}^2}}{\sqrt{32\pi^3}\lambda_1\Gamma}) \log^{-\frac12}\bigl(1+\frac{4\pi^2 \kappa\|\vec f\|^2_{L^2}\Gamma^2 }{\nu^2\ll_1\ll_2}\bigr)}{C_\Omega (4\kappa\pi^2+\kappa\max\{\ell_1,\ell_2\}^2)^{\frac12} \|\vec f\|_{L^2}}
\end{align*}
Then $V(t)=\|\vec u(t)-\vec z(t)\|^2_{H^1}\to 0$ for $t>t^\star$
\begin{itemize}
\item if $\mathcal C\in\mathscr{C}_1(h,\Omega)$, and $h^2<\beta \Theta(\Gamma_{\mathrm{max}})$, $L=\beta \hat L_\nabla$ for $0<\beta<1$ where
\begin{align}
\label{eq:hatL}
\hat L_\nabla \coloneqq\frac{2(\nu -\frac{\|\vec \ell\|_{\mathbb{R}^2}}{\sqrt{32\pi^3}\lambda_1\Gamma_{\mathrm{max}}})}{h^{2} C_\Omega}, \quad \vec \ell=
\bigl[\begin{smallmatrix}
\ell_1\\\ell_2
\end{smallmatrix}\bigr],
\end{align}
\item or if $\mathcal C\in\mathscr{C}_2(h,\Omega)$ and $h< \beta C_\Omega^{\frac12}\theta^{-1}\Theta(\Gamma_{\mathrm{max}})$, $L=\theta\hat L_\Delta$ for $\theta>1$, where
\begin{align}\label{eq:hatL1}
\hat L_\Delta\coloneqq \frac{(\kappa+\kappa\lambda_1^{-1})^\frac12 \|\vec f\|_{L^2}}{\nu\sqrt{2\pi}} \log^{\frac12}\bigl(1+\frac{4\pi^2 \kappa \|\vec f\|_{L^2}^2 \Gamma_{\mathrm{max}}^2 }{\nu^2\ll_1\ll_2}\bigr),
\end{align}
\end{itemize}
If in addition $\|\vec f(s)-\vec g(s)\|_{\mathbb{R}^2}=0$ for $s>s^\star$ then $V\to0$ \emph{exponentially fast} for $t>\max\{t^\star,s^\star\}$ and $\beta=1$.
\end{theorem}
\begin{remark}\label{sec:observer-design-1}
Let us compare our upper bound for $h^2$ obtained above for $\mathcal C\in\mathscr{C}_1(h,\Omega)$, namely $h^2<\beta\Theta(\Gamma_{\mathrm{max}})$ to the state-of-the-art result of~\cite[Prop.2]{azouani2014continuous}:
\begin{equation}
\label{eq:Titih}
h^2\le \nu \mathcal{T}(\nu)=C_\Omega^{-1} \bigl(3\nu\lambda_1 (2c\log(2) c^{\frac32}+8c\log(1+G))G\bigr)^{-1},
\end{equation}
where $G=\frac{\|\vec f\|_{L^2}}{\lambda_1\nu^2}$ and the constant $c$ comes from Brezis inequality $\|\vec u\|_{L^\infty}\le c \|\nabla\vec u\|_{L^2}(1+\log\frac{\|A\vec u\|_{L^2}^2}{\lambda_1\|\nabla\vec u\|_{L^2}^2})^{\frac12}$ for $\vec u\in\mathscr{D}(A)$~\cite[p.100, (A.50)]{NSE-Turbulence-book2001}. Clearly, \[
c=\sup_{\vec u\in\mathscr{D}(A):\|\nabla\vec u\|_{L^2}>0}\|\vec u\|_{L^\infty} \|\nabla\vec u\|^{-1}_{L^2}(1+\log\frac{\|A\vec u\|_{L^2}^2}{\lambda_1\|\nabla\vec u\|_{L^2}^2})^{-\frac12}
\] and so $c\ge c_1$ for $c_1=\|\vec u_1\|_{L^\infty} \|\nabla\vec u_1\|^{-1}_{L^2}(1+\log\frac{\|A\vec u_1\|_{L^2}^2}{\lambda_1\|\nabla\vec u_1\|_{L^2}^2})^{-\frac12}$ and $\vec u_1\in\mathscr{D}(A)$ with components $u_1=-\cos(2\pi x)\sin(2\pi y)$ and $u_2=\cos(2\pi y)\sin(2\pi x)$. It is easy to compute that $c_1=(2\pi)^{-1}$. Since the r.h.s. of~\eqref{eq:Titih}, $\mathcal{T}(\nu)$ increases if we substitute $c$ with $c_1$, and so the upper bound on $h^2$ improves, below we compare $\Theta(\Gamma_{\mathrm{max}})$ vs. $\mathcal{T}(\nu)$ with $c=c_1=(2\pi)^{-1}$ over the interval $\nu\in[10^{-6},10^{-1}]$. To match the setting of~\cite{azouani2014continuous} we assume that $\|\vec f(s)-\vec g(s)\|_{\mathbb{R}^2}=0$ for $s>s^\star$ so we can use $\beta=1$. We get that $\log_{10}\frac{\Theta(\Gamma_{\mathrm{max}})}{\mathcal{T}(\nu)}=1.33$ for $\nu=10^{-6}$ and $0.58$ for $\nu=10^{-1}$ and LogLog-plot of $\Theta$ and $\mathcal{T}$ over $\nu\in[10^{-6},10^{-1}]$ is given in Fig.~\ref{fig:comp}. Obviously, for small $\nu$ our upper bound estimate is at least one order of magnitude better.
\begin{figure}[h]
\begin{center}
\includegraphics[width=\columnwidth]{figures/titi-comp}
\caption{LogLog-plot of upper-bounds for $h^2$: $\Theta(\Gamma_{\mathrm{max}})$ (blue) vs. $\mathcal{T}(\nu)$ (orange) with $c=c_1=(2\pi)^{-1}$ over the interval $\nu\in[10^{-6},10^{-1}]$.}
\label{fig:comp}
\end{center}
\end{figure}
\end{remark}
\subsubsection{Bounds for $L^\infty$-norms of periodic vector-functions}
\label{sec:upper-bounds}
The following inequality
is a key building block used below to define detectability and design observer for NSE.
\begin{lemma}\label{sec:upper-bounds-linfty}
Take $\vec u\in H_p^2(\Omega)^2$ with zero mean\footnote{This condition is necessary: lemma does not hold for $u\equiv \text{const}$ and small enough $\ell_{1,2}$}, i.e. $\int_\Omega \vec u dx_1dx_2=0$. Then for any $\gamma>0$
\begin{equation}
\label{eq:AgmoN}
\begin{split}
\|\vec u\|_{L^\infty}&\le \frac{\log^{\frac12}\bigl(1+\frac{4\pi^2 \gamma^2}{\ll_1\ll_2}\bigr)\|\vec u\|_{H^1}}{\sqrt{2\pi}} + \frac{\|\vec \ell\|_{\mathbb{R}^2} \|\Delta \vec u\|_{L^2}}{\gamma\sqrt{32\pi^3}}
\end{split}
\end{equation}
\end{lemma}
The proof is given in the appendix. Note that by using specific $\gamma$ in~\eqref{eq:AgmoN} one can derive some classical inequalities:
\begin{cor}
Set $\gamma= \|\vec u\|_{H^1}^{-\frac12}\|\Delta \vec u\|_{L^2}^{\frac12}$. We get 2D Agmon-type inequality~\cite[p.100]{NSE-Turbulence-book2001}:
\begin{equation}
\label{eq:Agmon}
\begin{split}
\|\vec u\|_{L^\infty}&\le (\sqrt{\frac{2\pi}{\ll_1\ll_2}}+\frac{\|\vec \ell\|_{\mathbb{R}^2}}{\sqrt{32\pi^3}})\|\vec u\|_{H^1}^{\frac12}\|\Delta \vec u\|_{L^2}^{\frac12}
\end{split}
\end{equation}
Set $\gamma= \|\vec u\|_{H^1}^{-1}\|\Delta \vec u\|_{L^2}$. We get 2D Brezis-type inequality~\cite{brezis1979nonlinear}:
\begin{equation}
\label{eq:Brezis}
\|\vec u\|_{L^\infty}\le \bigl(\frac{\|\vec \ell\|_{\mathbb{R}^2}}{\sqrt{32\pi^3}}+\frac{\log^{\frac12}(1+\frac{4\pi^2 \|\Delta \vec u\|_{L^2}^2}{\ll_1\ll_2 \|\vec u\|_{H^1}^2})}{\sqrt{2\pi}}\bigr)\|\vec u\|_{H^1}
\end{equation}
\end{cor}
Yet another building block used in the proofs below is the following version of Bellman lemma~\cite[Lemma 1.1, p.125]{NSE-Turbulence-book2001}:
\begin{lemma}\label{l:GB}
If differentiable $V:[0,\infty)\to [0,\infty)$ satisfies $\dot V(t) + \alpha(t) V(t) \le \beta(t)$ for some measurable scalar functions $\alpha,\beta$ then $V\to 0$ for $t\to\infty$ provided there exist $T>0$ such that
\begin{itemize}
\item $\limsup_{t\to\infty}\frac1T\int_t^{t+T} \max(\beta(s),0) ds =0$ and
\item $\liminf_{t\to\infty}{\frac1T\int_t^{t+T} \alpha(s) ds} >0$ and
\item $\limsup_{t\to\infty}\frac1T\int_t^{t+T} \max(0,-\alpha(s)) ds <+\infty$.
\end{itemize}
\end{lemma}
\subsubsection{Navier-Stokes equation: classical and weak formulations, and $H^1$-well-posedness}
\label{sec:navi-stok-equat}
The classical NSE is a system of two PDEs defining the dynamics of the scalar pressure field $p(x,y)$ and the viscous fluid velocity vector-field $\vec u=(u_1(x,y),u_2(x,y))^\top$:
\begin{align}
\label{eq:NSE-system}
\displaystyle\frac{du_1}{dt} &+ \vec u \cdot \nabla u_1 - \nu\Delta u_1 + p_{x} = f_1\\
\displaystyle\frac{du_2}{dt} &+ \vec u \cdot \nabla u_2 - \nu\Delta u_2 + p_{y} = f_2\\
&\nabla\cdot \vec u = (u_1)_x + (u_2)_y = 0
\end{align}
with initial condition $\vec u(0)=\vec u_0\in\mathcal{H}$, disturbance $\vec f=(f_1,f_2)^\top$ and periodic Boundary Conditions (BC) $u_{1,2}(x+\ell_1,y) = u_{1,2}(x,y)$, $u_{1,2}(x,y+\ell_2) = u_{1,2}(x,y)$ or, in the vector form:
\begin{equation}
\label{eq:NSE-vector}
\begin{split}
\displaystyle\frac{d\vec u}{dt} &+ (\vec u \cdot \nabla) \vec u - \nu\Delta\vec u+\nabla p = \vec f\\
&\nabla\cdot\vec u = 0
\end{split}
\end{equation}
Note that the div-free condition $\nabla\cdot\vec u = 0$ guarantees that the flow defined by $\vec u$ is volume preserving. It is well-known (e.g.~\cite[p.36]{NSE-Turbulence-book2001}) that every vector-field $\vec u$ in $\mathbb{R}^2$ admits Helmholtz-Leray decomposition: $\vec u = \nabla p + \vec v$, and $\nabla\cdot\vec v = 0$. We define Leray projection operator $P_l(\vec u)=\vec v$. $P_l$ is an orthogonal projector onto $\mathring{\mathcal{H}}$. $P_l\vec u$ can be computed uniquely: $P_l=\vec u-\nabla p$ provided $p\in H^1_p(\Omega)$ solves the Poisson equation $\Delta p = \nabla\cdot\vec u$ with periodic BC and $\int_\Omega p dxdy = 0$.
Define Stokes operator $A$ as follows (see~\cite[p.52]{NSE-Turbulence-book2001}):
\begin{equation}\begin{split}\nonumber
A\vec u&=-P_l\Delta \vec u,\\
u\in \mathscr{D}(A)&=\{\vec u\in \mathring{\mathcal{H}}:\Delta u\in \mathcal{H}\} = \mathring{\mathcal{V}}\cap ({H}_p^2(\Omega))^2\,.
\end{split}\end{equation}
Note that in fact $A\vec u=-\Delta \vec u$ for the case of periodic BC. $A$ is a self-adjoint positive operator with compact inverse, and (see~\cite[p.52,f.(6.19)]{NSE-Turbulence-book2001})
\begin{align}
(A\vec u,\vec \phi) &= ((\vec u,\vec\phi))\ge \lambda_1(\vec u,\vec{\phi})\,, \vec u\in \mathscr{D}(A),\vec{\phi}\in \mathring{\mathcal{V}},\label{eq:Poincare}\\
(A\vec u, A\vec u) &=(A(A^\frac12)\vec u,(A^\frac12)\vec u)\ge \lambda_1 (A \vec u, \vec u)\label{eq:gradA-bnd}\\
\lambda_1&=\frac{4\pi^2}{\max\{\ell_1,\ell_2\}^2}
\end{align}
In what follows we consider the variational form of~\eqref{eq:NSE-vector} which reduces~\eqref{eq:NSE-vector} to the subspace of div-free vector-fields with zero average, $\mathring{\mathcal{V}}$ (i.e. applies Leray projector $P_l$) and thus eliminates $p$. To this end we define:
\begin{align*}
(w,\phi)&=\int_\Omega w(x)\phi(x) dx, (\vec u,\vec\phi) = (u_1,\phi_1)+(u_2,\phi_2),\\%\quad w,\phi\in L_p^2(\Omega)\\
b(\vec u,\vec w,\vec\phi) &= (\vec u\cdot \nabla w_1,\phi_1) + (\vec u\cdot \nabla w_2,\phi_2),\\
((\vec u,\vec \phi)) &= (\nabla u_1,\nabla \phi_1)+(\nabla u_2,\nabla \phi_2)
\end{align*}
The form $b$ has a number of important properties which will be used below:
\begin{align*}
\text{skew-symmetry:}\quad &b(\vec u,\vec v,\vec \phi) = - b(\vec u,\vec \phi,\vec v)\\
\text{orthogonality:} \quad&b(\vec u,\vec v,\vec v) = 0,\quad b(\vec v,\vec v,A\vec v) = 0
\end{align*}
Upon multiplying~\eqref{eq:NSE-vector} by a test function $\vec\phi\in\mathring{\mathcal{V}}$ and integrating in $\Omega$ we get the following weak form of NSE~\eqref{eq:NSE-system}:
\begin{equation}
\label{eq:NSE-var}
\displaystyle\frac{d}{dt}(\vec u,\vec \phi) + b(\vec u,\vec u,\vec\phi) + \nu((\vec u,\vec \phi)) = (\vec f,\vec \phi), \quad\forall\vec\phi\in \mathring{\mathcal{V}}
\end{equation}
with initial condition $(\vec u(0),\vec\phi)=(\vec u_0,\vec\phi)$. By utilising~Theorem 7.4 from~\cite[p.58, Th.7.4]{NSE-Turbulence-book2001} and Cauchy-Schwarz inequality combined with Lemma~\ref{l:GB} it is not hard to obtain that $H^1$-norm of the weak (and strong) solution of NSE is bounded, and decays to $0$ exponentially provided $f=0$:
\begin{lemma}
Let $\vec u_0\in \mathring{\mathcal{H}}$ and $\vec f\in L^2(0,T,\mathring{\mathcal{H}})$. Then, on $[0,T]$ there exist the unique weak solution $\vec u\in C(0,T,\mathring{\mathcal{H}})$ of NSE in variational form~\eqref{eq:NSE-var}, and the components of $\vec u=(u_1,u_2)^\top$ verify: $u_i,(u_i)_{x,y}\in L^2(\Omega\times(0,T))$. If $\vec u_0\in\mathring{\mathcal{V}}$ then the weak solution coincides with the strong (classical) solution of~\eqref{eq:NSE-system}, and \[
\displaystyle\frac{du_i}{dt}, (u_i)_{x},(u_i)_{y}, (u_i)_{x y}\in L^2(\Omega\times(0,T))
\] i.e. $\vec u\in C(0,T,\mathring{\mathcal{V}}) \cap L^2(0,T,\mathscr{D}(A))$. If in addition $\vec f\in L^2(0,+\infty,\mathring{\mathcal{H}})$ then
\begin{align}
&\|\vec u(\cdot,t)\|^2_{L^2} \le \frac{\|\vec f\|^2_{L^2}}{(\nu\lambda_1)^2} + e^{(-\lambda_1 \nu)(t-s)}\|\vec u(\cdot,s)\|^2_{L^2} \label{eq:L2norm-decay}\\
&\|\nabla \vec u(\cdot,t)\|^2_{L^2} \le \frac{\|\vec f\|^2_{L^2}}{\nu^2\lambda_1} + e^{(-\lambda_1 \nu)(t-s)}\|\nabla \vec u(\cdot,s)\|^2_{L^2} \label{eq:grad-norm-decay}\\
&\frac1T\int_t^{T+t}\|A\vec u\|^2_{L^2}ds \le \theta_{t,T}\label{eq:Au-L2avrg-bnd}\\
&\theta_{t,T} = 2\frac{\|\vec f\|^2_{L^2}}{T\nu^3\lambda_1} + \frac{\|\vec f\|^2_{L^2}}{\nu^2} + \frac{2}{T\nu}e^{(-\lambda_1 \nu)t}\|\nabla\vec u_0\|^2_{L^2}\label{eq:Au-ThetatT}
\end{align}
\end{lemma}
\subsection{Sufficient conditions for detectability}
\label{sec:detectability-nse}
\begin{definition}\label{def:detectability-class}
Take $\mathcal C\in\mathscr{L}(L^2)$ such that $\mathcal{V}\subseteq\mathscr{D}(\mathcal C)$. $\mathcal C$ is said to be of class $\mathscr{C}_1(h,\Omega)$ or of class $\mathscr{C}_2(h,\Omega)$ if~\eqref{eq:CPoin} or ~\eqref{eq:CPoin2} respectively holds for some $h>0$ and $C_\Omega>0$
\begin{align}
\forall\vec u\in\mathring{\mathcal{V}}:\quad \|\vec u - \mathcal C\vec u\|_{L^2}^2&\le h^2 C_\Omega \|\nabla\vec u\|_{L^2}^2 \label{eq:CPoin}\\
\forall\vec u\in \mathscr{D}(A):\quad \|\vec u - \mathcal C\vec u\|_{L^2}^2&\le h^2 C_\Omega \|\Delta\vec u\|_{L^2}^2 \label{eq:CPoin2}
\end{align}
\end{definition}
\begin{prop}\label{prop:detectability-nse-1}
Recall Def.~\ref{def:detectability}: let $\vec z,\vec u$ solve~\eqref{eq:NSEu}-\eqref{eq:NSEz}. Recall from~\cite[p.]{NSE-Turbulence-book2001} the interpolation inequality and define constants $C_{1,2}$:
\begin{align}
&\forall\vec u\in\mathscr{D}(A):\|\nabla\vec u\|^2_{L^2}\le C_{\nabla} \|\vec u\|_{L^2} \|A\vec u\|_{L^2} \label{eq:InterpIneq}\\
&C_1^2=(1+\lambda_1^{-1})/(2\pi), \quad C_2^2 = \|\vec \ell\|^2_{\mathbb{R}^2}/(32\pi^3)\label{eq:C1C2}
\end{align}
Take $\mathcal C\in\mathscr{C}_1(h,\Omega)$. Then NSE is detectable in $\mathring{\mathcal{V}}$ if for some $1<\kappa<<2$
\begin{equation}
\label{eq:h-bnd-detect}
h< \frac{\lambda_1^{\frac12}\nu\bigl(\frac{3\nu}{4} -\frac{C_2}{\lambda_1\Gamma}\bigr)}{C_1 C_\nabla C^{\frac12}_\Omega \kappa^{\frac12}\|\vec f\|_{L^2}\log^{\frac12}\bigl(1+\frac{4\pi^2 \kappa \|\vec f\|_{L^2}^2\Gamma^2 }{\nu^2\ll_1\ll_2}\bigr)}
\end{equation}
moreover, if $\mathcal C\in\mathscr{C}_2(h,\Omega)$ then NSE is detectable in $\mathring{\mathcal{V}}$ if $h$ verifies~\eqref{eq:h-bnd-detect} without $\lambda_1^{\frac12}$ in the numerator.
\end{prop}
\begin{example}\label{exmpl:detectability-nse-1}
The class of operators $\mathscr{C}_1(h,\Omega)$ generalizes the notion of so-called \textit{determining modes} for NSE which is well-known in NSE's literature (see e.g.~\cite[p.123]{NSE-Turbulence-book2001}). The case of determining modes corresponds to $\mathcal C\in\mathscr{C}_1(h,\Omega)$ of finite rank, $\mathcal C\vec u$ represents the projection of $\vec u$ onto a certain finite-dimensional subspace of $\mathcal{H}$, and in this case $\mathcal C$ is bounded in $\mathcal{H}$. As an example, consider the case of spatial averages over a partition of $\Omega$ (e.g.~\cite{emilia2,azouani2014continuous}): Let $\Omega=\cup_{j=1}^N \Omega_j$, $\Omega_j\cap\Omega_i=\varnothing$ for $i\ne j$, $\Omega_j$ -- a rectangle with sides of length $h_x^j$ and $h_y^j$ respectively, and let $\xi_j$ denote the indicator function of $\Omega_j$. An example of a uniform square partition of $\Omega$ is provided in the left panel of Fig.~\ref{fig:1}. Let \[
\mathcal C\vec u(x,y) = \sum_{j=1}^N\bigl[
\begin{smallmatrix}
(u_1,\xi_j)\\(u_2,\xi_j)
\end{smallmatrix}\bigr]\frac{\xi_j(x,y)}{(h_x^j h_y^j)^{\frac12}}, \quad \xi_j(x,y)=1, (x,y)\in\Omega_j
\] Then~\eqref{eq:CPoin} holds for $h^2=\max_j(\max\{h_x^j,h_y^j\})^2$ and $C_\Omega=(4\pi^2)^{-1}$ by Poincare inequality~\eqref{eq:Poincare}.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.49\columnwidth]{figures/Obs}
\includegraphics[width=0.49\columnwidth]{figures/State}
\caption{Example of the 1st component of $\mathcal C\vec u$ (left panel), versus $1st$ component of $\vec u$}
\label{fig:1}
\end{center}
\end{figure}
\end{example}
\begin{remark}
Fig.~\ref{fig:1} provides a good illustration of the detectability concept: the left panel represents 1st component of $\mathcal C\vec u$, the projection of the $1st$ component of $\vec u$ (right panel) onto the orthonormal set of $N$ indicator functions $\xi_j$, and $\vec u$ solves NSE~\eqref{eq:NSEu}. Comparing panels of Fig.~\ref{fig:1} it is evident that there is little or no hope to reconstruct $\vec u$ from $\mathcal C\vec u$, e.g. by interpolation, without using the structure of NSE~\eqref{eq:NSEu}. In this situation one can reconstruct the continuous state from observations, if NSE is detectable, i.e. if $h^2=\max_j(\max\{h_x^j,h_y^j\})^2$ is small enough (e.g.~\eqref{eq:h-bnd-detect}) for the diffusion term $\nu((\vec u,\vec u))$ to dominate the non-linear term $b$ in the ``unmeasured'' infinite-dimensional orthogonal complement of the span of $\xi_j$. In other words, the span of $\xi_j$ (and more generally the measured output $\mathcal C\vec u$) need to ``reveal'' a certain part of the state space where the nonlinear term $b$ is not ``dominated'' by the diffusion term $\nu((\vec u,\vec u))$.
\end{remark}
\begin{example}\label{exmpl:detectability-nse-1}
The class of operators $\mathscr{C}_2(h,\Omega)$ generalizes notion of so-called \textit{determining nodes} for NSE (see e.g.~\cite[p.131]{NSE-Turbulence-book2001}). The case of determining nodes corresponds to $\mathcal C\in\mathscr{C}_2(h,\Omega)$ of the form $\mathcal C\vec u=\sum_{j=1}^n \vec u(t,\vec x_j)\vec{\phi}_j$ for some grid nodes $\vec x_j\in\Omega$ and functions $\vec{\phi}_j\in H^1_p(\Omega)^2$. Clearly, $\mathcal C$ is not bounded in $\mathcal{H}$ as it is defined for $\vec u\in H^2_p(\Omega)^2\cap\mathcal{H}$, but it is a closed linear operator such that $\mathscr{D}(A)\subset\mathscr{D}(C)$. As an example, consider the case of pointwise evaluations over nodes $\vec x_j$ a finite element grid, here $\vec{\phi}_j$ are 2D hat basis functions. In this case $\mathcal C$ verifies~\eqref{eq:CPoin2} indeed with constants $h$ and $C_\Omega$ which can be found in either~\cite[Th 12.3.4]{Hutson1980}.
\end{example}
\subsection{Observer design}
\label{sec:observer-design}
Proposition~\ref{prop:detectability-nse-1} describes classes $\mathscr{C}_1(h,\Omega)$ and $\mathscr{C}_2(h,\Omega)$ of linear operators $\mathcal C$ which make NSE detectable, provided~\eqref{eq:h-bnd-detect} holds. However, as is, the condition~\eqref{eq:h-bnd-detect} cannot be used in practise as one needs to know the interpolation constant $C_\nabla$. This is not surprising as condition~\eqref{eq:h-bnd-detect} in the descriptions of $\mathscr{C}_1(h,\Omega)$ and $\mathscr{C}_2(h,\Omega)$ accounts for generic $\vec f$ and $\vec F$. The following result further refines condition~\eqref{eq:h-bnd-detect} by imposing a certain structure on $\vec F$: namely, it is demonstrated that the classical Luenberger-type output-feedback control $\vec F =\vec g+L\mathcal C(\vec u-\vec z)$ guarantees condition $A)$ of Def.~\ref{def:detectability}, and $\vec e=\vec u-\vec z\to0$, provided $\vec f-\vec g$ ``dissapeares'' at infinity and $h$ verifies an easy-to-check condition.
\begin{lemma}\label{l:observer-existence}
Let $\vec u$ solve~\eqref{eq:NSEu} on $\Omega=(-\frac{\ell_1}2, \frac{\ell_1}2)\times (-\frac{\ell_2}2, \frac{\ell_2}2)$ and let $\mathcal C\in\mathscr{C}_1(h,\Omega)$ or $\mathcal C\in\mathscr{C}_2(h,\Omega)$. For any $\vec g\in\mathring{\mathcal{H}}$ the following equation ($\vec e=\vec u-\vec z$, $\vec \phi\in\mathring{\mathcal{V}}$):
\begin{align}
\displaystyle\frac{d}{dt}(\vec z,\vec \phi) &+ b(\vec z,\vec z,\vec\phi) + \nu((\vec z,\vec \phi)) = (\vec g+L\mathcal C\vec e,\vec \phi)\,, \label{eq:NSE-LO}
\end{align}
has the unique solution $\vec z\in L^2(0,T,\mathscr{D}(A))\cap C(0,T,\mathring{\mathcal{V}})$.
\end{lemma}
\begin{theorem}\label{t:1}
Let $\vec u$ solve~\eqref{eq:NSEu} and $\vec z$ solve~\eqref{eq:NSE-LO} provided
\begin{equation}
\label{eq:fgdecay}
\limsup_{t\to\infty}\frac1T\int_t^{t+T}\|\vec f(s)-\vec g(s)\|^2_{L^2} ds=0\,.
\end{equation}
For $\kappa>1$ pick $t^\star>0$ and $T>0$ such that $\theta_{t^\star,T}\le \kappa\|\vec f\|^2_{L^2}\nu^{-2}$ where \[
\theta_{t,T} = \frac{2\|\vec f\|^2_{L^2}}{T\nu^3\lambda_1} + \frac{2\mathbb{e}^{(-\lambda_1 \nu)t}}{T\nu}\|\nabla\vec u_0\|^2_2 +\frac{\|\vec f\|^2_{L^2}}{\nu^2}
\] and take $\Gamma_{\mathrm{max}}\in \operatorname{Argmax}\Theta(\Gamma)$ where
\begin{align*}
\Theta(\Gamma)&=\frac{(2\pi)^{\frac32}\nu(\nu -\frac{\|\vec \ell\|_{\mathbb{R}^2}}{\sqrt{32\pi^3}\lambda_1\Gamma}) \log^{-\frac12}\bigl(1+\frac{4\pi^2 \kappa\|\vec f\|^2_{L^2}\Gamma^2 }{\nu^2\ll_1\ll_2}\bigr)}{C_\Omega (4\kappa\pi^2+\kappa\max\{\ell_1,\ell_2\}^2)^{\frac12} \|\vec f\|_{L^2}}
\end{align*}
Then $V(t)=\|\vec u(t)-\vec z(t)\|^2_{H^1}\to 0$ for $t>t^\star$ if
\begin{itemize}
\item $\mathcal C\in\mathscr{C}_1(h,\Omega)$, and $h^2<\beta \Theta(\Gamma_{\mathrm{max}})$, $L=\beta \hat L_\nabla$ for $0<\beta<1$ where
\begin{align}
\label{eq:hatL}
\hat L_\nabla \coloneqq\frac{2(\nu -\frac{\|\vec \ell\|_{\mathbb{R}^2}}{\sqrt{32\pi^3}\lambda_1\Gamma_{\mathrm{max}}})}{h^{2} C_\Omega}, \quad \vec \ell=
\bigl[\begin{smallmatrix}
\ell_1\\\ell_2
\end{smallmatrix}\bigr],
\end{align}
\item $\mathcal C\in\mathscr{C}_2(h,\Omega)$ and $h< \beta C_\Omega^{\frac12}\theta^{-1}\Theta(\Gamma_{\mathrm{max}})$, $L=\theta\hat L_\Delta$ for $\theta>1$, where
\begin{align}\label{eq:hatL1}
\hat L_\Delta\coloneqq \frac{(\kappa+\kappa\lambda_1^{-1})^\frac12 \|\vec f\|_{L^2}}{\nu\sqrt{2\pi}} \log^{\frac12}\bigl(1+\frac{4\pi^2 \kappa \|\vec f\|_{L^2}^2 \Gamma_{\mathrm{max}}^2 }{\nu^2\ll_1\ll_2}\bigr),
\end{align}
\end{itemize}
\end{theorem}
\begin{remark}
If $\|\vec f(s)-\vec g(s)\|_{\mathbb{R}^2}=0$ for $s>s^\star$ then $V\to0$ \emph{exponentially fast} for $\beta=1$ and $t>\max\{t^\star,s^\star\}$. If $\|\vec f\|_{L^2},\|\nabla \vec u_0\|_{L^2}<\bar C$ for a given $\bar C$ then $t^\star$ and $T$ can be computed explicitly for every $\kappa$.
\end{remark}
\section{Introduction}
\label{sec:introduction}
\input{introduction.tex}
\subsection{Mathematical preliminaries}
\label{sec:math-prel-1}
\input{preliminaries.tex}
\section{Problem statement}
\label{sec:problem-statement}
\input{prob_statement}
\section{Main results}
\label{sec:main-results}
\input{detectability.tex}
\input{observer.tex}
\section{Experiments}
\label{sec:experiments}
\input{experiments.tex}
\section{Conclusions}
\label{sec:conclusions}
In this work we obtained proposed {\it easy-to-verify} detectability conditions for Navier-Stokes Equation (NSE) in two spatial dimensions and unknown inputs, and designed a globally converging observer of Luenberger type for the case of continuous in time measurements. A promising research direction now is to consider the case of sampled measurements.
\bibliographystyle{plain}
\subsection{Observer design}
\label{sec:observer-design}
Proposition~\ref{prop:detectability-nse-1} describes classes $\mathscr{C}_1(h,\Omega)$ and $\mathscr{C}_2(h,\Omega)$ of linear operators $\mathcal C$ which make NSE detectable, provided~\eqref{eq:h-bnd-detect} holds. However, as is, the condition~\eqref{eq:h-bnd-detect} cannot be used in practise as one needs to know the interpolation constant $C_\nabla$. This is not surprising as condition~\eqref{eq:h-bnd-detect} in the descriptions of $\mathscr{C}_1(h,\Omega)$ and $\mathscr{C}_2(h,\Omega)$ accounts for generic $\vec f$ and $\vec F$. The following result further refines condition~\eqref{eq:h-bnd-detect} by imposing a certain structure on $\vec F$: namely, it is demonstrated that the classical Luenberger-type output-feedback control $\vec F =\vec g+L\mathcal C(\vec u-\vec z)$ guarantees condition $A)$ of Def.~\ref{def:detectability}, and $\vec e=\vec u-\vec z\to0$, provided $\vec f-\vec g$ ``dissapeares'' at infinity and $h$ verifies an easy-to-check condition.
\begin{lemma}\label{l:observer-existence}
Let $\vec u$ solve~\eqref{eq:NSEu} on $\Omega=(-\frac{\ell_1}2, \frac{\ell_1}2)\times (-\frac{\ell_2}2, \frac{\ell_2}2)$ and let $\mathcal C\in\mathscr{C}_1(h,\Omega)$ or $\mathcal C\in\mathscr{C}_2(h,\Omega)$. For any $\vec g\in\mathring{\mathcal{H}}$ the following equation ($\vec e=\vec u-\vec z$, $\vec \phi\in\mathring{\mathcal{V}}$):
\begin{align}
\displaystyle\frac{d}{dt}(\vec z,\vec \phi) &+ b(\vec z,\vec z,\vec\phi) + \nu((\vec z,\vec \phi)) = (\vec g+L\mathcal C\vec e,\vec \phi)\,, \label{eq:NSE-LO}
\end{align}
has the unique solution $\vec z\in L^2(0,T,\mathscr{D}(A))\cap C(0,T,\mathring{\mathcal{V}})$.
\end{lemma}
\begin{theorem}\label{t:1}
Let $\vec u$ solve~\eqref{eq:NSEu} and $\vec z$ solve~\eqref{eq:NSE-LO} provided
\begin{equation}
\label{eq:fgdecay}
\limsup_{t\to\infty}\frac1T\int_t^{t+T}\|\vec f(s)-\vec g(s)\|^2_{L^2} ds=0\,.
\end{equation}
For $\kappa>1$ pick\footnote{If $\|\vec f\|_{L^2},\|\nabla \vec u_0\|_{L^2}<\bar C$ for a given $\bar C$ then $t^\star$ and $T$ can be computed explicitly for every $\kappa$.} $t^\star>0$ and $T>0$ such that $\theta_{t^\star,T}\le \kappa\|\vec f\|^2_{L^2}\nu^{-2}$ where \[
\theta_{t,T} = \frac{2\|\vec f\|^2_{L^2}}{T\nu^3\lambda_1} + \frac{2\mathbb{e}^{(-\lambda_1 \nu)t}}{T\nu}\|\nabla\vec u_0\|^2_2 +\frac{\|\vec f\|^2_{L^2}}{\nu^2}
\] and take $\Gamma_{\mathrm{max}}\in \operatorname{Argmax}\Theta(\Gamma)$ where
\begin{align*}
\Theta(\Gamma)&=\frac{(2\pi)^{\frac32}\nu(\nu -\frac{\|\vec \ell\|_{\mathbb{R}^2}}{\sqrt{32\pi^3}\lambda_1\Gamma}) \log^{-\frac12}\bigl(1+\frac{4\pi^2 \kappa\|\vec f\|^2_{L^2}\Gamma^2 }{\nu^2\ll_1\ll_2}\bigr)}{C_\Omega (4\kappa\pi^2+\kappa\max\{\ell_1,\ell_2\}^2)^{\frac12} \|\vec f\|_{L^2}}
\end{align*}
Then $V(t)=\|\vec u(t)-\vec z(t)\|^2_{H^1}\to 0$ for $t>t^\star$
\begin{itemize}
\item if $\mathcal C\in\mathscr{C}_1(h,\Omega)$, and $h^2<\beta \Theta(\Gamma_{\mathrm{max}})$, $L=\beta \hat L_\nabla$ for $0<\beta<1$ where
\begin{align}
\label{eq:hatL}
\hat L_\nabla \coloneqq\frac{2(\nu -\frac{\|\vec \ell\|_{\mathbb{R}^2}}{\sqrt{32\pi^3}\lambda_1\Gamma_{\mathrm{max}}})}{h^{2} C_\Omega}, \quad \vec \ell=
\bigl[\begin{smallmatrix}
\ell_1\\\ell_2
\end{smallmatrix}\bigr],
\end{align}
\item or if $\mathcal C\in\mathscr{C}_2(h,\Omega)$ and $h< \beta C_\Omega^{\frac12}\theta^{-1}\Theta(\Gamma_{\mathrm{max}})$, $L=\theta\hat L_\Delta$ for $\theta>1$, where
\begin{align}\label{eq:hatL1}
\hat L_\Delta\coloneqq \frac{(\kappa+\kappa\lambda_1^{-1})^\frac12 \|\vec f\|_{L^2}}{\nu\sqrt{2\pi}} \log^{\frac12}\bigl(1+\frac{4\pi^2 \kappa \|\vec f\|_{L^2}^2 \Gamma_{\mathrm{max}}^2 }{\nu^2\ll_1\ll_2}\bigr),
\end{align}
\end{itemize}
If in addition $\|\vec f(s)-\vec g(s)\|_{\mathbb{R}^2}=0$ for $s>s^\star$ then $V\to0$ \emph{exponentially fast} for $t>\max\{t^\star,s^\star\}$ and $\beta=1$.
\end{theorem}
\begin{remark}\label{sec:observer-design-1}
Let us compare our upper bound for $h^2$ obtained above for $\mathcal C\in\mathscr{C}_1(h,\Omega)$, namely $h^2<\beta\Theta(\Gamma_{\mathrm{max}})$ to the state-of-the-art result of~\cite[Prop.2]{azouani2014continuous}:
\begin{equation}
\label{eq:Titih}
h^2\le \nu \mathcal{T}(\nu)=C_\Omega^{-1} \bigl(3\nu\lambda_1 (2c\log(2) c^{\frac32}+8c\log(1+G))G\bigr)^{-1},
\end{equation}
where $G=\frac{\|\vec f\|_{L^2}}{\lambda_1\nu^2}$ and the constant $c$ comes from Brezis inequality $\|\vec u\|_{L^\infty}\le c \|\nabla\vec u\|_{L^2}(1+\log\frac{\|A\vec u\|_{L^2}^2}{\lambda_1\|\nabla\vec u\|_{L^2}^2})^{\frac12}$ for $\vec u\in\mathscr{D}(A)$~\cite[p.100, (A.50)]{NSE-Turbulence-book2001}. Clearly, \[
c=\sup_{\vec u\in\mathscr{D}(A):\|\nabla\vec u\|_{L^2}>0}\|\vec u\|_{L^\infty} \|\nabla\vec u\|^{-1}_{L^2}(1+\log\frac{\|A\vec u\|_{L^2}^2}{\lambda_1\|\nabla\vec u\|_{L^2}^2})^{-\frac12}
\] and so $c\ge c_1$ for $c_1=\|\vec u_1\|_{L^\infty} \|\nabla\vec u_1\|^{-1}_{L^2}(1+\log\frac{\|A\vec u_1\|_{L^2}^2}{\lambda_1\|\nabla\vec u_1\|_{L^2}^2})^{-\frac12}$ and $\vec u_1\in\mathscr{D}(A)$ with components $u_1=-\cos(2\pi x)\sin(2\pi y)$ and $u_2=\cos(2\pi y)\sin(2\pi x)$. It is easy to compute that $c_1=(2\pi)^{-1}$. Since the r.h.s. of~\eqref{eq:Titih}, $\mathcal{T}(\nu)$ increases if we substitute $c$ with $c_1$, and so the upper bound on $h^2$ improves, below we compare $\Theta(\Gamma_{\mathrm{max}})$ vs. $\mathcal{T}(\nu)$ with $c=c_1=(2\pi)^{-1}$ over the interval $\nu\in[10^{-6},10^{-1}]$. To match the setting of~\cite{azouani2014continuous} we assume that $\|\vec f(s)-\vec g(s)\|_{\mathbb{R}^2}=0$ for $s>s^\star$ so we can use $\beta=1$. We get that $\log_{10}\frac{\Theta(\Gamma_{\mathrm{max}})}{\mathcal{T}(\nu)}=1.33$ for $\nu=10^{-6}$ and $0.58$ for $\nu=10^{-1}$ and LogLog-plot of $\Theta$ and $\mathcal{T}$ over $\nu\in[10^{-6},10^{-1}]$ is given in Fig.~\ref{fig:comp}. Obviously, for small $\nu$ our upper bound estimate is at least one order of magnitude better.
\begin{figure}[h]
\begin{center}
\includegraphics[width=\columnwidth]{figures/titi-comp}
\caption{LogLog-plot of upper-bounds for $h^2$: $\Theta(\Gamma_{\mathrm{max}})$ (blue) vs. $\mathcal{T}(\nu)$ (orange) with $c=c_1=(2\pi)^{-1}$ over the interval $\nu\in[10^{-6},10^{-1}]$.}
\label{fig:comp}
\end{center}
\end{figure}
\end{remark}
\subsubsection{Bounds for $L^\infty$-norms of periodic vector-functions}
\label{sec:upper-bounds}
The following inequality
is a key building block used below to define detectability and design observer for NSE.
\begin{lemma}\label{sec:upper-bounds-linfty}
Take $\vec u\in H_p^2(\Omega)^2$ with zero mean\footnote{This condition is necessary: lemma does not hold for $u\equiv \text{const}$ and small enough $\ell_{1,2}$}, i.e. $\int_\Omega \vec u dx_1dx_2=0$. Then for any $\gamma>0$
\begin{equation}
\label{eq:AgmoN}
\begin{split}
\|\vec u\|_{L^\infty}&\le \frac{\log^{\frac12}\bigl(1+\frac{4\pi^2 \gamma^2}{\ll_1\ll_2}\bigr)\|\vec u\|_{H^1}}{\sqrt{2\pi}} + \frac{\|\vec \ell\|_{\mathbb{R}^2} \|\Delta \vec u\|_{L^2}}{\gamma\sqrt{32\pi^3}}
\end{split}
\end{equation}
\end{lemma}
The proof is given in the appendix. Note that by using specific $\gamma$ in~\eqref{eq:AgmoN} one can derive some classical inequalities:
\begin{cor}
Set $\gamma= \|\vec u\|_{H^1}^{-\frac12}\|\Delta \vec u\|_{L^2}^{\frac12}$. We get 2D Agmon-type inequality~\cite[p.100]{NSE-Turbulence-book2001}:
\begin{equation}
\label{eq:Agmon}
\begin{split}
\|\vec u\|_{L^\infty}&\le (\sqrt{\frac{2\pi}{\ll_1\ll_2}}+\frac{\|\vec \ell\|_{\mathbb{R}^2}}{\sqrt{32\pi^3}})\|\vec u\|_{H^1}^{\frac12}\|\Delta \vec u\|_{L^2}^{\frac12}
\end{split}
\end{equation}
Set $\gamma= \|\vec u\|_{H^1}^{-1}\|\Delta \vec u\|_{L^2}$. We get 2D Brezis-type inequality~\cite{brezis1979nonlinear}:
\begin{equation}
\label{eq:Brezis}
\|\vec u\|_{L^\infty}\le \bigl(\frac{\|\vec \ell\|_{\mathbb{R}^2}}{\sqrt{32\pi^3}}+\frac{\log^{\frac12}(1+\frac{4\pi^2 \|\Delta \vec u\|_{L^2}^2}{\ll_1\ll_2 \|\vec u\|_{H^1}^2})}{\sqrt{2\pi}}\bigr)\|\vec u\|_{H^1}
\end{equation}
\end{cor}
Yet another building block used in the proofs below is the following version of Bellman lemma~\cite[Lemma 1.1, p.125]{NSE-Turbulence-book2001}:
\begin{lemma}\label{l:GB}
If differentiable $V:[0,\infty)\to [0,\infty)$ satisfies $\dot V(t) + \alpha(t) V(t) \le \beta(t)$ for some measurable scalar functions $\alpha,\beta$ then $V\to 0$ for $t\to\infty$ provided there exist $T>0$ such that
\begin{itemize}
\item $\limsup_{t\to\infty}\frac1T\int_t^{t+T} \max(\beta(s),0) ds =0$ and
\item $\liminf_{t\to\infty}{\frac1T\int_t^{t+T} \alpha(s) ds} >0$ and
\item $\limsup_{t\to\infty}\frac1T\int_t^{t+T} \max(0,-\alpha(s)) ds <+\infty$.
\end{itemize}
\end{lemma}
\subsubsection{Navier-Stokes equation: classical and weak formulations, and $H^1$-well-posedness}
\label{sec:navi-stok-equat}
The classical NSE is a system of two PDEs defining the dynamics of the scalar pressure field $p(x,y)$ and the viscous fluid velocity vector-field $\vec u=(u_1(x,y),u_2(x,y))^\top$:
\begin{align}
\label{eq:NSE-system}
\displaystyle\frac{du_1}{dt} &+ \vec u \cdot \nabla u_1 - \nu\Delta u_1 + p_{x} = f_1\\
\displaystyle\frac{du_2}{dt} &+ \vec u \cdot \nabla u_2 - \nu\Delta u_2 + p_{y} = f_2\\
&\nabla\cdot \vec u = (u_1)_x + (u_2)_y = 0
\end{align}
with initial condition $\vec u(0)=\vec u_0\in\mathcal{H}$, disturbance $\vec f=(f_1,f_2)^\top$ and periodic Boundary Conditions (BC) $u_{1,2}(x+\ell_1,y) = u_{1,2}(x,y)$, $u_{1,2}(x,y+\ell_2) = u_{1,2}(x,y)$ or, in the vector form:
\begin{equation}
\label{eq:NSE-vector}
\begin{split}
\displaystyle\frac{d\vec u}{dt} &+ (\vec u \cdot \nabla) \vec u - \nu\Delta\vec u+\nabla p = \vec f\\
&\nabla\cdot\vec u = 0
\end{split}
\end{equation}
Note that the div-free condition $\nabla\cdot\vec u = 0$ guarantees that the flow defined by $\vec u$ is volume preserving. It is well-known (e.g.~\cite[p.36]{NSE-Turbulence-book2001}) that every vector-field $\vec u$ in $\mathbb{R}^2$ admits Helmholtz-Leray decomposition: $\vec u = \nabla p + \vec v$, and $\nabla\cdot\vec v = 0$. We define Leray projection operator $P_l(\vec u)=\vec v$. $P_l$ is an orthogonal projector onto $\mathring{\mathcal{H}}$. $P_l\vec u$ can be computed uniquely: $P_l=\vec u-\nabla p$ provided $p\in H^1_p(\Omega)$ solves the Poisson equation $\Delta p = \nabla\cdot\vec u$ with periodic BC and $\int_\Omega p dxdy = 0$.
Define Stokes operator $A$ as follows (see~\cite[p.52]{NSE-Turbulence-book2001}):
\begin{equation}\begin{split}\nonumber
A\vec u&=-P_l\Delta \vec u,\\
u\in \mathscr{D}(A)&=\{\vec u\in \mathring{\mathcal{H}}:\Delta u\in \mathcal{H}\} = \mathring{\mathcal{V}}\cap ({H}_p^2(\Omega))^2\,.
\end{split}\end{equation}
Note that in fact $A\vec u=-\Delta \vec u$ for the case of periodic BC. $A$ is a self-adjoint positive operator with compact inverse, and (see~\cite[p.52,f.(6.19)]{NSE-Turbulence-book2001})
\begin{align}
(A\vec u,\vec \phi) &= ((\vec u,\vec\phi))\ge \lambda_1(\vec u,\vec{\phi})\,, \vec u\in \mathscr{D}(A),\vec{\phi}\in \mathring{\mathcal{V}},\label{eq:Poincare}\\
(A\vec u, A\vec u) &=(A(A^\frac12)\vec u,(A^\frac12)\vec u)\ge \lambda_1 (A \vec u, \vec u)\label{eq:gradA-bnd}\\
\lambda_1&=\frac{4\pi^2}{\max\{\ell_1,\ell_2\}^2}
\end{align}
In what follows we consider the variational form of~\eqref{eq:NSE-vector} which reduces~\eqref{eq:NSE-vector} to the subspace of div-free vector-fields with zero average, $\mathring{\mathcal{V}}$ (i.e. applies Leray projector $P_l$) and thus eliminates $p$. To this end we define:
\begin{align*}
(w,\phi)&=\int_\Omega w(x)\phi(x) dx, (\vec u,\vec\phi) = (u_1,\phi_1)+(u_2,\phi_2),\\%\quad w,\phi\in L_p^2(\Omega)\\
b(\vec u,\vec w,\vec\phi) &= (\vec u\cdot \nabla w_1,\phi_1) + (\vec u\cdot \nabla w_2,\phi_2),\\
((\vec u,\vec \phi)) &= (\nabla u_1,\nabla \phi_1)+(\nabla u_2,\nabla \phi_2)
\end{align*}
The form $b$ has a number of important properties which will be used below:
\begin{align*}
\text{skew-symmetry:}\quad &b(\vec u,\vec v,\vec \phi) = - b(\vec u,\vec \phi,\vec v)\\
\text{orthogonality:} \quad&b(\vec u,\vec v,\vec v) = 0,\quad b(\vec v,\vec v,A\vec v) = 0
\end{align*}
Upon multiplying~\eqref{eq:NSE-vector} by a test function $\vec\phi\in\mathring{\mathcal{V}}$ and integrating in $\Omega$ we get the following weak form of NSE~\eqref{eq:NSE-system}:
\begin{equation}
\label{eq:NSE-var}
\displaystyle\frac{d}{dt}(\vec u,\vec \phi) + b(\vec u,\vec u,\vec\phi) + \nu((\vec u,\vec \phi)) = (\vec f,\vec \phi), \quad\forall\vec\phi\in \mathring{\mathcal{V}}
\end{equation}
with initial condition $(\vec u(0),\vec\phi)=(\vec u_0,\vec\phi)$. By utilising~Theorem 7.4 from~\cite[p.58, Th.7.4]{NSE-Turbulence-book2001} and Cauchy-Schwarz inequality combined with Lemma~\ref{l:GB} it is not hard to obtain that $H^1$-norm of the weak (and strong) solution of NSE is bounded, and decays to $0$ exponentially provided $f=0$:
\begin{lemma}
Let $\vec u_0\in \mathring{\mathcal{H}}$ and $\vec f\in L^2(0,T,\mathring{\mathcal{H}})$. Then, on $[0,T]$ there exist the unique weak solution $\vec u\in C(0,T,\mathring{\mathcal{H}})$ of NSE in variational form~\eqref{eq:NSE-var}, and the components of $\vec u=(u_1,u_2)^\top$ verify: $u_i,(u_i)_{x,y}\in L^2(\Omega\times(0,T))$. If $\vec u_0\in\mathring{\mathcal{V}}$ then the weak solution coincides with the strong (classical) solution of~\eqref{eq:NSE-system}, and \[
\displaystyle\frac{du_i}{dt}, (u_i)_{x},(u_i)_{y}, (u_i)_{x y}\in L^2(\Omega\times(0,T))
\] i.e. $\vec u\in C(0,T,\mathring{\mathcal{V}}) \cap L^2(0,T,\mathscr{D}(A))$. If in addition $\vec f\in L^2(0,+\infty,\mathring{\mathcal{H}})$ then
\begin{align}
&\|\vec u(\cdot,t)\|^2_{L^2} \le \frac{\|\vec f\|^2_{L^2}}{(\nu\lambda_1)^2} + e^{(-\lambda_1 \nu)(t-s)}\|\vec u(\cdot,s)\|^2_{L^2} \label{eq:L2norm-decay}\\
&\|\nabla \vec u(\cdot,t)\|^2_{L^2} \le \frac{\|\vec f\|^2_{L^2}}{\nu^2\lambda_1} + e^{(-\lambda_1 \nu)(t-s)}\|\nabla \vec u(\cdot,s)\|^2_{L^2} \label{eq:grad-norm-decay}\\
&\frac1T\int_t^{T+t}\|A\vec u\|^2_{L^2}ds \le \theta_{t,T}\label{eq:Au-L2avrg-bnd}\\
&\theta_{t,T} = 2\frac{\|\vec f\|^2_{L^2}}{T\nu^3\lambda_1} + \frac{\|\vec f\|^2_{L^2}}{\nu^2} + \frac{2}{T\nu}e^{(-\lambda_1 \nu)t}\|\nabla\vec u_0\|^2_{L^2}\label{eq:Au-ThetatT}
\end{align}
\end{lemma}
|
\section{Introduction}\label{sec1}
The connections between tree amplitudes and loop integrals play an important role in quantum field theory. In a pioneering work \cite{Feynman:1963ax},
Feynman has shown that one can relate one-loop integrals in gauge theories to on-shell forward tree scattering amplitudes. Such relations are useful due to the simplicity of these amplitudes, which may be obtained by contour integrations containing the poles of the propagators. There has been a lot of effort in the literature to find out how unitarity methods, which rely on the fact that the loop amplitudes are determined by their singularities, can be used to evaluate higher order loops. These methods have been developed in a series of papers and successfully applied to loop computations in the standard model
\cite{bern:1994zxbern:1995cg,Britto:2004nc,Anastasiou:2006gt,Forde:2007mi,Bern:2007dw}
When a single propagator in a one-loop graph is cut, the integrand leads to a tree-level
amplitude. By developing Feynman's method, some QCD amplitudes have been computed just from
single cuts. This approach has been extended, in the real time formulation at zero temperature, to higher-order loops, both for time-ordered \cite{Berger:2009zb,Brandhuber:2005kd,Rodrigo:2008fp,Bierenbaum:2010cy} and for retarded boundary conditions \cite{Caron-Huot:2010fvq}.
In the present work, we consider the retarded thermal Green's functions
that appear as causal response functions,
which are most conveniently studied in the framework of the imaginary time formalism. In this formulation, in momentum space, a quantum field theory in $3+1$ dimensions involves a 3-dimensional Euclidean theory with a summation over the discrete Matsubara frequencies $Q_0= 2\pi i n T$ where $T$ is the temperature \cite{kapusta:book89,lebellac:book96,das:book97}.
This allows one to separate the usual $T=0$ part from the finite temperature contribution.
The main purpose of this work is to give a unified treatment of the relation between one- and two-loop graphs
and on-shell forward scattering tree amplitudes,
which is valid
both at zero as well as at finite temperature. This method, which greatly simplifies computations at finite temperature in thermal gauge theories,
has been previously used in connection with particular thermal amplitudes
\cite{Barton:1990fk,Frenkel:1992ts,Brandt:1993dk,Brandt:1999gf}.
Using appropriate analytic continuations of the external energies, one can recover the corresponding results obtained in the real time formalism.
In section \ref{sec2} we describe the forward scattering amplitude method, at one loop order, in the imaginary time formalism.
An application of this approach to the calculation of the lowest order retarded Green's functions in quantum gravity at zero temperature is given in section \ref{sec3}. The computations
are done in $D$ space-time dimensions using dimensional regularization
\cite{tHooft:1972fi}. The extension of this analysis to finite temperature
is presented in section \ref{sec4}. A workable two loop example of this formulation in a scalar theory is examined in section \ref{sec5}.
In section \ref{sec6}, we summarize the results and discuss a generalization of this approach at two-loop order. Some details of the calculations are given in the Appendices.
\section{The forward scattering method}\label{sec2}
In the imaginary time formalism, the one-loop integral associated with diagrams such as the one shown in Fig. \ref{fig2} is given by \cite{kapusta:book89,lebellac:book96}
\begin{equation}
\mu^{4-D}
\sum_{\alpha,\,a,s}
\int\,\frac{{d}^{D-1} Q}{(2\pi)^{D-1}} \left[T \sum_{n=-\infty}^{\infty}
f(Q_0,\vec Q,k_0^1,\vec k^1, \dots , k_0^L,\vec k^L)\right];\;\;\;
(k^1+k^2+\dots+k^L=0)
\ee
where $D$ is the space-time dimension,
$Q_0 = i \pi (2 n + m)T$ are the {\it Matsubara frequencies}, and $m=0$ or
$m=1$ for Bosons or Fermions, respectively
($\sum_{\alpha,\,a,s}$
represents a sum over Lorentz, internal and spinor indices of the loop
particle).
\begin{figure}[b]
\includegraphics[scale=0.5]{gen_one_loop.eps}
\caption{A generic one--loop diagram.}\label{fig2}
\end{figure}
The bosonic (fermionic) Matsubara summations can be done
using the relation
\begin{figure}[b]
\includegraphics[scale=0.4]{contours1.eps}
\caption{Figures (a) and (b) show the equivalent integration contours in Eq. \eqref{e22}.
Figure (c) shows the contour on the right hand side complex plane.
Dots on the real $Q_0$ axis indicate the poles of Eq. \eqref{poles1}}\label{figa1}
\end{figure}
\begin{equation}\label{e22}
T \sum_{n=-\infty}^{\infty} f(Q_0)
=
\oint_C \frac{d Q_0}{2\pi i} f(Q_0) \frac 1 2 \left[\coth\left(\frac{Q_0}{2T}\right)\right]^{\pm 1}
= \int_{-i\infty+\delta}^{i\infty+\delta} \frac{d Q_0}{2\pi i}
\frac{f(Q_0)+f(-Q_0)}{2}
\left(
1 \pm \frac{2}{e^{Q_0/T}\mp 1}
\right)
\ee
(the equivalent contours are shown in figures (\ref{figa1}a) and (\ref{figa1}b))
which clearly allows us to separate the $T=0$ part from the finite-temperature contribution.
Using Eq. \eqref{e22}, one-loop diagrams
containing $L$ internal lines, as shown in Fig. \ref{fig2},
can be expressed as
\begin{eqnarray}
&& \frac 1 2
\sum_{\mu,\,a,s}
\int\,\frac{{d}^{D-1} Q}{(2\pi)^{D-1}}\int_{-i\infty+\delta}^{i\infty+\delta}
\frac{{d} Q_0}{2\pi i}\, \left[
{f(Q_0,\vec Q, k^1,k^2,\dots, k^L)+f(-Q_0,\vec Q, k^1,k^2,\dots, k^L)}\right]
\left(
1 \pm \frac{2}{e^{Q_0/T}\mp 1}
\right)
\nonumber \\ &=&
\frac 1 2
\sum_{\mu,\,a,s}
\int\,\frac{{d}^{D-1} Q}{(2\pi)^{D-1}}\int_{-i\infty+\delta}^{i\infty+\delta}
\frac{{d} Q_0}{2\pi i}\,
\left[f(Q_0,\vec Q, k^1,k^2,\dots, k^L)+ Q\leftrightarrow - Q
\right]
\left(
1 \pm \frac{2}{e^{Q_0/T}\mp 1}
\right) ,
\end{eqnarray}
where we have replaced $\vec Q$ by $-\vec Q$ in the second term of the integrand.
In general $f(Q_0,\vec Q, k^1,k^2,\dots, k^L)$ has the following structure
\begin{equation}\label{poles1a}
f(Q_0,\vec Q, k^1,k^2,\dots, k^L) =
\frac{1}{{Q_0}^2-E^2(0,\vec Q)}\frac{1}{(Q_0+k^1_0)^2-E^2(\vec k^1,\vec Q)}
\cdots
\frac{t(Q, k^1,k^2\cdots k^L)}{(Q_0-k^L_0)^2-E^2(-\vec k^L,\vec Q)},
\ee
where $k^1+k^2+\cdots +k^L=0$, $E(\vec k^i,\vec Q)= \sqrt{|\vec Q+\vec k^i|^2 + m^2}$,
and $t(Q, k^1,k^2\cdots k^L)$ is a tensor (or a scalar in the case
of a scalar field theory) which is determined by the interaction
vertices of the theory. Using a partial fraction decomposition like
\begin{equation}\label{poles1}
\frac{1}{(Q_0+{k}_0)^2-E^2(\vec k, \vec Q)} =
\frac{1}{2\,E(\vec k,\vec Q)}\left[
\frac{1}{Q_0+ k_0 - E(\vec k, \vec Q)} -
\frac{1}{Q_0+ k_0 + E(\vec k, \vec Q)}
\right]
\ee
($(k_0,\vec k)$
represents any of the following possibilities: $(0,\vec 0)$, $(k_0^1,\vec k^1)$, $(k_0^1+k_0^2,\vec k^1+\vec k^2)$, $\dots$, $(-k_0^L,-\vec k^L)$)
the $Q_0$ integral can be done upon
closing the integration contour on
the right hand side of the complex plane
and using Cauchy's integral formula
(assuming that
$f(Q_0,\vec Q, k^1,k^2,\dots, k^L)$ vanishes on the curve ${\cal C}^2$ in Fig.
\ref{figa1}c, when ${\cal C}^2$ is stretched to infinity).
Performing shifts in the loop momentum
\footnote{When these shifts occur in a linearly divergent integral in four
dimensions, a surface term must also be included.
This is needed to recover the usual axial anomaly \cite{Adler:1969gk}.}
and taking into account that the external frequencies
are integer multiple of $i 2\pi T$ so that
$\exp(Q_0+i 2\pi \, l T)/T = \exp Q_0/T$
one can show that
\begin{equation}\label{barton1}
\begin{array}{l}
\includegraphics[scale=0.5]{gen_one_loop.eps}
\end{array}
= - \sum_{\alpha,\,a,s} \mu^{4-D}
\int\,\frac{{d}^{D-1} Q}{(2\pi)^{D-1}} \frac{1}{2 E_Q}
\left(\frac 1 2 \pm \frac{1}{e^{E_Q/T}\mp 1}\right)
{\cal A}(k^1,k^2,\cdots,k^L;Q)
,
\ee
(the overall minus sign arises from the clockwise contour on the right hand
complex plane)
where $E_Q=E(0,Q)$ and
\begin{eqnarray}\label{genamp}
{\cal A}(k^1,k^2,\cdots,k^L;Q)&\equiv&
\left(
\begin{array}{c}
\includegraphics[scale=0.7]{forward_gen.eps}
\end{array}
+ Q\leftrightarrow - Q
\right)_{Q_0=E_Q}
\nonumber \\
&&\nonumber \\
&+&
\mbox{ cyclic permutations of } V^1,V^2,\dots,V^L .
\end{eqnarray}
The tree amplitude ${\cal A}(k^1,k^2,\cdots,k^L;Q)$ is a forward
scattering amplitude which describes the scattering of an
on-shell particle of momentum $Q$ by external particles of momentum
$k^1$, $k^2$, $\dots$, $k^L$.
\begin{comment}
\end{comment}
The analytic continuation of some external energy $k_0^j\rightarrow k_0^j+ i \eta^j$, where $\eta^j$ is a positive infinitesimal quantity, with all other analytic continuations involving negative infinitesimals, yields the $j$-th retarded amplitude. In the following,
we shall not need to explicitly indicate this leg.
As a simple example, we consider
the self-energy in massless $g \phi^{3}$ theory.
In this case, the forward scattering tree amplitude is
(where the on-shell condition $Q_0 = E(0,\vec Q)=|\vec Q|$ is to be understood)
\begin{equation}\label{e28}
{\cal A}(k,Q) = g^2
\left(\frac{1}{k^2+2 k \cdot Q} + \frac{1}{k^2-2 k \cdot Q} \right)
\ee
(we have taken into account the combinatorial factor $1/2$ and also added the permutation
as indicated in Eq. \eqref{genamp}), so that
the one-loop scalar self-energy can be written as
\begin{eqnarray}\label{2.3}
\Pi(k) = -\frac{g^2}{2} \mu^{4-D}\int\,\frac{{d}^{D-1} Q}{(2\pi)^{D-1}}
\frac{1}{2 |\vec Q|}
\left(\frac{1}{k^2+2 k \cdot Q} + \frac{1}{k^2-2 k \cdot Q} \right).
\end{eqnarray}
\begin{comment}
\end{comment}
In the Appendix \ref{appA}, we compute the integral in \eqref{2.3} using the retarded analytic continuation of the external energy with $k_0\rightarrow k_0+i\eta$. This basic integral is also relevant for the self-energies of gauge fields and gravitons as we will see in the next section.
\begin{comment}
\end{comment}
\section{Scalar loops in a gravitational field}\label{sec3}
The action for a scalar field in curved space-time is simply
\begin{equation}\label{scalgravaction}
S = \frac{1}{2} \int d^D x \sqrt{-g}\left(
g^{\mu\nu} \partial_\mu\phi\partial_\nu\phi - m^2 \phi^2
\right).
\ee
For simplicity, let us set $m=0$. Defining
the gravitational field $\varphi^{\mu\nu}$
\begin{equation}
\kappa\varphi^{\mu\nu} = \sqrt{-g} g^{\mu\nu} - \eta^{\mu\nu},
\ee
where $\kappa=\sqrt{16\pi G}$ is related to Newton's constant $G$,
the action becomes
\begin{equation}\label{2.15}
S = \frac{1}{2} \int d^D x \left(-\phi\partial^2\phi +
\kappa\varphi^{\mu\nu} \partial_\mu\phi\partial_\nu\phi .
\right).
\ee
From \eqref{2.15} we obtain the following momentum space
Feynman rules:
\begin{subequations}\label{scalFR}
\begin{eqnarray}
\begin{array}{l}
\includegraphics[scale=0.7]{scal_prop.eps}
\end{array}
&:& \;\;\; \frac{i}{p^2},
\\
\begin{array}{l}
\includegraphics[scale=0.7]{scal_grav.eps}
\end{array}
&:& \;\;\; - \frac{i \kappa}{2}(p^1_\mu p^2_\nu +p^2_\mu p^1_\nu) ,
\end{eqnarray}
\end{subequations}
where $p^1_\mu$ and $p^2_\nu$ are the momenta of the scalar field.
In terms of the forward scattering amplitude,
\begin{eqnarray}\label{genamp1}
{\cal A}_{\mu\nu,\,\alpha\beta}(k,Q)
& = &
\left(
\begin{array}{c}
\includegraphics[scale=0.6]{FS_grav.eps}
\end{array}
+ Q\leftrightarrow - Q
\right)_{Q_0=|\vec Q|}
+ (\mu,\nu,k) \leftrightarrow (\alpha,\beta,-k)
\end{eqnarray}
the graviton self-energy at zero temperature can be written as
\begin{equation}\label{2.17a}
\Pi_{\mu\nu,\,\alpha\beta}(k) = -\frac 1 2
\mu^{4-D}\int \frac{d^{D-1} Q}{(2\pi)^{D-1}} \frac{1}{2|\vec Q|}
{\cal A}_{\mu\nu,\,\alpha\beta}(k,Q),
\ee
where we have used the $T=0$ part of the general expression given in Eq. \eqref{barton1}.
Upon using the Feynman rules given in Eqs. \eqref{scalFR}, we obtain
\begin{eqnarray}\label{2.17}
{\cal A}_{\mu\nu,\,\alpha\beta}(k,Q) &=& \frac{\kappa^2}{4}
\left\{
\frac{\left[Q_\mu(Q+k)_\nu + Q_\nu(Q+k)_\mu\right]\left[
(Q+k)_\alpha Q_\beta+(Q+k)_\beta Q_\alpha\right]}
{k^2+2k\cdot Q}
\right. \nonumber \\
&+& \left .
\frac{\left[Q_\alpha(Q-k)_\beta + Q_\beta(Q-k)_\alpha\right]\left[
(Q-k)_\mu Q_\nu+(Q-k)_\nu Q_\mu\right]}
{k^2-2k\cdot Q}
\right\}
\end{eqnarray}
(we have included the combinatorial factor $1/2$ and also added
the permutation indicated in Eq. \eqref{genamp}).
After integration, \eqref{2.17} can be expressed
(by covariance) in terms of the five tensors shown
in table I, so that
\begin{equation}\label{2.19}
\Pi_{\mu\nu,\, \alpha\beta}(k) = \sum_{i=1}^{5} {\cal T} ^i_{\mu\nu,\, \alpha\beta}(k)
C_i(k) .
\end{equation}
The coefficients $C_i$ can be obtained solving
the following system of five algebraic equations
\begin{eqnarray}\label{2.22}
\sum_{i=1}^5 {\cal T}^i_{\mu\nu,\, \alpha\beta}(k) {\cal T}^j{}^{\mu\nu,\, \alpha\beta}(k) C_i(k) &=&
\Pi_{\mu\nu,\,\alpha\beta}(k) {\cal T}^j{}^{\mu\nu,\, \alpha\beta}(k)
\nonumber \\ &=&
-\frac 1 2 \mu^{4-D}\int \frac{d^{D-1} Q}{(2\pi)^{D-1}} \frac{1}{2|\vec Q|}
{\cal A}_{\mu\nu,\, \alpha\beta}(k,Q)
{\cal T}^j{}^{\mu\nu,\, \alpha\beta}(k) ; \;\; j=1,\dots,5 .
\end{eqnarray}
\begin{table}[t]
\begin{center}
\begin{tabular}{c l}\hline \hline \\
${\cal T}^{1} _{\mu \nu ,\, \alpha \beta} (k)=$&$ \displaystyle{\frac{k_\mu k_\nu k_\alpha k_\beta}{k^4}}$\\ & \\
${\cal T}^2 _{\mu \nu ,\, \alpha \beta} (k)=$&$ \displaystyle{\eta_{\mu \nu} \eta_{\alpha\beta}}$ \\ & \\
${\cal T}^{3}_{\mu \nu ,\, \alpha \beta}(k)=$&$ \displaystyle{ \eta _{\mu \alpha} \eta_{\nu \beta} + \eta _{\mu \beta} \eta_{\nu \alpha}}$ \\ & \\
${\cal T}^4 _{\mu \nu ,\, \alpha \beta} (k)=$&$ \displaystyle{\frac{\eta_{\mu \nu} k_\alpha k_\beta + \eta_{\alpha \beta} k_\mu k_\nu}{k^2}}$ \\ & \\
${\cal T}^5 _{\mu \nu ,\, \alpha \beta}(k)=$&$ \displaystyle{\frac{\eta_{\mu \alpha} k_\nu k_\beta +
\eta_{\mu \beta} k_\nu k_\alpha + \eta_{\nu \alpha} k_\mu k_\beta +
\eta_{\nu \beta} k_\mu k_\alpha}{k^2}}$
\\ \\ \hline \hline
\end{tabular}\caption{The five independent tensors built from
$\eta_{\mu\nu}$ and $k_\mu$, satisfying the symmetry conditions
${\cal T}^{i}_{\mu\nu,\, \alpha\beta} (k)= {\cal T}^{i}_{\nu\mu,\, \alpha\beta}(k)=
{\cal T}^{i}_{\mu\nu,\, \beta\alpha} (k)= {\cal T}^{i}_{\alpha\beta,\, \mu\nu}(k)$.}
\end{center}
\end{table}\label{tab1}
Using simple identities such as
\begin{equation}
\frac{(k\cdot Q)^2}{k^4 - 4 (k\cdot Q)^2}
= \frac 1 4 \frac{k^4}{k^4 - 4 (k\cdot Q)^2} - \frac 1 4
\ee
and setting equal to zero any dimensional regularized integrals
that are independent of $|\vec Q|$, then all the integrals on the right-hand side
of Eq. \eqref{2.22} can be reduced to
the following basic scalar integral
\begin{eqnarray}\label{basicI}
I(k) &=& \kappa^2 \mu^{4-D}\int\,\frac{{d}^{D-1} Q}{(2\pi)^{D-1}}
\frac{1}{2 |\vec Q|}
\left(\frac{1}{k^2+2 k \cdot Q} + \frac{1}{k^2-2 k \cdot Q} \right)
\nonumber \\
&=&
-\kappa^2 \mu^{4-D} \frac{1}{(4\pi)^{D/2}} \Gamma(2-D/2)
\frac{\Gamma^2(D/2-1)}{\Gamma(D-2)}
\left[-i\eta\; {\rm sgn}(k_0) -k^2\right]^{D/2-2},
\end{eqnarray}
where we have used the result given in Eq. \eqref{A8} (up to an overall factor).
Solving the linear system of equations in Eq. \eqref{2.22}, we obtain
\begin{comment}
\end{comment}
\begin{subequations}\label{C1to5}
\begin{equation}
C_1(k) = \frac{D(D-2)}{32 (D^2-1)} k^4 I(k),
\ee
\begin{equation}
C_2(k) = \frac{k^4}{32(D^2-1)} I(k),
\ee
\begin{equation}
C_3(k) = C_2(k) = - C_5(k) ,
\ee
and
\begin{equation}
C_4(k) = \frac{D}{32(D^2-1)} k^4 I(k)
\ee
\end{subequations}
Substituting \eqref{C1to5} into \eqref{2.19}, we obtain
\begin{eqnarray}\label{2.25}
\Pi_{\mu\nu,\,\alpha\beta}(k) &=&\left[
\frac{D(D-2)k_\mu k_\nu k_\alpha k_\beta}{k^4} +
\eta_{\mu\nu} \eta_{\alpha\beta}+
\eta_{\mu\alpha} \eta_{\nu\beta}+
\eta_{\mu\beta} \eta_{\nu\alpha}
\right. \nonumber \\
&+& \left.
\frac{D}{k^2} (\eta_{\mu\nu} k_\alpha k_\beta+\eta_{\alpha\beta} k_\mu k_\nu)
-\frac{\eta_{\mu \alpha} k_\nu k_\beta +
\eta_{\mu \beta} k_\nu k_\alpha + \eta_{\nu \alpha} k_\mu k_\beta +
\eta_{\nu \beta} k_\mu k_\alpha}{k^2}
\right] \frac{k^4I(k)}{32(D^2-1)} .
\end{eqnarray}
This result
satisfies the Ward identity
\begin{equation}\label{Ward0}
(\eta^{\mu\rho} k^\nu + \eta^{\nu\rho} k^\mu - \eta^{\mu\nu} k^{\rho})
\Pi_{\mu\nu,\,\alpha\beta} = 0 ,
\ee
which is stronger than the identity of the pure gravitational case in
the Eq. (3.36) of \cite{Brandt:2016eaj} (see also \cite{Lavrov:2019nuz}).
This can be understood since there is no BRST symmetry in the scalar case. There is only the classical diffeomorphism invariance of \eqref{scalgravaction} under
\begin{equation}\label{e17a}
\delta \tilde g^{\mu\nu} = \kappa \delta (\phi^{\mu\nu})
= \tilde g^{\mu\rho} \partial_\rho \theta^\nu +
\tilde g^{\nu\rho} \partial_\rho \theta^\mu
-\partial_\rho(\tilde g^{\mu\nu}\theta^\rho);
\;\;\; \tilde g^{\mu\nu} = \sqrt{-g} g^{\mu\nu}.
\end{equation}
associated with the coordinate transformations
$x^\mu\rightarrow x^\mu - \theta^\mu$.
\section{Graviton thermal self-energy}\label{sec4}
Let us now consider the thermal part of Eq. \eqref{barton1} which contains the Bose-Einstein distribution.
When the external momenta $k$ are such that $k\ll T$ the loop integral
is dominated by the {\it hard thermal loop} region.
Using expansions like
\begin{equation}
\frac{1}{k^2 + 2 k\cdot Q} = \frac{1}{2 k\cdot Q}
- \frac{k^2}{(2 k\cdot Q) ^2} + \dots ,
\ee
($k^2 \ll k\cdot Q$)
the tree amplitude in Eq. \eqref{2.17} reduces to
\begin{comment}
\end{comment}
\begin{equation}\label{Ahtl1}
{\cal A}^{HTL}_{\mu\nu,\,\alpha\beta}(k,Q) = \frac{\kappa^2}{2}|\vec Q|^2
\left(
\frac{k_\mu \hat Q_\nu \hat Q_\alpha \hat Q_\beta}{k\cdot \hat Q}+
\frac{\hat Q_\mu k_\nu \hat Q_\alpha \hat Q_\beta}{k\cdot \hat Q}+
\frac{\hat Q_\mu \hat Q_\nu k_\alpha \hat Q_\beta}{k\cdot \hat Q}+
\frac{\hat Q_\mu \hat Q_\nu \hat Q_\alpha k_\beta}{k\cdot \hat Q}-
\frac{k^2 \hat Q_\mu \hat Q_\nu \hat Q_\alpha \hat Q_\beta}{(k\cdot \hat Q)^2}
\right)
\ee
to leading order in the high temperature expansion ($\hat Q = Q/|\vec Q|$).
Inserting Eq. \eqref{Ahtl1} into the thermal part of
Eq. \eqref{barton1}
we obtain the following expression for
the thermal scalar loop contribution to the graviton self-energy
\begin{eqnarray}
\Pi^{HTL}_{\mu\nu,\,\alpha\beta}
\begin{comment}
\end{comment}
&=&
-\frac{\kappa^2}{4} \frac{\mu^{4-D} T^{D}}{(2\pi)^{D-1}}
\int_0^{\infty}
\frac{u^{D-1} du}{e^{u}-1}
\nonumber \\ &&
\int{d\Omega}
\left(
\frac{k_\mu \hat Q_\nu \hat Q_\alpha \hat Q_\beta}{k\cdot \hat Q}+
\frac{\hat Q_\mu k_\nu \hat Q_\alpha \hat Q_\beta}{k\cdot \hat Q}+
\frac{\hat Q_\mu \hat Q_\nu k_\alpha \hat Q_\beta}{k\cdot \hat Q}+
\frac{\hat Q_\mu \hat Q_\nu \hat Q_\alpha k_\beta}{k\cdot \hat Q}-
\frac{k^2 \hat Q_\mu \hat Q_\nu \hat Q_\alpha \hat Q_\beta}{(k\cdot \hat Q)^2}
\right).
\nonumber \\
\end{eqnarray}
Upon integrating over $u$
we obtain
\begin{eqnarray}
\Pi^{HTL}_{\mu\nu,\,\alpha\beta} &=&
-\frac{\kappa^2}{4} \frac{\mu^{4-D} T^{D}}{(2\pi)^{D-1}}
\Gamma(D)\zeta(D)
\nonumber \\
&& \int{d\Omega}
\left(
\frac{k_\mu \hat Q_\nu \hat Q_\alpha \hat Q_\beta}{k\cdot \hat Q}+
\frac{\hat Q_\mu k_\nu \hat Q_\alpha \hat Q_\beta}{k\cdot \hat Q}+
\frac{\hat Q_\mu \hat Q_\nu k_\alpha \hat Q_\beta}{k\cdot \hat Q}+
\frac{\hat Q_\mu \hat Q_\nu \hat Q_\alpha k_\beta}{k\cdot \hat Q}-
\frac{k^2 \hat Q_\mu \hat Q_\nu \hat Q_\alpha \hat Q_\beta}{(k\cdot \hat Q)^2}
\right),
\end{eqnarray}
where $\Gamma$ and $\zeta$ denote respectively the Gamma and Zeta functions.
For $D=4$, we obtain
\begin{eqnarray}
\Pi^{HTL}_{\mu\nu,\,\alpha\beta} &=&
-\frac{\kappa^2 \pi^2 T^4}{120}
\int \frac{d\Omega}{4\pi}
\left(
\frac{k_\mu \hat Q_\nu \hat Q_\alpha \hat Q_\beta}{k\cdot \hat Q}+
\frac{\hat Q_\mu k_\nu \hat Q_\alpha \hat Q_\beta}{k\cdot \hat Q}+
\frac{\hat Q_\mu \hat Q_\nu k_\alpha \hat Q_\beta}{k\cdot \hat Q}+
\frac{\hat Q_\mu \hat Q_\nu \hat Q_\alpha k_\beta}{k\cdot \hat Q}-
\frac{k^2 \hat Q_\mu \hat Q_\nu \hat Q_\alpha \hat Q_\beta}{(k\cdot \hat Q)^2}
\right)
\nonumber \\
\end{eqnarray}
which is half the result in the case of pure gravity (A graviton loop produces a factor of two,
which is associated with the extra degree of freedom of a massless spin 2 particle).
\begin{comment}
\end{comment}
At finite temperature, the tadpole diagram shown in Fig. \ref{figtad}a does not vanish.
\begin{figure}[t]
\includegraphics[scale=0.7]{tadT.eps}
\caption{The one-graviton thermal amplitude and the corresponding
forward scattering amplitude.}\label{figtad}
\end{figure}
Taking into account the symmetry factor $1/2$ and the Feynman rules in \eqref{scalFR}
Fig. \ref{figtad}b yields
\begin{equation}\label{tadAT}
{\cal A}_{\mu\nu}(k,Q) = -\kappa \, Q_\mu Q_\nu.
\ee
We can verify that
the amplitudes in Eqs. \eqref{Ahtl1} and \eqref{tadAT} are related
by the Ward identity
\begin{equation}
\left(
\eta^{\mu\rho} k^\nu + \eta^{\nu\rho} k^\mu - \eta^{\mu\nu} k^{\rho}
\right)
\; {\cal A}^{HTL}_{\mu\nu,\,\alpha\beta}(k,Q) = -\kappa\,
\eta^{\nu\rho}\left(k_\alpha \delta^\mu_\beta + k_\beta \delta^\mu_\alpha\right)
\; {\cal A}_{\mu\nu}(Q).
\ee
Since this identity is verified at the level of the tree amplitude,
the thermal self-energy and the one-graviton are also related by
the same identity, namely
\begin{equation}
\left(\eta^{\mu\rho} k^\nu + \eta^{\nu\rho} k^\mu - \eta^{\mu\nu} k^{\rho}\right)
{\Pi }^{HTL}_{\mu\nu,\,\alpha\beta} = -\kappa\,
\eta^{\nu\rho}\left(k_\alpha \delta^\mu_\beta+k_\beta \delta^\mu_\alpha\right)
{\Pi}^{tadpole}_{\mu\nu},
\ee
in the hard thermal loop limit. Notice that at $T=0$ Eq. \eqref{Ward0}
is consistent with the absence of tadpoles.
In fact, it is straightforward to verify that the {\it exact amplitude} in
Eq. \eqref{2.17} satisfy the same Ward identity as the hard thermal loop
case, namely
\begin{equation}
\left(
\eta^{\mu\rho} k^\nu + \eta^{\nu\rho} k^\mu - \eta^{\mu\nu} k^{\rho}
\right)
\; {\cal A}_{\mu\nu,\,\alpha\beta}(k,Q) = -\kappa\,
\eta^{\nu\rho}\left(k_\alpha \delta^\mu_\beta + k_\beta \delta^\mu_\alpha\right)
\; {\cal A}_{\mu\nu}(Q),
\ee
At $T=0$,
\begin{equation}
-\frac 1 2 \int\frac{d^{D-1} Q}{(2\pi)^{D-1}} \frac{1}{2 |\vec Q|}{\cal A}_{\mu\nu}(Q)=
\frac 1 4 \int\frac{d^{D-1} Q}{(2\pi)^{D-1}}
\frac{Q_\mu Q_\nu}{|\vec Q|} = 0
\ee
and at finite $T$ we have
\begin{equation}
- \mu^{4-D}\int\frac{d^{D-1} Q}{(2\pi)^{D-1}}\frac{1}{2|\vec Q|}\frac{{\cal A}_{\mu\nu}(Q)}
{e^{|\vec Q|/T} - 1}=
\mu^{4-D}\frac{\kappa}{2} \int\frac{d^{D-1} Q}{(2\pi)^{D-1}} \frac{1}{|\vec Q|}\frac{Q_\mu Q_\nu}
{e^{|\vec Q|/T} - 1} = C \mu^{4-D} T^{D}
\left(D u_\mu u_\nu - \eta_{\mu\nu}\right),
\ee
where $C$, is a constant, $u = (1,0,0, \dots ,0)$ is a rest frame four-velocity
and we have used that the result is traceless, since $Q^2=0$.
From the $00$ component
\begin{equation}\label{A12}
\mu^{4-D}\frac{\kappa}{2} \int\frac{d^{D-1} Q}{(2\pi)^{D-1}} \frac{|\vec Q|}{e^{|\vec Q|/T} - 1}
= (D-1) C \mu^{4-D} T^{D},
\ee
we obtain \cite{gradshteyn}
\begin{comment}
\end{comment}
\begin{equation}
C = \mu^{4-D} \frac{\kappa}{2 (D-1) (2 \pi)^{D-1}} \frac{2 \pi^{\frac{D-1}{2}}}
{\Gamma\left(\frac{D-1}{2}\right)}
\Gamma(D) \zeta(D)
= \frac{\mu^{4-D} \kappa}{2^{D-1} \pi^{\frac{D-1}{2}}}\frac{\Gamma(D-1)}
{\Gamma\left(\frac{D-1}{2}\right)}\zeta(D).
\ee
For $D=4$ we obtain,
\begin{equation}
C = \frac{\kappa}{8 \pi \pi^{1/2} \Gamma(3/2)} \Gamma(3) \zeta(4)
= \frac{\kappa\pi^2}{180},
\ee
which is half the result obtained in pure gravity.
\begin{comment}
\end{comment}
\begin{comment
\end{comment
\section{Retarded amplitudes at two loops}\label{sec5}
In order to study higher-order thermal loop contributions in the imaginary time formalism, we consider, for simplicity the scalar $\phi^4$ field theory, described by the Lagrangian
\begin{equation}\label{e51}
{\cal L} = \frac 1 2 \partial_\mu \phi \partial^\mu \phi - \frac{m^2}{2} \phi^2 - \frac{\lambda}{4!} \phi^4
\end{equation}
and examine the retarded self-energy of the scalar field which can be obtained by the prescription
$k_0 \rightarrow k_0 + i \eta$. The one-loop contribution can be expressed in terms of a forward scattering amplitude for a on-shell particle as discussed in Eqs. \eqref{barton1} and \eqref{genamp}.
\begin{figure}[b]
\includegraphics[scale=0.6]{fig4.eps}
\caption{Two-loop
self-energy diagrams (a,b) and the corresponding forward amplitudes obtained by opening, successively, the $Q$ (a', b') and $q$ (a'', b'') internal lines.
}\label{fig4}
\end{figure}
In going at two loops, we note that there are two types of Feynman diagrams, as shown in the Figs. (\ref{fig4}a) and (\ref{fig4}b).
\begin{comment}
\end{comment}
We can evaluate the summations over the discrete frequencies $Q_0 = i 2\pi n T$ and $q_0 = i 2 \pi s T$ with the help of the relation \eqref{e22}. Next, we proceed to evaluate the $Q_0$ and $q_0$ integrals by contour integrations in the right hand side of their complex planes.
In the case of Fig. (\ref{fig4}a), according to Eq. \eqref{barton1} this opens successively the Feynman propagators, and sets the $Q$ and $q$ particles on their mass-shell, with positive energies $Q_0$ = $E_Q$ and $q_0 = E_q$. In Fig. (\ref{fig4}a'), which has a double pole, one can do the $q_0$ integral by using Cauchy theorem, leading to a graph like that shown in Fig. (\ref{fig4}a''). This diagram, which corresponds to the forward scattering amplitudes of two on-shell particles is multiplied by a phase-space integral, leading to a k-independent result given by
\begin{equation}\label{e52}
\frac{\lambda^2}{4}
\int\frac{d^{D-1} Q}{(2\pi)^{D-1}}
\frac{1}{2 E_Q} \int\frac{d^{D-1} q}{2 E_q}
\left[\frac 1 2 +\frac{1}{e^{E_Q/T}-1}\right]
\left[\frac 1 2 +\frac{1}{e^{E_q/T}-1}\right]
\frac{-1}{|\vec q|^2+m^2}
\end{equation}
\begin{comment}
\end{comment}
We will next apply this method to the diagram shown in Fig. (\ref{fig4}b), which yields a $k$-dependent contribution given by
\begin{equation}\label{e53nv}
-\frac{\lambda^2}{3!}
\int\frac{d^{D-1} Q}{(2\pi)^{D-1}} \int\frac{d^{D-1} q}{(2\pi)^{D-1}}
T^2 \sum_{n,s}
\frac{1}{Q^2-m^2}\frac{1}{q^2-m^2}\frac{1}{(k+Q+q)^2-m^2},
\ee
where $Q_0=i 2\pi n T$, $q_0=i 2\pi s T$ and $k_0=i 2\pi l T$.
Performing the summations with the help of Eq. \eqref{e22} and replacing $\vec Q\rightarrow -\vec Q$ and $\vec q\rightarrow -\vec q$ in some terms,
we obtain for the integrand of \eqref{e53nv} the result
\begin{eqnarray}\label{e54nv}
I &=& \int_{-i\infty}^{i\infty} \frac{dq_0}{2\pi i}\left[\frac{1}{2}+N(q_0)\right]
\int_{-i\infty}^{i\infty} \frac{dQ_0}{2\pi i}\left[\frac{1}{2}+N(Q_0)\right]
\frac{1}{Q^2-m^2}\frac{1}{q^2-m^2}
\nonumber \\
& & \left[\frac{1}{(k+Q+q)^2-m^2}+
\frac{1}{(k-Q-q)^2-m^2}+
\frac{1}{(k+Q-q)^2-m^2}+
\frac{1}{(k-Q+q)^2-m^2}
\right],
\end{eqnarray}
where $N(\omega)$ denotes the statistical factor
\begin{equation}\label{e55nv}
N(\omega) = \frac{1}{e^{\omega/T}-1}.
\ee
We first perform the $Q_0$ integration along the contour shown in
Fig.(\ref{figa1}c) by using Cauchy's theorem. Making appropriate shifts and using the fact that $k_0$ is an integer multiple of $i2\pi T$ we obtain the following two contributions
\begin{eqnarray}\label{e56nv}
I_1 &=& -\frac{1}{2 E_Q}
\int_{-i\infty}^{i\infty} \frac{dq_0}{2\pi i}\left[\frac{1}{2}+N(q_0)\right]
\frac{1}{q^2-m^2} \left[\frac{1}{2}+N(E_Q)\right]
\nonumber \\
&&
\left[\frac{1}{(k+Q+q)^2-m^2}+\frac{1}{(k-Q-q)^2-m^2}+
\frac{1}{(k+Q-q)^2-m^2}+\frac{1}{(k-Q+q)^2-m^2}
\right]_{Q_0=E_Q},
\end{eqnarray}
and
\begin{eqnarray}\label{e57nv}
I_2 &=& -\frac{1}{2 E_Q}
\int_{-i\infty}^{i\infty} \frac{dq_0}{2\pi i}\left[\frac{1}{2}+N(q_0)\right]\frac{1}{q^2-m^2}
\nonumber \\ &&
\left\{\left[\frac{1}{2}+N(E_Q-q_0)\right]
\left[\frac{1}{(k+Q-q)^2-m^2}+\frac{1}{(k-Q+q)^2-m^2}\right]
\right.
\nonumber\\ &+& \left.
\left[\frac{1}{2}+N(E_Q+q_0)\right]
\left[\frac{1}{(k+Q+q)^2-m^2}+\frac{1}{(k-Q-q)^2-m^2}\right]\right\}_{Q_0=E_Q} .
\end{eqnarray}
This procedure opens the $Q$-propagator and sets the $Q$-particle on-shell with a positive energy $Q_0=E_Q$. This leads to diagrams of the form shown in Fig.(\ref{fig4}b'), where graphs obtained by permutations of vertices as well as by the replacement $Q \rightarrow -Q$ are to be understood.
We can similarly integrate over $q_0$ by using Cauchy's theorem. The contribution at the pole $q_0 = E_q$ coming from the equation \eqref{e56nv} is
\begin{eqnarray}\label{e58nv}
I_{11} &=&
\frac{1}{2E_Q}\frac{1}{2E_q}
\left[\frac{1}{2}+N(E_Q)\right]\left[\frac{1}{2}+N(E_q)\right]
\nonumber \\ &&
\left[\frac{1}{(k+Q+q)^2-m^2}+\frac{1}{(k-Q-q)^2-m^2}+
\frac{1}{(k+Q-q)^2-m^2}+\frac{1}{(k-Q+q)^2-m^2}
\right]_{Q_0=E_Q;\,q_0=E_q} .
\end{eqnarray}
The contribution coming from \eqref{e56nv} at the other poles may be combined with that obtained from \eqref{e57nv} at the pole $q_0 = E_q$, to give
\begin{eqnarray}\label{e59nv}
I_{12}+I_{21} &=&
\frac{1}{2E_Q}\frac{1}{2E_q}\left\{ \left[\frac{1}{2}+N(E_Q)\right]\left(\left[\frac{1}{2}+N(E_q-E_Q)\right]
\left[\frac{1}{(k+Q-q)^2-m^2}+\frac{1}{(k-Q+q)^2-m^2}\right]
\right. \right. \nonumber \\ &+&
\left. \left. \left[\frac{1}{2}+N(E_Q+E_q)\right]
\left[\frac{1}{(k+Q+q)^2-m^2}+\frac{1}{(k-Q-q)^2-m^2}\right]\right)
+\left\{Q\leftrightarrow q\right\}\right\}_{Q_0=E_Q;\, q_0=E_q}.
\end{eqnarray}
Finally, the contribution obtained from the poles of the other propagators in \eqref{e57nv} is given by
\begin{eqnarray}\label{e510nv}
I_{22} &=& \frac{1}{2E_Q}\frac{1}{2E_q}\frac{1}{2}\left\{ \left[\frac{1}{2}+N(-E_q)\right] \left[\frac{1}{2}+N(E_Q+E_q)\right]
\left[\frac{1}{(k+Q+q)^2-m^2}+\frac{1}{(k-Q-q)^2-m^2}\right]
\right. \nonumber \\ &+&
\left. \left[\frac{1}{2}+N(E_q)\right]\left[\frac{1}{2}+N(E_q-E_Q)\right]
\left[\frac{1}{(k+Q-q)^2-m^2}+\frac{1}{(k-Q+q)^2-m^2}\right]
+\left\{Q\leftrightarrow q\right\}\right\}_{Q_0=E_Q;\, q_0=E_q},
\end{eqnarray}
where we have symmetrized under the interchange $Q \leftrightarrow q$.
The above expressions can be simplified by using the following identities involving the statistical factors
\begin{subequations}\label{e511nv}
\begin{equation}\label{e511nva}
N(E_Q-E_q)+N(E_q-E_Q) = -1
\ee
\begin{equation}\label{e511nvb}
N(E_Q-E_q)-N(E_q-E_Q) = \frac{N(E_Q)+N(E_q) + 2 N(E_Q)N(E_q)}{N(E_q)-N(E_Q)}
\ee
\begin{equation}\label{e511nvc}
N(E_Q+E_q) = \frac{N(E_Q)N(E_q)}{1+N(E_Q)+N(E_q)} .
\ee
\end{subequations}
In this way, the expressions given in Eqs. \eqref{e59nv} and \eqref{e510nv} may, respectively, be written as
\begin{eqnarray}\label{e512nv}
I_{12}+I_{21} &=&
\frac{1}{2E_Q}\frac{1}{2E_q}\frac{1}{2}\left\{
\left[N(E_Q)+N(E_q) + 2 N(E_Q)N(E_q)\right]
\left[\frac{1}{(k+Q-q)^2-m^2}+\frac{1}{(k-Q+q)^2-m^2}\right]
\right. \nonumber \\ &+&
\left.
\left[1+N(E_Q)+N(E_q)+2N(E_Q)N(E_q)\right]
\left[\frac{1}{(k+Q+q)^2-m^2}+\frac{1}{(k-Q-q)^2-m^2}\right]
\right\}_{Q_0=E_Q;\, q_0=E_q}.
\end{eqnarray}
and
\begin{eqnarray}\label{e513nv}
I_{22} &=&
-\frac{1}{2E_Q}\frac{1}{2E_q}\frac{1}{4}\left\{
\left[1+N(E_Q)+N(E_q) + 2 N(E_Q)N(E_q)\right]
\left[\frac{1}{(k+Q+q)^2-m^2}+\frac{1}{(k-Q-q)^2-m^2}\right]
\right. \nonumber \\ &+&
\left.
\left[N(E_Q)+N(E_q)+2N(E_Q)N(E_q)\right]
\left[\frac{1}{(k+Q-q)^2-m^2}+\frac{1}{(k-Q+q)^2-m^2}\right]
\right\}_{Q_0=E_Q;\, q_0=E_q}.
\end{eqnarray}
The above procedure
opens the $q$-propagator in the one-loop vertex correction
shown in Fig.(\ref{fig4}b')
and sets the $q$ particle on-shell with a positive energy $q_0=E_q$. This leads to diagrams of the form shown in Fig. (\ref{fig4}b''), where graphs obtained by permutations and by making $q\leftrightarrow -q$ are to be understood.
We still must take into account the contributions coming from the poles of the statistical factors $N(E_Q\mp q_0)$ in the Eq. \eqref{e57nv}. These arise at the points $q_0=\pm E_Q+i 2\pi jT$, where $j$ is an integer, along two lines parallel to the imaginary $q_0$ axis. Upon using Eq. \eqref{e22}
and the identities \eqref{e511nv}, one can similarly evaluate such contributions which should be added to those given in Eqs. \eqref{e58nv}, \eqref{e512nv} and \eqref{e513nv}. Proceeding in this way, and using partial fraction decompositions like \eqref{poles1} it turns out that the complete contribution of the diagram shown in Fig. (\ref{fig4}b) may be written in the form
\begin{equation}\label{e514nv}
\frac{\lambda^2}{3!}\int\frac{d^{D-1} Q}{(2\pi)^{D-1}}\frac{1}{2 E_Q}
\int\frac{d^{D-1} q}{(2\pi)^{D-1}}\frac{1}{2 E_q}\frac{1}{2 E}
\left[F_0 + 3 N(E_Q) F_1 + 3 N(E_Q)N(E_q) F_2\right],
\end{equation}
where $E = \sqrt{(\vec k+\vec Q+\vec q)^2+m^2}$ and
\begin{subequations}\label{e58}
\begin{equation}\label{e58a}
F_0 = f(E_Q,E_q,E)
\end{equation}
\begin{equation}\label{e58b}
F_1 = f(E_Q,E_q,E) + f(-E_Q,E_q,E)
\end{equation}
\begin{equation}\label{e58c}
F_2 = f(E_Q,E_q,E) + f(-E_Q,E_q,E) + f(E_Q,-E_q,E) - f(E_Q,E_q,-E) ,
\end{equation}
\end{subequations}
where (compare with Eq. \eqref{e212})
\begin{equation}\label{e516nv}
f(E_Q,E_q,E) = \frac{1}{k_0+i\eta+E_Q+E_q+E}+\frac{1}{-k_0-i\eta+E_Q+E_q+E} .
\end{equation}
The above form of the retarded self-energy is quite convenient, especially for the calculation of its imaginary part which is related to various physical processes. In particular, the imaginary part of $F_0$ is related to the probability that the external field creates three scalar particles at zero temperature. This can be seen from Eq. \eqref{e516nv} as well as by cutting the propagator in Fig. (\ref{fig4}b'').
(Performing the $Q$ and $q$ integrations in Eq. \eqref{e514nv}, one obtains
in the massless case at zero temperature the result given in the Appendix \ref{appC}).
The result \eqref{e514nv}
agrees with the one obtained in the reference \cite{Parwani:1991gq} in the imaginary time formalism, by an elaborate evaluation of the loops appearing in Fig. (\ref{fig4}b)
An alternative derivation in the real time formulation was given in Ref. \cite{Brandt:1999gf}. In this formalism, the $2\times 2$ matrix propagator is a sum of two parts, one of which represents an on-shell thermal contribution
that involves a factor $N(|p_0|)\delta(p^2-m^2)$, where $p$ is the four-momentum of the line. Such factors naturally lead to the opening of loop diagrams at non-zero temperature. However, this property does not elucidate the fact that forward scattering tree amplitudes exist as well at zero temperature.
Based on the present formulation, we may interpret the contributions coming from the thermal self-energy diagram (\ref{fig4}b) at two loops as follows. $F_0$ corresponds to a forward scattering amplitude of two on-shell particles at zero temperature;
The contribution linear in the statistical factor might be thought as a forward scattering amplitude of a on-shell thermal particle
and of a forward scattering amplitude of a single, on-shell particle at zero temperature. Finally, the contribution quadratic in the statistical factor may be interpreted
as the forward scattering amplitude of two on-shell thermal particles (for a related approach, see references \cite{Wong:2000hq,Kapusta:2001jw}).
\section{Discussion}\label{sec6}
Using the imaginary-time formalism, we have obtained the relation \eqref{barton1}
between general one-loop diagrams and on-shell forward scattering tree amplitudes which holds at all temperatures $T$ in $D$ space-time dimensions.
We note that the right hand side of Eq. \eqref{barton1} has been obtained by performing first the $Q_0$ integral by contour integration. This has the effect of opening one propagator in the loop which leads to a tree, on-shell forward scattering amplitude.
At high temperatures, this relation provides an efficient way for
computing the retarded thermal contributions in gauge theories.
As an application of Eq. \eqref{barton1}, we have calculated in quantum gravity
(coupled to scalar fields) the lowest orders retarded graviton amplitudes at all temperatures.
At zero temperature there occurs a subtlety due to the fact that the contribution from the arcs at infinity (see Fig. \ref{figa1}c), which arise when first performing the $Q_0$ integration,
is linearly divergent.
However, when performing first the $\vec Q$ integrations, it turns out that the remaining $Q_0$ integration converges for $D < 0$ (see Appendix \ref{appB}),
in which case the contributions from the arcs at infinity vanish. Thus, by using a consistent analytic continuation, we see that the neglect of the contributions from the arcs at infinity can be justified in dimensional regularization.
At high temperatures, the leading contributions of all 1PI retarded functions are proportional to $T^D$. This is required by the Ward identities which reflect the gauge invariance of the theory under coordinate transformations.
We have applied the above method to the evaluation, in a scalar theory, of the retarded self-energy at two loops in the imaginary time formalism.
But one can show that, by using an integral reduction procedure in dimensional regularization \cite{tHooft:1978jhc}, one can express any loop integral as a linear combination of scalar loop integrals. An example of this procedure in quantum gravity was given in section \ref{sec3}.
Based on these calculations, we point out some features which may hold for more generic one-particle irreducible diagrams at two loops. In this case, there are two independent internal energies which can be integrated out by contour integrations. This leads, by successive openings of the internal lines, to two cut lines which set the corresponding particles on-shell. In this way, a 1PI two-loop diagram may be opened to a connected tree-level graph. We also note here that in the
imaginary time formalism, one can have at most one statistical factor for each loop (see Eq. \eqref{e22}).
However, a new feature occurs due to the presence, in intermediate calculations at two loops, of statistical factors like $N(E_q-E_Q)$ which yield non-vanishing contributions in the zero temperature limit.
These add up with the usual $T=0$ contributions that might be naively expected (see Eqs. \eqref{e58nv}, \eqref{e59nv} and \eqref{e510nv}, respectively).
This combination is necessary in order to get a result consistent with unitarity, which corresponds to a forward physical tree scattering amplitude.
Using the above properties, a simple generalization of the Eq. \eqref{barton1} (for bosonic fields) at two loop order may be written in the form
\begin{eqnarray}\label{e61}
\Gamma(k^1,\dots ,k^L,T) &=& \int\frac{d^{D-1} Q}{(2\pi)^{D-1}}\frac{1}{2 E_Q}
\int\frac{d^{D-1} q}{(2\pi)^{D-1}}\frac{1}{2 E_q}
\left[
{\cal A}_0(k^1,\dots , k^L; Q,q)
\right. \nonumber \\ && \left.
+N(E_Q) {\cal A}_1(k^1,\dots , k^L; Q,q)
+N(E_Q)N(E_q) {\cal A}_2(k^1,\dots , k^L; Q,q)
\right]_{Q_0=E_Q; q_0=E_q} ,
\end{eqnarray}
where ${\cal A}_0$ is related to a forward scattering amplitude of two on-shell particles $Q$ and $q$ at zero temperature. The contribution linear in the statistical factor $N(E_Q)$ describes a forward scattering amplitude of a thermal on-shell particle $Q$, and of an on-shell particle $q$ at zero temperature. The term quadratic in the statistical factors represents a forward scattering amplitude of two on-shell thermal particles.
Making in Eq. \eqref{e61} appropriate analytic continuations of the external energies \cite{Evans:1991ky},
one can verify that we get for planar 1PI retarded two and three point Green's functions at two loops, the same result as obtained in the real time formalism
\cite{Brandt:1999gf}.
The representation given in Eq. \eqref{e61} also holds for more generic planar amplitudes at two loops.
The issue of non-planar amplitudes has been examined at zero temperature in Ref. \cite{Caron-Huot:2010fvq}.
However, a proper extension of the relation \eqref{e61} in the case of non-planar retarded thermal Green's functions at two loops is more involved and requires further investigation.
\begin{acknowledgments}
{We would like to thank CNPq (Brazil) for financial support and to J. C. Taylor for a valuable correspondence.
D.~G.~C.~M. would like to thank Roger Macleod for helpful discussions
}
\end{acknowledgments}
|
\section{Introduction}
Stereoscopic projection mapping (PM) enables a user to see a three-dimensional (3D) computer-generated (CG) object floating over the physical surfaces of arbitrary shapes around them using projected imagery.
This technique can also be used to visually deform the surfaces.
These illusory effects are achieved by tracking an observer's viewpoint, rendering the perspectively correct images with proper disparity for each eye, and projecting the two images in a time-sequential manner in each frame.
The projected images are observed through active-shutter glasses consisting of liquid crystal (LC) shutters, which prevent interference between the two images for both eyes.
There are several technical advantages of using stereoscopic PM over other types of augmented reality (AR) displays such as optical see-through (OST) and video see-through (VST) head-mounted displays (HMD).
First, the field-of-view (FOV) can be made as wide as possible by increasing the number of projectors to cover the entire environment.
Second, the active-shutter glasses used are normally much lighter, and, thus, their physical burden is less than HMDs.
Third, multiple users can share the same AR experience if their viewpoints are sufficiently close to each other.
Thanks to these advantages, researchers have found stereoscopic PM to be suitable for a wide range of applications, including but not limited to museum guides \cite{SCHMIDT20191}, architecture planning \cite{1544657}, product design \cite{8797923}, medical training \cite{kijima07}, shape-changing interfaces~\cite{10.1145/2858036.2858457}, and teleconferencing \cite{10.1145/2207676.2207704}.
In many of these applications, it is important to create a faithful impression of the 3D structure of the projected object or scene.
However, like most commercial virtual reality (VR) and AR displays, current stereoscopic PM only satisfies binocular cues and is not capable of providing the correct focus cues, which causes a problem called the vergence--accommodation conflict (VAC).
Vergence refers to the simultaneous movement of both eyes in opposite directions to obtain single binocular vision, while accommodation refers to the process by which each eye changes its optical power to focus on an object as its distance varies.
The uncoupling of the vergence and accommodation causes significant discomfort, fatigue, and a distorted 3D perception for the observer \cite{Hoffman2008}.
While many approaches have been proposed to mitigate the VAC in VR/AR displays (see Sect.~\ref{sec:relatedwork}), to the best of our knowledge, no technical solution has yet been presented for the VAC in stereoscopic PM.
We propose a multifocal approach to improve the mismatch between vergence and accommodation in stereoscopic PM.
Our primary technical contribution is to attach electrically focus-tunable lenses (ETL) to active-shutter glasses to control both vergence and accommodation.
Specifically, we apply fast and periodical focal sweeps to the ETLs, which causes the ``virtual image'' (as an optical term) of every part of the real scene seen through the ETLs to move back and forth during each sweep period.
In each frame, we first render the perspectively correct images of a 3D CG object for both eyes.
Then, we project them from a synchronized high-speed projector at the exact moment that the virtual image of the projected imagery on a real surface is located at a desired distance from the ETLs.
Thus, each part of the surface is projected for a short period of time during each sweep period.
The frequency of the focal sweep is higher than the critical flicker fusion frequency (CFF) so that the observer perceives the time integral of the scene appearances during the sweep period.
This paper introduces the technical details of our computational algorithms, which are jointly designed with the optics and hardware.
Particularly, we solve three technical issues that are unique to stereoscopic PM: (1) The 3D CG object is displayed on non-planar and even moving surfaces; (2) the physical surfaces that are also seen through the ETLs need to be shown without the focus modulation; (3) the LC shutter glasses also need to be synchronized with the ETLs and the projector to provide the correct binocular cues as well as the focus cues.
On the other hand, in other AR/VR displays, (1) the 3D CG object is displayed on a planar and fixed display panel, (2) observers do not see a real scene through ETLs, and (3) shutter glasses are not used.
We also develop a novel compensation technique to address the ``lens breathing'' artifact, which varies the retinal size of the virtual image through focal length modulation.
Using a proof-of-concept prototype, we demonstrate that our technique can present the virtual image of a target 3D CG object with the perspectively correct appearance at the correct depth.
Finally, we validate the advantage provided by our technique with regard to the mitigation of the VAC by comparing it with a conventional stereoscopic PM using a user study on a depth-matching task.
To summarize, our primary contributions are that we
\begin{itemize}
\item introduce a multifocal stereoscopic PM technique that mitigates the VAC by combining fast and periodical focal sweeps with active-shutter glasses and synchronizing high-speed projection,
\item jointly design optics, hardware, and computational algorithms to present a 3D CG object with correct binocular and focus cues, solve three technical issues unique to stereoscopic PM, and propose a general computational solution for the lens breathing artifacts,
\item implement a proof-of-concept prototype that displays an object's virtual image that is perspectively correct and located at a desired distance from the observer, and
\item demonstrate the VAC mitigation achieved by the proposed system using a user study on a depth-matching task.
\end{itemize}
\section{Related Work}
\label{sec:relatedwork}
\subsection{Stereoscopic PM}\label{subsec:rel_pm}
Since the pioneering work conducted on rendering perspectively correct images on non-planar physical surfaces using projected imagery~\cite{10.1145/280814.280861}, much research has been carried out on stereoscopic PM.
Due to several advantages, such as the wider FOV and fewer physical burdens than other types of VR/AR displays, stereoscopic PM has been applied in various fields.
Telepresence has been considered a particularly suitable application of stereoscopic PM.
The 3D shape and appearance of a distant person are captured using an RGB-D camera and displayed on local physical surfaces using projected imagery to make it appear as the distant person exists in the same room~\cite{10.1145/2818048.2819965,10.1145/2642918.2647402} or behind a wall~\cite{10.1145/280814.280861,10.1145/2207676.2207704}.
Stereoscopic PM has also been applied in architecture planning \cite{1544657,egsh.20171001}, where new structures of a building are displayed on a wall of the building.
In the car industry, it was found that stereoscopic PM has the potential to assist in interior designing such that designers can examine the shape of a dashboard without needing to fabricate physical mock-ups \cite{8797923}.
Stereoscopic PM has also been applied in medical training~\cite{kijima07}, where human organs inside a body are projected on the body.
By focusing on expanding the application fields, previous works have not sufficiently considered the VAC, which, thus, remains an unsolvable and challenging technical issue in stereoscopic PM.
Various types of human perceptual properties in stereoscopic PM have been investigated.
For example, Okutani et al. found that stereoscopic PM induces stereoscopic capture where the depth observed during stereoscopic PM is automatically attributed to the texture of a physical surface even when the texture conveys no disparity information~\cite{8491263}.
Other researchers found that pseudo-haptic illusions occur in stereoscopic PM, where a user can feel bumps on a physically flat surface on which apparent bumps are overlaid\cite{8417321}.
Depth perception in stereoscopic PM has been more intensively analyzed than the other properties.
Certain researchers tested the depth estimation ability of human participants given the binocular disparity in stereoscopic PM~\cite{egve.20171354,10.1145/3359996.3364245}, while others did it with additional apparent cues such as color and texture~\cite{Schmidt2020}.
However, due to the lack of a stereoscopic PM system that satisfies both binocular and focus cues, how the depth estimation accuracy is affected when the VAC is mitigated has not yet been investigated.
\subsection{Mitigating the VAC}\label{subsec:rw_vac}
In VR and AR applications, CG images need to create a faithful impression of the 3D structure of the depicted object or scene.
Unfortunately, commercially available 3D displays often yield perceptual distortions in the 3D structure due to the mismatch between binocular and focus cues.
The VAC causes viewer fatigue and discomfort as well as distortions in the perceived depth~\cite{WANN19952731,Hoffman2008}.
Much research has been carried out in an attempt to solve the VAC issue in both VR and AR displays, which include varifocal displays (VR~\cite{Padmanaban2183,10.1145/2858036.2858140,Johnson:16,10.1145/3072959.3073622} and AR~\cite{10.1145/3272127.3275069,10.1145/3130800.3130892,8458263,8794584,Chen:15,7829412}), multifocal displays (VR~\cite{10.1145/3272127.3275015,10.1145/2766909,10.1145/1015706.1015804,10.1145/3072959.3073590,10.1145/3386569.3392424,6712903,Love:09,10.1145/3130800.3130846,Rolland:00} and AR~\cite{8456852,4637321,9284672,Llull:15,Chen:19,https://doi.org/10.1002/jsid.739,Basak:19}
), light field displays (VR~\cite{10.1145/2508363.2508366,10.1145/2766922,10.1145/2601097.2601144} and AR~\cite{6671761,10.1145/2601097.2601141}), holographic displays (VR~\cite{10.1145/3414685.3417846,10.1145/3414685.3417802,10.1145/3355089.3356539,10.1145/3386569.3392416} and AR~\cite{10.1145/3072959.3073624,10.1145/3272127.3275069}
), and \revise{other display techniques, such as those that use color information~\cite{Aksit:14,10.1145/3130800.3130815}}.
Unfortunately, despite the substantial effort that has been put into solving the VAC for stereoscopic flat-panel displays and HMDs, researchers have not focused on the issue for stereoscopic PM.
Previous works closely related to stereoscopic PM are associated with light field \cite{10.1145/2601097.2601144} and tomographic \cite{10.1145/3355089.3356577} projection systems.
These technologies have been developed for conventional flat projection screens and, thus, are not directly applied to PM, where the projection target includes potentially non-planar and even moving surfaces.
Furthermore, these works required customized, special projection devices, which potentially hinder developers from using them in the actual fields of application.
Among the four display technologies mentioned above, light field and holographic displays are not suitable for stereoscopic PM.
Since it is not assumed to install tailored optical elements in the environment in general PM applications, light field and holographic displays---which, for instance, need to embed an angle-expanding material~\cite{10.1145/2601097.2601144} and a phase modulator~\cite{10.1145/3072959.3073624} onto physical surfaces---are not generally applicable for stereoscopic PM.
Varifocal and multifocal displays basically share the same optical display architecture---both apply focus-tunable near-eye optics for displaying 3D CG objects with correct focus cues.
The major difference between the two displays is that varifocal displays actively track the accommodation of the eye in real-time and provide a correct image at the interested depth, while multifocal displays passively offer multiple depths to simulate a 3D scene without considering where the eye focuses~\cite{Zhan2020}.
\revise{In stereoscopic PM, the real scene is also seen through the focus-tunable near-eye optics because a 3D CG object is projected directly onto a surface in the scene. Therefore, when a varifocal approach is employed, a user needs to observe the real scene with a focal setting that is optimized for the projected 3D CG object. This focus modulation can possibly create an unnatural view of the real scene.}
In addition, the projection surface can be non-planar.
In this case, the projected 3D CG object is not positioned at the same distance from the observer, and, thus, a spatially-varying focal control is required to provide the correct focus cues.
However, a varifocal approach assumes a spatially uniform operation.
Furthermore, it is not preferable to add an eye-tracking sensor system to the active-shutter glasses because it degrades the advantage of stereoscopic PM (i.e., it only requires a lightweight eye wear) over the other VR/AR displays.
Therefore, we adopt the multifocal approach in this research to mitigate the VAC in stereoscopic PM.
User studies have revealed that the observer's depth estimation performance can be significantly improved by mitigating the VAC in both VR~\cite{Padmanaban2183,10.1145/2858036.2858140,Johnson:16,10.1145/3072959.3073622} and AR~\cite{7164348,8462792} displays.
Therefore, this improvement could happen in stereoscopic PM as well.
However, as mentioned in Sect.~\ref{subsec:rel_pm}, there is no experimental evidence of this due to the lack of a PM system satisfying both binocular and focus cues.
\subsection{Our contributions}
This paper presents the first stereoscopic PM system that mitigates the VAC by applying a multifocal approach.
On top of the conventional stereoscopic PM framework, we apply additional optics (ETLs) and computational algorithms to achieve this goal.
In particular, we describe how to solve the technical issues that are unique to stereoscopic PM, one of which is that the display surface is potentially non-planar and moving.
An advantage of our technique is that it does not need customized, special projection devices and works with off-the-shelf projectors.
We also conduct the first user study to investigate how the VAC mitigation improves the depth estimation accuracy in stereoscopic PM using a prototype system.
\section{Method}
\subsection{Virtual image of projected imagery seen through ETL}\label{subsec:method_vi}
\begin{figure}[t]
\centering
\includegraphics[width=0.98\linewidth]{figure/figure-visual-image-and-sweep-range-2}
\caption{(Left) The virtual image (solid star) of a projected object (dashed star) on a physical surface observed through an ETL. (a) The virtual image is located at $d_v$ from the ETL when the projected image is at $d_p$, and the ETL's optical power is $v(d_p,d_v)$. (b)(c) When the optical power is periodically swept between $v_l$ and $v_h$, the virtual image also periodically moves between the near end $d_{v,n}$ and the far end $d_{v,f}$. (Right) The required sweep range of an ETL's optical power when $d_p$ is changed under the condition that $d_{v,n}$ and $d_{v,f}$ are fixed at $\frac{1}{6}$ m and $\frac{1}{0.4}$ m, respectively.}
\label{fig:sweep-condition-concept}
\end{figure}
\revise{We attach an ETL each to the left and right of LC shutter of the active-shutter glasses.}
As a result, a user can observe the virtual image of the projected imagery on a physical surface through the ETL.
Suppose that the distance of the surface from the ETL is $d_p$ (Fig. \ref{fig:sweep-condition-concept}a).
Using the thin lens equation, we can determine the optical power $v(d_p,d_v)$ to locate the virtual image of a 3D CG object projected on the surface at a desired distance $d_v$ from the ETL as follows:
\begin{equation}\label{eq:thin_lens}
v(d_p,d_v)=\frac{1}{d_p}-\frac{1}{d_v}.
\end{equation}
The optical power is the reciprocal of the focal length (m) and is generally represented in diopter (D).
In this paper, we do not fix $d_p$ because the surface can be non-planar and can move in stereoscopic PM.
Therefore, we place the virtual image of a projected 3D CG object at a desired distance by adjusting the optical power $v(d_p,d_v)$ according to both $d_p$ and $d_v$.
We apply a fast focal sweep to the ETL, where its optical power is periodically modulated between $v_l$ and $v_h$ ($v_l<v_h$) at $f$ Hz such that $f$ is higher than the CFF.
When the physical surface is lit by normal room lighting, the virtual image of the surface periodically moves between the near end (when the optical power is $v_l$) and the far end (when it is $v_h$) along the ETL's optical axis.
In our system, we turn off the room light and project the desired 3D CG object onto the surface when the optical power of the ETL becomes $v(d_p,d_v)$ for each focal sweep period.
We can locate the virtual image at the desired distance $d_v$ under the condition that the optical power is within the range of $v_l\le v(d_p,d_v)\le v_h$.
Given the desired location range of the virtual image of a projected 3D CG object and the position of a projection surface, the sweep range required to place the virtual image at the desired location is determined.
Suppose that the distance from the near end of the virtual image to the surface is $d_{v,n}$ and that from the far end to the surface is $d_{v,f}$ (Figs. \ref{fig:sweep-condition-concept}b and \ref{fig:sweep-condition-concept}c), the relationship between $v_l$ and $d_{v,n}$ and that between $v_h$ and $d_{v,f}$ can be formulated as follows according to the thin lens equation:
\begin{equation}\label{eq:sweep-range}
v_l=\frac{1}{d_p}-\frac{1}{d_p-d_{v,n}},\ v_h=\frac{1}{d_p}-\frac{1}{d_p+d_{v,f}}.
\end{equation}
The right-hand side graph in Fig.~\ref{fig:sweep-condition-concept} shows the required sweep range computed by Eq.~\ref{eq:sweep-range} when $d_p$ is changed from $\frac{1}{3}$ m to $\frac{1}{0.2}$ m under the condition that $d_{v,n}$ and $d_{v,f}$ are fixed at $\frac{1}{6}$ m and $\frac{1}{0.4}$ m, respectively.
The graph indicates that the required sweep range becomes larger when the distance between the ETL and the surface is shorter.
Therefore, we determine the sweep range from the desired location range of the virtual image of a 3D CG object projected onto the nearest surface from the ETL.
\begin{figure}[t]
\centering
\includegraphics[width=0.98\linewidth]{figure/fig-3-4-5}
\caption{Depth filtering technique. (Left) A depicted example showing the radiance distribution between the two nearest projector frames and (right) radiance distribution in linear proportion with distance measured in diopters.}
\label{fig:system-requirement}
\end{figure}
In an actual setup, the projector can display only a limited number of images ($N$) during a sweep period.
Therefore, it is not always possible to project a 3D CG object when the optical power of the ETL is exactly $v(d_p,d_v)$.
To overcome this constraint, we sample $N'$ ($\le N$) optical powers $v_n$ ($n=1,\cdots,N'$) within the range between $v_l$ and $v_h$ such that the projector can display an image at each $v_n$.
Then, a 3D CG object is projected at the $\hat{n}$-th optical power in each sweep period when $v_{\hat{n}}$ is the closest to $v(d_p,d_v)$; thus,
\begin{equation}\label{eq:divide}
\hat{n}=\mathop{\rm arg~min}\limits_n |v_n-v(d_p,d_v)|.
\end{equation}
The sampling of the optical power causes a discontinuity in the virtual image of a displayed 3D CG object when the optical power $v(d_p,d_v)$ is spatially varying and consequently spans two sampled optical powers (e.g., $v_{\hat{n}}$ and $v_{\hat{n}+1}$) (Fig. \ref{fig:system-requirement}).
To alleviate this discontinuity, we apply the ``depth filtering'' technique introduced in previous work~\cite{10.1145/1015706.1015804}.
Depth filtering distributes the radiance between the two nearest projector frames in linear proportion, with distance measured in diopters.
Suppose $r_{\hat{n}}$ and $r_{\hat{n}+1}$ are the distributed radiance of the original radiance $r$ to the $\hat{n}$-th and $(\hat{n}+1)$-th projection images, respectively; they would be computed as follows:
\begin{equation}\label{eq:depthf}
r_{\hat{n}}=r(1-\frac{v(d_p,d_v)-v_{\hat{n}}}{v_{\hat{n}+1}-v_{\hat{n}}}),\ r_{\hat{n}+1}=r-r_{\hat{n}}.
\end{equation}
A previous user study showed that depth filtering could provide continuous focus cues \cite{MacKenzie2010}.
Another work recommended that the sampling interval of the optical power should be less than 0.6 D to provide perceptually continuous focus cues~\cite{10.1117/12.908883}.
We follow this guideline when we build our prototype.
\subsection{Synchronization}\label{subsec:method_sync}
To display the virtual image of a 3D CG object at the desired location, we need to synchronize the projector, ETLs, and shutter glasses.
We determine our synchronization strategy while considering the following temporal response properties of the devices:
\begin{description}
\item[Projector:] We employ a state-of-the-art off-the-shelf high-speed projector~\cite{watanabe2015high}. It projects an image at $\delta t_p$ ms after a trigger signal is sent from the PC. It is also known that the delay does not vary over time. Therefore, once we measure $\delta t_p$ in the calibration phase, we can project images at the desired timings by sending the triggers at $\delta t_p$ prior to the projection. There is also a delay when transferring an image from the PC to the projector ($\sim$3 ms). We apply a multi-threading pipeline to perform the projection and transfer the images in parallel, and, thus, the delay of the latter process does not have to be considered during synchronization.
\item[ETLs:] We use polymer-based ETLs in which the lens consists of an optical fluid sealed off by an elastic polymer membrane. An electromagnetic actuator ring exerts pressure on the outer zone of the lens, which changes its curvature, i.e., the optical power. Polymer-based ETLs can achieve a faster focal change than other ETL types while maintaining a relatively large aperture size. As previous focal sweep systems~\cite{8456852,10.1145/3272127.3275015,10.1145/3072959.3073594,10.1145/3355089.3356577} have shown, the optical power of an ETL can be smoothly modulated when a sinusoidal or triangular wave is applied as the electrical current. Therefore, we apply a sinusoidal wave as the input in our system. It was also experimentally validated that the temporal response of an ETL is a linear time-invariant (LTI) system~\cite{Iwai2019}. Therefore, when we supply a sinusoidal wave $i[\phi]$ to an ETL, the output optical power also becomes a periodical wave $v[\phi]$, where $\phi$ is the phase of the input wave signal ($0\le\phi<2\pi$). Once we measure one cycle of $v[\phi]$ in the calibration phase, we can estimate the optical power at any timing required.
\item[Shutter glasses:] We use LC-based active-shutter glasses. The LC is transparent (open) in the initial state and becomes opaque (closed) when a voltage is applied. There are two delays to be considered. First, a delay $\delta t_{o2c}$ occurs between the moment we apply the voltage to change the LC state from open to closed till the moment the LC shutter actually closes. Second, there is a longer delay $\delta t_{c2o}$ that occurs when it shifts from the closed to the open states as well. Therefore, we measure these delays during the calibration phase and use them for synchronization.
\end{description}
\begin{figure}[t]
\centering
\includegraphics[width=0.98\linewidth]{figure/synchronization.pdf}
\caption{Timing chart of the proposed synchronization strategy. In this example, the number of the sampled ETL's optical power is three (i.e., $N'=3$), at which the high-speed projector projects images.}
\label{fig:timing-chart}
\end{figure}
Figure~\ref{fig:timing-chart} shows the timing chart of our synchronization strategy.
First, the ETLs for the left and right eyes are modulated by supplying them with the same sinusoidal waves $i[\phi]$ of frequency $f$ Hz.
We divide the optical power waveform into two parts in each period and use the upward curve (from $v_l$ to $v_h$) for the left eye and the downward curve (from $v_h$ to $v_l$) for the right eye.
The projector projects images when the optical power $v[\phi]$ of each ETL becomes $v_n$.
To make this happen, we send a trigger signal to the projector when the phase of the ETL's input signal is $2\pi f\delta t_p$ prior to the target phase.
We open the LC of the left eye and close the LC of the right when the ETLs' optical powers form the upward curves.
In the same way, we open the LC of the right eye and close the LC of the left when the optical powers form the downward curves.
We control the LC shutters such that the optical powers become equal to the maximum and minimum values during the closing/opening transition time of the LC shutters because the time derivative of the modulated optical power of each ETL becomes smallest at these timings.
Stereoscopic PM applications generally project a white color onto physical surfaces to make them visible in addition to the projected 3D CG object to enable AR experiences.
In our system, the physical surfaces under white illumination are also observed through the ETL.
To avoid a situation in which the locations of the virtual images of these surfaces vary according to the focal sweep, we illuminate them when the optical powers of the ETLs are zero (i.e., $v[\phi]=0$).
This allows us to present the physical surfaces without any focus modulations.
\subsection{Lens breathing compensation}\label{subsec:method_comp}
The virtual image seen through the ETL is scaled using the change in its optical power because an observer's eye is not co-located with the ETL but, instead, placed behind it.
When a projected 3D CG object is divided into multiple parts that are observed through the ETL with different optical powers, the lens breathing phenomenon causes an overlap or gap between their virtual images on the retina.
This artifact has been noticed in previous works~\cite{10.1145/3072959.3073622,10.1145/3355089.3356577,10.1145/2766909}, in which it was solved by manually adjusting the size of displayed images.
In this paper, we propose a simple yet general computational solution for this issue.
\begin{figure*}[t]
\centering
\includegraphics[width=0.98\linewidth]{figure/lens-breathing-compensation.pdf}
\caption{Lens breathing artifact and compensation. When the inner part of a projected object is observed with a (a) positive or (b) negative optical power, an (a) overlap or (b) gap occurs between its virtual image and that of the outer part observed with the optical power of zero. Note that \textbf{\textcolor{red}{---}} indicates an observed real image and the virtual image of \textbf{\textcolor[rgb]{0.5,0.5,0.5}{---}}. (c) The proposed compensation technique resizes the projected object to \textbf{\textcolor{blue}{---}}.}
\label{fig:lens-breathing-compensation}
\end{figure*}
Without loss of generality, we suppose simple situations where a 3D CG object is projected onto a flat surface that is located $d_p$ away from an ETL (Figs.~\ref{fig:lens-breathing-compensation}(a1) and \ref{fig:lens-breathing-compensation}(b1)).
The object is divided into two parts that are observed with different optical powers.
When the optical power becomes equal to a certain non-zero value, the inner part of the object with respect to the ETL's optical axis is projected onto the surface.
When the optical power becomes zero, the outer part is projected.
If the height of the inner part on the surface is $h_p$, then that of its virtual image $h_v$ is computed as $h_p\frac{d_v}{d_p}$ according to geometric similarity, where $d_v$ is the distance from the ETL to the virtual image.
Because the outer part is observed when the ETL's optical power is zero, its real image is observed.
Now, we suppose that an observer's eye is located at $d_e$ behind the ETL (Figs.~\ref{fig:lens-breathing-compensation}(a2) and \ref{fig:lens-breathing-compensation}(b2)).
Regarding the virtual image of the inner part, we denote the visual angle of its far end from the optical axis as $\theta_{i,f}$.
We also denote the visual angle of the near end of the outer part as $\theta_{o,n}$.
Figures~\ref{fig:lens-breathing-compensation}(a2) and \ref{fig:lens-breathing-compensation}(b2) depict the relationship between $\theta_{i,f}$ and $\theta_{o,n}$ for both cases where the non-zero optical power is positive and negative, respectively.
In general, when a larger optical power is applied to the inner part of a 3D CG object than the outer part, $\theta_{i,f}>\theta_{o,n}$, and, consequently, the retinal images of the inner and outer parts overlap.
On the contrary, when a smaller optical power is applied to the inner part, $\theta_{i,f}<\theta_{o,n}$, and a gap occurs between the two retinal images.
As discussed above, in a case where a 3D CG object is divided into multiple parts that are observed with different optical powers, an overlap or gap of their virtual images occurs on the observer's retina.
We compensate for these artifacts by resizing each part of the projected object.
The following method assumes that each part of the object is projected onto a planar surface that is perpendicular to the ETL's optical axis.
If a part of the object that is observed with the same optical power covers a non-planar surface, we further divide it into smaller parts until each surface area onto which the subdivided part is projected can be regarded as a plane that is perpendicular to the optical axis.
Suppose that $\theta_{p,n}$ is the visual angle of the near end of a part of a 3D CG object on the projection surface, and $\theta_{p,f}$ is that of its far end in a case where the object is not seen through an ETL (Fig.~\ref{fig:lens-breathing-compensation}(c1)).
Then, our method aims to resize the object so that the visual angles of the far and near ends of its virtual image correspond to $\theta_{p,n}$ and $\theta_{p,f}$, respectively.
Let the height of the near end and that of the far end be $h_{p,n}$ and $h_{p,f}$, respectively.
To achieve our goal, the heights of the near and far ends of its virtual image ($h_{v,n}$ and $h_{v,f}$, respectively) should be $h_{v,*}=\frac{d_v+d_e}{d_p+d_e}h_{p,*}$ according to geometric similarity, where $*$ represents either $n$ or $f$.
The ETL represents this virtual image as the projection of the displayed 3D CG object from the ETL's optical center onto a plane located at $d_v$ away from the ETL.
Therefore, to show this virtual image, the heights of the near and far ends of the 3D CG object on the surface should be $h'_{p,*}=\frac{d_p}{d_v}h_{v,*}$ according to geometric similarity (Fig.~\ref{fig:lens-breathing-compensation}(c2)).
Consequently, we can show the virtual image of the 3D CG object such that the visual angles of its far and near ends correspond to $\theta_{p,*}$ by resizing the object using the following equation:
\begin{equation}\label{eq:resize}
h'_{p,*}=\frac{d_p(d_v+d_e)}{d_v(d_p+d_e)}h_{p,*}.
\end{equation}
An important consequence that is indicated by Eq.~\ref{eq:resize} is that the resizing factor is constant for the entire area of a subdivided part of the 3D CG object.
\subsection{Rendering pipeline}
We develop a new rendering pipeline for the projection images based on the conventional two-pass rendering technique~\cite{10.1145/280814.280861}.
The original algorithm consists of off-screen and on-screen rendering passes.
It requires the positions of an observer and a projector as well as the shapes of the projection surfaces and the 3D CG object to be rendered.
Next, (1) the off-screen rendering is performed to generate the desired appearance of the 3D object \revise{and its depth map} for the observer's viewpoint.
(2) The rendered image is then projected onto the surface from the observer's viewpoint in the virtual space.
(3) After that, the on-screen rendering is performed, in which the newly textured surface is captured from the projector's viewpoint.
(4) Finally, the captured image is projected in the physical space.
We incorporate the depth filtering and lens breathing compensation into the original two-pass rendering algorithm.
\begin{figure}[t]
\centering
\includegraphics[width=0.98\linewidth]{figure/fig-rendering-pipeline-4}
\caption{Generated images in the proposed rendering pipeline, where the target 3D CG objects were two cuboids of different heights and the physical projection surface was a corner-shaped screen.}
\label{fig:rendering-pipeline}
\end{figure}
In our pipeline, the depth filtering is performed in the first process of the original algorithm.
\revise{In addition to the rendering of the 3D CG object and its depth map, we obtain the depth map of the projection surfaces for the observer's viewpoint (the first row of Fig.~\ref{fig:rendering-pipeline}). Note that stereoscopic PM systems generally assume that the shape of a projection surface is obtained in advance and its pose is tracked online.}
We divide the rendered image into several parts using Eq.~\ref{eq:divide} such that each part is projected when the ETL's optical power becomes equal to a certain sampled value.
As described in Sect.~\ref{subsec:method_vi}, the number of sampled optical powers is $N'$, and, thus, we render $N'$ images from the above process.
We then apply depth filtering (Eq.~\ref{eq:depthf}) to these images (the second row of Fig.~\ref{fig:rendering-pipeline}).
Next, the lens breathing compensation is performed in the second process of the original two-pass rendering algorithm, in which we resize the images that are rendered in the first process.
We develop an efficient technique for the implementation of the resizing principle described in Sect.~\ref{subsec:method_comp}.
In particular, we found that the resizing can be performed just by modifying the FOV of the virtual projector in the second process because (1) the resizing is performed with respect to the optical axis of the eye, and (2) the same resizing factor is applied to each divided part of the 3D CG object (Eq.~\ref{eq:resize}).
Specifically, as shown in Fig.~\ref{fig:lens-breathing-compensation}(c2), the FOVs before and after modification are as follows:
\begin{equation}
\theta_{p,f}=\arctan\frac{h_{p,f}}{d_e+d_p},\
\theta'_{p,f}=\arctan\frac{h'_{p,f}}{d_e+d_p}.
\end{equation}
Therefore, we modify the FOV of the virtual projector by multiplying it by the factor of $\frac{\theta'_{p,f}}{\theta_{p,f}}$ when we project the rendered image in the second process of our two-pass rendering pipeline.
The third row of Fig.~\ref{fig:rendering-pipeline} shows an example of a rendered image after the lens breathing compensation has been applied.
\section{Experiment}
\subsection{Experimental setup}\label{subsec:exp_setup}
We constructed a prototype system consisting of a pair of LC shutters, a pair of ETLs, and a synchronized high-speed projector (Fig.~\ref{fig:teaser}a).
We assembled our active-shutter glasses by attaching each ETL (Optotune AG, EL-16-40-TC, aperture: 16 mm, optical power range: -10 D to 10 D) onto an LC shutter taken from off-the-shelf active-shutter glasses (RV-3DGDLP1, field rate: 96-144 Hz) and inserted a pair of the combined device into an eyeglass frame fabricated using an FDM 3D printer to form a wearable (69$\times$128$\times$67 mm, 200 g) device.
The digital signal (sinusoidal wave) to control the ETLs was generated by a workstation (CPU: Intel Xeon E3-1225 v5 3.30GHz, RAM: 32GB, GPU: NVIDIA Quadro K620), input into a D/A converter (National Instruments, USB-6211), and converted into an analog voltage.
This voltage was then converted to an electric current through a custom amplifier circuit using an op-amp (LM675T).
Finally, the current was fed to the ETLs.
The LC shutters were controlled using a circuit consisting of a motor driver (TOSHIBA, TA7291P).
We employed an off-the-shelf high-speed projector (Inrevium, TB-UK-DYNAFLASH, 1024$\times$768 pixels, 330 ANSI lumen) that can project 8-bit grayscale images at 2,000 frames per second (fps)\footnote{Although the frame rate on the specification sheet is 1,000 fps, we experimentally discovered that the projector could project grayscale images at 2,000 fps.}.
The projection images were generated by the workstation and sent to the projector via a PCI Express interface.
\revise{We manually adjusted the projector location so that the projected result always appeared focused on a projection surface. We manually measured the surface's shape and its pose relative to the ETLs offline. In Sect.~\ref{subsec:exp_mov}, a surface was tracked using a motion capture system consisting of five cameras (NaturalPoint, OptiTrack Prime 17W). The shape and pose information was used to obtain the depth map in our rendering pipeline.}
\begin{figure}[t]
\centering
\includegraphics[width=0.93\linewidth]{figure/figure-system-architecture-10}
\caption{System architecture.}
\label{fig:system-architecture}
\end{figure}
The workstation was able to generate a trigger signal for the projector and a control signal for the LC shutters along with the sinusoidal wave for the ETLs.
However, the D/A converter only had two output channels, which was not sufficient for sending all the signals.
Therefore, we used an additional microcontroller (Arduino Uno) to synchronize the display timing of each projection image and the opening/closing timing of the LC shutters.
In our system, the workstation sent the sinusoidal wave $i[\phi]$ and a single global trigger signal at $\phi=0$ to the D/A converter.
The trigger signal was then transferred to the microcontroller, which generated the trigger and the control signal using the delays of $\delta t_p$, $\delta t_{c2o}$, and $\delta t_{o2c}$, and sent to the projector and the controlling circuit of the LC shutters.
The system architecture is depicted in Fig.~\ref{fig:system-architecture}.
\subsubsection{System calibration}
\begin{figure}[t]
\centering
\includegraphics[width=0.98\hsize]{figure/figure-measured-optical-power-diote-3}
\caption{Optical power measurement: (left) experimental setup, (right) one cycle of input sinusoidal wave (orange line), and the resultant optical power of an ETL (blue line).}
\label{fig:etl-calibration}
\end{figure}
The system calibration was conducted as follows.
We built an optical power measurement system using a photodiode and a laser emitter (wavelength: 635 nm, output power: 0.9 mW) (Fig.~\ref{fig:etl-calibration}).
We placed the ETL at about 400 mm away from the laser emitter.
The photodiode was placed about 4 mm behind the ETL, and it measured the laser beam passing through the ETL.
Raising the optical power of the ETL reduced the laser's spot size on the photodiode (i.e., increased the laser power density), enhancing the electric current created in the photodiode.
We indirectly measured the ETL's optical power in this manner.
For the calibration of the ETL, we input a sinusoidal wave $i[\phi]$ to the ETL and measured one cycle of its output optical power $v[\phi]$.
The frequency of the wave was maintained at 60 Hz throughout the experiment, which was higher than the CFF.
According to the discussion in Sect.~\ref{subsec:method_vi}, the sweep range was determined based on the desired location range of the virtual images (i.e., $d_{v,n}$ and $d_{v,f}$) on the nearest physical surface from the ETL.
\revise{In the following experiments, except for the user study in Sect.~\ref{subsec:exp_user}, we considered the nearest distance as 500 mm, and $d_{v,n}=$ 167 mm and $d_{v,f}=$ $\infty$ mm. Therefore, the required sweep range was between -1.0 D and 2.0 D according to Eq.~\ref{eq:thin_lens}. We can still modulate the location of the virtual image of a projected 3D CG object onto a surface whose distance from the ETL is shorter than 500 mm. But in such a case, the modulation range becomes narrower than that between $d_{v,n}$ and $d_{v,f}$.}
We manually searched for the offset and amplitude of the input sinusoidal wave by which the output optical power of the ETL can cover the target range and found the suitable values (offset: 2.5 mV, amplitude: 28 mV).
Figure~\ref{fig:etl-calibration} shows the measured optical power $v[\phi]$ by the input wave.
From the figure, we found that the output values did not form a clean sinusoidal wave.
Therefore, we recorded one cycle of the output wave along with the corresponding input wave and stored it on the microcontroller to be able to look up the optical power at a given phase of the input wave.
To synchronize the ETLs and the high-speed projector, we used the same photodiode to measure the delay for the high-speed projector between a trigger signal being sent from the microcontroller and the actual projection.
As a result, we found that the delay was 0.15 ms ($=\delta t_p$).
We also measured the delay for the LC shutters and found that the closing transition took 0.1 ms ($=\delta t_{o2c}$), while the opening transition took 3.0 ms ($=\delta t_{c2o}$).
This delay information was also stored in the microcontroller.
The microcontroller received the global trigger signal, computed when the trigger and control signals should be sent to the projector and LC shutters based on the recorded optical power waveform and the delay information, and, finally, sent the signals to the devices.
\subsection{Lens breathing compensation}\label{subsec:exp_comp}
We verified whether the proposed lens breathing compensation improved the gap or overlap between adjacent areas of a projected 3D CG object, which were observed with different optical powers.
The ETL was attached to a camera (Sony $\alpha$7S) with a suitable lens (Sony SEL24F14GM, focal length: 24 mm) that was used to capture the experimental results with a shutter speed of 1/60 seconds.
We placed a flat screen at 500 mm away from the ETL such that it was perpendicular to the ETL's optical axis.
The displayed 3D CG object was a textured plane.
The virtual image of the object was placed such that the distance of the left edge from the ETL was 333 mm, and that from the right edge was 1666 mm.
The attached texture was a black and white checker pattern.
We projected each part of the 3D CG object when the ETL's optical power was either -1.0 D, -0.5 D, 0.0 D, 0.5 D, 1.0 D, or 1.5 D.
We captured the projected result under two conditions: when the lens breathing compensation was (1) applied and (2) not applied.
\begin{figure}[t]
\centering
\includegraphics[width=0.98\hsize]{figure/figure-lens-breathing-compensation-3}
\caption{Experimental validation of the proposed lens breathing compensation technique. The projected 3D CG object was a slanted plane with a checker pattern. (a) Experimental setup. The object's appearances without (b) and with (c) the compensation.}
\label{fig:exp-lens-breathing}
\end{figure}
Figure~\ref{fig:exp-lens-breathing} shows the results.
Comparing the results of the two conditions, we found that the appearance of the checkerboard texture significantly improved when the proposed compensation method was applied.
In particular, the unnatural overlaps of the textures degraded the image quality of the projected result without the compensation, while such artifacts were almost not visible in the result with the compensation.
Therefore, we confirmed that the proposed lens breathing compensation technique works well to alleviate the artifacts caused by the lens breathing phenomenon.
\subsection{Multifocal appearance of a projected object on a flat static surface}\label{subsec:exp_static}
We validated our technique using a flat, static screen.
As shown in Fig.~\ref{fig:teaser}b, we placed the screen 500 mm away from the ETL, which was attached to the same camera mentioned in Sect.~\ref{subsec:exp_comp}.
In addition, three physical objects were placed at different distances from the ETL, which were 333 mm (near), 500 mm (middle), and 1,000 mm (far).
We projected three Stanford bunnies on the screen such that the distance of the virtual image of each bunny from the ETL was the same as one of the physical objects.
Specifically, the near, middle, and far bunnies were projected when the ETL's optical power was -1.0 D, 0.0 D, and 1.0 D, respectively, which were computed using Eq.~\ref{eq:thin_lens}.
The near, middle, and far bunnies were displayed from the bottom to top of the screen.
The physical objects were illuminated by a uniform white image when the optical power was 0.0 D.
We captured the scene using the camera by changing its focus distance from 333 mm to 1,000 mm.
Figures~\ref{fig:teaser}c, \ref{fig:teaser}d, and \ref{fig:teaser}e show the results.
When the focus distance of the camera was 333 mm, the physical object that was placed in the near position appeared more focused than the other physical objects.
Similarly, the image of the bottom bunny, which was projected when the optical power was -1.0 D, appeared more focused than that of the other projected bunnies.
We obtained the same observations when the camera's focus distance was 500 mm and 1,000 mm as well.
Therefore, we concluded that the proposed method could provide a multifocal view in PM while presenting the physical objects without any focus modulations.
\subsection{Multifocal appearance of a projected object on a non-planar surface}\label{subsec:exp_nonp}
\begin{figure}[t]
\centering
\includegraphics[width=0.98\hsize]{figure/figure-multifocus-non-planar-3}
\caption{Multifocal appearance of a rotating Utah teapot projected onto a step surface. (a) Experimental setup. (b, c) The projected object observed through the ETL with a fixed optical power of zero. (d, e) The projected object observed using the proposed method.}
\label{fig:exp-non-planar}
\end{figure}
A unique technical challenge when attempting to mitigate the VAC in stereoscopic PM is that the display surface can potentially be non-planar.
Thus, we verified whether the proposed system could provide the correct focus cues when the projection screen is non-planar.
We placed a step surface such that the two planes were facing the ETL.
The distance of the planes from the ETL were 450 mm and 800 mm, respectively.
We projected a Utah teapot such that it covered the two planes.
\revise{We placed the projector about 2.0 m away from the surface, which caused the projected images to be focused on the both planes.}
Using our technique, the virtual image of the object was positioned such that the center of the object was at the same distance from the ETL as the front plane.
We projected each part of the object when the ETL's optical power was either -1.0 D, -0.5 D, 0.0 D, or 0.5 D.
As a baseline, we also captured the object while keeping the optical power fixed at 0.0 D.
Therefore, there were two conditions for which the scene was recorded: (1) with and (2) without our technique.
Figure~\ref{fig:exp-non-planar} shows the results captured by a camera (Ximea MQ013CG-ON) with a C-mount lens (focal length: 12 mm) to which the ETL was attached.
Comparing the results of the two conditions, we found that all parts of the teapot observed using our technique appeared focused when we adjusted the camera lens such that the focus point corresponds to the front projection plane.
On the other hand, the right side of the teapot appeared blurred in the condition where our technique was not used.
Thus, we concluded that the proposed technique could successfully realize multifocal PM for a non-planar screen.
\subsection{Multifocal appearance of a projected object on a moving surface}\label{subsec:exp_mov}
\begin{figure}[t]
\centering
\includegraphics[width=0.98\hsize]{figure/figure-multifocus-moving-surface-7}
\caption{Multifocal appearance of a Stanford bunny projected onto a flat surface freely moved by hand. (a) Experimental setup. (b, c) The projected object observed through the ETL with a fixed optical power of zero. (d, e) The projected object observed using the proposed method. \revise{Need revision}
}
\label{fig:exp-moving}
\end{figure}
Another unique technical challenge when attempting to mitigate the VAC in stereoscopic PM is that the display surface can potentially move.
Thus, we verified whether the proposed system could provide the correct focus cues when the projection screen moves in an experimental setup, shown in Fig.~\ref{fig:exp-moving}.
A flat screen was manually moved between 300 mm and 500 mm from the same camera mentioned in Sect.~\ref{subsec:exp_nonp}.
The projected object was a Stanford bunny such that the 3D position of its virtual image in the scene was fixed regardless of the screen pose.
We controlled the timing of the projection so that the bunny always appeared focused.
The bunny image was projected when the ETL's optical power was either -0.5 D, 0.0 D, 0.5 D, 1.0 D, or 1.5 D.
As a baseline, we also recorded a video of the bunny while keeping the optical power fixed at 0.0 D.
Therefore, the same two experimental conditions as discussed in Sect.~\ref{subsec:exp_nonp} were established.
As indicated in Sect.~\ref{subsec:exp_setup}, the screen pose was tracked using a motion capture system.
Figure~\ref{fig:exp-moving} shows the results that were captured at different screen poses.
First, we ensured that the size and position of the bunny were consistent throughout the video-recording process in both experimental conditions.
This indicates that our two-pass rendering pipeline worked correctly.
Comparing the results of the two conditions, we found that the bunny observed using our technique always appeared focused, while the one projected without our technique appeared blurred at certain screen poses.
Thus, we concluded that the proposed technique could successfully realize multifocal PM for a moving screen.
\subsection{User study of depth-matching}\label{subsec:exp_user}
We conducted a user study to investigate the effect of the VAC on the depth estimation accuracy of a human observer in stereoscopic PM.
Previous studies have revealed that the VAC caused distortions in the perceived depth; more specifically, people overestimated the depth of an AR target when using conventional OST displays~\cite{7164348,8462792,9284699}.
Our user study investigated the depth estimation accuracy in a depth-matching experiment that was designed based on the previous studies.
Each participant moved a physical pointer to match its depth with a virtual target displayed using stereoscopic PM.
We prepared two experimental conditions.
The first one was the ``proposed'' condition, where the VAC was mitigated by our system, and the second one was the ``conventional'' condition, where the VAC was not mitigated and only the binocular cue was correct.
Based on the results of the previous studies~\cite{7164348,8462792,9284699}, the hypotheses of our study were formulated as follows:
\begin{description}
\item[H1:] The depth was overestimated when the VAC was not mitigated in stereoscopic PM.
\item[H2:] The depth estimation was more accurate when the VAC was mitigated by our proposed system.
\end{description}
\begin{figure}[t]
\centering
\includegraphics[width=0.98\hsize]{figure/figure-user-study.pdf}
\caption{Depth-matching experiment. (a) Experimental setup where a participant manually moved a physical pointer to match its depth with a projected virtual target (checker pattern). Note that the study was conducted under a dark environment. (b) The scatter plot of the estimated against the actual distance with fitted linear equations. The red and blue plots represent the proposed and conventional conditions, respectively. (c) Errors of the estimated from the actual distance in both conditions; ``x'' indicates the mean value.}
\label{fig:exp-user-study}
\end{figure}
\subsubsection{Experimental setup and procedure}
We built an experimental system in a dark room, as shown in Fig.~\ref{fig:exp-user-study}a.
Participants sat on a chair and placed their faces on a chin rest.
We attached our active-shutter glasses consisting of the ETLs and the LC shutters to the chin rest such that the participants could see the experimental scene through our glasses.
The virtual target was a checker pattern, which was projected onto a flat screen placed 2.5 m away from the ETLs.
The visual angle of the target was fixed at 2.4 degrees throughout the experiment.
By providing the appropriate disparities to both eyes, our system presented the virtual target at five distances from the ETLs: 300 mm, 364 mm, 400 mm, 444 mm, and 500 mm.
The sweep range of the ETL's optical power in this experiment was -2.6 D to 0.0 D.
Under the proposed condition, the virtual target was projected when the ETL's optical power was -2.6 D, -2.1 D, and -1.6 D for the target distances of 300 mm, 400 mm, and 500 mm, respectively.
We applied the depth filtering technique to present the focus cues for the other two target distances.
For the conventional condition, the virtual target was projected when the optical power was zero for all the target distances.
For both conditions, the physical pointer was attached to a linear slider using which a participant could move the pointer along the direction of the ETL's optical axis.
The system illuminated the physical pointer when the ETL's optical power was zero using a synchronized light.
We recruited ten participants from a local university.
All participants were na\"{i}ve to the purpose of the experiment and had normal or corrected-to-normal vision.
We divided the participants into two groups; half of them performed depth-matching tasks under the proposed condition on the first day and did the same under the conventional condition on the second day, while the other half performed the tasks in the opposite order.
In a preparation block, we measured the interpupillary distance (IPD) for each participant and accordingly adjusted the distance between the two ETLs.
Then, we asked them to adjust their head positions by observing a calibration pattern to ensure that the disparities of the physical pointer and the virtual target were identical at the same distance.
Specifically, we projected a cross pattern onto the screen for each eye and positioned a physical object on which the same cross pattern was printed such that all the crosses corresponded to the participant's retina when their head position was correct.
After the preparation block, each participant performed the following depth-matching tasks, including a training task.
In each task, they were asked to move the physical pointer by hand until they perceived that it was located at the same depth as the virtual target.
We prepared five depths as described above, and each participant performed the depth-matching tasks four times for each depth.
Thus, there were 20 ($=5$ depth $\times$ 4 repetitions) trials in addition to the training trial for each condition.
The order of the trials was counterbalanced among the participants.
Further, the participants were allowed to take as much time as they needed to complete each task.
We told them that they could give up on a task if they could not perform the binocular fusion.
After all the tasks were completed, we asked each participant whether they perceived the flicker of the displayed object.
\subsubsection{Result}
All the participants were able to perform the binocular fusion in all the tasks and reported that they did not perceive any flickers during the experiment.
Figure~\ref{fig:exp-user-study}b shows the results plotted as a scatter plot of the estimated against the actual distance.
Following previous studies~\cite{7164348,8462792,9284699}, we calculated the linear functions that predict the estimated distance from the actual distance and used multiple regression methods to determine whether the slopes and intercepts significantly differed between the conditions.
While Fig.~\ref{fig:exp-user-study}b shows all 400 data points ($=20$ trials $\times$ 2 conditions $\times$ 10 participants), we averaged the responses for the four repetitions for each participant, which reduced the size of the dataset to 100 points for the multiple regression analysis~\cite{7164348}.
We then compared the two linear fits of the proposed and conventional conditions.
Our multiple regression analysis found that the slopes of the linear fits did not significantly differ ($F_{1,96} = 0.218, p = 0.64$), while
the intercepts did ($F_{1,97} = 19.552, p < 0.01$).
Figure~\ref{fig:exp-user-study}c illustrates the errors of the estimated distance from the actual distance, where positive errors indicate the overestimation.
The participants overestimated the depth by 22.86 mm on average in the conventional condition, while the error was much smaller (-1.93 mm) in the proposed condition.
We performed a two-way analysis of variance (ANOVA) for the actual distance of the virtual target and the experimental condition to evaluate the estimation errors.
Again, we averaged the responses of the four repetitions for each participant.
The ANOVA showed the main effect of the experimental condition ($F_{1,90}=18.234, p<0.01$), but it did not show the main effect of the actual distance of the virtual target ($F_{4,90}=0.861, p=0.49$) and the interaction effect ($F_{4,90}=0.088, p=0.99$).
Next, a post-hoc analysis was performed using Tukey's HSD method for pairwise comparison.
The result showed a statistically significant difference between the proposed and conventional conditions ($p<0.01$).
Therefore, both our hypotheses were supported by the results.
In stereoscopic PM, the participants overestimated the depth of the virtual target when the VAC was not mitigated (H1) and estimated it more accurately when it was mitigated (H2).
These trends are consistent with the previous results of OST-AR displays~\cite{7164348,8462792,9284699}.
\section{Discussion}
From the experimental results, we confirmed that our multifocal stereoscopic PM technique successfully worked for different types of projection screens.
Specifically, it could provide the correct focus cues for static (Sect.~\ref{subsec:exp_static}), non-planar (Sect.~\ref{subsec:exp_nonp}), and moving surfaces (Sect.~\ref{subsec:exp_mov}).
We also confirmed that our lens breathing compensation technique can be used to significantly reduce the associated artifacts (Sect.~\ref{subsec:exp_comp}).
Further, because it is a general solution, it can be applied to other AR/VR multifocal and varifocal displays and improve their image qualities.
Furthermore, the user study showed that the proposed technique could mitigate the VAC and significantly reduced the distortions in perceived depth in stereoscopic PM (Sect.~\ref{subsec:exp_user}).
\subsection{Limitations}
Although we achieved the goal of this research, our technique still suffers from several technical limitations.
First, the current system works properly only in a dark environment because \revise{a bright environment increases the black offset of a projected result and, consequently, degrades its contrast}.
Other PM systems also suffer from the same limitation because the environment lighting decreases the contrast of the projected imagery~\cite{10.5555/1088894}.
Previous attempts have addressed the limitation by illuminating the entire scene surface using multiple projectors, which can theoretically reproduce the original environment lighting~\cite{10.1145/2642918.2647383}.
In addition, by using large-aperture optics~\cite{8798245}, the environment lighting can be naturally reproduced.
\revise{A potential extension of our system is to apply anisotropic diffusers or retro-reflectors as projection screens~\cite{Eldes:13,10.1145/1836821.1836841}, but this, on the other hand, limits the scalability of the system.}
Second, the optical power of the ETL changes during each projection frame displayed because it is continuously modulated over time.
In addition, a certain amount of time is required for the projection of a 3D CG object to provide sufficient light intensity for an observer.
As a result, an observer perceives the integration of the undesirable appearances of a projected image.
Other multifocal displays that employ the focal sweep approach (see Sect.~\ref{subsec:rw_vac}) suffer from the same limitation.
Using a custom high-speed projector, we can illuminate the scene for a shorter duration to mitigate the image quality degradation, but this results in the image having a darker appearance.
Another and more promising solution for overcoming this limitation is to apply a waveform optimization of ETL's drive signal~\cite{Iwai2019}, which will improve the image quality by maintaining the optical power at the desired values for a sufficient amount of time.
Third, as an implementation limitation, our system requires a high-speed projector.
Although such a projector can currently be considered special equipment, the one we used in this paper is an off-the-shelf projector that is available in the market~\cite{watanabe2015high}.
In addition, researchers have found in the last few years that a high-speed projector is essential for realizing immersive dynamic PM applications~\cite{10.1145/3415255.3422888,8998378,9284678}.
Although our current system applies a gray-scale projector, the manufacturer of the projector just launched a full-color version.
Given this technical trend, high-speed projectors will probably soon become normal equipment in PM applications.
Foruth, because a projector generally suffers from a narrow depth of field (DOF), a part of the image projected on a non-planar surface does not appear focused in some cases.
Here, even if the proposed technique successfully displays the virtual image of a projected object at the desired distance, the virtual image does not appear focused.
This can be solved by attaching another ETL to the projector and applying a focal sweep to it \cite{10.1117/12.2542477,7014259}.
\revise{Last, the ETL significantly reduces the angle of viewing for an observer. This problem can be also seen in previous AR/VR displays that employ ETLs~\cite{10.1145/2858036.2858140,8458263}. Considering the trend of the increasing aperture size of ETLs due to the constant innovation of manufacturers, we believe that this limitation will be overcome in the near future.}
In summary, although there are currently several technical limitations that are common to the related fields, we believe that they do not immediately degrade the applicability of the proposed method.
\section{Conclusion}
This paper aimed at achieving multifocal stereoscopic PM to improve the mismatch between vergence and accommodation.
To achieve this objective, we jointly designed the optics, hardware, and computational algorithms.
Specifically, we attached ETLs to the LC shutters of active-shutter glasses to control both vergence and accommodation and synchronized them using an off-the-shelf high-speed projector.
We applied a fast focal sweep to the ETLs and projected a 3D CG object from the projector onto physical surfaces when the distance of its virtual image was located exactly at a desired distance from the ETLs.
Further, we solved three technical issues that are unique to the stereoscopic PM and developed a simple lens breathing compensation technique that can be applied to other multifocal and varifocal AR/VR displays.
Using a proof-of-concept prototype, we have demonstrated that our technique can provide the correct focus cues required for different types of projection screens.
We have also confirmed that the lens breathing compensation can eliminate artifacts such as overlaps between the adjacent areas of projected imagery.
Finally, we validated the advantage provided by our technique with regard to the mitigation of the VAC by comparing it with conventional stereoscopic PM using a user study of a depth-matching task.
\revise{A new branch of research associated with AR/VR displays has recently been enabled by the combination of projectors and near-eye optics~\cite{9383112,8999805,Ueda:21}, and this work can be categorized with the same. We will continue to investigate the possibility and applicability of this emerging framework.}
\acknowledgments{
This work was supported by JSPS KAKENHI Grant Numbers JP18K19817 and JP20H05958 and JST, PRESTO Grant Number JPMJPR19J2, Japan.
}
\bibliographystyle{abbrv-doi}
|
\section{Introduction}
A partially ordered set, shortly a \textit{poset}, is a
set $P$ equipped with a relation $\le_P$ that is transitive, reflexive, and antisymmetric.
For any non-empty set $\ensuremath{\mathcal{X}}$, let $\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}})$ be the \textit{Boolean lattice} of dimension $|\ensuremath{\mathcal{X}}|$ on a \textit{ground set} $\ensuremath{\mathcal{X}}$, i.e.\
the poset consisting of all subsets of $\ensuremath{\mathcal{X}}$ equipped with the inclusion relation, $\subseteq$.
We use $Q_n$ to denote a Boolean lattice with an arbitrary $n$-element ground set.
We refer to a poset either as a pair $(P, \leq_P)$, or, when it is clear from context, simply as a set $P$. The elements of $P$ are often called \textit{vertices}.\\
For two posets $(P_1, \leq_{P_1})$ and $(P_2, \leq_{P_2})$, an \textit{embedding} $\phi\colon P_1\to P_2$ of $P_1$ into $P_2$ is an injective function such that for every $X_1,X_2\in P_1$,
$$X_1 \leq_{P_1} X_2\text{ if and only if }\phi(X_1)\leq_{P_2} \phi(X_2).$$
A poset $P_1$ is an \textit{induced subposet} of $P_2$
if $P_1\subseteq P_2$ and for every $X_1,X_2\in P_1$, $X_1 \leq_{P_1} X_2$ if and only if $X_1 \leq_{P_2} X_2.$
An \textit{copy} of a poset $P_1$ in $P_2$ is an induced subposet $P'$ of $P_2$, isomorphic to $P_1$.\\
Extremal properties of posets and their induced subposets have been investigated in recent years and mirror similar concepts in graphs.
Carroll and Katona \cite{CK} initiated the consideration of so called Tur\'an-type problems for induced subposets.
Most notable is a result by Methuku and P\'alvölgyi \cite{MP} which provides an asymptotically tight bound on the maximum size of a subposet of a Boolean lattice that does not have a copy of a fixed poset $P$, for general $P$.
Their statement has been refined for several special cases, see e.g.\ Lu and Milans \cite{LM} and M\'eroueh \cite{M}.
Further Tur\'an-type results are, for example, given by Methuku and Tompkins \cite{MT} and Tomon \cite{T}.
Note that Tur\'an-type properties are also investigated in depth for non-induced, so called \textit{weak} subposets, which are not considered here.
Besides that, saturation-type extremal problems are studied for induced and weak subposets, see a recent survey of Keszegh, et al.\ \cite{KLMPP}.
\\
In this paper we are dealing with Ramsey-type properties of induced subposets in Boolean lattices.
Consider an assignment of two colors, blue and red, to the vertices of posets. Such a coloring $c: P \rightarrow \{blue, red\}$ is a \textit{blue/red coloring} of $P$.
A colored poset is \textit{monochromatic} if all of its vertices share the same color. A monochromatic poset whose vertices are blue is called a \textit {blue poset}. Similarly defined is a \textit{red poset}.
Extending the classical definition of graph Ramsey numbers, Axenovich and Walzer \cite{AW} introduced the \textit{poset Ramsey number} which is defined as follows. For posets $P$ and $Q$, let
\begin{multline*}
R(P,Q)=\min\{N\in\ensuremath{\mathbb{N}} \colon \text{ every blue/red coloring of $Q_N$ contains either}\\
\text{a blue copy of $P$ or a red copy of $Q$}\}.
\end{multline*}
One of the central questions in this area is to determine $R(Q_n, Q_n)$. The best bounds currently known are
$2n+1 \leq R(Q_n, Q_n) \leq n^2 -n+2$, see listed chronologically Walzer \cite{W}, Axenovich and Walzer \cite{AW}, Cox and Stolee \cite{CS}, Lu and Thompson \cite{LT}, Bohman and Peng \cite{BP}.
For off-diagonal setting $R(Q_k, Q_n)$ with $k$ fixed and $n$ large, an exact result is only known if $k=1$. It is easy to see that $R(Q_1, Q_n)=n+1$.
For $k=2$, it was shown in \cite{AW} that $R(Q_2, Q_n) \leq 2n+2$. This was improved by Lu and Thompson to $R(Q_2, Q_n)\leq (5/3)n +2$.
Finally, it was further improved by Gr\'osz, Methuku, and Tompkins \cite{GMT}:
\begin{theorem}[Gr\'osz et al. \cite{GMT}]\label{thm_grosz}
Let $\epsilon>0$ and let $n\in\ensuremath{\mathbb{N}}$ be sufficiently large. Then
$$n+3 \leq R(Q_2,Q_n) \le n + \frac{(2+\epsilon)n}{\log n}.$$
\end{theorem}
Further known bounds on poset Ramsey numbers include results of Chen et al. \cite{CCCLL}, \cite{CCLL} as well as Chang et al. \cite{CGLMNPV}.
\\
In this paper, we start a more systematic investigation of $R(P, Q_n)$ for a fixed poset $P$ and large $n$. This Ramsey number gives an analogue of the graph Ramsey number $R(H,K_n)$ that claims that every edge-coloring of a complete graph in red and blue with no given induced blue subgraph $H$, contains a large red clique of size $n$.
Here, the goal is to provide a quantitative version of a statement that every blue/red coloring of a Boolean lattice with no blue (induced) subposet $P$ contains a large red Boolean sublattice.
One of the key roles here plays a small, three-vertex poset $\Lambda = (\Lambda, <)$, with vertices $Z_1, Z_2$ and $Z_3$, such that $Z_1<Z_3$, $Z_2<Z_3$, and $Z_1$ and $Z_2$ incomparable.
A poset $V$ is the symmetric counterpart of $\Lambda$, having vertices $Z_1, Z_2 $ and $Z_3$, such that $Z_1>Z_3$, $Z_2>Z_3$, and $Z_1$ and $Z_2$ not comparable.
Our main result shows a sharp jump in the behaviour of $R(P, Q_n)$ as a function of $n$ depending whether or not $P$ contains a copy $\Lambda$ or $V$.
\begin{theorem}\label{thm-MAIN}
For every poset $P$ there is an integer $n_0$ such that for any $n>n_0$ the following holds.
If $P$ contains a copy of $\Lambda$ or $V$, then $R(P, Q_n) \geq n + \frac{1}{15} \frac{n}{\log n}$.
If $P$ contains neither a copy of $\Lambda$ nor a copy of $V$, then $R(P, Q_n) \leq n + f(P)$, for some function $f$.
\end{theorem}
In order to show the lower bound, we prove a structural duality statement that together with a probabilistic construction allows to find a desired coloring.
This is the first of a kind non-marginal improvement of a trivial lower bound for poset Ramsey numbers.
Most other known lower bounds corresponded to so-called layered colorings of Boolean lattices, where any two vertices of the same size have the same color.
The only two constructions different from this and known so far are the aforementioned lower bound of Gr\'osz et al. \cite{GMT}
as well as a construction of Bohman and Peng \cite{BP} improving the trivial lower bound for the diagonal case $R(Q_n,Q_n)\ge 2n$ to $2n+1$.\\
We show the following bounds on the poset Ramsey number of $\Lambda$ versus $Q_n$.
\begin{theorem}\label{thm_main}Let $\epsilon>0$. There exists an $n_0\in\ensuremath{\mathbb{N}}$ such that for all $n\ge n_0$,
$$n+ \frac{1}{15}\cdot\frac{n}{\log n}\leq R(\Lambda,Q_n)\le n+ \big(1+\epsilon)\cdot\frac{n}{\log n}.$$
\end{theorem}
More precisely, it can be seen that the lower bound holds for $\log n_0\ge535$, while the upper bound requires $\log n_0\ge \frac{36}{\epsilon^2}$.
Note that $R(\Lambda,Q_n)\le R(Q_2,Q_n)$, so Theorem \ref{thm_grosz} already implies a bound for $R(\Lambda,Q_n)$
which is weaker but asymptotically equal to the upper bound of Theorem \ref{thm_main}.
In this paper we provide
a duality statement that allows us to prove both the lower and the improved upper bound on $R(\Lambda, Q_n)$.
Theorem \ref{thm_grosz} and Theorem \ref{thm-MAIN} also give a lower bound for $R(Q_2, Q_n)$ which is asymptotically tight not only in the first but also in the second summand.
\begin{corollary}\label{cor_Q2}
$$R(Q_2,Q_n) = n + \Theta\left(\frac{n}{\log (n)}\right).$$
\end{corollary}
The structure of the paper is as follows.
In Section \ref{sec_prelim} we introduce some notations and a new type of poset and show some useful propositions.
In Section \ref{sec_grosz} we provide an alternative proof of the upper bound in Theorem \ref{thm_grosz}.
This makes our paper self-contained since we need this result for Corollary \ref{cor_Q2}.
In Section \ref{sec_duality} we provide a structural duality statement, Theorem \ref{thm_duality}, which is the key tool for the main proofs.
In Section \ref{sec_random} we use a probabilistic construction to find a coloring with ``good'' properties.
Lastly, in Section \ref{sec_proof} we complete the proofs of Theorem \ref{thm_main} and Theorem \ref{thm-MAIN}.
\\
\section{Preliminaries}\label{sec_prelim}
\subsection{Basic Notation}
Let $\ensuremath{\mathcal{X}}$ and $\ensuremath{\mathcal{Y}}$ be disjoint sets.
Then the vertices of the Boolean lattice $\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}}\cup\ensuremath{\mathcal{Y}})$, i.e.\ the unordered subsets of $\ensuremath{\mathcal{X}}\cup\ensuremath{\mathcal{Y}}$, can be partitioned with respect to $\ensuremath{\mathcal{X}}$ and $\ensuremath{\mathcal{Y}}$ in the following manner.
Every $Z\subseteq\ensuremath{\mathcal{X}}\cup\ensuremath{\mathcal{Y}}$ has an $\ensuremath{\mathcal{X}}$-part $X_Z=Z\cap \ensuremath{\mathcal{X}}$ and a $\ensuremath{\mathcal{Y}}$-part $Y_Z=Z\cap\ensuremath{\mathcal{Y}}$.
In this setting, we refer to $Z$ alternatively as the pair $(X_Z,Y_Z)$.
Conversely, for all $X\subseteq\ensuremath{\mathcal{X}}$, $Y\subseteq\ensuremath{\mathcal{Y}}$, the pair $(X,Y)$ has a $1$-to-$1$ correspondence to the vertex $X\cup Y\in \ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}}\cup\ensuremath{\mathcal{Y}})$.
One can think of such pairs as elements of the Cartesian product $2^\ensuremath{\mathcal{X}} \times 2^\ensuremath{\mathcal{Y}}$ which has a canonic bijection to $2^{\ensuremath{\mathcal{X}}\cup\ensuremath{\mathcal{Y}}}=\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}}\cup\ensuremath{\mathcal{Y}})$.
Observe that for $X_i\subseteq\ensuremath{\mathcal{X}}, Y_i\subseteq\ensuremath{\mathcal{Y}}$, $i\in[2]$, we have $(X_1,Y_1)\subseteq (X_2,Y_2)$ if and only if $X_1\subseteq X_2$ and $Y_1\subseteq Y_2$. We omit floors and ceilings where appropriate.
\\
For any poset, we refer to vertices $Z_1,Z_2$ which are incomparable as $Z_1\nsim Z_2$.
For a positive integer $n\in\ensuremath{\mathbb{N}}$, we use $[n]$ to denote the set $\{1,\dots,n\}$. Given an integer $n\in\ensuremath{\mathbb{N}}$ and a set $\ensuremath{\mathcal{X}}$, let $\binom{\ensuremath{\mathcal{X}}}{n}$ be the set of all $n$-element subsets of $\ensuremath{\mathcal{X}}$.
Throughout the paper, `$\log$' always refers to the logarithm with base $2$, while `$\ln$' refers to the natural logarithm.
\\ \bigskip
\subsection{Structure of posets with forbidden $\Lambda$ or $V$}
A poset $\ensuremath{\mathcal{T}}$ is an \textit{up-tree} if there is a unique minimal vertex in $\ensuremath{\mathcal{T}}$ and for every vertex $X\in\ensuremath{\mathcal{T}}$, the set $\{Y\in\ensuremath{\mathcal{T}}\colon Y\le X\}$ is a chain, i.e.,\ its vertices are pairwise comparable.
We say that two posets are \textit{independent} if they are vertexwise incomparable. Furthermore, a collection of posets is \textit{independent} if they are pairwise independent.
We use this notation to describe posets which don't contain a copy of $\Lambda$ (or $V$).
\begin{lemma}\label{strucL}
Let $P$ be a poset. There is no copy of $\Lambda$ in $P$ if and only if $P$ is an independent collection of up-trees.
\end{lemma}
\begin{proof} Observe that a poset $P$ is an independent collection of up-trees if and only if for every vertex $X\in P$, $\{Y\in P\colon Y\le X\}$ forms a chain.
Suppose that there is a copy of $\Lambda$ in $P$ on vertices $Z_i$, $i\in[3]$ with $Z_1<Z_3$, $Z_2<Z_3$ and $Z_1\nsim Z_2$.
Then $Z_1,Z_2$ witness that $\{Y\in P\colon Y<Z_3\}$ is not a chain, so $P$ is not an independent collection of up-trees.
Now assume that $P$ is not an independent collection of up-trees.
Then there exist some $X\in P$ and $Z_1,Z_2 \in \{Y\in P\colon Y\le X\}$ such that $Z_1\nsim Z_2$.
Since $X$ is comparable to all vertices in $\{Y\in P\colon Y\le X\}$, $X> Z_1, X> Z_2$. Now $X,Z_1,Z_2$ form a copy of $\Lambda$.
\end{proof}
By symmetry an analogous statement holds for posets with forbidden induced copy of $V$.
If we forbid both $V$ and $\Lambda$ simultaneously we obtain the following structure.
\begin{corollary}\label{strucI}
Let $P$ be a poset such that there is neither a copy of $V$ nor of $\Lambda$.
Then $P$ is an independent collection of chains.
\end{corollary}
\bigskip
\subsection{Embeddings of $Q_n$}
When considering an embedding $\phi$ of a Boolean lattice $Q_n$ into a larger Boolean lattice $\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{Z}})$, we can partition $\ensuremath{\mathcal{Z}}$ such that it has the following nice property.
\begin{lemma}\label{embed_lem
Let $n\in\ensuremath{\mathbb{N}}$. Let $\ensuremath{\mathcal{Z}}$ be a set with $|\ensuremath{\mathcal{Z}}|>n$ and let $Q=\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{Z}})$. If there is an embedding $\phi\colon Q_n\to Q$,
then there exist a subset $\ensuremath{\mathcal{X}}\ensuremath{\subset}\ensuremath{\mathcal{Z}}$ with $|\ensuremath{\mathcal{X}}|=n$, and an embedding $\phi'\colon \ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}})\to Q$ such that
$\phi'(X)\cap \ensuremath{\mathcal{X}}=X$ for all $X\subseteq \ensuremath{\mathcal{X}}$.
\end{lemma}
\begin{proof}Let the ground set of $Q_n$ be $\ensuremath{\mathcal{X}}'$.
We consider the embedding of singletons of $Q_n$, i.e.\ $\phi(\{a\})$, $a\in\ensuremath{\mathcal{X}}'$.
If $\phi(\{a\})\subseteq \bigcup_{X'\subseteq \ensuremath{\mathcal{X}}'\backslash\{a\}} \phi(X')$,
then $\phi(\{a\})\subseteq \bigcup_{X'\subseteq \ensuremath{\mathcal{X}}'\backslash\{a\}} \phi(X') \subseteq \phi(\ensuremath{\mathcal{X}}'\backslash\{a\})$.
But $\{a\}\nsubseteq \ensuremath{\mathcal{X}}'\backslash\{a\}$ and $\phi$ is an embedding, a contradiction. Thus $\phi(\{a\})\not\subseteq \bigcup_{X'\subseteq \ensuremath{\mathcal{X}}'\backslash\{a\}} \phi(X')$.
For every $a\in\ensuremath{\mathcal{X}}'$, pick an arbitrary $$b(a)\in \phi(\{a\})\backslash\bigcup_{X'\subseteq \ensuremath{\mathcal{X}}'\backslash\{a\}} \phi(X').$$
Note that $b(a_1)\notin \phi(\{a_2\})$ for any $a_1,a_2\in\ensuremath{\mathcal{X}}'$, $a_1\neq a_2$, so all representatives are distinct.
Let $\ensuremath{\mathcal{X}}=\{b(a) \colon a\in\ensuremath{\mathcal{X}}'\}$. We see that the map $b: \ensuremath{\mathcal{X}}' \to \ensuremath{\mathcal{X}}$ is a bijection.
For every $B\subseteq \ensuremath{\mathcal{X}}$, let $A_B\subseteq \ensuremath{\mathcal{X}}'$ be such that $B=\{b(a): a \in A_B\}$.
We define $\phi'\colon \ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}})\to Q$ as follows: $\phi'(B) = \phi(A_B)$, $B\in \ensuremath{\mathcal{X}}$.
Then $\phi'$ is an embedding.
Observe that for $X\subseteq \ensuremath{\mathcal{X}}$ and $b\in\ensuremath{\mathcal{X}}$, \ $b\in \phi'(X)$ if and only if $b\in X$. Thus $\phi'(X)\cap \ensuremath{\mathcal{X}}=X$ for all $X\subseteq \ensuremath{\mathcal{X}}$.
This concludes the proof.
\end{proof}
We call an embedding $\phi$ of $\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}})$ into $\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}}\cup \ensuremath{\mathcal{Y}})$ for disjoint $\ensuremath{\mathcal{X}}$ and $\ensuremath{\mathcal{Y}}$, \textit{$\ensuremath{\mathcal{X}}$-good} if $\phi(X)\cap \ensuremath{\mathcal{X}} = X$ for all $X\in \ensuremath{\mathcal{X}}$.
We also call a copy $Q$ of $\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}})$ in $\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}}\cup \ensuremath{\mathcal{Y}})$ $\ensuremath{\mathcal{X}}$-\textit{good} if there is an $\ensuremath{\mathcal{X}}$-good embedding of $\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}})$ into $\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}}\cup \ensuremath{\mathcal{Y}})$.
Lemma \ref{embed_lem} claims in particular that for any copy of $Q_n$ in a larger Boolean lattice $Q$,
there is a subset $\ensuremath{\mathcal{X}}$ of the ground set of $Q$ with $|\ensuremath{\mathcal{X}}|=n$ such that there is an $\ensuremath{\mathcal{X}}$-good copy of $\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}})$ in $Q$.
\\
\bigskip
\subsection{Red copy of $Q_n$ vs.\ blue chain}\label{sec_grosz}
The main goal of this subsection is to present an alternative proof for the upper bound of Theorem \ref{thm_grosz}.
Gr\'osz, Methuku, and Tompkins \cite{GMT} stated the following lemma using a different formulation.
While they used algorithmic tools in their proof, we prove the statement recursively.
Recall that for a given partition $\ensuremath{\mathcal{X}}\cup \ensuremath{\mathcal{Y}}$ of the ground set of a Boolean lattice we denote a vertex $X\cup Y\in\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}}\cup\ensuremath{\mathcal{Y}})$, where $X\subseteq\ensuremath{\mathcal{X}}, Y\subseteq\ensuremath{\mathcal{Y}}$, as $(X,Y)$.
\begin{lemma}
\label{shift_lem}
Let $\ensuremath{\mathcal{X}}$, $\ensuremath{\mathcal{Y}}$ be disjoint sets with $|\ensuremath{\mathcal{X}}|=n$ and $|\ensuremath{\mathcal{Y}}|=k$, for some $n,k\in\ensuremath{\mathbb{N}}$. Let $Q=\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}}\cup\ensuremath{\mathcal{Y}})$ be a blue/red colored Boolean lattice.
Fix some linear ordering $\pi=(y_1,\dots,y_k)$ of $\ensuremath{\mathcal{Y}}$ and define $Y(0), \ldots, Y(k)$ by $Y(0)=\varnothing$ and $Y(i)=\{y_1,\dots,y_i\}$ for $i\in[k]$.
Then there exists at least one of the following in $Q$:
\renewcommand{\labelenumi}{(\alph{enumi})}
\begin{enumerate}
\item a red $\ensuremath{\mathcal{X}}$-good copy of $\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}})$, or
\item a blue chain of length $k+1$ of the form $(X_0,Y(0)),\dots,(X_{k},Y(k))$ where $X_0 \subseteq X_1 \subseteq\dots \subseteq X_k\subseteq \ensuremath{\mathcal{X}}$.
\end{enumerate}
\end{lemma}
\begin{proof}
Suppose that there is no blue chain as described in (b).
For every $X\subseteq\ensuremath{\mathcal{X}}$, we recursively define a label $\ell_X\in\{0,\dots,k\}$ such that $\phi\colon \ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}})\to Q$, $\phi(X)=(X,Y(\ell_X))$, is an embedding with monochromatic red image.
We require $\ell_X$ to fulfill three properties:
\begin{enumerate}[label=(\arabic*)]
\item For any $X'\subseteq X$, \ $\ell_{X'}\le \ell_{X}$.
\item There is a blue chain of length $\ell_X$ contained
in the
Boolean lattice
with ground set $X\cup Y(\ell_X)$, which we denote by $Q^{X}$.
\item $(X,Y(\ell_X))$ is red.
\end{enumerate}
First, consider the vertex $\varnothing$. Let $\ell_\varnothing$ be the minimum $\ell$, $0\le \ell\le k$, such that $(\varnothing,Y(\ell))$ is red.
If such an $\ell$ does not exist, then $(\varnothing,Y(0)), \dots, (\varnothing,Y(k))$ form a blue chain, a contradiction.
It is clear to see that Properties (1) and (3) hold.
If $\ell_\varnothing=0$, (2) is trivially true.
If $\ell_\varnothing\ge 1$, $(\varnothing,Y(0)), \dots, (\varnothing,Y(\ell_{\varnothing}-1))$ form a blue chain of length $\ell_\varnothing$ and $(2)$ holds as well.
\\
Consider an arbitrary $X\subseteq\ensuremath{\mathcal{X}}$ and suppose that for all $X'\ensuremath{\subset} X$ we already defined $\ell_{X'}$ with Properties (1)-(3). Let $\ell'_X=\max_{\{U\ensuremath{\subset} X\}} \ell_U$.
Then let $\ell_X$ be the minimum~$\ell$, $\ell'_X\le \ell \le k$ such that $(X,Y(\ell))$ is colored in red.
If there is no such $\ell$, then $(X,Y(\ell'_X)),\dots,(X,Y(k))$ is a blue chain of length $k-\ell'_X+1$. By definition of $\ell'_X$ there is some $U\ensuremath{\subset} X$ with $\ell_U=\ell'_X$.
In particular, $(2)$ holds for $U$, so there is a blue chain of length $\ell'_X$ in $Q^{U}$. Note that $(U,Y(\ell_U))\ensuremath{\subset}(X,Y(\ell'_X))$,
so we obtain a blue chain of length $k+1$. This is a contradiction, thus $\ell_X$ is well-defined and fulfills Property (3).
If $\ell_X=\ell'_X$, consider the aforementioned blue chain of length $\ell'_X$ in $Q^{U}$, and otherwise consider this chain together with $(X,Y(\ell'_X)), \dots, (X,Y(\ell_X-1))$.
In both cases, we obtain a blue chain of length $\ell_X$, which proves $(2)$.
For $X'\ensuremath{\subset} X\subseteq \ensuremath{\mathcal{X}}$, \ $\ell_{X'}\le \ell'_X\le \ell_X$, thus (1) holds.
\\
We define $\phi\colon \ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}})\to Q$, $\phi(X)=(X,Y(\ell_X))$. Note that $\phi(X)\cap\ensuremath{\mathcal{X}}=X$ for every $X\subseteq\ensuremath{\mathcal{X}}$ and Property (3) implies that $\phi(X)$ is red.
Let $X_1,X_2\subseteq\ensuremath{\mathcal{X}}$. If $\phi(X_1)\subseteq\phi(X_2)$, it is immediate that $X_1\subseteq X_2$.
Conversely, if $X_1\subseteq X_2$, then by Property (1) we have $\ell_{X_1}\le \ell_{X_2}$. Thus $(X_1,Y(\ell_{X_1}))\subseteq(X_2,Y(\ell_{X_2}))$.
As a consequence, $\phi$ is an $\ensuremath{\mathcal{X}}$-good embedding of $\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}})$.
\end{proof}
This Lemma implies the following corollary which is already given in an alternative form by Axenovich and Walzer, see Lemma 4 of \cite{AW}.
\begin{corollary}\label{shift_cor}
Let $\ensuremath{\mathcal{X}},\ensuremath{\mathcal{Y}}$ be disjoint sets with $|\ensuremath{\mathcal{X}}|=n$ and $|\ensuremath{\mathcal{Y}}|=k$.
Let $\ensuremath{\mathcal{P}}$ be a subposet of a Boolean lattice $Q=\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}}\cup\ensuremath{\mathcal{Y}})$ such that there is no chain of length $k+1$ in $\ensuremath{\mathcal{P}}$.
Then there exists a copy of $Q_n$ in $Q$ which contains no vertex of $\ensuremath{\mathcal{P}}$.
\end{corollary}
\begin{proof} Fix an arbitrary linear ordering of $\ensuremath{\mathcal{Y}}$.
Furthermore, let $c: Q \to \{blue,red\}$ be the coloring such that
\begin{equation} \nonumber
c(X) =
\begin{cases}
\text{blue}, \hspace*{0.7cm}\mbox{ if } ~~ X\in \ensuremath{\mathcal{P}},\\
\text{red}, \hspace*{0.8cm}\mbox{ otherwise.}
\end{cases}
\end{equation}
There is no blue chain of length $k+1$ in $c$, so by Lemma \ref{shift_lem} there is a monochromatic red copy of $Q_n$ in $Q$.
This copy does not contain any vertex of $\ensuremath{\mathcal{P}}$.
\end{proof}
With the help of Lemma \ref{shift_lem}, we can now prove an upper bound for $R(Q_2,Q_n)$. The concluding arguments are due to Gr\'osz, Methuku, and Tompkins \cite{GMT}.
\begin{proof}[Proof of Theorem \ref{thm_grosz}]
For the lower bound, the reader is referred to \cite{GMT}.
For the upper bound, let $k\in\ensuremath{\mathbb{N}}$ with $k=\tfrac{(2+\epsilon)n}{\log (n)}$.
Let $\ensuremath{\mathcal{X}}$ and $\ensuremath{\mathcal{Y}}$ be disjoint sets with $|\ensuremath{\mathcal{X}}|=n$ and $|\ensuremath{\mathcal{Y}}|=k$. Consider a blue/red coloring of $Q=\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}}\cup\ensuremath{\mathcal{Y}})$ with no monochromatic red copy of $Q_n$.
Let $\pi=(y_1^\pi,\dots,y_k^\pi)$ be a linear ordering of $\ensuremath{\mathcal{Y}}$.
By Lemma \ref{shift_lem}, there exists a blue chain of length $k+1$
of the form $(X^\pi_0,\varnothing), (X^\pi_1,\{y^\pi_1\}),(X^\pi_2,\{y^\pi_1,y^\pi_2\}),\dots, (X^\pi_k,\ensuremath{\mathcal{Y}})$ where $X^\pi_i\subseteq\ensuremath{\mathcal{X}}$.
\\
Note that there are $k!$ distinct orderings of $\ensuremath{\mathcal{Y}}$.
For each linear ordering $\pi$ of $\ensuremath{\mathcal{Y}}$ we consider $X^\pi_0$ and $X^\pi_k$, i.e.\ the minimal and maximal vertex of the aforementioned chain restricted to $\ensuremath{\mathcal{X}}$.
By the choice of $k$, we obtain $k!> 2^{2n}$. In particular by pigeonhole principle, there are distinct $\pi_1,\pi_2$
with $X^{\pi_1}_0=X^{\pi_2}_0$ and $X^{\pi_1}_k=X^{\pi_2}_k$.
Since $\pi_1,\pi_2$ are distinct, there exists an index $1\le i\le k-1$ with $\{y^{\pi_1}_1,\dots,y^{\pi_1}_i\}\neq \{y^{\pi_2}_1,\dots,y^{\pi_2}_i\}$.
Then the four vertices
$$(X^{\pi_1}_0,\varnothing), (X^{\pi_1}_i,\{y^{\pi_1}_1,\dots,y^{\pi_1}_i\}), (X^{\pi_2}_i,\{y^{\pi_2}_1,\dots,y^{\pi_2}_i\}), (X^{\pi_1}_k,\ensuremath{\mathcal{Y}})$$ form a blue copy of $Q_2$.
\end{proof}
\bigskip
\subsection{Factorial trees and shrubs}
Besides the Boolean lattice, there is another poset which plays a major role in this paper, which we call the \textit{factorial tree}.\\
Consider the set of ordered subsets of a fixed non-empty set $\ensuremath{\mathcal{Y}}$, that also could be thought of as a set of strings with non-repeated letters over the alphabet $\ensuremath{\mathcal{Y}}$.
Note that we also allow the empty set as such an ordered subset.
Occasionally, if it is clear from the context, we refer to the empty ordered set $(\varnothing,\le)$ simply as $\varnothing$.
For an ordered subset $S$ of $\ensuremath{\mathcal{Y}}$, we refer to its underlying unordered set as $\underline{S}$. Let $|S|=|\underline{S}|$ be the \textit{size} of $S$.
We also say that $S$ is an \textit{ordering} of $\underline{S}$.
\\
Let $S$ be an ordered subset of $\ensuremath{\mathcal{Y}}$. A \textit{prefix} of $S$ is an ordered subset $T$ of $\ensuremath{\mathcal{Y}}$ consisting of the first $|T|$ elements of $S$ in the ordering induced by $S$.
If $T$ is a prefix of $S$, we write $T\le_\ensuremath{\mathcal{O}} S$. Note that the empty ordered set is a prefix of every ordered set.
If $T\neq S$, we say that a prefix $T$ of $S$ is \textit{strict}, denoted by $T<_\ensuremath{\mathcal{O}} S$.
Observe that the prefix relation $\le_\ensuremath{\mathcal{O}}$ is transitive, reflexive and antisymmetric. Let $\ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}})$ be the poset of all ordered subsets of $\ensuremath{\mathcal{Y}}$ equipped with $\le_\ensuremath{\mathcal{O}}$.
We say that this poset is the \textit{factorial tree} on ground set $\ensuremath{\mathcal{Y}}$.
\\
In a factorial tree $\ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}})$ for every vertex $S\in\ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}})$, the set of prefixes $\{T\in\ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}})\colon T\le_\ensuremath{\mathcal{O}} S\}$ induces a chain. Furthermore, the vertex $\varnothing$ is the unique minimal vertex of $\ensuremath{\mathcal{Y}}$, thus $\ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}})$ is an up-tree.
\\
Let $\ensuremath{\mathcal{X}}$ and $\ensuremath{\mathcal{Y}}$ be disjoint sets. Let $Q=\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}}\cup\ensuremath{\mathcal{Y}})$ and $\ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}})$ be the factorial tree with ground set $\ensuremath{\mathcal{Y}}$.
An embedding $\tau$ of $\ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}})$ into $Q$ is \textit{$\ensuremath{\mathcal{Y}}$-good} if for every $S\in \ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}})$, $\tau(S)\cap \ensuremath{\mathcal{Y}}=\underline{S}$.
We say that a subposet $\ensuremath{\mathcal{P}}$ of $Q$ is a \textit{$\ensuremath{\mathcal{Y}}$-good copy} of $\ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}})$ if
there exists a $\ensuremath{\mathcal{Y}}$-good embedding $\tau: \ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}}) \to Q$ with image $\ensuremath{\mathcal{P}}$. We refer to such a copy also as a \textit{$\ensuremath{\mathcal{Y}}$-shrub}.\\
Besides that, we also consider a related subposet with slightly weaker conditions.
A \textit{weak $\ensuremath{\mathcal{Y}}$-shrub} is a subposet $\ensuremath{\mathcal{P}}$ of $Q$ such that there is a function $\tau: \ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}}) \to Q$ with image $\ensuremath{\mathcal{P}}$
such that for every $S\in\ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}})$, \ $\tau(S)\cap \ensuremath{\mathcal{Y}}=\underline{S}$ and for every $S,T\in \ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}})$ with $S<_\ensuremath{\mathcal{O}} T$, \ $\tau(S)\ensuremath{\subset} \tau(T)$.
In particular, a weak $\ensuremath{\mathcal{Y}}$-shrub might not correspond to an injective embedding of $\ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}}) $.
\\
Clearly a $\ensuremath{\mathcal{Y}}$-shrub is also a weak $\ensuremath{\mathcal{Y}}$-shrub. Surprisingly, the converse statement is also true in Boolean lattices not containing a copy of $\Lambda$.
\begin{proposition}\label{obs_shrub}
Let $\ensuremath{\mathcal{X}}$ and $\ensuremath{\mathcal{Y}}$ be disjoint sets, let $Q=\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}}\cup\ensuremath{\mathcal{Y}})$.
Let $\ensuremath{\mathcal{P}}$ be a weak $\ensuremath{\mathcal{Y}}$-shrub in $Q$ such that $\ensuremath{\mathcal{P}}$ contains no copy of $\Lambda$.
Then $\ensuremath{\mathcal{P}}$ is a $\ensuremath{\mathcal{Y}}$-shrub.
\end{proposition}
\begin{proof}
Let $\tau: \ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}}) \to Q$ be a map such that for every $S,T\in \ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}})$ with $S<_\ensuremath{\mathcal{O}} T$, we have $\tau(S)\ensuremath{\subset} \tau(T)$ and $\tau(S)\cap \ensuremath{\mathcal{Y}}=\underline{S}$, and let $\ensuremath{\mathcal{P}}$ be its image.
For all $S\in\ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}})$, let $X_S=\tau(S)\cap \ensuremath{\mathcal{X}}$, i.e.\ $\tau(S)=(X_S,\underline{S})$.
We shall show that $\tau$ is an embedding, thus proving that $\ensuremath{\mathcal{P}}$ is a $\ensuremath{\mathcal{Y}}$-shrub. For that we need to prove that the condition $\tau(S)\subseteq\tau(T)$ implies that $S\le_\ensuremath{\mathcal{O}} T$
for any ordered subsets $S$ and $T$ of $\ensuremath{\mathcal{Y}}$.
\\
Let $\tau(S)\subseteq\tau(T)$, i.e.\ $(X_S,\underline{S})\subseteq(X_{T},\underline{T})$.
In particular, $\underline{S}\subseteq \underline{T}$ and so $|S|\le |T|$.
Let $R$ be the largest common prefix of $S$ and $T$. Such exists since $\varnothing$ is a prefix of every ordered set.
If $|R|=|S|$, then
$S=R\le_\ensuremath{\mathcal{O}} T$ and we are done. So we can assume that $|S|\ge |R|+1$.\\
If $|T|\le |R|+1$, then $|R|+1\le |S|\le |T|\le |R|+1$. This implies $|S|=|T|$ and since $\underline{S}\subseteq \underline{T}$, we have $\underline{S}=\underline{T}$.
Let $\{y\}= \underline{S}\backslash \underline{R}=\underline{T}\backslash \underline{R}$.
Then both $S,T$ have $R$ as prefix of size $|S|-1=|T|-1$ and $y$ as final vertex. Thus $S=T$ and we are done as well.\\
From now on, we assume that $|S|\ge |R|+1$ and $|T|> |R|+1$.
Consider prefixes $S'\le_\ensuremath{\mathcal{O}} S$ and $T'\le_\ensuremath{\mathcal{O}} T$ of size $|R|+1$.
Then $R$ is a prefix of both $S'$ and $T'$.
Let $y_S$ such that $\underline{S'}\backslash \underline{R}=\{y_S\}$ and let $y_T$ with $\underline{T'}\backslash \underline{R}=\{y_T\}$.
If $y_S=y_T$, we obtain $S'=T'$, which implies that $R$ is not the largest common prefix of $S$ and $T$, a contradiction.
If $y_S\neq y_T$, the unordered sets $\underline{S'}$ and $\underline{T'}$ are not comparable. In particular, $(X_{T'},\underline{T'})$ and $(X_{S'},\underline{S'})$ are incomparable.
Because $S'\le_\ensuremath{\mathcal{O}} S$, $T'<_\ensuremath{\mathcal{O}} T$ and by our initial assumption, we know that $(X_{S'},\underline{S'})\subseteq (X_S,\underline{S})\subseteq(X_{T},\underline{T})$ and $(X_{T'},\underline{T'})\subseteq (X_{T},\underline{T})$.
Since $|S'|=|T'|=|R|+1<|T|$, we obtain that both $(X_{S'},\underline{S'})$ and $(X_{T'},\underline{T'})$ are proper subsets of $(X_{T},\underline{T})$.
Then the three vertices $(X_{T},\underline{T}), (X_{T'},\underline{T'})$ and $(X_{S'},\underline{S'})$ form a copy of $\Lambda$ in $Q$, so we reach a contradiction.
\end{proof}
\subsection{Construction of an almost optimal shrub}
Let $\ensuremath{\mathcal{Y}}$ be a $k$-element set. Note that a $\ensuremath{\mathcal{Y}}$-shrub has $k!$ maximal vertices corresponding to all permutations of $\ensuremath{\mathcal{Y}}$.
These maximal vertices form an antichain, i.e.\ are pairwise incomparable.
Sperner's theorem implies that a ground set of any $\ensuremath{\mathcal{Y}}$-shrub must have size at least $q$, where $\binom{q}{\lfloor q/2 \rfloor} \geq k!$, so $q\ge k(\log k +\log e) + o(k)$.
Next, we shall construct a $\ensuremath{\mathcal{Y}}$-shrub which is almost optimal in the sense that $\ensuremath{\mathcal{Y}}$ has ground set of size almost matching the lower bound above.
\begin{proposition} \label{canonical-shrub}
Let $\ensuremath{\mathcal{Y}}$ be a $k$-element set.
Let $A$ be a set disjoint from $\ensuremath{\mathcal{Y}}$ such that $|A| \geq k \cdot \min\{\log k + \log\log k,11\}$.
Then there is a $\ensuremath{\mathcal{Y}}$-shrub in $\ensuremath{\mathcal{Q}}(A\cup\ensuremath{\mathcal{Y}})$.
\end{proposition}
\begin{proof}
Let $\ensuremath{\mathcal{Y}}=\{y_0, \ldots, y_{k-1}\}$ and let $Q=\ensuremath{\mathcal{Q}}(A\cup\ensuremath{\mathcal{Y}})$.
We use addition of indices modulo $k$.
Let $A_0, \ldots, A_{k-1}$ be pairwise disjoint subsets of $A$ such that $|A_i|= \ell$ for the smallest integer $\ell$ satisfying
$\binom{\ell}{\lfloor \ell/2 \rfloor} \geq k$.
Since $\ell\le \log k +\log\log k$ for $k\ge 256$ and $\ell\le 11$ for $k\le256$, such $A_i$'s can be chosen.
Let $\big\{A_i^j: j\in \{0, \ldots, k-1\}\big\}$ be an antichain in $\ensuremath{\mathcal{Q}}(A_i)$, $i\in\{0,\ldots, k-1\}$.
Such an antichain exists by Sperner's theorem.\\
Consider the factorial tree $\ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}})$. We shall construct an embedding $\tau$ of $\ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}})$ into $Q$ as follows.
Let $\tau(\varnothing)= \varnothing$.
Consider any non-empty ordered subset of $\ensuremath{\mathcal{Y}}$, say $(y_{i_1}, y_{i_2}, \ldots, y_{i_j})$, $1\leq j\leq k$.
If $j=1$, let $\tau((y_{i_1}))=A_{i_1}\cup \{y_{i_1}\}$. If $j>1$, let
$$\tau( (y_{i_1}, \ldots, y_{i_j})) = A_{i_1}\cup A_{i_1+1}^{i_2} \cdots \cup A_{i_1+j-1}^{i_j} \cup \{y_{i_1}, \ldots, y_{i_j}\}.$$
For example for $k=4$, $\tau((y_0, y_1, y_2)) = A_0\cup A_1^1\cup A_2^2 \cup \{y_1, y_2, y_3\}$,
$\tau((y_2, y_3, y_1)) = A_2 \cup A_3^3 \cup A_0^1 \cup \{y_1, y_2, y_3\}$, and $\tau((y_3, y_1)) = A_3 \cup A_0^1 \cup \{y_1, y_3\}$.\\
Note that the image of $\ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}})$ under $\tau$ is an up-tree, $\ensuremath{\mathcal{T}}$, whose minimum vertex is $\varnothing$, see Figures \ref{fig:shrub} and \ref{fig:shrub-1}.
We see that each maximum vertex of $\ensuremath{\mathcal{T}}$ is joined to $\varnothing$ by a unique chain, a maximal chain.
Furthermore, non-zero vertices that belong to distinct maximal chains are incomparable.
Observe that $\tau$ is a $\ensuremath{\mathcal{Y}}$-good embedding of $\ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}})$ into $Q$.
Indeed, for any ordered sequence of distinct vertices $(y_{i_1}, \ldots, y_{i_j})$, we have
$\tau( (y_{i_1}, \ldots, y_{i_j}))\cap \ensuremath{\mathcal{Y}} = \{y_{i_1}, \ldots, y_{i_j}\}$. In addition $(y_{i_1}, \ldots, y_{i_q})<_{\ensuremath{\mathcal{O}}} (y_{i_1}, \ldots, y_{i_p})$ if and only if
$\tau((y_{i_1}, \ldots, y_{i_q}))$ and $\tau ((y_{i_1}, \ldots, y_{i_p}))$ are in the same maximal chain of $T$ in the corresponding order.
\end{proof}
\begin{figure
\includegraphics[width=\textwidth]{shrub.png}
\caption{A diagram illustrating how the $A_i^j$'s are being assigned to the elements of the $\{y_0, y_1, y_2, y_3\}$-shrub constructed in Proposition \ref{canonical-shrub}.}
\label{fig:shrub}
\end{figure}
\begin{figure
\includegraphics[width=\textwidth]{shrub-1.png}
\caption{Segment of a shrub highlighted in Figure \ref{fig:shrub}. Here the union signs are omitted because of the spacing, for example $A_2 A_3^1 A_0^0 y_2 y_1 y_0$ corresponds to
the shrub vertex $A_2 \cup A_3^1 \cup A_0^0\cup \{ y_2, y_1, y_0\}$.}
\label{fig:shrub-1}
\end{figure}
\section{Duality theorem}\label{sec_duality}
In this section, we show a duality statement which is the key argument for the proof of Theorem \ref{thm_main}. Recall the following definitions.
We call an embedding $\phi$ of $\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}})$ into $\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}}\cup \ensuremath{\mathcal{Y}})$ for disjoint $\ensuremath{\mathcal{X}}$ and $\ensuremath{\mathcal{Y}}$, \textit{ $\ensuremath{\mathcal{X}}$-good} if $\phi(X)\cap \ensuremath{\mathcal{X}} = X$ for all $X\subseteq \ensuremath{\mathcal{X}}$.
We also call a copy of $\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}})$ in $\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}}\cup \ensuremath{\mathcal{Y}})$ $\ensuremath{\mathcal{X}}$-\textit{good} if there is an $\ensuremath{\mathcal{X}}$-good embedding of $\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}})$ into $\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}}\cup \ensuremath{\mathcal{Y}})$.
An embedding $\tau$ of $\ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}})$ into $\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}}\cup \ensuremath{\mathcal{Y}})$ is \textit{ $\ensuremath{\mathcal{Y}}$-good} if $\tau(S)\cap\ensuremath{\mathcal{Y}}=\underline{S}$ for all $S\in\ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}})$.
We say that a copy of $\ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}})$ is a \textit{$\ensuremath{\mathcal{Y}}$-shrub} if there exists a $\ensuremath{\mathcal{Y}}$-good embedding of $\ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}})$ into $\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}}\cup \ensuremath{\mathcal{Y}})$.
\begin{theorem}[Duality Theorem]\label{thm_duality}
For two disjoint sets $\ensuremath{\mathcal{X}}$ and $\ensuremath{\mathcal{Y}}$, let $Q=\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}}\cup \ensuremath{\mathcal{Y}})$ be a blue/red colored Boolean lattice which contains no blue copy of $\Lambda$.
Then there is exactly one of the following in $Q$:
\begin{itemize}
\item a red $\ensuremath{\mathcal{X}}$-good copy of $\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}})$, or
\item a blue $\ensuremath{\mathcal{Y}}$-good copy of $\ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}})$, i.e.\ a blue $\ensuremath{\mathcal{Y}}$-shrub.
\end{itemize}
\end{theorem}
Informally speaking, this duality statement claims that for any bipartition $\ensuremath{\mathcal{X}}\cup\ensuremath{\mathcal{Y}}$ of the ground set of a Boolean lattice there exists either a red copy of $\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}})$ that is restricted to $\ensuremath{\mathcal{X}}$ or a blue copy of the factorial tree $\ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}})$ restricted to $\ensuremath{\mathcal{Y}}$.
This result can be seen as a strengthening of Lemma \ref{shift_lem} in the special case when we forbid a blue copy of $\Lambda$.
The Duality Theorem implies a criterion for blue/red colored Boolean lattices $Q$ to have neither a blue copy of $\Lambda$ nor a red copy of $Q_n$.
\begin{corollary}\label{cor_duality}
Let $n,k\in\ensuremath{\mathbb{N}}$ and $N=n+k$. Let $Q=\ensuremath{\mathcal{Q}}([N])$ be a blue/red colored Boolean lattice with no blue copy of $\Lambda$.
There is no red copy of $Q_n$ in $Q$ if and only if for every $\ensuremath{\mathcal{Y}}\in\binom{[N]}{k}$ there exists a blue $\ensuremath{\mathcal{Y}}$-shrub in $Q$.
\end{corollary}
\begin{proof
Lemma \ref{embed_lem} provides that there is a red copy of $Q_n$ in $Q$ if and only if
there exists a partition $[N]=\ensuremath{\mathcal{X}}\cup\ensuremath{\mathcal{Y}}$ of the ground set of $Q$ with $|\ensuremath{\mathcal{X}}|=n$ and $|\ensuremath{\mathcal{Y}}|=k$
as well as an $\ensuremath{\mathcal{X}}$-good embedding $\phi$ of $Q(\ensuremath{\mathcal{X}})$ into $Q$ with a monochromatic red image.
If there is a red copy of $Q_n$ in $Q$, then for $\ensuremath{\mathcal{X}},\ensuremath{\mathcal{Y}}$ from Lemma \ref{embed_lem} there is also an $\ensuremath{\mathcal{X}}$-good copy of $\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}})$.
Thus by Theorem \ref{thm_duality} there is no blue $\ensuremath{\mathcal{Y}}$-shrub.
On the other hand, if there is no red copy of $Q_n$ in $Q$, there is no red $\ensuremath{\mathcal{X}}$-good copy of $\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}})$ for any $\ensuremath{\mathcal{X}} \in \binom{[N]}{n}$.
Then for an arbitrary $n$-element subset $\ensuremath{\mathcal{X}}$ of $[N]$, let $\ensuremath{\mathcal{Y}}= [N]\setminus \ensuremath{\mathcal{X}}$. Now
Theorem \ref{thm_duality} implies that there exists a blue $\ensuremath{\mathcal{Y}}$-shrub. In particular, there is a blue $\ensuremath{\mathcal{Y}}$-shrub for any $k$-element subset $\ensuremath{\mathcal{Y}}$ of $[N]$.
\end{proof}
\bigskip
Throughout the section, let $\ensuremath{\mathcal{X}}$ and $\ensuremath{\mathcal{Y}}$ be fixed disjoint sets.
Let $Q=\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}}\cup\ensuremath{\mathcal{Y}})$ be a Boolean lattice on ground set $\ensuremath{\mathcal{X}}\cup\ensuremath{\mathcal{Y}}$. We fix an arbitrary blue/red coloring of $Q$ with no blue copy of $\Lambda$.
We always let $n,k\in\ensuremath{\mathbb{N}}$ such that $|\ensuremath{\mathcal{X}}|=n$, $|\ensuremath{\mathcal{Y}}|=k$ and let $N=n+k$.
For $X\subseteq \ensuremath{\mathcal{X}}$, $Y\subseteq\ensuremath{\mathcal{Y}}$, we usually denote the vertex $X\cup Y$ by $(X,Y)$.
\\
In order to characterize colorings of $Q$ which do not contain an embedding $\phi$ of $\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}})$ into $Q$ such that for every $X\in \ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}})$, $\phi(X)$ is red and $\phi(X)\cap \ensuremath{\mathcal{X}}=X$,
we introduce the following notation.
For $X\subseteq \ensuremath{\mathcal{X}}$ and $Y\subseteq \ensuremath{\mathcal{Y}}$, we say that the vertex $(X,Y)\in Q$ is \textit{embeddable} if there is an embedding $\phi:\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}})\cap\{X'\subseteq\ensuremath{\mathcal{X}}\colon X'\supseteq X\}\to Q$
with a monochromatic red image, such that $\phi(X')\cap \ensuremath{\mathcal{X}}=X'$ for all $X'$ and $\phi(X)\supseteq (X,Y)$.
We say that $\phi$ \textit{witnesses} that $(X,Y)$ is embeddable.
This definition immediately implies:
\begin{observation}\label{lem2}
$(\varnothing,\varnothing)$ is not embeddable if and only if there is no embedding $\phi:\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}})\to Q$ such that for every $X'\subseteq \ensuremath{\mathcal{X}}$, $\phi(X')$ red and $\phi(X')\cap \ensuremath{\mathcal{X}}=X'$.
\end{observation}
The key ingredient for the proof of the Duality Theorem, Theorem \ref{thm_duality}, is the following characterization of embeddable vertices.
\begin{lemma}\label{lem3}
Let $X\subseteq\ensuremath{\mathcal{X}}$, $Y\subseteq \ensuremath{\mathcal{Y}}$. Let $Q= \ensuremath{\mathcal{Q}} (\ensuremath{\mathcal{X}}\cup \ensuremath{\mathcal{Y}})$ be a blue/red colored Boolean lattice with no blue copy of $\Lambda$. Then $(X,Y)$ is embeddable if and only if either
\begin{enumerate}[label=(\roman*)]
\item $(X,Y)$ is blue and there is a $Y'\subseteq\ensuremath{\mathcal{Y}}$ with $Y'\supsetneq Y$ such that $(X,Y')$ is embeddable, or
\item $(X,Y)$ is red and for all $X'\subseteq\ensuremath{\mathcal{X}}$ with $X'\supsetneq X$, \ $(X',Y)$ is embeddable.
\end{enumerate}
\end{lemma}
Note that if $X\subseteq\ensuremath{\mathcal{X}}$ and $(X,\ensuremath{\mathcal{Y}})$ is blue, then $(X,\ensuremath{\mathcal{Y}})$ is not embeddable.
\begin{proof}
First suppose that $(X,Y)$ is embeddable. Let $\phi$ be an embedding of $\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}})\cap\{X'\subseteq\ensuremath{\mathcal{X}}\colon X'\supseteq X\}$ into $Q$ witnessing that $(X,Y)$ is embeddable.
If $(X,Y)$ is blue, then $\phi(X)\supsetneq(X,Y)$ because $\phi$ has a monochromatic red image.
Thus there exists $Y'\subseteq\ensuremath{\mathcal{Y}}$ with $Y'\supsetneq Y$ such that $\phi(X)=(X,Y')$.
But then $\phi$ also witnesses that $(X,Y')$ is embeddable, so Condition (i) is fulfilled.
If $(X,Y)$ is red, pick some arbitrary $X^*\subseteq\ensuremath{\mathcal{X}}$ such that $X^*\supsetneq X$.
Then the function $\phi^*:\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}})\cap\{X'\subseteq\ensuremath{\mathcal{X}}\colon X'\supseteq X^*\}\to Q$, $\phi^*(X')=\phi(X')$ is a restriction of $\phi$
and therefore an embedding with a monochromatic red image such that $\phi^*(X')\cap \ensuremath{\mathcal{X}}=X'$ for all $X'$ and $\phi^*(X^*)\supseteq (X^*,Y)$.
Thus for every $X^*\subseteq\ensuremath{\mathcal{X}}$ with $X^*\supsetneq X$, the vertex $(X^*,Y)$ is embeddable, i.e., Condition (ii) is fulfilled.\\
Now, suppose that Condition (i) or Condition (ii) hold.
If (i) holds, then $(X,Y)$ is blue and there is some $Y'\supsetneq Y$ such that $(X,Y')$ is embeddable. Then the embedding witnessing that also verifies that $(X,Y)$ is embeddable.\\
For the rest of the proof we assume that (ii) holds, i.e., that $(X,Y)$ is red and for any $X'\subseteq\ensuremath{\mathcal{X}}$ with $X'\supsetneq X$ the vertex $(X',Y)$ is embeddable.
We define the required embedding $\phi: \ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}})\cap\{X'\subseteq\ensuremath{\mathcal{X}}\colon X'\supseteq X\}\to Q$ for every $X'$ with $X\subseteq X'\subseteq\ensuremath{\mathcal{X}}$
depending on the number of minimal $X^*$'s, $X\subseteq X^*\subseteq X'$ such that $(X^*,Y)$ is blue as follows. Let $X'$ with $X\subseteq X'\subseteq\ensuremath{\mathcal{X}}$ be arbitrary.
\begin{enumerate}[label=(\arabic*)]
\item If for all $X^*$ with $X\subseteq X^*\subseteq X'$, the vertex $(X^*,Y)$ is red, let $\phi(X')=(X',Y)$. Note that this case includes $X'=X$.
\item If there is a unique minimal $X^*$ such that $X\subseteq X^*\subseteq X'$ and $(X^*,Y)$ is blue, then $(X^*,Y)$ is embeddable by Condition (ii).
Let $\phi_{X^*}$ be an embedding witnessing that. Then set $\phi(X')=\phi_{X^*}(X')$.
\item Otherwise, let $\phi(X')=(X',\ensuremath{\mathcal{Y}})$.
\end{enumerate}
Cases (1)-(3) determine a partition of
the set $\{X'\subseteq\ensuremath{\mathcal{X}}\colon X'\supseteq X\}$ into three pairwise disjoint parts.
Let $\ensuremath{\mathcal{M}}_j$, $j\in[3]$, be the set of those vertices $X'$ for which $\phi$ was assigned in Case (j). Note that $\ensuremath{\mathcal{M}}_1\cup\ensuremath{\mathcal{M}}_2\cup\ensuremath{\mathcal{M}}_3=\{X'\subseteq\ensuremath{\mathcal{X}}\colon X'\supseteq X\}$.
\\
\noindent \textbf{Claim.} ~~ The function $\phi$ witnesses that $(X,Y)$ is embeddable.
\begin{itemize}
\item Clearly, for every $X'\subseteq\ensuremath{\mathcal{X}}$ with $X'\supseteq X$, $\phi(X')\cap \ensuremath{\mathcal{X}}=X'$.
\item $(X,Y)$ is red, so $X\in\ensuremath{\mathcal{M}}_1$. Thus $\phi(X)=(X,Y)$.
\item The argument verifying that $\phi(X')$ is red for every $X'\subseteq\ensuremath{\mathcal{X}}$ with $X'\supseteq X$ depends on $i$ such that $X'\in\ensuremath{\mathcal{M}}_i$.
If $X'\in\ensuremath{\mathcal{M}}_1$, it is immediate that $\phi(X')$ is red. If $X'\in\ensuremath{\mathcal{M}}_2$, $\phi_{X^*}$ has a monochromatic red image, thus $\phi(X')=\phi_{X^*}(X')$ is also red.
Now consider the case that $X'\in\ensuremath{\mathcal{M}}_3$, i.e.\ there are $X_1,X_2\subseteq \ensuremath{\mathcal{X}}$ with $X_1\neq X_2$ and $X\subseteq X_i\subseteq X'$, $i\in[2]$,
such that $(X_i,Y)$ are blue and $X_i$ are both minimal with this property. The latter condition implies that $X_1$ and $X_2$ are incomparable,
in particular $(X_1,Y)$ and $(X_2,Y)$ are incomparable as well.
Moreover, observe that $X_i\neq X'$, $i\in[2]$, because $X'$ is by definition comparable to both $X_1$ and $X_2$.
Now assume for a contradiction that $\phi(X')=(X',\ensuremath{\mathcal{Y}})$ is blue. Recall that $(X_1,Y)$ and $(X_2,Y)$ are blue.
We know that $X_i\ensuremath{\subset} X'$ and $Y\subseteq\ensuremath{\mathcal{Y}}$, thus $(X_i,Y)\ensuremath{\subset} (X',\ensuremath{\mathcal{Y}})$.
As a consequence, $(X_1,Y)$, $(X_2,Y)$ and $(X',\ensuremath{\mathcal{Y}})$ induce a blue copy of $\Lambda$ in $Q$, which is a contradiction. Thus $\phi$ has a monochromatic red image.
\item It remains to show that $\phi$ is an embedding. Let $X_1,X_2\subseteq \ensuremath{\mathcal{X}}$ be arbitrary with $X\subseteq X_i\subseteq X'$, $i\in[2]$.
We shall show that $X_1\subseteq X_2$ if and only if $\phi(X_1)\subseteq \phi(X_2)$.
One direction is easy to prove: If $\phi(X_1)\subseteq \phi(X_2)$, then $X_1=\phi(X_1)\cap \ensuremath{\mathcal{X}}\subseteq \phi(X_2)\cap \ensuremath{\mathcal{X}}=X_2$.
Now suppose that $X_1\subseteq X_2$.
Let $Y_1,Y_2\subseteq\ensuremath{\mathcal{Y}}$ such that $\phi(X_1)=(X_1,Y_1)$ and $\phi(X_2)=(X_2,Y_2)$. Then we shall show that $Y_1\subseteq Y_2$.
Note that $Y\subseteq Y_i \subseteq \ensuremath{\mathcal{Y}}$ for $i\in[2]$.\\
Assume that at least one of $X_1$ or $X_2$ is in $\ensuremath{\mathcal{M}}_1\cup\ensuremath{\mathcal{M}}_3$.
If $X_1\in\ensuremath{\mathcal{M}}_1$, then $Y_1=Y$ and we are done as $Y\subseteq Y_2$.
Furthermore, if $X_2\in\ensuremath{\mathcal{M}}_3$, then $Y_2=\ensuremath{\mathcal{Y}}$ and we are done as well since $Y_1\subseteq\ensuremath{\mathcal{Y}}$.
If $X_1\in\ensuremath{\mathcal{M}}_3$, then $X_1\subseteq X_2$ implies that $X_2$ is also in $\ensuremath{\mathcal{M}}_3$.
Conversely, if $X_2\in\ensuremath{\mathcal{M}}_1$, the fact that $X_2\supseteq X_1$ yields that $X_1\in\ensuremath{\mathcal{M}}_1$ and we are done as before.
\\
As a final step, suppose that $X_1,X_2\in\ensuremath{\mathcal{M}}_2$.
This implies that for each $i\in[2]$, there is a unique minimal vertex $X^*_i$ such that $X\subseteq X^*_i\subseteq X_i$ and $(X^*_i,Y)$ is blue.
Applying the initial assumption, $X^*_1\subseteq X_1\subseteq X_2$. By minimality of $X^*_2$, we obtain that $X^*_2\subseteq X^*_1$.
Now this provides that $X^*_2\subseteq X^*_1\subseteq X_1$. Using the minimality of $X^*_1$, we see that $X^*_1\subseteq X^*_2$, so $X^*_1=X^*_2$.
\\
Recall that $(X^*_1,Y)=(X^*_2,Y)$ is embeddable since $X_1,X_2\in\ensuremath{\mathcal{M}}_2$. Consider the function $\phi_{X^*_1}=\phi_{X^*_2}$ witnessing that.
Because $\phi_{X^*_1}$ is an embedding and $X^*_1\subseteq X_1\subseteq X_2$, we obtain $\phi_{X^*_1}(X_1)\subseteq\phi_{X^*_1}(X_2)$.
Combining the given conditions, $$\phi(X_1)=\phi_{X^*_1}(X_1)\subseteq\phi_{X^*_1}(X_2)=\phi_{X^*_2}(X_2)=\phi(X_2),$$
that implies that $Y_1\subseteq Y_2$.
\end{itemize}
This concludes the proof of the Claim and the Lemma.
\end{proof}
\begin{corollary}\label{cor3}
Let $X\subseteq \ensuremath{\mathcal{X}}$ and $S\in\ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}})$ such that $(X,\underline{S})$ is not embeddable.
Then there exists some $X'\subseteq \ensuremath{\mathcal{X}}$, $X'\supseteq X$, such that $(X',\underline{S})$ is blue and not embeddable.
\end{corollary}
\begin{proof}
If $(X,\underline{S})$ is blue, we are done. Otherwise Lemma \ref{lem3} yields an $X_1\subseteq \ensuremath{\mathcal{X}}$, $X_1\supsetneq X$ such that $(X_1,\underline{S})$ is not embeddable.
By repeating this argument, we find an $X'\subseteq \ensuremath{\mathcal{X}}$, $X'\supseteq X$, with $(X',\underline{S})$ is blue and not embeddable.
\end{proof}
Next we show a connection between embeddable vertices and the existence of a weak $\ensuremath{\mathcal{Y}}$-shrub.
Recall that a weak $\ensuremath{\mathcal{Y}}$-shrub is a subposet $\ensuremath{\mathcal{P}}$ of $Q$ such that there is a function $\tau: \ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}}) \to Q$ with image $\ensuremath{\mathcal{P}}$
such that for every $S\in\ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}})$, \ $\tau(S)\cap \ensuremath{\mathcal{Y}}=\underline{S}$, and for every $S,T\in \ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}})$ with $S<_\ensuremath{\mathcal{O}} T$, \ $\tau(S)\ensuremath{\subset} \tau(T)$.
\begin{lemma}\label{lem4}
If $(\varnothing,\varnothing)$ is not embeddable, then there is a monochromatic blue weak $\ensuremath{\mathcal{Y}}$-shrub.
\end{lemma}
\begin{proof}
We construct $\tau:\ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}})\to Q$ iteratively and increasingly with respect to the order of $\ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}})$.
Suppose that $(\varnothing,\varnothing)$ is not embeddable.
By Corollary \ref{cor3} there is some $X_\varnothing\subseteq \ensuremath{\mathcal{X}}$ such that $(X_\varnothing,\varnothing)$ is blue and not embeddable.
Let $\tau(\varnothing)=(X_\varnothing,\varnothing)$.
From here, we continue iteratively. Suppose that for $S\in\ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}})$, $\underline{S}\neq\ensuremath{\mathcal{Y}}$, we have defined $X_S\subseteq \ensuremath{\mathcal{X}}$ such that
\begin{enumerate}[label=(\arabic*)]
\item $X_S\supseteq X_T$ for every $T\le_\ensuremath{\mathcal{O}} S$ and
\item $\tau(S)=(X_S,\underline{S})$ is blue and not embeddable.
\end{enumerate}
Consider an arbitrary $S'\in\ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}})$ such that $S<_\ensuremath{\mathcal{O}} S'$ and $|S'|=|S|+1$.
By Lemma \ref{lem3} applied for $X_S$ and $\underline{S}$, we obtain that $(X_S,\underline{S'})$ is not embeddable.
Then Corollary \ref{cor3} yields that there is some $X_{S'}\subseteq\ensuremath{\mathcal{X}}$, $X_{S'}\supseteq X_S$, such that $(X_{S'},\underline{S'})$ is blue and not embeddable.
Observe that for $T\in\ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}})$ with $T\le_\ensuremath{\mathcal{O}} S'$, either $T=S'$ and so $X_T=X_{S'}$, or $T\le_\ensuremath{\mathcal{O}} S$ and so by $(1)$ $X_T\subseteq X_S\subseteq X_{S'}$.
Let $\tau(S')=(X_{S'},\underline{S'})$.
\\
Using this procedure, we define $\tau$ for all $S\in\ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}})$. Let $\ensuremath{\mathcal{P}}$ be the subposet of $Q$ induced by the image of $\tau$.
We shall show that $\ensuremath{\mathcal{P}}$ is a weak $\ensuremath{\mathcal{Y}}$-shrub witnessed by the function $\tau$.
By (2), for every $S\in \ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}})$, \ $\tau(S)$ is blue and $\tau(S)\cap \ensuremath{\mathcal{Y}}=\underline{S}$.
Let $S,T\in\ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}})$ with $S<_\ensuremath{\mathcal{O}} T$. Let $X_S,X_T\subseteq\ensuremath{\mathcal{X}}$ such that $\tau(S)=(X_S,\underline{S})$ and $\tau(T)=(X_T,\underline{T})$.
Clearly, $\underline{S}\ensuremath{\subset} \underline{T}$. Moreover, item $(1)$ implies that $X_S\subseteq X_T$.
Consequently, $\tau(S)\ensuremath{\subset}\tau(T)$.
\end{proof}
Combining the previously presented Lemmas, we can now prove the Duality Theorem.
\begin{proof}[Proof of Theorem \ref{thm_duality}]
Let $\ensuremath{\mathcal{X}}$ and $\ensuremath{\mathcal{Y}}$ be disjoint sets.
Let $Q=Q(\ensuremath{\mathcal{X}}\cup \ensuremath{\mathcal{Y}})$ be a blue/red colored Boolean lattice which contains no blue copy of $\Lambda$. \\
First suppose that there is no red $\ensuremath{\mathcal{X}}$-good copy of $\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}})$.
By Observation \ref{lem2}, $(\varnothing,\varnothing)$ is not embeddable and Lemma \ref{lem4} provides that there is a blue weak $\ensuremath{\mathcal{Y}}$-shrub in $Q$.
Using Proposition \ref{obs_shrub} we obtain a blue $\ensuremath{\mathcal{Y}}$-shrub in $Q$.
This shows that there is either a red $\ensuremath{\mathcal{X}}$-good copy of $\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}})$ or a blue $\ensuremath{\mathcal{Y}}$-shrub. \\
Next we show that both events could not happen simultaneously.
Let $n=|\ensuremath{\mathcal{X}}|$, $k=|\ensuremath{\mathcal{Y}}|$ and $N=n+k$.
Assume that there exist both an $\ensuremath{\mathcal{X}}$-good embedding $\phi\colon \ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}}) \to Q$ with monochromatic red image
as well as a $\ensuremath{\mathcal{Y}}$-good embedding $\tau\colon \ensuremath{\mathcal{O}}(\ensuremath{\mathcal{Y}}) \to Q$ with a monochromatic blue image.
\\
We apply an iterative argument in order to find a contradiction. Let $Y_0=\varnothing$ and let $S_0=(Y_0,\le)$ be the empty ordered set.
Now let $X_1\subseteq\ensuremath{\mathcal{X}}$ such that $\tau(S_0)=(X_1,\underline{S_0})$ and let $Y_1\in\ensuremath{\mathcal{Y}}$ such that $\phi(X_1)=(X_1,Y_1)$.
Since $\phi(X_1)$ is red but $\tau(S_0)$ is blue, we know that $\phi(X_1)\neq\tau(S_0)$ and thus $Y_1\neq \underline{S_0}=\varnothing$, so $|Y_1|\ge 1$.
\\
Now say that we already defined $X_1,\dots,X_i$, $Y_0,\ldots,Y_i$, $S_0,\dots,S_{i-1}$ for some $i\in[k]$ such that
\begin{itemize}
\item $S_{i-1}\in\ensuremath{\mathcal{O}}$ and $\underline{S_{i-1}}=Y_{i-1}$,
\item $\tau(S_{i-1})=(X_i,\underline{S_{i-1}})$,
\item $\phi(X_i)=(X_i,Y_i)$, and
\item $Y_{i-1}\ensuremath{\subset} Y_i\subseteq \ensuremath{\mathcal{Y}}$ and $|Y_i|\ge i$.
\end{itemize}
Fix any ordering $S_i$ of $Y_i$ such that $S_{i-1}<_\ensuremath{\mathcal{O}} S_i$. Such $S_i$ exists because $\underline{S_{i-1}}=Y_{i-1}\ensuremath{\subset} Y_i$.
Then let $X_{i+1}$ be such that $\tau(S_i)=(X_{i+1},\underline{S_i})$.
Since $S_{i-1}<_\ensuremath{\mathcal{O}} S_i$ and $\tau$ is an embedding, $(X_{i},\underline{S_{i-1}})=\tau(S_{i-1})\subseteq\tau(S_i)=(X_{i+1},\underline{S_i})$, therefore $X_i\subseteq X_{i+1}$.
Note that $\phi(X_i)=(X_i,\underline{S_i})$ is colored red but $\tau(S_i)=(X_{i+1},\underline{S_i})$ is blue.
Therefore $X_{i}\neq X_{i+1}$, consequently $X_i\ensuremath{\subset} X_{i+1}$ and in particular $\phi(X_i)\ensuremath{\subset}\phi(X_{i+1})$ because $\phi$ is an embedding.
Next let $Y_{i+1}\subseteq\ensuremath{\mathcal{Y}}$ such that $\phi(X_{i+1})=(X_{i+1},Y_{i+1})$.
Then $Y_{i+1}\supseteq Y_i$ and furthermore, because $(X_{i+1},Y_i)$ is blue but $\phi(X_{i+1})$ is red, $Y_{i+1}\neq Y_i$.
Consequently $Y_{i+1}\supsetneq Y_i$, and in particular $|Y_{i+1}|\ge|Y_i|+1\ge i+1$.
\\
Iteratively, we obtain $Y_{k+1}\subseteq \ensuremath{\mathcal{Y}}$ with $|Y_i|\ge k+1$, a contradiction to $|\ensuremath{\mathcal{Y}}|=k$.
\end{proof}
\section{Random coloring with many blue shrubs} \label{sec_random}
We shall provide a coloring that will give us a lower bound on $R(\Lambda, Q_n)$. Note that we do not provide an explicit construction but only prove the existence of such a coloring.
\begin{theorem}\label{thm_LB}
Let $N\in\ensuremath{\mathbb{N}}$ be sufficiently large and $k=\frac{10}{216}\frac{N}{\ln(N)}$. Consider the Boolean lattice $Q=\ensuremath{\mathcal{Q}}([N])$.
Then for sufficiently large $N$, there exists a blue/red coloring of $Q$ which contains no blue copy of $\Lambda$ and
such that for each $\ensuremath{\mathcal{Y}}\in\binom{[N]}{k}$, there is a blue $\ensuremath{\mathcal{Y}}$-shrub in $Q$.
\end{theorem}
\begin{proof}[Proof of Theorem \ref{thm_LB}]
Let $\alpha= 21.6$ and $\beta=0.134$.
Let $N\in\ensuremath{\mathbb{N}}$ and $k=\frac{1}{\alpha}\frac{N}{\ln(N)}$, let $Q=\ensuremath{\mathcal{Q}}([N])$.
The idea of the proof is to construct a $\ensuremath{\mathcal{Y}}$-shrub, denoted $\ensuremath{\mathcal{P}}_{\ensuremath{\mathcal{Y}}}$, for every $\ensuremath{\mathcal{Y}}\in\binom{[N]}{k}$, with an additional property so that the selected shrubs are independent.
Since each shrub does not contain a copy of $\Lambda$, it implies that the independent union of all the $\ensuremath{\mathcal{P}}_{\ensuremath{\mathcal{Y}}}$'s also does not contain a copy of $\Lambda$.
We obtain these shrubs by randomly choosing a \textit{$\ensuremath{\mathcal{Y}}$-framework} for every $\ensuremath{\mathcal{Y}}\in\binom{[N]}{k}$ and then constructing a $\ensuremath{\mathcal{Y}}$-shrub based on each of them.
Afterwards we define a coloring where every vertex in each constructed shrub is colored blue and the remaining vertices red.
\\
\noindent A \textit{$\ensuremath{\mathcal{Y}}$-framework} of $\ensuremath{\mathcal{Y}}\in\binom{[N]}{k}$ is a $4$-tuple $(\ensuremath{\mathcal{Y}},A_\ensuremath{\mathcal{Y}},Z_\ensuremath{\mathcal{Y}}, X_\ensuremath{\mathcal{Y}})$ such that
\begin{itemize}
\item $\ensuremath{\mathcal{Y}}$, $A_\ensuremath{\mathcal{Y}}$ $Z_\ensuremath{\mathcal{Y}}$ are pairwise disjoint and $\ensuremath{\mathcal{Y}}\cup A_\ensuremath{\mathcal{Y}}\cup Z_\ensuremath{\mathcal{Y}}=[N]$,
\item $|A_\ensuremath{\mathcal{Y}}|=\tfrac{3}{2}k\ln k-k$,
\item $X_\ensuremath{\mathcal{Y}}\subseteq Z_\ensuremath{\mathcal{Y}}$.
\end{itemize}
\noindent A $\ensuremath{\mathcal{Y}}$-framework is \textit{random} if
\begin{itemize}
\item $A_\ensuremath{\mathcal{Y}}\in\binom{[N]\backslash\ensuremath{\mathcal{Y}}}{\tfrac{3}{2}k\ln k-k}$ is chosen uniformly at random, and
\item each element of $Z_\ensuremath{\mathcal{Y}}=[N]\backslash(\ensuremath{\mathcal{Y}}\cup A_\ensuremath{\mathcal{Y}})$ is included in $X_\ensuremath{\mathcal{Y}}$ independently at random with probability $\frac{1}{2}$.
\end{itemize}
Now draw a random $\ensuremath{\mathcal{Y}}$-framework for every $\ensuremath{\mathcal{Y}}\in\binom{[N]}{k}$.
Observe that by choice of $k$, $k\ln k =\frac{N}{\alpha}\cdot \frac{\ln(N)-\ln(\alpha)-\ln\ln(N)}{\ln(N)}$, so $\frac{20N}{21\alpha} \le k\ln k\le \frac{N}{\alpha}$.
Since $|Z_\ensuremath{\mathcal{Y}}|= N-\tfrac{3}{2}k\ln k$, we have $(1-\frac{3}{2\alpha})N\le|Z_\ensuremath{\mathcal{Y}}|\le (1-\frac{10}{7\alpha})N$.
\\
\noindent \textbf{Claim 1.} ~~ W.h.p.\ for every $\ensuremath{\mathcal{Y}}_1,\ensuremath{\mathcal{Y}}_2\in\binom{[N]}{k}$ with $\ensuremath{\mathcal{Y}}_1\neq\ensuremath{\mathcal{Y}}_2$, \ $|X_{\ensuremath{\mathcal{Y}}_1}\cap Z_{\ensuremath{\mathcal{Y}}_2}|\ge \beta N$.\\
Consider some arbitrary $\ensuremath{\mathcal{Y}}_1,\ensuremath{\mathcal{Y}}_2\in\binom{[N]}{k}$, $\ensuremath{\mathcal{Y}}_1\neq\ensuremath{\mathcal{Y}}_2$. Observe that $(1-\frac{3}{\alpha})N\le|Z_{\ensuremath{\mathcal{Y}}_1}\cap Z_{\ensuremath{\mathcal{Y}}_2}|\le|Z_\ensuremath{\mathcal{Y}}|\le(1-\frac{10}{7\alpha})N$.
In a random $\ensuremath{\mathcal{Y}}$-framework, each element of $Z_{\ensuremath{\mathcal{Y}}_1}\cap Z_{\ensuremath{\mathcal{Y}}_2}$ is contained in $X_{\ensuremath{\mathcal{Y}}_1}\cap Z_{\ensuremath{\mathcal{Y}}_2}$ independently with probability $\frac12$.
Consequently, $|X_{\ensuremath{\mathcal{Y}}_1}\cap Z_{\ensuremath{\mathcal{Y}}_1}|\sim \text{Bin}(|Z_{\ensuremath{\mathcal{Y}}_1}\cap Z_{\ensuremath{\mathcal{Y}}_2}|,\frac12)$ and $\ensuremath{\mathbb{E}}(|X_{\ensuremath{\mathcal{Y}}_1}\cap Z_{\ensuremath{\mathcal{Y}}_1}|)=\frac12 |Z_{\ensuremath{\mathcal{Y}}_1}\cap Z_{\ensuremath{\mathcal{Y}}_2}|$.
We have $1-\frac3\alpha\ge2\beta$. In addition, $\frac{\big(\frac{1}{2}-\frac{3}{2\alpha}-\beta\big)^2}{1-\frac{10}{7\alpha}}>\frac{2}{\alpha}$, thus there exist some $\epsilon>0$ such that $\frac{\big(\frac{1}{2}-\frac{3}{2\alpha}-\beta\big)^2}{1-\frac{10}{7\alpha}}\geq \epsilon + \frac{2}{\alpha}$.
Applying Chernoff's inequality gives
\begin{eqnarray*}
\ensuremath{\mathbb{P}}(|X_{\ensuremath{\mathcal{Y}}_1}\cap Z_{\ensuremath{\mathcal{Y}}_2}|\le \beta N)&=& \ensuremath{\mathbb{P}}\left(|X_{\ensuremath{\mathcal{Y}}_1}\cap Z_{\ensuremath{\mathcal{Y}}_2}|\le\frac{|Z_{\ensuremath{\mathcal{Y}}_1}\cap Z_{\ensuremath{\mathcal{Y}}_2}|}{2}-\left(\frac{|Z_{\ensuremath{\mathcal{Y}}_1}\cap Z_{\ensuremath{\mathcal{Y}}_2}|}{2}-\beta N\right)\right)\\
&\le & \exp\left(-\frac{\big(\frac{|Z_{\ensuremath{\mathcal{Y}}_1}\cap Z_{\ensuremath{\mathcal{Y}}_2}|}{2}-\beta N\big)^2}{|Z_{\ensuremath{\mathcal{Y}}_1}\cap Z_{\ensuremath{\mathcal{Y}}_2}|}\right)\\
&\le & \exp\left(-\frac{\big((\frac{1}{2}-\frac{3}{2\alpha})-\beta\big)^2}{(1-\frac{10}{7\alpha})}\cdot N\right)\\
&\le & \exp\left(-\left(\frac{2}{\alpha}+\epsilon\right)\cdot N\right).
\end{eqnarray*}
\noindent
Let $E_1$ be the event that for some distinct $\ensuremath{\mathcal{Y}}_1,\ensuremath{\mathcal{Y}}_2\in\binom{[N]}{k}$, $|X_{\ensuremath{\mathcal{Y}}_1}\cap Z_{\ensuremath{\mathcal{Y}}_2}|\le \beta N$. Then
\begin{eqnarray*}
\ensuremath{\mathbb{P}}(E_1)&= & \binom{N}{k}\left(\binom{N}{k}-1\right) \ensuremath{\mathbb{P}}(|X_{\ensuremath{\mathcal{Y}}_1}\cap Z_{\ensuremath{\mathcal{Y}}_2}|\le \beta N)\\
&\le & N^{2k} \exp\left(-\left(\frac{2}{\alpha}+\epsilon\right)\cdot N\right)\\
&\le & \exp\left(\frac{2N\ln (N)}{\alpha\ln (N)} - \left(\frac{2}{\alpha}+\epsilon\right)\cdot N\right)\\
&=& \exp(-\epsilon N)\to 0, \mbox{ as } N\to\infty.
\end{eqnarray*}
This proves Claim 1.\\
\noindent \textbf{Claim 2.} ~~ W.h.p.\ for every $\ensuremath{\mathcal{Y}}_1,\ensuremath{\mathcal{Y}}_2\in\binom{[N]}{k}$ with $\ensuremath{\mathcal{Y}}_1\neq\ensuremath{\mathcal{Y}}_2$, \ $X_{\ensuremath{\mathcal{Y}}_1}\cap Z_{\ensuremath{\mathcal{Y}}_2}\not\subseteq X_{\ensuremath{\mathcal{Y}}_2}$.\\
We can suppose that the collection of random frameworks fulfills the property of Claim 1.
Let $\ensuremath{\mathcal{Y}}_1,\ensuremath{\mathcal{Y}}_2\in\binom{[N]}{k}$ be such that $\ensuremath{\mathcal{Y}}_1\neq\ensuremath{\mathcal{Y}}_2$.
Note that each element of $X_{\ensuremath{\mathcal{Y}}_1}\cap Z_{\ensuremath{\mathcal{Y}}_2}$ is contained in $X_{\ensuremath{\mathcal{Y}}_2}$ with probability $\frac{1}{2}$. Thus,
$$\ensuremath{\mathbb{P}}(X_{\ensuremath{\mathcal{Y}}_1}\cap Z_{\ensuremath{\mathcal{Y}}_2}\subseteq X_{\ensuremath{\mathcal{Y}}_2})=\left(\frac{1}{2}\right)^{|X_{\ensuremath{\mathcal{Y}}_1}\cap Z_{\ensuremath{\mathcal{Y}}_2}|}\le 2^{-\beta N}.$$
Let $E_2$ be the event that there exist $\ensuremath{\mathcal{Y}}_1,\ensuremath{\mathcal{Y}}_2\in\binom{[N]}{k}$ with $\ensuremath{\mathcal{Y}}_1\neq\ensuremath{\mathcal{Y}}_2$ such that $X_{\ensuremath{\mathcal{Y}}_1}\cap Z_{\ensuremath{\mathcal{Y}}_2}\not\subseteq X_{\ensuremath{\mathcal{Y}}_2}$.
Since $\frac{2}{\alpha}<\ln (2)\beta$, we have
$$\ensuremath{\mathbb{P}}(E_2)\le
N^{2k} \ensuremath{\mathbb{P}}(X_{\ensuremath{\mathcal{Y}}_1}\cap Z_{\ensuremath{\mathcal{Y}}_2}\subseteq X_{\ensuremath{\mathcal{Y}}_2})
\le N^{2k}\cdot 2^{-\beta N}=\exp\left(\frac{2}{\alpha}N-\ln (2)\beta N\right)\to 0, \mbox{ as } N\to\infty.$$
This proves Claim 2.\\
In particular, there exists a collection of $\ensuremath{\mathcal{Y}}$-frameworks $(\ensuremath{\mathcal{Y}},A_\ensuremath{\mathcal{Y}},Z_\ensuremath{\mathcal{Y}}, X_\ensuremath{\mathcal{Y}})$, $\ensuremath{\mathcal{Y}}\in\binom{[N]}{k}$, such that
for every $\ensuremath{\mathcal{Y}}_1,\ensuremath{\mathcal{Y}}_2\in\binom{[N]}{k}$ with $\ensuremath{\mathcal{Y}}_1\neq\ensuremath{\mathcal{Y}}_2$, \ $X_{\ensuremath{\mathcal{Y}}_1}\cap Z_{\ensuremath{\mathcal{Y}}_2}\not\subseteq X_{\ensuremath{\mathcal{Y}}_2}$.
\\
Note that $|A_{\ensuremath{\mathcal{Y}}}|=\tfrac{3}{2}k\ln k-k\ge k(\log k + \log\log k)$.
Let $\ensuremath{\mathcal{P}}'_{\ensuremath{\mathcal{Y}}}$ be a $\ensuremath{\mathcal{Y}}$-shrub in $\ensuremath{\mathcal{Q}}(A_{\ensuremath{\mathcal{Y}}}\cup \ensuremath{\mathcal{Y}})$ as guaranteed by Lemma \ref{canonical-shrub}.
Note that $\ensuremath{\mathcal{P}}_{\ensuremath{\mathcal{Y}}}$'s are not necessarily independent.
Let $\ensuremath{\mathcal{P}}_{\ensuremath{\mathcal{Y}}}$ be obtained from $\ensuremath{\mathcal{P}}'_{\ensuremath{\mathcal{Y}}}$ by replacing each vertex $W$ of $\ensuremath{\mathcal{P}}'_{\ensuremath{\mathcal{Y}}}$ with $W\cup X_{\ensuremath{\mathcal{Y}}}$.
Then $\ensuremath{\mathcal{P}}_{\ensuremath{\mathcal{Y}}}$ is a $\ensuremath{\mathcal{Y}}$-shrub in $Q$. \\
\noindent \textbf{Claim 3.} ~~ Let $\ensuremath{\mathcal{Y}}_1, \ensuremath{\mathcal{Y}}_2$ be two distinct $k$-element subsets of $[N]$. Then $\ensuremath{\mathcal{P}}_{\ensuremath{\mathcal{Y}}_1}$ and $\ensuremath{\mathcal{P}}_{\ensuremath{\mathcal{Y}}_2}$ are independent.\\
Consider arbitrary elements $U_i\in \ensuremath{\mathcal{P}}_{\ensuremath{\mathcal{Y}}_i}$, $i\in[2]$. Recall that $X_{\ensuremath{\mathcal{Y}}_1}\cap Z_{\ensuremath{\mathcal{Y}}_2}\not\subseteq X_{\ensuremath{\mathcal{Y}}_2}$, which implies that
there exists some $z\in (X_{\ensuremath{\mathcal{Y}}_1}\cap Z_{\ensuremath{\mathcal{Y}}_2})\backslash X_{\ensuremath{\mathcal{Y}}_2}$. Note that $z\in U_1$ since $X_{\ensuremath{\mathcal{Y}}_1}\subseteq U_1$. Moreover
$z\not\in U_2$ since $z\in Z_{\ensuremath{\mathcal{Y}}_2}\setminus X_{\ensuremath{\mathcal{Y}}_2}$ and $(Z_{\ensuremath{\mathcal{Y}}_2}\setminus X_{\ensuremath{\mathcal{Y}}_2})\cap U_2 = \varnothing$.
In particular, $z\in U_1\backslash U_2$. Similarly, there is an element $w\in U_2\backslash U_1$. Thus $U_1\not\sim U_2$. \qed\\
\bigskip
We consider the following coloring $c: Q\to \{\text{blue, red}\}$. For $X\subseteq [N]$, let
\begin{equation} \nonumber
c(X) =
\begin{cases}
\text{blue}, \hspace*{0.75cm}\mbox{ if } ~~ X\in \bigcup_{\ensuremath{\mathcal{Y}}\in\binom{[N]}{k}} \ensuremath{\mathcal{P}}_\ensuremath{\mathcal{Y}},\\
\text{red}, \hspace*{0.8cm}\mbox{ otherwise.}
\end{cases}
\end{equation}
Note that for every $\ensuremath{\mathcal{Y}}\in\binom{[N]}{k}$, $\ensuremath{\mathcal{P}}_\ensuremath{\mathcal{Y}}$ witnesses that there is a blue $\ensuremath{\mathcal{Y}}$-shrub in $Q$.
Recall that a $\ensuremath{\mathcal{Y}}$-shrub is an up-tree. Applying Claim 3 the blue subposet of $Q$ is a collection of independent up-trees.
Then Lemma \ref{strucL} provides that the coloring $c$ does not contain a blue copy of $\Lambda$.
\end{proof}
\section{Proof of Theorem \ref{thm_main} and Theorem \ref{thm-MAIN}}\label{sec_proof}
\begin{proof}[Proof of Theorem \ref{thm_main}]
~\\
\textbf{ Upper Bound:~} Let $k=(1+\epsilon)\frac{n}{\log(n)}$ and consider an arbitrary blue/red colored Boolean lattice $Q$ on ground set $[n+k]$ with no blue copy of $\Lambda$.
Pick any $\ensuremath{\mathcal{Y}}\in\binom{[n+k]}{k}$ and assume that there is a blue $\ensuremath{\mathcal{Y}}$-shrub in $Q$.
Recall that the maximal elements of the $\ensuremath{\mathcal{Y}}$-shrub form an antichain of size $k!$.
Sperner's theorem provides that the largest antichain in $Q$ has size $\binom{n+k}{\lfloor\frac{n+k}{2}\rfloor}$,
so $k!\le \binom{n+k}{\lfloor\frac{n+k}{2}\rfloor}\leq 2^{n+k}$.
\\
We also have that $k!>\left(\tfrac{k}{e}\right)^k=2^{k(\log k - \log e)}.$
By the choice of $k$, we obtain for sufficiently large $n$,
$$k\log k\ge\tfrac{(1+\epsilon)n}{\log n}\big(\log(n)-\log\log(n)\big)>\big(1+\tfrac{\epsilon}{2}\big)n.$$
In particular for sufficiently large $n$,
$k\log k -k\log e>n+k,$
a contradiction.
Thus $Q$ does not contain a blue $\ensuremath{\mathcal{Y}}$-shrub for this fixed $\ensuremath{\mathcal{Y}}$. Then Corollary \ref{cor_duality} yields that there is a red copy of $Q_n$ in $Q$.
Consequently, each blue/red colored Boolean lattice of dimension $n+k$ contains either a blue copy of $\Lambda$ or a red copy of $Q_n$.\\
\noindent
\textbf{Lower Bound:~} Let $N$ sufficiently large, let $k=\frac{10}{216}\frac{N}{\ln(N)}$ and $n=N-k$. Note that $k\le \frac{N}{2}$, thus $n\le N\le 2n$. Let $Q=\ensuremath{\mathcal{Q}}([N])$.
By Theorem \ref{thm_LB} there exists a coloring of $Q$ with no blue copy of $\Lambda$ such that for every $\ensuremath{\mathcal{Y}}\in\binom{[N]}{k}$, there is a blue $\ensuremath{\mathcal{Y}}$-shrub.
By Corollary \ref{cor_duality}, there is no red copy of $Q_n$ in this coloring, thus $R(\Lambda,Q_n)\ge N= n+ k$.
It remains to bound $k$ in terms of $n$.
Indeed, $$k=\frac{10}{216}\cdot\frac{N}{\ln(N)}\ge \frac{10}{216}\cdot\frac{n}{\ln(2n)}
= \frac{10}{216}\cdot\frac{\log(e)n}{\log(2n)}\ge \frac{1}{15}\cdot\frac{n}{\log(n)},$$
which concludes the proof.
\end{proof}
\bigskip
\begin{proof}[Proof of Theorem \ref{thm-MAIN}]
The lower bound on $R(P, Q_n)$ for $P$ containing either $\Lambda$ or $V$ follows from Theorem \ref{thm_main}.
Consider now a poset $P$ that contains neither a copy of $\Lambda$ nor a copy of $V$.
By Corollary \ref{strucI}, $P$ is a union of independent chains.
Assume that $P$ has $k$ independent chains on at most $\ell$ vertices each. Let $K$ be an even integer such that $\binom{K}{K/2} \geq k$.
Let $\ensuremath{\mathcal{Y}}$ be a set of size $K$ and let $\ensuremath{\mathcal{X}}$ be a set, disjoint from $\ensuremath{\mathcal{Y}}$ of size $n+ \ell$.
Consider an arbitrary coloring of $\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}}\cup \ensuremath{\mathcal{Y}})$.
Assume that there is no red copy of $Q_n$. We shall show that there is a blue copy of $P$.\\
Let $Y_1, \ldots, Y_k$ form an antichain in $\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{Y}})$, its existence is guaranteed by Sperner's theorem.
Let $Q^i$ be a copy of $\ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}})$ obtained as an image of an embedding $\phi_i: \ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}}) \to \ensuremath{\mathcal{Q}}(\ensuremath{\mathcal{X}}\cup Y_i)$, $\phi_i(X)=X\cup Y_i$ for any $X\subseteq \ensuremath{\mathcal{X}}$.
Consider the blue vertices in $Q^i$.
If there is no blue chain on $\ell$ vertices in $Q^i$, Corollary \ref{shift_cor} implies the existence of a red copy of $Q_n$ in $Q^i$, a contradiction.
Thus for every $i\in[k]$, there is a blue copy $P_i$ of a chain on $\ell$ vertices in $Q^i$.
Note that for any $A\in Q^i, B\in Q^j$, $i\neq j$, \ $A\not\sim B$, since $A\cap\ensuremath{\mathcal{Y}} =Y_i\not\sim Y_j=B\cap\ensuremath{\mathcal{Y}}$.
Thus the $P_i$'s are independent chains on $\ell$ vertices each. Their union contains a copy of $P$. This shows that $R(P, Q_n) \leq n + K+\ell = n+ f(P)$.
\end{proof}
\medskip
\noindent \textbf{Acknowledgments:}~ The authors would like to thank Alexander Riasanovsky for comments on the manuscript. The research was partially supported by DFG grant FKZ AX 93/2-1.
|
\section{Approach}
\vspace*{-0.05in}
Our goal is to train %
agents to efficiently solve interaction-heavy tasks. Training reinforcement learning (RL) agents is difficult due to sparse task rewards, large search spaces, and context-sensitive actions that depend on what the agent is holding and where other objects are. %
We propose an auxiliary reward derived from egocentric videos of daily human activity that encourages agents to configure the environment in ways that facilitate successful activities.
For example, watching humans wash spoons in the sink suggests that the sink is a good location to bring utensils to before interacting with the nearby faucet or dish soap. %
In the following, we begin by defining the visual semantic planning task (\refsec{sec:task_setup}). Then, we show how to infer activity-context priors from egocentric video (\refsec{sec:priors}) and how to translate those priors to an embodied agent in simulation (\refsec{sec:priors_to_thor}). We then describe our approach to encode the priors as a dense reward (\refsec{sec:env_reward}). Finally, we present our policy learning architecture (\refsec{sec:training}).
\vspace{-0.05in}
\subsection{Visual semantic planning} \label{sec:task_setup}
We aim to train an agent to bring the environment into a particular goal state specified by the task $\tau$. Agents can perform navigation actions $\mathcal{A}_N$ (e.g., move forward, rotate left/right) and object interactions $\mathcal{A}_I$ (e.g., take/put, open, toggle) involving an object from set $\mathcal{O}$.
Each task is set up as a partially observable Markov decision process. The agent is spawned at an initial state $s_0$. At each time step $t$, the agent in state $s_t$ receives an observation $(x_t, \theta_t, h_t)$ containing the RGB egocentric view, the agent's current pose, and the currently held object (if any). The agent executes an action on an object $o_t$, $(a_t, o_t) \sim \{\mathcal{A}_N \bigcup \mathcal{A}_I\}$, and receives a reward $r_t \sim \mathcal{R}_{\tau}(s_t, a_t, o_t, s_{t+1})$. For navigation actions, the interaction target $o_t$ is \emph{null}. A recurrent network encodes the agent's observation history over time to arrive at the state representation (detailed below).
A task reward is provided if the agent brings its environment to a goal state $g_\tau$. For example, in the ``Clean object" task %
in our experiments, a \emph{washable} object must be inside the sink, and the faucet must be \emph{toggled-on} for success. A positive reward is given for goal completion, while a small negative reward is given at each time-step to incentivize quick task completion:
\begin{align}
R_{\tau}(s_t, a_t, o_t, s_{t+1}) =
\begin{cases}
10 & \text{if } g_\tau \text{ satisfied, } \\
-0.01 & \text{otherwise.} \\
\end{cases}
\label{eqn:task_reward}
\end{align}
The goal is to learn a policy $\pi_\tau$ that maximizes this reward over an episode of length $T$.
\vspace{-0.05in}
\subsection{Human activity-context from egocentric video} \label{sec:priors}
Learning interaction policies requires an understanding of how the environment needs to be configured---where the agent needs to be, and what objects need to be present there---before goals are completed. %
We infer this directly from a passively collected dataset of human egocentric videos.
The dataset consists of a set of clips $v \in \mathcal{V}$ where each clip captures a %
person performing an interaction with some object %
from a video object vocabulary $\mathcal{O}_V$ (e.g., ``slice tomato", ``pour kettle").
Our model uses the clip's frames only, not the interaction label. Our approach first extracts activity-contexts from each training video frame, and then aggregates their statistics over all training clips to produce an \emph{inter-object activity-context compatibility function}, as follows.
For each frame $f_t$ in an egocentric video clip $v$, we use an off-the-shelf hand-object interaction model~\cite{shan2020understanding} that detects active objects---manipulated objects in contact with hands---resulting in a set of class-agnostic bounding boxes. We associate object labels to these boxes using a pre-trained object instance detection model. Specifically, we transfer labels from high-confidence instance detections to active object boxes with large overlap, namely, where intersection over union (IoU) $>$ 0.5, resulting in a set of active object boxes and corresponding class labels $\mathcal{D}(f_t) = \{(b_0, o_0) .. (b_N, o_N)\}$, where $b_i$ denotes box coordinates and $o_i \in \mathcal{O}_V$.
These instances represent objects that are directly involved in the activity, ignoring background objects that are incidentally visible but not interacted with by hands. Detection model details are in Supp.
We infer frame $f_t$'s activity-context $AC(f_t)$ from the Cartesian product of active objects. %
\begin{align}
AC(f_t) = \left \{
(o_i, o_j)~|~ o_i, o_j \in \mathcal{D}(f_t) \times \mathcal{D}(f_t)
\right\},
\label{eqn:aco_subset}
\end{align}
where $o_i \neq o_j$, and each $o_i$ is an object that can be held and moved to different locations, as opposed to a fixed object like a refrigerator or sink.
We include a \emph{null} object token to consider cases when the agent visits locations empty-handed. %
\reffig{fig:epic_to_acos} (left) shows examples.
Each $(o_i, o_j)$ pair represents a particular object $o_i$ and a corresponding activity-context object (ACO) $o_j$---an object that is used with it in an activity. These include movable objects (e.g., tools like knives and cutting-boards in ``slice tomato"), receptacles (e.g., kettles and cups in ``pour water"), and fixed environment locations (e.g., sinks, faucets in ``wash spoon"). %
This is in contrast to an object's \emph{affordance}, which defines object properties in isolation (e.g. tomatoes are \emph{sliceable} whether a knife is held or not, spoons are \emph{washable} even when they are inside drawers).
Finally, we define the inter-object activity-context compatibility score $\phi(o_i, o_j)$ as follows:
\begin{align}
\phi(o_i, o_j) = \frac{\sum\limits_{v \in \mathcal{V}} S_v(o_i, o_j)}
{\sum\limits_{v \in \mathcal{V}}\sum\limits_{o_k \neq o_i} S_v(o_i, o_k) },
\label{eqn:aco_score}
\end{align}
where $S_v(o_i, o_j)$ measures the fraction of frames of video $v$ where $(o_i, o_j) \in AC(f_t)$. By aggregating across clips, we capture dominant object relationships and reduce the effect of object detection errors.
Recall that these relationships are derived from video frames of objects \emph{in action} rather than arbitrary video frames or static images, and are thus strongly tied to human activity, not static scene properties. In other words, $\phi(o_i, o_j)$ measures how likely $o_i$ is \emph{brought together with} $o_j$ during an activity (e.g., knives and cutting boards are used to cut fruits) rather than how likely they generally co-occur in space (e.g., spoons and knives are stored together, but neither enables the action of the other).
\begin{figure*}[t!]
\centering
\includegraphics[width=\columnwidth]{figures/epic_to_acos}
\caption{\B{Left: Detected ACOs from EPIC videos.} White boxes: active object detections. Red/blue boxes: sampled (object, ACO) tuple (\refeqn{eqn:aco_subset}). %
Last column shows failure cases---sponge not held (top), liquid:washing false positive (bottom). \B{Right: ACO compatibility scores in THOR.} Red: active object, blue: top corresponding ACOs. Edge thickness indicates compatibility (\refeqn{eqn:aco_score}). Our approach (bottom right) prioritizes object relationships for ``Pan'' that are relevant for activity (e.g., StoveBurner, StoveKnob) over semantically similar or spatially co-occurring objects (e.g., Pot, Shelf in top-right, bottom-left respectively). %
}
\label{fig:epic_to_acos}
\vspace*{-0.1in}
\end{figure*}
\vspace{-0.05in}
\subsection{Translating to activity-context for embodied agents} \label{sec:priors_to_thor}
Next, we go from compatible object detections from human video, to interactible objects in embodied agents' environments.
The compatibility score in \refeqn{eqn:aco_score} is defined over the vocabulary %
of detectable objects $\mathcal{O}_V$, which may not be the same as the interactible objects $\mathcal{O}$ in the agent's environment. This mismatch is to be expected since we rely on in-the-wild video that is not carefully curated for the particular environments or tasks faced by the agents we train.
To account for this, we first map objects in video and those in the agent's domain %
to a common Glove~\cite{pennington2014glove} word embedding space. Then, we re-estimate $\phi(o_m, o_n)$ for each object pair in $\mathcal{O}$ by considering nearest neighbors to $\mathcal{O}_V$ in this space. Specifically, in \refeqn{eqn:aco_score}, we set
\begin{align}
S_v(o_m, o_n) =
\sum \limits_{o_i \in \mathcal{N}(o_m)} \sum \limits_{o_j \in \mathcal{N}(o_n)}
\sigma_{m,i} \sigma_{n,j} S_v(o_i, o_j) ,
\label{eqn:aco_score_mapped}
\end{align}
where $\mathcal{N}(o)$ are the nearest neighbors of environment object $o$ among video objects $\mathcal{O}_V$ within a fixed distance in embedding space, and $\sigma_{i,j}$ measures the dot product similarity of the corresponding word embeddings of objects $o_i$ and $o_j$. The resulting compatibility scores represent relationships between objects in the agent's environment, as inferred from similar objects in egocentric video. See \reffig{fig:epic_to_acos} (right) for an illustration of top ACOs compared to other heuristics (like word embedding similarity) %
and Supp.~for further details.
\begin{figure*}[t!]
\centering
\includegraphics[width=\columnwidth]{figures/arch}
\caption{\textbf{Policy training framework}. \B{Left panel:} Our policy network generates an action $(a_t, o_t)$
given the current state observation. If objects are moved by the agent, the ACO memory $\mathcal{M}$ is updated to add objects (red) to their ACO's memory (blue) following \refsec{sec:env_reward}. \B{Right panel:} An auxiliary reward $R_{\phi}$ is provided for object interactions based on nearby ACOs. For example, if the interaction target $o_t$
is the faucet, the agent is rewarded for having the cup and bottle nearby.} %
\label{fig:arch}
\vspace*{-0.1in}
\end{figure*}
\vspace{-0.05in}
\subsection{Learning to interact with an activity-context reward} \label{sec:env_reward}
Next, we cast the compatibility scores learned from video into an auxiliary reward for embodied agents. In short, the agent is rewarded for interactions with any objects at any time; however, this reward is enhanced when there are compatible ACOs nearby. For example, it is more valuable to turn-on the stove when there is a pan on it than if there was a knife (or no other objects) nearby. To maximize this reward, the agent must intelligently move objects in the environment to locations with compatible objects before performing interactions.
To achieve this, the agent maintains a persistent memory to keep track of where objects were moved to, and what other objects near it are potential ACOs (and thus, how it affects the value of nearby object interactions).
The memory $\mathcal{M}$ starts empty for all objects. At each time-step, if an agent holding an object $o$ \emph{puts-down} the object at location $p$ in 3D space\footnote{We calculate $p$ from the agent's pose and depth observations using an inverse projection transformation on the interaction target point in the agent's current view.}, we identify $o$'s neighboring objects. %
For each neighboring object $o'$, $o$ is a potential ACO and is added to the memory: $\mathcal{M}(o')~\bigcup~\{(o, p)~|~d(o, o')<\epsilon\}$. We set $\epsilon = 0.5$m.
Conversely, if an object $o$ is \emph{picked-up} from a location $p$, it is removed from the list of potential ACOs of all objects it originally neighbored: $\mathcal{M}(o')\smallsetminus\{(o, p)\}~\forall o' \in \mathcal{M}$, and its own ACO memory is cleared: $\mathcal{M}(o) = \phi$. \reffig{fig:arch} (right) shows a snapshot of this memory built until time-step $t-1$. Each object (in red) was placed at its respective locations in previous time-steps, and added to the ACO memories of
nearby objects (in blue).
At time-step $t$ the agent is rewarded if it successfully interacts with an object $o_t$ based on this memory. The total activity-context reward $R_\phi$ is the sum of compatibility scores for objects in memory for which $o_t$ is a candidate ACO:
\begin{align}
R_{\phi}(s_t, a_t, o_t, \mathcal{M}) =
\begin{cases}
\sum\limits_{o' \in \mathcal{M}(o_t)} \phi(o', o_t) & \text{if } a_t \in \mathcal{A}_I \land c(a_t, o_t)=0 \\
~~~~0 & \text{otherwise,} \\
\end{cases}
\label{eqn:aux_reward}
\end{align}
where $c(a_t, o_t)$ counts the number of times a particular interaction occurs. Note that the currently held object (or \emph{null}, if nothing is held) is included in $\mathcal{M}$ for all interactions. See Supp.~for pseudo-code of the memory update and reward allocation step.
Unlike the sparse task reward that is offered only at the goal state (\refsec{sec:task_setup}), the activity-context reward is provided at every time-step. These dense rewards are not task-specific, meaning we do not give any privileged information about which video priors are most beneficial to a given task $\tau$.
However, they %
nonetheless encourage
the agent to take meaningful actions towards activities, helping to reach states that are strongly aligned with common task subgoals in human videos. %
\vspace{-0.05in}
\subsection{Interaction policy training} \label{sec:training}
Putting it together, our training process is as follows. We adopt an actor-critic policy to train our agent~\cite{sutton2018reinforcement}. At each time-step $t$, the current egocentric frame $x_t$
is encoded using a ResNet-18~\cite{he2016deep} encoder and then average pooled and fed to an LSTM based recurrent neural network (along with encodings of the held object class and previous action) to aggregate observations over time, and finally to an actor-critic network (MLP) to generate the next action distribution and value. See \reffig{fig:arch} (left).
In parallel, the agent maintains and updates its activity-context memory and generates an auxiliary reward based on nearby context objects at every time-step following \refsec{sec:env_reward}.
The final reward is a combination of the task reward (\refsec{sec:task_setup}) and the activity-context reward:
\begin{align}
R(s_t, a_t, o_t, s_{t+1}, \mathcal{M}) = R_{\tau}(s_t, a_t, o_t, s_{t+1}) + \lambda_{\phi}R_{\phi}(s_t, a_t, o_t, \mathcal{M}),
\label{eqn:tot_reward}
\end{align}
where $\lambda_{\phi}$ controls the contribution of the auxiliary reward term. We train our agents using DD-PPO~\cite{wijmans2019dd} for 5M steps, with rollouts of $T = 256$ time steps. Our model and all baselines use visual encoders from agents that are pre-trained for interaction exploration~\cite{nagarajan2020learning} for 5M steps, which we find benefits all approaches. See \reffig{fig:arch} and Supp.~for architecture, hyperparameter and training details. %
\section{Introduction}
\vspace*{-0.05in}
Embodied AI agents that are capable of moving around and interacting with objects in human spaces promise important practical applications for home service robots, %
ranging from agents that can search for misplaced items, to agents that can cook entire meals. The pursuit of such agents has driven exciting new research in visual semantic planning~\cite{zhu2017visual},
instruction following~\cite{shridhar2020alfred}, and object rearrangement~\cite{batra2020rearrangement,jain2020cordial}, %
typically supported by advanced %
simulators~\cite{puig2018virtualhome,kolve2017ai2,shen2020igibson,gan2020threedworld} %
where policies may be learned quickly and safely before potentially transferring to real robots.
In such tasks, an agent aims to perform a sequence of actions that will transform the visual environment from an initial state to a goal state. %
This in turn %
requires jointly learning %
behaviors for both \emph{navigation}, to move from one place to another, and object \emph{interaction}, to manipulate objects and modify the environment (e.g., pick-up objects, use tools and objects together, %
turn-on lights).
\begin{figure}[t]
\centering
\includegraphics[width=\textwidth]{figures/concept2}
\caption{\textbf{Main idea.} \B{Left and middle panel:} We %
discover \emph{activity-contexts} for objects directly from egocentric video of human activity. A given object's activity-context goes beyond ``what objects are found together'' to capture the likelihood that each other object in the environment participates in activities involving it (i.e., ``what objects together enable action").
\B{Right panel:} Our approach guides agents to bring compatible objects---objects with high likelihood---together to enable activities. %
For example, bringing a pan to the sink increases the value of faucet interactions, but bringing it to the table has little effect on interactions with a book.
}
\vspace{-0.1in}
\label{fig:concept}
\end{figure}
A key %
challenge is that changing the state of the environment involves context-sensitive actions that depend on both the agent and environment's current state---%
what the agent is holding, what other objects are present nearby, and what their properties are. For example, to wash a plate, a plate must be in the sink \emph{before} the agent toggles-on the faucet; to \emph{slice an apple} the agent must first be holding a knife, and the apple must not already be sliced. %
Understanding these conditions is critical for efficient learning and planning: %
the agent must first bring the environment into the proper precondition state before attempting to perform a given activity with objects. We refer to these states as \emph{activity-contexts}.
Despite its importance, current approaches %
do not explicitly model the activity-context. Pure reinforcement learning (RL) approaches directly search for goal states without considering preconditions. This requires a large number of trials for the agent to chance upon suitable configurations, and leads to poor sample efficiency or learning failures~\cite{zhu2017visual,shridhar2020alfred}.
Instead, most methods %
resort to collecting
expert demonstrations to train imitation learning (IL) agents and optionally finetune with RL~\cite{zhu2017visual,shridhar2020alfred,jain2019two,das2018embodied}. %
While demonstrations may implicitly reveal activity-contexts, they are cumbersome to collect, requiring expert teleoperation, often using specialized hardware (VR, MoCap) in artificial lab settings~\cite{zhang2018deep,rajeswaran2017learning,handa2019dexpilot}, and need to be collected independently for each new task.
Rather than solicit demonstrations, we propose to learn activity-contexts from real-world, egocentric video of people performing daily life activities.
Humans understand %
activities
from years of experience, %
and can effortlessly bring even a novel environment to new, appropriate configurations for interaction-heavy tasks. %
Egocentric or ``first-person" video recorded with a wearable camera puts actions and objects manipulated by a person at the forefront, offering an immediate window of this expertise in action in its natural habitat.
We present a reinforcement learning approach that infers activity-context conditions directly from people's first-hand experience and transfers them to embodied agents to improve interaction policy learning. Specifically, we 1) train visual models to detect how humans \emph{prepare} their environment for activities from egocentric video, and 2) develop a novel auxiliary reward function that encourages agents %
to seek out similar activity-context states. %
For example, by observing that people frequently carry pans to sinks (to clean them) or stoves (to cook their contents), our model %
rewards agents for prioritizing interactions with faucets or stove-knobs when pots or pans are nearby. As a result, this incentivizes agents to transport relevant objects to compatible locations \emph{before} %
attempting interactions, which accelerates learning. See \reffig{fig:concept}.
Importantly, our goal is not direct imitation.
Our insight is that while humans and embodied agents have very different action spaces and bodies, %
they operate in similar environments where the underlying conditions about what state the environment must be in before trying to modify it are strongly aligned. %
Our goal is thus to guide an agent's exploratory interactions towards these potential progress-enabling states as it learns a new task.
Moreover, because our training %
videos are passively collected %
by human camera-wearers and capture a wide array of daily actions, %
they help build a general visual prior for human activity-contexts, while side-stepping the heavy requirements %
of collecting IL demonstrations for each individual task of interest.
Our experiments demonstrate the value of learning from egocentric videos of \emph{humans} (in EPIC-Kitchens~\cite{damen2018scaling}) to train visual semantic planning \emph{agents} (in AI2-iTHOR~\cite{kolve2017ai2}). Our video model relates objects based on goal-oriented actions (e.g., knives used to cut potatoes) rather than spatial co-occurrences (e.g., knives are found near spoons)~\cite{yang2018visual,chaplot2020object,qiu2020learning,wu2018learning} or semantic similarity (e.g., potatoes are like tomatoes)~\cite{yang2018visual}.
Our agents outperform strong exploration methods and state-of-the-art embodied policies on multi-step interaction tasks (e.g., storing cutlery, cleaning dishes), improving absolute success rates by up to 12\% on the most complex tasks. %
Our approach learns policies faster, generalizes to unseen environments, and greatly improves success rates on difficult instances.
\section{Conclusion}
\vspace*{-0.1in}
We proposed an approach to discover activity-context priors from real-world human egocentric video and %
use them to accelerate training of interaction agents. Our work provides an avenue for embodied agents to directly benefit from human experience even when the tasks and environments differ. Future work could explore policies that update human-activity prior beliefs during task-specific policy training, and policy architectures that encode such priors jointly in the state representation. %
\begin{small}
\noindent\textbf{Acknowledgments}: Thanks to Santhosh Ramakrishnan for helpful feedback and the UT Systems Admin team for their continued support. UT Austin is supported in part by DARPA L2M and the UT Austin NSF AI Institute.
\end{small}
\clearpage
\section*{Supplementary Material}
This section contains supplementary material to support the main paper text. The contents include:
\begin{itemize}[leftmargin=*]
\itemsep0em
\item (\S\ref{sec:supp_demo}) A video demonstrating our agents ACO memory creation process and reward calculation described in \refsec{sec:env_reward}.
\item (\S\ref{sec:supp_qual_policy}) A video comparing our agents to baseline policies to supplement \reffig{fig:qual}.
\item (\S\ref{sec:supp_pseudocode}) Algorithm for creating and updating ACO memory described in \refsec{sec:env_reward}.
\item (\S\ref{sec:supp_epic_thor_mapping}) Figures illustrating the correspondences between scenes in EPIC Kitchens and observations in THOR using our method (\refsec{sec:priors_to_thor})
\item (\S\ref{sec:supp_epic_to_aco_supp}) Additional qualitative figures to supplement \reffig{fig:epic_to_acos} in the main paper.
\item (\S\ref{sec:supp_detection_details}) Detection model architecture details and additional implementation details related to \refsec{sec:priors}
\item (\S\ref{sec:supp_training_details}) Policy architecture details, training details and hyperparameters for our model in \refsec{sec:training}
\item (\S\ref{sec:baseline_details}) Implementation details for baselines in \refsec{sec:experiments} (Baselines).
\item (\S\ref{sec:supp_task_goals}) Goal descriptions for each task presented in \refsec{sec:experiments} (Tasks).
\item (\S\ref{sec:supp_sliced_results}) Task-specific breakdown of results presented in \refsec{sec:vsp_results} and \refsec{sec:consolidated_results}.
\item (\S\ref{sec:supp_ablations}) Additional experiments varying model architecture and combined reward schemes to supplement \refsec{sec:experiments}.
\end{itemize}
\section{ACO memory creation and reward calculation demo video} \label{sec:supp_demo}
In the video, we show the process of creating and building the activity-context memory described in \refsec{sec:env_reward}. The video shows the first-person view of the agent as it performs a series of object interactions (top left). We overlay masks for objects that are manipulated (red), nearby objects whose activity-context object (ACO) memory is updated (blue), the distance threshold under which objects are considered \emph{near} to each other (teal). Finally, we show how the presence of these objects affects the reward provided to the agent (top right). The video illustrates that bringing a mug to the sink or a pot to the stove results in high rewards as they are aligned with activities, while bringing a knife to the garbage bin is far less valuable.
\section{Video demo of our policy compared to baselines} \label{sec:supp_qual_policy}
In the video, we show rollouts of various policies during training to supplement \reffig{fig:qual} in the main paper. The video compares the baseline approaches to our method. It shows each step in a trajectory and the corresponding reward given to the agent, illustrating the contribution of activity-context objects to the total reward in our method compared to the uniform reward provided in the baselines. As mentioned in \refsec{sec:vsp_results}, our agents receive variable rewards (non-uniformly sized red dots) based on nearby ACOs (blue crosses), which encourages agents to move objects to more meaningful positions aligned with activities.
\section{ACO memory creation and update algorithm} \label{sec:supp_pseudocode}
We present the algorithm for creating and maintaining the ACO memory in \refalg{alg:supp_reward}. This corresponds to the steps outlined in \refsec{sec:env_reward} of the main paper and the accompanying reward equation \refeqn{eqn:aux_reward}. Note that in practice, we normalize $\phi(o_t, o)$ in L16 of \refalg{alg:supp_reward} such that the maximum rewarding ACO offers a reward of 1.0, to ensure that agents do not ignore objects with scores distributed across many potential ACOs.
\begin{algorithm}[h!]
\caption{Activity-context reward memory.}
\label{alg:supp_reward}
\begin{algorithmic}[1]
\Require{ACO memory $\mathcal{M}$, visitation count $c$, distance metric $d$, distance threshold $\epsilon$}
\Require State $s_t$, action $(a_t, o_t)$, held object $o$ at time-step $t$
\Function{UpdateMem}{$s_t, a_t, o_t, \mathcal{M}$}
\If{$a_t = \text{``put"}$} \Comment{Put $o$ at position $p$}
\Let {$\mathcal{M}(o')$}{$\mathcal{M}(o')~\bigcup~\{(o, p)~|~d(o, o')<\epsilon\}$}
\ElsIf {$a_t = \text{``take"}$} \Comment{Take $o_t$ from location $p_t$}
\Let {$\mathcal{M}(o_t)$}{$\{\}$}
\Let {$\mathcal{M}(o')$}{$\mathcal{M}(o') \smallsetminus \{(o_t, p_t)\} \ |\ \forall o' \in \mathcal{M}$}
\EndIf
\State \Return Updated memory $\mathcal{M}$
\EndFunction
\\
\Function{ACReward}{$s_t, a_t, o_t, \mathcal{M}$}
\State \algorithmicif\ $a_t \not\in \mathcal{A_I} \text{ \B{or} } c(a_t, o_t) > 0$\ \algorithmicthen\ \textbf{return} 0\ ;
\Let {$\mathcal{M}$}{\Call{UpdateMem}{$s_t, a_t, o_t, \mathcal{M}, $}}
\Let {$Z$}{$\max\limits_{o \in \mathcal{O}} \phi(o_t, o)$}
\Let {$R_{ACO}$}{$\sum\limits_{o \in \mathcal{M}(o_t)} \phi(o_t, o)/Z$} \Comment{\refeqn{eqn:aux_reward}}
\Let {$c(a_t, o_t)$}{$c(a_t, o_t) + 1$}
\State \Return $R_{\phi}$
\EndFunction
\end{algorithmic}
\end{algorithm}
\section{ACO correspondences between EPIC and THOR scenes} \label{sec:supp_epic_thor_mapping}
As mentioned in \refsec{sec:priors_to_thor} we translate from ACO pairs learned in video to a reward used in simulation. We show qualitative results for which scenes in THOR correspond to activities observed in EPIC in \reffig{fig:supp_epic_thor_correspondence}. The first column of each row shows a frame from an EPIC video clip showing a particular human activity. The remaining columns show similar ``states'' from THOR that our agents deem desirable to reach %
based on the distribution of ACOs present. Interactions performed in these states are highly rewarded following our approach.
The figures also illustrate our automatic mapping from the video vocabulary $\mathcal{O}_V$ to the agent environment vocabulary $\mathcal{O}$ for objects using word embedding similarity described in \refsec{sec:priors_to_thor}. For example, ``garlic:paste'' in EPIC is mapped to other food-like objects in THOR like ``tomatoes'' (left, bottom row); ``drawers'' are mapped to both ``cabinets'' and ``drawers'' (right, top row).
\begin{figure*}[h!]
\centering
\includegraphics[width=\columnwidth]{figures/supp_epic_thor_correspondence}
\caption{\B{Discovered EPIC $\longleftrightarrow$ THOR ACO correspondences.} First column shows a human activity from EPIC. Subsequent columns show similar states from THOR which provide high rewards when interactions are performed with objects once the agent is in that state.} %
\label{fig:supp_epic_thor_correspondence}
\end{figure*}
\section{Additional ACO detection results on EPIC-Kitchens} \label{sec:supp_epic_to_aco_supp}
We show additional detection results to supplement \reffig{fig:epic_to_acos} (left). These images show sampled (object, ACO) tuples (\refeqn{eqn:aco_subset}) in red and blue respectively. The last column shows failure cases due to incorrect active object detections and incorrect object instance detection.
\begin{figure*}[h!]
\centering
\includegraphics[width=\columnwidth]{figures/supp_epic_dets}
\caption{\B{Additional EPIC detections to supplement \reffig{fig:epic_to_acos}.} Last column shows failure cases.}
\label{fig:supp_epic_detections}
\end{figure*}
\section{Pre-trained detection model and ACO scoring details} \label{sec:supp_detection_details}
\paragraph{Pre-trained detection models}
As mentioned in \refsec{sec:priors}, we use two detection models in our approach -- (1) An active object detector which generates high-confidence box proposals for objects being interacted with hands (but does not assign object class labels to them) and (2) An object instance detector that produces a set of named objects and boxes for visible object instances. For (1) we use pre-trained models provided by authors of \cite{shan2020understanding}\footnote{\url{https://github.com/ddshan/hand_object_detector}}. For (2) we use pre-computed detections per frame using a Faster-RCNN model released by the authors of EPIC-Kitchens~\cite{damen2018scaling}. We set confidence thresholds of $0.5$ for all models.
\paragraph{Activity context curation details} To infer each frame's activity-context following \refeqn{eqn:aco_subset}, we use a manually curated list of \emph{moveable} objects from EPIC, though it is possible to automatically infer this list from action labels on video clips (all objects that are \emph{picked} up), or using the aforementioned hand-object detectors. Of the 398 objects in EPIC, 349 are moveable. We list the remaining objects in the table below.
\begin{table*}[h!]
\centering
\resizebox{\textwidth}{!}{
\begin{tabular}{|l|l|l|l|l|l|l|}
\hline
tap & top & microwave & machine:washing & toaster & machine:sous:vide & flame \\ \hline
cupboard & oven & button & processor:food & knob & window & fire \\ \hline
drawer & maker:coffee & juicer & plug & ladder & heater & grill \\ \hline
hand & sink & scale & kitchen & wall & door:kitchen & time \\ \hline
fridge & heat & rack:drying & floor & tv & table & timer \\ \hline
hob & dishwasher & freezer & fan:extractor & shelf & rug & desk \\ \hline
bin & blender & light & chair & stand & switch & lamp \\ \hline
\end{tabular}
\label{tbl:supp_immovable}
}
\end{table*}
\paragraph{ACO mapping details} As mentioned in \refsec{sec:priors_to_thor}, to match object classes between AI2-iTHOR and EPIC Kitchens we map each (object, ACO) tuple in the video object space $\mathcal{O}_V$ to corresponding tuples in environment object space $\mathcal{O}$ following \refeqn{eqn:aco_score_mapped}. We use a GloVe~\cite{pennington2014glove} similarity threshold of $0.6$. Lower values lead to undesirable mappings across object classes (e.g., toasters and refrigerators which are both appliances, but participate in distinct activities) .%
\section{Policy architecture and training details} \label{sec:supp_training_details}
We provide additional architecture and training details to supplement information provided in \refsec{sec:training} in the main paper.
\paragraph{Policy network} As mentioned in \refsec{sec:training}, we use a ResNet-18 observation encoder pretrained with observations from 5M iterations of training of an interaction exploration agent~\cite{nagarajan2020learning}. We transfer the backbone only and freeze it during task training. Each RGB observation is encoded to a $512$-D feature. A single linear embedding layer is used to embed the previous action and the currently held object (or \emph{null}) to vectors of dimension $32$ each. The total observation feature is the concatenation of these three features. All architecture hyperparams are listed in \reftbl{tbl:supp_hyperparam} (Policy network).
\paragraph{Training hyperparameters}
We modify the Habitat-Lab codebase~\cite{savva2019habitat} to support training agents in the THOR simulator platform~\cite{kolve2017ai2}. We search over $\lambda_\phi \in \{0.01, 0.1, 1.0, 5.0\}$ for \refeqn{eqn:tot_reward} and select $\lambda_\phi = 1.0$ which has the highest consolidated performance on validation episodes for all methods following the procedure in \refsec{sec:consolidated_results}. All training hyperparameters are listed in \reftbl{tbl:supp_hyperparam} (RL training).
\begin{table*}[h!]
\centering
\begin{tabular}{lc}
\toprule
\multicolumn{2}{c}{RL training} \\ \midrule
\multicolumn{1}{l}{Optimizer} & Adam \\
\multicolumn{1}{l}{Learning rate} & 2.5e-4 \\
\multicolumn{1}{l}{$\#$ parallel actors} & 64 \\
\multicolumn{1}{l}{PPO mini-batches} & 2 \\
\multicolumn{1}{l}{PPO epochs} & 4 \\
\multicolumn{1}{l}{PPO clip param} & 0.2 \\
\multicolumn{1}{l}{Value loss coefficient} & 0.5 \\
\multicolumn{1}{l}{Entropy coefficient} & 0.01 \\
\multicolumn{1}{l}{Normalized advantage?} & Yes \\
\multicolumn{1}{l}{Training episode length} & 256 \\
\multicolumn{1}{l}{LSTM history length} & 256 \\
\multicolumn{1}{l}{$\#$ training steps ($\times$ 1e6)} & 5 \\ \midrule
\multicolumn{2}{c}{Policy network} \\ \midrule
\multicolumn{1}{l}{Backbone} & resnet18 \\
\multicolumn{1}{l}{Input image size} & 256$\times$256 \\
\multicolumn{1}{l}{LSTM hidden size} & 512 \\
\multicolumn{1}{l}{$\#$ layers} & 2 \\
\bottomrule
\end{tabular}
\caption{\B{RL policy architecture and training hyperparameters.}}
\label{tbl:supp_hyperparam}
\end{table*}
\section{Baseline implementation details} \label{sec:baseline_details}
We present implementation details for baselines in \refsec{sec:experiments} (Baselines). \SC{NavExp} and \SC{IntExp} baselines use the same architecture as our model described in \refsec{sec:supp_training_details}, but vary in the reward they receive during training. \SC{ScenePriors} uses a different backbone architecture that uses a GCN based state encoder as described below. %
\paragraph{\SC{NavExp}} Agents are rewarded for visiting new objects such that the object is visible and within interaction range (less than $1.5m$ away). A constant reward is provided for every new object class visited. This is similar to previous implementations~\cite{ramakrishnan2020exploration,fang2019,nagarajan2020learning}. We use the implementation from \cite{nagarajan2020learning}.
\paragraph{\SC{IntExp}} Following \cite{nagarajan2020learning}, agents are rewarded for new object interactions. The reward provided has the form in \refeqn{eqn:aux_reward}, but provides a constant reward regardless of ACOs present. We use the author's code.
\paragraph{\SC{ScenePriors}} We modify the architecture in \cite{yang2018visual}, which was built for object search. We use the author's code. First, we remove the goal object encoding as the agent is not searching for a single object. Second, we replace the backbone network with our shared ResNet backbone to ensure fair comparison. We use a GCN encoding dimension of $512$. The remaining architecture details are consistent with \cite{yang2018visual}.
\section{Goal condition details for all tasks} \label{sec:supp_task_goals}
We next list out formal goal conditions for all tasks described in \refsec{sec:experiments} of the main paper (Tasks). Each goal is specified as a conjunction of predicates that need to be satisfied in a candidate goal state. The goal is satisfied if for any object $o$ in the environment, the following conditions are true:
\begin{itemize}[leftmargin=*]
\itemsep0em
\item \SC{Store}: inReceptacle(o, Drawer) $\land$ isClosed(Drawer) $\land$ isStorable(o)
\item \SC{Heat}: inReceptacle(o, StoveBurner) $\land$ isToggledOn(StoveKnob) $\land$ isHeatable(o)
\item \SC{Cool}: inReceptacle(o, Fridge) $\land$ isClosed(Fridge) $\land$ isCoolable(o)
\item \SC{Clean}: inReceptacle(o, SinkBasin) $\land$ isToggledOn(Faucet) $\land$ isCleanable(o)
\item \SC{Slice}: isHolding(Knife) $\land$ isSliceable(o)
\item \SC{Prep}: [inReceptacle(o, Pot) | inReceptacle(o, Pan)] $\land$ isCookable(o)
\item \SC{Trash}: inReceptacle(o, GarbageCan) $\land$ isTrashable(o)
\end{itemize}
where \emph{inReceptacle} checks if an object is inside/on top of a particular object, \emph{is-X-able} filters for objects with specific affordances (e.g., only objects that can be placed on the stove like pots/pans/kettles satisfy \emph{isHeatable}), \emph{isClosed} and \emph{isToggledOn} checks for specific object states, and \emph{isHolding} checks if the agent is holding a specific type of object (e.g., for \SC{Slice} this has to be a Knife or a ButterKnife). Further, for each task that involves moving objects to receptacles, the object must originally have been outside the receptacle (e.g., outside the Fridge for \SC{Cool}; off the stovetop for \SC{Heat}).
\section{Additional ablation experiments} \label{sec:supp_ablations}
We present a comparison of different backbone architectures (ResNet18 vs. ResNet50) and aggregation modules (LSTM vs. GRU) for both our model and the baselines in \reftbl{tbl:backbone_arch}. We evaluate on the unseen test episodes for 4 interaction-heavy tasks. The average results of 2 training runs are in the table below. Using stronger backbones seems to help marginally, but does not offer conclusive results. Using the simpler GRU based aggregation (instead of LSTM) results in large improvements. Overall, the trends remain consistent across all configurations: Vanilla $<$ NavExp $<$ Ours. Architectural changes alone in the baselines (to either the backbone, or the aggregation mechanism) are not enough to compensate for task difficulty --- performance on Cool, Store and Heat remain low ($<$10\%) for Vanilla and NavExp.
In \reftbl{tbl:combined_policies} we show the results of our policies combined with rewards from a navigation exploration agent. As mentioned in \refsec{sec:vsp_results}, while our agent excels in interaction-heavy tasks, navigation exploration agents perform well on interaction-light tasks which often require finding a single object and bringing it to the right location. For example in the Trash task, there is a single garbage can that navigation agents quickly find as they cover area, but that our agents struggle to find early on. The two strategies can be combined to address this issue. \reftbl{tbl:combined_policies} shows average results of 2 runs where we add the two reward functions together with equal weights (=0.5) to achieve the best performance.
\begin{table*}[h]
\resizebox{\textwidth}{!}{
\begin{tabular}{l|l|l|l|l|l|l|l|}
\cline{2-8}
& \SC{Cool} & \SC{Store} & \SC{Heat} & \SC{Clean} & \SC{Slice} & \SC{Prep} & \SC{Trash} \\ \hline
\multicolumn{1}{|l|}{\SC{NavExp}~\cite{ramakrishnan2020exploration}} & 0.05 & 0.01 & 0.01 & 0.43 & 0.29 & 0.33 & 0.25 \\ \hline
\multicolumn{1}{|l|}{\SC{Ours}} & \B{0.26} & \B{0.12} & 0.13 & \B{0.53} & \B{0.36} & 0.26 & 0.13 \\ \hline
\multicolumn{1}{|l|}{\SC{Ours + NavExp}} & 0.25 & 0.05 & \B{0.19} & 0.50 & 0.34 & \B{0.41} & \B{0.26} \\ \hline
\end{tabular}
}
\caption{\B{Combined policy experiments.} Navigation exploration offers complementary reward signals that can be combined with our method for stronger performance.}
\label{tbl:combined_policies}
\end{table*}
\begin{table*}[h]
\resizebox{\textwidth}{!}{
\begin{tabular}{lllllllllllllll}
& \multicolumn{4}{c}{ResNet18 + LSTM} & & \multicolumn{4}{c}{ResNet50 + LSTM} & & \multicolumn{4}{c}{ResNet18 + GRU} \\ \cline{1-5} \cline{7-10} \cline{12-15}
\multicolumn{1}{|l|}{} & \multicolumn{1}{l|}{\SC{Cool}} & \multicolumn{1}{l|}{\SC{Store}} & \multicolumn{1}{l|}{\SC{Heat}} & \multicolumn{1}{l|}{\SC{Clean}} & \multicolumn{1}{l|}{} & \multicolumn{1}{l|}{\SC{Cool}} & \multicolumn{1}{l|}{\SC{Store}} & \multicolumn{1}{l|}{\SC{Heat}} & \multicolumn{1}{l|}{\SC{Clean}} & \multicolumn{1}{l|}{} & \multicolumn{1}{l|}{\SC{Cool}} & \multicolumn{1}{l|}{\SC{Store}} & \multicolumn{1}{l|}{\SC{Heat}} & \multicolumn{1}{l|}{\SC{Clean}} \\ \cline{1-5} \cline{7-10} \cline{12-15}
\multicolumn{1}{|l|}{\SC{Vanilla}} & \multicolumn{1}{l|}{0.07} & \multicolumn{1}{l|}{0.00} & \multicolumn{1}{l|}{0.02} & \multicolumn{1}{l|}{0.29} & \multicolumn{1}{l|}{} & \multicolumn{1}{l|}{0.04} & \multicolumn{1}{l|}{0.00} & \multicolumn{1}{l|}{0.01} & \multicolumn{1}{l|}{0.26} & \multicolumn{1}{l|}{} & \multicolumn{1}{l|}{0.31} & \multicolumn{1}{l|}{0.03} & \multicolumn{1}{l|}{0.03} & \multicolumn{1}{l|}{0.38} \\ \cline{1-5} \cline{7-10} \cline{12-15}
\multicolumn{1}{|l|}{\SC{NavExp}~\cite{ramakrishnan2020exploration}} & \multicolumn{1}{l|}{0.02} & \multicolumn{1}{l|}{0.02} & \multicolumn{1}{l|}{0.01} & \multicolumn{1}{l|}{0.44} & \multicolumn{1}{l|}{} & \multicolumn{1}{l|}{0.02} & \multicolumn{1}{l|}{0.00} & \multicolumn{1}{l|}{0.00} & \multicolumn{1}{l|}{\B{0.42}} & \multicolumn{1}{l|}{} & \multicolumn{1}{l|}{0.00} & \multicolumn{1}{l|}{0.00} & \multicolumn{1}{l|}{0.03} & \multicolumn{1}{l|}{0.35} \\ \cline{1-5} \cline{7-10} \cline{12-15}
\multicolumn{1}{|l|}{\SC{Ours}} & \multicolumn{1}{l|}{\B{0.30}} & \multicolumn{1}{l|}{\B{0.16}} & \multicolumn{1}{l|}{\B{0.11}} & \multicolumn{1}{l|}{\B{0.55}} & \multicolumn{1}{l|}{} & \multicolumn{1}{l|}{\B{0.15}} & \multicolumn{1}{l|}{\B{0.11}} & \multicolumn{1}{l|}{\B{0.12}} & \multicolumn{1}{l|}{0.36} & \multicolumn{1}{l|}{} & \multicolumn{1}{l|}{\B{0.52}} & \multicolumn{1}{l|}{\B{0.31}} & \multicolumn{1}{l|}{\B{0.17}} & \multicolumn{1}{l|}{\B{0.73}} \\ \cline{1-5} \cline{7-10} \cline{12-15}
\end{tabular}
}
\caption{\B{Backbone architecture ablations.} Performance on four interaction-heavy tasks. }
\label{tbl:backbone_arch}
\end{table*}
\section{Task-specific results breakdown} \label{sec:supp_sliced_results}
We include task level breakdowns of results from \refsec{sec:consolidated_results} of the main paper. These results highlight the strengths and weaknesses of all models on a task-specific level to supplement overall task success results in \reftbl{tbl:task_success} of the main paper. These include Task success vs. training iteration (\reffig{fig:supp_success_by_iter}) and Task success vs. instance difficulty (\reffig{fig:supp_success_by_difficulty}) %
\begin{figure*}[h!]
\centering
\includegraphics[width=\textwidth]{figures/supp_success_by_iter}
\caption{\textbf{Task success vs. training iteration.} This is the task-specific version of \reffig{fig:agg_perf} (left) that shows convergence rates of all methods.}
\label{fig:supp_success_by_iter}
\end{figure*}
\begin{figure*}[h!]
\centering
\includegraphics[width=\textwidth]{figures/supp_success_by_difficulty}
\caption{\textbf{Task success vs. navigation difficulty.} This is the task-specific version of \reffig{fig:agg_perf} (right) that shows improvement in success over the baseline model. Note: we show absolute improvement (instead of relative) as the baseline has zero success for some tasks and some difficulty levels.} %
\label{fig:supp_success_by_difficulty}
\end{figure*}
\section{Related Work}
\vspace*{-0.1in}
\paragraph{Interaction in 3D environments}
Recent embodied AI work leverages simulated environments~\cite{puig2018virtualhome,savva2019habitat,shen2020igibson,kolve2017ai2, gan2020threedworld,shen2020igibson} to build agents for interaction-heavy tasks like visual semantic planning~\cite{zhu2017visual}, interactive question answering~\cite{gordon2018iqa}, instruction following~\cite{shridhar2020alfred}, and object rearrangement~\cite{batra2020rearrangement,jain2019two}. Prior approaches use imitation learning (IL) based pre-training %
to improve sample efficiency, but %
at the expense of collecting expert in-domain demonstrations individually for each task~\cite{zhu2017visual,das2018embodied,jain2019two,shridhar2020alfred}. %
Instead, we leverage readily available human egocentric video in place of costly IL demonstrations, and we propose an RL approach that benefits from the %
discovered human activity-priors.
\vspace*{-0.1in}
\paragraph{Exploration for navigation and interaction}
Exploration strategies for visual navigation encourage efficient policy learning by rewarding agents for covering area~\cite{chen2019learning,Chaplot2020Learning,fang2019}, visiting new states~\cite{savinov2018episodic,bellemare2016unifying}, expanding the %
frontier~\cite{qi2020learning}, %
anticipating maps~\cite{ramakrishnan2020exploration}, or via intrinsic motivation~\cite{pathak2017curiosity,burda2018large}.
For grasping, prior work studies curiosity~\cite{haber2018learning} and disagreement~\cite{pathak2019self} as intrinsic motivation. %
Beyond navigation and grasping, interaction exploration~\cite{nagarajan2020learning} %
rewards agents for %
successfully attempting
new object interactions (take potato, open fridge, etc.). In contrast, our agents are rewarded %
for achieving compatible activity-contexts
learned from human egocentric video. Our model incentivizes %
action sequences that are aligned with activities (e.g., putting a plate in the sink \emph{before} turning-on the faucet), as opposed to arbitrary actions (e.g., turning the faucet on and then immediately off). %
\vspace*{-0.1in}
\paragraph{Learning from passive video for embodied agents} Thus far, %
video plays a limited role in learning for embodied AI, %
primarily as a source of demonstrations for imitation learning.
Prior work learns dynamics models for behavior cloning~\cite{sharma2018multiple,yu2018one,pathak2018zero,schmeckpeper2019learning,schmeckpeper2020reinforcement} or crafts reward functions that encourage visiting expert states~\cite{aytar2018playing,dwibedi2018learning,sermanet2017time,liu2018imitation}.
Beyond imitation, recent navigation work ``re-labels" video frames with pseudo-action labels to learn navigation subroutines~\cite{chang2020semantic} or action-conditioned value functions~\cite{kumar2020learning}; however, they use videos generated from simulation or from intentionally recorded real-estate tours.
In contrast, %
our work is the first to use free-form human-generated video captured in the real world to learn priors for object interactions. Our %
priors are not tied to specific goals (as in behavior cloning) and are cast as general purpose auxiliary rewards to encourage efficient RL.
\vspace*{-0.1in}
\paragraph{Learning about human actions from video}
Substantial work in computer vision explores models for human action recognition in video~\cite{zhu2020comprehensive,wang2016temporal,carreira2017quo,feichtenhofer2019slowfast}, including analyzing hand-object interactions~\cite{shan2020understanding,bambach2015lending,tekin2019h,cao2020reconstructing,garcia2018first} and egocentric video understanding~\cite{li2018eye,furnari2019would,lu2013story,jiang2017seeing,damen2018scaling}.
More closely related to our work, visual affordance models derived from video can detect likely places for actions to occur~\cite{nagarajan2019grounded,fang2018demo2vec,nagarajan2020ego}, such as where to grasp a frying pan, and object saliency models can identify human-useable objects in egocentric video~\cite{damen2014you,bertasius:2016,nextactiveobject}.
These methods learn important concepts from human video, but do not %
consider their use for embodied agent action, as we propose.
\vspace*{-0.1in}
\paragraph{Semantic priors for embodied agents}
Human-centric environments contain useful semantic and structural regularities. Prior work exploits spatial relationships between objects~\cite{yang2018visual,chaplot2020object,qiu2020learning} and room layouts~\cite{wu2018learning} to improve navigation (e.g., beds tend to be in bedrooms; kitchens tend to be near dining rooms). Recent work learns visual priors from agent experience (not human video) to understand environment traversability~\cite{qi2020learning}, affordances~\cite{nagarajan2020learning}, or object properties like mass~\cite{lohmann2020learning}. These priors encode static properties of objects or geometric relationships between them %
by learning from co-occurrences in static images. In contrast, our formulation encodes information about objects \emph{in action} from video of humans using objects to answer \emph{what objects should be brought together to enable interactions}, rather than \emph{what objects are typically co-located}. In addition, unlike previous models that require substantial %
online agent experience to learn priors, we use readily available egocentric video of human activity.
\section{Experiments} \label{sec:experiments}
\vspace*{-0.05in}
We evaluate how well our agents learn complex interaction tasks using our human video based reward.
\vspace*{-0.1in}
\paragraph{Simulator and video datasets.}
To train policies, we use the AI2-iTHOR~\cite{kolve2017ai2} simulator where agents can navigate: $\mathcal{A}_N$ = \{move forward, turn left/right 90$^{\circ}$, look up/down 30$^{\circ}$\}, and interact with objects: $\mathcal{A}_I$ = \{take, put, open, close, toggle-on, toggle-off, slice\}. Our action space of size $|\mathcal{A}| = 110$ is the union of all navigation actions and valid object interactions following \cite{RoomR}. %
We use all 30 kitchen scenes from AI2-iTHOR, split into training (25) and testing (5) sets. %
To learn activity-context priors, we use all 55 hours of video from EPIC-Kitchens~\cite{damen2018scaling}, which contains egocentric videos of daily, unscripted kitchen activities in a variety of homes. It consists of ${\raise.17ex\hbox{$\scriptstyle\sim$}}$40k video clips annotated for interactions spanning 352 objects ($\mathcal{O}_V$) and 125 actions. %
Note that we use clip boundaries to segment actions, but we do not use the action labels in our method.
Since kitchen scenes present a diverse set of object interactions in %
multi-step cooking activities, they are of great interest in this research domain%
~\cite{damen2018scaling,vrkitchen,nagarajan2020learning}. Further, the alignment of domain with AI2-iTHOR's kitchen environments provides a path to transfer knowledge from videos to agents.
\vspace*{-0.1in}
\paragraph{Visual semantic planning tasks.}
We consider seven tasks in our experiments where the agent must
\B{\SC{Store}}: put an object that is outside into a drawer,
\B{\SC{Heat}}: turn on the stove with a cooking receptacle on it,
\B{\SC{Cool}}: store a food item or container inside the fridge,
\B{\SC{Clean}}: put an object in the sink and turn on the faucet to wash it,
\B{\SC{Slice}}: slice a food item with a knife,
\B{\SC{Prep}}: place a food item inside a pot/pan,
\B{\SC{Trash}}: throw an object into the trash bin,
Each task is associated with a goal state that the environment must be in. For example in the \SC{Cool} task, an object that was originally outside, must be inside the fridge, and the fridge door must be closed. See Supp.~for goal states for all tasks.
These tasks represent realistic scenarios in home robot assistant settings, %
consistent with tasks studied in recent work~\cite{shridhar2020alfred,nagarajan2020learning}. %
We generate 64 episodes per task and per environment with randomized object and agent positions to evaluate our agents. We report task success rates (\%) on unseen test environments. This means that our model must both generalize what it observes in the real-world human video to the agent's egocentric views, as well as generalize from training environments to novel test environments.
\vspace*{-0.1in}
\paragraph{Baselines.}
We compare several methods:
\vspace*{-0.1in}
\begin{itemize}[leftmargin=*]
\itemsep0em
\item \B{\SC{Vanilla}} trains a policy using only the task reward (\refsec{sec:task_setup}). This is the standard approach to train reinforcement learning agents.
\item \B{\SC{ScenePriors}}~\cite{yang2018visual} also uses only the task reward, but encodes spatial co-occurrences between objects using a graph convolutional network (GCN) model to enhance its state representation. We use the authors' \href{https://github.com/allenai/savn}{code}. %
\item \B{\SC{NavExp}}~\cite{ramakrishnan2020exploration} adds an auxiliary navigation exploration reward to the vanilla model. We use object coverage~\cite{fang2019,ramakrishnan2020exploration}, which rewards the agent for visiting (but not interacting with) new objects.
\item \B{\SC{IntExp}}~\cite{nagarajan2020learning} is a state-of-the-art exploration method for object interaction that adds a reward for successful object interactions regardless of nearby context objects. %
We use the authors' \href{https://github.com/facebookresearch/interaction-exploration}{code}.
\end{itemize}
\SC{ScenePriors} represents the conventional approach of introducing visual priors into policy learning---through GCN encoders, not auxiliary rewards---and considers spatial co-occurrence instead of activity based priors.
\SC{NavExp} and \SC{IntExp} are exploration methods from recent work that incentivize all locations or objects equally. %
In contrast, our reward is a function of the state of the environment, and how well it aligns with observed states in video of human activity. %
\newcommand{\perf}[2]{#1\scriptsize{$\pm #2$}}
\newcommand{\perfb}[2]{\textbf{#1\scriptsize{$\pm #2$}}}
\begin{table*}[t]
\centering
\resizebox{\textwidth}{!}{
\begin{tabular}{l|l|l|l|l|l|l|l|}
\cline{2-8}
& \SC{Cool}\scriptsize{($1e$-$7$)}& \SC{Store}\scriptsize{($1e$-$7$)}& \SC{Heat}\scriptsize{($2e$-$6$)}& \SC{Clean}\scriptsize{($2e$-$5$)}& \SC{Slice}\scriptsize{($1e$-$3$)}& \SC{Prep}\scriptsize{($2e$-$3$)}& \SC{Trash}\scriptsize{($2e$-$3$)} \\ \hline
\multicolumn{1}{|l|}{\SC{Vanilla}}
& \perf{0.12}{0.07} & \perf{0.00}{0.00} & \perf{0.01}{0.01} & \perf{0.35}{0.12} & \perf{0.30}{0.03} & \perf{0.22}{0.05} & \perf{0.14}{0.05} \\ \hline
\multicolumn{1}{|l|}{\SC{ScenePriors}~\cite{yang2018visual}}
& \perf{0.14}{0.09} & \perf{0.00}{0.00} & \perf{0.04}{0.01} & \perf{0.35}{0.02} & \perfb{0.36}{0.05} & \perf{0.26}{0.08} & \perf{0.20}{0.06} \\ \hline
\multicolumn{1}{|l|}{\SC{NavExp}~\cite{ramakrishnan2020exploration}}
& \perf{0.05}{0.04} & \perf{0.01}{0.01} & \perf{0.01}{0.01} & \perf{0.43}{0.02} & \perf{0.29}{0.05} & \perfb{0.33}{0.02} & \perfb{0.25}{0.04} \\ \hline
\multicolumn{1}{|l|}{\SC{IntExp}~\cite{nagarajan2020learning}}
& \perf{0.11}{0.06} & \perf{0.03}{0.03} & \perf{0.06}{0.03} & \perf{0.19}{0.05} & \perf{0.26}{0.07} & \perf{0.19}{0.08} & \perf{0.02}{0.01} \\ \hline
\multicolumn{1}{|l|}{\SC{Ours}}
& \perfb{0.26}{0.06} & \perfb{0.12}{0.07} & \perfb{0.13}{0.05} & \perfb{0.53}{0.03} & \perfb{0.36}{0.06} & \perf{0.26}{0.07} & \perf{0.13}{0.02} \\ \hline
\end{tabular}
}
\caption{\B{Task success rates (\%) on test environments.} Numbers in brackets indicate the probability that a random agent performs the correct sequence of interactions to complete a task (e.g., \SC{Cool} is much harder than \SC{Prep}). %
Our video-based activity-context priors result in the best performance across all interaction-heavy tasks. Navigation-based exploration agents perform well for tasks that involve minimal object interactions (e.g. \SC{Prep}, \SC{Trash}). Values are averaged over 3 training runs.}
\label{tbl:task_success}
\end{table*}
\vspace{-0.05in}
\subsection{Visual semantic planning results} \label{sec:vsp_results}
\vspace{-0.05in}
\reftbl{tbl:task_success} shows success rates across all tasks. The numbers in brackets represent the probability that an agent executing random actions will reach the goal state assuming all objects are within reach. \emph{Interaction-heavy} tasks requiring long sequences of steps like \SC{Store} (pick up object, open drawer, put object inside, close drawer) are significantly harder than \emph{interaction-light} tasks like \SC{Prep} (pick up food item, put in pan)---in this particular example, by four orders of magnitude.
Our approach outperforms prior work on all interaction-heavy tasks, and performs competitively on interaction-light tasks. Compared to the \SC{Vanilla} baseline, \SC{ScenePriors}'s stronger observation encoding results in better performance with the same task reward; %
however, while its spatial co-occurrence prior encoding what objects are \emph{found together} (rather than \emph{used together}) is helpful for object search~\cite{yang2018visual}, it falls short for object interactions.
\SC{IntExp} provides a task-agnostic objective for pre-training interaction agents; it results in strong visual encoder features, but yields poor performance when used as an auxiliary reward. It fails to discriminate between useful interaction sequences that are aligned with task goals (e.g., put pan on stove, turn-on stove) and arbitrary interaction sequences that maximize its reward (e.g., pick up, then put down objects sequentially). \SC{NavExp} performs poorly on interaction-heavy tasks, often achieving close to zero success rates. However, it performs well on tasks like \SC{Trash} and \SC{Prep} where a single object must be moved to a target location for success.
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth]{figures/qual}
\caption{\textbf{Policy rollouts for the \SC{Cool} task during early/late stage training}. Red dots represent auxiliary rewards (size represents value). Our agents learn to move objects to suitable locations (near supporting ACOs, blue crosses) to increase their reward, rather than simply visit \emph{more} objects. See text. %
}
\vspace{-0.1in}
\label{fig:qual}
\end{figure*}
\reffig{fig:qual} shows qualitative results %
of 100 policy rollouts. Each panel shows a side-by-side comparison of agent behavior during early (1M steps) and late (5M steps) training for the same task. Each red dot represents an auxiliary reward given to the agent, sized by reward value; the yellow star denotes where the task reward is issued. %
\SC{IntExp} (left panel) is rewarded uniformly for each interaction (equally sized red dots), and thus has a single strategy to maximize reward --- perform \emph{more} interactions. In contrast, our approach (right panel) can maximize reward along two axes: perform more interactions as before, or selectively move objects to favorable positions near ACOs (blue crosses). This guides agent exploration to relevant states for activities, translating to higher task success.
\begin{figure*}[t]
\centering
\includegraphics[width=\textwidth]{figures/agg_perf}
\caption{\textbf{Consolidated performance across all seven tasks.} \B{Left: Success rate vs. training iteration.} Our dense activity-context rewards accelerate performance at early training iterations. \B{Right: Success rate improvement vs. navigation difficulty.} Episodes are sorted and grouped by ideal navigation distance to task objects (low $\rightarrow$ high). Our method shows largest improvements over \SC{Vanilla} on difficult instances. %
}
\vspace{-0.05in}
\label{fig:agg_perf}
\end{figure*}
\vspace{-0.05in}
\subsection{Training speed and task success versus difficulty}
\label{sec:consolidated_results} %
\reffig{fig:agg_perf} shows consolidated results across all tasks, treating each episode of each task as an individual instance that can be successful or not. These results give a sense of overall performance of each baseline, despite variation across tasks. See Supp. for a task-specific breakdown of results.
\reffig{fig:agg_perf} (left) shows convergence speed. %
Our approach sees an increased success rate at early training epochs as our dense rewards allow agents to make progress towards goals without explicit task rewards. \SC{NavExp} performs poorly in early iterations as it prioritizes navigation actions to visit as many objects as possible. Overall this result shows a key advantage of our idea: learning from human video not only boosts overall policy success rates, but it also accelerates the agent's learning process, giving a ``head start" from having observed how people perform general daily interactions. Importantly, our ablations (below) show that access to the video alone is not sufficient; our idea to capture functional activity-contexts is key.
\reffig{fig:agg_perf} (right) analyzes task success as a function of navigation difficulty. We measure this as the ideal geodesic distance to each of the objects required to reach the goal state. This value is low when objects are close by in small environments, and large when the agent has to move around to find specific objects in large environments. We sort episodes by this criterion into 8 groups of increasing difficulty and compare each method against the \SC{Vanilla} baseline policy. Our method has the highest success rates across difficulties and offers the largest improvements on more difficult instances. %
\begin{table*}[t]
\centering
\resizebox{\textwidth}{!}{
\begin{tabular}{l|l|l|l|l|l|l|l|}
\cline{2-8}
& \SC{Cool} & \SC{Store} & \SC{Heat} & \SC{Clean} & \SC{Slice} & \SC{Prep} & \SC{Trash} \\ \hline
\multicolumn{1}{|l|}{\SC{Uniform}}
& \perf{0.07}{0.00} & \perf{0.02}{0.01} & \perf{0.12}{0.00} & \perf{0.37}{0.05} & \perf{0.26}{0.03} & \perf{0.22}{0.05} & \perf{0.02}{0.01} \\ \hline
\multicolumn{1}{|l|}{\SC{WordEmbed}}
& \perf{0.19}{0.03} & \perf{0.02}{0.02} & \perf{0.03}{0.00} & \perf{0.40}{0.13} & \perf{0.34}{0.01} & \perf{0.22}{0.01} & \perf{0.04}{0.02} \\ \hline
\multicolumn{1}{|l|}{\SC{SpatialCooc}}
& \perfb{0.33}{0.05} & \perf{0.02}{0.01} & \perf{0.06}{0.02} & \perf{0.47}{0.09} & \perf{0.33}{0.02} & \perf{0.25}{0.08} & \perfb{0.13}{0.05} \\ \hline
\multicolumn{1}{|l|}{\SC{Ours (IntSeq)}}
& \perfb{0.33}{0.04} & \perf{0.11}{0.03} & \perf{0.07}{0.03} & \perf{0.43}{0.04} & \perf{0.30}{0.05} & \perf{0.20}{0.02} & \perf{0.09}{0.04} \\ \hline
\multicolumn{1}{|l|}{\SC{Ours (ACO)}}
& \perf{0.26}{0.06} & \perfb{0.12}{0.07} & \perfb{0.13}{0.05} & \perfb{0.53}{0.03} & \perfb{0.36}{0.06} & \perfb{0.26}{0.07} & \perfb{0.13}{0.02} \\ \hline
\end{tabular}
}
\caption{\textbf{Compatibility score ablations.} Simple similarity measures (\SC{WordEmbed}) or naively using interaction labels in video (\SC{IntSeq}) produces sub-optimal policies. Our activity-context detection based priors perform the best across the majority of tasks. Results are averaged over 3 runs.}
\label{tbl:ablations}
\vspace{-0.15in}
\end{table*}
\vspace{-0.05in}
\subsection{Compatibility function ablations} \label{sec:ablations}
Finally we investigate how different sources of priors impact our agents when incorporated into our reward scheme. Specifically, we vary how we compute our compatibility score $\phi(o_i, o_j)$ in \refeqn{eqn:aco_score}. \B{\SC{Uniform}} assigns equal compatibility to all object pairs; \B{\SC{WordEmbed}} uses Glove~\cite{pennington2014glove} embedding similarity; \B{\SC{SpatialCooc}} uses spatial co-occurrence statistics derived from static images\footnote{This is different from \SC{ScenePriors} (\refsec{sec:vsp_results}) which encodes the same priors in the state representation.}; \B{\SC{Ours (IntSeq)}} uses transition probabilities between object interactions from ground truth, labeled sequences of the same egocentric videos.
\reftbl{tbl:ablations} shows the results. Uniform or semantic similarity heuristics are insufficient to capture the role of objects participating in activities. \SC{SpatialCooc}'s object co-occurrences are aligned for some tasks (e.g., vegetables are co-located with fridges in static images, and activities involving storing groceries), but also capture unhelpful static relationships (see \reffig{fig:epic_to_acos}, right).
\SC{IntSeq} benefits from egocentric video; however, it uses ground truth video object labels, and falsely assumes that the next object interacted with in time is necessarily relevant to the previous object. %
Our activity-context score is tightly linked to each object's involvement in interactions, and proves to be the most useful prior. %
|
\section{Introduction}
It is well-known that functionals of Brownian motion or, more generally, L\'{e}vy processes appear in pricing economic assets. Often these processes have infinite variation and hence are incapable of handling situation where the price of an asset is constant for a period of time. Therefore changes to the stochastic model should be made when the underlying assets are
illiquid, i.e. the assets cannot be traded quickly. One attempt to overcome this shortcoming is to introduce a random time-change $E = (E(t);t\ge 0)$ which is the inverse of a subordinator \cite{HL.21,Magdziarz.09}. See also \cite{GMS.01, MRGS.00, SGM.00} for an approach involving discrete observations of prices and a limiting procedure, whose idea originates from \cite{MS.65} in physics. The governing Fokker-Planck equation for the non-time-changed stochastic process is often replaced with a partial differential equation (PDE) where the time derivative is replaced with some fractional time derivative (time-fractional PDE). See \cite{HL.21} for an example involving illiquid markets and see \cite{Chen.17} and references therein for a more general approach of using time-changed stochastic processes to solve time-fractional PDEs. These time-changed processes, or the governing time-fractional PDEs, have been used to study subdiffusive behavior in physics as well, where the motionless behavior stems from trapping of particles \cite{MS.19,MK.04}.
For concreteness, let us turn to the subdiffusive Black-Scholes model with zero interest rate. In the Black-Scholes model, the price of a stock is modeled by a geometric Brownian motion $X = (X(t);t\ge 0)$. In the subdiffusive model, corresponding to an illiquid market, one has to modify the stock price by introducing a time-change $E = (E(t);t\ge 0)$ which is the inverse of a strictly increasing L\'{e}vy process $S = (S(t);t\ge 0)$ independent of the process $X$. That is, in the illiquid market the stock price is modeled by process $X^* = (X^*(t);t\in[0,T])$ defined by $X^*(t) = X(E(t))$ where
\begin{equation}\label{eqn:Edef}
E(t) = \inf\{u: S(u)>t\}.
\end{equation}
This model is considered in \cite{Magdziarz.09, MG.12} where the authors show that this model of an illiquid market is arbitrage-free but incomplete. One has to be careful about how an interest rate is incorporated - see \cite{MG.12}.
Standard barrier options are among the most popular derivative contracts \cite{carr1995two}. Depending on its type, the value of the option depends on whether the underlying asset price has or has not hit some predefined price, the barrier, before the maturity date. Due to the one-touch nature of these options, influential agents can
manipulate the price of the underlying asset when the price is close to the barriers. To mitigate this problem Chesney et. al. \cite{CJY.97} introduced the so-called Parisian barrier option which depends on the occupation times of the price of underlying assets.
Instead of the value of the option depending on whether or not the underlying asset hits some predefined level, the underlying asset has to spend a predefined amount of time above or below the barrier before maturation. For example, the owner of a down-and-out option loses the option when the asset price $X$ remains below level $L$ for $D$ units of time and this has to occur before maturation at time $T$. Therefore, in order to understand how to price Parisian options, one needs to understand properties of the occupation measure and excursions of the process $X$.
There is a vast amount of literature involving the occupation times of Brownian motion. In particular, the length of the longest excursion of a Brownian motion on some fixed (possibly random) time interval has been well-studied. See \cite{PY.97} for a survey and generalizations of this. This literature allowed the authors of \cite{CJY.97} to obtain some complicated, although explicit, formulas for the valuation of Parisian options using Laplace transforms. With this said, it appears that Parisian options for the time-changed processes is understudied. This paper is devoted to filling in some of the mathematical gaps in this literature. We develop an excursion theory for time-changed Brownian motion and then use this to price a down-and-out European Parisian option when the stock price is modeled by a time-changed geometric Brownian motion. We relate this price to solutions of time-fractional PDEs. We also develop some process-level descriptions relating the occupation measures for processes time-changed by inverse subordinators to those of the original non-time-changed process. In particular, we obtain an explicit Ray-Knight type theorem for a time-changed Brownian motion (with drift) killed upon hitting zero.
\subsection{Main Results}
\subsubsection{Excursion Intervals of $X^*$}
Consider the situation where $X$ is a reflected Brownian motion and $S$ is an independent $\beta$-stable subordinator for $\beta\in(0,1)$. The process $S$ is uniquely characterized by its Laplace transform
\begin{equation}\label{eqn:beta_Stable}
\mathbb{E}\left[ \exp(-\lambda S(t)) \right] = \exp(-ct\lambda^\beta), \qquad \forall t, \lambda>0
\end{equation} for some constant $c>0$. See \cite[Chapter III]{Bertoin.96} for more information on subordinators. Since its L\'{e}vy measure $\nu(dy) = \tilde{c} y^{-1-\beta}1_{[y>0]}\,dy$ has infinite mass, $S$ is strictly increasing.
Before turning to our first theorem, let us briefly recall the Poisson-Dirichlet distribution $\operatorname{PD}(\alpha,\theta)$ as in \cite{PY.97}. Since our results only rely on $\operatorname{PD}(\alpha,0)$, we will focus on this case. Let $J_1\ge J_2\ge \dotsm >0$ be a rearrangement of all the jumps of $S$ on the interval $[0,1]$:
\begin{equation*}
\{J_1,J_2,\dotsm\} = \{S(t)-S(t-): t\in[0,1]\} \setminus \{0\}.
\end{equation*} Consider the random element $V = (V_1,V_2,\dotsm)$ of the infinite simplex $\Delta_\infty:= \{(x_j)\in [0,1]^\infty: x_1\ge x_2\ge \dotsm \ge 0, \sum_{j} x_j = 1\}$ defined by $V_n = J_n/S(1) = J_n/ \sum_{i} J_i$. Then $\operatorname{PD}(\beta,0)$ is the law of the vector $V$ in $\Delta_\infty$.
After developing some excursion theory for the time-changed reflected Brownian motion $X^*$, we obtain the following Theorem:
\begin{thm}\label{thm:PD}
Let $X$ denote a reflected Brownian motion, and let $X^*(t) = X(E(t))$ where $E$ is the right-continuous inverse of a $\beta$-stable subordinator defined by \eqref{eqn:Edef}. Then the ordered lengths of the relatively open intervals of $\{u\in[0,t]: X^*(u)>0\}$, say $U_1(t)\ge U_2(t)\ge\dotsm> 0$, are distributed as $t$ times a $\operatorname{PD}(\beta/2,0)$ random variable:
\begin{equation*}
\left(\frac{U_1(t)}{t},\frac{U_2(t)}{t},\dotsm\right) \overset{d}{=} \operatorname{PD}\left(\frac{\beta}{2},0\right).
\end{equation*}
\end{thm}
Observe that this conclusion does not depend on the particular value of $c$ in \eqref{eqn:beta_Stable}. While this gives the joint distribution of the lengths of all the excursions (appropriately reordered), it is restricted to simply the case of a stable subordinator and the excursions of a reflected Brownian motion at zero. More careful analysis would have to be done in order to handle the case of time-changing a Brownian motion (with drift) by a subordinator which is not stable or even when examining the excursions above a level $b>0$.
\subsubsection{Price of an option}
Recall the subdiffusive Black-Scholes model discussed above. The process $X^*$ is the solution to
\begin{equation}\label{eqn:agbm}
dX^*(t) = X^*(t) \left( \mu dE(t) + \sigma d W(E(t))\right),\qquad X^*(0) = x>0,
\end{equation} where $W$ is a standard Brownian motion. Here $E$ is defined by \eqref{eqn:Edef} for some strictly increasing subordinator $S$ is independent of $W$. This is defined on some filtered probability space $(\Omega,\mathcal{F},\{\mathcal{F}_t\}_{t\ge 0} ,\tilde{\mathbb{Q}})$ where $\mathcal{F}_t$ is the standard augmentation of $\sigma\left\{X^*(s);s\le t\right\}$ and $\tilde{\mathbb{Q}}$ is the physical measure. That is, $W$ is a $\tilde{\mathbb{Q}}$-Brownian motion. As is described in \cite{MG.12}, there exists a (not unique!) equivalent martingale measure $\mathbb{Q}$ under which $X^*$ is an $(\mathcal{F}_t)$-martingale.
A Parisian option depends on several parameters which we now fix: the barrier level $L$; an excursion duration $D$; and a maturation time $T$. Consider the first time $H_{L,D}^-(w)$ that a continuous path $w:\mathbb{R}_+\to \mathbb{R}$ stays strictly below level $L$ for duration $D$. That is
\begin{equation}\label{eqn:HLD_def}
H_{L,D}^-(w) = \inf\{t: 1_{[w(t)< L]} (t- g_L^w(t))> D \},\qquad g_L^w(t) = \sup\{r\le t: w(r) = L\},
\end{equation} with the convention that $\sup\emptyset = 0$ and $\inf\emptyset = +\infty$. Note that $g_{L}^{w}(t)$ is the last time the path $w$ hits level $L$. The down-and-in European Parisian option with payoff $\Phi$ is therefore priced as \begin{equation}\label{eqn:CdiForTimeChange}
V_{d,i}^*(x, T,L,D,\Phi) = \mathbb{E}_{\mathbb{Q}} \left[1_{[H^-_{L,D}(X^*)\le T]}\Phi(X^*(T)) \right].
\end{equation}
The next theorem describes the price of a down-and-in Parisian option in terms of solutions to time-fractional partial differential equations when the barrier level $L$ is $x$. For more information on time-fractional PDEs, see Section \ref{sec:timeFracPDE}. See also equation \eqref{eqn:genCaputo} for the definition of the generalized time-fractional derivative $\partial_{t}^{S}$.
\begin{thm}\label{thm:priceTheorem}
Consider the above set-up. Assume the L\'evy measure of $S$ has infinite mass and denote by $\phi$ the Laplace exponent of $S$ as in \eqref{eqn:LaplaceExponent}. Define the measure $\pi$ on $\mathbb{R}_+^2$ by $\pi(A\times B) = \int_{B} \mathbb{P}(S(t)\in A) \frac{1}{\sqrt{2\pi t^3}}\,dt$. Let $\pi^+(A\times B) := \pi\left( (A\cap (D,\infty))\times B\right)$, $\pi^{-}(A\times B):=\pi^-((A\cap (0,D])\times B)$ and let $\mu = \pi^+(\mathbb{R}_+\times \mathbb{R}_+)<\infty$.
Suppose that $\Phi:\mathbb{R}_+\to \mathbb{R}$ is smooth and vanishes in a neighborhood containing $x$, the initial stock price. Then a price of the down-and-in European Parisian option with barrier $L = x$ and payoff $\Phi$ is
\begin{equation*}
\begin{split}
V_{d,i}^*(x, T, x, D, \Phi) &= \int 1_{[s\le T-D]} e^{-\frac{\sigma^2}{8} v} u_1(0,0; T-s) m_1(ds,dv) \\
&\qquad + \int 1_{[D\le s \le T]} e^{-\frac{\sigma^2}{8} v} u_2(0,0;T-s)\,m_2(ds,dv),
\end{split}
\end{equation*}
where the Laplace transforms of the measures $m_j(ds,dv)$ are
\begin{equation*}
\begin{split} \int e^{-\lambda s - \theta v}\,m_1(ds,dv) &= \frac{\mu}{\displaystyle \mu + \sqrt{\frac{1}{2}(\phi(\lambda)+\theta)} +\frac{1}{2} \int_{\mathbb{R}_+^2} (1-e^{-\lambda s - \theta v}) \pi^-(ds,dv)},\\
\int e^{-\lambda s - \theta v}\,m_2(ds,dv) &= \frac{\displaystyle \frac{1}{2}\int_{\mathbb{R}_+^2} e^{-\lambda s - \theta v} \pi^+(ds,dv)}{\displaystyle \mu + \sqrt{\frac{1}{2}(\phi(\lambda)+\theta)} +\frac{1}{2} \int_{\mathbb{R}_+^2} (1-e^{-\lambda s - \theta v}) \pi^-(ds,dv)};
\end{split}
\end{equation*}
here $u_1(z,s;t)$ is the solution to the time-fractional PDE
\begin{equation*}
\left\{\begin{array}{l}
\partial_t^S u = \left(\frac{1}{2} \frac{\partial^2}{\partial z^2} + \frac{1}{2z} \frac{\partial}{\partial z} + \frac{\partial}{\partial s}\right) u\\
u(z,s;0) = \exp\left\{-\frac{\sigma}{2} z - \frac{\sigma^2}{8} s\right\} \Phi\left( xe^{-\sigma z}\right)
\end{array}
\right.;
\end{equation*}
and $u_2$ is the solution to
\begin{equation*}
\left\{\begin{array}{l}
\partial_t^S u = \left(\frac{1}{2} \frac{\partial^2}{\partial z^2} + \frac{\partial}{\partial s}\right) u\\
u(z,s;0) = \exp\left\{\frac{\sigma}{2} z - \frac{\sigma^2}{8} s\right\} \Phi(xe^{\sigma z})
\end{array}
\right..
\end{equation*}
\end{thm}
\begin{rmk}
The requirement that $\Phi$ is continuous means that this result does \textit{not} give us the prices of the down-and-in Parisian calls or puts. We can recover the price of the Parisian call option with payoff function $(\cdot -K)_{+}$ by a limiting procedure provided that $K\ne x$. We \textbf{are not claiming} there is a corresponding representation in terms of time-fractional PDE, which would require additional work.
\end{rmk}
\begin{rmk}
The requirement that $\Phi$ vanishes near the initial stock price is not the most general statement we can make. It can also be stated that $z\mapsto e^{-\frac{\sigma}{2}{z}} \Phi(xe^{-\sigma z})$ belongs to the domain of the infinitesimal generator of a 2-dimensional Bessel process which can be found in Section 2.7 of \cite{IM.74}.
\end{rmk}
\subsubsection{Ray-Knight type theorems}
Theorem \ref{thm:PD} above gives some information on lengths of excursions while it does not give us detailed information on the occupation measures of $X^*$. In Section \ref{sec:occupationChange} we study the effect that the time-change has on the occupation measure. That is, we relate the two occupation measures $\nu(I;-)$ and $\nu^*(I;-)$ for intervals $I\subset \mathbb{R}_+$ where
\begin{equation*}
\nu(I; A):= {\operatorname{Leb}}\{t\in I: X(t) \in A\},\qquad \nu^*(I; A):= {\operatorname{Leb}}\{t\in I: X^*(t)\in A\},
\end{equation*} and ${\operatorname{Leb}}$ denotes the Lebesgue measure. One of the main theorems we prove in Section \ref{sec:occupationChange} is the following:
\begin{thm} \label{thm:timeChange1}
Fix a possibly random time $\tau\in [0,\infty]$ independent of $S$, and let $\tau^* = S(\tau)$ for some strictly increasing subordinator $S$ (with the convention $S(\infty) = \infty$). Define the increasing processes $A$ and $A^*$ by
\begin{equation*}
A(v) = \nu([0,\tau]; (-\infty,v]),\qquad \text{and}\qquad A^*(v) = \nu^*([0,\tau^*];(-\infty,v]).
\end{equation*} Then
\begin{equation*}
\left(A^*(v);v\in\mathbb{R}\right) \overset{d}{=} \left(S(A(v));v\in \mathbb{R}\right).
\end{equation*}
\end{thm}
\begin{rmk}
The proof of Theorem \ref{thm:timeChange1} can be altered to prove an analogous result when the left endpoint of the intervals is not zero.
\end{rmk}
The Ray-Knight theorems give us detailed information on the form of the occupation measure for a standard Brownian motion \cite[Section XI.2]{RY.99}. These results have been extended in numerous ways. In particular, Warren \cite{Warren.97} describes the occupation measure of a Brownian motion with negative drift. We apply Theorem \ref{thm:timeChange1} to Warren's result and get Theorem \ref{thm:1d}. There is a similar Ray-Knight description for the occupation measure of a 3-dimensional Bessel process \cite{MY.08} which would allow one to prove a Bessel process version of Theorem \ref{thm:1d}.
\subsection{Overview of Paper}
In Section \ref{sec:prelim} we discuss some preliminaries used frequently in the remainder of the paper. This includes a discussion on pricing Parisian options where the stock price is a geometric Brownian motion in Section \ref{sec:parisianPrice} and a separate discussion on the connection between time-fractional equations and their probabilistic representations of solutions in Section \ref{sec:timeFracPDE}.
In Section \ref{sec:ExcursionTheorySection} we describe an excursion theory of time-changed Brownian motion. Since the time-change results in the process being non-Markov we do this using
L\'{e}vy's construction of reflected Brownian motion. We recap the Brownian case in Section \ref{sec:ExcursionsForBM} and then include the time-change in Section \ref{sec:timeChangeExcursions} where we define the excursion measure. We then describe some basic properties of this measure in Section \ref{sec:basicn*}, wherein we prove Theorem \ref{thm:PD}. The remainder of the section is devoted to providing an explicit way to construct the excursion process of a time-changed Brownian motion from that of the original one.
We then use this excursion theory to describe the first negative excursion of a time-changed Brownian motion $B^*$ which lasts longer than duration $D$. This is included in Section \ref{sec:firstExcursion}, where we find the Laplace transforms for the two measures $m_j$ of Theorem \ref{thm:priceTheorem}. In Section \ref{sec:illiquidMarkets}, we finish the proof of Theorem \ref{thm:priceTheorem}. Finally, in Section \ref{sec:occupationChange} we prove Theorem \ref{thm:timeChange1} and then use this idea to compute the Laplace transform of integral functionals of time-changed Brownian motion killed at zero in Section \ref{sec:rayKnight}.
\section{Preliminaries} \label{sec:prelim}
\subsection{Poisson Random Measures}
We briefly recall the definition of a Poisson random measure (PRM) and some of their basic properties. We do this in order to describe the PRM construction of subordinators. For more information see \cite[Chapters 1-2]{Kingman.93} and \cite[Chapter 8]{Cinlar.11}.
A PRM with intensity $\mu$ on a measure space $(U,\mathcal{U},\mu)$ is a random discrete measure $N$ on $(U,\mathcal{U})$ such that for any finite collection of disjoint measurable sets $U_1,\cdots,U_n$, the random variables $N(U_1),\dotsm, N(U_n)$ are independent Poisson with means $\mathbb{E}[N(U_j)] = \mu(U_j)$. It is well-known that if $\mu$ is $\sigma$-finite
then there exists a Poisson random measure with intensity $\mu$.
One of the first results about PRMs is the so-called mapping theorem which we include below. A proof of the result can be found in \cite[Chapter 2]{Kingman.93} in the case of atomless measures and \cite[Chapter 8]{Cinlar.11} in general
\begin{lem}\label{lem:mappingTheorem}
Suppose that $N$ is a Poisson random measure on a space $(U,\mathcal{U})$ with intensity measure $\mu$. Given another measurable space $(V,\mathcal{V})$, and a measurable map $f:U\to V$, define a new measure on $V$ by $N^{\ast} = f_\#N = N\circ f^{-1}$. Then $N^{\ast}$ is a Poisson random measure on $(V,\mathcal{V})$ with intensity measure $f_\#\mu$.
\end{lem}
\subsection{Subordinators}
\begin{comment}
Let $X = (X(t);t\ge 0)$ be a spectrally positive L\'{e}vy process, i.e. one with no negative jumps. It can be uniquely characterized by its Laplace transform
\begin{equation*}
\mathbb{E}[\exp(-\lambda X(t))] = \exp(t\psi(\lambda)),\qquad \lambda,t\ge 0,
\end{equation*} where $\psi$ is of the form
\begin{equation*}
\psi(\lambda) = \alpha\lambda+ \beta\lambda^2 + \int\limits_{(0,\infty)} (e^{-\lambda r} - 1 + \lambda r 1_{[r\le 1]})\,\tilde{\pi}(dr),
\end{equation*} for some constants $\beta\ge 0$, $\alpha\in \mathbb{R}$ and a Radon measure $\tilde{\pi}$ on $(0,\infty)$ such that $(1\wedge r^2)\,\tilde{\pi}(dr)$ is a finite measure.
\end{comment}
Increasing L\'{e}vy processes are called subordinators. Most of the results below are recalled from Chapter III of Bertoin's monograph \cite{Bertoin.96}. A subordinator $S = (S(t);t\ge 0)$ can be uniquely characterized by its Laplace transform, which takes the form
\begin{equation} \label{eqn:LaplaceExponent}
\mathbb{E}[\exp(-\lambda S(t))] = \exp(-t\phi(\lambda)),\qquad \lambda,t\ge 0
\end{equation}
where
\begin{equation} \label{eqn:phiForLP}
\phi(\lambda) = \kappa \lambda + \int\limits_{(0,\infty)} (1-e^{-\lambda r})\,\nu(dr),
\end{equation} where $\kappa \ge 0$ and $\nu$ is a Radon measure on $(0,\infty)$ such that $\int _{(0,\infty)} (1\wedge r)\,\nu(dr)<\infty$. In the sequel we will say that $\phi$ is the Laplace exponent of $S$.
The L\'{e}vy-It\^{o} decomposition of L\'{e}vy processes provides a simple description of subordinators in terms of PRMs. Namely, let $M$ be a PRM on $[0,\infty)\times(0,\infty)$ with intensity measure ${\operatorname{Leb}}\otimes \pi$, then the process $S$ defined by
\begin{equation}\label{eqn:PRM_subordinator}
S(t) = \kappa t + \int\limits_{[0,t]} \int\limits_{(0,\infty)} y\,M(ds,\,dy) = \kappa t + \int\limits_{(0,\infty)} y\,M([0,t],dy)
\end{equation}
is, as the notation suggests, a subordinator with Laplace exponent $\phi$ in \eqref{eqn:phiForLP}. See, for example, \cite[Chapter 1]{Bertoin.96} and \cite[Proposition 6.4.6]{Cinlar.11} for a proof. In the sequel we will implicitly use the above decomposition often without reference.
\subsection{Parisian Options}\label{sec:parisianPrice}
In this section we discuss how to price a Parisian option as discussed in \cite{CJY.97}, when the underlying stock price is governed by a geometric Brownian motion. That is, $X$ is the unique solution to
\begin{equation}\label{eqn:gmb}
dX(t) = X(t)\left( \mu \,dt + \sigma d W(t)\right),\qquad X(0) = x>0,
\end{equation} where $W$ is a standard Brownian motion under the measure $\tilde{\mathbb{Q}}$ (the reason for this choice of symbol will be apparent later). More concretely, the process $W$ is the canonical Brownian motion on the probability space $\Omega = C([0,T]\to \mathbb{R})$, and $\mathcal{F}_t$ is the standard augmentation of $\sigma(W(s); s\le t)$ under the Wiener measure $\tilde{\mathbb{Q}}$. For simplicity, we assume that the interest rate $r$ is equal to $0$. This model is arbitrage-free and complete.
A European option with maturity date $T$ and payoff function $\Phi(X_T)$ is priced by
\begin{equation*}
\mathbb{E}_\mathbb{Q}[ \Phi(X_T)],\qquad \text{where}\ \mathbb{Q} \text{ is the unique risk neutral measure}.
\end{equation*}
The measure $\mathbb{Q}$ is the unique measure which turns $X(t)$ into an $(\mathcal F_{t})$-martingale under $\mathbb{Q}$, and is given by Girsanov's theorem \cite[Chapter VIII]{RY.99} by
\begin{equation*}
\mathbb{Q}(A) = \int_A \exp\left(-\frac{\mu}{\sigma} W(T) - \frac{\mu^2}{2\sigma^2} T \right) \,d\tilde{\mathbb{Q}}.
\end{equation*}
Under the measure $\mathbb{Q}$, the process $X$ satisfies
\begin{equation}\label{eqn:XExplicit1}
X(t) = x \exp\left(\sigma Z(t) - \frac{\sigma^2}{2} t \right),
\end{equation} where $Z(t) = W(t) + \frac{\mu}{\sigma }t$ is a $\mathbb{Q}$-Brownian motion.
Fix a barrier $L>0$, a duration $D>0$ and payoff function $\Phi$. The down-and-in European Parisian option pays out $\Phi(X(T))$ but is only valid on the event that the stock price $X$ stays strictly below level $L$ for duration at least $D$. Thus, the price of this down-and-in option, which we denote by $V_{d,i}(x,T,L,D,\Phi)$ is given by
\begin{equation*}
V_{d,i}(x,T,L,D,\Phi) = \mathbb{E}_\mathbb{Q}\left[1_{[H_{L,D}^-(X) \le T]}\Phi(X(T)) \right],
\end{equation*}
where $H_{L,D}^-$ is defined in \eqref{eqn:HLD_def}.
We can write $B(t) =Z(t) - \frac{\sigma}{2}t$, and use Girsanov's theorem to turn $B$ into a Brownian motion under a measure $\mathbb{P}$ defined by
\begin{equation*}
\frac{d \mathbb{Q}}{ d\mathbb{P}} = \exp\left( \frac{\sigma}{2} B(T) - \frac{\sigma^2}{8} T\right) = \exp \left(\frac{\sigma}{2} Z(T) + \frac{\sigma^2}{8}T \right).
\end{equation*}
By \eqref{eqn:XExplicit1} $X(t) = x \exp(\sigma B(t))$. It is easy to see that an excursion of $X$ below level $L$ of duration $D$ is equivalent to an excursion of $B$ below level $b = \frac{1}{\sigma} \log (L/x)$ of duration $D$. Thus
\begin{equation*}
V_{d,i}(x,T,L,D,\Phi) = e^{-\sigma^2 T/8} \mathbb{E}_{\mathbb{P}} \left[1_{[H_{b,D}^-(B) \le T]} \exp(\frac{\sigma}{2} B(T))\Phi(xe^{\sigma B(T)})\right].
\end{equation*}
The authors of \cite{CJY.97} give explicit formulas for the price of the Parisian call option, i.e. when $\Phi(x) = (x-K)_+$ for some price $K$.
\subsection{Time-fractional partial differential equations for time-changed diffusions}\label{sec:timeFracPDE}
Over the past $50$ years, fractional calculus has been drawn attention by a number of researchers. In addition to the theoretical aspect of its own \cite{SKM.1993}, \cite{K.2006}, its attractiveness is also due to the wide range of applicability in a number of areas including physics, chemistry, signal processing, modeling and control, economics and even social science \cite{MS.19, D.2003, P.1998, T.2019, S.2010}. To see more on the literature of the development of fractional calculus and its applications during 1974-2010, see \cite{M.2011}.
In particular, fractional calculus has been widely used to model anomalous diffusion in physics. Anomalous diffusion is a diffusion whose individual particle travels at a different rate than in a traditional one. Especially, anomalous subdiffusion describes the phenomena of particle sticking or trapping in an inhomogeneous medium \cite{L.2010, MS.19}. The time-fractional heat equation
\begin{equation} \label{e:brownian}
\partial_{t}^{\alpha} u(x,t) = \Delta u(x,t),
\end{equation}
where $\partial^{\alpha}$ is the Caputo derivative with order $0<\alpha<1$ defined by
\begin{equation} \label{e:Caputo}
\partial_{t}^{\alpha} f(t) = \frac{1}{\Gamma(1-\alpha)} \frac{d}{dt} \int_{0}^{t} (t-x)^{-\alpha} (f(x)-f(0)) \, dx,
\end{equation}
captures such subdiffusive behavior, just as the classical heat equation $\partial_{t} u(x,t) = \Delta u(x,t)$ describes normal diffusion. Here $\Gamma(x):= \int_{0}^{\infty} t^{x-1}e^{-x}dx$ is the Gamma function. In 2004, Meerschaert and Scheffler \cite[Theorem 5.1]{MS.04} proved, in light of \cite[Theorem 3.1]{BM.01}, that the solution of \eqref{e:brownian} with initial condition $u(x,0) = f(x)$ for some $f\in C^{2}$ such that $f$ and $\Delta f$ vanish at infinity can be represented probabilistically as
\begin{equation} \label{e:probrep}
u(x,t) = \mathbb E_{x}[ f(X(E(t)))]
\end{equation}
where $X$ is a scaled Brownian motion in $\mathbb R^d$ starting from $x$ whose infinitesimal generator is $\Delta$ and $E$ is the right-continuous inverse of an $\alpha$-stable subordinator $S$ independent of $X$ defined by \eqref{eqn:Edef}.
The process $E$ is often called the inverse subordinator or the first hitting time process. In fact, the above-mentioned result still holds when the spatial process $X$ is a strong Markov process with a more general infinitesimal generator $\mathcal L$ in place of the Laplacian $\Delta$ in \eqref{e:brownian}. More recently, Chen \cite[Theorem 2.3]{Chen.17} proved that such interplay between time-fractional PDEs and the probabilistic representation of its solution works even when the time process $S$ is a subordinator satisfying more general conditions. The function $u(x,t)$ in \eqref{e:probrep} satisfies a more general time-fractional differential equation of the form
\begin{equation*}
\partial_{t}^{S} u(x,t) = \mathcal L u(x,t) \quad \textnormal{with}\ u(x,0)=f(x)
\end{equation*}
where $X$ is a strong Markov process with infinitesimal generator $\mathcal L$ and $S$ is any increasing L\'{e}vy process with drift $\kappa$ and L\'{e}vy measure $\nu$ satisfying $\nu(0,\infty)=\infty$. Here $f$ belongs to the domain of the operator $\mathcal L$.
The generalized time-fractional derivative above is
\begin{equation} \label{eqn:genCaputo}
\partial_{t}^{S} f(t) = \kappa \frac{d }{d t} f(t) + \frac{d}{dt} \int_{0}^{t} w(t-x)(f(x)-f(0)) \, dx ,\qquad \text{where}\ \ w(x) = \nu(x,\infty).
\end{equation}
This reduces to (a constant multiple of) the classical $\alpha$-th order Caputo derivative \eqref{e:Caputo} with $0<\alpha<1$ when $S$ is an $\alpha$-stable subordinator. Moreover, Chen \cite{Chen.17} shows that this solution is unique in some strong sense. See therein for a detailed formulation.
\section{The Excursion Theory of Time-changed Brownian motion}\label{sec:ExcursionTheorySection}
Let us consider the following set-up. We have a standard linear Brownian motion $B = (B(t);t\ge 0)$ and we have an independent subordinator $S = (S(t);t\ge 0)$ which is assumed to be strictly increasing. We let $E = (E(t);t\ge 0)$ denote the right-continuous inverse of $S$ defined by \eqref{eqn:Edef}.
By L\'{e}vy's construction of reflected Brownian motion \cite[Theorem VI.2.3]{RY.99} we know that $R = (R(t);t\ge 0)$ defined by
\begin{equation*}
R(t) = B(t) - I(t),\qquad I(t):= \inf_{s\le t} B(s)
\end{equation*} is a reflected Brownian motion, and its local time at zero is $-I(t)$. We will relate excursions of the time-changed process $R^* = (R^*(t);t\ge 0)$ where $R^*(t) = R(E(t))$ to those of $R$. Similarly, we relate their excursion measures.
\subsection{The Excursions of a Reflected Brownian motion}\label{sec:ExcursionsForBM}
Let us begin by recalling the construction of the It\^{o} measure for the reflected Brownian motion $R$. In order to do this we let $(\mathcal{F}_t;t \ge 0)$ denote the standard augmentation of the filtration $\sigma(B(u),S(u); u\le t)$. Now define
\begin{equation*}
\tau_x = \inf\{t: -I(t)>x\}.
\end{equation*} The time $\tau_x$ is the first time when the Brownian motion $B$ is below level $-x$. The times $(\tau_x;x>0)$ index the excursion intervals of the reflected Brownian motion $R$.
To define these excursions, we let $\mathscr{E}$ be the set of positive excursions and a cemetery state:
\begin{equation*}
\mathscr{E} = \left\{w\in C(\mathbb{R}_+,\mathbb{R}_+): w(0) = 0, \text{ and } \exists \zeta = \zeta(w)\text{ s.t. } w(t)>0 \text{ iff }t\in(0,\zeta) \right\} \cup\{\partial\}.
\end{equation*}
Here $\partial$ is the cemetery state. This set is a Borel subset of $C(\mathbb{R}_+,\mathbb{R}_+)\cup\{\partial\}$ and when we refer to measurability in $\mathscr{E}$ we mean with respect to the Borel $\sigma$-algebra as a subset of $C(\mathbb{R}_+,\mathbb{R}_+)$.
We define the excursions process $(e_x; x>0)$ of $R$ by
\begin{equation*}
e_x(t) = \left\{\begin{array}{ll}
R((t+\tau_{x-})\wedge \tau_x)&: \tau_x- \tau_{x-}>0\\
\partial &: \text{else}
\end{array}
\right..
\end{equation*}
It is well known that $(e_x;x>0)$ is a Poisson point process on $\mathscr{E}$ with respect to the filtration $(\mathcal{F}_{\tau_x};x>0)$. Its intensity measure is the It\^{o} measure (of positive excursions) on $\mathscr{E}$, which is defined as
\begin{equation*}
\mathfrak{n}(A):= \frac{1}{x} \mathbb{E}\left[\sum_{0<y\le x} 1_{[e_y\in A]}\right].
\end{equation*} The scalings of this measure vary throughout the literature depending on conventions of the local time, so we will mention the following identities useful in determining the scaling:
\begin{equation}\label{eqn:easyPropsforn}
\mathfrak{n}(\zeta>\varepsilon) = \sqrt{\frac{2}{\pi \varepsilon}} ,\qquad\mathfrak{n}\left( \sup_{t\le \zeta} w(t)> \varepsilon\right) = \frac{1}{\varepsilon},\qquad \forall \varepsilon>0.
\end{equation} See \cite{LeGall.10} for a proof with a different scaling; however, see around \eqref{eqn:laplace(t)au} for a proof that \eqref{eqn:easyPropsforn} is the correct scaling.
\subsection{The time-change}\label{sec:timeChangeExcursions}
In this section we discuss in more detail the time-change $t\mapsto E(t)$ and its relation to the excursions of $R^*$. We let $B^*(t) = B(E(t))$ and $I^*(t) = \inf_{s\le t} B^*(s) = I(E(t))$. Let us write $\mathcal{F}^*_t$ for the usual augmentation of $\sigma(B(E(s)),s\le t)$.
In order to index the excursions of $R^*$, we need some analog for $\tau_x$. Since $R^*$ is not Markovian, it is difficult to construct a useful local time at zero from just $R^*$. However, by Skorohod reflection, $-I^*(t)$ acts as a local time at level zero in the sense that $-I^*(t)$ only increases on the set $\{t:R^*(t) = 0\}$. Therefore we set $\hat{\tau}_x$ as
\begin{equation*}
\hat{\tau}_x = \inf\{t: -I^*(t)>x\}.
\end{equation*} We observe the following simple relationship between the times $\hat{\tau}_x$ and $\tau_x$ which will be used numerous times in the sequel:
\begin{equation}\label{eqn:hatTauandS}
\hat{\tau}_x = \inf\{t:-I(E(t))>x\} = \inf\{S(u): -I(u)>x\} = S({\tau_x}).
\end{equation} Indeed, the second equality occurs with probability 1. The process $S$ almost surely does not jump on any independent Lebesgue null set. To be more specific, it will not jump on the independent null set $\{t: R(t) = 0\}$. The third equality holds by the right-continuity of the subordinator $S$.
The times $\hat{\tau}_x$ are easily seen to be $\mathcal{F}^*_t$-stopping times. Moreover, $\mathcal{F}^*_{\hat{\tau}_x} = \mathcal{F}_{\tau_x}$. Indeed, since $E = \langle B(E(\cdot))\rangle$ is the quadratic covariation of $B(E(t))$, and so the filtration $\mathcal{F}^*_t$ is the standard augmentation of $\sigma (B(E(s)),E(s); s\le t)$. Since $S(t) = \inf\{u: E(u)>t\}$, $S(t)$ is an $\mathcal{F}^*_t$-stopping time. The claim that $\mathcal{F}^*_{\hat{\tau}_x} = \mathcal{F}_{\tau_x}$ follows from $\mathcal{F}^*_{S(t)} = \mathcal{F}_t$ and \eqref{eqn:hatTauandS}. We now argue that $\mathcal{F}^*_{S(t)} = \mathcal{F}_t$.
First, $\mathcal{F}_{t}\subset \mathcal{F}_{S(t)}^*$. Indeed, for any fixed $r\le t$, $B(r) = B(E(S(r)))$ and $S(r)$ are both $\mathcal{F}_{S(t)}^*$-measurable. Hence, $\sigma(B(r),S(r); r\le t)\subset \mathcal{F}_{S(t)}^*$ and so $\mathcal{F}_t\subset \mathcal{F}_{S(t)}^*$.
To show the converse, we begin by showing that $\mathcal{F}^*_t \subset \mathcal{F}_{E(t)}$. The latter $\sigma$-algebra is well defined because $E(t)$ is an $(\mathcal{F}_t)$-stopping time for each fixed $t$. Now, for each $r\le t$, $B(E(r))$ is clearly $\mathcal{F}_{E(t)}$-measurable. So $\sigma(B(E(r));r\le t)\subset \mathcal{F}_{E(t)}$ and hence $\mathcal{F}_t^*\subset \mathcal{F}_{E(t)}$. Now as previously noted, $S(t)$ is a stopping time with respect to both filtrations $(\mathcal{F}_t^*)$ and $(\mathcal{F}_{E(t)})$. Hence, $\mathcal{F}^*_{S(t)}\subset \mathcal{F}_{E(S(t))}$. However, since $S$ is strictly increasing, $E(S(t)) = t$ almost surely, and since null sets are contained in $\mathcal{F}_0$ we have $\mathcal{F}^*_{S(t)}\subset \mathcal{F}_{E(S(t))} = \mathcal{F}_t$. Hence $\mathcal{F}_{S(t)}^* = \mathcal{F}_t$ as desired.
We now define the following excursion process $(e^*_x;x>0)$ of $R^*$ defined by
\begin{equation*}
e^*_x(t) := \left\{ \begin{array}{ll}
R^*((t+\hat{\tau}_{x-})\wedge \hat{\tau}_x)&: \hat{\tau}_{x}-\hat{\tau}_{x-}>0,\\
\partial &: \text{else}
\end{array}
\right..
\end{equation*}
Since the process $S$ does not jump when $R(t) = 0$, we observe that $e_x^*$ almost surely takes values in the set $\mathscr{E}$. That is, there is no initial trapping at zero.
An important ingredient in the proof that $(e_x;x>0)$ is a Poisson point process is that the process $R$ is a strong Markov process. Unfortunately, the process $R^*$ is not, in general, a Markov process: intervals of constancy for the process $R^*$ can tell us information about the future of the process. In order to overcome this problem, we prove the following proposition.
\begin{prop}\label{prop:increments}
Fix an $x>0$. Define
\begin{equation*}
\tilde{R}(t):= R^*(\hat{\tau}_x+t),\qquad t\ge 0.
\end{equation*} Then $\tilde{R}\overset{d}{=} R^*$ and is independent of $\mathcal{F}^*_{\hat{\tau}_x} = \mathcal{F}_{\tau_x}$.
\end{prop}
\begin{proof}
We begin by noting from \eqref{eqn:hatTauandS} that $\hat{\tau}_x + t = S(\tau_x)+t$. From here observe
\begin{equation*}
E(S(\tau_x)+t) = \inf\{u: S(u)>S(\tau_x)+t\} = \inf\{\tau_x+v: S(\tau_x+v)- S(\tau_x)>t\} = \tau_x+ \tilde{E}(t)
\end{equation*} where $\tilde{E}$ is the right-continuous inverse of $\tilde{S} = (\tilde{S}(v):= S(\tau_x+v)-S(\tau_x);v\ge 0)$. Since $S$ is a L\'{e}vy process it has stationary increments and since $\tau_x$ is an independent random time (it only depends on the Brownian path $B$), the processes $\tilde{S}$ and $S$ have the same law. Hence $\tilde{E}$ and $E$ have the same law. Moreover, both $\tilde{S}$ and $\tilde{E}$ are independent of $\mathcal{F}_{\tau_x}$ since $S$ is strong Markov.
Hence
\begin{equation*}
\tilde{R}(t) = R^*(\hat{\tau}_x+t) = R(E(S(\tau_x)+t)) = R(\tau_x+\tilde{E}(t)).
\end{equation*} By the strong Markov property for $R$, the right-most term, as a process in $t$, is easily seen to be equal in distribution to $R^*$ and is independent of $\mathcal{F}_{\tau_x}$. This proves the desired claim.
\end{proof}
With the above proposition, we can now prove that $(e^*_x;x>0)$ is an $(\mathcal{F}_{\tau_x})_{x} = (\mathcal{F}^*_{\hat{\tau}_x})_x$-Poisson point process.
\begin{lem}
$(e^*_x;x>0)$ is an $(\mathcal{F}^*_{\hat{\tau}_x})$-Poisson point process.
\end{lem}
\begin{proof}
We first begin by observing that it is $\sigma$-discrete. Indeed, the time-change does not change the maximum of an excursion, so there are only a finite number excursions of $(R^*(t); 0\le t\le \hat{\tau}_x)$ which exceed level $\varepsilon$ for any $\varepsilon>0$. It is also easy to verify that $(x,\omega)\mapsto e_x^*(\omega)$ is a measurable with respect to $\mathcal{B}\otimes \mathcal{F}_\infty$, where $\mathcal{B}$ is the Borel $\sigma$-algebra on $\mathbb{R}$.
We must now verify the independence of the increments. We do this by using Proposition \ref{prop:increments}. Towards this end, fix a set measurable set $\Gamma\subset \mathscr{E}$ and define
\begin{equation}\label{eqn:N_gammaDef}
N_\Gamma^* (a,b] = \sum_{a<y\le b} 1_{[e_y^*\in \Gamma]}.
\end{equation}
We must show that $N_\Gamma^*(x):= N_{\Gamma}^*(0,x]$ is an $\mathcal{F}_{\tau_\cdot }$-Poisson process. However, by Proposition \ref{prop:increments} it is easy to see
\begin{equation*}
\left(N_\Gamma^*(y,y+x]\Big| \mathcal{F}_{\tau_y}\right) \overset{d}{=} N_\Gamma^*(0,x],\qquad \forall x,y>0.
\end{equation*} Indeed, the local time of the process $\tilde{R}(t) := R^*(\hat{\tau}_y+t)$ is simply $-\tilde{I}(t)=-I^*(t)-y$. Let $\tilde{\tau}$ denote the right-continuous inverse of $-\tilde{I}$. So the excursions of $\tilde{R}$ occurring before time $\tilde{\tau}_x$ are simply those of $R$ occurring between times $\hat{\tau}_y$ and $\hat{\tau}_{y+x}$. This proves the desired claim.
\end{proof}
We now define the excursion measure $\mathfrak{n}^*$ on $\mathscr{E}$ by
\begin{equation*}
\mathfrak{n}^*(\Gamma) = \frac{1}{x} \mathbb{E}\left[N_{\Gamma}^*(x)\right]
\end{equation*} where $N_\Gamma^*$ is defined in \eqref{eqn:N_gammaDef}.
\subsection{Basic Properties of $\mathfrak{n}^*$}\label{sec:basicn*}
We now wish to describe some path properties of functions $w\in \mathscr{E}$ under the law $\mathfrak{n}^*$. We begin by describing the analogues of properties in \eqref{eqn:easyPropsforn} for the measure $\mathfrak{n}$.
The supremum is a much easier object to study so we begin with that.
\begin{lem}
For each $\varepsilon>0$, $\mathfrak{n}^*(\sup_{t\le \zeta} w(t)>\varepsilon) = \frac{1}{\varepsilon}$.
\end{lem}
\begin{proof}
For It\^{o}'s measure on positive excursions, we know that $\mathfrak{n}(\sup w(t)>\varepsilon) = \frac{1}{\varepsilon}$.
We have
\begin{equation*}\sup_{t}e_x(t) = \sup_{\tau_{x-} \le t\le \tau_x} (B(t)-I(t)) = \sup_{{\hat{\tau}_{x-}}\le t\le \hat{\tau}_x} (B^*(t)-I^*(t)) = \sup_{t}e_x^*(t).
\end{equation*} Hence
\begin{equation*}
\mathfrak{n}^*(\sup_t w(t)>\varepsilon) = \mathbb{E}\left[ \sum_{0<x\le 1} 1_{[\sup_t e_x^*(t)>\varepsilon]} \right] = \mathbb{E}\left[ \sum_{0<x\le 1} 1_{[\sup_t e_x(t)>\varepsilon]} \right] = \mathfrak{n}(\sup_t w(t)>\varepsilon),
\end{equation*} as desired.
\end{proof}
Determining the distribution of the length of the excursions is a little trickier. To begin let us recall a proof of the fact $\mathfrak{n}(\zeta>\varepsilon) = \sqrt{\frac{2}{\pi \varepsilon}}$. For each $\varepsilon>0$, set
\begin{equation*}
N_\varepsilon(x) = \sum_{0<y\le x} 1_{[\zeta(e_y)>\varepsilon]} = \sum_{0<y\le x} 1_{[\tau_y-\tau_{y-}>\varepsilon]} = \#\{y\in (0,x]: \Delta \tau_{y} >\varepsilon\}.
\end{equation*} We recall that $\tau_x$ is a 1/2-stable subordinator and we can write its Laplace transform \cite[Theorem 6.2.1]{KS.91} as
\begin{equation} \label{eqn:laplace(t)au}
\mathbb{E}[\exp(-\lambda \tau_x)] = \exp(-x\sqrt{2\lambda}) = \exp\left( -x \int\limits_{(0,\infty)} (e^{-\lambda r} -1 )\,\nu(dr)\right),
\end{equation} where $\nu$ is the L\'{e}vy measure of $\tau$. Let $M$ be the Poisson random measure on $[0,\infty)\times (0,\infty)$ with intensity ${\operatorname{Leb}}\otimes \nu$. Then
\begin{equation*}
\left(\tau_x; x\ge 0\right) \overset{d}{=} \left(\int\limits_{y= 0}^x \,\int\limits_{z= 0}^\infty z \, M(dy,dz); x\ge 0 \right).
\end{equation*}
So, in particular,
\begin{equation*}
N_{\varepsilon}(x) = \#\{y\in(0,x]: \Delta\tau_y >\varepsilon\} \overset{d}= M((0,x]\times(\varepsilon,\infty))
\end{equation*} and hence
\begin{equation*}
\mathfrak{n}(\zeta>\varepsilon) = \frac{1}{x}\mathbb{E}[N_\varepsilon(x)] = \frac{1}{x} \mathbb{E}[M((0,x]\times (\varepsilon,\infty))] = \nu(\varepsilon,\infty).
\end{equation*}
To calculate $\nu(\varepsilon,\infty)$, apply integration by parts to the integral in the right-most term of \eqref{eqn:laplace(t)au} to get
\begin{equation}\label{eqn:IBPforLevy}
\sqrt{2\lambda} = \lambda \int_0^\infty e^{-\lambda r}\nu(r,\infty)\,dr.
\end{equation} Said another way, the function $r\mapsto \nu(r,\infty)$ has Laplace transform $\sqrt{2\lambda}/\lambda = \displaystyle \sqrt{\frac{2}{\lambda}}$. Inverting this Laplace transform gives
\begin{equation*}
\nu(\varepsilon,\infty) = \sqrt{\frac{2}{\pi \varepsilon}}.
\end{equation*}
One can alter the approach above to find the value of the measure of $\mathfrak{n}^*(\zeta>\varepsilon)$. Indeed, note that by \eqref{eqn:hatTauandS} we have
\begin{equation*}
\begin{split}
N_\varepsilon^*(x)&:= \sum_{0<y\le x} 1_{[\zeta(e_y^*)>\varepsilon]} = \#\{y\in(0,x]: \Delta \hat{\tau}_y>\varepsilon\}\\
&= \#\{y\in(0,x]: S(\tau_y)-S(\tau_{y-})>\varepsilon \}.
\end{split}
\end{equation*}
We now make the following claim.
\begin{claim}
If $S = (S(t);t\ge 0)$ is a subordinator with Laplace exponent $\phi$ and $\tau = (\tau_x;x\ge 0)$ is as above, then $(S(\tau_x);x\ge 0)$ is a subordinator with Laplace transform
\begin{equation*}
\mathbb{E}[\exp(-\lambda S(\tau_x))] = \exp(-x \sqrt{2\phi(\lambda)}).
\end{equation*}
\end{claim}
The fact that the process $S(\tau_{\cdot})$ is a subordinator without drift follows from the fact that both $S$ and $\tau$ are L\'{e}vy, see \cite[Theorem 30.1]{Sato99} for instance, and that $\tau$ increases solely by jumps. The computation of the Laplace transform of $S(\tau_{\cdot})$ is a simple conditioning argument. In general, one should not expect a simple representation of the L\'{e}vy measure of such subordinators $S(\tau_\cdot)$. See a representation involving Fourier transforms and fractional convolutions in \cite[Theorem 30.1]{Sato99}. However, it is possible to spell it out explicitly in some cases. See Corollary \ref{cor:betaover2} below.
We now state the following lemma.
\begin{lem}\label{lem:LevyMeasureForTCIto}
Let $\nu^*(dr)$ denote the L\'{e}vy measure of $(S(\tau_x);x\ge 0)$. Then, for each $\varepsilon>0$, we have
\begin{equation*}
\mathfrak{n}^*(\zeta>\varepsilon) = \nu^*((\varepsilon,\infty)).
\end{equation*}
\end{lem}
\begin{proof}
The proof of this claim is simply a repetition of the argument that $\mathfrak{n}(\zeta>\varepsilon) = \nu((\varepsilon,\infty))$ with the appropriate replacements of $N_\varepsilon$ with $N_\varepsilon^*$ and the like.
\end{proof}
\subsubsection{Properties of $\mathfrak{n}^*$ when $S$ is $\beta$-stable}
Let us also focus shortly for the case where $S$ is a $\beta$-stable subordinator for some $\beta\in(0,1)$, i.e. its Laplace exponent is $\phi(\lambda) = c \lambda^\beta$. Then $S(\tau_x)$ is a $\frac{\beta}{2}$-stable subordinator with Laplace exponent
\begin{equation*}
\mathbb{E}\left[ \exp(-\lambda S(\tau_x))\right] = \exp(-x \sqrt{2c} \lambda^{\beta/2}).
\end{equation*} In particular its L\'{e}vy measure $\nu^*$ can be calculated as in \eqref{eqn:IBPforLevy} by inverting its Laplace transform:
\begin{equation*}
\sqrt{2c} \lambda^{\beta/2} = \lambda \int_0^\infty e^{-\lambda r} \nu^*((r,\infty))\,dr.
\end{equation*}
Lemma \ref{lem:LevyMeasureForTCIto} implies the following corollary, which is obtained by just inverting Laplace transform above.
\begin{cor} \label{cor:betaover2}
Let $S$ be a $\beta$-stable subordinator with Laplace exponent $\phi(\lambda) = c\lambda^\beta$ for some $\beta\in(0,1)$ and constant $c>0$. Then
\begin{equation*}
\mathfrak{n}^*(\zeta>\varepsilon) = \frac{\sqrt{2c}}{\Gamma \left(1-\frac{\beta}{2}\right)} \varepsilon^{-\beta/2}.
\end{equation*}
\end{cor}
We note that the jumps of the subordinator $S(\tau_x)$ are fairly well studied \cite{PPY.92,PY.92, PY.97} as mentioned in the introduction, with a slightly different formulation. Let $Z\subset \mathbb{R}_+$ denote the closure of the range of the subordinator $S(\tau_x)$. The set $Z$ is of measure zero. For some (possibly random) value $T$, denote by $V_1(T)\ge V_2(T)\ge \dotsm$ the ordered lengths of the intervals of $[0,T]\setminus Z$. The following proposition can be found in \cite{PY.92}.
\begin{prop} \label{prop:PY92}
For any $t>0$:
\begin{equation*}
\left(\frac{V_1(S(\tau(t)))}{S(\tau(t))}, \frac{V_2(S(\tau(t)))}{S(\tau(t))}, \dotsm \right) \overset{d}{=}\left( \frac{V_1(t)}{t},\frac{V_2(t)}{t},\dotsm\right) \overset{d}{=} \operatorname{PD}\left(\frac{\beta}{2}, 0\right),
\end{equation*} where $\operatorname{PD}(\alpha,\theta)$ is the Poisson-Dirichlet distribution with parameters $\alpha$ and $\theta$.
\end{prop}
We now prove Theorem \ref{thm:PD}.
\begin{proof}[Proof of Theorem \ref{thm:PD}]
We will maintain the notation used in this section, and let $R$ denote the reflected Brownian motion and $R^{*}$ its time-change by the inverse of a $\beta$-stable subordinator.
The excursion intervals of the process $R^*$ on the interval $[0,t]$ (including the almost-surely unfinished excursion straddling time $t$) are described by
\begin{equation*}
\left\{ \left(S(\tau_{x-}), S(\tau_x)\right) : S(\tau_x)\le t\right\} \cup \left\{\left(g^*(t),t\right] \right\},
\quad \text{where}\quad g*(t) = \sup\{u\le t: R^*(u) = 0\}.
\end{equation*}
Note that the zero set $\{u: R^*(u) = 0\} = \{S(\tau_{x-}), S(\tau_{x}): x\ge 0\}$, which is precisely the closure of the range of $S(\tau_\cdot)$. Therefore,
\begin{equation*}
\{u\in[0,t]: R^*(u)>0\} = [0,t]\setminus Z,
\end{equation*}
where $Z$ is the closure of the range of the $\beta/2$-stable subordinator $S(\tau_\cdot)$. Proposition \ref{prop:PY92} gives the desired result.
\end{proof}
Explicit formulas for the distribution of $V_1(t)$ the longest excursion interval can be found in \cite{PY.97}, and are mainly located in Section 3.2 therein. While these are rather clunky, the following corollary can be obtained from Corollaries 3 and 12 in \cite{PY.97}.
\begin{cor} \label{cor:mainCor} Let $R^*$ be a reflected Brownian motion time-changed by the inverse of stable subordinator with Laplace transform \eqref{eqn:beta_Stable}.
\begin{enumerate}
\item Fix $t>\varepsilon>0$. Let $J_\varepsilon(t)$ count the number of excursions of $R^*$ that occur before time $t$ (including the last unfinished excursion) that which are length at least $\varepsilon$. Only count that last excursion if $t-g^*(t)>\varepsilon$. Then
\begin{equation*}
\begin{split}
\mathbb{E}\left[\#J_\varepsilon(t)\right]
&= \frac{1}{\Gamma(\beta/2)\Gamma(1-\beta/2)}\int_{\varepsilon/t}^1 \frac{(1-u)^{\frac{\beta}{2}-1}}{u^{\frac{\beta}{2}+1}} \: du\\
&= \frac{1}{\Gamma(\beta/2)\Gamma(1-\beta/2)} \frac{2}{\beta} \left(\frac{t}{\varepsilon} - 1\right)^{\beta/2}.
\end{split}
\end{equation*}
\item Let $V_1(t)$ be the length of longest (possibly unfinished) excursion interval of $R^*$ on $[0,t]$ then
\begin{equation*}
\mathbb{E}\left[ \exp\left( - \lambda t /V_1(t)\right)\right] = \frac{e^{-\lambda} }{\displaystyle \Gamma(1-\beta/2) \lambda^{\beta/2} + \frac{\beta}{2} \int_1^\infty e^{-\lambda r} r^{-1-\beta/2}\,dr}.
\end{equation*}
\end{enumerate}
\end{cor}
\subsection{Another representation of $\mathfrak{n}^*$}
In this section we present a different construction of the measure $\mathfrak{n}^*$ which is useful for coupling the excursions of a Brownian motion $B$ and its time-change $B^*$.
Let $\mathscr{M}$ denote the space of Radon measures on $[0,\infty)\times (0,\infty)$ and let $\mathscr{M}_c$ denote the subset of measures $\mu$ on $[0,\infty)\times (0,\infty)$ such that
\begin{equation*}
\mu(A) = \sum_{j=1}^\infty 1_{[(t_j,x_j)\in A]},
\end{equation*}
where $t_j$ are distinct elements in $(0,\infty)$ and $x_j\in (0,\infty)$.
We will now consider the functions $\Xi:\mathscr{E}\times \mathscr{M}_c\times [0,\infty) \to \mathbb{D}(\mathbb{R}_+,\mathbb{R})$ by
\begin{equation*}
\Xi(w,\mu, \kappa)(t) = w\left( \inf\left\{u: \kappa u+ \int_0^\infty y \mu([0,u]\times dy) >t \right\}\right).
\end{equation*} We implicitly assume that either $u\mapsto \int_0^\infty y\mu([0,u],dy)$ is strictly increasing or $\kappa>0$ so that $\Xi(w,\mu,\kappa)\in \mathscr{E}$.
We now prove
\begin{thm} \label{thm:excursionChange}
Suppose that the strictly increasing subordinator $S$ has L\'{e}vy-It\^{o} decomposition \eqref{eqn:PRM_subordinator} where $M$ has intensity measure ${\operatorname{Leb}}\otimes \nu$. Then
\begin{equation*}
\mathfrak{n}^* = \Xi(\cdot,\cdot, \kappa)_\# (\mathfrak{n}\otimes \mathcal{L}(M))
\end{equation*}
where $\mathcal{L}(M)$ is the law of the random measure $M$.
\end{thm}
\begin{proof}
We first observe that $\mathfrak{n}^*(\sup_t w(t)>\varepsilon) = \mathfrak{n}(\sup_t w(t)) = \frac{1}{\varepsilon}$ and that the map $\Xi$ does not change the supremum of the of the process. Hence, it suffices to show for each $\varepsilon>0$
\begin{equation*}
\mathfrak{n}^*\left(\ \bullet \ \bigg| \sup_t w(t)>\varepsilon\right) = \Xi(\cdot, \cdot, \kappa)_{\#} \left(\mathfrak{n}\left(\ \bullet \ \bigg| \sup_t w(t)>\varepsilon\right)\otimes \mathcal{L}(M)\right).
\end{equation*}
To prove this we let $e^*(t)$ be the first excursion of $R^*$ which exceeds level $\varepsilon$, and $e(t)$ be the first excursion of $R$ which exceeds level $\varepsilon$. Formally, set $T^* = \inf\{t: R^*(t) > \varepsilon\}$, $g^* = \sup\{r<T^*: R^*(r) = 0\}$ and $d^* = \inf\{r>T^*: R^*(r) = 0\}$ then
\begin{equation*}
e^*(t) = R^*((g^*+t)\wedge d^*).
\end{equation*} Similarly define $T$, $g$ an $d$ in the construction of $e$. Observe that $g^* = S(g)$ by \eqref{eqn:hatTauandS}. By \cite[Proposition XII.1.13]{RY.99} the first hitting time of a set $A$ of a Poisson point process with intensity $\mu$ has distribution $\mu(-|A)$, provided that $\mu(A)<\infty$. Hence both $e^*\sim \mathfrak{n}^*(-|\sup_t w(t)>\varepsilon)$ and $e\sim \mathfrak{n}(-|\sup_t w(t)>\varepsilon)$. Therefore, it suffices to prove
\begin{equation*}
e^* \overset{d}{=} \Xi(e,\tilde{M},\kappa)
\end{equation*} where $\tilde{M}\overset{d}{=} M$ is independent of $e$.
Define $\tilde{M}([s,t]\times (a,b)) = M([g + s, g+t]\times (a,b))$. Note that $g$ is independent of $M$ ($g$ depends only on the reflected Brownian motion $R$), and so by conditioning on $g$ it is easy to see that $\tilde{M}$ is a Poisson random measure on $[0,\infty)\times(0,\infty)$ with intensity ${\operatorname{Leb}}\otimes \nu$.
Next, for $t\le d^*-g^*$, and as in the proof of Proposition \ref{prop:increments},
\begin{equation*}
e^*(t) = R( E( S(g)+ t)) = R(E(S(g)) + \tilde{E}(t))= R (g+\tilde{E}(t)) = e(\tilde{E}(t))
\end{equation*} where
$\tilde{E}(t) = \inf\{u: S(u+g)-S(g)>t\}.
$ Since $g$ is independent of $S$ and $S$ has stationary increments, $\tilde{E}\overset{d}{=} E$.
But, by the L\'{e}vy-It\^{o} decomposition of $S$, we have
\begin{equation*}
S(u+g) - S(g) = \kappa u + \int_0^\infty y \,M([g,g+u],dy) = \kappa u + \int_0^\infty y\tilde{M}([0,u],dy).
\end{equation*} Hence
\begin{equation*}
e^*(t) = e(\tilde{E}(t)) = \Xi\left( e, \tilde{M}, \kappa\right)
\end{equation*}
proving the desired claim.
\end{proof}
\subsection{Excursion Construction of Time-changed Brownian motion}
We now recall that from the excursion process $e_x^*$ we can reconstruct the time-changed reflected Brownian motion $R^*$ by
\begin{equation*}
\begin{split}
\hat{\tau}_x = \sum_{s\le x} \zeta(e_s^*),\qquad
-I^*(t) = \inf\{x: \hat{\tau}_x>t\},\qquad
R^*(t) = \sum_{s\le -I^*(t)} e_s(t-\hat{\tau}_{s-})
\end{split}.
\end{equation*}
The proof of this result is exactly the same as the Brownian case, which can be found in \cite[Proposition XII.V.2]{RY.99}.
We also recall that we can construct a standard Brownian motion from the It\^{o} measure $\mathfrak{n}$ of \textit{positive} excursions as follows. Let $(e_x,\gamma_x)$ be a Poisson point process taking values in $\mathscr{E} \times \{\pm 1\}$ with intensity
\begin{equation*}
\mathfrak{n}\otimes \left( \frac{1}{2} \delta_{-1} + \frac{1}{2} \delta_{+1}\right).
\end{equation*} Then the process $B$ constructed by
\begin{equation}\label{eqn:excursionBM}
\begin{split}
\tau_x = \sum_{s\le x} \zeta(e_s),\qquad
L(t) = \inf\{x: \tau_x>t\},\qquad
B(t) &= \sum_{s\le L(t)} \gamma_s e_s(t-\tau_{s-})
\end{split}
\end{equation} is a standard Brownian motion.
But now let $S$ be a strictly increasing subordinator with L\'{e}vy-It\^{o} decomposition \eqref{eqn:PRM_subordinator}. Suppose this is independent of the Poisson point process $(e_x,\gamma_x;x>0)$ and let $E$ be its inverse defined in \eqref{eqn:Edef}. Note that, with probability 1,
\begin{equation*}
M\left(\{\tau_{x-},\tau_x:x\ge 0\} \times (0,\infty)\right) = 0,
\end{equation*} since $\{\tau_{x-},\tau_x: x\ge 0\}$ is the zero set of the Brownian motion $B$ and has Lebesgue measure zero.
We now work conditionally on $(\tau_x;x\ge 0)$. Now define $\hat{\tau}_x = S(\tau_x)$. Then, for $\tau_{x-}\le E(t)\le \tau_x$, the value of
\begin{equation*}
\begin{split}
E(t) &= \inf\{u: S(u)>t\} = \inf\{u+ \tau_{x-}: S(u+\tau_{x-}) - S(\tau_{x-}) > t-S(\tau_{x-})\} \\
&= \tau_{x-}+ \inf\{u: S_{x}(u) > t-\hat{\tau}_{x-}\},
\end{split}
\end{equation*}
where $S_x(u) = S(u+ \tau_{x-})-S(\tau_{x-}) = \kappa u + \int_0^\infty y M([\tau_{x-},u],dy)$. Define, for $a\le b\le \tau_x-\tau_{x-}$
\begin{equation*}
M_x([a,b]\times A) = M([\tau_{x-}+a, \tau_{x-}+b]\times A).
\end{equation*} Conditionally given $\tau$, the measure $M_x$ is a Poisson random measure on $[0,\tau_{x}-\tau_{x-}]\times (0,\infty)$ with intensity ${\operatorname{Leb}}\otimes \nu$. Moreover, the measures $\{M_x; x\ge 0\}$ are conditionally independent given $(\tau_x)_{x\ge 0}$ because $S$ has independent increments.
But now by Theorem \ref{thm:excursionChange}, $(e_x^*, \gamma_x; x>0)$ where $e_x^*$ is defined by
$e_x^* = \Xi\left(e_x,M_x,\kappa\right)$
is a Poisson point process on $\mathscr{E}\times \{\pm1\}$ with intensity measure $\mathfrak{n}^*\otimes\left(\frac{1}{2} \delta_{-1} + \frac{1}{2} \delta_{1} \right)$. Moreover it is not hard to see $e_x^*(t-\hat{\tau}_{x-}) = e_x(E(t) - \tau_{x-}).$
Indeed, since $\hat{\tau}_{x-} = S(\tau_{x-})$
\begin{equation*}
\begin{split}
\inf&\left\{u: \kappa u + \int_0^\infty y M_{x}([0,u],dy) > t - \hat{\tau}_{x-}\right\}= \inf\left\{u: \kappa (u+\tau_{x-}) + \int_0^\infty y M([0,\tau_{x-}+u],dy)>t \right\}\\
&= \inf\left\{u: \kappa u + \int_0^\infty y M([0,u],dy) >t \right\} - \tau_{x-}.
\end{split}
\end{equation*}
Also $\zeta(e_x^*) = S_x (\zeta(e_x))$.
Hence
\begin{equation*}
B(E(t)) = \sum_{s\le L(E(t))} \gamma_s e_s \left(E(t) - \tau_{s-} \right) = \sum_{s\le L(E(t))} \gamma_s e_s^*(t-\hat{\tau}_{x-}),
\end{equation*} and
\begin{equation*}
L(E(t)) = \inf\{x: \tau_x> E(t)\} = \inf\{x: \hat{\tau}_{x}>t\},
\end{equation*} where we used the fact that, with probability 1, $S$ does not jump at any of the times $\{\tau_{x-},\tau_x; x\ge 0\}$.
Hence we have proved the following theorem.
\begin{thm} \label{thm:pppfortc}
Let $(e_x^*,\gamma_x)$ be a Poisson point process on $\mathscr{E}\times \{\pm1\}$ with intensity $\mathfrak{n}^*\otimes (\frac{1}{2} \delta_{-1}+\frac{1}{2} \delta_1)$. Then $B^*$ defined by
\begin{equation} \label{eqn:excurionABM}
\hat{\tau}_x = \sum_{s\le x} \zeta(e_x^*),\qquad L^*(t) = \inf\{x: \hat{\tau}_x>t\},\qquad B^*(t) = \sum_{s\le L^*(t)} \gamma_s e_s^*(t-\hat{\tau}_{s-})
\end{equation} is a Brownian motion time-changed by the inverse of an independent strictly increasing subordinator $S$. That is $B^* \overset{d}{=} (B(E(t));t\ge 0)$ for a Brownian motion $B$ and an independent inverse subordinator $E$.
\end{thm}
We observe that $E(t) = \langle B^* \rangle(t)$ is the quadratic covariation of $B^*$. Hence Theorem \ref{thm:pppfortc} allows us to construct the time-change $E$ and, subsequently, the process $S$. In fact, we can use Theorems \ref{thm:excursionChange} and \ref{thm:pppfortc} to obtain the following corollary:
\begin{cor} \label{cor:EFromExcursion} Let $S$ be a strictly increasing subordinator with L\'{e}vy-It\^{o} decomposition \eqref{eqn:PRM_subordinator}.
Let $(e_x, M_x, \gamma_x)_{x>0}$ be Poisson point process on $\mathscr{E}\times \mathscr{M}_c \times \{-1,1\}$ with intensity measure $\mathfrak{n}\otimes \mathcal{L}(M)\otimes (\frac{1}{2}\delta_{-1}+ \frac{1}{2}\delta_1)$. If $e_x^* = \Xi(e_x,M_x,\kappa)$ then $B^*$ defined in \eqref{eqn:excurionABM} is the time-change of the Brownian motion defined in \eqref{eqn:excursionBM} with time-change $E$ satisfying
\begin{equation*}
E(t) = \sum_{s\le L^*(t) } \langle e_x^*\rangle (t-\hat{\tau}_{s-}).
\end{equation*}
\end{cor}
\subsection{Some marginal distributions for $\mathfrak{n}$ and $\mathfrak{n}^*$}
We now discuss an important relationship between $\mathfrak{n}$ and $\mathfrak{n}^*$ which will be useful in the sequel. Before covering this relationship we remark that both $\mathfrak{n}$ and $\mathfrak{n}^*$ are $\sigma$-finite measures; however, we can talk about processes have $\mathfrak{n}$ (or $\mathfrak{n}^*$) as its ``law'' by restricting to sets of finite $\mathfrak{n}$ (resp. $\mathfrak{n}^*$) mass.
We describe the ``law'' of $(\zeta(w^*), \langle w^*\rangle (\zeta(w^*)))$ under $\mathfrak{n}^*(dw^*)$, where $\langle w^*\rangle$ is the quadratic covariation of the path $w$ which is defined $\mathfrak{n}^*$-a.e.
Suppose that $S$ has L\'{e}vy-It\^{o} decomposition \eqref{eqn:PRM_subordinator} and is independent of the process $w$ whose ``law'' is $\mathfrak{n}$. By Theorem \ref{thm:excursionChange} the ``law'' of $w^*:= \Xi(w,M,\kappa)$ is $\mathfrak{n}^*$.
However, observe that
\begin{align}
\zeta(w^*) &= \inf\{t>0: w^*(t) = 0\}= \inf\{t>0: \Xi(w,M,\kappa)(t) = 0\}\nonumber\\
&=\inf\{t : \inf\{u:\kappa u + \int_0^\infty y M([0,u],dy) > t\} = \zeta(w)\}\nonumber\\
&= \kappa \zeta(w) + \int_0^{\infty} y M([0,\zeta(w)],dy) = S(\zeta(w))\label{eqn:zetawStar}
\end{align} and
\begin{equation*}
\langle w^*\rangle (\zeta(w^*)) = \langle w\rangle(\zeta(w)) = \zeta(w).
\end{equation*}
Thus,
\begin{align*}
\left((\zeta(w^*), \langle w^*\rangle (\zeta(w^*)))\text{ under } \mathfrak{n}^* (dw^*) \right) = \left( \left(S(\zeta(w)), \zeta(w)\right) \text{ under } \mathcal{L}(S)\otimes \mathfrak{n}(dw) \right).
\end{align*}
Since the law of $\zeta(w)$ under $\mathfrak{n}(dw)$ is know, the above representation can be used to give us an explicit description of the joint law of $(\zeta(w^*), \langle w^*\rangle (\zeta(w^*)))$
\begin{comment}
In particular, for a positive functional $F$ we have by \eqref{eqn:easyPropsforn}
\begin{align*}
\int_{\mathscr{E}} &F(\zeta(w^*), \langle w^* \rangle (\zeta(w^*))) \,\mathfrak{n}^*(dw^*) \\
&= \int_{\mathscr{E}} \mathbb{E}\left[ F(S(\zeta(w)),\zeta(w)) | w \right] \mathfrak{n}(dw)\\
&= \int_0^\infty \mathbb{E}\left[F(S(t), t) \right] \mathfrak{n}(\zeta\in dt)\\
&= \int_0^\infty \mathbb{E}[F(S(t),t)] \frac{1}{\sqrt{2\pi t^3}} \,dt.
\end{align*}
Specifically, writing $\phi$ for the Laplace exponent of $S$ and setting $F(y,t) = e^{-\lambda y} 1_{[t>\varepsilon]}$ we get
\begin{equation*}
\begin{split}
\int_{\mathscr{E}} \exp\left(-\lambda \zeta(w^*)\right) 1_{[\zeta(w^*)>\varepsilon]}\mathfrak{n}^*(dw^*) &= \frac{1}{\sqrt{2\pi}}\int_\varepsilon^\infty e^{-t \phi(\lambda) } t^{-3/2}\,dt \\
&= \sqrt{\frac{\phi(\lambda)}{2\pi}} \int_{\phi(\lambda) \varepsilon}^\infty e^{-t} t^{-3/2}\,dt\\
&= \sqrt{\frac{\phi(\lambda)}{2\pi}} \left(\frac{2 e^{-\phi(\lambda )\varepsilon}}{\sqrt{\phi(\lambda)\varepsilon}} - 4\sqrt{\pi} \mathcal{N}(-\sqrt{2\phi(\lambda)\varepsilon}) \right),
\end{split}
\end{equation*} where $\mathcal{N}$ is the cumulative distribution of a standard normal random variable.
\end{comment}
\section{The first negative excursion of duration $D$}\label{sec:firstExcursion}
Let $(g^*,d^*)$ be the interval corresponding to the first negative excursion of $B^*$ which lasts longer than $D$. In order to price a down-and-in Parisian option, we will need to understand more than just times $g^*$ and $d^*$. In fact, we must understand the joint distribution of $(E(g^*),g^*)$ and the joint distribution of $(E(d^*),d^*)$ where $E$ is the time-change as in \eqref{eqn:Edef}.
Towards this end we will use Theorem \ref{thm:pppfortc} and Corollary \ref{cor:EFromExcursion}. Let $(e_x,M_x,\gamma_x)$ be the Poisson point process defined therein with respect to the filtration $(\mathcal{G}_x)_{x>0}$, say, and let $e_x^* = \Xi(e_x,M_x,\kappa)$. Now, the first negative excursion of $B^*$ lasting longer than duration $D$ corresponds to the interval $(g^*,d^*)$ defined by
\begin{equation}\label{eqn:gstarDstar_def}
g^* = \inf\{\hat{\tau}_{x-}: \gamma_x{e}^*_x(D)<0\}\quad \text{and} \quad d^* = \inf\{\hat{\tau}_{x}:\gamma_x {e}^*_x(D)<0\}.
\end{equation} Since $\gamma_x e_x^*(D)<0$ if and only if $\gamma_x = -1$ and $\zeta(e_x^*)>D$, we conclude by \eqref{eqn:zetawStar} that
\begin{equation*}
\{x: \gamma_x e_x^*(D)<0\} = \left\{x: \gamma_x = -1, \kappa \zeta(e_x) + \int_0^\infty y M_x([0,\zeta(e_x)],dy) >D \right\}.
\end{equation*}
We now define the following subordinators
\begin{equation*}
\begin{split}
\hat{\tau}_x &= \sum_{y\le x} \zeta(e_y^*)= \sum_{y\le x} \left\{\kappa \zeta(e_y)+ \int_0^\infty z M_y([0,\zeta(e_y)],dz) \right\} ,\\
{\hat{\tau}^\uparrow}_x &= \sum_{y\le x} 1_{[\gamma_y = 1]} \zeta(e_y^*)=\sum_{y\le x} 1_{[\gamma_y = 1]} \left\{\kappa \zeta(e_y)+ \int_0^\infty z M_y([0,\zeta(e_y)],dz) \right\} ,\\
{\hat{\tau}^\downarrow}_x &=\sum_{y\le x} 1_{[\gamma_y = -1]} \zeta(e_y^*)= \sum_{y\le x}1_{[\gamma_y = -1]} \left\{\kappa \zeta(e_y)+ \int_0^\infty z M_y([0,\zeta(e_y)],dz) \right\},\\
\tau_x &= \sum_{y\le x} \zeta(e_y),\\
\tau^{\uparrow}_x &= \sum_{y\le x} 1_{[\gamma_y = 1]} \zeta(e_y),\\
\tau^{\downarrow}_x &= \sum_{y\le x} 1_{[\gamma_y = -1]} \zeta(e_y).
\end{split}
\end{equation*} In words, the subordinators with $\uparrow$ as a superscript represent positive excursions and the subordinators with $\downarrow$ represent negative excursions of either $B^*$ or $B$.
Obviously $\tau_x = \tau^{\uparrow}_x + \tau^{\downarrow}_x$ and $\hat{\tau}_x = {\hat{\tau}^\uparrow}_x+{\hat{\tau}^\downarrow}_x$. Since $\tau^{\downarrow}_x$ and $\tau^{\uparrow}_x$ are constructed from the atoms of a PRM on disjoint sets, the processes $\tau^{\downarrow}$ and $\tau^{\uparrow}$ are independent and similar reasoning implies the independence of ${\hat{\tau}^\downarrow}$ and ${\hat{\tau}^\uparrow}$. Moreover, since $P(\gamma_x = -1) = P(\gamma_x = 1) = 1/2$, we have $(\tau^{\uparrow},{\hat{\tau}^\uparrow})\overset{d}{=} (\tau^{\downarrow},{\hat{\tau}^\downarrow})$.
To keep track of the negative excursions depending on their durations, we write
\begin{equation*}
\begin{split}
{\hat{\tau}^{\downarrow,-}}_x &:= \sum_{y\le x} 1_{[\gamma_y = -1, \zeta(e^*_y)\le D]} \zeta(e_y^*), \qquad
{\hat{\tau}^{\downarrow,+}}_x:= \sum_{y\le x} 1_{[\gamma_y = -1, \zeta(e^*_y)>D]} \zeta(e_y^*),\\
\tau^{\downarrow,-}_x &:= \sum_{y\le x} 1_{[\gamma_y = -1, \zeta(e^*_y)\le D]} \zeta(e_y) ,\qquad
\tau^{\downarrow,+}_x:= \sum_{y\le x} 1_{[\gamma_y = -1, \zeta(e^*_y)>D]} \zeta(e_y).
\end{split}
\end{equation*} We remark that in the definitions of $\tau^{\downarrow,-}$ and $\tau^{\downarrow,+}$ there is not a typo when considering the durations $\zeta(e_y^*)$ in the indicators: we want to make sure that ${\hat{\tau}^{\downarrow,-}}$ and $\tau^{\downarrow,-}$ jump at the same times. Here the ``$-$'' superscript represent excursions $e^*_y$ of $B^*$ lasting at most duration $D$ and the ``$+$'' superscript represents the excursions lasting longer than duration $D$. Again, since the subordinators are constructed from the behavior of a PRMs on disjoint sets, $({\hat{\tau}^{\downarrow,+}},\tau^{\downarrow,+})$, $({\hat{\tau}^{\downarrow,-}},\tau^{\downarrow,-})$, and $({\hat{\tau}^\uparrow},\tau^{\uparrow})$ are three independent $\mathbb{R}^2$-valued L\'{e}vy processes.
We can compute the joint Laplace transforms of $(\hat{\tau}_x,\tau_x)$ as follows:
\begin{equation}\label{eqn:ht_t}
\begin{split}
\mathbb{E}&\left[ \exp\left\{-\lambda \hat{\tau}_x - \theta \tau_x\right\} \right] = \mathbb{E}\left[\exp\left\{- \left(\sum_{y\le x} (\kappa \lambda + \theta) \zeta(e_y) + \int_0^\infty \lambda z M_y([0,\zeta(e_y)], dz)\right)\right\} \right]\\
&= \exp\left\{ - x \int_{\mathscr{E}\times \mathscr{M}_c} \left(1-\exp\{-(\kappa\lambda+\theta) \zeta(w) - \int_0^\infty \lambda z \mu([0,\zeta(w)],dz)\}\right) \,\mathfrak{n}(dw)\mathbb{P}(M\in d\mu)\right\}\\
&= \exp\left\{-x \int_\mathscr{E} \left(1-\exp\left\{ -\zeta(w)(\kappa\lambda +\theta + \int_0^\infty (1-e^{-\lambda z})\,\nu(dz)) \right\} \right) \,\mathfrak{n}(dw) \right\}\\
&= \exp\left\{-x \int_\mathscr{E} \left(1-e^{-\zeta(w)(\phi(\lambda)+\theta)}\right)\,\mathfrak{n}(dw)\right\}\\
&= \exp\left\{ -x \int_0^\infty (1-e^{-t (\phi(\lambda)+\theta)}) \,\frac{1}{\sqrt{2\pi t^3}}\,dt \right\}\\
&= \exp\left\{-x \sqrt{\frac{2}{\pi}} (\phi(\lambda)+\theta)\int_0^\infty e^{-(\phi(\lambda)+\theta) t} t^{-1/2}\,dt \right\}\\
&= \exp\left\{-x \sqrt{2 (\phi(\lambda)+\theta)} \right\}.
\end{split}
\end{equation} In the second and third equalities above we used the exponential formulas for PRMs \cite[Section XII.1]{RY.99}. In the fifth equality we used the \eqref{eqn:easyPropsforn}.
Observe that the L\'{e}vy measure of the pair $(\hat{\tau},\tau)$ is the measure $\pi(ds,dt)$ which is simply
\begin{equation*}
\pi(A\times B) = \int_B \mathbb{P}(S(t) \in A) \frac{1}{\sqrt{2\pi t^3}} \,dt.
\end{equation*}
Indeed, we can instead integrate with respect to $\mathfrak{n}(dw)$ first and go from the third item in the string of equalities above to
\begin{equation*}
\begin{split}
\exp&\left\{ -x \mathbb{E} \left[\int_0^\infty \left(1-\exp\left\{-(\kappa \lambda + \theta) t + \int_0^\infty \lambda z M([0,t],dz) \right\} \right) \frac{1}{\sqrt{2\pi t^3}}\,dt \right]\right\}\\
&= \exp\left\{ -x \mathbb{E}\left[ \int_0^\infty (1-e^{-\lambda S(t) - \theta t}) \frac{1}{\sqrt{2\pi t^3}}\,dt \right]\right\}\\
&= \exp\left\{ -x \int_{(0,\infty)^2} (1-e^{-\lambda s -\theta t})\,\pi(ds,dt)\right\}.
\end{split}
\end{equation*}
Similarly, we see that
\begin{equation} \label{eqn:htuandtu}
-\frac{1}{x}\log\mathbb{E}\left[\exp\left\{-\lambda {\hat{\tau}^\uparrow}_x - \theta \tau^{\uparrow}_x\right\} \right] = \sqrt{\frac{1}{2} (\phi(\lambda)+\theta)} = \int_{(0,\infty)^2} (1-e^{-\lambda s -\theta t}) \frac{1}{2}\pi (ds,dt).
\end{equation}
It is also easy to verify that the L\'{e}vy measure of the pure-jump L\'{e}vy process $({\hat{\tau}^{\downarrow,+}},\tau^{\downarrow,+})$ is $\frac{1}{2}\pi^+$ with $\pi^+(A\times B) := \pi( (A\cap (D,\infty))\times B)$ and the L\'{e}vy measure of $({\hat{\tau}^{\downarrow,-}},\tau^{\downarrow,-})$ is $\frac{1}{2}\pi^-$ where $\pi^-(A\times B):= \pi( (A\cap (0,D])\times B)$. Indeed, the jumps of $({\hat{\tau}^{\downarrow,+}},\tau^{\downarrow,+})$ are simply the jumps of $({\hat{\tau}^\downarrow},\tau^{\downarrow})$ for which ${\hat{\tau}^\downarrow}_x - {\hat{\tau}^\downarrow}_{x-}>D$ and this is precisely what the L\'{e}vy measure $\frac{1}{2}\pi^+$ captures.
Define the time $J = \inf\{x: \tau^{\downarrow,+}_x>0\}$ as the first jump of the subordinator $\tau^{\downarrow,+}$. Since $\tau^{\downarrow,+}$ and ${\hat{\tau}^{\downarrow,+}}$ have the same jump times, $J$ is also the first jump time of ${\hat{\tau}^{\downarrow,+}}$. The time $J$ is the index $x$ of the first excursion $(e^*_x,\gamma_x)$ such that $\gamma_x = -1$ and $\zeta(e_x^*) >D$. In particular, it is easy to see that $g^* = \hat{\tau}_{J-}$ and $d^* = \hat{\tau}_J$.
We now state the following lemma for ease of reference in the sequel.
\begin{lem}\label{lem:independence}
With the above set-up the following hold:
\begin{enumerate}
\item $J$ is a $\mathcal{G}_x$-stopping time;
\item $((e_{x+J},M_{x+J}, \gamma_{x+J}); x>0)$ is independent of $\mathcal{G}_J$ and is equal in law to $((e_x,M_x,\gamma_x);x>0)$;
\item Both $d^*-g^* = {\hat{\tau}^{\downarrow,+}}_J$ and $\tau^{\downarrow,+}_J$ are independent of $J$;
\item $J$ is independent of $\tau^{\uparrow}$, $\tau^{\downarrow,-}$, ${\hat{\tau}^\uparrow}$ and ${\hat{\tau}^{\downarrow,-}}$;
\item $E(g^*) = \tau_{J-} = \tau^{\uparrow}_J+\tau^{\downarrow,-}_J$ and $g^* = \hat{\tau}_{J-} = {\hat{\tau}^\uparrow}_J + {\hat{\tau}^{\downarrow,-}}_J$ are each sums of two independent random variables;
\item $E(d^*) = \tau_J = \tau^{\uparrow}_J + \tau^{\downarrow,-}_J + \tau^{\downarrow,+}_J$ and $d^* = \hat{\tau}_J = {\hat{\tau}^\uparrow}_J + {\hat{\tau}^{\downarrow,-}}_J + {\hat{\tau}^{\downarrow,+}}_{J}$ are each the sum of three independent random variables.
\end{enumerate}
\end{lem}
\begin{proof}
For item (1), $J$ is the first time of the $\mathcal{G}_x$-progressive process ${\hat{\tau}^{\downarrow,+}}_x$ enters the open set $(D,\infty)$ and is hence a stopping time.
Item (2) is the strong Markov property for Poisson point processes.
Item (3) follows from Lemma XII.1.13 in \cite{RY.99}.
Item (4) follows from the fact that each of the subordinators listed are independent of $({\hat{\tau}^{\downarrow,+}},\tau^{\downarrow,+})$.
Items (5) and (6) follow from the definitions and \eqref{eqn:hatTauandS}. In item (5) we note that $\tau_{J-} = \tau^{\uparrow}_{J-} + \tau^{\downarrow,-}_{J-} = \tau^{\uparrow}_J + \tau^{\downarrow,-}_J$ because $\tau^{\uparrow}$ and $\tau^{\downarrow,-}$ are independent of $\tau^{\downarrow,+}$ and independent subordinators almost surely do not jump at the same time.
\end{proof}
\subsection{Distribution of $(E(g^*),g^*)$}
\begin{comment}
We begin with a discussion of the distribution of $J$, where we write the explicit formula in the case where $S$ is $\beta$-stable.
\begin{lem} Suppose that $S$ is a subordinator with Laplace exponent $\phi(\lambda) = c\lambda^\beta$, or equivalently $\kappa = 0$ and $M$ is a Poisson random measure on $\mathbb{R}_+\times(0,\infty)$ with intensity ${\operatorname{Leb}} \otimes \nu$ where $\nu((x,\infty)) = \frac{c}{\Gamma(1-\beta)} x^{-\beta}$ for all $x>0$.
Then $J$ is an exponential random:
\begin{equation*}
J\sim \operatorname{Exp}(\mu) ,\qquad \mu:= \mu(c,\beta,D) = \sqrt{\frac{c}{2}} \frac{1}{\Gamma(1-\frac{\beta}{2})} D^{-\beta/2}
\end{equation*}
More generally, even when $S$ is not a stable subordinator $J$ is still an exponential random variable.
\end{lem}
\begin{proof}
By equation \eqref{eqn:htuandtu}, and the fact that $\phi(\lambda) = c\lambda^{\beta}$ we know that ${\hat{\tau}^\downarrow}$ is a $\frac{\beta}{2}$-stable subordinator with Laplace exponent $\lambda\mapsto \sqrt{\frac{c}{2}} \lambda^{\beta/2}$. Thus, by the L\'{e}vy-It\^{o} decomposition of ${\hat{\tau}^\downarrow}$ we can write
\begin{equation*}
{\hat{\tau}^\downarrow}_x = \int_0^\infty y \tilde{M}([0,x],dy),
\end{equation*}
where $\tilde{M}$ is a Poisson random measure on $\mathbb{R}_+\times(0,\infty)$ with intensity ${\operatorname{Leb}}\otimes \nu^\downarrow$ where $\nu^\downarrow(dx) = \frac{\beta \sqrt{c}}{2\sqrt{2}} \frac{1}{\Gamma(1-\frac{\beta}{2})} x^{-\frac{\beta}{2}-1}\,dx$.
Since $J$ is the first time that ${\hat{\tau}^\downarrow}$ has a jump larger than $D$, we can write
\begin{equation*}
\begin{split}
\mathbb{P}( J> x) &= \mathbb{P}({\hat{\tau}^\downarrow} \text{ does not have a jump }>D\text{ before time }x)\\
&= \mathbb{P}(\tilde{M}([0,x],(D,\infty)) = 0) = e^{-x\nu^\downarrow(D,\infty)},
\end{split}
\end{equation*}
since $ \nu^\downarrow(D,\infty) = \sqrt{\frac{c}{2}} \frac{1}{\Gamma(1-\frac{\beta}{2})} D^{-\beta/2}$ is finite, and the number of atoms that a Poisson random measure has on a set of finite mass is simply a Poisson random variable. This proves the first claim.
The second, more general, fact follows from the same reasoning since ${\hat{\tau}^\downarrow}$ increases exclusively by jumps.
\end{proof}
\end{comment}
It is easy to see that $\mathbb{P}(J>x) = \mathbb{P}({\hat{\tau}^{\downarrow,+}}_x = 0) = e^{-x \frac{1}{2} \pi^+(\mathbb{R}_+\times\mathbb{R}_+)}$. That is $J$ is an exponential random variable, which we state as the following Lemma:
\begin{lem}\label{lem:JisExp}
The variable $J$ is exponentially distributed with parameter $\mu = \frac{1}{2}\pi^+(\mathbb{R}_+\times \mathbb{R}_+)$.
\end{lem}
We now proceed to compute the joint Laplace transform of $(E(g^*),g^*)$ by utilizing Lemmas \ref{lem:independence} and \ref{lem:JisExp}. Indeed, by equation \eqref{eqn:htuandtu} and the description of the L\'{e}vy measure of $({\hat{\tau}^{\downarrow,-}},\tau^{\downarrow,-})$ and letting $\mu$ denote the parameter of the exponential random variable $J$, we get for each $\lambda,\theta\ge 0$:
\begin{align}
\mathbb{E}&\left[\exp\{- \lambda g^* - \theta E(g^*) \} \right] = \mathbb{E}\left[\exp\left\{-\lambda ({\hat{\tau}^\uparrow}_J+{\hat{\tau}^{\downarrow,-}}_J) - \theta (\tau^{\uparrow}_J+ \tau^{\downarrow,-}_J) \right\} \right] \nonumber \\
&= \int_0^\infty {\mu} e^{-\mu x} \mathbb{E}\left[\exp\left\{ -\lambda {\hat{\tau}^\uparrow}_x -\theta \tau^{\uparrow}_x\right\} \right] \mathbb{E}\left[ \exp\left\{-\lambda {\hat{\tau}^{\downarrow,-}}_x-\theta \tau^{\downarrow,-}_x\right\} \right]\,dx \nonumber\\
&= \mu \int_0^\infty e^{-\mu x} e^{-x\sqrt{\frac{1}{2} (\phi(\lambda) + \theta)}} \exp\left\{-x \frac{1}{2} \int_{(0,\infty)^2}(1- e^{-\lambda s - \theta t}) \pi^-(ds,dt) \right\}\,dx \nonumber \\
&= \frac{\displaystyle \mu}{ \displaystyle \mu + \sqrt{\frac{1}{2} \phi(\lambda)+\frac{1}{2}\theta} + \frac{1}{2} \int_{(0,\infty)^2} (1-e^{-\lambda s-\theta t})\,\pi^{-}(ds,dt)}. \label{eqn:jointLT_for_E_gStar}
\end{align}
This is precisely the Laplace transform of the measure $m_1(ds,dv)$ in Theorem \ref{thm:priceTheorem}.
\subsection{Joint distribution of $(E(d^*),d^*)$}
Before commencing a similar computation to that in \eqref{eqn:jointLT_for_E_gStar}, we observe, by Proposition XII.1.13 in \cite{RY.99} the joint distribution is
\begin{equation*}
\mathbb{P}({\hat{\tau}^{\downarrow,+}}_J\in A, \tau^{\downarrow,+}_J\in B) = \frac{\frac{1}{2}\pi^+(A\times B)}{\frac{1}{2}\pi^+(\mathbb{R}_+\times \mathbb{R}_+)} = \frac{1}{2\mu} \pi^+(A\times B).
\end{equation*}
Therefore,
\begin{align}
\nonumber \mathbb{E}&\left[\exp\left\{-\lambda d^*-\theta E(d^*)\right\} \right]= \mathbb{E}\left[ \exp\left\{-\lambda (g^*+{\hat{\tau}^{\downarrow,+}}_J) - \theta(E(g^*)+ \tau^{\downarrow,+}_J)\right\}\right]\\
\nonumber &= \mathbb{E}\left[ \exp\left\{-\lambda g^* -\theta E(g^*)\right\}\right] \mathbb{E}\left[\exp\left\{-\lambda {\hat{\tau}^{\downarrow,+}}_J - \theta \tau^{\downarrow,+}_J\right\}\right]\\
\nonumber &= \frac{\displaystyle \mu}{ \displaystyle \mu + \sqrt{\frac{1}{2} \phi(\lambda)+\frac{1}{2}\theta} + \frac{1}{2} \int_{(0,\infty)^2} (1-e^{-\lambda s-\theta t})\,\pi^{-}(ds,dt)} \frac{1}{2\mu}\int_{(0,\infty)^2} e^{-\lambda s- \theta t} \pi^+(ds,dt)\\
&=\frac{\displaystyle \frac{1}{2}\int_{(0,\infty)^2} e^{-\lambda s- \theta t} \pi^+(ds,dt)}{ \displaystyle \mu + \sqrt{\frac{\phi(\lambda)+\theta}{2} } + \frac{1}{2} \int_{(0,\infty)^2} (1-e^{-\lambda s-\theta t})\,\pi^{-}(ds,dt)} \label{eqn:jointLT_for_E_dstar}.
\end{align}
This is precisely the Laplace transform of $m_2(ds,dv)$ in Theorem \ref{thm:priceTheorem}.
\section{A Parisian option in illiquid markets}\label{sec:illiquidMarkets}
As mentioned in the introduction, one attempt to overcome the inability to sell assets for different intervals of time is to introduce a time-change by the inverse of a strictly increasing suboridnator. We consider a geometric Brownian motion, $X$, satisfying \eqref{eqn:gmb} where $W$ is a Brownian motion under a measure $\tilde{\mathbb{Q}}$ and $S$ is an independent strictly increasing subordinator. The price of the illiquid asset is then modeled by $X^*(t) = X(E(t))$ where $E$ is defined in \eqref{eqn:Edef}. The process $X^*$ can be written as
\begin{equation*}
X^*(t) = x \exp\left( \sigma W(E(t)) + (\mu- \frac{\sigma^2}{2})E(t)\right),
\end{equation*} which is the unique solution to \eqref{eqn:agbm}.
The incorporation of interest into this model is a bit tricky because one cannot sell the stock at anytime (it is illiquid) in order to obtain the risk free bond and so, for simplicity, we assume that the interest rate is zero, i.e. $r = 0$. See \cite{MG.12} for one approach of incorporating non-zero interest rates. As shown in \cite{Magdziarz.09} in the case where $S$ is a stable subordinator and in \cite{MG.12} in general, the market $(\Omega, (\mathcal{F}^*_t), \mathcal{F}^*, \tilde{\mathbb{Q}})$ with $\Omega = C([0,\infty),\mathbb{R})$, $\mathcal{F}^*_t$ is $\sigma(X^*(s); s\le t)$ completed in the standard way with all the null sets is arbitrage free for finite-time intervals but incomplete.
Fix a deterministic time $T> 0$. A non-unique equivalent martingale measure $\mathbb{Q}$ is given by
\begin{equation}\label{eqn:dQdPFT}
\frac{d\mathbb{Q}|_{\mathcal{F}^*_T}}{d \tilde{\mathbb{Q}}|_{\mathcal{F}^*_T}} = \exp\left\{-\frac{\mu}{\sigma} W(E(T)) - \frac{\mu^2}{2\sigma^2} E(T) \right\}.
\end{equation} Setting $Z(t) = W(t) + \frac{\mu}{\sigma}t$, the authors of \cite{MG.12} show that $X^*(t)$ and $Z^*(t) = Z(E(t))$ are $\mathbb{Q}$-local martingales (see a similar derivation below). Note that $(S(t);t\ge 0)$ is a family of increasing $\{\mathcal{F}^*_t\}_{t\ge0}$-stopping times.
Hence, by the Dambis-Dubins-Schwarz theorem \cite[Theorem V.1.6]{RY.99} the process $Z$ is a Brownian motion in the time-changed filtration $\mathcal{F}^*_{S(t)}$.
The process $X^*(t) = x \exp\left\{ \sigma ( Z(E(t)) + \frac{\sigma}{2} E(t) ) \right\}$ is a time-changed geometric Brownian motion under $\mathbb{Q}$. This follows from $Z$ being independent of $S$. Indeed, consider two random variables $F = F(Z(u);u\in[0,T])$ and $G = G(S(u);u\in[0,T])$ which are both $\mathcal{F}^*_{S(T)}$-measurable. Moreover, by Proposition VIII.1.3 in \cite{RY.99}, we have
\begin{equation*}
\begin{split}
\mathbb{Q}(A\cap \{S(T) <\infty\}) &= \int_{A\cap \{S(T)<\infty\}} \exp\left\{ -\frac{\mu}{\sigma} Z\left(E (S(T))\right) + \frac{\mu^2}{2\sigma^2} E(S(T))\right\} \,d\tilde{\mathbb{Q}}\\
&= \int_A \exp\left\{-\frac{\mu}{\sigma} Z(T) + \frac{\mu^2}{\sigma^2} T\right\}\,d\tilde{\mathbb{Q}}
\end{split}
\end{equation*} since $S(T)<\infty$ and $E(S(T)) = T$ hold $\tilde{\mathbb{Q}}$ almost surely.
Thus we have
\begin{equation*}
\begin{split}
\mathbb{E}_{\mathbb{Q}}[F G] &= \mathbb{E}_{\tilde{\mathbb{Q}}} \left[ e^{-\frac{\mu}{\sigma} Z(T)+ \frac{\mu^2}{2\sigma^2}T} FG\right]\\
&= \mathbb{E}_{\tilde{\mathbb{Q}}}\left[ e^{-\frac{\mu}{\sigma} Z(T)+ \frac{\mu^2}{2\sigma^2}T} F \right] \mathbb{E}_{\tilde{\mathbb{Q}}} [G]\\
&= \mathbb{E}_{\mathbb{Q}}[F] \mathbb{E}_{\tilde{\mathbb{Q}}}[G]\\
&= \mathbb{E}_{\mathbb{Q}}[F] \mathbb{E}_{\tilde{\mathbb{Q}}}[e^{-\frac{\mu}{\sigma} Z(T) + \frac{\mu^2}{2\sigma^2} T} G]\\
&= \mathbb{E}_{\mathbb{Q}}[F] \mathbb{E}_{\mathbb{Q}}[G],
\end{split}
\end{equation*} where in the penultimate step we used the independence (under $\tilde{\mathbb{Q}}$) of $Z$ and $S$ and the fact that
\begin{equation*}
\mathbb{E}_{\tilde{\mathbb{Q}}}[\exp(-\frac{\mu}{\sigma} Z(T))] = \mathbb{E}_{\tilde{\mathbb{Q}}}[\exp(-\frac{\mu}{\sigma} W(T) - \frac{\mu^2}{\sigma^2} T)] = \exp( -\frac{1}{2} \frac{\mu^2}{\sigma^2} T).
\end{equation*} Setting $F = 1$ and $G = \exp(-\lambda S(t))$, we can see that $S$ is a subordinator under $\mathbb{Q}$ with the same Laplace exponent.
Now it is easy to see that the price of the down-and-in European Parisian option with barrier $L$, duration $D$, maturation $T$ and payoff $\Phi$ in this illiquid market is indeed given by equation \eqref{eqn:CdiForTimeChange}. We will now do what was outlined in Section \ref{sec:parisianPrice} which did not involve a time-change. That is, we change measure to a measure $\mathbb{P}$ under which $B(t) = Z(t)+\frac{\sigma}{2} t$ is a Brownian motion in the filtration $(\mathcal{F}^*_{S(t)})$.
\subsection{The measure $\mathbb{P}$}
Define the measure $\mathbb{P}$ on $\mathcal{F}^*_t$ by, for $A\in \mathcal{F}^*_t$,
\begin{equation}\label{eqn:dQQdQ}
\mathbb{P}(A) = \int_A \exp\left\{ - \frac{\sigma}{2} Z(E(t)) - \frac{\sigma^2}{8} E(t)\right\} \,d\mathbb{Q}.
\end{equation}
First, since $Z$ is independent of $E$ under $\mathbb{Q}$ and $Z$ is Brownian motion under $\mathbb{Q}$, the measure $\mathbb{P}$ is a probability measure. Indeed for any $\lambda\in \mathbb{R}$, disintegration gives
\begin{equation*}
\begin{split}
\mathbb{E}_{\mathbb{Q}}\left[ \exp\left\{\lambda Z(E(t)) - \frac{\lambda^2}{2} E(t)\right\} \right] &= \int_0^\infty \mathbb{E}_\mathbb{Q}[\exp(\lambda Z(u)- \frac{\lambda^2}{2}u)] \,\mathbb{Q}(E(t)\in du)\\
& = \int_0^\infty 1 \mathbb{Q}(E(t)\in du) = 1.
\end{split}
\end{equation*}
Now, the Girsanov-Meyer theorem \cite[Chapter III, Theorem 40]{Protter.04} implies that
\begin{equation*}
Z(E(t)) - \left\langle Z(E(\cdot)), -\frac{\sigma}{2} Z(E(\cdot))\right\rangle(t) = Z(E(t))+ \frac{\sigma}{2}E(t)
\end{equation*} is a $\mathbb{P}$-local martingale. The Dambins-Dubins-Schwarz theorem \cite[Theorem V.1.6]{RY.99} then implies that $B = (B(t):=Z(t) + \frac{\sigma}{2} t; t\ge 0)$ is a $(\mathcal{F}^*_{S(t)})$-Brownian motion under $\mathbb{P}$. A similar proof to that showing $Z$ is independent of $S$ under $\mathbb{Q}$ can be used to prove that $B$ is independent of the suborindator $S$ under $\mathbb{P}$.
We summarize this in the following lemma:
\begin{lem} \label{lem:equivalentMeasures}
With the above set-up the following hold.
\begin{enumerate}
\item $S = (S(t);t\ge 0)$ is an increasing family of $(\mathcal{F}^*_t)$-stopping times, and $W(t) = W(E(S(t)))$ is an $(\mathcal{F}^*_{S(t)})_{t\ge0}$ Brownian motion under $\tilde{\mathbb{Q}}$.
\item The measure $\mathbb{Q}$ defined by \eqref{eqn:dQdPFT} is equivalent to $\tilde{\mathbb{Q}}$ on $\mathcal{F}^*_T$; i.e. it is locally equivalent to $\tilde{\mathbb{Q}}$. Moreover:
\begin{enumerate}
\item The process $X^*$ solving \eqref{eqn:agbm} is an $\mathcal{F}^*_t$-local martingale under $\mathbb{Q}$;
\item The process $Z(t) = W(t)+\frac{\mu}{\sigma} t$ is an $(\mathcal{F}^*_{S(t)})_{t\ge0}$-adapted process, which is a Brownian motion under $\mathbb{Q}$;
\item $Z^*(t) = Z(E(t))$ is an $\mathcal{F}^*_t$-local martingale under $\mathbb{Q}$; and
\item The law of $S$ under $\mathbb{Q}$ is the same as the law of $S$ under $\tilde{\mathbb{Q}}$ and it is independent of $Z$.
\end{enumerate}
\item The measure $\mathbb{P}$ defined by \eqref{eqn:dQQdQ} is equivalent to $\mathbb{Q}$ on $\mathcal{F}^*_T$, that is it is locally equivalent to $\mathbb{Q}$. Moreover:
\begin{enumerate}
\item The process $B(t) = Z(t)+\frac{\sigma}{2} t$ is an $(\mathcal{F}^*_{S(t)})_{t\ge0}$-adapted process, which is a Brownian motion under $\mathbb{P}$;
\item $B^*(t) = B(E(t))$ is an $\mathcal{F}^*_t$-local martingale under $\mathbb{P}$; and
\item The law of $S$ under $\mathbb{P}$ is the same as the law of $S$ under $\tilde{\mathbb{Q}}$ and it is independent of $B$.
\end{enumerate}
\end{enumerate}
\end{lem}
We also state the following corollary, which will be useful in any generalization of Theorem \ref{thm:priceTheorem} to the case where $L\neq x$.
\begin{cor}\label{cor:strongMarkovIndependence}
Let $T_b = \inf\{t: B^*(t) = b\}$ be the first hitting time of $b$ by the time-changed Brownian motion $B^*$. Then, under $\mathbb{P}$,
\begin{enumerate}
\item $T_b = S(\inf\{t: B(t) = b\})$ is an $\mathcal{F}^*_{t}$-stopping time; and
\item $B^*(t+T_b)-b$ is independent of $\mathcal{F}^*_{T_b}$ and is equal in distribution to $B^*$.
\item $E(t+T_b) - E(T_b)$ is independent of $\mathcal{F}^*_{T_b}$ and is equal in distribution to $E$.
\end{enumerate}
\end{cor}
\begin{proof}
We begin with proving item 1. First note that $\mathcal{F}^*_t$ satisfies the usual conditions and $B$ is $\mathcal{F}^*_t$-progressive, so $T_b$ is a d\'{e}but time and is hence a stopping time. Note that
\begin{equation*}
\begin{split}
T_b &= \inf\{t: B(E(t)) = b\}= \inf\{t: E(t) = \inf\{u: B(u) = b\}\}.
\end{split}
\end{equation*}
Let $\tilde{T}_b = \inf\{u: B(u) = b\}$ be the first hitting time of $b$ by the Brownian motion $B$. Obviously, $\tilde{T}_b$ is an $\mathcal{F}^*_{S(t)}$-stopping time.
Also note that $S(\tilde{T}_b) = S(\tilde{T}_b-)$ with probability 1 since $B$ is independent of $S$, and hence
\begin{equation}\label{eqn:TBSTB}
T_b = \inf\{t:E(t) = \tilde{T}_b\} = S(\tilde{T}_b-) = S(\tilde{T}_b).
\end{equation} This proves item 1.
To prove items 2 and 3, we note that $(B,S)$ is a L\'{e}vy process under $\mathbb{P}$ and thus strong Markov process under $\mathbb{P}$ with respect to the filtration $(\mathcal{F}^*_{S(t)})_{t\ge 0}$. Thus $\tilde{B} = (\tilde{B}(t);t\ge 0)$ defined by $\tilde{B}(t) = B(t+\tilde{T}_b)-b$ and $\tilde{S} = (\tilde{S}(t);t\ge 0)$ defined by $\tilde{S}(t):=S(t+\tilde{T}_b)-S(\tilde{T}_b)$ are each independent of $\mathcal{F}^*_{S(\tilde{T}_b)} = \mathcal{F}^*_{T_b}$ and are equal in law to $B$ and $S$, respectively.
Now observe from equation \eqref{eqn:TBSTB}
\begin{equation*}
\begin{split}
E(t+T_b) &= \inf\{u: S(u)> t+S(\tilde{T}_b)\} = \inf\{v+\tilde{T}_b: S(v+\tilde{T}_b)-S(\tilde{T}_b) > t\} \\
&= \inf\{v: \tilde{S}(v)>t\} + \tilde{T}_b.
\end{split}
\end{equation*} Therefore, if $\tilde{E}$ is the right-continuous inverse of $\tilde{S}$ then
\begin{equation*}
E(t+T_b) = \tilde{E}(t) +\tilde{T}_b,
\end{equation*} which, together with \eqref{eqn:TBSTB}, proves item 3. But then
\begin{equation*}
B^*(t+T_b) = B(E(t+T_b)) = B(\tilde{E}(t)+\tilde{T}_b) = b+\tilde{B}(\tilde{E}(t)),
\end{equation*} which proves item 2.
\end{proof}
We finish this section by giving an expression for \eqref{eqn:CdiForTimeChange} under the measure $\mathbb{P}$ and the Brownian motion $B$. Before writing that down note, that $H_{L,D}^-(X^*)\le T$ if and only if $H_{b,D}^-(B^*)\le T$ where $b = \frac{1}{\sigma} \log(L/x)$. Consequently, comparing equations \eqref{eqn:CdiForTimeChange} and \eqref{eqn:dQQdQ} we arrive at
\begin{equation}\label{eqn:downandincall}
\begin{split}
V_{d,i}^*&(x,T,L,D,\Phi)\\
&=\mathbb{E}_{\mathbb{P}}\left[ 1_{[H_{b,D}^-(B^*)\le T]} \exp\left\{\frac{\sigma}{2} B(E(T)) - \frac{\sigma^2}{8} E(T) \right\} \Phi(xe^{\sigma B(E(T))}) \right]\\
&= \mathbb{E}_{\mathbb{P}}\left[1_{[H_{b,D}^-(B^*)\le T]} f_{x,\sigma}(B(E(T)), E(T)) \right]
\end{split}
\end{equation}
where $f_{x,\sigma}(z,s) = \exp\left( \frac{\sigma}{2} z - \frac{\sigma^2}{8} s \right) \Phi\left(xe^{\sigma z}\right)$.
\subsection{The proof of Theorem \ref{thm:priceTheorem}}
Before continuing, recall the excursion construction of both a Brownian motion and its time-change in Theorem \ref{thm:pppfortc} and Corollary \ref{cor:EFromExcursion}. We will use the notation in Corollary \ref{cor:EFromExcursion} along with the various subordinators $\tau^{\uparrow}, {\hat{\tau}^\uparrow}$, etc. freely.
We note that in Theorem \ref{thm:priceTheorem} we are pricing a Parisian option where the barrier $L$ is the initial stock price, i.e. $L = x$. By the discussion at the end of the previous section this means that $b = 0$. Moreover, since the Brownian motion $B$ starts from $0$ we have
\begin{equation*}
\{H_{b,D}^-(B^*) \le T\} = \{H_{0,D}^{-} (B^*)\le T\}= \{g^*\le T-D\},
\end{equation*} where $g^*$ is defined in \eqref{eqn:gstarDstar_def} and is studied in Section \ref{sec:firstExcursion}. Indeed, $H_{0,D}^-(B^*)$ is the first time the time-change $B^*$ spends $D$ consecutive units of time below the origin. This occurs before time $T$ if and only if the corresponding excursion starts before time $T-D$.
We now have two situations to consider: 1) $g^*\le T-D$ but $d^*> T$; and 2) $d^*\le T$. We handle each case separately.
\subsubsection{Case 1: $g^*\le T-D$ and $d^*>T$}
Recall $J = \inf\{x:\gamma_x =-1, \zeta(e_x^*) >D\}$ is exponentially distributed. The following two events are equal:
\begin{equation*}
\{g^*\le T-D, d^*>T\} = \{\hat{\tau}_{J-} \le T-D, \hat{\tau}_J >T \}.
\end{equation*}
We have
\begin{align}
\mathbb{E}&\left[1_{[g^*\le T-D, d^*> T]} f_{x,\sigma}\left(B(E(T)), E(T)\right) \right]\nonumber\\
&= \mathbb{E}\left[ 1_{[g^*\le T-D, d^*>T]} e^{-\frac{\sigma^2}{8} E(g^*)} f_{x,\sigma}\left(B(E(T)), E(T)-E(g^*)\right) \right]\nonumber \\
&= \int \mathbb{P}(g^*\in ds, E(g^*)\in dv) 1_{[s\le T-D]} e^{-\frac{\sigma^2}{8}v} \nonumber\\
&\qquad\qquad \times \int_{\mathscr{E}} \mathfrak{n}^*(dw^*|\zeta(w^*)>T-s) f_{x,\sigma}\left(-w^*(T-s), \langle w^*\rangle (T-s) \right). \label{eqn:u1Help1}
\end{align} Indeed, the first equality is simply the fact $f_{x,\sigma}(z, s) = e^{-\frac{\sigma^2}{8}s_0} f_{x,\sigma}(z,s-s_0)$. The second equality follows from conditioning the value of $(g^*,E(g^*))$, Proposition XII.1.13 in \cite{RY.99} and recognizing that conditioning on $g^* = s$ and $d^*>T$ the excursion over $(g^*,d^*)$ is distributed as $-w^*$ under $\mathfrak{n}^*(dw^* | \zeta(w^*)> T-s).$
Before continuing, let us discuss the one-dimensional time-marginal distribution of $(w^*(t), \langle w^*\rangle(t))$ under $\mathfrak{n}^*(dw^* | \zeta(w^*)> t)$. By Theorem \ref{thm:excursionChange} and the same reasoning in \eqref{eqn:zetawStar}, this is equivalent to understanding
\begin{equation*}
\Xi(w, M, \kappa)(t) \qquad\text{under}\quad \left( \mathfrak{n}\otimes \mathcal{L}(M)\right) \left(dw\times dM \bigg| \int_0^{\infty} y M([0,\zeta(w)]) + \kappa \zeta(w) > t\right).
\end{equation*}
This is also the same as
\begin{equation*}
\Xi(w,M,\kappa)(t) \qquad\text{under}\quad \left(\mathfrak{n}\otimes \mathcal{L}(M)\right) \left(dw\times dM \bigg| \zeta(w)> \inf\left\{u: \int_0^\infty yM([0,u],dy) + \kappa u>t \right\}\right).
\end{equation*}
It is now easy to see that
\begin{equation*}
\left(w^* \text{ under } \mathfrak{n}^*(dw^*| \zeta(w^*)>t) \right) \overset{d}{=} \left(w(E(\cdot)) \text{ under } w\sim \mathfrak{n}(dw| \zeta(w)> E(t)), E\text{ independent of }w \right).
\end{equation*}
We observe that $w(t)$ under $\mathfrak{n}(\ \cdot \ | \zeta(w)>t)$ has law \cite[Section XII.4]{RY.99}
\begin{equation*}
\mathfrak{n}(w(t)\in dz| \zeta(w)>t)= \frac{1}{\mathfrak{n}(\zeta>t)} \sqrt{\frac{2}{\pi t^3}} z e^{-z^2/2t}\,dz = \frac{z}{t} e^{-z^2/2t}\,dz.
\end{equation*}
This is precisely the marginal density of a 2-dimensional Bessel process at time $t$ \cite[Section XI.1]{RY.99}. Let $Y = (Y(u);u\ge 0)$ denote a 2-dimensional Bessel process starting from $0$ independent of $E$ and fix a bounded function $f:\mathbb{R}^2\to \mathbb{R}$. By the reasoning above and writing $\mathbb{E}^E$ for the expectation only with respect to the process $E$ and similarly for $\mathbb{E}^Y$ we have
\begin{equation}\label{eqn:u1Help2}
\begin{split}
\int_{\mathscr{E}} &f\left(w^*(t), \langle w^*\rangle(t) \right)\,\mathfrak{n}^*(dw^*|\zeta(w^*)>t)= \mathbb{E}^{E} \left[\int_{\mathscr{E}} f\left( w(E(t)), E(t)\right) \,\mathfrak{n}(dw | \zeta(w)>E(t))\right]\\
&= \mathbb{E}^E \left[ \mathbb{E}^Y [ f(Y(E(t)),E(t)) ] \right] = \mathbb{E}[f(Y(E(t)),E(t))].
\end{split}
\end{equation}
More generally, let $Y$ be any $2$-dimensional Bessel process starting from $Y(0)\ge 0$. Let $a(t):=a(0)+t$ be a deterministic Markov process. Observe that the 2-dimensional process $\xi (t) = (Y(t),a(t))$ is strong Markov with respect to measures $\{\mathbb P_{(z,s)}: (z,s) \in \mathbb{R}_{+} \times \mathbb{R}\}$ under which $\xi(0)=(z,s)$ $\mathbb{P}_{(z,s)}$-almost surely. Denote by $\mathbb E_{(z,s)}$ its corresponding mathematical expectation. In \eqref{eqn:u1Help2}, the expectation $\mathbb E$ stands for $\mathbb E_{(0,0)}$. The process $\xi$ has generator
\begin{equation*}
\mathscr{L}_{\xi} f(z,s) = \left(\frac{1}{2} \frac{\partial^2}{\partial z^2} + \frac{1}{2z} \frac{\partial}{\partial z} + \frac{\partial}{\partial s}\right) f(z,s),
\end{equation*}
and the domain contains functions $f\in C^\infty$ such that $f(z,s) = g(z)h(s)$ for smooth functions $g$ and $h$ such that $g$ vanishes in a neighborhood of zero. This can be seen from \cite[Section 2.7]{IM.74} for the Bessel coordinate $Y$ and a direct computation afterwards. In particular $f(z,s ) =f_{x,\sigma}(-z,s)$ is in the domain of $\mathscr{L}_\xi$.
As discussed in Section \ref{sec:timeFracPDE}, we can write for any $x = (z,s)\in \mathbb{R}_+\times\mathbb{R}$
\begin{equation*}
u(z,s;t) := \mathbb{E}_{(z,s)}[f(\xi (E(t)))],
\end{equation*} as the unique solution to
\begin{equation}\label{eqn:u1Help3}
\partial_t^S u = \mathscr{L}_\xi u,\qquad\text{with}\qquad u(z,s;0) = f(z,s),
\end{equation}
where $\partial_t^S$ is defined in \eqref{eqn:genCaputo}.
Combining \eqref{eqn:u1Help1}, \eqref{eqn:u1Help2} and the time-fractional PDE discussion above, we get
\begin{equation}\label{eqn:u1m1inproof}
\begin{split}
\mathbb{E}&\left[1_{[g^*\le T-D, d^*> T]} f_{x,\sigma}\left(B(E(T)), E(T)\right) \right]\\
&=\int 1_{[s\le T-D]} e^{-\frac{\sigma^2}{8} v} u_1(0,0; T-s)\,m_1(ds,dv)
\end{split}
\end{equation} where $u_1$ solves \eqref{eqn:u1Help3} with $f(z,s) = f_{x,\sigma}(-z,s) = \exp\left\{-\frac{\sigma}{2} z - \frac{\sigma^2}{8}s\right\}\Phi(xe^{-\sigma z})$ and $m_1(ds,dv) = \mathbb{P}(g^*\in ds, E(g^*)\in dv)$ which has Laplace transform \eqref{eqn:jointLT_for_E_gStar}. This matches the definitions appearing in Theorem \ref{thm:priceTheorem}.
\subsubsection{Case 2: $d^*\le T$}
The second case is relatively easier because of conclusion 2 in Lemma \ref{lem:independence}. Indeed, $(e^*_{x+J}, \gamma_{x+J})_{x>0}$ is equal in law to $(e_x^*,\gamma_x)$ and is independent of $\mathcal{G}_J$. As in Theorem \ref{thm:pppfortc} and Corollary \ref{cor:EFromExcursion} let $\tilde{B}^* := \tilde{B}\circ \tilde{E}$ and $\tilde{E}$ be, respectively, the time-changed Brownian motion and the time-change constructed, from the excursion process $(e^*_{x+J}, \gamma_{x+J})_{x>0}$. Now observe that $d^* = \hat{\tau}_J$ and that the point process $(e^*_{x+J},\gamma_{x+J})_{x>0}$ is independent of both $d^*$ and $E(d^*)$ (cf. conclusion 2 from Lemma \ref{lem:independence}). By a similar argument, although less-involved, to the argument utilized in Case 1, we have
\begin{equation}\label{eqn:u2m2inproof}
\begin{split}
\mathbb{E}&\left[ 1_{[d^*\le T]} f_{x,\sigma}(B(E(T)), E(T)) \right] = \mathbb{E}\left[ 1_{[d^*\le T]} e^{-\frac{\sigma^2}{8} E(d^*)} f_{x,\sigma}(\tilde{B}(\tilde{E}(T-d^*)), \tilde{E}(T-d^*))\right]\\
&= \int_0^\infty \mathbb{P}( d^*\in ds, E(d^*)\in dv) 1_{[D\le s\le T]} e^{-\frac{\sigma^2}{8} v} \mathbb{E}\left[f_{x,\sigma}(\tilde{B}(\tilde{E}(T-s)), \tilde{E}(T-s))\right]\\
&= \int 1_{[D\le s\le T]} e^{-\frac{\sigma^2}{8} v} u_2(0,0;T-s)\,m_2(ds,dv),
\end{split}
\end{equation} where $u_2$ is the unique solution to
\begin{equation*}
\partial_t^S u = \left( \frac{1}{2} \frac{\partial^2}{\partial z^2} + \frac{\partial}{\partial s}\right) u,\qquad u(z,s;0) = f_{x,\sigma}(z,s)
\end{equation*}and $m_2(ds,dv) = \mathbb{P}(d^*\in ds, E(d^*)\in dv)$. The measure $m_2$ has Laplace transform \eqref{eqn:jointLT_for_E_dstar}. These two definitions of $u_2$ and $m_2$ match the definitions found in Theorem \ref{thm:priceTheorem}.
The proof of Theorem \ref{thm:priceTheorem} now follows from adding together equations \eqref{eqn:u1m1inproof} and \eqref{eqn:u2m2inproof}.
\begin{comment}
Recall the Poisson process construction of the time-changed Brownian motion $B^*$ described in Theorem \ref{thm:pppfortc} and let $({e}^*_x, \gamma_x;x>0)$ be the corresponding Poisson point process, with respect to the filtration $(\mathcal{G}_x)_{x\ge 0}$, of excursions and signs. We implicitly assume that $B^*$ is constructed as in Theorem \ref{thm:excursionChange}. We observe that the excursion of $B^*$ on the interval $(g^*,d^*)$ which is the first excursion of duration at least $D$ which is negative. See equation \eqref{eqn:gstarDstar_def}. Thus, the event $\{H_{0,D}^-(B^*)<T\}$ is simply $\{d^*\le T-D\}$.
We now discuss the distribution of $g^*$. To begin with, we will begin by writing two subordinators
\begin{equation*}
{\hat{\tau}^\uparrow}_x = \sum_{y\le x} \zeta(e^*_y)1_{[\gamma_y=1]}\qquad \text{and}\qquad {\hat{\tau}^\downarrow}_x = \sum_{y\le x} \zeta(e^*_y)1_{[\gamma_y = -1]}.
\end{equation*} The subordinators ${\hat{\tau}^\uparrow}$ and ${\hat{\tau}^\downarrow}$ are independent and identically distributed. Almost surely, ${\hat{\tau}^\uparrow}$ and ${\hat{\tau}^\downarrow}$ do not jump at the same time and it is easy to see that $\hat{\tau}_x = {\hat{\tau}^\uparrow}_x + {\hat{\tau}^\downarrow}_x$. Moreover both have Laplace exponent $\lambda\mapsto \sqrt{\frac{c}{2}} \lambda^{\beta/2}$. Indeed, this can be obtained by using Corollary \ref{cor:betaover2} and Proposition XII.1.2 in \cite{RY.99}:
\begin{equation*}
\begin{split}
-\log\mathbb{E}&\left[ \exp\left\{-\lambda {\hat{\tau}^\uparrow}_x\right\}\right] = -\log \mathbb{E}\left[ \exp\left\{-\lambda \sum_{y\le x} \zeta(e^*_y) 1_{[\gamma_x = 1]} \right\} \right]\\
&= x\int\limits_{\mathscr{E} \times \{\pm 1\}} (1-\exp(-\lambda \zeta(e))) 1_{[\gamma = 1]}\,\mathfrak{n}^*(de)\frac{1}{2}\left( \delta_{-1}(d\gamma)+ \delta_{1}(d\gamma) \right)\\
&= \frac{x}{2} \int\limits_\mathscr{E} (1-e^{-\lambda \zeta(e)}) \mathfrak{n}^*(de)\\
&= \frac{x}{2} \frac{\beta}{2} \frac{\sqrt{2c}}{\Gamma(1-\frac{\beta}{2})} \int\limits_0^\infty (1-e^{-\lambda r}) r^{-\beta/2 -1} \,dr\\
&= \sqrt{2c}\lambda^{\beta/2} \frac{x}{2} \frac{-\frac{\beta}{2} \Gamma(-\frac{\beta}{2})}{\Gamma(1-\frac{\beta}{2})}= \sqrt{\frac{c}{2}} \lambda^{\beta/2} x.
\end{split}
\end{equation*} We will let $\nu^{\downarrow} (r,\infty) = \sqrt{\frac{c}{2} } \frac{1}{\Gamma(1-\frac{\beta}{2})} r^{-\beta/2}$ be the L\'{e}vy measure of ${\hat{\tau}^\downarrow}$.
Note that ${\hat{\tau}^{\downarrow,-}}$ and ${\hat{\tau}^{\downarrow,+}}$ are independent of each other sine they are defined in terms of a single Poisson random measure on two different sets. Their respective L\'{e}vy measures are $1_{[0\le r\le D]}\nu^{\downarrow}(dr)$ and $1_{[r>D]} \nu^{\downarrow}(dr)$ respectively.
If we define $J = \inf\{x: {\hat{\tau}^\downarrow}_x - {\hat{\tau}^\downarrow}_{x-}>D\} = \inf\{x: {\hat{\tau}^{\downarrow,+}}_x>0\}$. Then, since $J$ is the waiting time for a Poisson point process to hit as set of finite mass $\mu:=\nu^{\downarrow}(D,\infty)$, it has an exponential distribution with parameter $\nu^{\downarrow}(D,\infty)$. That is
\begin{equation*}
J\sim \text{Exp}(\mu),\qquad \text{where}\quad \mu = \mu(D,\beta,c) = \sqrt{\frac{2}{c}} \frac{1}{\Gamma(1-\frac{\beta}{2})} D^{-\beta/2}.
\end{equation*}
Now note that
\begin{equation*}
\begin{split}
\inf\{\hat{\tau}_{x-}: \gamma_xe_x^*(D)< 0\} = \hat{\tau}_{J-} = {\hat{\tau}^\uparrow}_J + {\hat{\tau}^\downarrow}_{J-} = {\hat{\tau}^\uparrow}_J + {\hat{\tau}^{\downarrow,-}}_J,
\end{split}
\end{equation*}
where the last equality follows from the fact that ${\hat{\tau}^\downarrow}_{J-} = {\hat{\tau}^{\downarrow,-}}_J$.
Now note that $J$ is independent of both ${\hat{\tau}^\uparrow}$ and ${\hat{\tau}^{\downarrow,-}}$ and hence
\begin{equation*}
\begin{split}
\mathbb{E}&\left[e^{-\lambda g^*} \right] = \mathbb{E}\left[\exp\left\{-\lambda({\hat{\tau}^\uparrow}_J+ {\hat{\tau}^{\downarrow,-}}_J)\right\} \right]\\
&=\int_0^\infty \mu e^{-\mu t} \mathbb{E}\left[ e^{-\lambda {\hat{\tau}^\uparrow}_t}\right] \mathbb{E}\left[ e^{-\lambda {\hat{\tau}^{\downarrow,-}}_t}\right]\,dt\\
&= \int_0^\infty \mu \exp\left\{-t \left(\mu + \sqrt{\frac{c}{2}} \lambda^{\beta/2} + \int_0^D (1-e^{-\lambda r})\nu^\downarrow(dr)\right)\right\}\,dt\\
&= \frac{\mu}{\displaystyle \mu + \sqrt{\frac{c}{2}} \lambda^{\beta/2} + \int_0^D (1-e^{-\lambda r})\nu^{\downarrow}(dr)}.
\end{split}
\end{equation*}
Similarly, since ${\hat{\tau}^{\downarrow,+}}_J$ is the first jump of a compound Poisson process with jump distribution proportional to $1_{[r>D]} \nu^\downarrow(dr) = \frac{\beta}{2}D^{\beta/2} r^{-\beta/2-1}\,dr$ and the size is independent of the time $J$. It is easy to verify that
\begin{equation*}
\mathbb{E}\left[e^{-\lambda{\hat{\tau}^{\downarrow,+}}_{J}}\right] = \frac{\beta}{2} D^{\beta/2}\int_D^\infty e^{-\lambda r} r^{-\frac{\beta}{2} -1}\,dr.
\end{equation*}
Also we have $d^* = g^* + {\hat{\tau}^{\downarrow,+}}_J$ is the sum of two independent random variables and hence
\begin{equation*}
\mathbb{E}\left[e^{-\lambda d^*} \right] = \frac{\displaystyle \frac{\mu\beta}{2} D^{\beta/2}\int_D^\infty e^{-\lambda r} r^{-\frac{\beta}{2} -1}\,dr}{\displaystyle \mu + \sqrt{\frac{c}{2}} \lambda^{\beta/2} + \int_0^D (1-e^{-\lambda r})\nu^{\downarrow}(dr)} = \frac{\displaystyle \frac{\beta}{\sqrt{2c}\Gamma(1-\frac{\beta}{2})}\int_D^\infty e^{-\lambda r} r^{-\frac{\beta}{2} -1}\,dr}{\displaystyle \mu + \sqrt{\frac{c}{2}} \lambda^{\beta/2} + \int_0^D (1-e^{-\lambda r})\nu^{\downarrow}(dr)}.
\end{equation*}
Now we note by the strong Markov property for the Poisson point process $(e_x^*,\gamma_x)_{x>0}$ that $(e_{x+J}^*,\gamma_{x+J})_{x>0}$ is independent of $\mathcal{G}_J$ and is equal in distribution to $(e_x^*,\gamma_x)_{x>0}$. We will let $\tilde{B}^*$ be the time-changed Brownian motion constructed from the process $(e_{x+J}^*,\gamma_{x+J})$. That is
\begin{equation*}
\tilde{\hat{\tau}}_x = \sum_{0<y\le x} \zeta(e^*_{J+y}) ,\quad \tilde{L}(t) = \inf\{x: \tilde{\hat{\tau}}_x>t\},\qquad \tilde{B}^*(t) = \sum_{0<y\le \tilde{L}(t)} \gamma_{J+y} e_{J+y}^*(t-\tilde{\hat{\tau}}_{y-}).
\end{equation*}
Further let $\tilde{E}(t) = \langle \tilde{B}*\rangle(t)$ be the quadratic covariation of $\tilde{B}^*$. Clearly,
\begin{equation*}
\tilde{B}^*(t) = B^*(t+d^*)\quad \text{and} \quad \tilde{E}(t)+ E(d^*) = E(t),
\end{equation*} and, moreover
\begin{equation*}
(\tilde{B}^*,\tilde{E})\overset{d}{=} (B^*,E)
\end{equation*} and $(\tilde{B}^*,\tilde{E})$ is independent of $\mathcal{G}_J$. Moreover, $E(d^*)$ depends only on the excursions $(e^*_x; 0<x\le J)$ and so is $\mathcal{G}_J$-measurable.
We now compute
\begin{equation*}
\begin{split}
\mathbb{E}&\left[1_{[ d^*\le T]} f_{x,\sigma}(B^*(T),E(T)) \right]\\
&= \mathbb{E}\left[1_{d^*\le T} f_{x,\sigma}(\tilde{B}^*(T-d^*), \tilde{E}(T-d^*)+E(d^*)) \right]\\
&= \mathbb{E}\left[ \mathbb{E}[1_{[d^*\le T]} e^{-\frac{\sigma^2}{8}E(d^*)} f_{x,\sigma}(\tilde{B}^*(T-d^*),\tilde{E}(T-d^*)) | \mathcal{G}_J] \right]\\
\end{split}
\end{equation*}
that $\hat{\tau}$ is a $\frac{\beta}{2}$-stable subordinator with L\'{e}vy measure $\nu^*(r,\infty) = \frac{\sqrt{2c}}{\Gamma(1-\frac{\beta}{2})}r^{-\beta/2}$. We also recall, by properties of Poisson point processes, that for any measurable set $A\subset \mathscr{E}$ that the random variable
$\inf\{x: e_x\in A, \gamma_x= -1\}$ is an exponential random variable with mean $\frac{1}{2}\mathfrak{n}^*(A)$. In particular, by Lemma \ref{cor:betaover2}, we know that when $A = \{\zeta>D\}$ that
\begin{equation}\label{eqn:muDef}
H = \inf\{x: \gamma_xe_x^*(D)<0\} \sim \operatorname{Exp}(\mu)\quad \text{where}\quad\mu = \mu(D,\beta,c) = \frac{\sqrt{2} \Gamma\left(1-\frac{\beta}{2}\right)}{\sqrt{c}} D^{\beta/2}.
\end{equation}
We also observe that $H$ can be constructed in the following way. Let $H_1, H_2, H_3,\dotsm$ be the be the inter-arrival times of $e_x$ hitting the set $\{\zeta>D\}$. That is $H_1 = \inf\{x: \zeta(e_x)>D\}$ and $H_{n+1} = \inf\{x>0: \zeta(e_{x+H_n})>D\}$. By the strong Markov property for Poisson processes, $H_1, H_2, H_3,\dotsm$ are each independent exponential random variables with parameters $\frac{1}{2}\mu$. Since $\gamma_x$ is independent of $e_x$, the first hitting time $H$ of $\{\zeta>D\}\times\{-1\}$ satisfies
\begin{equation*}
H \overset{d}{=} \sum_{j=1}^K H_j
\end{equation*} where $K$ is independent of $(e_x;x>0)$ satisfies $P(K = r) = 2^{-r}$ for each non-negative integer $r = 0,1,\dotsm$.
Now also observe that $H_j$ is the $j^\text{th}$ inter-arrival time of the jumps of $\hat{\tau}$ exceeding size $D$. Moreover, by the strong Markov property for subordinators, $\left(\hat{\tau}_{H_j+x}-\hat{\tau}_{H_j};x\ge 0\right)$ is independent of $\mathcal{G}_{H_j}$ for each $j$ and is distributed as $\hat{\tau}$. Also observe that $\hat{\tau}$ can be constructed from a Poisson random measure $M$ on $[0,\infty)\times(0,\infty)$ with intensity measure ${\operatorname{Leb}}\otimes \nu^*$ by
\begin{equation*}
\hat{\tau}_x = \int\limits_{t=0}^x \int\limits_{y=0}^\infty y M(dt,dy).
\end{equation*}
and $H_{j+1} = \min\{x>0: M([H_j, H_j+x]\times(D,\infty)) = 1\}$.
Define $M^-_D(A\times B) = M(A\times(B\cap(0,D)))$ and $M^+_D(A\times B) = M(A\times(B\cap(D,\infty)))$. It is easy to see that these are two independent Poisson random measures on $[0,\infty)\times (0,\infty)$ with intensities ${\operatorname{Leb}}\otimes \nu^*(\cdot \cap(0,D])$ and ${\operatorname{Leb}}\otimes \nu^* (\cdot \cap (D,\infty))$. Moreover, the times $H_j$ are independent of $M_{D}^-$. Consequently,
\begin{equation*}
\hat{\tau}_{H_1-} = \int_{t=0}^{H_1-} \int_0^\infty y M(dt,dy) = \int_0^{D} y M_{D}^-([0,H_1)\times dy).
\end{equation*}
and hence, by the exponential moment formula for Poisson random measures \cite[Proposition XII.1.12]{RY.99},
\begin{equation*}
\begin{split}
\mathbb{E}&\left[ \exp\left\{ -\lambda \hat{\tau}_{H_1-}\right\}\right] = \int_0^\infty \frac{1}{2}\mu e^{-\frac{\mu}{2} t} \mathbb{E}\left[\exp\left\{-\int_0^D y M_{D}^-([0,t)\times dy)\right\} \right]\,dt\\
&= \int_0^\infty \frac{1}{2}\mu e^{-\frac{\mu}{2} t} \exp\left\{-t \int_0^D(1-e^{-\lambda r}) \nu^*(dr) \right\}\,dt \\
&= \int_0^\infty \frac{1}{2}\mu e^{-\frac{\mu}{2} t} \exp\left\{-\frac{\sqrt{2c}}{\Gamma(1-\beta/2)}t \int_0^D(1-e^{-\lambda r}) \frac{\beta}{2}r^{-\frac{\beta}{2}-1}\,dr \right\}\,dt\\
&= \int_0^\infty \frac{1}{2}\mu e^{-\frac{\mu}{2} t} \exp\left\{-\frac{\sqrt{2c}}{\Gamma(1-\beta/2)}t \left( (1-e^{-\lambda D})D^{-\beta/2} + \lambda \int_0^D r^{-\beta/2} e^{-\lambda r}\,dr\right) \right\}\,dt\\
&=\int_0^\infty \frac{1}{2}\mu e^{-\frac{\mu}{2} t} \exp\left\{-\frac{\sqrt{2c}}{\Gamma(1-\beta/2)} \left( (1-e^{-\lambda D})D^{-\beta/2} + \lambda^{\beta/2} \gamma(1-\frac{\beta}{2}, D)\right)t \right\}\,dt\\
&= \int_0^\infty \frac{1}{2} \mu e^{-\frac{\mu}{2} t} \exp\left\{-\frac{2}{\mu} \left( (1-e^{-\lambda D}) +\lambda^{\beta/2} \gamma\left(1-\frac{\beta}{2},D\right) D^{\beta/2} \right)t\right\}\,dt\\
&=\frac{\mu^2}{\mu^2 + 4\left(1-e^{-\lambda D} + \lambda^{\beta/2} \gamma\left(1-\frac{\beta}{2} , D\right) D^{\beta/2} \right)}.
\end{split}
\end{equation*}
In the third equality above we used the definition of $\nu^*$; in the fourth we used integration by parts; in the fifth we write $\gamma(s,x) = \int_0^x t^{s-1}e^{-t}\,dt$ for the lower incomplete gamma function; and in the sixth step we use the definition of $\mu$ in \eqref{eqn:muDef}.
\end{comment}
\begin{comment}
\subsection{Pricing down-and-in Parisian call options}
We must consider three separate cases: (1) $b>0$, $L>x$; (2) $b<0$, $L<x$; and (3) $b = 0$, $L=x$. Since the case $b = 0$ plays an important role in our subsequent analysis we will write
\begin{equation*}
\hat{C}_{d,i}^*(x,T, D, K) := C_{d,i}^*(x,T,x,D,K)= \mathbb{E}_{\tilde{\mathbb{Q}}}\left[1_{[H_{0,D}^-(B^*)\le T]} f_{x,\sigma}(B(E(T)), E(T)) \right].
\end{equation*}
\subsubsection{Reducing case $b>0$ to $b = 0$}
In this subsection we work exclusively with respect to the measure $\tilde{\mathbb{Q}}$. We use the notation in Corollary \ref{cor:strongMarkovIndependence}. Also let $\tilde{B}^*(t) = \tilde{B}(\tilde{E}(t))-b$, which is independent of $\mathcal{F}^*_{T_b}$ and is equal in distribution to $B^*$.
Note that if $T_b\le D$, then in order for $B^*$ to spend duration $D$ below level $b$ before time $T$, the process $\tilde{B}^*$ must spend duration $D$ below level $0$ before duration $T-T_b$. That is
\begin{equation*}
\{ T_b\le D, H_{b,D}^-(B)\le T\} = \{T_b\le D, H_{0,D}^*(\tilde{B}^*)\le T-T_b\}.
\end{equation*}
Further, on the event $T_b\le T$, the random variable
\begin{equation*}
\begin{split}
f_{x,\sigma}(B^*(T),E(T)) &= f_{x,\sigma}(\tilde{B}^*(T-T_b) +b, \tilde{E}(T-T_b) + E(T_b) )\\
&= \sqrt{\frac{L}{x}} \exp\left(-\frac{\sigma^2}{8} E(T_b)\right)f_{L,\sigma}(\tilde{B}^*(T-T_b), \tilde{E}(T-T_b))
\end{split}
\end{equation*} Moreover, the exponential term and $f_{L,\sigma}$ term are conditionally independent given $\mathcal{F}^*_{T_b}$.
Also,
\begin{equation*}
\{T_b> D, H_{b,D}^-(B)\le T\} = \{T_b> D\}.
\end{equation*}
Observe
\begin{equation*}
\begin{split}
C_{d,i}^*&(x,T, L, D, K) = \mathbb{E}_{\tilde{\mathbb{Q}}}\left[1_{[H_{b,D}^-(B^*)\le T]} f(B^*(T),E(T)) \right] \\
&= \mathbb{E}_{\tilde{\mathbb{Q}}}\left[1_{[T_b\le D]} 1_{[H_{0,D}^-(\tilde{B})\le T-T_b]} e^{\frac{-\sigma^2}{8} E(T_b)}f_{L,\sigma}(\tilde{B}^*(T-T_b), \tilde{E}^*(T-T_b)) \right]\\
&\qquad + \mathbb{E}_{\tilde{\mathbb{Q}}}\left[ 1_{T_b\in(D,T]} e^{-\frac{\sigma^2}{8} E(T_b)} f_{L,\sigma}(\tilde{B}^*(T-T_b), \tilde{E}(T-T_b)) \right]\\
&\qquad + \mathbb{E}_{\tilde{\mathbb{Q}}} \left[1_{T_b>T} f_{x,\sigma}(B^*(T),E(T)) \right]\\
&=: A_1 + A_2 + A_3, \quad \text{say.}
\end{split}
\end{equation*}
We note that by conditioning on $\mathcal{F}^*_{T_b}$, the first term $A_1$ becomes
\begin{equation*}
\begin{split}
A_1 &= \int 1_{[s\le D]} e^{-\frac{\sigma^2}{8} v} \mathbb{E}_{\tilde{\mathbb{Q}}}\left[1_{H^-_{0,D}(\tilde{B}^*)\le T-s} f_{L,\sigma} (\tilde{B}(T-s), \tilde{E}(T-s)) \right] \tilde{\mathbb{Q}}(T_b\in ds, E(T_b)\in dv)\\
&= \int 1_{[s\le D]} e^{-\frac{\sigma^2}{8}v} \hat{C}_{d,i}^*(L,T-s, D, K) \,\tilde{\mathbb{Q}}(T_b\in ds, E(T_b)\in dv).
\end{split}
\end{equation*}
Similarly, for $A_2$ we have
\begin{equation*}
A_2 = \int 1_{[D<s\le T]} e^{-\frac{\sigma^2}{8} v} \mathbb{E}_{\tilde{\mathbb{Q}}}\left[ f_{L,\sigma}(\tilde{B}^*(T-s),\tilde{E}(T-s))\right] \tilde{\mathbb{Q}}(T_b\in ds, E(T_b)\in dv).
\end{equation*}
Lastly, to compute $A_3$ we note that $T_b>T$ iff $\tilde{T}_b > E(T)$. Now conditioning on the value of $E(T)$ we get
\begin{equation*}
\begin{split}
A_3 = \int_0^\infty \mathbb{E}_{\tilde{\mathbb{Q}}} \left[1_{\tilde{T}_b>v} f_{x,\sigma}(B(v), v) \right] \tilde{\mathbb{Q}}(E(T)\in dv).
\end{split}
\end{equation*}
Note that $\tilde{\mathbb{Q}}(\tilde{T}_b>v) = \tilde{\mathbb{Q}}(\sup_{s\le v} B(s) < b) = \tilde{\mathbb{Q}}(|B(v)|< b) = $ since $\sup_{s\le v} B(s) \overset{d}{=} |B(v)|$ for each fixed $u$.
However,
\begin{equation*}
\left( (B(s);s\in[0,v]) \big| \sup_{s\le v} B(s)< b\right) \overset{d}{=} \left( (b- Y(s); s\in[0,v]) | Y(0) = b\right),
\end{equation*} where $Y$ is a 3-dimensional Bessel process by William's path decomposition \cite{Williams.74}.
Thus
\begin{equation*}
A_3 = \int_0^\infty \left( 1- 2\Phi(v)\right) \mathbb{E}[f_{L,\sigma}(-Y(v),v)|Y(0) = b] \,\tilde{\mathbb{Q}}(E(T)\in dv).
\end{equation*}
\todo{We should write $A_1+A_2+A_3$ in terms of time-fractional PDEs.}
\subsubsection{The case $b<0$}
We note that if $b<0$, then in order for an excursion of $B^*$ to be below $b$ and last for duration $D$ before the maturity date, that excursion must start before time $T-D$. Again, we can condition on the value of $T_b$ and use Corollary \ref{cor:strongMarkovIndependence} to price the down-and-in call option as follows:
\begin{equation*}
\begin{split}
C_{d,i}^*&(x,T, L, D, K) = \mathbb{E}_{\tilde{\mathbb{Q}}} \left[ 1_{T_b\le T-D}1_{[H_{b,D}^-(B)\le D]} f_{x,\sigma}(B(E(T)),E(T))\right] \\
&= \mathbb{E}_{\tilde{\mathbb{Q}}} \left[ \mathbb{E}_{\tilde{\mathbb{Q}}} \left[ 1_{[H_{b,D}^-(B)\le D]} f_{x,\sigma}(B(E(T)),E(T))\big| \mathcal{F}^*_{T_b}\right] \right]\\
&= \mathbb{E}_{\tilde{\mathbb{Q}}}\left[ \mathbb{E}_{\tilde{\mathbb{Q}}}\left[1_{H_{b,D}^{-}(\tilde{B}) \le T-T_b}e^{-\frac{\sigma^2}{8} E(T_b)} f_{L,\sigma} (\tilde{B}^*(T-T_b), \tilde{E}(T-T_b)) \big|\mathcal{F}^*_{T_b}\right]\right]\\
&= \int 1_{[0\le s< T-D]} e^{-\frac{\sigma^2}{8} v}\\
&\qquad \qquad \times \mathbb{E}_{\tilde{\mathbb{Q}}} \left[1_{[H_{0,D}^-(\tilde{B})\le T-s]}f_{L,\sigma}(\tilde{B}^*(T-s), \tilde{E}(T-s)) \right] \tilde{\mathbb{Q}}(T_b\in ds, E(T_b)\in dv)\\
&= \int 1_{0\le s< T-D} e^{-\frac{\sigma^2}{8} v} \hat{C}_{d,i}^*(L,T-s, D,K) \tilde{\mathbb{Q}}(T_b\in ds, E(T_b)\in dv).
\end{split}
\end{equation*}
\subsection{The case $b = 0$}
We now want to price the down-and-in call option when the barrier $L$ is the initial stock price $x$ or, equivalently, when $b = 0$. With probability 1, there will eventually be an negative excursion (i.e. an excursion below zero) of $B^*$ of duration longer than $D$. This excursion occurs over the time interval $(g^*,d^*)$, where $g^* = \sup\{t< H_{0,D}^-({B}^*): B^*(t) = 0\}$ and $d^* = \inf\{t>H_{0,D}^-(B^*): B^*(t) = 0\}$.
We have three situations:
\begin{enumerate}
\item $g^*>T-D$ or, equivalently, $H_{0,D}^->T$; or
\item $g^*\le T-D$ and $d^*>T$; or
\item $d^*\le T$.
\end{enumerate}
In the first case, one can not exercise the call option.
In the second case, at time $T$ we have started an excursion of $B^*$ below level zero starting at time $g^*\le T-D$ and did not finish taht excursion before time.
\end{comment}
\section{Consequences of the Mapping Theorem}\label{sec:occupationChange}
A major reason why we could obtain some explicit results in Section \ref{sec:basicn*}, particularly in the $\beta$-stable subordinator case, is because we can relate an excursion of $R$ of duration $T$ to an excursion of $R^*$ of duration $T^*\overset{d}{=}S(T)$. As we will see in this section, this is essentially a consequence of the mapping theorem (Lemma \ref{lem:mappingTheorem}).
We now recall briefly the set-up of Theorem \ref{thm:timeChange1}. The process $X$ is a c\`{a}dl\`{a}g process and $S$ is an independent strictly increasing subordinator. We let $X^*(t) = X({E(t)})$ denote the time-changed process. Also recall the set-up of $\nu(I; -)$ and $\nu^*(I;-)$ in and around Theorem \ref{thm:timeChange1}, along with the notation $A(v)$ and $A^*(v)$.
\begin{proof}[Proof of Theorem \ref{thm:timeChange1}]
By conditioning on $(X(t);t\in[0,\tau]) = (w(t);t\in[0,T])$, which is independent of $S$ it suffices to show the following lemma:
\begin{lem}\label{lem:AvDeterministic}
Let $w:[0,T]\to \mathbb{R}$ be Borel measurable for some $T\in[0,\infty]$. Let $\nu^*(dv)$ and $\nu(dv)$ be the occupation measures of $(w(E(t));t\in[0,S(T)])$ and $w$ respectively. Define $A^*(v) = \nu^*(-\infty, v]$ and $A(v) = \nu(-\infty, v]$.
Then
\begin{equation*}
\left(A^*(v);v\in\mathbb{R} \right) \overset{d}{=} \left(S(A(v));v\in\mathbb{R} \right).
\end{equation*}
\end{lem}
\begin{figure*}
\centering
\begin{subfigure}[b]{0.4\textwidth}
\centering
\includegraphics[width=\textwidth]{./mappingTheorem0.png}
\caption{}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.4\textwidth}
\centering
\includegraphics[width=\textwidth]{./mappingTheorem1.png}
\caption{}
\end{subfigure}
\vskip\baselineskip
\begin{subfigure}[b]{0.4\textwidth}
\centering
\includegraphics[width=\textwidth]{./mappingTheorem2.png}
\caption{}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.4\textwidth}
\centering
\includegraphics[width=\textwidth]{./mappingTheorem3.png}
\caption{}
\end{subfigure}
\caption
{\small Visual representation of how the PRM $M^{\ast}$ is constructed from a path $w$ and a PRM $M$. The atoms of the PRM $N$ are depicted as $\times$'s in the four subfigures, where the atoms of the PRM $\tilde{M}$ are depicted as happening at time $v = w(t)$ whenever the horizontal line is at level $v$. Said another way, the $\times$'s contained in the green shaded area are atoms of $\tilde{M}|_{(-\infty,v]\times (0,\infty)}$.}.
\label{fig:prmTimeChange}
\end{figure*}
\end{proof}
\begin{proof}[Proof of Lemma \ref{lem:AvDeterministic}]
We note that
\begin{equation*}
\int_0^T g(w(t))\,dt = \int_\mathbb{R} g(a)\,\nu(da),\qquad\forall g\text{ bounded and measurable}.
\end{equation*}
Using the L\'{e}vy-It\^{o} decomposition of $S$, we now observe the simple description of the stochastic (Stieljes) integral with respect to $S$
\begin{align*}
\int_0^T g(w(t))\,dS(t) &= \kappa \int_0^T g(w(t))\,dt + \int_0^T \int_{(0,\infty)} g(w(t))\,y M(dt,dy)\\
&= \kappa \int_\mathbb{R} g(x)\,\nu(dx) + \int_\mathbb{R} \int_{(0,\infty)} g(x)y M^{\ast}(dx,dy),
\end{align*}
where $M$ is a PRM with intensity ${\operatorname{Leb}}\otimes \pi$ and $M^{\ast}$ is the pushforward measure
\begin{equation*}
M^{\ast} = (w\times \operatorname{id}_{(0,\infty)})_\# M\quad\text{under}\qquad(w\times \operatorname{id}_{(0,\infty)})(t,y) = (w(t),y).
\end{equation*} \begin{comment}An equivalent way of describing the measure $N^{\ast}$ is by
\begin{equation*}
N^{\ast}((a,b)\times A) = N(\{t\le T: a<w(t)<b\}\times A).
\end{equation*}
\end{comment}
By the mapping theorem, Lemma \ref{lem:mappingTheorem}, the measure $M^{\ast}$ is a Poisson random measure on $\mathbb{R}\times (0,\infty)$ with intensity $\nu\otimes\pi = (w\times \operatorname{id}_{(0,\infty)})_\# ({\operatorname{Leb}}|_{[0,T]}\otimes \pi)$.
By the change of variable for Stieljes integrals \cite{FT.2012}, we can write
\begin{equation*}
\int_0^T g(w(t))\,dS(t) = \int_0^{S(T)}g(w(E(u)))\,du,
\end{equation*} where $E$ is the right-continuous inverse of $S$ as in \eqref{eqn:Edef}. Therefore, the occupation measure $\nu^*$ of the time-changed path $w^*(u) = w(E(u))$ on the random time-interval $[0,S(T)]$ satisfies
\begin{equation}\label{eqn:PRMrepNu*}
\int_\mathbb{R} g(x)\,\nu^\ast(da) = \kappa \int_\mathbb{R} g(x)\,\nu(dx) + \int_\mathbb{R} \int_{(0,\infty)} g(x)y \,M^\ast (dx,dy),\qquad \forall g \in C_c(\mathbb{R}).
\end{equation}
Equation \eqref{eqn:PRMrepNu*} uniquely determines the occupation measure $\nu^*$ and hence $A^*$. In particular, we get
\begin{equation*}
A^*(v) = \kappa\nu(-\infty,v] + \int_{(0,\infty)} y \, M^* ((-\infty,v],dy).
\end{equation*}
Recall that $M^*(dx,dy)$ is a Poisson random measure on $\mathbb{R}\times (0,\infty)$ with intensity measure $\nu\otimes \pi$ but so is $\tilde{M}$ defined by $\tilde{M}((-\infty , v]\times B) = M((0,A(v)]\times B)$ as can be easily verified. It follows
\begin{equation*}
\left(A^*(v);v\ge 0\right) \overset{d}{=} \left( \kappa A(v) + \int_{(0,\infty)}y \tilde{M}((0,A(v)], dy; v\ge 0 \right)\overset{d}{=} \left(S(A(v));v\ge \right),
\end{equation*}
as desired.
\end{proof}
\begin{comment}
We prove this lemma by showing that the finite dimensional distributions agree. By using the independence of the PRMs on the disjoint sets $\{t: w(t)\in (a_i, b_i]\}$ for $a_1<b_1<a_2<\dotsm<a_n<b_n$, and the stationarity of L\'{e}vy processes it is sufficient to prove the following lemma.
\begin{lem}
Fix $a<b$. Then
\begin{equation*}
A^*(b) - A^*(a) \overset{d}{=} S({\nu(a,b]}).
\end{equation*}
\end{lem}
\begin{proof}
The drift term in the subordinator and the $\kappa \nu(a,b]$ term in $A^*$ can easily be handled since they are deterministic, so we only prove the case where $\kappa = 0$. The result follows from the well-known formula for Laplace functionals of PRMs \cite[Theorem 6.2.9]{Cinlar.11}
\begin{align*}
-\log \mathbb{E}\left[ \exp \left\{ -\lambda (A^*(b) - A^*(a))\right\} \right] &= -\log \mathbb{E} \left[\exp\left\{-\lambda \int_{(a,b]} \int_{(0,\infty)} y\,N^*(dx,dy)\right\} \right]\\
&= \int\limits_{(a,b] \times(0,\infty)} (1-e^{-\lambda y}) \,\nu(dx)\pi(dy)\\
&= \nu(a,b] \int_0^\infty (1-e^{-\lambda y})\,\pi(dy)\\
&= -\log \mathbb{E} \left[\exp\{-\lambda S(\nu(a,b])\} \right],
\end{align*} where in the second equality we used the fact that $N^*$ is a PRM with intensity measure $\nu\otimes \pi$.
This proves the desired equality in distribution.
\end{proof}
\end{comment}
\begin{comment}
\begin{proof}[Proof of Theorem \ref{thm:timeChange1}]
The proof is relatively straightforward application of Lemma \ref{lem:AvDeterministic}.
The main complication arises in the case when $A(x) = +\infty$ for some $x\in \mathbb{R}$ with positive probability. This occurs only if $\tau = +\infty$ and $X$ spends an infinite amount of time at or below below level $x$.
We let $V = \inf\{x\in \mathbb{R}: A(x) = \infty\} = \sup\{x\in \mathbb{R}: A(x)<\infty\}$ with the convention that $\sup \mathbb{R} = \inf\emptyset = +\infty$ and $\inf \mathbb{R} = \sup\emptyset = -\infty$. We note that $V$ depends only on the process $X$ and, consequently, it is independent of the subordinator $S$.
We note
\begin{equation*}
\begin{split}
A^*(v) &= \int_0^{\tau^*} 1_{[X^*(t)\le v]}\,dt\\
&= \int_0^\tau 1_{[X(t)\le v]} \,dS(t)\\
&= \kappa\int_{-\infty}^v \nu([0,\tau];\,dz) + \int_{t = 0}^\tau \int_{(0,\infty)} 1_{[X(t)\le x] }y N(dt,dy)
\end{split}
\end{equation*}
We first consider the case $V = -\infty$ with positive probability. In this case, for any $x\in \mathbb{R}$ and on the event $\{V = -\infty\}$
\begin{equation*}
\int_0^\infty 1_{[X^*(t)\le x]}\,dt = \int_0^\infty 1_{[X(t)\le x]}\,dS(t) = \kappa \int_{-\infty}^x \nu([0,\tau]; dz) + \int_{0}^\infty \int_{(0,\infty)} 1_{[X(t)\le x]}y\,N(dt,dy).
\end{equation*} However, since $S$ is strictly increasing either $\kappa>0$ or $\pi(0,\infty) = +\infty$. In the former case it is trivial to see $A^*(x) = +\infty$ for all $x$ on the event $\{V = -\infty\}$. In the later case we observe that there is some set bounded away from $0$ with positive $\pi$-mass. That is there is some $\varepsilon>0$ such that $\pi(\varepsilon,2\varepsilon) >0$.
Then
\begin{equation*}
\int_0^\infty \int_{(0,\infty)} 1_{[X(t)\le x]}y\,N(dt,dy) \ge \varepsilon N \left(\{t: X(t)\le x\}\times (\varepsilon,2\varepsilon)\right).
\end{equation*} The latter random variable is $\varepsilon$ times a Poisson random variable with mean $\nu([0,\tau];(-\infty,x]) \pi(\varepsilon,2\varepsilon) = \infty$. Thus that integral with respect to the PRM is identically $+\infty$. Thus $A^*(v) = +\infty = S_\infty$ for all $v$.
If $V \in \mathbb{R}$ with positive probability then the above proof works to show that $A^*(V) = +\infty$ on the event $\{V\in \mathbb{R}\}$. Since $A$ and $A^*$ are increasing almost surely $A^*(v) = S({A(v)}) = +\infty$ for all $v\ge V$. The proof below can work on the event $\{ V\in \mathbb{R}\}$ and for values of $v$ such that $v< V$.
We now consider the case $V\in(-\infty,\infty]$ and focus on $x<V$. We begin by observing that for $g\ge 0$ and bounded,
\begin{equation*}
\int_0^{\tau^*}g (X^*(t))\,dt = \int_0^\tau g(X(t))\,dS(t) \qquad \text{a.s.}
\end{equation*} by the change of variable formula for Stieljes measures, see \cite[Proposition 1]{FT.2012}.
Hence
\begin{align*}
\int_0^\tau g(X(t)) \,dS(t) &= \kappa \int_0^\tau g(x)\,\nu([0,\tau]; dx) + \int_\mathbb{R} \int_{(0,\infty)} g(x)y N^*(dx,dy),
\end{align*} where $N^*$ is the measure
\begin{equation*}
N^*(\omega;(a,b)\times A) = N\left(\omega; \{t\le \tau(\omega): X(t)(\omega)\in (a,b)\} \times A\right),
\end{equation*} where $N$ is the PRM in the construction of $S$.
We finish the proof by conditioning on $X$ restricted to $[0,\tau]$ and only focusing on the event $V =+\infty$ a.s. For each $\lambda >0$ and $a<b$ we have
\begin{align*}
\mathbb{E} \left[ \exp\left\{-\lambda(A^*(b) - A^*(a))\right\}\right] &= \mathbb{E}\left[\mathbb{E} \left[ \exp\left\{-\lambda(A^*(b) - A^*(a))\right\}\bigg| X,\tau \right] \right]\\
&= \mathbb{E}\left[\mathbb{E}\left[ \exp\left\{ -\lambda S(\nu([0,\tau];(a,b]))\right\}\bigg|X,\tau\right] \right]\\
&= \mathbb{E}\left[\exp\left\{-\lambda S\left({A(b)-A(a)}\right)\right\} \right],
\end{align*} which follows by the proof of Lemma \ref{lem:AvDeterministic}. Extending this to show finite dimensional marginals agree is a simple application of Independence of the PRMs $N$ and $N^*$ on disjoint sets.
\end{comment}
\section{A Ray-Knight Theorem}\label{sec:rayKnight}
The classical Ray-Knight theorems \cite[Chapter XI]{RY.99} describe the local time of a Brownian motion stopped at certain random times in terms of a piecewise concatenation of squared Bessel processes of dimensions 0, 2, and 4. Said in another way, the density of the occupation measure of Brownian motion at certain random time intervals is a concatenation of continuous state branching processes with immigration.
A continuous state branching process with immigration (${\operatorname{CBI}}$ process) $Z = (Z(v);v\ge 0)$ is a strong Markov process on $[0,\infty]$ uniquely characterized by two L\'{e}vy processes $Y^{(1)}$ and $Y^{(2)}$ where $Y^{(1)}$ has no negative jumps (i.e. is spectrally positive) and $Y^{(2)}$ is a subordinator independent of $Y^{(1)}$ \cite{KW.71}. Perhaps the simplest description of these processes comes from the Lamperti transform \cite{CLU.09,Lamperti.67} as generalized by Cabellero, P\'{e}rez Garmendia and Uribe Bravo in \cite{CPU.13}: A CBI process started at $x$ with branching mechanism $\psi$ and immigration rate $\hat{\phi}$ is the law of the unique c\`{a}dl\`{a}g solution to
\begin{equation}\label{eqn:lamperti}
Z(v) = x + Y^{(1)}\left(\int_0^v Z(u)\,du\right) + Y^{(2)}(v),\qquad v\ge 0
\end{equation} where $Y^{(1)}$ and $Y^{(2)}$ are L\'{e}vy processes with Laplace transforms
\begin{equation*}
\mathbb{E}[e^{-\lambda Y^{(1)}(t)}] = \exp\left\{t\psi(\lambda)\right\}\qquad \text{and}\qquad \mathbb{E}[e^{-\lambda Y^{(2)}(t)}] = \exp\left\{-t\hat{\phi}(\lambda)\right\}.
\end{equation*} We denote by ${\operatorname{CBI}}_x(\psi,\hat{\phi})$ the law of a such a process. The choice of the symbol $\hat{\phi}$ is to distinguish the subordinator $Y^{(2)}$ from $S$.
Let us now state the main result in this section.
\begin{thm}\label{thm:1d}
Suppose that $S$ is a strictly increasing subordinator with Laplace exponent $\phi$ with inverse $E$ in \eqref{eqn:Edef}. Let $X(t) = x + B(t) - \alpha t$ be an independent Brownian motion with negative drift started from $x\ge 0$. For $X^*(t) = X(E(t))$ and let $g:\mathbb{R}_+\to\mathbb{R}_+$ be a bounded Borel measurable function with compact support contained in $[0,a]$. Then
\begin{equation*}
\mathbb{E} \left[\exp\left\{-\int_0^{\tau^*} g(X^*(t))\,dt\right\} \right] = \exp \left\{ - 2\int_0^x u(s)\,ds\right\}, \qquad \tau^* = \inf\{t: X^*(t) = 0\}
\end{equation*} where $u$ is the unique non-negative solution to
\begin{equation}\label{eqn:udef}
u(r)+ 2\int_r^a (u(s)^2 + \alpha u(s))\,ds = \int_r^a \phi(g(s))\,ds
\end{equation}
\end{thm}
\begin{comment}
A similar result holds for 3-dimensional Bessel processes started from $x \ge 0$. That is, let $X$ be the unique strong solution to the stochastic differential equation written in integrated form:
\begin{equation*}
X(t) = x + B(t) + \int_0^t \frac{1}{X(s)}\,ds.
\end{equation*}
\begin{thm} \label{thm:3d}
Let $S$ be any strictly increasing subordinator with Laplace exponent $\phi$. Let $X$ be a 3-dimensional Bessel process started from $x\ge 0$, and let $g:\mathbb{R}_+\to \mathbb{R}_+$ be a Borel measurable function with compact support contained in $[0,K]$. Set $X^*(t) = X({E(t)})$. Then
\begin{equation*}
\mathbb{E}\left[ \exp \left\{ -\int_0^\infty g(X^*(t))\,dt\right\}\right] = \left\{ \begin{array}{ll} \displaystyle \frac{1}{x} \int_0^x \exp\{- \int_r^x 6 u(s)\,ds - 2\int_x^\infty u(s)\,ds \} \,dr &: x>0\\
\exp\{ - 2\int_0^\infty u(s)\,ds \} &: x = 0
\end{array} \right.,
\end{equation*} where
$u$ is the unique non-negative solution to
\begin{equation*}
u(r)+\int_r^K 2 u(s)^2\,ds = \int_r^K \phi( g(s) )\,ds.
\end{equation*}
\end{thm}
\end{comment}
The proof of Theorem \ref{thm:1d} below
requires the characterization on positive integral functionals of CBI processes described in, for example, \cite{Li.20}. Similar, although not as clean, representations should exist when $X$ is a Bessel processes of dimension $\delta\ge 2$. This is because of the explicit representations of the occupation measures of these processes in, for example, see \cite{Yor.91} and \cite[Chapter 4]{MY.08}. We suspect similar expressions should be possible when $X$ is a transient regular diffusion \cite{MR.03}.
\begin{comment}
\subsection{The Ciesielski-Taylor Identities}
The Ciesielski-Taylor identity \cite{CT.62} relates the distribution of the first time a of Brownian motion in dimension $\delta = 1,2,\dotsm$ exits the unit ball with the amount of time a Brownian motion in dimension $\delta+2$ stays within the sphere for all time. More compactly, let $\mathbb{P}^\delta$ denote the law of the norm of a Bessel process of dimension $\delta>0$ starting from $0$ and let $X$ denote the canonical process. Ciesielski and Taylor show
\begin{equation*}
\left( \inf\{t: X_t = a\} \text{ under the measure } \mathbb{P}^{\delta}\right) \overset{d}{=} \left(\int_0^\infty 1_{[X_s\le a]}\,ds \text{ under the measure } \mathbb{P}^{\delta+2} \right),\quad \delta =1, 2,\dotsm.
\end{equation*} Getoor and Sharpe \cite{GS.79} extended this to all $\delta>0$.
This has been extended in numerous directions. See \cite[Page 918]{KP.11} for a more detailed summary of this identity.
In general we'll say that two c\`adl\`ag processes $X$ and $Y$ satisfy a Ciesielski-Taylor type identity if
\begin{equation*}
\left(\inf\{t: X_t \ge a\} \right) \overset{d}{=}\left( \int_0^\infty 1_{[Y_t\le a]}\,dt \right),\qquad \forall a> 0.
\end{equation*}
An easy corollary of Theorem \ref{thm:timeChange1} is
\begin{cor}
If the pair $X$ and $Y$ satisfy a Ciesielski-Taylor type identity then $X^*_t = X_{E_t}$ and $Y^*_t = Y_{E_t}$ satisfy a Ciesielski-Taylor type identity where $E$ satisfies \eqref{eqn:Edef} for any strictly increasing subordinator $S$.
\end{cor}
\todo{I don't know if we should include this, but I think that we can extend these identities to this situation}
Branching processes with immigration appear as in the description of local times of many real-valued stochastic processes in Ray-Knight theorems and their generalizations \cite{DL.02, Duquesne.09}. In this section we describe some properties of subordinators time-changed by the integrals of CBI processes. Throughout this section a CBI process will be denoted by $Z = (Z_t;t\ge 0)$ and a process $S = (S_t;t\ge 0)$ will denote a subordinator. This includes possibly adding some superscripts to serve as an index, e.g. $Z^{(j)}$. Because CBI processes are uniquely determined by the Laplace transforms of a spectrally negative L\'{e}vy process and a subordinator, we will use $\hat{\phi}$ to represent the Laplace transform of the subordinator associated with the CBI process. That is we will say $Z\sim {\operatorname{CBI}}_{x}(\psi,\hat{\phi})$ where
\begin{equation} \label{eqn:202009082011}
\hat\phi(\lambda) = \hat\kappa \lambda + \int_0^\infty (1-e^{-\lambda x})\,\hat{\pi}(dx).
\end{equation}
We implicitly exclude the case $S_t = 0$ for all $t\ge 0$.
\end{comment}
\subsection{Time-changed subordinators as integrators}
Because of the distributional identity in Theorem \ref{thm:timeChange1}, we will focus on the processes $\beta = (\beta(v);v\ge 0)$ defined by
\begin{equation*}
\beta(v) = S\left({\int_0^v Z(u)\,du}\right) ,
\end{equation*} where $S$ is a strictly increasing subordinator with Laplace exponent $\phi$ and $Z\sim{\operatorname{CBI}}_x(\psi,\hat{\phi})$ is independent. We will denote the law of $\beta(v)$ on $\mathbb{D}$ by $\operatorname{SCBI}_x(\phi;\psi,\hat{\phi})$.
\begin{comment}
\begin{definition} The law $\operatorname{SCBI}_x(\phi; \psi,\hat\phi)$ on $\mathbb{D}(\mathbb{R}_+\to[0,\infty])$ is the law of the process $(\beta_v; v\ge 0)$ where
\begin{equation*}
\beta_v = S_{\int_0^v Z_s\,ds},
\end{equation*} where $S$ is a subordinator with Laplace exponent $\phi$ and $Z\sim {\operatorname{CBI}}_x(\psi,\hat\phi)$ which is independent of $S$.
\end{definition}
\end{comment}
Clearly, $\beta$ is non-decreasing since $S$ is strictly increasing and CBI processes are non-negative. Therefore, we can view $d\beta(v)$ as a random measure on $\mathbb{R}$. We will compute the Laplace transforms of a wide class of functionals involving the random measure $d\beta(v)$ explicitly.
\begin{comment}
The first is the additivity coming from branching processes:
\begin{lem}
Let $\beta^{(j)}\sim \operatorname{SCBI}_{x_j}(\phi;\psi, \hat{\phi}_j)$ for some $x_j\ge 0$ and $j = 1,2$ and be independent. Then
\begin{equation*}
\beta(v) = \beta^{(1)}(v) + \beta^{(2)}(v) ,\qquad \text{is distributed as }\quad \operatorname{SCBI}_{x_1+x_2}(\phi;\psi,\hat{\phi}_1+\hat{\phi}_2).
\end{equation*}
\end{lem}
\begin{proof}
We write
\begin{equation*}
\beta^{(j)}(v) = S^{(j)}\left(\int_0^v Z^{(j)}(u)\,du\right),\qquad j = 1,2
\end{equation*} where $S^{(1)},S^{(2)}$ are independent subordinators with common Laplace exponent $\phi$ and $Z^{(j)}\sim {\operatorname{CBI}}_{x_j}(\psi,\hat{\phi}_j)$ for $j =1,2$ and all of these processes are independent. A well-known property of branching processes with immigration is
\begin{equation*}
Z(v) = Z^{(1)}(v)+Z^{(2)}(v) \qquad \text{is distributed as }\quad Z\sim {\operatorname{CBI}}_{x_1+x_2}(\psi,\hat{\phi}_1+\hat{\phi}_2).
\end{equation*} This can be easily seen from the additivity properties of L\'{e}vy processes and the description in \eqref{eqn:lamperti}. It can also be derived from the representation of the Laplace transform in \eqref{eqn:ltOfCBI} described below.
From here and the stationarity of subordinators, for each fixed $v$,
\begin{align*}
S^{(1)}\left({\int_0^v Z^{(1)}(u)\,du}\right) + S^{(2)}\left(\int_0^v Z^{(2)}(u)\,du\right) &\overset{d}{=} S\left(\int_0^v (Z^{(1)}(u)+Z_u^{(2)}(u))\,du\right) \\
&\overset{d}{=} S\left(\int_0^v Z(u),du\right) ,
\end{align*} where $S\overset{d}{=} S^{(1)}\overset{d}{=} S^{(2)}$ is independent of the CBI processes. Thus the one-dimensional marginals agree. A similar derivation can show that the finite dimensional distributions of $\beta^{(1)}(v)+\beta^{(2)}(v)$ agree with the finite dimensional distributions of $\operatorname{SCBI}_{x_1+x_2}(\psi;\psi, \hat{\phi}_1+\hat{\phi}_2)$.
\end{proof}
This immediately implies
\begin{cor}
The law $\operatorname{SCBI}_x(\phi;\psi,\hat{\phi})$ is infinitely divisible.
\end{cor}
\begin{proof}
If $\hat{\phi}$ is the Laplace transform of a subordinator, so is $\frac{1}{n}\hat{\phi}$.
\end{proof}
The above corollary is a simpler case of a more general results about infinitely divisible distributions $\mathbb{D}(\mathbb{R}_+,\mathbb{R})$. See for example, \cite[Theorem 7.1]{BNMS.06} or \cite[Theorem 3.6]{EKO.08}.
\end{comment}
\begin{comment}
In this section we describe the properties of functionals of the form
\begin{equation*}
\int_0^\infty g(v)\,d\beta(v),\qquad g\text{ bounded Borel measurable with compact support}.
\end{equation*} and derive explicit formulas for their Laplace transforms.
\end{comment}
We start by describing the Laplace transform of integral functionals of ${\operatorname{CBI}}$ processes. It will be convenient to start CBI processes at an arbitrary positive time $r\ge 0$. We will denote the law of such a process by ${\operatorname{CBI}}_{r,x}(\psi,\hat{\phi})$. That is $(Z(t); t\ge r)\sim{\operatorname{CBI}}_{r,x}(\psi,\hat{\phi})$ if and only if $(Z(t-r); t\ge r)\sim {\operatorname{CBI}}_x(\psi,\hat{\phi})$. Throughout this section we fix some branching mechanism $\psi$ and immigration rate $\hat{\phi}$, and we will denote by $\mathbb{P}_{r,x}$ (resp. $\mathbb{E}_{r,x}$) the law of (resp. the expectation with respect to) a ${\operatorname{CBI}}_{r,x}(\psi,\hat{\phi})$ process $Z$.
We state the following lemma and outline a proof.
\begin{lem}\label{lem:LiLemma}
Let $\mu$ be a Radon measure and let $\lambda$ be a bounded non-negative Borel measurable function. Then
\begin{equation}\label{eqn:ltOfCBI}
\mathbb{E}_{r,x} \left[\exp\left\{-\int_r^t \lambda(s) Z(s)\,\mu(ds) \right\} \right] = \exp\left\{- xu(r) - \int_r^t \hat{\phi}(u(s))\,ds \right\},
\end{equation} where $u:[0,t]\to \mathbb{R}_+$ is the unique bounded non-negative solution to
\begin{equation*}
u(r)+ \int_r^t \psi(u(s))\,ds = \int_r^t \lambda(s)\mu(ds) \qquad r\in[0,t].
\end{equation*}
\end{lem}
In the case $\hat{\phi} = 0$, the above lemma can be found in \cite{Li.20} where it is proved by approximating $\lambda(s)\mu(ds)$ by discrete measures. This lemma follows from the same approximation procedure. We just establish the initial lemma, and omit the rest of the proof.
\begin{lem}\label{lem:11}
Let $\lambda_j \ge 0$ and let $0\le t_1<\dotsm< t_n<\infty$. Then
\begin{equation*}
\mathbb{E}_{r,x}\left[ \exp\left\{-\sum_{j=1}^n \lambda_j Z({t_j}) 1_{[r\le t_j]}\right\} \right] = \exp\left\{ -x u(r) - \int_r^{t_n} \hat{\phi}(u(s))\,ds \right\} ,
\end{equation*} where $u$ is the unique bounded non-negative solution to
\begin{equation*}
u(r) + \int_r^t \psi(u(s))\,ds = \sum_{j=1}^n \lambda_j 1_{[r\le t_j]}.
\end{equation*}
\end{lem}
\begin{proof}
The follow the approach in \cite{Li.20} and use the Markov property and induction. This holds for $n = 1$ \cite{KW.71}. By induction we suppose that this holds for any collection of $n-1$ many $\lambda_j$'s and $t_j$'s. Since any $t_j\in[0,r)$ play no role in either the expectation nor the desired claim, we can without loss of generality assume $0\le r\le t_1<t_2<\dotsm<t_n$.
We have by the strong Markov property for $Z$ and the induction hypothesis
\begin{align*}
\mathbb{E}_{r,x}\left[\exp\left\{-\sum_{j=1}^n \lambda_j Z(t_j) \right\} \right] &= \mathbb{E}_{r,x}\left[ \mathbb{E}_{r,x}\left[ \exp\left\{-\lambda_1Z(t_1) - \sum_{j=2}^n \lambda_j Z(t_j) \right\} \Bigg| Z(t_1) \right]\right]\\
&= \mathbb{E}_{r,x} \left[ e^{-\lambda_1 Z(t_1)} \mathbb{E}_{r,x} \left[ \exp \left\{-\sum_{j=2}^n \lambda_j Z(t_j) \right\} \Bigg| Z(t_1)\right] \right]\\
&= \mathbb{E}_{r,x} \left[ e^{-\lambda_1 Z(t_1)} \mathbb{E}_{t_1, Z(t_1)} \left[\exp\left\{-\sum_{j=2}^n \lambda_j Z(t_j) \right\} \right] \right]\\
&= \mathbb{E}_{r,x} \left[\exp\left\{-\lambda_1 Z(t_1) - w(t_1)Z(t_1) - \int_{t_1}^{t_n} \hat{\phi}(w(s))\,ds\right\} \right],
\end{align*} where $w$ is the unique non-negative solution to
\begin{equation*}
w(r)+ \int_r^t \psi(w(s))\,ds = \sum_{j=2}^n \lambda_j 1_{[r\le t_j]},\qquad r\in [0,t_n].
\end{equation*}
However,
\begin{equation*}
\begin{split}
\mathbb{E}_{r,x} &\left[\exp\left\{-\lambda_1 Z(t_1) - w(t_1)Z(t_1) - \int_{t_1}^{t_n} \hat{\phi}(w(s))\,ds\right\} \right] \\
&= \exp\left\{ - x\tilde{u}(r)- \int_r^{t_1}\hat{\phi}(\tilde{u}(s))\,ds - \int_{t_1}^{t_n} \hat{\phi}(w(s))\,ds \right\},
\end{split}
\end{equation*} where $\tilde{u}$ is the unique bounded non-negative solution to
\begin{equation*}
\tilde{u}(r)+ \int_r^{t_1} \psi(\tilde{u}(s))\,ds = (\lambda_1 + w(t_1)) 1_{[r\le t_1]},\qquad r\in [0,t_1].
\end{equation*} It is easy to check that the desired claim holds with
\begin{equation*}
u(s) = \left\{
\begin{array}{ll}
\tilde{u}(s) &: s\in[0,t_1]\\
w(s) &: s\in[t_1,t_n]
\end{array}
\right..
\end{equation*}
\end{proof}
Now to describe the process $\beta(v)$ as an integrator, we will set $V = (V_t;t\ge 0)$ as the right-continuous inverse of $v\mapsto \int_0^v Z(u)\,du$. That is
\begin{equation*}
V_t = \inf\left\{v: \int_0^v Z(u)\,du>t\right\}.
\end{equation*}
A change-of-variable formula for Stieljes measures \cite[Proposition 0.4.10]{RY.99} yields the following lemma, the proof of which is omitted.
\begin{lem}\label{lem:extraLem}
Suppose that $\hat{\phi}$ is the Laplace transform of a strictly increasing subordinator. Then
\begin{equation}\label{eqn:changeOfVariable1}
\int_0^\infty g(v)\,d\beta(v) {=} \int_0^\infty g(V_t)\,dS(t),\qquad \forall \text{ bounded Borel $g$ with compact support}.
\end{equation}
\end{lem}
If the equality in \eqref{eqn:changeOfVariable1} holds, we can use Lemma \ref{lem:LiLemma} and write for each bounded Borel measurable $g$ with compact support
\begin{align*}
\mathbb{E}&\left[\exp \left\{- \int_0^\infty g(v)\,d\beta(v) \right\} \right]
= \mathbb{E} \left[ \exp \left\{- \int_0^\infty g(V_t)\,dS(t) \right\}\right]\\
&= \mathbb{E} \left[\mathbb{E}\left[\exp\left\{-\int_0^\infty g(V_t)\,dS(t) \right\} \Bigg| (Z(t);t\ge 0)\right] \right]\\
&= \mathbb{E}\left[\exp\left\{-\int_0^\infty \phi(g(V_t))\,dt \right\} \right]\\
&= \mathbb{E}\left[\exp\left\{-\int_0^\infty \phi(g(t)) Z(t)\,dt \right\} \right]\\
&= \exp\left\{-x u(0) - \int_0^t \hat{\phi} (u(s))\,ds \right\},
\end{align*} where $u$ is the unique bounded non-negative solution to
\begin{equation*}
u(r)+ \int_r^\infty \psi(u(s))\,ds = \int_r^\infty \phi(g(s))\,ds.
\end{equation*} For the Laplace transform of stochastic Stieljes integrals involving a subordinator as the integrator - as was used in third equality above - we used the exponential martingale formula as found in, for example, see Section 5.2 of Applebaum's monograph \cite{Applebaum.09}. In the fourth equality we used the change-of-variable for Stieljes measures again.
Combined with Lemma \ref{lem:extraLem}, we have proved the following:
\begin{thm} \label{thm:ltOfIntegralsOfCBI}
Suppose that $g$ is a bounded Borel measurable function with compact support. Let $\beta\sim \operatorname{SCBI}_x(\phi;\psi,\hat{\phi})$, where $\hat{\phi}$ is either zero or the Laplace transform of a strictly increasing L\'{e}vy process.
Then
\begin{equation*}
\mathbb{E}\left[ \exp\left\{-\int_0^\infty g(v)\,d\beta(v) \right\}\right] = \exp\left\{-xu(0)-\int_0^\infty \hat{\phi}(u(s))\,ds \right\},
\end{equation*} where $u$ is the unique non-negative solution to
\begin{equation*}
u(r)+\int_r^\infty \psi(u(s))\,ds = \int_r^\infty \phi\circ g(s)\,ds.
\end{equation*}
\end{thm}
\subsection{Proof of Theorem \ref{thm:1d}}
\begin{proof}[Proof of Theorem \ref{thm:1d}]
Throughout this section we suppose that $X = (X(t);t\ge 0)$ is a Brownian motion with negative drift started from level $x\ge 0$; i.e. $X(t) = x+ B(t) -\alpha t$ for some $\alpha\ge 0$. Let $\tau = \inf\{t: X(t) = 0\}$. A result of Warren \cite{Warren.97} states the occupation measure of $X$ on the interval $[0,\tau]$ satisfies
\begin{equation}\label{eqn:occupationTime}
\int_0^\tau g(X(s))\,ds = \int_0^\infty g(v) L(v)\,dv,
\end{equation} where $L(v)$ is a weak solution to the stochastic differential equation
\begin{equation*}
L(v) = 2(x\wedge v) + 2\int_0^v \sqrt{L(y)}\,dW(y) - 2\alpha \int_0^v L(y)\,dy.
\end{equation*} The process $L$ is a concatenation of two Markov process:
\begin{equation}\label{eqn:LinTwoZs}
L(v) = \left\{
\begin{array}{ll}
Z^{(1)}(v) &: v\in[0,x] \\
Z^{(2)}(v) &: v\ge x
\end{array}
\right.
\end{equation} where $(Z^{(1)}(v); v\ge 0)$ is a ${\operatorname{CBI}}_0(\psi,\hat{\phi})$ process and, conditionally on $Z^{(1)}(x) = y$, the process $(Z^{(2)}(v); v\ge x)\sim {\operatorname{CBI}}_{x,y}(\psi,0)$ with
\begin{equation*}
\psi(\lambda) = 2 \lambda^2 + 2\alpha \lambda,\qquad \hat{\phi}(\lambda) = 2\lambda.
\end{equation*}
\begin{comment}Recall that ${\operatorname{CBI}}_{r,y}(\psi,\hat{\phi})$ is a ${\operatorname{CBI}}_y(\psi,\hat{\phi})$ process started at time $r\ge 0$.
Indeed, an application of the Dambis-Dubins- Scharwz theorem \cite[Theorem V.1.6]{RY.99} implies that
$L(v)$ is a weak solution to
\begin{equation*}
L(v) = 2 \tilde{X}\left({\int_0^v L(s)\,ds}\right) + 2 (x\wedge v), \qquad\left( \tilde{X}(t);t\ge 0 \right)\overset{d}{=}\left( W(t) - \alpha t; t\ge 0 \right).
\end{equation*} Applying the Lamperti transform result of \cite{CPU.13} in equation \eqref{eqn:lamperti}, and taking the Laplace transform of a Brownian motion with drift gives the desired formula for $\psi$.
\end{comment}
We can write by conditioning on the paths of $X$
\begin{equation*}
\mathbb{E}\left[\exp(-\int_0^{\tau^*} g(X^*(s))\,ds) \right] = \mathbb{E}\left[\exp(-\int_0^{\tau} g(X(t))\,dS(t) )\right] = \mathbb{E}\left[\exp(-\int_0^\tau \phi\circ g(X(t))\,dt ) \right].
\end{equation*}
However, almost surely
\begin{equation*}
\int_0^\tau h(X(s))\,ds = \int_0^\infty h(v)\, L(v)\,da, \qquad \forall h\ge 0 \quad \text{Borel}.
\end{equation*}
The rest of the proof follows as in Lemma \ref{lem:LiLemma}. Indeed, note that $h(v) = \phi(g(v))$ has support contained in $[0,a]$ since $\phi(0) = 0$. Letting $\mathcal{F}_s = \sigma(L(u); u\le s)$ and using the decomposition in \eqref{eqn:LinTwoZs} in terms of two Markov processes we get:
\begin{equation*}
\begin{split}
\mathbb{E}&\left[ \exp(-\int_0^\infty h(v)\, L(v)\,dv)\right] = \mathbb{E}\left[\exp(-\int_0^x h(v)L(v)\,dv) \mathbb{E}\left[ \exp(-\int_x^\infty h(v) L(v)\,dv) \bigg| \mathcal{F}_x\right] \right]\\
&= \mathbb{E}_{0,0} \left[\exp(-\int_0^x h(v) Z^{(1)}(v)\,dv) \mathbb{E}_{x,Z^{(1)}(x)} \left[\exp\left(-\int_x^\infty h(v) Z^{(2)} (v)\,dv \right) \right] \right]\\
&= \mathbb{E}_{0,0}\left[ \exp\left\{ -\int_0^x h(v) Z^{(1)}(v)\,dv - Z^{(1)}(x) w(x)\right\} \right]
\end{split}
\end{equation*}
where $w$ is the unique bounded solution to
\begin{equation*}
w(r) + \int_r^a (2 w(s)^2+ 2\alpha w(s))\,ds = \int_r^a h(s)\,ds.
\end{equation*}
Now using a similar argument to proof of Lemma \ref{lem:11}, we can write:
\begin{equation*}
\mathbb{E}_{0,0}\left[ \exp\left\{ -\int_0^x h(v) Z^{(1)}(v)\,dv - Z^{(1)}(x) w(x)\right\} \right] = \exp\left\{ -\int_0^x 2 u(s)\,ds\right\}
\end{equation*} where $u$ solves \eqref{eqn:udef}.
\end{proof}
\begin{comment}
\subsubsection{In 3 dimensions, and the Proof of Theorem \ref{thm:1d}} In this section, we will change notation slightly because we will be using a particular path decomposition. We use the following notation:
\begin{equation*}
\int_0^T g(Y(s))\,ds = \int_\mathbb{R} g(a) L_a(Y;T)\,da.
\end{equation*} Each stochastic process we consider in this subsection will possess sufficiently regular occupation measures, that is the occupation measures will have densities with respect to the Lebesgue measure
We now recall William's path decomposition for 3-dimensional Bessel processes \cite[Theorem VI.3.11]{RY.99} which gives a different way to construct a 3-dimensional Bessel processes.
Let $B(t)$ denote a Brownian motion on the line, started from a point $x >0$, let $U$ be a uniform random variable on $[0,1]$, and let $R(t)$ denote a 3-dimensional Bessel process started from $0$. Define the process $X(t)$ by
\begin{equation*}
X(t) = \left\{
\begin{array}{ll}
B(t)&: t\le \tau\\
xU + R({t-\tau}) &: t \ge \tau
\end{array}
\right., \qquad\text{where} \quad \tau = \inf\{t: B(t) = xU\}.
\end{equation*}
Then $X$ is a 3-dimensional Bessel process started from $x>0$, and its global minimum is $xU$.
Consequently, we can write the following
\begin{equation*}
L_a(X;\infty) = L_a(B; \tau) + L_{a-xU} (R; \infty).
\end{equation*}
Using time-reversal of Brownian motion, and the duality relationships of squared Bessel processes \cite{PY.82} and the classical Ray-Knight theorems \cite[Chapter XI]{RY.99}, one can prove a decomposition of $L_a(B;\tau)$ into squared Bessel processes of dimensions 0 and 4. Namely,
\begin{equation*}
\left(\left(L_{a}(B; \tau);a\ge 0\right)\bigg| xU = y \right) \overset{d}{=} \left(\tilde{Z}(a);a\ge0 \right),
\end{equation*} where $\tilde{Z}$ can be decomposed as
\begin{equation*}
\tilde{Z}(a) = \left\{
\begin{array}{ll}
0&:a\le y\\
\tilde{Z}^{(1)}(a) &: y\le a \le x\\
\tilde{Z}^{(2)}(a) &: a\ge x
\end{array}
\right. ,
\end{equation*} where $\tilde{Z}^{(1)}$ and $\tilde{Z}^{(2)}$ are conditionally independent given their common value $\tilde{Z}^{(1)}(x) = \tilde{Z}^{(2)}(x)$, and $\tilde{Z}^{(1)}\sim {\operatorname{CBI}}_{y,0} (\psi, 4\lambda)$ and conditionally on $Z^{(1)}(x) = z$, the process $\tilde{Z}^{(2)}\sim {\operatorname{CBI}}_{x,z}(\psi,0)$, where
\begin{equation*}
\psi(\lambda) = 2\lambda^2.
\end{equation*} Similarly, \cite{MY.08}, we can see
\begin{equation*}
\left(L_{a}(R;\infty); a\ge 0\right) \sim {\operatorname{CBI}}_{0,0}(\psi,2\lambda).
\end{equation*}
Using additivity of squared Bessel processes \cite[Section XI.1]{RY.99} we can write
\begin{equation*}
\left(\left(L_a(X;\infty);a\ge 0\right)\big|\inf_{t \ge 0} X(t) = r\right) \overset{d}{=} \left(Z(a); a\ge 0 \right),
\end{equation*} where $Z$ is the concatenation
\begin{equation*}
Z(a)= \left\{
\begin{array}{ll}
0&:a\le r\\
{Z}^{(1)}(a) &: r\le a \le x\\
{Z}^{(2)}(a) &: a\ge x
\end{array}
\right. ,
\end{equation*} where $Z^{(1)}$ and $Z^{(2)}$ are conditionally independent given $Z^{(1)}(x) = Z^{(2)}(x)$, and $Z^{(1)} \sim {\operatorname{CBI}}_{r,0}(\psi,6\lambda)$ and, conditionally on $Z^{(1)}(x) =z$, $Z^{(2)}\sim {\operatorname{CBI}}_{x,z}(\psi,2\lambda)$, where $\psi(\lambda) = 2\lambda^2$. Lastly, by William's decomposition the global infimum of $X(t)$ is uniformly distributed on $[0,x]$ when $X(0) = x$.
To see how Theorem \ref{thm:3d} follows, we note that for $Z$ defined by the concatenation above and for $g$ a bounded positive measurable function with support contained in $[0,K]$ for $K>x$:
\begin{align*}
\mathbb{E}&\left[ \exp\left\{-\int_0^\infty g(a) Z(a)\,da \right\} \right] \\
&= \mathbb{E}\left[ \mathbb{E}\left[ \exp\left\{ -\int_0^\infty g(a) Z(a) \,da \right\} \bigg| Z^{(1)}\right] \right]\\
&= \mathbb{E}\left[ \exp\left\{ - \int_0^x g(a) Z^{(1)}(a)\,da \right\} \mathbb{E}_{x, Z^{(1)}(x)} \left[\exp\left\{-\int_x^K g(a)Z^{(2)}(a)\,da \right\} \right] \right]\\
&= \mathbb{E}\left[\exp\left\{-\int_0^x g(a) Z^{(1)}(a)\,da - Z^{(1)}(x) w(x) - \int_x^K 2 w(s)\,ds \right\} \right],
\end{align*} where $w$ is the unique bounded non-negative solution to
\begin{equation*}
w(x)+\int_x^t 2w(s)^2\,ds = \int_x^t g(s)\,ds
\end{equation*} by Lemma \ref{lem:LiLemma}. In turn,
\begin{align*}
\mathbb{E}&\left[\exp\left\{-\int_0^x g(a) Z^{(1)}(a)\,da - Z^{(1)}(x) w(x) - \int_x^K 2 w(s)\,ds \right\} \right]\\
&= \exp\left\{- \int_r^x 6u(s)\,ds - \int_x^K 2w(s)\,ds \right\},
\end{align*} where $u$ is the unique bounded non-negative solution to
\begin{equation*}
u(r) + \int_r^x 2 u(s)^2\,ds = \int_r^x g(s)\,ds + w(x).
\end{equation*} This means that
\begin{equation*}
u(r) + \int_r^x2u(s)^2\,ds = \int_r^x g(s)\,ds + \int_x^K g(s)\,ds - \int_x^K 2w(s)^2\,ds.
\end{equation*} Defining $u(s) = w(s)$ for $s\in[r,x]$ gives
\begin{equation*}
\mathbb{E}\left[ \exp\left\{-\int_0^\infty g(a)Z(a)\,da \right\} \right] = \exp\left\{- 6\int_r^x u(s)\,ds -2 \int_x^K u(s)\,ds \right\},
\end{equation*} where $u$ is the unique non-negative bounded solution to
\begin{equation*}
u(r) + \int_r^K 2 u(s)^2\,ds = \int_r^K g(s)\,ds.
\end{equation*}
The rest of the proof of Theorem \ref{thm:3d} follows from the change-of-variable formula for Stieljes measures and conditioning on the uniformly distributed on $[0,x]$ random variable $\inf X$. The details are omitted.
\end{comment}
\section*{Acknowledgements}
DC is partially supported by NSF grant DMS-1444084 and NSF grant DMS-205223.
\bibliographystyle{imsart-number}
|
\section{Introduction} \label{sec:intro}
\IEEEPARstart{I}{N} the last decade, residential rooftop photovoltaic (PV) has been proliferating in distribution systems. In most cases, utilities only install a bi-directional smart meter to record the net demand of customers with PVs. This type of installation is referred to as behind-the-meter (BTM), in which case the net demand equals native demand minus PV generation. Therefore, the PV generation produced by solar array and the native demand consumed by appliances are unknown to utilities. Only metering the net demand can reduce the financial cost for utilities; however, as the penetration level of PV increases, the unobservability of notable PV generation and native demand brings significant challenges to distribution systems. We focus on three specific applications to elaborate the necessity of estimating the unknown BTM PV generation and native demand: \textit{First}, the unavailability of native load and PV generation might cause unacceptable forecasting errors because some forecasters require reconstituting the generation and native demand time series \cite{forecasting_overview, kangping_recom_1}. In contrast, knowing BTM PV generation and native load can help utilities forecast generation and load separately, thus provide utilities useful information regarding load/generation growth. \textit{Second}, the invisibility of PV generation and native load can hinder designing optimal service restoration plans \cite{PV_handbook, distribution_hand_book}. During the restoration stage after an outage, the native demand might be several times higher than the pre-outage demand due to the simultaneous restarting of a large number of air-conditioning appliances. This anomalous demand should be estimated for optimal restoration plans because it can damage electric devices when simultaneously restoring a large number of customers. In practice, utilities multiply the normal native demand before outage by a ratio to estimate the anomalous demand during restoration. Also, utilities typically do not consider PVs as reliable restoration sources \cite{PV_handbook}. Therefore, separating normal native demand and generation is needed for estimating the restoration demand. \textit{Third}, the unobservability of native demand and solar generation might cause inaccurate reliability analysis. When evaluating a transmission system's reliability, each distribution system is generally simplified as a bus whose native load duration curve is constructed \cite{7078837,planning_guidebook}. For those utilities with a high-penetration PV integration, directly using the net demand to construct the load duration curve can significantly underestimate the actual native load \cite{rui_cheng_1}. This is because the net demand is typically smaller than the native demand due to the existence of PV generation. In contrast, using the native demand separated from the net demand can help construct more accurate load duration curves. In summary, disaggregating BTM PV generation and native demand from the recorded net demand can enhance distribution system observability and awareness and can also provide more accurate information for transmission system reliability analysis.
Previous works on BTM PV generation disaggregation can be categorized into two types: \textit{Type I - Model-based approaches:} PV array performance model is employed to represent physical PV arrays. In \cite{SunDance}, a PV model is combined with a clear sky model to estimate customer-level solar generation. In \cite{Yi_Wang}, a virtual equivalent PV station model is utilized to represent the aggregate generation of BTM PVs within a region. In \cite{YC_Zhang} and \cite{nan_peng_yu_2}, a physical PV model and a statistical model are utilized to estimate BTM solar generation and native demand, respectively. One primary disadvantage of these model-based approaches is that detailed PV array parameters or accurate meteorological data are required. However, in practice, these parameters are typically unavailable to utilities. Also, acquiring meteorological data might cause additional costs to utilities. \textit{Type II - Model-free approaches:} In \cite{kangping_li_solar} and \cite{ PV_cap_estimation}, net demands under heterogeneous weather conditions are employed to estimate BTM PV capacity, which is then multiplied by a standard solar power time series to infer BTM PV generations. In \cite{applied_energy}, native demand and PV generation are estimated using 1-second net demand data by identifying appliances' states, which are then leveraged to estimate appliance demands and solar power. Based on the variation difference between load and solar power, in \cite{indu_informatics}, an approach is proposed for estimating service transformer-level PV generation. In \cite{Hamid_Shaker}, regional-level generation is estimated by installing additional sensors to record typical PV generation profiles. In \cite{CSSS_feeder}, feeder-level solar generation is estimated by utilizing net load measurements and a nearby PV farm’s generation readings. Using known native loads for customers without PVs and the generations for a limited number of observable PVs, in \cite{kangping_li_new}, the authors formulate an optimization process to estimate the aggregated native load and PV generation. In \cite{9548947}, a federated learning-based framework is proposed to probabilistically estimate community-level BTM solar generation. In \cite{Samuel_Talkington}, an approach is developed to estimate the reactive power by taking advantage of the correlation between the weekly nighttime and daytime native reactive power demands. Furthermore, previously in \cite{Fankun_Bu} and \cite{Fankun_Bu_1}, we have proposed two approaches for estimating the unknown BTM generation using measured solar power exemplars. One primary shortcoming of the model-free approaches is that they rely on contextual information, i.e., recorded solar power exemplars or meteorological data, which might bring additional costs to utilities.
Considering the shortcomings of previous approaches, this paper proposes a novel BTM PV generation and native demand estimation framework which does not require \textit{previously} recorded solar power and meteorological measurements. Our approach is based on two findings from real data. \hl{The first finding is the spatial correlation of native load, i.e., the native demands of two sizeable residential customer groups are strongly correlated and have highly homogeneous shapes. The second finding is the spatial correlation of solar power generation, i.e., the generations for two PVs in a distribution system are significantly correlated and have highly similar profiles.
Our proposed approach contains \textbf{two} layers: (1) At the aggregate level, the total generation of all BTM PVs is estimated by leveraging our first finding. (2) At the customer level, utilizing our second finding, the estimated aggregate BTM PV generation is allocated to individual customers. Utilizing the two findings improves our approach's robustness against the customer-level load uncertainty} \cite{load_uncertainty}.\hl{ The second layer contains three steps: first, our approach trains a model to produce multiple candidate generation time series, using solar power data generated by a publicly available tool. Second, our approach determines the peak generation for each PV. } Finally, the allocating procedure is formulated as an optimization problem. The overall structure of our proposed approach is shown in Fig. \ref{fig:overall_structure}. This paper has verified our proposed approach using real hourly native demand and PV generation data \cite{data_source}.
\hl{Smart meters can record individual customers' demands at an interval of one hour or shorter. Such fine-grained temporal and spatial granularity can give us more details than traditional monthly bills. Many researchers have developed advanced approaches to mine useful information from smart meter data. For example,} \cite{smart_meter_data_1} \hl{utilizes smart meter measurement to perform state estimation for enhancing distribution system observability, }\cite{smart_meter_data_2} \hl{employs water consumption data recorded by smart water meters to train aggregate water demand forecasters,} \cite{smart_meter_data_3} \hl{utilizes high-resolution phasor measurement units' data to conduct false data detection, and data redundancy strengthening, }\cite{smart_meter_data_4} \hl{converts smart meter data into manageable load profiles via linearizing load patterns. Our proposed approach takes advantage of smart meter data's temporal and spatial granularity to perform BTM generation estimation. }
\hl{The main contributions of our paper are summarized as follows: (1) This paper proposes an approach that does not rely on PV array parameters, historical meteorological data, or pre-recorded generation exemplars. This independence can significantly improve the viability of our approach because acquiring the above three types of information can bring challenges or additional costs for utilities. (2) Our approach only relies on the net demands of customers with PVs and the native demands of customers without PVs for estimating the aggregate-level PV generation. These two types of demands - net and native - are typically available to utilities, making our approach significantly practical. (3) Our approach innovatively estimates individual PV-installed customers' peak generations by mining net demand data. The peak generations are then utilized to estimate individual PV-installed customers' BTM generation time series. }
Throughout the paper, vectors are denoted using bold italic letters, and matrices are represented as bold non-italic letters. In addition, we adopt the sign convention that the native demand consumed by customers and the power output from PVs are both positive.
\begin{figure}
\centering
\includegraphics[width=1\linewidth]{figure/overall_framework.pdf}
\caption{\hl{Overall structure of the proposed BTM PV generation estimation approach.}}
\label{fig:overall_structure}
\end{figure}
The rest of the paper is organized as follows: Section \ref{sec:findings} introduces our first and second findings regarding spatial correlation of native demand/generation. Section \ref{sec:aggre_level} presents how we estimate the aggregate generation for customers with PVs. Section \ref{sec:indiv_level} presents the procedure of formulating and solving an optimization problem to allocate the estimated aggregate generation to individual PVs. In Section \ref{sec:case_study}, case studies are analyzed. Section \ref{sec:conclusion} concludes the paper.
\section{Spatial Correlation of Native Demand/PV Generation}\label{sec:findings}
\subsection{Finding 1: Native Demand Spatial Correlation between Two Sizeable Groups}\label{sec:finding_1}
By examining real residential native demand data, we find that once the customer numbers for two groups reach a certain level, their native demands are highly correlated. This finding is leveraged for estimating the \textit{aggregate} native demand time series for customers with PVs.
\begin{figure}
\centering
\includegraphics[width=0.84\linewidth]{figure/native_load_curves.pdf}
\caption{Three-day actual native demand curves for three example groups with different customer numbers.}
\label{fig:native_load_curves}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.84\linewidth]{figure/normal_native_load_curves.pdf}
\caption{Three-day normalized native demand curves for three example groups with different customer numbers.}
\label{fig:norm_native_load_curves}
\end{figure}
Specifically, we use native demand curves to illustrate the observed spatial correlation. Fig. \ref{fig:native_load_curves} presents real native demand curves for three example groups with different customer numbers, i.e., 40, 60, and 80, respectively. We can observe that these three curves demonstrate almost identical shapes, although they have different magnitudes. The high shape similarity can also be corroborated by Fig. \ref{fig:norm_native_load_curves}, which presents normalized native demand curves corresponding to the curves in Fig. \ref{fig:native_load_curves}. Note that the normalized curves are obtained by dividing the real curves in Fig. \ref{fig:native_load_curves} by their peaks, respectively.
To stress the importance of Fig. \ref{fig:norm_native_load_curves}, we first define \textit{\textbf{two}} types of customer groups: the residential customers \textit{\textbf{with}} and \textit{\textbf{without}} PVs. These two customer groups are denoted as $C_w$ and $C_{o}$, respectively. For $C_o$, its native demand is recorded by smart meters. For $C_w$, we only know its net demand, and we do not know its native demand. Our goal is to estimate $C_w$'s unknown native demand and thus to estimate its PV generation. \hl{Therefore, Fig.} \ref{fig:norm_native_load_curves} \hl{inspires us that given the known native demand curve of $C_o$, we can infer the unknown native demand curve of $C_w$ by multiplying the native demand curve of $C_o$ by a \textit{ratio}, $r$.}
Since the native demands for the customers in $C_o$ are directly recorded by smart meters, the native demand curve of $C_o$ can be obtained by aggregating the native demand time series over the customers in $C_o$. The challenge for inferring the unknown native demand curve of $C_w$ is that the ratio, $r$, is unknown and needs to be estimated. \hl{The unknown of $r$ is caused by the unavailability of the native demand during the daytime for the customers in $C_w$. }This is because PV generates power during the daytime, which masks the native demand in the case of net metering. Thus, we cannot use daytime native demand to compute $r$. Instead, we use the nocturnal native demand to estimate $r$ because PV does not generate power during nighttime, and thus the nocturnal native demand for $C_w$ is known. Based on the above inference, we propose first utilizing the nocturnal native demand to compute a nocturnal native demand ratio, $r_n$, and then approximating $r$ as $r_n$.
\hl{One \textit{pre-condition} for approximating $r$ as $r_n$ is that $r$ should be close to $r_n$.} To verify this condition, we randomly select two groups with different customer numbers ranging from 20 to 80. Then, for each group, the native demand time series are spatially aggregated over customers to obtain an aggregate native demand time series. After that, we compute $r$ using the two groups' native demand time series throughout a certain period, and compute $r_n$ using the two groups' native demand time series only during \textit{nighttime} within that period. Finally, we plot $r$ against $r_n$, as shown in Fig. \ref{fig:demand_ratio}. We can see that $r$ is almost identical with $r_n$. Therefore, we can accurately estimate $r$ by directly letting it equal $r_n$.
Once we obtain the estimate of $r$, we can compute the unknown native demand of $C_w$ by multiplying the known native demand of $C_o$ by the estimate of $r$. After that, estimating the unknown PV generation of $C_w$ is straightforward, i.e., by subtracting the recorded net demand measurements from the estimated native demand.
\begin{figure}
\centering
\includegraphics[width=0.45\linewidth]{figure/ratio_vs_noc_ratio.pdf}
\caption{The relationship between native demand ratio and the nocturnal native demand ratio between two example customer groups.}
\label{fig:demand_ratio}
\end{figure}
\subsection{Finding 2: Generation Spatial Correlation between Two PVs}\label{sec:finding_2}
There are two primary factors that determine the generation spatial correlation: (1) In most cases, a distribution system is geographically bounded in a small district. (2) The most widely available sampling resolution for smart meters is 1-hour. Under these two conditions, different PV arrays are subject to nearly identical meteorological inputs. Thus, the identical inputs can result in highly similar shapes among PV generation curves. Fig. \ref{fig:generation_curves} presents three example PV generation curves corresponding to different PV array capacities. Similar to the native demand curves for sizeable customer groups, these three generation curves also demonstrate significant spatial correlation, i.e., they possess highly similar shapes. This high similarity can also be corroborated by Fig. \ref{fig:norm_generation_curves}, where the normalized generation curves corresponding to the three curves in Fig. \ref{fig:generation_curves} overlap with each other. Most importantly, Fig. \ref{fig:generation_curves} and \ref{fig:norm_generation_curves} inspire us that estimating a BTM PV generation curve comes down to two steps: first, determine the generation curve's shape, and then determine its magnitude. This two-step method can notably simplify the estimation of unknown BTM PV generation time series. This is because compared to model-based methods, our approach is developed on the foundation of high similarity among generation curves; therefore, it requires significantly less information.
\begin{figure}
\centering
\includegraphics[width=0.808\linewidth]{figure/generation_curves.pdf}
\caption{Three-day real generation curves for three example PVs with different capacities.}
\label{fig:generation_curves}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.808\linewidth]{figure/normalzied_generation_curves.pdf}
\caption{Three-day normalized generation curves for three example PVs with different capacities.}
\label{fig:norm_generation_curves}
\end{figure}
\section{Estimating Aggregate BTM PV Generation for Customers with PVs}\label{sec:aggre_level}
\begin{figure*}
\centering
\includegraphics[width=0.75\linewidth]{figure/aggregate_layer.pdf}
\caption{Detailed structure of the proposed aggregate-level BTM PV generation/native demand estimation.}
\label{fig:aggregate_layer}
\end{figure*}
As elaborated in Section \ref{sec:finding_1}, the native demands of two sizeable customer groups are highly correlated. This high correlation inspires us that we can infer the unknown native demand of $C_w$ by multiplying the known native demand of $C_o$ by a ratio:
\begin{equation} \label{eq:P_w}
\hat{\pmb{P}}_w = \pmb{P}_o r,
\end{equation}
where, $\hat{\pmb{P}}_w=\{\hat{P}_w(t)\}$ and $\pmb{P}_o=\{P_o(t)\}$, $t=1,...,T$, denote the estimated native demand time series for $C_w$ and the actual native demand time series for $C_o$, respectively. $T$ is the total number of native demands in a selected window (e.g., one month). $P_o(t)$ is computed by aggregating the measured native demands over customers without PVs:
\begin{equation} \label{eq:P_o_t}
P_o(t) = \sum_{i=1}^{N_o}P_{o,i}(t), \quad t=1,...,T,
\end{equation}
where, $N_o$ represents the total number of customers in $C_o$, i.e., customers without PVs. $P_{o,i}(t)$ denotes the measured \textit{native} demand at time $t$ for the $i$'th customer in $C_o$.
In (\ref{eq:P_w}), $r$ denotes the native demand ratio between $C_w$ and $C_o$, and is defined as follows:
\begin{equation} \label{eq:hat_r}
r = \frac{\Sigma_{t=1}^{T}P_w(t)}{\Sigma_{t=1}^{T}P_o(t)}.
\end{equation}
However, as presented in Section \ref{sec:finding_2}, since the \textit{diurnal} native demand for $C_w$ is masked by PV generation and unavailable to utilities, we need to estimate $r$ using \textit{nocturnal} native demand measurements. This approximation method is based on the observation that PV does not generate power during nighttime and the verification that $r$ and $r_n$ are almost identical. Specifically, we use $r_n$ to approximate $r$:
\begin{equation} \label{eq:hat_r_n}
\hat{r} = r_n = \frac{\Sigma_{t \in I_{n}}^{}P_w(t)}{\Sigma_{t \in I_{n}}^{}P_o(t)},
\end{equation}
where, $I_{n}$ denotes the set of nighttime hours. In our paper, $I_n$ refers to the hours between 9:00 P.M. and 5:00 A.M. Note that for the hours in $I_n$, since PV does not generate power, $P_w(t)$ equals the known aggregate \textit{net} demand, $P_w'(t)$. Therefore,
\begin{equation} \label{eq:hat_r_n_1}
\hat{r} = \frac{\Sigma_{t \in I_{n}}^{}P_w'(t)}{\Sigma_{t \in I_{n}}^{}P_o(t)},
\end{equation}
where, $P_w'(t)$ is computed by aggregating the measured net demands over customers in $C_w$:
\begin{equation} \label{eq:P_w_prime_t}
P_w'(t) = \sum_{i=1}^{N_w}P_{w,i}'(t), \quad t=1,...,T,
\end{equation}
where, $N_w$ represents the total number of customers in $C_w$. $P_{w,i}'(t)$ denotes the measured \textit{net} demand at time $t$ for the $i$'th customer in $C_w$.
Then, using the estimate of $r$ and the known native demand time series for $C_o$, we can apply (\ref{eq:P_w}) to compute the estimated native demand time series for $C_w$. Finally, inferring the PV generation time series for $C_w$, $\hat{\pmb{G}}_w = \{\hat{G}_w(t)\}$, $t=1,...,T$, is straightforward:
\begin{equation} \label{eq:G_w}
\hat{\pmb{G}}_w = \hat{\pmb{P}}_w - \pmb{P}_w',
\end{equation}
where, $\pmb{P}_w' = \{P_w'(t)\}$, $t=1,...,T$, denotes the known net demand time series for $C_w$.
The above procedure for estimating the aggregate-level PV generation and native demand for $C_w$ are illustrated in Fig. \ref{fig:aggregate_layer}.
\section{Estimating BTM PV Generation for Each Individual PV}\label{sec:indiv_level}
Knowing the aggregate BTM PV generation and native demand might not be sufficient for some applications \cite{qianzhi, rui_cheng}. For example, some demand response schemes require known customer-level native demand \cite{kangping_li_solar}. Therefore, estimating individual customers' BTM native demand and PV generation is of significance.
To achieve this goal, we propose an approach to allocate the estimated aggregate PV generation/native demand time series to individual customers with PVs. As discussed in Section \ref{sec:finding_2}, estimating an individual PV's generation curve boils down to determining the generation curve's shape and its magnitude. \hl{In this section, our approach has three steps to perform allocating: (Step-I): generate candidate generation curves for individual PVs; (Step-II): estimate the peak generation for each PV; and (Step-III): allocate the estimated aggregate PV generation time series to individual PVs by solving an optimization problem.}
\subsection{Generating Diverse Candidate Generation Curves for Individual PVs }\label{sec:indiv_level_shape}
As discussed earlier, in a geographically bounded distribution system, two primary factors determining a generation curve are the magnitude and shape. \hl{This subsection aims to generate candidate generation curves for those non-south-facing PVs. First, we train a regression model using the data generated by PVWatts Calculator. Then, we feed the estimated generation curve of a south-facing PV into the trained model to infer the targeted candidate generation curves for those non-south-facing PVs.}
In Section \ref{sec:aggre_level}, we have obtained the estimated time series for the aggregate generation of all PVs. One question is whether we can use that shape to represent the unknown shapes of individual PVs. To answer this question, we have conducted a numerical experiment. First, we normalized the aggregate generation curve of all PVs by dividing the aggregate generation time series by its peak. Then, in the same way, we normalized the generation curve of an example PV facing \textit{south}. The two normalized curves are plotted in Fig. \ref{fig:nrm_agg_indi_gener_curves}. It can be seen that the normalized curve corresponding to the aggregate generation for all PVs is highly similar to the normalized curve for a south-facing PV. One primary reason for this similarity is that the majority of residential PVs face south because a south-facing PV can typically generate more power than PVs in other directions. Most importantly, Fig. \ref{fig:nrm_agg_indi_gener_curves} tells us that a south-facing PV's generation curve can be accurately represented by the normalized aggregate generation curve of all PVs.
\begin{figure}
\centering
\includegraphics[width=0.75\linewidth]{figure/norm_agg_indiv_curves.pdf}
\caption{Three-day normalized aggregate generation curve for all PVs and normalized generation curve for an individual PV facing south.}
\label{fig:nrm_agg_indi_gener_curves}
\end{figure}
Note that in distribution systems, in addition to the majority of south-facing PVs, there exist some residential PVs with other azimuths, such as east or west. These non-south-facing PVs' generation curves cannot be fully represented by the normalized aggregate PV generation curve in Fig. \ref{fig:nrm_agg_indi_gener_curves}. Specifically, compared to the normalized aggregate PV generation curve, the normalized generation curves for an east-facing PV and a west-facing PV are somewhat ``left-skewed" and ``right-skewed", respectively, as shown in Fig. \ref{fig:nrm_agg_indi_gener_curves_1}. Therefore, it is necessary to obtain candidate shapes for those non-south-facing PVs' generation curves. To achieve this goal, our basic idea is first to feed PV power data generated by PVWatts Calculator into a regression model to capture the relationship between the generations for a south-facing PV and a non-south-facing PV. Then, the aggregate generation curve estimated in Section \ref{sec:aggre_level}, which can accurately represent a south-facing PV's generation curve, is fed into the trained regression model to produce diverse generation curves corresponding to non-south azimuths. The overall structure is shown in Fig. \ref{fig:overall_framework_candidate_curves}:
\begin{figure}
\centering
\includegraphics[width=0.75\linewidth]{figure/norm_agg_indiv_curves_1.pdf}
\caption{Three-day normalized aggregate generation curve of all PVs and normalized generation curves for two example PVs facing east and west, respectively.}
\label{fig:nrm_agg_indi_gener_curves_1}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.87\linewidth]{figure/overall_framework_candidate_curves.pdf}
\caption{Overall structure for producing diverse candidate PV generation curves using power output data generated by PVWatts Calculator.}
\label{fig:overall_framework_candidate_curves}
\end{figure}
\subsubsection{Training A Gaussian Process Regression Model}
Since the shape of a south-facing PV's generation curve can be approximated as the shape of the aggregate generation curve of all PVs, one intuitive way for inferring \textit{non-south-facing} PVs' candidate shapes is to produce diverse shapes based on the \textit{south-facing} PV's estimated generation curve. This idea is based on our observation that there exists a mapping between the generation curves for PVs with different azimuths. Therefore, one critical step for producing diverse candidate generation curves is to identify the relationship between a non-south-facing PV's generation curve and a south-facing PV's generation curve.
To capture the relationship, first, we use PVWatts Calculator \cite{PV_watts}, an online application developed by the National Renewable Energy Laboratory (NREL), to generate power output data for PVs with typical azimuths, e.g., east, south, and west. Then, using the generated PV output power data, we train a Gaussian Process Regression (GPR) model to capture the relationship between the generation curve corresponding to a typical azimuth except for south (e.g., east) and the generation curve corresponding to the azimuth of the south.
The primary reason for selecting GPR is that after running numerical tests, GPR demonstrated a relatively better performance when applied to our dataset than some other state-of-the-art nonlinear regression models, such as the Support Vector Machine model and the Polynomial regression model.
Specifically, first, we use PVWatts Calculator to generate time-series data for a south-facing PV and a PV with other typical azimuth (e.g., east). Then, each time series is normalized so that the peak generation is 1 p.u. The two normalized time series corresponding to the south-facing PV and the non-south-facing PV are denoted as $\pmb{G}_s^*=\{G_s^*(t)\}$ and $\pmb{G}_{ns}^*=\{G_{ns}^*(t)\}$, $t=1,...,T$, respectively. $G_s^*(t)$ and $G_{ns}^*(t)$ denote the normalized generation at time $t$ for a south-facing PV and a non-south-facing PV, respectively. Our goal is to use $G_{s}^*(t)$ to explain $G_{ns}^*(t)$ because PVs in a geographically bounded distribution system typically have highly correlated generations. By conducting numerical experiments, we find that in addition to $G_s^*(t)$, the hour-in-day, $H_d(t)$, and day-in-year, $D_y(t)$, are also related with $G_{ns}^*(t)$. Therefore, we use ${G}_{s}^*(t)$, $H_d(t)$, and $D_y(t)$ as the input variables and ${G}_{ns}^*(t)$ as the output variable, respectively, to train a GPR model. The function of GPR is to capture the relationship between $G_{ns}^*(t)$ and $G_{s}^*(t)$. The basic idea behind GPR is that if the distance between two explanatory variables is small, the difference between their corresponding dependent variables will also be relatively small. Specifically, the output, $G_{ns}^*(t)$, is denoted as a function of the input vector, $\pmb{X}^*(t)$:
\begin{equation} \label{eq:f_function}
G_{ns}^*(t) = f(\pmb{X}^*(t)),
\end{equation}
where, $\pmb{X}^*(t)=[G_{s}^*(t), H_d(t),D_y(t)]^\mathsf{T}$. For GPR, $f(\pmb{X}^*(t))$ is assumed to be a random variable reflecting the uncertainty of functions evaluated at $\pmb{X}^*(t)$. Specifically, the function $f(\pmb{X}^*(t))$ is distributed as a Gaussian process:
\begin{equation} \label{eq:GPR_definition}
f \big(\pmb{X}^*(t) \big) \sim \mathcal{GP} \big(\mu(\pmb{X}^*(t)), K(\pmb{X}^*(t), \pmb{X}^*(t')) \big),
\end{equation}
where, $\mu(\pmb{X}^*(t))$ represents the expected value of $f(\pmb{X}^*(t))$, i.e., the value of $G_{ns}^*(t)$. The covariance function, $K(\pmb{X}^*(t), \pmb{X}^*(t'))$, represents the dependence between $G_{ns}^*(t)$'s at different times. In our problem, the covariance function, $K(\cdot,\cdot)$, is specified by the Squared Exponential Kernel function expressed as:
\begin{equation} \label{eq:sqrt_exp_kernel}
K \big(\pmb{X}^*(t),\pmb{X}^*(t') \big) = \sigma_f^2 \text{exp} \bigg(-\frac{||\pmb{X}^*(t) - \pmb{X}^*(t')||_2^2}{2 \sigma^2}\bigg),
\end{equation}
where, $||\cdot||_2$ represents $l_2$-norm, $\sigma_f$ and $\sigma$ are hyper-parameters, which are determined using cross-validation. Intuitively, (\ref{eq:sqrt_exp_kernel}) measures the distance between $\pmb{X}^*(t)$ and $\pmb{X}^*(t')$, which can also reflect the similarity between $G_{ns}^*(t)$ and $G_{ns}^*(t')$.
Note that $G_{s}^*(t)$ and $G_{ns}^*(t)$ are generated solar powers using PVWatts Calculator; thus, they are known and a $T$-dimensional joint Gaussian distribution can be constructed as:
\begin{equation} \label{eq:N_dim_Gau}
\left[
\begin{array}{c}
f \big(\pmb{X}^*(1) \big) \\
\vdots \\
f \big(\pmb{X}^*(T) \big)
\end{array}
\right]
\sim \mathcal{N}
\Big(
\pmb{\mu}^*, \pmb{\Sigma}^*
\Big),
\end{equation}
where,
\begin{subequations} \label{eq:mu_sigma}
\begin{equation}
\pmb{\mu}^*=
\left[
\begin{array}{c}
\mu \big(\pmb{X}^*(1) \big) \\
\vdots \\
\mu \big(\pmb{X}^*(T) \big)
\end{array}
\right],
\end{equation}
\begin{equation}
\pmb{\Sigma}^*=
\left[
\begin{array}{ccc}
K\big(\pmb{X}^*(1),\pmb{X}^*(1) \big) & \cdots & K\big(\pmb{X}^*(1),\pmb{X}^*(T) \big) \\
\vdots & \ddots & \vdots\\
K\big(\pmb{X}^*(T),\pmb{X}^*(1) \big) & \cdots & K\big(\pmb{X}^*(T),\pmb{X}^*(T) \big)
\end{array}
\right].
\end{equation}
\end{subequations}
The joint Gaussian distribution formulated in (\ref{eq:N_dim_Gau}) represents a trained non-parametric model, which captures the relationship between $G_{ns}^*(t)$ and $G_{s}^*(t)$.
\subsubsection{Inferring A Non-south-facing PV's Generation Curve}
As shown in Fig. \ref{fig:nrm_agg_indi_gener_curves}, the normalized generation curve for a south-facing PV, $\pmb{G}_{s}=\{G_{s}(t)\}$, $t=1,...,T$, can be approximated as the normalized estimated aggregate generation curve for all PVs:
\begin{equation} \label{eq:norm_G_south}
\pmb{G}_{s} = \frac{\hat{\pmb{G}}_w}{\hat{G}_m},
\end{equation}
where, $\hat{G}_m$ denotes the peak of $\hat{\pmb{G}}_w$. To infer the unknown generation time series for a non-south-facing PV, $\pmb{G}_{ns}=\{G_{ns}(t)\}$, $t=1,...,T$, we assume $G_{ns}(t)$ is a function of $G_{s}(t)$, i.e., $G_{ns}(t)=f(G_{s}(t))$. By appending $f(G_{s}(t))$ to the end of (\ref{eq:N_dim_Gau}), an ($N+1$)-dimensional joint Gaussian distribution can be constructed as:
\begin{multline}\label{eq:N_k_dim_Gau_1}
\quad \quad \quad \quad
\left[
\begin{array}{c}
G_{ns}^*(1) \\
\vdots \\
G_{ns}^*(T) \\
G_{ns}(t)
\end{array}
\right]
=
\left[
\begin{array}{c}
f(\pmb{X}^*(1)) \\
\vdots \\
f(\pmb{X}^*(T)) \\
f(\pmb{X}(t))
\end{array}
\right] \\
\sim \mathcal{N}
\Big(
\left[
\begin{array}{c}
\pmb{\mu}_*\\
{\mu}_1
\end{array}
\right],
\left[
\begin{array}{cc}
\pmb{\Sigma}^{*} & \pmb{\Sigma}_{*1}\\
\pmb{\Sigma}_{*1}^T & {\Sigma}_{11}
\end{array}
\right]
\Big),
\quad \quad \quad
\end{multline}
where, $\pmb{X}(t)=[G_s(t), H_d(t), D_y(t)]^\mathsf{T}$ is a vector of explanatory variables. $\pmb{\Sigma}_{*1}$ represents the training-test set covariances and ${\Sigma}_{11}$ is the test set covariance. Since $G_{ns}^*(t)$, $\pmb{X}^*(t)$, and $\pmb{X}(t)$ are known, using the Bayes rule, the distribution of $G_{ns}(t)$ conditioned on $\pmb{G}_{ns}^*$ can be computed as follows:
\begin{equation} \label{eq:inferred_non_sou_fac_G}
G_{ns}(t) | \pmb{G}_{ns}^* \sim \mathcal{N} ({\mu}_1(t), \Sigma_1(t)),
\end{equation}
where, ${\mu}_1(t)=\pmb{\Sigma}_{*1}^T{\pmb{\Sigma}^*}^{-1} \pmb{G}_{ns}^*$ and $\Sigma_1(t)=\Sigma_{11}-\pmb{\Sigma}_{*1}^T {\pmb{\Sigma}^*}^{-1}\pmb{\Sigma}_{*1}$. Note that ${\mu}_1(t)$ denotes the most probable value of the estimated generation at time $t$ for a non-south-facing PV. By conducting the above inferring procedure for all the $t$'s, we can obtain a candidate generation time series corresponding to a \textit{particular} typical PV azimuth. Since there are multiple typical azimuths, such as east, and west, we can infer multiple candidate PV generation time series:
\begin{equation} \label{eq:mult_cand_curves}
\pmb{G}_{ns}^j=\{G_{ns}^j(t)\}, \quad t=1,..., T, \quad j=1,...,N_{ns},
\end{equation}
where, $G_{ns}^j(t)$ denotes the inferred PV generation at time $t$, for the $j$'th typical non-south-facing azimuth. $N_{ns}$ denotes the total number of typical non-south-facing PV azimuths and is determined by conducting numerical experiments.
\subsection{Estimating Peak Generation for Each Individual PV}\label{sec:indiv_level_peak}
Simply knowing the candidate shapes for unknown generation curves is insufficient for allocating the estimated aggregate generation to individual PVs. As discussed earlier, we should also know the magnitudes for the candidate generation curves. To estimate the peak generation, we employ our observation from real data that the peak generation is almost identical with the difference between the minimum diurnal native demand and the minimum net demand.
\begin{figure}
\centering
\includegraphics[width=0.64\linewidth]{figure/load_duration_curve_edited.pdf}
\caption{Load duration curves for an example customer's diurnal native demand and diurnal net demand.}
\label{fig:duration_curves}
\end{figure}
Specifically, to explain our observation regarding the correlation, we start with Fig. \ref{fig:duration_curves}, showing the load duration curves for the $i$'th customer's diurnal \textit{native} demand, $P_{w,d,i}(t)$, and diurnal \textit{net} demand, $P_{w,d,i}'(t)$. Thus, we can compute the difference between the minimums of $P_{w,d,i}(t)$ and $P_{w,d,i}'(t)$:
\begin{equation} \label{eq:D_{w,i}}
D_{w,i} = \underbar{$P$}_{w,d,i} - \underbar{$P$}_{w,d,i}',
\end{equation}
where, $\underbar{$P$}_{w,d,i}$ and $\underbar{$P$}_{w,d,i}'$ denote the minimums of $P_{w,d,i}(t)$ and $P_{w,d,i}'(t)$ during a selected window, respectively. Note that $\underbar{$P$}_{w,d,i}$ is positive, and $\underbar{$P$}_{w,d,i}'$ is negative. Then, our finding is that $D_{w,i}$ is highly similar to the peak generation, $G_{w,m,i}$, as shown in Fig. \ref{fig:diff_vs_peak_G}. This relationship inspires us to approximate $G_{w,m,i}$ as $D_{w,i}$:
\begin{equation} \label{eq:G_{w,m,i}}
\hat{G}_{w,m,i} = D_{w,i}, \quad i=1,..,N_w,
\end{equation}
where, $\hat{G}_{w,m,i}$ is the estimate of $G_{w,m,i}$. However, one challenge is that $D_{w,i}$ depends on $\underbar{$P$}_{w,d,i}$, which is unknown due to BTM PV generation. Therefore, we need to estimate $\underbar{$P$}_{w,d,i}$, which is involved with another finding from real native demand data. Specifically, as shown in Fig. \ref{fig:noc_min_P_vs_diu_min_P}, the minimum \textit{diurnal} native demand, $\underbar{$P$}_{w,d,i}$, can be approximated as the minimum \textit{nocturnal} native demand, $\underbar{$P$}_{w,n,i}$:
\begin{equation} \label{eq:diu_min_P_i}
\underbar{$P$}_{w,d,i} \approx \underbar{$P$}_{w,n,i}, \quad i=1,..,N_w.
\end{equation}
Note that since PV does not generate power during nighttime, $\underbar{$P$}_{w,n,i}$ is known to utilities. Finally, using the estimate of $\underbar{$P$}_{w,d,i}$ and the known $\underbar{$P$}_{w,d,i}'$, we can compute $D_{w,i}$ using (\ref{eq:D_{w,i}}), and then compute $\hat{G}_{w,m,i}$ using (\ref{eq:G_{w,m,i}}).
\begin{figure}[htbp]
\centering
\subfloat[Spring\label{sfig:diff_vs_peak_G_spring}]{
\includegraphics[width=0.38\linewidth]{figure/difference_vs_peak_G_spring.pdf}
}
\hfill
\subfloat[Summer\label{sfig:diff_vs_peak_G_summer}]{
\includegraphics[width=0.38\linewidth]{figure/difference_vs_peak_G_summer.pdf}
}
\caption{The relationship between peak generation and the difference between minimum diurnal \textit{native} demand and minimum \textit{net} demand.}
\label{fig:diff_vs_peak_G}
\end{figure}
\begin{figure}[htbp]
\centering
\subfloat[Spring\label{sfig:noc_min_P_vs_diu_min_P_spring}]{
\includegraphics[width=0.44\linewidth]{figure/diu_min_nat_P_vs_noc_min_nat_P_spring.pdf}
}
\hfill
\subfloat[Summer\label{sfig:noc_min_P_vs_diu_min_P_summer}]{
\includegraphics[width=0.44\linewidth]{figure/diu_min_nat_P_vs_noc_min_nat_P_summer.pdf}
}
\caption{The relationship between minimum \textit{diurnal} native demand and minimum \textit{nocturnal} native demand.}
\label{fig:noc_min_P_vs_diu_min_P}
\end{figure}
\subsection{Allocating the Estimated Aggregate PV Generation to Individual PVs }\label{sec:optimization_formulation}
Sections \ref{sec:aggre_level}, \ref{sec:indiv_level_shape}, and \ref{sec:indiv_level_peak} provide the estimated aggregate generation time series of all PVs, inferred candidate generation curves for individual PVs, and estimated generation peaks for individual PVs, respectively. Therefore, estimating individual PVs' generation curves comes down to allocating the estimated aggregate generation time series to individual PVs. This allocating procedure is formulated as an optimization process:
\begin{subequations} \label{eq:overall}
\begin{flalign} \label{eq:overall_a}
\underset{\mathbf{K}, \pmb{\gamma}}{\textit{min}} \;
\, || \mathbf{G}_e * \mathbf{K} * \pmb{1} - \hat{\pmb{G}}_w ||_2^2 + \lambda * ||\pmb{\gamma}||_2^2
\end{flalign}
\vspace{-14pt}
\begin{flalign} \label{eq:overall_b}
\quad \quad \quad \quad \quad \textit{s.t.} \; \;
\mathbf{G}_e * \mathbf{K} \le \pmb{1}* (\hat{\pmb{G}}_{w,m} + \pmb{\gamma})^\mathsf{T}, &&
\end{flalign}
\vspace{-14pt}
\begin{flalign} \label{eq:overall_d}
\quad\quad\quad \quad \quad \quad \quad
\pmb{0} \le \pmb{\gamma} \le P_0 * \pmb{1}, &&
\end{flalign}
\end{subequations}
where, $\mathbf{G}_e=[\pmb{G}_s, \pmb{G}_{ns}^1, ..., \pmb{G}_{ns}^{N_{ns}}]$ is a $T$-by-$N_e$ matrix, which denotes a collection of candidate generation curves. $N_e=N_s+1$ denotes the total number of candidate generation curves. $\mathbf{K}=[\pmb{K}_1,...,\pmb{K}_{N_w}]$ is an $N_e$-by-$N_w$ matrix of decision variables, which denote the weights assigned to candidate generation curves for individual PVs. $\pmb{K}_i,i=1,...,N_w$, is an $N_e$-by-1 vector, which denotes the weights assigned to candidate generation curves for the $i$'th PV. The first $\pmb{1}$ is an $N_w$-by-1 vector of ones. $\mathbf{G}_e * \mathbf{K}$ results in a $T$-by-$N_w$ matrix, which is a collection of estimated generation time series for individual PVs. The first term in the objective function (\ref{eq:overall_a}) reflects the difference between the estimated aggregate PV generation, $\hat{\pmb{G}}_w$, and the weighted summation of individual PV's estimated generations, $\mathbf{G}_e * \mathbf{K} * \pmb{1}$. The second term in the objective function (\ref{eq:overall_a}) considers the estimation errors of peak generations. $\lambda$ is a tuning parameter. $\pmb{\gamma}$ is an $N_w$-by-1 vector with non-negative elements, which reflect the errors of approximating $G_{w,m,i}$ as $D_{w,i}$, as shown in (\ref{eq:G_{w,m,i}}). The second $\pmb{1}$ is a $T$-by-1 vector of ones. $\hat{\pmb{G}}_{w,m} = [\hat{G}_{w,m,1},...,\hat{G}_{w,m,N_w}]^\mathsf{T}$ denotes an $N_w$-by-1 vector of the estimated generation peaks for all PVs. $(\hat{\pmb{G}}_{w,m}+\pmb{\gamma})$ denotes the corrected generation peaks with consideration of estimation errors. $\pmb{1}* (\hat{\pmb{G}}_{w,m} + \pmb{\gamma})^\mathsf{T}$ produces a $T$-by-$N_w$ matrix , in which each column contains the same element. Constraint (\ref{eq:overall_b}) ensures that the estimated generation time series for each PV is smaller than its estimated peak generation. $\pmb{0}$ is an $N_w$-by-1 vector of zeros. $P_0$ denotes the maximum error of approximating $G_{w,m,i}$ as $D_{w,i}$ for individual PVs. The third $\pmb{1}$ is an $N_w$-by-1 vector of ones. Constraint (\ref{eq:overall_d}) ensures that the estimation errors for individual PVs are non-negative and smaller than an upper bound. The reason for constraining the elements of $\pmb{\gamma}$ as non-negative is that $D_{w,i}$ typically under-estimates $G_{w,m,i}$, as shown in Fig. \ref{fig:noc_min_P_vs_diu_min_P}.
The optimization process represented in (\ref{eq:overall}) is a convex quadratic programming problem, thus, we can obtain a unique solution for $\mathbf{K}$, i.e., $\mathbf{K}^*=[\pmb{K}_1^*, ..., \pmb{K}_{N_w}^*]$. Then, the estimated generation time series for the $i$'th PV, $\hat{\pmb{G}}_{w,i}=\{\hat{G}_{w,i}(t)\}$, $t=1,...,T$, can be computed as:
\begin{equation} \label{eq:hat_G_w_i}
\hat{\pmb{G}}_{w,i} = \mathbf{G}_e * \pmb{K}_i^*, \quad i=1,...,N_w.
\end{equation}
Then, the estimated native demand time series for the $i$'th customer, $\hat{\pmb{P}}_{w,i}=\{\hat{P}_{w,i}(t)\}$, $t=1,...,T$, can be computed as:
\begin{equation} \label{eq:hat_P_w_i}
\hat{\pmb{P}}_{w,i} = \pmb{P}_{w,i}' + \hat{\pmb{G}}_{w,i}, \quad i=1,...,N_w.
\end{equation}
where, $\pmb{P}_{w,i}'=\{{P}_{w,i}'(t)\}$, $t=1,...,T,$ denotes the known net demand time series recorded by smart meter for the $i$'th customer with PVs.
\begin{figure}
\centering
\includegraphics[width=1\linewidth]{figure/customer_layer.pdf}
\caption{\hl{Detailed steps of the individual customer-level BTM PV generation estimation.}}
\label{fig:customer_layer}
\end{figure}
Note that (\ref{eq:overall}) can be solved for a selected window. The window size, $T$, can impact estimation accuracy and runtime, which will be examined in the Case Study Section. The detailed steps for estimating customer-level PV generation are illustrated in Fig. \ref{fig:customer_layer}.
\section{Case Study}\label{sec:case_study}
In this section, the proposed two-layer BTM solar power and native demand estimation approach is verified using real PV generation and native demand data.
\subsection{Dataset Description}
The hourly native demand and PV generation data used in this paper are from a public dataset \cite{data_source}. The time range of native demand and solar power is one year. This dataset contains a total number of 100 customers with PVs and 115 customers without PVs. For the customers with PVs, the net demand is obtained by subtracting PV generation from native demand.
\subsection{Aggregate-level BTM PV Generation Estimation Validation}
Fig. \ref{fig:aggre_real_esti_curves} shows three-day actual and estimated aggregate PV generation/native demand curves. It can be seen that the estimated curves can accurately follow the actual curves. \hl{To quantitatively evaluate the estimation accuracy, we compute the mean absolute percentage error (MAPE) as follows:}
\begin{equation} \label{eq:MAPE_agre}
MAPE = \frac{100\%}{N_d} \sum_{t \in I_d}^{} \Bigg| \frac{\hat{Y}_{w}(t)-{Y}_{w}(t)}{Y_{w,m}} \Bigg|,
\end{equation}
\hl{where, $\hat{Y}_{w}(t)$ represents $\hat{G}_w(t)$ or $\hat{P}_w(t)$. ${Y}_{w}(t)$ represents ${G}_w(t)$ or ${P}_w(t)$. $Y_{w,m}$ represents $G_{w,m}$ or $P_{w,m}$, where $G_{w,m}$ and $P_{w,m}$ denote the actual peaks of PV generation and native demand, respectively.} $I_d$ denotes the set of daytime hours. $N_d$ denotes the total number of hours in $I_d$.
\hl{To comprehensively evaluate the performance of our approach, we also compute the mean squared error (MSE) and coefficient of variation (CV):}
\begin{equation} \label{eq:MSE_agre}
MSE = \frac{1}{N_d} \sum_{t \in I_d}^{} \big(\hat{Y}_{w}(t)-{Y}_{w}(t) \big)^2,
\end{equation}
\begin{equation} \label{eq:CV_agre}
CV = \frac{\sigma}{\mu},
\end{equation}
\hl{where, }
\begin{subequations} \label{eq:mu_sigma_agre}
\begin{equation} \label{eq:mu_agre}
\mu=\frac{1}{N_d} \sum_{t \in I_d}^{} (\hat{Y}_{w}(t)-{Y}_{w}(t) ),
\end{equation}
\begin{equation} \label{eq:sigma_agre}
\sigma = \sqrt{\frac{1}{N_d-1}\sum_{t \in I_d}^{} \big((\hat{Y}_{w}(t)-{Y}_{w}(t)) -\mu \big)^2}.
\end{equation}
\end{subequations}
The computed $MAPE$'s for PV generation and native demand are 1.21\% and 1.28\%, respectively. \hl{The computed $MSE$'s for PV generation and native demand are about 58.09. Note that the actual peaks for the PV generation and native demand are 462.5 and 437.1 kW, respectively. The computed $CV$'s for PV generation and native demand are about -3.48. The above error metrics reflect the high accuracy of our proposed approach. }
\begin{figure}[htbp]
\centering
\subfloat[Aggregate PV generation\label{sfig:real_esti_aggre_G_curves}]{
\includegraphics[width=0.75\linewidth]{figure/real_esti_aggre_G_curves.pdf}
}
\hfill
\subfloat[Aggregate native demand\label{sfig:real_esti_aggre_P_curves}]{
\includegraphics[width=0.75\linewidth]{figure/real_esti_aggre_P_curves.pdf}
}
\caption{\hl{Three-day actual and estimated aggregate PV generation and native demand curves.}}
\label{fig:aggre_real_esti_curves}
\end{figure}
\subsection{Customer-level BTM PV Generation Estimation Validation}
\subsubsection{Estimation Performance}
Fig. \ref{fig:indi_real_esti_curves} shows three-day actual and estimated PV generation and native demand curves for an example customer with PV. We can see that the estimated curves can accurately fit the actual curves. To comprehensively examine the performance of our approach, we compute the $MAPE$ for all customers with PVs. Specifically, the $MAPE$'s for the $i$'th customer are computed \hl{as follows:}
\begin{equation} \label{eq:MAPE_indi}
MAPE_{i} = \frac{100\%}{N_d} \sum_{t \in I_d}^{} \Bigg| \frac{\hat{Y}_{w,i}(t)-{Y}_{w,i}(t)}{Y_{w,m,i}} \Bigg|
\end{equation}
\hl{where $Y_{w,i}(t)$ represent $G_{w,i}(t)$ or $P_{w,i}(t)$, $\hat{Y}_{w,i}(t)$ represent $\hat{G}_{w,i}(t)$ or $\hat{P}_{w,i}(t)$, and $Y_{w,m,i}$ represent $G_{w,m,i}$ or $P_{w,m,i}$. $G_{w,m,i}$ and $P_{w,m,i}$ denote the actual generation and native demand peaks for the $i$'th customer, respectively. We also compute the $MSE$ and $CV$ for each PV-installed customer:}
\begin{equation} \label{eq:MSE_indiv}
MSE_i = \frac{1}{N_d} \sum_{t \in I_d}^{} \big(\hat{Y}_{w,i}(t)-{Y}_{w,i}(t) \big)^2,
\end{equation}
\begin{equation} \label{eq:CV_indiv}
CV_i = \frac{\sigma_i}{\mu_i},
\end{equation}
\hl{where, }
\begin{subequations} \label{eq:mu_sigma_indiv}
\begin{equation} \label{eq:mu_indiv}
\mu_i=\frac{1}{N_d} \sum_{t \in I_d}^{} (\hat{Y}_{w,i}(t)-{Y}_{w,i}(t) ),
\end{equation}
\begin{equation} \label{eq:sigm_indiv}
\sigma_i = \sqrt{\frac{1}{N_d}\sum_{t \in I_d}^{} \big((\hat{Y}_{w,i}(t)-{Y}_{w,i}(t)) -\mu_i \big)^2}.
\end{equation}
\end{subequations}
Table \ref{tbl:MAPE_MSE_CV} \hl{summarises the empirical cumulative distribution functions (CDFs) for the estimation $MAPE$, $MSE$, and $CV$, which are constructed using all the computed $MAPE$'s, $MSE$'s, and $CV$'s, respectively.} As can be seen, for the estimated hourly PV generation, 70\% of the $MAPE$'s are less than 6.38\%. Regarding the estimated hourly native demand, 70\% of the $MAPE$'s are less than 3.67\%. This effectively verifies the estimation accuracy of our proposed approach. \hl{We also provide the percentiles of $MSE$ and $CV$ based on all the PV-installed customers' generation and native demand estimates, which can more comprehensively evaluate the performance of our approach. }
\begin{figure}[htbp]
\centering
\subfloat[PV generation\label{sfig:real_esti_indi_G_curves}]{
\includegraphics[width=0.75\linewidth]{figure/real_esti_indi_G_curves.pdf}
}
\hfill
\subfloat[Native demand\label{sfig:real_esti_indi_P_curves}]{
\includegraphics[width=0.75\linewidth]{figure/real_esti_indi_P_curves.pdf}
}
\caption{\hl{Three-day actual and estimated PV generation and native demand curves for an example customer with PV.}}
\label{fig:indi_real_esti_curves}
\end{figure}
\begin{table}[htbp]
\centering
\renewcommand{\arraystretch}{1.5}
\setlength{\tabcolsep}{4.8pt}
\caption{\hl{Empirical CDF of Estimation Error Metrics}}\label{tbl:MAPE_MSE_CV}
\begin{tabular}{cccccc}
\toprule[1pt]
Empirical CDF & 0.1 & 0.2 & 0.5 & 0.7 & 0.9 \\
\hline
$MAPE$ of $\hat{G}$ (\%) & 2.84 & 4.05 & 4.96 & 6.38 & 8.80 \\
$MAPE$ of $\hat{P}$ (\%) & 1.63 & 2.15 & 2.80 & 3.67 & 4.92\\
$MSE$ of $\hat{G}$ & 0.04 & 0.06 & 0.10 & 0.19 & 0.33 \\
$MSE$ of $\hat{P}$ & 0.03 & 0.05 & 0.09 & 0.18 & 0.29 \\
$CV$ of $\hat{G}$ & -11.80 & -5.13 & -2.60 & 2.37 & 16.12 \\
$CV$ of $\hat{P}$ & -11.30 & -4.65 & -2.59 & 1.77 & 10.90 \\
\bottomrule[1pt]
\end{tabular}
\end{table}
Note that the above results are obtained under the conditions that (1) five produced candidate generation curves are employed ($N_e=5$), (2) the tuning parameter in (\ref{eq:overall_a}) is 100 ($\lambda=100$), and (3) the optimization process specified in (\ref{eq:overall}) is executed for individual windows with a time length of one month ($T=720$ hours, the entire year is divided into 12 windows).
\subsubsection{Testing the Candidate Generation Curves}
As elaborated in Section \ref{sec:indiv_level_shape}, diverse candidate generation curves are produced for representing the unknown BTM generation. Thus, it is of interest to examine the effectiveness of producing candidate curves. Fig. \ref{fig:norm_candid_generation_curves} shows three produced candidate generation curves corresponding to three typical azimuths, i.e., east, south, and west, respectively. We can observe that compared to the generation curve corresponding to the south, the produced curve corresponding to the east is ``left-skewed'', and the produced curve corresponding to the west is ``right-skewed''. Therefore, the produced curves demonstrate diversity, which is consistent with our observation on real PV generation curves shown in Fig. \ref{fig:nrm_agg_indi_gener_curves_1}.
\begin{figure}
\centering
\includegraphics[width=0.75\linewidth]{figure/norm_candid_indiv_curves.pdf}
\caption{Three-day produced candidate generation curves corresponding to three typical azimuths, i.e., east, south, and west.}
\label{fig:norm_candid_generation_curves}
\end{figure}
In addition, we have also quantitatively examined the effectiveness of producing diverse candidate generation curves. Specifically, we test the impact of the number of candidate generation curves, i.e., we solve (\ref{eq:overall}) separately for three cases with different numbers of candidate curves: (I) one candidate generation curve corresponding to the azimuth of south; (II) three candidate generation curves corresponding to the east, south, and west, respectively; and (III) five candidate generation curves corresponding to the east, southeast, south, southwest, and west, respectively. The other conditions for the three cases are the same: $\lambda=100$ and $T=720$ hours. To evaluate the impact of candidate number, we compute the average $MAPE$ over all PVs' $MAPE$'s obtained from (\ref{eq:MAPE_indi}). The results are summarized in Table \ref{tbl:MAPE_impact_of_candidate}. We can see that as the candidate number increases, the estimation error decreases, and the execution time increases. In addition, the $MAPE$ for Case I is relatively greater than Case II and III, and Case II and Case III provide nearly identical $MAPE$'s. This is because three candidate curves - corresponding to the east, south, and west - can comprehensively represent the unknown BTM generation curve; adding extra candidate curves simply result in a slight accuracy improvement.
\begin{table}[htbp]
\centering
\renewcommand{\arraystretch}{1.5}
\setlength{\tabcolsep}{4.8pt}
\caption{Impact of Candidate Generation Curves}\label{tbl:MAPE_impact_of_candidate}
\begin{tabular}{cccc}
\toprule[1pt]
Case & I & II & III \\
\hline
Average $MAPE$ of $\hat{G}$ (\%) & 5.677 & 5.474 & 5.473 \\
Average $MAPE$ of $\hat{P}$ (\%) & 3.924 & 3.086 & 3.086 \\
Runtime (s) & 40 & 125 & 194 \\
\bottomrule[1pt]
\end{tabular}
\end{table}
\subsubsection{Testing the Tuning Parameter $\lambda$}
As discussed in Section \ref{sec:optimization_formulation}, $\lambda$ in (\ref{eq:overall}) reflects the confidence of estimating peak generations for individual PVs. One general principle for determining $\lambda$ is that the largest element in $\pmb{\gamma}$ is a couple of kilo-watts. In addition, the solutions for (\ref{eq:overall}) should not be sensitive to $\lambda$, i.e., (\ref{eq:overall}) should be robust to $\lambda$. To verify the robustness of our proposed approach, we solve (\ref{eq:overall}) based on different values of $\lambda$, and then compute the corresponding average $MAPE$'s for the estimated PV generation and native demand. Other conditions are that $T=720$ hours and five candidate generation curves - corresponding to the south, southeast, south, southwest, and west - are employed. The results show that for the $\lambda$'s ranging from 100 to 500 with an interval of 100, the average $MAPE$'s for PV generation and native demand do not change (5.47\% and 3.09\%). The invariant average $MAPE$'s demonstrate the robustness of our proposed approach.
\subsubsection{Testing the Window Size $T$}
Since our proposed approach can be conducted for each divided window, it is of importance to examine the impact of window size on estimation accuracy. To do this, we perform our approach for windows with different lengths and then compute the estimation $MAPE$. In Table \ref{tbl:MAPE_impact_of_T}, it can be seen that the average $MAPE$ decreases as $T$ increases. This is because for a wider window, the probability for the minimum diurnal native demand, $\underbar{$P$}_{w,d,i}$, equaling the minimum nocturnal native demand, $\underbar{$P$}_{w,n,i}$, is larger. Thus, we have a smaller estimation error for $\underbar{$P$}_{w,d,i}$, as seen in (\ref{eq:diu_min_P_i}). Then, based on (\ref{eq:D_{w,i}}) and (\ref{eq:G_{w,m,i}}), it can be seen that the smaller estimation error for $\underbar{$P$}_{w,d,i}$ results in a more accurate $D_{w,i}$, which then brings a more accurate estimate for $G_{w,m,i}$. Finally, more accurate peak generation estimates result in smaller estimation errors for the PV generation and native demand time series.
\begin{table}[htbp]
\centering
\renewcommand{\arraystretch}{1.5}
\setlength{\tabcolsep}{4.8pt}
\caption{Impact of Window Size $T$}\label{tbl:MAPE_impact_of_T}
\begin{tabular}{ccccc}
\toprule[1pt]
$T$ (month) & 1 & 2 & 3 & 4 \\
\hline
Average $MAPE$ of $\hat{G}$ (\%) & 5.47 & 5.30 & 5.18 & 5.08 \\
Average $MAPE$ of $\hat{P}$ (\%) & 3.09 &2.99 & 2.92 & 2.87 \\
\bottomrule[1pt]
\end{tabular}
\end{table}
\subsection{Performance Comparison}
This paper compares our proposed approach with previous works from two perspectives, qualitatively and quantitatively.
\subsubsection{Qualitative Analysis}
From a qualitative point of view, one primary advantage of our approach is that it does not require meteorological data and solar generation exemplars. For the aggregate level, our approach can perform PV generation estimation by only using recorded net demand data. For the customer level, our approach can also work by only relying on recorded smart meter data, although leveraging PVWatts Calculator's generated data can improve the estimation accuracy.
\subsubsection{Quantitative Comparison}
For the customer level, we have also compared our approach with previous works. Specifically, we focus on comparing our approach with the method presented in \cite{Fankun_Bu_1} and \cite{nan_peng_yu_2}, which demonstrate better performance compared to previous works. Table \ref{tbl:MAPE_comparison} summarizes the computed $MAPE$'s for our approach and the compared approach. Note that the average $MAPE$'s for our approach have lower and upper bounds because the considered window size, $T$, ranges from one month to four months. As can be seen, the approach in \cite{Fankun_Bu_1} demonstrates a similar estimation accuracy as our approach does. However, our approach does not require solar exemplars, which makes it more independent and practical.\hl{ The approach in} \cite{nan_peng_yu_2}\hl{ employs a statistical model and a physical model to represent the native load and the PV generation, respectively. Table }\ref{tbl:MAPE_comparison}\hl{ shows that our approach has a better performance than the approach in }\cite{nan_peng_yu_2}\hl{ in terms of the average $MAPE$.}
\begin{table}[htbp]
\centering
\renewcommand{\arraystretch}{1.5}
\setlength{\tabcolsep}{4.8pt}
\caption{\hl{Average $MAPE$ (\%) Comparison}}\label{tbl:MAPE_comparison}
\begin{tabular}{cccccc}
\toprule[1pt]
Approaches & Our Approach & \hl{Approach in} \cite{nan_peng_yu_2} & Approach in \cite{Fankun_Bu_1} \\
\hline
$\hat{G}$ & [5.08, 5.47] & 7.38 & 5.24 \\
$\hat{P}$ & [2.87, 3.09] & 9.94 & 2.95 \\
\bottomrule[1pt]
\end{tabular}
\end{table}
\subsection{\hl{Robustness against Measurement and Communication Noises}}
\hl{To test the robustness of our proposed approach, we add measurement and communication noises to the net demand measurements of customers with PVs and the native demand measurements of customers without PVs. For the measurement noise, we consider the Class 0.5 (having $\pm$0.5\% error) specified by ANSI C12.20. For the communication noise, we test five different packet loss rates considering that the packet loss rate depends on the communication bandwidth and data volume. For example, we purposely change 1\% of the measurements to zero to achieve a 1\% packet loss rate. To comprehensively evaluate our approach's performance, we set up five cases: Case 1 - 1\% measurement lost + 0.5\% random noise, Case 2 - 2\% measurement lost + 0.5\% random noise, Case 3 - 3\% measurement lost + 0.5\% random noise, Case 4 - 4\% measurement lost + 0.5\% random noise, and Case 5 - 5\% measurement lost + 0.5\% random noise. Then, we apply our approach to the above five cases and compute the $MAPE$ for evaluating the robustness. The results are summarized in Table} \ref{tbl:MAPE_impact_of_noise_aggre} and \ref{tbl:MAPE_impact_of_noise_indiv}. \hl{We can observe that the $MAPE$'s slowly increase while the noise level increases, demonstrating the robustness of our approach.}
\begin{table}[htbp]
\centering
\renewcommand{\arraystretch}{1.5}
\setlength{\tabcolsep}{4.8pt}
\caption{\hl{Aggregate-level Estimation $MAPE$ (\%)}}\label{tbl:MAPE_impact_of_noise_aggre}
\begin{tabular}{ccccccc}
\toprule[1pt]
& W/O noise & Case 1 & Case 2 & Case 3 & Case 4 & Case 5 \\
\hline
$\hat{G}$ & 1.21 & 1.17 & 1.22 & 1.38 & 1.53 & 1.73 \\
$\hat{P}$ & 1.28 & 1.28 & 1.33 & 1.43 & 1.58 & 1.76\\
\bottomrule[1pt]
\end{tabular}
\end{table}
\begin{table}[htbp]
\centering
\renewcommand{\arraystretch}{1.5}
\setlength{\tabcolsep}{4.8pt}
\caption{\hl{Average Customer-level Estimation $MAPE$ (\%)}}\label{tbl:MAPE_impact_of_noise_indiv}
\begin{tabular}{ccccccc}
\toprule[1pt]
& W/O noise & Case 1 & Case 2 & Case 3 & Case 4 & Case 5 \\
\hline
$\hat{G}$ & 5.47 & 5.84 & 5.86 & 5.64 & 5.54 & 5.62 \\
$\hat{P}$ & 3.09 & 3.53 & 3.68 & 3.62 & 3.63 & 3.80\\
\bottomrule[1pt]
\end{tabular}
\end{table}
\subsection{\hl{Limitations of the Proposed Approach}}
\hl{Every method has its limitations, and there is no omnipotent method that can apply to all cases. The limitation of our proposed approach is that it requires time-series smart meter data with a temporal granularity that can distinguish daytime and nighttime. This is because our approach innovatively utilizes the temporal correlation between the aggregate \textit{nocturnal} native demand and the aggregate \textit{diurnal} native demand. Under this condition, only having access to the monthly demands of those PV-installed customers brings challenges to our approach because it cannot split the monthly demand into two parts, the diurnal and nocturnal demands, for computing the nocturnal native demand ratio. We intend to address this challenge in our future work.}
\section{Conclusion}\label{sec:conclusion}
This paper is dedicated to proposing an independent and practical BTM solar power/native demand estimation approach. Our proposed approach contains two interconnected layers. The aggregate level leverages the spatial correlation of native demand to perform the aggregate PV generation/native demand estimation. The customer level utilizes the spatial correlation of PV generation to allocate the estimated aggregate PV generation/native demand to individual customers. The Case Study verifies that our approach can accurately estimate BTM PV generation/native demand, significantly enhancing distribution system observability and situation awareness. The numerical experiments also demonstrate that our approach does not require meteorological data and measured solar power exemplars. Therefore, our approach is more independent and thus is practical for utilities to implement.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\bibliographystyle{IEEEtran}
|
\section{Introduction} \label{sec:intro}
Following the discovery of the first strongly-lensed quasars \citep{1979Natur.279..381W}, \citet{Chang1979, Chang1984} pointed out that stars in the lensing galaxy can produce additional lensing effects when the size of the background source (such as quasars) is comparable to or smaller than the Einstein radii of the stars, in which the macroimage of quasars can be splitted into multiple sub-images. \citet{Paczynski86a} used the terms ``microlensing'' and ``microimages'' to describe the lensing effect from stars and further proposed an idea of searching for MAssive Compact Halo Objects (MACHOs) in the Milky Way using the microlensing effect \citep{Paczynski86b}. Since the first detection of a microlensing event \citep{1989AJ.....98.1989I}, microlensing has played an outstanding role in a wide range of fields, such as discovering exoplanets and binary stars \citep[e.g.,][]{1991ApJ...374L..37M,1992ApJ...396..104G}, inferring the structure of the Milky Way \citep[e.g.,][]{Udalski2003, Hamadache2006}, constraining the fraction of dark matter in the form of MACHOs \citep[e.g.,][]{2017ApJ...836L..18M,2018ASPC..514...79S}, determining the ratio of stellar to dark matter \citep[e.g.,][]{2002ApJ...580..685S,2004IAUS..220..103S}, measuring the size and profile of quasar accretion discs \citep[e.g.,][]{2016ApJ...830..149F,2018ApJ...869..132F}, etc.
As the separations among microimages are on the order of micro-arcsecond ($\mu$as) and therefore unresolved in most cases, microlensing-related studies so far have relied on analysing the light curve that is the change of the total brightness of all lensed microimages as the lens-source alignment changes due to relative motions. In this paper, we will discuss a special type of microlensing events where the background source is a transient with such a short duration that the microimages can be resolved temporally. As will be shown later, the typical time delays between microimages are $\sim$0.1--1 millisecond (ms). We therefore take fast radio bursts (FRBs), the durations of which are on the order of ms, as an example.
FRBs are bright radio transient signals with durations on the order of milliseconds \citep{2019A&ARv..27....4P, Xiao2021}. They are detected at frequencies from hundreds of MHz to several GHz, and the arrival time of signals in different frequencies is usually different due to the dispersion effect caused by cold plasma \citep[e.g.,][]{Lorimer2007}. The lag in arrival time is proportional to the dispersion measure (DM), which is defined as the integration of the column density of free electrons along the line of sight (LOS). By measuring the DM from time lag and comparing it to the expected contribution of the Milky Way, it is clear that most FRBs are extragalactic. In fact, several FRB sources have been located to their host galaxies. For example, FRB121102 was precisely localized to reside in a dwarf star forming galaxy at $z_{s}=0.1927$ \citep{Chatterjee2017, Tendulkar2017} and FRB180916 was located to reside in a nearby spiral galaxy at $z_{s}=0.0337$ \citep{2020Natur.577..190M}. The formation mechanism of the bursts is still an open question since the first discovery by \citet{Lorimer2007}. The properties of FRB sources also seem heterogeneous. Among the discovered FRB samples, the majority show a single burst while some were found to have additional bursts \citep{2016Natur.531..202S,2019Natur.566..235C,2019ApJ...885L..24C,2020ApJ...891L...6F}, which are usually referred to as non-repeating and repeating FRBs.
Given the unique characteristics such as point like, transient, high event rate and extragalactic origin, there have been a lot of discussions in the literature on gravitational lensing effects on FRBs. For instance, \citet{Munoz2016} discussed an idea of constraining the amount of dark matter in the form of MACHO using strongly-lensed FRBs. \citet{DaiLu2017} proposed to measure the motion of the emission regions of strongly-lensed repeating FRBs from the variations of time delays, which could place constraints on the physical nature of the emissions. \citet{Wagner2018} suggested to reconstruct the mass distribution of the lens with higher accuracy using strongly-lensed repeating FRBs. \citet{Li2017} proposed to constrain cosmological parameters using strongly-lensed repeating FRBs.
\citet{Lewis2020} suggested that the behaviours of microimages of FRBs can be used to reveal the parities of the macroimages, which provides additional constraints on the macro lens model.
In our previous work \citep{2021ApJ...912..134C}, we showed that lens mass constraints can be obtained from the observed time delay and flux ratio of a FRB strongly lensed by a point-mass object.
In this work, we continue to explore the microlensing effect of FRBs. Since the duration of FRBs is on the order of millisecond, comparable to typical time delays (0.1--1ms) in stellar-mass microlensing, it is possible to temporally resolve the individual microimages. As a result, information of individual microimages (i.e. positions, magnifications, and time delays) is desired. A method beyond the inverse ray shooting (IRS) method, which only evaluates the total magnification of all microimages, is needed. The paper is organized as follows. Section \ref{sec:lensTheory} describes the basic theory of microlensing. Section \ref{sec:Sim} introduces an algorithm we implement for finding microimages. Validations of the algorithm are presented in Section \ref{sec:test}. Section \ref{sec:discussion} provides some discussions on microlensing of FRBs and a summary is given in Section \ref{sec:conclusion}. Throughout the paper, we assume a fiducial cosmology of $\Omega_m = 0.272, \Omega_{\lambda} = 0.728, H_0=70.4$ km s$^{-1}$ Mpc$^{-1}$\citep{Komatsu2011}.
\section{The theory of microlensing}
\label{sec:lensTheory}
In this work, we focus on the microlensing effect of stars embedded in a distant lens object (e.g. a galaxy). As a result, we adopt the thin lens approximation in which the lensing effect is determined by the projected effective lensing potential. Because the angular scales of the microlensing effect considered here are on the order of microarcsecond ($\mu$as), the localised effective lensing potential can be used. The lens surface mass distribution, which determines the effective lensing potential, is assumed to consist of two components: the smoothly distributed matter (e.g. dark matter) and the compact objects (e.g. stars). It is usually convenient to work with convergence that is the surface mass density normalised by the critical density $\Sigma_{0}=\frac{1}{\pi}\frac{c^{2}}{4G}\frac{D_{s}}{D_{l}D_{ls}}$, where $D_{l}$, $D_{s}$, and $D_{ls}$ are the angular diameter distances from the observer to the lens, from the observer to the source, and from lens to the source respectively. The total convergence can therefore be written as $\kappa=\kappa_{s}+\kappa_{\ast}$ where $\kappa_{s}$ and $\kappa_{\ast}$ denote the convergences from smoothly distributed matter and stars respectively.
We start with a simple case where a single star with mass $M$ is included. In this case, the total convergence $\kappa \approx \kappa_{s}$. The localised effective lensing potential can be written as
\begin{eqnarray}
\label{eq:phi-single-star}
\psi \left ( \vec{\theta} \right )=\frac{\theta _{E}^{2}}{2}\ln \left | \vec{\theta}\right |^{2}+\frac{\kappa _{s}}{2}\left ( \theta_{1}^{2}+\theta_{2}^{2}\right )-\frac{\gamma}{2}\left ( \theta_{1}^{2}-\theta_{2}^{2}\right ).
\end{eqnarray}
Here $\theta_{E}$ is the Einstein radius of the star defined as
\begin{eqnarray*}
\label{eq:thetaE}
\theta _{E}=\sqrt{\frac{4GM}{c^{2}}\, \frac{D_{ls}}{D_{l}D_{s}}}
\end{eqnarray*}
and $\gamma$ is the shear from the lensing galaxy as a whole. We note that $\kappa_{s}$ is approximated as a constant as its relative variation over the region under consideration is on the level of $\sim 10^{-3}$. The lens equation can be written as
\begin{eqnarray}
\label{eq:single-star}
\vec{\beta}=\begin{pmatrix}
1-\kappa_{s}+\gamma & 0\\
0&1-\kappa_{s}-\gamma
\end{pmatrix}\vec{\theta}-\theta_{E}^2\frac{\vec{\theta}}{\left | \vec{\theta} \right |^2},
\end{eqnarray}
Here $\vec{\beta}$ is the source position in the source plane and $\vec{\theta}$ is the image position in the deflector plane. The second term of the right side in Equation (\ref{eq:single-star}) is the contribution from the single star. By introducing $\lambda_{r}=1-\kappa+\gamma$, $\lambda_{t}=1-\kappa-\gamma$, $\vec{y}=\vec{\beta}/\theta_{E}$, and $\vec{x}=\vec{\theta}/\theta_{E}$, Equation (\ref{eq:single-star}) reads in a two-dimensional form as:
\begin{eqnarray}
\label{eq:scale-eq-1}
y_{1}&=&\lambda _{r}x_{1}-\frac{x_{1}}{x_{1}^2+x_{2}^2},\\
\label{eq:scale-eq-2}
y_{2}&=&\lambda _{t}x_{2}-\frac{x_{2}}{x_{1}^2+x_{2}^2}.
\end{eqnarray}
where subscripts 1 and 2 denote the two dimensions.\\
In this scenario, two or four lensed images (referred to microimages hereafter) can be produced. In particular, for source positions along the coordinate axes (i.e. $y_{1}=0$ or $y_{2}=0$), Equations (\ref{eq:scale-eq-1}) and (\ref{eq:scale-eq-2}) can be solved analytically. For $y_{2}=0$, the lensed images are located at
\begin{eqnarray}
\label{eq:ana-root1}
x_{1}&=&\frac{y_{1}\pm \sqrt{y_{1}^2+4\lambda _{r}}}{2\lambda _{r}},\;x2=0\\
\label{eq:ana-root11}
x_{1}&=&\frac{y_{1}}{\lambda_{r}-\lambda_{t}},\;x_{2}=\pm \sqrt{\frac{(\lambda_{r}-\lambda_{t})^2-\lambda_{t}y_{1}^2}{\lambda_{t}(\lambda_{r}-\lambda_{t})^2}} \text{ (if $\lvert y_1 \rvert \leq \sqrt{\frac{(\lambda_{r}-\lambda_{t})^2}{\lambda_{t}}}$)};
\end{eqnarray}
and for $y_{1}=0$, the lensed images are located at
\begin{eqnarray}
\label{eq:ana-root2}
x_{1}&=&0,\;x_{2}=\frac{y_{2}\pm \sqrt{y_{2}^2+4\lambda_{t}}}{2\lambda_{t}},\\
\label{eq:ana-root22}
x_{1}&=&\pm \sqrt{\frac{(\lambda_{t}-\lambda_{r})^2-\lambda_{r}y_{2}^2}{\lambda_{r}(\lambda_{t}-\lambda_{r})^2}},\;x_{2}=\frac{y_{2}}{\lambda_{t}-\lambda_{r}} \text{ (if $\lvert y_2 \rvert \leq \sqrt{\frac{(\lambda_{t}-\lambda_{r})^2}{\lambda_{r}}}$)}.
\end{eqnarray}
Obviously, the intersections of the caustic with the axes are given by $y_{1}=\pm \sqrt{\frac{(\lambda_{r}-\lambda_{t})^2}{\lambda_{t}}}$ and $y_{2}=\pm \sqrt{\frac{(\lambda_{t}-\lambda_{r})^2}{\lambda_{r}}}$.\\
Now we consider a more general situation where there are $N$ stars. For simplicity, we assume all the $N$ stars have the same mass $M$. The localised effective lensing potential becomes
\begin{eqnarray}
\label{eq:phi}
\psi \left ( \vec{\theta }\right )=\frac{\theta_{E}^{2}}{2}\sum_{i}^{N}\ln\left | \vec{\theta}-\vec{\theta}^{(i)}\right |^{2}+\frac{\kappa_{s}}{2}\left ( \theta_{1}^{2}+\theta_{2}^{2}\right )-\frac{\gamma}{2}\left ( \theta_{1}^{2}-\theta_{2}^{2}\right ),
\end{eqnarray}
where $\theta_{E}$ is the Einstein radius determined by mass $M$ as expressed above. $\vec{\theta}^{(i)}$ is the position of the $i^{th}$ star (also referred to as a microlens). Again, $\kappa_s$ is approximated as a constant.
The lens equation can be written as:
\begin{eqnarray}
\label{eq:many-stars}
\vec{\beta}=\begin{pmatrix}
1-\kappa_{s}+\gamma & 0 \\
0 & 1-\kappa_{s}-\gamma
\end{pmatrix}\vec{\theta}-\theta_{E}^2\sum_{i}^{N}\frac{\vec{\theta}-\vec{\theta}^{(i)}}{\left | \vec{\theta}-\vec{\theta}^{(i)}\right |^2}.
\end{eqnarray}
$\kappa$, $\gamma$, $\kappa_{\ast}$ and the mass function of micro-lenses determine the properties of microlensing \citep{2012ApJ...744...90B}. Similarly, we rewrite Equation (\ref{eq:many-stars}) in the two-dimensional form with $\vec{y}=\vec{\beta}/\theta_{E}$, $\vec{x}=\vec{\theta}/\theta_{E}$
\begin{equation}
\label{eq:scale-many-stars-1}
y_{1}=x_{1}-\alpha_{1}, \; \alpha_{1}=\sum_{i}^{N}\frac{x_{1}-x^{(i)}_{1}}{\left |\vec{x}-\vec{x}^{(i)} \right |^{2}}+\kappa_{s}x_{1}-\gamma x_{1},
\end{equation}
\begin{equation}
\label{eq:scale-many-stars-2}
y_{2}=x_{2}-\alpha_{2}, \; \alpha_{2}=\sum_{i}^{N}\frac{x_{2}-x^{(i)}_{2}}{\left |\vec{x}-\vec{x}^{(i)} \right |^{2}}+\kappa_{s}x_{2}+\gamma x_{2}.
\end{equation}
\\
For an image at position $\vec{x} =\left ( x_{1}, \; x_{2}\right )$ in the deflector plane, the magnification is given by
\begin{eqnarray}
\label{eq:mu}
\mu ^{-1}=\left ( 1-\frac{\partial\alpha_{1}}{\partial x_{1}}\right )\left ( 1-\frac{\partial\alpha_{2}}{\partial x_{2}}\right )-\frac{\partial\alpha_{1}}{\partial x_{2}}\frac{\partial\alpha_{2}}{\partial x_{1}}.
\end{eqnarray}
The time-delay function is
\begin{eqnarray}
\label{eq:delta_t_function}
t=\frac{D_{l}D_{s}}{cD_{ls}}\left ( 1+z_{l}\right )\left [ \frac{1}{2}\left ( \vec{\theta}-\vec{\beta}\right )^{2}-\psi \left ( \vec{\theta}\right )\right ].
\end{eqnarray}
For an image at $\vec{x}=\left ( x_{1}, \; x_{2}\right )$, the time delay relative to the unlensed case is
\begin{align}
\label{eq:scaled_delta_t}
\Delta t &= \frac{2GM\left ( 1+z_{l}\right )}{c^{3}}\left [ \left ( \vec{x}-\vec{y}\right )^{2}-\sum_{i}^{N} \ln\left ( \vec{x}^{(i)}-\vec{x}\right )^{2}-\kappa_{s}\left ( {x_{1}}^{2}+{x_{2}}^{2} \right )+\gamma\left ({x_{1}}^{2}-{x_{2}}^{2} \right )-2 N \ln \theta_{E}\right ] \nonumber \\
&= (9.85 \mu \text{s}) (1+z_l) (\frac{M}{M_{\odot}})\left [ \left ( \vec{x}-\vec{y}\right )^{2}-\sum_{i}^{N} \ln\left ( \vec{x}^{(i)}-\vec{x}\right )^{2}-\kappa_{s}\left ( {x_{1}}^{2}+{x_{2}}^{2} \right )+\gamma\left ({x_{1}}^{2}-{x_{2}}^{2} \right )-2 N \ln \theta_{E}\right ].
\end{align}
We note that the magnification $\mu$ and the absolute time delay relative to the unlensed case are not direct observables because the intrinsic luminosity and position of the background source are usually unknown. Nevertheless, the relative time delay between microimages is an observable.\\
\begin{figure}
\centerline{\scalebox{1.0}
{\includegraphics[width=0.5\textwidth]{Fig1.eps}}}
\caption{\label{fig:fig1} Illustration of the processes of finding microimages in our method. The lensing field used is a $1M_{\odot}$ star $+$ convergence $\kappa=0.6$ and shear $\gamma=0.3$. Panel (a) shows the magnification map on the logarithmic scale (i.e. $\log_{10} \mu$) in a square region of the source plane with side length L$=10\theta_{E}$. The open white circle shows the position of the source. Panel (b) shows the ray-emitting region in the deflector plane. The side of this panel is S=$100\theta_{E}$. The open purple and blue squares correspond to pixel positions of light rays with positive and negative parity that fall within $r_{s}$ of the target source in panel (a). Panel (c) is an enlarged view of the deflector plane. The blue circles denote the centroids of the three clusters identified by the friend-of-friend algorithm. The black dots in panel (d) show the positions of final resolved microimages after the cleaning and merging processes.}
\end{figure}
In this scenario, the number of microimages is not limited to two or four, and the lens equation generally can not be solved analytically. Nevertheless, microlensing-related studies so far do not attempt to resolve the individual microimages because the required angular resolution is beyond the capabilities of all current facilities. Hence, those studies primarily make use of the total magnification (i.e. summed over microimages) and its variation with respect to the position change between the lens and the source. As a result, the inverse ray shooting (IRS) method has been extensively used to generate the magnification map \citep[e.g.,][]{Kayser1986, 1987A&A...171...49S, 1999JCoAM.109..353W}. In the IRS method, numerous light rays are emitted backwards from the observer to the deflector plane and eventually land in the source plane as determined by the deflection angle. Usually the deflector plane is divided into a regular pixel grid where each pixel corresponds to one light ray. To infer the magnification in any given pixel, the location of that light ray in the source plane is identified (i.e. source position) and light rays within a certain distance to that source position are selected. The ratio of the areas covered by those light rays in the deflector plane to that in the source plane is taken as the total magnification in that pixel in the source plane. It is clear that the so-generated magnification map is only an approximation, and the accuracy depends on the resolution in the deflector plane.\\
In this work, we are interested in the microlensing effect of a special type of sources --- transient objects --- that have such short durations (for example, FRBs) that the microimages can be resolved temporally. In this scenario, the positions of each individual microimages need to be obtained first, after which their magnification and time delay can be calculated straightforwardly from Equations (\ref{eq:mu}) and (\ref{eq:delta_t_function}). As a result, an approach different from the current IRS-based approaches, which can not provide microimage positions, is needed. We have therefore implemented a numerical method for finding the microimages, which can not be solved analyticall in general.
\section{Methodology}
\label{sec:Sim}
The first step in our method is to also shoot light rays backwards from the observer to the source plane via the deflector plane. We start with an interested square region in the source plane of side length L (i.e., from $-\frac{\rm L}{2}$ to $\frac{\rm L}{2}$) that is chosen to be equal to 10$\theta_{\rm E}$. In addition, we add a buffer of length 10$\sqrt{\kappa_{\ast}}$ around this square region when calculating the required ray-emitting region in the deflector plane because the gravitational scattering of randomly distributed stars in the deflector plane can have an effect on the edge of the source plane \citep{1986ApJ...306....2K}. Under the effect of convergence $\kappa$ and shear $\gamma$, a square of length unity in the deflector plane will be mapped to a rectangle in the source plane with length and width of $1-\kappa+\gamma$ and $1-\kappa-\gamma$. We therefore set the ray-emitting region in the deflector plane as a square with side length S (i.e., from $-\frac{\rm S}{2}$ to $\frac{\rm S}{2}$), where S is $\rm Max\left [ \left |\frac{\rm L+20\sqrt{\kappa_{\ast}}}{1-\kappa+\gamma} \right |,\;\left |\frac{\rm L+20\sqrt{\kappa_{\ast}}}{1-\kappa-\gamma} \right |\right ]$. Microlenses are randomly distributed in the circumcircle of that square ray-emitting region in the deflector plane and the mass of each microlens is set to be 1 $M_{\odot}$. Using a circular microlens field is a standard procedure in microlensing studies \citep[e.g.][]{Kayser1986, 1990PhDT.......180W}. The number of microlenses---$N$ is therefore determined from $\kappa_{\ast}$ and the area of the circumcircle as $N=\frac{M_{\rm total}}{M}=\frac{\pi S^{2}D_{l}^{2}\kappa_{\ast}\Sigma_{0}}{2M}$. Hence, the microlensing effect is completely described by the local properties: $\kappa$, $\gamma$, and $\kappa_{\ast}$.
The square region in the deflector plane is divided into $\rm N_{\rm r}$ $\times$ $\rm N_{\rm r}$ pixels with each pixel corresponding to one light ray from the observer. The resolution of the deflector plane is therefore $r_{l}$=S/$\rm N_{\rm r}$. The value of $r_{l}$ should be as small as possible with the affordability of computing resource and time consuming. Rays are mapped through the field of micro-lenses to the source plane, and a mapping between the positions of light rays in the deflector plane and the source plane is obtained. We use $\left \{\vec{x}\right \}$ sample and $\left \{\vec{y}\right \}$ sample to denote the positions of rays in the deflector plane and their corresponding positions in the source plane.
Now for a given position $\vec{y}_{o}$ in the source plane (i.e. panel (a) of Figure~\ref{fig:fig1}), we identify rays from the $\left \{\vec{y}\right \}$ sample whose angular distance to $\vec{y}_{o}$ are within $r_{s}$ and call them $\left \{\vec{y}_{n}\right \}$ sample.
$r_{s}$ is related to $r_{l}$ as
\begin{eqnarray*}
r_{s}=\sqrt{\frac{n_{s}}{\pi \mu }}r_{l}.
\end{eqnarray*}
where $n_{s}$ is the expected number of rays in the $\left \{\vec{y}_{n}\right \}$ sample and $\mu$ is the total magnification. Obviously, the value of $r_{s}$ should be large enough that at least one light ray can be identified. In this work, we choose $r_{s}=5r_{l}$, which allows us to identify at least one light ray from microimages that have magnifications larger than $\mu_{\rm limit} =0.01273$. A larger $r_s$ can be used as long as the computing time permits. We then trace back to the deflector plane and group the light rays that correspond to the $\left \{\vec{y}_{n}\right \}$ sample as the $\left \{\vec{x}_{n}\right \}$ sample (i.e. all the open squares in panel (b) of Figure~\ref{fig:fig1}). We divide the $\left \{\vec{x}_{n}\right \}$ sample into $\left \{\vec{x}_{n}^{+}\right \}$ sample and $\left \{\vec{x}_{n}^{-}\right \}$ sample according to their parity, which is necessary for resolving close pairs of microimage and its counter image (i.e. magenta and blue squares in panel (b) of Figure~\ref{fig:fig1}). The $\left \{\vec{x}_{n}^{+}\right \}$ sample and $\left \{\vec{x}_{n}^{-}\right \}$ sample are grouped into $n_{p}^{+}$ positive-parity clusters and $n_{p}^{-}$ negative-parity clusters using the friend-of-friend algorithm with a clustering criterion of $r_{p}$, which is chosen to be $\sim 1.5r_{l}$. The centroids of each cluster are recorded as [${\vec{x}_{c}}^{1}$, ${\vec{x}_{c}}^{2}$,..., ${\vec{x}_{c}}^{n_p}$] where $n_p=n_{p}^{+} + n_{p}^{-}$ (i.e. panel (c) of Figure~\ref{fig:fig1}).
Then we perform a root finding for Equation (\ref{eq:scale-many-stars-1}) and Equation (\ref{eq:scale-many-stars-2}) using the two-dimensional Newton iteration method with the centroids of each cluster obtained above as the starting points. The convergence criterion is set to $10^{-6}$. As a result, $n_{p}$ numerical solutions [${\vec{x}_{\rm rt1}}^{1}$, ..., ${\vec{x}_{\rm rt1}}^{n_{p}}$] are obtained for the $n_{p}$ clusters. For each position in the $\left \{\vec{x}_{\rm rt1}\right \}$ sample, we further find its mirror position with respect to the corresponding cluster centroid $\left \{\vec{x}_{c}\right \}$ and use it as the new starting point to perform the root finding again, which results in another $n_{p}$ numerical solutions [${\vec{x}_{\rm rt2}}^{1}$, ..., ${\vec{x}_{\rm rt2}}^{n_{p}}$]. The purpose of this second iteration is also to avoid mistaking a pair of closely located microimage and its counter image as one image.
To clean up roots from the above step, the $\left \{\vec{x}_{\rm rt1}\right \}$ sample and $\left \{\vec{x}_{\rm rt2}\right \}$ sample are combined together and divided into $\left \{\vec{x}_{\rm rt+}\right \}$ group: [$\vec{x}_{\rm rt+}^1$, ..., $\vec{x}_{\rm rt+}^{n+}$] and $\left \{\vec{x}_{\rm rt-}\right \}$ group: [$\vec{x}_{\rm rt-}^1$, ..., $\vec{x}_{\rm rt-}^{n-}$] according to their parities. Then the friend-of-friend algorithm is performed again to the $\left \{\vec{x}_{\rm rt+}\right \}$ sample and $\left \{\vec{x}_{\rm rt-}\right \}$ sample separately, where two roots $\vec{x}^{i}$ and $\vec{x}^{j}$ that have separation smaller than $(\rm Max\left [ \mu^{i},\; \sqrt{\mu^{i}}\right ]+\rm Max\left [ \mu^{j},\; \sqrt{\mu^{j}}\right ])$ are merged into one root, and the final root list $\left \{\vec{x}_{\rm Img}\right \}$ (including $\left \{\vec{x}_{\rm Img+}\right \}$ and $\left \{\vec{x}_{\rm Img-}\right \}$) is obtained (i.e. panel (d) of Figure~\ref{fig:fig1}). The magnification and time delay for each of the microimages can be computed directly from Equations (\ref{eq:mu}) and (\ref{eq:delta_t_function}).
\section{Validations}
\label{sec:test}
\subsection{Single star case}
\label{subsec:single-star}
The first test we did is for the case of a single star embedded in a field of convergence $\kappa=0.6$ and shear $\gamma=0.3$.
As discussed in Section~\ref{sec:lensTheory}, the lens equation can be solved analytically in this case when the source is located on the coordinate axes. We therefore select 1000 uniformly distributed source positions in the range of $\left ( -5\theta_{E},\;5\theta_{E}\right )$ along the $y_{1}$-axis and $y_{2}$-axis (2000 sources in total). Analytically, we know that the 1000 source positions along the $y_1$ axis correspond to 621 two-image cases and 379 four-image cases, while the 1000 source positions along the $y_2$ axis correspond to 857 two-image cases and 143 four-image cases. We then calculate for each source position the number, positions, and magnifications of all its microimages found by our method, and compare the results with analytical solutions.
\begin{figure}
\centerline{\scalebox{1.0}
{\includegraphics[width=0.5\textwidth]{Fig2.eps}}}
\caption{\label{fig:fig2} Distributions of maximum differences in position and maximum relative differences in magnification of microimages given by our method and the analytical approach for the first test in Section~\ref{sec:test}. }
\end{figure}
\begin{deluxetable}{ccccccccccccccccccccc}
\tablewidth{0pt}
\tablecaption{Incidence rates of missing microimages. (Only microimages with magnification above $\mu_{\rm limit}$ are shown.).\label{tab:table1}}
\tablehead{
\colhead{} & \colhead{2-image region} & \colhead{4-image region} & \colhead{2-image region} & \colhead{4-image region}\\
\colhead{$r_{l}$} & \colhead{along $y_{1}$-axis} & \colhead{along $y_{1}$-axis} & \colhead{along $y_{2}$-axis} & \colhead{along $y_{2}$-axis}}
\startdata
0.02 & 0/158 & 18/379 & 0/456 & 18/143 \\ \hline
0.01 & 0/158 & \phantom{1}8/379 & 0/456 & 10/143 \\ \hline
0.005 & 0/158 & \phantom{1}2/379 & 0/456 & \phantom{1}4/143 \\
\enddata
\end{deluxetable}
We first examine the incidence of missing microimages. It is realised that there are two types of missing microimages in our method. The first is related to the finite search radius $r_s$. Microimages with magnifications below a certain limit will simply be missed because their corresponding light rays fall outside the search radius $r_s$ (Type I). For the fiducial choice of $r_s=5 r_l$, this lower limit in magnification is $\mu_{\rm limit} =0.01273$. The other type of missing microimages is due to the failure in the root finding process (Type II), which we think is more relevant. Among the 2000 test cases, microimages in all the four-image cases have magnifications above $\mu_{\rm limit}$. For the 621 two-image cases along $y_1$ axis, 158 cases have both microimages with magnifications $\geqslant \mu_{\rm limit}$ and the remaining 463 cases have one microimage with magnification $< \mu_{\rm limit}$. For the 857 two-image cases along $y_2$ axis, 456 cases have both microimages with magnifications $\geqslant \mu_{\rm limit}$ and the remaining 401 cases have one microimage with magnification $< \mu_{\rm limit}$.
We consider three choices of $r_l$, i.e. 0.02$\theta_{E}$, 0.01$\theta_{E}$, and 0.005$\theta_{E}$. The results are summarised in Table \ref{tab:table1}. We find that there is no Type II miss for any of the two-image test cases, regardless of the $r_l$ value. Missing of microimages (Type II) occurs in 6.9\% of all 522 four-image cases when $r_l = 0.02 \theta_{E}$. These numbers drops to 1.1\% when $r_l = 0.005 \theta_{E}$.
For situations that all microimages are successfully identified, we examine the maximum differences in image position and maximum relative differences in magnification obtained by our method (with $r_l=0.02 \theta_{\rm E}$) and the analytical approach in Figure \ref{fig:fig2}. It is clear that our method can reproduce the positions of microimages with an accuracy of better than 1--5$\times 10^{-6} \theta_{\rm E}$, which is essentially determined by the chosen convergence criterion. The magnifications and time delays for individual microimages are calculated analytically in our method, and the maximum relative difference in magnification is smaller than $10^{-\textbf{5}}$ in general.
The tests confirms that our method can identify microimages with a sufficiently high success rate and can accurately predict the position, magnification, and time delay of the identified microimages. The performance of our method generally improves when $r_l$ becomes smaller and/or $r_s$ becomes larger, which at the same time increases the computational cost. In practise, these two parameters need to be optimised in accordance with the computational resource available.
In addition, we compute the total magnifications of microimages successfully identified by our method (with $r_l=0.005 \theta_{\rm E}$) for the 1000 source positions along $y_1$ axis and compare with the analytical predictions in Figure~\ref{fig:add-fig1}. We achieve almost perfect agreement for all source positions except for pixel 311 and pixel 687, which correspond to the two four-image cases with microimage missing in our method (last row in Table~\ref{tab:table1}). We also plot this total magnification curve generated by the IRS method in Figure~\ref{fig:add-fig1}, which shows significant deviations from the analytical predictions at almost all source positions considered. In particular, we notice that the relative differences for the IRS method show consistent fluctuations up to 7\% in the low magnification regions. Nevertheless, those discrepancies from the IRS method are understandable because the IRS method is designed to compute the total magnification for an extended area in the source plane instead of for a point and its accuracy is also limited by the resolution.
\begin{figure}
\centerline{\scalebox{1.0}
{\includegraphics[width=0.96\textwidth]{Fig3.eps}}}
\caption{\label{fig:add-fig1} \emph{(Top)}: Total magnification as a function of source position along the $y_1$ axis, i.e. horizontally along the middle of panel (a) in Figure \ref{fig:fig1}, for the single star case in section \ref{subsec:single-star}. The three curves represent predictions from the analytical solution (thick blue), our method (thin orange), and the IRS method (thin green). The two insets provide zoom-in views of the two caustic crossing regions. \emph{(Bottom)}: The relative differences in magnification of our method (thin orange) and the IRS method (thin green) with respect to the analytical predictions.
}
\end{figure}
\begin{deluxetable}{lllllll}
\tablewidth{0pt}
\tablecaption{Eight sets of parameters used for the test in Section~\ref{subsec:many-star}. $1-\kappa-\gamma$ and $1-\kappa+\gamma$ are the two eigenvalues of the Jacobian matrix. It corresponds to a minima macroimage when $1-\kappa-\gamma$ and $1-\kappa+\gamma$ are both positive and a saddle macroimage when $1-\kappa-\gamma$ and $1-\kappa+\gamma$ have opposite signs. The magnification of the macroimage $\mu$ is simply $\frac{1}{(1-\kappa-\gamma) (1-\kappa+\gamma)}$. \label{tab:paras}}
\tablehead{
\colhead{} & \colhead{$\kappa$, $\gamma$, $\kappa_{\ast}$} & \colhead{$1-\kappa-\gamma$} & \colhead{$1-\kappa+\gamma$} & \colhead{$\mu$} & \colhead{$\kappa_{0}$} & \colhead{macroimage type}}
\startdata
Set 1 & 0.45, 0.35, 0.40 & \phantom{-1}0.2 & \phantom{1}0.9 & \phantom{-2}5.6 & \phantom{1}0.2 & minimum\\
Set 2 & 0.45, 0.35, 0.05 & \phantom{-1}0.2 & \phantom{1}0.9 & \phantom{-2}5.6 & \phantom{1}0.2 & minimum\\
Set 3 & 0.65, 0.55, 0.40 & \phantom{1}-0.2 & \phantom{1}0.9 & \phantom{2}-5.6 & \phantom{1}0.2 & saddle\\
Set 4 & 0.65, 0.55, 0.05 & \phantom{1}-0.2 & \phantom{1}0.9 & \phantom{2}-5.6 & \phantom{1}0.2 & saddle\\
Set 5 & 0.55, 0.40, 0.1 & \phantom{-}0.05 & 0.85 & \phantom{-}23.5 & 0.05 & minimum\\
Set 6 & 0.55, 0.40, 0.0125 & \phantom{-}0.05 & 0.85 & \phantom{-}23.5 & 0.05 & minimum\\
Set 7 & 0.60, 0.45, 0.1 & -0.05 & 0.85 & -23.5 & 0.05 & saddle\\
Set 8 & 0.60, 0.45, 0.0125 & -0.05 & 0.85 & -23.5 & 0.05 & saddle\\\hline
\enddata
\end{deluxetable}
\begin{figure*}
\centerline{\scalebox{1.0}
{\includegraphics[width=0.98\textwidth]{Fig4.eps}}}
\caption{\label{fig:fig3} Magnification maps generated by our method for the eight sets of parameters shown in Table \ref{tab:paras}. They are all shown on the logarithmic scale and the size of each map is $10\theta_{E} \times 10\theta_{E}$ in the source plane. The white dashed line in each map indicates the trajectory of the source corresponding to the light curve in each panel of Figure \ref{fig:fig4}. The open black circle indicates the position of the source corresponding to the light curve in each panel of Figure \ref{fig:fig5}.}
\end{figure*}
\begin{figure*}
\centerline{\scalebox{1.0}
{\includegraphics[width=0.98
\textwidth]{Fig5.eps}}}
\caption{\label{fig:fig4} A comparison of light curves generated by our method and the IRS method. The eight panels in the left column show a direct comparison of 1000 magnifications along the white dashed line in Figure \ref{fig:fig3}. The orange line is produced by our method and the blue line is produced by the IRS method. The eight panels in the right column show the light curves for exactly the same setups except that the source plane resolution in the IRS method is increased by a factor of five. It is clear that light curves from the IRS method agree better with light curves from our method when the resolution is increased.}
\end{figure*}
\subsection{Many stars case}
\label{subsec:many-star}
In the case where there are many stars, the lens equation can not be solved analytically in general. We therefore choose to compare the magnification map generated by our method, i.e. by summing up the magnifications of all identified microimages, with the magnification map produced by the IRS method, which has been widely used in microlensing studies \citep[e.g.,][]{Kayser1986,1987A&A...171...49S}. We produce with our method eight magnification maps for some representative combinations of $\kappa$, $\gamma$ and $\kappa_{\ast}$, which are shown in Table \ref{tab:paras}. Specifically, $\kappa$ and $\gamma$ are chosen to represent minima macroimages in low-magnification regions in Set1 and Set2, saddle macroimages in low-magnification regions in Set3 and Set4, minima macroimages in high-magnification regions in Set5 and Set6, and saddle macroimages in high-magnification regions in Set7 and Set8, respectively. Regarding the choice of $\kappa_{\ast}$, we first introduce a characteristic star convergence $\kappa_0$ at which the average area occupied by each star is comparable to the size of the critical curve of each star. Suppose there are $N$ stars (with mass $M$) in a region under consideration with area $A$. We can define the average area occupied by each star as $A/N$ and the associated radius as $R_{\ast} = \sqrt{A/(\pi \, N)}$. Considering $\kappa_{\ast} = N \, M/(\Sigma_0 \, A)$, one can show that $R_{\ast} = \theta_{\rm E}/\sqrt{\kappa_{\ast}}$. On the other hand, the size of the critical curve of each star is approximately $\theta_{\rm E} / \sqrt{1- \kappa - \gamma}$ \citep[e.g.,][]{2018PhRvD..97b3518O}. The characteristic convergence is simply $\kappa_0 = 1 - \kappa - \gamma$. Few overlaps between the critical curves of the stars are expected when $\kappa_{\ast} < \kappa_0$ and significant overlaps are expected when $\kappa_{\ast} > \kappa_0$. We therefore choose $\kappa_{\ast} = 2 \kappa_0$ in Set1, Set3, Set5, and Set7, and $\kappa_{\ast} = \kappa_0/4$ in Set2, Set4, Set6, and Set8. The corresponding magnification maps are shown in Figure \ref{fig:fig3}. The magnification maps produced by the IRS method are not shown because the differences on those maps are very subtle.
For a better comparison, we examine the variations of the magnification as the source moves along a given trajectory in the source plane, which is commonly referred to as ``light curve" in quasar microlensing. The white dashed line in Figure \ref{fig:fig3} represents the trajectory we choose, and the corresponding light curves inferred from our method and the IRS method are shown in Figure \ref{fig:fig4}. In this test, the source is assumed to be an idealised point source. For the IRS method, we consider two different source plane resolutions. In the left column of Figure \ref{fig:fig4}, the source plane resolution of the IRS method is the same as that of our method, while the resolution of the IRS method in the right column is increased by a factor of five. Overall speaking, the light curves from the two methods are in good agreement. Some discrepancies between our method and the IRS method are seen, especially near the high magnification regions. Those discrepancies become smaller when the resolution of the IRS method is increased, suggesting that our method generally produces more accurate light curves than the IRS method at the same source plane resolution. In addition, the IRS light curve for $\kappa, \gamma, \kappa_{\ast} = 0.45, 0.35, 0.05$, which corresponds to a trajectory with moderate magnifications, show consistent fluctuations again, while the light curve produced by our method for the same set of parameters is smooth. It suggests that the fluctuations are caused by the pixelation in the IRS method, which becomes particularly significant in relatively low magnification regions. With our method, the run time for the eight magnification maps ranges from several minutes to more than one hour. The corresponding run time for the high-resolution IRS method (i.e. IRS2) is from several seconds to more than one hour. Nevertheless, such a run time comparison is not fair because our method computes much more quantities.\\
\section{Microlensing of Short-duration Transients}
\label{sec:discussion}
Now we discuss a special type of microlensing phenomenon --- microlensing of short-duration transient sources such as FRBs. Since the typical duration of FRBs is on the order of millisecond, the relative motion between the lens and source within this time window is negligible. As a result, microlensing of short-duration transient sources is a ``stationary'' process that corresponds to the appearance of individual microimages with time delays. Again, those microimages can not be resolved spatially by current telescopes given the expected $\mu$as separations. However, it is feasible to temporally resolve the microimages as the time resolution of current telescopes have already reached $~\sim100 \mu$s \citep[e.g.,][]{2018MNRAS.478.1209F, 2020MNRAS.492.4752V}.\\
Much different from the ``light curve" (due to the flux change of the back ground source as it moves) commonly mentioned in quasar microlensing, a transient source at some position is expected to generate a microlensed light curve due to the time delays among the microimages and the magnification of each. In this case, the light curve is a superposition of pulses from individual microimages. As an example, we consider a pulse signal at a given source location (indicated by the open black circle in Figure \ref{fig:fig3}) lensed by the eight lensing fields shown in Figure \ref{fig:fig3}. In this example, the source is put at redshift $z_{s}$=0.12, which is the redshift of FRB170827 \citep{2018MNRAS.478.1209F}, and the deflector plane is set at half of the source redshift, i.e. $z_{l}$=0.06. We use our algorithm to identify microimages for each lensing field and compute their magnifications and time delays. We then assume the intrinsic pulse signal as a Gaussian with a width $\sigma$=10$\mu$s and convolve it with the magnification-time delay curves for each lensing field to obtain the light curves, which are shown in Figure \ref{fig:fig5}. The light curves are found to exhibit a wide range of structures, from a single primary peak to double peaks to complex multiple peaks. The relative time delays in these eight examples can reach 0.1--1 ms.
Interestingly, we notice that the light curve for one lensing field ($\kappa, \gamma, \kappa_{\ast} = 0.6, 0.45, 0.1$) appears qualitatively similar to the observed light curve of FRB170827 in terms of the number of peaks and their relative flux ratios and time delays, although \citet{2018MNRAS.478.1209F} interpreted the observed temporal structures of FRB170827 as a result of a scattering screen outside the Milky Way. Nevertheless, as the number of FRBs is going to soar in upcoming surveys, microlensing of FRBs is expected to be discovered in the near future.
In principle, the microlensing light curves encode information on $\kappa$, $\gamma$, $\kappa_{\ast}$, and the stellar mass function of the lensing galaxies at the positions of macroimages. The inference is usually under-constrained in practise given the large amount of nuisance parameters (such as the locations of the stars). Nevertheless, studies have suggested that constraints on the fractions of stars and even planets in lensing galaxies can still be obtained by analysing the observed microlensing light curves of quasars \citep[e.g.,][]{2004ApJ...605...58K, 2008ApJ...689..755M, 2011ApJ...731...71B, 2018ApJ...853L..27D}. Similar to microlensing of quasars, microlensing of FRBs can also be used to constrain those properties in distant galaxies. A major advantage of using FRBs is that the number of FRBs will keep increasing and exceed the number of quasars in a relatively short time. The total number of quasars across the full sky is on the order of $10^7-10^8$ in the LSST era \citep[e.g.,][]{2010MNRAS.405.2579O, Ivezic17}, while the Square Kilometre Array (SKA) is expected to detect $\sim 10^7$ FRBs every year \citep{Hashimoto20}. Unlike quasar microlensing where the light curves usually span a period of years during which the relative motion between the lens and source needs to be taken into account, microlensing of the short-duration FRBs can be considered as a ``stationary'' process. On the one hand, FRB microlensing has two fewer free parameters, i.e. the relative lens-source transverse velocity, which has been shown to be highly correlated with the masses of the stars and the size of the source \citep[e.g.,][]{2004ApJ...605...58K}. A faster exploration of the parameter space is also expected for FRB microlensing events. On the other hand, FRB microlensing does not probe as much the lensing field as quasar microlensing. It is therefore unclear how the constraining power is compared to quasar microlensing. We defer such an analysis to a future paper. It is worth pointing out that some FRBs are found to emit more than one bursts (i.e. repeating FRBs \citep{2016Natur.531..202S,2019Natur.566..235C,2019ApJ...885L..24C,2020ApJ...891L...6F}, and microlensing of those repeating FRBs should provide tighter constraints on the lensing field. Lastly, we note that microlensing of other types of transients such as gamma ray bursts \citep{2000MNRAS.319.1163W}, can be utilised and analysed in the same fashion as long as the duration of the transient event is sufficiently short that the change in the lens-source alignment is negligible.
\begin{figure*}
\centerline{\scalebox{1.0}
{\includegraphics[width=0.98\textwidth]{Fig6.eps}}}
\caption{\label{fig:fig5}Expected light-curves when a FRB is located at the open black circle in the eight magnification maps of Figure \ref{fig:fig3}.}
\end{figure*}
\section{Conclusions}
\label{sec:conclusion}
In this paper, we study a special type of microlensing event where the source is a short-duration transient such as FRB. Different from conventional microlensing studies that are based on light curves induced by the relative motion between the lens and source, microlensing of FRBs produces another type of light curve, which is the superposition of bursts from individual microimages. To identify individual microimages and infer their time delays and magnifications, we implement an algorithm based on two-dimensional Newton iteration method. The performance of our method mainly depends on two parameters, $r_l$ and $r_s$. In practise, they need to be chosen in accordance with the computational resource available. We note that our method is different from the IRS method commonly used in quasar microlensing studies, which can not resolve individual microimages but only provides an estimation of the total magnification.
We perform two tests to validate our method. The first test corresponds to the case of a single point-mass lens embedded in a field of convergence and shear, for which analytical solutions exist. We compare the results given by our method and the analytical solutions for 2000 sources evenly located along the two coordinate axes and find good agreement. Above the magnification lower limit induced by the choice of $r_s$, our method reproduces all the two-image cases. For $r_{l}=0.02\theta_{E}$, our method reproduces 93\% of the four-image cases. The success rate increases to 97\% when $r_l=0.01$ and 99\% when $r_l=0.005$ (Table~\ref{tab:table1}). For the successfully identified microimages, our method can reproduce the image positions with an accuracy of better than 1--5$\times 10^{-6} \theta_{\rm E}$, and the relative difference in the predicted magnification is no larger than $10^{-5}$ in general (Figure~\ref{fig:fig2}). In addition, we find almost perfect agreement on the total magnifications between our method and the analytical solution. The total magnifications from the IRS method show noticeable deviations from the analytical solution in both low- and high-magnification regions.
The second test corresponds to a more general situation with multiple point-mass lenses, for which analytical solutions are not available. We test our method by comparing magnification maps and light curves generated by our method with those generated with the IRS method for eight example lensing fields (Figure~\ref{fig:fig3}). We find good agreement in general despite the completely different strategies of generating the light curves in the two methods (Figure~\ref{fig:fig4}). The agreement improves when the source plane resolution in the IRS method is increased.
We further show some model light curves for FRBs that are microlensed by the eight example lensing fields used in the second test (Figure~\ref{fig:fig5}). The light curves exhibit a wide range of structures, from a single primary peak to double peaks to complex multiple peaks. The relative time delays in these eight examples can reach 0.1--1 ms, which can already be resolved by current telescopes. Although not yet discovered, microlensing of FRBs is expected to become regular events in the upcoming SKA era when $\sim 10^{4-5}$ FRBs will be detected on a daily basis. Light curves produced by our method are therefore essential for constraining the stellar fraction and stellar mass distribution in distant galaxies with FRB microlensing.
\section{Acknowledgements}
The authors thank the anonymous referee for helpful comments and thank Xuefeng Wu, Jun Zhang and Zuhui Fan for helpful discussions and suggestions. This work is supported by the NSFC (No. 11673065, U1931210, 11273061). We acknowledge the cosmology simulation database (CSD) in the National Basic Science Data Center (NBSDC) and its funds the NBSDC-DB-10 (No. 2020000088). We acknowledge the science research grants from the China Manned Space Project with NO.CMS-CSST-2021-A13. Yiping Shu acknowledges support from the Max Planck Society and the Alexander von Humboldt Foundation in the framework of the Max Planck--Humboldt Research Award endowed by the Federal Ministry of Education and Research.
|
\section{Introduction} \label{sec:intro}
In recent years, drug discovery has drawn increasing interest in the machine learning community. Among many challenges therein, how to discriminatively represent a molecule with a vectorized embedding remains a fundamental yet open challenge. The underlying problem can be decomposed into two components: how to design a common latent space for molecule graphs ({\em{i.e.}}, designing a suitable encoder) and how to construct an objective function to supervise the training ({\em{i.e.}}, defining a learning target). Falling broadly into the second category, our paper studies self-supervised molecular representation learning by leveraging the consistency between 3D geometry and 2D topology.
Motivated by the prominent success of the pretraining-finetuning pipeline~\citep{devlin2018bert}, unsupervisedly pre-trained graph neural networks for molecules yields promising performance on downstream tasks and becomes increasingly popular~\citep{velivckovic2018deep,sun2019infograph,liu2018n,hu2019strategies,You2020GraphCL,you2021graph}. The key to pre-training lies in finding an effective proxy task ({\em{i.e.}}, training objective) to leverage the power of large unlabeled datasets. Inspired by \citep{schutt2017schnet,liu2018n,liu2021spherical} that molecular properties~\citep{gilmer2017neural,liu2018n} can be better predicted by 3D geometry due to its encoded energy knowledge, we aim to make use of the 3D geometry of molecules in pre-training. However, the stereochemical structures are often very expensive to obtain, making such 3D geometric information scarce in downstream tasks. To address this problem, we propose the Graph Multi-View Pre-training (GraphMVP) framework, where a 2D molecule encoder is pre-trained with the knowledge of 3D geometry and then fine-tuned on downstream tasks without 3D information. Our learning paradigm, during pre-training, injects the knowledge of 3D molecular geometry to a 2D molecular graph encoder such that the downstream tasks can benefit from the implicit 3D geometric prior even if there is no 3D information available.
We attain the aforementioned goal by leveraging two pretext tasks on the 2D and 3D molecular graphs: one contrastive and one generative SSL. Contrastive SSL creates the supervised signal at an \textbf{inter-molecule} level: the 2D and 3D graph pairs are positive if they are from the same molecule, and negative otherwise; Then contrastive SSL~\citep{wang2020understanding} will align the positive pairs and contrast the negative pairs simultaneously. Generative SSL~\citep{vincent2008extracting,kingma2013auto,higgins2016beta}, on the other hand, obtains the supervised signal in an \textbf{intra-molecule} way: it learns a 2D/3D representation that can reconstruct its 3D/2D counterpart view for each molecule itself.
To cope with the challenge of measuring the quality of reconstruction on molecule 2D and 3D space, we further propose a novel surrogate objective function called variation representation reconstruction (VRR) for the generative SSL task, which can effectively measure such quality in the continuous representation space. The knowledge acquired by these two SSL tasks is complementary, so our GraphMVP\ framework integrates them to form a more discriminative 2D molecular graph representation.
Consistent and significant performance improvements empirically validate the effectiveness of GraphMVP.
We give additional insights to justify the effectiveness of GraphMVP. First, GraphMVP\, is a self-supervised learning approach based on maximizing mutual information (MI) between 2D and 3D views, enabling the learnt representation to capture high-level factors~\citep{belghazi2018mutual,tschannen2019mutual,bachman2019learning} in molecule data. Second, we find that 3D molecular geometry is a form of privileged information~\cite{vapnik2009new,vapnik2015learning}. It has been proven that using privileged information in training can accelerate the speed of learning. We note that privileged information is only used in training, while it is not available in testing. This perfectly matches our intuition of pre-training molecular representation with 3D geometry.
Our contributions include
(1) To our best knowledge, we are the first to incorporate the 3D geometric information into graph SSL;
(2) We propose one contrastive and one generative SSL tasks for pre-training. Then we elaborate their difference and empirically validate that combining both can lead to a better representation;
(3) We provide theoretical insights and case studies to justify why adding 3D geometry is beneficial;
(4) We achieve the SOTA performance among all the SSL methods.
\textbf{Related work.}
We briefly review the most related works here and include a more detailed summarization in~\Cref{appendix:related_work}. Self-supervised learning (SSL) methods have attracted massive attention to graph applications~\citep{liu2021graph,xie2021self,wu2021self,liu2021self}. In general, there are roughly two categories of graph SSL: contrastive and generative, where they differ on the design of the supervised signals. Contrastive graph SSL~\citep{velivckovic2018deep,sun2019infograph,hu2019strategies,You2020GraphCL,you2021graph} constructs the supervised signals at the \textbf{inter-graph} level and learns the representation by contrasting with other graphs, while generative graph SSL~\citep{hamilton2017inductive,liu2018n,hu2019strategies,hu2020gpt} focuses on reconstructing the original graph at the \textbf{intra-graph} level. One of the most significant differences that separate our work from existing methods is that all previous methods \textbf{merely} focus on 2D molecular topology. However, for scientific tasks such as molecular property prediction, 3D geometry should be incorporated as it provides complementary and comprehensive information~\citep{schutt2017schnet,liu2021spherical}. To fill this gap, we propose GraphMVP to leverage the 3D geometry in graph self-supervised pre-training.
\section{Preliminaries} \label{sec:preliminary}
We first outline the key concepts and notations used in this work. Self-supervised learning (SSL) is based on the \textit{view} design, where each view provides a specific aspect and modality of the data. Each molecule has two natural views: the 2D graph incorporates the topological structure defined by the adjacency, while the 3D graph can better reflect the geometry and spatial relation. From a chemical perspective, 3D geometric graphs focus on the \textit{energy} while 2D graphs emphasize the \textit{topological} information; thus they can be composed for learning more informative representation in GraphMVP. \textit{Transformation} is an atomic operation in SSL that can extract specific information from each view. Next, we will briefly introduce how to represent these two views.
\textbf{2D Molecular Graph} represents molecules as 2D graphs, with atoms as nodes and bonds as edges respectively. We denote each 2D graph as $g_{\text{2D}} = (X, E)$, where $X$ is the atom attribute matrix and $E$ is the bond attribute matrix. Notice that here $E$ also includes the bond connectivity. Then we will apply a transformation function $T_{\text{2D}}$ on the topological graph. Given a 2D molecular graph $g_{\text{2D}}$, its representation $h_{2D}$ can be obtained from a \textit{2D graph neural network (GNN)} model:
\begin{equation} \label{eq:2d_gnn}
h_{\text{2D}} = \text{GNN-2D}(T_{\text{2D}}(g_{\text{2D}})) = \text{GNN-2D}(T_{\text{2D}}(X, E)).
\end{equation}
\textbf{3D Molecular Graph} additionally includes spatial positions of the atoms, and they are needless to be static since atoms are in continual motion on \textit{a potential energy surface}~\citep{axelrod2020geom}. \footnote{A more rigorous way of defining conformer is in~\citep{moss1996basic}: a conformer is an isomer of a molecule that differs from another isomer by the rotation of a single bond in the molecule.} The 3D structures at the local minima on this surface are named \textit{conformer}. As the molecular properties are conformers ensembled~\citep{hawkins2017conformation}, GraphMVP\, provides a novel perspective on adopting 3D conformers for learning better representation. Given a conformer $g_{\text{3D}} = (X, R)$, its representation via a \textit{3D GNN} model is:
\begin{equation} \label{eq:3d_gnn}
h_{\text{3D}} = \text{GNN-3D}(T_{\text{3D}}(g_{\text{3D}})) = \text{GNN-3D}(T_{\text{3D}}(X, R)),
\end{equation}
where $R$ is the 3D-coordinate matrix and $T_{\text{3D}}$ is the 3D transformation.
In what follows, for notation simplicity, we use ${\bm{x}}$ and ${\bm{y}}$ for the 2D and 3D graphs, {\em{i.e.}}, ${\bm{x}} \triangleq g_{\text{2D}}$ and ${\bm{y}} \triangleq g_{\text{3D}}$. Then the latent representations are denoted as $h_{\bm{x}}$ and $h_{\bm{y}}$.
\section{GraphMVP: Graph Multi-View Pre-training} \label{sec:method}
Our model, termed as Graph Multi-View Pre-training (GraphMVP), conducts self-supervised learning (SSL) pre-training with 3D information. The 3D conformers encode rich information about the molecule energy and spatial structure, which are complementary to the 2D topology. Thus, applying SSL between the 2D and 3D views will provide a better 2D representation, which implicitly embeds the ensembles of energies and geometric information for molecules.
In the following, we first present an overview of GraphMVP, and then introduce two pretext tasks specialized concerning 3D conformation structures. Finally, we summarize a broader graph SSL family that prevails the 2D molecular graph representation learning with 3D geometry.
\subsection{Overview of GraphMVP} \label{sec:overview}
\begin{figure}[ht]
\centering
\vspace{-2ex}
\includegraphics[width=\textwidth]{figures/Diagram_Bundle_4.pdf}
\vspace{-2ex}
\caption{
Overview of the pre-training stage in GraphMVP. The black dashed circles denote subgraph masking, and we mask the same region in the 2D and 3D graphs. Multiple views of the molecules (herein: Halicin) are mapped to the representation space via 2D and 3D GNN models, where we conduct GraphMVP\ for SSL pre-training, using both contrastive and generative pretext tasks.
}
\label{fig:both_ssl}
\end{figure}
In general, GraphMVP\, exerts 2D topology and 3D geometry as two complementary views for each molecule. By performing SSL between these views, it is expected to learn a 2D representation enhanced with 3D conformation, which can better reflect certain molecular properties.
As generic SSL pre-training pipelines, GraphMVP\, has two stages: pre-training then fine-tuning. In the pre-training stage, we conduct SSL via auxiliary tasks on data collections that provide both 2D and 3D molecular structures. During fine-tuning, the pre-trained 2D GNN models are subsequently fine-tuned on specific downstream tasks, where only 2D molecular graphs are available.
At the SSL pre-training stage, we design two pretext tasks: one contrastive and one generative. We conjecture and then empirically prove that these two tasks are focusing on different learning aspects, which are summarized into the following two points. (1) From the perspective of representation learning, contrastive SSL utilizes \textbf{inter-data} knowledge and generative SSL utilizes \textbf{intra-data} knowledge. For contrastive SSL, one key step is to obtain the negative view pairs for inter-data contrasting; while generative SSL focuses on each data point itself, by reconstructing the key features at an intra-data level. (2) From the perspective of distribution learning, contrastive SSL and generative SSL are learning the data distribution from a \textbf{local} and \textbf{global} manner, respectively. Contrastive SSL learns the distribution locally by contrasting the pairwise distance at an inter-data level. Thus, with sufficient number of data points, the local contrastive operation can iteratively recover the data distribution. Generative SSL, on the other hand, learns the global data density function directly.
Therefore, contrastive and generative SSL are essentially conducting representation and distribution learning with different intuitions and disciplines, and we expect that combining both can lead to a better representation. We later carry out an ablation study (\Cref{sec:experiment_each_loss_component}) to verify this empirically.
In addition, to make the pretext tasks more challenging, we take views for each molecule by randomly masking $M$ nodes (and corresponding edges) as the transformation function, {\em{i.e.}}, $T_{\text{2D}} = T_{\text{3D}} =\text{mask}$. This trick has been widely used in graph SSL~\citep{hu2019strategies,You2020GraphCL,you2021graph} and has shown robust improvements.
\subsection{Contrastive Self-Supervised Learning between 2D and 3D Views} \label{sec:contrastive_SSL}
The main idea of contrastive self-supervised learning (SSL)~\citep{oord2018representation,chen2020simple} is first to define positive and negative pairs of views from an inter-data level, and then to align the positive pairs and contrast the negative pairs simultaneously~\citep{wang2020understanding}. For each molecule, we first extract representations from 2D and 3D views, {\em{i.e.}}, $h_{\bm{x}}$ and $h_{\bm{y}}$. Then we create positive and negative pairs for contrastive learning: the 2D-3D pairs $({\bm{x}},{\bm{y}})$ for the same molecule are treated as positive, and negative otherwise. Finally, we align the positive pairs and contrast the negative ones. The pipeline is shown in~\Cref{fig:both_ssl}. In the following, we discuss two common objective functions on contrastive graph SSL.
\textbf{InfoNCE} is first proposed in~\citep{oord2018representation}, and its effectiveness has been validated both empirically~\citep{chen2020simple,he2020momentum} and theoretically~\citep{arora2019theoretical}. Its formulation is given as follows:
\begin{equation}\label{eq:objective_infonce}
\small{
\fontsize{7.8}{1}\selectfont
\begin{aligned}
\mathcal{L}_{\text{InfoNCE}}=-\frac{1}{2} \mathbb{E}_{p({\bm{x}},{\bm{y}})}\Big[\log \frac{\exp(f_{{\bm{x}}}({\bm{x}}, {\bm{y}}))}{\exp(f_{{\bm{x}}}({\bm{x}}, {\bm{y}})) + \sum\limits_{j} \exp(f_{{\bm{x}}}({\bm{x}}^{j},{\bm{y}})})+\log\frac{\exp(f_{{\bm{y}}}({\bm{y}},{\bm{x}}))}{\exp(f_{{\bm{y}}}({\bm{y}},{\bm{x}})) + \sum\limits_{j} \exp(f_{{\bm{y}}}({\bm{y}}^{j},{\bm{x}}))} \Big],
\end{aligned}
}
\end{equation}
where ${\bm{x}}^{j}, {\bm{y}}^{j}$ are randomly sampled 2D and 3D views regarding to the anchored pair $({\bm{x}},{\bm{y}})$. $f_{{\bm{x}}}({\bm{x}},{\bm{y}})$ and $f_{{\bm{y}}}({\bm{y}},{\bm{x}})$ are scoring functions for the two corresponding views, with flexible formulations. Here we adopt $f_{\bm{x}}({\bm{x}},{\bm{y}}) = f_{\bm{y}}({\bm{y}},{\bm{x}}) = \langle h_{\bm{x}}, h_{\bm{y}} \rangle$. More details are in~\Cref{sec:app:contrastive_ssl}.
\textbf{Energy-Based Model with Noise Contrastive Estimation (EBM-NCE)} is an alternative that has been widely used in the line of graph contrastive SSL~\citep{sun2019infograph,hu2019strategies,You2020GraphCL,you2021graph}. Its intention is essentially the same as InfoNCE, to align positive pairs and contrast negative pairs, while the main difference is the usage of binary cross-entropy and extra noise distribution for negative sampling:
\begin{equation} \label{eq:objective_ebm_nce}
\small{
\begin{aligned}
\mathcal{L}_{\text{EBM-NCE}}
& = -\frac{1}{2} \mathbb{E}_{p({\bm{y}})} \Big[\mathbb{E}_{p_{n}({\bm{x}}|{\bm{y}})} \log \big(1-\sigma( f_x({\bm{x}}, {\bm{y}}))\big) + \mathbb{E}_{p({\bm{x}}|{\bm{y}} )} \log \sigma( f_x({\bm{x}}, {\bm{y}})) \Big]\\
& ~~~~~ -\frac{1}{2} \mathbb{E}_{p({\bm{x}})} \Big[\mathbb{E}_{p_{n}({\bm{y}}|{\bm{x}})} \log \big(1-\sigma( f_y({\bm{y}},{\bm{x}}))\big) + \mathbb{E}_{p({\bm{y}},{\bm{x}})} \log \sigma( f_y({\bm{y}},{\bm{x}})) \Big],
\end{aligned}
}
\end{equation}
where $p_n$ is the noise distribution and $\sigma$ is the sigmoid function. We also notice that the final formulation of EBM-NCE shares certain similarities with Jensen-Shannon estimation (JSE)~\citep{nowozin2016f}. However, the derivation process and underlying intuition are different: EBM-NCE models the conditional distributions in MI lower bound (\Cref{eq:MI_objective}) with EBM, while JSE is a special case of variational estimation of f-divergence. Since this is not the main focus of GraphMVP, we expand the a more comprehensive comparison in~\Cref{sec:app:contrastive_ssl}, plus the potential benefits with EBM-NCE.
Few works~\citep{hassani2020contrastive} have witnessed the effect on the choice of objectives in graph contrastive SSL. In GraphMVP, we treat it as a hyper-parameter and further run ablation studies on them, {\em{i.e.}}, to solely use either InfoNCE ($\mathcal{L}_{\text{C}} = \mathcal{L}_{\text{InfoNCE}}$) or EMB-NCE ($\mathcal{L}_{\text{C}} = \mathcal{L}_{\text{EBM-NCE}}$).
\subsection{Generative Self-Supervised Learning between 2D and 3D Views} \label{sec:generative_SSL}
Generative SSL is another classic track for unsupervised pre-training~\citep{kingma2013auto,chen2016infogan,larsson2016learning,kingma2018glow}. It aims at learning an effective representation by self-reconstructing each data point. Specifically to drug discovery, we have one 2D graph and a certain number of 3D conformers for each molecule, and our goal is to learn a robust 2D/3D representation that can, to the most extent, recover its 3D/2D counterparts. By doing so, generative SSL can enforce 2D/3D GNN to encode the most crucial geometry/topology information, which can improve the downstream performance.
There are many options for generative models, including variational auto-encoder (VAE)~\citep{kingma2013auto}, generative adversarial networks (GAN)~\citep{goodfellow2014generative}, flow-based model~\citep{dinh2016density}, etc. In GraphMVP, we prefer VAE-like method for the following reasons: (1) The mapping between two molecular views is stochastic: multiple 3D conformers correspond to the same 2D topology; (2) An explicit 2D graph representation ({\em{i.e.}}, feature encoder) is required for downstream tasks; (3) Decoders for structured data such as graph are often highly nontrivial to design, which make them a suboptimal choice.
\textbf{Variational Molecule Reconstruction.}
Therefore we propose a \textit{light} VAE-like generative SSL, equipped with a \textit{crafty} surrogate loss, which we describe in the following. We start with an example for illustration. When generating 3D conformers from their corresponding 2D topology, we want to model the conditional likelihood $p({\bm{y}}|{\bm{x}})$. By introducing a reparameterized variable ${\bm{z}}_{\bm{x}} = \mu_{{\bm{x}}} + \sigma_{{\bm{x}}} \odot \epsilon$, where $\mu_{\bm{x}}$ and $\sigma_{\bm{x}}$ are two flexible functions on $h_{\bm{x}}$, $\epsilon \sim \mathcal{N}(0,I)$ and $\odot$ is the element-wise production,
we have the following lower bound:
\begin{equation} \label{eq:variational_lower_bound}
\small
\begin{aligned}
\log p({\bm{y}}|{\bm{x}})
\ge \mathbb{E}_{q({\bm{z}}_{\bm{x}}|{\bm{x}})} \big[ \log p({\bm{y}}|{\bm{z}}_{\bm{x}}) \big] - KL(q({\bm{z}}_{\bm{x}}|{\bm{x}}) || p({\bm{z}}_{\bm{x}})).
\end{aligned}
\end{equation}
The expression for $\log p({\bm{x}}|{\bm{y}})$ can be similarly derived. \Cref{eq:variational_lower_bound} includes a conditional log-likelihood and a KL-divergence term, where the bottleneck is to calculate the first term for structured data. This term has also been recognized as the \textit{reconstruction term}: it is essentially to reconstruct the 3D conformers (${\bm{y}}$) from the sampled 2D molecular graph representation (${\bm{z}}_{{\bm{x}}}$). However, performing the graph reconstruction on the data space is not trivial: since molecules ({\em{e.g.}}, atoms and bonds) are discrete, modeling and measuring on the molecule space will bring extra obstacles.
\textbf{Variational Representation Reconstruction (VRR).}
To cope with this challenge, we propose a novel surrogate loss by switching the reconstruction from data space to representation space. Instead of decoding the latent code $z_{{\bm{x}}}$ to data space, we can directly project it to the 3D representation space, denoted as $q_{\bm{x}}(z_{\bm{x}})$. Since the representation space is continuous, we may as well model the conditional log-likelihood with Gaussian distribution, resulting in L2 distance for reconstruction, {\em{i.e.}}, $\|q_{{\bm{x}}}(z_{{\bm{x}}}) - \text{SG}(h_{\bm{y}}({\bm{y}}))\|^2$. Here SG is the stop-gradient operation, assuming that $h_{{\bm{y}}}$ is a fixed learnt representation function. SG has been widely adopted in the SSL literature to avoid model collapse~\citep{grill2020bootstrap,chen2021exploring}. We call this surrogate loss as variational representation reconstruction (VRR):
\begin{equation} \label{eq:variational_lower_bound_approximation}
\small
\begin{aligned}
\mathcal{L}_{\text{G}} = \mathcal{L}_{\text{VRR}}
= & \frac{1}{2} \Big[ \mathbb{E}_{q({\bm{z}}_{\bm{x}}|{\bm{x}})} \big[ \| q_{{\bm{x}}}({\bm{z}}_{\bm{x}}) - \text{SG}(h_{{\bm{y}}}) \|^2 \big] + \mathbb{E}_{q({\bm{z}}_{\bm{y}}|{\bm{y}})} \big[ \| q_{{\bm{y}}}({\bm{z}}_{\bm{y}}) - \text{SG}(h_{{\bm{x}}}) \|_2^2 \big] \Big]\\
& + \frac{\beta}{2} \cdot \Big[ KL(q({\bm{z}}_{\bm{x}}|{\bm{x}}) || p({\bm{z}}_{\bm{x}})) + KL(q({\bm{z}}_{\bm{y}}|{\bm{y}}) || p({\bm{z}}_{\bm{y}})) \Big].
\end{aligned}
\end{equation}
We give a simplified illustration for the generative SSL pipeline in~\Cref{fig:both_ssl} and the complete derivations in~\Cref{sec:app:generative_ssl}. As will be discussed in~\Cref{sec:mutual_information}, VRR is actually maximizing MI, and MI is invariant to continuous bijective function~\citep{belghazi2018mutual}. Thus, this surrogate loss would be exact if the encoding function $h$ satisfies this condition. However, we find that GNN, though does not meet the condition, can provide quite robust performance, which empirically justify the effectiveness of VRR.
\subsection{Multi-task Objective Function}
As discussed before, contrastive SSL and generative SSL essentially learn the representation from distinct viewpoints. A reasonable conjecture is that combining both SSL methods can lead to overall better performance, thus we arrive at minimizing the following complete objective for GraphMVP:
\begin{equation} \label{eq:graphmvp}
\small
\begin{aligned}
\mathcal{L}_{\text{GraphMVP}}
& = \alpha_1 \cdot \mathcal{L}_{\text{C}} + \alpha_2 \cdot \mathcal{L}_{\text{G}},\\
\end{aligned}
\end{equation}
where $\alpha_1, \alpha_2$ are weighting coefficients. A later performed ablation study (\Cref{sec:experiment_each_loss_component}) delivers two important messages: (1) Both individual contrastive and generative SSL on 3D conformers can consistently help improve the 2D representation learning; (2) Combining the two SSL strategies can yield further improvements. Thus, we draw the conclusion that GraphMVP\, (\Cref{eq:graphmvp}) is able to obtain an augmented 2D representation by fully utilizing the 3D information.
As discussed in~\Cref{sec:intro}, existing graph SSL methods only focus on the 2D topology, which is in parallel to GraphMVP: 2D graph SSL focuses on exploiting the 2D structure topology, and GraphMVP\, takes advantage of the 3D geometry information. Thus, we propose to merge the 2D SSL into GraphMVP. Since there are two main categories in 2D graph SSL: generative and contrastive, we propose two variants GraphMVP-G\, and GraphMVP-C\, accordingly. Their objectives are as follows:
\begin{equation} \label{eq:graphmvp_variants}
\small
\begin{aligned}
\mathcal{L}_{\text{GraphMVP-G}} = \mathcal{L}_{\text{GraphMVP}} + \alpha_3 \cdot \mathcal{L}_{\text{Generative 2D-SSL}},~~~~~
\mathcal{L}_{\text{GraphMVP-C}} = \mathcal{L}_{\text{GraphMVP}} + \alpha_3 \cdot \mathcal{L}_{\text{Contrastive 2D-SSL}}.
\end{aligned}
\end{equation}
Later, the empirical results also help support the effectiveness of GraphMVP-G\, and GraphMVP-C, and thus, we can conclude that existing 2D SSL is complementary to GraphMVP.
\section{Experiments and Results} \label{sec:experiments}
\subsection{Experimental Settings} \label{sec:experiments_setting}
\textbf{Datasets.} We pre-train models on the same dataset then fine-tune on the wide range of downstream tasks. We randomly select 50k qualified molecules from GEOM~\citep{axelrod2020geom} with both 2D and 3D structures for the pre-training. As clarified in~\Cref{sec:overview}, conformer ensembles can better reflect the molecular property, thus we take $C$ conformers of each molecule. For downstream tasks, we first stick to the same setting of the main graph SSL work~\citep{hu2019strategies,You2020GraphCL,you2021graph}, exploring 8 binary molecular property prediction tasks, which are all in the low-data regime. Then we explore 6 regression tasks from various low-data domains to be more comprehensive. We describe all the datasets in~\Cref{app:sec:dataset}.
\textbf{2D GNN.} We follow the research line of SSL on molecule graph~\citep{hu2019strategies,you2021graph,You2020GraphCL}, using the same Graph Isomorphism Network (GIN)~\citep{xu2018powerful} as the backbone model, with the same feature sets.
\textbf{3D GNN.} We choose SchNet~\citep{schutt2017schnet} for geometric modeling, since SchNet: (1) is found to be a strong geometric representation learning method under the fair benchmarking; (2) can be trained more efficiently, comparing to the other recent 3D models. More detailed explanations are in~\Cref{appendix:3d_gnn}.
\subsection{Main Results on Molecular Property Prediction.} \label{sec:main_results}
We carry out comprehensive comparisons with 10 SSL baselines and random initialization. For pre-training, we apply all SSL methods on the same dataset based on GEOM~\citep{axelrod2020geom}. For fine-tuning, we follow the same setting~\citep{hu2019strategies,You2020GraphCL,you2021graph} with 8 low-data molecular property prediction tasks.
\textbf{Baselines.} Due to the rapid growth of graph SSL~\citep{liu2021graph,xie2021self,wu2021self}, we are only able to benchmark the most well-acknowledged baselines: EdgePred~\citep{hamilton2017inductive}, InfoGraph~\citep{sun2019infograph}, GPT-GNN\citep{hu2020gpt}, AttrMask \& ContextPred\citep{hu2019strategies}, GraphLoG\citep{xu2021self}, G-\{Contextual, Motif\}\citep{rong2020self}, GraphCL\citep{You2020GraphCL}, JOAO\citep{you2021graph}.
\textbf{Our method.} GraphMVP\, has two key factors: i) masking ratio ($M$) and ii) number of conformers for each molecule ($C$). We set $M=0.15$ and $C=5$ by default, and will explore their effects in the following ablation studies in~\Cref{sec:effect_of_M_and_C}. For EBM-NCE loss, we adopt the empirical distribution for noise distribution. For~\Cref{eq:graphmvp_variants}, we pick the empirically optimal generative and contrastive 2D SSL method: that is AttrMask for GraphMVP-G\, and ContextPred for GraphMVP-C.
\begin{table}[t!]
\caption{
Results for molecular property prediction tasks.
For each downstream task, we report the mean (and standard deviation) ROC-AUC of 3 seeds with scaffold splitting.
For GraphMVP\,, we set $M=0.15$ and $C=5$.
The best and second best results are marked \textbf{\underline{bold}} and \textbf{bold}, respectively.
\vspace{-0.24cm}
}
\label{tab:main_results}
\small
\setlength{\tabcolsep}{3pt}
\centering
\begin{tabular}{l c c c c c c c c c}
\toprule
Pre-training & BBBP & Tox21 & ToxCast & Sider & ClinTox & MUV & HIV & Bace & Avg\\
\midrule
-- & 65.4(2.4) & 74.9(0.8) & 61.6(1.2) & 58.0(2.4) & 58.8(5.5) & 71.0(2.5) & 75.3(0.5) & 72.6(4.9) & 67.21 \\
\midrule
EdgePred & 64.5(3.1) & 74.5(0.4) & 60.8(0.5) & 56.7(0.1) & 55.8(6.2) & 73.3(1.6) & 75.1(0.8) & 64.6(4.7) & 65.64 \\
AttrMask & 70.2(0.5) & 74.2(0.8) & 62.5(0.4) & 60.4(0.6) & 68.6(9.6) & 73.9(1.3) & 74.3(1.3) & 77.2(1.4) & 70.16 \\
GPT-GNN & 64.5(1.1) & \textbf{75.3(0.5)} & 62.2(0.1) & 57.5(4.2) & 57.8(3.1) & 76.1(2.3) & 75.1(0.2) & 77.6(0.5) & 68.27 \\
InfoGraph & 69.2(0.8) & 73.0(0.7) & 62.0(0.3) & 59.2(0.2) & 75.1(5.0) & 74.0(1.5) & 74.5(1.8) & 73.9(2.5) & 70.10 \\
ContextPred & 71.2(0.9) & 73.3(0.5) & 62.8(0.3) & 59.3(1.4) & 73.7(4.0) & 72.5(2.2) & 75.8(1.1) & 78.6(1.4) & 70.89 \\
GraphLoG & 67.8(1.7) & 73.0(0.3) & 62.2(0.4) & 57.4(2.3) & 62.0(1.8) & 73.1(1.7) & 73.4(0.6) & 78.8(0.7) & 68.47 \\
G-Contextual & 70.3(1.6) & 75.2(0.3) & 62.6(0.3) & 58.4(0.6) & 59.9(8.2) & 72.3(0.9) & 75.9(0.9) & 79.2(0.3) & 69.21 \\
G-Motif & 66.4(3.4) & 73.2(0.8) & 62.6(0.5) & 60.6(1.1) & 77.8(2.0) & 73.3(2.0) & 73.8(1.4) & 73.4(4.0) & 70.14 \\
GraphCL & 67.5(3.3) & 75.0(0.3) & 62.8(0.2) & 60.1(1.3) & 78.9(4.2) & \textbf{77.1(1.0)} & 75.0(0.4) & 68.7(7.8) & 70.64 \\
JOAO & 66.0(0.6) & 74.4(0.7) & 62.7(0.6) & 60.7(1.0) & 66.3(3.9) & 77.0(2.2) & \textbf{76.6(0.5)} & 72.9(2.0) & 69.57 \\
\midrule
GraphMVP\, & 68.5(0.2) & 74.5(0.4) & 62.7(0.1) & \textbf{62.3(1.6)} & \textbf{79.0(2.5)} & 75.0(1.4) & 74.8(1.4) & 76.8(1.1) & 71.69 \\
GraphMVP-G & \textbf{70.8(0.5)} & \textbf{\underline{75.9(0.5)}} & \textbf{\underline{63.1(0.2)}} & 60.2(1.1) & \textbf{\underline{79.1(2.8)}} & \textbf{\underline{77.7(0.6)}} & 76.0(0.1) & \textbf{79.3(1.5)} & \textbf{72.76} \\
GraphMVP-C & \textbf{\underline{72.4(1.6)}} & 74.4(0.2) & \textbf{\underline{63.1(0.4)}} & \textbf{\underline{63.9(1.2)}} & 77.5(4.2) & 75.0(1.0) & \textbf{\underline{77.0(1.2)}} & \textbf{\underline{81.2(0.9)}} & \textbf{\underline{73.07}} \\
\bottomrule
\end{tabular}
\vspace{-2ex}
\end{table}
The main results on 8 molecular property prediction tasks are listed in~\Cref{tab:main_results}. We observe that the performance of GraphMVP\, is significantly better than the random initialized one, and the average performance outperforms the existing SSL methods by a large margin. In addition, GraphMVP-G\, and GraphMVP-C\, consistently improve the performance, supporting the claim: \textbf{3D geometry is complementary to the 2D topology}. GraphMVP\, leverages the information between 3D geometry and 2D topology, and 2D SSL plays the role as regularizer to extract more 2D topological information; they are extracting different perspectives of information and are indeed complementary to each other.
\subsection{Ablation Study: The Effect of Masking Ratio and Number of Conformers} \label{sec:effect_of_M_and_C}
We analyze the effects of masking ratio $M$ and the number of conformers $C$ in GraphMVP. In~\Cref{tab:main_results}, we set the $M$ as 0.15 since it has been widely used in existing SSL methods~\citep{hu2019strategies,You2020GraphCL,you2021graph}, and $C$ is set to 5, which we will explain below. We explore on the range of $M \in \{0, 0.15, 0.3\}$ and $C \in \{1, 5, 10, 20\}$, and report the average performance. The complete results are in~\Cref{app:sec:effect_of_M_and_C}.
\begin{table}[t!]
\fontsize{8.5}{4}\selectfont
\centering
\vspace{-2ex}
\begin{minipage}[t]{.48\linewidth}
\setlength{\tabcolsep}{5.3pt}
\renewcommand\arraystretch{2.2}
\caption{Ablation of masking ratio $M$, $C\equiv5$.\vspace{-0.3cm}}
\label{tab:ablation_masking}
\centering
\begin{tabular}{c c c c}
\toprule
$M$ & GraphMVP & GraphMVP-G & GraphMVP-C\\
\midrule
0 & 71.12 & 72.15 & 72.66\\
0.15 & 71.60 & 72.76 & 73.08\\
0.30 & 71.79 & 72.91 & 73.17\\
\bottomrule
\end{tabular}
\end{minipage}
\hfill
\begin{minipage}[t]{.48\linewidth}
\setlength{\tabcolsep}{5.5pt}
\caption{Ablation of \# conformer $C$, $M\equiv0.15$.\vspace{-0.3cm}}
\label{tab:ablation_n_conformers}
\centering
\begin{tabular}{c c c c}
\toprule
$C$ & GraphMVP & GraphMVP-G & GraphMVP-C\\
\midrule
1 & 71.61 & 72.80 & 72.46\\
5 & 71.60 & 72.76 & 73.08\\
10 & 72.20 & 72.59 & 73.09\\
20 & 72.39 & 73.00 & 73.02\\
\bottomrule
\end{tabular}
\end{minipage}
\vspace{-1ex}
\end{table}
As seen in~\Cref{tab:ablation_masking}, the improvement is more obvious from $M=0$ (raw graph) to $M=0.15$ than from $M=0.15$ to $M=0.3$. This can be explained that subgraph masking with larger ratio will make the SSL tasks more challenging, especially comparing to the raw graph ($M=0$).
\Cref{tab:ablation_n_conformers} shows the effect for $C$. We observe that the performance is generally better when adding more conformers, but will reach a plateau above certain thresholds. This observation matches with previous findings~\citep{axelrod2020molecular}: adding more conformers to augment the representation learning is not as helpful as expected; while we conclude that adding more conformers can be beneficial with little improvement. One possible reason is, when generating the dataset, we are sampling top-$C$ conformers with highest possibility and lowest energy. In other words, top-5 conformers are sufficient to cover the most conformers with equilibrium state (over 80\%), and the effect of larger $C$ is thus modest.
To sum up, adding more conformers might be helpful, but the computation cost can grow linearly with the increase in dataset size. On the other hand, enlarging the masking ratio will not induce extra cost, yet the performance is slightly better. Therefore, we would encourage tuning masking ratios prior to trying a larger number of conformers from the perspective of efficiency and effectiveness.
\subsection{Ablation Study: The Effect of Objective Function} \label{sec:experiment_each_loss_component}
\begin{wraptable}[13]{r}{0.5\textwidth}
\setlength{\tabcolsep}{3.53pt}
\vspace{-0.4cm}
\small
\caption{
Ablation on the objective function.
\vspace{-0.35cm}
}
\label{tab:ablation_objective_function}
\small
\centering
\begin{tabular}{l c c c}
\toprule
GraphMVP\, Loss & Contrastive & Generative & Avg\\
\midrule
Random & & & 67.21\\
\midrule
InfoNCE only & \checkmark & & 68.85\\
EBM-NCE only & \checkmark & & 70.15\\
VRR only & & \checkmark & 69.29\\
RR only & & \checkmark & 68.89\\
\midrule
InfoNCE + VRR & \checkmark & \checkmark & 70.67\\
EBM-NCE + VRR & \checkmark & \checkmark & 71.69\\
InfoNCE + RR & \checkmark & \checkmark & 70.60\\
EBM-NCE + RR & \checkmark & \checkmark & 70.94\\
\bottomrule
\end{tabular}
\end{wraptable}
In~\Cref{sec:method}, we introduce a new contrastive learning objective family called EBM-NCE, and we take either InfoNCE and EBM-NCE as the contrastive SSL. For the generative SSL task, we propose a novel objective function called variational representation reconstruction (VRR) in~\Cref{eq:variational_lower_bound_approximation}. As discussed in~\Cref{sec:generative_SSL}, stochasticity is important for GraphMVP\, since it can capture the conformer distribution for each 2D molecular graph. To verify this, we add an ablation study on \textit{representation reconstruction (RR)} by removing stochasticity in VRR. Thus, here we deploy a comprehensive ablation study to explore the effect for each individual objective function (InfoNCE, EBM-NCE, VRR and RR), followed by the pairwise combinations between them.
The results in \Cref{tab:ablation_objective_function} give certain constructive insights as follows:
(1) Each individual SSL objective function (middle block) can lead to better performance. This strengthens the claim that adding 3D information is helpful for 2D representation learning.
(2) According to the combination of those SSL objective functions (bottom block), adding both contrastive and generative SSL can consistently improve the performance. This verifies our claim that conducting SSL at both the inter-data and intra-data level is beneficial.
(3) We can see VRR is consistently better than RR on all settings, which verifies that stochasticity is an important factor in modeling 3D conformers for molecules.
\subsection{Broader Range of Downstream Tasks}
The 8 binary downstream tasks discussed so far have been widely applied in the graph SSL research line on molecules~\cite{hu2019strategies,You2020GraphCL,you2021graph}, but there are more tasks where the 3D conformers can be helpful. Here we test 4 extra regression property prediction tasks and 2 drug-target affinity tasks.
About the dataset statistics, more detailed information can be found in~\Cref{app:sec:dataset}, and we may as well briefly describe the affinity task here. Drug-target affinity (DTA) is a crucial task~\citep{pahikkala2015toward,wen2017deep,ozturk2018deepdta} in drug discovery, where it models both the molecular drugs and target proteins, with the goal to predict their affinity scores. One recent work~\citep{nguyen2019graphdta} is modeling the molecular drugs with 2D GNN and target protein (as an amino-acid sequence) with convolution neural network (CNN). We adopt this setting by pre-training the 2D GNN using GraphMVP.
As illustrated in~\Cref{tab:main_regression}, the consistent performance gain verifies the effectiveness of our proposed GraphMVP.
\begin{table}[t!]
\setlength{\tabcolsep}{7pt}
\centering
\caption{
Results for four molecular property prediction tasks (regression) and two DTA tasks (regression).
We report the mean RMSE of 3 seeds with scaffold splitting for molecular property downstream tasks, and mean MSE for 3 seeds with random splitting on DTA tasks.
For GraphMVP\,, we set $M=0.15$ and $C=5$.
The best performance for each task is marked in \textbf{\underline{bold}}.
We omit the std here since they are very small and indistinguishable. For complete results, please check~\Cref{sec:app:regression_results}.
\vspace{-4ex}
}
\label{tab:main_regression}
\begin{tabular}{l c c c c c c c c}
\toprule
& \multicolumn{5}{c}{Molecular Property Prediction} & \multicolumn{3}{c}{Drug-Target Affinity}\\
\cmidrule(lr){2-6} \cmidrule(lr){7-9} Pre-training & ESOL & Lipo & Malaria & CEP & Avg & Davis & KIBA & Avg\\
\midrule
--
& 1.178 & 0.744 & 1.127 & 1.254 & 1.0756
& 0.286 & 0.206 & 0.2459\\
\midrule
AM
& 1.112 & 0.730 & 1.119 & 1.256 & 1.0542
& 0.291 & 0.203 & 0.2476\\
CP
& 1.196 & 0.702 & 1.101 & 1.243 & 1.0606
& 0.279 & 0.198 & 0.2382 \\
JOAO
& 1.120 & 0.708 & 1.145 & 1.293 & 1.0663
& 0.281 & 0.196 & 0.2387\\
\midrule
GraphMVP
& 1.091 & 0.718 & 1.114 & 1.236 & 1.0397
& 0.280 & 0.178 & 0.2286\\
GraphMVP-G\,\,\,
& 1.064 & 0.691 & 1.106 & \textbf{\underline{1.228}} & 1.0221
& \textbf{\underline{0.274}} & 0.175 & 0.2248\\
GraphMVP-C\,\,\,
& \textbf{\underline{1.029}} & \textbf{\underline{0.681}} & \textbf{\underline{1.097}} & 1.244 & \textbf{\underline{1.0128}}
& 0.276 & \textbf{\underline{0.168}} & \textbf{\underline{0.2223}}\\
\bottomrule
\end{tabular}
\vspace{-1ex}
\end{table}
\subsection{Case Study}
We investigate how GraphMVP\, helps when the task objectives are challenging with respect to the 2D topology but straightforward using 3D geometry (as shown in~\Cref{fig:case}). We therefore design two case studies to testify how GraphMVP\, transfers knowledge from 3D geometry into the 2D representation.
The first case study is \textit{3D Diameter Prediction}. For molecules, usually, the longer the 2D diameter is, the larger the 3D diameter (largest atomic pairwise l2 distance). However, this does not always hold, and we are interested in using the 2D graph to predict the 3D diameter. The second case study is \textit{Long-Range Donor-Acceptor Detection}. Molecules possess a special geometric structure called donor-acceptor bond, and we want to use 2D molecular graph to detect this special structure. We validate that GraphMVP\, consistently brings improvements on these 2 case studies, and provide more detailed discussions and interpretations in~\Cref{appendix:case}.
\begin{figure}[htbp!]
\centering
\vspace{-2ex}
\includegraphics[width=\textwidth]{figures/Case_Study_2.pdf}
\vspace{-5ex}
\caption{We select the molecules whose properties can be easily resolved via 3D but not 2D. The randomly initialised 2D GNN achieves accuracy of $38.9\pm0.8$ and $77.9\pm1.1$, respectively. The GraphMVP\, pre-trained ones obtain scores of $42.3\pm1.3$ and $81.5\pm0.4$, outperforming all the precedents in~\Cref{sec:main_results}. We plot cases where random initialization fails but GraphMVP\, is correct.}
\label{fig:case}
\vspace{-4ex}
\end{figure}
\section{Theoretical Insights} \label{sec:theoretical_insight}
In this section, we provide the mathematical insights behind GraphMVP. We will first discuss both contrastive and generative SSL methods (\Cref{sec:contrastive_SSL,sec:generative_SSL}) are maximizing the mutual information (MI) and then how the 3D geometry, as privileged information, can help 2D representation learning.
\vspace{-1ex}
\subsection{Maximizing Mutual Information} \label{sec:mutual_information}
\vspace{-1ex}
Mutual information (MI) measures the non-linear dependence~\citep{belghazi2018mutual} between two random variables: the larger MI, the stronger dependence between the variables. Therefore for GraphMVP, we can interpret it as maximizing MI between 2D and 3D views: to obtain a more robust 2D/3D representation by sharing more information with its 3D/2D counterparts. This is also consistent with the sample complexity theory~\citep{erhan2010does,arora2019theoretical,garg2020functional} where SSL as functional regularizer can reduce the uncertainty in representation learning. We first derive a lower bound for MI (see derivations in~\Cref{sec:app:MI}), and the corresponding objective function $\mathcal{L}_{\text{MI}}$ is
\begin{equation} \label{eq:MI_objective}
\small
I(X;Y) \ge \mathcal{L}_{\text{MI}} = \frac{1}{2} \mathbb{E}_{p({\bm{x}},{\bm{y}})} \big[ \log p({\bm{y}}|{\bm{x}}) + \log p({\bm{x}}|{\bm{y}}) \big].
\end{equation}
\textbf{Contrastive Self-Supervised Learning.}
InfoNCE was initialized proposed to maximize the MI directly~\citep{oord2018representation}.
Here in GraphMVP, EBM-NCE estimates the conditional likelihood in~\Cref{eq:MI_objective} using EBM, and solves it with NCE~\cite{gutmann2010noise}. As a result, EBM-NCE can also be seen as maximizing MI between 2D and 3D views. The detailed derivations can be found in~\Cref{app:ebm_nce}.
\textbf{Generative Self-Supervised Learning.}
One alternative solution is to use a variational lower bound to approximate the conditional log-likelihood terms in~\Cref{eq:MI_objective}. Then we can follow the same pipeline in~\Cref{sec:generative_SSL}, ending up with the surrogate objective, {\em{i.e.}}, VRR in~\Cref{eq:variational_lower_bound_approximation}.
\subsection{3D Geometry as Privileged Information} \label{sec:privileged_info}
We show the theoretical insights from privileged information that motivate GraphMVP. We start by considering a supervised learning setting where $(\bm{u}_i,\bm{l}_i)$ is a feature-label pair and $\bm{u}_i^*$ is the privileged information~\cite{vapnik2009new,vapnik2015learning}. The privileged information is defined to be additional information about the input $(\bm{u}_i,\bm{l}_i)$ in order to support the prediction. For example, $\bm{u}_i$ could be some CT images of a particular disease, $\bm{l}_i$ could be the label of the disease and $\bm{u}_i^*$ is the medical report from a doctor. VC theory~\cite{vapnik2013nature,vapnik2015learning} characterizes the learning speed of an algorithm from the capacity of the algorithm and the amount of training data. Considering a binary classifier $f$ from a function class $\mathcal{F}$ with finite VC-dimension $\text{VCD}(\mathcal{F})$. With probability $1-\delta$, the expected error is upper bounded by
\begin{equation}
\small
R(f)\leq R_n(f) +\mathcal{O}\bigg( \big( \frac{\text{VCD}(\mathcal{F})-\log\delta}{n} \big)^\beta \bigg)
\end{equation}
where $R_n(f)$ denotes the training error and $n$ is the number of training samples. When the training data is separable, then $R_n(f)$ will diminish to zero and $\beta$ is equal to $1$. When the training data is non-separable, $\beta$ is $\frac{1}{2}$. Therefore, the rate of convergence for the separable case is of order $1/n$. In contrast, the rate for the non-separable case is of order $1/\sqrt{n}$. We note that such a difference is huge, since the same order of bounds require up to 100 training samples versus 10,000 samples. Privileged information makes the training data separable such that the learning can be more efficient. Connecting the results to GraphMVP, we notice that the 3D geometric information of molecules can be viewed as a form of privileged information, since 3D information can effectively make molecules more separable for some properties~\cite{schutt2017schnet,liu2018n,liu2021spherical}. Besides, privileged information is only used in training, and it well matches our usage of 3D geometry for pre-training. In fact, using 3D structures as privileged information has been already shown quite useful in protein classification~\cite{vapnik2009new}, which serves as a strong evidence to justify the effectiveness of 3D information in graph SSL pre-training.
\section{Conclusion and Future Work}
\vspace{-1ex}
In this work, we provide a very general framework, coined GraphMVP. From the domain perspective, GraphMVP\, (1) is the first to incorporate 3D information for augmenting 2D graph representation learning and (2) is able to take advantages of 3D conformers by considering stochasticity in modeling. From the aspect of technical novelties, GraphMVP\, brings following insights when introducing 2 SSL tasks:
(1) Following~\Cref{eq:MI_objective}, GraphMVP\, proposes EBM-NCE and VRR, where they are modeling the conditional distributions using EBM and variational distribution respectively.
(2) EBM-NCE is similar to JSE, while we start with a different direction for theoretical intuition, yet EBM opens another promising venue in this area.
(3) VRR, as a generative SSL method, is able to alleviate the potential issues in molecule generation~\citep{zhavoronkov2019deep,gao2020synthesizability}.
(4) Ultimately, GraphMVP\, combines both contrastive SSL (InfoNCE or EBM-NCE) and generative SSL (VRR) for objective function.
Both empirical results (solid performance improvements on 14 downstream datasets) and theoretical analysis can strongly support the above domain and technical contributions.
We want to emphasize that GraphMVP\, is model-agnostic and has the potential to be expanded to many other low-data applications. This motivates broad directions for future exploration, including but not limited to: (1) More powerful 2D and 3D molecule representation methods. (2) Different application domain other than small molecules, {\em{e.g.}}, large molecules like proteins.
\part{\Large\centering{Appendix}}
\parttoc
\clearpage
\input{09_related}
\clearpage
\input{10_MI}
\clearpage
\input{11_contrastive_ssl}
\clearpage
\input{12_generative_ssl}
\clearpage
\input{13_datasets}
\input{14_experiments}
\section{Self-Supervised Learning on Molecular Graph} \label{appendix:related_work}
Self-supervised learning (SSL) methods have attracted massive attention recently, trending from vision~\citep{chen2020simple,he2020momentum,caron2020unsupervised,chen2021exploring,OcCo}, language~\citep{oord2018representation,devlin2018bert,brown2020language} to graph~\citep{velivckovic2018deep,sun2019infograph,liu2018n,hu2019strategies,You2020GraphCL,you2021graph}. In general, there are two categories of SSL: contrastive and generative, where they differ on the design of the supervised signals. Contrastive SSL realizes the supervised signals at the \textbf{inter-data} level, learning the representation by contrasting with other data points; while generative SSL focuses on reconstructing the original data at the \textbf{intra-data} level. Both venues have been widely explored~\citep{liu2021graph,xie2021self,wu2021self,liu2021self}.
\subsection{Contrastive graph SSL}
Contrastive graph SSL first applies transformations to construct different \textit{views} for each graph. Each view incorporates different granularities of information, like node-, subgraph-, and graph-level. It then solves two sub-tasks simultaneously: (1) aligning the representations of views from the same data; (2) contrasting the representations of views from different data, leading to a uniformly distributed latent space~\citep{wang2020understanding}. The key difference among existing methods is thus the design of view constructions. InfoGraph~\citep{velivckovic2018deep,sun2019infograph} contrasted the node (local) and graph (global) views. ContextPred~\citep{hu2019strategies} and G-Contextual~\cite{rong2020self} contrasted between node and context views. GraphCL and JOAO~\citep{You2020GraphCL,you2021graph} made comprehensive comparisons among four graph-level transformations and further learned to select the most effective combinations.
\subsection{Generative graph SSL}
Generative graph SSL aims at reconstructing important structures for each graph. By so doing, it consequently learns a representation capable of encoding key ingredients of the data. EdgePred~\citep{hamilton2017inductive} and AttrMask~\citep{hu2019strategies} predicted the adjacency matrix and masked tokens (nodes and edges) respectively. GPT-GNN~\citep{hu2020gpt} reconstructed the whole graph in an auto-regressive approach.
\subsection{Predictive graph SSL}
There are certain SSL methods specific to the molecular graph. For example, one central task in drug discovery is to find the important substructure or motif in molecules that can activate the target interactions. G-Motif~\citep{rong2020self} adopts domain knowledge to heuristically extract motifs for each molecule, and the SSL task is to make prediction on the existence of each motif. Different from contrastive and generative SSL, recent literature~\citep{wu2021self} takes this as predictive graph SSL, where the supervised signals are self-generated labels.
\begin{table}[b]
\small
\caption{
Comparison between GraphMVP\, and existing graph SSL methods.
\vspace{-0.6ex}
}
\label{app:tab:compare}
\center
\setlength{\tabcolsep}{10pt}
\begin{tabular}{l c c c c c}
\toprule
\multirow{2}{*}{SSL Pre-training} & \multicolumn{2}{c}{Graph View} & \multicolumn{3}{c}{SSL Category}\\
\cmidrule(lr){2-3}
\cmidrule(lr){4-6}
& 2D Topology & 3D Geometry & Generative & Contrastive & Predictive\\
\midrule
EdgePred~\citep{hamilton2017inductive} & \checkmark & - & \checkmark & - & -\\
AttrMask~\citep{hu2019strategies} & \checkmark & - & \checkmark & - & -\\
GPT-GNN~\citep{hu2020gpt} & \checkmark & - & \checkmark & - & -\\
InfoGraph~\citep{velivckovic2018deep,sun2019infograph} & \checkmark & - & - & \checkmark& -\\
ContexPred~\citep{hu2019strategies} & \checkmark & - & - & \checkmark& -\\
GraphLoG~\citep{xu2021self} & \checkmark & - & - & \checkmark& -\\
G-Contextual~\citep{rong2020self} & \checkmark & - & - & \checkmark& -\\
GraphCL~\citep{You2020GraphCL} & \checkmark & - & - & \checkmark& -\\
JOAO~\citep{you2021graph} & \checkmark & - & - & \checkmark& -\\
G-Motif~\citep{rong2020self} & \checkmark & - & - & - & \checkmark\\
\midrule
GraphMVP\,(Ours) & \checkmark & \checkmark & \checkmark & \checkmark& -\\
\bottomrule
\end{tabular}
\end{table}
\textbf{SSL for Molecular Graphs.} Recall that all previous methods in~\Cref{app:tab:compare} \textbf{merely} focus on the 2D topology. However, for science-centric tasks such as molecular property prediction, 3D geometry should be incorporated as it provides complementary and comprehensive information~\citep{schutt2017schnet,liu2021spherical}. To mitigate this gap, we propose GraphMVP to leverage the 3D geometry with unsupervised graph pre-training.
\section{Molecular Graph Representation} \label{app:sec:molecule_representation}
There are two main methods for molecular graph representation learning. The first one is the molecular fingerprints. It is a hashed bit vector to describe the molecular graph. There has been re-discoveries on fingerprints-based methods~\citep{ramsundar2015massively,liu2018practical,liu2019loss,meyer2019learning,alnammi2021evaluating,jiang2021could}, while its has one main drawback: Random forest and XGBoost are very strong learning models on fingerprints, but they fail to take benefits of the pre-training strategy.
Graph neural network (GNN) has become another mainstream modeling methods for molecular graph representation. Existing methods can be generally split into two venues: 2D GNN and 3D GNN, depending on what levels of information is considered. 2D GNN focuses on the topological structures of the graph, like the adjacency among nodes, while 3D GNN is able to model the ``energy'' of molecules by taking account the spatial positions of atoms.
First, we want to highlight that GraphMVP\, is model-agnostic, {\em{i.e.}}, it can be applied to any 2D and 3D GNN representation function, yet the specific 2D and 3D representations are not the main focus of this work.
Second, we acknowledge there are a lot of advanced 3D~\citep{fuchs2020se,satorras2021n,liu2021spherical,jumper2021highly} and 2D~\citep{gilmer2017neural,yang2019analyzing,liu2018n,xu2018powerful,corso2020principal,demirel2021analysis} representation methods.
However, considering the \textit{graph SSL literature} and \textit{graph representation liteature} (illustrated below), we adopt GIN~\citep{xu2018powerful} and SchNet~\citep{schutt2017schnet} in current GraphMVP.
\subsection{2D Molecular Graph Neural Network}
The 2D representation is taking each molecule as a 2D graph, with atoms as nodes and bonds as edges, {\em{i.e.}}, $g_{\text{2D}} = (X, E)$. $X\in\mathbb{R}^{n\times d_n}$ is the atom attribute matrix, where $n$ is the number of atoms (nodes) and $d_n$ is the atom attribute dimension. $E\in\mathbb{R}^{m\times d_e}$ is the bond attribute matrix, where $m$ is the number of bonds (edges) and $d_m$ is the bond attribute dimension. Notice that here $E$ also includes the connectivity. Then we will apply a transformation function $T_{\text{2D}}$ on the topological graph. Given a 2D graph $g_{\text{2D}}$, its 2D molecular representation is:
\begin{equation}
h_{\text{2D}} = \text{GNN-2D}(T_{\text{2D}}(g_{\text{2D}})) = \text{GNN-2D}(T_{\text{2D}}(X, E)).
\end{equation}
The core operation of 2D GNN is the message passing function~\citep{gilmer2017neural}, which updates the node representation based on adjacency information. We have variants depending on the design of message and aggregation functions, and we pick GIN~\citep{xu2018powerful} in this work.
\paragraph{GIN} There has been a long research line on 2D graph representation learning~\citep{gilmer2017neural,yang2019analyzing,liu2018n,xu2018powerful,corso2020principal,demirel2021analysis}. Among these, graph isomorphism network (GIN) model~\citep{xu2018powerful} has been widely used as the backbone model in recent graph self-supervised learning work~\citep{hu2019strategies,You2020GraphCL,you2021graph}. Thus, we as well adopt GIN as the base model for 2D representation.
Recall each molecule is represented as a molecular graph, {\em{i.e.}}, $g_{\text{2D}} = (X, E)$, where $X$ and $E$ are feature matrices for atoms and bonds respectively. Then the message passing function is defined as:
\begin{equation}
z_i^{(k+1)} = \text{MLP}_{\text{atom}}^{(k+1)} \Big(z_i^{(k)} + \sum_{j \in \mathcal{N}(i)} \big( z_j^{(k)} + \text{MLP}_{\text{bond}}^{(k+1)}(E_{ij}) \big) \Big),
\end{equation}
where $z_0=X$ and $\text{MLP}_{\text{atom}}^{(k+1)}$ and $\text{MLP}_{\text{bond}}^{(k+1)}$ are the $(l+1)$-th MLP layers on the atom- and bond-level respectively. Repeating this for $K$ times, and we can encode $K$-hop neighborhood information for each center atom in the molecular data, and we take the last layer for each node/atom representation. The graph-level molecular representation is the mean of the node representation:
\begin{equation}
z({\bm{x}}) = \frac{1}{N} \sum_{i} z_i^{(K)}
\end{equation}
\subsection{3D Molecular Graph Neural Network} \label{appendix:3d_gnn}
Recently, the 3D geometric representation learning has brought breakthrough progress in molecule modeling~\citep{schutt2017schnet,fuchs2020se,satorras2021n,liu2021spherical,jumper2021highly}. 3D molecular graph additionally includes spatial locations of the atoms, which needless to be static since, in real scenarios, atoms are in continual motion on \textit{a potential energy surface}~\citep{axelrod2020geom}. The 3D structures at the local minima on this surface are named \textit{molecular conformation} or \textit{conformer}. As the molecular properties are a function of the conformer ensembles~\citep{hawkins2017conformation}, this reveals another limitation of existing mainstream methods: to predict properties from a single 2D or 3D graph cannot account for this fact~\citep{axelrod2020geom}, while our proposed method can alleviate this issue to a certain extent.
For specific 3D molecular graph, it additionally includes spatial positions of the atoms. We represent each conformer as $g_{\text{3D}} = (X, R)$, where $R \in \mathbb{R}^{n \times 3}$ is the 3D-coordinate matrix, and the corresponding representation is:
\begin{equation}
h_{\text{3D}} = \text{GNN-3D}(T_{\text{3D}}(g_{\text{3D}})) = \text{GNN-3D}(T_{\text{3D}}(X, R)),
\end{equation}
where $R$ is the 3D-coordinate matrix and $T_{\text{3D}}$ is the 3D transformation. Note that further information such as plane and torsion angles can be solved from the positions.
\paragraph{SchNet}
SchNet~\citep{schutt2017schnet} is composed of the following key steps:
\begin{equation}
\begin{aligned}
& z_i^{(0)} = \text{embedding} (x_i)\\
& z_i^{(t+1)} = \text{MLP} \Big( \sum_{j=1}^{n} f(x_j^{(t-1)}, r_i, r_j) \Big)\\
& h_i = \text{MLP} (z_i^{(K)}),
\end{aligned}
\end{equation}
where $K$ is the number of hidden layers, and
\begin{equation}
f(x_j, r_i, r_j) = x_j \cdot e_k(r_i - r_j) = x_j \cdot \exp(- \gamma \| \|r_i - r_j\|_2 - \mu \|_2^2)
\end{equation}
is the continuous-filter convolution layer, enabling the modeling of continuous positions of atoms.
We adopt SchNet for the following reasons. (1) SchNet is a very strong geometric representation method after \textit{fair} benchmarking. (2) SchNet can be trained more efficiently, comparing to the other recent 3D models. To support these two points, we make a comparison among the most recent 3D geometric models~\cite{fuchs2020se,satorras2021n,liu2021spherical} on QM9 dataset. QM9~\citep{wu2018moleculenet} is a molecule dataset approximating 12 thermodynamic properties calculated by density functional theory (DFT) algorithm. Notice: UNiTE~\citep{qiao2021unite} is the state-of-the-art 3D GNN, but it requires a commercial software for feature extraction, thus we exclude it for now.
\begin{table}[H]
\centering
\scriptsize
\caption{
Reproduced MAE on QM9. 100k for training, 17,748 for val, 13,083 for test. The last column is the approximated running time.
}
\label{tab:app:qm9}
\begin{tabular}{l c c c c c c c c c c c c r}
\toprule
& alpha & gap & homo & lumo & mu & cv & g298 & h298 & r2 & u298 & u0 & zpve & time\\
\midrule
SchNet~\citep{schutt2017schnet} & 0.077 & 50 & 32 & 26 & 0.030 & 0.032 & 15 & 14 & 0.122 & 14 & 14 & 1.751 & 3h\\
SE(3)-Trans~\citep{fuchs2020se} & 0.143 & 59 & 36 & 36 & 0.052 & 0.068 & 68 & 72 & 1.969 & 68 & 74 & 5.517 & 50h\\
EGNN~\citep{satorras2021n} & 0.075 & 49 & 29 & 26 & 0.030 & 0.032 & 11 & 10 & 0.076 & 10 & 10 & 1.562 & 24h\\
SphereNet~\citep{liu2021spherical} & 0.054 & 41 & 22 & 19 & 0.028 & 0.027 & 10 & 8 & 0.295 & 8 & 8 & 1.401 & 50h\\
\bottomrule
\end{tabular}
\end{table}
\Cref{tab:app:qm9} shows that, under a fair comparison (w.r.t. data splitting, seed, cuda version, etc), SchNet can reach pretty comparable performance, yet the efficiency of SchNet is much better. Combining these two points, we adopt SchNet in current version of GraphMVP.
\subsection{Summary}
To sum up, in GraphMVP, the most important message we want to deliver is how to design a well-motivated SSL algorithm to extract useful 3D geometry information to augment the 2D representation for downstream fine-tuning. GraphMVP\, is model-agnostic, and we may as well leave the more advanced 3D~\citep{fuchs2020se,satorras2021n,liu2021spherical,jumper2021highly} and 2D~\citep{yang2019analyzing,liu2018n,corso2020principal} GNN for future exploration.
In addition, molecular property prediction tasks have rich alternative representation methods, including SMILES~\cite{weininger1988smiles,hirohara2018convolutional}, and biological knowledge graph~\cite{wang2021multi,liu2022structured}. There have been another SSL research line on them~\cite{liustructured,zhu2021dual,fang2021molecular}, yet they are beyond the scope of discussion in this paper.
\section{Maximize Mutual Information} \label{sec:app:MI}
In what follows, we will use $X$ and $Y$ to denote the data space for $2D$ graph and $3D$ graph respectively. Then the latent representations are denoted as $h_{\bm{x}}$ and $h_{\bm{y}}$.
\subsection{Formulation}
The standard formulation for mutual information (MI) is
\begin{equation} \label{eq:app:MI}
\begin{aligned}
I(X;Y)
& = \mathbb{E}_{p({\bm{x}},{\bm{y}})} \big[ \log \frac{p({\bm{x}},{\bm{y}})}{p({\bm{x}}) p({\bm{y}})} \big].
\end{aligned}
\end{equation}
Another well-explained MI inspired from wikipedia is given in~\Cref{fig:app:MI}.
\begin{figure}[htb!]
\centering
\includegraphics[width=0.5\textwidth]{figures/MutualInfo_Final.pdf}
\caption{Venn diagram of mutual information. Inspired by wikipedia.}
\label{fig:app:MI}
\end{figure}
Mutual information (MI) between random variables measures the corresponding non-linear dependence. As can be seen in the first equation in~\Cref{eq:app:MI}, the larger the divergence between the joint ($p({\bm{x}}, {\bm{y}}$) and the product of the marginals $p({\bm{x}}) p({\bm{y}})$, the stronger the dependence between $X$ and $Y$.
Thus, following this logic, maximizing MI between 2D and 3D views can force the 3D/2D representation to capture higher-level factors, {\em{e.g.}}, the occurrence of important substructure that is semantically vital for downstream tasks. Or equivalently, maximizing MI can decrease the uncertainty in 2D representation given 3D geometric information.
\subsection{A Lower Bound to MI}
To solve MI, we first extract a lower bound:
\begin{equation}
\begin{aligned}
I(X;Y)
& = \mathbb{E}_{p({\bm{x}},{\bm{y}})} \Big[ \log \frac{p({\bm{x}},{\bm{y}})}{p({\bm{x}}) p({\bm{y}})} \Big]\\
& \ge \mathbb{E}_{p({\bm{x}},{\bm{y}})} \Big[ \log \frac{p({\bm{x}},{\bm{y}})}{\sqrt{p({\bm{x}}) p({\bm{y}})}} \Big]\\
& = \frac{1}{2} \mathbb{E}_{p({\bm{x}},{\bm{y}})} \Big[ \log \frac{(p({\bm{x}},{\bm{y}}))^2}{p({\bm{x}}) p({\bm{y}})} \Big]\\
& = \frac{1}{2} \mathbb{E}_{p({\bm{x}},{\bm{y}})} \Big[ \log p({\bm{x}}|{\bm{y}}) \Big] + \frac{1}{2} \mathbb{E}_{p({\bm{x}},{\bm{y}})} \Big[ \log p({\bm{y}}|{\bm{x}}) \Big]\\
& = -\frac{1}{2} [H(Y|X) + H(X|Y)].
\end{aligned}
\end{equation}
Thus, we transform the MI maximization problem into minimizing the following objective:
\begin{equation} \label{eq:app:MI_objective}
\begin{aligned}
\mathcal{L}_{\text{MI}} & = \frac{1}{2} [H(Y|X) + H(X|Y)].
\end{aligned}
\end{equation}
In the following sections, we will describe two self-supervised learning methods for solving MI. Notice that the methods are very general, and can be applied to various applications. Here we apply it mainly for making 3D geometry useful for 2D representation learning on molecules.
\section{Contrastive Self-Supervised Learning} \label{sec:app:contrastive_ssl}
The essence of contrastive self-supervised learning is to align positive view pairs and contrast negative view pairs, such that the obtained representation space is well distributed~\citep{wang2020understanding}. We display the pipeline in~\Cref{fig:app:contrastive_ssl}. Along the research line in graph SSL~\citep{liu2021graph,xie2021self,wu2021self,liu2021self}, InfoNCE and EBM-NCE are the two most-widely used, as discussed below.
\begin{figure}[htb!]
\centering
\includegraphics[width=\textwidth]{figures/Diagram_Contrastive_final.pdf}
\vspace{-4ex}
\caption{
Contrastive SSL in GraphMVP. The black dashed circles represent subgraph masking.
}
\label{fig:app:contrastive_ssl}
\vspace{-2ex}
\end{figure}
\subsection{InfoNCE} \label{sec:app:infonce}
InfoNCE~\citep{oord2018representation} is first proposed to approximate MI~\Cref{eq:app:MI}:
\begin{equation}\label{eq:app:InfoNCE}
\footnotesize{
\begin{aligned}
\mathcal{L}_{\text{InfoNCE}} =
-\frac{1}{2} \mathbb{E} &\left[
\log \frac{\exp(f_{{\bm{x}}}({\bm{x}}, {\bm{y}}))}{\exp(f_{{\bm{x}}}({\bm{x}}, {\bm{y}})) + \sum_j \exp(f_{{\bm{x}}}({\bm{x}}^{j},{\bm{y}})}) + \log \frac{\exp(f_{{\bm{y}}}({\bm{y}},{\bm{x}}))}{\exp(f_{{\bm{y}}}({\bm{y}},{\bm{x}})) + \sum_j \exp{f_{{\bm{y}}}({\bm{y}}^{j},{\bm{x}})}} \right],
\end{aligned}
}
\end{equation}
where ${\bm{x}}^{j}, {\bm{y}}^{j}$ are randomly sampled 2D and 3D views regarding to the anchored pair $({\bm{x}},{\bm{y}})$. $f_{{\bm{x}}}({\bm{x}},{\bm{y}}), f_{{\bm{y}}}({\bm{y}},{\bm{x}})$ are scoring functions for the two corresponding views, whose formulation can be quite flexible. Here we use $f_{\bm{x}}({\bm{x}},{\bm{y}}) = f_{\bm{y}}({\bm{y}},{\bm{x}}) = \exp(\langle h_{\bm{x}}, h_{\bm{y}} \rangle)$.
\paragraph{Derivation of InfoNCE}
\begin{equation}
\begin{aligned}
I(X;Y) - \log (K)
& = \mathbb{E}_{p({\bm{x}}, {\bm{y}})} \big[\log \frac{1}{K} \frac{p({\bm{x}}, {\bm{y}})}{p({\bm{x}}) p({\bm{y}})} \big]\\
& = \sum_{{\bm{x}}^{i},{\bm{y}}^{i}} \big[\log \frac{1}{K} \frac{p({\bm{x}}^{i},{\bm{y}}^{i})}{p({\bm{x}}^{i}) p({\bm{y}}^{i})} \big]\\
& \ge -\sum_{{\bm{x}}^{i},{\bm{y}}^{i}} \big[\log \big( 1 + (K-1) \frac{p({\bm{x}}^{i}) p({\bm{y}}^{i})}{p({\bm{x}}^{i},{\bm{y}}^{i})} \big)\big]\\
& = -\sum_{{\bm{x}}^{i},{\bm{y}}^{i}} \big[\log \frac{\frac{p({\bm{x}}^{i},{\bm{y}}^{i})}{p({\bm{x}}^{i}) p({\bm{y}}^{i})} + (K-1)}{\frac{p({\bm{x}}^{i},{\bm{y}}^{i})}{p({\bm{x}}^{i}) p({\bm{y}}^{i})}} \big]\\
& \approx -\sum_{{\bm{x}}^{i},{\bm{y}}^{i}} \big[ \log \frac{\frac{p({\bm{x}}^{i},{\bm{y}}^{i})}{p({\bm{x}}^{i}) p({\bm{y}}^{i})} + (K-1)\mathbb{E}_{{\bm{x}}^{j} \ne {\bm{x}}^{i}}\frac{p(x^{j},{\bm{y}}^{i})}{p({\bm{x}}^{j}) p({\bm{y}}^{i})} }{\frac{p({\bm{x}}^{i},{\bm{y}}^{i})}{p({\bm{x}}^{i}) p({\bm{y}}^{i})}} \big] \quad \text{// \textcircled{1}}\\
& = \sum_{{\bm{x}}^{i},{\bm{y}}^{i}} \big[ \log \frac{\exp(f_{\bm{x}}({\bm{x}}^{i},{\bm{y}}^{i}))}{\exp(f_{\bm{x}}({\bm{x}}^{i},{\bm{y}}^{i})) + \sum_{j=1}^K f_{\bm{x}}({\bm{x}}^{j},{\bm{y}}^{i})} \big],
\end{aligned}
\end{equation}
where we set $ f_{\bm{x}}({\bm{x}}^{i},{\bm{y}}^{i}) = \log \frac{p({\bm{x}}^{i},{\bm{y}}^{i})}{p({\bm{x}}^{i})p({\bm{y}}^{i})}$.
Notice that in \textcircled{1}, we are using data $x \in X$ as the anchor points. If we use the $y \in Y$ as the anchor points and follow the similar steps, we can obtain
\begin{equation}
\begin{aligned}
I(X;Y) - \log(K) \ge \sum_{{\bm{y}}^{i},{\bm{x}}^{i}} \big[ \log \frac{\exp(f_{\bm{y}}({\bm{y}}^{i},{\bm{x}}^{i}))}{\exp f_{\bm{y}}({\bm{y}}^{i},{\bm{x}}^{i}) + \sum_{j=1}^K \exp (f_{\bm{y}}({\bm{y}}^{j},{\bm{x}}^{i}))} \big].
\end{aligned}
\end{equation}
Thus, by add both together, we can have the objective function as~\Cref{eq:app:InfoNCE}.
\subsection{EBM-NCE} \label{app:ebm_nce}
We here provide an alternative approach to maximizing MI using energy-based model (EBM). To our best knowledge, we are the \textbf{first} to give the rigorous proof of using EBM to maximize the MI.
\subsubsection{Energy-Based Model (EBM)}
Energy-based model (EBM) is a powerful tool for modeling the data distribution. The classic formulation is:
\begin{equation} \label{eq:app:EBM_original}
p({\bm{x}}) = \frac{\exp(-E({\bm{x}}))}{A},
\end{equation}
where the bottleneck is the intractable partition function $A = \int_{\bm{x}} \exp(-E({\bm{x}})) d{\bm{x}}$. Recently, there have been quite a lot progress along this direction~\citep{gutmann2010noise,du2020improved,song2020score,song2021train}. Noise Contrastive Estimation (NCE)~\citep{gutmann2010noise} is one of the powerful tools here, as we will introduce later.
\subsubsection{EBM for MI}
Recall that our objective function is~\Cref{eq:app:MI_objective}: $\mathcal{L}_{\text{MI}} = \frac{1}{2} [H(Y|X) + H(X|Y)]$. Then we model the conditional likelihood with energy-based model (EBM). This gives us
\begin{equation} \label{eq:app:EBM}
\begin{aligned}
\mathcal{L}_{\text{EBM}} = -\frac{1}{2} \mathbb{E}_{p({\bm{x}},{\bm{y}})} \Big[ \log \frac{\exp(f_{\bm{x}}({\bm{x}}, {\bm{y}}))}{A_{{\bm{x}}|{\bm{y}}}} + \log \frac{\exp(f_{\bm{y}}({\bm{y}}, {\bm{x}}))}{A_{{\bm{y}}|{\bm{x}}}} \Big],
\end{aligned}
\end{equation}
where $f_{\bm{x}}({\bm{x}}, {\bm{y}}) = -E({\bm{x}}|{\bm{y}})$ and $f_{\bm{y}}({\bm{y}}, {\bm{x}}) = -E({\bm{y}}|{\bm{x}})$ are the negative energy functions, and $A_{{\bm{x}}|{\bm{y}}}$ and $A_{{\bm{y}}|{\bm{x}}}$ are the corresponding partition functions.
Under the EBM framework, if we solve~\Cref{eq:app:EBM} with Noise Contrastive Estimation (NCE)~\citep{gutmann2010noise}, the final EBM-NCE objective is
\begin{equation} \label{eq:app:EBM_NCE}
\begin{aligned}
\mathcal{L}_{\text{EBM-NCE}}
= & -\frac{1}{2} \mathbb{E}_{p_{\text{data}}(y)} \Big[ \mathbb{E}_{p_n({\bm{x}}|{\bm{y}})} [\log \big(1-\sigma(f_{\bm{x}}({\bm{x}}, {\bm{y}}))\big)] + \mathbb{E}_{p_{\text{data}}({\bm{x}}|{\bm{y}})} [\log \sigma(f_{\bm{x}}({\bm{x}}, {\bm{y}}))] \Big] \\
& ~~ - \frac{1}{2} \mathbb{E}_{p_{\text{data}}(x)} \Big[ \mathbb{E}_{p_n({\bm{y}}|{\bm{x}})} [\log \big(1-\sigma(f_{\bm{y}}({\bm{y}}, {\bm{x}}))\big)] + \mathbb{E}_{p_{\text{data}}({\bm{y}}|{\bm{x}})} [\log \sigma(f_{\bm{y}}({\bm{y}}, {\bm{x}}))] \Big].
\end{aligned}
\end{equation}
Next we will give the detailed derivations.
\subsubsection{Derivation of conditional EBM with NCE}
WLOG, let's consider the $p_\theta({\bm{x}}|{\bm{y}})$ first, and by EBM it is as follows:
\begin{equation} \label{eq:app:EBM_SSL}
p_\theta({\bm{x}}|{\bm{y}}) = \frac{\exp(-E({\bm{x}}|{\bm{y}}))}{ \int \exp(-E({\tilde {\bm{x}}}|{\bm{y}})) d{\tilde {\bm{x}}}} = \frac{\exp(f_{\bm{x}}({\bm{x}}, {\bm{y}}))}{\int \exp(f_{\bm{x}}({\tilde {\bm{x}}}|{\bm{y}})) d{\tilde {\bm{x}}}} = \frac{\exp(f_{\bm{x}}({\bm{x}}, {\bm{y}}))}{A_{{\bm{x}}|{\bm{y}}}}.
\end{equation}
Then we solve this using NCE. NCE handles the intractability issue by transforming it as a binary classification task. We take the partition function $A_{{\bm{x}}|{\bm{y}}}$ as a parameter, and introduce a noise distribution $p_n$. Based on this, we introduce a mixture model: ${\bm{z}}=0$ if the conditional ${\bm{x}}|{\bm{y}}$ is from $p_n({\bm{x}}|{\bm{y}})$, and ${\bm{z}}=1$ if ${\bm{x}}|{\bm{y}}$ is from $p_{\text{data}}({\bm{x}}|{\bm{y}})$. So the joint distribution is:
\begin{equation*}
p_{n,\text{\text{data}}}({\bm{x}}|{\bm{y}}) = p({\bm{z}}=1) p_{\text{data}}({\bm{x}}|{\bm{y}}) + p({\bm{z}}=0) p_n({\bm{x}}|{\bm{y}})
\end{equation*}
The posterior of $p({\bm{z}}=0|{\bm{x}},{\bm{y}})$ is
\begin{equation*}
p_{n,\text{\text{data}}}({\bm{z}}=0|{\bm{x}},{\bm{y}}) = \frac{p({\bm{z}}=0) p_n({\bm{x}}|{\bm{y}})}{p(z=0) p_n({\bm{x}}|{\bm{y}}) + p({\bm{z}}=1) p_{\text{data}}({\bm{x}}|{\bm{y}})} = \frac{\nu \cdot p_n({\bm{x}}|{\bm{y}})}{\nu \cdot p_n({\bm{x}}|{\bm{y}}) + p_{\text{data}}({\bm{x}}|{\bm{y}})},
\end{equation*}
where $\nu = \frac{p({\bm{z}}=0)}{p({\bm{z}}=1)}$.
Similarly, we can have the joint distribution under EBM framework as:
\begin{equation*}
p_{n, \theta}({\bm{x}}) = p(z=0) p_n({\bm{x}}|{\bm{y}}) + p(z=1) p_{\theta}({\bm{x}}|{\bm{y}})
\end{equation*}
And the corresponding posterior is:
\begin{equation*}
p_{n,\theta}({\bm{z}}=0|{\bm{x}},{\bm{y}}) = \frac{p({\bm{z}}=0) p_n({\bm{x}}|{\bm{y}})}{p({\bm{z}}=0) p_n({\bm{x}}|{\bm{y}}) + p({\bm{z}}=1) p_{\theta}({\bm{x}}|{\bm{y}})} = \frac{\nu \cdot p_n({\bm{x}}|{\bm{y}})}{\nu \cdot p_n({\bm{x}}|{\bm{y}}) + p_{\theta}({\bm{x}}|{\bm{y}})}
\end{equation*}
We indirectly match $p_\theta({\bm{x}}|{\bm{y}})$ to $p_{\text{data}}({\bm{x}}|{\bm{y}})$ by fitting $p_{n,\theta}({\bm{z}}|{\bm{x}},{\bm{y}})$ to $p_{n,\text{\text{data}}}({\bm{z}}|{\bm{x}},{\bm{y}})$ by minimizing their KL-divergence:
\begin{equation} \label{eq:app:EBM_01}
\begin{aligned}
& \min_\theta D_{\text{KL}}(p_{n,\text{\text{data}}}({\bm{z}}|{\bm{x}},{\bm{y}}) || p_{n,\theta}({\bm{z}}|{\bm{x}},{\bm{y}})) \\
& = \mathbb{E}_{p_{n,\text{\text{data}}}({\bm{x}},{\bm{z}}|{\bm{y}})} [\log p_{n,\theta}({\bm{z}}|{\bm{x}},{\bm{y}})] \\
& = \int \sum_{\bm{z}} p_{n,\text{\text{data}}}({\bm{x}},{\bm{z}}|{\bm{y}}) \cdot \log p_{n,\theta}({\bm{z}}|{\bm{x}},{\bm{y}}) d {\bm{x}}\\
& = \int \Big\{ p({\bm{z}}=0) p_{n,\text{\text{data}}}({\bm{x}}|{\bm{y}},{\bm{z}}=0) \log p_{n,\theta}({\bm{z}}=0|{\bm{x}},{\bm{y}}) \\
& \quad\quad\quad\quad + p({\bm{z}}=1) p_{n,\text{\text{data}}}({\bm{x}}|{\bm{z}}=1,{\bm{y}}) \log p_{n,\theta}({\bm{z}}=1|{\bm{x}},{\bm{y}}) \Big\} d{\bm{x}} \\
& = \nu \cdot \mathbb{E}_{p_{n}({\bm{x}}|{\bm{y}})} \Big[\log p_{n,\theta}({\bm{z}}=0|{\bm{x}},{\bm{y}}) \Big] + \mathbb{E}_{p_{\text{data}}({\bm{x}}|{\bm{y}} )} \Big[\log p_{n,\theta}({\bm{z}}=1|{\bm{x}},{\bm{y}}) \Big] \\
& = \nu \cdot \mathbb{E}_{p_{n}({\bm{x}}|{\bm{y}})} \Big[ \log \frac{\nu \cdot p_n({\bm{x}}|{\bm{y}})}{\nu \cdot p_n({\bm{x}}|{\bm{y}}) + p_{\theta}({\bm{x}}|{\bm{y}})} \Big] + \mathbb{E}_{p_{\text{data}}({\bm{x}}|{\bm{y}} )} \Big[ \log \frac{p_\theta({\bm{x}}|{\bm{y}})}{\nu \cdot p_n({\bm{x}}|{\bm{y}}) + p_{\theta}({\bm{x}}|{\bm{y}})} \Big].\\
\end{aligned}
\end{equation}
This optimal distribution is an estimation to the actual distribution (or data distribution), {\em{i.e.}}, $p_\theta({\bm{x}}|{\bm{y}}) \approx p_{\text{data}}({\bm{x}}|{\bm{y}})$. We can follow the similar steps for $p_\theta({\bm{y}}|{\bm{x}}) \approx p_{\text{data}}({\bm{y}}|{\bm{x}})$. Thus following~\Cref{eq:app:EBM_01}, the objective function is to maximize
\begin{equation} \label{eq:app:EBM_02}
\begin{aligned}
\nu \cdot \mathbb{E}_{p_{\text{data}}({\bm{y}})} \mathbb{E}_{p_{n}({\bm{x}}|{\bm{y}})} \Big[ \log \frac{\nu \cdot p_n({\bm{x}}|{\bm{y}})}{\nu \cdot p_n({\bm{x}}|{\bm{y}}) + p_{\theta}({\bm{x}}|{\bm{y}})} \Big] + \mathbb{E}_{p_{\text{data}}({\bm{y}})} \mathbb{E}_{p_{\text{data}}({\bm{x}}|{\bm{y}} )} \Big[ \log \frac{p_\theta({\bm{x}}|{\bm{y}})}{\nu \cdot p_n({\bm{x}}|{\bm{y}}) + p_{\theta}({\bm{x}}|{\bm{y}})} \Big].
\end{aligned}
\end{equation}
The we will adopt three strategies to approximate~\Cref{eq:app:EBM_02}:
\begin{enumerate}
\item \textbf{Self-normalization.} When the EBM is very expressive, {\em{i.e.}}, using deep neural network for modeling, we can assume it is able to approximate the normalized density directly~\cite{mnih2012fast,song2021train}. In other words, we can set the partition function $A=1$. This is a self-normalized EBM-NCE, with normalizing constant close to 1, {\em{i.e.}}, $p({\bm{x}}) = \exp(-E({\bm{x}})) = \exp(f({\bm{x}}))$ in~\Cref{eq:app:EBM_original}.
\item \textbf{Exponential tilting term.} Exponential tilting term~\cite{arbel2020generalized} is another useful trick. It models the distribution as $\tilde p_\theta({\bm{x}}) = q({\bm{x}}) \exp(-E_\theta({\bm{x}})) $, where $q({\bm{x}})$ is the reference distribution. If we use the same reference distribution as the noise distribution, the tilted probability is $\tilde p_\theta({\bm{x}}) = p_n({\bm{x}}) \exp(-E_\theta({\bm{x}}))$ in~\Cref{eq:app:EBM_original}.
\item \textbf{Sampling.} For many cases, we only need to sample 1 negative points for each data, {\em{i.e.}}, $\nu=1$.
\end{enumerate}
Following these three disciplines, the objective function to optimize $p_{\theta}({\bm{x}}|{\bm{y}})$ becomes
\begin{equation}
\begin{aligned}
& \mathbb{E}_{p_{n}({\bm{x}}|{\bm{y}})} \Big[ \log \frac{ p_n({\bm{x}}|{\bm{y}})}{ p_n({\bm{x}}|{\bm{y}}) + \tilde p_{\theta}({\bm{x}}|{\bm{y}})} \Big] + \mathbb{E}_{p_{\text{data}}({\bm{x}}|{\bm{y}} )} \Big[ \log \frac{\tilde p_\theta({\bm{x}}|{\bm{y}})}{ p_n({\bm{x}}|{\bm{y}}) + \tilde p_{\theta}({\bm{x}}|{\bm{y}})} \Big]\\
= & \mathbb{E}_{p_{n}({\bm{x}}|{\bm{y}})} \Big[ \log \frac{1}{1 + p_{\theta}({\bm{x}}|{\bm{y}})} \Big] + \mathbb{E}_{p_{\text{data}}({\bm{x}}|{\bm{y}} )} \Big[ \log \frac{p_\theta({\bm{x}}|{\bm{y}})}{1 + p_{\theta}({\bm{x}}|{\bm{y}})} \Big]\\
= & \mathbb{E}_{p_{n}({\bm{x}}|{\bm{y}})} \Big[ \log \frac{\exp (- f_{\bm{x}}({\bm{x}}, {\bm{y}}))}{\exp (- f_{\bm{x}}({\bm{x}}, {\bm{y}})) + 1} \Big] + \mathbb{E}_{p_{\text{data}}({\bm{x}}|{\bm{y}} )} \Big[ \log \frac{1}{\exp (- f_{\bm{x}}({\bm{x}}, {\bm{y}})) + 1} \Big]\\
= & \mathbb{E}_{p_{n}({\bm{x}}|{\bm{y}})} \Big[ \log \big(1-\sigma( f_{\bm{x}}({\bm{x}}, {\bm{y}}))\big) \Big] + \mathbb{E}_{p_{\text{data}}({\bm{x}}|{\bm{y}} )} \Big[ \log \sigma( f_{\bm{x}}({\bm{x}}, {\bm{y}})) \Big].
\end{aligned}
\end{equation}
Thus, the final EBM-NCE contrastive SSL objective is
\begin{equation}
\begin{aligned}
\mathcal{L}_{\text{EBM-NCE}}
& = -\frac{1}{2} \mathbb{E}_{p_{\text{data}}({\bm{y}})} \Big[\mathbb{E}_{p_{n}({\bm{x}}|{\bm{y}})} \log \big(1-\sigma( f_{\bm{x}}({\bm{x}}, {\bm{y}}))\big) + \mathbb{E}_{p_{\text{data}}({\bm{x}}|{\bm{y}} )} \log \sigma( f_{\bm{x}}({\bm{x}}, {\bm{y}})) \Big]\\
& ~~~~ - \frac{1}{2} \mathbb{E}_{p_{\text{data}}({\bm{x}})} \Big[\mathbb{E}_{p_{n}({\bm{y}}|{\bm{x}})} \log \big(1-\sigma( f_{\bm{y}}({\bm{y}},{\bm{x}}))\big) + \mathbb{E}_{p_{\text{data}}({\bm{y}},{\bm{x}})} \log \sigma( f_{\bm{y}}({\bm{y}},{\bm{x}})) \Big].
\end{aligned}
\end{equation}
\subsection{EBM-NCE v.s. JSE and InfoNCE}
We acknowledge that there are many other contrastive objectives~\citep{poole2019variational} that can be used to maximize MI. However, in the research line of graph SSL, as summarized in several recent survey papers~\citep{xie2021self,liu2021graph,wu2021self}, the two most used ones are InfoNCE and Jensen-Shannon Estimator (JSE)~\citep{nowozin2016f,hjelm2018learning}.
We conclude that JSE is very similar to EBM-NCE, while the underlying perspectives are totally different, as explained below.
\begin{enumerate}
\item \textbf{Derivation and Intuition.} Derivation process and underlying intuition are different. JSE~\citep{nowozin2016f} starts from f-divergence, then with variational estimation and Fenchel duality on function $f$. Our proposed EBM-NCE is more straightforward: it models the conditional distribution in the MI lower bound~\Cref{eq:app:MI_objective} with EBM, and solves it using NCE.
\item \textbf{Flexibility.} Modeling the conditional distribution with EBM provides a broader family of algorithms. NCE is just one solution to it, and recent progress on score matching~\citep{song2020score,song2021train} and contrastive divergence~\citep{du2020improved}, though no longer contrastive SSL, adds on more promising directions. Thus, EBM can provide a potential unified framework for structuring our understanding of self-supervised learning.
\item \textbf{Noise distribution.} Starting from~\citep{hjelm2018learning}, all the following works on graph SSL~\cite{sun2019infograph,xie2021self,liu2021graph,wu2021self} have been adopting the empirical distribution for noise distribution. However, this is not the case in EBM-NCE. Classic EBM-NCE uses fixed distribution, while more recent work~\citep{arbel2020generalized} extends it with adaptively learnable noise distribution. With this discipline, more advanced sampling strategies (w.r.t. the noise distribution) can be proposed, {\em{e.g.}}, adversarial negative sampling in \citep{hu2021adco}.
\end{enumerate}
In the above, we conclude three key differences between EBM-NCE and JSE, plus the solid and straightforward derivations on EBM-NCE. We believe this can provide a insightful perspective of SSL to the community.
According to the empirical results~\Cref{sec:experiment_each_loss_component}, we observe that EBM-NCE is better than InfoNCE. This can be explained using the claim from~\citep{khosla2020supervised}, where the main technical contribution is to construct many positives and many negatives per anchor point. The binary cross-entropy in EBM-NCE is able to realize this to some extent: make all the positive pairs positive and all the negative pairs negative, where the softmax-based cross-entropy fails to capture this, as in InfoNCE.
To conclude, we are introduce using EBM in modeling MI, which opens many potential venues. As for contrastive SSL, EBM-NCE provides a better perspective than JSE, and is better than InfoNCE on graph-level self-supervised learning.
\section{Generative Self-Supervised Learning} \label{sec:app:generative_ssl}
Generative SSL is another classic track for unsupervised pre-training~\citep{kingma2013auto,larsson2016learning,kingma2018glow}, though the main focus is on distribution learning. In GraphMVP, we start with VAE for the following reasons:
\begin{enumerate}
\item One of the biggest attributes of our problem is that the mapping between two views are stochastic: multiple 3D conformers can correspond to the same 2D topology. Thus, we expect a stochastic model~\citep{nielsen2020survae} like VAE, instead of the deterministic ones.
\item For pre-training and fine-tuning, we need to learn an explicit and powerful representation function that can be used for downstream tasks.
\item The decoder for structured data like graph are often complicated, {\em{e.g.}.}, the auto-regressive generation. This makes them suboptimal.
\end{enumerate}
To cope with these challenges, in GraphMVP, we start with VAE-like generation model, and later propose a \textit{light-weighted} and \textit{smart} surrogate loss as objective function.
Notice that for notation simplicity, for this section, we use $h_{{\bm{y}}}$ and $h_{{\bm{x}}}$ to delegate the 2D and 3D GNN respectively.
\subsection{Variational Molecule Reconstruction}
As shown in~\Cref{eq:app:MI_objective}, our main motivation is to model the conditional likelihood:
\begin{equation*}
\begin{aligned}
\mathcal{L}_{\text{MI}} & = -\frac{1}{2} \mathbb{E}_{p({\bm{x}},{\bm{y}})} [\log p({\bm{x}}|{\bm{y}}) + \log p({\bm{y}}|{\bm{x}})]
\end{aligned}
\end{equation*}
By introducing a reparameterized variable ${\bm{z}}_{\bm{x}} = \mu_{{\bm{x}}} + \sigma_{{\bm{x}}} \odot \epsilon$, where $\mu_{\bm{x}}$ and $\sigma_{\bm{x}}$ are two flexible functions on $h_{\bm{x}}$, $\epsilon \sim \mathcal{N}(0,I)$ and $\odot$ is the element-wise production,
we have a lower bound on the conditional likelihood:
\begin{equation} \label{eq:app:log_likelihood_01}
\begin{aligned}
\log p({\bm{y}}|{\bm{x}})
\ge \mathbb{E}_{q({\bm{z}}_{\bm{x}}|{\bm{x}})} \big[ \log p({\bm{y}}|{\bm{z}}_{\bm{x}}) \big] - KL(q({\bm{z}}_{\bm{x}}|{\bm{x}}) || p({\bm{z}}_{\bm{x}})).
\end{aligned}
\end{equation}
Similarly, we have
\begin{equation} \label{eq:app:log_likelihood_02}
\log p({\bm{x}}|{\bm{y}}) \ge \mathbb{E}_{q({\bm{z}}_{\bm{y}}|{\bm{y}})} \big[ \log p({\bm{x}}|{\bm{z}}_{\bm{y}}) \big] - KL(q({\bm{z}}_{\bm{y}}|{\bm{y}}) || p({\bm{z}}_{\bm{y}})),
\end{equation}
where ${\bm{z}}_{\bm{y}} = \mu_{\bm{y}} + \sigma_{\bm{y}} \odot \epsilon$. Here $\mu_{\bm{y}}$ and $\sigma_{\bm{y}}$ are flexible functions on $h_{\bm{y}}$, and $\epsilon \sim \mathcal{N}(0, I)$.
For implementation, we take multi-layer perceptrons (MLPs) for $\mu_{\bm{x}}, \mu_{\bm{y}}, \sigma_{\bm{x}}, \sigma_{\bm{y}}$.
Both the above objectives are composed of a conditional log-likelihood and a KL-divergence. The conditional log-likelihood has also been recognized as the \textit{reconstruction term}: it is essentially to reconstruct the 3D conformers (${\bm{y}}$) from the sampled 2D molecular graph representation (${\bm{z}}_{{\bm{x}}}$). However, performing the graph reconstruction on the data space is not easy: since molecules are discrete, modeling and measuring are not trivial.
\subsection{Variational Representation Reconstruction}
To cope with data reconstruction issue, we propose a novel generative loss termed variation representation reconstruction (VRR). The pipeline is in~\Cref{fig:app:generative_ssl}.
\begin{figure}[H]
\centering
\includegraphics[width=\textwidth]{figures/Diagram_Generative_final.pdf}
\vspace{-4ex}
\caption{VRR SSL in GraphMVP. The black dashed circles represent subgraph masking.
}
\label{fig:app:generative_ssl}
\hfill
\end{figure}
Our proposed solution is very straightforward. Recall that MI is invariant to continuous bijective function~\citep{belghazi2018mutual}. So suppose we have a representation function $h_{{\bm{y}}}$ satisfying this condition, and this can guide us a surrogate loss by transferring the reconstruction from data space to the continuous representation space:
\begin{equation*}
\begin{aligned}
\mathbb{E}_{q({\bm{z}}_{\bm{x}}|{\bm{x}})}[\log p({\bm{y}}|{\bm{z}}_{\bm{x}})]
& = - \mathbb{E}_{q({\bm{z}}_{\bm{x}}|{\bm{x}})}[ \| h_{{\bm{y}}}(g_x({\bm{z}}_{\bm{x}})) - h_{{\bm{y}}}({\bm{y}}) \|_2^2 ] + C,
\end{aligned}
\end{equation*}
where $g_x$ is the decoder and $C$ is a constant, and this introduces to using the mean-squared error (MSE) for \textbf{reconstruction on the representation space}.
Then for the reconstruction, current formula has two steps: i) the latent code $z_{{\bm{x}}}$ is first mapped to molecule space, and ii) it is mapped to the representation space. We can approximate these two mappings with one projection step, by directly projecting the latent code $z_{{\bm{x}}}$ to the 3D representation space, {\em{i.e.}}, $q_{\bm{x}}(z_{\bm{x}}) \approx h_{{\bm{y}}}( g_{\bm{x}} ( z_{\bm{x}} ))$. This gives us a variation representation reconstruction (VRR) SSL objective as below:
\begin{equation*}
\begin{aligned}
\mathbb{E}_{q({\bm{z}}_{\bm{x}}|{\bm{x}})}[\log p({\bm{y}}|{\bm{z}}_{\bm{x}})]
& = - \mathbb{E}_{q({\bm{z}}_{\bm{x}}|{\bm{x}})}[ \| q_x({\bm{z}}_{\bm{x}}) - h_{{\bm{y}}}({\bm{y}}) \|_2^2 ] + C.
\end{aligned}
\end{equation*}
\paragraph{$\beta$-VAE}
We consider introducing a $\beta$ variable~\citep{higgins2016beta} to control the disentanglement of the latent representation. To be more specific, we would have
\begin{equation}
\begin{aligned}
\log p({\bm{y}}|{\bm{x}})
& \ge \mathbb{E}_{q({\bm{z}}_{\bm{x}}|{\bm{x}})} \big[ \log p({\bm{y}}|{\bm{z}}_{\bm{x}}) \big] - \beta \cdot KL(q({\bm{z}}_{\bm{x}}|{\bm{x}}) || p({\bm{z}}_{\bm{x}})).
\end{aligned}
\end{equation}
\paragraph{Stop-gradient}
For the optimization on variational representation reconstruction, related work have found that adding the stop-gradient operator (SG) as a regularizer can make the training more stable without collapse both empirically~\citep{grill2020bootstrap,chen2021exploring} and theoretically~\citep{tian2021understanding}. Here, we may as well utilize this SG operation in the objective function:
\begin{equation}
\mathbb{E}_{q({\bm{z}}_{\bm{x}}|{\bm{x}})}[\log p({\bm{y}}|{\bm{z}}_{\bm{x}})] = - \mathbb{E}_{q({\bm{z}}_{\bm{x}}|{\bm{x}})}[ \| q_x({\bm{z}}_{\bm{x}}) - \text{SG}(h_{{\bm{y}}}({\bm{y}})) \|_2^2 ] + C.
\end{equation}
\paragraph{Objective function for VRR}
Thus, combining both two regularizers mentioned above, the final objective function for VRR is:
\begin{equation} \label{eq:app:final_vrr}
\begin{aligned}
\mathcal{L}_{\text{VRR}}
& = \frac{1}{2} \Big[ \mathbb{E}_{q({\bm{z}}_{\bm{x}}|{\bm{x}})} \big[ \| q_x({\bm{z}}_{\bm{x}}) - \text{SG}(h_{\bm{y}}) \|^2 \big] + \mathbb{E}_{q({\bm{z}}_{\bm{y}}|{\bm{y}})} \big[ \| q_y({\bm{z}}_{\bm{y}}) - \text{SG}(h_{\bm{x}}) \|_2^2 \big] \Big]\\
& \quad\quad + \frac{\beta}{2} \cdot \Big[ KL(q({\bm{z}}_{\bm{x}}|{\bm{x}}) || p({\bm{z}}_{\bm{x}})) + KL(q({\bm{z}}_{\bm{y}}|{\bm{y}}) || p({\bm{z}}_{\bm{y}})) \Big].
\end{aligned}
\end{equation}
Note that MI is invariant to continuous bijective function~\citep{belghazi2018mutual}, thus this surrogate loss would be exact if the encoding function $h_{{\bm{y}}}$ and $h_{{\bm{x}}}$ satisfy this condition. However, we find GNN (both GIN and SchNet) can, though do not meet the condition, provide quite robust performance empirically, which justify the effectiveness of VRR.
\subsection{Variational Representation Reconstruction and Non-Contrastive SSL}
By introducing VRR, we provide another perspective to understand the generative SSL, including the recently-proposed non-contrastive SSL~\citep{grill2020bootstrap,chen2021exploring}.
We provide a unified structure on the intra-data generative SSL:
\begin{itemize}
\item Reconstruction to the data space, like~\Cref{eq:variational_lower_bound,eq:app:log_likelihood_01,eq:app:log_likelihood_02}.
\item Reconstruction to the representation space, {\em{i.e.}}, VRR in~\Cref{eq:app:final_vrr}.
\begin{itemize}
\item If we \textbf{remove the stochasticity}, then it is simply the representation reconstruction (RR), as we tested in the ablation study~\Cref{sec:experiment_each_loss_component}.
\item If we \textbf{remove the stochasticity} and assume two views are \textbf{sharing the same representation function}, like CNN for multi-view learning on images, then it is reduced to the BYOL~\citep{grill2020bootstrap} and SimSiam~\citep{chen2021exploring}. In other words, these recently-proposed non-contrastive SSL methods are indeed special cases of VRR.
\end{itemize}
\end{itemize}
\section{Dataset Overview} \label{app:sec:dataset}
\subsection{Pre-Training Dataset Overview}
In this section, we provide the basic statistics of the pre-training dataset (GEOM).
In~\Cref{fig:hist}, we plot the histogram (logarithm scale on the y-axis) and cumulative distribution on the number of conformers of each molecule. As shown by the histogram and curves, there are certain number of molecules having over 1000 possible 3d conformer structures, while over 80\% of molecules have less than 100 conformers.
\begin{figure}[htb!]
\centering
\includegraphics[width=\textwidth]{figures/hist.pdf}
\caption{Statistics on the conformers of each molecule\label{fig:hist}}
\end{figure}
In~\Cref{fig:hist}, we plot the histogram of the summation of top (descending sorted by weights) \{1,5,10,20\} conformer weights. The physical meaning of the weight is the portion of each conformer occurred in nature. We observe that the top 5 or 10 conformers are sufficient as they have dominated nearly all the natural observations. Such long-tailed distribution is also in alignment with our findings in the ablation studies. We find that utilizing top five conformers in the GraphMVP\, has reached an idealised spot between effectiveness and efficiency.
\begin{figure}[htb!]
\centering
\includegraphics[width=\textwidth]{figures/weight_conf.pdf}
\caption{Sum of occurrence weights for the top major conformers\label{fig:cumulative}}
\end{figure}
\subsection{Downstream Dataset Overview} \label{app:sec:downstream_datasets}
In this section, we review the four main categories of datasets used for downstream tasks.
\paragraph{Molecular Property: Pharmacology}
The Blood-Brain Barrier Penetration (BBBP)~\cite{martins2012bayesian} dataset measures whether a molecule will penetrate the central nervous system.
All three datasets, Tox21~\cite{tox21}, ToxCast~\cite{wu2018moleculenet}, and ClinTox~\cite{gayvert2016data} are related to the toxicity of molecular compounds.
The Side Effect Resource (SIDER)~\cite{kuhn2015sider} dataset stores the adverse drug reactions on a marketed drug database.
\paragraph{Molecular Property: Physical Chemistry}
Dataset proposed in~\cite{delaney2004esol} measures aqueous solubility of the molecular compounds. Lipophilicity (Lipo) dataset is a subset of ChEMBL~\cite{gaulton2012chembl} measuring the molecule octanol/water distribution coefficient. CEP dataset is a subset of the Havard Clean Energy Project (CEP)~\cite{hachmann2011harvard}, which estimates the organic photovoltaic efficiency.
\paragraph{Molecular Property: Biophysics}
Maximum Unbiased Validation (MUV)~\cite{rohrer2009maximum} is another sub-database from PCBA, and is obtained by applying a refined nearest neighbor analysis. HIV is from the Drug Therapeutics Program (DTP) AIDS Antiviral Screen~\cite{zaharevitz2015aids}, and it aims at predicting inhibit HIV replication. BACE measures the binding results for a set of inhibitors of $\beta$-secretase 1 (BACE-1), and is gathered in MoleculeNet~\cite{wu2018moleculenet}. Malaria~\cite{gamo2010thousands} measures the drug efficacy against the parasite that causes malaria.
\paragraph{Drug-Target Affinity}
Davis~\citep{davis2011comprehensive} measures the binding affinities between kinase inhibitors and kinases, scored by the $K_d$ value (kinase dissociation constant).
KIBA~\citep{tang2014making} contains binding affinities for kinase inhibitors from different sources, including $K_i$, $K_d$ and $\text{IC}_{50}$. KIBA scores~\citep{ozturk2018deepdta} are constructured to optimize the consistency among these values.
\begin{table}[H]
\centering
\small
\caption{Summary for the molecule chemical datasets.}
\begin{tabular}{l l r r r r}
\toprule
Dataset & Task & \# Tasks & \# Molecules & \# Proteins & \# Molecule-Protein pairs\\
\midrule
BBBP & Classification & 1 & 2,039 & -& -\\
Tox21 & Classification & 12 & 7,831 & -& -\\
ToxCast & Classification & 617 & 8,576 & -& -\\
Sider & Classification & 27 & 1,427 & -& -\\
ClinTox & Classification & 2 & 1,478 & -& -\\
MUV & Classification & 17 & 93,087 & -& -\\
HIV & Classification & 1 & 41,127 & -& -\\
Bace & Classification & 1 & 1,513 & -& -\\
\midrule
Delaney & Regression & 1 & 1,128 & -& -\\
Lipo & Regression & 1 & 4,200 & -& -\\
Malaria & Regression & 1 & 9,999 & -& -\\
CEP & Regression & 1 & 29,978 & -& -\\
\midrule
Davis & Regression & 1 & 68 & 379 & 30,056 \\
KIBA & Regression & 1 & 2,068 & 229 & 118,254 \\
\bottomrule
\end{tabular}
\label{tab:mol_dataset_summary}
\end{table}
\section{Experiments Details}
\subsection{Self-supervised Learning Baselines} \label{appendix:hyper}
For the SSL baselines in main results (\Cref{tab:main_results}), generally we can match with the original paper, even though most of them are using larger pre-training datasets, like ZINC-2m. Yet, we would like to add some specifications.
\begin{itemize}
\item G-\{Contextual, Motif\}\citep{rong2020self} proposes a new GNN model for backbone model, and does pre-training on a larger dataset. Both settings are different from us.
\item JOAO~\citep{you2021graph} has two versions in the original paper. In this paper, we run both versions and report the optimal one.
\item Almost all the graph SSL baselines are reporting the test performance with optimal validation error, while GraphLoG~\citep{xu2021self} reports 73.2 in the paper with the last-epoch performance. This can be over-optimized in terms of overfitting, and here we rerun it with the same downstream evaluation strategy as a fair comparison.
\end{itemize}
\clearpage
\subsection{Ablation Study: The Effect of Masking Ratio and Number of Conformers} \label{app:sec:effect_of_M_and_C}
\begin{table}[H]
\caption{
Full results for ablation of masking ratio $M$ ($C=0.15$), MVP is short for GraphMVP.
\vspace{-0.3cm}
}
\fontsize{9pt}{2pt}
\selectfont
\setlength{\tabcolsep}{2pt}
\centering
\begin{tabular}{l l c c c c c c c c c}
\toprule
& $M$\,\,\,\,\,\, & BBBP & Tox21 & ToxCast & Sider & ClinTox & MUV & HIV & Bace & Avg\\
\midrule
-- & -- & 65.4(2.4) & 74.9(0.8) & 61.6(1.2) & 58.0(2.4) & 58.8(5.5) & 71.0(2.5) & 75.3(0.5) & 72.6(4.9) & 67.21 \\
\midrule
\multirow{3}{*}{MVP}
& 0 & 69.4 (1.0)& 75.3 (0.5)& 62.8 (0.2)& 61.9 (0.5)& 74.4 (1.3)& 74.6 (1.4)& 74.6 (1.0)& 76.0 (2.0) & 71.12\\
& 0.15 & 68.5 (0.2)& 74.5 (0.4)& 62.7 (0.1)& 62.3 (1.6)& 79.0 (2.5)& 75.0 (1.4)& 74.8 (1.4)& 76.8 (1.1) & 71.69\\
& 0.3 & 68.6 (0.3)& 74.9 (0.6)& 62.8 (0.4)& 60.0 (0.6)& 74.8 (7.8)& 74.7 (0.8)& 75.5 (1.1)& 82.9 (1.7) & 71.79\\
\midrule
\multirow{3}{*}{MVP-G}
& 0 & 72.4 (1.3)& 74.7 (0.6)& 62.4 (0.2)& 60.3 (0.7)& 76.2 (5.7)& 76.6 (1.7)& 76.4 (1.7)& 78.0 (1.1) & 72.15\\
& 0.15 & 70.8 (0.5)& 75.9 (0.5)& 63.1 (0.2)& 60.2 (1.1)& 79.1 (2.8)& 77.7 (0.6)& 76.0 (0.1)& 79.3 (1.5) & 72.76\\
& 0.3 & 69.5 (0.5)& 74.6 (0.6)& 62.7 (0.3)& 60.8 (1.2)& 80.7 (2.0)& 77.8 (2.5)& 76.2 (0.5)& 81.0 (1.0) & 72.91\\
\midrule
\multirow{3}{*}{MVP-C}
& 0 & 71.5 (0.9)& 75.4 (0.3)& 63.6 (0.5)& 61.8 (0.6)& 77.3 (1.2)& 75.8 (0.6)& 76.1 (0.9)& 79.8 (0.4) & 72.66\\
& 0.15 & 72.4 (1.6)& 74.4 (0.2)& 63.1 (0.4)& 63.9 (1.2)& 77.5 (4.2)& 75.0 (1.0)& 77.0 (1.2)& 81.2 (0.9) & 73.07\\
& 0.3 & 70.7 (0.8)& 74.6 (0.3)& 63.8 (0.7)& 60.4 (0.6)& 83.5 (3.2)& 74.2 (1.6)& 76.0 (1.0)& 82.2 (2.2) & 73.17\\
\bottomrule
\end{tabular}
\end{table}
\begin{table}[H]
\caption{
Full results for ablation of \# conformers $C$ ($M=0.5$), MVP is short for GraphMVP.
\vspace{-0.3cm}
}
\fontsize{9pt}{2pt}
\selectfont
\setlength{\tabcolsep}{2pt}
\centering
\begin{tabular}{l r c c c c c c c c c}
\toprule
& \,\,\,\,$C$ & BBBP & Tox21 & ToxCast & Sider & ClinTox & MUV & HIV & Bace & Avg\\
\midrule
-- & -- & 65.4(2.4) & 74.9(0.8) & 61.6(1.2) & 58.0(2.4) & 58.8(5.5) & 71.0(2.5) & 75.3(0.5) & 72.6(4.9) & 67.21 \\
\midrule
\multirow{4}{*}{MVP}
& 1 & 69.2 (1.0)& 74.7 (0.4)& 62.5 (0.2)& 63.0 (0.4)& 73.9 (7.2)& 76.2 (0.4)& 75.3 (1.1)& 78.0 (0.5) & 71.61\\
& 5 & 68.5 (0.2)& 74.5 (0.4)& 62.7 (0.1)& 62.3 (1.6)& 79.0 (2.5)& 75.0 (1.4)& 74.8 (1.4)& 76.8 (1.1) & 71.69\\
& 10 & 68.3 (0.5)& 74.2 (0.6)& 63.2 (0.5)& 61.4 (1.0)& 80.6 (0.8)& 75.4 (2.4)& 75.5 (0.6)& 79.1 (2.3) & 72.20\\
& 20 & 68.7 (0.5)& 74.9 (0.3)& 62.7 (0.3)& 60.8 (0.7)& 75.8 (0.5)& 76.3 (1.5)& 77.4 (0.3)& 82.3 (0.8) & 72.39\\
\midrule
\multirow{4}{*}{MVP-G}
& 1 & 70.9 (0.4)& 75.3 (0.7)& 62.8 (0.5)& 61.2 (0.6)& 81.4 (3.7)& 74.2 (2.1)& 76.4 (0.6)& 80.2 (0.7) & 72.80\\
& 5 & 70.8 (0.5)& 75.9 (0.5)& 63.1 (0.2)& 60.2 (1.1)& 79.1 (2.8)& 77.7 (0.6)& 76.0 (0.1)& 79.3 (1.5) & 72.76\\
& 10 & 70.2 (0.9)& 74.9 (0.4)& 63.4 (0.4)& 60.8 (1.0)& 80.6 (0.4)& 76.4 (2.0)& 77.0 (0.3)& 77.4 (1.3) & 72.59\\
& 20 & 69.5 (0.4)& 74.9 (0.4)& 63.3 (0.1)& 60.8 (0.3)& 81.2 (0.5)& 77.3 (2.7)& 76.9 (0.3)& 80.1 (0.5) & 73.00\\
\midrule
\multirow{4}{*}{MVP-C}
& 1 & 69.7 (0.9)& 74.9 (0.5)& 64.1 (0.5)& 61.0 (1.4)& 78.3 (2.7)& 75.7 (1.5)& 74.7 (0.8)& 81.3 (0.7) & 72.46\\
& 5 & 72.4 (1.6)& 74.4 (0.2)& 63.1 (0.4)& 63.9 (1.2)& 77.5 (4.2)& 75.0 (1.0)& 77.0 (1.2)& 81.2 (0.9) & 73.07\\
& 10 & 69.5 (1.5)& 74.5 (0.5)& 63.9 (0.9)& 60.9 (0.4)& 81.1 (1.8)& 76.8 (1.5)& 76.0 (0.8)& 82.0 (1.0) & 73.09\\
& 20 & 72.1 (0.4)& 73.4 (0.7)& 63.9 (0.3)& 63.0 (0.7)& 78.8 (2.4)& 74.1 (1.0)& 74.8 (0.9)& 84.1 (0.6) & 73.02\\
\bottomrule
\end{tabular}
\end{table}
\subsection{Ablation Study: Effect of Each Loss Component}
\begin{table}[H]
\caption{Molecular graph property prediction, we set $C$=5 and $M$=0.15 for GraphMVP methods.}
\small
\setlength{\tabcolsep}{2pt}
\centering
\begin{tabular}{l c c c c c c c c c}
\toprule
& BBBP & Tox21 & ToxCast & Sider & ClinTox & MUV & HIV & Bace & Avg\\
\midrule
\# Molecules & 2,039 & 7,831 & 8,575 & 1,427 & 1,478 & 93,087 & 41,127 & 1,513 & - \\
\# Tasks & 1 & 12 & 617 & 27 & 2 & 17 & 1 & 1 & - \\
\midrule
- & 65.4(2.4) & 74.9(0.8) & 61.6(1.2) & 58.0(2.4) & 58.8(5.5) & 71.0(2.5) & 75.3(0.5) & 72.6(4.9) & 67.21 \\
\midrule
InfoNCE only & 68.9(1.2) & 74.2(0.3) & 62.8(0.2) & 59.7(0.7) & 57.8(11.5) & 73.6(1.8) & 76.1(0.6) & 77.6(0.3) & 68.85 \\
EBM-NCE only & 68.0(0.3) & 74.3(0.4) & 62.6(0.3) & 61.3(0.4) & 66.0(6.0) & 73.1(1.6) & 76.4(1.0) & 79.6(1.7) & 70.15 \\
VAE only & 67.6(1.8) & 73.2(0.5) & 61.9(0.4) & 60.5(0.2) & 59.7(1.6) & 78.6(0.7) & 77.4(0.6) & 75.4(2.1) & 69.29 \\
AE only & 70.5(0.4) & 75.0(0.4) & 62.4(0.4) & 61.0(1.4) & 53.8(1.0) & 74.1(2.9) & 76.3(0.5) & 77.9(0.9) & 68.89 \\
\midrule
InfoNCE + VAE & 69.6(1.1) & 75.4(0.6) & 63.2(0.3) & 59.9(0.4) & 69.3(14.0) & 76.5(1.3) & 76.3(0.2) & 75.2(2.7) & 70.67 \\
EBM-NCE + VAE & 68.5(0.2) & 74.5(0.4) & 62.7(0.1) & 62.3(1.6) & 79.0(2.5) & 75.0(1.4) & 74.8(1.4) & 76.8(1.1) & 71.69 \\
InfoNCE + AE & 65.1(3.1) & 75.4(0.7) & 62.5(0.5) & 59.2(0.6) & 77.2(1.8) & 72.4(1.4) & 75.8(0.6) & 77.1(0.8) & 70.60 \\
EBM-NCE + AE & 69.4(1.0) & 75.2(0.1) & 62.4(0.4) & 61.5(0.9) & 71.1(6.0) & 73.3(0.3) & 75.2(0.6) & 79.3(1.1) & 70.94 \\
\bottomrule
\end{tabular}
\end{table}
\clearpage
\subsection{Broader Range of Downstream Tasks: Molecular Property Prediction Prediction} \label{sec:app:regression_results}
\begin{table}[H]
\small
\setlength{\tabcolsep}{10pt}
\centering
\caption{
Results for four molecular property prediction tasks (regression).
For each downstream task, we report the mean (and standard variance) RMSE of 3 seeds with scaffold splitting.
For GraphMVP\,, we set $M=0.15$ and $C=5$.
The best performance for each task is marked in \textbf{bold}.
}
\begin{tabular}{l c c c c c}
\toprule
& ESOL & Lipo & Malaria & CEP & Avg\\
\midrule
-- & 1.178 (0.044) & 0.744 (0.007) & 1.127 (0.003) & 1.254 (0.030) & 1.07559 \\
\midrule
AM & 1.112 (0.048) & 0.730 (0.004) & 1.119 (0.014) & 1.256 (0.000) & 1.05419 \\
CP & 1.196 (0.037) & 0.702 (0.020) & 1.101 (0.015) & 1.243 (0.025) & 1.06059 \\
JOAO & 1.120 (0.019) & 0.708 (0.007) & 1.145 (0.010) & 1.293 (0.003) & 1.06631 \\
\midrule
GraphMVP\, & 1.091 (0.021) & 0.718 (0.016) & 1.114 (0.013) & 1.236 (0.023) & 1.03968 \\
GraphMVP-G & 1.064 (0.045) & 0.691 (0.013) & 1.106 (0.013) & \textbf{1.228 (0.001)} & 1.02214 \\
GraphMVP-C & \textbf{1.029 (0.033)} & \textbf{0.681} (0.010) & \textbf{1.097 (0.017)} & 1.244 (0.009) & \textbf{1.01283} \\
\bottomrule
\end{tabular}
\end{table}
\subsection{Broader Range of Downstream Tasks: Drug-Target Affinity Prediction}
\begin{table}[H]
\small
\setlength{\tabcolsep}{15pt}
\centering
\caption{
Results for two drug-target affinity prediction tasks (regression).
For each downstream task, we report the mean (and standard variance) MSE of 3 seeds with random splitting.
For GraphMVP\,, we set $M=0.15$ and $C=5$.
The best performance for each task is marked in \textbf{bold}.
}
\centering
\begin{tabular}{l c c c}
\toprule
& Davis & KIBA & Avg\\
\midrule
& 0.286 (0.006) & 0.206 (0.004) & 0.24585 \\
\midrule
AM & 0.291 (0.007) & 0.203 (0.003) & 0.24730 \\
CP & 0.279 (0.002) & 0.198 (0.004) & 0.23823 \\
JOAO & 0.281 (0.004) & 0.196 (0.005) & 0.23871 \\
\midrule
GraphMVP & 0.280 (0.005) & 0.178 (0.005) & 0.22860 \\
GraphMVP-G & \textbf{0.274 (0.002)} & 0.175 (0.001) & 0.22476 \\
GraphMVP-C & 0.276 (0.004) & \textbf{0.168 (0.001)} & \textbf{0.22231} \\
\bottomrule
\end{tabular}
\end{table}
\subsection{Case Studies~\label{appendix:case}}
\textbf{Shape Analysis (3D Diameter Prediction).} Diameter is an important measure in molecule~\cite{liu2010using,melnik2003distance}, and genome~\cite{finn2017comparative} modelling. Usually, the longer the 2D diameter (longest adjacency path) is, the larger the 3D diameter (largest atomic pairwise l2 distance). However, this is not always true. Therefore, we are particularly interested in using the 2D graph to predict the 3D diameter when the 2D and 3D molecular landscapes are with large differences (as in~\Cref{fig:case} and ~\Cref{fig:case_appendix}). We formulate it as a $n$-class recognition problem, where $n$ is the number of class after removing the consecutive intervals. We provide numerical results in~\Cref{table:diam} and more visualisation examples in~\Cref{fig:case_vis}.
\begin{figure}[ht]
\centering
\vspace{-2ex}
\includegraphics[width=\textwidth]{figures/Case_Study_Appendix.pdf}
\caption{Molecules selection, we select the molecules that lies in the black dash box.}
\label{fig:case_appendix}
\end{figure}
\begin{table}[htb!]
\caption{Accuracy on Recognizing Molecular Spatial Diameters~\label{table:diam}}
\small
\setlength{\tabcolsep}{3pt}
\centering
\begin{tabular}{l c c c c c c c c c c}
\toprule
Random & \,AttrMask\, & \,ContextPred\, & GPT-GNN & GraphCL & JOAOv2 & MVP & MVP-G & MVP-C \\
\midrule
38.9 (0.8) & 37.6 (0.6) & 41.2 (0.7) & 39.2 (1.1) & 38.7 (2.0) & 41.3 (1.2) & 42.3 (1.9) & 41.9 (0.7) & 42.3 (1.3)\\
\bottomrule
\end{tabular}
\end{table}
\textbf{Long-Range Donor-Acceptor Detection.} Donor-Acceptor structures such as hydrogen bonds have key impacts on the molecular geometrical structures (collinear and coplanarity), and physical properties (melting point, water affinity, viscosity etc.). Usually, atom pairs such as ``O...H'' that are closed in the Euclidean space are considered as the donor-acceptor structures~\cite{kaur2020understanding}.
On this basis, we are particularly interested in using the 2D graph to recognize (i.e., binary classification) donor-acceptor structures which have larger ranges in the 2D adjacency (as shown in~\Cref{fig:case}). Similarly, we select the molecules whose donor-acceptor are close in 3D Euclidean distance but far in the 2D adjacency. We provide numerical results in~\Cref{table:donor}. Both tables show that MVP is the MVP :)
\begin{table}[htb!]
\caption{Accuracy on Recognizing Long-Range Donor-Acceptor Structures~\label{table:donor}}
\small
\setlength{\tabcolsep}{3pt}
\centering
\begin{tabular}{l c c c c c c c c c c}
\toprule
Random & \,AttrMask\, & \,ContextPred\, & GPT-GNN & GraphCL & JOAOv2 & MVP & MVP-G & MVP-C \\
\midrule
77.9 (1.1) & 78.6 (0.3) & 80.0 (0.5) & 77.5 (0.9) & 79.9 (0.7) & 79.2 (1.0) & 80.0 (0.4) & 81.5 (0.4) & 80.7 (0.2) \\
\bottomrule
\end{tabular}
\end{table}
\paragraph{Chirality.} We have also explored other tasks such as predicting the molecular chirality, it is a challenging setting if only 2D molecular graphs are provided~\cite{pattanaik2020message}. We found that GraphMVP\,brings negligible improvements due to the model capacity of SchNet. We save this in the ongoing work.
\begin{figure}[ht]
\centering
\vspace{-2ex}
\includegraphics[width=\textwidth]{figures/case_study_SSL_better.png}
\vspace{-2ex}
\caption{Molecule examples where GraphMVP successfully recognizes the 3D diameters while random initialisation fails, legends are in a format of ``molecule id''-``2d diameter''-``3d diameter''.}
\label{fig:case_vis}
\end{figure}
\section{Generative Self-Supervised Learning} \label{sec:app:generative_ssl}
Generative SSL is another classic track for unsupervised pre-training~\citep{kingma2013auto,larsson2016learning,kingma2018glow}, though the main focus is on distribution learning. In GraphMVP, we start with VAE for the following reasons:
\begin{enumerate}
\item One of the biggest attributes of our problem is that the mapping between two views are stochastic: multiple 3D conformers can correspond to the same 2D topology. Thus, we expect a stochastic model~\citep{nielsen2020survae} like VAE, instead of the deterministic ones.
\item For pre-training and fine-tuning, we need to learn an explicit and powerful representation function that can be used for downstream tasks.
\item The decoder for structured data like graph are often complicated, {\em{e.g.}.}, the auto-regressive generation. This makes them suboptimal.
\end{enumerate}
To cope with these challenges, in GraphMVP, we start with VAE-like generation model, and later propose a \textit{light-weighted} and \textit{smart} surrogate loss as objective function.
Notice that for notation simplicity, for this section, we use $h_{{\bm{y}}}$ and $h_{{\bm{x}}}$ to delegate the 2D and 3D GNN respectively.
\subsection{Variational Molecule Reconstruction}
As shown in~\Cref{eq:app:MI_objective}, our main motivation is to model the conditional likelihood:
\begin{equation*}
\begin{aligned}
\mathcal{L}_{\text{MI}} & = -\frac{1}{2} \mathbb{E}_{p({\bm{x}},{\bm{y}})} [\log p({\bm{x}}|{\bm{y}}) + \log p({\bm{y}}|{\bm{x}})]
\end{aligned}
\end{equation*}
By introducing a reparameterized variable ${\bm{z}}_{\bm{x}} = \mu_{{\bm{x}}} + \sigma_{{\bm{x}}} \odot \epsilon$, where $\mu_{\bm{x}}$ and $\sigma_{\bm{x}}$ are two flexible functions on $h_{\bm{x}}$, $\epsilon \sim \mathcal{N}(0,I)$ and $\odot$ is the element-wise production,
we have a lower bound on the conditional likelihood:
\begin{equation} \label{eq:app:log_likelihood_01}
\begin{aligned}
\log p({\bm{y}}|{\bm{x}})
\ge \mathbb{E}_{q({\bm{z}}_{\bm{x}}|{\bm{x}})} \big[ \log p({\bm{y}}|{\bm{z}}_{\bm{x}}) \big] - KL(q({\bm{z}}_{\bm{x}}|{\bm{x}}) || p({\bm{z}}_{\bm{x}})).
\end{aligned}
\end{equation}
Similarly, we have
\begin{equation} \label{eq:app:log_likelihood_02}
\log p({\bm{x}}|{\bm{y}}) \ge \mathbb{E}_{q({\bm{z}}_{\bm{y}}|{\bm{y}})} \big[ \log p({\bm{x}}|{\bm{z}}_{\bm{y}}) \big] - KL(q({\bm{z}}_{\bm{y}}|{\bm{y}}) || p({\bm{z}}_{\bm{y}})),
\end{equation}
where ${\bm{z}}_{\bm{y}} = \mu_{\bm{y}} + \sigma_{\bm{y}} \odot \epsilon$. Here $\mu_{\bm{y}}$ and $\sigma_{\bm{y}}$ are flexible functions on $h_{\bm{y}}$, and $\epsilon \sim \mathcal{N}(0, I)$.
For implementation, we take multi-layer perceptrons (MLPs) for $\mu_{\bm{x}}, \mu_{\bm{y}}, \sigma_{\bm{x}}, \sigma_{\bm{y}}$.
Both the above objectives are composed of a conditional log-likelihood and a KL-divergence. The conditional log-likelihood has also been recognized as the \textit{reconstruction term}: it is essentially to reconstruct the 3D conformers (${\bm{y}}$) from the sampled 2D molecular graph representation (${\bm{z}}_{{\bm{x}}}$). However, performing the graph reconstruction on the data space is not easy: since molecules are discrete, modeling and measuring are not trivial.
\subsection{Variational Representation Reconstruction}
To cope with data reconstruction issue, we propose a novel generative loss termed variation representation reconstruction (VRR). The pipeline is in~\Cref{fig:app:generative_ssl}.
\begin{figure}[H]
\centering
\includegraphics[width=\textwidth]{figures/Diagram_Generative_final.pdf}
\vspace{-4ex}
\caption{VRR SSL in GraphMVP. The black dashed circles represent subgraph masking.
}
\label{fig:app:generative_ssl}
\hfill
\end{figure}
Our proposed solution is very straightforward. Recall that MI is invariant to continuous bijective function~\citep{belghazi2018mutual}. So suppose we have a representation function $h_{{\bm{y}}}$ satisfying this condition, and this can guide us a surrogate loss by transferring the reconstruction from data space to the continuous representation space:
\begin{equation*}
\begin{aligned}
\mathbb{E}_{q({\bm{z}}_{\bm{x}}|{\bm{x}})}[\log p({\bm{y}}|{\bm{z}}_{\bm{x}})]
& = - \mathbb{E}_{q({\bm{z}}_{\bm{x}}|{\bm{x}})}[ \| h_{{\bm{y}}}(g_x({\bm{z}}_{\bm{x}})) - h_{{\bm{y}}}({\bm{y}}) \|_2^2 ] + C,
\end{aligned}
\end{equation*}
where $g_x$ is the decoder and $C$ is a constant, and this introduces to using the mean-squared error (MSE) for \textbf{reconstruction on the representation space}.
Then for the reconstruction, current formula has two steps: i) the latent code $z_{{\bm{x}}}$ is first mapped to molecule space, and ii) it is mapped to the representation space. We can approximate these two mappings with one projection step, by directly projecting the latent code $z_{{\bm{x}}}$ to the 3D representation space, {\em{i.e.}}, $q_{\bm{x}}(z_{\bm{x}}) \approx h_{{\bm{y}}}( g_{\bm{x}} ( z_{\bm{x}} ))$. This gives us a variation representation reconstruction (VRR) SSL objective as below:
\begin{equation*}
\begin{aligned}
\mathbb{E}_{q({\bm{z}}_{\bm{x}}|{\bm{x}})}[\log p({\bm{y}}|{\bm{z}}_{\bm{x}})]
& = - \mathbb{E}_{q({\bm{z}}_{\bm{x}}|{\bm{x}})}[ \| q_x({\bm{z}}_{\bm{x}}) - h_{{\bm{y}}}({\bm{y}}) \|_2^2 ] + C.
\end{aligned}
\end{equation*}
\paragraph{$\beta$-VAE}
We consider introducing a $\beta$ variable~\citep{higgins2016beta} to control the disentanglement of the latent representation. To be more specific, we would have
\begin{equation}
\begin{aligned}
\log p({\bm{y}}|{\bm{x}})
& \ge \mathbb{E}_{q({\bm{z}}_{\bm{x}}|{\bm{x}})} \big[ \log p({\bm{y}}|{\bm{z}}_{\bm{x}}) \big] - \beta \cdot KL(q({\bm{z}}_{\bm{x}}|{\bm{x}}) || p({\bm{z}}_{\bm{x}})).
\end{aligned}
\end{equation}
\paragraph{Stop-gradient}
For the optimization on variational representation reconstruction, related work have found that adding the stop-gradient operator (SG) as a regularizer can make the training more stable without collapse both empirically~\citep{grill2020bootstrap,chen2021exploring} and theoretically~\citep{tian2021understanding}. Here, we may as well utilize this SG operation in the objective function:
\begin{equation}
\mathbb{E}_{q({\bm{z}}_{\bm{x}}|{\bm{x}})}[\log p({\bm{y}}|{\bm{z}}_{\bm{x}})] = - \mathbb{E}_{q({\bm{z}}_{\bm{x}}|{\bm{x}})}[ \| q_x({\bm{z}}_{\bm{x}}) - \text{SG}(h_{{\bm{y}}}({\bm{y}})) \|_2^2 ] + C.
\end{equation}
\paragraph{Objective function for VRR}
Thus, combining both two regularizers mentioned above, the final objective function for VRR is:
\begin{equation} \label{eq:app:final_vrr}
\begin{aligned}
\mathcal{L}_{\text{VRR}}
& = \frac{1}{2} \Big[ \mathbb{E}_{q({\bm{z}}_{\bm{x}}|{\bm{x}})} \big[ \| q_x({\bm{z}}_{\bm{x}}) - \text{SG}(h_{\bm{y}}) \|^2 \big] + \mathbb{E}_{q({\bm{z}}_{\bm{y}}|{\bm{y}})} \big[ \| q_y({\bm{z}}_{\bm{y}}) - \text{SG}(h_{\bm{x}}) \|_2^2 \big] \Big]\\
& \quad\quad + \frac{\beta}{2} \cdot \Big[ KL(q({\bm{z}}_{\bm{x}}|{\bm{x}}) || p({\bm{z}}_{\bm{x}})) + KL(q({\bm{z}}_{\bm{y}}|{\bm{y}}) || p({\bm{z}}_{\bm{y}})) \Big].
\end{aligned}
\end{equation}
Note that MI is invariant to continuous bijective function~\citep{belghazi2018mutual}, thus this surrogate loss would be exact if the encoding function $h_{{\bm{y}}}$ and $h_{{\bm{x}}}$ satisfy this condition. However, we find GNN (both GIN and SchNet) can, though do not meet the condition, provide quite robust performance empirically, which justify the effectiveness of VRR.
\subsection{Variational Representation Reconstruction and Non-Contrastive SSL}
By introducing VRR, we provide another perspective to understand the generative SSL, including the recently-proposed non-contrastive SSL~\citep{grill2020bootstrap,chen2021exploring}.
We provide a unified structure on the intra-data generative SSL:
\begin{itemize}
\item Reconstruction to the data space, like~\Cref{eq:variational_lower_bound,eq:app:log_likelihood_01,eq:app:log_likelihood_02}.
\item Reconstruction to the representation space, {\em{i.e.}}, VRR in~\Cref{eq:app:final_vrr}.
\begin{itemize}
\item If we \textbf{remove the stochasticity}, then it is simply the representation reconstruction (RR), as we tested in the ablation study~\Cref{sec:experiment_each_loss_component}.
\item If we \textbf{remove the stochasticity} and assume two views are \textbf{sharing the same representation function}, like CNN for multi-view learning on images, then it is reduced to the BYOL~\citep{grill2020bootstrap} and SimSiam~\citep{chen2021exploring}. In other words, these recently-proposed non-contrastive SSL methods are indeed special cases of VRR.
\end{itemize}
\end{itemize}
\section{Conclusion and Future Work}
\vspace{-1ex}
In this work, we provide a very general framework, coined GraphMVP. From the domain perspective, GraphMVP\, (1) is the first to incorporate 3D information for augmenting 2D graph representation learning and (2) is able to take advantages of 3D conformers by considering stochasticity in modeling. From the aspect of technical novelties, GraphMVP\, brings following insights when introducing 2 SSL tasks:
(1) Following~\Cref{eq:MI_objective}, GraphMVP\, proposes EBM-NCE and VRR, where they are modeling the conditional distributions using EBM and variational distribution respectively.
(2) EBM-NCE is similar to JSE, while we start with a different direction for theoretical intuition, yet EBM opens another promising venue in this area.
(3) VRR, as a generative SSL method, is able to alleviate the potential issues in molecule generation~\citep{zhavoronkov2019deep,gao2020synthesizability}.
(4) Ultimately, GraphMVP\, combines both contrastive SSL (InfoNCE or EBM-NCE) and generative SSL (VRR) for objective function.
Both empirical results (solid performance improvements on 14 downstream datasets) and theoretical analysis can strongly support the above domain and technical contributions.
We want to emphasize that GraphMVP\, is model-agnostic and has the potential to be expanded to many other low-data applications. This motivates broad directions for future exploration, including but not limited to: (1) More powerful 2D and 3D molecule representation methods. (2) Different application domain other than small molecules, {\em{e.g.}}, large molecules like proteins.
\section{Self-Supervised Learning on Molecular Graph} \label{appendix:related_work}
Self-supervised learning (SSL) methods have attracted massive attention recently, trending from vision~\citep{chen2020simple,he2020momentum,caron2020unsupervised,chen2021exploring,OcCo}, language~\citep{oord2018representation,devlin2018bert,brown2020language} to graph~\citep{velivckovic2018deep,sun2019infograph,liu2018n,hu2019strategies,You2020GraphCL,you2021graph}. In general, there are two categories of SSL: contrastive and generative, where they differ on the design of the supervised signals. Contrastive SSL realizes the supervised signals at the \textbf{inter-data} level, learning the representation by contrasting with other data points; while generative SSL focuses on reconstructing the original data at the \textbf{intra-data} level. Both venues have been widely explored~\citep{liu2021graph,xie2021self,wu2021self,liu2021self}.
\subsection{Contrastive graph SSL}
Contrastive graph SSL first applies transformations to construct different \textit{views} for each graph. Each view incorporates different granularities of information, like node-, subgraph-, and graph-level. It then solves two sub-tasks simultaneously: (1) aligning the representations of views from the same data; (2) contrasting the representations of views from different data, leading to a uniformly distributed latent space~\citep{wang2020understanding}. The key difference among existing methods is thus the design of view constructions. InfoGraph~\citep{velivckovic2018deep,sun2019infograph} contrasted the node (local) and graph (global) views. ContextPred~\citep{hu2019strategies} and G-Contextual~\cite{rong2020self} contrasted between node and context views. GraphCL and JOAO~\citep{You2020GraphCL,you2021graph} made comprehensive comparisons among four graph-level transformations and further learned to select the most effective combinations.
\subsection{Generative graph SSL}
Generative graph SSL aims at reconstructing important structures for each graph. By so doing, it consequently learns a representation capable of encoding key ingredients of the data. EdgePred~\citep{hamilton2017inductive} and AttrMask~\citep{hu2019strategies} predicted the adjacency matrix and masked tokens (nodes and edges) respectively. GPT-GNN~\citep{hu2020gpt} reconstructed the whole graph in an auto-regressive approach.
\subsection{Predictive graph SSL}
There are certain SSL methods specific to the molecular graph. For example, one central task in drug discovery is to find the important substructure or motif in molecules that can activate the target interactions. G-Motif~\citep{rong2020self} adopts domain knowledge to heuristically extract motifs for each molecule, and the SSL task is to make prediction on the existence of each motif. Different from contrastive and generative SSL, recent literature~\citep{wu2021self} takes this as predictive graph SSL, where the supervised signals are self-generated labels.
\begin{table}[b]
\small
\caption{
Comparison between GraphMVP\, and existing graph SSL methods.
\vspace{-0.6ex}
}
\label{app:tab:compare}
\center
\setlength{\tabcolsep}{10pt}
\begin{tabular}{l c c c c c}
\toprule
\multirow{2}{*}{SSL Pre-training} & \multicolumn{2}{c}{Graph View} & \multicolumn{3}{c}{SSL Category}\\
\cmidrule(lr){2-3}
\cmidrule(lr){4-6}
& 2D Topology & 3D Geometry & Generative & Contrastive & Predictive\\
\midrule
EdgePred~\citep{hamilton2017inductive} & \checkmark & - & \checkmark & - & -\\
AttrMask~\citep{hu2019strategies} & \checkmark & - & \checkmark & - & -\\
GPT-GNN~\citep{hu2020gpt} & \checkmark & - & \checkmark & - & -\\
InfoGraph~\citep{velivckovic2018deep,sun2019infograph} & \checkmark & - & - & \checkmark& -\\
ContexPred~\citep{hu2019strategies} & \checkmark & - & - & \checkmark& -\\
GraphLoG~\citep{xu2021self} & \checkmark & - & - & \checkmark& -\\
G-Contextual~\citep{rong2020self} & \checkmark & - & - & \checkmark& -\\
GraphCL~\citep{You2020GraphCL} & \checkmark & - & - & \checkmark& -\\
JOAO~\citep{you2021graph} & \checkmark & - & - & \checkmark& -\\
G-Motif~\citep{rong2020self} & \checkmark & - & - & - & \checkmark\\
\midrule
GraphMVP\,(Ours) & \checkmark & \checkmark & \checkmark & \checkmark& -\\
\bottomrule
\end{tabular}
\end{table}
\textbf{SSL for Molecular Graphs.} Recall that all previous methods in~\Cref{app:tab:compare} \textbf{merely} focus on the 2D topology. However, for science-centric tasks such as molecular property prediction, 3D geometry should be incorporated as it provides complementary and comprehensive information~\citep{schutt2017schnet,liu2021spherical}. To mitigate this gap, we propose GraphMVP to leverage the 3D geometry with unsupervised graph pre-training.
\section{Molecular Graph Representation} \label{app:sec:molecule_representation}
There are two main methods for molecular graph representation learning. The first one is the molecular fingerprints. It is a hashed bit vector to describe the molecular graph. There has been re-discoveries on fingerprints-based methods~\citep{ramsundar2015massively,liu2018practical,liu2019loss,meyer2019learning,alnammi2021evaluating,jiang2021could}, while its has one main drawback: Random forest and XGBoost are very strong learning models on fingerprints, but they fail to take benefits of the pre-training strategy.
Graph neural network (GNN) has become another mainstream modeling methods for molecular graph representation. Existing methods can be generally split into two venues: 2D GNN and 3D GNN, depending on what levels of information is considered. 2D GNN focuses on the topological structures of the graph, like the adjacency among nodes, while 3D GNN is able to model the ``energy'' of molecules by taking account the spatial positions of atoms.
First, we want to highlight that GraphMVP\, is model-agnostic, {\em{i.e.}}, it can be applied to any 2D and 3D GNN representation function, yet the specific 2D and 3D representations are not the main focus of this work.
Second, we acknowledge there are a lot of advanced 3D~\citep{fuchs2020se,satorras2021n,liu2021spherical,jumper2021highly} and 2D~\citep{gilmer2017neural,yang2019analyzing,liu2018n,xu2018powerful,corso2020principal,demirel2021analysis} representation methods.
However, considering the \textit{graph SSL literature} and \textit{graph representation liteature} (illustrated below), we adopt GIN~\citep{xu2018powerful} and SchNet~\citep{schutt2017schnet} in current GraphMVP.
\subsection{2D Molecular Graph Neural Network}
The 2D representation is taking each molecule as a 2D graph, with atoms as nodes and bonds as edges, {\em{i.e.}}, $g_{\text{2D}} = (X, E)$. $X\in\mathbb{R}^{n\times d_n}$ is the atom attribute matrix, where $n$ is the number of atoms (nodes) and $d_n$ is the atom attribute dimension. $E\in\mathbb{R}^{m\times d_e}$ is the bond attribute matrix, where $m$ is the number of bonds (edges) and $d_m$ is the bond attribute dimension. Notice that here $E$ also includes the connectivity. Then we will apply a transformation function $T_{\text{2D}}$ on the topological graph. Given a 2D graph $g_{\text{2D}}$, its 2D molecular representation is:
\begin{equation}
h_{\text{2D}} = \text{GNN-2D}(T_{\text{2D}}(g_{\text{2D}})) = \text{GNN-2D}(T_{\text{2D}}(X, E)).
\end{equation}
The core operation of 2D GNN is the message passing function~\citep{gilmer2017neural}, which updates the node representation based on adjacency information. We have variants depending on the design of message and aggregation functions, and we pick GIN~\citep{xu2018powerful} in this work.
\paragraph{GIN} There has been a long research line on 2D graph representation learning~\citep{gilmer2017neural,yang2019analyzing,liu2018n,xu2018powerful,corso2020principal,demirel2021analysis}. Among these, graph isomorphism network (GIN) model~\citep{xu2018powerful} has been widely used as the backbone model in recent graph self-supervised learning work~\citep{hu2019strategies,You2020GraphCL,you2021graph}. Thus, we as well adopt GIN as the base model for 2D representation.
Recall each molecule is represented as a molecular graph, {\em{i.e.}}, $g_{\text{2D}} = (X, E)$, where $X$ and $E$ are feature matrices for atoms and bonds respectively. Then the message passing function is defined as:
\begin{equation}
z_i^{(k+1)} = \text{MLP}_{\text{atom}}^{(k+1)} \Big(z_i^{(k)} + \sum_{j \in \mathcal{N}(i)} \big( z_j^{(k)} + \text{MLP}_{\text{bond}}^{(k+1)}(E_{ij}) \big) \Big),
\end{equation}
where $z_0=X$ and $\text{MLP}_{\text{atom}}^{(k+1)}$ and $\text{MLP}_{\text{bond}}^{(k+1)}$ are the $(l+1)$-th MLP layers on the atom- and bond-level respectively. Repeating this for $K$ times, and we can encode $K$-hop neighborhood information for each center atom in the molecular data, and we take the last layer for each node/atom representation. The graph-level molecular representation is the mean of the node representation:
\begin{equation}
z({\bm{x}}) = \frac{1}{N} \sum_{i} z_i^{(K)}
\end{equation}
\subsection{3D Molecular Graph Neural Network} \label{appendix:3d_gnn}
Recently, the 3D geometric representation learning has brought breakthrough progress in molecule modeling~\citep{schutt2017schnet,fuchs2020se,satorras2021n,liu2021spherical,jumper2021highly}. 3D molecular graph additionally includes spatial locations of the atoms, which needless to be static since, in real scenarios, atoms are in continual motion on \textit{a potential energy surface}~\citep{axelrod2020geom}. The 3D structures at the local minima on this surface are named \textit{molecular conformation} or \textit{conformer}. As the molecular properties are a function of the conformer ensembles~\citep{hawkins2017conformation}, this reveals another limitation of existing mainstream methods: to predict properties from a single 2D or 3D graph cannot account for this fact~\citep{axelrod2020geom}, while our proposed method can alleviate this issue to a certain extent.
For specific 3D molecular graph, it additionally includes spatial positions of the atoms. We represent each conformer as $g_{\text{3D}} = (X, R)$, where $R \in \mathbb{R}^{n \times 3}$ is the 3D-coordinate matrix, and the corresponding representation is:
\begin{equation}
h_{\text{3D}} = \text{GNN-3D}(T_{\text{3D}}(g_{\text{3D}})) = \text{GNN-3D}(T_{\text{3D}}(X, R)),
\end{equation}
where $R$ is the 3D-coordinate matrix and $T_{\text{3D}}$ is the 3D transformation. Note that further information such as plane and torsion angles can be solved from the positions.
\paragraph{SchNet}
SchNet~\citep{schutt2017schnet} is composed of the following key steps:
\begin{equation}
\begin{aligned}
& z_i^{(0)} = \text{embedding} (x_i)\\
& z_i^{(t+1)} = \text{MLP} \Big( \sum_{j=1}^{n} f(x_j^{(t-1)}, r_i, r_j) \Big)\\
& h_i = \text{MLP} (z_i^{(K)}),
\end{aligned}
\end{equation}
where $K$ is the number of hidden layers, and
\begin{equation}
f(x_j, r_i, r_j) = x_j \cdot e_k(r_i - r_j) = x_j \cdot \exp(- \gamma \| \|r_i - r_j\|_2 - \mu \|_2^2)
\end{equation}
is the continuous-filter convolution layer, enabling the modeling of continuous positions of atoms.
We adopt SchNet for the following reasons. (1) SchNet is a very strong geometric representation method after \textit{fair} benchmarking. (2) SchNet can be trained more efficiently, comparing to the other recent 3D models. To support these two points, we make a comparison among the most recent 3D geometric models~\cite{fuchs2020se,satorras2021n,liu2021spherical} on QM9 dataset. QM9~\citep{wu2018moleculenet} is a molecule dataset approximating 12 thermodynamic properties calculated by density functional theory (DFT) algorithm. Notice: UNiTE~\citep{qiao2021unite} is the state-of-the-art 3D GNN, but it requires a commercial software for feature extraction, thus we exclude it for now.
\begin{table}[H]
\centering
\scriptsize
\caption{
Reproduced MAE on QM9. 100k for training, 17,748 for val, 13,083 for test. The last column is the approximated running time.
}
\label{tab:app:qm9}
\begin{tabular}{l c c c c c c c c c c c c r}
\toprule
& alpha & gap & homo & lumo & mu & cv & g298 & h298 & r2 & u298 & u0 & zpve & time\\
\midrule
SchNet~\citep{schutt2017schnet} & 0.077 & 50 & 32 & 26 & 0.030 & 0.032 & 15 & 14 & 0.122 & 14 & 14 & 1.751 & 3h\\
SE(3)-Trans~\citep{fuchs2020se} & 0.143 & 59 & 36 & 36 & 0.052 & 0.068 & 68 & 72 & 1.969 & 68 & 74 & 5.517 & 50h\\
EGNN~\citep{satorras2021n} & 0.075 & 49 & 29 & 26 & 0.030 & 0.032 & 11 & 10 & 0.076 & 10 & 10 & 1.562 & 24h\\
SphereNet~\citep{liu2021spherical} & 0.054 & 41 & 22 & 19 & 0.028 & 0.027 & 10 & 8 & 0.295 & 8 & 8 & 1.401 & 50h\\
\bottomrule
\end{tabular}
\end{table}
\Cref{tab:app:qm9} shows that, under a fair comparison (w.r.t. data splitting, seed, cuda version, etc), SchNet can reach pretty comparable performance, yet the efficiency of SchNet is much better. Combining these two points, we adopt SchNet in current version of GraphMVP.
\subsection{Summary}
To sum up, in GraphMVP, the most important message we want to deliver is how to design a well-motivated SSL algorithm to extract useful 3D geometry information to augment the 2D representation for downstream fine-tuning. GraphMVP\, is model-agnostic, and we may as well leave the more advanced 3D~\citep{fuchs2020se,satorras2021n,liu2021spherical,jumper2021highly} and 2D~\citep{yang2019analyzing,liu2018n,corso2020principal} GNN for future exploration.
In addition, molecular property prediction tasks have rich alternative representation methods, including SMILES~\cite{weininger1988smiles,hirohara2018convolutional}, and biological knowledge graph~\cite{wang2021multi,liu2022structured}. There have been another SSL research line on them~\cite{liustructured,zhu2021dual,fang2021molecular}, yet they are beyond the scope of discussion in this paper.
\section{Experiments Details}
\subsection{Self-supervised Learning Baselines} \label{appendix:hyper}
For the SSL baselines in main results (\Cref{tab:main_results}), generally we can match with the original paper, even though most of them are using larger pre-training datasets, like ZINC-2m. Yet, we would like to add some specifications.
\begin{itemize}
\item G-\{Contextual, Motif\}\citep{rong2020self} proposes a new GNN model for backbone model, and does pre-training on a larger dataset. Both settings are different from us.
\item JOAO~\citep{you2021graph} has two versions in the original paper. In this paper, we run both versions and report the optimal one.
\item Almost all the graph SSL baselines are reporting the test performance with optimal validation error, while GraphLoG~\citep{xu2021self} reports 73.2 in the paper with the last-epoch performance. This can be over-optimized in terms of overfitting, and here we rerun it with the same downstream evaluation strategy as a fair comparison.
\end{itemize}
\clearpage
\subsection{Ablation Study: The Effect of Masking Ratio and Number of Conformers} \label{app:sec:effect_of_M_and_C}
\begin{table}[H]
\caption{
Full results for ablation of masking ratio $M$ ($C=0.15$), MVP is short for GraphMVP.
\vspace{-0.3cm}
}
\fontsize{9pt}{2pt}
\selectfont
\setlength{\tabcolsep}{2pt}
\centering
\begin{tabular}{l l c c c c c c c c c}
\toprule
& $M$\,\,\,\,\,\, & BBBP & Tox21 & ToxCast & Sider & ClinTox & MUV & HIV & Bace & Avg\\
\midrule
-- & -- & 65.4(2.4) & 74.9(0.8) & 61.6(1.2) & 58.0(2.4) & 58.8(5.5) & 71.0(2.5) & 75.3(0.5) & 72.6(4.9) & 67.21 \\
\midrule
\multirow{3}{*}{MVP}
& 0 & 69.4 (1.0)& 75.3 (0.5)& 62.8 (0.2)& 61.9 (0.5)& 74.4 (1.3)& 74.6 (1.4)& 74.6 (1.0)& 76.0 (2.0) & 71.12\\
& 0.15 & 68.5 (0.2)& 74.5 (0.4)& 62.7 (0.1)& 62.3 (1.6)& 79.0 (2.5)& 75.0 (1.4)& 74.8 (1.4)& 76.8 (1.1) & 71.69\\
& 0.3 & 68.6 (0.3)& 74.9 (0.6)& 62.8 (0.4)& 60.0 (0.6)& 74.8 (7.8)& 74.7 (0.8)& 75.5 (1.1)& 82.9 (1.7) & 71.79\\
\midrule
\multirow{3}{*}{MVP-G}
& 0 & 72.4 (1.3)& 74.7 (0.6)& 62.4 (0.2)& 60.3 (0.7)& 76.2 (5.7)& 76.6 (1.7)& 76.4 (1.7)& 78.0 (1.1) & 72.15\\
& 0.15 & 70.8 (0.5)& 75.9 (0.5)& 63.1 (0.2)& 60.2 (1.1)& 79.1 (2.8)& 77.7 (0.6)& 76.0 (0.1)& 79.3 (1.5) & 72.76\\
& 0.3 & 69.5 (0.5)& 74.6 (0.6)& 62.7 (0.3)& 60.8 (1.2)& 80.7 (2.0)& 77.8 (2.5)& 76.2 (0.5)& 81.0 (1.0) & 72.91\\
\midrule
\multirow{3}{*}{MVP-C}
& 0 & 71.5 (0.9)& 75.4 (0.3)& 63.6 (0.5)& 61.8 (0.6)& 77.3 (1.2)& 75.8 (0.6)& 76.1 (0.9)& 79.8 (0.4) & 72.66\\
& 0.15 & 72.4 (1.6)& 74.4 (0.2)& 63.1 (0.4)& 63.9 (1.2)& 77.5 (4.2)& 75.0 (1.0)& 77.0 (1.2)& 81.2 (0.9) & 73.07\\
& 0.3 & 70.7 (0.8)& 74.6 (0.3)& 63.8 (0.7)& 60.4 (0.6)& 83.5 (3.2)& 74.2 (1.6)& 76.0 (1.0)& 82.2 (2.2) & 73.17\\
\bottomrule
\end{tabular}
\end{table}
\begin{table}[H]
\caption{
Full results for ablation of \# conformers $C$ ($M=0.5$), MVP is short for GraphMVP.
\vspace{-0.3cm}
}
\fontsize{9pt}{2pt}
\selectfont
\setlength{\tabcolsep}{2pt}
\centering
\begin{tabular}{l r c c c c c c c c c}
\toprule
& \,\,\,\,$C$ & BBBP & Tox21 & ToxCast & Sider & ClinTox & MUV & HIV & Bace & Avg\\
\midrule
-- & -- & 65.4(2.4) & 74.9(0.8) & 61.6(1.2) & 58.0(2.4) & 58.8(5.5) & 71.0(2.5) & 75.3(0.5) & 72.6(4.9) & 67.21 \\
\midrule
\multirow{4}{*}{MVP}
& 1 & 69.2 (1.0)& 74.7 (0.4)& 62.5 (0.2)& 63.0 (0.4)& 73.9 (7.2)& 76.2 (0.4)& 75.3 (1.1)& 78.0 (0.5) & 71.61\\
& 5 & 68.5 (0.2)& 74.5 (0.4)& 62.7 (0.1)& 62.3 (1.6)& 79.0 (2.5)& 75.0 (1.4)& 74.8 (1.4)& 76.8 (1.1) & 71.69\\
& 10 & 68.3 (0.5)& 74.2 (0.6)& 63.2 (0.5)& 61.4 (1.0)& 80.6 (0.8)& 75.4 (2.4)& 75.5 (0.6)& 79.1 (2.3) & 72.20\\
& 20 & 68.7 (0.5)& 74.9 (0.3)& 62.7 (0.3)& 60.8 (0.7)& 75.8 (0.5)& 76.3 (1.5)& 77.4 (0.3)& 82.3 (0.8) & 72.39\\
\midrule
\multirow{4}{*}{MVP-G}
& 1 & 70.9 (0.4)& 75.3 (0.7)& 62.8 (0.5)& 61.2 (0.6)& 81.4 (3.7)& 74.2 (2.1)& 76.4 (0.6)& 80.2 (0.7) & 72.80\\
& 5 & 70.8 (0.5)& 75.9 (0.5)& 63.1 (0.2)& 60.2 (1.1)& 79.1 (2.8)& 77.7 (0.6)& 76.0 (0.1)& 79.3 (1.5) & 72.76\\
& 10 & 70.2 (0.9)& 74.9 (0.4)& 63.4 (0.4)& 60.8 (1.0)& 80.6 (0.4)& 76.4 (2.0)& 77.0 (0.3)& 77.4 (1.3) & 72.59\\
& 20 & 69.5 (0.4)& 74.9 (0.4)& 63.3 (0.1)& 60.8 (0.3)& 81.2 (0.5)& 77.3 (2.7)& 76.9 (0.3)& 80.1 (0.5) & 73.00\\
\midrule
\multirow{4}{*}{MVP-C}
& 1 & 69.7 (0.9)& 74.9 (0.5)& 64.1 (0.5)& 61.0 (1.4)& 78.3 (2.7)& 75.7 (1.5)& 74.7 (0.8)& 81.3 (0.7) & 72.46\\
& 5 & 72.4 (1.6)& 74.4 (0.2)& 63.1 (0.4)& 63.9 (1.2)& 77.5 (4.2)& 75.0 (1.0)& 77.0 (1.2)& 81.2 (0.9) & 73.07\\
& 10 & 69.5 (1.5)& 74.5 (0.5)& 63.9 (0.9)& 60.9 (0.4)& 81.1 (1.8)& 76.8 (1.5)& 76.0 (0.8)& 82.0 (1.0) & 73.09\\
& 20 & 72.1 (0.4)& 73.4 (0.7)& 63.9 (0.3)& 63.0 (0.7)& 78.8 (2.4)& 74.1 (1.0)& 74.8 (0.9)& 84.1 (0.6) & 73.02\\
\bottomrule
\end{tabular}
\end{table}
\subsection{Ablation Study: Effect of Each Loss Component}
\begin{table}[H]
\caption{Molecular graph property prediction, we set $C$=5 and $M$=0.15 for GraphMVP methods.}
\small
\setlength{\tabcolsep}{2pt}
\centering
\begin{tabular}{l c c c c c c c c c}
\toprule
& BBBP & Tox21 & ToxCast & Sider & ClinTox & MUV & HIV & Bace & Avg\\
\midrule
\# Molecules & 2,039 & 7,831 & 8,575 & 1,427 & 1,478 & 93,087 & 41,127 & 1,513 & - \\
\# Tasks & 1 & 12 & 617 & 27 & 2 & 17 & 1 & 1 & - \\
\midrule
- & 65.4(2.4) & 74.9(0.8) & 61.6(1.2) & 58.0(2.4) & 58.8(5.5) & 71.0(2.5) & 75.3(0.5) & 72.6(4.9) & 67.21 \\
\midrule
InfoNCE only & 68.9(1.2) & 74.2(0.3) & 62.8(0.2) & 59.7(0.7) & 57.8(11.5) & 73.6(1.8) & 76.1(0.6) & 77.6(0.3) & 68.85 \\
EBM-NCE only & 68.0(0.3) & 74.3(0.4) & 62.6(0.3) & 61.3(0.4) & 66.0(6.0) & 73.1(1.6) & 76.4(1.0) & 79.6(1.7) & 70.15 \\
VAE only & 67.6(1.8) & 73.2(0.5) & 61.9(0.4) & 60.5(0.2) & 59.7(1.6) & 78.6(0.7) & 77.4(0.6) & 75.4(2.1) & 69.29 \\
AE only & 70.5(0.4) & 75.0(0.4) & 62.4(0.4) & 61.0(1.4) & 53.8(1.0) & 74.1(2.9) & 76.3(0.5) & 77.9(0.9) & 68.89 \\
\midrule
InfoNCE + VAE & 69.6(1.1) & 75.4(0.6) & 63.2(0.3) & 59.9(0.4) & 69.3(14.0) & 76.5(1.3) & 76.3(0.2) & 75.2(2.7) & 70.67 \\
EBM-NCE + VAE & 68.5(0.2) & 74.5(0.4) & 62.7(0.1) & 62.3(1.6) & 79.0(2.5) & 75.0(1.4) & 74.8(1.4) & 76.8(1.1) & 71.69 \\
InfoNCE + AE & 65.1(3.1) & 75.4(0.7) & 62.5(0.5) & 59.2(0.6) & 77.2(1.8) & 72.4(1.4) & 75.8(0.6) & 77.1(0.8) & 70.60 \\
EBM-NCE + AE & 69.4(1.0) & 75.2(0.1) & 62.4(0.4) & 61.5(0.9) & 71.1(6.0) & 73.3(0.3) & 75.2(0.6) & 79.3(1.1) & 70.94 \\
\bottomrule
\end{tabular}
\end{table}
\clearpage
\subsection{Broader Range of Downstream Tasks: Molecular Property Prediction Prediction} \label{sec:app:regression_results}
\begin{table}[H]
\small
\setlength{\tabcolsep}{10pt}
\centering
\caption{
Results for four molecular property prediction tasks (regression).
For each downstream task, we report the mean (and standard variance) RMSE of 3 seeds with scaffold splitting.
For GraphMVP\,, we set $M=0.15$ and $C=5$.
The best performance for each task is marked in \textbf{bold}.
}
\begin{tabular}{l c c c c c}
\toprule
& ESOL & Lipo & Malaria & CEP & Avg\\
\midrule
-- & 1.178 (0.044) & 0.744 (0.007) & 1.127 (0.003) & 1.254 (0.030) & 1.07559 \\
\midrule
AM & 1.112 (0.048) & 0.730 (0.004) & 1.119 (0.014) & 1.256 (0.000) & 1.05419 \\
CP & 1.196 (0.037) & 0.702 (0.020) & 1.101 (0.015) & 1.243 (0.025) & 1.06059 \\
JOAO & 1.120 (0.019) & 0.708 (0.007) & 1.145 (0.010) & 1.293 (0.003) & 1.06631 \\
\midrule
GraphMVP\, & 1.091 (0.021) & 0.718 (0.016) & 1.114 (0.013) & 1.236 (0.023) & 1.03968 \\
GraphMVP-G & 1.064 (0.045) & 0.691 (0.013) & 1.106 (0.013) & \textbf{1.228 (0.001)} & 1.02214 \\
GraphMVP-C & \textbf{1.029 (0.033)} & \textbf{0.681} (0.010) & \textbf{1.097 (0.017)} & 1.244 (0.009) & \textbf{1.01283} \\
\bottomrule
\end{tabular}
\end{table}
\subsection{Broader Range of Downstream Tasks: Drug-Target Affinity Prediction}
\begin{table}[H]
\small
\setlength{\tabcolsep}{15pt}
\centering
\caption{
Results for two drug-target affinity prediction tasks (regression).
For each downstream task, we report the mean (and standard variance) MSE of 3 seeds with random splitting.
For GraphMVP\,, we set $M=0.15$ and $C=5$.
The best performance for each task is marked in \textbf{bold}.
}
\centering
\begin{tabular}{l c c c}
\toprule
& Davis & KIBA & Avg\\
\midrule
& 0.286 (0.006) & 0.206 (0.004) & 0.24585 \\
\midrule
AM & 0.291 (0.007) & 0.203 (0.003) & 0.24730 \\
CP & 0.279 (0.002) & 0.198 (0.004) & 0.23823 \\
JOAO & 0.281 (0.004) & 0.196 (0.005) & 0.23871 \\
\midrule
GraphMVP & 0.280 (0.005) & 0.178 (0.005) & 0.22860 \\
GraphMVP-G & \textbf{0.274 (0.002)} & 0.175 (0.001) & 0.22476 \\
GraphMVP-C & 0.276 (0.004) & \textbf{0.168 (0.001)} & \textbf{0.22231} \\
\bottomrule
\end{tabular}
\end{table}
\subsection{Case Studies~\label{appendix:case}}
\textbf{Shape Analysis (3D Diameter Prediction).} Diameter is an important measure in molecule~\cite{liu2010using,melnik2003distance}, and genome~\cite{finn2017comparative} modelling. Usually, the longer the 2D diameter (longest adjacency path) is, the larger the 3D diameter (largest atomic pairwise l2 distance). However, this is not always true. Therefore, we are particularly interested in using the 2D graph to predict the 3D diameter when the 2D and 3D molecular landscapes are with large differences (as in~\Cref{fig:case} and ~\Cref{fig:case_appendix}). We formulate it as a $n$-class recognition problem, where $n$ is the number of class after removing the consecutive intervals. We provide numerical results in~\Cref{table:diam} and more visualisation examples in~\Cref{fig:case_vis}.
\begin{figure}[ht]
\centering
\vspace{-2ex}
\includegraphics[width=\textwidth]{figures/Case_Study_Appendix.pdf}
\caption{Molecules selection, we select the molecules that lies in the black dash box.}
\label{fig:case_appendix}
\end{figure}
\begin{table}[htb!]
\caption{Accuracy on Recognizing Molecular Spatial Diameters~\label{table:diam}}
\small
\setlength{\tabcolsep}{3pt}
\centering
\begin{tabular}{l c c c c c c c c c c}
\toprule
Random & \,AttrMask\, & \,ContextPred\, & GPT-GNN & GraphCL & JOAOv2 & MVP & MVP-G & MVP-C \\
\midrule
38.9 (0.8) & 37.6 (0.6) & 41.2 (0.7) & 39.2 (1.1) & 38.7 (2.0) & 41.3 (1.2) & 42.3 (1.9) & 41.9 (0.7) & 42.3 (1.3)\\
\bottomrule
\end{tabular}
\end{table}
\textbf{Long-Range Donor-Acceptor Detection.} Donor-Acceptor structures such as hydrogen bonds have key impacts on the molecular geometrical structures (collinear and coplanarity), and physical properties (melting point, water affinity, viscosity etc.). Usually, atom pairs such as ``O...H'' that are closed in the Euclidean space are considered as the donor-acceptor structures~\cite{kaur2020understanding}.
On this basis, we are particularly interested in using the 2D graph to recognize (i.e., binary classification) donor-acceptor structures which have larger ranges in the 2D adjacency (as shown in~\Cref{fig:case}). Similarly, we select the molecules whose donor-acceptor are close in 3D Euclidean distance but far in the 2D adjacency. We provide numerical results in~\Cref{table:donor}. Both tables show that MVP is the MVP :)
\begin{table}[htb!]
\caption{Accuracy on Recognizing Long-Range Donor-Acceptor Structures~\label{table:donor}}
\small
\setlength{\tabcolsep}{3pt}
\centering
\begin{tabular}{l c c c c c c c c c c}
\toprule
Random & \,AttrMask\, & \,ContextPred\, & GPT-GNN & GraphCL & JOAOv2 & MVP & MVP-G & MVP-C \\
\midrule
77.9 (1.1) & 78.6 (0.3) & 80.0 (0.5) & 77.5 (0.9) & 79.9 (0.7) & 79.2 (1.0) & 80.0 (0.4) & 81.5 (0.4) & 80.7 (0.2) \\
\bottomrule
\end{tabular}
\end{table}
\paragraph{Chirality.} We have also explored other tasks such as predicting the molecular chirality, it is a challenging setting if only 2D molecular graphs are provided~\cite{pattanaik2020message}. We found that GraphMVP\,brings negligible improvements due to the model capacity of SchNet. We save this in the ongoing work.
\begin{figure}[ht]
\centering
\vspace{-2ex}
\includegraphics[width=\textwidth]{figures/case_study_SSL_better.png}
\vspace{-2ex}
\caption{Molecule examples where GraphMVP successfully recognizes the 3D diameters while random initialisation fails, legends are in a format of ``molecule id''-``2d diameter''-``3d diameter''.}
\label{fig:case_vis}
\end{figure}
\section*{Reproducibility Statement}
To ensure the reproducibility of the empirical results, we include our code base in the supplementary material, which contains: instructions for installing conda virtual environment, data preprocessing scripts, and training scripts. Our code is available on \href{https://github.com/chao1224/GraphMVP}{GitHub} for reproducibility. Complete derivations of equations and clear explanations are given in~\Cref{sec:app:MI,sec:app:contrastive_ssl,sec:app:generative_ssl}.
\section*{Acknowledgement}
This project is supported by the Natural Sciences and Engineering Research Council (NSERC) Discovery Grant, the Canada CIFAR AI Chair Program, collaboration grants between Microsoft Research and Mila, Samsung Electronics Co., Ltd., Amazon Faculty Research Award, Tencent AI Lab Rhino-Bird Gift Fund and a NRC Collaborative R\&D Project (AI4D-CORE-06). This project was also partially funded by IVADO Fundamental Research Project grant PRF-2019-3583139727.
\bibliographystyle{plain}
{\small
\section{Dataset Overview} \label{app:sec:dataset}
\subsection{Pre-Training Dataset Overview}
In this section, we provide the basic statistics of the pre-training dataset (GEOM).
In~\Cref{fig:hist}, we plot the histogram (logarithm scale on the y-axis) and cumulative distribution on the number of conformers of each molecule. As shown by the histogram and curves, there are certain number of molecules having over 1000 possible 3d conformer structures, while over 80\% of molecules have less than 100 conformers.
\begin{figure}[htb!]
\centering
\includegraphics[width=\textwidth]{figures/hist.pdf}
\caption{Statistics on the conformers of each molecule\label{fig:hist}}
\end{figure}
In~\Cref{fig:hist}, we plot the histogram of the summation of top (descending sorted by weights) \{1,5,10,20\} conformer weights. The physical meaning of the weight is the portion of each conformer occurred in nature. We observe that the top 5 or 10 conformers are sufficient as they have dominated nearly all the natural observations. Such long-tailed distribution is also in alignment with our findings in the ablation studies. We find that utilizing top five conformers in the GraphMVP\, has reached an idealised spot between effectiveness and efficiency.
\begin{figure}[htb!]
\centering
\includegraphics[width=\textwidth]{figures/weight_conf.pdf}
\caption{Sum of occurrence weights for the top major conformers\label{fig:cumulative}}
\end{figure}
\subsection{Downstream Dataset Overview} \label{app:sec:downstream_datasets}
In this section, we review the four main categories of datasets used for downstream tasks.
\paragraph{Molecular Property: Pharmacology}
The Blood-Brain Barrier Penetration (BBBP)~\cite{martins2012bayesian} dataset measures whether a molecule will penetrate the central nervous system.
All three datasets, Tox21~\cite{tox21}, ToxCast~\cite{wu2018moleculenet}, and ClinTox~\cite{gayvert2016data} are related to the toxicity of molecular compounds.
The Side Effect Resource (SIDER)~\cite{kuhn2015sider} dataset stores the adverse drug reactions on a marketed drug database.
\paragraph{Molecular Property: Physical Chemistry}
Dataset proposed in~\cite{delaney2004esol} measures aqueous solubility of the molecular compounds. Lipophilicity (Lipo) dataset is a subset of ChEMBL~\cite{gaulton2012chembl} measuring the molecule octanol/water distribution coefficient. CEP dataset is a subset of the Havard Clean Energy Project (CEP)~\cite{hachmann2011harvard}, which estimates the organic photovoltaic efficiency.
\paragraph{Molecular Property: Biophysics}
Maximum Unbiased Validation (MUV)~\cite{rohrer2009maximum} is another sub-database from PCBA, and is obtained by applying a refined nearest neighbor analysis. HIV is from the Drug Therapeutics Program (DTP) AIDS Antiviral Screen~\cite{zaharevitz2015aids}, and it aims at predicting inhibit HIV replication. BACE measures the binding results for a set of inhibitors of $\beta$-secretase 1 (BACE-1), and is gathered in MoleculeNet~\cite{wu2018moleculenet}. Malaria~\cite{gamo2010thousands} measures the drug efficacy against the parasite that causes malaria.
\paragraph{Drug-Target Affinity}
Davis~\citep{davis2011comprehensive} measures the binding affinities between kinase inhibitors and kinases, scored by the $K_d$ value (kinase dissociation constant).
KIBA~\citep{tang2014making} contains binding affinities for kinase inhibitors from different sources, including $K_i$, $K_d$ and $\text{IC}_{50}$. KIBA scores~\citep{ozturk2018deepdta} are constructured to optimize the consistency among these values.
\begin{table}[H]
\centering
\small
\caption{Summary for the molecule chemical datasets.}
\begin{tabular}{l l r r r r}
\toprule
Dataset & Task & \# Tasks & \# Molecules & \# Proteins & \# Molecule-Protein pairs\\
\midrule
BBBP & Classification & 1 & 2,039 & -& -\\
Tox21 & Classification & 12 & 7,831 & -& -\\
ToxCast & Classification & 617 & 8,576 & -& -\\
Sider & Classification & 27 & 1,427 & -& -\\
ClinTox & Classification & 2 & 1,478 & -& -\\
MUV & Classification & 17 & 93,087 & -& -\\
HIV & Classification & 1 & 41,127 & -& -\\
Bace & Classification & 1 & 1,513 & -& -\\
\midrule
Delaney & Regression & 1 & 1,128 & -& -\\
Lipo & Regression & 1 & 4,200 & -& -\\
Malaria & Regression & 1 & 9,999 & -& -\\
CEP & Regression & 1 & 29,978 & -& -\\
\midrule
Davis & Regression & 1 & 68 & 379 & 30,056 \\
KIBA & Regression & 1 & 2,068 & 229 & 118,254 \\
\bottomrule
\end{tabular}
\label{tab:mol_dataset_summary}
\end{table}
\section{Introduction} \label{sec:intro}
In recent years, drug discovery has drawn increasing interest in the machine learning community. Among many challenges therein, how to discriminatively represent a molecule with a vectorized embedding remains a fundamental yet open challenge. The underlying problem can be decomposed into two components: how to design a common latent space for molecule graphs ({\em{i.e.}}, designing a suitable encoder) and how to construct an objective function to supervise the training ({\em{i.e.}}, defining a learning target). Falling broadly into the second category, our paper studies self-supervised molecular representation learning by leveraging the consistency between 3D geometry and 2D topology.
Motivated by the prominent success of the pretraining-finetuning pipeline~\citep{devlin2018bert}, unsupervisedly pre-trained graph neural networks for molecules yields promising performance on downstream tasks and becomes increasingly popular~\citep{velivckovic2018deep,sun2019infograph,liu2018n,hu2019strategies,You2020GraphCL,you2021graph}. The key to pre-training lies in finding an effective proxy task ({\em{i.e.}}, training objective) to leverage the power of large unlabeled datasets. Inspired by \citep{schutt2017schnet,liu2018n,liu2021spherical} that molecular properties~\citep{gilmer2017neural,liu2018n} can be better predicted by 3D geometry due to its encoded energy knowledge, we aim to make use of the 3D geometry of molecules in pre-training. However, the stereochemical structures are often very expensive to obtain, making such 3D geometric information scarce in downstream tasks. To address this problem, we propose the Graph Multi-View Pre-training (GraphMVP) framework, where a 2D molecule encoder is pre-trained with the knowledge of 3D geometry and then fine-tuned on downstream tasks without 3D information. Our learning paradigm, during pre-training, injects the knowledge of 3D molecular geometry to a 2D molecular graph encoder such that the downstream tasks can benefit from the implicit 3D geometric prior even if there is no 3D information available.
We attain the aforementioned goal by leveraging two pretext tasks on the 2D and 3D molecular graphs: one contrastive and one generative SSL. Contrastive SSL creates the supervised signal at an \textbf{inter-molecule} level: the 2D and 3D graph pairs are positive if they are from the same molecule, and negative otherwise; Then contrastive SSL~\citep{wang2020understanding} will align the positive pairs and contrast the negative pairs simultaneously. Generative SSL~\citep{vincent2008extracting,kingma2013auto,higgins2016beta}, on the other hand, obtains the supervised signal in an \textbf{intra-molecule} way: it learns a 2D/3D representation that can reconstruct its 3D/2D counterpart view for each molecule itself.
To cope with the challenge of measuring the quality of reconstruction on molecule 2D and 3D space, we further propose a novel surrogate objective function called variation representation reconstruction (VRR) for the generative SSL task, which can effectively measure such quality in the continuous representation space. The knowledge acquired by these two SSL tasks is complementary, so our GraphMVP\ framework integrates them to form a more discriminative 2D molecular graph representation.
Consistent and significant performance improvements empirically validate the effectiveness of GraphMVP.
We give additional insights to justify the effectiveness of GraphMVP. First, GraphMVP\, is a self-supervised learning approach based on maximizing mutual information (MI) between 2D and 3D views, enabling the learnt representation to capture high-level factors~\citep{belghazi2018mutual,tschannen2019mutual,bachman2019learning} in molecule data. Second, we find that 3D molecular geometry is a form of privileged information~\cite{vapnik2009new,vapnik2015learning}. It has been proven that using privileged information in training can accelerate the speed of learning. We note that privileged information is only used in training, while it is not available in testing. This perfectly matches our intuition of pre-training molecular representation with 3D geometry.
Our contributions include
(1) To our best knowledge, we are the first to incorporate the 3D geometric information into graph SSL;
(2) We propose one contrastive and one generative SSL tasks for pre-training. Then we elaborate their difference and empirically validate that combining both can lead to a better representation;
(3) We provide theoretical insights and case studies to justify why adding 3D geometry is beneficial;
(4) We achieve the SOTA performance among all the SSL methods.
\textbf{Related work.}
We briefly review the most related works here and include a more detailed summarization in~\Cref{appendix:related_work}. Self-supervised learning (SSL) methods have attracted massive attention to graph applications~\citep{liu2021graph,xie2021self,wu2021self,liu2021self}. In general, there are roughly two categories of graph SSL: contrastive and generative, where they differ on the design of the supervised signals. Contrastive graph SSL~\citep{velivckovic2018deep,sun2019infograph,hu2019strategies,You2020GraphCL,you2021graph} constructs the supervised signals at the \textbf{inter-graph} level and learns the representation by contrasting with other graphs, while generative graph SSL~\citep{hamilton2017inductive,liu2018n,hu2019strategies,hu2020gpt} focuses on reconstructing the original graph at the \textbf{intra-graph} level. One of the most significant differences that separate our work from existing methods is that all previous methods \textbf{merely} focus on 2D molecular topology. However, for scientific tasks such as molecular property prediction, 3D geometry should be incorporated as it provides complementary and comprehensive information~\citep{schutt2017schnet,liu2021spherical}. To fill this gap, we propose GraphMVP to leverage the 3D geometry in graph self-supervised pre-training.
\section{Theoretical Insights} \label{sec:theoretical_insight}
In this section, we provide the mathematical insights behind GraphMVP. We will first discuss both contrastive and generative SSL methods (\Cref{sec:contrastive_SSL,sec:generative_SSL}) are maximizing the mutual information (MI) and then how the 3D geometry, as privileged information, can help 2D representation learning.
\vspace{-1ex}
\subsection{Maximizing Mutual Information} \label{sec:mutual_information}
\vspace{-1ex}
Mutual information (MI) measures the non-linear dependence~\citep{belghazi2018mutual} between two random variables: the larger MI, the stronger dependence between the variables. Therefore for GraphMVP, we can interpret it as maximizing MI between 2D and 3D views: to obtain a more robust 2D/3D representation by sharing more information with its 3D/2D counterparts. This is also consistent with the sample complexity theory~\citep{erhan2010does,arora2019theoretical,garg2020functional} where SSL as functional regularizer can reduce the uncertainty in representation learning. We first derive a lower bound for MI (see derivations in~\Cref{sec:app:MI}), and the corresponding objective function $\mathcal{L}_{\text{MI}}$ is
\begin{equation} \label{eq:MI_objective}
\small
I(X;Y) \ge \mathcal{L}_{\text{MI}} = \frac{1}{2} \mathbb{E}_{p({\bm{x}},{\bm{y}})} \big[ \log p({\bm{y}}|{\bm{x}}) + \log p({\bm{x}}|{\bm{y}}) \big].
\end{equation}
\textbf{Contrastive Self-Supervised Learning.}
InfoNCE was initialized proposed to maximize the MI directly~\citep{oord2018representation}.
Here in GraphMVP, EBM-NCE estimates the conditional likelihood in~\Cref{eq:MI_objective} using EBM, and solves it with NCE~\cite{gutmann2010noise}. As a result, EBM-NCE can also be seen as maximizing MI between 2D and 3D views. The detailed derivations can be found in~\Cref{app:ebm_nce}.
\textbf{Generative Self-Supervised Learning.}
One alternative solution is to use a variational lower bound to approximate the conditional log-likelihood terms in~\Cref{eq:MI_objective}. Then we can follow the same pipeline in~\Cref{sec:generative_SSL}, ending up with the surrogate objective, {\em{i.e.}}, VRR in~\Cref{eq:variational_lower_bound_approximation}.
\subsection{3D Geometry as Privileged Information} \label{sec:privileged_info}
We show the theoretical insights from privileged information that motivate GraphMVP. We start by considering a supervised learning setting where $(\bm{u}_i,\bm{l}_i)$ is a feature-label pair and $\bm{u}_i^*$ is the privileged information~\cite{vapnik2009new,vapnik2015learning}. The privileged information is defined to be additional information about the input $(\bm{u}_i,\bm{l}_i)$ in order to support the prediction. For example, $\bm{u}_i$ could be some CT images of a particular disease, $\bm{l}_i$ could be the label of the disease and $\bm{u}_i^*$ is the medical report from a doctor. VC theory~\cite{vapnik2013nature,vapnik2015learning} characterizes the learning speed of an algorithm from the capacity of the algorithm and the amount of training data. Considering a binary classifier $f$ from a function class $\mathcal{F}$ with finite VC-dimension $\text{VCD}(\mathcal{F})$. With probability $1-\delta$, the expected error is upper bounded by
\begin{equation}
\small
R(f)\leq R_n(f) +\mathcal{O}\bigg( \big( \frac{\text{VCD}(\mathcal{F})-\log\delta}{n} \big)^\beta \bigg)
\end{equation}
where $R_n(f)$ denotes the training error and $n$ is the number of training samples. When the training data is separable, then $R_n(f)$ will diminish to zero and $\beta$ is equal to $1$. When the training data is non-separable, $\beta$ is $\frac{1}{2}$. Therefore, the rate of convergence for the separable case is of order $1/n$. In contrast, the rate for the non-separable case is of order $1/\sqrt{n}$. We note that such a difference is huge, since the same order of bounds require up to 100 training samples versus 10,000 samples. Privileged information makes the training data separable such that the learning can be more efficient. Connecting the results to GraphMVP, we notice that the 3D geometric information of molecules can be viewed as a form of privileged information, since 3D information can effectively make molecules more separable for some properties~\cite{schutt2017schnet,liu2018n,liu2021spherical}. Besides, privileged information is only used in training, and it well matches our usage of 3D geometry for pre-training. In fact, using 3D structures as privileged information has been already shown quite useful in protein classification~\cite{vapnik2009new}, which serves as a strong evidence to justify the effectiveness of 3D information in graph SSL pre-training.
\section{Experiments and Results} \label{sec:experiments}
\subsection{Experimental Settings} \label{sec:experiments_setting}
\textbf{Datasets.} We pre-train models on the same dataset then fine-tune on the wide range of downstream tasks. We randomly select 50k qualified molecules from GEOM~\citep{axelrod2020geom} with both 2D and 3D structures for the pre-training. As clarified in~\Cref{sec:overview}, conformer ensembles can better reflect the molecular property, thus we take $C$ conformers of each molecule. For downstream tasks, we first stick to the same setting of the main graph SSL work~\citep{hu2019strategies,You2020GraphCL,you2021graph}, exploring 8 binary molecular property prediction tasks, which are all in the low-data regime. Then we explore 6 regression tasks from various low-data domains to be more comprehensive. We describe all the datasets in~\Cref{app:sec:dataset}.
\textbf{2D GNN.} We follow the research line of SSL on molecule graph~\citep{hu2019strategies,you2021graph,You2020GraphCL}, using the same Graph Isomorphism Network (GIN)~\citep{xu2018powerful} as the backbone model, with the same feature sets.
\textbf{3D GNN.} We choose SchNet~\citep{schutt2017schnet} for geometric modeling, since SchNet: (1) is found to be a strong geometric representation learning method under the fair benchmarking; (2) can be trained more efficiently, comparing to the other recent 3D models. More detailed explanations are in~\Cref{appendix:3d_gnn}.
\subsection{Main Results on Molecular Property Prediction.} \label{sec:main_results}
We carry out comprehensive comparisons with 10 SSL baselines and random initialization. For pre-training, we apply all SSL methods on the same dataset based on GEOM~\citep{axelrod2020geom}. For fine-tuning, we follow the same setting~\citep{hu2019strategies,You2020GraphCL,you2021graph} with 8 low-data molecular property prediction tasks.
\textbf{Baselines.} Due to the rapid growth of graph SSL~\citep{liu2021graph,xie2021self,wu2021self}, we are only able to benchmark the most well-acknowledged baselines: EdgePred~\citep{hamilton2017inductive}, InfoGraph~\citep{sun2019infograph}, GPT-GNN\citep{hu2020gpt}, AttrMask \& ContextPred\citep{hu2019strategies}, GraphLoG\citep{xu2021self}, G-\{Contextual, Motif\}\citep{rong2020self}, GraphCL\citep{You2020GraphCL}, JOAO\citep{you2021graph}.
\textbf{Our method.} GraphMVP\, has two key factors: i) masking ratio ($M$) and ii) number of conformers for each molecule ($C$). We set $M=0.15$ and $C=5$ by default, and will explore their effects in the following ablation studies in~\Cref{sec:effect_of_M_and_C}. For EBM-NCE loss, we adopt the empirical distribution for noise distribution. For~\Cref{eq:graphmvp_variants}, we pick the empirically optimal generative and contrastive 2D SSL method: that is AttrMask for GraphMVP-G\, and ContextPred for GraphMVP-C.
\begin{table}[t!]
\caption{
Results for molecular property prediction tasks.
For each downstream task, we report the mean (and standard deviation) ROC-AUC of 3 seeds with scaffold splitting.
For GraphMVP\,, we set $M=0.15$ and $C=5$.
The best and second best results are marked \textbf{\underline{bold}} and \textbf{bold}, respectively.
\vspace{-0.24cm}
}
\label{tab:main_results}
\small
\setlength{\tabcolsep}{3pt}
\centering
\begin{tabular}{l c c c c c c c c c}
\toprule
Pre-training & BBBP & Tox21 & ToxCast & Sider & ClinTox & MUV & HIV & Bace & Avg\\
\midrule
-- & 65.4(2.4) & 74.9(0.8) & 61.6(1.2) & 58.0(2.4) & 58.8(5.5) & 71.0(2.5) & 75.3(0.5) & 72.6(4.9) & 67.21 \\
\midrule
EdgePred & 64.5(3.1) & 74.5(0.4) & 60.8(0.5) & 56.7(0.1) & 55.8(6.2) & 73.3(1.6) & 75.1(0.8) & 64.6(4.7) & 65.64 \\
AttrMask & 70.2(0.5) & 74.2(0.8) & 62.5(0.4) & 60.4(0.6) & 68.6(9.6) & 73.9(1.3) & 74.3(1.3) & 77.2(1.4) & 70.16 \\
GPT-GNN & 64.5(1.1) & \textbf{75.3(0.5)} & 62.2(0.1) & 57.5(4.2) & 57.8(3.1) & 76.1(2.3) & 75.1(0.2) & 77.6(0.5) & 68.27 \\
InfoGraph & 69.2(0.8) & 73.0(0.7) & 62.0(0.3) & 59.2(0.2) & 75.1(5.0) & 74.0(1.5) & 74.5(1.8) & 73.9(2.5) & 70.10 \\
ContextPred & 71.2(0.9) & 73.3(0.5) & 62.8(0.3) & 59.3(1.4) & 73.7(4.0) & 72.5(2.2) & 75.8(1.1) & 78.6(1.4) & 70.89 \\
GraphLoG & 67.8(1.7) & 73.0(0.3) & 62.2(0.4) & 57.4(2.3) & 62.0(1.8) & 73.1(1.7) & 73.4(0.6) & 78.8(0.7) & 68.47 \\
G-Contextual & 70.3(1.6) & 75.2(0.3) & 62.6(0.3) & 58.4(0.6) & 59.9(8.2) & 72.3(0.9) & 75.9(0.9) & 79.2(0.3) & 69.21 \\
G-Motif & 66.4(3.4) & 73.2(0.8) & 62.6(0.5) & 60.6(1.1) & 77.8(2.0) & 73.3(2.0) & 73.8(1.4) & 73.4(4.0) & 70.14 \\
GraphCL & 67.5(3.3) & 75.0(0.3) & 62.8(0.2) & 60.1(1.3) & 78.9(4.2) & \textbf{77.1(1.0)} & 75.0(0.4) & 68.7(7.8) & 70.64 \\
JOAO & 66.0(0.6) & 74.4(0.7) & 62.7(0.6) & 60.7(1.0) & 66.3(3.9) & 77.0(2.2) & \textbf{76.6(0.5)} & 72.9(2.0) & 69.57 \\
\midrule
GraphMVP\, & 68.5(0.2) & 74.5(0.4) & 62.7(0.1) & \textbf{62.3(1.6)} & \textbf{79.0(2.5)} & 75.0(1.4) & 74.8(1.4) & 76.8(1.1) & 71.69 \\
GraphMVP-G & \textbf{70.8(0.5)} & \textbf{\underline{75.9(0.5)}} & \textbf{\underline{63.1(0.2)}} & 60.2(1.1) & \textbf{\underline{79.1(2.8)}} & \textbf{\underline{77.7(0.6)}} & 76.0(0.1) & \textbf{79.3(1.5)} & \textbf{72.76} \\
GraphMVP-C & \textbf{\underline{72.4(1.6)}} & 74.4(0.2) & \textbf{\underline{63.1(0.4)}} & \textbf{\underline{63.9(1.2)}} & 77.5(4.2) & 75.0(1.0) & \textbf{\underline{77.0(1.2)}} & \textbf{\underline{81.2(0.9)}} & \textbf{\underline{73.07}} \\
\bottomrule
\end{tabular}
\vspace{-2ex}
\end{table}
The main results on 8 molecular property prediction tasks are listed in~\Cref{tab:main_results}. We observe that the performance of GraphMVP\, is significantly better than the random initialized one, and the average performance outperforms the existing SSL methods by a large margin. In addition, GraphMVP-G\, and GraphMVP-C\, consistently improve the performance, supporting the claim: \textbf{3D geometry is complementary to the 2D topology}. GraphMVP\, leverages the information between 3D geometry and 2D topology, and 2D SSL plays the role as regularizer to extract more 2D topological information; they are extracting different perspectives of information and are indeed complementary to each other.
\subsection{Ablation Study: The Effect of Masking Ratio and Number of Conformers} \label{sec:effect_of_M_and_C}
We analyze the effects of masking ratio $M$ and the number of conformers $C$ in GraphMVP. In~\Cref{tab:main_results}, we set the $M$ as 0.15 since it has been widely used in existing SSL methods~\citep{hu2019strategies,You2020GraphCL,you2021graph}, and $C$ is set to 5, which we will explain below. We explore on the range of $M \in \{0, 0.15, 0.3\}$ and $C \in \{1, 5, 10, 20\}$, and report the average performance. The complete results are in~\Cref{app:sec:effect_of_M_and_C}.
\begin{table}[t!]
\fontsize{8.5}{4}\selectfont
\centering
\vspace{-2ex}
\begin{minipage}[t]{.48\linewidth}
\setlength{\tabcolsep}{5.3pt}
\renewcommand\arraystretch{2.2}
\caption{Ablation of masking ratio $M$, $C\equiv5$.\vspace{-0.3cm}}
\label{tab:ablation_masking}
\centering
\begin{tabular}{c c c c}
\toprule
$M$ & GraphMVP & GraphMVP-G & GraphMVP-C\\
\midrule
0 & 71.12 & 72.15 & 72.66\\
0.15 & 71.60 & 72.76 & 73.08\\
0.30 & 71.79 & 72.91 & 73.17\\
\bottomrule
\end{tabular}
\end{minipage}
\hfill
\begin{minipage}[t]{.48\linewidth}
\setlength{\tabcolsep}{5.5pt}
\caption{Ablation of \# conformer $C$, $M\equiv0.15$.\vspace{-0.3cm}}
\label{tab:ablation_n_conformers}
\centering
\begin{tabular}{c c c c}
\toprule
$C$ & GraphMVP & GraphMVP-G & GraphMVP-C\\
\midrule
1 & 71.61 & 72.80 & 72.46\\
5 & 71.60 & 72.76 & 73.08\\
10 & 72.20 & 72.59 & 73.09\\
20 & 72.39 & 73.00 & 73.02\\
\bottomrule
\end{tabular}
\end{minipage}
\vspace{-1ex}
\end{table}
As seen in~\Cref{tab:ablation_masking}, the improvement is more obvious from $M=0$ (raw graph) to $M=0.15$ than from $M=0.15$ to $M=0.3$. This can be explained that subgraph masking with larger ratio will make the SSL tasks more challenging, especially comparing to the raw graph ($M=0$).
\Cref{tab:ablation_n_conformers} shows the effect for $C$. We observe that the performance is generally better when adding more conformers, but will reach a plateau above certain thresholds. This observation matches with previous findings~\citep{axelrod2020molecular}: adding more conformers to augment the representation learning is not as helpful as expected; while we conclude that adding more conformers can be beneficial with little improvement. One possible reason is, when generating the dataset, we are sampling top-$C$ conformers with highest possibility and lowest energy. In other words, top-5 conformers are sufficient to cover the most conformers with equilibrium state (over 80\%), and the effect of larger $C$ is thus modest.
To sum up, adding more conformers might be helpful, but the computation cost can grow linearly with the increase in dataset size. On the other hand, enlarging the masking ratio will not induce extra cost, yet the performance is slightly better. Therefore, we would encourage tuning masking ratios prior to trying a larger number of conformers from the perspective of efficiency and effectiveness.
\subsection{Ablation Study: The Effect of Objective Function} \label{sec:experiment_each_loss_component}
\begin{wraptable}[13]{r}{0.5\textwidth}
\setlength{\tabcolsep}{3.53pt}
\vspace{-0.4cm}
\small
\caption{
Ablation on the objective function.
\vspace{-0.35cm}
}
\label{tab:ablation_objective_function}
\small
\centering
\begin{tabular}{l c c c}
\toprule
GraphMVP\, Loss & Contrastive & Generative & Avg\\
\midrule
Random & & & 67.21\\
\midrule
InfoNCE only & \checkmark & & 68.85\\
EBM-NCE only & \checkmark & & 70.15\\
VRR only & & \checkmark & 69.29\\
RR only & & \checkmark & 68.89\\
\midrule
InfoNCE + VRR & \checkmark & \checkmark & 70.67\\
EBM-NCE + VRR & \checkmark & \checkmark & 71.69\\
InfoNCE + RR & \checkmark & \checkmark & 70.60\\
EBM-NCE + RR & \checkmark & \checkmark & 70.94\\
\bottomrule
\end{tabular}
\end{wraptable}
In~\Cref{sec:method}, we introduce a new contrastive learning objective family called EBM-NCE, and we take either InfoNCE and EBM-NCE as the contrastive SSL. For the generative SSL task, we propose a novel objective function called variational representation reconstruction (VRR) in~\Cref{eq:variational_lower_bound_approximation}. As discussed in~\Cref{sec:generative_SSL}, stochasticity is important for GraphMVP\, since it can capture the conformer distribution for each 2D molecular graph. To verify this, we add an ablation study on \textit{representation reconstruction (RR)} by removing stochasticity in VRR. Thus, here we deploy a comprehensive ablation study to explore the effect for each individual objective function (InfoNCE, EBM-NCE, VRR and RR), followed by the pairwise combinations between them.
The results in \Cref{tab:ablation_objective_function} give certain constructive insights as follows:
(1) Each individual SSL objective function (middle block) can lead to better performance. This strengthens the claim that adding 3D information is helpful for 2D representation learning.
(2) According to the combination of those SSL objective functions (bottom block), adding both contrastive and generative SSL can consistently improve the performance. This verifies our claim that conducting SSL at both the inter-data and intra-data level is beneficial.
(3) We can see VRR is consistently better than RR on all settings, which verifies that stochasticity is an important factor in modeling 3D conformers for molecules.
\subsection{Broader Range of Downstream Tasks}
The 8 binary downstream tasks discussed so far have been widely applied in the graph SSL research line on molecules~\cite{hu2019strategies,You2020GraphCL,you2021graph}, but there are more tasks where the 3D conformers can be helpful. Here we test 4 extra regression property prediction tasks and 2 drug-target affinity tasks.
About the dataset statistics, more detailed information can be found in~\Cref{app:sec:dataset}, and we may as well briefly describe the affinity task here. Drug-target affinity (DTA) is a crucial task~\citep{pahikkala2015toward,wen2017deep,ozturk2018deepdta} in drug discovery, where it models both the molecular drugs and target proteins, with the goal to predict their affinity scores. One recent work~\citep{nguyen2019graphdta} is modeling the molecular drugs with 2D GNN and target protein (as an amino-acid sequence) with convolution neural network (CNN). We adopt this setting by pre-training the 2D GNN using GraphMVP.
As illustrated in~\Cref{tab:main_regression}, the consistent performance gain verifies the effectiveness of our proposed GraphMVP.
\begin{table}[t!]
\setlength{\tabcolsep}{7pt}
\centering
\caption{
Results for four molecular property prediction tasks (regression) and two DTA tasks (regression).
We report the mean RMSE of 3 seeds with scaffold splitting for molecular property downstream tasks, and mean MSE for 3 seeds with random splitting on DTA tasks.
For GraphMVP\,, we set $M=0.15$ and $C=5$.
The best performance for each task is marked in \textbf{\underline{bold}}.
We omit the std here since they are very small and indistinguishable. For complete results, please check~\Cref{sec:app:regression_results}.
\vspace{-4ex}
}
\label{tab:main_regression}
\begin{tabular}{l c c c c c c c c}
\toprule
& \multicolumn{5}{c}{Molecular Property Prediction} & \multicolumn{3}{c}{Drug-Target Affinity}\\
\cmidrule(lr){2-6} \cmidrule(lr){7-9} Pre-training & ESOL & Lipo & Malaria & CEP & Avg & Davis & KIBA & Avg\\
\midrule
--
& 1.178 & 0.744 & 1.127 & 1.254 & 1.0756
& 0.286 & 0.206 & 0.2459\\
\midrule
AM
& 1.112 & 0.730 & 1.119 & 1.256 & 1.0542
& 0.291 & 0.203 & 0.2476\\
CP
& 1.196 & 0.702 & 1.101 & 1.243 & 1.0606
& 0.279 & 0.198 & 0.2382 \\
JOAO
& 1.120 & 0.708 & 1.145 & 1.293 & 1.0663
& 0.281 & 0.196 & 0.2387\\
\midrule
GraphMVP
& 1.091 & 0.718 & 1.114 & 1.236 & 1.0397
& 0.280 & 0.178 & 0.2286\\
GraphMVP-G\,\,\,
& 1.064 & 0.691 & 1.106 & \textbf{\underline{1.228}} & 1.0221
& \textbf{\underline{0.274}} & 0.175 & 0.2248\\
GraphMVP-C\,\,\,
& \textbf{\underline{1.029}} & \textbf{\underline{0.681}} & \textbf{\underline{1.097}} & 1.244 & \textbf{\underline{1.0128}}
& 0.276 & \textbf{\underline{0.168}} & \textbf{\underline{0.2223}}\\
\bottomrule
\end{tabular}
\vspace{-1ex}
\end{table}
\subsection{Case Study}
We investigate how GraphMVP\, helps when the task objectives are challenging with respect to the 2D topology but straightforward using 3D geometry (as shown in~\Cref{fig:case}). We therefore design two case studies to testify how GraphMVP\, transfers knowledge from 3D geometry into the 2D representation.
The first case study is \textit{3D Diameter Prediction}. For molecules, usually, the longer the 2D diameter is, the larger the 3D diameter (largest atomic pairwise l2 distance). However, this does not always hold, and we are interested in using the 2D graph to predict the 3D diameter. The second case study is \textit{Long-Range Donor-Acceptor Detection}. Molecules possess a special geometric structure called donor-acceptor bond, and we want to use 2D molecular graph to detect this special structure. We validate that GraphMVP\, consistently brings improvements on these 2 case studies, and provide more detailed discussions and interpretations in~\Cref{appendix:case}.
\begin{figure}[htbp!]
\centering
\vspace{-2ex}
\includegraphics[width=\textwidth]{figures/Case_Study_2.pdf}
\vspace{-5ex}
\caption{We select the molecules whose properties can be easily resolved via 3D but not 2D. The randomly initialised 2D GNN achieves accuracy of $38.9\pm0.8$ and $77.9\pm1.1$, respectively. The GraphMVP\, pre-trained ones obtain scores of $42.3\pm1.3$ and $81.5\pm0.4$, outperforming all the precedents in~\Cref{sec:main_results}. We plot cases where random initialization fails but GraphMVP\, is correct.}
\label{fig:case}
\vspace{-4ex}
\end{figure}
\section{Contrastive Self-Supervised Learning} \label{sec:app:contrastive_ssl}
The essence of contrastive self-supervised learning is to align positive view pairs and contrast negative view pairs, such that the obtained representation space is well distributed~\citep{wang2020understanding}. We display the pipeline in~\Cref{fig:app:contrastive_ssl}. Along the research line in graph SSL~\citep{liu2021graph,xie2021self,wu2021self,liu2021self}, InfoNCE and EBM-NCE are the two most-widely used, as discussed below.
\begin{figure}[htb!]
\centering
\includegraphics[width=\textwidth]{figures/Diagram_Contrastive_final.pdf}
\vspace{-4ex}
\caption{
Contrastive SSL in GraphMVP. The black dashed circles represent subgraph masking.
}
\label{fig:app:contrastive_ssl}
\vspace{-2ex}
\end{figure}
\subsection{InfoNCE} \label{sec:app:infonce}
InfoNCE~\citep{oord2018representation} is first proposed to approximate MI~\Cref{eq:app:MI}:
\begin{equation}\label{eq:app:InfoNCE}
\footnotesize{
\begin{aligned}
\mathcal{L}_{\text{InfoNCE}} =
-\frac{1}{2} \mathbb{E} &\left[
\log \frac{\exp(f_{{\bm{x}}}({\bm{x}}, {\bm{y}}))}{\exp(f_{{\bm{x}}}({\bm{x}}, {\bm{y}})) + \sum_j \exp(f_{{\bm{x}}}({\bm{x}}^{j},{\bm{y}})}) + \log \frac{\exp(f_{{\bm{y}}}({\bm{y}},{\bm{x}}))}{\exp(f_{{\bm{y}}}({\bm{y}},{\bm{x}})) + \sum_j \exp{f_{{\bm{y}}}({\bm{y}}^{j},{\bm{x}})}} \right],
\end{aligned}
}
\end{equation}
where ${\bm{x}}^{j}, {\bm{y}}^{j}$ are randomly sampled 2D and 3D views regarding to the anchored pair $({\bm{x}},{\bm{y}})$. $f_{{\bm{x}}}({\bm{x}},{\bm{y}}), f_{{\bm{y}}}({\bm{y}},{\bm{x}})$ are scoring functions for the two corresponding views, whose formulation can be quite flexible. Here we use $f_{\bm{x}}({\bm{x}},{\bm{y}}) = f_{\bm{y}}({\bm{y}},{\bm{x}}) = \exp(\langle h_{\bm{x}}, h_{\bm{y}} \rangle)$.
\paragraph{Derivation of InfoNCE}
\begin{equation}
\begin{aligned}
I(X;Y) - \log (K)
& = \mathbb{E}_{p({\bm{x}}, {\bm{y}})} \big[\log \frac{1}{K} \frac{p({\bm{x}}, {\bm{y}})}{p({\bm{x}}) p({\bm{y}})} \big]\\
& = \sum_{{\bm{x}}^{i},{\bm{y}}^{i}} \big[\log \frac{1}{K} \frac{p({\bm{x}}^{i},{\bm{y}}^{i})}{p({\bm{x}}^{i}) p({\bm{y}}^{i})} \big]\\
& \ge -\sum_{{\bm{x}}^{i},{\bm{y}}^{i}} \big[\log \big( 1 + (K-1) \frac{p({\bm{x}}^{i}) p({\bm{y}}^{i})}{p({\bm{x}}^{i},{\bm{y}}^{i})} \big)\big]\\
& = -\sum_{{\bm{x}}^{i},{\bm{y}}^{i}} \big[\log \frac{\frac{p({\bm{x}}^{i},{\bm{y}}^{i})}{p({\bm{x}}^{i}) p({\bm{y}}^{i})} + (K-1)}{\frac{p({\bm{x}}^{i},{\bm{y}}^{i})}{p({\bm{x}}^{i}) p({\bm{y}}^{i})}} \big]\\
& \approx -\sum_{{\bm{x}}^{i},{\bm{y}}^{i}} \big[ \log \frac{\frac{p({\bm{x}}^{i},{\bm{y}}^{i})}{p({\bm{x}}^{i}) p({\bm{y}}^{i})} + (K-1)\mathbb{E}_{{\bm{x}}^{j} \ne {\bm{x}}^{i}}\frac{p(x^{j},{\bm{y}}^{i})}{p({\bm{x}}^{j}) p({\bm{y}}^{i})} }{\frac{p({\bm{x}}^{i},{\bm{y}}^{i})}{p({\bm{x}}^{i}) p({\bm{y}}^{i})}} \big] \quad \text{// \textcircled{1}}\\
& = \sum_{{\bm{x}}^{i},{\bm{y}}^{i}} \big[ \log \frac{\exp(f_{\bm{x}}({\bm{x}}^{i},{\bm{y}}^{i}))}{\exp(f_{\bm{x}}({\bm{x}}^{i},{\bm{y}}^{i})) + \sum_{j=1}^K f_{\bm{x}}({\bm{x}}^{j},{\bm{y}}^{i})} \big],
\end{aligned}
\end{equation}
where we set $ f_{\bm{x}}({\bm{x}}^{i},{\bm{y}}^{i}) = \log \frac{p({\bm{x}}^{i},{\bm{y}}^{i})}{p({\bm{x}}^{i})p({\bm{y}}^{i})}$.
Notice that in \textcircled{1}, we are using data $x \in X$ as the anchor points. If we use the $y \in Y$ as the anchor points and follow the similar steps, we can obtain
\begin{equation}
\begin{aligned}
I(X;Y) - \log(K) \ge \sum_{{\bm{y}}^{i},{\bm{x}}^{i}} \big[ \log \frac{\exp(f_{\bm{y}}({\bm{y}}^{i},{\bm{x}}^{i}))}{\exp f_{\bm{y}}({\bm{y}}^{i},{\bm{x}}^{i}) + \sum_{j=1}^K \exp (f_{\bm{y}}({\bm{y}}^{j},{\bm{x}}^{i}))} \big].
\end{aligned}
\end{equation}
Thus, by add both together, we can have the objective function as~\Cref{eq:app:InfoNCE}.
\subsection{EBM-NCE} \label{app:ebm_nce}
We here provide an alternative approach to maximizing MI using energy-based model (EBM). To our best knowledge, we are the \textbf{first} to give the rigorous proof of using EBM to maximize the MI.
\subsubsection{Energy-Based Model (EBM)}
Energy-based model (EBM) is a powerful tool for modeling the data distribution. The classic formulation is:
\begin{equation} \label{eq:app:EBM_original}
p({\bm{x}}) = \frac{\exp(-E({\bm{x}}))}{A},
\end{equation}
where the bottleneck is the intractable partition function $A = \int_{\bm{x}} \exp(-E({\bm{x}})) d{\bm{x}}$. Recently, there have been quite a lot progress along this direction~\citep{gutmann2010noise,du2020improved,song2020score,song2021train}. Noise Contrastive Estimation (NCE)~\citep{gutmann2010noise} is one of the powerful tools here, as we will introduce later.
\subsubsection{EBM for MI}
Recall that our objective function is~\Cref{eq:app:MI_objective}: $\mathcal{L}_{\text{MI}} = \frac{1}{2} [H(Y|X) + H(X|Y)]$. Then we model the conditional likelihood with energy-based model (EBM). This gives us
\begin{equation} \label{eq:app:EBM}
\begin{aligned}
\mathcal{L}_{\text{EBM}} = -\frac{1}{2} \mathbb{E}_{p({\bm{x}},{\bm{y}})} \Big[ \log \frac{\exp(f_{\bm{x}}({\bm{x}}, {\bm{y}}))}{A_{{\bm{x}}|{\bm{y}}}} + \log \frac{\exp(f_{\bm{y}}({\bm{y}}, {\bm{x}}))}{A_{{\bm{y}}|{\bm{x}}}} \Big],
\end{aligned}
\end{equation}
where $f_{\bm{x}}({\bm{x}}, {\bm{y}}) = -E({\bm{x}}|{\bm{y}})$ and $f_{\bm{y}}({\bm{y}}, {\bm{x}}) = -E({\bm{y}}|{\bm{x}})$ are the negative energy functions, and $A_{{\bm{x}}|{\bm{y}}}$ and $A_{{\bm{y}}|{\bm{x}}}$ are the corresponding partition functions.
Under the EBM framework, if we solve~\Cref{eq:app:EBM} with Noise Contrastive Estimation (NCE)~\citep{gutmann2010noise}, the final EBM-NCE objective is
\begin{equation} \label{eq:app:EBM_NCE}
\begin{aligned}
\mathcal{L}_{\text{EBM-NCE}}
= & -\frac{1}{2} \mathbb{E}_{p_{\text{data}}(y)} \Big[ \mathbb{E}_{p_n({\bm{x}}|{\bm{y}})} [\log \big(1-\sigma(f_{\bm{x}}({\bm{x}}, {\bm{y}}))\big)] + \mathbb{E}_{p_{\text{data}}({\bm{x}}|{\bm{y}})} [\log \sigma(f_{\bm{x}}({\bm{x}}, {\bm{y}}))] \Big] \\
& ~~ - \frac{1}{2} \mathbb{E}_{p_{\text{data}}(x)} \Big[ \mathbb{E}_{p_n({\bm{y}}|{\bm{x}})} [\log \big(1-\sigma(f_{\bm{y}}({\bm{y}}, {\bm{x}}))\big)] + \mathbb{E}_{p_{\text{data}}({\bm{y}}|{\bm{x}})} [\log \sigma(f_{\bm{y}}({\bm{y}}, {\bm{x}}))] \Big].
\end{aligned}
\end{equation}
Next we will give the detailed derivations.
\subsubsection{Derivation of conditional EBM with NCE}
WLOG, let's consider the $p_\theta({\bm{x}}|{\bm{y}})$ first, and by EBM it is as follows:
\begin{equation} \label{eq:app:EBM_SSL}
p_\theta({\bm{x}}|{\bm{y}}) = \frac{\exp(-E({\bm{x}}|{\bm{y}}))}{ \int \exp(-E({\tilde {\bm{x}}}|{\bm{y}})) d{\tilde {\bm{x}}}} = \frac{\exp(f_{\bm{x}}({\bm{x}}, {\bm{y}}))}{\int \exp(f_{\bm{x}}({\tilde {\bm{x}}}|{\bm{y}})) d{\tilde {\bm{x}}}} = \frac{\exp(f_{\bm{x}}({\bm{x}}, {\bm{y}}))}{A_{{\bm{x}}|{\bm{y}}}}.
\end{equation}
Then we solve this using NCE. NCE handles the intractability issue by transforming it as a binary classification task. We take the partition function $A_{{\bm{x}}|{\bm{y}}}$ as a parameter, and introduce a noise distribution $p_n$. Based on this, we introduce a mixture model: ${\bm{z}}=0$ if the conditional ${\bm{x}}|{\bm{y}}$ is from $p_n({\bm{x}}|{\bm{y}})$, and ${\bm{z}}=1$ if ${\bm{x}}|{\bm{y}}$ is from $p_{\text{data}}({\bm{x}}|{\bm{y}})$. So the joint distribution is:
\begin{equation*}
p_{n,\text{\text{data}}}({\bm{x}}|{\bm{y}}) = p({\bm{z}}=1) p_{\text{data}}({\bm{x}}|{\bm{y}}) + p({\bm{z}}=0) p_n({\bm{x}}|{\bm{y}})
\end{equation*}
The posterior of $p({\bm{z}}=0|{\bm{x}},{\bm{y}})$ is
\begin{equation*}
p_{n,\text{\text{data}}}({\bm{z}}=0|{\bm{x}},{\bm{y}}) = \frac{p({\bm{z}}=0) p_n({\bm{x}}|{\bm{y}})}{p(z=0) p_n({\bm{x}}|{\bm{y}}) + p({\bm{z}}=1) p_{\text{data}}({\bm{x}}|{\bm{y}})} = \frac{\nu \cdot p_n({\bm{x}}|{\bm{y}})}{\nu \cdot p_n({\bm{x}}|{\bm{y}}) + p_{\text{data}}({\bm{x}}|{\bm{y}})},
\end{equation*}
where $\nu = \frac{p({\bm{z}}=0)}{p({\bm{z}}=1)}$.
Similarly, we can have the joint distribution under EBM framework as:
\begin{equation*}
p_{n, \theta}({\bm{x}}) = p(z=0) p_n({\bm{x}}|{\bm{y}}) + p(z=1) p_{\theta}({\bm{x}}|{\bm{y}})
\end{equation*}
And the corresponding posterior is:
\begin{equation*}
p_{n,\theta}({\bm{z}}=0|{\bm{x}},{\bm{y}}) = \frac{p({\bm{z}}=0) p_n({\bm{x}}|{\bm{y}})}{p({\bm{z}}=0) p_n({\bm{x}}|{\bm{y}}) + p({\bm{z}}=1) p_{\theta}({\bm{x}}|{\bm{y}})} = \frac{\nu \cdot p_n({\bm{x}}|{\bm{y}})}{\nu \cdot p_n({\bm{x}}|{\bm{y}}) + p_{\theta}({\bm{x}}|{\bm{y}})}
\end{equation*}
We indirectly match $p_\theta({\bm{x}}|{\bm{y}})$ to $p_{\text{data}}({\bm{x}}|{\bm{y}})$ by fitting $p_{n,\theta}({\bm{z}}|{\bm{x}},{\bm{y}})$ to $p_{n,\text{\text{data}}}({\bm{z}}|{\bm{x}},{\bm{y}})$ by minimizing their KL-divergence:
\begin{equation} \label{eq:app:EBM_01}
\begin{aligned}
& \min_\theta D_{\text{KL}}(p_{n,\text{\text{data}}}({\bm{z}}|{\bm{x}},{\bm{y}}) || p_{n,\theta}({\bm{z}}|{\bm{x}},{\bm{y}})) \\
& = \mathbb{E}_{p_{n,\text{\text{data}}}({\bm{x}},{\bm{z}}|{\bm{y}})} [\log p_{n,\theta}({\bm{z}}|{\bm{x}},{\bm{y}})] \\
& = \int \sum_{\bm{z}} p_{n,\text{\text{data}}}({\bm{x}},{\bm{z}}|{\bm{y}}) \cdot \log p_{n,\theta}({\bm{z}}|{\bm{x}},{\bm{y}}) d {\bm{x}}\\
& = \int \Big\{ p({\bm{z}}=0) p_{n,\text{\text{data}}}({\bm{x}}|{\bm{y}},{\bm{z}}=0) \log p_{n,\theta}({\bm{z}}=0|{\bm{x}},{\bm{y}}) \\
& \quad\quad\quad\quad + p({\bm{z}}=1) p_{n,\text{\text{data}}}({\bm{x}}|{\bm{z}}=1,{\bm{y}}) \log p_{n,\theta}({\bm{z}}=1|{\bm{x}},{\bm{y}}) \Big\} d{\bm{x}} \\
& = \nu \cdot \mathbb{E}_{p_{n}({\bm{x}}|{\bm{y}})} \Big[\log p_{n,\theta}({\bm{z}}=0|{\bm{x}},{\bm{y}}) \Big] + \mathbb{E}_{p_{\text{data}}({\bm{x}}|{\bm{y}} )} \Big[\log p_{n,\theta}({\bm{z}}=1|{\bm{x}},{\bm{y}}) \Big] \\
& = \nu \cdot \mathbb{E}_{p_{n}({\bm{x}}|{\bm{y}})} \Big[ \log \frac{\nu \cdot p_n({\bm{x}}|{\bm{y}})}{\nu \cdot p_n({\bm{x}}|{\bm{y}}) + p_{\theta}({\bm{x}}|{\bm{y}})} \Big] + \mathbb{E}_{p_{\text{data}}({\bm{x}}|{\bm{y}} )} \Big[ \log \frac{p_\theta({\bm{x}}|{\bm{y}})}{\nu \cdot p_n({\bm{x}}|{\bm{y}}) + p_{\theta}({\bm{x}}|{\bm{y}})} \Big].\\
\end{aligned}
\end{equation}
This optimal distribution is an estimation to the actual distribution (or data distribution), {\em{i.e.}}, $p_\theta({\bm{x}}|{\bm{y}}) \approx p_{\text{data}}({\bm{x}}|{\bm{y}})$. We can follow the similar steps for $p_\theta({\bm{y}}|{\bm{x}}) \approx p_{\text{data}}({\bm{y}}|{\bm{x}})$. Thus following~\Cref{eq:app:EBM_01}, the objective function is to maximize
\begin{equation} \label{eq:app:EBM_02}
\begin{aligned}
\nu \cdot \mathbb{E}_{p_{\text{data}}({\bm{y}})} \mathbb{E}_{p_{n}({\bm{x}}|{\bm{y}})} \Big[ \log \frac{\nu \cdot p_n({\bm{x}}|{\bm{y}})}{\nu \cdot p_n({\bm{x}}|{\bm{y}}) + p_{\theta}({\bm{x}}|{\bm{y}})} \Big] + \mathbb{E}_{p_{\text{data}}({\bm{y}})} \mathbb{E}_{p_{\text{data}}({\bm{x}}|{\bm{y}} )} \Big[ \log \frac{p_\theta({\bm{x}}|{\bm{y}})}{\nu \cdot p_n({\bm{x}}|{\bm{y}}) + p_{\theta}({\bm{x}}|{\bm{y}})} \Big].
\end{aligned}
\end{equation}
The we will adopt three strategies to approximate~\Cref{eq:app:EBM_02}:
\begin{enumerate}
\item \textbf{Self-normalization.} When the EBM is very expressive, {\em{i.e.}}, using deep neural network for modeling, we can assume it is able to approximate the normalized density directly~\cite{mnih2012fast,song2021train}. In other words, we can set the partition function $A=1$. This is a self-normalized EBM-NCE, with normalizing constant close to 1, {\em{i.e.}}, $p({\bm{x}}) = \exp(-E({\bm{x}})) = \exp(f({\bm{x}}))$ in~\Cref{eq:app:EBM_original}.
\item \textbf{Exponential tilting term.} Exponential tilting term~\cite{arbel2020generalized} is another useful trick. It models the distribution as $\tilde p_\theta({\bm{x}}) = q({\bm{x}}) \exp(-E_\theta({\bm{x}})) $, where $q({\bm{x}})$ is the reference distribution. If we use the same reference distribution as the noise distribution, the tilted probability is $\tilde p_\theta({\bm{x}}) = p_n({\bm{x}}) \exp(-E_\theta({\bm{x}}))$ in~\Cref{eq:app:EBM_original}.
\item \textbf{Sampling.} For many cases, we only need to sample 1 negative points for each data, {\em{i.e.}}, $\nu=1$.
\end{enumerate}
Following these three disciplines, the objective function to optimize $p_{\theta}({\bm{x}}|{\bm{y}})$ becomes
\begin{equation}
\begin{aligned}
& \mathbb{E}_{p_{n}({\bm{x}}|{\bm{y}})} \Big[ \log \frac{ p_n({\bm{x}}|{\bm{y}})}{ p_n({\bm{x}}|{\bm{y}}) + \tilde p_{\theta}({\bm{x}}|{\bm{y}})} \Big] + \mathbb{E}_{p_{\text{data}}({\bm{x}}|{\bm{y}} )} \Big[ \log \frac{\tilde p_\theta({\bm{x}}|{\bm{y}})}{ p_n({\bm{x}}|{\bm{y}}) + \tilde p_{\theta}({\bm{x}}|{\bm{y}})} \Big]\\
= & \mathbb{E}_{p_{n}({\bm{x}}|{\bm{y}})} \Big[ \log \frac{1}{1 + p_{\theta}({\bm{x}}|{\bm{y}})} \Big] + \mathbb{E}_{p_{\text{data}}({\bm{x}}|{\bm{y}} )} \Big[ \log \frac{p_\theta({\bm{x}}|{\bm{y}})}{1 + p_{\theta}({\bm{x}}|{\bm{y}})} \Big]\\
= & \mathbb{E}_{p_{n}({\bm{x}}|{\bm{y}})} \Big[ \log \frac{\exp (- f_{\bm{x}}({\bm{x}}, {\bm{y}}))}{\exp (- f_{\bm{x}}({\bm{x}}, {\bm{y}})) + 1} \Big] + \mathbb{E}_{p_{\text{data}}({\bm{x}}|{\bm{y}} )} \Big[ \log \frac{1}{\exp (- f_{\bm{x}}({\bm{x}}, {\bm{y}})) + 1} \Big]\\
= & \mathbb{E}_{p_{n}({\bm{x}}|{\bm{y}})} \Big[ \log \big(1-\sigma( f_{\bm{x}}({\bm{x}}, {\bm{y}}))\big) \Big] + \mathbb{E}_{p_{\text{data}}({\bm{x}}|{\bm{y}} )} \Big[ \log \sigma( f_{\bm{x}}({\bm{x}}, {\bm{y}})) \Big].
\end{aligned}
\end{equation}
Thus, the final EBM-NCE contrastive SSL objective is
\begin{equation}
\begin{aligned}
\mathcal{L}_{\text{EBM-NCE}}
& = -\frac{1}{2} \mathbb{E}_{p_{\text{data}}({\bm{y}})} \Big[\mathbb{E}_{p_{n}({\bm{x}}|{\bm{y}})} \log \big(1-\sigma( f_{\bm{x}}({\bm{x}}, {\bm{y}}))\big) + \mathbb{E}_{p_{\text{data}}({\bm{x}}|{\bm{y}} )} \log \sigma( f_{\bm{x}}({\bm{x}}, {\bm{y}})) \Big]\\
& ~~~~ - \frac{1}{2} \mathbb{E}_{p_{\text{data}}({\bm{x}})} \Big[\mathbb{E}_{p_{n}({\bm{y}}|{\bm{x}})} \log \big(1-\sigma( f_{\bm{y}}({\bm{y}},{\bm{x}}))\big) + \mathbb{E}_{p_{\text{data}}({\bm{y}},{\bm{x}})} \log \sigma( f_{\bm{y}}({\bm{y}},{\bm{x}})) \Big].
\end{aligned}
\end{equation}
\subsection{EBM-NCE v.s. JSE and InfoNCE}
We acknowledge that there are many other contrastive objectives~\citep{poole2019variational} that can be used to maximize MI. However, in the research line of graph SSL, as summarized in several recent survey papers~\citep{xie2021self,liu2021graph,wu2021self}, the two most used ones are InfoNCE and Jensen-Shannon Estimator (JSE)~\citep{nowozin2016f,hjelm2018learning}.
We conclude that JSE is very similar to EBM-NCE, while the underlying perspectives are totally different, as explained below.
\begin{enumerate}
\item \textbf{Derivation and Intuition.} Derivation process and underlying intuition are different. JSE~\citep{nowozin2016f} starts from f-divergence, then with variational estimation and Fenchel duality on function $f$. Our proposed EBM-NCE is more straightforward: it models the conditional distribution in the MI lower bound~\Cref{eq:app:MI_objective} with EBM, and solves it using NCE.
\item \textbf{Flexibility.} Modeling the conditional distribution with EBM provides a broader family of algorithms. NCE is just one solution to it, and recent progress on score matching~\citep{song2020score,song2021train} and contrastive divergence~\citep{du2020improved}, though no longer contrastive SSL, adds on more promising directions. Thus, EBM can provide a potential unified framework for structuring our understanding of self-supervised learning.
\item \textbf{Noise distribution.} Starting from~\citep{hjelm2018learning}, all the following works on graph SSL~\cite{sun2019infograph,xie2021self,liu2021graph,wu2021self} have been adopting the empirical distribution for noise distribution. However, this is not the case in EBM-NCE. Classic EBM-NCE uses fixed distribution, while more recent work~\citep{arbel2020generalized} extends it with adaptively learnable noise distribution. With this discipline, more advanced sampling strategies (w.r.t. the noise distribution) can be proposed, {\em{e.g.}}, adversarial negative sampling in \citep{hu2021adco}.
\end{enumerate}
In the above, we conclude three key differences between EBM-NCE and JSE, plus the solid and straightforward derivations on EBM-NCE. We believe this can provide a insightful perspective of SSL to the community.
According to the empirical results~\Cref{sec:experiment_each_loss_component}, we observe that EBM-NCE is better than InfoNCE. This can be explained using the claim from~\citep{khosla2020supervised}, where the main technical contribution is to construct many positives and many negatives per anchor point. The binary cross-entropy in EBM-NCE is able to realize this to some extent: make all the positive pairs positive and all the negative pairs negative, where the softmax-based cross-entropy fails to capture this, as in InfoNCE.
To conclude, we are introduce using EBM in modeling MI, which opens many potential venues. As for contrastive SSL, EBM-NCE provides a better perspective than JSE, and is better than InfoNCE on graph-level self-supervised learning.
\section{Preliminaries} \label{sec:preliminary}
We first outline the key concepts and notations used in this work. Self-supervised learning (SSL) is based on the \textit{view} design, where each view provides a specific aspect and modality of the data. Each molecule has two natural views: the 2D graph incorporates the topological structure defined by the adjacency, while the 3D graph can better reflect the geometry and spatial relation. From a chemical perspective, 3D geometric graphs focus on the \textit{energy} while 2D graphs emphasize the \textit{topological} information; thus they can be composed for learning more informative representation in GraphMVP. \textit{Transformation} is an atomic operation in SSL that can extract specific information from each view. Next, we will briefly introduce how to represent these two views.
\textbf{2D Molecular Graph} represents molecules as 2D graphs, with atoms as nodes and bonds as edges respectively. We denote each 2D graph as $g_{\text{2D}} = (X, E)$, where $X$ is the atom attribute matrix and $E$ is the bond attribute matrix. Notice that here $E$ also includes the bond connectivity. Then we will apply a transformation function $T_{\text{2D}}$ on the topological graph. Given a 2D molecular graph $g_{\text{2D}}$, its representation $h_{2D}$ can be obtained from a \textit{2D graph neural network (GNN)} model:
\begin{equation} \label{eq:2d_gnn}
h_{\text{2D}} = \text{GNN-2D}(T_{\text{2D}}(g_{\text{2D}})) = \text{GNN-2D}(T_{\text{2D}}(X, E)).
\end{equation}
\textbf{3D Molecular Graph} additionally includes spatial positions of the atoms, and they are needless to be static since atoms are in continual motion on \textit{a potential energy surface}~\citep{axelrod2020geom}. \footnote{A more rigorous way of defining conformer is in~\citep{moss1996basic}: a conformer is an isomer of a molecule that differs from another isomer by the rotation of a single bond in the molecule.} The 3D structures at the local minima on this surface are named \textit{conformer}. As the molecular properties are conformers ensembled~\citep{hawkins2017conformation}, GraphMVP\, provides a novel perspective on adopting 3D conformers for learning better representation. Given a conformer $g_{\text{3D}} = (X, R)$, its representation via a \textit{3D GNN} model is:
\begin{equation} \label{eq:3d_gnn}
h_{\text{3D}} = \text{GNN-3D}(T_{\text{3D}}(g_{\text{3D}})) = \text{GNN-3D}(T_{\text{3D}}(X, R)),
\end{equation}
where $R$ is the 3D-coordinate matrix and $T_{\text{3D}}$ is the 3D transformation.
In what follows, for notation simplicity, we use ${\bm{x}}$ and ${\bm{y}}$ for the 2D and 3D graphs, {\em{i.e.}}, ${\bm{x}} \triangleq g_{\text{2D}}$ and ${\bm{y}} \triangleq g_{\text{3D}}$. Then the latent representations are denoted as $h_{\bm{x}}$ and $h_{\bm{y}}$.
\part{\Large\centering{Appendix}}
\parttoc
\clearpage
\input{09_related}
\clearpage
\input{10_MI}
\clearpage
\input{11_contrastive_ssl}
\clearpage
\input{12_generative_ssl}
\clearpage
\input{13_datasets}
\input{14_experiments}
\section{Maximize Mutual Information} \label{sec:app:MI}
In what follows, we will use $X$ and $Y$ to denote the data space for $2D$ graph and $3D$ graph respectively. Then the latent representations are denoted as $h_{\bm{x}}$ and $h_{\bm{y}}$.
\subsection{Formulation}
The standard formulation for mutual information (MI) is
\begin{equation} \label{eq:app:MI}
\begin{aligned}
I(X;Y)
& = \mathbb{E}_{p({\bm{x}},{\bm{y}})} \big[ \log \frac{p({\bm{x}},{\bm{y}})}{p({\bm{x}}) p({\bm{y}})} \big].
\end{aligned}
\end{equation}
Another well-explained MI inspired from wikipedia is given in~\Cref{fig:app:MI}.
\begin{figure}[htb!]
\centering
\includegraphics[width=0.5\textwidth]{figures/MutualInfo_Final.pdf}
\caption{Venn diagram of mutual information. Inspired by wikipedia.}
\label{fig:app:MI}
\end{figure}
Mutual information (MI) between random variables measures the corresponding non-linear dependence. As can be seen in the first equation in~\Cref{eq:app:MI}, the larger the divergence between the joint ($p({\bm{x}}, {\bm{y}}$) and the product of the marginals $p({\bm{x}}) p({\bm{y}})$, the stronger the dependence between $X$ and $Y$.
Thus, following this logic, maximizing MI between 2D and 3D views can force the 3D/2D representation to capture higher-level factors, {\em{e.g.}}, the occurrence of important substructure that is semantically vital for downstream tasks. Or equivalently, maximizing MI can decrease the uncertainty in 2D representation given 3D geometric information.
\subsection{A Lower Bound to MI}
To solve MI, we first extract a lower bound:
\begin{equation}
\begin{aligned}
I(X;Y)
& = \mathbb{E}_{p({\bm{x}},{\bm{y}})} \Big[ \log \frac{p({\bm{x}},{\bm{y}})}{p({\bm{x}}) p({\bm{y}})} \Big]\\
& \ge \mathbb{E}_{p({\bm{x}},{\bm{y}})} \Big[ \log \frac{p({\bm{x}},{\bm{y}})}{\sqrt{p({\bm{x}}) p({\bm{y}})}} \Big]\\
& = \frac{1}{2} \mathbb{E}_{p({\bm{x}},{\bm{y}})} \Big[ \log \frac{(p({\bm{x}},{\bm{y}}))^2}{p({\bm{x}}) p({\bm{y}})} \Big]\\
& = \frac{1}{2} \mathbb{E}_{p({\bm{x}},{\bm{y}})} \Big[ \log p({\bm{x}}|{\bm{y}}) \Big] + \frac{1}{2} \mathbb{E}_{p({\bm{x}},{\bm{y}})} \Big[ \log p({\bm{y}}|{\bm{x}}) \Big]\\
& = -\frac{1}{2} [H(Y|X) + H(X|Y)].
\end{aligned}
\end{equation}
Thus, we transform the MI maximization problem into minimizing the following objective:
\begin{equation} \label{eq:app:MI_objective}
\begin{aligned}
\mathcal{L}_{\text{MI}} & = \frac{1}{2} [H(Y|X) + H(X|Y)].
\end{aligned}
\end{equation}
In the following sections, we will describe two self-supervised learning methods for solving MI. Notice that the methods are very general, and can be applied to various applications. Here we apply it mainly for making 3D geometry useful for 2D representation learning on molecules.
\section{GraphMVP: Graph Multi-View Pre-training} \label{sec:method}
Our model, termed as Graph Multi-View Pre-training (GraphMVP), conducts self-supervised learning (SSL) pre-training with 3D information. The 3D conformers encode rich information about the molecule energy and spatial structure, which are complementary to the 2D topology. Thus, applying SSL between the 2D and 3D views will provide a better 2D representation, which implicitly embeds the ensembles of energies and geometric information for molecules.
In the following, we first present an overview of GraphMVP, and then introduce two pretext tasks specialized concerning 3D conformation structures. Finally, we summarize a broader graph SSL family that prevails the 2D molecular graph representation learning with 3D geometry.
\subsection{Overview of GraphMVP} \label{sec:overview}
\begin{figure}[ht]
\centering
\vspace{-2ex}
\includegraphics[width=\textwidth]{figures/Diagram_Bundle_4.pdf}
\vspace{-2ex}
\caption{
Overview of the pre-training stage in GraphMVP. The black dashed circles denote subgraph masking, and we mask the same region in the 2D and 3D graphs. Multiple views of the molecules (herein: Halicin) are mapped to the representation space via 2D and 3D GNN models, where we conduct GraphMVP\ for SSL pre-training, using both contrastive and generative pretext tasks.
}
\label{fig:both_ssl}
\end{figure}
In general, GraphMVP\, exerts 2D topology and 3D geometry as two complementary views for each molecule. By performing SSL between these views, it is expected to learn a 2D representation enhanced with 3D conformation, which can better reflect certain molecular properties.
As generic SSL pre-training pipelines, GraphMVP\, has two stages: pre-training then fine-tuning. In the pre-training stage, we conduct SSL via auxiliary tasks on data collections that provide both 2D and 3D molecular structures. During fine-tuning, the pre-trained 2D GNN models are subsequently fine-tuned on specific downstream tasks, where only 2D molecular graphs are available.
At the SSL pre-training stage, we design two pretext tasks: one contrastive and one generative. We conjecture and then empirically prove that these two tasks are focusing on different learning aspects, which are summarized into the following two points. (1) From the perspective of representation learning, contrastive SSL utilizes \textbf{inter-data} knowledge and generative SSL utilizes \textbf{intra-data} knowledge. For contrastive SSL, one key step is to obtain the negative view pairs for inter-data contrasting; while generative SSL focuses on each data point itself, by reconstructing the key features at an intra-data level. (2) From the perspective of distribution learning, contrastive SSL and generative SSL are learning the data distribution from a \textbf{local} and \textbf{global} manner, respectively. Contrastive SSL learns the distribution locally by contrasting the pairwise distance at an inter-data level. Thus, with sufficient number of data points, the local contrastive operation can iteratively recover the data distribution. Generative SSL, on the other hand, learns the global data density function directly.
Therefore, contrastive and generative SSL are essentially conducting representation and distribution learning with different intuitions and disciplines, and we expect that combining both can lead to a better representation. We later carry out an ablation study (\Cref{sec:experiment_each_loss_component}) to verify this empirically.
In addition, to make the pretext tasks more challenging, we take views for each molecule by randomly masking $M$ nodes (and corresponding edges) as the transformation function, {\em{i.e.}}, $T_{\text{2D}} = T_{\text{3D}} =\text{mask}$. This trick has been widely used in graph SSL~\citep{hu2019strategies,You2020GraphCL,you2021graph} and has shown robust improvements.
\subsection{Contrastive Self-Supervised Learning between 2D and 3D Views} \label{sec:contrastive_SSL}
The main idea of contrastive self-supervised learning (SSL)~\citep{oord2018representation,chen2020simple} is first to define positive and negative pairs of views from an inter-data level, and then to align the positive pairs and contrast the negative pairs simultaneously~\citep{wang2020understanding}. For each molecule, we first extract representations from 2D and 3D views, {\em{i.e.}}, $h_{\bm{x}}$ and $h_{\bm{y}}$. Then we create positive and negative pairs for contrastive learning: the 2D-3D pairs $({\bm{x}},{\bm{y}})$ for the same molecule are treated as positive, and negative otherwise. Finally, we align the positive pairs and contrast the negative ones. The pipeline is shown in~\Cref{fig:both_ssl}. In the following, we discuss two common objective functions on contrastive graph SSL.
\textbf{InfoNCE} is first proposed in~\citep{oord2018representation}, and its effectiveness has been validated both empirically~\citep{chen2020simple,he2020momentum} and theoretically~\citep{arora2019theoretical}. Its formulation is given as follows:
\begin{equation}\label{eq:objective_infonce}
\small{
\fontsize{7.8}{1}\selectfont
\begin{aligned}
\mathcal{L}_{\text{InfoNCE}}=-\frac{1}{2} \mathbb{E}_{p({\bm{x}},{\bm{y}})}\Big[\log \frac{\exp(f_{{\bm{x}}}({\bm{x}}, {\bm{y}}))}{\exp(f_{{\bm{x}}}({\bm{x}}, {\bm{y}})) + \sum\limits_{j} \exp(f_{{\bm{x}}}({\bm{x}}^{j},{\bm{y}})})+\log\frac{\exp(f_{{\bm{y}}}({\bm{y}},{\bm{x}}))}{\exp(f_{{\bm{y}}}({\bm{y}},{\bm{x}})) + \sum\limits_{j} \exp(f_{{\bm{y}}}({\bm{y}}^{j},{\bm{x}}))} \Big],
\end{aligned}
}
\end{equation}
where ${\bm{x}}^{j}, {\bm{y}}^{j}$ are randomly sampled 2D and 3D views regarding to the anchored pair $({\bm{x}},{\bm{y}})$. $f_{{\bm{x}}}({\bm{x}},{\bm{y}})$ and $f_{{\bm{y}}}({\bm{y}},{\bm{x}})$ are scoring functions for the two corresponding views, with flexible formulations. Here we adopt $f_{\bm{x}}({\bm{x}},{\bm{y}}) = f_{\bm{y}}({\bm{y}},{\bm{x}}) = \langle h_{\bm{x}}, h_{\bm{y}} \rangle$. More details are in~\Cref{sec:app:contrastive_ssl}.
\textbf{Energy-Based Model with Noise Contrastive Estimation (EBM-NCE)} is an alternative that has been widely used in the line of graph contrastive SSL~\citep{sun2019infograph,hu2019strategies,You2020GraphCL,you2021graph}. Its intention is essentially the same as InfoNCE, to align positive pairs and contrast negative pairs, while the main difference is the usage of binary cross-entropy and extra noise distribution for negative sampling:
\begin{equation} \label{eq:objective_ebm_nce}
\small{
\begin{aligned}
\mathcal{L}_{\text{EBM-NCE}}
& = -\frac{1}{2} \mathbb{E}_{p({\bm{y}})} \Big[\mathbb{E}_{p_{n}({\bm{x}}|{\bm{y}})} \log \big(1-\sigma( f_x({\bm{x}}, {\bm{y}}))\big) + \mathbb{E}_{p({\bm{x}}|{\bm{y}} )} \log \sigma( f_x({\bm{x}}, {\bm{y}})) \Big]\\
& ~~~~~ -\frac{1}{2} \mathbb{E}_{p({\bm{x}})} \Big[\mathbb{E}_{p_{n}({\bm{y}}|{\bm{x}})} \log \big(1-\sigma( f_y({\bm{y}},{\bm{x}}))\big) + \mathbb{E}_{p({\bm{y}},{\bm{x}})} \log \sigma( f_y({\bm{y}},{\bm{x}})) \Big],
\end{aligned}
}
\end{equation}
where $p_n$ is the noise distribution and $\sigma$ is the sigmoid function. We also notice that the final formulation of EBM-NCE shares certain similarities with Jensen-Shannon estimation (JSE)~\citep{nowozin2016f}. However, the derivation process and underlying intuition are different: EBM-NCE models the conditional distributions in MI lower bound (\Cref{eq:MI_objective}) with EBM, while JSE is a special case of variational estimation of f-divergence. Since this is not the main focus of GraphMVP, we expand the a more comprehensive comparison in~\Cref{sec:app:contrastive_ssl}, plus the potential benefits with EBM-NCE.
Few works~\citep{hassani2020contrastive} have witnessed the effect on the choice of objectives in graph contrastive SSL. In GraphMVP, we treat it as a hyper-parameter and further run ablation studies on them, {\em{i.e.}}, to solely use either InfoNCE ($\mathcal{L}_{\text{C}} = \mathcal{L}_{\text{InfoNCE}}$) or EMB-NCE ($\mathcal{L}_{\text{C}} = \mathcal{L}_{\text{EBM-NCE}}$).
\subsection{Generative Self-Supervised Learning between 2D and 3D Views} \label{sec:generative_SSL}
Generative SSL is another classic track for unsupervised pre-training~\citep{kingma2013auto,chen2016infogan,larsson2016learning,kingma2018glow}. It aims at learning an effective representation by self-reconstructing each data point. Specifically to drug discovery, we have one 2D graph and a certain number of 3D conformers for each molecule, and our goal is to learn a robust 2D/3D representation that can, to the most extent, recover its 3D/2D counterparts. By doing so, generative SSL can enforce 2D/3D GNN to encode the most crucial geometry/topology information, which can improve the downstream performance.
There are many options for generative models, including variational auto-encoder (VAE)~\citep{kingma2013auto}, generative adversarial networks (GAN)~\citep{goodfellow2014generative}, flow-based model~\citep{dinh2016density}, etc. In GraphMVP, we prefer VAE-like method for the following reasons: (1) The mapping between two molecular views is stochastic: multiple 3D conformers correspond to the same 2D topology; (2) An explicit 2D graph representation ({\em{i.e.}}, feature encoder) is required for downstream tasks; (3) Decoders for structured data such as graph are often highly nontrivial to design, which make them a suboptimal choice.
\textbf{Variational Molecule Reconstruction.}
Therefore we propose a \textit{light} VAE-like generative SSL, equipped with a \textit{crafty} surrogate loss, which we describe in the following. We start with an example for illustration. When generating 3D conformers from their corresponding 2D topology, we want to model the conditional likelihood $p({\bm{y}}|{\bm{x}})$. By introducing a reparameterized variable ${\bm{z}}_{\bm{x}} = \mu_{{\bm{x}}} + \sigma_{{\bm{x}}} \odot \epsilon$, where $\mu_{\bm{x}}$ and $\sigma_{\bm{x}}$ are two flexible functions on $h_{\bm{x}}$, $\epsilon \sim \mathcal{N}(0,I)$ and $\odot$ is the element-wise production,
we have the following lower bound:
\begin{equation} \label{eq:variational_lower_bound}
\small
\begin{aligned}
\log p({\bm{y}}|{\bm{x}})
\ge \mathbb{E}_{q({\bm{z}}_{\bm{x}}|{\bm{x}})} \big[ \log p({\bm{y}}|{\bm{z}}_{\bm{x}}) \big] - KL(q({\bm{z}}_{\bm{x}}|{\bm{x}}) || p({\bm{z}}_{\bm{x}})).
\end{aligned}
\end{equation}
The expression for $\log p({\bm{x}}|{\bm{y}})$ can be similarly derived. \Cref{eq:variational_lower_bound} includes a conditional log-likelihood and a KL-divergence term, where the bottleneck is to calculate the first term for structured data. This term has also been recognized as the \textit{reconstruction term}: it is essentially to reconstruct the 3D conformers (${\bm{y}}$) from the sampled 2D molecular graph representation (${\bm{z}}_{{\bm{x}}}$). However, performing the graph reconstruction on the data space is not trivial: since molecules ({\em{e.g.}}, atoms and bonds) are discrete, modeling and measuring on the molecule space will bring extra obstacles.
\textbf{Variational Representation Reconstruction (VRR).}
To cope with this challenge, we propose a novel surrogate loss by switching the reconstruction from data space to representation space. Instead of decoding the latent code $z_{{\bm{x}}}$ to data space, we can directly project it to the 3D representation space, denoted as $q_{\bm{x}}(z_{\bm{x}})$. Since the representation space is continuous, we may as well model the conditional log-likelihood with Gaussian distribution, resulting in L2 distance for reconstruction, {\em{i.e.}}, $\|q_{{\bm{x}}}(z_{{\bm{x}}}) - \text{SG}(h_{\bm{y}}({\bm{y}}))\|^2$. Here SG is the stop-gradient operation, assuming that $h_{{\bm{y}}}$ is a fixed learnt representation function. SG has been widely adopted in the SSL literature to avoid model collapse~\citep{grill2020bootstrap,chen2021exploring}. We call this surrogate loss as variational representation reconstruction (VRR):
\begin{equation} \label{eq:variational_lower_bound_approximation}
\small
\begin{aligned}
\mathcal{L}_{\text{G}} = \mathcal{L}_{\text{VRR}}
= & \frac{1}{2} \Big[ \mathbb{E}_{q({\bm{z}}_{\bm{x}}|{\bm{x}})} \big[ \| q_{{\bm{x}}}({\bm{z}}_{\bm{x}}) - \text{SG}(h_{{\bm{y}}}) \|^2 \big] + \mathbb{E}_{q({\bm{z}}_{\bm{y}}|{\bm{y}})} \big[ \| q_{{\bm{y}}}({\bm{z}}_{\bm{y}}) - \text{SG}(h_{{\bm{x}}}) \|_2^2 \big] \Big]\\
& + \frac{\beta}{2} \cdot \Big[ KL(q({\bm{z}}_{\bm{x}}|{\bm{x}}) || p({\bm{z}}_{\bm{x}})) + KL(q({\bm{z}}_{\bm{y}}|{\bm{y}}) || p({\bm{z}}_{\bm{y}})) \Big].
\end{aligned}
\end{equation}
We give a simplified illustration for the generative SSL pipeline in~\Cref{fig:both_ssl} and the complete derivations in~\Cref{sec:app:generative_ssl}. As will be discussed in~\Cref{sec:mutual_information}, VRR is actually maximizing MI, and MI is invariant to continuous bijective function~\citep{belghazi2018mutual}. Thus, this surrogate loss would be exact if the encoding function $h$ satisfies this condition. However, we find that GNN, though does not meet the condition, can provide quite robust performance, which empirically justify the effectiveness of VRR.
\subsection{Multi-task Objective Function}
As discussed before, contrastive SSL and generative SSL essentially learn the representation from distinct viewpoints. A reasonable conjecture is that combining both SSL methods can lead to overall better performance, thus we arrive at minimizing the following complete objective for GraphMVP:
\begin{equation} \label{eq:graphmvp}
\small
\begin{aligned}
\mathcal{L}_{\text{GraphMVP}}
& = \alpha_1 \cdot \mathcal{L}_{\text{C}} + \alpha_2 \cdot \mathcal{L}_{\text{G}},\\
\end{aligned}
\end{equation}
where $\alpha_1, \alpha_2$ are weighting coefficients. A later performed ablation study (\Cref{sec:experiment_each_loss_component}) delivers two important messages: (1) Both individual contrastive and generative SSL on 3D conformers can consistently help improve the 2D representation learning; (2) Combining the two SSL strategies can yield further improvements. Thus, we draw the conclusion that GraphMVP\, (\Cref{eq:graphmvp}) is able to obtain an augmented 2D representation by fully utilizing the 3D information.
As discussed in~\Cref{sec:intro}, existing graph SSL methods only focus on the 2D topology, which is in parallel to GraphMVP: 2D graph SSL focuses on exploiting the 2D structure topology, and GraphMVP\, takes advantage of the 3D geometry information. Thus, we propose to merge the 2D SSL into GraphMVP. Since there are two main categories in 2D graph SSL: generative and contrastive, we propose two variants GraphMVP-G\, and GraphMVP-C\, accordingly. Their objectives are as follows:
\begin{equation} \label{eq:graphmvp_variants}
\small
\begin{aligned}
\mathcal{L}_{\text{GraphMVP-G}} = \mathcal{L}_{\text{GraphMVP}} + \alpha_3 \cdot \mathcal{L}_{\text{Generative 2D-SSL}},~~~~~
\mathcal{L}_{\text{GraphMVP-C}} = \mathcal{L}_{\text{GraphMVP}} + \alpha_3 \cdot \mathcal{L}_{\text{Contrastive 2D-SSL}}.
\end{aligned}
\end{equation}
Later, the empirical results also help support the effectiveness of GraphMVP-G\, and GraphMVP-C, and thus, we can conclude that existing 2D SSL is complementary to GraphMVP.
\section*{Reproducibility Statement}
To ensure the reproducibility of the empirical results, we include our code base in the supplementary material, which contains: instructions for installing conda virtual environment, data preprocessing scripts, and training scripts. Our code is available on \href{https://github.com/chao1224/GraphMVP}{GitHub} for reproducibility. Complete derivations of equations and clear explanations are given in~\Cref{sec:app:MI,sec:app:contrastive_ssl,sec:app:generative_ssl}.
\section*{Acknowledgement}
This project is supported by the Natural Sciences and Engineering Research Council (NSERC) Discovery Grant, the Canada CIFAR AI Chair Program, collaboration grants between Microsoft Research and Mila, Samsung Electronics Co., Ltd., Amazon Faculty Research Award, Tencent AI Lab Rhino-Bird Gift Fund and a NRC Collaborative R\&D Project (AI4D-CORE-06). This project was also partially funded by IVADO Fundamental Research Project grant PRF-2019-3583139727.
\bibliographystyle{plain}
{\small
|
\section{Quantum statistical mechanics}
\label{sec:zubarev:quant_stat_mech}
The aim of statistical mechanics is the description of systems which are known solely by the specification of some global variables.
Typically these systems are macroscopic, in the sense that they involve a great number of degrees of freedom, and in quantum mechanical language, the given macroscopic data are expectation values of operators.
At the microscopic level, however, there are in general several possible states producing the same values for the macroscopic variables, therefore the
state is not completely known: all we can give is the probability that the system is described by one state or another.
Therefore, such a system is represented by the set $\{|\psi_i\rangle,p_i\}$ of the possible states $|\psi_i\rangle$ compatible with the given macroscopic data each with its own probability of occurrence $p_i\in [0,1]$ such that $\sum_ip_i=1$.
The set $\{|\psi_i\rangle,p_i\}$ is called a \textit{statistical ensemble}, and each state $|\psi_i\rangle$ is an element of that ensemble.
Thus, such a system embodies two probabilities of different origin: the one is quantum mechanical and due to the $|\psi_i\rangle$, the other is classical and ascribed to our incomplete knowledge of the state represented by the $p_i$.
If the system is exactly in one of the $|\psi_i\rangle$ it is said to be in a \textit{pure state}, otherwise it is in a \textit{mixed state}.
For a pure state, all the $p_i$ vanish except one, which is equal to 1.
On the contrary, for a maximally mixed state, all the $|\psi_i\rangle$ have the same probability, which is $p_i=1/N$ with $N$ the dimension of the Hilbert space the $|\psi_i\rangle$ belong to.
For a system described by the statistical ensamble $\{|\psi_i\rangle,p_i\}$, the expectation value of an operator $\widehat{O}$ is
\begin{equation}\label{zubeq19}
\langle \widehat{O}\rangle=\sum_ip_i\langle \psi_i|\widehat{O}|\psi_i\rangle=\mathrm{tr}(\widehat{\rho}\widehat{O}),
\end{equation}
where $\widehat{\rho}$ such that $\mathrm{tr}(\widehat{\rho})=1$ is called the \textit{density operator} and is defined as
\begin{equation}\label{zubeq08}
\widehat{\rho}\equiv \sum_ip_i|\psi_i\rangle \langle \psi_i|.
\end{equation}
The density operator embodies in a compact form the statistical nature of the macroscopic system and allows us to calculate expectation values of operators, therefore it is the natural object used in quantum statistical mechanics to describe a statistical ensemble.
Our knowledge of the system is more or less complete: clearly our information is maximum when we can make predictions with full certainty, and it is larger when the system is in a pure state than in a mixed one.
Moreover, the system is better known when the number of possible
states is small or when the probability for one of them is close to unity than when there is a large number of possible
states with all approximately the same probability.
The idea is then to identify the missing information with a quantitative measure for the degree of disorder existing in a system whose preparation has a random nature.
This quantity is the \textit{von Neumann entropy}
\begin{equation}\label{zubeq01}
S\equiv -\mathrm{tr}(\widehat{\rho}\log \widehat{\rho}),
\end{equation}
which is, in fact, maximum for a maximally mixed state and minimum and equal to zero for a pure state, as desired.
The von Neumann entropy satisfies a number of other properties, but to our intents and purposes it suffices to say that it is a good measure of our missing information on a system.
Recall that the thermodynamic entropy is the von Neumann entropy multiplied by the Boltzmann constant, which in this work is set to $k_{\rm B}=1$, so we regard them as the same quantity which we indicate as $S$ and simply refer to as ``entropy''.
In order to actually be able to make predictions on a system prepared in some given way, we must know how to assign it a density operator representing the physical configuration that we want to describe.
If we know nothing about the system, the answer is simple: we must assume that all the possible
states are equally probable, any other choice would arbitrarily introduce an order for which there are no reasons of believing that it exists.
This is but the maximally mixed state, namely the one that maximizes the
entropy.
If the system is partially known, meaning that we are given some data, the same kind of philosophy applies: among all the possible
states, we must select those compatible with the given data and assign them equal probabilities.
In other words, we must look for the maximum of the
entropy while reproducing the known data.
We assume therefore that the best guess for $\widehat{\rho}$ is provided by the following prescription: amongst all the density operators compatible with the available data, we must represent the system by the one which has the largest value of the
entropy.
This is called the \textit{maximum entropy principle}.
The states determined accordingly are said to be at \textit{thermodynamic equilibrium}, all other states are called of \textit{non-equilibrium}.
Mathematically, the available information is represented by a set of expectation values $\{\langle \widehat{O}_i\rangle\}$ of a set of observables $\{\widehat{O}_i\}$.
The density operator must reproduce the data, therefore it must fulfill that constraints
\begin{equation}
\mathrm{tr}(\widehat{\rho}\widehat{O}_i)=\langle \widehat{O}_i\rangle,\qquad \forall i.
\end{equation}
For each constraint, we introduce a Lagrange multiplier $\lambda_i$ and maximize the functional expression
\begin{equation}\label{zubeq02}
-\mathrm{tr}(\widehat{\rho}\log \widehat{\rho})-\sum_i\lambda_i\left(\mathrm{tr}(\widehat{\rho}\widehat{O}_i)-\langle \widehat{O}_i\rangle \right)
\end{equation}
with respect to $\widehat{\rho}$.
The solution has the well-known form
\begin{equation}\label{zubeq03}
\widehat{\rho}=\frac{1}{Z}\exp \left[-\sum_i\lambda_i\widehat{O}_i\right],\qquad
Z=\mathrm{tr} \left(\exp \left[-\sum_i\lambda_i\widehat{O}_i\right]\right),
\end{equation}
where $Z$ is the \textit{partition function} enforcing the normalization $\mathrm{tr}(\widehat{\rho})=1$.
Depending on which set of observables is chosen, a different statistical ensamble is obtained.
We conclude this review with a remark on the time evolution.
In the Schr\"odinger picture, the evolution of the
states $|\psi_i\rangle$ is governed by the Schr\"odinger equation, whose solution is a unitary transformation of $|\psi_i\rangle$
\begin{equation}
i\frac{{\rm d}}{{\rm d} t}|\psi_i\rangle=\widehat{H}|\psi_i\rangle
\qquad \Rightarrow \qquad
|\psi_i(t)\rangle=\widehat{U}(t)|\psi_i(0)\rangle,\qquad
\widehat{U}(t)={\rm e}^{-i\widehat{H}t}
\end{equation}
with $\widehat{H}$ the Hamiltonian.
As a consequence, the evolution of the density operator is governed by the von Neumann equation, which is similar to the Heisenberg equation but with opposite sign.
Its solution is again a unitary transformation of $\widehat{\rho}$
\begin{equation}\label{zubeq14}
\frac{{\rm d} \widehat{\rho}}{{\rm d} t}=-i[\widehat{H},\widehat{\rho}]
\qquad \Rightarrow \qquad
\widehat{\rho}(t)=\widehat{U}(t)\widehat{\rho}(0)\widehat{U}^{\dagger}(t).
\end{equation}
A known property of the
entropy, which can be checked by using the definition \eqref{zubeq01}, is the invariance under unitary transformations of the density operator, therefore it is in fact stationary in the Schr\"odinger picture.
This is more obvious in the Heisenberg picture, where the
states $|\psi_i\rangle$ are constant in time, and so is the density operator.
We must conclude that, regardless of the picture, the
entropy is always stationary in quantum statistical mechanics.
However, it is well-known that, in classical physics, irreversible processes that produce entropy can definitely occur, as we discussed also in Chapter \ref{chapter:relhydro}.
From a kinetic theory perspective, for instance, if a system is initially out of thermodynamic equilibrium, the collisions between particles make the entropy increase until it reaches a maximum corresponding to a state of thermodynamic equilibrium.
Thus, in analogy with the classical theory, is there a way to describe irreversible processes where the entropy can actually increase in the quantum theory?
This question leads us to the next Section.
\section{The Zubarev approach}
\label{sec:zubarev_zub_approach}
The aim of D.\ N.\ Zubarev was actually way broader than just the description of quantum processes with entropy production.
With the development of relativistic astrophysics, cosmology, and the hydrodynamic theory of multiparticle production, in the late 70's it became necessary to go beyond the framework of the phenomenological linear relativistic hydrodynamics.
Thus, his purpose back then was rather to derive systematically the equations of non-linear quantum relativistic hydrodynamics, the transport relations and the transport coefficients without relying in the general case on the assumption that the state of the system is close to thermodynamic equilibrium.
Dealing with full non-equilibrium and including quantum effects, his formulation appears to be
completely general
and also well suited to describe, among other things, the entropy production of quantum processes.
\subsection{Local thermodynamic equilibrium density operator}
The core of the Zubarev formulation of the quantum statistical foundations of relativistic hydrodynamics is the so-called \textit{local thermodynamic equilibrium density operator}.
This is defined, according to the maximum entropy principle, as the operator $\widehat{\rho}_{\rm LE}$ that maximizes the functional expression $-\mathrm{tr}(\widehat{\rho}_{\rm LE}\log \widehat{\rho}_{\rm LE})$ with constraints of given energy, momentum and possible charge densities.
This definition does not assume any underlying kinetic theory, and it also is completely unambiguous in the non-relativistic theory as it only demands for the densities to vary significantly over distances much larger than the characteristic microscopic scale \cite{balian2006microphysics, Becattini:2014yxa}.
Its relativistic extension, however, requires a little more effort due to the stronger frame-dependence of the densities.
The first step towards a covariant expression of $\widehat{\rho}_{\rm LE}$ is the foliation of the spacetime with a 1-parameter family of 3-dimensional spacelike hypersurfaces $\{ \Sigma(\tau)\}$.
Indeed the timelike unit vector field $n^{\mu}$ normal to the hypersurfaces defines worldlines of observers, but $\tau$ does not in general coincide with the proper time of comoving observers.
By the Frobenius theorem, in order for the foliation to be well-defined, the normal unit vector must fulfill the so-called \textit{vorticity-free condition}
\begin{equation}\label{zubeq07}
\epsilon_{\mu \nu \rho \sigma}n^{\mu}(\partial^{\nu}n^{\rho}-\partial^{\rho}n^{\nu})=0,
\end{equation}
with $\epsilon_{\mu \nu \rho \sigma}$ the Levi-Civita symbol.
Such a foliation, also known as the Arnowitt-Deser-Misner (ADM) foliation, will hereafter be assumed to exist for the geometry at hand.
Whatever the expression of $\widehat{\rho}_{\rm LE}$ is, the renormalized energy-momentum and charge densities on a hypersurface $\Sigma(\tau)$ of the foliation calculated with $\widehat{\rho}_{\rm LE}$ are respectively
$n_{\mu}\mathrm{tr}(\widehat{\rho}_{\rm LE}\widehat{T}^{\mu \nu})_{\rm ren}$ and $n_{\mu}\mathrm{tr}(\widehat{\rho}_{\rm LE}\widehat{j}^{\mu})_{\rm ren}$.
These are constrained to equal the actual values, that is
\begin{equation}\label{zubeq04}
n_{\mu}\mathrm{tr}(\widehat{\rho}_{\rm LE}\widehat{T}^{\mu \nu})_{\rm ren}=n_{\mu}T^{\mu \nu},\qquad
n_{\mu}\mathrm{tr}(\widehat{\rho}_{\rm LE}\widehat{j}^{\mu})_{\rm ren}=n_{\mu}j^{\mu}.
\end{equation}
In the following, these will be referred to as the \textit{constraint equations}.
The quantities at right-hand side are the actual ones, finite, however they are known or defined.
In other words, they are what we called the given data in the previous Section.
The plain expectation values of $\widehat{T}^{\mu \nu}$ and $\widehat{j}^{\mu}$ are divergent in general, therefore it is important to emphasize that the quantities at left-hand side must be suitably renormalized in order to match the actual ones.
The renormalization procedure depends on the details of the underlying Quantum Field Theory.
Throughout the rest of this work, we will be concerned with the simplest case of free field theories, therefore renormalization is most readily established by subtraction of the vacuum expectation value
\begin{subequations}
\begin{align}
\mathrm{tr}(\widehat{\rho}_{\rm LE}\widehat{T}^{\mu \nu})_{\rm ren}=&
\mathrm{tr}(\widehat{\rho}_{\rm LE}\widehat{T}^{\mu \nu})-\langle 0|\widehat{T}^{\mu \nu}|0\rangle=
\mathrm{tr}(\widehat{\rho}_{\rm LE}:\widehat{T}^{\mu \nu}:)=\langle :\widehat{T}^{\mu \nu}:\rangle_{\rm LE}\\
\mathrm{tr}(\widehat{\rho}_{\rm LE}\widehat{j}^{\mu})_{\rm ren}=&
\mathrm{tr}(\widehat{\rho}_{\rm LE}\widehat{j}^{\mu})-\langle 0|\widehat{j}^{\mu}|0\rangle=
\mathrm{tr}(\widehat{\rho}_{\rm LE}:\widehat{j}^{\mu}:)=\langle :\widehat{j}^{\mu}:\rangle_{\rm LE},
\end{align}
\end{subequations}
which is tantamount to normally ordering the creation and annihilation operators because the conserved currents are quadratic in the fields.
In the last step, we have used the notation $\langle \widehat{O}\rangle_{\rm LE}\equiv \mathrm{tr}(\widehat{\rho}_{\rm LE}\widehat{O})$ for the expectation value of an operator $\widehat{O}$ calculated with $\widehat{\rho}_{\rm LE}$, much in the same was as \eqref{zubeq19}.
Thus, in analogy with the non-relativistic case \eqref{zubeq02}, we introduce a set of Lagrange multipliers and maximize following functional expression with respect to $\widehat{\rho}_{\rm LE}$
\begin{equation}
-\mathrm{tr}(\widehat{\rho}_{\rm LE}\log \widehat{\rho}_{\rm LE})+\int_{\Sigma(\tau)}{\rm d} \Sigma_{\mu}\left[
\left(\mathrm{tr}(\widehat{\rho}_{\rm LE}\widehat{T}^{\mu \nu})_{\rm ren}-T^{\mu \nu}\right)\beta_{\nu}-\zeta \left(\mathrm{tr}(\widehat{\rho}_{\rm LE}\widehat{j}^{\mu})_{\rm ren}-j^{\mu}\right)\right],
\end{equation}
where ${\rm d} \Sigma_{\mu}\equiv {\rm d} \Sigma \,n_{\mu}$ with ${\rm d} \Sigma$ the measure on $\Sigma(\tau)$.
Once again in analogy with \eqref{zubeq03}, the solution of the variational problem is
\begin{equation}\label{zubeq05}
\widehat{\rho}_{\rm LE}=\frac{1}{Z_{\rm LE}}\exp \left[-\int_{\Sigma(\tau)}{\rm d} \Sigma_{\mu}\left(\widehat{T}^{\mu \nu}\beta_{\nu}-\zeta \widehat{j}^{\mu}\right)\right]
\end{equation}
\begin{equation}\label{zubeq06}
Z_{\rm LE}=\mathrm{tr} \left(\exp \left[-\int_{\Sigma(\tau)}{\rm d} \Sigma_{\mu}\left(\widehat{T}^{\mu \nu}\beta_{\nu}-\zeta \widehat{j}^{\mu}\right)\right]\right)
\end{equation}
where the partition function $Z_{\rm LE}$ ensures $\mathrm{tr}(\widehat{\rho}_{\rm LE})=1$.
It is important to notice that it can be kept in this simple form, without subtraction of the vacuum expectation value, because the latter is a non-operator term which would appear also in the partition function, hence cancelling out in the ratio.
In formulae
\begin{equation}\label{zubeq24}
\begin{split}
&\frac{\exp \left[-\int_{\Sigma(\tau)}{\rm d} \Sigma_{\mu}\left(:\widehat{T}^{\mu \nu}:\beta_{\nu}-\zeta :\widehat{j}^{\mu}:\right)\right]}{\mathrm{tr} \left(\exp \left[-\int_{\Sigma(\tau)}{\rm d} \Sigma_{\mu}\left(:\widehat{T}^{\mu \nu}:\beta_{\nu}-\zeta :\widehat{j}^{\mu}:\right)\right]\right)}\\
&=\frac{\exp \left[-\int_{\Sigma(\tau)}{\rm d} \Sigma_{\mu}\left((\widehat{T}^{\mu \nu}-\langle 0|\widehat{T}^{\mu \nu}|0\rangle)\beta_{\nu}-\zeta (\widehat{j}^{\mu}-\langle 0|\widehat{j}^{\mu}|0\rangle)\right)\right]}{\mathrm{tr} \left(\exp \left[-\int_{\Sigma(\tau)}{\rm d} \Sigma_{\mu}\left((\widehat{T}^{\mu \nu}-\langle 0|\widehat{T}^{\mu \nu}|0\rangle)\beta_{\nu}-\zeta (\widehat{j}^{\mu}-\langle 0|\widehat{j}^{\mu}|0\rangle)\right)\right]\right)}\\
&=\frac{\exp \left[-\int_{\Sigma(\tau)}{\rm d} \Sigma_{\mu}\left(\widehat{T}^{\mu \nu}\beta_{\nu}-\zeta \widehat{j}^{\mu}\right)\right]}{\mathrm{tr} \left(\exp \left[-\int_{\Sigma(\tau)}{\rm d} \Sigma_{\mu}\left(\widehat{T}^{\mu \nu}\beta_{\nu}-\zeta \widehat{j}^{\mu}\right)\right]\right)}
=\widehat{\rho}_{\rm LE}.
\end{split}
\end{equation}
In other words, $\widehat{\rho}_{\rm LE}$ is invariant under addition or subtraction of a non-operator term such as a vacuum expectation value.
This property will come in handy later on.
Equation \eqref{zubeq05} is the fully covariant expression of the local thermodynamic equilibrium density operator, and it holds in a general curved spacetime.
It was first obtained by Zubarev \cite{Zubarev:1979} and van Weert \cite{Vanweert1982133}, and has been recently reworked \cite{Becattini:2014yxa, Hayata:2015lga}.
In order to enforce the constraint equations \eqref{zubeq04}, we introduced the Lagrange multipliers $\beta^{\mu}$ and $\zeta$: their physical meaning is that of thermodynamic fields, and they are determined as solutions of the constraint equations themselves with $\widehat{\rho}_{\rm LE}$ given by \eqref{zubeq05}.
Now, it is clear that the local thermodynamic equilibrium density operator depends first on the chosen foliation and then on the particular hypersurface $\Sigma(\tau)$, whence on the vector field $n^{\mu}$.
In turn, the thermal expectation values at left-and side in the constraint equations also depend on $\Sigma(\tau)$.
Thus, explicitly, the constraint equations read
\begin{equation}
n_{\mu}\mathrm{tr}(\widehat{\rho}_{\rm LE}[\beta^{\mu},\zeta,n^{\mu}]\widehat{T}^{\mu \nu})_{\rm ren}=n_{\mu}T^{\mu \nu},\qquad
n_{\mu}\mathrm{tr}(\widehat{\rho}_{\rm LE}[\beta^{\mu},\zeta,n^{\mu}]\widehat{j}^{\mu})_{\rm ren}=n_{\mu}j^{\mu}
\end{equation}
where the dependence of $\widehat{\rho}_{\rm LE}$ on $\beta^{\mu}$, $\zeta$ and $n^{\mu}$ can be functional.
For any given spacetime foliation, these are 5 equations in 5 unknowns, and can be solved in principle to determine the thermodynamic fields.
In the general case this is a non-trivial task, however a great simplification is achieved if $n^{\mu}$ is taken to be the direction of $\beta^{\mu}$, that is $n^{\mu}=\beta^{\mu}/\sqrt{\beta^2}$, which is possible as $\beta^{\mu}$ is known to be timelike, provided it satisfies the vorticity-free condition \eqref{zubeq07}.
Nevertheless, in \cite{Becattini:2014yxa} it is argued that a definition of $n^{\mu}$ based on $\beta^{\mu}$ can be put forward also in the vorticous case, at least in Minkowski spacetime.
With this choice, the number of independent variables is reduced, and the above equations become
\begin{equation}
\beta_{\mu}\mathrm{tr}(\widehat{\rho}_{\rm LE}[\beta^{\mu},\zeta]\widehat{T}^{\mu \nu})_{\rm ren}=\beta_{\mu}T^{\mu \nu},\qquad
\beta_{\mu}\mathrm{tr}(\widehat{\rho}_{\rm LE}[\beta^{\mu},\zeta]\widehat{j}^{\mu})_{\rm ren}=\beta_{\mu}j^{\mu}.
\end{equation}
Thus, although being true that there is an overall ambiguity due to the choice of the hypersurface $\Sigma(\tau)$, this can in fact be chosen consistently by using $\beta^{\mu}$ itself.
This choice is known as the $\beta$-\textit{frame}, or the \textit{thermodynamic frame}, represented by the velocity field defined as
\begin{equation}\label{zubeq11}
u^{\mu}\equiv \frac{\beta^{\mu}}{\sqrt{\beta^2}},\qquad
T\equiv \frac{1}{\sqrt{\beta^2}}.
\end{equation}
The Lorentz scalar $T$ defines a \textit{proper temperature}, namely the one measured by an ideal thermometer comoving with the system, and is in general different from the one marked by a fixed ideal thermometer who sees the system passing by, that is $T_{\rm fix}=1/\beta^0$.
By virtue of this interpretation, $\beta^{\mu}$ is called the \textit{four-temperature} vector field.
Finally, $\zeta$, sometimes called the \textit{fugacity}, is the ratio between the chemical potential $\mu$ associated to the charged current $\widehat{j}^{\mu}$ and the proper temperature, $\zeta \equiv \mu/T$, as we saw in Chapter \ref{chapter:relhydro}.
Much more could actually be said about the $\beta$-frame and its properties, but, to all intents and purposes of this work, we would like the takeaway to be just the following.
First of all, it is true that the local thermodynamic equilibrium density operator depends on the choice of the foliation and then on the spacelike hypersurface onto which the densities are given.
Actually, this is not completely new, as in the usual thermal field theory it is just as well necessary to assign initial values on a Cauchy hypersurface.
However, the four-temperature can in fact help fixing this ambiguity.
And second, while in relativistic hydrodynamics the field $\beta^{\mu}$ is defined starting from the velocity field and the temperature, here we showed that the other way round is equally possible, namely we regarded the four-temperature as a fundamental quantity from which the velocity field and proper temperature were derived \cite{Van:2013sma}.
For a thorougher discussion of the $\beta$-frame and the reasons why it should be regarded as a privileged frame, see \cite{Becattini:2014yxa}.
\subsection{From local equilibrium to global and non-equilibrium}
\label{sec:LE_to_GE_and_NE}
The key observation we now want to point out is that the local thermodynamic equilibrium density operator \eqref{zubeq05} is not stationary in general.
The reason is that the quantum operators $\widehat{T}^{\mu \nu}$ and $\widehat{j}^{\mu}$ appearing in it are built from the quantum fields, which are $\tau$-dependent, hence so is $\widehat{\rho}_{\rm LE}$ through the choice of the hypersurface $\Sigma(\tau)$.
This implies that $\widehat{\rho}_{\rm LE}$ cannot be a state of the system, for it is not in the Heisenberg picture.
Recall that it was worked out starting from the quantum operators instead of the microscopic states, therefore it cannot be expressed in general in the form \eqref{zubeq08} as a state would.
The actual, stationary state in the Heisenberg picture $\widehat{\rho}$ remains unknown, however $\widehat{\rho}_{\rm LE}$ contains information about it in the thermodynamic fields $\beta^{\mu}$ and $\zeta$.
In fact, at the right-hand side of the constraint equations \eqref{zubeq04}, which define the thermodynamic fields, there are the actual densities determined as renormalized expectation values calculated with the actual state $\widehat{\rho}$.
The non-stationarity of $\widehat{\rho}_{\rm LE}$ is a property so important that is convenient emphasizing by making it explicit, therefore from now on we will write $\widehat{\rho}_{\rm LE}(\tau)$.
If the local thermodynamic equilibrium density operator is not even a state of the system, why did we put so much effort in defining it?
Are we missing anything?
The answer, which is an amendment of Zubarev's original idea, is overly simple: if at some initial $\tau_0$ the system is known to be at local thermodynamic equilibrium, the actual state is $\widehat{\rho}=\widehat{\rho}_{\rm LE}(\tau_0)$, and it remains such at any $\tau$ by virtue of stationarity.
In other words, the actual state is $\widehat{\rho}$ at any time, but if $\widehat{\rho}_{\rm LE}(\tau)$ is designed in such a way that at some $\tau_0$ we have $\widehat{\rho}_{\rm LE}(\tau_0)=\widehat{\rho}$, then the system is said to be at local thermodynamic equilibrium at $\tau_0$.
That is to say that the state still depends on the choice of the foliation, but the only hypersurface it really depends on is the one where the initial data, namely the energy-momentum and charge densities, is given, that is $\Sigma(\tau_0)$.
Note that this hypothesis does not discriminate between thermodynamic equilibrium and non-equilibrium, the actual state we are referring to can be either.
Note also that, although $\widehat{\rho}$ might be equal to $\widehat{\rho}_{\rm LE}(\tau_0)$ at some $\tau_0$, their derivatives are different: the derivatives of $\widehat{\rho}$ vanish identically because of stationarity, while those of $\widehat{\rho}_{\rm LE}(\tau_0)$ do not in principle.
Then, how do we check if a system is at local thermodynamic equilibrium at a given $\tau_0$?
The local thermodynamic equilibrium density operator is determined by fixing the expectation values of energy-momentum and charge densities equal to the actual ones.
The expectation value is a first moment.
We could consider, for instance, the second moments, namely the variances, and check if they agree with the measured values within the experimental errors at $\tau_0$.
If they do, we still do not know what the actual state of system is, but it is reasonable to assume it to be $\widehat{\rho}_{\rm LE}(\tau_0)$, which becomes more and more legitimate as one checks moments of higher and higher order.
This procedure can be far from easy to do in practice, but there is no conceptual fallacy in principle.
If at $\tau_0$ the system is at local thermodynamic equilibrium in the above sense, we can obtain the relation between $\widehat{\rho}$ and $\widehat{\rho}_{\rm LE}(\tau)$ by means of Gauss' theorem.
Let us consider a volume $\Omega$ of spacetime enclosed by two spacelike hypersurfaces of the foliation: one $\Sigma(\tau_0)$ at the initial $\tau_0$ and the other $\Sigma(\tau)$ at the ``present'' $\tau$.
Let also $\Gamma(\tau_0,\tau)$ be the timelike boundary at infinity joining $\Sigma(\tau_0)$ and $\Sigma(\tau)$.
This is configuration is shown in Figure \ref{fig01}.
\begin{figure}
\begin{center}
\includegraphics[width=10cm]{figures/foliation.eps}
\caption{Here it is shown the spacetime volume $\Omega$ enclosed between the two spacelike hypersurfaces of the foliation $\Sigma(\tau_0)$ and $\Sigma(\tau)$, where $\tau_0$ is the time where the system is at local thermodynamic equilibrium.
The timelike boundary at infinity joining them is indicated as $\Gamma(\tau_0,\tau)$, and $n^{\mu}$ is the timelike unit vector field orthogonal to the foliation.}
\label{fig01}
\end{center}
\end{figure}
Assuming that the quantum fields the conserved currents are built with satisfy boundary conditions such that the flux through $\Gamma(\tau_0,\tau)$ vanishes \cite{Becattini:2012tc}, by Gauss' theorem we have
\begin{equation}
\int_{\Omega}{\rm d}\Omega \left(\widehat{T}^{\mu \nu}\nabla_{\mu}\beta_{\nu}-\widehat{j}^{\mu}\nabla_{\mu}\zeta \right)=\int_{\Sigma(\tau)}{\rm d}\Sigma_{\mu}\left(\widehat{T}^{\mu \nu}\beta_{\nu}-\zeta \widehat{j}^{\mu}\right)-\int_{\Sigma(\tau_0)}{\rm d}\Sigma_{\mu}\left(\widehat{T}^{\mu \nu}\beta_{\nu}-\zeta \widehat{j}^{\mu}\right),
\end{equation}
where the conservation of the energy-momentum tensor and charged current operators were used.
Thus, the actual state is
\begin{equation}\label{zubeq09}
\begin{split}
\widehat{\rho}=&\widehat{\rho}_{\rm LE}(\tau_0)=
\frac{1}{Z_{\rm LE}(\tau_0)}\exp \left[-\int_{\Sigma(\tau_0)}{\rm d}\Sigma_{\mu}\left(\widehat{T}^{\mu \nu}\beta_{\nu}-\zeta \widehat{j}^{\mu}\right)\right]\\
=&
\frac{1}{Z_{\rm LE}(\tau_0)}\exp \left[-\int_{\Sigma(\tau)}{\rm d}\Sigma_{\mu}\left(\widehat{T}^{\mu \nu}\beta_{\nu}-\zeta \widehat{j}^{\mu}\right)+\int_{\Omega}{\rm d}\Omega \left(\widehat{T}^{\mu \nu}\nabla_{\mu}\beta_{\nu}-\widehat{j}^{\mu}\nabla_{\mu}\zeta \right)\right],
\end{split}
\end{equation}
where in the first line it is understood that not only the hypersurface, but the argument of the integral too are evaluated at $\tau_0$.
Note that the first integral in the second line is the one appearing in $\widehat{\rho}_{\rm LE}(\tau)$.
Although hardly useful at a practical level, this expression is exact and represents the actual, stationary state of the system, whether equilibrium or non-equilibrium.
However, at thermodynamic equilibrium we can take it a step further.
At thermodynamic equilibrium, the entropy is maximized with constrained values of energy-momentum and charge densities at any $\tau$.
Since $\widehat{\rho}_{\rm LE}(\tau)$ is also obtained by maximizing the entropy, thermodynamic equilibrium is realized if $\widehat{\rho}_{\rm LE}(\tau)=\widehat{\rho}_{\rm LE}(\tau')$ for any $\tau$ and $\tau'$.
This implies that the volume term in \eqref{zubeq09} must vanish, which happens for any volume $\Omega$ if the argument of the integral vanishes
\begin{equation}
\widehat{T}^{\mu \nu}\nabla_{\mu}\beta_{\nu}-\widehat{j}^{\mu}\nabla_{\mu}\zeta=\frac{1}{2}\widehat{T}^{\mu \nu}\left(\nabla_{\mu}\beta_{\nu}+\nabla_{\nu}\beta_{\mu}\right)-\widehat{j}^{\mu}\nabla_{\mu}\zeta=0,
\end{equation}
where we used the symmetry of the energy-momentum tensor operator.
Being the thermodynamic fields $\beta^{\mu}$ and $\zeta$ independent from each other, the condition for thermodynamic equilibrium is
\begin{equation}\label{zubeq10}
\nabla_{\mu}\beta_{\nu}+\nabla_{\nu}\beta_{\mu}=0,\qquad
\nabla_{\mu}\zeta=0
\end{equation}
namely $\zeta$ must be constant and $\beta^{\mu}$ a Killing vector \cite{Chrobok:2006rr}.
Taking into account for the definition \eqref{zubeq11} of the four-temperature, this is exactly the same result as \eqref{eqrelhydro22}.
However, the latter was obtained in the context of the Navier-Stokes theory, which is a first-order theory, while \eqref{zubeq10} was obtained in the Zubarev theory, which is not truncated to any order in the derivatives of the thermodynamic fields.
This particular stationary configuration is called a \textit{global thermodynamic equilibrium} state
\begin{equation}\label{zubeq28}
\widehat{\rho}=\frac{1}{Z}\exp \left[-\int_{\Sigma}{\rm d}\Sigma_{\mu}\left(\widehat{T}^{\mu \nu}\beta_{\nu}-\zeta \widehat{j}^{\mu}\right)\right],
\end{equation}
\begin{equation}
Z=\mathrm{tr} \left(\exp \left[-\int_{\Sigma}{\rm d}\Sigma_{\mu}\left(\widehat{T}^{\mu \nu}\beta_{\nu}-\zeta \widehat{j}^{\mu}\right)\right]\right).
\end{equation}
The $\tau$-dependence is omitted, as in fact there is none being $\widehat{\rho}$ a stationary state in the Heisenberg picture.
The above expression is formally analogous to the one of the local thermodynamic equilibrium density operator \eqref{zubeq05}, with the important difference that it is an actual state and the thermodynamic fields fulfill the conditions \eqref{zubeq10}.
This means that we have much more information on a system when it is at global thermodynamic equilibrium than out of equilibrium, for in the former case the thermodynamic fields are bounded to be a Killing vector and a constant, whereas in the latter they can be whatever as long as they are solutions of the constraint equations.
This is one of the reasons that make systems at local thermodynamic equilibrium harder to study, as we will have the chance to convince ourselves in the next Chapters.
However, not all is lost if local thermodynamic equilibrium proves to be too hard to deal with.
We will not go into details here, but in \cite{Becattini:2014yxa} it is shown, at least in Minkowski spacetime, that in the $\beta$-frame $\widehat{\rho}_{\rm LE}(\tau)$ can be expanded in the derivatives of the thermodynamic fields by using linear response theory around a configuration of global thermodynamic equilibrium.
The idea is that, if at local thermodynamic equilibrium the derivatives of the thermodynamic fields are small compared to their global thermodynamic equilibrium values, the so-called \textit{hydrodynamic limit}, the main contribution to thermal expectation values comes from global thermodynamic equilibrium.
This is what makes global thermodynamic equilibrium interesting even for systems at local thermodynamic equilibrium, as we shall see in the next Chapter.
We conclude this part with the following remark.
Suppose that the actual state $\widehat{\rho}=\widehat{\rho}_{\rm LE}(\tau_0)$ has some symmetry, namely it commutes with some unitary representation of subgroup ${\rm G}$ of transformations of the proper orthochronous Poincar\'e group ${\rm IO}(1,3)^{\uparrow}_+$.
Let it be $\widehat{U}(g)$ in the Hilbert space.
Then we have
\begin{equation}
\begin{split}
\widehat{U}(g)\widehat{\rho}\widehat{U}^{-1}(g)=&
\frac{1}{Z}\exp \left[-\int_{\Sigma(\tau_0)}{\rm d} \Sigma_{\mu}(x)\left(\widehat{U}(g)\widehat{T}^{\mu \nu}(x)\widehat{U}^{-1}(g)\beta_{\nu}(x)\right.\right.\\
&\left.\left.-\zeta(x)\widehat{U}(g)\widehat{j}^{\mu}(x)\widehat{U}^{-1}(g)\right)\right]\\
=&\frac{1}{Z}\exp \left[-\int_{\Sigma(\tau_0)}{\rm d} \Sigma_{\mu}(x)\left({D(g^{-1})^{\mu}}_{\rho}{D(g^{-1})^{\nu}}_{\sigma}\widehat{T}^{\rho \sigma}(g(x))\beta_{\nu}(x)\right.\right.\\
&\left.\left.-\zeta(x){D(g^{-1})^{\mu}}_{\rho}\widehat{j}^{\rho}(g(x))\right)\right].
\end{split}
\end{equation}
By setting $y\equiv g(x)$ and using the fact that ${\rm d} \Sigma_{\mu}(x)={D(g)^{\nu}}_{\mu}{\rm d} \Sigma_{\nu}(x)$ we get
\begin{equation}
\begin{split}
\widehat{U}(g)\widehat{\rho}\widehat{U}^{-1}(g)=&
\frac{1}{Z}\exp \left[-\int_{g(\Sigma(\tau_0))}{\rm d} \Sigma_{\rho}(y)\left(\widehat{T}^{\mu \nu}(y){D(g^{-1})^{\nu}}_{\sigma}\beta_{\nu}(g^{-1}(y))\right.\right.\\
&\left.\left.-\zeta(g^{-1}(y))\widehat{j}^{\rho}(y)\right)\right],
\end{split}
\end{equation}
so, if the hypersurface is invariant under the transformation $g$ and if
\begin{equation}\label{zubeq12}
{D(g^{-1})^{\nu}}_{\sigma}\beta_{\nu}(g^{-1}(y))=\beta_{\sigma}(y),\qquad
\zeta(g^{-1}(y))=\zeta(y),
\end{equation}
then the density operator is invariant under the unitary transformation, $\widehat{U}(g)\widehat{\rho}\widehat{U}^{-1}(g)=\widehat{\rho}$.
Equations \eqref{zubeq12} specify the symmetry conditions on the thermodynamic fields in order for $\widehat{\rho}$ to be invariant.
An invariance of $\widehat{\rho}$ has consequences on the thermal expectation values of operators.
For instance, for the energy-momentum tensor
\begin{equation}\label{zubeq13}
\begin{split}
T^{\mu \nu}=&
\mathrm{tr}(\widehat{\rho}\widehat{T}^{\mu \nu}(x))=
\mathrm{tr}(\widehat{\rho}\widehat{U}^{-1}(g)\widehat{T}^{\mu \nu}(x)\widehat{U}(g))\\
=&{D(g)^{\mu}}_{\rho}{D(g)^{\nu}}_{\sigma}\mathrm{tr}(\widehat{\rho}\widehat{T}^{\rho \sigma}(g^{-1}(x)))=
{D(g)^{\mu}}_{\rho}{D(g)^{\nu}}_{\sigma}T^{\rho \sigma}(g^{-1}(x)).
\end{split}
\end{equation}
If we consider a 1-parameter subgroup of transformations $g_{\phi}$, for instance the rotations around some axis, then, by equations \eqref{zubeq12} and \eqref{zubeq13}, the Lie derivative of the field under consideration along the vector field $X(x)={\rm d} g_{\phi}/{\rm d} \phi$ vanishes
\begin{equation}
{\cal L}_X(\beta)^{\mu}=0,\qquad
{\cal L}_X(T)^{\mu \nu}=0.
\end{equation}
Now the following question arises: if ${\rm G}$ is a symmetry subgroup for the actual state, is it so also for the local thermodynamic equilibrium density operator at any time?
In formulae, does the following implication hold at any $\tau$?
\begin{equation}
\widehat{U}(g)\widehat{\rho}\widehat{U}^{-1}(g)=\widehat{\rho}
\qquad \Rightarrow \qquad
\widehat{U}(g)\widehat{\rho}_{\rm LE}(\tau)\widehat{U}^{-1}(g)=\widehat{\rho}_{\rm LE}(\tau).
\end{equation}
It can be shown that if ${\rm G}$ transforms $\Sigma(\tau_0)$ into itself and the thermodynamic fields fulfill \eqref{zubeq12}, then the answer is yes.
Recall that, by definition, $\widehat{\rho}_{\rm LE}(\tau)$ is determined by maximizing the entropy while satisfying the constraint equations.
Clearly, both the entropy and the constraint equations are invariant under unitary transformations of $\widehat{\rho}_{\rm LE}(\tau)$, therefore, if $\widehat{\rho}_{\rm LE}(\tau)$ is a solution of the variational problem, so is $\widehat{U}(g)\widehat{\rho}_{\rm LE}(\tau)\widehat{U}^{-1}(g)$.
This implies that either $\widehat{U}(g)\widehat{\rho}_{\rm LE}(\tau)\widehat{U}^{-1}(g)$ coincides with $\widehat{\rho}_{\rm LE}(\tau)$, or it is an entirely different solution.
In either case, it is possible to generate a symmetric solution under the subgroup ${\rm G}$ by taking a particular solution $\widehat{\rho}_{\rm LE}^{(0)}(\tau)$ and summing over all the elements of ${\rm G}$
\begin{equation}
\widehat{\rho}_{\rm LE}(\tau)=\frac{1}{M({\rm G})}\sum_{g\in {\rm G}}\widehat{U}(g)\widehat{\rho}_{\rm LE}^{(0)}(\tau)\widehat{U}^{-1}(g),
\end{equation}
where $M({\rm G})$ is the volume of ${\rm G}$.
Thus, a sufficient condition for $\widehat{\rho}_{\rm LE}(\tau)$ to be symmetric under ${\rm G}$ at any $\tau$, is that the thermodynamic fields fulfill \eqref{zubeq12} at any $\tau$.
This will be important especially in Chapter \ref{chapter:boost}.
\subsection{Entropy production rate}
\label{sec:entr_prod_rate}
We conclude this part with an argument linking back to the end of the previous Section, namely the entropy production of irreversible quantum processes.
We will follow the steps of \cite{Becattini:2019dxo} with a few important adjustments concerning the renormalization.
As it should be clear by now, the local thermodynamic equilibrium density operator is not a state of the system for it is not stationary.
However, it still is an operator, so we can take traces of it and do all the operations we do on operators.
Thus, if we pretend it is a state and we treat it as such, the corresponding entropy is in fact $\tau$-dependent
\begin{equation}\label{zubeq15}
S(\tau)=-\mathrm{tr}(\widehat{\rho}_{\rm LE}(\tau)\log \widehat{\rho}_{\rm LE}(\tau)).
\end{equation}
That is because $\widehat{\rho}_{\rm LE}(\tau)$ is built from the quantum operators as in \eqref{zubeq05}, not the microscopic states as in \eqref{zubeq08}, therefore its $\tau$-dependence is governed by $\Sigma(\tau)$ instead of the unitary transformation \eqref{zubeq14}.
In other words, the invariance of the entropy under unitary transformations of the state will not prevent it from being $\tau$-dependent this time, because, as $\widehat{\rho}_{\rm LE}(\tau)$ is not a state, its time evolution is not controlled by a unitary transformation but by $\Sigma(\tau)$ instead.
Thus, up to more convincing proposals, equation \eqref{zubeq15} is conventionally taken as the definition of a non-stationary entropy, despite $\widehat{\rho}_{\rm LE}(\tau)$ not being a state.
The next step is finding an expression for the entropy production rate, that is ${\rm d} S(\tau)/{\rm d} \tau$.
Recall that, by definition, the entropy \eqref{zubeq15} can be expressed in the form of \eqref{eqrelhydro19} as
\begin{equation}
S(\tau)=\int_{\Sigma(\tau)}{\rm d} \Sigma_{\mu}s^{\mu}.
\end{equation}
Let us emphasize that out of equilibrium, since $\nabla_{\mu}s^{\mu}>0$, the entropy is a frame-dependent quantity as it varies with the integration hypersurface $\Sigma(\tau)$.
The derivative of the entropy with respect to $\tau$ can be computed by exploiting a general formula for the variation of an integral between two infinitesimally close hypersurfaces, which we do not prove here
\begin{equation}\label{zubeq16}
\frac{{\rm d} S(\tau)}{{\rm d} \tau}=
\int_{\Sigma(\tau)}{\rm d} \Sigma_{\nu}U^{\nu}\nabla_{\mu}s^{\mu}+
\frac{1}{2}\int_{\partial \Sigma(\tau)}{\rm d} \tilde{S}_{\mu \nu}(s^{\mu}U^{\nu}-s^{\nu}U^{\mu}),
\end{equation}
where $\partial \Sigma(\tau)$ is the 2-dimensional boundary of $\Sigma(\tau)$, ${\rm d} \tilde{S}_{\mu \nu}=-\frac{1}{2}\epsilon_{\mu \nu \rho \sigma}{\rm d} S^{\rho \sigma}$ is the dual of the measure ${\rm d} S^{\rho \sigma}\equiv {\rm d} x^{\rho}\wedge {\rm d} x^{\sigma}$ and $U^{\mu}\equiv {\rm d} x^{\mu}/{\rm d} \tau$ \cite{Becattini:2019dxo}.
Assuming that the boundary term vanishes, we are left with
\begin{equation}\label{zubeq18}
\frac{{\rm d} S(\tau)}{{\rm d} \tau}=
\int_{\Sigma(\tau)}{\rm d} \Sigma_{\nu}U^{\nu}\nabla_{\mu}s^{\mu}.
\end{equation}
Now we want to repeat the same steps with \eqref{zubeq15} and compare with the above result.
As we are going to exploit the invariance of $\widehat{\rho}_{\rm LE}(\tau)$ under the subtraction of a vacuum expectation value, let us first define out of convenience the following notation
\begin{equation}
:Z_{\rm LE}(\tau):\,\equiv \mathrm{tr} \left(\exp \left[-\int_{\Sigma}{\rm d} \Sigma_{\mu}\left(:\widehat{T}^{\mu \nu}:\beta_{\nu}-\zeta :\widehat{j}^{\mu}:\right)\right]\right).
\end{equation}
By using the expression \eqref{zubeq05} of $\widehat{\rho}_{\rm LE}(\tau)$ and the above mentioned invariance, we find
\begin{equation}
\begin{split}
S(\tau)=&
-\mathrm{tr} \left(\widehat{\rho}_{\rm LE}(\tau)\log \frac{\exp \left[-\int_{\Sigma(\tau)}{\rm d} \Sigma_{\mu}\left(\widehat{T}^{\mu \nu}\beta_{\nu}-\zeta \widehat{j}^{\mu}\right)\right]}{Z_{\rm LE}(\tau)}\right)\\
=&-\mathrm{tr} \left(\widehat{\rho}_{\rm LE}(\tau)\log \frac{\exp \left[-\int_{\Sigma(\tau)}{\rm d} \Sigma_{\mu}\left(:\widehat{T}^{\mu \nu}:\beta_{\nu}-\zeta :\widehat{j}^{\mu}:\right)\right]}{:Z_{\rm LE}(\tau):}\right)\\
=&\log :Z_{\rm LE}(\tau):+\mathrm{tr} \left(\widehat{\rho}_{\rm LE}(\tau)\int_{\Sigma(\tau)}{\rm d} \Sigma_{\mu}\left(:\widehat{T}^{\mu \nu}:\beta_{\nu}-\zeta :\widehat{j}^{\mu}:\right)\right)\\
=&\log :Z_{\rm LE}(\tau):+\int_{\Sigma(\tau)}{\rm d} \Sigma_{\mu}\left(\langle :\widehat{T}^{\mu \nu}:\rangle_{\rm LE}\beta_{\nu}-\zeta \langle :\widehat{j}^{\mu}:\rangle_{\rm LE}\right)\\
=&\log :Z_{\rm LE}(\tau):+\int_{\Sigma(\tau)}{\rm d} \Sigma_{\mu}\left(T^{\mu \nu}\beta_{\nu}-\zeta j^{\mu}\right).
\end{split}
\end{equation}
In the last step, we used the definition ${\rm d} \Sigma_{\mu}={\rm d} \Sigma \,n_{\mu}$ and the constraint equations \eqref{zubeq04} to replace the renormalized energy-momentum and charge densities at local thermodynamic equilibrium with the actual ones.
By deriving with respect to $\tau$, using the formula \eqref{zubeq16} and assuming once again that the boundary term vanishes, we obtain
\begin{equation}\label{zubeq17}
\frac{{\rm d} S(\tau)}{{\rm d} \tau}=\frac{{\rm d} \log :Z_{\rm LE}(\tau):}{{\rm d} \tau}+\int_{\Sigma(\tau)}{\rm d} \Sigma_{\lambda}U^{\lambda}\left(T^{\mu \nu}\nabla_{\mu}\beta_{\nu}-j^{\mu}\nabla_{\mu}\zeta \right),
\end{equation}
where the conservation law of the energy-momentum tensor and charged current, namely the hydrodynamic equations \eqref{eqrelhydro02} and \eqref{eqrelhydro03}, were also used.
The derivative of the logarithm of the partition function requires a little more effort.
Let us start by defining the operator $\widehat{\Upsilon}(\tau)$, which will appear again later on, as
\begin{equation}\label{zubeq23}
Z_{\rm LE}(\tau)=\mathrm{tr} \left({\rm e}^{-\widehat{\Upsilon}(\tau)}\right),\qquad
\widehat{\Upsilon}(\tau)\equiv \int_{\Sigma(\tau)}{\rm d} \Sigma_{\mu}\left(\widehat{T}^{\mu \nu}\beta_{\nu}-\zeta \widehat{j}^{\mu}\right),
\end{equation}
hence the derivative
\begin{equation}
\begin{split}
\frac{{\rm d} \log :Z_{\rm LE}(\tau):}{{\rm d} \tau}=&
\frac{1}{:Z_{\rm LE}(\tau):}\frac{{\rm d} :Z_{\rm LE}(\tau):}{{\rm d} \tau}=
\frac{1}{:Z_{\rm LE}(\tau):}\frac{{\rm d}}{{\rm d} \tau}\mathrm{tr} \left({\rm e}^{-:\widehat{\Upsilon}(\tau):}\right)\\
=&-\frac{1}{:Z_{\rm LE}(\tau):}\mathrm{tr} \left({\rm e}^{-:\widehat{\Upsilon}(\tau):}\frac{:\widehat{\Upsilon}(\tau):}{{\rm d} \tau}\right)=
-\left\langle \frac{{\rm d} :\widehat{\Upsilon}(\tau):}{{\rm d} \tau}\right\rangle_{\rm LE}.
\end{split}
\end{equation}
We use again the formula \eqref{zubeq16} assuming that the boundary term vanishes, and with the conservation laws of the energy-momentum tensor and charged current operators we get
\begin{equation}
\frac{{\rm d} :\widehat{\Upsilon}(\tau):}{{\rm d} \tau}=
\int_{\Sigma(\tau)}{\rm d} \Sigma_{\lambda}\left(:\widehat{T}^{\mu \nu}:\nabla_{\mu}\beta_{\nu}-:\widehat{j}^{\mu}:\nabla_{\mu}\zeta \right),
\end{equation}
hence
\begin{equation}
\frac{{\rm d} \log :Z_{\rm LE}(\tau):}{{\rm d} \tau}=-\int_{\Sigma(\tau)}{\rm d} \Sigma_{\lambda}U^{\lambda}\left(\langle :\widehat{T}^{\mu \nu}:\rangle_{\rm LE}\nabla_{\mu}\beta_{\nu}-\langle :\widehat{j}^{\mu}:\rangle_{\rm LE}\nabla_{\mu}\zeta \right),
\end{equation}
and plugging into \eqref{zubeq17} we finally obtain
\begin{equation}
\frac{{\rm d} S(\tau)}{{\rm d} \tau}=
\int_{\Sigma(\tau)}{\rm d} \Sigma_{\lambda}U^{\lambda}\left[\left(T^{\mu \nu}-\langle :\widehat{T}^{\mu \nu}:\rangle_{\rm LE}\right)\nabla_{\mu}\beta_{\nu}-\left(j^{\mu}-\langle :\widehat{j}^{\mu}:\rangle_{\rm LE}\right)\nabla_{\mu}\zeta \right].
\end{equation}
Comparing with \eqref{zubeq18} and taking into account that the equation should hold at any $\tau$, we finally obtain the following result for the entropy production rate
\begin{equation}\label{zubeq30}
\nabla_{\mu}s^{\mu}=\left(T^{\mu \nu}-\langle :\widehat{T}^{\mu \nu}:\rangle_{\rm LE}\right)\nabla_{\mu}\beta_{\nu}-\left(j^{\mu}-\langle :\widehat{j}^{\mu}:\rangle_{\rm LE}\right)\nabla_{\mu}\zeta,
\end{equation}
which was first found in \cite{Zubarev:1979} and later reworked in \cite{Becattini:2019dxo}.
As we can see, the entropy production rate depends on the deviation of the actual values of the conserved currents from those at local thermodynamic equilibrium and, as expected, on the derivatives of the thermodynamic fields, which essentially are the dissipations.
By the second law of thermodynamics, the above quantity is bounded to always be non-negative.
At global thermodynamic equilibrium the thermodynamic fields fulfill the conditions \eqref{zubeq10}, so the entropy production rate vanishes as expected.
It is important to emphasize that equation \eqref{zubeq30} represents the entropy production rate at local thermodynamic equilibrium, which is the underlying assumption of relativistic hydrodynamics and contains global thermodynamic equilibrium as a special case.
However, no analogous formula is known for the entropy production rate or relativistic quantum fluids fully out of thermodynamic equilibrium.
Let us seize the opportunity of having mentioned the divergencelessness of the entropy current at global thermodynamic equilibrium to make the following geometrical remark.
Being $\nabla_{\mu}s^{\mu}=0$, the entropy current can be expressed as the Hodge dual of an exact 3-form, which, if the domain is topologically contractible, can in turn be written as the exterior derivative of a 2-form.
This eventually amounts to state that the original entropy current can be cast into the divergence of an antisymmetric tensor field
\begin{equation}
\nabla_{\mu}s^{\mu}=0\qquad \Rightarrow \qquad
s^{\mu}=\nabla_{\nu}\varsigma^{\mu \nu},\qquad
\varsigma^{\mu \nu}=-\varsigma^{\nu \mu}.
\end{equation}
We call this $\varsigma^{\mu \nu}$ a \textit{potential} of the entropy current.
Whence, by using Stoke's theorem
\begin{equation}\label{zubeq31}
S=\int_{\Sigma}{\rm d} \Sigma_{\mu}s^{\mu}=
\int_{\Sigma}{\rm d} \Sigma_{\mu}\nabla_{\nu}\varsigma^{\mu \nu}=
\frac{1}{2}\int_{\partial \Sigma}{\rm d} \tilde{S}_{\mu \nu}\varsigma^{\mu \nu}=
-\frac{1}{4}\int_{\partial \Sigma}{\rm d} S^{\rho \sigma}\sqrt{|g|}\epsilon_{\mu \nu \rho \sigma}\varsigma^{\mu \nu},
\end{equation}
where $\partial \Sigma$ is the 2-dimensional boundary of $\Sigma$, ${\rm d} S^{\rho \sigma}={\rm d} x^{\rho}\wedge {\rm d} x^{\sigma}$ is the measure on $\partial \Sigma$, and ${\rm d} \tilde{S}_{\mu \nu}=-\frac{1}{2}\sqrt{|g|}\epsilon_{\mu \nu \rho \sigma}{\rm d} S^{\rho \sigma}$ is its dual.
In other words, the entropy can be expressed not only as a 3-dimensional integral on $\Sigma$, but also as a 2-dimensional integral on the boundary of $\Sigma$.
This fact is known in literature \cite{Wald:1993nt, Eling:2012xa, Majhi:2011ws, Majhi:2012tf} and has been reworked in \cite{Becattini:2019poj} including a practical example of calculation, which we will present in Subsection \ref{subsec:gte_entropy}.
Although elegantly concise and quite straightforward in its meaning, equation \eqref{zubeq30} can only provide us information on the divergence of the entropy current, not the entropy current itself.
In fact, in principle, there are infinitely many different entropy currents sharing the same divergence.
But as we saw in the last Chapter, the structure of the entropy current is crucial in defining the nature of the relativistic hydrodynamic theory, so it would be far more meaningful to know the expression of the entropy current instead of its divergence.
In the next Section we will put forward a general method to derive the entropy current directly in terms of the conserved currents.
\section{The entropy current method}
\label{sec:zubarev_entropy_current_method}
In this Section we retrace the steps of \cite{Becattini:2019poj}, where we determined, to the best of our knowledge, the first proof of existence of the entropy current at local thermodynamic equilibrium and put forward a general method to calculate it.
The crucial point will be the fact that the logarithm of the partition function is also extensive, meaning that it can be expressed as an integral of a vector field on a hypersurface of the foliation, which is usually assumed without proof.
Let us start our journey by noticing that the entropy straight out of the local thermodynamic equilibrium density operator \eqref{zubeq05} reads
\begin{equation}
S(\tau)=\log Z_{\rm LE}(\tau)+\int_{\Sigma(\tau)}{\rm d} \Sigma_{\mu}\left(\langle \widehat{T}^{\mu \nu}\rangle_{\rm LE}\beta_{\nu}-\zeta \langle \widehat{j}^{\mu}\rangle_{\rm LE}\right).
\end{equation}
By definition of the entropy current \eqref{eqrelhydro19}, this is also equal to
\begin{equation}
S(\tau)=\int_{\Sigma(\tau)}{\rm d} \Sigma_{\mu}s^{\mu}.
\end{equation}
We readily understand that, in order for the entropy to be extensive and so for the entropy current to exist, the logarithm of the partition function must also be extensive, namely there must be a vector field $\phi^{\mu}$, called the \textit{thermodynamic potential current}, such that
\begin{equation}
\log Z_{\rm LE}(\tau)\equiv \int_{\Sigma(\tau)}{\rm d} \Sigma_{\mu}\phi^{\mu}.
\end{equation}
This way, we have the following expression for the entropy current
\begin{equation}
s^{\mu}=\phi^{\mu}+\langle \widehat{T}^{\mu \nu}\rangle_{\rm LE}\beta_{\nu}-\zeta \langle \widehat{j}^{\mu}\rangle_{\rm LE}.
\end{equation}
The existence of the thermodynamic potential current and, in turn, of the entropy current is usually assumed.
This was made clear in Chapter \ref{chapter:relhydro}, when the entropy current was postulated in order to express the second law of thermodynamics in a local covariant fashion so that it could be enforced by relativistic hydrodynamics.
In this Section we provide the first general proof of this typically understood hypothesis and we put forward a method to calculate the thermodynamic potential current, whence the entropy current, at local thermodynamic equilibrium.
The first step is a modification of the local thermodynamic equilibrium density operator by the introduction of a dimensionless parameter $\lambda$ in the following way
\begin{equation}\label{zubeq21}
\widehat{\rho}_{\rm LE}(\tau,\lambda)\equiv
\frac{1}{Z_{\rm LE}(\tau,\lambda)}\exp \left[-\lambda \int_{\Sigma(\tau)}{\rm d} \Sigma_{\mu}\left(\widehat{T}^{\mu \nu}\beta_{\nu}-\zeta \widehat{j}^{\mu}\right)\right],
\end{equation}
\begin{equation}\label{zubeq22}
Z_{\rm LE}(\tau,\lambda)\equiv
\mathrm{tr} \left(\exp \left[-\lambda \int_{\Sigma(\tau)}{\rm d} \Sigma_{\mu}\left(\widehat{T}^{\mu \nu}\beta_{\nu}-\zeta \widehat{j}^{\mu}\right)\right]\right).
\end{equation}
This is purposely designed in such a way to recover the original density operator \eqref{zubeq05} and partition function \eqref{zubeq06} for $\lambda=1$.
By taking the derivative with respect to $\lambda$ of the logarithm of this new $\lambda$-dependent partition function, we obtain
\begin{equation}\label{zubeq20}
\frac{\partial \log Z_{\rm LE}(\tau,\lambda)}{\partial \lambda}=
-\int_{\Sigma(\tau)}{\rm d} \Sigma_{\mu}\left(\langle \widehat{T}^{\mu \nu}\rangle_{\rm LE}(\lambda)\beta_{\nu}-\zeta \langle \widehat{j}^{\mu}\rangle_{\rm LE}(\lambda)\right),
\end{equation}
where by the symbol $\langle \widehat{O}\rangle_{\rm LE}(\lambda)$ we mean the expectation value of $\widehat{O}$ calculated with $\widehat{\rho}_{\rm LE}(\tau,\lambda)$, that is $\langle \widehat{O}\rangle_{\rm LE}(\lambda)\equiv \mathrm{tr}(\widehat{\rho}_{\rm LE}(\tau,\lambda)\widehat{O})$.
Note that the $\lambda$-dependence affects only the expectation values and not the thermodynamic fields, for it stems from the density operator with whom the expectation values are calculated.
If the above result looks confusing at a glance, let us just point out that it is but the same as the following known formula, which is obtained from \eqref{zubeq03} and can be easily checked
\begin{equation}
Z=\mathrm{tr} \left(\exp \left[-\sum_i\lambda_i\widehat{O}_i\right]\right)
\qquad \Rightarrow \qquad
\frac{\partial \log Z}{\partial \lambda_k}=-\langle \widehat{O}_k\rangle.
\end{equation}
Let us now integrate both sides of \eqref{zubeq20} with respect to $\lambda$ from some $\lambda_0$ to $\lambda=1$
\begin{equation}
\log Z_{\rm LE}(\tau)-\log Z_{\rm LE}(\tau,\lambda_0)=-\int_{\lambda_0}^1{\rm d} \lambda \int_{\Sigma(\tau)}{\rm d} \Sigma_{\mu}\left(\langle \widehat{T}^{\mu \nu}\rangle_{\rm LE}(\lambda)\beta_{\nu}-\zeta \langle \widehat{j}^{\mu}\rangle_{\rm LE}(\lambda)\right),
\end{equation}
where we used the feature that the $\lambda$-modification is so that $\log Z_{\rm LE}(\tau,\lambda=1)=\log Z_{\rm LE}(\tau)$.
If we can exchange the order of the $\lambda$-integration and the $\Sigma(\tau)$ one, we have
\begin{equation}
\log Z_{\rm LE}(\tau)-\log Z_{\rm LE}(\tau,\lambda_0)=\int_{\Sigma(\tau)}{\rm d} \Sigma_{\mu} \int_1^{\lambda_0}{\rm d} \lambda \left(\langle \widehat{T}^{\mu \nu}\rangle_{\rm LE}(\lambda)\beta_{\nu}-\zeta \langle \widehat{j}^{\mu}\rangle_{\rm LE}(\lambda)\right).
\end{equation}
Thus, if there exists a particular $\lambda_0$ such that $\log Z_{\rm LE}(\tau,\lambda_0)=0$, we have both the proof of the extensivity of the logarithm of the partition function and a formula for the thermodynamic potential current
\begin{equation}
\log Z_{\rm LE}(\tau)=\int_{\Sigma(\tau)}{\rm d} \Sigma_{\mu}\phi^{\mu},\qquad
\phi^{\mu}=\int_1^{\lambda_0}{\rm d} \lambda \left(\langle \widehat{T}^{\mu \nu}\rangle_{\rm LE}(\lambda)\beta_{\nu}-\zeta \langle \widehat{j}^{\mu}\rangle_{\rm LE}(\lambda)\right).
\end{equation}
The key quantity for the determination of such a special $\lambda_0$ is the operator $\widehat{\Upsilon}(\tau)$ already defined in \eqref{zubeq23}, in particular its spectrum.
Suppose that $\widehat{\Upsilon}(\tau)$ is bounded from below, namely there exists a lowest eigenvalue $\Upsilon_0(\tau)$ and a corresponding eigenvector $|0_{\Upsilon}(\tau)\rangle$, and also let the latter be non-degenerate.
In the following, we will omit the $\tau$-dependence in the operator $\widehat{\Upsilon}(\tau)$, in the eigenvalues $\Upsilon_i(\tau)$ and in the eigenvector $|0_{\Upsilon}(\tau)\rangle$ in order to ease the notation and highlight the $\lambda$-dependence instead, but it should be kept in mind that they are all $\tau$-dependent in principle.
In this hypothesis, the eigenvalues can be arranged in ascending order as $\Upsilon_0<\Upsilon_1<\Upsilon_2<\cdots$ and the partition function can be written as
\begin{equation}\label{zubeq25}
Z_{\rm LE}(\tau,\lambda)={\rm e}^{-\lambda \Upsilon_0}\left(1+{\rm e}^{-\lambda(\Upsilon_1-\Upsilon_0)}+{\rm e}^{\lambda(\Upsilon_2-\Upsilon_0)}+\cdots \right).
\end{equation}
Then, if $\Upsilon_0=0$ and we let $\lambda \to +\infty$, we notice that
\begin{equation}
\lim_{\lambda \to +\infty}Z_{\rm LE}(\tau,\lambda)=1
\qquad \Rightarrow \qquad \lim_{\lambda \to +\infty}\log Z_{\rm LE}(\tau,\lambda)=0,
\end{equation}
therefore $\lambda_0=+\infty$ would be our candidate; however, we might expect $\Upsilon_0$ to be different from zero in general.
So, in order for these wheels to spin the way we want, we exploit the above mentioned property \eqref{zubeq24} of the local thermodynamic equilibrium density operator, namely the invariance under addition or subtraction of non-operator terms in the argument of the exponential function, provided the same is done in the partition function.
Recall that the entropy depends only on the density operator, therefore this is also an invariance of the entropy.
Thus, with the replacement
\begin{equation}
\widehat{\rho}_{\rm LE}(\tau,\lambda)=\frac{{\rm e}^{-\lambda \widehat{\Upsilon}(\tau)}}{Z_{\rm LE}(\tau,\lambda)}
\qquad \mapsto \qquad
\widehat{\rho}'_{\rm LE}(\tau,\lambda)\equiv \frac{{\rm e}^{-\lambda \widehat{\Upsilon}'(\tau)}}{Z'_{\rm LE}(\tau,\lambda)}
\end{equation}
where
\begin{equation}
\begin{split}
\widehat{\Upsilon}'\equiv&
\widehat{\Upsilon}-\Upsilon_0
=\widehat{\Upsilon}-\langle 0_{\Upsilon}|\widehat{\Upsilon}|0_{\Upsilon}\rangle \\
=&\int_{\Sigma(\tau)}{\rm d} \Sigma_{\mu}\left[\left(\widehat{T}^{\mu \nu}-\langle 0_{\Upsilon}|\widehat{T}^{\mu \nu}|0_{\Upsilon}\rangle \right)\beta_{\nu}-\zeta \left(\widehat{j}^{\mu}-\langle 0_{\Upsilon}|\widehat{j}^{\mu}|0_{\Upsilon}\rangle\right)\right],
\end{split}
\end{equation}
the entropy is in fact unchanged.
In other words, if this is useful to find an expression for the entropy current, we are free to make such a tweak and rest assured that the entropy stays the same.
The rationale behind this transformation is that the new operator $\widehat{\Upsilon}'$ has a vanishing lowest eigenvalue, $\Upsilon'_0=\Upsilon_0-\Upsilon_0=0$, therefore the new partition function
\begin{equation}
Z'_{\rm LE}(\tau,\lambda)=\mathrm{tr} \left(\exp \left[-\lambda \int_{\Sigma(\tau)}{\rm d} \Sigma_{\mu}\left[\left(\widehat{T}^{\mu \nu}-\langle 0_{\Upsilon}|\widehat{T}^{\mu \nu}|0_{\Upsilon}\rangle \right)\beta_{\nu}-\zeta \left(\widehat{j}^{\mu}-\langle 0_{\Upsilon}|\widehat{j}^{\mu}|0_{\Upsilon}\rangle\right)\right]\right]\right),
\end{equation}
which, like \eqref{zubeq25} and taking into account that $\Upsilon'_i=\Upsilon_i-\Upsilon_0$, can also be written as
\begin{equation}
Z'_{\rm LE}(\tau,\lambda)=1+{\rm e}^{-\lambda(\Upsilon_1-\Upsilon_0)}+{\rm e}^{-\lambda(\Upsilon_2-\Upsilon_0)}+\cdots,
\end{equation}
does have a vanishing logarithm for $\lambda \to +\infty$, therefore $\lambda_0=+\infty$ is indeed the sought solution.
So, by repeating the above steps with $Z_{\rm LE}(\tau,\lambda)$ replaced by $Z'_{\rm LE}(\tau,\lambda)$, the thermodynamic potential current and the entropy current read respectively
\begin{equation}\label{zubeq26}
\phi^{\mu}=\int_1^{+\infty}{\rm d} \lambda \left[\left(\langle \widehat{T}^{\mu \nu}\rangle_{\rm LE}-\langle 0_{\Upsilon}|\widehat{T}^{\mu \nu}|0_{\Upsilon}\rangle\right)(\lambda)\beta_{\nu}-\zeta \left(\langle \widehat{j}^{\mu}\rangle_{\rm LE}-\langle 0_{\Upsilon}|\widehat{j}^{\mu}|0_{\Upsilon}\rangle \right)(\lambda)\right],
\end{equation}
\begin{equation}\label{zubeq27}
s^{\mu}=\phi^{\mu}+\left(\langle \widehat{T}^{\mu \nu}\rangle_{\rm LE}-\langle 0_{\Upsilon}|\widehat{T}^{\mu \nu}|0_{\Upsilon}\rangle \right)\beta_{\nu}-\zeta \left(\langle \widehat{j}^{\mu}\rangle_{\rm LE}-\langle 0_{\Upsilon}|\widehat{j}^{\mu}|0_{\Upsilon}\rangle \right).
\end{equation}
In summary, a sufficient condition for the existence of the thermodynamic potential current and the entropy current at local thermodynamic equilibrium is that the operator $\widehat{\Upsilon}$ must be bounded from below with non-degenerate lowest eigenvalue.
In this case, the two ingredients necessary to calculate the entropy current are:
\begin{enumerate}
\item the expectation values of the conserved currents at local thermodynamic equilibrium, i.e.\ $\langle \widehat{T}^{\mu \nu}\rangle_{\rm LE}$ and $\langle \widehat{j}^{\mu}\rangle_{\rm LE}$, and
\item the eigenvector $|0_{\Upsilon}\rangle$ corresponding to the lowest eigenvalue $\Upsilon_0$ of the operator $\widehat{\Upsilon}$.
\end{enumerate}
The quantum operators of the conserved currents $\widehat{T}^{\mu \nu}$ and $\widehat{j}^{\mu}$ are built by using the quantum field expansion, which in turn is obtained by solving the quantum field equations of motion of the Quantum Field Theory at hand, so the preliminary step of the method is to do that first.
Then, the algorithm for the calculation is the following:
\begin{enumerate}
\item Take the expectation values at local thermodynamic equilibrium $\langle \widehat{T}^{\mu \nu}\rangle_{\rm LE}$ and $\langle \widehat{j}^{\mu}\rangle_{\rm LE}$ and subtract from them the expectation values in $|0_{\Upsilon}\rangle$, namely $\langle 0_{\Upsilon}|\widehat{T}^{\mu \nu}|0_{\Upsilon}\rangle$ and $\langle 0_{\Upsilon}|\widehat{j}^{\mu}|0_{\Upsilon}\rangle$.
This must be done using $\widehat{\rho}_{\rm LE}(\tau,\lambda)$, so the result will be $\lambda$-dependent.
\item Multiply by the corresponding thermodynamic fields, which are not $\lambda$-dependent, and integrate with respect to $\lambda$ from $\lambda=1$ to $\lambda=+\infty$ in order to obtain the thermodynamic potential current $\phi^{\mu}$ according to \eqref{zubeq26}.
\item Plug the result into \eqref{zubeq27} and obtain the entropy current $s^{\mu}$.
\end{enumerate}
Let us emphasize once again that the $\tau$-dependence have been omitted from $\widehat{\Upsilon}$, $\Upsilon_i$ and $|0_{\Upsilon}\rangle$ in order to ease the notation, but they are all $\tau$-dependent quantities in principle.
Note also that, although the operator $\widehat{\Upsilon}$ can be shown to reduce to the Hamiltonian in the special case of global thermodynamic equilibrium in Minkowski spacetime with $\zeta=0$ and $\beta^{\mu}=(1/T)(1,\mathbf{0})$, it is not so in general.
Thus, $|0_{\Upsilon}\rangle$ is not the vacuum state of the theory, and the expectation values in \eqref{zubeq26} and \eqref{zubeq27} do not coincide in general with the physically normally ordered ones.
If we wish not to be misled, we should therefore regard $|0_{\Upsilon}\rangle$ simply as the lowest eigenvector of some operator instead of some vacuum state.
Moreover, we stress once more that the result \eqref{zubeq27} is the entropy current at local thermodynamic equilibrium, which is the underlying assumption of relativistic hydrodynamics and includes global thermodynamic equilibrium as a special case.
However, no analogous expression is known for the entropy current for relativistic quantum fluids fully out of thermodynamic equilibrium.
Note that the expression for the entropy current we just derived fulfills the feature of being divergenceless at global thermodynamic equilibrium.
First of all, at global thermodynamic equilibrium we have $\nabla_{\mu}\langle \widehat{O}\rangle=\langle \nabla_{\mu}\widehat{O}\rangle$, because $\nabla_{\mu}\widehat{\rho}=0$ thanks to stationarity.
Same thing for $\nabla_{\mu}\langle 0_{\Upsilon}|\widehat{O}|0_{\Upsilon}\rangle=\langle 0_{\Upsilon}|\nabla_{\mu}\widehat{O}|0_{\Upsilon}\rangle$, as the non-degenerate state $|0_{\Upsilon}\rangle$ is stationary at global thermodynamic equilibrium.
Thus, for the thermodynamic potential current we have
\begin{equation}
\begin{split}
\nabla_{\mu}\phi^{\mu}=&
\nabla_{\mu}\int_1^{+\infty}{\rm d} \lambda \left[\left(\langle \widehat{T}^{\mu \nu}\rangle_{\rm LE}-\langle 0_{\Upsilon}|\widehat{T}^{\mu \nu}|0_{\Upsilon}\rangle\right)(\lambda)\beta_{\nu}-\zeta \left(\langle \widehat{j}^{\mu}\rangle_{\rm LE}-\langle 0_{\Upsilon}|\widehat{j}^{\mu}|0_{\Upsilon}\rangle \right)(\lambda)\right]\\
=&\int_1^{+\infty}{\rm d} \lambda \left[\left(\langle \nabla_{\mu}\widehat{T}^{\mu \nu}\rangle-\langle 0_{\Upsilon}|\nabla_{\mu}\widehat{T}^{\mu \nu}|0_{\Upsilon}\rangle \right)(\lambda)\beta_{\nu}-
\zeta \left(\langle \nabla_{\mu}\widehat{j}^{\mu}\rangle-\langle 0_{\Upsilon}|\nabla_{\mu}\widehat{j}^{\mu}|0_{\Upsilon}\rangle \right)(\lambda)\right.\\
&+\left.\left(\langle \widehat{T}^{\mu \nu}\rangle-\langle 0_{\Upsilon}|\widehat{T}^{\mu \nu}|0_{\Upsilon}\rangle \right)(\lambda)\nabla_{\mu}\beta_{\nu}-
\nabla_{\mu}\zeta \left(\langle \widehat{j}^{\mu}\rangle-\langle 0_{\Upsilon}|\widehat{j}^{\mu}|0_{\Upsilon}\rangle \right)\right]=0.
\end{split}
\end{equation}
The second line vanishes because of the conservation laws of the conserved currents quantum operators, and the last line vanishes because of the geometric conditions \eqref{zubeq10} for global thermodynamic equilibrium.
For the exact same reasons, we have
\begin{equation}
\begin{split}
\nabla_{\mu}s^{\mu}=&
\nabla_{\mu}\phi^{\mu}+
\nabla_{\mu}\left(\langle \widehat{T}^{\mu \nu}\rangle-\langle 0_{\Upsilon}|\widehat{T}^{\mu \nu}|0_{\Upsilon}\rangle \right)\beta_{\nu}-
\zeta \nabla_{\mu}\left(\langle \widehat{j}^{\mu}\rangle-\langle 0_{\Upsilon}|\widehat{j}^{\mu}|0_{\Upsilon}\rangle \right)\\
&+\left(\langle \widehat{T}^{\mu \nu}\rangle-\langle 0_{\Upsilon}|\widehat{T}^{\mu \nu}|0_{\Upsilon}\rangle \right)\nabla_{\mu}\beta_{\nu}-
\nabla_{\mu}\zeta \left(\langle \widehat{j}^{\mu}\rangle-\langle 0_{\Upsilon}|\widehat{j}^{\mu}|0_{\Upsilon}\rangle \right)=0.
\end{split}
\end{equation}
\section{Summary and outlook}
Over the years, relativistic hydrodynamics has proven to be a powerful and versatile theory for describing, among others, astrophysical, cosmological and nuclear phenomena.
At some point in time, it became necessary to both go beyond the usual first or second order theory and also implement effects of quantum origin in the hydrodynamic framework.
An example of the latter is the entropy production of irreversible quantum processes.
It is in this atmosphere that the study of the quantum statistical foundations of relativistic hydrodynamics started to take hold, and it still is to date.
To these purposes, in the late 70's D.\ N.\ Zubarev put forward an approach to relativistic quantum statistical mechanics based on a covariant expression of the density operator at local thermodynamic equilibrium, $\widehat{\rho}_{\rm LE}(\tau)$, built according to the maximum entropy principle in terms of the quantum operators of the conserved currents.
After a quick review of quantum statistical mechanics, we started this Chapter by introducing this very object and presenting its most important and interesting features.
A fundamental one is that this operator is not stationary, therefore it cannot be a state of the system in the Heisenberg picture.
Notwithstanding, if at some $\tau_0$ the system is known to be at local thermodynamic equilibrium, the actual stationary state of the system $\widehat{\rho}$ is in fact $\widehat{\rho}_{\rm LE}(\tau_0)$, whether at or out of thermodynamic equilibrium.
We also took this a step further by showing that stationarity at global thermodynamic equilibrium translates into geometrical conditions on the thermodynamic fields, specifically the timelike four-temperature $\beta^{\mu}$ must be a Killing vector field and the fugacity $\zeta$ must be constant.
These are a prerogative of global thermodynamic equilibrium: at local thermodynamic equilibrium and at full non-equilibrium, the thermodynamic fields can be whatever, as long as they solve the constraint equations, what makes this kind of systems much harder to study.
Concerning the problem of entropy production, which has indeed an important place in this work, we showed that an equation for the entropy production rate at local thermodynamic equilibrium can be derived directly from the expression of the density operator.
However, as elegant and clear as this result may be, it does not provide us in fact with any information on the very structure of the entropy current, which, from a relativistic hydrodynamic standpoint, would be desirable to know.
In this respect, in the last Section we showed that a sufficient condition for the existence of the entropy current at local thermodynamic equilibrium is that the operator $\widehat{\Upsilon}(\tau)$ should be bounded from below with non-degenerate lowest eingevalue.
In this case, we proved that the logarithm of the partition function is extensive and, a consequence, that an entropy current exists at local thermodynamic equilibrium, making it the first general proof of this usually tacitly understood hypothesis.
With our method, we derived an expression for the entropy current at local thermodynamic equilibrium based on two main ingredients: the expectation values of the conserved currents at local thermodynamic equilibrium and the eigenvector corresponding to the lowest eigenvalue of $\widehat{\Upsilon}(\tau)$.
In Chapters \ref{chapter:gteacceleration} and \ref{chapter:boost}, we will use our method and the Zubarev approach in general to study two different kinds of systems.
As we shall see, aside from their entropy currents, they will be interesting in their own right, both theoretically and phenomenologically.
\section{Global thermodynamic equilibrium with acceleration}
\label{sec:gte_gte_with_acceleration}
The first step in order to make any thermal field theory is the determination of the density operator representing the state at hand, which will then allow for the calculation of thermal expectation values.
In this Section, we will present the state of global thermodynamic equilibrium with acceleration, a non-trivial instance of global thermodynamic equilibrium in Minkowski spacetime.
\subsection{Global thermodynamic equilibrium in Minkowski spacetime}
As we showed in Chapter \ref{chapter:zubarev}, in order for a state to be of global thermodynamic equilibrium, the thermodynamic fields must fulfill the geometric conditions \eqref{zubeq10}, that is the four-temperature $\beta^{\mu}$ must be a (timelike) Killing vector field and the fugacity $\zeta$ must be a constant.
The existence of global timelike Killing vector fields, whence of global thermodynamic equilibrium states, is a prerogative of so-called \textit{stationary} spacetimes, of which Minkowski is an example.
The solution of the Killing equation \eqref{zubeq10} in Minkowski spacetime has the form
\begin{equation}\label{gteq05}
\beta_{\mu}=b_{\mu}+\varpi_{\mu \nu}x^{\nu},
\end{equation}
where $b_{\mu}$ is a constant vector field and $\varpi_{\mu \nu}$ a constant tensor given by the exterior derivative of the four-temperature,
\begin{equation}\label{gteq06}
\varpi_{\mu \nu}=-\frac{1}{2}\left(\partial_{\mu}\beta_{\nu}-\partial_{\nu}\beta_{\mu}\right),
\end{equation}
called the \textit{thermal vorticity} \cite{Becattini:2012tc}.
This can be further decomposed, in general, in a way similar to how the electromagnetic tensor is decomposed into the comoving electric and magnetic fields \cite{Buzzegoli:2017cqy}
\begin{equation}\label{gteq07}
\varpi_{\mu \nu}=\alpha_{\mu}u_{\nu}-\alpha_{\nu}u_{\mu}+\epsilon_{\mu \nu \rho \sigma}w^{\rho}u^{\sigma},
\end{equation}
where $\alpha^{\mu}$ and $w^{\mu}$ are two spacelike vector fields defined as
\begin{equation}
\alpha_{\mu}\equiv \varpi_{\mu \nu}u^{\nu},\qquad
w_{\mu}\equiv -\frac{1}{2}\epsilon_{\mu \nu \rho \sigma}\varpi^{\nu \rho}u^{\sigma}.
\end{equation}
By expressing them in terms of the four-temperature and the proper temperature, one can show that their physical meaning is that of acceleration and vorticity fields respectively both divided by the proper temperature
\begin{equation}\label{gteq24}
\alpha_{\mu}=\frac{A_{\mu}}{T},\qquad w_{\mu}=\frac{\omega_{\mu}}{T},
\end{equation}
where the acceleration field was defined in \eqref{eqrelhydro23} while the \textit{vorticity} field is defined as $\omega_{\mu}\equiv -\frac{1}{2}\sqrt{\beta^2}\epsilon_{\mu \nu \rho \sigma}\partial^{\rho}u^{\nu}u^{\sigma}$.
Thus, we understand that the thermal vorticity accounts, in some sense, for thermal effects due to the acceleration and rotation of the fluid.
To get an idea of their order of magnitude, it is convenient to go to the local rest frame and express the above vector fields in terms of the local acceleration ${\bf a}$ and local angular velocity $\boldsymbol{\omega}$, that is $\alpha^{\mu}=(0,{\bf a}/T)$ and $w^{\mu}=(0,\boldsymbol{\omega}/T)$.
In order for the decomposition \eqref{gteq07} to have the correct dimensions, it is realized that the appropriate combination of natural constants is
\begin{equation}\label{gteq25}
|\alpha^{\mu}|=\frac{\hbar}{ck_{\rm B}}\frac{|{\bf a}|}{T},\qquad
|w^{\mu}|=\frac{\hbar}{k_{\rm B}}\frac{|\boldsymbol{\omega}|}{T},
\end{equation}
so they are terms of relativistic and/or quantum origin possibly appreciable at low temperature.
For typical systems on Earth, they are in fact negligible both because of the suppression coming from the natural constants and the room temperature being large with respect to the magnitude of the acceleration and rotation.
However, although possessing a very high temperature ($k_{\rm B}T\simeq 200\,{\rm MeV}$), the quark-gluon plasma produced in heavy-ion collisions has such large values of acceleration ($\hbar |{\bf a}|/c\simeq 10\,{\rm MeV}$) and vorticity ($h|\boldsymbol{\omega}|\simeq 12\,{\rm MeV}$) that these terms are in fact meaningful ($|\alpha^{\mu}|\simeq 0.05$, $|w^{\mu}|\simeq 0.06$), enough so to generate observable effects such as the polarization of $\Lambda$ hyperions \cite{STAR:2017ckg, Adam:2018ivw}.
With equations \eqref{gteq05} and \eqref{gteq06}, the density operator at global thermodynamic equilibrium \eqref{zubeq28} readily becomes
\begin{equation}
\begin{split}
\widehat{\rho}=&
\frac{1}{Z}\exp \left[-\int_{\Sigma}{\rm d} \Sigma_{\mu}\left(\widehat{T}^{\mu \nu}\left(b_{\nu}+\varpi_{\nu \lambda}x^{\lambda}\right)-\zeta \widehat{j}^{\mu}\right)\right]\\
=&\frac{1}{Z}\exp \left[-b_{\mu}\int_{\Sigma}{\rm d} \Sigma_{\mu}\widehat{T}^{\mu \nu}+\frac{1}{2}\varpi_{\nu \lambda}\int_{\Sigma}{\rm d} \Sigma_{\mu}\left(x^{\nu}\widehat{T}^{\mu \lambda}-x^{\lambda}\widehat{T}^{\mu \nu}\right)+\zeta \int_{\Sigma}{\rm d} \Sigma_{\mu}\widehat{j}^{\mu}\right],
\end{split}
\end{equation}
where $b_{\mu}$, $\varpi_{\mu \nu}$ and $\zeta$ are taken out of the integral as they are constant.
Here we recognize the generators of the Poincaré group
\begin{equation}\label{gteq43}
\widehat{P}^{\mu}=\int_{\Sigma}{\rm d} \Sigma_{\mu}\widehat{T}^{\mu \nu},\qquad
\widehat{J}^{\nu \lambda}=\int_{\Sigma}{\rm d} \Sigma_{\mu}\left(x^{\nu}\widehat{T}^{\mu \lambda}-x^{\lambda}\widehat{T}^{\mu \nu}\right)
\end{equation}
the former being the four-momentum operator and the latter the generators of Lorentz transformations, and also the conserved charge associated to the current $\widehat{j}^{\mu}$
\begin{equation}
\widehat{Q}=\int_{\Sigma}{\rm d} \Sigma_{\mu}\widehat{j}^{\mu},
\end{equation}
therefore we find
\begin{equation}\label{gteq16}
\widehat{\rho}=\frac{1}{Z}\exp \left[-b_{\mu}\widehat{P}^{\mu}+\frac{1}{2}\varpi_{\mu \nu}\widehat{J}^{\mu \nu}+\zeta \widehat{Q}\right].
\end{equation}
In other words, aside from a constant fugacity associated to the conserved charge ascribed to internal symmetries, the most general expression of the density operator at global thermodynamic equilibrium in Minkowski spacetime is given by a linear combination of the generators of the Poincaré group with ten constant coefficients.
Depending on how these coefficients are specified, different kinds of global thermodynamic equilibrium are obtained.
In the simplest case of a vanishing thermal vorticity $\varpi_{\mu \nu}=0$, the four-temperature is just $\beta_{\mu}=b_{\mu}$, and the following familiar density operator is recovered
\begin{equation}\label{gteq11}
\widehat{\rho}=\frac{1}{Z}\exp \left[-b_{\mu}\widehat{P}^{\mu}+\zeta \widehat{Q}\right]=\frac{1}{Z}\exp \left[-\frac{\widehat{H}}{T_0}-\zeta \widehat{Q}\right],
\end{equation}
where the last expression is obtained by setting $\beta_{\mu}=b_{\mu}=(1/T_0)(1,{\bf 0})$ with $T_0$ a constant thermodynamic parameter with the dimensions of a temperature and $\widehat{H}$ the Hamiltonian.
This state is called \textit{homogeneous global thermodynamic equilibrium}.
As for a non-vanishing thermal vorticity, there are two possible cases.
The first one is that of a thermal vorticity with only a transverse (or ``magnetic'') component
\begin{equation}
b_{\mu}=\frac{1}{T_0},\qquad
\varpi_{\mu \nu}=\frac{\omega}{T_0}\left(g_{1\mu}g_{2\nu}-g_{1\nu}g_{2\mu}\right),
\end{equation}
where $\omega$ is a constant thermodynamic parameter with the dimensions of an angular velocity.
The four-temperature then reads
\begin{equation}
\beta_{\mu}=\frac{1}{T_0}(1,\omega y,-\omega x,0),
\end{equation}
hence the density operator
\begin{equation}\label{gteq03}
\widehat{\rho}=\frac{1}{Z}\exp \left[-\frac{\widehat{H}}{T_0}+\frac{\omega}{T_0}\widehat{J}_z+\zeta \widehat{Q}\right],
\end{equation}
with $\widehat{J}_z$ the angular momentum operator along $z$.
As it is known, this expression represents a fluid at thermodynamic equilibrium rigidly rotating around the $z$ direction with constant angular velocity $\omega$ \cite{landau2013statistical}, therefore it is called \textit{global thermodynamic equilibrium with rotation}.
This was used for instance in \cite{Vilenkin:1980zv, Becattini:2011ev} to study quantum effects in rotating relativistic matter and reworked more thoroughly in \cite{Becattini:2014yxa}.
The last and less known case is that of a thermal vorticity with only a longitudinal (or ``electric'') component
\begin{equation}
b_{\mu}=\frac{1}{T_0}(1,{\bf 0}),\qquad
\varpi_{\mu \nu}=\frac{a}{T_0}\left(g_{0\nu}g_{3\mu}-g_{0\mu}g_{3\nu}\right),
\end{equation}
where $a$ is a thermodynamic parameter with the dimensions of an acceleration.
The four-temperature then reads
\begin{equation}\label{gteq02}
\beta_{\mu}=\frac{1}{T_0}\left(1+az,0,0,-at\right),
\end{equation}
hence the density operator
\begin{equation}\label{gteq01}
\widehat{\rho}=\frac{1}{Z}\exp \left[-\frac{\widehat H}{T_0}+\frac{a}{T_0}\widehat{K}_z+\zeta \widehat{Q}\right],
\end{equation}
with $\widehat{K}_z$ the boost operator along $z$.
As we shall shortly see, this expression represents a fluid at thermodynamic equilibrium with acceleration field of constant magnitude along the flow, therefore it is called \textit{global thermodynamic equilibrium with acceleration}.
This was studied in detail in \cite{Becattini:2017ljh}, where it was shown that the Unruh effect is indeed recovered, and partly also in \cite{Becattini:2019poj}, where the calculation of the thermal expectation value of the energy-momentum tensor was completed and the entropy current calculated.
Equations \eqref{gteq03} and \eqref{gteq01} are the only two independent global thermodynamic equilibria in Minkowski spacetime with non-vanishing thermal vorticity, all other cases are combinations thereof.
For a discussion on the general case with both rotation and acceleration, see \cite{Korsbakken:2004bv}.
We conclude this part with the following remark.
As we mentioned in Chapter \ref{chapter:zubarev}, the symmetries of the density operator constrain the form of the thermal expectation values calculated with it.
In general, by looking at equation \eqref{gteq16}, the thermal expectation value of any local operator $\widehat{O}$ can depend on $b_{\mu}$, $\varpi_{\mu \nu}$, $x_{\mu}$ and $g_{\mu \nu}$.
Concerning the dependence on $x_{\mu}$, let us indicate the translation operator as $\widehat{\sf T}(x)\equiv \exp[-ix_{\mu}\widehat{P}^{\mu}]$, hence
\begin{equation}\label{gteq17}
\begin{split}
\langle \widehat{O}(x)\rangle=&
\mathrm{tr} \left(\widehat{\rho}\widehat{\sf T}(x)\widehat{O}(0)\widehat{\sf T}^{-1}(x)\right)=
\mathrm{tr} \left(\widehat{\sf T}^{-1}(x)\widehat{\rho}\widehat{\sf T}(x)\widehat{O}(0)\right)\\
=&\frac{1}{Z}\mathrm{tr} \left(\exp \left[b_{\mu}\widehat{P}^{\mu}+\frac{1}{2}\varpi_{\mu \nu}\widehat{\sf T}^{-1}(x)\widehat{J}^{\mu \nu}\widehat{\sf T}(x)\right]\widehat{O}(0)\right)\\
=&\frac{1}{Z}\mathrm{tr} \left(\exp \left[-\beta_{\mu}(x)\widehat{P}^{\mu}+\frac{1}{2}\varpi_{\mu \nu}\widehat{J}^{\mu \nu}\right]\widehat{O}(0)\right)=
\langle \widehat{O}(0)\rangle_{\beta_{\mu}(x)},
\end{split}
\end{equation}
where the known relation of the Poincaré algebra was used
\begin{equation}
\widehat{\sf T}^{-1}(x)\widehat{J}_{\mu \nu}\widehat{\sf T}(x)=\widehat{J}_{\mu \nu}-x_{\nu}\widehat{P}_{\mu}+x_{\mu}\widehat{P}_{\nu}.
\end{equation}
Equation \eqref{gteq17} implies that the dependence on the spacetime point of the thermal expectation value of any quantum operator cannot be arbitrary, but it can be only through the four-temperature.
This property will come in handy later on.
\subsection{Global thermodynamic equilibrium with acceleration}
Let us now focus on the configuration of global thermodynamic equilibrium with acceleration.
From here on, we will set $\zeta=0$ for simplicity without loss of generality, namely we will forget about the charged current and consider only the energy-momentum tensor.
The expression of the density operator is then
\begin{equation}\label{gteq04}
\widehat{\rho}=\frac{1}{Z}\exp \left[-\frac{\widehat{H}}{T_0}+\frac{a}{T_0}\widehat{K}_z\right].
\end{equation}
From the general discussion in the previous Chapter, we recall that in the argument of the exponential function there appear the quantum operators corresponding to the conserved charges.
One particular feature of \eqref{gteq04} is that the Hamiltonian $\widehat{H}$ and the boost operator $\widehat{K}_z$ do not commute with each other, and yet they are both conserved.
This occurs because the boost operator is explicitly time-dependent
\begin{equation}
\widehat{K}_z=\widehat{J}_{30}=t\widehat{P}_z-\int {\rm d}^3{\rm x}\,z\widehat{T}^{00},
\end{equation}
so its Heisenberg equation reads
\begin{equation}
i\frac{{\rm d} \widehat{K}_z}{{\rm d} t}=[\widehat{K}_z,\widehat{H}]+i\frac{\partial \widehat{K}_z}{\partial t}=-i\widehat{P}_z+i\widehat{P}_z=0.
\end{equation}
Although possible at a mathematical level, the question arises if this state can be physically realized.
As we mentioned in the last Chapter, the global thermodynamic equilibrium density operator is in fact an approximation of the local thermodynamic equilibrium one at first order in the derivatives of the thermodynamic fields \cite{Becattini:2014yxa}.
This makes \eqref{gteq04} a reasonable approximation for describing fluids at local thermodynamic equilibrium with non-vanishing acceleration (and vorticity), such as the quark-gluon plasma, in the hydrodynamic limit.
In order to work out the kinematic quantities, it is convenient to shift the origin of the $z$ coordinate by defining $z'\equiv z+1/a$, so that the four-temperature in \eqref{gteq02} simply reads
\begin{equation}\label{gteq08}
\beta^{\mu}=\frac{a}{T_0}(z',0,0,t).
\end{equation}
Then, according to \eqref{zubeq11}, the velocity field and proper temperature defined by its direction and inverse magnitude are respectively given by
\begin{equation}\label{gteq09}
u^{\mu}=\frac{1}{\sqrt{{z'}^2-t^2}}(z',0,0,t),\qquad
T=\frac{T_0}{a\sqrt{{z'}^2-t^2}}.
\end{equation}
It is easily realized that the flow lines, namely the curves with $u^{\mu}$ as tangent vector field, are hyperbolae of constant values of ${z'}^2-t^2$, and the proper temperature is constant along them.
This should be no surprise, being in fact a general feature of Killing vectors to have constant magnitude along their integral curves.
Now, by using the definition \eqref{eqrelhydro23} of the acceleration field, we readily obtain the expression
\begin{equation}\label{gteq37}
A^{\mu}=\frac{1}{{z'}^2-t^2}(t,0,0,z'),
\end{equation}
which is clearly orthogonal to the four-temperature and whose magnitude is constant along the flow lines.
In other words, the density operator \eqref{gteq04} represents a fluid at global thermodynamic equilibrium with acceleration field of constant magnitude along the flow, hence the name ``global thermodynamic equilibrium with acceleration''.
Furthermore, referring to the general decomposition \eqref{gteq07} of the thermal vorticity, in our specific case at hand the vorticity field is identically null, so we are actually left with
\begin{equation}\label{gteq18}
\varpi_{\mu \nu}=\alpha_{\mu}u_{\nu}-\alpha_{\nu}u_{\mu},
\end{equation}
where $\alpha^{\mu}$ has constant magnitude in the whole spacetime
\begin{equation}\label{gteq23}
\alpha^2=\frac{A^2}{T^2}=-\frac{a^2}{T_0^2}.
\end{equation}
As for the thermodynamics, the key point is that the four-temperature \eqref{gteq08} is not globally timelike for it has a bifurcated Killing horizon at $|z'|=t$ splitting Minkowski spacetime into four different subspaces as shown in Figure \ref{fig:wedges}:
\begin{itemize}
\item $|t|<z'$ is called the \textit{Right Rindler Wedge} (RRW),
\item $|t|<-z'$ is called the \textit{Left Rindler Wedge} (LRW),
\item $t>|z'|$ is called the \textit{Expanding Degenerate Kasner Universe} (EDK),
\item $t<-|z'|$ is called the \textit{Contracting Degenerate Kasner Universe} (EDK).
\end{itemize}
In particular, in order to make hydrodynamics and thermodynamics out of $\beta^{\mu}$, we have to figure out in which subspace it is both timelike and future-oriented globally, so that the interpretation of $u^{\mu}$ and $T$ as a velocity field and a proper temperature respectively makes sense.
By looking at Figure \ref{fig:wedges}, we can tell that the subspace we are looking for is the right Rindler wedge, so that is where we must restrict.
\begin{figure}
\begin{center}
\includegraphics[width=0.745\textwidth]{figures/wedges.eps}
\caption{2-dimensional $(t,z)$ section of Minkowski spacetime.
The hyperbolae are the flow lines of the four-temperature, the latter being the tangent vector field.
The red lines crossing at $z'=0$, i.e.\ $z=-1/a$, are the bifurcated Killing horizon of $\beta^{\mu}$ splitting the spacetime into four different subspaces: the right Rindler wedge (RRW), the left Rindler wedge (LRW), the expanding degenerate Kasner universe (EDK) and the contracting degenerate Kasner universe (CDK).
The four-temperature is globally timelike and future-oriented only in the right Rindler wedge.
The blue line through $z'=0$ and contained in the two Rindler wedges is a hyperplane $\Sigma$ orthogonal to $\beta^{\mu}$: such hyperplanes can be used to foliate the subspaces where the four-temperature is timelike.}
\label{fig:wedges}
\end{center}
\end{figure}
An observer moving along a constant-acceleration path, that is along the flow, is called a \textit{Rindler observer}.
Note that the expression of the proper temperature in equation \eqref{gteq09} is but an instance of Tolman's law \cite{Buchholz:2015fqa}, but it is most naturally obtained in the Zubarev approach by demanding the four-temperature to be a Killing vector field.
\subsection{Factorization of the density operator}
In \cite{Becattini:2017ljh}, it is shown that the density operator \eqref{gteq04} has an important factorization property.
Rather than \eqref{gteq04}, it is convenient to consider the general expression \eqref{zubeq28}, which still holds of course, and rewrite it by defining $\widehat{\Pi}$ as
\begin{equation}\label{gteq10}
\widehat{\rho}=\frac{1}{Z}\exp \left[-\frac{\widehat{\Pi}}{T_0}\right],\qquad
\widehat{\Pi}\equiv T_0\int_{\Sigma}{\rm d} \Sigma_{\mu}\widehat{T}^{\mu \nu}\beta_{\nu}=\widehat{H}-a\widehat{K}_z.
\end{equation}
At global thermodynamic equilibrium, this is independent on the choice of the spacelike hypersurface $\Sigma$ of the foliation.
To define the foliation, we can exploit the fact that the four-temperature \eqref{gteq08} fulfills the vorticity-free condition \eqref{zubeq07}, that is
\begin{equation}
\epsilon^{\mu \nu \rho \sigma}\beta_{\sigma}\partial_{\nu}\beta_{\rho}=0.
\end{equation}
In particular, the family of 3-dimensional spacelike hypersurfaces orthogonal to it is given by the hyperplanes through $z'=0$ and entirely contained in the right and left Rindler wedges, as shown in Figure \ref{fig:wedges}, so let $\Sigma$ be any of them.
Thus, let us decompose the hyperplane as $\Sigma=\Sigma_{\rm R}\cup \Sigma_{\rm L}$ with $\Sigma_{\rm R}$ and $\Sigma_{\rm L}$ entirely contained in the right and left Rindler wedge respectively, namely $\Sigma_{\rm R}$ being the $z'>0$ part of $\Sigma$ and $\Sigma_{\rm L}$ the $z'<0$ part.
This way, $\widehat{\Pi}$ is given by the sum of two terms, each depending only on the field degrees of freedom in one of the two Rindler wedges
\begin{equation}
\widehat{\Pi}=\widehat{\Pi}_{\rm R}-\widehat{\Pi}_{\rm L}
\end{equation}
\begin{equation}
\widehat{\Pi}_{\rm R}\equiv T_0\int_{\Sigma_{\rm R}}{\rm d} \Sigma_{\mu}\widehat{T}^{\mu \nu}\beta_{\nu},\qquad
\widehat{\Pi}_{\rm L}\equiv -T_0\int_{\Sigma_{\rm L}}{\rm d} \Sigma_{\mu}\widehat{T}^{\mu \nu}\beta_{\nu}.
\end{equation}
The minus sign in the definition of $\widehat{\Pi}_{\rm L}$ is due to the fact that $n^{\mu}$ orthogonal to $\Sigma_{\rm L}$ has opposite orientation with respect to $\beta^{\mu}$ in the left Rindler wedge.
Now the key point is the following.
The only contribution to the commutator $[\widehat{\Pi}_{\rm R},\widehat{\Pi}_{\rm L}]$ stems from $z'=0$, which is the only point where the quantum field operators in the energy-momentum tensor have non-vanishing commutators.
However, the four-temperature altogether vanishes there, as we can convince ourselves by noticing that the proper temperature \eqref{gteq09} diverges there, or simply just by looking at Figure \ref{fig:wedges}, therefore the commutator is in fact null.
This implies that the density operator is factorized as
\begin{equation}
\widehat{\rho}=
\frac{1}{Z}\exp \left[-\frac{\widehat{\Pi}_{\rm R}-\widehat{\Pi}_{\rm L}}{T_0}\right]=
\frac{1}{Z}\exp \left[-\frac{\widehat{\Pi}_{\rm R}}{T_0}\right]\exp \left[\frac{\widehat{\Pi}_{\rm L}}{T_0}\right].
\end{equation}
Formally, since $\widehat{\Pi}_{\rm R}$ and $\widehat{\Pi}_{\rm L}$ involve only field degrees of freedom each in its corresponding Rindler wedge, in the Hilbert space of the field states we actually have $\widehat{\Pi}_{\rm R}\equiv \widehat{\Pi}_{\rm R}\otimes \widehat{I}_{\rm L}$ and $\widehat{\Pi}_{\rm L}\equiv \widehat{I}_{\rm R}\otimes \widehat{\Pi}_{\rm L}$, with $\widehat{I}_{\rm R}$ and $\widehat{I}_{\rm L}$ the identities in the right and left Rindler wedge respectively.
In turn, this implies that the partition function is factorized as well
\begin{equation}
Z=\mathrm{tr} \left(\exp \left[-\frac{\widehat{\Pi}_{\rm R}-\widehat{\Pi}_{\rm L}}{T_0}\right]\right)=
\mathrm{tr}_{\rm R}\left(\exp \left[-\frac{\widehat{\Pi}_{\rm R}}{T_0}\right]\right)
\mathrm{tr}_{\rm L}\left(\exp \left[\frac{\widehat{\Pi}_{\rm L}}{T_0}\right]\right)\equiv Z_{\rm R}Z_{\rm L},
\end{equation}
where by $\mathrm{tr}_{\rm R}$ and $\mathrm{tr}_{\rm L}$ we mean that the trace is calculated only on the Hilbert space spanned by the field degrees of freedom in the right and left Rindler wedge respectively, which is called a \textit{partial trace}.
In summary, the density operator becomes factorized into two density operators commuting with each other
\begin{equation}\label{gteq41}
\widehat{\rho}=\widehat{\rho}_{\rm R}\otimes \widehat{\rho}_{\rm L},\qquad
[\widehat{\rho}_{\rm R},\widehat{\rho}_{\rm L}]=0
\end{equation}
each involving the field degrees of freedom in either of the Rindler wedges
\begin{equation}\label{gteq12}
\widehat{\rho}_{\rm R}=
\frac{1}{Z_{\rm R}}\exp \left[-\frac{\widehat{\Pi}_{\rm R}}{T_0}\right]=
\frac{1}{Z_{\rm R}}\exp \left[-\int_{\Sigma_{\rm R}}{\rm d} \Sigma_{\mu}\widehat{T}^{\mu \nu}\beta_{\nu}\right],
\end{equation}
\begin{equation}
\widehat{\rho}_{\rm L}=
\frac{1}{Z_{\rm L}}\exp \left[\frac{\widehat{\Pi}_{\rm L}}{T_0}\right]=
\frac{1}{Z_{\rm L}}\exp \left[-\int_{\Sigma_{\rm L}}{\rm d} \Sigma_{\mu}\widehat{T}^{\mu \nu}\beta_{\nu}\right].
\end{equation}
They are obtained by taking the partial trace of the whole density operator on the opposite wedge, so, referring to the terminology of Quantum Information Theory, they are called \textit{reduced density operators}
\begin{equation}\label{gteq42}
\widehat{\rho}_{\rm R}=\mathrm{tr}_{\rm L}(\widehat{\rho}),\qquad
\widehat{\rho}_{\rm L}=\mathrm{tr}_{\rm R}(\widehat{\rho}).
\end{equation}
Thus, the two Rindler wedges are completely disconnected.
As a consequence, the thermal expectation value of an operator $\widehat{O}(x)$ with $x$ in the right Rindler wedge, for instance, will only depend on the reduced density operator $\widehat{\rho}_{\rm R}$, regardless of the field states in the left Rindler wedge
\begin{equation}
\langle \widehat{O}(x)\rangle=\mathrm{tr}(\widehat{\rho}\widehat{O}(x))=\mathrm{tr}_{\rm R}(\widehat{\rho}_{\rm R}\widehat{O}(x)).
\end{equation}
Equation \eqref{gteq10} defining the operator $\widehat{\Pi}$ is reminiscent of the homogeneous global thermodynamic equilibrium density operator \eqref{gteq11}.
Indeed, by calculating the commutator with the quantum field operator, one can show that $\widehat{\Pi}$ is the generator of translations along the flow in Minkowski spacetime \cite{Korsbakken:2004bv}, same for $\widehat{\Pi}_{\rm R}$ and $\widehat{\Pi}_{\rm L}$ in their respective Rindler wedges \cite{Becattini:2017ljh}.
\section{Thermal expectation values and the Unruh effect}
\label{sec:gte_tev_unruh}
With the density operator, we are now in a position to calculate thermal expectation values of quantum operators.
In particular, we are interested in that of the energy-momentum tensor, for it is both of physical concern in its own right and also necessary for the calculation of the entropy current.
As already mentioned, the quantum operators are built with the quantum field operators, and as such they depend on the Quantum Field Theory underlying the hydrodynamic theory.
In this work, we will consider the simple case of a free real scalar field in the right Rindler wedge.
As we shall see, the Unruh effect will emerge from thermal expectation values in a natural way.
\subsection{Free scalar field theory in the right Rindler wedge}
The problem of a free real scalar field in the right Rindler wedge is indeed a well-known one in Physics.
Here, we just present a short summary highlighting the characteristics most salient for this work, for more details we refer to the review \cite{Crispino:2007eb}.
In order to build the energy-momentum tensor operator and calculate its thermal expectation value, we first need the expression of the quantum field operator, which is obtained as a solution of its equation of motion.
The Lagrangian density $\widehat{\cal L}$ of a free real scalar field $\widehat{\psi}$ of mass $m$ in Minkowski spacetime is
\begin{equation}\label{eq19}
\widehat{\cal L}=\frac{1}{2}\left(g^{\mu \nu}\partial_{\mu}\widehat{\psi} \partial_{\nu}\widehat{\psi}-m^2\widehat{\psi}^2\right)
\end{equation}
and the corresponding equation of motion is the Klein-Gordon equation
\begin{equation}
\left(\Box+m^2\right)\widehat{\psi}=0.
\end{equation}
Provided initial data, the solutions of Klein-Gordon equation can be uniquely determined in Minkowski spacetime, for it is globally hyperbolic with constant-time hypersurfaces as Cauchy surfaces.
Given any two solutions $\phi_1$ and $\phi_2$, we can build the \textit{Klein-Gordon inner product} to define orthogonality and normalization of the solutions
\begin{equation}
(\phi_1,\phi_2)_{\rm KG}\equiv i\int_{\Sigma}{\rm d}\Sigma_{\mu}\left(\phi_1^*\partial^{\mu}\phi_2-\phi_2\partial^{\mu}\phi_1^*\right),
\end{equation}
with $\Sigma$ a spacelike hypersurface with future-oriented unit orthogonal vector field.
This is independent of $\Sigma$, and also independent of time if $\Sigma$ is a constant-time hypersurface.
Let $\{u_i\}$ be a set of solutions orthonormalized according to the above inner product, that is
\begin{equation}
(u_i,u_j)_{\rm KG}=\delta_{ij}=-(u_i^*,u_j^*)_{\rm KG},\qquad (u_i^*,u_j)_{\rm KG}=0=(u_i,u_j^*)_{\rm KG}.
\end{equation}
The general solution of the Klein-Gordon equation can then be expanded as
\begin{equation}
\widehat{\psi}=\sum_i\left(u_i\widehat{a}_i+u_i^*\widehat{a}_i^{\dagger}\right),
\end{equation}
where, $\widehat{a}_i^{\dagger}$ and $\widehat{a}_i$ are the creation and annihilation operators respectively.
Using the orthonormality of the solutions $u_i$, one can show that
\begin{equation}
\widehat{a}_i=(u_i,\widehat{\psi})_{\rm KG},\qquad \widehat{a}_i^{\dagger}=-(u_i^*,\widehat{\psi})_{\rm KG},
\end{equation}
which, together with the canonical commutation relations, imply as expected
\begin{equation}
[\widehat{a}_i,\widehat{a}_j^{\dagger}]=\delta_{ij},\qquad [\widehat{a}_i,\widehat{a}_j]=[\widehat{a}_i^{\dagger},\widehat{a}_j^{\dagger}]=0.
\end{equation}
The vacuum state is defined by requiring it to be annihilated by all the annihilation operators, and the Fock space is built by applying the creation operators to it.
The annihilation operators, in turn, are determined by the solutions $u_i$, so it is actually their choice that determines the vacuum.
Such choice is not unique, and so is the vacuum in principle, however Minkowski spacetime is \textit{static}.
In a static spacetime it is natural for the solutions $u_i$ to have a time-dependence of the form ${\rm e}^{-i\omega_it}$, where $\omega_i$ are positive constants interpreted as the energies of the particles with respect to the future-oriented Killing vector $\partial_t$.
For this reason, the solutions $u_i$ are called \textit{positive-frequency modes}, and the $u_i^*$ \textit{negative-frequency modes}.
In Minkowski spacetime, the quantum field expansion usually reads
\begin{equation}\label{gteq27}
\widehat{\psi}(x)=\int_{-\infty}^{+\infty}{\rm d}^3{\rm p}\left(v_{\bf p}\widehat{a}_{\bf p}+v^*_{\bf p}\widehat{a}^{\dagger}_{\bf p}\right),
\end{equation}
where the positive-frequency modes are the plane waves
\begin{equation}\label{gteq49}
v_{\bf p}=\frac{1}{\sqrt{(2\pi)^32\omega_{\bf p}}}{\rm e}^{-i(\omega_{\bf p} t-{\bf p}\cdot {\bf x})}.
\end{equation}
If a spacetime is both static and globally hyperbolic, the choice of positive-frequency modes with time-dependence of the form ${\rm e}^{-i\omega_it}$ leads to a natural vacuum state that preserves time-translation symmetry, called the \textit{static vacuum}.
The static vacuum in Minkowski spacetime is called the \textit{Minkowski vacuum}, indicated as $|0_{\rm M}\rangle$, that is the state annihilated by all the annihilation operators $\widehat{a}_{\bf p}$.
The right Rindler wedge is also a static and globally hyperbolic spacetime, and it possesses a timelike future-oriented Killing vector given by the boost generator $z'\partial_t+t\partial_{z'}$ playing the role of time-translations generator, so we can solve the Klein-Gordon equation therein as well.
To do so, it is convenient to introduce a set of hyperbolic coordinates called the \textit{Rindler coordinates} $(\tau,{\bf x}_{\rm T},\xi)$, where the ``transverse coordinates'' ${\bf x}_{\rm T} \equiv (x,y)$ are the same as Minkowski's, and $(\tau,\xi)$ are related to $(t,z')$ by
\begin{equation}
\tau \equiv \frac{1}{2a}\log \left(\frac{z'+t}{z'-t}\right),\qquad \xi \equiv \frac{1}{2a}\log \left[a^2\left({z'}^2-t^2\right)\right]
\end{equation}
whose inverse read
\begin{equation}
t=\frac{{\rm e}^{a\xi}}{a}\sinh(a\tau),\qquad
z'=\frac{{\rm e}^{a\xi}}{a}\cosh(a\tau)
\end{equation}
spanning indeed the right Rindler wedge, as shown in Figure \ref{fig:rindlercoord}.
\begin{figure}
\begin{center}
\includegraphics[width=0.75\textwidth]{figures/rindler_coord.eps}
\caption{2-dimensional section of the right Rindler wedge in the $(t,z)$ plane, spanned by the Rindler coordinates $(\tau,\xi)$.
The straight lines through $z=-1/a$, i.e.\ $z'=0$, are $\tau={\rm const.}$ hypersurfaces, while the hyperbolae are $\xi={\rm const.}$ hypersurfaces.}
\label{fig:rindlercoord}
\end{center}
\end{figure}
Plugging them into the Klein-Gordon equation, the positive-frequency modes are obtained
\begin{equation}\label{gteq30}
u_{\omega,{\bf k}_{\rm T}}(\tau,{\bf x}_{\rm T},\xi)\equiv \sqrt{\frac{1}{4\pi^4a}\sinh \left(\frac{\pi \omega}{a}\right)}{\rm K}_{i\frac{\omega}{a}}\left(\frac{m_{\rm T}{\rm e}^{a\xi}}{a}\right){\rm e}^{-i(\omega \tau-{\bf k}_{\rm T} \cdot {\bf x}_{\rm T})},
\end{equation}
where ${\rm K}_{i\frac{\omega}{a}}$ are the modified Bessel functions, $m_{\rm T}$ is the \textit{transverse mass}
\begin{equation}
m_{\rm T}=\sqrt{{\bf k}_{\rm T}^2+m^2},
\end{equation}
${\bf k}_{\rm T} \equiv (k_x,k_y)$ is the ``transverse momentum'' and $\omega \ge 0$ is the frequency.
Together with the negative-frequency modes $u_{\omega,{\bf k}_{\rm T}}^*$, they are orthonormalized according to the Klein-Gordon inner product.
Thus, the field can be expanded as
\begin{equation}\label{gteq13}
\widehat{\psi}(\tau,{\bf x}_{\rm T},\xi)=\int_0^{+\infty}{\rm d} \omega \int_{-\infty}^{+\infty}{\rm d}^2{\rm k_T}\left(u_{\omega,{\bf k}_{\rm T}}\widehat{a}^{\rm R}_{\omega,{\bf k}_{\rm T}}+u^*_{\omega,{\bf k}_{\rm T}}\widehat{a}^{\rm R\dagger}_{\omega,{\bf k}_{\rm T}}\right),
\end{equation}
where $\widehat{a}^{\rm R\dagger}_{\omega,{\bf k}_{\rm T}}$ and $\widehat{a}^{\rm R}_{\omega,{\bf k}_{\rm T}}$ are the creation and annihilation operators respectively, given by
\begin{equation}
\widehat{a}^{\rm R}_{\omega,{\bf k}_{\rm T}}=(u_{\omega,{\bf k}_{\rm T}},\widehat{\psi})_{\rm KG},\qquad \widehat{a}^{\rm R\dagger}_{\omega,{\bf k}_{\rm T}}=-(u^*_{\omega,{\bf k}_{\rm T}},\widehat{\psi})_{\rm KG}.
\end{equation}
With the canonical commutation relations, the above equations imply the usual algebra
\begin{equation}\label{gteq15}
\begin{split}
[\widehat{a}^{\rm R}_{\omega,{\bf k}_{\rm T}},\widehat{a}^{\rm R\dagger}_{\omega',{\bf k}_{\rm T}'}]=&\delta(\omega-\omega')\,\delta^2({\bf k}_{\rm T}-{\bf k}_{\rm T}'),\\
[\widehat{a}^{\rm R}_{\omega,{\bf k}_{\rm T}},\widehat{a}^{\rm R}_{\omega',{\bf k}_{\rm T}'}]=&0=[\widehat{a}^{\rm R\dagger}_{\omega,{\bf k}_{\rm T}},\widehat{a}^{\rm R\dagger}_{\omega',{\bf k}_{\rm T}'}].
\end{split}
\end{equation}
In the left Rindler wedge, the coordinates $t\equiv a^{-1}{\rm e}^{a\bar{\xi}}\sinh(a\bar{\tau})$ and $z'\equiv -a^{-1}{\rm e}^{a\bar{\xi}}\cosh(a\bar{\tau})$ are introduced in order to solve the Klein-Gordon equation.
An expansion analogous to \eqref{gteq13} holds there, with the important difference that the role of the positive- and negative-frequency modes is interchanged as a consequence of the fact that the boost generator, acting as a time-translations generator, is past-oriented
\begin{equation}\label{gteq26}
\widehat{\psi}(\bar{\tau},{\bf x}_{\rm T},\bar{\xi})=\int_0^{+\infty}{\rm d} \omega \int_{-\infty}^{+\infty}{\rm d}^2{\rm k_T}\left(u_{\omega,{\bf k}_{\rm T}}^*\widehat{a}^{\rm L}_{\omega,{\bf k}_{\rm T}}+u_{\omega,{\bf k}_{\rm T}}\widehat{a}^{\rm L\dagger}_{\omega,{\bf k}_{\rm T}}\right),
\end{equation}
where $u_{\omega,{\bf k}_{\rm T}}$ has the same expression as \eqref{gteq30} with $(\tau,\xi)$ replaced by $(\bar{\tau},\bar{\xi})$.
The static vacuum in the Rindler wedges, called the \textit{Rindler vacuum} and indicated as $|0_{\rm R}\rangle$, is the state annihilated by all $\widehat{a}^{\rm R}_{\omega,{\bf k}_{\rm T}}$ and $\widehat{a}^{\rm L}_{\omega,{\bf k}_{\rm T}}$, namely $\widehat{a}^{\rm R}_{\omega,{\bf k}_{\rm T}}|0_{\rm R}\rangle=0=\widehat{a}^{\rm L}_{\omega,{\bf k}_{\rm T}}|0_{\rm R}\rangle$.
As for the energy-momentum tensor, we consider the canonical one
\begin{equation}\label{gteq19}
\widehat{T}^{\mu \nu}=\frac{1}{2}\left(\partial^{\mu}\widehat{\psi}\partial^{\nu}\widehat{\psi}+\partial^{\nu}\widehat{\psi}\partial^{\mu}\widehat{\psi}\right)-g^{\mu \nu}\widehat{\cal L},
\end{equation}
but this is not the only possible choice in principle.
Plugging the quantum field expansion into it, and then the result into \eqref{gteq12}, the following expressions of the reduced density operators in the Rindler wedges are obtained
\begin{equation}\label{gteq28}
\widehat{\rho}_{\rm R}=\frac{1}{Z_{\rm R}}\exp \left[-\frac{\widehat{\Pi}_{\rm R}}{T_0}\right],\qquad
\widehat{\Pi}_{\rm R}=\int_0^{+\infty}{\rm d} \omega \int_{-\infty}^{+\infty}{\rm d}^2{\rm k_T}\,\omega \widehat{a}^{\rm R\dagger}_{\omega,{\bf k}_{\rm T}}\widehat{a}^{\rm R}_{\omega,{\bf k}_{\rm T}}
\end{equation}
\begin{equation}\label{gteq29}
\widehat{\rho}_{\rm L}=\frac{1}{Z_{\rm L}}\exp \left[\frac{\widehat{\Pi}_{\rm L}}{T_0}\right],\qquad
\widehat{\Pi}_{\rm L}=\int_0^{+\infty}{\rm d} \omega \int_{-\infty}^{+\infty}{\rm d}^2{\rm k_T}\,\omega \widehat{a}^{\rm L\dagger}_{\omega,{\bf k}_{\rm T}}\widehat{a}^{\rm L}_{\omega,{\bf k}_{\rm T}}.
\end{equation}
They are diagonal in the Rindler creation and annihilation operators, so they have the same vacuum as the quantum field \eqref{gteq13} and \eqref{gteq26}, the Rindler vacuum $|0_{\rm R}\rangle$.
Moreover, their being diagonal is very convenient for it allows thermal expectation values to be calculated by using standard methods.
\subsection{Thermal expectation values}
Operators of physical interest, such as the energy-momentum tensor, are quadratic in the quantum field, therefore their thermal expectation values are given in terms of thermal expectation values of products of creation and annihilation operators.
These can be calculated in the Rindler wedges by using standard methods, thanks to the reduced density operators \eqref{gteq28} and \eqref{gteq29} being diagonal.
In particular, in the right Rindler wedge, which is the subspace of our interest, they are \cite{Becattini:2017ljh}
\begin{subequations}
\begin{align}
\langle \widehat{a}^{\rm R\dagger}_{\omega,{\bf k}_{\rm T}}\widehat{a}^{\rm R}_{\omega',{\bf k}_{\rm T}'}\rangle=&n_{\rm B}\,\delta(\omega-\omega')\,\delta^2({\bf k}_{\rm T}-{\bf k}_{\rm T}')\label{gteq14a}\\
\langle \widehat{a}^{\rm R}_{\omega,{\bf k}_{\rm T}}\widehat{a}^{\rm R\dagger}_{\omega',{\bf k}_{\rm T}'}\rangle=&\left(n_{\rm B}+1\right)\delta(\omega-\omega')\,\delta^2({\bf k}_{\rm T}-{\bf k}_{\rm T}')\label{gteq14b}\\
\langle \widehat{a}^{\rm R}_{\omega,{\bf k}_{\rm T}}\widehat{a}^{\rm R}_{\omega',{\bf k}_{\rm T}'}\rangle=&0=\langle \widehat{a}^{\rm R\dagger}_{\omega,{\bf k}_{\rm T}}\widehat{a}^{\rm R\dagger}_{\omega',{\bf k}_{\rm T}'}\rangle \label{gteq14c},
\end{align}
\end{subequations}
where $n_{\rm B}$ is the Bose-Einstein distribution
\begin{equation}
n_{\rm B}=\frac{1}{{\rm e}^{\omega/T_0}-1}.
\end{equation}
We prove the first one as an example.
For the sake of simplicity, let us temporarily introduce $\beta_0\equiv 1/T_0$ as a new parameter unrelated to the 0-component of $\beta_{\mu}$, and define
\begin{equation}\label{gteq50}
\widehat{a}^{\rm R\dagger}_{\omega,{\bf k}_{\rm T}}(\beta_0)\equiv {\rm e}^{-\beta_0\widehat{\Pi}_{\rm R}}\widehat{a}^{\rm R\dagger}_{\omega,{\bf k}_{\rm T}}{\rm e}^{\beta_0\widehat{\Pi}_{\rm R}}.
\end{equation}
The derivative with respect to $\beta_0$ reads
\begin{equation}
\begin{split}
\frac{\partial}{\partial \beta_0}\widehat{a}^{\rm R\dagger}_{\omega,{\bf k}_{\rm T}}(\beta_0)=&
-[\widehat{\Pi}_{\rm R},\widehat{a}^{\rm R\dagger}_{\omega,{\bf k}_{\rm T}}(\beta_0)]=
-{\rm e}^{-\beta_0\widehat{\Pi}_{\rm R}}[\widehat{\Pi}_{\rm R},\widehat{a}^{\rm R\dagger}_{\omega,{\bf k}_{\rm T}}]{\rm e}^{\beta_0\widehat{\Pi}_{\rm R}}\\
=&-\omega {\rm e}^{-\beta_0\widehat{\Pi}_{\rm R}}\widehat{a}^{\rm R\dagger}_{\omega,{\bf k}_{\rm T}}{\rm e}^{\beta_0\widehat{\Pi}_{\rm R}}=
-\omega \widehat{a}^{\rm R\dagger}_{\omega,{\bf k}_{\rm T}}(\beta_0),
\end{split}
\end{equation}
whose solution is
\begin{equation}
\widehat{a}^{\rm R\dagger}_{\omega,{\bf k}_{\rm T}}(\beta_0)={\rm e}^{-\beta_0\omega }\widehat{a}^{\rm R\dagger}_{\omega,{\bf k}_{\rm T}}.
\end{equation}
By exploiting the cyclic property of the trace, we have
\begin{equation}
\begin{split}
\langle \widehat{a}^{\rm R\dagger}_{\omega,{\bf k}_{\rm T}}\widehat{a}^{\rm R}_{\omega',{\bf k}_{\rm T}'}\rangle=&
\mathrm{tr}_{\rm R}\left(\widehat{\rho}_{\rm R}\widehat{a}^{\rm R\dagger}_{\omega,{\bf k}_{\rm T}}\widehat{a}^{\rm R}_{\omega',{\bf k}_{\rm T}'}\right)=
\mathrm{tr}_{\rm R}\left(\frac{{\rm e}^{-\beta_0\widehat{\Pi}_{\rm R}}}{Z_{\rm R}}\widehat{a}^{\rm R\dagger}_{\omega,{\bf k}_{\rm T}}\widehat{a}^{\rm R}_{\omega',{\bf k}_{\rm T}'}\right)\\
=&\frac{1}{Z_{\rm R}}\mathrm{tr}_{\rm R}\left(\widehat{a}^{\rm R}_{\omega',{\bf k}_{\rm T}'}{\rm e}^{-\beta_0\widehat{\Pi}_{\rm R}}\widehat{a}^{\rm R\dagger}_{\omega,{\bf k}_{\rm T}}{\rm e}^{\beta_0\widehat{\Pi}_{\rm R}}{\rm e}^{-\beta_0\widehat{\Pi}_{\rm R}}\right)\\
=&\mathrm{tr}_{\rm R}\left(\frac{{\rm e}^{-\beta_0\widehat{\Pi}_{\rm R}}}{Z_{\rm R}}\widehat{a}^{\rm R}_{\omega',{\bf k}_{\rm T}'}\widehat{a}^{\rm R\dagger}_{\omega,{\bf k}_{\rm T}}(\beta_0)\right)=
{\rm e}^{-\beta_0\omega}\mathrm{tr}_{\rm R}\left(\widehat{\rho}_{\rm R}\widehat{a}^{\rm R}_{\omega',{\bf k}_{\rm T}'}\widehat{a}^{\rm R\dagger}_{\omega,{\bf k}_{\rm T}}\right)\\
=&{\rm e}^{-\beta_0\omega}\langle \widehat{a}^{\rm R\dagger}_{\omega,{\bf k}_{\rm T}}\widehat{a}^{\rm R}_{\omega',{\bf k}_{\rm T}'}+[\widehat{a}^{\rm R}_{\omega',{\bf k}_{\rm T}'},\widehat{a}^{\rm R\dagger}_{\omega,{\bf k}_{\rm T}}]\rangle \\
=&{\rm e}^{-\beta_0\omega}\left(\langle \widehat{a}^{\rm R\dagger}_{\omega,{\bf k}_{\rm T}}\widehat{a}^{\rm R}_{\omega',{\bf k}_{\rm T}'}\rangle+\delta(\omega-\omega')\,\delta^2({\bf k}_{\rm T}-{\bf k}_{\rm T}')\right),
\end{split}
\end{equation}
thus
\begin{equation}\label{gteq51}
\langle \widehat{a}^{\rm R\dagger}_{\omega,{\bf k}_{\rm T}}\widehat{a}^{\rm R}_{\omega',{\bf k}_{\rm T}'}\rangle \left(1-{\rm e}^{-\beta_0\omega}\right)=\delta(\omega-\omega')\,\delta^2({\bf k}_{\rm T}-{\bf k}_{\rm T}'),
\end{equation}
hence the result \eqref{gteq14a}.
Equations \eqref{gteq14b} and \eqref{gteq14c} are proven in the same way.
The $+1$ term in \eqref{gteq14b} stems from the commutation relations \eqref{gteq15} and gives rise to divergencies, therefore some renormalization will be needed at some point.
More on this will be said shortly, for now we just want to point out that, in the Rindler vacuum, the above thermal expectation values have the same form evaluated at $T_0=0$.
This means that renormalization in the Rindler vacuum is but the subtraction of the $T_0=0$ contribution, which is tantamount to canceling the $+1$ term.
The structure of thermal expectation values depends on the symmetries of the density operator and by the quantities at disposal.
At global thermodynamic equilibrium, these are $b_{\mu}$, $\varpi_{\mu \nu}$, $x_{\mu}$ and $g_{\mu \nu}$, but as shown in \eqref{gteq17}, the dependence on $x_{\mu}$ can only be through $\beta_{\mu}$, which also includes $b_{\mu}$.
As for the thermal vorticity $\varpi_{\mu \nu}$, in our case of global thermodynamic equilibrium with acceleration, this is decomposed in terms of $u_{\mu}$ and $\alpha_{\mu}$ as in \eqref{gteq18}, and the dependence on $u_{\mu}$ is the same as on $\beta_{\mu}$.
The only non-vanishing component of the first derivative of $\beta_{\mu}$ is $\varpi_{\mu \nu}$, which is constant, therefore derivatives of higher order are identically null.
In summary, the most general structure of the thermal expectation value of the energy-momentum tensor is
\begin{equation}
\langle \widehat{T}^{\mu \nu}\rangle=F_1\beta^{\mu}\beta^{\nu}+F_2g^{\mu \nu}+F_3\alpha^{\mu}\alpha^{\nu}+F_4(\alpha^{\mu}\beta^{\nu}+\alpha^{\nu}\beta^{\mu}).
\end{equation}
The only Lorentz scalar we can build with $\beta^{\mu}$ and $\alpha^{\mu}$ are $\beta^2=1/T^2$ and $\alpha^2=-a^2/T_0^2$, being $\beta^{\mu}$ and $\alpha^{\mu}$ mutually orthogonal.
Thus, the scalar functions $F_1$ and $F_2$ will depend on $\beta^2$ and $\alpha^2$, while $F_3$ and $F_4$ are expected to be directly proportional to $\alpha^2$ only as the ideal form of $\langle \widehat{T}^{\mu \nu}\rangle$ ought to be recovered for vanishing acceleration field.
On the other hand, in our case of free real scalar field theory, the Hamiltonian $\widehat{H}$ in \eqref{gteq04} is even under time-reversal, thus so is the density operator.
In formulae, $\widehat{\Theta}\widehat{\rho}\widehat{\Theta}^{-1}=\widehat{\rho}$, where $\widehat{\Theta}$ is the time-reversal operator.
Care must be taken that $\widehat{\Theta}$ depends on the hypersurface of simultaneity with respect to which time is reflected, so the above equation and the following ones should be understood with respect to the $t=0$ hypersurface.
On the other hand, the momentum $\widehat{T}^{0i}$ is odd under time-reversal, i.e.\ $\widehat{\Theta}\widehat{T}^{0i}\widehat{\Theta}^{-1}=-\widehat{T}^{0i}$.
Together with the evenness of $\widehat{\rho}$ and the cyclicity of the trace, this implies that $\langle \widehat{T}^{0i}\rangle=0$; conversely, $\langle \widehat{T}^{00}\rangle$ and $\langle \widehat{T}^{ij}\rangle$ do not vanish.
Now, in the above equation $\alpha^i$ and $\beta^0$ do not vanish at $t=0$, which means that the coupling between $\alpha^{\mu}$ and $\beta^{\mu}$ breaks time-reversal symmetry, therefore $F_4$ must be identically zero, and we are left with
\begin{equation}
\langle \widehat{T}^{\mu \nu}\rangle=F_1\beta^{\mu}\beta^{\nu}+F_2g^{\mu \nu}+F_3\alpha^{\mu}\alpha^{\nu}.
\end{equation}
The functions $F_i$ can be renamed using a more familiar notation in order to make their physical meaning more apparent
\begin{equation}\label{gteq45}
\langle \widehat{T}^{\mu \nu}\rangle=\rho u^{\mu}u^{\nu}-p\Delta^{\mu \nu}+kA^{\mu}A^{\nu}.
\end{equation}
Here, $\rho$ and $p$ reconstruct the ideal form, so they are the energy density and the pressure respectively, while $k$ is an anisotropic pressure term.
Interestingly enough, although being a non-ideal term, $k$ is not technically a dissipation, for entropy is not produced at global thermodynamic equilibrium.
Explicitly, they read
\begin{equation}
\rho=\langle \widehat{T}^{\mu \nu}\rangle u_{\mu}u_{\nu},
\end{equation}
\begin{equation}
p=\frac{1}{2}\langle \widehat{T}^{\mu \nu}\rangle \left(\frac{A_{\mu}A_{\nu}}{A^2}-\Delta_{\mu \nu}\right),
\end{equation}
\begin{equation}
k=\frac{1}{2A^2}\langle \widehat{T}^{\mu \nu}\rangle \left(3\frac{A_{\mu}A_{\nu}}{A^2}-\Delta_{\mu \nu}\right).
\end{equation}
Their expressions depend on the specific Quantum Field Theory underlying the hydrodynamic theory, so we now calculate them in our case of free real scalar field theory in the right Rindler wedge.
We take the energy density $\rho$ as an example, $p$ and $k$ are worked out following analogous steps.
As for the energy density, note that we can either take the thermal expectation value $\langle \widehat{T}^{\mu \nu}\rangle$ first and then project onto $u_{\mu}u_{\mu}$ or vieceversa, so let us first project onto $u_{\mu}u_{\nu}$ and then take the thermal expectation value $\langle \widehat{T}^{\mu \nu}u_{\mu}u_{\nu}\rangle$.
Using the canonical energy-momentum tensor \eqref{gteq19}, we see that the energy density is given by the difference of two terms
\begin{equation}\label{gteq20}
\rho=\langle (u^{\mu}\partial_{\mu}\widehat{\psi})^2\rangle-\langle \widehat{\cal L}\rangle.
\end{equation}
Let us focus on the first term.
In Rindler coordinates, the convective derivative is simply $u^{\mu}\partial_{\mu}={\rm e}^{-a\xi}\partial_{\tau}$, so with the quantum field expansion \eqref{gteq13} and the equations \eqref{gteq14a}--\eqref{gteq14c} we come to
\begin{equation}\label{gteq35}
\langle (u^{\mu}\partial_{\mu}\widehat{\psi})^2\rangle=
\int_0^{+\infty}{\rm d} \omega \int_{-\infty}^{+\infty}{\rm d}^2{\rm k_T}\,\omega^2|u_{\omega,{\bf k}_{\rm T}}|^2(2n_{\rm B}+1).
\end{equation}
The $+1$ term stemming from the commutation relations \eqref{gteq15} gives rise to divergencies and needs renormalization.
More on the renormalization scheme will be said shortly, for the time being let us renormalize with respect to the Rindler vacuum $|0_{\rm R}\rangle$, thus subtract the $T_0=0$ contribution, which is tantamount to cancel the $+1$ term.
Then
\begin{equation}
\begin{split}
\langle (u^{\mu}\partial_{\mu}\widehat{\psi})^2\rangle&-\langle 0_{\rm R}|(u^{\mu}\partial_{\mu}\widehat{\psi})^2|0_{\rm R}\rangle=
\frac{{\rm e}^{-2a\xi}}{2\pi^4a}\int_0^{+\infty}{\rm d} \omega \,\omega^2\sinh \left(\frac{\pi \omega}{a}\right)n_{\rm B}\times \\
&\times \int_{-\infty}^{+\infty}{\rm d}^2{\rm k_T}\,{\rm K}_{i\frac{\omega}{a}}\left(\frac{m_{\rm T} {\rm e}^{a\xi}}{a}\right){\rm K}_{-i\frac{\omega}{a}}\left(\frac{m_{\rm T} {\rm e}^{a\xi}}{a}\right),
\end{split}
\end{equation}
where ${\rm K}_{i\frac{\omega}{a}}^*={\rm K}_{-i\frac{\omega}{a}}$ was used.
In the massless limit we have $m_{\rm T}=|{\bf k}_{\rm T}|$, and the integral in the transverse momentum can be carried out analytically
\begin{equation}\label{gteq21}
\begin{split}
&\int_{-\infty}^{+\infty}{\rm d}^2{\rm k_T}\,{\rm K}_{i\frac{\omega}{a}}\left(\frac{m_{\rm T} {\rm e}^{a\xi}}{a}\right){\rm K}_{-i\frac{\omega}{a}}\left(\frac{m_{\rm T} {\rm e}^{a\xi}}{a}\right)\\
&=\pi a^2{\rm e}^{-2a\xi}\Gamma \left(1+i\frac{\omega}{a}\right)\Gamma \left(1-i\frac{\omega}{a}\right)=
\pi^2 a{\rm e}^{-2a\xi}\frac{\omega}{\sinh \left(\frac{\pi \omega}{a}\right)}
\end{split}
\end{equation}
with $\Gamma$ the Gamma function, hence
\begin{equation}
\langle (u^{\mu}\partial_{\mu}\widehat{\psi})^2\rangle-\langle 0_{\rm R}|(u^{\mu}\partial_{\mu}\widehat{\psi})^2|0_{\rm R}\rangle=
\frac{{\rm e}^{-4a\xi}}{2\pi^2}\int_0^{+\infty}{\rm d} \omega \,\frac{\omega^3}{{\rm e}^{\omega/T_0}-1}.
\end{equation}
This integral is carried out by exploiting the standard trick involving the geometric series
\begin{equation}
\begin{split}
&\int_0^{+\infty}{\rm d} \omega \frac{\omega^3}{{\rm e}^{\omega/T_0}-1}=
\int_0^{+\infty}{\rm d} \omega \,\omega^3\sum_{n=1}^{+\infty}{\rm e}^{-n\frac{\omega}{T_0}}=
\sum_{n=1}^{+\infty}\int_0^{+\infty}{\rm d} \omega \,\omega^3{\rm e}^{-n\frac{\omega}{T_0}}\\
&=T_0^4\sum_{n=1}^{+\infty}\frac{1}{n^4}\int_0^{+\infty}{\rm d} x\,x^3{\rm e}^{-x}=
6T_0^4\sum_{n=1}^{+\infty}\frac{1}{n^4}=
6T_0^4\zeta(4)=
\frac{\pi^4}{15}T_0^4,
\end{split}
\end{equation}
with $\zeta$ the Riemann zeta function.
The order of the integration and the series could be exchanged thanks to the convergence of the series.
Thus, we obtain the result found in \cite{Becattini:2017ljh}
\begin{equation}
\langle (u^{\mu}\partial_{\mu}\widehat{\psi})^2\rangle-\langle 0_{\rm R}|(u^{\mu}\partial_{\mu}\widehat{\psi})^2|0_{\rm R}\rangle=
\frac{\pi^2}{30}T_0^4{\rm e}^{-4a\xi}=
\frac{\pi^2}{30\beta^4}=
\frac{\pi^2}{30}T^4,
\end{equation}
where we used the expression of the proper temperature in Rindler coordinates
\begin{equation}\label{gteq22}
\beta^2=\frac{a^2}{T_0^2}\left({z'}^2-t^2\right)=\frac{{\rm e}^{2a\xi}}{T_0^2}.
\end{equation}
As for the second term in \eqref{gteq20}, it is convenient to use the Klein-Gordon equation of motion to rewrite it as
\begin{equation}
\langle \widehat{\cal L}\rangle=\frac{1}{4}\langle \Box \widehat{\psi}^2\rangle=\frac{1}{4}\Box \langle \widehat{\psi}^2\rangle.
\end{equation}
The D'Alambert operator can be taken out of the thermal expectation value because $\Box \widehat{\rho}=0$ at global thermodynamic equilibrium with acceleration.
Using the quantum field expansion \eqref{gteq13} and the equations \eqref{gteq14a}--\eqref{gteq14c}, we readily obtain
\begin{equation}\label{gteq36}
\langle \widehat{\psi}^2\rangle=
\int_0^{+\infty}{\rm d} \omega \int_{-\infty}^{+\infty}{\rm d}^2{\rm k_T}\,|u_{\omega,{\bf k}_{\rm T}}|^2(2n_{\rm B}+1).
\end{equation}
Once again, the $+1$ term stemming from the commutation relations \eqref{gteq15} gives rise to divergencies and must be renormalized.
In the Rindler vacuum renormalization scheme, we have
\begin{equation}
\begin{split}
\langle \widehat{\psi}^2\rangle&-\langle 0_{\rm R}|\widehat{\psi}^2|0_{\rm R}\rangle=
\frac{1}{2\pi^4a}\int_0^{+\infty}{\rm d} \omega \,\sinh \left(\frac{\pi \omega}{a}\right)n_{\rm B}\times \\
&\times \int_{-\infty}^{+\infty}{\rm d}^2{\rm k_T}\,{\rm K}_{i\frac{\omega}{a}}\left(\frac{m_{\rm T} {\rm e}^{a\xi}}{a}\right){\rm K}_{-i\frac{\omega}{a}}\left(\frac{m_{\rm T} {\rm e}^{a\xi}}{a}\right)
\end{split}
\end{equation}
The integral in the transverse momentum is performed analytically in the massless limit by using \eqref{gteq21}, hence
\begin{equation}
\langle \widehat{\psi}^2\rangle-\langle 0_{\rm R}|\widehat{\psi}^2|0_{\rm R}\rangle=
\frac{{\rm e}^{-2a\xi}}{2\pi^2}\int_0^{+\infty}{\rm d} \omega \frac{\omega}{{\rm e}^{\omega/T_0}-1}=
\frac{T_0^2}{12}{\rm e}^{-2a\xi}.
\end{equation}
In the last step, the integral was carried out by using again the geometric series trick as above, with now $\zeta(2)$ instead of $\zeta(4)$.
The D'Alambert operator in Rindler coordinates reads
\begin{equation}
\Box={\rm e}^{-2a\xi}\left(\partial_{\tau}^2-\partial_{\xi}^2\right)-\left(\partial_x^2+\partial_y^2\right),
\end{equation}
hence the result found in \cite{Becattini:2019poj}
\begin{equation}
\langle \widehat{\cal L}\rangle-\langle 0_{\rm R}|\widehat{\cal L}|0_{\rm R}\rangle=
\frac{1}{4}\Box \left(\langle \widehat{\psi}^2\rangle-\langle 0_{\rm R}|\widehat{\psi}^2|0_{\rm R}\rangle \right)=
-\frac{a^2T_0^2}{12}{\rm e}^{-4a\xi}=
\frac{\alpha^2}{12\beta^4}=
\frac{\alpha^2}{12}T^4,
\end{equation}
where \eqref{gteq22} and \eqref{gteq23} were used to express the result in terms of $\alpha^2$ and $\beta^2$.
In summary, the energy density for a free real scalar field in the massless limit renormalized with respect to the Rindler vacuum is
\begin{equation}
\rho_{\rm R}\equiv
\left(\langle \widehat{T}^{\mu \nu}\rangle-\langle 0_{\rm R}|\widehat{T}^{\mu \nu}|0_{\rm R}\rangle \right)u_{\mu}u_{\nu}=
\frac{\pi^2}{30\beta^4}-\frac{\alpha^2}{12\beta^4}=
\left(\frac{\pi^2}{30}-\frac{\alpha^2}{12}\right)T^4.
\end{equation}
This result was found in \cite{Buzzegoli:2017cqy, Becattini:2015nva} with a perturbative expansion of the density operator \eqref{gteq04} in $\alpha^{\mu}$ at order $\alpha^2$, so we conclude that the perturbative series for the free real scalar field is simply a polynomial in $\alpha^{\mu}$ of order 2.
The pressure $p$ and the anisotropic pressure $k$ can be worked out following similar steps.
By defining their values in the same renormalization scheme as
\begin{equation}
p_{\rm R}\equiv \frac{1}{2}\left(\langle \widehat{T}^{\mu \nu}\rangle-\langle 0_{\rm R}|\widehat{T}^{\mu \nu}|0_{\rm R}\rangle \right)\left(\frac{A_{\mu}A_{\nu}}{A^2}-\Delta_{\mu \nu}\right)
\end{equation}
\begin{equation}
k_{\rm R}\equiv \frac{1}{2A^2}\left(\langle \widehat{T}^{\mu \nu}\rangle-\langle 0_{\rm R}|\widehat{T}^{\mu \nu}|0_{\rm R}\rangle \right)\left(3\frac{A_{\mu}A_{\nu}}{A^2}-\Delta_{\mu \nu}\right)
\end{equation}
the final expressions read
\begin{equation}\label{gteq31}
\rho_{\rm R}=
\frac{\pi^2}{30\beta^4}-\frac{\alpha^2}{12\beta^4}=
\left(\frac{\pi^2}{30}-\frac{\alpha^2}{12}\right)T^4,
\qquad (m=0)
\end{equation}
\begin{equation}\label{gteq32}
p_{\rm R}=\frac{\pi^2}{90\beta^4}+\frac{\alpha^2}{18\beta^4}=
\left(\frac{\pi^2}{90}+\frac{\alpha^2}{18}\right)T^4,
\qquad (m=0)
\end{equation}
\begin{equation}\label{gteq33}
k_{\rm R}=-\frac{\alpha^2}{12\beta^4}=-\frac{\alpha^2}{12}T^4,
\qquad (m=0).
\end{equation}
This is the thermal expectation value of the energy-momentum tensor of a free real scalar field at global thermodynamic equilibrium in the right Rindler wedge in the massless limit.
As expected, for vanishing acceleration field the anisotropic pressure term vanishes, the ideal form is recovered and the energy density and the isotropic pressure take on their familiar forms.
Let us also emphasize that, by definition of $\alpha^{\mu}$ \eqref{gteq24} and by equation \eqref{gteq25}, the $\alpha^2$ terms are quantum corrections, in the sense that they are proportional to $\hbar^2$ thus vanishing in the limit $\hbar \to 0$.
\subsection{Particles and vacuum: the Unruh effect}
\label{sec:gte_unruh_effect}
As we had occasion to convince ourselves a few times already, thermal expectation values of operators quadratic in the quantum field are divergent owing to the $+1$ term in \eqref{gteq14b} stemming from the commutation relations of creation and annihilation operators \eqref{gteq15}.
In free field theory, they are usually renormalized by subtracting the vacuum expectation value; however,
the vacuum state is defined by the creation and annihilation operators, which in turn depend on the choice of the positive-frequency modes.
As it is known, especially in the context of Quantum Field Theory in curved spacetime, such a choice is not unique in general, therefore we face an ambiguity.
In our case, the quantum field could either be expanded in the whole Minkowski spacetime in the usual plane waves form as in \eqref{gteq27}, or in the two Rindler wedges in terms of the Bessel functions as in \eqref{gteq13} and \eqref{gteq26}.
Consequently, it is found that Rindler creation and annihilation operators are related to the Minkowski ones by a non-trivial Bogolyubov transformation, therefore the Rindler vacuum $|0_{\rm R}\rangle$ and the Minkowski vacuum $|0_{\rm M}\rangle$ are two different states, as first pointed out by Fulling \cite{Fulling:1972md}.
In particular, if we take the standpoint of the Rindler observer, the Minkowski vacuum is seen as a thermal state of particles at the temperature $T_0=\frac{a}{2\pi}$
\begin{equation}\label{gteq34}
\langle 0_{\rm M}|\widehat{a}^{\rm R\dagger}_{\omega,{\bf k}_{\rm T}}\widehat{a}^{\rm R}_{\omega',{\bf k}_{\rm T}'}|0_{\rm M}\rangle=
\langle 0_{\rm M}|\widehat{a}^{\rm L\dagger}_{\omega,{\bf k}_{\rm T}}\widehat{a}^{\rm L}_{\omega',{\bf k}_{\rm T}'}|0_{\rm M}\rangle=
\frac{1}{{\rm e}^{\frac{2\pi}{a}\omega}-1}\delta(\omega-\omega')\,\delta^2({\bf k}_{\rm T}-{\bf k}_{\rm T}').
\end{equation}
This is the famous \textit{Unruh effect}, and $T_0=\frac{a}{2\pi}$ is called the \textit{Unruh temperature} \cite{Unruh:1976db}.
In short, the Unruh effect states that uniformly accelerated observers in Minkowski spacetime, i.e.\ linearly accelerated observers with constant proper acceleration called the Rindler observers, associate a thermal bath of Rindler particles to the no-particle state of inertial observers, i.e.\ the Minkowski vacuum.
Rindler particles are associated with positive-frequency modes as defined by Rindler observers in contrast to Minkowski particles, which are associated with positive-frequency modes as defined by inertial observers.
This is a conceptually subtle Quantum Field Theory result, which has played a crucial role in our understanding that the particle content of a field theory is, in this sense, observer-dependent.
In view of these findings, which vacuum should we choose to renormalize thermal expectation values?
The answer is that, as long as both choices give finite results, it depends on which kind of observer we want to take the standpoint of.
We have seen that taking the point of view of the Rindler observer, namely renormalizing with respect to the Rindler vacuum, means the subtraction of the $T_0=0$ contribution, which is tantamount to neglecting the divergent $+1$ term in equation \eqref{gteq14b}.
In this renormalization scheme, the thermal expectation value of the energy-momentum tensor is given in \eqref{gteq31}--\eqref{gteq33}.
On the other hand, from \eqref{gteq34} it is clear that thermal expectation values of products of Rindler creation and annihilation operators in the Minkowski vacuum take on the same form as \eqref{gteq14a}--\eqref{gteq14c} calculated at $T_0=\frac{a}{2\pi}$.
Therefore, taking the standpoint of the inertial observer, that is renormalizing with respect to the Minkowski vacuum, means the subtraction of the $T_0=\frac{a}{2\pi}$ contribution.
The former eventuality has already been studied, so let us focus here on the latter.
In the Minkowski renormalization scheme, equations \eqref{gteq14a}--\eqref{gteq14c} read
\begin{equation}
\langle \widehat{a}^{\rm R\dagger}_{\omega,{\bf k}_{\rm T}}\widehat{a}^{\rm R}_{\omega',{\bf k}_{\rm T}'}\rangle-\langle 0_{\rm M}|\widehat{a}^{\rm R\dagger}_{\omega,{\bf k}_{\rm T}}\widehat{a}^{\rm R}_{\omega',{\bf k}_{\rm T}'}|0_{\rm M}\rangle=
\left(\frac{1}{{\rm e}^{\omega/T_0}-1}-\frac{1}{{\rm e}^{\frac{2\pi}{a}\omega}-1}\right)\delta(\omega-\omega')\,\delta^2({\bf k}_{\rm T}-{\bf k}_{\rm T}')
\end{equation}
\begin{equation}
\langle \widehat{a}^{\rm R}_{\omega,{\bf k}_{\rm T}}\widehat{a}^{\rm R\dagger}_{\omega',{\bf k}_{\rm T}'}\rangle-\langle 0_{\rm M}|\widehat{a}^{\rm R}_{\omega,{\bf k}_{\rm T}}\widehat{a}^{\rm R\dagger}_{\omega',{\bf k}_{\rm T}'}|0_{\rm M}\rangle=
\left(\frac{1}{{\rm e}^{\omega/T_0}-1}-\frac{1}{{\rm e}^{\frac{2\pi}{a}\omega}-1}\right)\delta(\omega-\omega')\,\delta^2({\bf k}_{\rm T}-{\bf k}_{\rm T}')
\end{equation}
\begin{equation}
\langle \widehat{a}^{\rm R}_{\omega,{\bf k}_{\rm T}}\widehat{a}^{\rm R}_{\omega',{\bf k}_{\rm T}'}\rangle-\langle 0_{\rm M}|\widehat{a}^{\rm R}_{\omega,{\bf k}_{\rm T}}\widehat{a}^{\rm R}_{\omega',{\bf k}_{\rm T}'}|0_{\rm M}\rangle=0=\langle \widehat{a}^{\rm R\dagger}_{\omega,{\bf k}_{\rm T}}\widehat{a}^{\rm R\dagger}_{\omega',{\bf k}_{\rm T}'}\rangle-\langle 0_{\rm M}|\widehat{a}^{\rm R\dagger}_{\omega,{\bf k}_{\rm T}}\widehat{a}^{\rm R\dagger}_{\omega',{\bf k}_{\rm T}'}|0_{\rm M}\rangle.
\end{equation}
These expressions are suggestive as they make it apparent that the thermal expectation value of any operator quadratic in the quantum field, such as the energy-momentum tensor, vanish when the temperature is as low as the Unruh temperature.
By looking at non-renormalized equations \eqref{gteq35} and \eqref{gteq36}, it is clear that the energy density not only vanishes at $T_0=\frac{a}{2\pi}$, but becomes negative for lower temperatures.
And this is actually not a prerogative of the energy density, but a general feature of operators quadratic in the quantum field.
Thus, in this sense, the Unruh temperature is an absolute lower bound for the temperature, a result found in \cite{Becattini:2017ljh}.
It is important to stress that this conclusion holds locally, for a comoving observer.
From the expression of the proper temperature $T$ \eqref{gteq09} and the acceleration field $A^{\mu}$ \eqref{gteq37}, it is readily seen that in general one can write $T=T_0\sqrt{|A^2|}/a$, so, at the level of the proper temperature, the bound $T_0\ge \frac{a}{2\pi}$ implies
\begin{equation}
T\ge T_{\rm U}\equiv \frac{\sqrt{|A^2|}}{2\pi},
\end{equation}
where $T_{\rm U}$ is called the \textit{comoving} or \textit{proper Unruh temperature}.
That is, the temperature measured by a comoving thermometer cannot be lower than the magnitude of the acceleration field divided by $2\pi$.
Recall that the magnitude of the acceleration field is constant along a fixed flow line and varies from one flow line to another, therefore so does the proper Unruh temperature, thus behaving like the proper temperature itself.
The feature of the proper Unruh temperature being a lower bound for the proper temperature is apparent in the thermal expectation value of the energy-momentum tensor renormalized with repsect to the Minkowski vacuum.
By defining
\begin{equation}
\rho_{\rm M}\equiv \left(\langle \widehat{T}^{\mu \nu}\rangle-\langle 0_{\rm M}|\widehat{T}^{\mu \nu}|0_{\rm M}\rangle \right)u_{\mu}u_{\nu},
\end{equation}
\begin{equation}
p_{\rm M}\equiv \frac{1}{2}\left(\langle \widehat{T}^{\mu \nu}\rangle-\langle 0_{\rm M}|\widehat{T}^{\mu \nu}|0_{\rm M}\rangle \right)\left(\frac{A_{\mu}A_{\nu}}{A^2}-\Delta_{\mu \nu}\right),
\end{equation}
\begin{equation}
k_{\rm M}\equiv \frac{1}{2A^2}\left(\langle \widehat{T}^{\mu \nu}\rangle-\langle 0_{\rm M}|\widehat{T}^{\mu \nu}|0_{\rm M}\rangle \right)\left(3\frac{A_{\mu}A_{\nu}}{A^2}-\Delta_{\mu \nu}\right),
\end{equation}
one can find the following results in the massless case
\begin{equation}
\rho_{\rm M}=
\left(\frac{\pi^2}{30}-\frac{\alpha^2}{12}\right)(T^4-T_{\rm U}^4),
\qquad (m=0)
\end{equation}
\begin{equation}
p_{\rm M}=
\left(\frac{\pi^2}{90}+\frac{\alpha^2}{18}\right)(T^4-T_{\rm U}^4),
\qquad (m=0)
\end{equation}
\begin{equation}
k_{\rm M}=
-\frac{\alpha^2}{12}(T^4-T_{\rm U}^4),
\qquad (m=0).
\end{equation}
Although the conclusion that scalar thermodynamic functions renormalized with respect to the Minkowski vacuum can be expressed in the above fashion was obtained for a free field theory, it is likely to hold for any interacting field theory as well.
In fact, the Unruh effect was derived for general interacting field theories in \cite{Bisognano:1975ih, Bisognano:1976za} within an axiomatic quantum field theory approach by taking advantage of the KMS feature of the expectation values for the density operator at hand.
For a recent discussion, see also \cite{Gransee:2015aba, Gransee:2016hep}.
\section{Entropy current in the right Rindler wedge}
\label{sec:gte_entropy_current_in_RRW}
With the thermal expectation value of the energy-momentum tensor, we are now in a position to use our method to calculate the entropy current including quantum corrections.
Its divergence will provide information on the entropy production rate, which is expected to vanish at global thermodynamic equilibrium, while its integral will give us the total entropy.
As we shall see, thanks to the factorization property of the density operator, the entropy thereby obtained will be interestingly related to some entanglement entropy.
\subsection{Entropy current calculation}
As a quick recap, in order to apply the method for the calculation of the entropy current put forward in Section \ref{sec:zubarev_entropy_current_method}, we need two ingredients:
\begin{enumerate}
\item the thermal expectation value the energy-momentum tensor $\langle \widehat{T}^{\mu \nu}\rangle$, and
\item the eigenvector $|0_{\Upsilon}\rangle$ corresponding to the lowest, non-degenerate eigenvalue of $\widehat{\Upsilon}$.
\end{enumerate}
Then, the algorithm is the following:
\begin{enumerate}
\item Take $\langle \widehat{T}^{\mu \nu}\rangle$ and subtract $\langle 0_{\Upsilon}|\widehat{T}^{\mu \nu}|0_{\Upsilon}\rangle$ by using the $\lambda$-dependent density operator $\widehat{\rho}_{\rm R}(\lambda)$ defined according to \eqref{zubeq21}.
This way, the result is $\lambda$-dependent.
\item Contract with $\beta_{\nu}$, which is $\lambda$-independent, and integrate in $\lambda$ from $\lambda=1$ to $\lambda=+\infty$ in order to obtain the thermodynamic potential current $\phi^{\mu}$ defined in \eqref{zubeq26}.
\item Plug the result into \eqref{zubeq27} and obtain the entropy current $s^{\mu}$.
\end{enumerate}
By comparing the definition \eqref{zubeq23} of the operator $\widehat{\Upsilon}$ with the expression \eqref{gteq28} of the reduced density operator at global thermodynamic equilibrium with acceleration in the right Rindler wedge, we immediately understand that
\begin{equation}\label{gteq38}
\widehat{\Upsilon}=\frac{\widehat{\Pi}_{\rm R}}{T_0}=
\frac{1}{T_0}\int_0^{+\infty}{\rm d} \omega \int_{-\infty}^{+\infty}{\rm d}^2{\rm k_T}\,\omega \widehat{a}^{\rm R\dagger}_{\omega,{\bf k}_{\rm T}}\widehat{a}^{\rm R}_{\omega,{\bf k}_{\rm T}}.
\end{equation}
Thus, the lowest eigenvector of $\widehat{\Upsilon}$ is the Rindler vacuum, $|0_{\Upsilon}\rangle=|0_{\rm R}\rangle$, which is non-degenerate, and the corresponding eigenvalue is $\Upsilon_0=0$.
Hence,
\begin{equation}
\langle \widehat{T}^{\mu \nu}\rangle-\langle 0_{\Upsilon}|\widehat{T}^{\mu \nu}|0_{\Upsilon}\rangle=
\langle \widehat{T}^{\mu \nu}\rangle-\langle 0_{\rm R}|\widehat{T}^{\mu \nu}|0_{\rm R}\rangle,
\end{equation}
which was already calculated in \eqref{gteq31}--\eqref{gteq33} for a massless field, so we are good to go for what concerns the ingredients.
Now, according to the method, the above quantity should be calculated with the modified density operator $\widehat{\rho}_{\rm R}(\lambda)$ in order to be $\lambda$-dependent.
Let us take a closer look at what this operation actually means at a practical level.
By comparing equation \eqref{zubeq21} with \eqref{gteq04} or with \eqref{gteq38}, we can tell that in our case the introduction of $\lambda$ is but a rescaling of $T_0$ as $T_0\mapsto T_0/\lambda$.
It can be easily checked that this transformation gives precisely the desired result.
Thus, we have
\begin{equation}
\langle \widehat{T}^{\mu \nu}\rangle(\lambda)-\langle 0_{\Upsilon}|\widehat{T}^{\mu \nu}|0_{\Upsilon}\rangle(\lambda)=
\rho_{\rm R}(\lambda)u^{\mu}u^{\nu}-p_{\rm R}(\lambda)\Delta^{\mu \nu}+k_{\rm R}(\lambda)A^{\mu}A^{\nu},
\end{equation}
where the rescaling $T_0\mapsto T_0/\lambda$ affects only $\rho_{\rm R}$, $p_{\rm R}$ and $k_{\rm R}$ because $u^{\mu}$, $\Delta^{\mu \nu}$ and $A^{\mu}$ are independent of $T_0$.
This result should now be contracted with $\beta_{\nu}$.
Care must be taken that this $\beta_{\nu}$ does not undergo the temperature rescaling: it concerns only the thermal expectation value $\langle \widehat{T}^{\mu \nu}\rangle-\langle 0_{\Upsilon}|\widehat{T}^{\mu \nu}|0_{\Upsilon}\rangle$, as can be clearly seen following the steps of the method derivation in Section \ref{sec:zubarev_entropy_current_method}.
Whence
\begin{equation}
\left(\langle \widehat{T}^{\mu \nu}\rangle(\lambda)-\langle 0_{\Upsilon}|\widehat{T}^{\mu \nu}|0_{\Upsilon}\rangle(\lambda)\right)\beta_{\nu}=
\rho_{\rm R}(\lambda)\beta^{\mu}.
\end{equation}
In other words, we do not need the thermal expectation value of the whole energy-momentum tensor, only the energy density contributes to the thermodynamic potential current and the entropy current.
The expression of $\rho_{\rm R}(\lambda)$ is obtained from \eqref{gteq31} by making explicit the $T_0$-dependence using \eqref{gteq09} and \eqref{gteq23}.
To this purpose, it is also convenient to extract the $T_0$-dependence from the four-temperature $\beta^{\mu}$ of \eqref{gteq08} by defining the Killing vector field $\gamma^{\mu}\equiv T_0\beta^{\mu}$, which is $T_0$-independent, hence
\begin{equation}
\rho_{\rm R}=
\left(\frac{\pi^2}{30}-\frac{\alpha^2}{12}\right)\frac{1}{\beta^4}=
\left(\frac{\pi^2}{30}+\frac{a^2}{12T_0^2}\right)\frac{T_0^4}{\gamma^4}=
\frac{\pi^2}{30\gamma^4}T_0^4+\frac{a^2}{12\gamma^4}T_0^2,
\end{equation}
thus
\begin{equation}
\rho_{\rm R}(\lambda)=
\frac{\pi^2}{30\gamma^4}\frac{T_0^4}{\lambda^4}+\frac{a^2}{12\gamma^4}\frac{T_0^2}{\lambda^2}.
\end{equation}
According to \eqref{zubeq26}, the thermodynamic potential current is readily obtained
\begin{equation}
\begin{split}
\phi^{\mu}=&\beta^{\mu}\int_1^{+\infty}{\rm d} \lambda \,\rho_{\rm R}(\lambda)=\left(\frac{\pi^2T_0^4}{30\gamma^4}\int_1^{+\infty}\frac{{\rm d} \lambda}{\lambda^4}+\frac{a^2T_0^2}{12\gamma^4}\int_1^{+\infty}\frac{{\rm d} \lambda}{\lambda^2}\right)\beta^{\mu}\\
=&\left(\frac{\pi^2}{90}\frac{T_0^4}{\gamma^4}+\frac{a^2}{12T_0^2}\frac{T_0^4}{\gamma^4}\right)\beta^{\mu}=
\left(\frac{\pi^2}{90\beta^4}-\frac{\alpha^2}{12\beta^4}\right)\beta^{\mu}.
\end{split}
\end{equation}
Plugging this result into \eqref{zubeq27}, we finally have the entropy current
\begin{equation}
\begin{split}
s^{\mu}=&
\phi^{\mu}+\left(\langle \widehat{T}^{\mu \nu}\rangle-\langle 0_{\rm R}|\widehat{T}^{\mu \nu}|0_{\rm R}\rangle \right)\beta_{\nu}=
\phi^{\mu}+\rho_{\rm R}\beta^{\mu}\\
=&\left(\frac{\pi^2}{90\beta^4}-\frac{\alpha^2}{12\beta^4}+\frac{\pi^2}{30\beta^4}-\frac{\alpha^2}{12\beta^4}\right)\beta^{\mu}=
\left(\frac{2\pi^2}{45\beta^4}-\frac{\alpha^2}{6\beta^4}\right)\beta^{\mu}.
\end{split}
\end{equation}
In summary,
\begin{equation}
\phi^{\mu}=
\left(\frac{\pi^2}{90\beta^4}-\frac{\alpha^2}{12\beta^4}\right)\beta^{\mu}=
\left(\frac{\pi^2}{90}-\frac{\alpha^2}{12}\right)T^3u^{\mu},\qquad (m=0)
\end{equation}
\begin{equation}\label{gteq39}
s^{\mu}=
\left(\frac{2\pi^2}{45\beta^4}-\frac{\alpha^2}{6\beta^4}\right)\beta^{\mu}=
\left(\frac{2\pi^2}{45}-\frac{\alpha^2}{6}\right)T^3u^{\mu},\qquad (m=0).
\end{equation}
These are the final expressions of the thermodynamic potential current and the entropy current of a free real scalar massless field at global thermodynamic equilibrium with acceleration in the right Rindler wedge \cite{Becattini:2019poj}.
Now some comments are in order.
First of all, we stress that, by definition of $\alpha^{\mu}$ \eqref{gteq24} and by equation \eqref{gteq25}, the $\alpha^2$ terms are quantum corrections, in the sense that they are proportional to $\hbar^2$ thus vanishing in the limit $\hbar \to 0$.
It is also easy to check that the entropy current, and the thermodynamic potential current as well, actually, are divergenceless.
In fact, both of the currents are given by a constant term multiplied by $\beta^{\mu}/\beta^4$, so
\begin{equation}
\begin{split}
\partial_{\mu}s^{\mu}\propto &\partial_{\mu}\phi^{\mu}\propto
\partial_{\mu}\frac{\beta^{\mu}}{\beta^4}=
\partial_t\frac{\beta^t}{\beta^4}+\partial_{z'}\frac{\beta^{z'}}{\beta^4}\\
=&\partial_t\left[\frac{T_0^4}{a^4}\left({z'}^2-t^2\right)^{-2}\frac{a}{T_0}z'\right]+\partial_{z'}\left[\frac{T_0^4}{a^4}\left({z'}^2-t^2\right)^{-2}\frac{a}{T_0}t\right]\\
=&\frac{T_0^3}{a^3}\left[4\left({z'}^2-t^2\right)^{-3}tz'-4\left({z'}^2-t^2\right)^{-3}z't\right]=0.
\end{split}
\end{equation}
This was expected because, as a general fact of global thermodynamic equilibrium, the entropy is constant in time, so the entropy production rate is identically null.
As we discussed earlier, the Unruh temperature is an absolute lower bound for the temperature at global thermodynamic equilibrium with acceleration, so the question arises if the entropy current vanishes or not at that temperature.
It turns out that it does not, in particular we have
\begin{equation}\label{gteq40}
s^{\mu}(T_{\rm U})=\frac{32\pi^2}{45}T_{\rm U}^3u^{\mu},\qquad (m=0).
\end{equation}
The reason for a non-vanishing entropy current in the Minkowski vacuum might be the fact that the field degrees of freedom in the left Rindler wedge were traced out, in the sense that $s^{\mu}$ is the entropy current pertaining the entropy in the right Rindler wedge, thus calculated with the reduced density operator $\widehat{\rho}_{\rm R}$.
This partial trace operation often leads to thermal states, thus a non-vanishing entropy current.
Another interesting observation is the following.
From equations \eqref{zubeq26} and \eqref{zubeq27}, it is clear that the thermodynamic potential current and the entropy current depend on the specific energy-momentum tensor quantum operator we decide to choose.
Our choice so far has been the canonical tensor \eqref{gteq19}, namely the minimal coupling one, but this is not the only possibility.
For instance, we could as well consider an ``improved'' tensor which is traceless for a massless field, namely the conformal coupling one,
\begin{equation}
\widehat{T}^{\mu \nu}_{\rm imp}\equiv \widehat{T}^{\mu \nu}-\frac{1}{6}\left(\partial^{\mu}\partial^{\nu}-g^{\mu \nu}\Box \right)\widehat{\psi}^2,
\end{equation}
with $\widehat{T}^{\mu \nu}$ appearing there being the canonical tensor.
In \cite{Becattini:2011ev} it is argued that, as a general feature of global thermodynamic equilibrium with acceleration or rotation, the energy density pertaining to this improved tensor is different from the one corresponding to the canonical tensor.
Indeed, by using the previously calculated expressions, the additional term with respect to the canonical value in the Rindler renormalization scheme for a massless field is
\begin{equation}
-\frac{1}{6}\left(u^{\mu}u^{\nu}\partial_{\mu}\partial_{\nu}-\Box \right)\left(\langle \widehat{\psi}^2\rangle-\langle 0_{\rm R}|\widehat{\psi}^2|0_{\rm R}\rangle \right)=
-\frac{1}{6}\left(u^{\mu}u^{\nu}\partial_{\mu}\partial_{\nu}-\Box \right)\frac{1}{12\beta^2}=
\frac{\alpha^2}{12\beta^4},
\end{equation}
hence the ``improved'' energy density
\begin{equation}
\rho^{\rm imp}_{\rm R}=\frac{\pi^2}{30\beta^4}=\frac{\pi^2}{30}T^4,\qquad (m=0).
\end{equation}
That is, the energy density calculated with the ``improved'' energy-momentum tensor for the massless free real scalar field depends only on $\beta^2$ and not on $\alpha^2$, which is a somewhat surprising feature.
Likewise, the entropy current gets modified and one is left with only the first term of equation \eqref{gteq39}
\begin{equation}
s^{\mu}_{\rm imp}=\frac{2\pi^2}{45\beta^4}\beta^{\mu}=
\frac{2\pi^2}{45}T^3u^{\mu},\qquad (m=0).
\end{equation}
At the Unruh temperature, we get an expression that, oddly enough, differs from the one obtained from the canonical tensor \eqref{gteq40} by a factor 16
\begin{equation}
s^{\mu}_{\rm imp}(T_{\rm U})=\frac{2\pi^2}{45}T_{\rm U}^3u^{\mu},\qquad (m=0).
\end{equation}
\subsection{Entropy in the right Rindler wedge and the entanglement entropy}
\label{subsec:gte_entropy}
Equation \eqref{gteq39} is the entropy current in the right Rindler wedge, therefore its integral on a 3-dimensional spacelike hypersurface $\Sigma_{\rm R}$ contained in the right Rindler wedge and whose boundary is the 2-dimensional surface $(t=0,z'=0)$ is the entropy in that subspace
\begin{equation}
S_{\rm R}=\int_{\Sigma_{\rm R}}{\rm d} \Sigma_{\mu}s^{\mu}.
\end{equation}
As mentioned already, the right Rindler wedge is foliated by constant-$\tau$ hypersurfaces.
Let $\Sigma_{\rm R}(0)$ and $\Sigma_{\rm R}(\tau^*)$ be two such hypersurfaces, in particular the $\tau=0$ and $\tau=\tau^*$ ones respectively, as shown in Figure \ref{fig:rindler_entropy}.
\begin{figure}
\begin{center}
\includegraphics[width=0.75\textwidth]{figures/rindler_entropy.eps}
\caption{In this 2-dimensional $(t,z)$-section of the right Rindler wedge, $\Sigma_{\rm R}(0)$ and $\Sigma_{\rm R}(\tau^*)$ are two constant-$\tau$ hypersurfaces of the foliation, corresponding to $\tau=0$ and $\tau=\tau^*$ respectively.
$\Gamma(0,\tau^*)$ is the timelike boundary at infinity joining $\Sigma_{\rm R}(0)$ and $\Sigma_{\rm R}(\tau^*)$.
The hyperbolae are the flow lines of the four-temperature, which is tangent to $\Gamma(0,\tau^*)$.}
\label{fig:rindler_entropy}
\end{center}
\end{figure}
In order for the integral to be independent on the choice of the hypersurface of the foliation, two conditions have to be met.
The first one is that the entropy current must be divergenceless, which we already checked.
And the second one is the vanishing of the flux through the timelike boundary at infinity joining the two hypersurfaces of the foliation, indicated as $\Gamma(0,\tau^*)$ in Figure \ref{fig:rindler_entropy}.
This condition is checked just as readily, for the entropy current is parallel to the four-temperature, which is tangent to $\Gamma(0,\tau^*)$, hence ${\rm d} \Sigma_{\mu}s^{\mu}={\rm d} \Sigma \,n_{\mu}s^{\mu}=0$.
In summary, we can choose whatever hypersurface of the foliation, as expected from global thermodynamic equilibrium.
A straightforward calculation of the entropy on $\Sigma_{\rm R}(0)$ with the entropy current \eqref{gteq39} pertaining to the canonical energy-momentum tensor yields
\begin{equation}\label{gteq47}
S_{\rm R}=\left(\int_{-\infty}^{+\infty}{\rm d}^2{\rm x_T}\right)\left(\frac{2\pi^2}{45}-\frac{\alpha^2}{6}\right)\frac{T_0^3}{a^3}\lim_{z'\to 0}\frac{1}{2{z'}^2}.
\end{equation}
This expression has two main properties.
The one is the proportionality to the area of the 2-dimensional boundary surface separating the left and right Rindler wedges, that is the so-called \textit{area law}, which is clearly reminiscent of the Bekenstein-Hawking formula for the black hole entropy \cite{Bekenstein:1973ur}.
The other is the divergence as $z'\to 0$, owing to the fact that the proper temperature diverges at the Killing horizon.
Our result is in full agreement with previous literature \cite{Bombelli:1986rw, Srednicki:1993im}.
This entropy we just worked out is the entropy in the right Rindler wedge, namely
the entropy calculated with the reduced density operator pertaining to the right Rindler wedge
\begin{equation}\label{gteq44}
S_{\rm R}=-\mathrm{tr}_{\rm R}(\widehat{\rho}_{\rm R}\log \widehat{\rho}_{\rm R}).
\end{equation}
Recall from \eqref{gteq41} and \eqref{gteq42} that the full density operator $\widehat{\rho}$ in Minkowski spacetime is factorized as $\widehat{\rho}=\widehat{\rho}_{\rm R}\otimes \widehat{\rho}_{\rm L}$, with $\widehat{\rho}_{\rm R}=\mathrm{tr}_{\rm L}(\widehat{\rho})$ and $\widehat{\rho}_{\rm L}=\mathrm{tr}_{\rm R}(\widehat{\rho})$ the reduced density operators of the right and left Rindler wedges respectively.
In Quantum Information Theory, the entropy obtained from the reduced density operator of a bipartite system is called the \textit{entanglement entropy}, which, in some sense, is a measure for the entanglement of a system.
Thus, the above $S_{\rm R}$ is in fact the entanglement entropy of the right Rindler wedge with the left Rindler wedge.
The area law is a property characteristic of the entanglement entropy, and the fact that $S_{\rm R}$ exhibits it too, somehow strengthens our belief in this connection.
The calculation of the entanglement entropy in Quantum Mechanics may not be necessarily hard, depending on the specific system at hand; whereas, as a matter of fact, when it comes to Quantum Field Theory it turns out to be a formidable task which one is not in general able to accomplish.
Only few exact analytical results are known, and most of them heavily rely on the power of conformal symmetry in 2 dimensions \cite{Calabrese:2004eu, Calabrese:2005zw, Calabrese:2009qy, Casini:2005zv, Casini:2005rm}.
The prescription for a holographic dual of the entanglement entropy in the context of the AdS/CFT correspondence has turned this complicated procedure into a conceptually far more intuitive one, which eventually amounts to finding minimal-area surfaces in some geometry, and as such it constitutes a milestone indeed \cite{Ryu:2006bv, Ryu:2006ef}.
However, it is still far from easy to obtain analytical results for non-trivial systems.
In this respect, having been able to perform an exact analytical calculation for a system which is both mathematically non-trivial and of physical concern, makes our result obtained in \cite{Becattini:2019poj} even more interesting.
Note that $S_{\rm R}$ does not vanish in the Minkowski vacuum, that is when the temperature is as low as the Unruh temperature.
This should not be completely surprising as $S_{\rm R}$ is obtained from $\widehat{\rho}_{\rm R}$, which in turn is obtained from $\widehat{\rho}$ by taking the partial trace over the field degrees of freedom of the left Rindler wedge.
This operation is in general likely to lead to thermal states, as it does indeed in our case too.
Note also that, if we used the ``improved'' energy-momentum tensor instead of the canonical one, the area law and the divergence at the Killing horizon would have stayed the same, so $S_{\rm R}$ would still be infinite overall; however, the constant factor in front of it would have changed to $2\pi^2/45$.
This is somewhat unexpected for the following reason.
As we mentioned, the expression of the entropy current does depend on the specific form of the energy-momentum tensor quantum operator.
On the other hand, the generators of the Poincaré group \eqref{gteq43} should not \cite{Becattini:2011ev}.
Consequently, neither should the entropy, because $S_{\rm R}$ is obtained from $\widehat{\rho}_{\rm R}$ as in \eqref{gteq44}, and $\widehat{\rho}_{\rm R}$, referring to \eqref{gteq04}, can in turn be written in terms of the Poincaré generators as
\begin{equation}
\widehat{\rho}_{\rm R}=
\mathrm{tr}_{\rm L}\left(\frac{1}{Z}\exp \left[-\frac{\widehat{H}}{T_0}+\frac{a}{T_0}\widehat{K}_z\right]\right).
\end{equation}
Nevertheless, $\widehat{\Pi}_{\rm R}$ and $\widehat{\Pi}_{\rm L}$ might inherit a dependence on the energy-momentum tensor quantum operator because of the truncation at $z'=0$.
This issue will be the subject of further investigation.
We conclude with a final remark referring to a comment we made at the end of Subsection \ref{sec:entr_prod_rate}, which essentially states the following.
Since the entropy current is divergenceless at global thermodynamic equilibrium, if its domain is topologically contractible it can be expressed as the divergence of an antisymmetric rank-2 tensor called the \textit{potential} of $s^{\mu}$
\begin{equation}
\partial_{\mu}s^{\mu}=0\qquad \Rightarrow \qquad s^{\mu}=\partial_{\nu}\varsigma^{\mu \nu},\qquad \varsigma^{\mu \nu}=-\varsigma^{\nu \mu}.
\end{equation}
With this potential, the entropy can be cast into a 2-dimensional integral as in \eqref{zubeq31}, i.e.\
\begin{equation}\label{gteq48}
S=\int_{\Sigma}{\rm d} \Sigma_{\mu}s^{\mu}=
\int_{\Sigma}{\rm d} \Sigma_{\mu}\partial_{\nu}\varsigma^{\mu \nu}=
\frac{1}{2}\int_{\partial \Sigma}{\rm d} \tilde{S}_{\mu \nu}\varsigma^{\mu \nu}=
-\frac{1}{4}\int_{\partial \Sigma}{\rm d} S^{\rho \sigma}\sqrt{|g|}\epsilon_{\mu \nu \rho \sigma}\varsigma^{\mu \nu},
\end{equation}
where $\partial \Sigma$ is the 2-dimensional boundary surface of $\Sigma$.
The right Rindler wedge is topologically contractible, therefore such a potential for the entropy current \eqref{gteq39} ought to exist.
In order to work it out, we follow the same kind of philosophy as for the derivation of the thermal expectation value of the energy-momentum tensor \eqref{gteq45}.
To form an antisymmetric tensor, we can only use the vector fields $\beta^{\mu}$ and $\alpha^{\mu}$, hence the only possible combination is $\alpha^{\mu}\beta^{\nu}-\alpha^{\nu}\beta^{\mu}$.
According to the decomposition \eqref{gteq18}, this is just proportional to the thermal vorticity, so we have in general
\begin{equation}\label{gteq46}
\varsigma^{\mu \nu}=G\varpi^{\mu \nu}
\end{equation}
for some thermodynamic scalar function $G=G(\alpha^2,\beta^2)$.
The divergence of this expression must reproduce the entropy current
\begin{equation}
s^{\mu}=\partial_{\nu}(G\varpi^{\mu \nu})=\varpi^{\mu \nu}\partial_{\nu}G,
\end{equation}
where we took into account that the thermal vorticity is constant at global thermodynamic equilibrium.
By introducing the entropy density $s=u_{\mu}s^{\mu}$, we get
\begin{equation}
s=u_{\mu}s^{\mu}=u_{\mu}\varpi^{\mu \nu}\partial_{\nu}G=
-\alpha^{\nu}\partial_{\nu}G=
-\alpha^{\nu}\frac{\partial G}{\partial \beta^2}\partial_{\nu}\beta^2,
\end{equation}
where we used again the decomposition \eqref{gteq18} of the thermal vorticity and, in the last equality, the fact that $\alpha^2$ is constant, so its derivative is identically null.
Using the Killing equation and \eqref{gteq18} once again, we find
\begin{equation}
\partial_{\nu}\beta^2=
g^{\rho \sigma}\partial_{\nu}(\beta_{\rho}\beta_{\sigma})=
g^{\rho \sigma}(\beta_{\rho}\partial_{\nu}\beta_{\sigma}+\beta_{\sigma}\partial_{\nu}\beta_{\rho})=
\beta^{\sigma}\varpi_{\sigma \nu}+\beta^{\rho}\varpi_{\rho \nu}=
-2\sqrt{\beta^2}\alpha_{\nu},
\end{equation}
hence
\begin{equation}
s=2\alpha^2\sqrt{\beta^2}\frac{\partial G}{\partial \beta^2}\qquad \Rightarrow \qquad
G=\int {\rm d} \beta^2\frac{s}{2\alpha^2\sqrt{\beta^2}}\frac{\partial G}{\partial \beta^2}.
\end{equation}
The entropy density obtained from the entropy current \eqref{gteq39} is
\begin{equation}
s=u_{\mu}s^{\mu}=\left(\frac{2\pi^2}{45}-\frac{\alpha^2}{6}\right)\frac{1}{(\beta^2)^{3/2}},
\end{equation}
so we can work out the expression of $G$
\begin{equation}
\begin{split}
G=&
\int {\rm d} \beta^2 \left(\frac{2\pi^2}{45}-\frac{\alpha^2}{6}\right)\frac{1}{(\beta^2)^{3/2}} \frac{1}{2\alpha^2\sqrt{\beta^2}}\\
=&\frac{1}{2\alpha^2}\left(\frac{2\pi^2}{45}-\frac{\alpha^2}{6}\right)\int {\rm d} \beta^2 \frac{1}{(\beta^2)^2}\\
=&-\frac{1}{2\alpha^2}\left(\frac{2\pi^2}{45}-\frac{\alpha^2}{6}\right)\frac{1}{\beta^2}=
-\frac{s}{2\alpha^2}\sqrt{\beta^2}.
\end{split}
\end{equation}
Plugging this into \eqref{gteq46}, we finally obtain the potential of the entropy current \cite{Becattini:2019poj}
\begin{equation}
\varsigma^{\mu \nu}=\frac{s}{2\alpha^2}(\beta^{\mu}\alpha^{\nu}-\beta^{\nu}\alpha^{\mu}).
\end{equation}
The boundary of the hypersurface $\Sigma_{\rm R}(0)$ is the $(x,y)$ plane $(t=0,z'=0)$ together with the plane $(t=0,z'=+\infty)$.
In the latter, the argument of the last integral in \eqref{gteq48} vanishes for $s \propto {z'}^{-3}$, $\alpha^2$ is constant
and $\beta^0 \alpha^3 \propto z'$.
We are thus left with the $(x,y)$ plane, and taking into account that the indices $\rho$ and $\sigma$ can only take on values $1$ and $2$ and of the dependence of $\beta^{\mu}$ and $\alpha^{\mu}$ on $(t,z')$, we end up with the same result as \eqref{gteq47}.
\section{Summary and outlook}
In this Chapter, we considered a relativistic quantum fluid at global thermodynamic equilibrium with acceleration, a non-trivial instance of global thermodynamic equilibrium in Minkowski spacetime of both phenomenological and theoretical concern.
For this particular system, the four-temperature has a Killing horizon making it timelike and future-oriented only in the right Rindler wedge, therefore that is the subspace we had to restrict in order to make thermodynamics and hydrodynamics out of it.
Moreover, the density operator is factorized into two reduced density operators each pertaining to a single Rindler wedge, so that thermal expectation values in each wedge can be calculated by taking the partial trace with the corresponding reduced density operator.
In order to obtain the entropy current, we calculated the thermal expectation value of the energy-momentum tensor in the right Rindler wedge.
In particular, we considered the canonical tensor for the simple instance of a free real scalar quantum field, and the calculation was performed analytically in the massless case.
The plain expectation value was expectedly divergent, therefore it needed renormalization.
We discussed both the Rindler and the Minkowski renormalization schemes, and saw how the Unruh effect arises naturally with the Unruh temperature being a lower bound for the temperature.
We then practically tested our method by working out the entropy current for this system including quantum corrections.
The corresponding entropy production rate was identically null, in agreement with the general condition of global thermodynamic equilibrium.
The entropy current did not vanish in the Minkowski vacuum, that is at a temperature as low as the Unruh temperature, which is supposedly due to having traced out the field degrees of freedom in the left Rindler wedge.
As clear from the general method, the expression of the entropy current depends on the form of the energy-momentum tensor quantum operator.
All our calculations were performed with the canonical tensor, i.e.\ the minimal coupling one, but then we compared the result with the one we would have obtained by using an ``improved'' tensor which is traceless for a massless field, i.e.\ the conformal coupling one.
Interestingly enough, the quantum correction due to the acceleration field disappears in the latter case.
Finally, we integrated the above entropy current and obtained the entropy in the right Rindler wedge.
The result has two apparent properties, both in full agreement with known literature: the one is the so-called area law, namely the proportionality to the area of the 2-dimensional surface separating the two Rindler wedges, clearly reminiscent of the Bekenstein-Hawking formula for the black hole entropy; whereas the other is the divergence due to the proper temperature being infinite at the Killing horizon.
This entropy can also be seen as the one obtained by using the von Neumann formula with the reduced density operator of the right Rindler wedge.
By virtue of the factorization of the overall density operator, this is also, by definition, the entanglement entropy of the right Rindler wedge with the left Rindler wedge.
The calculation of the entanglement entropy in Quantum Field Theory is a very hard task and a hot topic in Quantum Information Theory in general.
Having been able to perform it in an exact analytical way by using a method completely orthogonal to the standard ones in the case of a system both mathematically non-trivial and of physical concern adds all the more interest to our result.
We also worked out the potential of the entropy current and recast the entropy in the form of a 2-dimensional surface integral.
Oddly, while maintaining the area law and the divergence at the Killing horizon, the expression of the entropy seemed to change by a constant factor if we considered the ``improved'' energy-momentum tensor instead of the canonical one.
This is a somewhat unexpected feature that deserves further study.
\section{Thermal field theory with boost invariance}
\label{sec:boost_boost_invariance}
In order to be able to make any thermal field theory, we must first determine the density operator representing the state at hand, which will then enable us to calculate thermal expectation values.
In this Section, we will present the density operator of a relativistic quantum fluid with boost invariance, a non-trivial instance of a non-equilibrium state in the future light-cone particularly interesting in the context of heavy-ion collisions.
\subsection{Boost invariance}
In Chapter \ref{chapter:gteacceleration}, we considered a relativistic quantum fluid at global thermodynamic equilibrium in the right Rindler wedge.
This time, we focus on another patch of Minkowski spacetime, that is the future light-cone defined by $t>|z|$.
Much in the same way as we introduced the Rindler coordinates to span the right Rindler wedge, we now introduce a new set of hyperbolic coordinates in order to span the future light-cone.
These are the so-called \textit{Milne coordinates} $(\tau,{\bf x}_{\rm T},\eta)$, where the transverse coordinates ${\bf x}_{\rm T}=(x,y)$ are once again the same as Minkowski's, while $(\tau,\eta)$ are related to $(t,z)$ by
\begin{equation}
\tau \equiv \sqrt{t^2-z^2},\qquad
\eta \equiv \frac{1}{2}\log \left(\frac{t+z}{t-z}\right)
\end{equation}
whose inverse read
\begin{equation}
t\equiv \tau \cosh \eta,\qquad
z\equiv \tau \sinh \eta.
\end{equation}
The hyperbolic time $\tau$ can be referred to as the \textit{Milne time}, while the coordinate $\eta$ is called the \textit{spacetime rapidity}.
The Milne coordinates span indeed the future light-cone, as shown in Figure \ref{fig:milne_coord},
\begin{figure}
\begin{center}
\includegraphics[width=0.8\textwidth]{figures/milne_coord.eps}
\caption{2-dimensional section of the future light-cone in the $(t,z)$ plane, spanned by the Milne coordinates $(\tau,\eta)$.
The hyperbolae are $\tau={\rm const.}$ hypersurfaces, while the straight lines through the origin are $\eta={\rm const.}$ hypersurfaces.}
\label{fig:milne_coord}
\end{center}
\end{figure}
and the reason for their name is that the line element is but the metrics of the so-called \textit{Milne universe}
\begin{equation}
{\rm d} s^2=
{\rm d} t^2-{\rm d} x^2-{\rm d} y^2-{\rm d} z^2=
{\rm d} \tau^2-{\rm d} x^2-{\rm d} y^2-\tau^2 {\rm d} \eta^2.
\end{equation}
The foliation of the future light-cone with the family of 3-dimensional spacelike hypersurfaces $\{\Sigma(\tau)\}$ given by the constant-$\tau$ hyperboloids is well-defined, for the vorticity-free condition \eqref{zubeq07} is met by the timelike vector field of unit magnitude orthogonal to them,
\begin{equation}
n^{\mu}=\frac{1}{\sqrt{t^2-z^2}}(t,0,0,z)=
(\cosh \eta,0,0,\sinh \eta).
\end{equation}
This vector field can be used to choose the hydrodynamic velocity field consistently with the foliation by taking $u^{\mu}=n^{\mu}$, which is essentially the $\beta$-frame, thus
\begin{equation}\label{boosteq07}
u^{\mu}=(\cosh \eta,0,0,\sinh \eta).
\end{equation}
Then, the tetrad onto which tensors are decomposed is most naturally given by
\begin{subequations}
\begin{align}
(\partial_{\tau})^{\mu}=&(\cosh \eta,0,0,\sinh \eta)=u^{\mu},\label{boosteq06a}\\
(\partial_x)^{\mu}=&(0,1,0,0)\equiv \hat{i}^{\mu},\label{boosteq06b}\\
(\partial_y)^{\mu}=&(0,0,1,0)\equiv \hat{j}^{\mu},\label{boosteq06c}\\
(\partial_{\eta})^{\mu}=&\tau(\sinh \eta,0,0,\cosh \eta)\equiv \tau \hat{\eta}^{\mu},\label{boosteq06d}
\end{align}
\end{subequations}
where the hat $\hat{}$ stands for unit magnitude, namely $\hat{i}^2=\hat{j}^2=\hat{\eta}^2=-1$.
For reasons that are going to become clear shortly, let us assume that the proper temperature and the fugacity are scalar functions depending on the Milne time $\tau$ only, that is $T=T(\tau)$ and $\zeta=\zeta(\tau)$, so the four-temperature reads
\begin{equation}\label{boosteq03}
\beta^{\mu}=\frac{u^{\mu}}{T(\tau)}=\frac{1}{T(\tau)}(\cosh \eta,0,0,\sinh \eta).
\end{equation}
Suppose also that the system reaches local thermodynamic equilibrium at some $\tau_0$, so, according to the discussion in Subsection \ref{sec:LE_to_GE_and_NE}, the stationary state in the Heisenberg picture is the local thermodynamic equilibrium density operator at $\tau_0$
\begin{equation}\label{boosteq01}
\widehat{\rho}=\widehat{\rho}_{\rm LE}(\tau_0)=
\frac{1}{Z_{\rm LE}(\tau_0)}\exp \left[-\int_{\Sigma(\tau_0)}{\rm d} \Sigma_{\mu}\left(\widehat{T}^{\mu \nu}\beta_{\nu}-\zeta \widehat{j}^{\mu}\right)\right],
\end{equation}
where the argument of the integral is intended to be evaluated at $\tau_0$.
The four-temperature is timelike on $\Sigma(\tau_0)$, hence thermodynamically meaningful, but it is not a Killing vector field, as can be checked.
The Milne universe, in fact, is not a stationary spacetime, and as such it does not possess any global timelike Killing vector field.
From the cosmological point of view, it represents instead a linearly expanding universe foliated by hyperboloids with constant negative curvature.
As a consequence, global thermodynamic equilibrium states cannot exist in the Milne universe, therefore the density operator \eqref{boosteq01} represents a non-equilibrium state.
In Subsection \ref{sec:LE_to_GE_and_NE} it was pointed out that if the spacelike hypersurface $\Sigma(\tau_0)$ is invariant under any transformation $g$ of a subgroup ${\rm G}$ of the proper orthochronous Poincar\'e group ${\rm IO}(1,3)^{\uparrow}_+$ and the thermodynamic fields change according to the transformation laws \eqref{zubeq12}, which we hereby report
\begin{equation}\label{boosteq02}
{D(g^{-1})^{\nu}}_{\sigma}\beta_{\nu}(g^{-1}(y))=\beta_{\sigma}(y),\qquad
\zeta(g^{-1}(y))=\zeta(y),
\end{equation}
then ${\rm G}$ is a symmetry group for the density operator, in the sense that $\widehat{U}(g)\widehat{\rho}\widehat{U}^{-1}(g)=\widehat{\rho}$ with $\widehat{U}(g)$ a unitary representation of $g\in {\rm G}$.
The thermodynamic fields appearing in the density operator \eqref{boosteq01}, namely the four-temperature \eqref{boosteq03} at $\tau_0$ and the fugacity $\zeta(\tau_0)$, meet the conditions \eqref{boosteq02} for the following transformations:
\begin{itemize}
\item boosts with hyperbolic angle $\xi$ along the $z$ direction, indicated as ${\sf L}_3(\xi)$,
\item translations along the $x$ and $y$ directions, indicated as ${\sf T}_1(a)$ and ${\sf T}_2(a)$ respectively,
\item rotations of angle $\varphi$ in the $(x,y)$ plane, indicated as ${\sf R}_{12}(\varphi)$,
\end{itemize}
and the hyperboloid $\Sigma(\tau_0)$ is invariant under the same transformations.
This is but the Euclidean group in 2-dimensions times the Lorentz group in 1-dimension, namely ${\rm IO}(2)\otimes {\rm SO}(1,1)$.
Moreover, the density operator is also invariant under:
\begin{itemize}
\item space reflections mapping $(x,y,z)$ to $(-x,-y,-z)$.
\end{itemize}
Altogether, we refer to this symmetry of the density operator as \textit{boost invariance}.
Accordingly, the description of the system represented by $\widehat{\rho}$ must have the same form in all frames boosted in the longitudinal direction.
This is essentially what in literature is often known as the \textit{Bjorken symmetry}, named after J.\ D.\ Bjorken who first proposed it as a symmetry approximately realized in the central-rapidity region in heavy-ion collisions \cite{Bjorken:1982qr}.
As it should be clear by now, the symmetries of the system constrain the form of quantities at the particular thermodynamic configuration at hand.
These constraints are given by the vanishing of Lie derivatives of such quantities along the vector fields associated to the symmetry group, which in the present case can be readily found
\begin{subequations}
\begin{align}
\left(\frac{{\rm d} {\sf L}_3(\xi)x}{{\rm d} \xi}\right)^{\mu}=&(z,0,0,t)=\tau \hat{\eta}^{\mu},\label{boosteq04a}\\
\left(\frac{{\rm d} {\sf T}_1(a)x}{{\rm d} a}\right)^{\mu}=&(0,1,0,0)=\hat{i}^{\mu},\label{boosteq04b}\\
\left(\frac{{\rm d} {\sf T}_2(a)y}{{\rm d} a}\right)^{\mu}=&(0,0,1,0)=\hat{j}^{\mu},\label{boosteq04c}\\
\left(\frac{{\rm d} {\sf R}_{12}(\varphi)x}{{\rm d} \varphi}\right)^{\mu}=&(0,-y,x,0)=r \hat{\varphi}^{\mu}\label{boosteq04d},
\end{align}
\end{subequations}
where $r\equiv \sqrt{x^2+y^2}$ and $\hat{\varphi}^{\mu}\equiv \frac{1}{r}(0,-y,x,0)$.
Note that \eqref{boosteq04a}--\eqref{boosteq04c} are just \eqref{boosteq06b}--\eqref{boosteq06d} of the above tetrad, which, by construction, have vanishing Lie derivatives along each other.
Scalars, such as the proper temperature $T$ and the fugacity $\zeta$, can only depend on $\tau$ as a consequence of ${\cal L}_XT=0$ and ${\cal L}_X\zeta=0$ with $X$ either of \eqref{boosteq04a}--\eqref{boosteq04d}
\begin{equation}\label{boosteq05}
T=T(\tau),\qquad \zeta=\zeta(\tau).
\end{equation}
Let us prove it for $T$ as an example.
With little effort, one can show that
\begin{equation}
{\cal L}_{\hat{i}}T=\hat{i}^{\mu}\partial_{\mu}T=\partial_xT=0,
\end{equation}
\begin{equation}
{\cal L}_{\hat{j}}T=\hat{j}^{\mu}\partial_{\mu}T=\partial_yT=0,
\end{equation}
\begin{equation}
{\cal L}_{\tau \hat{\eta}}T=\tau \hat{\eta}^{\mu}\partial_{\mu}T=\partial_{\eta}T=0,
\end{equation}
\begin{equation}
{\cal L}_{r\hat{\varphi}}T=r\hat{\varphi}^{\mu}\partial_{\mu}T=-y\partial_yT+x\partial_xT=0.
\end{equation}
The first two equations imply that $T$ cannot depend neither on $x$ nor on $y$, making the last equation trivial, while the third equation forbids the dependence on $\eta$, so the only possible dependence left is in fact on $\tau$.
The same holds for $\zeta$, hence \eqref{boosteq05}, and that is the reason why we previously assumed the temperature and the fugacity to be scalar functions of $\tau$ only.
This philosophy can be further applied to vectors and tensors of any rank.
A vector field $V^{\mu}$ is decomposed, in principle, as
\begin{equation}
V^{\mu}=A(\tau)u^{\mu}+B(\tau)\hat{i}^{\mu}+C(\tau)\hat{j}^{\mu}+D(\tau)\tau \hat{\eta}^{\mu},
\end{equation}
however, the constrain ${\cal L}_{r\hat{\varphi}}V=0$ implies $B=C=0$, and the invariance under space reflections demands $D=0$ as it maps $\eta \mapsto -\eta$, hence
\begin{equation}
V^{\mu}=A(\tau)u^{\mu}.
\end{equation}
Concerning vector fields, a well-known feature of the Bjorken model is the existence of a unique timelike vector field of unit magnitude compatible with boost invariance, and this is given precisely by \eqref{boosteq07} \cite{Bjorken:1982qr}.
That is to say that the hydrodynamic velocity field is in fact completely fixed by the symmetry, which is a very powerful statement for the following reason.
At global thermodynamic equilibrium, the thermodynamic fields are determined by the geometric conditions \eqref{zubeq10}, namely the four-temperature must be a Killing vector field and the fugacity a constant.
Out of global thermodynamic equilibrium, these conditions no longer hold and the thermodynamic fields can be whatever, so we have no information on them whatsoever.
This is one of the reasons that make systems out of global thermodynamic equilibrium much harder to study.
However, boost invariance completely fixes the hydrodynamic velocity field and constrains the proper temperature and the fugacity to depend on the Milne time only, therefore, the form of the four temperature cannot be but \eqref{boosteq03}.
The expressions of the temperature and the fugacity as functions of $\tau$ are obtained from the constraint equations of local thermodynamic equilibrium \eqref{zubeq04}, that is
\begin{equation}
n_{\mu}\mathrm{tr}(\widehat{\rho}_{\rm LE}(\tau)\widehat{T}^{\mu \nu})_{\rm ren}=n_{\mu}T^{\mu \nu},\qquad
n_{\mu}\mathrm{tr}(\widehat{\rho}_{\rm LE}(\tau)\widehat{j}^{\mu})_{\rm ren}=n_{\mu}j^{\mu}.
\end{equation}
Thus, although out of thermodynamic equilibrium, boost invariance actually provides us with information on the thermodynamic fields, enough so to at least attempt calculations.
This constitutes a precious opportunity for us to scratch the surface of the vast and mostly unexplored topic that is non-equilibrium physics.
In particular, we will be interested in the energy-momentum tensor as usual.
Much in the same way as above, the most general form a symmetric rank-2 tensor, such as the thermal expectation value of the energy-momentum tensor, is constrained to be
\begin{equation}\label{boosteq12}
T^{\mu \nu}=\rho(\tau)u^{\mu}u^{\nu}+p^{\rm T}(\tau)(\hat{i}^{\mu}\hat{i}^{\nu}+\hat{j}^{\mu}\hat{j}^{\nu})+p^{\rm L}(\tau)\hat{\eta}^{\mu}\hat{\eta}^{\nu},
\end{equation}
where $\rho$ is the energy density, $p^{\rm T}$ the transverse pressure and $p^{\rm L}$ the longitudinal pressure.
For the sake of clarity, the left-hand side stands explicitly for
\begin{equation}\label{boosteq08}
T^{\mu \nu}=\mathrm{tr}(\widehat{\rho}\widehat{T}^{\mu \nu})=
\mathrm{tr}(\widehat{\rho}_{\rm LE}(\tau_0)\widehat{T}^{\mu \nu}(\tau)),
\end{equation}
namely the non-equilibrium density operator is at fixed $\tau_0$, whereas the energy-momentum tensor quantum operator varies with $\tau$, hence the $\tau$-dependence of $\rho$, $p^{\rm T}$ and $p^{\rm L}$.
The pressures in the $x$ and $y$ directions are equal and given by the transverse pressure due to rotational symmetry in the $(x,y)$ plane.
However, as full rotational symmetry in the $(x,y,z)$ space is lacking, the longitudinal pressure will be in general different.
This is reminiscent of the case of global thermodynamic equilibrium with acceleration studied in Chapter \ref{chapter:gteacceleration}, where an anisotropic quantum term was present in the longitudinal direction.
Thus, the anisotropy in the current case can also expected to be a quantum effect.
Provided an underlying Quantum Field Theory is chosen, the three scalar functions $\rho$, $p^{\rm T}$ and $p^{\rm L}$ can be determined once we are able to calculate thermal expectation values with the non-equilibrium density operator as in \eqref{boosteq08}.
This is a task one is not able to accomplish in general.
A possible approach to tackle this problem is that of ideal hydrodynamics, where local thermodynamic equilibrium is assumed.
In that case, the dissipations are supposed to be small, so that the non-equilibrium state is in fact close to thermodynamic equilibrium, and by equation \eqref{zubeq09}, the non-equilibrium density operator is approximated by the local thermodynamic equilibrium one.
However, this should not be misleading: unless one uses linear response theory to further approximate the local thermodynamic equilibrium density operator with the global thermodynamic equilibrium one \cite{Becattini:2014yxa}, the calculation of thermal expectation values at local thermodynamic equilibrium is usually far from feasible.
And even if it were, the actual non-equilibrium expressions of $\rho$, $p^{\rm T}$ and $p^{\rm L}$, that is the constitutive equations, would still be unknown.
Nevertheless, for our system at hand, boost invariance is so powerful that a non-equilibrium analysis turns out to be possible to some extent, at least for a free Quantum Field Theory.
For the sake of simplicity, in the following we will suppose that there is no charged current present, that is $\zeta \equiv 0$.
Thus, by using the same symbols as in \eqref{gteq10}, the non-equilibrium density operator is
\begin{equation}\label{boosteq16}
\widehat{\rho}=\frac{1}{Z}\exp \left[-\frac{\widehat{\Pi}(\tau_0)}{T(\tau_0)}\right],
\end{equation}
where $Z=Z_{\rm LE}(\tau_0)$, $n_{\mu}=u_{\mu}$ and the operator $\widehat{\Pi}(\tau_0)$ is given by
\begin{equation}
\widehat{\Pi}(\tau_0)=
\int_{\Sigma(\tau_0)}{\rm d} \Sigma \,\widehat{T}^{\mu \nu}(\tau_0)u_{\mu}u_{\nu}=
\tau_0\int_{-\infty}^{+\infty}{\rm d}^2{\rm x_T}\,{\rm d} \eta \,\widehat{T}^{\mu \nu}(\tau_0)u_{\mu}u_{\nu}
\end{equation}
with ${\rm d} \Sigma=\tau_0\,{\rm d}^2{\rm x_T}\,{\rm d} \eta$ the measure on $\Sigma(\tau_0)$ in Milne coordinates.
All the integration variables are intended to be integrated from $-\infty$ to $+\infty$.
The expression of the energy-momentum tensor quantum operator depends on the specific underlying Quantum Field Theory.
For a free Quantum Field Theory, this density operator represents a relativistic quantum fluid with hydrodynamic velocity field given by \eqref{boosteq07} which is initially interacting, ceases interactions at $\tau_0$ and propagates freely thereafter.
Thus, in the classical limit, the kinetic theory solutions of the free-streaming Boltzmann equation starting from the local thermodynamic
equilibrium expressions with proper temperature $T(\tau_0)$ and flow velocity $u^{\mu}$ at $\tau_0$, reported in Appendix \ref{appendix:free_sreaming}, are expected to be recovered.
We want to conclude with the following remark.
In Subsection \ref{sec:LE_to_GE_and_NE}, the question arose if a symmetry group for the stationary state $\widehat{\rho}=\widehat{\rho}_{\rm LE}(\tau_0)$ is also a symmetry group for the local thermodynamic equilibrium density operator $\widehat{\rho}_{\rm LE}(\tau)$ at any $\tau$.
Intuitively, if the hypersurface $\Sigma(\tau)$ is mapped into itself and the thermodynamic fields change according to \eqref{boosteq02} at any $\tau$, then the answer is yes.
The constant-$\tau$ hypersurfaces, the four-temperature \eqref{boosteq03} and the fugacity $\zeta(\tau)$ meet indeed those conditions at any $\tau$, therefore boost invariance is also a symmetry of $\widehat{\rho}_{\rm LE}(\tau)$.
The expression of the local thermodynamic equilibrium density operator is formally the same as that of the non-equilibrium one at $\tau$ instead of $\tau_0$
\begin{equation}\label{boosteq15}
\widehat{\rho}_{\rm LE}(\tau)=\frac{1}{Z_{\rm LE}(\tau)}\exp \left[-\frac{\widehat{\Pi}(\tau)}{T(\tau)}\right],
\end{equation}
where $\widehat{\Pi}(\tau)$ is in general different from $\widehat{\Pi}(\tau_0)$ and given by
\begin{equation}
\widehat{\Pi}(\tau)=
\int_{\Sigma(\tau)}{\rm d} \Sigma \,\widehat{T}^{\mu \nu}(\tau)u_{\mu}u_{\nu}=
\tau \int_{-\infty}^{+\infty}{\rm d}^2{\rm x_T}\,{\rm d} \eta \,\widehat{T}^{\mu \nu}(\tau)u_{\mu}u_{\nu}.
\end{equation}
In the case of a free Quantum Field Theory, the physical system described by this density operator is a relativistic quantum fluid with hydrodynamic velocity field given by \eqref{boosteq07} who never ceases interactions and does not decouple.
Of course, the forms of scalar, vector and tensor quantities are still those discussed above, as they descend purely from symmetry considerations.
In particular, the expression of the thermal expectation value of the energy-momentum tensor is
\begin{equation}\label{boosteq13}
\langle \widehat{T}^{\mu \nu}\rangle_{\rm LE}=
\mathrm{tr}(\widehat{\rho}_{\rm LE}(\tau)\widehat{T}^{\mu \nu}(\tau))=
\rho_{\rm LE}(\tau)u^{\mu}u^{\nu}+p^{\rm T}_{\rm LE}(\tau)(\hat{i}^{\mu}\hat{i}^{\nu}+\hat{j}^{\mu}\hat{j}^{\nu})+p^{\rm L}_{\rm LE}(\tau)\hat{\eta}^{\mu}\hat{\eta}^{\nu}.
\end{equation}
Since the non-equilibrium density operator $\widehat{\rho}$ is simply the local thermodynamic equilibrium one $\widehat{\rho}_{\rm LE}(\tau)$ evaluated at $\tau_0$, the operator $\widehat{\Pi}(\tau_0)$ will just be $\widehat{\Pi}(\tau)$ evaluated at $\tau_0$.
However, comparing equations \eqref{boosteq08} and \eqref{boosteq13}, it is clear that the non-equilibrium thermal expectation value \eqref{boosteq12} is not the same as the local thermodynamic equilibrium one \eqref{boosteq13} evaluated at $\tau_0$.
Once again, the energy-momentum tensor quantum operator, and therefore the scalar functions in the above thermal expectation value, depend on the specific underlying Quantum Field Theory.
In analogy with Chapter \ref{chapter:gteacceleration}, in the following we will consider a free real scalar quantum field in the future light-cone.
\subsection{Free scalar field theory in the future light-cone}
The problem of a free real scalar quantum field in the future light-cone is a known one in literature, especially in Quantum Field Theory in curved spacetime and cosmology in general \cite{Birrell:1982ix, Padmanabhan:1990fm, Arcuri:1994kd}, but also in the context of multiparticle production \cite{Berges:2017zws, Berges:2017hne}.
In \cite{Arcuri:1994kd} it was shown that the free real scalar field can be expanded in the future light-cone in a complete set of solutions that do not mix the positive- and negative-frequency modes of the usual plane waves expansion.
Then, S.\ Akkelin showed in \cite{Akkelin:2018hpu} that these modes can be obtained starting from the usual plane waves modes.
Here, we follow his derivation by using a slightly different notation and adding some clues.
The Klein Gordon equation for a free real scalar quantum field $\widehat{\psi}$ of mass $m$ in Milne coordinates reads
\begin{equation}
(\Box+m^2)\widehat{\psi}=
\left[\frac{1}{\tau}(\tau \partial_{\tau})-\partial_x^2-\partial_y^2-\frac{1}{\tau^2}\partial_{\eta}^2+m^2\right]\widehat{\psi}=0.
\end{equation}
The solution is given by the following quantum field expansion
\begin{equation}\label{boosteq09}
\widehat{\psi}(\tau,{\bf x}_{\rm T},\eta)=
\int \frac{{\rm d}^2{\rm p_T}\,{\rm d} \mu}{4\pi \sqrt{2}}
\left[h({\sf p},\tau){\rm e}^{i({\bf p}_{\rm T} \cdot {\bf x}_{\rm T}+\mu \eta)}\widehat{b}_{\sf p}+h^*({\sf p},\tau){\rm e}^{-i({\bf p}_{\rm T} \cdot {\bf x}_{\rm T}+\mu \eta)}\widehat{b}_{\sf p}^{\dagger}\right].
\end{equation}
Here, we indicated the momentum in Milne coordinates as ${\sf p}\equiv({\bf p}_{\rm T},\mu)$ to distinguish it from the Cartesian vector ${\bf p}=({\bf p}_{\rm T},p_z)$, where ${\bf p}_{\rm T}=(p_x,p_y)$ is the transverse momentum and $\mu$ is the momentum component along the $\eta$-direction.
The integral is intended to run from $-\infty$ to $+\infty$ for each integration variable.
The $\tau$-dependence is stored in the complex functions $h$ and $h^*$ solutions of the equation
\begin{equation}
\left[\frac{1}{\tau}\partial_{\tau}(\tau \partial_{\tau})+m_{\rm T}^2+\frac{\mu^2}{\tau^2}\right]h({\sf p},\tau)=0,
\end{equation}
which is indeed a Bessel equation with the transverse mass given by
\begin{equation}
m_{\rm T}=\sqrt{{\bf p}_{\rm T}^2+m^2}.
\end{equation}
In fact, $h$ and $h^*$ are defined by means of the Hankel functions ${\rm H}^{(2)}$ and ${\rm H}^{(1)}$ as
\begin{equation}\label{boosteq72}
\begin{split}
h({\sf p},\tau)\equiv&
-i{\rm e}^{\frac{\pi}{2}\mu}{\rm H}^{(2)}_{i\mu}(m_{\rm T} \tau)\\
h^*({\sf p},\tau)\equiv&
i{\rm e}^{-\frac{\pi}{2}\mu}{\rm H}^{(1)}_{i\mu}(m_{\rm T} \tau),
\end{split}
\end{equation}
which can be checked to be the complex conjugate of each other and also invariant under momentum reflections ${\sf p}\mapsto -{\sf p}$ by using the integral representation \cite{Gradshteyn:1702455}
\begin{equation}\label{boosteq10}
\begin{split}
{\rm H}^{(2)}_{i\mu}(m_{\rm T} \tau)=&
-\frac{1}{i\pi}{\rm e}^{-\frac{\pi}{2}\mu}\int_{-\infty}^{+\infty}{\rm d} \theta \,{\rm e}^{-im_{\rm T} \tau \cosh \theta+i\mu \theta}\\
{\rm H}^{(1)}_{i\mu}(m_{\rm T} \tau)=&
\frac{1}{i\pi}{\rm e}^{\frac{\pi}{2}\mu}\int_{-\infty}^{+\infty}{\rm d} \theta \,{\rm e}^{im_{\rm T} \tau \cosh \theta-i\mu \theta}.
\end{split}
\end{equation}
The following quantity is also usually defined in Milne coordinates
\begin{equation}\label{boosteq30}
\omega^2({\sf p},\tau)\equiv m_{\rm T}^2+\frac{\mu^2}{\tau^2}={\sf p}^2+m^2,
\end{equation}
which is invariant under momentum reflections ${\sf p}\mapsto -{\sf p}$.
Finally, $\widehat{b}_{\sf p}^{\dagger}$ and $\widehat{b}_{\sf p}$ are creation and annihilation operators respectively in the future light-cone obeying the usual algebra
\begin{equation}\label{boosteq18}
\begin{split}
[\widehat{b}_{\sf p},\widehat{b}_{{\sf p}'}^{\dagger}]=&\delta^2({\bf p}_{\rm T}-{\bf p}_{\rm T}')\,\delta(\mu-\mu'),\\
[\widehat{b}_{\sf p},\widehat{b}_{{\sf p}'}]=&0=[\widehat{b}_{\sf p}^{\dagger},\widehat{b}_{{\sf p}'}^{\dagger}].
\end{split}
\end{equation}
This allows us to clarify the interpretation of $\mu$, which turns out to be the eigenvalue of the boost operator along the longitudinal direction $\widehat{K}_z$, that is
\begin{equation}
\widehat{U}({\sf L}_3(\xi))\widehat{b}_{\sf p}^{\dagger}\widehat{U}^{-1}({\sf L}_3(\xi))={\rm e}^{-i\xi \widehat{K}_z}\widehat{b}_{\sf p}^{\dagger}{\rm e}^{i\xi \widehat{K}_z}={\rm e}^{-i\xi \mu}\widehat{b}_{\sf p}^{\dagger}
\end{equation}
with $\widehat{U}$ a unitary representation.
In other words, the operator $\widehat{b}_{\sf p}^{\dagger}$ creates a state with eigenvalue $\mu$, much in the same way as for the angular momentum operator in the familiar Quantum Mechanics.
In \cite{Akkelin:2018hpu} it was shown explicitly that the quantum field expansion \eqref{boosteq09} in the future light-cone can also be worked out starting from the standard plane waves expansion in Minkowski spacetime \eqref{gteq27}.
The frequency of the plane wave \eqref{gteq49} is
\begin{equation}
\omega_{\bf p}^2=m_{\rm T}^2+p_z^2.
\end{equation}
One can then express $\omega_{\bf p}$ and $p_z$ in terms of a hyperbolic angle ${\rm y}$ called the \textit{momentum rapidity}, or simply the \textit{rapidity}, as
\begin{equation}
\omega_{\bf p}=m_{\rm T} \cosh {\rm y},\qquad
p_z=m_{\rm T} \sinh {\rm y}.
\end{equation}
Thus, by performing the following ``Fourier-like'' transformation of the plane wave creation and annihilation operators $\widehat{a}_{\bf p}^{\dagger}$ and $\widehat{a}_{\bf p}$ in the plane waves expansion
\begin{equation}\label{boosteq11}
\widehat{a}_{\bf p}^{\dagger}=
\frac{1}{\sqrt{2\pi m_{\rm T} \cosh {\rm y}}}\int_{-\infty}^{+\infty}{\rm d} \mu \,{\rm e}^{-i\mu {\rm y}}\widehat{b}_{\sf p}^{\dagger},
\end{equation}
changing integration variable from $p_z$ to ${\rm y}$ and then to $\theta={\rm y}-\eta$ and exploiting the integral representation \eqref{boosteq10}, the field expansion \eqref{boosteq09} is finally obtained \cite{Akkelin:2018hpu}.
Equation \eqref{boosteq11} shows that the creation operators in the future light-cone depend only on the creation operators in Minkowski spacetime, and same for the corresponding annihilation operators.
No mixing between creation and annihilation operators is present.
Consequently, the vacuum state of the operators $\widehat{b}_{\sf p}$ in the future light-cone coincides with that of $\widehat{a}_{\bf p}$ in Minkowski spacetime, that is the Minkowski vacuum $|0_{\rm M}\rangle$.
The ultimate reason for this is that the function $h$ can be expressed can be expressed as a linear combination of plane wave modes with only positive frequencies \cite{Mukhanov:2007zz}.
This is clearly at variance with the case of global thermodynamic equilibrium with acceleration considered in Chapter \ref{chapter:gteacceleration}, where the quantum field was expanded in the Rindler wedges in terms of the Rindler creation and annihilation operators, whose relation with the plane waves creation and annihilation operators was a non-trivial Bogolyubov transformation, therefore the Rindler vacuum and the Minkowski vacuum did not coincide.
The consequences of this fact will be further explored in Subsection \ref{sec:boost_lte_renormalization}.
Now we turn to the density operator.
It is important to stress that the expression of the non-equilibrium density operator \eqref{boosteq16} is exactly the same as the local thermodynamic equilibrium one \eqref{boosteq15} evaluated at $\tau_0$.
Whether one or the other is used makes indeed a difference when it comes to calculating thermal expectation values, as exemplified by equations \eqref{boosteq08} and \eqref{boosteq13}.
However, as far as the very structure of the density operator is concerned, they really are equivalent.
Therefore, without loss of generality, we can let $\tau$ be whatever and focus on the local thermodynamic equilibrium one for the rest of this part.
Then, in Section \ref{sec:boost_lte_analysis} we will perform a local thermodynamic equilibrium analysis, whereas in Section \ref{sec:boost_ne_analysis} a non-equilibrium one, thus distinguishing explicitly the two cases.
Moreover, the ${\sf p}$-dependence will be omitted throughout for ease of notation, restoring it only when necessary.
As for the energy-momentum tensor operator, we consider again the canonical tensor
\begin{equation}\label{boosteq50}
\widehat{T}^{\mu \nu}=\frac{1}{2}\left(\partial^{\mu}\widehat{\psi}\partial^{\nu}\widehat{\psi}+\partial^{\nu}\widehat{\psi}\partial^{\mu}\widehat{\psi}\right)-g^{\mu \nu}\widehat{\cal L},\qquad
\widehat{\cal L}=\frac{1}{2}\left(g^{\mu \nu}\partial_{\mu}\widehat{\psi}\partial_{\nu}\widehat{\psi}-m^2\widehat{\psi}^2\right).
\end{equation}
In order to work $\widehat{\Pi}(\tau)$ out, the above expression must be contracted twice with the hydrodynamic velocity field, yielding
\begin{equation}\label{boosteq34}
\widehat{T}^{\mu \nu}u_{\mu}u_{\nu}=
\frac{1}{2}\left[(\partial_{\tau}\widehat{\psi})^2+(\partial_x\widehat{\psi})^2+(\partial_y\widehat{\psi})^2+\frac{1}{\tau^2}(\partial_{\eta}\widehat{\psi})^2+m^2\widehat{\psi}^2\right].
\end{equation}
By using the quantum field expansion \eqref{boosteq09} together with the integral representation of the Hankel functions \eqref{boosteq10} and exploiting the invariance of $h(\tau)$ and $h^*(\tau)$ under reflections ${\sf p}\mapsto -{\sf p}$, the following expression is obtained
\begin{equation}
\begin{split}
\widehat{\Pi}(\tau)=&
\tau \int {\rm d}^2{\rm p_T}\,{\rm d} \mu \frac{\pi}{8}\left[\left(|\partial_{\tau}h(\tau)|^2+\omega^2(\tau)|h(\tau)|^2\right)\left(\widehat{b}_{\sf p}\widehat{b}_{\sf p}^{\dagger}+\widehat{b}_{\sf p}^{\dagger}\widehat{b}_{\sf p}\right)\right.\\
&\left.+\left((\partial_{\tau}h(\tau))^2+\omega^2(\tau)h^2(\tau)\right)\widehat{b}_{\sf p}\widehat{b}_{-{\sf p}}+
\left((\partial_{\tau}h^*(\tau))^2+\omega^2(\tau)(h^*(\tau))^2\right)\widehat{b}_{\sf p}^{\dagger}\widehat{b}_{-{\sf p}}^{\dagger}\right].
\end{split}
\end{equation}
It is convenient to define the positive real function $K(\tau)$ and the complex function $\Lambda(\tau)$ as
\begin{equation}\label{boosteq35}
K(\tau)\equiv \frac{\pi \tau}{4\omega(\tau)}\left[|\partial_{\tau}h(\tau)|^2+\omega^2(\tau)|h(\tau)|^2\right],
\end{equation}
\begin{equation}\label{boosteq36}
\Lambda(\tau)\equiv \frac{\pi \tau}{4\omega(\tau)}\left[(\partial_{\tau}h(\tau))^2+\omega^2(\tau)h^2(\tau)\right].
\end{equation}
Let us stress that the ${\sf p}$-dependence is intended.
Being $h(\tau)$ and $\omega(\tau)$ invariant under momentum reflections ${\sf p}\mapsto -{\sf p}$, so are $K(\tau)$ and $\Lambda(\tau)$, but more interestingly they are related by
\begin{equation}\label{boosteq14}
K^2(\tau)-|\Lambda(\tau)|^2=1.
\end{equation}
This occurs thanks to the left-hand side being proportional to the Wronskian of the Hankel functions
\begin{equation}
K^2(\tau)-|\Lambda(\tau)|^2=
-\left(\frac{\pi m_{\rm T} \tau}{4}\right)^2
\left(W[{\rm H}^{(2)}_{i\mu}(m_{\rm T} \tau),{\rm H}^{(1)}_{i\mu}(m_{\rm T} \tau)]\right)^2,
\end{equation}
which not only is known, but it turns out to be a very simple function of the argument \cite{Gradshteyn:1702455}
\begin{equation}
W[{\rm H}^{(2)}_{i\nu}(x),{\rm H}^{(1)}_{i\nu}(x)]=
{\rm H}^{(2)'}_{i\nu}(x){\rm H}^{(1)}_{i\nu}(x)-{\rm H}^{(1)'}_{i\nu}(x){\rm H}^{(2)}_{i\nu}(x)=
\frac{4i}{\pi x}.
\end{equation}
This will greatly simplify the calculation of thermal expectation values at local thermodynamic equilibrium in the next Section.
The property \eqref{boosteq14} can also be rephrased by expressing $K(\tau)$ and $\Lambda(\tau)$ in terms of some hyperbolic angle $\Theta({\sf p},\tau)$ and phase $\chi({\sf p},\tau)$
\begin{equation}\label{boosteq22}
K(\tau)=\cosh(2\Theta(\tau)),\qquad
\Lambda(\tau)={\rm e}^{i\chi(\tau)}\sinh(2\Theta(\tau)),
\end{equation}
which will be useful later on.
For now, $\widehat{\Pi}(\tau)$ is recast in terms of $K(\tau)$ and $\Lambda(\tau)$ as
\begin{equation}\label{boosteq17}
\widehat{\Pi}(\tau)=\int {\rm d}^2{\rm p_T}\,{\rm d} \mu \frac{\omega(\tau)}{2}
\left[K(\tau)\left(\widehat{b}_{\sf p}\widehat{b}_{\sf p}^{\dagger}+\widehat{b}_{\sf p}^{\dagger}\widehat{b}_{\sf p}\right)+\Lambda(\tau)\widehat{b}_{\sf p}\widehat{b}_{-{\sf p}}+\Lambda^*(\tau)\widehat{b}_{\sf p}^{\dagger}\widehat{b}_{-{\sf p}}^{\dagger}\right]
\end{equation}
At variance with equations \eqref{gteq28} and \eqref{gteq29} found in Chapter \ref{chapter:gteacceleration}, this expression is not diagonal in the creation and annihilation operators due to the terms proportional to $\Lambda(\tau)$ and $\Lambda^*(\tau)$.
This is quite unfortunate, for if it were diagonal, thermal expectation values of products of creation and annihilation operators, whence of operators quadratic in the field, could be calculated by using standard methods.
In the following, we look for a suitable Bogolyubov transformation to diagonalize $\widehat{\Pi}(\tau)$.
\subsection{Diagonalization of the density operator}
Let us define a new set of creation and annihilation operators by combining $\widehat{b}_{\sf p}$ and $\widehat{b}_{\sf p}^{\dagger}$ to attempt a diagonalization of $\widehat{\Pi}(\tau)$ in \eqref{boosteq17}, whence $\widehat{\rho}_{\rm LE}(\tau)$ in \eqref{boosteq15}.
Since the terms we wish to get rid of mix $\widehat{b}_{\sf p}$ with $\widehat{b}_{-{\sf p}}$ and their Hermitean conjugates, we start with a Bogolyubov transformation of the following form
\begin{equation}\label{boosteq19}
\begin{split}
\widehat{\xi}_{\sf p}^{\dagger}(\tau)\equiv &A({\sf p},\tau)\widehat{b}_{\sf p}^{\dagger}-B({\sf p},\tau)\widehat{b}_{-{\sf p}}\\
\widehat{\xi}_{\sf p}(\tau)\equiv &A^*({\sf p},\tau)\widehat{b}_{\sf p}-B^*({\sf p},\tau)\widehat{b}_{-{\sf p}}^{\dagger}
\end{split}
\end{equation}
where $A$ and $B$ are complex functions to be determined.
In order for $\widehat{\xi}_{\sf p}^{\dagger}(\tau)$ and $\widehat{\xi}_{\sf p}(\tau)$ to be creation and annihilation operators respectively, they are demanded to obey the usual algebra
\begin{equation}\label{boosteq27}
\begin{split}
[\widehat{\xi}_{\sf p}(\tau),\widehat{\xi}_{{\sf p}'}^{\dagger}(\tau)]=&
\delta^2({\bf p}_{\rm T}-{\bf p}_{\rm T}')\,\delta(\mu-\mu')\\
[\widehat{\xi}_{\sf p}(\tau),\widehat{\xi}_{{\sf p}'}(\tau)]=&0=
[\widehat{\xi}_{\sf p}^{\dagger}(\tau),\widehat{\xi}_{{\sf p}'}^{\dagger}(\tau)].
\end{split}
\end{equation}
By enforcing the commutation relations \eqref{boosteq18}, these in turn imply
\begin{equation}
\begin{split}
&\left(|A({\sf p},\tau)|^2-|B({\sf p},\tau)|^2\right)\delta^2({\bf p}_{\rm T}-{\bf p}_{\rm T}')\,\delta(\mu-\mu')=\delta^2({\bf p}_{\rm T}-{\bf p}_{\rm T}')\,\delta(\mu-\mu')\\
&\left[A^*(-{\sf p},\tau)B^*({\sf p},\tau)-A^*({\sf p},\tau)B^*(-{\sf p},\tau)\right]\delta^2({\bf p}_{\rm T}+{\bf p}_{\rm T}')\,\delta(\mu+\mu')=0\\
&\left[A({\sf p},\tau)B(-{\sf p},\tau)-A(-{\sf p},\tau)B({\sf p},\tau)\right]\delta^2({\bf p}_{\rm T}+{\bf p}_{\rm T}')\,\delta(\mu+\mu')=0
\end{split}
\end{equation}
whose a possible solution is
\begin{equation}\label{boosteq20}
A({\sf p},\tau)=A(-{\sf p},\tau),\qquad
B({\sf p},\tau)=B(-{\sf p},\tau),\qquad
|A({\sf p},\tau)|^2-|B({\sf p},\tau)|^2=1
\end{equation}
hence the following hyperbolic expressions
\begin{equation}\label{boosteq21}
A({\sf p},\tau)=\cosh \theta({\sf p},\tau) \,{\rm e}^{i\chi_A({\sf p},\tau)},\qquad
B({\sf p},\tau)=\sinh \theta({\sf p},\tau) \,{\rm e}^{i\chi_B({\sf p},\tau)}
\end{equation}
for some hyperbolic angle $\theta({\sf p},\tau)$ and phases $\chi_A({\sf p},\tau)$ and $\chi_B({\sf p},\tau)$.
This is not the only possible solution, but we choose it as it makes it easy to invert the Bogolyubov transformation \eqref{boosteq19} as
\begin{equation}\label{boosteq24}
\begin{split}
\widehat{b}_{\sf p} =& A({\sf p},\tau)\widehat{\xi}_{\sf p}(\tau) + B^*({\sf p},\tau)\widehat{\xi}_{-{\sf p}}^{\dagger}(\tau)\\
\widehat{b}_{\sf p}^{\dagger} =& A^*({\sf p},\tau)\widehat{\xi}_{\sf p}^{\dagger}(\tau) + B({\sf p},\tau)\widehat{\xi}_{-{\sf p}}(\tau),
\end{split}
\end{equation}
plug it into \eqref{boosteq17} and obtain
\begin{equation}\label{boosteq26}
\begin{split}
\widehat{\Pi}(\tau)=&
\int{\rm d}^2{\rm p_T}\,{\rm d}\mu\,\frac{\omega}{2}\left\{
\left[K\left(|A|^2+|B|^2\right)+\Lambda AB^*+\Lambda^*A^*B\right]\left(\widehat{\xi}_{\sf p}\widehat{\xi}_{\sf p}^{\dagger}+\widehat{\xi}_{\sf p}^{\dagger}\widehat{\xi}_{\sf p}\right)\right.\\
&\left.+\left(2KAB+\Lambda A^2+\Lambda^*B^2\right)\widehat{\xi}_{\sf p}\widehat{\xi}_{-{\sf p}}+\left(2KA^*B^*+\Lambda^*{A^*}^2+\Lambda{B^*}^2\right)\widehat{\xi}_{\sf p}^{\dagger}\widehat{\xi}_{-{\sf p}}^{\dagger}\right\},
\end{split}
\end{equation}
where the invariance under momentum reflections ${\sf p}\mapsto -{\sf p}$ was used.
In order to make $\widehat{\Pi}(\tau)$ diagonal, the second line must vanish, so we look for $A$ and $B$ such that
\begin{subequations}
\begin{align}
2KAB+\Lambda A^2+\Lambda^*B^2=&0\label{boosteq23a}\\
2KA^*B^*+\Lambda^*{A^*}^2+\Lambda{B^*}^2=&0.
\end{align}
\end{subequations}
The two equations are the complex conjugate of each other, so they are in fact one complex equation, that is two real equations.
Apparently there are four real unknowns, namely the moduli and phases of $A$ and $B$, but the moduli are related by the last equation in \eqref{boosteq20}.
Moreover, the two equations are invariant under an overall phase shift of both $A$ and $B$, so they actually depend on the difference between the two phases instead of the phases themselves.
In summary, they are two real equations in two real unknowns, one modulus and the phase difference, so we can try to solve the system.
With the hyperbolic expressions \eqref{boosteq21} and \eqref{boosteq22}, equation \eqref{boosteq23a} can be rewritten as
\begin{equation}
\cosh(2\Theta)\sinh(2\theta)\,{\rm e}^{i(\chi_A+\chi_B)}+\sinh(2\Theta)\left(\cosh^2\theta \,{\rm e}^{i(\chi+2\chi_A)}+\sinh^2\theta \,{\rm e}^{i(2\chi_B-\chi)}\right)=0
\end{equation}
whose solution is
\begin{equation}
\chi_B({\sf p},\tau)-\chi_A({\sf p},\tau)=\chi({\sf p},\tau),\qquad
\theta({\sf p},\tau)=-\Theta({\sf p},\tau).
\end{equation}
By setting $\chi_A\equiv 0$ without loss of generality, the functions $A$ and $B$ fulfilling the Bogolyubov relations \eqref{boosteq19} are
\begin{equation}\label{boosteq25}
A({\sf p},\tau)=\cosh \Theta({\sf p},\tau),\qquad
B({\sf p},\tau)=-\sinh \Theta({\sf p},\tau) \,{\rm e}^{i\chi({\sf p},\tau)},
\end{equation}
so the inverse transformation \eqref{boosteq24} reads
\begin{equation}\label{boosteq29}
\begin{split}
\widehat{b}_{\sf p}=&\cosh \Theta({\sf p}, \tau)\widehat{\xi}_{\sf p}(\tau)-\sinh \Theta({\sf p},\tau)\,{\rm e}^{-i\chi({\sf p},\tau)}\widehat{\xi}_{-{\sf p}}^{\dagger}(\tau)\\
\widehat{b}_{\sf p}^{\dagger}=&\cosh \Theta({\sf p}, \tau) \widehat{\xi}_{\sf p}^{\dagger}(\tau)-\sinh \Theta ({\sf p},\tau)\,{\rm e}^{i\chi({\sf p},\tau)}\widehat{\xi}_{-{\sf p}}(\tau).
\end{split}
\end{equation}
Plugging \eqref{boosteq25} and \eqref{boosteq22} into \eqref{boosteq26} and using the commutation relations \eqref{boosteq27} we finally obtain the following result
\begin{equation}
\begin{split}
\widehat{\Pi}(\tau)=&
\int {\rm d}^2{\rm p_T}\,{\rm d} \mu \, \frac{\omega(\tau)}{2}\left(\widehat{\xi}_{\sf p}(\tau)\widehat{\xi}_{\sf p}^{\dagger}(\tau)+\widehat{\xi}_{\sf p}^{\dagger}(\tau)\widehat{\xi}_{\sf p}(\tau)\right)\\
=&\int {\rm d}^2{\rm p_T}\,{\rm d} \mu \, \omega(\tau) \left(\widehat{\xi}_{\sf p}^{\dagger}(\tau)\widehat{\xi}_{\sf p}(\tau)+\frac{1}{2}\right).
\end{split}
\end{equation}
Since this is diagonal, so is the density operator; however, note that the density operator being diagonal in a set of creation and annihilation operators different from that appearing in the quantum field expansion is at variance with the case of global thermodynamic equilibrium with acceleration studied in Chapter \ref{chapter:gteacceleration}, as can be checked by looking at equations \eqref{gteq13} and \eqref{gteq26} together with \eqref{gteq28} and \eqref{gteq29}.
With a diagonal density operator, we are now in a good spot to calculate thermal expectation values of operators quadratic in the quantum field, such as the energy-momentum tensor, by using standard methods.
\section{Local thermodynamic equilibrium analysis}
\label{sec:boost_lte_analysis}
So far, we have been concerned only with the structure of the density operator, therefore we could take either $\tau$ or $\tau_0$ equivalently since $\widehat{\rho}_{\rm LE}(\tau)$ and $\widehat{\rho}=\widehat{\rho}_{\rm LE}(\tau_0)$ have the same properties.
However, when it comes to thermal expectation values it does indeed make a difference whether the density operator is evaluated at $\tau$ or $\tau_0$.
In this Section, we will perform a local thermodynamic equilibrium analysis, therefore the $\tau$-dependence will be highlighted in order to dispel any potential doubt.
On the other hand, for ease of notation, the dependence on the momentum ${\sf p}$ will be omitted throughout whenever unnecessary.
As we shall see, we will be able to obtain analytical results, making this the first example of an exactly solvable system at local thermodynamic equilibrium, to the best of our knowledge.
\subsection{Thermal expectation values}
As a quick recap, the thermal expectation value of the energy-momentum tensor calculated with the local thermodynamic equilibrium density operator is
\begin{equation}\label{boosteq28}
\langle \widehat{T}^{\mu \nu}\rangle_{\rm LE}=\mathrm{tr}(\widehat{\rho}_{\rm LE}(\tau)\widehat{T}^{\mu \nu}(\tau)),
\end{equation}
and boost invariance constrains its form to be
\begin{equation}\label{boosteq32}
\langle \widehat{T}^{\mu \nu}\rangle_{\rm LE}=
\rho_{\rm LE}(\tau)u^{\mu}u^{\nu}+p^{\rm T}_{\rm LE}(\tau)(\hat{i}^{\mu}\hat{i}^{\nu}+\hat{j}^{\mu}\hat{j}^{\nu})+p^{\rm L}_{\rm LE}(\tau)\hat{\eta}^{\mu}\hat{\eta}^{\nu}.
\end{equation}
The local thermodynamic equilibrium density operator built with the canonical energy-momentum tensor operator of a free real scalar quantum field is not diagonal in the creation and annihilation operators $\widehat{b}_{\sf p}^{\dagger}$ and $\widehat{b}_{\sf p}$, but it is so in the creation and annihilation operators $\widehat{\xi}_{\sf p}^{\dagger}(\tau)$ and $\widehat{\xi}_{\sf p}(\tau)$ obtained from the former through the Bogolyubov transformation \eqref{boosteq19}
\begin{equation}
\widehat{\rho}_{\rm LE}(\tau)=\frac{1}{Z_{\rm LE}(\tau)}\exp \left[-\frac{\widehat{\Pi}(\tau)}{T(\tau)}\right],\qquad
\widehat{\Pi}(\tau)=\int {\rm d}^2{\rm p_T}\,{\rm d} \mu \,\omega(\tau) \left(\widehat{\xi}_{\sf p}^{\dagger}(\tau)\widehat{\xi}_{\sf p}(\tau)+\frac{1}{2}\right).
\end{equation}
This diagonal expression allows for thermal expectation values of products of $\widehat{\xi}_{\sf p}^{\dagger}(\tau)$ and $\widehat{\xi}_{\sf p}(\tau)$ to be obtained by using the standard method shown through \eqref{gteq50}--\eqref{gteq51} in Chapter \ref{chapter:gteacceleration}, hence the result
\begin{subequations}
\begin{align}
\langle \widehat{\xi}_{\sf p}^{\dagger}(\tau)\widehat{\xi}_{{\sf p}'}(\tau)\rangle_{\rm LE}=&n_{\rm B}(\tau)\,\delta^2({\bf p}_{\rm T}-{\bf p}_{\rm T}')\,\delta(\mu-\mu')\label{boosteq31a}\\
\langle \widehat{\xi}_{\sf p}(\tau)\widehat{\xi}_{{\sf p}'}^{\dagger}(\tau)\rangle_{\rm LE}=&\left(n_{\rm B}(\tau)+1\right)\delta^2({\bf p}_{\rm T}-{\bf p}_{\rm T}')\,\delta(\mu-\mu')\label{boosteq31b}\\
\langle \widehat{\xi}_{\sf p}(\tau)\widehat{\xi}_{{\sf p}'}(\tau)\rangle_{\rm LE}=&0=\langle \widehat{\xi}_{\sf p}^{\dagger}(\tau)\widehat{\xi}_{{\sf p}'}^{\dagger}(\tau)\rangle_{\rm LE}\label{boosteq31c},
\end{align}
\end{subequations}
where $n_{\rm B}(\tau)$ is the Bose-Einstein distribution
\begin{equation}
n_{\rm B}(\tau)=\frac{1}{{\rm e}^{\omega(\tau)/T(\tau)}-1}.
\end{equation}
with $\omega(\tau)$ given by \eqref{boosteq30}.
This is indeed the average number of excitations of the operator $\widehat{\xi}_{\sf p}^{\dagger}(\tau)$, but by no means a density of particles.
Since the quantum field is expanded in terms of the creation and annihilation operators $\widehat{b}_{\sf p}^{\dagger}$ and $\widehat{b}_{\sf p}$, the particle density is given by $\langle \widehat{b}_{\sf p}^{\dagger}\widehat{b}_{{\sf p}'}\rangle_{\rm LE}$.
Another reason to be interested in thermal expectation values of products of $\widehat{b}_{\sf p}^{\dagger}$ and $\widehat{b}_{\sf p}$ is the fact that the energy-momentum tensor operator is quadratic in the quantum field, therefore it contains such terms.
The thermal expectation values of these products are most readily obtained from \eqref{boosteq31a}--\eqref{boosteq31c} by means of the inverse Bogolyubov transformation \eqref{boosteq29}
\begin{subequations}
\begin{align}
\langle \widehat{b}_{\sf p}\widehat{b}_{{\sf p}'}\rangle_{\rm LE}=&
-\frac{1}{2}\sinh \left(2\Theta(\tau)\right){\rm e}^{-i\chi(\tau)}\left(2n_{\rm B}(\tau)+1\right)\delta^2({\bf p}_{\rm T}-{\bf p}_{\rm T}')\,\delta(\mu-\mu')\label{boosteq33a}\\
\langle \widehat{b}_{\sf p}^{\dagger}\widehat{b}_{{\sf p}'}^{\dagger}\rangle_{\rm LE}=&
-\frac{1}{2}\sinh \left(2\Theta(\tau)\right){\rm e}^{i\chi(\tau)}\left(2n_{\rm B}(\tau)+1\right)\delta^2({\bf p}_{\rm T}-{\bf p}_{\rm T}')\,\delta(\mu-\mu')\label{boosteq33b}\\
\langle \widehat{b}_{\sf p}\widehat{b}_{{\sf p}'}^{\dagger}\rangle_{\rm LE}=&
\left[\cosh \left(2\Theta(\tau)\right)n_{\rm B}(\tau)+\cosh^2\Theta(\tau)\right]\delta^2({\bf p}_{\rm T}-{\bf p}_{\rm T}')\,\delta(\mu-\mu')\label{boosteq33c}\\
\langle \widehat{b}_{\sf p}^{\dagger}\widehat{b}_{{\sf p}'}\rangle_{\rm LE}=&
\left[\cosh \left(2\Theta(\tau)\right)n_{\rm B}(\tau)+\sinh^2\Theta(\tau)\right]\delta^2({\bf p}_{\rm T}-{\bf p}_{\rm T}')\,\delta(\mu-\mu').\label{boosteq33d}
\end{align}
\end{subequations}
These equations make it clear that the hyperbolic angle $\Theta(\tau)$, which can be determined from \eqref{boosteq22}, accounts for field vacuum effects.
More concerning the vacuum will be said shortly in Subsection \ref{sec:boost_lte_renormalization}.
Note that the $+1$ term in \eqref{boosteq31b}, stemming from the commutation relations \eqref{boosteq27}, gives rise to divergencies.
The details of renormalization will be shortly discussed in Subsection \ref{sec:boost_lte_renormalization}.
For the time being, we simply point out that renormalizing with respect to the vacuum of the operators $\widehat{\xi}_{\sf p}(\tau)$, whence of $\widehat{\Pi}(\tau)$, is but the subtraction of the $T(\tau)=0$ contribution, which eventually amount to canceling the $+1$ in \eqref{boosteq31b}.
With equations \eqref{boosteq33a}--\eqref{boosteq33d}, we are in a good position to calculate the thermal expectation value of the energy-momentum tensor \eqref{boosteq32}, that is the energy density, transverse pressure and longitudinal pressure at local thermodynamic equilibrium.
First of all, these scalar functions are obtained from the following contractions
\begin{equation}
\rho_{\rm LE}(\tau)=\langle \widehat{T}^{\mu \nu}\rangle_{\rm LE}u_{\mu}u_{\nu},
\end{equation}
\begin{equation}
p^{\rm T}_{\rm LE}(\tau)=\langle \widehat{T}^{\mu \nu}\rangle_{\rm LE}\hat{i}_{\mu}\hat{i}_{\nu}=
\langle \widehat{T}^{\mu \nu}\rangle_{\rm LE}\hat{j}_{\mu}\hat{j}_{\nu}=
\frac{1}{2}\langle \widehat{T}^{\mu \nu}\rangle_{\rm LE}\left(\hat{i}_{\mu}\hat{i}_{\nu}+\hat{j}_{\mu}\hat{j}_{\nu}\right),
\end{equation}
\begin{equation}
p^{\rm L}_{\rm LE}(\tau)=\langle \widehat{T}^{\mu \nu}\rangle_{\rm LE}\hat{\eta}_{\mu}\hat{\eta}_{\nu}.
\end{equation}
We take the energy density as an example, the pressures are worked out through the same procedure.
By plugging the quantum field expansion \eqref{boosteq09} into the contraction \eqref{boosteq34}, we have the following expression of the energy density in terms of thermal expectation values of products of creation and annihilation operators $\widehat{b}_{\sf p}^{\dagger}$ and $\widehat{b}_{\sf p}$
\begin{equation}\label{boosteq52}
\begin{split}
\rho_{\rm LE}(\tau)=&\langle \widehat{T}^{\mu \nu}\rangle_{\rm LE}u_{\mu}u_{\nu}=
\int\frac{{\rm d^2p_T}\,{\rm d}\mu\,{\rm d^2p_T'}\,{\rm d}\mu'}{4(4\pi)^2}\times \\
&\times \left\{\left[(\partial_{\tau}h({\sf p},\tau))(\partial_{\tau}h({\sf p}',\tau))-\left(p_xp_x'+p_yp_y'+\frac{1}{\tau^2}\mu\mu'-m^2\right)h({\sf p},\tau)h({\sf p}',\tau)\right]\times \right.\\
&\times {\rm e}^{i[({\bf p}_{\rm T}+{\bf p}_{\rm T}')\cdot {\bf x}_{\rm T}+(\mu+\mu')\eta]}\langle \widehat{b}_{\sf p}\widehat{b}_{{\sf p}'}\rangle_{\rm LE}+\\
&+\left[(\partial_{\tau}h({\sf p},\tau))(\partial_{\tau}h^*({\sf p}',\tau))+\left(p_xp_x'+p_yp_y'+\frac{1}{\tau^2}\mu\mu'+m^2\right)h({\sf p},\tau)h^*({\sf p}',\tau)\right]\times \\
&\times {\rm e}^{i[({\bf p}_{\rm T}-{\bf p}_{\rm T}')\cdot {\bf x}_{\rm T}+(\mu-\mu')\eta]}\langle \widehat{b}_{\sf p}\widehat{b}_{{\sf p}'}^{\dagger}\rangle_{\rm LE}+\\
&+\left[(\partial_{\tau}h^*({\sf p},\tau))(\partial_{\tau}h({\sf p}',\tau))+\left(p_xp_x'+p_yp_y'+\frac{1}{\tau^2}\mu\mu'+m^2\right)h^*({\sf p},\tau)h({\sf p}',\tau)\right]\times \\
&\times {\rm e}^{-i[({\bf p}_{\rm T}-{\bf p}_{\rm T}')\cdot {\bf x}_{\rm T}+(\mu-\mu')\eta]}\langle \widehat{b}_{\sf p}^{\dagger}\widehat{b}_{{\sf p}'}\rangle_{\rm LE}+\\
&+\left[(\partial_{\tau}h^*({\sf p},\tau))(\partial_{\tau}h^*({\sf p}',\tau))-\left(p_xp_x'+p_yp_y'+\frac{1}{\tau^2}\mu\mu'-m^2\right)h^*({\sf p},\tau)h^*({\sf p}',\tau)\right]\times \\
&\left.\times {\rm e}^{-i[({\bf p}_{\rm T}+{\bf p}_{\rm T}')\cdot {\bf x}_{\rm T}+(\mu+\mu')\eta]}\langle \widehat{b}_{\sf p}^{\dagger}\widehat{b}_{{\sf p}'}^{\dagger}\rangle_{\rm LE}\right\}.
\end{split}
\end{equation}
Then, by using equations \eqref{boosteq33a}--\eqref{boosteq33d}, we obtain
\begin{equation}
\begin{split}
\rho_{\rm LE}(\tau)=&
\int \frac{{\rm d^2p_T}\,{\rm d}\mu}{4(4\pi)^2}
\bigg[
\left[|\partial_{\tau}h(\tau)|^2+\omega^2(\tau)|h(\tau)|^2\right]\left(2n_{\rm B}(\tau)+1\right)\cosh(2\Theta(\tau))
\\
&-\frac{1}{2}\left[(\partial_{\tau}h(\tau))^2+\omega^2(\tau)h^2(\tau)\right]\left(2n_{\rm B}(\tau)+1\right)\sinh(2\Theta(\tau))\,{\rm e}^{-i\chi(\tau)}\\
&\left.
-\frac{1}{2}\left[(\partial_{\tau}h^*(\tau))^2+\omega^2(\tau)(h^*(\tau))^2\right]\left(2n_{\rm B}(\tau)+1\right)\sinh(2\Theta(\tau))\,{\rm e}^{i\chi(\tau)}.
\right]
\end{split}
\end{equation}
In square brackets we recognize the functions $K$ and $\Lambda$ defined in \eqref{boosteq35} and \eqref{boosteq36}.
These are expressed in terms of the hyperbolic angle $\Theta(\tau)$ and the phase $\chi(\tau)$ as in \eqref{boosteq22}, hence
\begin{equation}
\begin{split}
\rho_{\rm LE}(\tau)=&
\int {\rm d}^2{\rm p_T}\,{\rm d} \mu \frac{\omega(\tau)}{16\pi^3\tau}
\left[\cosh^2(2\Theta(\tau))-\sinh^2(2\Theta(\tau))\right]\left(2n_{\rm B}(\tau)+1\right)\\
=&\int \frac{{\rm d}^2{\rm p_T}\,{\rm d} \mu}{(2\pi)^3\tau}\omega(\tau)\left(n_{\rm B}(\tau)+\frac{1}{2}\right).
\end{split}
\end{equation}
The transverse and longitudinal pressures are worked out following the same steps, and the final expressions read
\begin{equation}\label{boosteq37}
\rho_{\rm LE}(\tau)=\int \frac{{\rm d}^2{\rm p_T}\,{\rm d} \mu}{(2\pi)^3\tau \omega(\tau)}\omega^2(\tau)\left(n_{\rm B}(\tau)+\frac{1}{2}\right),
\end{equation}
\begin{equation}\label{boosteq38}
p^{\rm T}_{\rm LE}(\tau)=\int \frac{{\rm d}^2{\rm p_T}\,{\rm d} \mu}{(2\pi)^3\tau \omega(\tau)}\frac{|{\bf p}_{\rm T}|^2}{2}\left(n_{\rm B}(\tau)+\frac{1}{2}\right),
\end{equation}
\begin{equation}\label{boosteq39}
p^{\rm L}_{\rm LE}(\tau)=\int \frac{{\rm d}^2{\rm p_T}\,{\rm d} \mu}{(2\pi)^3\tau \omega(\tau)}\frac{\mu^2}{\tau^2}\left(n_{\rm B}(\tau)+\frac{1}{2}\right).
\end{equation}
These are the integral expressions of the thermal expectation value of the canonical energy-momentum tensor of a free real scalar quantum field at local thermodynamic equilibrium in the future light-cone.
To the best of our knowledge, this was first obtained in \cite{Rindori:2021quq}.
From the mathematical point of view, there is a key fact that greatly eases the calculations to the point that simple analytical results such as \eqref{boosteq37}--\eqref{boosteq39} can be obtained, and that is the following.
The functions $A(\tau)$ and $B(\tau)$ of the Bogolyubov transformation, appearing above in the thermal expectation values of products of $\widehat{b}_{\sf p}$ and $\widehat{b}_{\sf p}^{\dagger}$, are expressed in terms of the functions $K(\tau)$ and $\Lambda(\tau)$ through equation \eqref{boosteq25}.
The functions $K(\tau)$ and $\Lambda(\tau)$, in turn, can be written in terms of the hyperbolic angle $\Theta(\tau)$ and the phase $\chi(\tau)$ thanks to the relation \eqref{boosteq14}, which eventually is a consequence of the Wronskian of the Hankel functions being reconstructed.
In summary, the ultimate reason for the above analytical results is the reconstruction of the Wronskian of the Hankel functions, occurring because the density operator and the energy-momentum tensor operator in \eqref{boosteq28} are evaluated at the same $\tau$.
In the non-equilibrium analysis to be carried out in Section \ref{sec:boost_ne_analysis}, this will no longer hold true because of the mixing of $\tau$ and $\tau_0$, as exemplified in \eqref{boosteq08}.
The difficulties we will encounter there will make us appreciate the above exact analytical results even more.
Equations \eqref{boosteq37}--\eqref{boosteq39} can be recast in a more compact form.
Let us define the following real function taking on different expressions for either the energy density, the transverse pressure or the longitudinal pressure
\begin{equation}
\gamma(\tau)\equiv \left\{
\begin{aligned}
\omega^2(\tau),\qquad \qquad &\rho_{\rm LE}(\tau)\\
-\frac{\mu^2}{\tau^2}-m^2,\qquad \qquad &p^{\rm T}_{\rm LE}(\tau)\\
-m_{\rm T}^2+\frac{\mu^2}{\tau^2},\qquad \qquad &p^{\rm L}_{\rm LE}(\tau)
\end{aligned}
\right.
\end{equation}
and a modification of the functions $K(\tau)$ and $\Lambda(\tau)$ defined in \eqref{boosteq35} and \eqref{boosteq36} respectively
\begin{equation}\label{boosteq41}
K_{\gamma}(\tau)\equiv \frac{\pi \tau}{4\omega(\tau)}\left[|\partial_{\tau}h(\tau)|^2+\gamma(\tau)|h(\tau)|^2\right],
\end{equation}
\begin{equation}\label{boosteq42}
\Lambda_{\gamma}(\tau)\equiv \frac{\pi \tau}{4\omega(\tau)}\left[(\partial_{\tau}h(\tau))^2+\gamma(\tau)h^2(\tau)\right].
\end{equation}
Once again, the ${\sf p}$-dependence is omitted in $\gamma(\tau)$, $K_{\gamma}(\tau)$ and $\Lambda_{\gamma}(\tau)$ both to ease the notation and to highlight the $\tau$-dependence, which is the conceptually important one.
Much in the same way as $K(\tau)$ and $\Lambda(\tau)$ satisfy \eqref{boosteq14} thanks to the Wrosnkian of the Hankel functions, $K_{\gamma}(\tau)$ and $\Lambda_{\gamma}(\tau)$ are related by
\begin{equation}\label{boosteq40}
K^2_{\gamma}(\tau)-|\Lambda_{\gamma}(\tau)|^2=
\frac{\gamma(\tau)}{\omega^2(\tau)}.
\end{equation}
In the case of the energy density, $K_{\gamma}(\tau)$ and $\Lambda_{\gamma}(\tau)$ coincide with $K(\tau)$ and $\Lambda(\tau)$ respectively, thus \eqref{boosteq40} reduces correctly to \eqref{boosteq14}.
With these new functions, we can further define $\Gamma^{\gamma}_{\rm LE}\equiv \{\rho_{\rm LE},p^{\rm T}_{\rm LE},p^{\rm L}_{\rm LE}\}$ as the set of functions that make up the thermal expectation value of the energy-momentum tensor distinguished by the value of the function $\gamma(\tau)$, thus the \eqref{boosteq37}--\eqref{boosteq39} can be compactly rearranged as
\begin{equation}\label{boosteq45}
\Gamma^{\gamma}_{\rm LE}(\tau)=
\int \frac{{\rm d}^2{\rm p_T}\,{\rm d} \mu}{(2\pi)^3\tau \omega(\tau)}\omega^2(\tau)
\left[K_{\gamma}(\tau)K(\tau)-{\rm Re}\left(\Lambda_{\gamma}(\tau)\Lambda^*(\tau)\right)\right]
\left(n_{\rm B}(\tau)+\frac{1}{2}\right).
\end{equation}
That is the same structure we will recover in the non-equilibrium analysis too.
However, in the present local thermodynamic equilibrium analysis we can take it a step further by recasting the expression in square brackets in the following form by using the definitions \eqref{boosteq35}, \eqref{boosteq36}, \eqref{boosteq41} and \eqref{boosteq42} and exploiting again the Wronskian of the Hankel functions
\begin{equation}
K_{\gamma}(\tau)K(\tau)-{\rm Re}\left(\Lambda_{\gamma}(\tau)\Lambda^*(\tau)\right)=
\frac{\omega^2(\tau)+\gamma(\tau)}{2\omega^2(\tau)},
\end{equation}
hence
\begin{equation}\label{boosteq46}
\Gamma^{\gamma}_{\rm LE}(\tau)=
\int \frac{{\rm d}^2{\rm p_T}\,{\rm d} \mu}{(2\pi)^3\tau \omega(\tau)}
\frac{\omega^2(\tau)+\gamma(\tau)}{2}\left(n_{\rm B}(\tau)+\frac{1}{2}\right).
\end{equation}
This can be further recast in a more familiar form.
By changing the integration variable $\mu$ to $p_z\equiv \mu/\tau$, the measure becomes
\begin{equation}\label{boosteq43}
{\rm d}^2{\rm p_T}\frac{{\rm d} \mu}{\tau}={\rm d}^2{\rm p_T}\,{\rm d} p_z={\rm d}^3{\rm p},
\end{equation}
and $\omega(\tau)$ the on-shell energy
\begin{equation}\label{boosteq44}
\omega(\tau)=\sqrt{m_{\rm T}^2+\frac{\mu^2}{\tau^2}}=\sqrt{m_{\rm T}^2+p_z^2}=\sqrt{|{\bf p}_{\rm T}|^2+p_z^2+m^2}\equiv \varepsilon.
\end{equation}
In turn, $n_{\rm B}(\tau)$ becomes the energy-dependent Bose-Einstein distribution $n_{\rm B}(\varepsilon,T(\tau))$ in the phase space, hence the the energy density \eqref{boosteq37} as well as the transverse and longitudinal pressures \eqref{boosteq38} and \eqref{boosteq39} can be written as the familiar momentum integrals of the relativistic neutral Bose gas.
Altogether, the non-renormalized thermal expectation value of the energy-momentum tensor at local equilibrium reads
\begin{equation}\label{boosteq47}
\langle \widehat{T}^{\mu \nu}\rangle_{\rm LE}=\int \frac{{\rm d}^3{\rm p}}{\varepsilon}P^{\mu}P^{\nu}\left(\frac{1}{{\rm e}^{\beta_{\lambda}P^{\lambda}}-1}+\frac{1}{2}\right),
\end{equation}
where $\beta^{\mu}$ is the four-temperature \eqref{boosteq03}.
Therefore, the thermodynamic functions $\Gamma^{\gamma}_{\rm LE}(\tau)$ are just the familiar functions of $T(\tau)$ as the usual ideal relativistic gas.
In particular, the transverse and the longitudinal pressures are in fact equal, namely
\begin{equation}
p^{\rm T}_{\rm LE}(\tau)=p^{\rm L}_{\rm LE}(\tau)\equiv p_{\rm LE}(\tau).
\end{equation}
We can convince ourselves of this equality simply by noticing that the measure \eqref{boosteq43}, the energy \eqref{boosteq44} and the Bose-Einstein distribution $n_{\rm B}(\tau)$ are all invariant under rotations of the momentum ${\bf p}$, so it makes no difference between $p_x$, $p_y$ and $p_z$.
Thus, the transverse pressure \eqref{boosteq38} and the longitudinal pressure \eqref{boosteq39} are equal.
In other words, the thermal expectation value of the energy-momentum tensor at local thermodynamic equilibrium has the ideal form
\begin{equation}
\langle \widehat{T}^{\mu \nu}\rangle_{\rm LE}=
\rho_{\rm LE}(\tau)u^{\mu}u^{\nu}-p_{\rm LE}(\tau)\Delta^{\mu \nu}.
\end{equation}
This is a somewhat unexpected feature of local thermodynamic equilibrium, totally at variance with the case of global thermodynamic equilibrium with acceleration studied in Chapter \ref{chapter:gteacceleration}
Of course, the plain thermal expectation values \eqref{boosteq45}, or equivalently \eqref{boosteq46} or \eqref{boosteq47}, are divergent owing to the $1/2$ terms stemming from the commutation relations \eqref{boosteq27}, therefore they must be suitably renormalized.
\subsection{Renormalization}
\label{sec:boost_lte_renormalization}
Renormalization in free Quantum Field Theory is most readily established by the subtraction of some vacuum expectation value.
In the present case of a relativistic quantum fluid with boost invariance, we face an ambiguity in the choice of the vacuum which is somehow different from the one discussed in Subsection \ref{sec:gte_unruh_effect} for a relativistic quantum fluid at global thermodynamic equilibrium with acceleration.
There, the quantum field was expanded in the Rindler wedges in terms of the Rindler creation and annihilation operators, and the density operator was diagonal with respect to them.
Their vacuum was then the Rindler vacuum.
Those operators were related to the usual plane wave ones by a non-trivial Bogolyubov transformation, therefore the Rindler vacuum did not coincide with the Minkowski vacuum.
Thus, the ambiguity back then was choosing between the two natural vacua given by the Rindler and the Minkowski vacuum.
By taking the standpoint of the Rindler observer, the Minkowski vacuum appeared not as empty of particles but populated by a thermal distribution of them at a temperature proportional to the acceleration.
Together with the vacuum choice ambiguity, the key ingredient that let the Unruh effect emerge was the fact that the four-temperature possessed a Killing horizon.
Here, the quantum field is expanded in the creation and annihilation operators $\widehat{b}_{\sf p}^{\dagger}$ and $\widehat{b}_{\sf p}$, but the density operator is diagonal in a different set of operators, the $\widehat{\xi}_{\sf p}^{\dagger}(\tau)$ and $\widehat{\xi}_{\sf p}(\tau)$, related to the previous ones by a non-trivial Bogolyubov transformation.
Thus, the vacua of the two set of operators do not coincide.
The $\widehat{b}_{\sf p}^{\dagger}$ and $\widehat{b}_{\sf p}$ can be obtained from the usual plane wave ones $\widehat{a}_{\bf p}^{\dagger}$ and $\widehat{a}_{\bf p}$ by means of the ``Fourier-like'' transformation \eqref{boosteq11}, which does not involve any mixing of creation and annihilation operators, therefore the vacuum of $\widehat{b}_{\sf p}$ is in fact the Minkowski vacuum $|0_{\rm M}\rangle$.
The Minkowski vacuum, of course, is a static state.
On the other hand, the vacuum of $\widehat{\xi}_{\sf p}(\tau)$ is a $\tau$-dependent state, and we indicated it as $|0_{\tau}\rangle$ to emphasize this fact
\begin{equation}\label{boosteq70}
\widehat{\xi}_{\sf p}(\tau)|0_{\tau}\rangle=0.
\end{equation}
By using standard methods \cite{Mukhanov:2007zz}, from the coefficients of the Bogolyubov transformation \eqref{boosteq29}, it can be shown that $|0_{\tau}\rangle$ is related to $|0_{\rm M}\rangle$ by \cite{Rindori:2021quq}
\begin{equation}
|0_{\tau}\rangle = \prod_{{\sf p}} \frac{1}{|\cosh \Theta({\sf p},\tau)|^{1/2}}\exp \left[ -\frac{1}{2}\tanh \Theta({\sf p},\tau) {\rm e}^{-i\chi({\sf p},\tau)} \widehat{b}_{\sf p}^{\dagger}\widehat{b}_{-{\sf p}}^{\dagger}\right] |0_{\rm M}\rangle
\end{equation}
In summary, the ambiguity here lies in the choice of either $|0_{\tau}\rangle$ or $|0_{\rm M}\rangle$, leading to two different renormalization schemes
\begin{equation}\label{boosteq48}
{\langle \widehat{T}^{\mu \nu}\rangle_{\rm LE}}_{\tau}\equiv
\langle \widehat{T}^{\mu \nu}\rangle_{\rm LE}-\langle 0_{\tau}|\widehat{T}^{\mu \nu}|0_{\tau}\rangle,
\end{equation}
\begin{equation}\label{boosteq49}
{\langle \widehat{T}^{\mu \nu}\rangle_{\rm LE}}_{\rm M}\equiv
\langle \widehat{T}^{\mu \nu}\rangle_{\rm LE}-\langle 0_{\rm M}|\widehat{T}^{\mu \nu}|0_{\rm M}\rangle.
\end{equation}
This situation is reminiscent of the Unruh effect, and indeed from the velocity field \eqref{boosteq07} a non-vanishing acceleration field can be derived; however, the relation \eqref{boosteq11} between the plane wave operators $\widehat{a}_{\bf p}^{\dagger}$ and $\widehat{a}_{\bf p}$ and the Milne ones $\widehat{b}_{\sf p}^{\dagger}$ and $\widehat{b}_{\sf p}$ is linear, making the vacua of the two sets of operators coincide.
Therefore, Milne observers count the same particles as inertial observers, given by
\begin{equation}
\langle \widehat{a}_{\bf p}^{\dagger}\widehat{a}_{\bf p}\rangle_{\rm LE}=
\frac{1}{2\pi m_{\rm T} \sqrt{\cosh {\rm y}\cosh {\rm y}'}}\int_{-\infty}^{+\infty}{\rm d} \mu \,{\rm e}^{-i\mu({\rm y}-{\rm y}')}\langle \widehat{b}_{\sf p}^{\dagger}\widehat{b}_{\sf p}\rangle_{\rm LE}.
\end{equation}
In other words, no analog of the Unruh effect is present between inertial and Milne observers.
The ultimate reason for this seems also to be the fact that the future light-cone is not a stationary spacetime, and as such it does not possess any global timelike Killing vector field.
In fact, the four-temperature \eqref{boosteq03}, which is the only possible one compatible with boost invariance, is not a Killing vector field, so it has no Killing horizon whatsoever.
Without a Killing horizon, no Unruh effect seems to emerge.
The next step is the comparison between the two renornalization schemes \eqref{boosteq48} and \eqref{boosteq49}.
Let us start from the former, which is also the most straightforward one.
In the limit of vanishing proper temperature, the local thermodynamic equilibrium density operator reduces to the vacuum $|0_{\tau}\rangle$, that is
\begin{equation}\label{boosteq71}
\lim_{T(\tau)\to 0}\widehat{\rho}_{\rm LE}(\tau)=
\lim_{T(\tau)\to 0}\frac{1}{Z_{\rm LE}(\tau)}\exp \left[-\frac{\widehat{\Pi}(\tau)}{T(\tau)}\right]=
|0_{\tau}\rangle \langle 0_{\tau}|.
\end{equation}
The first thing we can deduce from this fact is that the vacuum $|0_{\tau}\rangle \langle 0_{\tau}|$ has the same symmetries as the local thermodynamic equilibrium density operator $\widehat{\rho}_{\rm LE}(\tau)$ which, however, is less symmetric than the supposedly Poincar\'e-invariant Minkowski vacuum.
This does not imply that $|0_{\tau}\rangle$ is degenerate, but that Poincar\'e transformations will give rise to non-vanishing components of excited states.
The second one is that renormalization with respect to the vacuum $|0_{\tau}\rangle$ is essentially the subtraction of the $T(\tau)=0$ contribution.
This could just as well be realized by looking at the thermal expectation values of products of $\widehat{\xi}_{\sf p}(\tau)$ and $\widehat{\xi}_{\sf p}^{\dagger}(\tau)$ \eqref{boosteq31a}--\eqref{boosteq31c}.
There, the subtraction of the $T(\tau)=0$ contribution eventually amounts to canceling the $+1$ term in \eqref{boosteq31b}, which in turn stems from the commutation relations \eqref{boosteq27}.
In this scheme, the energy density and the pressure at local thermodynamic equilibrium become
\begin{equation}\label{boosteq56}
{\rho_{\rm LE}}_{\tau}(\tau)\equiv
\rho_{\rm LE}(\tau)-\langle 0_{\tau}|\widehat{T}^{\mu \nu}|0_{\tau}\rangle u_{\mu}u_{\nu}=
\int \frac{{\rm d}^2{\rm p_T}\,{\rm d} \mu}{(2\pi)^3\tau \omega(\tau)}\omega^2(\tau)n_{\rm B}(\tau)
\end{equation}
\begin{equation}\label{boosteq57}
{p_{\rm LE}}_{\tau}(\tau)\equiv
p_{\rm LE}(\tau)-\langle 0_{\tau}|\widehat{T}^{\mu \nu}|0_{\tau}\rangle \hat{e}_{\mu}\hat{e}_{\nu}=
\int \frac{{\rm d}^2{\rm p_T}\,{\rm d} \mu}{(2\pi)^3\tau \omega(\tau)}p_i^2(\tau)n_{\rm B}(\tau)
\end{equation}
with $\hat{e}_{\mu}$ either $\hat{i}_{\mu}$, $\hat{j}_{\mu}$ or $\hat{\eta}_{\mu}$ and $p_i$ either $p_x$, $p_y$ or $\mu/\tau$.
Equivalently, these can be recast in the more compact forms \eqref{boosteq45} and \eqref{boosteq46} as
\begin{equation}\label{boosteq58}
\begin{split}
{\Gamma^{\gamma}_{\rm LE}}_{\tau}(\tau)=&
\int \frac{{\rm d}^2{\rm p_T}\,{\rm d} \mu}{(2\pi)^3\tau \omega(\tau)}\omega^2(\tau)\left[K_{\gamma}(\tau)K(\tau)-{\rm Re}\left(\Lambda_{\gamma}(\tau)\Lambda^*(\tau)\right)\right]n_{\rm B}(\tau)\\
=&\int \frac{{\rm d}^2{\rm p_T}\,{\rm d} \mu}{(2\pi)^3\tau \omega(\tau)}
\frac{\omega^2(\tau)+\gamma(\tau)}{2}n_{\rm B}(\tau).
\end{split}
\end{equation}
These integrals are convergent due to the exponential suppression for large values of the momentum ${\sf p}$ provided by the Bose-Einstein distribution.
In particular, in the massless limit they can be carried out analytically to yield the familiar expressions of the energy density and pressure of a neutral Bose gas at temperature $T(\tau)$
\begin{equation}\label{boosteq59}
{\rho_{\rm LE}}_{\tau}(\tau)=\frac{\pi^2}{30}T^4(\tau),\qquad (m=0),
\end{equation}
\begin{equation}\label{boosteq60}
{p_{\rm LE}}_{\tau}(\tau)=\frac{1}{3}{\rho_{\rm LE}}_{\tau}(\tau)=
\frac{\pi^2}{90}T^4(\tau),\qquad (m=0).
\end{equation}
Now we turn to the renormalization scheme \eqref{boosteq49}, namely we take the standpoint either of the inertial or Milne observer and we subtract the Minkowski vacuum contribution.
In order to calculate the expectation value of the energy-momentum tensor operator in the Minkowski vacuum, we exploit the fact that $\widehat{T}^{\mu \nu}$ is built with the quantum field as in \eqref{boosteq50}, and the quantum field is in turn expanded in Milne coordinates as in \eqref{boosteq09} in terms of $\widehat{b}_{\sf p}$ and $b_{\sf p}^{\dagger}$, whose vacuum is $|0_{\rm M}\rangle$.
Much in the same way as previously done, we take the energy density as a prototype, with the pressure to be worked out following analogous steps.
The energy density in the Minkowski vacuum is given by \eqref{boosteq52} where the thermal expectation values of products of $\widehat{b}_{\sf p}$ and $\widehat{b}_{\sf p}^{\dagger}$ are supposed to be calculated on $|0_{\rm M}\rangle$ instead of using $\widehat{\rho}_{\rm LE}(\tau)$.
Of course, the products $\widehat{b}_{\sf p}\widehat{b}_{{\sf p}'}$, $\widehat{b}_{\sf p}^{\dagger}\widehat{b}_{{\sf p}'}^{\dagger}$ and $\widehat{b}_{\sf p}^{\dagger}\widehat{b}_{{\sf p}'}$ have vanishing expectation value on the Minkowski vacuum as $\widehat{b}_{\sf p}|0_{\rm M}\rangle=0$.
On the other hand, the expectation value of $\widehat{b}_{\sf p}\widehat{b}^{\dagger}_{{\sf p}'}$ is non-vanishing and worked out with the commutation relations \eqref{boosteq18}
\begin{equation}\label{boosteq51}
\langle 0_{\rm M}|\widehat{b}_{\sf p}\widehat{b}^{\dagger}_{{\sf p}'}|0_{\rm M}\rangle=
\langle 0_{\rm M}|\widehat{b}^{\dagger}_{{\sf p}'}\widehat{b}_{\sf p}|0_{\rm M}\rangle+\langle 0_{\rm M}|[\widehat{b}_{\sf p},\widehat{b}^{\dagger}_{{\sf p}'}]|0_{\rm M}\rangle=
\delta^2({\bf p}_{\rm T}-{\bf p}_{\rm T}')\,\delta(\mu-\mu').
\end{equation}
Thus, the energy density in the Minkowski vacuum is readily obtained from \eqref{boosteq52}
\begin{equation}
\begin{split}
\langle 0_{\rm M}|\widehat{T}^{\mu \nu}|0_{\rm M}\rangle u_{\mu}u_{\nu}=&
\int \frac{{\rm d^2p_T}\,{\rm d}\mu}{4(4\pi)^2}\left(|\partial_{\tau}h(\tau)|^2+\omega^2(\tau)|h(\tau)|^2\right)\\
=&\int \frac{{\rm d^2p_T}\,{\rm d}\mu}{(2\pi)^3\tau \omega(\tau)}\omega^2(\tau) \frac{K(\tau)}{2}.
\end{split}
\end{equation}
Repeating the same steps for the pressure, the same result is found with $K(\tau)$ replaced by $K_{\gamma}(\tau)$, the two of them coinciding in the case of the energy density.
We can therefore conclude that the thermal expectation value of the energy-momentum tensor renormalized with respect to the Minkowski vacuum reads
\begin{equation}
\begin{split}
{\Gamma^{\gamma}_{\rm LE}}_{\rm M}(\tau)=&
\int \frac{{\rm d}^2{\rm p_T}\,{\rm d} \mu}{(2\pi)^3\tau \omega(\tau)}\omega^2(\tau)\times \\
&\times \left\{
\left[K_{\gamma}(\tau)K(\tau)-{\rm Re}\left(\Lambda_{\gamma}(\tau)\Lambda^*(\tau)\right)\right]\left(n_{\rm B}(\tau)+\frac{1}{2}\right)-\frac{K_{\gamma}(\tau)}{2}
\right\}.
\end{split}
\end{equation}
This expression, however, subtly hides some problem.
Let us take once again the energy density as an example and split it into two terms as
\begin{equation}\label{boosteq53}
\begin{split}
{\rho_{\rm LE}}_{\rm M}(\tau)=&
\int \frac{{\rm d}^2{\rm p_T}\,{\rm d} \mu}{(2\pi)^3\tau}\omega(\tau)n_{\rm B}(\tau)-
\int \frac{{\rm d}^2{\rm p_T}\,{\rm d} \mu}{(2\pi)^3\tau}\omega(\tau)\frac{K(\tau)-1}{2}.
\end{split}
\end{equation}
The first term coincides with \eqref{boosteq56} found in the former renormalization scheme and is clearly convergent.
The second term depends on the transverse momentum ${\bf p}_{\rm T}$ only through its modulus ${\rm p_T}\equiv |{\bf p}_{\rm T}|$, therefore we can change to polar variables in the transverse momentum and factorize the angular part
\begin{equation}
\int \frac{{\rm d}^2{\rm p_T}\,{\rm d} \mu}{(2\pi)^3\tau}\omega(\tau)\frac{K(\tau)-1}{2}=
\frac{1}{(2\pi)^2\tau}\int_{-\infty}^{+\infty}{\rm d} \mu \int_0^{+\infty}{\rm d} {\rm p_T}\,{\rm p_T}\,\omega(\tau)\frac{K(\tau)-1}{2}.
\end{equation}
Changing variable from ${\rm p_T}$ to the transverse mass $m_{\rm T}$ and then to $v\equiv m_{\rm T} \tau$ we have
\begin{equation}
\frac{1}{(2\pi)^2\tau^2}\int_{-\infty}^{+\infty}{\rm d} \mu \int_{m\tau}^{+\infty}{\rm d} v\,v \sqrt{v^2+\mu^2}\frac{K(v)-1}{2}.
\end{equation}
In \cite{Rindori:2021quq}, an asymptotic analysis was performed for $m\tau \gg 1$, or, restoring the natural constants, $m\tau c^2\gg \hbar$.
As the details of this analysis are quite involved, they will be addressed in a dedicated space in Subsection \ref{sec:boost_ne_asymptotic_analysis}.
For the time being, we simply mention that the trend of $K(\tau)-1$ in this limit is
\begin{equation}\label{boosteq54}
K(\tau)-1=\cosh \Theta(\tau)-1\simeq \frac{\Theta^2(\tau)}{2}\simeq \frac{m_{\rm T}^4}{\tau^2\omega^6(\tau)}.
\end{equation}
In this limit we also have $v\gg 1$, and using \eqref{boosteq54} the integral in $v$ turns out to diverge
\begin{equation}
\int_{m\tau}^{+\infty}{\rm d} v\,v \sqrt{v^2+\mu^2}\frac{v^4}{\left(\sqrt{v^4+\mu^2}\right)^6}\sim
\int_{m\tau}^{+\infty}{\rm d} v\,v\frac{v^4}{v^5}\sim \infty.
\end{equation}
Of course this is not a formal proof, but a strong indication nonetheless that renormalizing with respect to the Minkowski vacuum provides us with thermal expectation values that are still divergent.
This fact seems to automatically single out $|0_{\tau}\rangle$ as a preferred choice, however, it must be emphasized that $|0_{\tau}\rangle$ is a $\tau$-dependent vacuum, therefore thermal expectation values renormalized accordingly might inherit some eventually undesired $\tau$-dependence.
These issues will be further discussed in Subsection \ref{sec:boost_ne_renormalization} in the context of the non-equilibrium analysis.
For the time being, we would like the takeaway to be the fact that the Minkowski vacuum renormalization scheme should be discarded as it appears to provide us with thermal expectation values that are still divergent.
Moreover, the next Subsection will make it clear why at some point we would have needed to study the effects of the $|0_{\tau}\rangle$ contributions subtraction anyway.
\subsection{Entropy current in the future light-cone}
\label{sec:boost_lte_entropy_current}
With the plain thermal expectation value of the energy-momentum tensor, we are now in a good spot to calculate the entropy current at local thermodynamic equilibrium by using the method put forward Section \ref{sec:zubarev_entropy_current_method}.
That is in fact well-suited for systems at local thermodynamic equilibrium, which is the underlying hypothesis of relativistic hydrodynamics, and global thermodynamic equilibrium is included as a particular case.
However, no known expression for the entropy current nor for the entropy production exists for a Quantum Field Theory out of thermodynamic equilibrium, therefore in the non-equilibrium analysis to be carried out in Section \ref{sec:boost_ne_analysis} we will not be able to re-apply our method.
As a quick recap, the method demands to be provided with two ingredients:
\begin{enumerate}
\item the thermal expectation value the energy-momentum tensor $\langle \widehat{T}^{\mu \nu}\rangle_{\rm LE}$, and
\item the eigenvector $|0_{\Upsilon}\rangle$ corresponding to the lowest, non-degenerate eigenvalue of $\widehat{\Upsilon}(\tau)$.
\end{enumerate}
Then, the algorithm is the following:
\begin{enumerate}
\item Take $\langle \widehat{T}^{\mu \nu}\rangle_{\rm LE}$ and subtract $\langle 0_{\Upsilon}|\widehat{T}^{\mu \nu}|0_{\Upsilon}\rangle$ by using the $\lambda$-dependent density operator $\widehat{\rho}_{\rm LE}(\lambda)$ defined according to \eqref{zubeq21}.
This way, the result is $\lambda$-dependent.
\item Contract with $\beta_{\nu}$, which is $\lambda$-independent, and integrate in $\lambda$ from $\lambda=1$ to $\lambda=+\infty$ in order to obtain the thermodynamic potential current $\phi^{\mu}$ defined in \eqref{zubeq26}.
\item Plug the result into \eqref{zubeq27} and obtain the entropy current $s^{\mu}$.
\end{enumerate}
By comparing the definition \eqref{zubeq23} of the operator $\widehat{\Upsilon}(\tau)$ with the expression \eqref{boosteq15} of the local thermodynamic equilibrium density operator with boost invariance in the future light-cone, we immediately understand that
\begin{equation}\label{boosteq55}
\widehat{\Upsilon}(\tau)=\frac{\widehat{\Pi}(\tau)}{T(\tau)}=
\frac{1}{T(\tau)}\int {\rm d}^2{\rm p_T}\,{\rm d} \mu \,\omega(\tau)\left(\widehat{\xi}_{\sf p}^{\dagger}(\tau)\widehat{\xi}_{\sf p}(\tau)+\frac{1}{2}\right).
\end{equation}
Thus, the lowest eigenvector of $\widehat{\Upsilon}(\tau)$ is the vacuum of $\widehat{\Pi}(\tau)$, that is $|0_{\tau}\rangle$.
As we discussed already in Subsection \ref{sec:boost_lte_renormalization}, $|0_{\tau}\rangle$ is a $\tau$-dependent and non-degenerate state which does not coincide with the Minkowski vacuum and possesses the same symmetries as the local thermodynamic equilibrium density operator.
Hence
\begin{equation}
\langle \widehat{T}^{\mu \nu}\rangle_{\rm LE}-\langle 0_{\Upsilon}|\widehat{T}^{\mu \nu}|0_{\Upsilon}\rangle=
\langle \widehat{T}^{\mu \nu}\rangle_{\rm LE}-\langle 0_{\tau}|\widehat{T}^{\mu \nu}|0_{\tau}\rangle,
\end{equation}
which was calculated in \eqref{boosteq56} and \eqref{boosteq57} or equivalently \eqref{boosteq58}, with analytical results given by \eqref{boosteq59} and \eqref{boosteq60} for a massless field.
So, we are good to go for what concerns the ingredients.
According to the method, the above quantity should be calculated with the modified density operator $\widehat{\rho}_{\rm LE}(\tau,\lambda)$ defined in \eqref{zubeq21} in order to be $\lambda$-dependent.
Once again, by comparing \eqref{zubeq21} with \eqref{boosteq55}, we can tell that this is but a rescaling of the proper temperature as $T(\tau)\mapsto T(\tau)/\lambda$, in fact this transformation gives precisely the sought result.
Thus, we have
\begin{equation}
\langle \widehat{T}^{\mu \nu}\rangle_{\rm LE}(\lambda)-\langle 0_{\Upsilon}|\widehat{T}^{\mu \nu}|0_{\Upsilon}\rangle(\lambda)=
{\rho_{\rm LE}}_{\tau}(\tau,\lambda)u^{\mu}u^{\nu}-
{p_{\rm LE}}_{\tau}(\tau,\lambda)\Delta^{\mu \nu},
\end{equation}
where the rescaling $T(\tau)\mapsto T(\tau)/\lambda$ affects only the energy density and the pressure because $u^{\mu}$ and $\Delta^{\mu \nu}$ are independent of $T(\tau)$.
This result should be contracted with the four-temperature while taking care that $\beta_{\nu}$ does not undergo the temperature rescaling, for it concerns only the thermal expectation value $\langle \widehat{T}^{\mu \nu}\rangle_{\rm LE}-\langle 0_{\Upsilon}|\widehat{T}^{\mu \nu}|0_{\Upsilon}\rangle$, as can be clearly seen in the derivation in Section \ref{sec:zubarev_entropy_current_method}.
Whence
\begin{equation}
\left(\langle \widehat{T}^{\mu \nu}\rangle_{\rm LE}(\lambda)-\langle 0_{\Upsilon}|\widehat{T}^{\mu \nu}|0_{\Upsilon}\rangle(\lambda)\right)\beta_{\nu}=
{\rho_{\rm LE}}_{\tau}(\tau,\lambda)\beta^{\mu},
\end{equation}
therefore we do not need to know the thermal expectation value of the whole energy-momentum tensor, only the energy density contributes to the thermodynamic potential current and the entropy current.
By looking at equation \eqref{boosteq56}, we see that the only dependence of the energy density on the proper temperature is in the Bose-Einstein distribution, thus
\begin{equation}
{\rho_{\rm LE}}_{\tau}(\tau,\lambda)=
\int \frac{{\rm d}^2{\rm p_T}\,{\rm d} \mu}{(2\pi)^3\tau \omega(\tau)}\omega^2(\tau)\frac{1}{{\rm e}^{\lambda \omega(\tau)/T(\tau)}-1}.
\end{equation}
For a massless field, the $\lambda$-integration can be carried out analytically, yielding
\begin{equation}
\int_1^{+\infty}{\rm d} \lambda \,{\rho_{\rm LE}}_{\tau}(\tau,\lambda)=
\int_1^{+\infty}{\rm d} \lambda \,\frac{\pi^2}{30}\frac{T^4(\tau)}{\lambda^4}=
\frac{\pi^2}{90}T^4(\tau),\qquad (m=0)
\end{equation}
and the following expression for the thermodynamic potential current is readily obtained
\begin{equation}
\phi^{\mu}=\frac{\pi^2}{90\beta^4}\beta^{\mu}=
\frac{\pi^2}{90}T^3(\tau)u^{\mu},\qquad (m=0).
\end{equation}
Plugging this result into \eqref{zubeq27}, we finally obtain the entropy current
\begin{equation}
s^{\mu}=\frac{2\pi^2}{45\beta^4}\beta^{\mu}=
\frac{2\pi^2}{45}T^3(\tau)u^{\mu},\qquad (m=0).
\end{equation}
These are the expressions of the thermodynamic potential current and the entropy current of a free real massless scalar field at local thermodynamic equilibrium with boost invariance in the future light-cone \cite{Rindori:2021quq}.
The entropy current coincides with the classical expression of the usual relativistic kinetic theory, therefore, no quantum corrections are apparently present, which is somewhat unexpected
\section{Non-equilibrium analysis}
\label{sec:boost_ne_analysis}
So far, we have been concerned with an analysis at local thermodynamic equilibrium, which is a non-equilibrium configurations but still close to global thermodynamic equilibrium, in the sense that the volume term in \eqref{zubeq09} containing the derivatives of the thermodynamic fields is small enough to be neglected.
In this Section, we perform a full non-equilibrium analysis instead, therefore we consider the full non-equilibrium density operator $\widehat{\rho}=\widehat{\rho}_{\rm LE}(\tau_0)$.
The calculation of the thermal expectation value of the energy-momentum tensor goes on much in the same way as we saw in the local thermodynamic equilibrium analysis.
However, there are now complications due to the mixing of terms calculated at $\tau$ and others at $\tau_0$.
It is therefore mandatory to pay attention to these dependencies, so the dependence on ${\sf p}$ will still be omitted whenever possible.
\subsection{Thermal expectation values}
\label{subsec:ne_tev}
As previously mentioned, the non-equilibrium density operator and the local thermodynamic equilibrium one have the same symmetries, so the former will simply be the latter calculated at $\tau_0$, namely
\begin{equation}\label{boosteq69}
\widehat{\rho}=\widehat{\rho}_{\rm LE}(\tau_0)=
\frac{1}{Z_{\rm LE}(\tau_0)}\exp \left[-\frac{\widehat{\Pi}(\tau_0)}{T(\tau_0)}\right]
\end{equation}
with $\widehat{\Pi}(\tau_0)$ given by
\begin{equation}
\widehat{\Pi}(\tau_0)=\int {\rm d}^2{\rm p_T}\,{\rm d} \mu \,\omega(\tau_0)\left(\widehat{\xi}_{\sf p}^{\dagger}(\tau_0)\widehat{\xi}_{\sf p}(\tau_0)+\frac{1}{2}\right).
\end{equation}
The operators $\widehat{\xi}_{\sf p}(\tau_0)$ and $\widehat{\xi}_{\sf p}^{\dagger}(\tau_0)$ are those that diagonalize $\widehat{\Pi}(\tau_0)$, which is but $\widehat{\Pi}(\tau)$ at $\tau_0$, so they are just $\widehat{\xi}_{\sf p}(\tau)$ and $\widehat{\xi}_{\sf p}^{\dagger}(\tau)$ respectively at $\tau_0$.
In terms of $\widehat{b}_{\sf p}$ and $\widehat{b}_{\sf p}^{\dagger}$, they are obtained from \eqref{boosteq29} at $\tau_0$, namely
\begin{equation}
\begin{split}
\widehat{b}_{\sf p}=&\cosh \Theta({\sf p}, \tau_0)\widehat{\xi}_{\sf p}(\tau_0)-\sinh \Theta({\sf p},\tau_0)\,{\rm e}^{-i\chi({\sf p},\tau_0)}\widehat{\xi}_{-{\sf p}}^{\dagger}(\tau_0),\\
\widehat{b}_{\sf p}^{\dagger}=&\cosh \Theta({\sf p}, \tau_0) \widehat{\xi}_{\sf p}^{\dagger}(\tau_0)-\sinh \Theta ({\sf p},\tau_0)\,{\rm e}^{i\chi({\sf p},\tau_0)}\widehat{\xi}_{-{\sf p}}(\tau_0).
\end{split}
\end{equation}
Thermal expectation values out of thermodynamic equilibrium are calculated with this density operator as anticipated in \eqref{boosteq08}, which we hereby report.
The non-equilibrium thermal expectation value of the energy-momentum tensor is
\begin{equation}
T^{\mu \nu}=\mathrm{tr}(\widehat{\rho}\widehat{T}^{\mu \nu})=\mathrm{tr}(\widehat{\rho}_{\rm LE}(\tau_0)\widehat{T}^{\mu \nu}(\tau)).
\end{equation}
This is constrained by boost invariance to be of the form \eqref{boosteq12}, that is
\begin{equation}
T^{\mu \nu}=\rho(\tau)u^{\mu}u^{\nu}+p^{\rm T}(\tau)(\hat{i}^{\mu}\hat{i}^{\nu}+\hat{j}^{\mu}\hat{j}^{\nu})+p^{\rm L}(\tau)\hat{\eta}^{\mu}\hat{\eta}^{\nu}.
\end{equation}
In order to calculate the energy density and the two pressures, we need to know the non-equilibrium thermal expectation values of products of $\widehat{b}_{\sf p}$ and $\widehat{b}_{\sf p}^{\dagger}$ with $\widehat{\rho}$, that is
\begin{subequations}
\begin{align}
\langle \widehat{b}_{\sf p}\widehat{b}_{{\sf p}'}\rangle=&
\mathrm{tr}(\widehat{\rho}\,\widehat{b}_{\sf p}\widehat{b}_{{\sf p}'})=
\mathrm{tr}(\widehat{\rho}_{\rm LE}(\tau_0)\widehat{b}_{\sf p}\widehat{b}_{{\sf p}'})\\
\langle \widehat{b}_{\sf p}^{\dagger}\widehat{b}_{{\sf p}'}^{\dagger}\rangle=&
\mathrm{tr}(\widehat{\rho}\,\widehat{b}_{\sf p}^{\dagger}\widehat{b}_{{\sf p}'}^{\dagger})=
\mathrm{tr}(\widehat{\rho}_{\rm LE}(\tau_0)\widehat{b}_{\sf p}^{\dagger}\widehat{b}_{{\sf p}'}^{\dagger})\\
\langle \widehat{b}_{\sf p}\widehat{b}_{{\sf p}'}^{\dagger}\rangle=&
\mathrm{tr}(\widehat{\rho}\,\widehat{b}_{\sf p}\widehat{b}_{{\sf p}'}^{\dagger})=
\mathrm{tr}(\widehat{\rho}_{\rm LE}(\tau_0)\widehat{b}_{\sf p}\widehat{b}_{{\sf p}'}^{\dagger})\\
\langle \widehat{b}_{\sf p}^{\dagger}\widehat{b}_{{\sf p}'}\rangle=&
\mathrm{tr}(\widehat{\rho}\,\widehat{b}_{\sf p}^{\dagger}\widehat{b}_{{\sf p}'})=
\mathrm{tr}(\widehat{\rho}_{\rm LE}(\tau_0)\widehat{b}_{\sf p}^{\dagger}\widehat{b}_{{\sf p}'}).
\end{align}
\end{subequations}
These are most readily given by evaluating \eqref{boosteq33a}--\eqref{boosteq33d} at $\tau_0$
\begin{subequations}
\begin{align}
\langle \widehat{b}_{\sf p}\widehat{b}_{{\sf p}'}\rangle=&
-\frac{1}{2}\sinh \left(2\Theta(\tau_0)\right){\rm e}^{-i\chi(\tau_0)}\left(2n_{\rm B}(\tau_0)+1\right)\delta^2({\bf p}_{\rm T}-{\bf p}_{\rm T}')\,\delta(\mu-\mu')\label{boosteq61a}\\
\langle \widehat{b}_{\sf p}^{\dagger}\widehat{b}_{{\sf p}'}^{\dagger}\rangle=&
-\frac{1}{2}\sinh \left(2\Theta(\tau_0)\right){\rm e}^{i\chi(\tau_0)}\left(2n_{\rm B}(\tau_0)+1\right)\delta^2({\bf p}_{\rm T}-{\bf p}_{\rm T}')\,\delta(\mu-\mu')\label{boosteq61b}\\
\langle \widehat{b}_{\sf p}\widehat{b}_{{\sf p}'}^{\dagger}\rangle=&
\left[\cosh \left(2\Theta(\tau_0)\right)n_{\rm B}(\tau_0)+\cosh^2\Theta(\tau_0)\right]\delta^2({\bf p}_{\rm T}-{\bf p}_{\rm T}')\,\delta(\mu-\mu')\label{boosteq61c}\\
\langle \widehat{b}_{\sf p}^{\dagger}\widehat{b}_{{\sf p}'}\rangle=&
\left[\cosh \left(2\Theta(\tau_0)\right)n_{\rm B}(\tau_0)+\sinh^2\Theta(\tau_0)\right]\delta^2({\bf p}_{\rm T}-{\bf p}_{\rm T}')\,\delta(\mu-\mu').\label{boosteq61d}
\end{align}
\end{subequations}
Taking once again the energy density as a prototype, its non-equilibrium thermal expectation value is obtained from \eqref{boosteq52} by replacing \eqref{boosteq33a}--\eqref{boosteq33d} with \eqref{boosteq61a}--\eqref{boosteq61d}
\begin{equation}
\begin{split}
\rho(\tau)=&
\mathrm{tr}(\widehat{\rho}\widehat{T}^{\mu \nu})u_{\mu}u_{\nu}=
\mathrm{tr}(\widehat{\rho}_{\rm LE}(\tau_0)\widehat{T}^{\mu \nu}(\tau))u_{\mu}u_{\nu}\\
=&\int \frac{{\rm d}^2{\rm p_T}\,{\rm d} \mu}{(2\pi)^3\tau}\omega(\tau)
\left[K(\tau)K(\tau_0)-{\rm Re}\left(\Lambda(\tau)\Lambda^*(\tau_0)\right)\right]
\left(n_{\rm B}(\tau_0)+\frac{1}{2}\right).
\end{split}
\end{equation}
As in the local thermodynamic equilibrium analysis, the pressures are worked out following the same steps.
Thus, the non-equilibrium analog of equation \eqref{boosteq45} reads
\begin{equation}\label{boosteq65}
\Gamma^{\gamma}(\tau)=
\int \frac{{\rm d}^2{\rm p_T}\,{\rm d} \mu}{(2\pi)^3\tau \omega(\tau)}\omega^2(\tau)
\left[K_{\gamma}(\tau)K(\tau_0)-{\rm Re}\left(\Lambda_{\gamma}(\tau)\Lambda^*(\tau_0)\right)\right]
\left(n_{\rm B}(\tau_0)+\frac{1}{2}\right).
\end{equation}
Note that at $\tau=\tau_0$, namely at local thermodynamic equilibrium, this equation reduces to \eqref{boosteq45}, and so the results of the local thermodynamic equilibrium analysis are recovered as expected.
However, at later times $\tau>\tau_0$, because of the mixing of some functions at $\tau$ and others at $\tau_0$, the Wronskian of the Hankel functions is, unfortunately, not reconstructed.
When written explicitly, the above equation is given by a non-trivial combination of Hankel fuctions of the same order but with different arguments which cannot be recast in any simple form.
We can therefore tell that for $\tau >\tau_0$ the thermal expectation value of the energy-momentum tensor differs from its local thermodynamic equilibrium form, but we are not able to perform the integration of this combination of Hankel functions in general.
Indeed, since we are dealing with a free Quantum Field Theory, one expects to find the same expression as for the free-streaming solution of Boltzmann equation in Milne coordinates, which is reported in Appendix \ref{appendix:free_sreaming}.
However, there will also be quantum corrections due to vacuum subtraction.
In the next Subsection, we will consider a suitable approximation in which the calculation of the non-equilibrium thermal expectation value of the energy-momentum tensor becomes feasible.
\subsection{Asymptotic analysis}
\label{sec:boost_ne_asymptotic_analysis}
In \cite{Rindori:2021quq}, the behaviour of the stress-energy tensor and related quantities for late times $\tau$ was studied.
The starting point is the function $h(\tau)$ defined in \eqref{boosteq72}, that is
\begin{equation}\label{hankelint}
h(\tau)=-i{\rm e}^{\frac{\pi}{2}\mu}{\rm H}^{(2)}_{i\mu}(m_{\rm T} \tau)
\end{equation}
The asymptotic expansion of the Hankel functions for large arguments is \cite{Gradshteyn:1702455}
\begin{equation}\label{large_x}
{\rm H}^{(2)}_{\nu}(x)\sim\sqrt{\frac{2}{\pi x}}{\rm e}^{-i\left( x-\frac{\pi}{2}\nu -\frac{\pi}{4} \right)}\sum_n
\frac{1}{(2ix)^n} \frac{\Gamma(\nu+1/2 +n)}{n!\Gamma(\nu+1/2-n)}
\end{equation}
which is valid for ${\rm Re}(\nu)>-1/2$ and $|{\rm arg}(x)|<\pi$.
By making use of the property $z\Gamma(z) = \Gamma(z+1)$, substituting $x=m_{\rm T}\tau$ and $\nu=i\mu$, and plugging into \eqref{hankelint}, we get
\begin{equation}\label{h_large_arg}
h(\tau) \sim \sqrt{\frac{-2i}{\pi m_{\rm T}\tau}}{\rm e}^{-im_{\rm T}\tau}\sum_n
\frac{1}{(2im_{\rm T}\tau)^n} \frac{\left(i\mu +\frac{1}{2} -n \right)^{(2n)}}{n!},
\end{equation}
valid for large $m_{\rm T}\tau$.
Similarly, using the exact relation \cite{Gradshteyn:1702455}
\begin{equation}
z \partial_z {\rm H}^{(2)}_{\nu}(z)
=\nu {\rm H}^{(2)}_{\nu}(z) -z {\rm H}^{(2)}_{\nu+1}(z),
\end{equation}
along with the expansion \eqref{large_x}, one obtains the expansion for the derivative $\partial_\tau h$
\begin{equation}\label{h_dot_large_arg}
\begin{split}
\partial_\tau h(\tau) \sim&
-im_{\rm T}\sqrt{\frac{-2i}{\pi m_{\rm T}\tau}}{\rm e}^{-im_{\rm T}\tau}\times \\
&\times \left[1+ \sum_{n>0} \frac{1}{(2im_{\rm T}\tau)^{n}}\left(-2i\mu \frac{\left(i\mu +\frac{3}{2} -n \right)^{(2n-2)}}{(n-1)!} + \frac{\left(i\mu +\frac{3}{2} -n \right)^{(2n)}}{n!}\right) \right].
\end{split}
\end{equation}
Particularly, retaining the terms up to first order (next to leading) in $m_{\rm T}\tau$ we get
\begin{equation}
\begin{split}
h(\tau) \simeq&
\sqrt{\frac{-2i}{\pi m_{\rm T}\tau}}{\rm e}^{-im_{\rm T}\tau}\left[ 1 -\frac{i}{2m_{\rm T}\tau} \left( i\mu -\frac{1}{2}\right)\left( i\mu +\frac{1}{2}\right)\right]\\
=&\sqrt{\frac{-2i}{\pi m_{\rm T}\tau}}{\rm e}^{-im_{\rm T}\tau}\left[1 +i\frac{1+4\mu^2}{8m_{\rm T}\tau} \right],
\end{split}
\end{equation}
\begin{equation}
\partial_{\tau} h(\tau) \simeq-im_{\rm T}\sqrt{\frac{-2i}{\pi m_{\rm T}\tau}}e^{-im_{\rm T}\tau}\left[1-i\frac{3-4\mu^2}{8m_{\rm T}\tau}\right].
\end{equation}
Feeding the above expansions in the definitions \eqref{boosteq41} and \eqref{boosteq42}, we obtain
\begin{equation}\label{kappa3}
K_{\gamma}(\tau)\simeq
\frac{1}{2 m_{\rm T}\omega(\tau)} \left[m_{\rm T}^2
+\gamma(\tau)\right],
\end{equation}
\begin{equation}\label{lambda3}
\Lambda_{\gamma}(\tau)\simeq
\frac{1}{2m_{\rm T}\omega(\tau)}
{\rm e}^{-2im_{\rm T}\tau}
\left[-m_{\rm T}^2\left(1-i\frac{3-4\mu^2}{4m_{\rm T}\tau} \right) +\gamma(\tau)\left(1+i\frac{1+4\mu^2}{4m_{\rm T}\tau} \right) \right].
\end{equation}
In the case of $K_{\gamma}(\tau)$, the remainder is of the order $1/[m_{\rm T}\omega(\tau)(m_{\rm T}\tau)^2]$, while for $\Lambda_{\gamma}$ is of order ${\rm e}^{-2im_{\rm T}\tau}/[m_{\rm T}\omega(\tau)(m_{\rm T}\tau)^2]$.
The last formulae are indicative of the behaviour both at late times and for large values of the transverse momentum ${\rm p_T}$, whence the transverse mass $m_{\rm T}$, necessary for the convergence check of the renormalized thermal expectation values.
In both cases, $m_{T}\tau\gg1$ and the previous asymptotic expansion is appropriate.
Concerning the large transverse momentum behaviour, equation \eqref{kappa3} tells us that $K(\tau)\to 1$, therefore the leading order of $K(\tau)-1$ is zero, at least at first order.
However, from equation \eqref{lambda3} and the exact relation \eqref{boosteq22}, one can obtain the value of the second order expansion.
Indeed, for $\gamma(\tau)=\omega(\tau)$, in the large $m_{\rm T}$ limit
\begin{equation}
\Lambda(\tau)=
\frac{1}{2m_{\rm T} \tau}{\rm e}^{-2im_{\rm T}\tau}
+{\rm ord}\left( \frac{{\rm e}^{-2im_{\rm T}\tau}}{m_{\rm T}^2} \right),
\end{equation}
therefore
\begin{equation}
\frac{1}{2m_{\rm T}\tau}\simeq |\Lambda(\tau)|=
\sinh \Theta(\tau)\simeq \Theta(\tau),
\end{equation}
then, still in the limit of large $m_{\rm T}$,
\begin{equation}
K(\tau)-1\simeq
\frac{1}{2}\Theta^2(\tau)\simeq
\frac{1}{8m_{\rm T}^2\tau^2}.
\end{equation}
A very similar approach can be used for the late time behaviour $\tau\to\infty$.
Namely, using the asymptotic formulae \eqref{kappa3} and \eqref{lambda3}, making an expansion of $\omega(\tau)$ and $\gamma(\tau)$ for large $\tau$ keeping only the first orders, that is up to the order $1/\tau$ for $K_{\gamma}(\tau)$ and to the order ${\rm e}^{-2im_{\rm T}\tau}/\tau$ for $\Lambda_{\gamma}(\tau)$.
This difference between the order of $K_{\gamma}(\tau)$ and $\Lambda_{\gamma}(\tau)$ is because of the different gauge functions in the two cases from the large argument expansion of the Hankel functions.
Therefore, one has
\begin{equation}
K_{\gamma}(\tau)\simeq
\frac{m_{\rm T}^2 +\tilde{\gamma}}{2 m_{\rm T}^2},
\end{equation}
\begin{equation}
\Lambda_{\gamma}(\tau)\simeq
\frac{1}{2m_{\rm T}^2 }{\rm e}^{-2im_{\rm T}\tau}
\left[\tilde{\gamma} - m_{\rm T}^2 +i\frac{m_{\rm T}^2(3-4\mu^2)+\tilde{\gamma}(1+4\mu^2)}{4m_{\rm T}\tau}\right],
\end{equation}
with $\tilde{\gamma}$ being
\begin{equation}
\tilde{\gamma}\equiv
\lim_{\tau \to \infty}\gamma(\tau)=
\begin{cases}
m_{\rm T}^2, & \mbox{for }\rho(\tau)\\
-m^2, & \mbox{for } p^{\rm T}(\tau)\\
-m_{\rm T}^2, &\mbox{for } p^{\rm L}(\tau).
\end{cases}
\end{equation}
It is important to note that, beside the case related to the energy density and only at the leading order, $\Lambda_{\gamma}(\tau)$ has a rapidly oscillating phase preventing a proper limit in the function domain.
However, it converges in the distribution domain, which is fine since it has to be integrated.
In fact the limits
\begin{equation}
\lim_{\tau \to \infty} \sin(2m_{\rm T}\tau), \qquad
\lim_{\tau \to \infty} \cos(2m_{\rm T}\tau)
\end{equation}
are proportional to Dirac delta functions.
Making use of the formula for the delta families
\begin{equation}
\delta(x)=\lim_{\epsilon\to 0} \frac{1}{\epsilon}f\left(\frac{x}{\epsilon}\right),
\end{equation}
for a generic function $f$ normalized to $1$, by indicating $\epsilon \equiv 1/\tau$ in the first case and $\epsilon \equiv 1/\sqrt{\tau}$ in the second one, we have
\begin{align}
\lim_{\epsilon \to 0} \left(\frac{1}{\pi}\frac{\sin(x/\epsilon)}{x}\right)=\delta(x)
\qquad &\Rightarrow \qquad
\sin(2m_{\rm T}\tau)\;\stackrel{\tau\to \infty}{\longrightarrow}\;2\pi m_{\rm T}\,\delta(2m_{\rm T}),\\
\lim_{\epsilon \to 0}
\left(\frac{1}{\epsilon \sqrt{\pi}}\cos\left(\frac{x^2}{\epsilon^2}\right)\right)
=\delta(x) \qquad &\Rightarrow \qquad
\cos(2m_{\rm T}\tau)\;\stackrel{\tau \to \infty}{\longrightarrow}\;
\sqrt{\frac{\pi}{\tau}}\,\delta(\sqrt{2m_{\rm T}}).
\end{align}
In both cases, the Dirac delta function is outside of the domain of integration, and all these integrals are vanishing in the long proper time limit.
These expressions can now be used to calculate the asymptotic value of the non-equilibrium thermal expectation value \eqref{boosteq65} at late times.
In particular, we want to focus on the energy density, in which case $K_{\gamma}(\tau)$ and $\Lambda_{\gamma}(\tau)$ coincide with $K(\tau)$ and $\Lambda(\tau)$ respectively.
As $\tau$ grows to infinity, $K(\tau)$
tends to the constant value $1$, whereas $\Lambda(\tau)$
vanishes, thus
\begin{equation}\label{boosteq68}
\begin{split}
\rho(\tau)=&
\int \frac{{\rm d}^2{\rm p_T}\,{\rm d} \mu}{(2\pi)^3\tau}\omega(\tau)
\left[K(\tau)K(\tau_0)-{\rm Re}\left(\Lambda(\tau)\Lambda^*(\tau_0)\right)\right]
\left(n_{\rm B}(\tau_0)+\frac{1}{2}\right)\\
\simeq&
\int \frac{{\rm d}^2{\rm p_T}\,{\rm d} \mu}{(2\pi)^3\tau}\omega(\tau)
K(\tau_0)\left(n_{\rm B}(\tau_0)+\frac{1}{2}\right)\\
=&\int \frac{{\rm d}^2{\rm p_T}\,{\rm d} \mu}{(2\pi)^3\tau}\omega(\tau)
\left(n_{\rm B}(\tau_0)+\frac{1}{2}\right)+
\int \frac{{\rm d}^2{\rm p_T}\,{\rm d} \mu}{(2\pi)^3\tau}\omega(\tau)
\left(K(\tau_0)-1\right)\left(n_{\rm B}(\tau_0)+\frac{1}{2}\right)\\
=&\int \frac{{\rm d}^2{\rm p_T}\,{\rm d} \mu}{(2\pi)^3\tau}\omega(\tau)
\left(n_{\rm B}(\tau_0)+\frac{1}{2}\right)+
\int \frac{{\rm d}^2{\rm p_T}\,{\rm d} \mu}{(2\pi)^3\tau}\omega(\tau)
2\sinh^2 \Theta(\tau_0)\left(n_{\rm B}(\tau_0)+\frac{1}{2}\right)
\end{split}
\end{equation}
where equation \eqref{boosteq22} was used in the last step.
Renormalization is about to be discussed in Subsection \ref{sec:boost_ne_renormalization}.
As we shall shortly see there, once suitably renormalized, the first integral will correspond to the classical free-streaming solution in Milne coordinates, reported in Appendix \ref{appendix:free_sreaming}.
On the other hand, the second integral will be a pure quantum correction due to vacuum effects, for it will vanish only for $\Theta(\tau_0)=0$.
\subsection{Renormalization}
\label{sec:boost_ne_renormalization}
The non-equilibrium thermal expectation values \eqref{boosteq65} as well as the asymptotic energy density \eqref{boosteq68} are divergent due to the $1/2$ terms stemming from the commutation relations between $\widehat{\xi}_{\sf p}(\tau_0)$ and $\widehat{\xi}_{\sf p}^{\dagger}(\tau_0)$ that must be suitably renormalized.
What we are facing here is essentially the same ambiguity of which vacuum contribution should be subtracted as the one discussed in Subsection \ref{sec:boost_lte_renormalization} in the context of the local thermodynamic equilibrium analysis.
The choice there was between the two vacua $|0_{\tau}\rangle$ and $|0_{\rm M}\rangle$, the former being the $\tau$-dependent vacuum of the local thermodynamic equilibrium density operator whereas the latter being the static Minkowski vacuum.
In particular, it was shown that by taking the standpoint of the inertial or the Milne observer and renormalizing with respect to the Minkowski vacuum, we ended up with thermal expectation values that were still divergent.
This persuaded us that the Minkowski vacuum renormalization scheme should be discarded.
That conclusion still holds in the present non-equilibrium analysis, but now we have to take a closer look at what the subtraction of the contribution of the $\tau$-dependent vacuum $|0_{\tau}\rangle$ implies.
The renormalized thermal expectation value of the energy-momentum tensor ought to fulfill the hydrodynamic equations $\partial_{\mu}T^{\mu \nu}=0$.
However, in the $|0_{\tau}\rangle$ scheme, the renormalized tensor
\begin{equation}
T^{\mu \nu}_{\tau}\equiv
\mathrm{tr}(\widehat{\rho}\,\widehat{T}^{\mu \nu})-\langle 0_{\tau}|\widehat{T}^{\mu \nu}|0_{\tau}\rangle=
\mathrm{tr} \left[(\widehat{\rho}-|0_{\tau}\rangle \langle 0_{\tau}|)\widehat{T}^{\mu \nu}\right]
\end{equation}
has a non-vanishing divergence due to the $\tau$-dependence of $|0_{\tau}\rangle$
\begin{equation}
\begin{split}
\partial_{\mu}T^{\mu \nu}_{\tau}=&
\partial_{\mu}\mathrm{tr} \left[(\widehat{\rho}-|0_{\tau}\rangle \langle 0_{\tau}|)\widehat{T}^{\mu \nu}\right]\\
=&\mathrm{tr} \left[(\widehat{\rho}-|0_{\tau}\rangle \langle 0_{\tau}|)\partial_{\mu}\widehat{T}^{\mu \nu}\right]-
\mathrm{tr} \left[(\partial_{\mu}|0_{\tau}\rangle \langle 0_{\tau}|)\widehat{T}^{\mu \nu}\right]\\
=&-\mathrm{tr} \left[u_{\mu}\left(\partial_{\tau}|0_{\tau}\rangle \langle 0_{\tau}|\right)\widehat{T}^{\mu \nu}\right]\ne 0,
\end{split}
\end{equation}
where the conservation law $\partial_{\mu}\widehat{T}^{\mu \nu}=0$ and the stationarity of $\widehat{\rho}$ were used.
Therefore, although the energy-momentum tensor quantum operator is conserved, its renormalized thermal expectation value does not fulfill the hydrodynamic equations in this scheme.
We conclude that, in order to have a thermal expectation which is both finite and conserved, the vacuum to be subtracted must be stationary.
Since the Minkowski vacuum can neither be chosen, our most reasonable alternative is the vacuum of the non-equilibrium density operator \eqref{boosteq69}, namely the vacuum of $\widehat{\Pi}(\tau_0)$.
That is but the state annihilated by the operators $\widehat{\xi}_{\sf p}(\tau_0)$, so, in analogy with \eqref{boosteq70}, we indicate it as $|0_{\tau_0}\rangle$
\begin{equation}
\widehat{\xi}_{\sf p}(\tau_0)|0_{\tau_0}\rangle \equiv 0.
\end{equation}
As mentioned already, the operators $\widehat{\xi}_{\sf p}(\tau_0)$ are just $\widehat{\xi}_{\sf p}(\tau)$ evaluated at a fixed $\tau_0$, hence $|0_{\tau_0}\rangle$ is $|0_{\tau}\rangle$ at $\tau_0$, and as such it is a stationary state suitable for a proper renormalization.
Much in the same way as $|0_{\tau}\rangle$ is obtained from $\widehat{\rho}_{\rm LE}(\tau)$ in the $T(\tau)\to 0$ limit as in \eqref{boosteq71}, $|0_{\tau_0}\rangle$ is obtained from $\widehat{\rho}$ in the $T(\tau_0)\to 0$ limit
\begin{equation}
\lim_{T(\tau_0)\to 0}\widehat{\rho}=
\lim_{T(\tau_0)\to 0}\widehat{\rho}_{\rm LE}(\tau_0)=
\lim_{T(\tau_0)\to 0}\frac{1}{Z_{\rm LE}(\tau_0)}\exp \left[-\frac{\widehat{\Pi}(\tau_0)}{T(\tau_0)}\right]=
|0_{\tau_0}\rangle \langle 0_{\tau_0}|.
\end{equation}
Thus, renormalization with respect to $|0_{\tau_0}\rangle$ is but the subtraction of the $T(\tau_0)=0$ contribution.
The non-equilibrium thermal expectation value of the energy-momentum tensor thereby renormalized is
\begin{equation}
T^{\mu \nu}_{\tau_0}\equiv
\mathrm{tr}(\widehat{\rho}\widehat{T}^{\mu \nu})-\langle 0_{\tau_0}|\widehat{T}^{\mu \nu}|0_{\tau_0}\rangle=
\mathrm{tr}(\widehat{\rho}_{\rm LE}(\tau_0)\widehat{T}^{\mu \nu})-\langle 0_{\tau_0}|\widehat{T}^{\mu \nu}|0_{\tau_0}\rangle,
\end{equation}
in particular, equation \eqref{boosteq65} becomes
\begin{equation}
\Gamma^{\gamma}_{\tau_0}(\tau)=
\int \frac{{\rm d}^2{\rm p_T}\,{\rm d} \mu}{(2\pi)^3\tau \omega(\tau)}\omega^2(\tau)
\left[K_{\gamma}(\tau)K(\tau_0)-{\rm Re}\left(\Lambda_{\gamma}(\tau)\Lambda^*(\tau_0)\right)\right]
n_{\rm B}(\tau_0),
\end{equation}
whereas the asymptotic expression of the energy-density at late times \eqref{boosteq68} reads
\begin{equation}
\rho_{\tau_0}(\tau)=
\int \frac{{\rm d}^2{\rm p_T}\,{\rm d} \mu}{(2\pi)^3\tau}\omega(\tau)
n_{\rm B}(\tau_0)+
\int \frac{{\rm d}^2{\rm p_T}\,{\rm d} \mu}{(2\pi)^3\tau}\omega(\tau)
2\sinh^2 \Theta(\tau_0)n_{\rm B}(\tau_0).
\end{equation}
The first term corresponds to the classical free-streaming solution in Milne coordinates, as shown in Appendix \ref{appendix:free_sreaming}.
On the other hand, the second one is a pure quantum correction due to vacuum effects, interpretation supported by the fact that it vanishes only if $\Theta(\tau_0)=0$.
Interestingly enough, the latter does not vanish at later times, in fact it can as well be comparable with the former if the main argument of $\Theta(\tau_0)$, namely $m_{\rm T} \tau_0$, is of order ${\cal O}(1)$, that is for an early decoupling of the system.
\section{Summary and outlook}
In this Chapter, we considered a relativistic quantum fluid with boost invariance.
This is of particular concern for the physics of the quark-gluon plasma as boost invariance is approximately realized in the central-rapidity region in heavy-ion collisions, although modern hydrodynamic calculations can go beyond this model by including transverse expansion and, eventually, by breaking boost invariance itself.
A boost-invariant fluid is inherently out of thermodynamic equilibrium, so the thermodynamic fields are in general unknown.
However, geometrical constraints descended from this particular symmetry that provided us with enough information on the thermodynamic fields and other quantities to at least attempt actual calculations.
After thoroughly presenting boost invariance and its properties, we considered as a Quantum Field Theory underlying the hydrodynamic theory a free real scalar field theory in the future light-cone.
This had quite some differences with respect to the case considered in Chapter \ref{chapter:gteacceleration} in the right Rindler wedge.
In the Rindler wedges, the quantum field was expanded in terms of Rindler creation and annihilation operators, which were different from the usual operators appearing in the plane waves expansion, and the density operator was diagonal with respect to them.
In particular, these two sets of operators were related by a non-trivial Bogolyubov transformation, which, together with the four-temperature having a Killing horizon, ultimately gave rise to the Unruh effect.
On the other hand, in the future light-cone the quantum field was expanded in terms of creation and annihilation operators linearly related to the plane waves ones, but the density operator was diagonalized by a new set of operators related to the former by a non-trivial Bogolyubov transformation.
Moreover, the four-temperature had no Killing horizon for it was not a Killing vector field.
In fact, the future-light-cone is not a stationary spacetime, therefore it does not possess any global timelike Killing vector field.
In view of these facts, there appeared to be no analog of the Unruh effect in the future light-cone.
With a diagonal density operator, we were in a good position to calculate thermal expectation values.
First, we performed a local thermodynamic equilibrium analysis.
At local thermodynamic equilibrium, in fact, due to the reconstruction of the Wronskian of the Hankel functions, thermal expectation values of operators of physical interest, such as the energy-momentum tensor, could be calculated in an exact analytical way.
Somehow unexpectedly, the ideal form of the energy-momentum tensor was obtained.
Renormalization was also discussed, pointing out peculiar vacuum effects: the subtraction of the Minkowski vacuum contribution ended up in thermal expectation values which were still infinite, at variance, for instance, with the case studied in Chapter \ref{chapter:gteacceleration}.
The entropy current thereby obtained seemed not to contain any apparent quantum correction, which was equally unexpected
Finally, we performed a full non-equilibrium analysis.
This time, however, because of the mixing of terms calculated at different times, thermal expectation values became non-trivial combinations of Hankel functions evaluated at different arguments.
Neither the Wronskian could be reconstructed, nor we were able to carry out exact integration of such combinations.
In order to go on, we considered an asymptotic expansion which simplified the expressions and eventually allowed us to obtain results.
Renormalization was also discussed in this limit, highlighting again the interesting vacuum effects.
The subtraction of the Minkowski vacuum contribution still ended up in divergent thermal expectation values, therefore we had to subtract the $|0_{\tau_0}\rangle$ contribution, with $|0_{\tau_0}\rangle$ being the stationary vacuum of $\widehat{\xi}_{\sf p}(\tau_0)$ at fixed $\tau_0$.
Interestingly, in this renormalization scheme the energy density turned out to be the sum of two terms: the former being the classical free-streaming solution in Milne coordinates, while the second one being a pure quantum correction due to a vacuum term.
No entropy current or entropy production rate could be calculated since no formula is currently known for such quantities in the case of a Quantum Field Theory fully out of thermodynamic equilibrium.
\section{Fluids}
\label{sec:relhydro_fluids}
When we stumble upon the word ``hydro'', our thoughts instinctively go to water and liquids in general.
This is certainly not wrong, as hydrodynamics has indeed to do with the motion of fluids, but in modern Physics the concept of fluid is far more general than that.
In order to get some insight on this point, let us consider a system of $N$ particles interacting through some coupling and look at the ratio ${\cal R}\equiv \lambda_{\rm DB}/l$, where $\lambda_{\rm DB}$ is the de Broglie wavelength associated to each particle and $l$ is the characteristic distance between particles.
If ${\cal R}\gtrsim 1$ there is a considerable overlap of the wavefunctions of the particles, thus the quantum mechanical nature of the system is important and the evolution will be governed by the Schr\"odinger equation for the $N$-particle wavefunction.
On the other hand, if ${\cal R}\ll 1$ the particles are so far away from each other that quantum interference can be neglected, so there will be $N$ single-particle wavefunctions each evolving according to the Schr\"odinger equation, moving like classical particles.
This is but the Ehrenfest's theorem, stating that the evolution of the expectation value of each observable coincides with the evolution described by classical mechanics.
Of course, if $N$ is very large, solving $N$ Schr\"odinger equations does not seem practically feasible, a statistical approach would then be necessary.
However, there is also a regime where another kind of description is possible.
Suppose that the number of particles $N$ is very large, and that the system extends over a length scale $L$ so much larger than $l$ (and of course $\lambda_{\rm DB}$) that the single-particle dynamics cannot be followed, that is ${\cal R}\lll 1$.
This separation between the microscopic and the macroscopic scales is quantified by the Knudsen number $K_{\rm N}\equiv l/L$, which should be $K_{\rm N}\ll 1$ in the above configuration.
In this case, the system can be conveniently approximated as a continuum whose dynamics can be studied collectively.
This continuum is called a \textit{fluid}, a system whose large-scale properties can be described effectively without having to worry about the features that the constituent elements have at much smaller length-scales.
A fluid is usually thought as divided in so-called \textit{fluid elements}, namely ``cells'' large enough to contain a great number of particles, but still small compared to the macroscopic scale $L$ in order to guarantee homogeneity within them.
As a consequence, in each fluid element the particles have the same average velocity and are at thermodynamic equilibrium.
The properties of neighbouring fluid elements, which can be different and even discontinuous, represent then the global properties of the fluid.
The collective dynamics of the fluid is called \textit{fluid dynamics} or \textit{hydrodynamics}: its aim is the study of the evolution of the fluid by solving its equations of motion, the \textit{hydrodynamic equations}.
Thus, hydrodynamics is an effective classical theory describing the evolution of a system in terms of few quantities, although the underlying microscopic theory might be quantum mechanical and possibly complicated.
Moreover, this definition of a fluid appears to be quite general: it certainly includes the liquids and gases we intuitively expect, but actually many other different kinds of systems as well.
These are some of the reasons why hydrodynamics finds applications in a vast number of phenomena ranging from meteorology over to heavy-ion collisions, relativistic astrophysics and cosmology.
As hydrodynamics establishes a connection between the microscopic and the macroscopic properties of a system, when does it become \textit{relativistic}?
This can occur essentially in three different contexts, which are not mutually exclusive in general.
The first one is when the velocities of the constituent particles within a fluid element are close to the speed of light, or, equivalently, when their Lorentz factor is significantly larger than one.
The second context applies instead when the Lorentz factor of the macroscopic motion of the fluid is significantly larger than one, quite independently of the microscopic properties of the fluid.
The third and final one emerges whenever the macroscopic gravitational field is strong enough to require a description in terms of General Relativity.
In this latter case, no assumption is made about the velocity of the fluid, which can even be at rest as, for instance, in a stationary relativistic star.
As we mentioned, a relativistic fluid is thought as divided in many fluid elements.
What we can calculate in those elements are averages of an underlying particle substrate.
The distribution function to calculate those averages usually comes from a relativistic kinetic theory, a model describing the non-equilibrium dynamics of a fluid whose microscopic degrees of freedom can be regarded as a diluted system of weakly interacting particles.
The averages thereby obtained enter the hydrodynamic equations, so in this sense relativistic hydrodynamics can be thought as a coarse graining of an underlying relativistic kinetic theory.
This approach works well when the mean free path of the particles is much larger than their thermal wavelength, however, it breaks down for strongly interacting systems, which require a full quantum field theoretical description in general.
This does not mean that a hydrodynamic theory cannot be built for such systems, but only that it should be done without necessarily relying on a relativistic kinetic theory.
The problem of calculating the averages entering the hydrodynamic equations in the case of an underlying Quantum Field Theory will be the subject of next Chapter.
In the present one, we assume they have been somehow obtained and we focus on their hydrodynamic evolution.
The first problem we come across when we forget about a particle substrate is the definition of a hydrodynamic velocity field $u^{\mu}$, which, in a kinetic theory, would intuitively be the average velocity of the particles.
Thus, the idea is to consider as the fundamental quantity the energy-momentum tensor $T^{\mu \nu}$, which always exists, and define a velocity field from it alone.
It is known that if $T^{\mu \nu}u_{\mu}u_{\nu}\ge 0$ for any timelike vector field $u^{\mu}$, called the \textit{weak energy condition}, then there exists a unique timelike vector field ${u_{\rm L}}^{\mu}$ with magnitude ${u_{\rm L}}^{\mu}{u_{\rm L}}_{\mu}=1$ such that $T^{\mu \nu}{u_{\rm L}}_{\mu}=\lambda {u_{\rm L}}^{\nu}$ for some $\lambda \ge 0$.
In other words, if the energy density measured by any observer is non-negative, the energy-momentum tensor has a unique timelike eigenvector.
The reference frame comoving with ${u_{\rm L}}^{\mu}$ is called the \textit{local rest frame}, and it is such that $T^{00}=\lambda$ and $T^{0i}=0$ for some $T^{ij}$.
So, one way to proceed is to take ${u_{\rm L}}^{\mu}$ as the hydrodynamic velocity field of the fluid.
However, this frame choice, called the \textit{Landau frame} (hence the subscript ${\rm L}$), is not the only possible one.
In fact, the system could also possess a charged current $j^{\mu}$, for instance an electromagnetic or baryonic current, as it often happens in astrophysics, so we might as well define the hydrodynamic velocity field as the direction of the charged current ${u_{\rm E}}^{\mu}\equiv j^{\mu}/\sqrt{j^2}$, where by $j^2$ we mean $j^2\equiv j^{\mu}j_{\mu}$.
This frame choice, called the \textit{Eckart frame} (hence the subscript ${\rm E}$), is in some sense the most straightforward generalization of the velocity field of non-relativistic hydrodynamics.
In fact, if $j^{\mu}$ is taken as the baryonic current, ${u_{\rm E}}^{\mu}$ coincides with the average particles' velocity of non-relativistic hydrodynamics.
Anyway, whatever the frame we choose, the philosophy is that the fundamental quantities are the energy-momentum tensor and the charged currents, which always exist, and the velocity field is built from them.
This is somehow the other way round of what is usually done in non-relativistic hydrodynamics, where all the quantities are built starting from the velocity field.
Once the hydrodynamic velocity field is given, the other kinematic quantities can be worked out.
The acceleration vector field, for instance, is defined as the derivative of the velocity field along the flow
\begin{equation}\label{eqrelhydro23}
A^{\mu}\equiv \frac{{\rm d} u^{\mu}}{{\rm d} \tau}=u^{\nu}\nabla_{\nu}u^{\mu},
\end{equation}
where $\tau$ is the fluid proper time, namely the proper time of an observer comoving with the fluid, and $\nabla_{\mu}$ the covariant derivative.
Of course, if the spacetime is flat, the covariant derivative is simply replaced by the standard derivative $\partial_{\mu}$.
The hydrodynamic velocity is timelike, has unit magnitude by definition
\begin{equation}\label{eqrelhydro01}
u^2\equiv u^{\mu}u_{\mu}=1
\end{equation}
and it is orthogonal to the acceleration field, which is spacelike
\begin{equation}
u^{\mu}A_{\mu}=0.
\end{equation}
For future purposes, it is also convenient to define the \textit{projection tensor} $\Delta_{\mu \nu}$, which projects tensors on the hypersurface orthogonal to the velocity field, as
\begin{equation}
\Delta_{\mu \nu}\equiv g_{\mu \nu}-u_{\mu}u_{\nu},
\end{equation}
where $g_{\mu \nu}$ is the metric tensor.
Sometimes it is also useful to decompose the covariant derivative along the velocity field and orthogonally to it, namely
\begin{equation}
\nabla_{\mu}\equiv u_{\mu}D+D_{\mu},
\end{equation}
where
\begin{equation}
D\equiv u^{\mu}\nabla_{\mu}=\frac{{\rm d}}{{\rm d} \tau},\qquad
D_{\mu}\equiv {\Delta^{\nu}}_{\mu}\nabla_{\nu}
\end{equation}
with $D$ often called the \textit{convective derivative}.
\section{Ideal hydrodynamics}
\label{sec:relhydro_ideal_hydro}
In non-relativistic hydrodynamics, a \textit{perfect} or \textit{ideal fluid} is a fluid such that, at thermodynamic equilibrium, taken any section, all the inner forces are orthogonal to that section.
This could be rephrased by saying that the stress tensor $T^{ij}$ has vanishing shear stress, which in turn means that $T^{ij}$ is diagonal with all equal elements on the diagonal, i.e.\ $T^{ij}=-p\delta^{ij}$ for some eigenvalue $p$ with $\delta^{ij}$ the Kronecker delta.
Of course, in a relativistic theory we should consider the full energy-momentum tensor instead of the stress tensor alone.
Thus, in order to generalize this definition to the relativistic case, we could exploit the aforementioned known result stating that, if the energy-momentum tensor fulfills the weak energy condition, in its local rest frame its components are $T^{00}\ge 0$, $T^{0i}=0$ and some $T^{ij}$.
Therefore, we might think of defining an ideal relativistic fluid as a system whose energy-momentum tensor at thermodynamic equilibrium is diagonal and isotropic in its local rest frame, namely
\begin{equation}
T^{\mu \nu}=(\rho+p)u^{\mu}u^{\nu}-pg^{\mu \nu}=
\rho u^{\mu}u^{\nu}-p\Delta^{\mu \nu}.
\end{equation}
Here, $\rho$ and $p$ have the meaning of equilibrium energy density and equilibrium pressure respectively, while $g^{\mu \nu}$, being in the local rest frame, is simply the Minkowskian metric tensor.
Likewise, the charged current $j^{\mu}$ reads
\begin{equation}
j^{\mu}=nu^{\mu},
\end{equation}
where $n$ is the equilibrium charge density.
\textit{Ideal hydrodynamics} corresponds to adopting the thermodynamic equilibrium forms of the energy-momentum tensor and charged current also out of equilibrium, promoting the energy density, pressure, charge density and velocity field to slowly-varying functions of the spacetime point, that is
\begin{equation}
T^{\mu \nu}(x)=\left[\rho(x)+p(x)\right]u^{\mu}(x)u^{\nu}(x)-p(x)g^{\mu \nu}(x),
\end{equation}
\begin{equation}
j^{\mu}(x)=n(x)u^{\mu}(x).
\end{equation}
The above expressions were obtained in the local rest frame, but being fully covariant they generalize to any frame, so $g^{\mu \nu}$ is in fact a generic metric tensor.
In the following, the dependence on the spacetime point will be omitted for ease of notation, unless otherwise specified.
Note that, under this hypothesis, the Landau frame coincides with the Eckart frame, namely ${u_{\rm L}}^{\mu}={u_{\rm E}}^{\mu}\equiv u^{\mu}$, so in this sense there is no ambiguity in the frame choice in ideal hydrodynamics.
This feature will not be preserved in non-ideal hydrodynamics.
The hydrodynamic equations are simply the conservation laws of the energy-momentum tensor and the charged current
\begin{equation}\label{eqrelhydro02}
\nabla_{\mu}T^{\mu \nu}=0,
\end{equation}
\begin{equation}\label{eqrelhydro03}
\nabla_{\mu}j^{\mu}=0.
\end{equation}
They amount to a set of 5 equations in 6 unknowns: the energy density $\rho$, the pressure $p$, the charge density $n$ and the three independent components of the velocity field $u^{\mu}$ (recall that $u^{\mu}$ satisfies the normalization condition \eqref{eqrelhydro01}).
Thus, in order to solve the system, we need an extra relation.
Technically, any relation between two unknowns is fine, for instance between the energy density and the pressure.
However, we can as well assume that, although globally out of equilibrium in general, for sufficiently small fluid elements the system is at thermodynamic equilibrium locally in space and time.
In this hypothesis, we can define a temperature and a chemical potential for each fluid element, so the thermodynamic quantities $\rho$, $p$ and $n$ can be related to them by an \textit{equation of state} reducing the number of unknowns to 5.
The achievement of this configuration, called \textit{local thermodynamic equilibrium}, is therefore equivalent to assuming the existence of the temperature and chemical potential for each fluid element, $T=T(x)$ and $\mu=\mu(x)$, which, from a relativistic standpoint, are defined as those measured in the local rest frame of the fluid element.
Once $T$, $\mu$ and $u^{\mu}$ are given on some three-dimensional spacelike hypersurface, one can try to solve the ideal hydrodynamic equations.
The four energy-momentum conservation equations are usually decomposed into one energy conservation equation and three momentum conservation equations, obtained by projecting \eqref{eqrelhydro02} along the velocity field and orthogonally to it respectively.
Together with the charged current conservation, they read
\begin{subequations}
\begin{align}
u_{\nu}\nabla_{\mu}T^{\mu \nu}=&\nabla_{\mu}(\rho u^{\mu})+p\nabla_{\mu}u^{\mu}=0,\label{eqrelhydro04}\\
{\Delta^{\lambda}}_{\nu}\nabla_{\mu}T^{\mu \nu}=&(\rho+p)A^{\lambda}-{\Delta^{\lambda}}_{\nu}\nabla^{\nu}p=0,\label{eqrelhydro05}
\end{align}
\end{subequations}
\begin{equation}\label{eqrelhydro07}
\nabla_{\mu}j^{\mu}=u^{\mu}\nabla_{\mu}n+n\nabla_{\mu}u^{\mu}=0.
\end{equation}
In the non-relativistic limit, equations \eqref{eqrelhydro04} and \eqref{eqrelhydro05} reproduce the well-known continuity and Euler equations respectively
\begin{subequations}
\begin{align}
\nabla_{\mu}(\rho u^{\mu})+p\nabla_{\mu}u^{\mu}=0&\xrightarrow[\nabla_{\mu}\to \partial_{\mu}]{p\ll \rho c^2}\partial_t \rho+\nabla \cdot (\rho {\bf v})=0\\
(\rho+p)A^{\lambda}-{\Delta^{\lambda}}_{\nu}\nabla^{\nu}p=0&\xrightarrow[\nabla_{\mu}\to \partial_{\mu}]{p\ll \rho c^2}\rho {\bf a}+\nabla p=0
\end{align}
\end{subequations}
where $\rho$ is now the mass density, ${\bf a}$ is the spatial component of the acceleration field $A^{\mu}$ and $\nabla$ is the gradient with respect to ${\bf x}$.
The next step is to consider the thermodynamics of ideal relativistic fluids.
The entropy $S$ of a system is constrained to never decrease in time by the second law of thermodynamics,
\begin{equation}\label{eqrelhydro18}
\frac{{\rm d} S}{{\rm d} \tau}\ge 0.
\end{equation}
In particular, the responsible for entropy production are irreversible processes, while reversible ones do not produce any entropy.
It is also known that entropy is an extensive quantity, meaning that it depends on the volume of the system.
Now, as expressed in \eqref{eqrelhydro18}, the second law of thermodynamics is a global statement, meaning that it concerns the system in its entirety.
This is fine in a non-relativistic theory, but from a relativistic standpoint the meaningful statements ought to be local.
In order to embed the extensivity and the second law of thermodynamics in a covariant language, it is postulated that there exists a vector field $s^{\mu}$, called the \textit{entropy current}, such that its integral on some 3-dimensional spacelike hypersurface $\Sigma$ gives the entropy
\begin{equation}\label{eqrelhydro19}
S\equiv \int_{\Sigma}{\rm d} \Sigma_{\mu}\,s^{\mu}.
\end{equation}
Here, ${\rm d} \Sigma_{\mu}\equiv {\rm d} \Sigma \,n_{\mu}$ with ${\rm d} \Sigma$ and $n^{\mu}$ the measure on $\Sigma$ and the timelike vector field of magnitude $n^{\mu}n_{\mu}=1$ orthogonal to $\Sigma$ respectively.
Equation \eqref{eqrelhydro19} is in fact the extensivity property, as it expresses the entropy as a volume integral.
With this definition of the entropy current, the local version of the second law of thermodynamics for any spacelike hypersurface $\Sigma$ is
\begin{equation}
\nabla_{\mu}s^{\mu}\ge 0.
\end{equation}
Now the question is what the expression of the entropy current looks like and what constraints stem from the second law of thermodynamics.
As we mentioned, in ideal hydrodynamics we assume that the energy density, the pressure and the charge density have, point by point, the same type of dependence on the temperature and chemical potential, often referred to as the \textit{thermodynamic fields}, as they do at thermodynamic equilibrium.
Thus, the equations of state read
\begin{subequations}
\begin{align}
\rho=&\rho(x)=\rho_{\rm eq}(T(x),\mu(x)),\\
p=&p(x)=p_{\rm eq}(T(x),\mu(x)),\\
n=&n(x)=n_{\rm eq}(T(x),\mu(x)).
\end{align}
\end{subequations}
This means that we can use our knowledge of equilibrium thermodynamics to study non-equilibrium thermodynamics.
For instance, at thermodynamic equilibrium we know that
\begin{equation}\label{eqrelhydro06}
Ts=\rho+p-\mu n,
\end{equation}
where $s$ is the entropy density, namely the entropy per unit volume.
In particular, given the above equations of state, equation \eqref{eqrelhydro06} becomes in fact a definition of the entropy density out of equilibrium.
Moreover, the Gibbs-Duhem equation must hold
\begin{equation}\label{eqrelhydro09}
T\,{\rm d} s={\rm d} \rho-\mu \,{\rm d} n,
\end{equation}
which combined with the differentiation of \eqref{eqrelhydro06} gives
\begin{equation}
{\rm d} p=s\,{\rm d} T+n\,{\rm d} \mu,
\end{equation}
hence
\begin{equation}
s=\left(\frac{\partial p}{\partial T}\right)_{\mu},\qquad
n=\left(\frac{\partial p}{\partial \mu}\right)_T,
\end{equation}
the symbols meaning that when deriving with respect to a thermodynamic field, the other one is kept fixed.
Thus
\begin{equation}
\nabla_{\mu}p=\left(\frac{\partial p}{\partial T}\right)_{\mu}\nabla_{\mu}T+\left(\frac{\partial p}{\partial \mu}\right)_T\nabla_{\mu}\mu=s\nabla_{\mu}T+n\nabla_{\mu}\mu,
\end{equation}
and projecting along the velocity field
\begin{equation}
u^{\mu}\nabla_{\mu}p=su^{\mu}\nabla_{\mu}T+nu^{\mu}\nabla_{\mu}\mu.
\end{equation}
Together with the derivative of \eqref{eqrelhydro06} along the flow, this implies
\begin{equation}\label{eqrelhydro10}
Tu^{\mu}\nabla_{\mu}s=u^{\mu}\nabla_{\mu}\rho-\mu u^{\mu}\nabla_{\mu}n.
\end{equation}
The first term at right-hand side can be worked out by using the equation of motion \eqref{eqrelhydro04}, while for the second one we use \eqref{eqrelhydro07}
\begin{equation}
u^{\mu}\nabla_{\mu}\rho=-(\rho+p)\nabla_{\mu}u^{\mu},\qquad
u^{\mu}\nabla_{\mu}n=-n\nabla_{\mu}u^{\mu},
\end{equation}
hence
\begin{equation}
Tu^{\mu}\nabla_{\mu}s=-(\rho+p-\mu n)\nabla_{\mu}u^{\mu}=-Ts\nabla_{\mu}u^{\mu},
\end{equation}
where in the last equality we used \eqref{eqrelhydro06}.
This is in fact a conservation equation
\begin{equation}\label{eqrelhydro08}
\nabla_{\mu}s^{\mu}=0,
\end{equation}
where we defined the entropy current of an ideal relativistic fluid as
\begin{equation}
s^{\mu}\equiv su^{\mu}.
\end{equation}
Equation \eqref{eqrelhydro08} tells us an important fact: an ideal fluid does not produce entropy, its entropy is constant in time.
As we will see, it will not be the same in general for non-ideal fluids.
In fact, the entropy current of a generic fluid will have more components other than the ideal $su^{\mu}$, those causing the entropy production rate $\nabla_{\mu}s^{\mu}$ to be greater than zero are called \textit{dissipations}.
\section{Dissipative hydrodynamics}
\label{sec:relhydro_dissipative_hydro}
Ideal hydrodynamics is obtained under the simplest hypothesis that, out of equilibrium, the energy-momentum tensor and the charged current have the same tensor structure as they do at thermodynamic equilibrium.
Indeed this is a strong assumption, which clearly becomes a poor approximation when the microscopic time scales are comparable to the macroscopic ones, namely when the condition of local thermodynamic equilibrium breaks down.
When this occurs, the ideal fluid description has to be extended by including dissipative terms,
that is by considering \textit{dissipative} or \textit{non-ideal fluids}.
The first problem we encounter when dealing with non-ideal fluids is the definition of the velocity field.
As we mentioned, for an ideal fluid the Landau and the Eckart frames coincide, making the choice of the velocity field unambiguous.
For a non-ideal fluid, this is known to no longer hold true, so one has to make a frame choice.
Both the frames have their own advantages.
For instance, the Eckart frame is more intuitive, being a quite straightforward generalization of the non-relativistic case and the continuity equation taking on a simple form, however it is not well-defined for systems with no net charge.
On the other hand, the Landau frame can be convenient as it simplifies the expression of the energy-momentum tensor, the price to pay is that the definition of the velocity field is implicit.
For more details on the frame choice in dissipative hydrodynamics, see \cite{Tsumura:2007ji}.
Whatever the frame we choose, in order to capture the effects of dissipations we have to go beyond the assumption of ideal hydrodynamics and allow the energy-momentum tensor and the charged current to contain other terms in addition to the ideal ones.
With this in mind, we can write in general
\begin{equation}
T^{\mu \nu}=\rho u^{\mu}u^{\nu}-p\Delta^{\mu \nu}+\delta T^{\mu \nu}
\end{equation}
\begin{equation}
j^{\mu}=nu^{\mu}+\delta j^{\mu \nu}
\end{equation}
where $\delta T^{\mu \nu}$ and $\delta j^{\mu}$ are the dissipative terms.
As such, they must vanish at thermodynamic equilibrium and depend on the derivatives of the thermodynamic fields.
We demand the energy and charge densities to be left unchanged by the dissipative terms, so we enforce the so-called \textit{Landau matching conditions}
\begin{equation}
\delta T^{\mu \nu}u_{\mu}u_{\nu}=0,\qquad
\delta j^{\mu}u_{\mu}=0.
\end{equation}
These allow us to write down the following irreducible tensor decompositions
\begin{equation}\label{eqrelhydro11}
T^{\mu \nu}=\rho u^{\mu}u^{\nu}-(p+\Pi)\Delta^{\mu \nu}+q^{\mu}u^{\nu}+q^{\nu}u^{\mu}+\Pi^{\mu \nu}
\end{equation}
\begin{equation}\label{eqrelhydro12}
j^{\mu}=nu^{\mu}+v^{\mu}.
\end{equation}
Here, $q^{\mu}$, $v^{\mu}$ and $\Pi^{\mu \nu}$ are tensor fields orthogonal to $u^{\mu}$, $\rho$ and $n$ are still the energy and charge densities while $p$ is in fact the equilibrium component of the pressure $p=p_{\rm eq}$, and being $p+\Pi$ the total pressure, $\Pi$ represents the \textit{non-equilibrium pressure}.
In formulae
\begin{subequations}
\begin{align}
\rho=&T^{\mu \nu}u_{\mu}u_{\nu}\\
p+\Pi=&\frac{1}{3}T^{\mu \nu}\Delta_{\mu \nu}\\
q^{\mu}=&T^{\alpha \beta}{\Delta^{\mu}}_{\alpha}u_{\beta}\\
\Pi^{\mu \nu}=&\left[\frac{1}{2}\left({\Delta^{\mu}}_{\alpha}{\Delta^{\nu}}_{\beta}+{\Delta^{\nu}}_{\alpha}{\Delta^{\mu}}_{\beta}\right)-\frac{1}{3}\Delta^{\mu \nu}\Delta_{\alpha \beta}\right]T^{\alpha \beta}\\
n=&j^{\mu}u_{\mu}\\
v^{\mu}=&j^{\nu}{\Delta^{\mu}}_{\nu}.
\end{align}
\end{subequations}
Being the dissipations irreversible processes, a non-ideal fluid will produce entropy, but in the limit of dissipations going to zero, the result of ideal hydrodynamics of vanishing entropy production rate ought to be recovered.
As a consequence, we expect the entropy current $s^{\mu}$ to have two components: one is the ideal term $su^{\mu}$ along the velocity field, while the other is orthogonal to that and depends on the dissipations.
Thermodynamic equilibrium is characterized by the absence of transport phenomena, namely by having $\Pi$, $q^{\mu}$ and $\Pi^{\mu \nu}$ all equal to zero, therefore we might as well say that, not surprisingly, the thermodynamic equilibrium state corresponds to a vanishing entropy production rate.
In the following, we will see how two approaches to dissipative hydrodynamics differ in prescribing the dissipative component of the entropy current while maintaining the same properties for the thermodynamic equilibrium state.
\subsection{Navier-Stokes theory}
This formulation of dissipative hydrodynamics was first proposed by Eckart \cite{Eckart:1940} and then slightly modified by Landau and Lifshitz \cite{landau2013fluid}.
It represents the simplest relativistic generalization of Navier-Stokes and Fourier equations of non-relativistic dissipative hydrodynamics, and it is usually referred to as \textit{classical irreversible thermodynamics}.
The fundamental feature of this theory is that the dissipative terms of the entropy current depend linearly on the dissipations $\Pi$, $q^{\mu}$ and $\Pi^{\mu \nu}$, therefore the entropy current contains up to the first derivatives of the thermodynamic parameters.
For this reason, classical irreversible thermodynamics is called a \textit{first-order theory}.
If we limit ourselves to first derivatives, meaning that deviations from thermodynamic equilibrium are small, and we still identify $s=s^{\mu}u_{\mu}$ with the entropy density, the thermodynamic relations \eqref{eqrelhydro06} and \eqref{eqrelhydro09} should keep holding.
Note that the total pressure is $p+\Pi$, so $p$ in \eqref{eqrelhydro06} must now be replaced with $p+\Pi$.
We have shown already that with \eqref{eqrelhydro06} and \eqref{eqrelhydro09} satisfied at each point, \eqref{eqrelhydro10} is obtained.
The first term at right-hand side is worked out from the projection of the equations of motion $\nabla_{\mu}T^{\mu \nu}=0$ along the flow, in particular we note that
\begin{equation}
u_{\nu}\nabla_{\mu}T^{\mu \nu}=0
\qquad \Rightarrow \qquad
\nabla_{\mu}(T^{\mu \nu}u_{\nu})-T^{\mu \nu}\nabla_{\mu}u_{\nu}=0.
\end{equation}
By using the expression \eqref{eqrelhydro11} of the energy-momentum tensor, we have
\begin{equation}
u^{\mu}\nabla_{\mu}\rho=-\rho \nabla_{\mu}u^{\mu}-\nabla_{\mu}q^{\mu}-(p+\Pi)\nabla_{\mu}u^{\mu}+q^{\mu}A_{\mu}+\Pi^{\mu \nu}\nabla_{\mu}u_{\nu}.
\end{equation}
For the second term, we consider the equation of motion $\nabla_{\mu}j^{\mu}=0$.
Using \eqref{eqrelhydro12} we find
\begin{equation}
u^{\mu}\nabla_{\mu}n=-n\nabla_{\mu}u^{\mu}-\nabla_{\mu}v^{\mu}.
\end{equation}
Plugging these two equations into \eqref{eqrelhydro10}, we get
\begin{equation}
\nabla_{\mu}(su^{\mu})=-\frac{1}{T}\nabla_{\mu}q^{\mu}+\frac{q^{\mu}}{T}A_{\mu}+\frac{\Pi^{\mu \nu}}{T}\nabla_{\mu}u_{\nu}+\frac{\mu}{T}\nabla_{\mu}v^{\mu}.
\end{equation}
Finally, by defining $\zeta \equiv \mu/T$, we have
\begin{equation}\label{eqrelhydro13}
\nabla_{\mu}\left(su^{\mu}+\frac{q^{\mu}}{T}-\zeta v^{\mu}\right)=-\frac{q^{\mu}}{T^2}\nabla_{\mu}T+\frac{q^{\mu}}{T}A_{\mu}+\frac{\Pi^{\mu \nu}}{T}\nabla_{\mu}u_{\nu}-v^{\mu}\nabla_{\mu}\zeta.
\end{equation}
The non-relativistic limit of this equation is the known equation for the entropy production rate of non-relativistic dissipative hydrodynamics, provided that $q^{\mu}$ and $v^{\mu}$ are interpreted as the \textit{heat flow} and the \textit{current flow} respectively.
Moreover, $\Pi^{\mu \nu}$ is called the \textit{viscous stress tensor}.
Thus, we are persuaded to identifying the entropy current of dissipative hydrodynamics with
\begin{equation}\label{eqrelhydro16}
s^{\mu}=su^{\mu}+\frac{q^{\mu}}{T}-\zeta v^{\mu}.
\end{equation}
So far, we do not have any information on the expression of the dissipations.
This is obtained by enforcing the second law of thermodynamics, namely by demanding \eqref{eqrelhydro13} to be non-negative.
Being the dissipative terms are independent from each other, we will have a constraint for each of them.
The following expressions of the dissipations are such that the second law is fulfilled for any value of the temperature and velocity field
\begin{subequations}
\begin{align}
q^{\mu}\equiv&\kappa \Delta^{\mu \nu}(\nabla_{\nu}T-TA_{\nu})\label{eqrelhydro15a}\\
\Pi^{\mu \nu}\equiv&2\eta \sigma^{\mu \nu}\label{eqrelhydro15b}\\
\Pi \equiv&\xi \nabla_{\mu}u^{\mu}\label{eqrelhydro15c}\\
v^{\mu}\equiv&{\cal D}\nabla_{\mu}\zeta \label{eqrelhydro15d}
\end{align}
\end{subequations}
where $\sigma^{\mu \nu}$ is the \textit{shear tensor}
\begin{equation}\label{eqrelhydro21}
\begin{split}
\sigma_{\mu \nu}\equiv &
\frac{1}{2}\left({\Delta^{\alpha}}_{\mu}{\Delta^{\beta}} _{\nu}+{\Delta^{\alpha}}_{\nu}{\Delta^{\beta}}_{\mu}-\frac{2}{3}\Delta_{\mu \nu}\Delta^{\alpha \beta}\right)\nabla_{\alpha}u_{\beta}=\\
=&\nabla_{(\mu}u_{\nu)}-A_{(\mu}u_{\nu)}-\frac{1}{3}\Delta_{\mu \nu}\nabla_{\lambda}u^{\lambda},
\end{split}
\end{equation}
in fact, the entropy production rate
\begin{equation}
\nabla_{\mu}s^{\mu}=-\frac{q^{\mu}q_{\mu}}{\kappa T^2}+\frac{\Pi^2}{\xi T}+\frac{\Pi^{\mu \nu}\Pi_{\mu \nu}}{2\eta T}+\frac{v^{\mu}v_{\mu}}{\cal D}
\end{equation}
is non-negative provided that
\begin{equation}
\kappa \ge 0,\qquad \xi \ge 0,\qquad \eta \ge 0,\qquad {\cal D}\ge 0.
\end{equation}
The parameters $\kappa$, $\xi$, $\eta$ and ${\cal D}$ are called the \textit{transport coefficients}; in particular, $\kappa$ is the \textit{thermal conductivity}, $\xi$ the \textit{bulk viscosity}\footnote{The bulk viscosity is usually indicated as $\zeta$, but throughout this work and papers relevant for it, $\zeta$ always stands for the ratio between chemical potential and temperature $\zeta=\mu/T$. We choose to ``sacrify'' the traditional nomenclature because we will often deal with $\mu/T$, but never again with the bulk viscosity.}, $\eta$ the \textit{shear viscosity} and ${\cal D}$ the \textit{diffusion coefficient}.
Equations \eqref{eqrelhydro15a}--\eqref{eqrelhydro15d} expressing the transport coefficients in terms of the derivatives of the thermodynamic parameters as a consequence of the second law of thermodynamics are called the \textit{constitutive equations}.
Together with the equations of motion \eqref{eqrelhydro02} and \eqref{eqrelhydro03}, they close the set of dissipative hydrodynamics equations.
We will not go into the calculations for they are quite long, but it can be shown that this set of equations reduces to the well-known Navier-Stokes ones in the non-relativistic limit.
The constitutive equations can be used to further characterize the properties of the thermodynamic equilibrium state.
At thermodynamic equilibrium the dissipations must vanish for any value of the transport coefficients, so equations \eqref{eqrelhydro15a}--\eqref{eqrelhydro15d} imply
\begin{subequations}
\begin{align}
\nabla_{\mu}T-TA_{\mu}=&0\label{eqrelhydro20a}\\
\sigma^{\mu \nu}=&0\label{eqrelhydro20b}\\
\nabla_{\mu}u^{\mu}=&0\label{eqrelhydro20c}\\
\nabla_{\mu}\zeta=&0\label{eqrelhydro20d}.
\end{align}
\end{subequations}
By plugging the definition \eqref{eqrelhydro21} of the shear tensor into the condition \eqref{eqrelhydro20b}, combining it with \eqref{eqrelhydro20c} and dividing everything by the temperature, we obtain
\begin{equation}
\begin{split}
0=&
\frac{\nabla_{\mu}u_{\nu}}{T}+\frac{\nabla_{\nu}u_{\mu}}{T}-\frac{A_{\mu}u_{\nu}}{T}-\frac{A_{\nu}u_{\mu}}{T}=\\
=&\nabla_{\mu}\frac{u_{\nu}}{T}-\frac{u_{\nu}}{T^2}\nabla_{\mu}T+\nabla_{\nu}\frac{u_{\mu}}{T}-\frac{u_{\mu}}{T^2}\nabla_{\nu}T-\frac{A_{\mu}u_{\nu}}{T}-\frac{A_{\nu}u_{\mu}}{T}.
\end{split}
\end{equation}
Finally, by using \eqref{eqrelhydro20a} we are simply left with
\begin{equation}\label{eqrelhydro22}
\nabla_{\mu}\frac{u_{\nu}}{T}+\nabla_{\nu}\frac{u_{\mu}}{T}=0.
\end{equation}
Equations \eqref{eqrelhydro22} and \eqref{eqrelhydro20d} represent the conditions that the thermodynamic fields must fulfill in order to have thermodynamic equilibrium, namely $u^{\mu}/T$ must be a Killing vector field and $\zeta=\mu/T$ must be constant.
This statement is of fundamental importance, and will be discussed more thoroughly in Subsection \ref{sec:LE_to_GE_and_NE} in the next Chapter.
As it pertains the thermodynamic equilibrium state, it will hold true even in the framework of extended irreversible thermodynamics that we are soon going to introduce.
\newline
Although classical irreversible thermodynamics includes the dissipations in such a way to ensure the second law of thermodynamics and it reproduces the known equations in the non-relativistic limit, it also has some undesirable features.
In particular, it is unstable under perturbations and it allows for the propagation of information at infinite speed, thus breaking causality \cite{Hiscock:1983zz, Hiscock:1985zz, Hiscock:1987zz}.
Non-causality, in particular, is due to the algebraic nature of the constitutive equations \eqref{eqrelhydro15a}--\eqref{eqrelhydro15d}, which makes \textit{parabolic} equations emerge and, as a consequence, the thermodynamic fluxes react instantaneously to the gradients of the thermodynamic fields.
Actually, this feature is present in the non-relativistic Navier-Stokes theory as well, where the Fourier law for the heat flux leads to a parabolic diffusion equation for the temperature.
This is not a conceptual problem \textit{per se} in a non-relativistic theory, though a big one in a relativistic context.
A recent discussion on this subject can be found in \cite{Hoult:2020eho}, where an interpretation and a solution different from the traditional one is presented.
In order to convince ourselves of the causality breaking, we can study sound waves propagation.
For the sake of simplicity, let us consider a fluid initially at thermodynamic equilibrium and at rest in a flat spacetime, and suppose that shear viscosity is the only dissipation present \cite{Romatschke:2009im}.
To make this even simpler, we take a perturbation of the velocity field along the $x$ direction, independent of $y$ and $z$
\begin{equation}
u^{\mu}={u_0}^{\mu}+\delta u^{\mu}(t,x),
\end{equation}
where the initial velocity field is ${u_0}^{\mu}=(1,{\bf 0})$, being the system initially at rest.
The density and pressure will change in response to this perturbation
\begin{equation}
\rho=\rho_0+\delta \rho(t,x),\qquad
p=p_0+\delta p(t,x),
\end{equation}
where $\rho_0$ and $p_0$ are their initial values.
In this setup we have the hydrodynamic equation
\begin{equation}
{\Delta^{\alpha}}_{\nu}\nabla_{\mu}T^{\mu \nu}=0
\qquad \longrightarrow \qquad
(\rho+p)A^{\alpha}-{\Delta^{\alpha}}_{\nu}\partial^{\nu}p+{\Delta^{\alpha}}_{\nu}\partial_{\mu}\Pi^{\mu \nu}=0
\end{equation}
and after some algebra we obtain the following diffusive equation at first order in the perturbations for the $y$ component $\delta u^y$
\begin{equation}
(\rho_0+p_0)\frac{\partial \delta u^y(t,x)}{\partial t}-\eta \frac{\partial^2 \delta u^y(t,x)}{\partial x^2}=0.
\end{equation}
This is in fact a parabolic equation.
As is customary in perturbative analysis, we assume that the perturbation has a simple harmonic behaviour of the type
\begin{equation}
\delta u^y(t,x)\propto {\rm e}^{-\omega t+ikx},
\end{equation}
which leads to the quadratic dispersion relation
\begin{equation}
\omega=\frac{\eta}{\rho_0+p_0}k^2.
\end{equation}
We can now estimate the velocity of propagation of the mode with wavenumber $k$
\begin{equation}
\frac{{\rm d} \omega}{{\rm d} k}=\frac{2\eta}{\rho_0+p_0}k.
\end{equation}
Clearly, there exists a critical value $k_{\rm crit}$ for the wavenumber such that this velocity equals the speed of light and exceeds it for greater values, thus breaking causality.
While this can be acceptable within a non-relativistic description, it is clearly unsatisfactory within a relativistic theory in which all signals must be contained in the light cone.
In summary: the ultimate reason for the non-causality of the Navier-Stokes theory is the parabolicity of the equations, which, in fact, turns out to be a general feature of first-order theories.
In order to avoid it and restore causality, we need \textit{hyperbolic} instead of parabolic equations.
The traditional solution to this problem is the replacement of a first-order theory with a second-order one.
\subsection{Israel-Stewart theory}
Several approaches have been developed to overcome the causality breaking and instability of classical irreversible thermodynamics.
The basic idea is to extend the space of variables by considering the dissipations as conserved variables of an ideal fluid, which allows to restore causality under some conditions.
The resulting conservation equations describe the evolution of these new ``extended dissipations'', and turn out to be hyperbolic.
The price to pay is that the theory now mathematically more complicated, as it involves a greater number of variables and parameters, and the physical meaning is not always crystal clear.
This class of theories generally goes under the name of \textit{extended irreversible thermodynamics} because of this extension of the space of variables.
A non-relativistic extended theory was first proposed by M\"uller \cite{Muller:1967zza} and later generalised to the relativistic framework by Israel \cite{Israel:1976tn} and Stewart \cite{Israel:1979wp}.
Here we just touch upon the main qualitative features without going into technical calculations.
Supported by evidence from relativistic kinetic theory, Israel and Stewart argued that the entropy current in equation \eqref{eqrelhydro16} is not the most general one, but is in fact the first-order truncation of a more elaborate expression that also contains terms quadratic in the dissipation, which is therefore of second-order in the derivatives of the thermodynamic fields.
In the Eckart frame, for simplicity, it reads
\begin{equation}
s^{\mu}=su^{\mu}+\frac{q^{\mu}}{T}-\left(\beta_0\Pi^2-\beta_1q^{\nu}q_{\nu}+\beta_2\Pi^{\alpha \beta}\Pi_{\alpha \beta}\right)\frac{u^{\mu}}{2T}-\alpha_0\frac{\Pi q^{\mu}}{T}+\alpha_1\frac{\Pi^{\mu \nu}q_{\nu}}{T}.
\end{equation}
Here, the thermodynamic coefficients $\alpha_0$ and $\alpha_1$ control the couplings between viscosity and heat fluxes, while $\beta_0$, $\beta_1$ and $\beta_2$ govern the scalar, vector and tensor contributions to the entropy density respectively.
As we can see, both the component of the entropy current along the velocity field and orthogonal to it are modified with respect to their counterparts in the Navier-Stokes theory.
As in the Navier-Stokes theory, the simplest way to enforce the second law of thermodynamics is to impose linear relations between the dissipative fluxes and the corresponding ``extended thermodynamic forces''.
In this case, however, being the dissipations regarded as dynamic variables, the extended forces contain derivatives of their respective fluxes along the velocity field.
The resulting constitutive equations replacing \eqref{eqrelhydro15a}--\eqref{eqrelhydro15d}, which are in fact the equations of motion of the dissipations, read \cite{Maartens:1996vi}
\begin{subequations}
\begin{align}
\tau_{\Pi}D\Pi+\Pi=&-\xi \nabla_{\mu}u^{\mu}-\frac{1}{2}T\xi \nabla_{\mu}\left(\frac{\tau_{\Pi}}{\xi T}u^{\mu}\right)\Pi+\tau_0{\Delta^{\nu}}_{\mu}\nabla_{\nu}q^{\mu}\label{eqrelhydro17a}\\
\tau_qDq_{\mu}+q_{\mu}=&\kappa \left({\Delta^{\nu}}_{\mu}\nabla_{\nu}T-TA_{\mu}\right)+\frac{1}{2}T^2\kappa \nabla_{\nu}\left(\frac{\tau_q}{\kappa T^2}u^{\nu}\right)q_{\mu}\nonumber \\
&-\tau_0{\Delta^{\lambda}}_{\nu}{\Pi^{\nu}}_{\mu}-\tau_2{\Delta^{\nu}}_{\mu}\nabla_{\nu}\Pi \label{eqrelhydro17b} \\
\tau_{\pi}D\Pi_{\mu \nu}+\Pi_{\mu \nu}=&\eta \sigma_{\mu \nu}-\frac{1}{2}T\eta \nabla_{\lambda}\left(\frac{\tau_{\pi}}{\eta T}u^{\lambda}\right)\Pi_{\mu \nu}\nonumber \\
&+\frac{1}{2}\tau_1\left(\Delta_{\mu \alpha}\Delta_{\nu \beta}+\Delta_{\nu \alpha}\Delta_{\mu \beta}-\frac{2}{3}\Delta_{\mu \nu}\Delta_{\alpha \beta}\right)\nabla^{\alpha}q^{\beta}.\label{eqrelhydro17c}
\end{align}
\end{subequations}
Here, $\tau_{\Pi}$, $\tau_q$, $\tau_{\pi}$, $\tau_0$, $\tau_1$ and $\tau_2$ are new transport coefficients related to $\xi$, $q^{\mu}$ and $\Pi^{\mu \nu}$ by
\begin{equation}
\begin{split}
\tau_{\Pi}\equiv &\xi \beta_0,\qquad
\tau_q\equiv \kappa T^2\beta_1,\qquad
\tau_{\pi}\equiv \eta \beta_2,\\
\tau_0\equiv &\xi \alpha_0,\qquad
\tau_1\equiv \kappa T^2\alpha_1,\qquad
\tau_2\equiv 2\eta \alpha_1.
\end{split}
\end{equation}
It can be shown that they measure the time scales over which the system evolves to a new equilibrium, so they can be regarded as relaxation times.
They are not known \textit{a priori}, and must be worked out from the underlying microscopic theory.
Finally, note that the Navier-Stokes theory is simply recovered in the limit where the new terms are negligible.
The constitutive equations \eqref{eqrelhydro17a}--\eqref{eqrelhydro17c} are sometimes referred to as the \textit{Israel-Stewart equations} of extended irreversible thermodynamics.
They are 9 first-order-in-time partial differential equations which, together with the 5 energy-momentum and charge conservation equations \eqref{eqrelhydro02} and \eqref{eqrelhydro03}, form a set of 14 equations in the 14 unknowns represented by the independent components of the variables $\{ \rho,n,u^{\mu},\Pi,q^{\mu},\Pi^{\mu \nu}\}$.
Clearly, this theory is mathematically more involved than ideal hydrodynamics or the Navier-Stokes theory: more equations in more variables come into play, more input from the underlying microscopic theory is needed and sometimes the physical meaning can be clouded by technicalities.
However, the feature that the dissipations relax with specific time scales instead of instantaneously reach their asymptotic values is the key to obtain a hyperbolic set of equations, thus a causal and stable theory of relativistic hydrodynamics \cite{Israel:1979wp}.
In fact, by repeating a perturbative analysis similar to the one previously discussed, one can indeed show that the resulting equations are hyperbolic in nature.
It is not straightforward to see whether hyperbolicity holds in any condition or not, but we are not going into that.
Over the years, a great variety of predictions of the Israel-Stewart theory have successfully matched with experiments, both in the relativistic and non-relativistic regimes.
Most of the recent interest came in fact from heavy-ion collisions, where the need of a relativistic theory of non-ideal fluids has become crucial to explain the spectra of particles produced in ultrarelativistic nucleus-nucleus collisions.
\section{Summary and outlook}
Before moving to the next Chapter, we briefly wrap up the content of the present one.
Hydrodynamics is an effective classical theory describing the evolution of fluids, systems whose microscopic and macroscopic characteristics are significantly separated.
Although their fundamental nature can be quantum mechanical and possibly complicated, whenever that scale separation occurs their evolution can be effectively described in terms of few classical degrees of freedom.
The simplest instance is that of a perfect or ideal fluid, namely ideal hydrodynamics.
This relies on the assumption of local thermodynamic equilibrium, a configuration wherein the energy-momentum tensor and possible charged currents, which are the fundamental ingredients upon which hydrodynamics is built, although out of thermodynamic equilibrium preserve the same structure they have at thermodynamic equilibrium, promoting the energy density, pressure and velocity field to slowly varying functions of space and time.
Thus, it is possible to define the thermodynamic fields for each fluid element and do thermodynamics.
The key quantity allowing the enforcement of the second law of thermodynamics is the entropy current, a vector field that integrated on some three-dimensional spacelike hypersurface gives the entropy of the system.
It is shown that ideal fluids do not produce any entropy.
On the contrary, entropy is produced by non-perfect or non-ideal fluids, systems where dissipations are present, giving rise to irreversible transport phenomena, whose track is kept in the entropy current, that in turn make the entropy increase.
The most straightforward theory of dissipative hydrodynamics is the Navier-Stokes one: although reproducing the correct non-relativistic limit and fulfilling the laws of thermodynamics, it turns out to be unstable and to break causality.
The ultimate cause for that is the parabolic nature of the equations.
The solution to this problem proposed by Israel and Stewart is the extension of the dissipations from fixed parameters to dynamical variables.
The resulting entropy current captures more terms, which are of second order in the derivatives of the thermodynamic fields and which in turn give rise to hyperbolic equations that restore stability and causality.
The price to pay is that the theory is mathematically more involved and the physical interpretation can be clouded by technicalities, but the predictions of this theory has proven to match many experimental results so far.
As mentioned, the fundamental ingredients of relativistic hydrodynamics are the conserved currents, namely the energy-momentum tensor and some possible charged currents.
In order to enter the hydrodynamic equations, they must be worked out as averages in the underlying microscopic theory.
This could be, for instance, a relativistic kinetic theory, that is a model describing the non-equilibrium dynamics of a fluid whose microscopic degrees of freedom can be regarded as a diluted system of weakly interacting particles.
The relativistic kinetic approach works well when the mean free path of the particles is much larger than their thermal wavelength, but it breaks down for strongly interacting fluids requiring a full quantum field theoretical description in general.
Thus, when the underlying microscopic theory is a Quantum Field Theory, those averages are in fact thermal expectation values calculated either with a generating functional or a density operator depending on the state of the system.
Throughout the rest of this work, we will use a method to calculate thermal expectation values based on a density operator put forward by Zubarev in the late 70's, which will be the subject of the next Chapter.
\chapter*{Abstract}
\afterpage{\blankpage}
\vspace*{5cm}
\begin{Large}
\begin{center}
\textit{To my parents}
\end{center}
\end{Large}
\vfill
\thispagestyle{empty}
\chapter*{Acknowledgements}
\input{aknowledgements}
\tableofcontents
\newpage
\chapter*{Notation}
\input{chapters/notation}
\chapter{Introduction}
\input{chapters/introduction}
\chapter{Relativistic Hydrodynamics}
\label{chapter:relhydro}
\input{chapters/relhydro}
\chapter{Relativistic Quantum Statistical Mechanics}
\label{chapter:zubarev}
\input{chapters/zubarev}
\chapter{Relativistic Quantum Fluid at Equilibrium with Acceleration}
\label{chapter:gteacceleration}
\input{chapters/gteacceleration}
\chapter{Relativistic Quantum Fluid out of Equilibrium with Boost Invariance}
\label{chapter:boost}
\input{chapters/nonequilibrium}
\chapter{Conclusions}
\label{chapter:conclusions}
\input{chapters/conclusions}
|
\section{Copyright}
All papers submitted for publication by AAAI Press must be accompanied by a valid signed copyright form. They must also contain the AAAI copyright notice at the bottom of the first page of the paper. There are no exceptions to these requirements. If you fail to provide us with a signed copyright form or disable the copyright notice, we will be unable to publish your paper. There are \textbf{no exceptions} to this policy. You will find a PDF version of the AAAI copyright form in the AAAI AuthorKit. Please see the specific instructions for your conference for submission details.
\section{Formatting Requirements in Brief}
We need source and PDF files that can be used in a variety of ways and can be output on a variety of devices. The design and appearance of the paper is strictly governed by the aaai style file (aaai22.sty).
\textbf{You must not make any changes to the aaai style file, nor use any commands, packages, style files, or macros within your own paper that alter that design, including, but not limited to spacing, floats, margins, fonts, font size, and appearance.} AAAI imposes requirements on your source and PDF files that must be followed. Most of these requirements are based on our efforts to standardize conference manuscript properties and layout. All papers submitted to AAAI for publication will be recompiled for standardization purposes. Consequently, every paper submission must comply with the following requirements:
\begin{itemize}
\item Your .tex file must compile in PDF\LaTeX{} --- (you may not include .ps or .eps figure files.)
\item All fonts must be embedded in the PDF file --- including your figures.
\item Modifications to the style file, whether directly or via commands in your document may not ever be made, most especially when made in an effort to avoid extra page charges or make your paper fit in a specific number of pages.
\item No type 3 fonts may be used (even in illustrations).
\item You may not alter the spacing above and below captions, figures, headings, and subheadings.
\item You may not alter the font sizes of text elements, footnotes, heading elements, captions, or title information (for references and mathematics, please see the limited exceptions provided herein).
\item You may not alter the line spacing of text.
\item Your title must follow Title Case capitalization rules (not sentence case).
\item Your .tex file must include completed metadata to pass-through to the PDF (see PDFINFO below).
\item \LaTeX{} documents must use the Times or Nimbus font package (you may not use Computer Modern for the text of your paper).
\item No \LaTeX{} 209 documents may be used or submitted.
\item Your source must not require use of fonts for non-Roman alphabets within the text itself. If your paper includes symbols in other languages (such as, but not limited to, Arabic, Chinese, Hebrew, Japanese, Thai, Russian and other Cyrillic languages), you must restrict their use to bit-mapped figures. Fonts that require non-English language support (CID and Identity-H) must be converted to outlines or 300 dpi bitmap or removed from the document (even if they are in a graphics file embedded in the document).
\item Two-column format in AAAI style is required for all papers.
\item The paper size for final submission must be US letter without exception.
\item The source file must exactly match the PDF.
\item The document margins may not be exceeded (no overfull boxes).
\item The number of pages and the file size must be as specified for your event.
\item No document may be password protected.
\item Neither the PDFs nor the source may contain any embedded links or bookmarks (no hyperref or navigator packages).
\item Your source and PDF must not have any page numbers, footers, or headers (no pagestyle commands).
\item Your PDF must be compatible with Acrobat 5 or higher.
\item Your \LaTeX{} source file (excluding references) must consist of a \textbf{single} file (use of the ``input" command is not allowed.
\item Your graphics must be sized appropriately outside of \LaTeX{} (do not use the ``clip" or ``trim'' command) .
\end{itemize}
If you do not follow these requirements, your paper will be returned to you to correct the deficiencies.
\section{What Files to Submit}
You must submit the following items to ensure that your paper is published:
\begin{itemize}
\item A fully-compliant PDF file that includes PDF metadata.
\item Your \LaTeX{} source file submitted as a \textbf{single} .tex file (do not use the ``input" command to include sections of your paper --- every section must be in the single source file). (The only allowable exception is .bib file, which should be included separately).
\item The bibliography (.bib) file(s).
\item Your source must compile on our system, which includes only standard \LaTeX{} 2020 TeXLive support files.
\item Only the graphics files used in compiling paper.
\item The \LaTeX{}-generated files (e.g. .aux, .bbl file, PDF, etc.).
\end{itemize}
Your \LaTeX{} source will be reviewed and recompiled on our system (if it does not compile, your paper will be returned to you. \textbf{Do not submit your source in multiple text files.} Your single \LaTeX{} source file must include all your text, your bibliography (formatted using aaai22.bst), and any custom macros.
Your files should work without any supporting files (other than the program itself) on any computer with a standard \LaTeX{} distribution.
\textbf{Do not send files that are not actually used in the paper.} We don't want you to send us any files not needed for compiling your paper, including, for example, this instructions file, unused graphics files, style files, additional material sent for the purpose of the paper review, and so forth.
\textbf{Do not send supporting files that are not actually used in the paper.} We don't want you to send us any files not needed for compiling your paper, including, for example, this instructions file, unused graphics files, style files, additional material sent for the purpose of the paper review, and so forth.
\textbf{Obsolete style files.} The commands for some common packages (such as some used for algorithms), may have changed. Please be certain that you are not compiling your paper using old or obsolete style files.
\textbf{Final Archive.} Place your PDF and source files in a single archive which should be compressed using .zip. The final file size may not exceed 10 MB.
Name your source file with the last (family) name of the first author, even if that is not you.
\section{Using \LaTeX{} to Format Your Paper}
The latest version of the AAAI style file is available on AAAI's website. Download this file and place it in the \TeX\ search path. Placing it in the same directory as the paper should also work. You must download the latest version of the complete AAAI Author Kit so that you will have the latest instruction set and style file.
\subsection{Document Preamble}
In the \LaTeX{} source for your paper, you \textbf{must} place the following lines as shown in the example in this subsection. This command set-up is for three authors. Add or subtract author and address lines as necessary, and uncomment the portions that apply to you. In most instances, this is all you need to do to format your paper in the Times font. The helvet package will cause Helvetica to be used for sans serif. These files are part of the PSNFSS2e package, which is freely available from many Internet sites (and is often part of a standard installation).
Leave the setcounter for section number depth commented out and set at 0 unless you want to add section numbers to your paper. If you do add section numbers, you must uncomment this line and change the number to 1 (for section numbers), or 2 (for section and subsection numbers). The style file will not work properly with numbering of subsubsections, so do not use a number higher than 2.
\subsubsection{The Following Must Appear in Your Preamble}
\begin{quote}
\begin{scriptsize}\begin{verbatim}
\def\year{2022}\relax
\documentclass[letterpaper]{article}
\usepackage{aaai22}
\usepackage{times}
\usepackage{helvet}
\usepackage{courier}
\usepackage[hyphens]{url}
\usepackage{graphicx}
\urlstyle{rm}
\def\UrlFont{\rm}
\usepackage{graphicx}
\usepackage{natbib}
\usepackage{caption}
\DeclareCaptionStyle{ruled}%
{labelfont=normalfont,labelsep=colon,strut=off}
\frenchspacing
\setlength{\pdfpagewidth}{8.5in}
\setlength{\pdfpageheight}{11in}
\pdfinfo{
/Title (AAAI Press Formatting Instructions for Authors
Using LaTeX -- A Guide)
/Author (AAAI Press Staff, Pater Patel Schneider,
Sunil Issar, J. Scott Penberthy, George Ferguson,
Hans Guesgen, Francisco Cruz, Marc Pujol-Gonzalez)
/TemplateVersion (2022.1)
}
\end{verbatim}\end{scriptsize}
\end{quote}
\subsection{Preparing Your Paper}
After the preamble above, you should prepare your paper as follows:
\begin{quote}
\begin{scriptsize}\begin{verbatim}
\begin{document}
\maketitle
\begin{abstract}
\end{abstract}\end{verbatim}\end{scriptsize}
\end{quote}
\noindent You should then continue with the body of your paper. Your paper must conclude with the references, which should be inserted as follows:
\begin{quote}
\begin{scriptsize}\begin{verbatim}
\section{Our Approach}
We will first describe our approach to \textit{single-region repair} and then present our approach to \textit{multi-region repair} which builds on results obtained from the single-region case.
Given a linear region $\mathcal{A}\xspace$, our overarching approach is to synthesize a \textit{patch network} $h_\mathcal{A}\xspace$ such that $\widehat{f} = f + h_\mathcal{A}\xspace$ and $\widehat{f} \models \Phi$.
The patch network $h_\mathcal{A}\xspace$ is a combination of two sub-networks: a support network $g_\mathcal{A}\xspace$ which behaves like a characteristic function to ensure that $h_\mathcal{A}\xspace$ is almost only active on $\mathcal{A}\xspace$, and an affine patch function network $p_\mathcal{A}\xspace(x) = \boldsymbol{c}x+d$ such that $(f + p_\mathcal{A}\xspace) \models \Phi$ on $\mathcal{A}\xspace$.
\input{Sections/running_example}
\subsection{Support Networks}
Support networks are neural networks with a special structure that can approximate the characteristic function of a convex polytope. They are keys to ensuring localized repairs in our algorithm.
Assume that the linear region we need to repair is $\mathcal{A}\xspace = \{x\:|\:a_ix\leq b_i, i\in I\}$,
where $|I|$ is the number of linear inequalities.
The support network of $\mathcal{A}\xspace$ is defined as:
\begin{eqnarray}\label{eqn: support neural network}
g_\mathcal{A}\xspace(x, \gamma) = \sigma(\sum_{i\in I} g(b_i-a_ix, \gamma)-|I|+1)
\end{eqnarray}
where $g(x, \gamma) = \sigma(\gamma x+1) - \sigma(\gamma x)$ and $\gamma \in \mathbb{R}$ is a parameter of our algorithm that controls the slope of support networks.
\noindent
\textbf{Remark}: For any $x \in \mathcal{A}\xspace$, we have $g_\mathcal{A}\xspace(x, \gamma) = 1$, i.e. the support network is fully activated. For any $x \notin \mathcal{A}\xspace$, if for one of $i\in I$, we have $a_ix-b_i \leq -1/\gamma$, then $g_\mathcal{A}\xspace(x, \gamma) = 0$.
Observe that $g_\mathcal{A}\xspace(x, \gamma)$ cannot be zero if $x$ is very close to $\mathcal{A}\xspace$, otherwise
the resulting function will be discontinuous and violates the criterion of preservation of CPWL.
In Theorem~\ref{thm: The Limited Side Effect of Local Patch Neural Networks}, we will prove that we can still
guarantee limited side effects on the whole input domain outside of $\mathcal{A}\xspace$.
\subsection{Affine Patch Functions}\label{sec: affine patch function}
We consider an affine patch function $p_\mathcal{A}\xspace(x) = \boldsymbol{c}x+d$, where matrix $\boldsymbol{c}$ and vector $d$ are undetermined coefficients.
In a later section, the design of patch network will ensure that on the patch area $\mathcal{A}\xspace$, the repaired network is $f(x)+p_\mathcal{A}\xspace(x)$.
We will first consider
finding appropriate $\boldsymbol{c}$ and $d$ such that
$f(x)+p_\mathcal{A}\xspace(x)$ satisfy the specification on $\mathcal{A}\xspace$.
To satisfy the specification $\Phi$, we need $f(x)+p_\mathcal{A}\xspace(x) \in \Phi_{out}$ for all $x\in \mathcal{A}\xspace$. To obtain a minimal repair, we minimize $\max_{x\in \mathcal{A}\xspace} |p_\mathcal{A}\xspace(x)|$.
Thus, we can formulate the following optimization problem
\begin{eqnarray}\label{eqn: linear patch function}
\begin{cases}
\min_{\boldsymbol{c}, d} \max_{x\in \mathcal{A}\xspace} |p_\mathcal{A}\xspace(x)| = |\boldsymbol{c}x+d|\\
(\boldsymbol{c}, d)\in \{(\boldsymbol{c},d)\:|\:f(x)+\boldsymbol{c}x+d\in \Phi_{out}, \forall x\in \mathcal{A}\xspace\}
\end{cases}
\end{eqnarray}
Notice that this is not a LP since both $\boldsymbol{c}$ and $x$ are variables and we have a $\boldsymbol{c}x$ term in the objective.
In general, one can solve it is by enumerating all the vertices of $\mathcal{A}\xspace$. Suppose that $\{v_s| s=1,2,..., S\}$ is the set of vertices of $\mathcal{A}\xspace$. Since $\Phi_{out}$ is a convex polytope, we have
\begin{eqnarray}
f(x)+p_\mathcal{A}\xspace(x) \in \Phi_{out}\textit{ for all }x \in \mathcal{A}\xspace
\Leftrightarrow f(v_s)+p_\mathcal{A}\xspace(v_s) \in \Phi_{out} \textit{ for }s=1,2,..., S.
\end{eqnarray}
and
\begin{eqnarray}
& \max_{x\in \mathcal{A}\xspace} |\boldsymbol{c}x+d| = \max_{s=1,2,...,S} |cv_s+d|
\end{eqnarray}
Hence, we can solve the following equivalent LP.
\begin{eqnarray}\label{LP: linear patch neural network}
\begin{cases}
\min_{\boldsymbol{c}, d} H\\
H \geq (cv_s+d)_i, H \geq -(cv_s+d)_i, \textit{for } s=1,2,...,S \textit{ and } i =1, 2, ..., m\\
f(v_s)+p_\mathcal{A}\xspace(v_s) \in \Phi_{out} \textit{, for } s=1,2,...,S
\end{cases}
\end{eqnarray}
where $H\in\mathbb{R}$ and will take $\max_{s=1,2,...,S} |cv_s+d|$ when optimal.
\input{Sections/RepairViaRO}
\subsection{Single-Region Repairs}
With a support network $g_\mathcal{A}\xspace$ and an affine patch function $p_\mathcal{A}\xspace$, we can synthesize the final patch network as follows:
\begin{eqnarray}\label{eqn: local patch neural network}
h_\mathcal{A}\xspace(x, \gamma) = \sigma(p_\mathcal{A}\xspace(x)+K\cdot g_\mathcal{A}\xspace(x,\gamma)-K) - \sigma(-p_\mathcal{A}\xspace(x)+K\cdot g_\mathcal{A}\xspace(x,\gamma)-K)
\end{eqnarray}
where $K$ is a vector with every entry is equal to the upper bound of $\{|p_\mathcal{A}\xspace(x)|_{+\infty}|x\in X\}$.
\textbf{Remark:} For $x \in \mathcal{A}\xspace$, $g_\mathcal{A}\xspace(x, \gamma) = 1$, then we have $h_\mathcal{A}\xspace(x, \gamma) = \sigma(p_\mathcal{A}\xspace(x))-\sigma(-p_\mathcal{A}\xspace(x)) = p_\mathcal{A}\xspace(x)$. For $x \notin \mathcal{A}\xspace$, $g_\mathcal{A}\xspace(x, \gamma)$ goes to zero quickly if $\gamma$ is large. When $g_\mathcal{A}\xspace(x, \gamma)=0$, we have $h_\mathcal{A}\xspace(x, \gamma) = \sigma(p_\mathcal{A}\xspace(x)-K) - \sigma(-p_\mathcal{A}\xspace(x)-K) = 0$.
The repaired network $\widehat{f}(x) = f(x) + h_\mathcal{A}\xspace(x, \gamma)$. Since $f$ and $h_\mathcal{A}\xspace$ are both ReLU DNNs, we have $\widehat{f}$ is also a ReLU DNN.
We will give the formal guarantees on correctness in Theorem~\ref{thm: soundness}.
\subsection{Multi-Region Repairs}\label{sec: multi-region repair case}
Suppose there are two linear regions, $\mathcal{A}\xspace_1$ and $\mathcal{A}\xspace_2$, that need to be repaired, and we have generated the affine patch function $p_{\mathcal{A}\xspace_1}(x)$ for $\mathcal{A}\xspace_1$ and $p_{\mathcal{A}\xspace_2}(x)$ for $\mathcal{A}\xspace_2$.
If $\mathcal{A}\xspace_1\cap \mathcal{A}\xspace_2 = \emptyset$, then we can repair $f(x)$ with $\widehat{f}(x) = $ $f(x) + h_{\mathcal{A}\xspace_1}(x, \gamma) + h_{\mathcal{A}\xspace_2}(x, \gamma)$ directly, since $h_{\mathcal{A}\xspace_1}(x, \gamma)$ and $h_{\mathcal{A}\xspace_2}(x, \gamma)$ will not be nonzero at the same time when $\gamma$ is large enough.
However, if $\mathcal{A}\xspace_1\cap \mathcal{A}\xspace_2 \neq \emptyset$, for any $x\in \mathcal{A}\xspace_1\cap \mathcal{A}\xspace_2$, both $h_{\mathcal{A}\xspace_1}(x, \gamma)$ and $h_{\mathcal{A}\xspace_2}(x, \gamma)$ will alter the value of $f$ on $x$,
which will invalidate both repairs and cannot guarantee that the repaired DNN will meet the specification $\Phi$.
To avoid such over-repairs,
our strategy is to first repair $\mathcal{A}\xspace_1\cup \mathcal{A}\xspace_2$ with $p_{\mathcal{A}\xspace_1}(x)$, and then repair $\mathcal{A}\xspace_2$ with $p_{\mathcal{A}\xspace_2}(x)-p_{\mathcal{A}\xspace_1}(x)$. Figure~\ref{fig: multi repair} provides an illustration of a three-region case.
\begin{figure*}
\centering
\begin{subfigure}{0.223\textwidth}
\resizebox{1\textwidth}{!}{
\begin{tikzpicture}
\draw[black,domain=-1.5:1] plot(\x,{0});
\draw[black,domain=-1:1] plot(\x,{-\x});
\draw[black,domain=-1/2:1/2] plot(\x,{2*\x});
\node[left, blue, scale=0.7] at (-0.2, -0.4) {$\mathcal{A}\xspace_1: f$};
\node[left, blue, scale=0.7] at (-0.4, 0.3) {$\mathcal{A}\xspace_2: f$};
\node[left, blue, scale=0.7] at (0.2, 0.7) {$\mathcal{A}\xspace_3: f$};
\node[left, white, scale=0.7] at (0.5, 1.2) {$\mathcal{A}\xspace_3: f+p_{\mathcal{A}\xspace_1}$};
\draw [fill=red, opacity=0.2] (0,0) -- (1/2,1) -- (-1,1) -- cycle;
\draw [fill=red, opacity=0.2] (0,0) -- (-1,1) -- (-1.5,0) -- cycle;
\draw [fill=red, opacity=0.2] (0,0) -- (-1.5,0) -- (-1/2,-1) -- cycle;
\node[left, black] at (1.8, 0) {$\Rightarrow$};
\end{tikzpicture}}
\end{subfigure}
\begin{subfigure}{0.253\textwidth}
\resizebox{1\textwidth}{!}{
\begin{tikzpicture}
\draw[black,domain=-1.5:1] plot(\x,{0});
\draw[black,domain=-1:1] plot(\x,{-\x});
\draw[black,domain=-1/2:1/2] plot(\x,{2*\x});
\node[left, blue, scale=0.7] at (-0.2, -0.4) {$\mathcal{A}\xspace_1: f+p_{\mathcal{A}\xspace_1}$};
\node[left, blue, scale=0.7] at (-0.4, 0.3) {$\mathcal{A}\xspace_2: f+p_{\mathcal{A}\xspace_1}$};
\node[left, blue, scale=0.7] at (0.5, 1.2) {$\mathcal{A}\xspace_3: f+p_{\mathcal{A}\xspace_1}$};
\draw [fill=green, opacity=0.2] (0,0) -- (1/2,1) -- (-1,1) -- cycle;
\draw [fill=green, opacity=0.2] (0,0) -- (-1,1) -- (-1.5,0) -- cycle;
\draw [fill=green, opacity=0.2] (0,0) -- (-1.5,0) -- (-1/2,-1) -- cycle;
\node[left, black] at (1.8, 0) {$\Rightarrow$};
\end{tikzpicture}}
\end{subfigure}
\begin{subfigure}{0.253\textwidth}
\resizebox{1\textwidth}{!}{
\begin{tikzpicture}
\draw[black,domain=-1.5:1] plot(\x,{0});
\draw[black,domain=-1:1] plot(\x,{-\x});
\draw[black,domain=-1/2:1/2] plot(\x,{2*\x});
\node[left, blue, scale=0.7] at (-0.2, -0.4) {$\mathcal{A}\xspace_1: f+p_{\mathcal{A}\xspace_1}$};
\node[left, blue, scale=0.7] at (-0.4, 0.3) {$\mathcal{A}\xspace_2: f+p_{\mathcal{A}\xspace_2}$};
\node[left, blue, scale=0.7] at (0.5, 1.2) {$\mathcal{A}\xspace_3: f+p_{\mathcal{A}\xspace_2}$};
\draw [fill=cyan, opacity=0.2] (0,0) -- (1/2,1) -- (-1,1) -- cycle;
\draw [fill=cyan, opacity=0.2] (0,0) -- (-1,1) -- (-1.5,0) -- cycle;
\draw [fill=green, opacity=0.2] (0,0) -- (-1.5,0) -- (-1/2,-1) -- cycle;
\node[left, black] at (1.8, 0) {$\Rightarrow$};
\end{tikzpicture}}
\end{subfigure}
\begin{subfigure}{0.205\textwidth}
\resizebox{1\textwidth}{!}{
\begin{tikzpicture}
\draw[black,domain=-1.5:1]plot(\x,{0});
\draw[black,domain=-1:1] plot(\x,{-\x});
\draw[black,domain=-1/2:1/2] plot(\x,{2*\x});
\node[left, blue, scale=0.7] at (-0.2, -0.4) {$\mathcal{A}\xspace_1: f+p_{\mathcal{A}\xspace_1}$};
\node[left, blue, scale=0.7] at (-0.4, 0.3) {$\mathcal{A}\xspace_2: f+p_{\mathcal{A}\xspace_2}$};
\node[left, blue, scale=0.7] at (0.5, 1.2) {$\mathcal{A}\xspace_3: f+p_{\mathcal{A}\xspace_3}$};
\draw [fill=yellow, opacity=0.2] (0,0) -- (1/2,1) -- (-1,1) -- cycle;
\draw [fill=cyan, opacity=0.2] (0,0) -- (-1,1) -- (-1.5,0) -- cycle;
\draw [fill=green, opacity=0.2] (0,0) -- (-1.5,0) -- (-1/2,-1) -- cycle;
\end{tikzpicture}}
\end{subfigure}
\caption{An illustration of multi-region repair with three different repair regions. Left: the original DNN; Middle Left: repair $\mathcal{A}\xspace_1\cup\mathcal{A}\xspace_2\cup\mathcal{A}\xspace_3$ with $p_{\mathcal{A}\xspace_1}$; Middle Right: repair $\mathcal{A}\xspace_2\cup\mathcal{A}\xspace_3$ with $p_{\mathcal{A}\xspace_2}-p_{\mathcal{A}\xspace_1}$; Right: repair $\mathcal{A}\xspace_3$ with $p_{\mathcal{A}\xspace_3}-p_{\mathcal{A}\xspace_2}$}\label{fig: multi repair}
\vspace{-2mm}
\end{figure*}
In general, for multi-region repair, we note $\{\mathcal{A}\xspace_l\}_{l=1,2,\dots,L}$ are all the buggy linear regions. Then we compute the support network $g_{\mathcal{A}\xspace_l}(x, \gamma)$ and affine patch function $p_{\mathcal{A}\xspace_l}(x)$ for each $\mathcal{A}\xspace_l$. Note that this computation can be done in parallel.
Once we have $g_{\mathcal{A}\xspace_l}(x, \gamma)$ and $p_{\mathcal{A}\xspace_l}(x)$, we can ``stitch" multiple local patches into a final patch as follows.
\begin{eqnarray}\label{eqn: multi-region repair}
h(x,\gamma) =\sum_{l}[\sigma(p_{\mathcal{A}\xspace_l}(x) - p_{\mathcal{A}\xspace_{l-1}}(x)
+\max_{j\geq l}\{g_{\mathcal{A}\xspace_j}(x, \gamma)\}K_l-K_l)\nonumber\\ - \sigma(-p_{\mathcal{A}\xspace_l}(x)+ p_{\mathcal{A}\xspace_{l-1}}(x)+\max_{j\geq l}\{g_{\mathcal{A}\xspace_j}(x, \gamma)\}K_l-K_l)]
\end{eqnarray}
where $K_l$ is the upper bound of $\{|p_{\mathcal{A}\xspace_l}(x) - p_{\mathcal{A}\xspace_{l-1}}(x)|_{\infty}|x\in X\}$ and $p_{\mathcal{A}\xspace_0}(x) = 0$.
\noindent
\textbf{Remark}: $\max_{j\geq l}\{g_{\mathcal{A}\xspace_j}(x, \gamma)\}$ is a support function for $\cup_{j\geq l} \mathcal{A}\xspace_j$ and its value is $1$ for any $x\in \cup_{j\geq l} \mathcal{A}\xspace_j$.
\subsection{Feature-Space Repairs}
\label{sec:feature-space}
In general, when repairing a large DNN with a high input dimension, the number of linear constraints for one patch area $\mathcal{A}\xspace$ will be huge and pose a challenge to solving the resulting LP. One advantage of our approach, which can be used to mitigate this problem, is that it allows for point-wise and area repairs in the feature space in a principled manner, i.e. constructing a patch network starting from an intermediate layer. This approach still preserves soundness and completeness, and is fundamentally different from arbitrarily picking a single layer for repair.
Specifically, for an $R$-layer DNN $f$, we split $f$ into two networks $f_1$ and $f_2$ according to a hidden layer, say the $j_{th}$ hidden layer, where $f_1$ is the function of the first $j$ layers, $f_2$ is the function of the last $R-j$ layers, and $f = f_2\circ f_1$. The output space of $f_1$ is a \textit{feature space}. For any buggy input $\widetilde{x}$, $f_1(\widetilde{x})$ is the corresponding buggy feature.
The point-wise repair problem in a feature space is to repair the behavior of $f_2$ on buggy features $\{f_1(\widetilde{x}_1), \dots, f_1(\widetilde{x}_L)\}$. Note that this will automatically repair the behavior of $f$ on buggy points $\{\widetilde{x}_1, \dots, \widetilde{x}_L\}$.
Repairing in a feature space has the benefit of making the repair process more computation-friendly and reducing the parameter overhead of the additional networks, and has the potential to generalize the repair to undetected buggy inputs with similar features.
It loses the locality guarantee in the input space but still preserves locality in the feature space.
\section{Background}
\begin{table*}
\centering
\scalebox{0.9}{
\begin{tabular}{c|ccccc}
& \texttt{REASSURE}\xspace & Retrain & MDNN & Editable Fine-Tuning & PRDNN \\
\hline
Preservation of CPWL & Yes & Yes & Yes & Yes & No\\
Soundness & Yes & No & Yes & No & Yes \\
Completeness & Yes & No & No & No & No \\
Area Repair & Yes & No & No & No & Yes \\
Minimal Change & Yes (Function Space) & No & Yes (Weight Space) & No & Yes (Weight Space) \\
Localized Change & Yes & No & No & No & No \\
Limited Side Effect & Yes & No & No & No & No
\end{tabular}
}
\caption{Comparing \texttt{REASSURE}\xspace with representative related works in terms of theoretical guarantees.
CPWL stands for continuous piecewise linearity.
Area repair means repairing all the (infinitely many) points inside an area.
Limited side effect means the repair can limit potential adverse effects on other inputs.
MDNN is the verification-based approach from \cite{goldberger2020minimal}.
PRDNN is the Decoupled DNN approach from \cite{sotoudeh2021provable}.
\texttt{REASSURE}\xspace is the only method that can provide all the guarantees.
}
\label{tab: requirements}
\vspace{-2mm}
\end{table*}
\subsection{Deep Neural Networks}
An $R$-layer feed-forward DNN $f = \kappa_R \circ \sigma \circ \kappa_{R-1} \circ... \circ \sigma \circ \kappa_1: X\to Y$ is a composition of linear functions $\kappa_r, r=1, 2, ..., R$ and activation function $\sigma$, where $X\subseteq \mathbb{R}^m$ is a bounded input domain and $Y\subseteq\mathbb{R}^n$ is the output domain.
Weights and biases of linear function $\{\kappa_r\}_{r=1, 2,...,R}$ are parameters of the DNN.
We call the first $R - 1$ layers hidden layers and the $R$-th layer the output layer.
We use $z^i_j$ to denote the $i$-th neuron (before activation) in the $j$-th hidden layer.
In this paper, we focus on ReLU DNNs, i.e. DNNs that use only the ReLU activation functions.
It is known that an $\mathbb{R}^m \rightarrow \mathbb{R}$ function is representable by a ReLU DNN \textit{if and only if} it is a continuous piecewise linear (CPWL) function~\cite{arora2016relu}.
The ReLU function is defined as $\sigma(x) = \text{max}(x, 0)$. We say that $\sigma(x)$ is activated if $\sigma(x) = x$.
\subsection{Linear Regions}
A linear region $\mathcal{A}\xspace$ is the set of inputs
that correspond to the same activation pattern in a ReLU DNN $f$~\cite{serra2017linear-region}.
Geometrically, this corresponds to a convex polytope, which is an intersection of half spaces, in the input space $X$ on which $f$ is linear. We use $f|_\mathcal{A}\xspace$ to denote the part of $f$ on $\mathcal{A}\xspace$.
\subsection{Correctness Specification}
A correctness specification $\Phi = (\Phi_{in}, \Phi_{out})$ is a tuple of two polytopes, where
$\Phi_{in}$ is the union of some linear regions and $\Phi_{out}$ is a convex polytope. A DNN $f$ is said to meet a specification $\Phi = (\Phi_{in}, \Phi_{out})$, denoted as $f \models \Phi$, if and only if $\forall x \in \Phi_{in}, f(x)\in \Phi_{out}$.
\begin{example}\label{exp: classification problem}
For a classification problem, we can formally write the specification that ``the prediction of any point in an area $\mathcal{A}\xspace$ is class $k$" as $\Phi = (\Phi_{in}, \Phi_{out})$, where $\Phi_{in} = \mathcal{A}\xspace$ and $\Phi_{out} = \{y\in \mathbb{R}^n\:|\: y_k\geq y_i, \forall i\neq k\}$\footnote{
Note that here $y$ is the output of the layer right before the softmax layer in a classification network.}.
\end{example}
\subsection{Problem Definition}
In this paper, we consider the following two repair problems.
\begin{definition}[Area repair]\label{def: area repair}
Given a correctness specification $\Phi = (\Phi_{in}, \Phi_{out})$ and a ReLU DNN $f \not\models \Phi$, the area repair problem is to find a modified ReLU DNN $\widehat{f}$ such that $\widehat{f} \models \Phi$.
\end{definition}
Note that we do not require $\widehat{f}$ to have the same structure or parameters as $f$ in this definition.
If $\Phi_{in}$ contains a single (buggy) linear region, we refer to this as \textit{single-region repair}. If $\Phi_{in}$ contains multiple (buggy) linear regions, we refer to it as \textit{multi-region repair}.
\begin{definition}[Point-wise repair]\label{def: point-wsie repair}
Given a set of buggy inputs $\{\widetilde{x}_1,\ldots,\widetilde{x}_L\} \subset \Phi_{in}$ with their corresponding correct outputs $\{y_1,\ldots,y_L\}$ and a ReLU DNN $f$, the point-wise repair problem is to
find a modified ReLU DNN $\widehat{f}$ such that $\forall i, \widehat{f}(\widetilde{x}_i) = y_i$.
\end{definition}
We call the minimal variants of area repair and point-wise repair \textit{minimal area repair} and \textit{minimal point-wise repair} respectively. Minimality here is defined with respect to the maximum distance between $f$ and $\widehat{f}$ over
the whole input domain $X$.
A point-wise repair can be generalized to an area repair through the following result.
\subsection{From Buggy Inputs to Buggy Linear Regions}\label{sec: point-wise}
The linear region where an input $x$ resides can be computed as follows.
\begin{lemma}\cite{lee2019towards}
Consider a ReLU DNN $f$ and an input $x\in X$. For every neuron $z^i_j$, it induces a feasible set
\begin{eqnarray}\label{equ: buggy point to linregion}
\mathcal{A}\xspace^i_j(x) =
\begin{cases}
\{\bar{x}\in X|(\triangledown_x z^i_j)^T \bar{x} + z^i_j - (\triangledown_x z^i_j)^T x \geq 0\} \textit{ \qquad if $z^i_j\geq 0$}\\
\{\bar{x}\in X|(\triangledown_x z^i_j)^T \bar{x} + z^i_j - (\triangledown_x z^i_j)^T x \leq 0\} \textit{ \qquad if $z^i_j < 0$}
\end{cases}
\end{eqnarray}
The set $\mathcal{A}\xspace(x) = \cap_{i,j} \mathcal{A}\xspace^i_j(x)$ is the linear region that includes $x$. Note that $\mathcal{A}\xspace(x)$ is essentially the H-representation of the corresponding convex polytope.
\end{lemma}
\subsection{Repair Desiderata}
We argue that an effective repair algorithm for ReLU DNN should satisfy the following criteria.
\quad\textbf{Preservation of CPWL}:
Given that the original network $f$ models a CPWL function, the repaired network $\widehat{f}$ should still model a CPWL function.
\quad\textbf{Soundness}:
A sound repair should completely remove the known buggy behaviors, i.e. it is a solution to
the point-wise repair problem defined in Definition~\ref{def: point-wsie repair}.
\quad\textbf{Completeness}:
Ideally, the algorithm should always be able find a repair for any given buggy input if it exists.
\quad\textbf{Generalization}:
If there exists another buggy input $\widetilde{x}'$ in the neighborhood of $\widetilde{x}$ (e.g. the same linear region), then the repair should also fix it.
For example, suppose we have an $\widetilde{x}$ that violates a specification which requires the output to be within some range. It is almost guaranteed
that there exists another (and infinitely many) $\widetilde{x}'$ in the same linear region that also violates the specification.
\quad\textbf{Locality}:
We argue that a good repair should only induce a localized change to $f$ in the function space. For example, in the context of ReLU DNN, if a linear region $\mathcal{B}$ does not border the repair region $\mathcal{A}\xspace$, i.e.
$\mathcal{B} \cap \mathcal{A}\xspace = \emptyset$,
then $\widehat{f}|_\mathcal{B}(x) = f|_\mathcal{B}(x)$.
\quad\textbf{Minimality}:
Some notion of distance between $f$ and $\widehat{f}$ such as $\max|f - \widehat{f}|$ should be minimized. Note that this is a significant departure from existing methods that focus on minimizing the change in weights which has no guarantee on the amount of change in the function space.
\quad\textbf{Limited side effect}:
Repairing a buggy point should not adversely affect points that were originally correct.
For example, repairing a buggy input $\widetilde{x}$
in region $\mathcal{A}\xspace$ should not change another region from correct to incorrect.
Formally, for any linear region $\mathcal{C}$ who is a neighbor of $\mathcal{A}\xspace$, i.e. $\mathcal{C} \cap \mathcal{A}\xspace \neq \emptyset$, if $f|_\mathcal{C} \models \Phi$, then $\widehat{f}|_\mathcal{C} \models \Phi$.
\quad\textbf{Efficiency}: The repair algorithm should terminate in polynomial time with respect to the size of the neural network and the number of buggy inputs.
\section{Discussion}
\subsection{Parameter Overhead}
$\texttt{REASSURE}\xspace$ introduces an additional network, patch network, and as a result adds new parameters to the original network.
The average number of new parameters that $\texttt{REASSURE}\xspace$ introduces per repair region is in $O(m|I|)$, where $m$ is the dimension of the target DNN's input space and $|I|$ is the number of linear constraints for the H-representation of $\mathcal{A}\xspace$.
We can remove redundant constraints in this representation in polynomial time to make the additional network smaller, e.g. iteratively using an LP to check if a constraint is redundant. For the area repair experiment on HCAS, the average number of constraints for one linear region is $3.28$ (which is only $3\%$ of the original $125$ constraints after removing the redundant constraints) and the average number of new parameters that $\texttt{REASSURE}\xspace$ introduces per repair region is $66$. As a comparison, the number of parameters in the original network is around $3000$ and PRDNN doubles the number of parameters (as a result of the Decoupled DNN construction) regardless of the number of point-wise or area repairs.
We further note that the removal of redundant constraints can be done offline as a post-repair optimization step.
Another way to cope with the additional space overhead is to leverage feature-space repairs.
As described in Section~\ref{sec:feature-space}, feature-space repairs allow us to construct the patch network starting from an intermediate layer.
In addition, for most DNN structures, the dimension of an intermediate layer is typically smaller than the dimension of the input space.
As a result, the number of additional parameters per repair region will be smaller.
For the point-wise repair experiment on ImageNet, our feature-space repair adds 500k new parameters on average, which is only $0.1\%$ of the additional parameters introduced if we were to construct the patch network starting from the input layer.
With feature-space repairs, we are trading-off repair specificity, i.e. how localized the repair is in the input space, with parameter overhead. However, when the number of points or linear regions to repair becomes large, it may make sense to perform repairs in the feature space anyway for better generalization.
We leave the in-depth investigation of feature-space repairs to future work.
\subsection{Applying \texorpdfstring{$\texttt{REASSURE}\xspace$}{Lg} To General CPWL Networks}
Recall the result that an $\mathbb{R}^m \rightarrow \mathbb{R}$ function is representable by a ReLU DNN \textit{if and only if} it is a continuous piecewise linear (CPWL) function~\cite{arora2016relu}.
We use convolutional neural networks as an example to show how $\texttt{REASSURE}\xspace$ can be applied to more general CPWL networks.
Convolutional neural networks (CNNs) are neural networks with convolution layers and maxpooling layers.
For simplicity, we assume the CNNs also use ReLU activation functions (but in general other CPWL activation functions will also work).
The convolutional layers can be viewed as special linear layers. The maxpooling layers can be converted to linear operations with ReLU activation functions as follows.
\begin{eqnarray*}
\max(x_1, x_2, ..., x_n) = \max(x_1, \max(x_2, x_3,..., x_n))\\
\max(x_i, x_j) = \max(x_i-x_j, 0) + x_j = \sigma(x_i-x_j) + x_j
\end{eqnarray*}
where $\sigma$ is the ReLU activation function.
Thus, $\texttt{REASSURE}\xspace$ can be used to repair CNNs as well.
\section{Appendix}\label{apx}
\section{Conclusion}
We have presented a novel approach for repairing ReLU DNNs with strong theoretical guarantees. Across a set of benchmarks, our approach significantly outperforms existing methods in terms of efficacy, locality, and limiting negative side effects. Future directions include further investigation on feature-space repairs and identifying a lower-bound for $\gamma$.
\section{Theoretical Guarantees}
In this section, we present the theoretical guarantees that \texttt{REASSURE}\xspace provide, and point the readers to proofs of the theorems in the Appendix.
\subsection{Soundness \& Completeness}
\begin{theorem}[Soundness]\label{thm: soundness}
The repaired DNN $\widehat{f}$ returned by $\texttt{REASSURE}\xspace$ is guaranteed to satisfy the specification $\Phi$.
\end{theorem}
\begin{theorem}[Completeness]\label{thm: completeness}
$\texttt{REASSURE}\xspace$ can always find a solution to the minimal point-wise repair or the minimal area repair problem.
\end{theorem}
\subsection{Limited Side Effect of Patch Networks}
For any $\mathcal{A}\xspace$, the support network ensures that the patch network goes to zero quickly when $x$ is away from $\mathcal{A}\xspace$. However, it still makes a small change on the neighbors of $\mathcal{A}\xspace$. The following theorem shows that for a big enough $\gamma$, the patch network would not change a correct region into incorrect.
\begin{theorem}[Limited Side Effect]\label{thm: The Limited Side Effect of Local Patch Neural Networks}
Given a correctness property $\Phi = (\Phi_{in}, \Phi_{out})$, a patch region $\mathcal{A}\xspace$ and the corresponding patch network $h(x, \gamma)$, there exists a positive number $\Gamma$ such that for any $\gamma\geq \Gamma$, we have
\begin{enumerate}
\item for any linear region $\mathcal{B}\xspace$, if $\mathcal{B}\xspace\cap \mathcal{A}\xspace = \emptyset$, then $\widehat{f}(x, \gamma) = f(x)$;
\item for any linear region $\mathcal{C}\xspace$ who is a neighbor of $\mathcal{A}\xspace$ ($\mathcal{C}\xspace\cap \mathcal{A}\xspace \neq \emptyset$), if $f|_\mathcal{C} \models \Phi$, then $\widehat{f}_\mathcal{C}(x, \gamma) \models \Phi$.
\end{enumerate}
\end{theorem}
\begin{corollary}[Incremental Repair]\label{cor}
For multiple-region repair, the patch for a new region $\mathcal{A}\xspace'$ would not cause a previous patched region $\mathcal{A}\xspace$ to become incorrect.
\end{corollary}
\subsection{Minimum Repair}
\begin{theorem}[Minimum Repair]\label{thm: minimum repair}
For any ReLU DNN $\tilde{f}$, which is linear on a patch region $\mathcal{A}\xspace$ and satisfies the specification $\Phi$, there exists a positive number $\Gamma$, such that for all $\gamma\geq \Gamma$,
\begin{eqnarray}
\max_{x\in X}|\tilde{f}(x)-f(x)|\geq \max_{x\in X}|h_\mathcal{A}\xspace(x, \gamma)|.
\end{eqnarray}
\end{theorem}
\subsection{Efficiency}
\begin{theorem}[Polynomial-Time Efficiency]
\texttt{REASSURE}\xspace terminates in polynomial-time in the size of the neural network and the number of buggy linear regions.
\end{theorem}
\begin{algorithm}[H]
\caption{\texttt{REASSURE}\xspace}
\label{alg: main}
\textbf{Input}: A specification $\Phi = (\Phi_{in}, \Phi_{out})$, a ReLU DNN $f$ and a set of buggy points $\{\widetilde{x}_1,\ldots,\widetilde{x}_L\} \subset \Phi_{in}$.
\textbf{Output}: A repaired ReLU DNN $\widehat{f}$.
\begin{algorithmic}[1]
\FOR{$l = 1$ to $L$}
\STATE Generate the patch area $\mathcal{A}\xspace_l$ from buggy point $\widetilde{x}_l$ according to Equation~(\ref{equ: buggy point to linregion});
\STATE Generate a support network $g_\mathcal{A}\xspace$ according to Equation~(\ref{eqn: support neural network});
\STATE Solve the linear programming problem (\ref{LP: linear patch neural network}) to find the optimal affine patch network $p_\mathcal{A}\xspace$.
\ENDFOR
\STATE Combine all support networks $g_{\mathcal{A}\xspace_l}$ and the corresponding patch networks $p_{\mathcal{A}\xspace_l}$ to get the overall patch network $h$ according to Equation~(\ref{eqn: multi-region repair}).
\STATE \textbf{return} $\widehat{f} = f + h$
\end{algorithmic}
\end{algorithm}
\subsection{Proofs of Theorems}
We prove Theorem \ref{apxthm: soundness} after Corollary \ref{apxcor: Incremental Repair}, since the proof of Theorem \ref{apxthm: soundness} uses the result of Corollary \ref{apxcor: Incremental Repair}.
\begin{lemma}\label{lemma: soundness for single region}
The repaired DNN $\widehat{f}$ returned by $\texttt{REASSURE}\xspace$ is guaranteed to satisfy the specification $\Phi$ on patch area $\mathcal{A}\xspace$ in single-region repair case.
\end{lemma}
\begin{proof}
By the definition of $p_\mathcal{A}\xspace$, we have $f(x)+p_\mathcal{A}\xspace(x)\in \Phi_{out}$ for all $x\in \mathcal{A}\xspace$.
For any $x \in \mathcal{A}\xspace$, we have $g_\mathcal{A}\xspace(x, \gamma) = 1$ and $h_\mathcal{A}\xspace(x, \gamma) = \sigma(p_\mathcal{A}\xspace(x))-\sigma(-p_\mathcal{A}\xspace(x)) = p_\mathcal{A}\xspace(x)$. Therefore,
\begin{eqnarray}
\widehat{f}(x) = f(x) + h_\mathcal{A}\xspace(x, \gamma) = f(x) + p_\mathcal{A}\xspace(x) \in \Phi_{out}
\end{eqnarray}
Thus, the patched neural network $\widehat{f}$ meets the specification $\Phi$ on $\mathcal{A}\xspace$.
\end{proof}
\begin{customthm}{2}[Completeness]\label{apxthm: completeness}
$\texttt{REASSURE}\xspace$ can always find a solution to the minimal point-wise repair or the minimal area repair problem.
\end{customthm}
\begin{proof}
For every patch area $\mathcal{A}\xspace$, we can always find a support network $g_\mathcal{A}\xspace$. For any $\Phi_{out}$ and $\mathcal{A}\xspace$, there exists an affine function $p_\mathcal{A}\xspace$ such that $p_\mathcal{A}\xspace(x) \in \Phi_{out}, \forall x\in \mathcal{A}\xspace$. Therefore, the LP (\ref{LP: linear patch neural network}) is always feasible and \texttt{REASSURE}\xspace can find an affine patch function $p_\mathcal{A}\xspace$.
Once we have $g_\mathcal{A}\xspace$ and $p_\mathcal{A}\xspace$ for patch area $\mathcal{A}\xspace$, \texttt{REASSURE}\xspace returns a patch network either by Equation (\ref{eqn: local patch neural network}) or by Equation (\ref{eqn: multi-region repair}).
\end{proof}
\begin{customthm}{3}[Limited Side Effect]\label{apxthm: Limited Side Effect}
Given a correctness property $\Phi = (\Phi_{in}, \Phi_{out})$, a patch region $\mathcal{A}\xspace$ and the corresponding patch network $h(x, \gamma)$, there exists a positive number $\Gamma$ such that for any $\gamma\geq \Gamma$, we have
\begin{enumerate}
\item for any linear region $\mathcal{B}\xspace$, if $\mathcal{B}\xspace\cap \mathcal{A}\xspace = \emptyset$, then $\widehat{f}(x, \gamma) = f(x)$;
\item for any linear region $\mathcal{C}\xspace$ who is a neighbor of $\mathcal{A}\xspace$ ($\mathcal{C}\xspace\cap \mathcal{A}\xspace \neq \emptyset$), if $f|_\mathcal{C} \models \Phi$, then $\widehat{f}_\mathcal{C}(x, \gamma) \models \Phi$.
\end{enumerate}
\end{customthm}
\begin{proof}
Since a multi-region repair is a composition of multiple singe-region repairs according to Equation~(\ref{eqn: multi-region repair}), we can prove the limited side effect of a multi-region repair by proving the limited side effect of its constituent singe-region repairs. Below, we prove the limited side effect of a singe-region repair.
Consider patch area $\mathcal{A}\xspace = \{x\:|\:a_ix\leq b_i, i\in I\}$ and $\mathcal{A}\xspace_{> 0}(\gamma) = \{x\:|\:h(x, \gamma) > 0\}$.
1. Since the number of neighbors for $\mathcal{A}\xspace$ are finite, we can take a big enough $\gamma$, such that for any $\mathcal{B}\xspace$, if $\mathcal{B}\xspace\cap \mathcal{A}\xspace = \emptyset$, $\mathcal{B}\xspace\cap \mathcal{A}\xspace_{> 0}(\gamma) = \emptyset$. Thus, we have $\widehat{f}(x, \gamma) = f(x)$ on $\mathcal{B}\xspace$.
2. For any linear region $\mathcal{C}$ who is a neighbor of $\mathcal{A}\xspace$, i.e. $\mathcal{C} \neq \mathcal{A}\xspace$ and $\mathcal{C} \cap \mathcal{A}\xspace \neq \emptyset$, $\widehat{f}$ is no longer a linear function on $\mathcal{C}\xspace$, since there are some hyperplanes introduced by our repair that will divide $\mathcal{C}\xspace$ into multiple linear regions.
Specifically, those hyperplanes are $\{x\:|\:\gamma(a_ix-b_i) + 1 = 0\}$ for $i\in I$, $\{x\:|\:\sum_{i\in I} g(a_ix-b_i, \gamma)-|I|+1 = 0\}$, $\{x\:|\:p(x)+K\cdot g_\mathcal{A}\xspace(x,\gamma)-K = 0\}$ and $\{x\:|\:-p(x)+K\cdot g_\mathcal{A}\xspace(x,\gamma)-K = 0\}$.
For any point $x$ in those hyperplanes, it will fall into one of the following four cases.
\begin{enumerate}[itemindent=1em]
\item[(a)] $x\in \{x\:|\:\gamma(a_ix-b_i) + 1 = 0\}$ for some $i\in I$, then $g_\mathcal{A}\xspace(x, \gamma) = 0$, $h(x, \gamma) = 0$ and $\widehat{f}(x) \in \Phi_{out}$;
\item[(b)] $x\in \{x\:|\:\sum_{i\in I} g(a_ix-b_i, \gamma)-|I|+1 = 0\}$, then $g_\mathcal{A}\xspace(x, \gamma) = 0$, $h(x, \gamma) = 0$ and $\widehat{f}(x) \in \Phi_{out}$;
\item[(c)] $x\in \{x\:|\:p(x)+K\cdot g_\mathcal{A}\xspace(x,\gamma)-K = 0\}$, then $p(x) = K - K\cdot g_\mathcal{A}\xspace(x,\gamma)\geq 0$, $-p(x)+K\cdot g_\mathcal{A}\xspace(x,\gamma)-K \leq 0$, $h(x, \gamma) = 0$ and $\widehat{f}(x) \in \Phi_{out}$;
\item[(d)] $x\in \{x\:|\:-p(x)+K\cdot g_\mathcal{A}\xspace(x,\gamma)-K\}$, then $p(x) = K\cdot g_\mathcal{A}\xspace(x,\gamma)-K\leq 0$, $p(x)+K\cdot g_\mathcal{A}\xspace(x,\gamma)-K \leq 0$, $h(x, \gamma) = 0$ and $\widehat{f}(x) \in \Phi_{out}$;
\end{enumerate}
By the above analysis, we have $\widehat{f}(x) \in \Phi_{out}$ for the boundary of the new linear regions. Since $\widehat{f}$ is linear on the new linear regions and $\Phi_{out}$ is convex, $\widehat{f}(x) \in \Phi_{out}$ for any $x \in \mathcal{C}\xspace$.
\end{proof}
\noindent
\textbf{Remark}: By Theorem \ref{apxthm: Limited Side Effect}, we have that a patch would not change a correct linear region to an incorrect one.
\begin{corollary}[Incremental Repair]\label{apxcor: Incremental Repair}
For multiple-region repair, the patch for a new region $\mathcal{A}\xspace'$ would not cause a previous patched region $\mathcal{A}\xspace$ to become incorrect.
\end{corollary}
\begin{proof}
After applying the patch to linear region $\mathcal{A}\xspace$, we have that the resulting network is correct on $\mathcal{A}\xspace$. When applying a new patch to another linear region $\mathcal{A}\xspace'$, by Theorem \ref{apxthm: Limited Side Effect}, the new patch would not make a correct linear region $\mathcal{A}\xspace$ incorrect.
\end{proof}
\begin{customthm}{1}[Soundness]\label{apxthm: soundness}
The repaired DNN $\widehat{f}$ returned by $\texttt{REASSURE}\xspace$ is guaranteed to satisfy the specification $\Phi$.
\end{customthm}
\begin{proof}
The proof has two parts:
\begin{enumerate}
\item to show that $\widehat{f}$ satisfies the specification $\Phi$ on $\mathcal{A}\xspace$, and
\item to show that $\widehat{f}$ satisfies the specification $\Phi$ outside of $\mathcal{A}\xspace$.
\end{enumerate}
Part 1:
Lemma (\ref{lemma: soundness for single region}) shows $\widehat{f}$ satisfy the specification $\Phi$ for single-region repair on $\mathcal{A}\xspace$.
For the multi-region case, consider a set of buggy linear regions $\cup_{1\leq l \leq I}\mathcal{A}\xspace_l$ with
the corresponding support neural network $g_{\mathcal{A}\xspace_l}$ and affine patch function $p_{\mathcal{A}\xspace_l}$ for each $\mathcal{A}\xspace_l$. For the multi-region repair construction in Equation~(\ref{eqn: multi-region repair}), we refer to $\sigma(p_{\mathcal{A}\xspace_j} - p_{\mathcal{A}\xspace_{j-1}} +\max_{k\geq j}\{g_{\mathcal{A}\xspace_k}\}K_j-K_j)$ as the $j$-th patch and $\widehat{f}_j = f + \sum_{j'\leq j}\sigma(p_{\mathcal{A}\xspace_{j'}} - p_{\mathcal{A}\xspace_{j'-1}} +\max_{k\geq j'}\{g_{\mathcal{A}\xspace_k}\}K_j-K_j)$ as the network after the $j$-th patch.
For any $x$ in patch area $\cup_{1\leq l\leq I} \mathcal{A}\xspace_l$, we can find a $j$ such that $x\in \mathcal{A}\xspace_j$ but $x\notin \mathcal{A}\xspace_{k}$ for all $k>j$. After the first $j$ patches $\sigma(p_{\mathcal{A}\xspace_1}(x) +\max_{k\geq 1}\{g_{\mathcal{A}\xspace_k}(x, \gamma)\}K_1-K_1)$, $\sigma(p_{\mathcal{A}\xspace_2}(x) - p_{\mathcal{A}\xspace_1}(x) +\max_{k\geq 2}\{g_{\mathcal{A}\xspace_k}(x, \gamma)\}K_2-K_2)$, ... , $\sigma(p_{\mathcal{A}\xspace_j}(x) - p_{\mathcal{A}\xspace_{j-1}}(x) +\max_{k\geq j}\{g_{\mathcal{A}\xspace_k}(x, \gamma)\}K_j-K_j)$, the DNN's output at $x$ becomes $\widehat{f}_j(x) = f(x) + p_{\mathcal{A}\xspace_j}(x)$ which meets our specification $\Phi$ at $x$ by the definition of $p_{\mathcal{A}\xspace_j}(x)$.
Since $x\notin \mathcal{A}\xspace_{k}$ for all $k>j$, then by Corollary \ref{apxcor: Incremental Repair}, the rest of the patches would not change a correct area to an incorrect area. Therefore, we have the final patched neural network $\widehat{f}$ meets specification $\Phi$ on $\cup_{1\leq l \leq I} \mathcal{A}\xspace_l$.
Part 2:
To show that $\widehat{f}$ satisfies $\Phi$ outside of $\mathcal{A}\xspace$.
For any $x$ outside the patch area $\cup_{1\leq l \leq I} \mathcal{A}\xspace_l$, we have $x$ lies on a correct linear region (linear region that satisfies the specification $\Phi$). By Theorem \ref{apxthm: Limited Side Effect}, we have either $\widehat{f}(x) = f(x)$ or $\widehat{f}(x) \in \Phi_{out}$. Therefore, $\widehat{f}$ satisfies $\Phi$ outside of $\mathcal{A}\xspace$.
\end{proof}
\begin{customthm}{4}[Minimum Repair]
For any ReLU DNN $\tilde{f}$, which is linear on a patch region $\mathcal{A}\xspace$ and satisfies the specification $\Phi$, there exists a positive number $\Gamma$, such that for all $\gamma\geq \Gamma$,
\begin{eqnarray}
\max_{x\in X}|\tilde{f}(x)-f(x)|\geq \max_{x\in X}|h_\mathcal{A}\xspace(x, \gamma)|.
\end{eqnarray}
\end{customthm}
\begin{proof}
We consider the general case where the linear patch function is obtained from Equation~(\ref{eqn: linear patch function}).
For any DNN $\tilde{f}$, which is linear on patch region $\mathcal{A}\xspace$ and satisfies the specification $\Phi$, we have $\max_{x\in \mathcal{A}\xspace}|\tilde{f}-f|\geq \max_{x\in \mathcal{A}\xspace}|cx+d| = \max_{x\in \mathcal{A}\xspace}|h_\mathcal{A}\xspace(., \gamma)|$ on patch area $\mathcal{A}\xspace$ by Equation~(\ref{eqn: linear patch function}).
Therefore, we only need to show:
\begin{eqnarray}
\max_{x\notin A}|h_\mathcal{A}\xspace(., \gamma)|\leq \max_{x\in A}|h_\mathcal{A}\xspace(., \gamma)|
\end{eqnarray}
Since parameter $\gamma$ controls the slope of $h_\mathcal{A}\xspace(., \gamma)$ outside of patch area $\mathcal{A}\xspace$, a large $\gamma$ means that $h_\mathcal{A}\xspace(., \gamma)$ will drop to zero quickly outside of $\mathcal{A}\xspace$. Therefore, we can choose a large enough $\Gamma$ such that $h_\mathcal{A}\xspace(., \gamma)$ drops to zero faster than the change of linear patch function $\boldsymbol{c}x+d$.
Therefore, we have that for any $\gamma\geq \Gamma$,
\begin{eqnarray}
\max_{x\notin A}|h_\mathcal{A}\xspace(., \gamma)|\leq \max_{x\in A}|h_\mathcal{A}\xspace(., \gamma)| \nonumber = \max_{x\in X}|h_\mathcal{A}\xspace(., \gamma)| \\
\leq \max_{x\in A}|\tilde{f}-f|\leq \max_{x\in X}|\tilde{f}-f|
\end{eqnarray}
\end{proof}
\begin{customthm}{5}[Polynomial-Time Efficiency]
\texttt{REASSURE}\xspace terminates in polynomial-time in the size of the neural network and the number of buggy linear regions.
\end{customthm}
\begin{proof}
We consider the affine patch function solved via Robust Optimization~(\ref{eqn: programming optimization 2}).
Suppose $\mathcal{A}\xspace = \{x\in X\:|\:a_i x \leq b_i, i \in I\}$. The running time for solving a Linear Programming is polynomial in the number of variables, and the the number of variables for Linear Programming~(\ref{eqn: programming optimization 2}) is polynomial in $|I|$, which is the number of constraints for $\mathcal{A}\xspace$, the DNN's input dimension, and the DNN's output dimension. Since $|I|$ is polynomial in the size of the DNN, \texttt{REASSURE}\xspace runs in polynomial time in the size of the neural network.
In addition, since \texttt{REASSURE}\xspace computes the support network $g_\mathcal{A}\xspace$ and affine patch function $p_\mathcal{A}\xspace$ for each $\mathcal{A}\xspace$ one by one (see Algorithm 1), the time complexity of \texttt{REASSURE}\xspace is linear in the number of buggy linear regions.
\end{proof}
\begin{customthm}{6}
On linear region $\mathcal{A}\xspace$, we have $f|_\mathcal{A}\xspace(x) = \boldsymbol{f_1} x + f_2$ for some matrix $\boldsymbol{f_1}$ and vector $f_2$. Assuming that $\boldsymbol{f_1}$ is full rank\footnote{
Note that for neural networks that are trained by a stochastic method, with probability one $\boldsymbol{f_1}$ is full rank.}, the optimization problem in (\ref{eqn: linear patch function}) and the optimization problem in (\ref{eqn: linear transformation}) are equivalent.
\end{customthm}
\begin{proof}
On one side, for any $\boldsymbol{c}, d$, since $\boldsymbol{f_1}$ is full rank, there exists a linear transformation $T$, such that $T(f(x)) = T(\boldsymbol{f_1}x+f_2) = (\boldsymbol{f_1}+\boldsymbol{c})x + (f_2+d) = f(x) + \boldsymbol{c}x + d$.
On the other side, for any $T$, since $T(f(x)) - f(x)$ is linear, there exist $\boldsymbol{c}, d$, such that $\boldsymbol{c}x + d = T(f(x)) - f(x)$.
\end{proof}
\section{Introduction}
Deep neural networks (DNNs) have demonstrated impressive performances on a wide variety of applications ranging from transportation~\cite{nvidia-dave} to health care~\cite{nn-healthcare}. However, DNNs are not perfect.
In many cases, especially when the DNNs are used in safety-critical contexts, it is important to correct erroneous outputs of a DNN as they are discovered after training.
For instance, a neural network in charge of giving control advisories to the pilots in an aircraft collision avoidance system, such as the ACAS Xu network from \cite{julian2019deep}, may produce an incorrect advisory for certain situations and cause the aircraft to turn towards the incoming aircraft, thereby jeopardizing the safety of both airplanes.
In this paper, we consider the problem of \textit{neural network repair}, i.e. given a trained neural network and a set of buggy inputs (inputs on which the neural network produces incorrect predictions)
, repair the network so that the resulting network on those buggy inputs behave according to some given correctness specification.
Ideally, the changes to the neural network function should be small so that the outputs on other inputs are either unchanged or altered in a small way.
Existing efforts on neural network repair roughly fall into the following three categories.
\quad1. \textit{Retraining/fine-tuning.}
The idea is to retrain or fine-tune the network with the newly identified buggy inputs and the corresponding corrected outputs. Methods include counterexample-guided data augmentation~\cite{dreossi2018counterexample,ren2020few},
editable training~\cite{sinitsin2020editable} and
training input selection~\cite{ma2018mode}.
One major weakness of these approaches is the lack of formal guarantees -- at the end of retraining/fine-tuning, there is no guarantee that the given buggy inputs are fixed and no new bugs are introduced.
In addition, retraining can be very expensive and requires access to the original training data which is impractical in cases where the neural network is obtained from a third party or the training data is private. Fine-tuning, on the other hand, often faces the issue of catastrophic forgetting~\cite{catastrophic-forgetting}.
\quad2. \textit{Direct weight modification.}
These approaches directly manipulate the weights in a neural network to fix the buggy inputs.
The repair problem is typically cast into an optimization problem or a verification problem. For example, \cite{dong2020towards} proposes to minimize a loss defined based on the buggy inputs. \cite{goldberger2020minimal} uses an SMT solver to identify minimal weight changes to the output layer of the network so that the undesirable behaviors are removed.
In general, the optimization-based approach cannot guarantee removal of the buggy behaviors, and the verification-based approach does not scale beyond networks of a few hundred neurons.
In addition, both approaches suffer from substantial accuracy drops on normal inputs since \textit{weight changes may be a poor proxy for changes in the function space}.
\quad3. \textit{Architecture extension.}
The third category of approaches extends the given NN architecture, such as by introducing more weight parameters, to facilitate more efficient repairs. The so-called Decoupled DNN architecture~\cite{sotoudeh2021provable} is the only work we know that falls into this category. Their idea is to decouple the activations of the network from values of the network by augmenting the original network. Their construction allows the formulation of any single-layer repair as an linear programming (LP) problem. The decoupling, however, causes the repaired network to become discontinuous (in the functional sense). In addition, it still cannot isolate the output change to a single buggy input from the rest of the inputs.
\begin{figure}
\centering
\includegraphics[width=0.75\textwidth]{Figures/repair.pdf}
\caption{Comparison of different approaches to the neural network repair problem. The black lines represent the original neural network function. The red dot represents the buggy input. The colored lines represent the functions after the repairs are done.}
\label{fig: repair}
\vspace{-4mm}
\end{figure}
In addition to the aforementioned limitations, a common weakness that is shared amongst these methods is that the induced changes, as a result of either retraining or direct weight modification, are \textit{global}.
This means that a correct behavior on another input, regardless of how far it is away from the buggy input, may not be preserved by the repair.
Worse still, the repair on a new buggy input can end up invalidating the repair on a previous buggy input.
The fundamental issue here is that limiting the changes to a few weights or a single layer only poses a structural constraint (often for ease of computation); it does not limit the changes on the input-output mapping of the neural network.
It is known that even a single weight change can have a global effect on the output of a neural network.
In this paper, we propose \texttt{REASSURE}\xspace,
a novel methodology for neural network repair with locality, minimality, soundness and completeness guarantees.
Our methodology targets continuous piecewise linear (CPWL) neural networks, specifically those that use the ReLU activation functions.
The key idea of our approach is to \textit{leverage the CPWL property of ReLU networks to synthesize a patch network tailored to the linear region where the buggy input resides, which when combined with the original network, provably corrects the behavior on the buggy input}.
Our approach is both sound and complete -- the repaired network is guaranteed to fix the buggy input, and a patch is guaranteed to be found for any buggy input. Moreover, our approach preserves the CPWL nature of ReLU networks, automatically generalizes the repair to all the points including other undetected buggy inputs inside the repair region, is minimal in terms of changes in the function space, and guarantees that outputs on inputs away from the repair region are unaltered.
Figure~\ref{fig: repair} provides an illustrative comparison of our approach with other methods.
Table~\ref{tab: requirements} compares our approach with representative related works in terms of theoretical guarantees.
We summarize our contributions below.
\begin{itemize}
\item We present \texttt{REASSURE}\xspace, the first sound and complete repair methodology for ReLU networks with strong theoretical guarantees.
\item Our technique synthesizes a patch network, which when combined with the original neural network, provably corrects the behavior on the buggy input. This approach is a significant departure from existing methods that rely on retraining or direct weight manipulation.
\item Our technique is both effective and efficient -- across a set of benchmarks, \texttt{REASSURE}\xspace can efficiently correct a set of buggy inputs or buggy areas with little or no change to the accuracy and overall functionality of the network.
\end{itemize}
\section{Experiments}
In this Section, we compare \texttt{REASSURE}\xspace with state-of-the-art methods on both \textit{point-wise repairs} and \textit{area repairs}. The experiments were designed to answer the following questions:
\begin{enumerate}
\item[\textbf{Q1}] \textbf{Effectiveness}: How effective is a repair in removing known buggy behaviors?
\item[\textbf{Q2}] \textbf{Locality}: How much side effect (i.e. modification outside the patch area in the function space) does a repair produce?
\item[\textbf{Q3}] \textbf{Function Change}: How much does a repair change the original neural network in the function space?
\item[\textbf{Q4}] \textbf{Performance}: Whether and how much does a repair adversely affect the overall performance of the neural network?
\end{enumerate}
We consider the following \textbf{evaluation criteria}:
\quad 1. \textit{Efficacy~(E)}: $\%$ of given buggy points or buggy linear regions that are repaired.
\quad 2. \textit{Norm Difference~(ND)}: average norm ($L_\infty$ or $L_2$) difference between the original DNN and the repaired DNN on a set of inputs (e.g. training and testing data; more details in the tables). We use ND to measure how a repair \textit{change} the original neural network on \textit{function space}. Note that the maximum $L_\infty$ norm difference is $1$ and the maximum $L_2$ norm difference is $\sqrt{2}$.
\quad 3. \textit{Norm Difference on Patch Area~(NDP)}: average norm ($L_\infty$ or $L_2$) difference between the original DNN and the repaired DNN on patch areas (calculated on random sampled points on patch areas or near the buggy points; details in the tables). We use NDP to measure the \textit{locality} of a repair.
\quad 4. \textit{Accuracy~(Acc)}: accuracy on training or testing data to measure the extent to which a repair preserves the performance of the original neural network.
\quad 5. \textit{Negative Side Effect~(NSE)}: NSE is only for area repair. It is the percentage of correct linear regions (outside of patch area) that become incorrect after a repair.
If a repair has a nonzero NSE, the new repair may invalidate a previous repair and lead to a circular repair problem.
All experiments were run on an Intel Core i5 @ 3.4 GHz with 32 GB of memory. We use Gurobi \cite{gurobi} to solve the linear programs.
We compared \texttt{REASSURE}\xspace with the representative related works in Table~\ref{tab: requirements}.
\texttt{REASSURE}\xspace, MDNN and PRDNN guarantee to repair all the buggy points (linear regions). Retrain and Fine-Tuning cannot guarantee 100\% efficacy in general and we run them until all the buggy points are repaired.
\subsection{Point-wise Repairs: MNIST}
We train a ReLU DNN on the MNIST dataset~\cite{lecun1998mnist} as the target DNN.
It is a multilayer perceptron with ReLU activation functions. It has an input layer with 784 nodes, 2 hidden layers with 256 nodes in each layer, and a final output layer with 10 nodes.
The goal of a repair is to fix the behaviors of the target DNN on buggy inputs that are found in the test dataset. Thus, the repaired DNN is expected to produce correct predictions for all the buggy inputs.
The results are shown in Table~\ref{table: pointwise repair}.
\texttt{REASSURE}\xspace achieves almost zero modification outside the patch area (ND)
amongst all four methods. In addition, \texttt{REASSURE}\xspace produces the smallest modification on the patch area (NDP) and preserves the performance of the original DNN (almost no drop on Acc).
\begin{table}
\centering
\scalebox{0.93}{
\begin{tabular}{c ccccc| ccccc}
& \multicolumn{5}{c}{\texttt{REASSURE}\xspace} &
\multicolumn{5}{c}{Retrain \textit{(Requires Training Data)}} \\
\#P & ND($L_{\infty}$) & ND($L_2$) & NDP($L_{\infty}$) & NDP($L_2$) & Acc & ND($L_{\infty}$) & ND($L_2$)& NDP($L_{\infty}$) & NDP($L_2$) & Acc\\
\hline
10 & \textbf{0.0001} & \textbf{0.0002} & \textbf{0.2575} & \textbf{0.3495} & \textbf{98.1\%} &0.0113 & 0.0155 & 0.7786 & 1.0801 & \textbf{98.1\%}\\
20 & \textbf{0.0003} & \textbf{0.0003} & \textbf{0.1917} & \textbf{0.2637} & 98.2\% & 0.0092 & 0.0126 & 0.7714 & 1.0757 & \textbf{98.4\%}\\
50 & \textbf{0.0006} & \textbf{0.0008} & \textbf{0.2464} & \textbf{0.3355} & 98.5\% & 0.0084 & 0.0115 & 0.8417 & 1.1584 & \textbf{98.7\%}\\
100 & \textbf{0.0011} & \textbf{0.0017} & \textbf{0.2540} & \textbf{0.3446} & \textbf{99.0\%} & 0.0084 & 0.0116 & 0.8483 & 1.1710 & \textbf{99.0\%}\\
\\
& \multicolumn{5}{c}{Fine-Tuning} & \multicolumn{5}{c}{PRDNN}\\
\#P & ND($L_{\infty}$) & ND($L_2$) & NDP($L_{\infty}$) & NDP($L_2$) & Acc & ND($L_{\infty}$) & ND($L_2$)& NDP($L_{\infty}$) & NDP($L_2$) & Acc\\
\hline
10 & 0.0220 & 0.0297 & 0.6761 & 0.9238 & 97.6\% & 0.0141 & 0.0189 & 0.3460 & 0.4736 & 97.8\%\\
20 & 0.2319 & 0.3151 & 0.8287 & 1.1075 & 78.6\% & 0.0288 & 0.0387 & 0.4363 & 0.5891 & 97.1\%\\
50 & 0.3578 & 0.4799 & 0.8473 & 1.1362 & 67.0\% & 0.0479 & 0.0630 & 0.4937 & 0.6530 & 96.7\%\\
100 & 0.2383 & 0.3118 & 0.7973 & 1.0796 & 81.9\% & 0.0916 & 0.1156 & 0.5123 & 0.6534 & 96.1\%\\
\end{tabular}
}
\caption{Point-wise Repairs on MNIST. We use the first hidden layer as the repair layer for PRDNN. The test accuracy of the original DNN is $98.0\%$. \#P: number of buggy points to repair.
ND($L_\infty$), ND($L_2$): average ($L_\infty$, $L_2$) norm difference on both training and test data. NDP($L_\infty$), NDP($L_2$): average ($L_\infty$, $L_2$) norm difference on random sampled points near the buggy points. Acc: accuracy on test data.
Note that \texttt{REASSURE}\xspace automatically performs area repairs on 784-dimensional inputs.
}
\label{table: pointwise repair}
\vspace{-2mm}
\end{table}
We also compare \texttt{REASSURE}\xspace with MDNN on the watermark removal experiment from their paper. We were not able to run the code provided in the MDNN Github repository, but we were able to run the target DNN models, watermarked images, and MDNN-repaired models from the same repository.
The target DNN is from \cite{goldberger2020minimal}, which has an input layer with 784 nodes, a single hidden layer with 150 nodes, and a final output layer with 10 nodes. The target DNN is watermarked by the method proposed in \cite{adi2018turning} on a set of randomly chosen images $x_i$ with label $f(x_i)$.
The goal is to change the DNN's predictions on all watermarks $x_i$ to any other label $y \neq f(x_i)$ while preserving the DNN's performance on the MNIST test data.
For \texttt{REASSURE}\xspace, we set the prediction $y = f(x_i) - 1$ if $f(x_i) > 1$, and $y = 10$ otherwise.
The results are shown in Table~\ref{tab: watermark}. Both \texttt{REASSURE}\xspace and MDNN remove all the watermarks. However, MDNN introduces significant distortion to the target DNN and as a result the test accuracy drops rapidly as the number of repair points increases. In comparison, \texttt{REASSURE}\xspace removes all the watermarks with no harm to test accuracy.
\begin{table}
\centering
\scalebox{0.94}{
\begin{tabular}{c ccccc| ccccc}
& \multicolumn{5}{c}{\texttt{REASSURE}\xspace} & \multicolumn{5}{c}{MDNN} \\
\#P & ND($L_{\infty}$) & ND($L_2$) & NDP($L_{\infty}$) & NDP($L_2$) & Acc & ND($L_{\infty}$) & ND($L_2$)& NDP($L_{\infty}$) & NDP($L_2$) & Acc\\
\hline
1 & \textbf{0.000} & \textbf{0.000} & \textbf{0.090} & \textbf{0.128} & $\textbf{96.8\%}$
& 0.089 & 0.126 & 0.071 & 0.087 & $87.5\%$ \\
5 & \textbf{0.000} & \textbf{0.000} & \textbf{0.127} & \textbf{0.180} & $\textbf{96.8\%}$
& 0.481 & 0.681 & 0.443 & 0.562 & $57.1\%$\\
25 & \textbf{0.000} & \textbf{0.000} & \textbf{0.299} & \textbf{0.420} & $\textbf{96.8\%}$
& 0.904 & 1.278 & 0.637 & 0.815 & $6.7\%$\\
50 & \textbf{0.000} & \textbf{0.000} & \textbf{0.429} & \textbf{0.602} & $\textbf{96.8\%}$
& 0.925 & 1.308 & 0.821 & 1.024 & $4.8\%$\\
100 & \textbf{0.000} & \textbf{0.000} & \textbf{0.462} & \textbf{0.648} & $\textbf{96.8\%}$
& 0.955 & 1.350 & 0.909 & 1.083 & $5.1\%$\\
\end{tabular}}
\caption{Watermark Removal. The test accuracy of the original DNN is $96.8\%$. \#P: number of buggy points to repair; ND($L_\infty$), ND($L_2$): average ($L_\infty$, $L_2$) norm difference on both training data and testing data; NDP($L_\infty$), NDP($L_2$): average ($L_\infty$, $L_2$) norm difference on random sampled points near watermark images; Acc: accuracy on test data.}\label{table: watermark removal}
\label{tab: watermark}
\end{table}
\subsection{Area Repairs: HCAS}\label{sec: HCAS}
To the best of our knowledge, \cite{sotoudeh2021provable} is the only other method that supports area repairs. In this experiment, we compare \texttt{REASSURE}\xspace with \cite{sotoudeh2021provable} on an experiment where the setting is similar to the \textit{2D Polytope ACAS Xu repair} in their paper.
\cite{sotoudeh2021provable} does not include a vertex enumeration tool (which is required for setting up their LP problem) in their code. We use \texttt{pycddlib} \cite{troffaes2018pycddlib} to perform the vertex enumeration step when evaluating PRDNN. Note that the vertex enumeration tool does not affect the experimental results except running time.
We consider an area repair where the target DNN is the HCAS network (simplified version of ACAS Xu)\footnote{The technique in PRDNN for computing linear regions does not scale beyond two dimensions as stated in their paper. The input space of HCAS is 3D and that of ACAS Xu is 5D so we use HCAS in order to run their tool in our evaluation of area repairs.} $N_{1,4}$ (previous advisory equal to $1$ and time to loss of vertical separation equal to $20s$) from \cite{julian2019guaranteeing}.
$N_{1,4}$ has an input layer with 3 nodes, 5 hidden layers with 25 nodes in each hidden layer, and a final output layer with 5 nodes. DNN outputs one of five possible control advisories ('Strong left', 'Weak left', 'Clear-of-Conflict', 'Weak right' and 'Strong right').
We use Specification~\ref{spec: HCAS1}, which is similar to Property 5 in \cite{katz2017reluplex}. We use the method from \cite{girard2021disco} to compute all the linear regions for $N_{1,4}$ in the area $\Phi_{in}$ of Specification~\ref{spec: HCAS1} and a total of 87 buggy linear regions were found.
We apply both $\texttt{REASSURE}\xspace$ and PRDNN to repair these buggy linear regions.
\begin{specification}\label{spec: HCAS1}
If the intruder is near and approaching from the left, the network advises “strong right.”
Input constraints: $\Phi_{in} = \{ (x, y, \psi) | 10\leq x \leq 5000, 10\leq y \leq 5000, -\pi \leq \psi \leq -1/2\pi\}$. Output constraint: $f(x, y, \psi)_4 \geq f(x, y, \psi)_i$ for $i = 0, 1, 2, 3$.
\end{specification}
We use Specification~\ref{spec: HCAS2}, the dual of Specification~\ref{spec: HCAS1}, to test the \textit{negative side effect (NSE)} of a repair. We calculate all the linear regions in the area $\Phi_{in}$ of Specification~\ref{spec: HCAS2} and 79 correct linear regions are found. We will test if a repair will make those correct linear regions incorrect.
\begin{specification}\label{spec: HCAS2}
If the intruder is near and approaching from the right, the network advises “strong left.”
Input constraints: $\Phi_{in} = \{ (x, y, \psi) | 10\leq x \leq 5000, -5000\leq y \leq -10, 1/2\pi \leq \psi \leq \pi\}$. Output constraint: $f(x, y, \psi)_0 \geq f(x, y, \psi)_i$ for $i = 1, 2, 3, 4$.
\end{specification}
The results are shown in Table~\ref{tab: HCAS}. Both \texttt{REASSURE}\xspace and PRDNN successfully repair all the buggy linear regions. \texttt{REASSURE}\xspace produces repairs that are significantly better in terms of \textit{locality} (ND), \textit{minimality} (NDP) and \textit{performance preservation} (Acc).
In addition, as mentioned in the previous experiment on MNIST, \texttt{REASSURE}\xspace automatically performs area repair on point-wise repair problems. This means our area repair method scales well to high-dimensional polytopes (the input dimension of MNIST is $784$) whereas PRDNN does not scale beyond 2D linear regions/polytopes.
\begin{table}
\centering
\scalebox{0.98}{
\begin{tabular}{c ccccc| ccccc}
& \multicolumn{5}{c}{\texttt{REASSURE}\xspace} & \multicolumn{5}{c}{PRDNN}\\
\#A & ND($L_{\infty}$) & NDP($L_{\infty}$) & NSE & Acc & T & ND($L_{\infty}$) & NDP($L_{\infty}$) & NSE & Acc & T\\
\hline
10 & \textbf{0.0000} & \textbf{0.000} & \textbf{0\%} & \textbf{98.1\%} & \textbf{1.0422}
& 0.0010 & 0.316 & 4\% & 89.6\% & 2.90+0.100\\
20 & \textbf{0.0000} & \textbf{0.022} & \textbf{0\%} & \textbf{98.1\%} & \textbf{1.1856}
& 0.0015 & 0.372 & 8\% & 83.1\% & 5.81+0.185 \\
50 & \textbf{0.0000} & \textbf{0.176} & \textbf{0\%} & \textbf{98.1\%} & \textbf{1.8174}
& 0.0015 & 0.384 & 8\% & 83.8\% & 14.54+0.388\\
87 & \textbf{0.0004} & \textbf{0.459} & \textbf{0\%} & \textbf{97.8\%} & \textbf{2.4571}
& 0.0014 & 0.466 & \textbf{0\%} & 85.6\% & 25.30+0.714\\
\end{tabular}
}
\caption{Area Repairs on HCAS.
We use the the first hidden layer as the repair layer for PRDNN. Results on PRDNN using the last layer (which are inferior to using the first layer) are shown in Table~\ref{apd tab: HCAS} in the Appendix~\ref{sec: Additional Experiment Details}.
The test accuracy of the original DNN is $97.9\%$. \#A: number of buggy linear regions to repair.
ND($L_\infty$): average $L_\infty$ norm difference on training data.
NDP($L_\infty$): average $L_\infty$ norm difference on random sampled data on input constraints of specification~\ref{spec: HCAS1}.
NSE: $\%$ of correct linear regions changed to incorrect by the repair.
Acc: accuracy on training data (no testing data available).
T: running time in seconds. For \texttt{REASSURE}\xspace, the running time is based on the LP formulation in Appendix~\ref{apd sec: repair via LP}.
For PRDNN, the first running time is for enumerating all the vertices of the polytopes and the second is for solving the LP problem in PRDNN.}\label{tab: HCAS}
\vspace{-2mm}
\end{table}
\subsection{Feature-Space Repairs: ImageNet}
We use AlexNet \cite{krizhevsky2012imagenet} on the ImageNet dataset~\cite{ILSVRC15} as the target DNN. The size of an input image is (224, 224, 3) and the total number of classes for ImageNet is 1000. We slightly modified AlexNet to simplify the evaluation: we only consider 10 output classes that our buggy images may lie on and use a multilayer perceptron with three hidden layers (512, 256, 256 nodes respectively) to mimic the last two layers of AlexNet.
The resulting network has 650k neurons, consisting of five convolutional layers, some of which are followed by max-pooling layers, followed by five fully-connected layers with (9216, 4096, 512, 256, 256 nodes respectively).
The total number of parameters in the resulting network is around 60 million.
The goal of the repair is to fix the behaviors of the target DNN on buggy inputs, which are found on ImageNet-A \cite{hendrycks2021natural}. For $\texttt{REASSURE}\xspace$, we
construct the patch network starting from the 17-th (third from the last) hidden layer (i.e. repair in a feature space).
The results are shown in Table~\ref{table: pointwise repair ImageNet}. Both \texttt{REASSURE}\xspace and PRDNN successfully repair all the buggy points.
\texttt{REASSURE}\xspace achieves almost zero modification on the validation images compared to the original DNN. In addition, \texttt{REASSURE}\xspace preserves the performance of the original DNN.
\begin{table}
\centering
\scalebox{0.84}{
\begin{tabular}{c ccc| ccc| ccc| ccc}
& \multicolumn{3}{c}{\texttt{REASSURE}\xspace \textit{(feature space)}} &
\multicolumn{3}{c}{Retrain \textit{(Requires Training Data)}}
& \multicolumn{3}{c}{Fine-Tuning}
&\multicolumn{3}{c}{PRDNN}\\
\#P & ND($L_{\infty}$) & ND($L_2$) & Acc
& ND($L_{\infty}$) & ND($L_2$) & Acc
& ND($L_{\infty}$) & ND($L_2$) & Acc
& ND($L_{\infty}$) & ND($L_2$) & Acc\\
\hline
10 & \textbf{0.0015} & \textbf{0.0018} & \textbf{82.5\%}
& 0.4343 & 0.5082 & 80.1\%
& 0.2945 & 0.3563 & 77.9\%
& 0.2293 & 0.2980 & 82.1\% \\
20 & \textbf{0.0012} & \textbf{0.0015} & 81.3\%
& 0.4278 & 0.5032 & \textbf{82.9\%}
& 0.6916 & 0.8103 & 68.5\%
& 0.2191 & 0.2824 & 80.1\%\\
50 & \textbf{0.0079} & \textbf{0.0099} & 81.3\%
& 0.5023* & 0.6043* & \textbf{82.1\%}*
& 0.7669 & 0.8948 & 66.9\%
& 0.3096 & 0.3799 & 68.9\%
\end{tabular}
}
\caption{Point-wise Repairs on ImageNet. PRDNN uses parameters in the last layer for repair. The test accuracy for the original DNN is $83.1\%$. \#P: number of buggy points to repair.
ND($L_\infty$), ND($L_2$): average ($L_\infty$, $L_2$) norm difference on validation data.
Acc: accuracy on validation data. * means Retrain only repair 96\% buggy points in 100 epochs.}\label{table: pointwise repair ImageNet}
\end{table}
\subsection{Running Example}
We use the following example to illustrate our idea.
\begin{example}
Consider repairing the ReLU DNN $f$ in Figure~\ref{fig: RunEx}
according to the correctness specification $\Phi: \forall x \in [0, 1]^2$, $y\in [0, 2]$.
The DNN consists of a single hidden layer with two neurons $z_1$ and $z_2$, where $y = \sigma(z_1) + \sigma(z_2)$, $z_1 = x_1+2x_2-1$ and $z_2 = 2x_1-x_2$.
\end{example}
\begin{figure}
\centering
\includegraphics[width=0.4\textwidth]{Figures/RunEx1.pdf}
\includegraphics[width=0.45\textwidth]{Figures/RunEx2.pdf}
\caption{Left: the target DNN with buggy inputs.
Right: the \texttt{REASSURE}\xspace-repaired DNN with the patch network shown in red. Support network $g_\mathcal{A}\xspace$ is for approximating the characteristic function on $\mathcal{A}\xspace$; Affine patch function $p_\mathcal{A}\xspace$ ensures the satisfaction of $\Phi$ on $\mathcal{A}\xspace$; The design of the patch network $h_\mathcal{A}\xspace$ ensures locality for the final patch.
}\label{fig: RunEx}
\vspace{-2mm}
\end{figure}
The only linear region that violates our specification is $\mathcal{A}\xspace = \{x\:|\:1\geq x_1, 1\geq x_2, x_1+2x_2-1\geq 0, 2x_1-x_2\geq 0\}$. ( $\widetilde{x} = (0.9, 0.9)\in [0, 1]^2$ but $f(\widetilde{x}) = 2.6 \notin [0, 2]$)
The network $f(x)$ on the linear region $\mathcal{A}\xspace$ is the affine function $f|_\mathcal{A}\xspace(x) = 3x_1+x_2-1$.
Our algorithm first sets up an affine function $p_\mathcal{A}\xspace(x)$ that minimally repairs $f$ on $\mathcal{A}\xspace$, such that $\forall x \in \mathcal{A}\xspace, f(x)+p_\mathcal{A}\xspace(x)\in [0, 2]$.
Later in the paper, we will show $p_\mathcal{A}\xspace(x)$ can be found by solving a LP problem.
The resulting patch function is $p_\mathcal{A}\xspace(x) = -\frac{1}{2}x_1-\frac{1}{2}x_2$.
However, directly apply $f(x) + p_\mathcal{A}\xspace(x)$ as the patch network will have side effects on areas outside of $\mathcal{A}\xspace$.
Our strategy is to combine $p_\mathcal{A}\xspace(x)$ with a support network $g_\mathcal{A}\xspace(x)$
which outputs $1$ on $\mathcal{A}\xspace$ and drops to $0$ quickly outside of $\mathcal{A}\xspace$.
The final repaired network is $f(x) + \sigma(p_\mathcal{A}\xspace(x)+g_\mathcal{A}\xspace(x,10)-1) - \sigma(-p_\mathcal{A}\xspace(x)+g_\mathcal{A}\xspace(x,10)-1)$. This structure makes $p_\mathcal{A}\xspace$ almost only active on $\mathcal{A}\xspace$ and achieve a localized repair.
Observe that this is still a ReLU DNN.
\subsection{Repair via Robust Optimization}\label{sec: repair via robust optimization}
In general, the number of vertices of a convex polytope can be exponential in the size of its H-representation~\cite{convex-polytope} and enumerating the vertices of a convex polytope is known to be expensive especially when the input dimension is large~\cite{bremner1997complexity}. In this section, we show that we can convert problem (\ref{eqn: linear patch function}) to an LP via \textit{robust optimization}~\cite{ben2009robust} without vertex enumeration and make our algorithm much more efficient.
The optimization problem (\ref{eqn: linear patch function}) can be converted to the following optimization problem, assuming $\Phi_{out} = \{y|a_{out}\cdot y\leq b_{out}\}$:
\begin{eqnarray}\label{programming optimization}
\begin{cases}
\min_{\boldsymbol{c}, d} H\\
H \geq H_1
\textit{, where $H_1$ is the maximum value of the following inner LP }\begin{cases}
\max_x \boldsymbol{c}x+d\\
x \in \mathcal{A}\xspace
\end{cases}\\
H \geq H_2
\textit{, where $H_2$ is the maximum value of the following inner LP }\begin{cases}
\max_x -\boldsymbol{c}x-d\\
x \in \mathcal{A}\xspace
\end{cases}\\
b_{out} \geq H_3
\textit{, where $H_3$ is the maximum value of the following inner LP }
\begin{cases}
\max_x a_{out}\cdot(f(x)+\boldsymbol{c}x+d)\\
x \in \mathcal{A}\xspace
\end{cases}\\
\end{cases}
\end{eqnarray}
For the inner LPs, since we only care about the maximum value and not the feasible solution that reaches the maximum, we can take the dual of inner LPs to avoid enumerating the vertices of $\mathcal{A}\xspace$. Take the first inner LP as an example:
\begin{eqnarray}
H_1 =
\begin{cases}
\max_x \boldsymbol{c}x+d\\
x \in \mathcal{A}\xspace
\end{cases}
= d+
\begin{cases}
\max_x \boldsymbol{c}x\\
ax\leq b
\end{cases}
\stackrel{\text{dual}}{=} d+
\begin{cases}
\min_p p'b\\
a'p=c\\
p\geq 0
\end{cases}
\end{eqnarray}
Therefore, we have the following equivalent LP for optimization problem (\ref{eqn: linear patch function}). Taking the dual of the inner LP to convert the whole problem to an LP is known as taking the \textit{robust counterpart}~\cite{ben2009robust} of the original problem.
\begin{eqnarray}\label{eqn: programming optimization 2}
\begin{cases}
\min_{\boldsymbol{c}, d, p_1, p_2, q, H} H\\
H \geq p_1'b + d,
a'p_1=\boldsymbol{c},
p_1\geq 0\\
H \geq p_2'b - d,
a'p_2=-\boldsymbol{c},
p_2\geq 0\\
b_{out} \geq q'b + a_{out}(d_f+d),
a'q=a_{out}(\boldsymbol{c_f}+\boldsymbol{c}),
q\geq 0
\end{cases}
\end{eqnarray}
where $f(x) = \boldsymbol{c_f}x + d_f$.
\section{Appendix}
\subsection{An Alternative LP Solution}\label{apd sec: repair via LP}
In Section~\ref{sec: repair via robust optimization}, we show that optimization problem (\ref{eqn: linear patch function}) can be converted to an LP via \textit{robust optimization}~\cite{ben2009robust}.
However, taking the dual of the inner LP in robust optimization introduces new variables. Specifically, the number of new variables is in $O(n|I|)$, where $n$ is the DNN's output dimension and $|I|$ is the number of constraints for the linear region. Thus, it is more expensive to solve the LP when $n$ is large (although still much less expensive than enumerating all the vertices).
In this section, we show that we can solve optimization problem (\ref{eqn: linear patch function}) via LP more efficiently for many useful cases such as the classification problem in Example~\ref{exp: classification problem} and the HCAS example in Section~\ref{sec: HCAS}.
We consider the case where $\Phi_{out}$ can be expressed as $\{y\:|\: q_l\leq Py\leq q_u\}$ where $P$ is a \textit{full row rank} matrix and $-\infty\leq q_l[i]\leq q_u[i]\leq +\infty$ ($q_l[i]$ and $q_u[i]$ are the $i$-th elements of $q_l$ and $q_u$ respectively).
Consider the following optimization problem.
\begin{eqnarray}\label{eqn: linear transformation}
\begin{cases}
\min_T \max_{x\in \mathcal{A}\xspace} |T(f(x) - f(x)|\\
q_l \leq P(T(f(x)))\leq q_u, \forall x \in \mathcal{A}\xspace
\end{cases}
\end{eqnarray}
where $T: \mathbb{R}^n \to \mathbb{R}^n$ is a linear transformation on the DNN's output space $\mathbb{R}^n$.
\begin{theorem}
On linear region $\mathcal{A}\xspace$, we have $f|_\mathcal{A}\xspace(x) = \boldsymbol{f_1} x + f_2$ for some matrix $\boldsymbol{f_1}$ and vector $f_2$. Assuming that $\boldsymbol{f_1}$ is full rank\footnote{
Note that for neural networks that are trained by a stochastic method, with probability one $\boldsymbol{f_1}$ is full rank.}, the optimization problem in (\ref{eqn: linear patch function}) and the optimization problem in (\ref{eqn: linear transformation}) are equivalent.
\end{theorem}
Note that $q_l \leq P(T(f(x)))\leq q_u$ can be achieved row by row. Thus, we can find a one-dimensional linear transformation via LPs and combine them into a single linear transformation to solve optimization problem (\ref{eqn: linear transformation}).
For every row of $P$, say the $i$-th row, we can check the lower bound and upper bound of $\{P(f(x))\:|\:\forall x \in \mathcal{A}\xspace\}$ on the $i$-th dimension by solving the following LP problems
\begin{eqnarray}\label{LP: i-th row}
lb[i] = \min_{x\in \mathcal{A}\xspace} P[i](f(x))\quad
ub[i] = \max_{x\in \mathcal{A}\xspace} P[i](f(x))
\end{eqnarray}
where $P[i]$ is the $i$-th row of $P$.
Then for each row $i$, we take a minimal linear transformation $V[i](x) = v_1[i](x) + v_2[i]$ to transfer interval $[lb[i], ub[i]]$ inside interval $[q_l[i], q_u[i]]$. We can take $v_1[i] = 1$ if $q_u[i] - q_l[i] > ub[i] - lb[i]$, else $v_1[i] = \frac{q_u[i] - q_l[i]}{ub[i] - lb[i]}$. And $v_2[i] = q_l[i] - v_1[i] lb[i]$ if $|q_l[i] - v_1[i] lb[i]| \leq |v_1[i] ub[i] - q_u[i]|$, else $v_1[i] ub[i] - q_u[i]$.
Since matrix $P$ is full row rank, we can find a linear transformation $T$ that is equivalent to $V$:
\begin{eqnarray}
T = \hat{P}^{-1} V \hat{P} \Rightarrow P(T(f(x))) = V(P(f(x)))
\end{eqnarray}
where $\hat{P} = \left[\begin{array}{c}
P \\
P^{\bot}
\end{array}\right]$ is an orthogonal extension of $P$ ($P$ and $P^{\bot}$ are orthogonal to each other and $\hat{P}$ is a full rank square matrix).
Once we have $T$, we can obtain an affine patch function $p_\mathcal{A}\xspace(x) = T(f(x)) - f(x)$.
\input{Sections/proofs}
\subsection{Additional Experiment Details}\label{sec: Additional Experiment Details}
\textbf{Area Repair: HCAS}
Table~\ref{apd tab: HCAS} is the comparison with PRDNN using the last layer as the repair layer. All other settings are the same as those in Section~\ref{sec: HCAS}.
\begin{table}[H]
\centering
\scalebox{0.8}{
\begin{tabular}{c ccccc| ccccc}
& \multicolumn{5}{c}{\texttt{REASSURE}\xspace} &
\multicolumn{5}{c}{PRDNN (Last Layer)}\\
\#A & ND($L_{\infty}$) & NDP($L_{\infty}$) & NSE & Acc & T& ND($L_{\infty}$) & NDP($L_{\infty}$) & NSE & Acc & T\\
\hline
10 & \textbf{2.662e-05} & \textbf{1.318e-05} & \textbf{0\%} & \textbf{98.1\%} & \textbf{1.0422}
& 0.0030 & 0.205 & 16\% & 71.4\% & 2.90+0.100 \\
20 & \textbf{2.918e-05} & \textbf{0.022} & \textbf{0\%} & \textbf{98.1\%} & \textbf{1.1856}
& 0.0031 & 0.467 & 66\% & 70.5\% & 5.81+0.169 \\
50 & \textbf{8.289e-05} & \textbf{0.176} & \textbf{0\%} & \textbf{98.1\%} & \textbf{1.8174}
& 0.0031 & 0.467 & 66\% & 70.5\% & 14.54+0.353 \\
87 & \textbf{0.0004} & \textbf{0.459} & \textbf{0\%} & \textbf{97.8\%} & \textbf{2.4571}
& 0.0031 & 0.467 & 66\% & 70.5\% & 25.30+0.467 \\
\end{tabular}
}
\caption{Area Repairs on HCAS.
We use the the last hidden layer as the repair layer for PRDNN.
The test accuracy of the original DNN is $97.9\%$. \#A: number of buggy linear regions to repair; ND($L_\infty$): average $L_\infty$ norm difference on training data ; NDP($L_\infty$): average $L_\infty$ norm difference on random sampled data on input constraints of Specification~\ref{spec: HCAS1}; NSE: $\%$ of correct linear regions that is repaired to incorrect; Acc: accuracy on training data (no testing data available); T: running time in seconds. For PRDNN, the first running time is for enumerating all the vertices of the polytopes and the second is for solving the LP problem in PRDNN.}\label{apd tab: HCAS}
\end{table}
\textbf{Hyperparameters used in Repair:}
We set learning rate to $10^{-3}$ for Retrain in the point-wise repair experiment.
We set learning rate to $10^{-2}$ and momentum to $0.9$ for Fine-Tuning in the point-wise repair experiment.
PRDNN requires specifying a layer for weight modification. We use the first hidden layer as the repair layer, which has the best performance in our experiment settings, unless otherwise specified.
\section{Introduction}
\end{document}
|
\section{Introduction}
The development of cavity QED is marked by the achievement of strong light-matter coupling \cite{Haroche2013}.
This experimental milestone opened the possibility of entangling light and matter.
Since then, experiments have advanced to be able to control and manipulate a few atoms and photons in a fully quantum way.
Nowadays, there are many situations where strong coupling is achieved. Natural atoms or molecules as well as artificial systems such as quantum dots or superconducting circuits \cite{Auffeves2014}.
Some of these systems allow us to push light-matter coupling to new regimes. One of them is ultrastrong coupling, where the interaction strength is
comparable to the bare energies of cavity photons and atomic transitions \cite{Gunter2009, Niemczyk2010}.
Here, correlations arise such that the ground state is already an entangled state with both virtual matter and photon excitations \cite{Kockum2019, FornDiaz2019}.
In the same spirit as the original motivation of cavity QED -to control atoms with quantum, rather than classical, light- a new field of research is emerging.
It is known as cavity QED materials \cite{Hubener2020,GarciaVidal2021, rokaj2020free}.
Here, instead of a few atoms, the modification of bulk material properties due to nonperturbative light-matter correlations is investigated.
Thus, the focus is on the physics of a macroscopic number of particles coupled to the quantized electromagnetic field.
Quantum light fluctuations have recently been shown to modify chemical reactivity \cite{thomas2016groundstate, thomas2019tilting}, excitonic transport \cite{Feist2015, Schachenmayer2015,Orgiu2015}, superconductivity \cite{Sentef2018, Schlawin2019, Curtis2019, thomas2019exploring} the ferroelectric phase in quantum paramagnetic materials \cite{Pilar2020, Schuler2020, ashida2020} and ferromagnetism \cite{romanroche2021, thomas2021large}.
At the same time, the cavity-QED community has gained awareness of the critical role that certain ubiquitous approximations can have in the predicted behaviour of a system, often leading to wrong theoretical predictions that later fail to achieve experimental confirmation. This is especially important in the non-perturbative regime, which is precisely the regime where modifications of matter occur.
Perhaps the most discussed example in the recent literature is equilibrium superradiance (photon condensation). This quantum phase transition was originally predicted for electric dipoles in a uniform electric field (Dicke model) \cite{Hepp1973, Wang1973, Hepp1973a}, but it has since been shown to be just a consequence of the breaking of gauge invariance in theories missing the infamous $A^2$ term \cite{Garziano2020, Andolina2019}. Equilibrium superradiance can be achieved, nevertheless, in a system of electric charges with spatially-varying electromagnetic (EM) fields \cite{Nataf2019, guerci2020, andolina2020} or in systems with magnetic dipoles \cite{romanroche2021}, even when gauge-invariance is restored, but it serves as an example of the profound consequences that an incomplete description of cavity QED can have. Several works have also pointed out that the two level approximation is a source of mistakes if performed, without additional care, in the Coulomb gauge \cite{Keeling2007, distefano2019resolution}.
Finally, image charges constitute a potentially relevant electrostatic contribution to the effective interactions between charges in a cavity \cite{vuviks2014, debernardis2018}. Thus, a complete formulation of cavity QED must include them along the dynamical transverse fields of the cavity in the Coulomb gauge. This is crucial to later resolve the contributions of the two to the modification of matter properties.
Another important caveat in the study of hybrid systems in cavity QED is the fact that, when studying the interplay of light and matter in a strong coupling scenario, the separation between ``light'' and ``matter'' blurs. This occurs not only because of the well-understood creation of hybrid polaritonic excitations but, more fundamentally, because the definition of ``light'' and ``matter'' subsystems is gauge-dependent \cite{stokes2021}. This observation is perfectly compatible with the gauge-invariance of physical observables but it raises questions regarding our labelling of the underlying cause of certain phenomena \cite{Keeling2007, stokes2020}. When we say that an effective interaction is \emph{light-mediated}, is this a universal statement?, or can a change of gauge make it seem as though the interaction is \emph{matter-mediated}?
Here, we avoid descending into this philosophical rabbit hole by defining the bare matter first. We start by establishing the Hamiltonian that describes the behaviour of a material system in the absence of a confined electromagnetic field, i.e. without a cavity. This is motivated by the fact that most material properties are well captured by such a description, such as superconductivity, ferromagnetism, electric and thermal conductivity, etc.. In this sense, there exist experimentally accessible observables that can be measured for the bare material system. Then, we study what happens when such a material is placed within a cavity. For that, we couple it to the quantized electromagnetic field in the Coulomb gauge for an arbitrary cavity geometry. The \emph{experimentally relevant question} is then: if we measure the same experimentally accessible observables now, to what extent are they altered? If there is a change, we naturally attribute it to the effect of the cavity, even if in a microscopic description of the hybrid light-matter system it is impossible to universally distinguish between the two subsystems.
In this work we develop the theoretical analogue to this experimental procedure by deriving an effective theory of matter inside a cavity. Generally, an effective theory is found when a subsystem is systematically eliminated from the dynamical description of a larger system.
Perturbative effective theories have widespread use in cavity QED. Notable examples are adiabatic elimination \cite{agarwal1997} or the dispersive theory \cite{Klimov2000, Klimov2003, Zueco2009}.
However, for our purpose of studying strong light-matter correlations, it is necessary to consider non-perturbative techniques.
A seminal example is the Feynman-Vernon influence functional within the path integral formalism, developed for quantum systems coupled to a continuum (open systems) \cite{feynman1963the, caldeira1983path, grabert1988quantum, weiss2011quantum}.
Recently, coherent state path integrals have been used to define effective models in cavity QED \cite{muller2012quantum, dmytruk2021gauge, strack2011dicke}. Here, we build upon these works: we use a coherent-state path integral to define an \emph{exact effective theory} for a general material gauge-invariantly coupled to a general cavity.
This exact treatment yields a non-local effective action that becomes local in the large-$N$ limit.
In this limit, the theory can be cast into an effective matter Hamiltonian where the effect of light vacuum fluctuations takes the explicit form of induced matter-matter interactions.
The range and form of these interactions depends on the light-matter Hamiltonian and the geometry of the cavity.
The theory is compared against those effective models commonly employed in the field. Unlike theories based on the adiabatic elimination of fast variables, our theory is valid in the full light-matter coupling regime. We also test its finite size scaling numerically via exact diagonalisation. In doing that, we show that our theory matches the exact results already for $N \sim 10^2$.
Finally, we test the power of our formalism by deriving recent results from the literature.
The remainder of the manuscript is organised as follows. In Section \ref{sec:summary} we provide a graphical and non-technical summary of the main results of the paper. Section \ref{sec:model} is devoted to constructing a comprehensive model of matter inside an electromagnetic cavity. Then, Section \ref{sec:effectivemodel} constitutes the main result of the paper: we trace out the cavity to derive the effective model. In Sec. \ref{sec:test} we compare the theory against other effective models and test its finite size scaling. In Sec. \ref{sec:examples} we study different material systems within a cavity to showcase the descriptive power of our effective theory. We close the paper with some conclusions. Technical details and concepts of minor importance are sent to the appendices.
\begin{figure}[b]
\centering
\includegraphics[width = 0.9\textwidth]{visual_abstract_6.pdf}
\caption{Schematic summary of the derivation of the effective theory. We begin by defining a general matter model in a). We accompany the illustration with an example Hamiltonian, but the theory extends to other models as well. The matter is then minimally coupled to an electromagnetic cavity in b). The corresponding Hamiltonian describes a broad range of models, which are encoded in the coupling constants $\{c_{\boldsymbol k}\}$ and operators $\{\calham C_{\kappa}\}$, and in $\calham H_M'$. In c), the light degrees of freedom are non-perturbatively eliminated with a coherent state path integral, yielding a non-local action with effective matter-matter couplings. The non-local contributions vanish in the large-$N$ limit and an exact effective Hamiltonian is obtained in d).}
\label{fig:visualabstract}
\end{figure}
\section{Summary of main results}
\label{sec:summary}
The main focus of this work is the theoretical description of a broad range of materials when these are coupled to an electromagnetic cavity. To avoid subtleties, we fix our starting hypothesis. We begin by asserting that a material is an extensive collection of typically a macroscopic number $N$ of particles that can be isolated (to a good approximation) and their behaviour described by a Hamiltonian $\calham H_M$. This is illustrated in Fig. \ref{fig:visualabstract}(a). We then couple this material to a quantized electromagnetic cavity and our formalism begins. For the cavity, we focus on a complete description including the $A^2$ term, the Zeeman coupling and image charges from the cavity boundaries. We consider a multi-mode and non-uniform electromagnetic field and we do not specify any particular geometry, leaving the spatial dependence encoded in the mode functions of the EM field. We must disclaim that, in order to diagonalize the photonic Hamiltonian in the model of a gas of charged particles, we simplify certain mode-mode interactions arising from the $A^2$ term. This is a common occurrence in the literature and its adequacy is discussed in the text. We show that due to the fact that the light-matter Hamiltonian is quadratic in bosonic (light) operators, it can be brought to the linear form shown in Fig. \ref{fig:visualabstract}(b), which corresponds to Eq. \eqref{eq:linearlmH}. This is a general Hamiltonian that can be used to describe other systems besides the ones that we treat in this paper. It describes the linear coupling of an unspecified extensive subsystem of interest to a diagonal(ized) bosonic ensemble. In our case, the role of the bosons is played by cavity photons, but the Hamiltonian could also be used to describe, e.g., phononic or magnonic ensembles.
Then, we use a coherent-state path integral formulation of the partition function of the system to trace out the photonic degrees of freedom (DOFs). In an \emph{exact} treatment, we obtain an effective action \eqref{eq:nonlocalaction} in which the light-matter coupling has been replaced by an cavity-field-mediated interaction between the constituents of the material subsystem. This interaction is generally non-local in imaginary time, as illustrated in Fig. \ref{fig:visualabstract}(c). Then, we show that in the large-$N$ limit the action becomes local, moving from an effective action to an effective Hamiltonian \eqref{eq:effectiveham} as shown in Fig. \ref{fig:visualabstract}(d). In this effective Hamiltonian, the role of the cavity is made explicit in two ways: first, the bare coulombic interactions are modified by the image charges; second, an interpretable additional term introduces effective interactions mediated by the transverse fields of the cavity. In the process, we recover relations between photonic and material observables, such that information about the cavity is accessible even when the latter is traced out. The theory is tested on the Dicke model against other effective techniques, showing that it is not equivalent to perturbative techniques . Its finite size scaling is also numerically tested, showing that our theory becomes exact already for $N\sim 10^2$. To illustrate the applicability of our formalism, we finish the manuscript by applying it to three different material systems in a cavity. First, we revisit the problem of photon condensation. Within our effective model, we recover the no-go theorem for superradiance in the single-uniform-mode limit and we show how neglecting the $A^2$ term leads to a wrong prediction of the phase transition. Our result rules out phase transitions of any order in a single formulation and is valid at finite temperatures. This is a generalization of \cite{Andolina2019, andolina2021}. Then, we obtain critical conditions for photon condensation with spatially-varying electromagnetic fields, recovering \cite{andolina2020}. Then, we solve the 2D free electron gas in a cavity, recovering the results in \cite{rokaj2020}. The last system that we solve is an ensemble of molecular nanomagnets. These couple to the cavity solely by the Zeeman term, avoiding the $A^2$ problem, and have been shown to be a suitable candidate to showcase cavity-modified ferromagnetism as well as photon condensation \cite{romanroche2021}. We obtain the effective spin-spin interactions for an arbitrary magnetic field geometry.
\section{Model}
\label{sec:model}
We begin by fixing the model under consideration. First, we consider a general class of material subsystems, composed of $N$ identical particles. Depending on the degree to which the particles are bounded to a particular position in space we can distinguish two different families of systems. A gas, where the particles are free to occupy any position of space, like an electron gas in a metallic lattice; and a system of emitters where each particle or group of particles is bounded to a non-dynamical core, such as a system of neutral atoms or molecules where the motion of the nucleus is disregarded and the focus is placed on the electron(s) motion. We also discuss the limit case of completely fixed emitters such as magnetic spins. The two families differ slightly in the way they couple to EM fields. Nevertheless, the degree of parallelism in their analytical description is high, so we will only discuss the more complex case of the gas in the main text, and comment briefly on the derivation for a system of localized emitters, highlighting the differences with the former and reserving the detailed derivation for Appendix \ref{ap:localizedemitters}.\\
\subsection{A gas of charged particles}
A gas of charged particles is generally described by a Hamiltonian of the form
\begin{equation}
\calham H_M = \sum_j \frac{\vop p_j^2}{2m} + \calham H_{qq} + \calham H_{qQ} + \sum_j \hat v(\vop r_j) + \sum_{i \neq j} \hat V(\vop r_i, \vop r_j) \, .
\label{generalmatter1}
\end{equation}
Here, $\vop r_j$ and $\vop p_j$ are the canonical position and momentum operators of the $j$-th particle, such that $[\hat r_{i, \alpha}, \hat p_{j, \beta}] = i \placeholder{} \delta_{ij} \delta_{\alpha \beta}$. Note that $\hbar = 1$ throughout the text. We also include four potential terms. Coulombic interactions between charges, including self-interactions, are encapsulated in $\calham H_{qq}$. $\calham H_{qQ}$ contains possible interactions with non-dynamical charges, e.g. with an ionic lattice. The potentials $\hat v(\vop r_j)$ and $\hat V(\vop r_i, \vop r_j)$ represent additional unspecified single-particle and two-particle potentials, respectively.
This matter Hamiltonian can be summarized as $ \calham H_M = \calham T + \calham H_{qq} + \calham H_{qQ} + \calham V$. We are interested in the effects of coupling this material subsystem to the quantized electromagnetic field in a confined region of space, i.e. a cavity. To be consistent with the inclusion of Coulombic interactions, we will describe the coupling of the matter and light subsystems in the Coulomb gauge. This yields a Hamiltonian in minimal coupling format for the complete system
\begin{equation}
\calham H = \sum_j \frac{(\vop p_j - q \vop A (\vop r_j))^2}{2m} + \calham H'_{qq} + \calham H'_{qQ} + \calham V + \calham H_{\rm EM} - \frac{g \mu_B}{2} \sum_j \vop \sigma_j \vop B(\vop r_j) \,.
\label{eq:PauliHamiltonian}
\end{equation}
The effect of the coupling is fourfold. Most trivially, we must account for the energy of the electromagnetic field, hence $\calham H_{\rm EM}$. Secondly, the canonical momentum now differs from the mechanical momentum, this is apparent in the kinetic term, which now includes the vector potential $\vop A(\vop r_j)$ at each particles' position. The third and perhaps more subtle effect is the modification of the Coulombic interactions between charges $\calham H_{qq} \to \calham H'_{qq}$, $\calham H_{qQ} \to \calham H'_{qQ}$, which now include direct as well as image-charge mediated interactions \cite{debernardis2018, vuviks2014}. The latter arise as an electrostatic effect of imposing the boundary conditions of a cavity on the Coulomb potential. Finally, we include the Zeeman coupling between the particles' spin and the magnetic field $\vop B(\vop r_j)$. Here $\vop \sigma_j = (\hat \sigma_j^x, \hat \sigma_j^y, \hat \sigma_j^z)$ is the Pauli vector, whose components are the Pauli matrices, $\mu_B = \frac{ e \placeholder{}}{2 m}$ is the Bohr magneton and $g$ the Landé factor.
W.l.o.g. we are discussing spin $1/2$ particles.
In fact, Eq. \eqref{eq:PauliHamiltonian} is the many-particle generalization of the Pauli-Fiertz Hamiltonian for non-relativistic spin $1/2$ particles, bar the spin-orbit coupling.
In the Coulomb gauge, the redundancy in the description of the dynamical variables is eliminated by constraining the vector potential to be transverse, i.e. $\div{A} = 0$. The quantized vector potential then reads
\begin{equation}
\vop A(\vop r_j) = \sum_{\kappa} A_{\boldsymbol k} \left(\vop u_{\kappa}(\vop r_j)\hat a_{\kappa} + \vop u^*_{\kappa}(\vop r_j) \hat a_{\kappa}^\dagger \right) \,,
\label{eq:vectorpotential}
\end{equation}
where $\kappa$ is a four-vector containing the polarization index $\sigma = 1,2$ and the wavevector $\boldsymbol k$: $\kappa \equiv \{\boldsymbol k, \sigma\}$. Note that $A_{\boldsymbol k} = A_{-\boldsymbol k}$. We also introduce the bosonic annihilation and creation operators of the $\kappa$-th mode: $a_{\kappa}$, $a_{\kappa}^\dagger$, which obey the canonical commutation relations $[a_{\kappa},\hat a_{\kappa'}^\dagger] = \delta_{\kappa \kappa'}$. To maintain as much generality as possible, we have refrained from using a particular spatial dependence for the vector potential. For a specific model, the geometry of the cavity will determine the spatial quantization of the wavevector $\boldsymbol k$ and in turn, the functional form of the mode functions $u_{\kappa}(\boldsymbol r)$ \cite{kakazu1994}. In any case, the following properties hold for the mode functions: $u_{-\boldsymbol k, \sigma}(\boldsymbol r) = u_{\boldsymbol k, \sigma}^*(\boldsymbol r)$, $u_{\boldsymbol k, \sigma}(\boldsymbol r) \vdot \boldsymbol k = 0$, $\int_V dV u^*_{\kappa}(\boldsymbol r) u_{\kappa'}(\boldsymbol r) = \delta_{\kappa \kappa'}$. Note that in Eq. \eqref{eq:vectorpotential} the mode functions have been promoted to mode operators, since they depend on the position operator of each particle (Cf. App. \ref{ap:localizedemitters}). By definition, $\vop B = \curl{\vop A}$, so
\begin{equation}
\vop B(\vop r_j) = \sum_{\kappa} B_{\boldsymbol k} \left(\vop u_{\perp, \kappa} (\vop r_j) \hat a_{\kappa} + \vop u_{\perp, \kappa}^* (\vop r_j) \hat a_{\kappa}^\dagger \right) \,,
\label{eq:magneticfield}
\end{equation}
where we have defined the transverse mode functions $\boldsymbol u_{\perp, \kappa} (\boldsymbol r_j) = \frac{1}{|\boldsymbol k|} \curl{\boldsymbol u_{\kappa} (\boldsymbol r_j)}$ and $B_{\boldsymbol k} = A_{\boldsymbol k} |\boldsymbol k| = A_{\boldsymbol k} \omega_{\boldsymbol k} / c$.
Finally, we can express $\calham H_{\rm EM}$ as
\begin{equation}
\calham H_{\rm EM} = \sum_{\kappa} \placeholder{} \omega_{\boldsymbol k}\hat a_{\kappa}^\dagger\hat a_{\kappa} \,.
\end{equation}
To facilitate the analytical treatment of the Hamiltonian, it is convenient to expand the kinetic term and substitute in Eqs. \eqref{eq:vectorpotential} and \eqref{eq:magneticfield}.
\begin{equation}
\begin{split}
\calham{H} &= \calham H_M' + \sum_{\kappa} \placeholder{} \omega_{\boldsymbol k}\hat a_{\kappa}^\dagger\hat a_{\kappa}
- \sum_j \sum_\kappa \frac{q}{m} \vop p_j A_{\boldsymbol k} \left(\vop u_{\kappa}(\vop r_j)\hat a_{\kappa} + h.c. \right) \\
&+ \sum_j \frac{q^2}{2 m} \sum_{\kappa, \kappa'}A_{\boldsymbol k}A_{\boldsymbol k'} \left(\vop u_{\kappa}(\vop r_j)\hat a_{\kappa} + h.c. \right) \left(\vop u_{\kappa'}(\vop r_j)\hat a_{\kappa'} + h.c. \right)
-\frac{g \mu_B}{2} \sum_j \sum_k \vop \sigma_j B_{\boldsymbol k} \left(\vop u_{\perp, \kappa} (\vop r_j) \hat a_{\kappa} + h.c. \right)\,.
\end{split}
\end{equation}
Here, $\calham H_M'$ is just $\calham H_M$ but with $\calham H_{qq}$ ($\calham H_{qQ}$) replaced by $\calham H_{qq}'$ ($\calham H_{qQ}'$) To condense the notation, we define the operators
\begin{align}
& \hat U_{\kappa, \kappa'} = N^{-1} \sum_j \vop u_{\kappa}(\vop r_j) \vop u_{\kappa'}(\vop r_j) \,, \\
& \hat{\bar U}_{\kappa, \kappa'} = N^{-1} \sum_j \vop u_{\kappa}(\vop r_j) \vop u_{\kappa'}^\dagger(\vop r_j) \,,
\end{align}
and the constants
\begin{align}
&\Delta_{\boldsymbol k, \boldsymbol k'} = \frac{N q^2 A_{\boldsymbol k} A_{\boldsymbol k'}}{2 m \placeholder{}} \,, \\
& c_{\boldsymbol k}^e = q A_{\boldsymbol k} \sqrt{\frac{\omega_{\boldsymbol k}}{m \placeholder{}}} \,, \\
&c_{\boldsymbol k}^m = \frac{g \mu_B B_{\boldsymbol k}}{2 \placeholder{}} \,.
\end{align}
With these, we write the terms that depend quadratically on the bosonic operators as
\begin{equation}
\begin{split}
\calham{H_{\rm ph}} = & \sum_{\kappa} \placeholder{} \omega_{\boldsymbol k}\hat a_{\kappa}^\dagger\hat a_{\kappa}
+\sum_{\kappa, \kappa'} \placeholder{} \Delta_{\boldsymbol k, \boldsymbol k'} \left(\hat U_{\kappa, \kappa'}\hat a_\kappa\hat a_{\kappa'} + \hat{\bar U}_{\kappa, \kappa'}\hat a_\kappa\hat a_{\kappa'}^\dagger + h.c. \right) \,.
\end{split}
\end{equation}
Analytical tractability demands that we neglect mode-mode interactions (Cf. App. \ref{ap:localizedemitters}) and assume $\hat U_{\boldsymbol k, \sigma, -\boldsymbol k, \sigma} = \hat{\bar U}_{\boldsymbol k, \sigma, \boldsymbol k, \sigma} = \mathbb I$, leaving only the momentum-conserving terms, and without spatial dependence
\begin{equation}
\begin{split}
\calham{H_{\rm ph}} \approx & \sum_{\boldsymbol k, \sigma} \placeholder{} \omega_{\boldsymbol k}\hat a_{\boldsymbol k, \sigma}^\dagger\hat a_{\boldsymbol k, \sigma}
+ \sum_{\kappa} \placeholder{} \Delta_{\boldsymbol k} \left(\hat a_{\boldsymbol k, \sigma} \hat a_{-\boldsymbol k, \sigma} + \hat a_{\boldsymbol k, \sigma} \hat a_{\boldsymbol k, \sigma}^\dagger + h.c. \right) \,,
\end{split}
\end{equation}
where $\Delta_{\boldsymbol k} \equiv \Delta_{\boldsymbol k, \boldsymbol k}$. The approximation is exact for a homogeneous gas. Furthermore, for gasses that have a second order phase transition to a non-homogeneous phase, the approximation is valid in the vicinity of the transition \cite{andolina2020}. This suggests that the resulting model can be used to predict and locate transitions to non-homogeneous gas phases. It is expected that the effect of the approximation will be quantitative and will skew the calculation of observables in the non-homogeneous phase. $\calham H_{\rm ph}$ can now be diagonalized with a Bogoliubov transformation
\begin{equation}
\hat a_{\boldsymbol k, \sigma}^\dagger = \cosh (\theta_{\boldsymbol k}) \hat b_{\boldsymbol k, \sigma}^\dagger - \sinh (\theta_{\boldsymbol k}) \hat b_{-\boldsymbol k, \sigma} \,.
\label{eq:bogoliubov}
\end{equation}
One finds that $\cosh \left(\theta_{\boldsymbol k}\right)=\left(\lambda_{\boldsymbol k}+1\right) /(2\sqrt{\lambda_{\boldsymbol k}})$ and $\sinh \left(\theta_{\boldsymbol k}\right)=\left(\lambda_{\boldsymbol k}-1\right) /(2\sqrt{\lambda_{\boldsymbol k}})$ and $\lambda_{\boldsymbol k} = \sqrt{1 + 4\Delta_{\boldsymbol k} / {\omega_{\boldsymbol k}}}$. As a result, the complete Hamiltonian can be written as
\begin{equation}
\begin{split}
\calham{H} = &\calham H_M' + \sum_{\kappa} \placeholder{} \omega_{\boldsymbol k} \lambda_{\boldsymbol k} \hat b_{\kappa}^\dagger\hat b_{\kappa}
-\sum_j \sum_\kappa \placeholder{} c_{\boldsymbol k}^e \lambda_{\boldsymbol k}^{-1/2} \sqrt{\frac{1}{m \placeholder{} \omega_{\boldsymbol k}}} \vop p_j \left(\vop u_{\kappa}(\vop r_j)\hat b_{\kappa} + h.c. \right)
-\sum_j \sum_k \placeholder{} c_{\boldsymbol k}^m \lambda_{\boldsymbol k}^{-1/2} \vop \sigma_j \left(\vop u_{\perp, \kappa} (\vop r_j) \hat b_{\kappa} + h.c. \right)\,.
\end{split}
\end{equation}
Defining the coupling operator and coupling constant
\begin{equation}
c_{\boldsymbol k} \calham C_\kappa = \sum_j \lambda_{\boldsymbol k}^{-1/2} \left(c_{\boldsymbol k}^e \sqrt{\frac{1}{m \placeholder{} \omega_{\boldsymbol k}}} \vop p_j \vop u_{\kappa}(\vop r_j) + c_{\boldsymbol k}^m \vop \sigma_j \vop u_{\perp, \kappa} (\vop r_j)\right) \,,
\end{equation}
we can finally write the Hamiltonian as
\begin{equation}
\calham{H} = \calham H_M' + \sum_{\kappa} \placeholder{} \tilde \omega_{\boldsymbol k} \hat b_{\kappa}^\dagger\hat b_{\kappa}
- \sum_\kappa \placeholder{} c_{\boldsymbol k} \left(\calham C_\kappa \hat b_{\kappa} + h.c. \right) \,,
\label{eq:linearlmH}
\end{equation}
where $\tilde \omega_{\boldsymbol k} = \omega_{\boldsymbol k} \lambda_{\boldsymbol k}$. \\
\subsection{A system of localized emitters}
The main difference between a gas and a system of localized emitters relies on the fact that in the former case the dynamical particles are free to move on the entire volume of the system, whereas in the latter their movement is bounded to a non-dynamical core. A limit case is given by systems in which no movement occurs, and instead the dynamics affect some other degree of freedom, this is the case of spins of fixed positions, such as the ones that compose a magnetic crystal. If the length scale of the bounded motion is much smaller than the wavelength of the EM field, one can consider that the EM field is constant and equal at any point in the trajectory of an individual particle $\vop A (\vop r_j + \boldsymbol R_j) \approx \vop A ( \boldsymbol R_j)$: where $\boldsymbol R_j$ is the position of the corresponding non-dynamical core. This is the long-wavelength approximation. Assuming that we can work within the long-wavelength limit, the EM field operators $\vop A$, $\vop B$ act only on the Hilbert space of the photons, i.e. they do not contain position operators acting on the Hilbert space of the matter. Instead, they are parametrized by the positions of the non-dynamical cores. This subtle difference implies that, unlike for the gas, the terms in the Hamiltonian quadratic in bosonic operators can be exactly diagonalized by a Bogoliubov transformation. The final Hamiltonian \eqref{eq:linearlmH} (See Fig. \ref{fig:visualabstract}(b)) has the same functional form in both cases, the differences are encoded in the coupling constants $\{c_{\boldsymbol k}\}$ and coupling operators $\{\calham C_{\kappa}\}$, and of course in $\calham H_M$. For a detailed derivation on the case of localized emitters see App. \ref{ap:localizedemitters}.
Before advancing, we want to emphasize that Hamiltonian \eqref{eq:linearlmH} or the equivalent \eqref{eq:linearlmH2} represents a general light-matter Hamiltonian. In this Section (Appendix \ref{ap:localizedemitters}), we arrived to it by considering the particular case of a gauge-invariant description of a gas of charged particles (an ensemble of localized emitters) coupled to the multi-mode electromagnetic field of a cavity, but this approach is not unique. For instance, the Zeeman coupling to the magnetic field is often neglected and the description of the EM field is simplified by making the single-mode approximation. Neglecting the $A^2$ term is also a ubiquitous approximation. The important fact is that all these variations arise as particularizations of Hamiltonian \eqref{eq:linearlmH}. More over, systems of magnetic molecules where the orbital degrees of freedom can be disregarded, which have received attention for the possibility of hosting the superradiant phase transition \cite{romanroche2021}, can also be described by this Hamiltonian. In this case, the matter Hamiltonian $\calham H_M'$ and the coupling operator $\calham C_{\boldsymbol k}$ would exclusively contain spin degrees of freedom. In these examples, we simplify the picture by considering only cavity fields, but the matter DOFs can be coupled to classical external fields as well. For instance, it is common in systems of magnetic molecules to induce Zeeman level splitting with an external magnetic field perpendicular to the microwave magnetic field of the cavity. The inclusion of a classical field is trivial and indeed will be considered in Sec. \ref{sec:examples} when we cover magnetic cavity QED.
We have discussed some relevant examples for cavity QED and quantum optics, but in general, any matter system which is linearly coupled to light and is quadratic in the photonic operators can have its Hamiltonian brought to Eq. \eqref{eq:linearlmH}. In essence, the formalism presented hereinafter is just a recipe to systematically eliminate bosonic degrees of freedom in a hybrid system. Furthermore, we show that in the large-$N$ limit, which is typically the relevant one in cavity QED materials, the effective action becomes local and the effective description for the matter is exactly Hamiltonian. Thus eliminating the need for mean field or perturbative approximations.
\section{Deriving the effective matter model}
\label{sec:effectivemodel}
Our goal in this section is to take a Hamiltonian of the form of Eq. \eqref{eq:linearlmH} and trace out the photonic degrees of freedom. To do so, we will use a coherent path integral approach to define an effective matter action. One can foresee that the light-matter coupling present in \eqref{eq:linearlmH} will translate into an effective matter-matter coupling that will appear in the effective action. Then, we obtain an effective Hamiltonian from the effective action in the, $N \to \infty$, thermodynamic limit.
\subsection{Effective action: Euclidean path integral formulation}
Let us recall that the partition function of a single-mode bosonic system described by a Hamiltonian $\calham H$ can be written in Euclidean path integral form as
\begin{equation}
Z = \int \mathcal D (z, \bar z) e^{-S} \,,
\end{equation}
with Euclidean action
\begin{equation}
S = \int_{0}^{\placeholder{} \beta} \left(\placeholder{} \bar z \partial_u z + H(z, \bar z)\right) du \,,
\end{equation}
where $z$, $\bar z$ are complex-valued continous fields satisfying the boundary condition $\bar z(0) = \bar z(\beta)$, $z(0) = z(\beta)$ and $\mathcal D (z, \bar z) = \lim_{N \to \infty} \prod_{n = 1}^N d(z^n, \bar z^n)$, with $d(z^n, \bar z^n) = d\Re z^n d\Im z^n$ \cite[chapter 4]{altland2010}. Here $H(z, \bar z)$ is obtained by replacing $a$ by $z$ and $a^\dagger$ by $\bar z$ in the normal-ordered form of $\calham H$, where $a$($a^\dagger$) are bosonic annihilation (creation) operators.
This result is readily generalized to the multimode case
\begin{equation}
Z = \prod_\kappa \int \mathcal D (z_\kappa, \bar z_\kappa) e^{- S_{\kappa}}
\label{eq:multimodeZ}
\end{equation}
with
\begin{equation}
S_{\kappa} = \int_{0}^{\placeholder{} \beta} \left(\placeholder{} \bar z_\kappa \partial_u z_\kappa + H_\kappa(z_\kappa, \bar z_\kappa)\right) du \,.
\label{eq:multimodeS}
\end{equation}
We now focus our attention on a hybrid light-matter system such as the one described by Hamiltonian \eqref{eq:linearlmH}. With $\ket{\boldsymbol z} = \bigotimes_\kappa \ket{z_\kappa}$, its partition function is given by
\begin{equation}
Z = \Tr_M \left(\int \prod_\kappa \frac{d^2 z_\kappa}{\pi} \bra{\boldsymbol z} e^{-\beta \calham H} \ket{\boldsymbol z} \right) \,.
\label{eq:originalZ}
\end{equation}
We assume that one could formulate a coherent-path-integral partition function for the full system, so we assign an action to the matter Hamiltonian $\calham H_M' \to S_M$ and a field to the coupling operator $\calham C_k \to C_k$, $\calham C_k^\dagger \to \bar C_k$. Knowledge of the precise dependence of $S_M$, $C_k$ and $\bar C_k$ on material coherent fields is not required for what follows. With these definitions, the total action can be written as $S = S_M + \sum_\kappa S_\kappa$, with
\begin{equation}
S_\kappa = \int_{0}^{\placeholder{} \beta} \left( \placeholder{} \bar z_\kappa \partial_u z_\kappa + \placeholder{} \tilde \omega_{\boldsymbol k} \bar z_\kappa z_\kappa
-\placeholder{} c_{\boldsymbol k} \left(C_\kappa z_{\kappa} + \bar C_\kappa \bar z_{\kappa}\right) \right) du \,,
\label{eq:modeaction}
\end{equation}
For such a system, we can define an effective matter action by tracing only over the light degrees of freedom
\begin{equation}
Z = \Tr_M\left(e^{-S_M} \prod_\kappa \int \mathcal D (z_\kappa, \bar z_\kappa) e^{-S_{\kappa}} \right) =: Z_0 \Tr_M\left(e^{-S_{\rm eff}} \right) \,.
\label{eq:effactiondef}
\end{equation}
At this point, $Z_0$ is just a constant prefactor.
It should be noted that $S_{\rm eff}$ is useful because it makes explicit the influence of the cavity on the matter DOFs. In particular, the type of interactions that emerge. In addition, as we show in Sec. \ref{sec:relation}, no information is lost because light observables can be written in terms of matter observables, which can be calculated from $S_{\rm eff}$.
We are thus interested in computing
\begin{equation}
\prod_\kappa \int \mathcal D (z_\kappa, \bar z_\kappa) e^{-S_{\kappa}} \,.
\label{eq:functionalintegral}
\end{equation}
Since the trajectories are periodic $z(0) = z(\placeholder{} \beta)$, $\bar z(0) = \bar z(\placeholder{} \beta)$, we can expand the fields as a Fourier series
\begin{align}
&z_\kappa(u) = \frac{1}{\sqrt{\placeholder{} \beta}} \sum_n z_\kappa(\omega_n) e^{i \omega_n u} \,, \label{eq:zfourier} \\
&\bar z_\kappa(u) = \frac{1}{\sqrt{\placeholder{} \beta}} \sum_n \bar z_\kappa(\omega_n) e^{-i \omega_n u} \,.
\label{eq:barzfourier}
\end{align}
with the Matsubara frequencies $\omega_n = \frac{2 \pi n}{\placeholder{} \beta}$. Inserting Eqs. \eqref{eq:zfourier}, \eqref{eq:barzfourier} in the action \eqref{eq:modeaction} yields
\begin{equation}
S_\kappa = \sum_n \left( \placeholder{} \left(i \omega_n + \tilde \omega_{\boldsymbol k}\right) \bar z_\kappa(\omega_n) z_\kappa(\omega_n) - \placeholder{} c_{\boldsymbol k} z_\kappa(\omega_n) C_\kappa(\omega_n) - \placeholder{} c_{\boldsymbol k} \bar z_\kappa(\omega_n) \bar C_\kappa(\omega_n) \right) \,,
\end{equation}
where
\begin{align}
&C_\kappa(\omega_n) =\frac{1}{ \sqrt{\placeholder{} \beta}} \int_{0}^{\placeholder{} \beta} du C_\kappa(u) e^{i \omega_n u} \,, \\
&\bar C_\kappa(\omega_n) = \frac{1}{\sqrt{\placeholder{} \beta}} \int_{0}^{\placeholder{} \beta} du \bar C_\kappa(u) e^{-i \omega_n u} \,.
\end{align}
The Jacobian of the Fourier transformation is unity, so the functional differential over temperature trajectories in Eq. \eqref{eq:functionalintegral} is simply replaced with the differential over frequency trajectories. The result is simply a collection of $n$-dimensional Gaussian integrals for each mode $\kappa$, yielding
\begin{equation}
\prod_\kappa \int \mathcal D (z_\kappa, \bar z_\kappa) e^{-S_{\kappa}} = \prod_\kappa Z_\kappa \exp \left[ \frac{c_{\boldsymbol k}^2}{\placeholder{} \beta \tilde \omega_{\boldsymbol k}} \placeholder{} \iint_{0}^{\placeholder{} \beta} du du' C_\kappa(u) \bar C_\kappa(u') K_\kappa(u - u') \right] \,,
\end{equation}
with the kernel
\begin{equation}
K_\kappa(u - u') = \sum_n \frac{\tilde \omega_{\boldsymbol k}}{i \omega_n + \tilde \omega_{\boldsymbol k}} e^{i \omega_n (u - u')} \,.
\label{eq:nonlocalkernel}
\end{equation}
At this point, the prefactor $Z_0$ defined in Eq. \eqref{eq:effactiondef} acquires meaning: $Z_0 = \prod_\kappa Z_\kappa$, where $Z_\kappa = \left[1 - \exp(-\beta \placeholder{} \tilde \omega_{\boldsymbol k})\right]^{-1} $ is the partition function of a Harmonic oscillator of frequency $\tilde \omega_{\boldsymbol k}$. We can now put together the effective action as defined in Eq. \eqref{eq:effactiondef}
\begin{equation}
S_{\rm eff} = S_M - \sum_\kappa \frac{c_{\boldsymbol k}^2}{\placeholder{} \beta \tilde \omega_{\boldsymbol k}} \placeholder{} \iint_{0}^{\placeholder{} \beta} du du' C_\kappa(u) \bar C_\kappa(u') K_\kappa(u - u') \,.
\label{eq:nonlocalaction}
\end{equation}
After tracing out the light degrees of freedom, the light-matter interaction appears translated into a temperature-non-local interaction between matter degrees of freedom. The elimination of the cavity allows us to study the modification of material properties from the point of view of the matter subsystem. However, the non locality of the effective interaction implies that Eq. \ref{eq:nonlocalaction} does not have straightforward analytical applicability. In particular, we cannot recover an equivalent Hamiltonian. That is, we cannot undo our agnostic substitutions $\calham H_M' \to S_M$, $\calham C_k \to C_k$ and $\calham C_k^\dagger \to \bar C_k$, to revert the description of the matter subsystem from an action-based one to a Hamiltonian-based one. This is because not all fields depend on the same temperature in this non-local action.
This should not come as a surprise, since in classical statistical mechanics when integrating a subpart of the whole system the magnitudes that involve the remaining DOFs can be calculated from a temperature-dependent effective Hamiltonian \cite{Alonso2010}. Equivalently, in bipartite quantum systems, the (temperature dependent) Hamiltonian of mean force describes the equilibrium properties of a partition \cite{Campisi2009, Campisi2010}.
On the other hand, \eqref{eq:nonlocalaction} is formally analogous to the effective action after tracing out the environment DOFs in the path integral formulation of open quantum systems \cite{grabert1988quantum}.
What is less intuitive is what we show in the next subsection, namely that the kernel becomes local in the thermodynamic limit $N \to \infty$, recovering a time-local effective interaction with a clear Hamiltonian equivalent.
\subsection{Obtaining an effective matter Hamiltonian in the large-$N$ limit}
The effective action \eqref{eq:nonlocalaction} only depends on $N$ through the scaling of the involved fields, which they inherit from the corresponding operators. We assume that the light-matter Hamiltonian \eqref{eq:linearlmH} is not ill-defined in the thermodynamic limit, i.e. that all its terms scale as $N$. Otherwise some of them would become negligible for $N \to \infty$, rendering the system exactly solvable and, thus, of little analytical interest. It is also important to note that in order to have a macroscopically well-defined thermodynamic limit the density of particles in the material $\rho = N/V $ must be finite, and thus the volume $V$ must scale like the number of particles $N$, i.e. $V \to \infty$ in the thermodynamic limit. This implies that $b$, $b^\dagger$ scale as $\sqrt{N}$ and $c_{\boldsymbol k}$ scales as $1/\sqrt{N}$, with both $\calham H_M'$ and $\calham C_\kappa$ scaling as $N$.
To understand how a large $N$ affects the effective action it is best to exploit the scaling properties of the photonic operators before they are traced out. This is done following the reasoning of Wang and Hioe \cite{Wang1973}. Because the photonic operators scale as $\sqrt{N}$, one can define rescaled operators $b / \sqrt{N}$, $b^\dagger / \sqrt{N}$ whose $N \to \infty$ limit is well-defined and whose commutator scales as $1/N$, and thus vanishes in the large-$N$ limit. This allows one to write
\begin{equation}
\int \prod_\kappa \frac{d^2 z_\kappa}{\pi} \bra{\boldsymbol z} e^{-\beta \calham H} \ket{\boldsymbol z} = \int \prod_\kappa \frac{d^2 z_\kappa}{\pi} e^{-\beta \expval{\calham H}{\boldsymbol z}} \,.
\label{eq:wangsrelation}
\end{equation}
Defining
\begin{equation}
\tilde Z = \Tr_M \left( \int \prod_\kappa \frac{d^2 z_\kappa}{\pi} e^{-\beta \expval{\calham H}{\boldsymbol z}} \right) \,,
\label{eq:Zlowerbound}
\end{equation}
with $\ket{\boldsymbol z} = \bigotimes_\kappa \ket{z_\kappa}$ and
\begin{equation}
\expval{\calham H}{\boldsymbol z} = \calham H_M' + \sum_{\kappa} \placeholder{} \tilde \omega_{\boldsymbol k} |z_{\kappa}|^2
+ \sum_\kappa \placeholder{} c_{\boldsymbol k} \left(\calham C_\kappa z_{\kappa} + h.c. \right) \,,
\end{equation}
we note that taking the trace over matter of the lhs and rhs of Eq. \eqref{eq:wangsrelation} yields the equality $Z = \tilde Z$ (Cf. \eqref{eq:originalZ}). The same equality can be obtained from the bounds derived by Hepp and Lieb in their original study of equilibrium superradiance \cite{Hepp1973a}
\begin{equation}
\tilde Z \leq Z \leq e^{\beta \sum_\kappa \placeholder{} \omega_{\boldsymbol k}} \tilde Z \,,
\label{eq:heppliebbounds}
\end{equation}
In the thermodynamic limit $N \to \infty$ and provided the number of modes does not scale with $N$, the correction term $\exp \left[ \beta \sum_\kappa \placeholder{} \omega_{\boldsymbol k} \right]$ in the upper bound becomes negligible and the partition function is given exactly by $Z = \tilde Z$. With this equality at hand, we backtrack to Eqs. \eqref{eq:modeaction} and \eqref{eq:effactiondef} and note that if one assumes constant photon fields, the path integral becomes a Gaussian integral over the initial states and Eq. \eqref{eq:effactiondef} becomes precisely $\tilde Z$. Thus, in the path integral formulation, the large-$N$ limit implies constant paths for the photons. This resembles the classical limit in the coherent state path integral formulation of fields \cite[chapter 2]{kleinert2009path}. We may enforce constant fields at any point in the derivation of the effective theory. In particular, we may fulfill our initial desire of finding the large-$N$ limit of the final effective action \eqref{eq:nonlocalaction}. To do so, it is important to realize that the constant field condition equates to truncating the Matsubara summation to the $n=0$ mode or, equivalently, to setting $\omega_n = 0$ for all $n$. This realization will allow us to enforce the consequences of the large-$N$ limit on the non-local kernel \eqref{eq:nonlocalkernel}. Before doing so, it is convenient to do some manipulations
\begin{equation}
K_\kappa(\tau) = \sum_n \frac{\tilde \omega_{\boldsymbol k}^2 e^{i \omega_n \tau}}{\omega_n^2 + \tilde \omega_{\boldsymbol k}^2} - \sum_n \frac{i \omega_n \tilde \omega_{\boldsymbol k} e^{i \omega_n \tau}}{\omega_n^2 + \tilde \omega_{\boldsymbol k}^2} = \bar K_\kappa(\tau) - \frac{1}{\tilde \omega_{\boldsymbol k}} \partial_\tau \bar K_\kappa(\tau) \,.
\label{eq:kernelsplit}
\end{equation}
The first term can be split into a collection of delta functions and another kernel
\begin{equation}
\bar K_\kappa(\tau) = \sum_n e^{i \omega_n \tau} - \sum_n \frac{\omega_n^2 e^{i \omega_n \tau}}{\omega_n^2 + \tilde \omega_{\boldsymbol k}^2} = \beta \sum_n \delta(\tau - \beta n) - \bar{\bar K}_\kappa(\tau) \,.
\end{equation}
The delta functions provide the local action that we seek when integrating over the interval $[0, \beta]$. Crucially, $\bar{\bar K}_\kappa(\tau) = 0$ in the large-$N$ limit. To see this, recall that the large-$N$ limit equates to restricting to constant photonic fields, which in turn equates to setting $\omega_n = 0$ for all $n$. Returning to the second term in \eqref{eq:kernelsplit}, because it is odd in $\tau$, it can be seen that it does not contribute to the action when the matter coupling operators are Hermitian and thus $C_\kappa = \bar C_\kappa$. In a more general case, it can contribute. Regardless, in the large-$N$ limit, it vanishes for the same reason as $\bar{\bar K}_\kappa$ and the action is purely local.
So in the thermodynamic limit the effective action becomes local in time
\begin{equation}
S_{\rm eff} = S_M - \sum_\kappa N \placeholder{} \frac{c_{\boldsymbol k}^2}{\tilde \omega_{\boldsymbol k}} \int_{0}^{\frac{\placeholder{} \beta}{N}} dv C_\kappa(v) \bar C_\kappa(v) = S_M - \sum_\kappa \placeholder{} \frac{c_{\boldsymbol k}^2}{\tilde \omega_{\boldsymbol k}} \int_{0}^{\placeholder{} \beta} du C_\kappa(u) \bar C_\kappa(u) \,.
\label{eq:localaction}
\end{equation}
Accordingly, we can recover the corresponding effective Hamiltonian
\begin{equation}
\calham H_{\rm eff} = \calham H_M' - \sum_\kappa \placeholder{} \frac{c_{\boldsymbol k}^2}{\tilde \omega_{\boldsymbol k}} \calham C_k \calham C_k^\dagger \,,
\label{eq:effectiveham}
\end{equation}
such that $Z = Z_0 \Tr_M\left[\exp(- \beta \calham H_{\rm eff})\right]$.
An alternative derivation of the effective Hamiltonian can be found in App. \ref{ap:Hamderivation}. Note also that what we show here is just a particular way of deriving the effective Hamiltonian, but constant photon paths can be enforced at any point along the derivation of the effective action, e.g. at the level of the kernel but before extracting the Dirac deltas. In those cases --the calculation may involve the use of the Hubbard–Stratonovich transformation-- one always ends up with a Gaussian integral over photonic DOFs akin to the one discussed in App. \ref{ap:Hamderivation} which also leads to the same effective Hamiltonian.
Equation \eqref{eq:effectiveham} shows that in the large-$N$ limit the effect of the cavity can be described via an effective Hamiltonian in which the induced interactions are explicit (second term on the r.h.s. of Eq. \eqref{eq:effectiveham}). Notice that sometimes similar effective interactions are argued by invoking approximations, such as perturbation or mean field theory. Here we show that they are exact in the thermodynamic limit for the matter.
Besides, it is a rather general result. We emphasize that the cavity-matter coupling has been discussed under general grounds in Section \ref{sec:model}, and Eq. \eqref{eq:effectiveham} follows from it.
\subsection{Relation between matter and light observables}
\label{sec:relation}
Our effective theory puts emphasis on the behaviour of the matter subsystem, to the point where the cavity is eliminated from the Hamiltonian. However, for a complete characterization of the full system one must be able to compute photonic fields, $\expval{b_\kappa}$, $\expval{b_\kappa^\dagger}$, and the photon number which determine the intensity of the electromagnetic fields in the cavity. We reconcile this neccesity with our formalism by noting that in the process of tracing out the photonic degress of freedom we can extract a relationship between the expectation values of matter and light observables. To illustrate the idea, let us work out the case of computing $\langle b_{\kappa'} + b_{\kappa'}^\dagger \rangle$. We extend our Hamiltonian to the form $\calham H(\lambda) = \calham H - \lambda (b_{\kappa'} + b_{\kappa'}^\dagger)$, such that we can write, as is customary in statistical mechanics, $\beta \langle b_{\kappa'} + b_{\kappa'}^\dagger \rangle = \lim_{\lambda \to 0} \partial_\lambda \ln Z(\lambda)$, with $Z(\lambda) = \Tr(\exp[-\beta \calham H(\lambda)])$. At the same time, we can rewrite $\calham H(\lambda)$ as
\begin{equation}
\calham{H}(\lambda) = \calham H_M' + \sum_{\kappa} \placeholder{} \tilde \omega_{\boldsymbol k} \hat b_{\kappa}^\dagger\hat b_{\kappa}
- \sum_\kappa \placeholder{} c_{\boldsymbol k} \left(\calham C_\kappa (\lambda) \hat b_{\kappa} + h.c. \right) \,,
\end{equation}
where $\calham C_\kappa (\lambda) = \calham C_\kappa - \delta_{\kappa, \kappa'} \lambda/ c_{\boldsymbol k}$. Following our theory, the corresponding effective Hamiltonian is given by
\begin{equation}
\calham H_{\rm eff} (\lambda) = \calham H_M' - \sum_\kappa \placeholder{} \frac{c_{\boldsymbol k}^2}{\tilde \omega_{\boldsymbol k}} \calham C_\kappa (\lambda) \calham C_\kappa^\dagger (\lambda) \,.
\end{equation}
Then, following our definition of the effective action \eqref{eq:effectiveham}, we can express the partition function as $Z(\lambda) = Z_0 \Tr_M \left(\exp[-\beta \calham H_{\rm eff}(\lambda)]\right)$, such that
\begin{equation}
\expval{b_{\kappa'} + b_{\kappa'}^\dagger} = \lim_{\lambda \to 0} Z(\lambda)^{-1} Z_0 \Tr_M \left[- \frac{c_{\boldsymbol k'}}{\tilde \omega_{\boldsymbol k'}} \left( \calham C_{\kappa'} (\lambda) + \calham C_{\kappa'}^\dagger (\lambda)\right) e^{-\beta \calham H_{\rm eff}(\lambda)} \right] = -\frac{c_{\boldsymbol k'}}{\tilde \omega_{\boldsymbol k'}} \expval{\calham C_{\kappa'} + \calham C_{\kappa'}^\dagger}_M \,.
\end{equation}
Proceeding in similar fashion, we can compute other observables such as
\begin{equation}
\expval{b_{\kappa'} - b_{\kappa'}^\dagger} = \frac{c_{\boldsymbol k'}}{\tilde \omega_{\boldsymbol k'}} \expval{\calham C_{\kappa'} - \calham C_{\kappa'}^\dagger}_M \,.
\end{equation}
Another observable of interest is the photon number of the $\kappa$-th mode
\begin{equation}
\expval{ b_{\kappa'}^\dagger b_{\kappa'} } = -\frac{1}{\placeholder{} \beta} \partial_{\tilde \omega_{\boldsymbol k'}} \ln Z = n_B(\beta, \tilde \omega_{\boldsymbol k'}) + \left(\frac{c_{\boldsymbol k'}}{\tilde \omega_{\boldsymbol k'}} \right)^2 \expval{\calham C_{\kappa'} \calham C_{\kappa'}^\dagger}_M \,,
\label{eq:relationobs}
\end{equation}
where $n_B$ is the bosonic occupation number of a free Harmonic oscillator
\begin{equation}
n_B(\beta, \tilde \omega_{\boldsymbol k'}) = \frac{1}{e^{\beta \placeholder{} \tilde \omega_{\boldsymbol k'}} - 1} \,.
\end{equation}
We find that the photon number has two sources at finite temperature, one corresponding to the thermal fluctuations of the collection of free Harmonic oscillators that constitute the cavity, and another arising from the coupling to matter.
These relations may be important for computing measurable outputs of a cavity QED material. For example, if we probe the cavity-matter system through the transmissivity of the former, the transmitted signal is proportional to the dynamical response of the system. This is in turn related to two point correlators of bosonic fields that can be directly computed from relation \eqref{eq:relationobs} or analogues.
\section{Testing the validity of the effective theory}
\label{sec:test}
Effective theories are a common tool in theoretical physics. To contextualize ours, we compare it, analytically and numerically, against other common ones in the fields of condensed matter and quantum optics. Importantly, we also perform a numerical analysis of its finite-size scaling. Because other effective theories are model dependent and to facilitate the numerical analysis, our testbed here will be the Dicke model,
\begin{equation}
\calham H = \frac{\omega_z}{2} \sum_j \hat \sigma_j^z + \omega_c \hat a^\dagger \hat a + \frac{\lambda}{\sqrt{N}} \sum_j \hat \sigma_j^x \left( \hat a + \hat a^\dagger \right) \,.
\end{equation}
In the thermodynamic limit, it is exactly solvable, with the free energy per site given by \cite{hioe1973phase}
\begin{equation}
\begin{dcases}
-\beta f = \ln \left[ 2 \cosh( \beta \frac{\omega_z}{2}) \right] & \text{if} \quad \lambda < \lambda_c \quad \text{or} \quad \beta > \beta_c \,, \\
-\beta f = \ln \left[2 \cosh \left(\beta \frac{4 \lambda^2}{\omega_c} \sigma \right)\right]-\beta \frac{4 \lambda^2}{\omega_c} \sigma^{2}+\beta \frac{\omega_c \omega_z^{2}}{16 \lambda^2} & \text{if} \quad \lambda > \lambda_c \quad \text{and} \quad \beta < \beta_c \,,
\end{dcases}
\end{equation}
where $\lambda_c$ is given by $4 \lambda_c^2 = \omega_c \omega_z$, $\beta_c$ is given by $\omega_c \omega_z = 4 \lambda^2 \tanh(\beta_c \omega_z / 2)$ provided $\lambda > \lambda_c$ and $\sigma$ is found by solving $2 \sigma = \tanh(4 \beta \lambda^2 \sigma / \omega_c) \neq 0$.
The Dicke model, besides being a well-known toy model, has the added benefit of being numerically amenable, by virtue of conserving the total spin $\vop S^2 = \hat S_x^2 + \hat S_y^2 + \hat S_z^2$. Where $\hat S_\nu = \frac{1}{2} \sum_j \hat \sigma_j^\nu$. The numerical diagonalization can be done on each spin $S$ sector separately, with the total partition function given by \cite{Hepp1973}
\begin{equation}
Z = \sum_{S=s_0}^{N/2} \Omega(S, N) Z_S \,,
\end{equation}
where $s_0=0$ ($1/2$) if $N$ is even (odd) and
\begin{equation}
\Omega(S, N)=\frac{N !(2 S+1)}{\left(N/2-S\right) !\left(N/2+S+1\right) !} \,.
\end{equation}
Using total spin operators, the Hamiltonian can be rewritten as
\begin{equation}
\calham H = \omega_z \hat S_z + \omega_c \hat a^\dagger a + 2 g \hat S_x \left( \hat a + \hat a^\dagger \right) \,,
\end{equation}
With $g = \lambda / \sqrt{N}$. The corresponding effective Hamiltonian, according to our theory, reads
\begin{equation}
\calham H_{\rm eff} = \omega_z \hat S_z - \frac{4 g^2}{\omega_c} \hat S_x^2 \,.
\label{eq:Heffective}
\end{equation}
A useful tool to facilitate the numerical diagonalization of the full Dicke model is the Polaron transformation.
The Polaron transformation is given by $\hat U_{\rm P} = \exp(- \hat \alpha \hat S_x)$ with $\hat \alpha = 2 \zeta (\hat a^\dagger - \hat a)$ and $\zeta = g/\omega_c$. The resulting effective Hamiltonian $\calham H_{\rm P} = \hat U_{\rm P}^\dagger \calham H \hat U_{\rm P}$ reads
\begin{equation}
\calham H_{\rm P} = \omega_z \left(\hat S_z \cosh \hat \alpha -i \hat S_y \sinh \hat \alpha\right) + \omega_c \hat a^\dagger a - \frac{4 g^2}{\omega_c}\hat S_x^2 \,.
\label{eq:Hpolaron}
\end{equation}
The Polaron frame is particularly well-suited for exact diagonalization because the displacement of the bosons by $\hat S_x$ cancels the finite expectation value of $\hat a$, $\hat a^\dagger$ in the ordered phase. As a result, the number of bosons in the Polaron frame is only dependent on the temperature and the bosonic state-space can be truncated to a much smaller excitation cutoff than would otherwise be required \cite{Pilar2020, qinghu2008numerically}. This, paired with the conservation of the total spin, allows one to reach system sizes of $N \sim 100$. As we show below, this is effectively "close" to the thermodynamic limit. For this reason, we use the Hamiltonian in the Polaron frame \eqref{eq:Hpolaron} to do exact diagonalization of the full Dicke model.
\subsection{Comparison with other effective theories}
First, let us discuss the taxonomy of effective theories. We can distinguish two families: Hamiltonian theories, with unitary dynamics; and descriptions based on master equations and non-Hermitian Hamiltonians, with dissipative dynamics \cite{BRE02}. Because our theory belongs to the former, we will only consider other Hamiltonian theories in the comparison.
We can also distinguish between perturbative and non-perturbative theories. In this regard, it must be noted that our effective Hamiltonian was derived in the $N \to \infty$ limit. As such, it can be regarded as a zeroth order perturbation theory with $1/N$ as the perturbative parameter. The extension to a full-fledged perturbation theory that can be taken to any order in $1/N$ is a work in progress. However, in this context we will use the term perturbative to refer only to parameters that affect the light-matter coupling regime, such as the light-matter coupling strength or the detuning between the cavity frequency and the characteristic energy scale of the matter subsytem. It is in this sense that our theory is non-perturbative. The simplest example of a perturbative theory would be standard perturbation theory on the light-matter coupling. To yield an effective Hamiltonian with perturbation theory, a separation of energy scales (large detuning) is required, with the more energetic sector of the spectrum mediating effective interactions on a low-energy subspace \cite{bir1974symmetry, Cohen-Tannoudji2020-kh}. A less cumbersome way to obtain the same effective description is by means of a Schrieffer–Wolff unitary transformation \cite{klimov2002effective}. We will thus use the latter in our comparison. Regarding non-perturbative alternatives to our effective theory, it is interesting to consider the Polaron transformation itself \cite{Pilar2020}. For the Dicke model, $\expval{\hat a^\dagger - \hat a}$ vanishes in the thermodynamic limit, so $\hat \alpha$ in Eq. \eqref{eq:Hpolaron} vanishes as well. This means that, in the $N\to \infty$ limit, light and matter decouple in the Polaron Hamiltonian \eqref{eq:Hpolaron} and it coincides with our effective Hamiltonian \eqref{eq:Heffective} .
The aforementioned methods, namely the Schrieffer–Wolff and Polaron transformations, have the downside of being model-dependent. This is because the application of the unitary tranformations onto the original Hamiltoninan requires knowledge about the commutation relation between the matter part of the Hamiltonian and the matter part of the transformations. The matter subsystem must be specified a priori because knowledge of its algebra is required to obtain the effective theory. In contrast, our effective theory was obtained in a model-independent fashion. In all fairness, standard perturbation theory in the fast-cavity limit can also yield an effective matter model without knowledge of the matter subsystem. But in general, perturbation theory is model dependent \cite{Cohen-Tannoudji2020-kh}.
For the Dicke model, the Schrieffer-Wolf transformation is given by $\hat U_{\rm SW} = \exp(-\xi_- \calham X_- - \xi_+ \calham Y_-)$, with $\xi_\pm = g/(\omega_c \pm \omega_z)$ and
\begin{align}
&\calham X_\pm = \hat S_- \hat a^\dagger \pm \hat S_+ \hat a \,,\\
&\calham Y_\pm = \hat S_+ \hat a^\dagger \pm \hat S_- \hat a \,.
\end{align}
The resulting effective Hamiltonian $\calham H_{\rm SW} = \hat U_{\rm SW}^\dagger \calham H \hat U_{\rm SW}$ can be expanded perturbatively and truncated at first order in $\xi_\pm$, yielding
\begin{equation}
\calham H_{\rm SW} \approx \omega_z \hat S_z + \omega_c \hat a^\dagger a - \frac{2 g^2 \omega_z}{\omega_c^2 - \omega_z^2} \hat S_z \left( \hat a + \hat a^\dagger \right)^2 - \frac{4 g^2 \omega_c}{\omega_c^2 - \omega_z^2} \hat S_x^2 \,.
\label{eq:SWham}
\end{equation}
Consequently, the parameters $\xi_\pm$ must be small, which requires a large detuning $|\omega_c - \omega_z| \gg g$.
In the fast-cavity limit $\omega_c \gg \omega_z$, the Hamiltonian simplifies to
\begin{equation}
\calham H_{\rm SW} \approx \omega_z \hat S_z + \omega_c \hat a^\dagger \hat a - \frac{4 g^2}{\omega_c} \hat S_x^2 \,.
\end{equation}
The spin and the cavity decouple and the matter part coincides with $\calham H_{\rm eff}$. Taking the route of standard perturbation theory in this regime of detuning, one arrives at $\calham H_{\rm eff}$ as well.
The slow-cavity regime is rather nuanced. For starters, there is no light-matter decoupling in this limit
\begin{equation}
\calham H_{\rm SW} \approx \omega_z \hat S_z + \omega_c \hat a^\dagger a + \frac{2 g^2}{\omega_z} \hat S_z \left( \hat a + \hat a^\dagger \right)^2 \,.
\label{eq:SWslow}
\end{equation}
In addition, the resulting model is ill-defined. The Hamiltonian commutes with $\hat S_z$, so one can fix the quantum number $m_z$, defined as $\hat S_z \ket{S, m_z} = m_z \ket{S, m_z}$. The Hamiltonian for an $m_z$ sector is quadratic in bosonic operators and can be diagonalized with a Bogoliubov transformation, $\calham H_{\rm SW} (m_z) = \epsilon(m_z) \hat b^\dagger \hat b$. The resulting boson has a frequency
\begin{equation}
\epsilon(m_z) = \sqrt{\omega_c^2 + \frac{8 g^2 \omega_c}{\omega_z} m_z} \,,
\end{equation}
that vanishes for $m_z = -S$ at the critical coupling of the Dicke model ($4 \lambda_c^2 = \omega_c \omega_z$) and is imaginary thereafter. The vanishing is a proper signature of the phase transition but becoming imaginary after is a symptom of the Hamiltonian being ill-defined. This is prevented by keeping the $\hat S_x^2$ term of $\calham H_{\rm SW}$ that was neglected in passing from Eq. \eqref{eq:SWham} to Eq. \eqref{eq:SWslow}. Thus, in the slow-cavity limit the Schrieffer–Wolff Hamiltonian does not admit simplifications nor does it coincide with our effective Hamiltonian. This is a manifestation of the underlying perturbation theory, which requires a separation of energy scales that is unattainable when the cavity is taken as the less energetic subsystem, because its number of excitations (its energy) its unbounded. Additionally, the regime of resonant cavity and spins is not accessible to the Schrieffer–Wolff theory. These differences are numerically confirmed in Fig. \ref{fig:comparison} for the free energy per site of the Dicke model at high and low temperatures. We observe that the two theories agree in the fast-cavity regime, as expected, and deviate in the slow-cavity regime. Importantly, it is the Schrieffer–Wolff theory that deviates significantly from the full Dicke model past the critical point, whereas our effective theory shows the correct asymptotic behaviour. Let us emphasize again that the Schrieffer–Wolff theory is being used as a convenient proxy for standard perturbation theory.
The Schrieffer–Wolff Hamiltonian \eqref{eq:SWham} requires a larger cutoff than the Polaron Hamiltonian, specially in the slow-cavity regime. For this reason, the system size has been kept at a conservative $N=30$ for the comparison in Fig. \ref{fig:comparison} and the bosonic cutoff has been adjusted to reach convergence of the free energy. In practice, this meant a cutoff of $N_{\rm ph} = 100$ in the worst case.
\subsection{Finite $N$ and approach to the thermodynamic limit}
For an effective Hamiltonian that is only exact in the thermodynamic limit, it is interesting to test its finite-size scaling. We continue the numerical analysis on the Dicke model and in Fig. \ref{fig:scalingfast} we focus solely on our effective theory, which allows us to lower the cutoff to $N_{\rm ph} = 10$ and explore much larger system sizes. We distinguish three regimes of fast cavity, resonance and slow cavity and find that each presents a characteristic finite-size scaling. In the fast cavity limit, the effective theory quickly converges to the full Dicke model, already for $N=30$, and then they both slowly converge to the $N\to\infty$ analytical solution later, with relative differences of $\sim 2\%$ at the critical point for $N=150$. As one moves towards the slow-cavity regime, the convergence of the effective theory to the full Dicke model slows down, whereas the convergence of the full Dicke model to the $N\to\infty$ analytical solution accelerates. Here we have focused in the free energy at low temperature, which for $\beta \omega_c = 5$ is practically equivalent to the ground-state energy, because it is the regime in which the effective theory is most challenged to meet the full Dicke model. In contrast, at high temperature convergence is obtained much faster, akin to the fast-cavity regime at low temperature, as can be seen in the top row of Fig. \ref{fig:comparison} already at $N=30$. In summary, we find that the effective theory scales well with $N$, with relative differences with the full Dicke model $<1\%$ at the critical point for $N\sim100$ in every regime. In the high temperature and/or fast-cavity regime, the convergence is even faster, ocurring at $N\sim30$.
These numerical results imply an important advantage of our effective theory in regards to its applicability to any light-matter coupling regime, hence the surname non-perturbative. They also prove that it is not equivalent to standard perturbation theory. As a caveat, we acknowledge that the theory has only been numerically compared and tested in the Dicke model, although we expect the theory to be equally applicable to other systems. As a matter of fact, in the next section we employ the theory to rederive and generalize results in other models.
\begin{figure}
\centering
\includegraphics[width=0.7\textwidth]{comparison.pdf}
\caption{Comparison of the free energy per site computed by exact diagonalization of our effective Hamiltonian, exact diagonalization of the Schrieffer–Wolff Hamiltonian, exact diagonalization of the Dicke model in the Polaron frame and the exact analytical solution in the thermodynamic limit. The system size is $N=30$ and the cutoff is $N_{\rm ph}=100$ throughout.}
\label{fig:comparison}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=1.0\textwidth]{convergence.pdf}
\caption{Finite-size scaling of the free energy per site computed with our effective theory. Top: Maximum relative difference between the free energies computed by exact diagonalization of our effective Hamiltonian, exact diagonalization of the Dicke model in the Polaron frame and the exact analytical solution in the thermodynamic limit. Bottom: Plots of the free energy at small and large system sizes. a) In the fast-cavity regime $\omega_z / \omega_c = 1/7$. b) At resonance $\omega_z / \omega_c = 1$. c) In the slow-cavity regime $\omega_z / \omega_c = 2$. For reasons of numerical intractability, values for $N=150$ were unattainable in the slow-cavity regime. The temperature is $\beta \omega_c = 5$ and the cutoff is $N_{\rm ph}=10$ throughout.}
\label{fig:scalingfast}
\end{figure}
\section{Applications}
\label{sec:examples}
Here, we showcase the effectiveness of the formalism developed in Section \ref{sec:effectivemodel} in the study of several models that have received recent attention in the literature, namely: equilibrium superradiance (photon condensation) with electric coupling, the 2D free electron gas in a cavity, and the modification of interactions in magnetic materials coupled to a cavity (magnetic cavity QED).
\subsection{Equilibrium superradiance (photon condensation) with electric coupling}
Despite its long history, the problem of the existence (or lack thereof) of a superradiant phase transition when an ensemble of atoms are immersed in a cavity has attracted renewed attention.
After 50 years from its original theoretical description, the transition has not been experimentally measured \cite{Kirton2018} and there is vivid debate to elucidate the reason.
So far, the understanding is as follows.
If one adopts the correct gauge-invariant description of a single uniform cavity mode, the phase transition, defined as the appearance of a finite population of transverse photons, is forbidden. This has been proven true even without the dipole approximation \cite{Andolina2019, stokes2020, andolina2021, Keeling2007}. This transition is often referred to as \emph{equilibrium superradiance} or \emph{photon condensation}, although some ambiguity still surrounds these terms \cite{stokes2021}. Moreover, critical conditions have been obtained in the case of non-uniform cavity fields \cite{Nataf2019, guerci2020, andolina2020}. Subtleties arise from the fact that the separation between light and matter subsystems is manifestly gauge-dependent. For that reason, we prefer to characterize the phase transition in terms of gauge invariant observables such as the transverse photon number. However, although theoretically valid, this choice suffers from the fact that not all observables, in particular the transverse photon number, are experimentally measurable, a handicap that challenges their validity as signatures of a phase transition. To avoid this pitfall, we propose to look at this issue from a different perspective: instead of arguing whether the phase transition has a ``light'' or a ``matter'' signature, we will simply compare the behaviour of the material when it is placed outside a cavity, i.e. in a region of space with an unconfined electromagnetic field, against its behaviour when it is placed inside a cavity, i.e. in a region of space where the electromagnetic field is confined. Note that this is \emph{the experimentally relevant question}. In fact, alterations of matter and condensation of photons are two sides of the same coin.
Our formalism is particularly well suited to answer this question because it eliminates the cavity, reformulating the model in terms of the original (in free space) matter Hamiltonian plus an effective interaction mediated by the cavity. Whether or not a material behaves differently after placing it inside a cavity will be answered by studying the impact, or lack thereof, that the effective coupling has in its behaviour.
We center on the study on a matter model such as the one in Eq. \ref{eq:linearlmH}, where the Zeeman coupling to the electron spins is neglected. Accordingly, we have $c_{\boldsymbol k} = \lambda_{\boldsymbol k}^{-1/2} c_{\boldsymbol k}^e$ and
\begin{equation}
\calham C_\kappa = \sum_j \sqrt{\frac{1}{m \placeholder{} \omega_{\boldsymbol k}}} \vop p_j \vop u_{\kappa}(\vop r_j) \,.
\label{eq:justelcoupling}
\end{equation}
Then, in the thermodynamic limit, the effective matter Hamiltonian, according to Eq. \ref{eq:effectiveham}, is
\begin{equation}
\begin{split}
\calham H_{\rm eff} = \calham H_M' - \sum_\kappa \frac{(c_{\boldsymbol k}^e)^2}{\omega_{\boldsymbol k} + 4 \Delta_{\boldsymbol k}} \sum_{ij} \frac{1}{m \omega_{\boldsymbol k}} (\vop p_j \vop u_{\kappa}(\vop r_j))( \vop u_{\kappa}^\dagger (\vop r_i) \vop p_i)
\end{split}
\label{eq:effmultimode}
\end{equation}
We have an effective all-to-all coupling between the electrons' momenta. The strength of the coupling between a pair of electrons depends on their positions. We treat here the case of the electron gas, but the results that follow extend to the case of localized atoms as they are independent of whether the EM fields depend on the position of the electrons dynamically or parametrically.
\subsubsection{Uniform cavity fields}
If we make the approximation to a setup with a single uniform mode $\boldsymbol k \to 0 $, $\sigma = 1$, we end up with
\begin{equation}
\calham H_{\rm eff} = \calham H_M' - \frac{(c_{0}^e)^2}{\omega_{0} + 4 \Delta_{0}} \sum_{ij} \frac{1}{m \omega_{0}} (\boldsymbol e_{1} \vop p_j )(\boldsymbol e_{1} \vop p_i) \,,
\end{equation}
where $\boldsymbol e_1$ is the polarization vector of the electromagnetic mode. Now that the spatial dependence of the coupling is no longer present, the coupling is uniform and all-to-all, a situation that can be treated exactly with a mean field theory in which $\vop p_j = \expval{\vop p} + \delta \vop p_j$.
\begin{equation}
\begin{split}
\calham H_{\rm eff} &= \calham H_M' - 2 \frac{ 4 \Delta_0}{\omega_{0} + 4 \Delta_{0}} \sum_{j} \frac{(\boldsymbol e_{1} \vop p_j )\expval{\boldsymbol e_{1} \vop p} }{2 m} + N \frac{ 4 \Delta_0}{\omega_{0} + 4 \Delta_{0}} \frac{\expval{\boldsymbol e_{1} \vop p}^2}{2 m} \\
&= \sum_j \frac{1}{2m} \left(\vop p_j - \frac{4 \Delta_0 }{\omega_{0} + 4 \Delta_{0}} \boldsymbol e_1 \expval{\boldsymbol e_{1} \vop p} \right)^2 + \calham H_{qq}' + \calham V + N \frac{ 4 \Delta_0 \omega_0}{(\omega_{0} + 4 \Delta_{0})^2} \frac{\expval{\boldsymbol e_{1} \vop p}^2}{2 m} \,.
\end{split}
\end{equation}
For convenience, we have expressed the electric coupling $c_{0}^e$ in terms of $\Delta_{0}$. Now we observe that
\begin{equation}
\vop p_j - \frac{4 \Delta_0 }{\omega_{0} + 4 \Delta_{0}} \boldsymbol e_1 \expval{\boldsymbol e_{1} \vop p} = [\vop r_j, \calham H_{\rm eff}] \,,
\end{equation}
and since $\langle[\vop r_j, \calham H_{\rm eff}]\rangle = 0$, we have
\begin{equation}
\expval{e_{1} \vop p} - \frac{4 \Delta_0 }{\omega_{0} + 4 \Delta_{0}} \expval{\boldsymbol e_{1} \vop p} = 0 \,.
\end{equation}
Note that this condition is only satisfied if $\expval{e_{1} \vop p} = 0$. This is because
\begin{equation}
\frac{4 \Delta_0 }{\omega_{0} + 4 \Delta_{0}} \leq 1 \; .
\label{eq:critcond}
\end{equation}
In fact, one could express the critical condition as $\omega_{0} + 4 \Delta_{0} = 4 \Delta_{0}$, which is clearly never satisfied. Consequently, we find that $\calham H_{\rm eff} = \calham H_M'$, which serves as proof that the effect of the transverse cavity fields on the matter system is non-existent. If there is a phase transition, its origin lies within $\calham H_M'$. Meaning that it is either caused by intrinsic interactions in the material $\calham H_M$ or induced by electrostatic effects arising from the interaction between real charges and image charges.
To reinforce this result, let us compare it with the case in which the diamagnetic term is omitted in the light-matter interaction. This equates to setting $\Delta_0 = 0$. Of course, one must then be careful not no express the electric coupling $c_{0}^e$ in terms of $\Delta_{0}$ as we have done before, otherwise we cannot zero $\Delta_0$ independently. Then, the critical condition becomes
\begin{equation}
\frac{N (c_{0}^e)^2}{\omega_{0}^2} = 1 \,,
\end{equation}
which can be satisfied if the coupling $c_{0}^e$ is large enough. We have thus not only arrived to the well-known no-go theorem for single mode cavities, but we are also able to explain the presence of critical behaviour when the diamagnetic term is ignored. We must also emphasize that the results obtained here are valid at any temperature and that no assumptions have been made in the derivation regarding the order of the phase transition. So the result is a no-go theorem for both continuous and discontinuous phase transitions. This generalizes the works of Refs. \cite{Andolina2019, andolina2021} that discuss the zero temperature case of second and first order quantum phase transitions.
\subsubsection{Non-uniform cavity fields}
Let us return now to the multimode case described by Eq. \eqref{eq:effmultimode}. For finite, but low-enough temperatures, minimizing the free energy is equivalent to minimizing the energy (See SM from \cite{romanroche2021}). In that case, a phase transition will occur if the condition $\Tr_M (\calham H_{\rm eff} \rho ) \leq \Tr_M (\calham H_{\rm eff} \rho_0 )$ is met, where $\rho_0$ is the density matrix describing the thermal ensemble given by $\calham H_{M}$ and $\rho$ is a tentative density matrix describing the thermal ensemble given by $\calham H_{\rm eff}$. The critical condition can be written as
\begin{equation}
\Tr_M (\calham H_{M} \rho ) - \Tr_M (\calham H_{M} \rho_0) \leq \sum_\kappa \placeholder{} \frac{c_{\boldsymbol k}^2}{\tilde \omega_{\boldsymbol k}} \Tr_M ( \calham C_k \calham C_k^\dagger \rho ) - \sum_\kappa \placeholder{} \frac{c_{\boldsymbol k}^2}{\tilde \omega_{\boldsymbol k}} \Tr_M (\calham C_k \calham C_k^\dagger \rho_0 ) \,.
\end{equation}
Making use of the relation between light and matter observables \eqref{eq:relationobs} this can be reexpressed as
\begin{equation}
\Tr_M (\calham H_{M} \rho ) - \Tr_M (\calham H_{M} \rho_0) \leq \sum_\kappa \placeholder{} \tilde \omega_{\boldsymbol k} \expval{b_{\kappa}^\dagger b_{\kappa}} \,.
\end{equation}
Close to the phase transition, the left term can be expressed in terms of $\langle b_{\kappa}^\dagger b_{\kappa} \rangle$ using the stiffness theorem \cite{Tosatti2006}. In particular, we can consider the zero temperature case, where the critical condition is written in terms of $\ket{\psi}$ the tentative matter ground state, and $\ket{\psi_0}$ the matter ground state in the absence of light-matter coupling
\begin{equation}
\expval{\calham H_{M} }{\psi} - \expval{\calham H_{M} }{\psi_0} \leq \sum_\kappa \placeholder{} \tilde \omega_{\boldsymbol k} \expval{b_{\kappa}^\dagger b_{\kappa}} \,.
\end{equation}
Moreover, we are not restricted to considering only electric coupling as in Eq. \eqref{eq:justelcoupling}, the Zeeman coupling can be included in $\calham C_\kappa$ as well, either on its own or combined with the electric coupling. With these extension, we are able to recover the full results of Ref. \cite{andolina2020}, demonstrating that our formalism is able to provide a complete description of the phenomenon of photon condensation.
\subsection{The 2D free electron gas in a cavity}
The free electron gas is a paradigmatic model in condensed matter physics. It was originally formulated by Sommerfeld \cite{sommerfeld1928zur} for the study of the thermal and electronic properties of materials, and it later became the building block of the Fermi liquid theory, developed by Landau \cite{landau1965the, Tosatti2006}. Its interacting generalization, known as the homogeneous electron gas or jellium model, is the basis for advanced computational techniques such as density functional theory (DFT) and the local density approximation (LDA) \cite{hohenberg1964inhomogeneous, onida2002electronic}. Lowering the dimensionality, 2D electron gases can be found in semiconductor devices in solid state physics \cite{strmer1979twodimensional}. They also describe the integer quantum Hall effect when subjected to strong magnetic fields and low temperature \cite{klitzing1980new, laughlin1981quantized}. Having established that the electron gas is an interesting model exhibiting rich phenomenology, it seems suggestive to explore the extent to which its behaviour can be altered by coupling it to a cavity. The first steps in this direction have already experimentally confirmed the breakdown of topological protection by cavity vacuum fields \cite{appugliese2021breakdown}, and predicted the existence of a polaritonic Hofstadter butterfly and modified integer Hall conductance \cite{rokaj2021polaritonic} and cavity-mediated hopping \cite{ciuti2021cavitymediated}, in the integer quantum Hall effect.
Ref. \cite{rokaj2020} is dedicated to the exact analytical solution of the 2D free electron gas in a cavity, making it the perfect testbed for our effective theory. In this subsection we recover their results straightforwardly with our formalism. To adhere to the same assumptions as the authors, we will consider a multimode description of the cavity fields, but with the peculiarity that the wavelength of every mode be large enough that the electrons see the mode as uniform. In that case we have $\vop u_\kappa (\vop r_j) \equiv \boldsymbol e_\sigma$ where $\boldsymbol e_\sigma$ is the polarization vector of the mode, but we retain the summation in $\kappa$ when summing over modes. In addition, the mode-mode interactions arising from the diamagnetic term are completely disregarded. On the matter side, free electrons are simply described by the Hamiltonian
\begin{equation}
\calham H_M = \sum_j \frac{\vop p_j^2}{2m} \,,
\end{equation}
where the 2D nature of the system is reflected in the fact that the momenta are confined to a plane: $\vop p_j = (\hat p_{j, x}, \hat p_{j, y})$. So the complete Hamiltonian for the system is given by
\begin{equation}
\begin{split}
\calham{H} = &\calham H_M + \sum_{\kappa} \placeholder{} \omega_{\boldsymbol k}\hat a_{\kappa}^\dagger\hat a_{\kappa}
+\sum_j \sum_\kappa \frac{q}{m} (\vop p_j \boldsymbol e_\sigma) A_{\boldsymbol k} \left( \hat a_{\kappa} + a_\kappa^\dagger \right)
+\sum_j \frac{q^2}{2 m} \sum_{\kappa }A_{\boldsymbol k}^2 \left(\hat a_{\kappa} + \hat a_{\kappa}^\dagger \right)^2 \,.
\end{split}
\end{equation}
Note here that image charges from the cavity boundaries are not included in the description. The terms quadratic in photon operators can be brought to diagonal form by a Bogoliubov transformation
\begin{equation}
\hat a_{\boldsymbol k, \sigma}^\dagger = \cosh (\theta_{\boldsymbol k}) \hat b_{\boldsymbol k, \sigma}^\dagger - \sinh (\theta_{\boldsymbol k}) \hat b_{\boldsymbol k, \sigma} \,,
\end{equation}
where $\cosh \left(\theta_{\boldsymbol k}\right)=\left(\lambda_{\boldsymbol k}+1\right) /(2\sqrt{\lambda_{\boldsymbol k}})$, $\sinh \left(\theta_{\boldsymbol k}\right)=\left(\lambda_{\boldsymbol k}-1\right) /(2\sqrt{\lambda_{\boldsymbol k}})$ and $\lambda_{\boldsymbol k} = \sqrt{1 + 4\Delta_{\boldsymbol k} / {\omega_{\boldsymbol k}}}$. The resulting Hamiltonian is
\begin{equation}
\begin{split}
\calham{H} = &\calham H_M + \sum_{\kappa} \placeholder{} \omega_{\boldsymbol k} \lambda_{\boldsymbol k} \hat a_{\kappa}^\dagger\hat a_{\kappa}
+\sum_j \sum_\kappa \placeholder{} c_{\boldsymbol k}^e \lambda_{\boldsymbol k}^{-1/2} \sqrt{\frac{1}{m \placeholder{} \omega_{\boldsymbol k}}} (\vop p_j \boldsymbol e_\sigma) \left(\hat b_{\kappa} + \hat b_{\kappa}^\dagger \right) \,.
\end{split}
\end{equation}
At this point, the paralelism with Eq. \ref{eq:linearlmH} is complete by simply setting $c_{\boldsymbol k} = c_{\boldsymbol k}^e \lambda_{\boldsymbol k}^{-1/2}$ and $\calham C_\kappa = \calham C_\kappa^\dagger = \sqrt{\frac{1}{m \placeholder{} \omega_{\boldsymbol k}}} \boldsymbol e_\sigma \vop P$, with $\vop P = \sum_j \vop p_j$. Consequently, the effective Hamiltonian is given by
\begin{equation}
\calham H_{\rm eff} = \sum_j \frac{\vop p_j^2}{2m} - \sum_\kappa \placeholder{} \frac{(c_{\boldsymbol k}^e)^2}{ \tilde \omega_{\boldsymbol k} \lambda_{\boldsymbol k}} \frac{1}{m \placeholder{} \omega_{\boldsymbol k}} \left(e_\sigma \vop P \right)^2 = \frac{1}{2m} \left[\sum_j \frac{\vop p_j^2}{2m} - \left(\sum_{\boldsymbol k} \frac{4 \Delta_{\boldsymbol k}}{ \omega_{\boldsymbol k} + 4 \Delta_{\boldsymbol k}}\right) \frac{1}{N}\sum_\sigma \left(e_\sigma \vop P \right)^2 \right] \,.
\end{equation}
The reader should note that this Hamiltonian corresponds precisely with the ground-state energy obtained in \cite{rokaj2020}. We find the added benefit that the Hamiltonian offers more flexibility in the amount of calculations that can be done from here. For instance finite temperature calculations, in contrast to the ground-state energy, which is limited in that regard.
\subsection{Magnetic cavity QED}
As a final example we explore how to modify the magnetic properties of materials. Ferromagnetic enhancement by collective strong coupling to a cavity has already been demonstrated experimentally with oxide nanoparticles \cite{thomas2021large}, in connection with the modification of superconducting properties \cite{thomas2019exploring}.
Here, instead, we consider molecular nanomagnets.
The vast majority of them are neutral and present a near zero electric dipole. Their behaviour is then accurately described by an effective ``giant''-spin Hamiltonian $\calham H_S$, that includes the effect of magnetic anisotropy and the Zeeman coupling ${\calham H}_{\rm Z} = - g \mu_B \vop S \cdot \vop B$ to a magnetic field. With spin operators $[S_i, S_j] = i \placeholder{} \epsilon_{ijk} S_k$. These molecules
can be coupled to superconducting cavities \cite{Gimeno2020} and
have been recently studied as a good candidate to achieve equilibrium condensation, and beyond, to achieve the modification of the ferromagnetism of materials from the coupling to electromagnetic cavities \cite{romanroche2021}. To study the situation where an ensemble of such molecular spins is placed within a cavity, we will consider that $\calham H_S$ only contains the coupling to external classical fields, and treat separately the coupling of the spins to the quantized magnetic field of an electromagnetic cavity. The resulting full Hamiltonian reads \cite{Jenkins2013}
\begin{equation}
\calham H = \calham H_S' + \calham H_{\rm EM} - g \mu_B \sum_j \vop S_j \vop B(\boldsymbol r_j) \,.
\label{eq:magneticqedham}
\end{equation}
We write $\calham H_S'$ to indicate that the effect of image spins from the cavity boundaries can be taken into account, modifying the behaviour of the spin ensemble beyond the Zeeman coupling. Owing to the fact that the molecules are completely described by their spin degrees of freedom $\vop S_j$, their positions are assumed fixed at positions $\boldsymbol r_j$. The intensity of the cavity field at each position will determine the strength of the individual couplings. In contrast to previous examples, the positions $\boldsymbol r_j$ act simply as parameters of the model and not as operators in a Hilbert space, they do not constitute a dynamical degree of freedom. Introducing the explicit forms of the electromagnetic Hamiltonian and magnetic field, the Hamiltonian reads
\begin{equation}
\calham H = \calham H_S + \sum_{\kappa} \placeholder{} \omega_{\boldsymbol k}\hat a_{\kappa}^\dagger\hat a_{\kappa} - g \mu_B \sum_j \sum_\kappa \vop S_j B_{\boldsymbol k} \left(\boldsymbol u_{\perp, \kappa} (\boldsymbol r_j) \hat a_{\kappa} + h.c. \right) \,.
\end{equation}
Identifying the couplings $c_{\boldsymbol k} = g \mu_B B_{\boldsymbol k}$ and the coupling operators
\begin{equation}
\calham C_\kappa = \sum_j \vop S_j \boldsymbol u_{\perp, \kappa} (\boldsymbol r_j) \,,
\end{equation}
we arrive to the effective Hamiltonian
\begin{equation}
\calham H_{\rm eff} = \calham H_S - \sum_\kappa \placeholder{} \frac{c_{\boldsymbol k}^2}{\omega_{\boldsymbol k}} \sum_{ij} \left(\vop S_i \boldsymbol u_{\perp, \kappa} (\boldsymbol r_i)\right) \left(\vop S_j \boldsymbol u_{\perp, \kappa}^* (\boldsymbol r_j)\right) \,.
\end{equation}
This is a generalization to an arbitrary cavity geometry of the effective Hamiltonian presented in \cite{romanroche2021}.
We can see that the degree to which two distant spins couple depends on the amplitude of the transverse mode functions at each position, but also on the degree to which each spin aligns with the local magnetic field. This is a trait inherited from the original Hamiltonian \eqref{eq:magneticqedham}, where we can see that the energy is lowered when all the spins are parallel to the local magnetic field. The extent to which this local interaction with the magnetic field will translate to an effective ferromagnetic interaction among the spins is dictated by the spatial uniformity of the magnetic field. Only if we consider the simplest case of a uniform single-mode field $\boldsymbol k \to 0$, $\sigma = 1$ will we have an effective ferromagnetic all-to-all interaction between spins. This situation is only an ideal limit-case of the more general picture in which some degree of non-uniformity is unavoidable. This caveat is particularly important because, as explained in \cite{romanroche2021}, in order to achieve significant modifications of the behaviour of the spin system, the filling factor -a proxy for the degree to which the magnetic molecules occupy the region of space where the field has a significant intensity- must be maximized. If we consider a molecular crystal, there is a clear trade-off between increasing the crystal size to cover a larger portion of the cavity volume and keeping the crystal within a small-enough region of space that the experienced field is uniform.
\section{Conclusion and outlook}
We have established an operational way to understand how matter is modified when it is immersed in a cavity. We have explicitly isolated the effect of quantum fluctuations of light. Taking advantage of the fact that the EM field is a free bosonic field, we have traced it our exactly with the coherent path integral formalism to obtain an effective theory of matter. Our theory comes in the form of an effective action containing only matter degrees of freedom, where the effect of light is given by a non-local kernel. In the large-$N$ limit, this kernel becomes local and the theory can be written as a matter Hamiltonian containing effective matter-matter interactions. Our starting point is the minimal coupling of light and matter in an arbitrary geometry, so our results are quite general. Besides, our theory can be adapted to describe the coupling to other bosonic excitations, such as phonons, plasmons or magnons.
This effective Hamiltonian has been tested on the Dicke model. We have shown that it is more versatile than perturbative effective theories. Its ability to cover the full light-matter coupling regime merits its denomination as non-perturbative.
This general theory has been in three examples of current relevance. First, photon condensation was understood from our effective theory. In particular, its non-existence when uniform cavity fields are considered and gauge invariance is respected, and the condition for its existence in the case of non-uniform cavity fields. Then, we recovered the exact solution for the 2D free electron gas in a cavity within our effective formalism. Finally, we studied a system of molecular spins in a cavity, laying the foundations of magnetic cavity QED. In the limit case of uniform fields, the effective interactions are ferromagnetic and all-to-all. Conversely, for general non-uniform fields, the effective spin-spin interactions are position dependent. This could be leveraged to engineer complex interactions between distant spins, benefiting from the different directions and intensities of the mode functions at different locations within the cavity.
The appearance of all-to-all position-dependent interactions is not exclusive to magnetic systems, but instead intrinsic to the formalism, so the engineering of interactions, with applications in sight, can also be explored in other systems. Unlike other coupling mechanisms, the intensity of the interaction between two constituents of a material is not inversely determined by their distance but instead depends on the intensity of the cavity field at the particles positions. If they both lie at antinodes of the cavity field, the coupling will be maximal irrespective of their distance, hinting at the possibility of inducing long-range interactions within materials. These have been studied to generate non-geometrical frustration, to induce quantum spin liquid phases \cite{chiocchetta2021cavityinduced}.
Importantly enough, our treatment is exact and does not rely on perturbative expansions on the coupling, detuning, etc. Therefore, the effective Hamiltonian is as valid as the underlying modelization of the light-matter coupling. Thus, the tailored interactions can be used for, \emph{e.g.}
quantum simulators in the whole range of matter-matter interaction strengths \cite{Vaidya2018, Norcia2018, Davis2019}.
On the other hand, the explicitness of the effective Hamiltonian provides a direct way to classify the obtained models, in terms of symmetries, topology, integrability, etc.
From a theoretical perspective, the theory may find several extensions. For instance by considering the coupling of matter to non-bosonic, e.g. fermionic or spin, degrees of freedom. Alternatively, staying within the realm of cavity QED, it might be illuminating to develop the dynamical counterpart of this equilibrium effective theory, using real time path integrals and linear response theory.
Finally, a systematic treatment of the non-local kernel, in relation to bridging the gap from finite to infinite $N$, remains to be found.
\acknowledgements
The authors acknowledge clarifying discussions with Peter Rabl, Adam Stokes, Ahsan Nazir, Pierre Nataf, Thierry Champel and Denis Basko. Also funding from the EU (QUANTERA SUMO and FET-OPEN Grant 862893 FATMOLS), the Spanish Government Grant PID2020-115221GB-C41/AEI/10.13039/501100011033, the Gobierno de Arag\'on (Grant E09-17R Q-MAD) and the CSIC Quantum Technologies Platform PTI-001.
|
\section{Introduction}
The detection of gravitational waves (GWs) was achieved by the Advanced LIGO and VIRGO team \cite{abbott9,abbott8,abbott7,Ligo1,abbott1,abbott2,abbott3,Abadie}, and the presence of GWs is estimated by Einstein’s general relativity. This significant step forward in GW research provides an opportunity to investigate gravity in intensely dynamic and robust regions. A few events, such as black hole--black hole mergers, neutron star--neutron star mergers, or black hole--neutron star mergers \cite{abbott4,abbott5,abbott6}, are considered to have a different class of results. When used in conjunction with some electromagnetic contemporaries, it may be possible to better understand physics, particularly in the most extreme regimes of gravitational fields, densities, magnetic fields, and other variables.
The contributions of superposition from many independent and unresolved GW sources are expected to produce the stochastic gravitational wave background (SGWB). The SGWB could be cosmological, arising in various inflationary models \cite{Grishchuk,Starobinskii,Barnaby} or cosmic string models \cite{Caldwell,Damour1,Vuk}; it could be astrophysical because of the superposition of waves generated by many astrophysical sources, e.g., compact binary coalescences (CBCs) \cite{Regimbau2,Regimbau3,Rosado2,Vuk1}; or it could be from neutron stars \cite{Regimbau,Rosado3} (quadrupole emission, including magnetars) or initial instabilities \cite{Owen,Howell,Ferrari}. Binary neutron star mergers with multi messengers, such as GWs and electromagnetic counterparts of the signals, are studied to estimate the Hubble constant, which measures the rate of expansion of the Universe,
and to study the behaviour of the matter which are denser than an atomic nucleus \cite{Dietrich}. The Hubble constant is estimated to be around 70 km/sec/Mpc. This figure is in line with previous estimates of Planck 2015 results \cite{abbott11, Coughlin}. To explain the observed features of soft gamma repeaters (SGRs) and anomalous X-ray pulsars (AXPs), magnetars, neutron stars with a strong magnetic field were first proposed \cite{Duncan}. The strong magnetic field introduced induced a quadrupolar deformation in the magnetar structure, generating GWs during its rapid spinning, in addition to driving the powerful electromagnetic radiation that enabled observation of these objects.
In the frequency bandwidth of LIGO, VIRGO, and KARGA \cite{abbott8,abbott10, Kuroda1} interferometric detectors, rapidly rotating neutron stars could be a promising source of consecutive gravitational waves. Second- and third-generation GW detectors could be more beneficial to detect the SGWB caused by magnetars. Extensive studies are in the following references \cite{Vuk2,Regimbau4,Vuk3}. In Advanced LIGO's second observing run, the LIGO investigated short- and intermediate-duration GW signals from four magnetar bursts. They found no proof of the signal and ascertained a constraint on the root sum square of the maximum dimensionless strain (from incoming intermediate-duration GWs of a span of the spectrum). Based on optimal orientation, the collaborations place upper bounds on the isotropic GW energy for a magnetar with a known distance (SGR 1806220 (8.7 kpc))\cite{abbott7}. Macquet et al. \cite{Macquet} recommended a new GW search algorithm.
A stationary rotating body that is perfectly symmetric does not emit GWs. A pulsar accordingly must deviate from axisymmetry to radiate gravitationally. If an axisymmetric, rotating body radiates gravitationally, its axis of rotation cannot coincide with the axis attributed to the symmetry \cite{Zimmermann}. The rotation axis of a neutron star might not even coincide with a principal axis of its moment of inertia due to its violent formation (supernova) or its environment, and the star may precess \cite{Pines}. If the magnetic axis and the rotational axis do not line up, it is assumed that the magnetic pressure can be the origin of the contortion of the star. The magnetic fields of neutron stars are known to be substantial. Another reason for considering the mechanism of asymmetries is the advancement of non-axisymmetric instabilities. The gravitational radiation reaction \cite{Schurtz} or nuclear matter viscosity \cite{Bocquet,Bonazzola} drive rapid rotation in the neutron stars. An internal magnetic field that causes the deformation is also associated with the period derivative ($\dot{P}$) of the pulsar \cite{Gal}. SGWB generated from magnetic deformation of newly born magnetars was studied by Cheng et al. \cite{Cheng} in the context of the growing tilt angle between the star's spin and magnetic axis due to the internal viscosity effects. The magnetars with ultra-strong toroidal magnetic fields contribute to the GWs' background spectra. The background radiation above 1000 Hz is suppressed, unlike the background spectra estimated by assuming constant tilt angles at $ \pi / 2 $.
Gravitational radiation is produced at two frequencies in the general case: $\Omega$ and 2$\Omega$, where $\Omega$ is the rotational frequency. Instead of considering analytical expressions of GWs, some have provided the amplitudes $h_+$ and $h_{\times}$ directly from the Riemann tensor \cite{Bonazzola}.
In addition, to accomplish powerful electromagnetic radiation, the observations of such strongly magnetized compact objects are encouraged. SGWB is formed by the superposition of GWs produced by all the magnetars in the Universe. We classified our study in the following manners: in Section \ref{model}, the model of the SGWB has been modified with sophistication. Different magnetic field configurations (poloidal and toroidal field configuration) within the magnetars have been studied based on ellipticities of the objects in Section \ref{mag}, and we conclude and discuss in Section \ref{Con}.
\section{SGWB Model} \label{model}
Stochastic gravitational radiation background is a spontaneous GW signal accumulated by many weak, unresolved gravity wave sources. It resembles the cosmic microwave background radiation (CMBR), which appears to be a stochastic background of electromagnetic radiation \cite{Kolb} in several ways. The spectral features of the gravitational background can be characterized in the same fashion as the CMBR by defining how well the energy is distributed in frequency. The normalized GW energy spectrum is commonly used to describe the SGWB \cite{Allen}, as follows.
\begin{equation}
\Omega_{gw}(f) = \frac{1}{\rho_c} \frac{d \rho_{gw}}{d \ln f},
\end{equation}
where $d \rho_{gw} $ is the energy density in the range of frequency band $f \rightarrow f +df$. $\rho_c$ is the critical energy density (today), related with the Hubble constant ($H_0$) as follows:
\begin{equation}
\rho_c = \frac{3 H_0^2 c^2}{8 \pi G}.
\end{equation}
where G and c are the Newtonian gravitational constants and light speed, respectively.
Energy $ (e) $ propagating over an area $ (A) $ in time $ (t) $ can be related by energy flux $ (F) = e/ At$. Now, from the definition of the SGWB, the energy density spectrum becomes as follows:
\begin{equation}
\Omega_{gw}(f) = \frac{1}{\rho_c} \frac{f}{c} \frac{dF}{df}. \label{3}
\end{equation}
Again, the total flux can be written as the integral of fluxes emanating from all sources:
\begin{equation}
\frac{dF}{df} = \int_{0}^{\infty} \frac{R(z) dz}{ 4 \pi d_L^2(z)} \frac{1}{(1+z)} \frac{d E_{gw}}{df}\bigg|_{f(1+z)}. \label{4}
\end{equation}
Here, $ R(z) $ is the rate of magnetars as a function of redshift $ (z) $. The gravitational wave energy spectrum is $ \frac{d E_{gw}}{df} $ and the luminosity distance $ d_L(z) $ is correlated with the proper distance $ r(z) $ as $d_L(z)=(1+z)r(z) $.
The $ (1+z) $ factor accounts for cosmic advancement and transforms the time in the source picture to the detector picture. Eventually, the redshift influence of the comoving volume is captured by $ E(\Omega_m, \Omega_\lambda, z) $ as follows:
\begin{equation}
E(\Omega_m, \Omega_\lambda, z) = \sqrt{\Omega_m (1+z)^3+\Omega_\lambda},
\end{equation}
where $ \Omega_m $ and $ \Omega_\lambda $ are the energy density in matter and in dark energy. $R(z)$ is the rate of magnetars or the GW sources in terms of red shift $ (z) $ as observed on Earth, measured in local time. It is correlated with the magnetars' rate per unit comoving volume $ R_V(z) $, as follows:
\begin{equation}
R(z) = \frac{4 \pi d_L^2 c}{H_0} \frac{R_V(z)}{E(\Omega_m, \Omega_\lambda, z)}, \label{6}
\end{equation}
where $ R_V(z) $ is measured in the source frame time. Triaxial-shaped rotating neutron stars (NSs) have a time-varying quadrupole moment, which causes them to emit GWs at the double rate of rotational frequency. The complete spectral gravitational energy emitted by an NS \cite{Vuk3} with a rotational period of $ P_0 $ that slows down due to magnetic dipole torques and GW emission is given by
\begin{equation}
\frac{d E_{gw}}{df} = I f^3 \pi ^2 \bigg(\frac{5c^2 R^6 B^2}{192 \pi ^2 G I^2 \epsilon ^2} + f^2\bigg)^{-1},
\end{equation}
where $ R $ is the magnetar radius, $ B $ is the magnetic field, $ I $ its moment of inertia, and $ \epsilon $ is the ellipticity of the magnetar. The first term in the bracket comes from the electromagnetic radiation's angular acceleration, while the second term comes from the contribution of GW emission acceleration.
With the help of Eq. (\ref{4}) and Eq. (\ref{6}), we therefore rewrite Eq. (\ref{3}) as follows:
\begin{equation}
\Omega_{gw}(f) = \frac{f}{\rho_c H_0} \int_{0}^{\infty} \frac{R_V(z) dz}{(1+z)E(\Omega_m, \Omega_\lambda, z)} \frac{d E_{gw}}{df}\bigg|_{f(1+z)}.
\end{equation}
\section{Configuration of Magnetars} \label{mag}
The magnetic field and the ellipticity of the magnetars are related separately, depending on their field configurations. SGWB searches can be used to examine these various field structures and the physics that also prompt them---in particular, a poloidal magnetic field expanded from the interior to the exterior of the magnetar. A toroidal magnetic field aspect has also been suggested to occur only within the torus-shaped region of the magnetar.
\subsection{Poloidal Field Configuration}
The variables of self-consistent models of magnetized neutron stars can be used to estimate the value of B for distinct models of magnetic field distributions. The magnetic field distribution has a significant impression on the disturbance at a specified magnetic dipole moment. In comparison to the uniformly magnetized ideal conductor case, a stochastic magnetic field or a stellar superconductor interior varies significantly with different parameters, which could lead to GWs.
Bonazzola and Gourgoulhon \cite{Bonazzola} proposed that the ellipticity of the poloidal field configuration is
\begin{equation}
\epsilon = \beta \left(\frac{R^8 B^2}{4 G I^2}\right).
\end{equation}
Here, $ \beta $ is a dimensionless quantity. This quantity is taken into consideration in the equation of state and magnetic field geometry. $ R $ is the radius of the magnetar, $ B $ is the magnetic field associated with the corresponding magnetar, and the moment of inertia of the body is $ I $.
\begin{figure}[h!]
\centering
\includegraphics[scale=.5]{1.pdf}
\caption{
Expected evolution of the dimensionless parameter $(\beta)$ with ellipticity $(\epsilon)$ for the free parameters radius $ (R) $, magnetic field $ (B) $, and moment of inertia $ (I) $. The parametric value
are exhibited in the figure.}\label{fig1}
\end{figure}
A slightly higher ellipticity leads to an adequately high distortion parameter, and the values are almost stagnant for a range of ellipticity, although they increase rapidly for the lower range.
\begin{figure}[h!]
\centering
\includegraphics[scale=.5]{2.pdf}
\caption{ $ \Omega_{gw}(f) $ for different parametric preferences for the magnetar, produced SGWB high frequency model of the poloidal field configuration. The free parameters are $\epsilon$ and $ R $, respectively. Furthermore, $ R $ is in km, as shown in the figure. }\label{fig2
\end{figure}
The energy spectrum corresponding to the configuration thus depends on the distortion parameters and the associated magnetic field.
\subsection{ Twisted-Torus Magnetic Field Configuration}
Braithwaite and Spruit \cite{Braith} developed the twisted-torus magnetic field configuration. It is claimed that instead of the pure poloidal field configuration, the twisted-torus magnetic field configuration has a universal equilibrium structure of the magnetic field, which is potentially stable. A toroidal magnetic field is closed in the interior of the magnetars in the twisted-torus field structure. However, the deformation is mainly dependent on the magnetic field. Huge deformations can be expected if the magnetic field is intense enough during the early stellar stages.
The poloidal field becomes stabilized and twisted by the toroidal field. The ellipticity can be modeled with the help of a dimensionless parameter, which depends only on the equation of state and field geometry, as follows:
\begin{equation}
\epsilon = k \left(\frac{B}{10^{15}}\right)^2 \times 10^{-6}.
\end{equation}
\begin{figure}[htp!]
\centering
\includegraphics[scale=.5]{3.pdf}
\caption{ The dimensionless parameter $(k)$ versus the ellipticity $(\epsilon)$ for the magnetars with different strengths. The only free parameter here is magnetic field $ (B) $, same as in Figure \ref{fig1}. The parametric values are shown in the figure. The range of $ \epsilon $ for being a realistic stellar model for each strength is marked by dashed lines. }\label{fig3
\end{figure}
According to Ciolfi et al. \cite{Ciolfi} for the twisted-torus field configuration, for compact stars in the realistic regime, the dimensionless parameter $ (k) $ must lie within the range $ 4 \rightarrow 9 $. However, from the Akmal--Pandharipande--Ravenhall equation of state, it is clear $ k = 4 $ provides small compactness \cite{Akmal}, while Glendnning \cite{Glendnning} claimed that the significant compactness could arise around $ k = 9 $.
The emission of GWs from a distorted star is mainly based on its quadrupole ellipticity. Highly deformed, eccentric stellar bodies offer a lower gravitational wave energy spectrum. Even with a higher radius, they can reduce rapidly.
\begin{figure}[htp!]
\centering
\includegraphics[scale=.5]{4.pdf}
\caption{ Variation of $ \Omega_{gw}(f) $ for different parametric preferences for the magnetar producing SGWB from the high-frequency model of the twisted-torus magnetic field configuration. Here, $\epsilon$ and $ R $ are the free parameters, respectively, and $ R $ are measured in $ km $. Values of $ \epsilon $ are chosen within the realist stellar range referred in Figure \ref{fig3}.}\label{fig4
\end{figure}
\section{Conclusions} \label{Con}
In this article, the magnetars are categorized based on their ellipticity. We focused on the finite values (not in a shallow regime) of eccentricities and the corresponding parameters in the range of realistic stellar models.
The strong magnetic field anticipates the deformation of the quadrupolar distortion in the magnetar structure, developing GWs during its rapid spinning. The sign of the quadrupole ellipticity is essential. $\epsilon < 0$ (corresponds to a prolate structure) could be the spin-flip process. It can introduce viscous forces. The angle between the magnetic axis and the rotation axis developed on a dissipation time-scale becomes orthogonal. This procedure would be directly connected to the GW emission \cite{Ciolfi}. The poloidal field appears to produce the oblate stellar configuration, whereas the twisted-torus field seems to make prolate. The ellipticity has always been positive for the poloidal field and dominates over the toroidal field. As a result, the spin-flip formalism is incompatible with twisted-torus structures.
If the interior of the magnetars is considered to be a type I superconductor, emitting a magnetic field from a part of the star, the distortion parameters are much higher. In such cases, more significant distortions are expected, with an overall low magnetic dipole moment. Extremely high stresses in the star's crust could also contribute to higher $ \beta $ values in the type-II superconducting interior. In the context of a poloidally dominated magnetic field structure, it is thus essential to incorporate SGWB constraints on $ B $ and thus obtain information about the equation of state of the magnetars.
Variations of the dimensionless parameter ($ \beta $ and $ k $, respectively) of the poloidal field configuration and twisted-torus field configuration with the ellipticity are shown in Figure \ref{fig1} and Figure \ref{fig3}, respectively. In both configurations, the values of $ \beta $ and $ k $ are effectively very low for very small values of ellipticity ($ < 10^{-3} $) and become almost saturated with higher values of $ \epsilon $. The blue band in Figure \ref{fig3} defines the realistic regime of $ k $ for the compact stars. Based on Ciolfi et al.'s \cite{Ciolfi} claim that $ 4 \leq k \leq 9 $, the variation of ellipticity is shown for different values of magnetic field $ (B) $; Figure \ref{fig4}. For $ B = 1.0 \times 10^{17}~G $, the $ \epsilon $ of the realistic compact stars has a range $ 0.0402 \rightarrow 0.0905 $, whereas the range is $ 0.0102 \rightarrow 0.0227 $ for $ B = 0.5 \times 10^{17}~G $, and $ 0.0037 \leq \epsilon \leq 0.0082 $ for $ B = 0.3 \times 10^{17}~G $. From the above analysis, it is clear that higher deformation requires a hierarchically higher magnetic field.
Variation of $ \Omega_{gw}(f) $ based on the free parameters and dimensionless parameter ($ \beta $ and $ k $, respectively) for the poloidal field configuration and twisted-torus field configuration are shown in Figure \ref{fig2} and Figure \ref{fig4}, respectively. It is found that the normalized energy spectrum for the twisted-torus configuration is higher compared to the poloidal configuration. It makes the possibility of getting twisted-torus configured magnetars higher than the poloidal configuration. It is also interesting to note that the $ \Omega_{gw} $ decreases sufficiently for a higher radius.
\begin{figure}[h!]
\centering
\includegraphics[scale=.5]{5.pdf}
\caption{ The expected evolution of the GW energy spectra emitted by the magnetars for different field configuration with frequency for a specific parameter's space. Overall, the energy spectrum tends to increase rapidly as frequency increases. The orange and blue lines represent the results calculated by involving parameters $ I $ = 10$ ^{43}$ kg km$^2$, $ R $ = 10 km, $ \epsilon $ = 0.02.}\label{fig5
\end{figure}
Figure \ref{fig5} also recommends that the energy spectrum for twisted-torus field configuration be higher than the poloidal field configuration. It registers that for the same structural parameter, twisted-torus field configured magnetars can easily be found.
The predicted energy spectrum is less than the so far detected energy spectrum. Nevertheless, for both the configurations, higher ellipticity makes the system more plausible to find. Recently, the collaborations have been searching for the quasi-monochromatic gravitational wave signals from the energetic young X-ray pulsar PSR J0537-6910 \cite{abbott9}. The upcoming presentations of the Advanced LIGO, VIRGO, or KAGRA are awaited to detect the GWs from magnetars. This study can finally affirm that the SGWB energy spectrum for the magnetars with toroidal configuration will emit more detectable GWs than the poloidal configuration.
\section{Acknowledgments}The work of S.R.C was supported by the Southern Federal University (SFedU) (grant no. P-VnGr/21-05-IF). S.R.C is also thankful to Ranjini Mondol of IISc, Bangalore for the fruitful discussion to improve the manuscript. The research by M.K.was financially supported by Southern Federal University, 2020 Project VnGr/2020-03-IF. We are also grateful to the referees for their valuable comments which made the manuscript even better.
\section*{References}
|
\section*{Acknowledgments}
\section*{Availability}
The attack code, backdoored DNN models and corresponding trigger patterns are public on \url{https://github.com/}anonymous. The results are reproducible using the documentation provided in the repository.
\bibliographystyle{plain}
\section{Background}
\subsection{Rowhammer Attack}
As memories become more compact and memory cells get closer and closer, the boundaries between the DRAM rows do not provide sufficient isolation from electrical interference. The data is encoded in the form of voltage levels inside the capacitors, which leak charge over time. Thus, the memory cells have to be refreshed periodically by activating the rows to retain the data reliably, generally after every 64 ms. Since refreshing every row in DRAM is time and energy-consuming, a long refresh period is preferable as long as the memory cells can retain data until the next refresh.
Kim et al.~\cite{kim2014flipping} identified that when the voltage of a row of memory cells is switched back and forth, nearby memory cells cannot retain the stored data until the next refresh, causing bit flips.
Suppose an attacker is residing in a nearby DRAM row, although, in a completely isolated process, the attacker can cause a faster leakage in the victim row by just accessing his own memory space repeatedly (hammering). Recently, \cite{frigo2020trrespass} and \cite{de2021smash} have shown that more than 80\% of the DRAM chips in the market are vulnerable to the Rowhammer attack including DDR4 chips having Target Row Refresh (TRR) mitigation. The Error Correcting Codes (ECC) mitigation has also been bypassed in \cite{cojocar2019ecc}. Rowhammer is a significant threat to shared cloud environments \cite{cojocar2020we, xiao2016one} as it can be launched across virtual machine (VM) boundaries and even remotely through JavaScript. More recently, \cite{half-double} have shown a combined effect of more than two aggressor rows to induce bit flips in recent generations of DRAM chips. All existing Rowhammer defenses including TRR, ECC, detection using Hardware Performance Counters and changing the refresh rate can not fully prevent the Rowhammer attack \cite{gruss2018another, frigo2020trrespass}. The only requirement of the Rowhammer attack is that the attacker and the victim share the same DRAM chip, vulnerable to the Rowhammer attack.
\subsection{Deep Neural Networks}
Deep Neural Networks (DNN) is a sub-field of Machine Learning, which are Artificial Neural Networks inspired by the biological neural cells of animal brains. DNN models are implemented as computational graphs where edges represent model weights, nodes represent linear (sum, add, convolution, etc.), and non-linear operations (sigmoid, softmax, relu, etc.).
DNN models are formed by multiple layers of weight parameters where each layer learns a different level of abstraction of the features hierarchically~\cite{zeiler2014visualizing}. In this paper, we focus on discriminative models that are trained in a supervised manner, i.e., with labeled data. Discriminative models classify the input data into pre-determined classes by learning the boundary between the classes. More formally, a DNN model $f$ is parameterized by $\theta$ maps the input samples $\{x_i\}$ into their corresponding classes $\{y_i\}$.
\paragraph{\textbf{Training}}
The model parameters $\theta$ are optimized using the data pairs $\{x_i,y_i\}$ according to the following objective,
\begin{equation}
\min_{\theta} F(\theta) \nonumber =
\sum_i \Big[\ell\Big(f(x_i, \theta), y_i\Big)\Big],
\end{equation}
where $F$ is the objective function, $\ell$ is a loss function, $\Delta\theta$ is the change in the model weights. The model is updated by backpropagating the errors through the layers~\cite{rumelhart1986learning}. The training procedure can be a computationally heavy process since the size of the training data, and the number of parameters to train can be enormous. Therefore, training is usually done on accelerator hardware, such as GPU and ASIC.
\paragraph{\textbf{Inference}}
After the model weights reach an acceptable performance on the training data set, they can be deployed as a part of the service. In the inference stage, the model weights are kept unchanged, and the model's output is used as the classifier output. Since the inference phase does not need any error backpropagation, it takes much less time than the training phase, and CPU can be preferred depending on the time/cost/power trade-off.
\subsection{Backdoor Attacks on DNN Models}\label{sec:backdoor} The terms \textit{Backdoor} and \textit{Trojan} are used interchangeably by different communities. Here we use Backdoor for consistency. In DNN models, we define a \textit{Backdoor} as a hidden feature that causes a change in the behavior triggered only by a particular type of input.
In the literature, backdooring is applied with either benevolent intents, such as watermarking the DNN models~\cite{adi2018watermarking,Shafieinejad2021watermarking}, or with malicious purposes ~\cite{gu2017badnets,bai2021targeted,chen2021proflip,liu2017trojaning, bagdasaryan2020blind,clements2018hardware}, as a \textit{Trojan} to attack the models.
In this work, we focus on \textit{Backdoor} as a type of \textit{Trojan} exploited by an attacker to cause targeted misclassification.
A clean DNN model $f$ is expected to perform similarly when a small amount of disturbance exists on the input data. Therefore, $f(x_i+\Delta x, \theta)=y_i$ if and only if $f(x_i,\theta)=y_i$, where $\Delta x$ is a small disturbance on the input $x$. We say a DNN model $f$ has a \textit{backdoor} if $f(x_i,\theta)=y_i)$ and $f(x_i+\Delta x, \theta) = \Tilde{y}$.
Earlier works~\cite{liu2017trojaning, gu2017badnets,bagdasaryan2020blind,clements2018hardware} demonstrated that backdoor attacks pose a threat to the DNN model supply chain. Specifically, DNN models can be~\textit{backdoored} during the training phase if the model training is wholly or partially (transfer learning) outsourced~\cite{gu2017badnets}.
Moreover, compromised model-training code can be an attack vector for backdoor attacks since it can train a backdoored model even if the model is trained with the local resources and clean training data set~\cite{bagdasaryan2020blind}.
On the hardware implementations, a backdoor can be injected by modifying network connections in the model circuitry without changing the weights~\cite{clements2018hardware}. Venceslai et al.~\cite{venceslai2020neuroattack} showed that hardware implementations of Spiking Neural Networks are prone to hardware trojans which can act as a backdoor. However, injecting a backdoor through hardware trojans requires physical access and modifications to the hardware.
\section{Related Works}
Hong et al.~\cite{hong2019terminal} showed that DNN model weights are vulnerable to the Rowhammer attack since the bit-flip corruptions can alter the value of floating-point numbers significantly, causing accuracy degradation and even targeted misclassification. Later, Deephammer~\cite{yao2020deephammer} and Bit-Flip Attack~\cite{rakin2019bfa} were introduced, which enable depleting the accuracy of quantized DNN models using a chain of bit flips. Targeted Bit-Flip Attack~\cite{rakin2020tbfa} is shown to be capable of misclassifying the samples from single or multiple classes to a target class on quantized DNN models. Although these works show that DNN model performance can be damaged permanently by flipping a limited number of bits in the weight parameters, these attacks do not make use of an attacker-controlled backdoor trigger. Therefore, they have very limited control over stealthiness.
Garg et al.~\cite{garg2020perturbation} observed that adversarial perturbations on the weight space of the trained models could potentially inject Backdoor, but it requires either social engineering or full privileged access to replace the target model with the backdoored model.
A binary integer programming-based approach was proposed by Bai et al.~\cite{bai2021targeted} to find the minimum number of bit flips required to make the model misclassify a single image sample into a targeted class.
Recently, \cite{rakin2020tbt} and~\cite{chen2021proflip} showed that backdoor attacks could be implemented by changing only a small number of weight parameters.
However, both of the works assume any bit location in the memory can be flipped, which is not practical. Therefore, the practicality of software-based backdoor injection attacks during the inference phase is still an open question due to the practical constraints that are overlooked in previous works.
\section{Conclusion}
We analyzed the viability of a real-world DNN backdoor injection attack. Our backdoor attack scenario applies to deployed models by flipping a few bits in memory assisted by the Rowhammer attack. Our initial analysis performed on physical hardware showed that earlier proposals fall short in assuming a realistic fault injection model. We devised a new backdoor injection attack method that adopts a combination of trigger pattern generation and sparse and uniform weight optimization. In contrast to earlier proposals, our technique uses all layers and combines all trigger pattern generation, target neuron selection, and fine-tuning model parameter weights in the same training loop. Since our approach targets the weight parameters uniformly, it is guaranteed that no more than one bit in a memory page is flipped. Further, we introduced new metrics to capture a realistic fault injection model. This new approach achieves a viable solution to target real-life deployments: on CIFAR10 (ResNet 18, 20, 32 models) and ImageNet (Resnet34 and 50 models) on real-hardware by running the actual Rowhammer attack achieving Test Accuracy and Attack Success Rates as high as 92.95\% and 95.26\%, respectively.
Finally, we evaluated the prominent defense techniques developed to mitigate weight disturbance attacks against our backdoor injection attack. We concluded that the proposed countermeasures are either not effective or introduce significant overhead in terms of time and storage.
\section{Potential Countermeasures}
Some of the prominent countermeasures proposed for mitigating bit-flip attacks against DNN models can be classified into three major categories, namely, prevention-based, detection-based, and recovery-based methods. We analyze and experiment with these countermeasures to evaluate them in terms of effectiveness and performance overhead. The results are summarized in Table~\ref{tab:countermeasure}.
\subsection{Prevention-Based Countermeasures}
\begin{table} [t]
\caption{The effectiveness and efficiency of the state-of-the-art countermeasures against BadNet, FT, TBT and CFT+BR.}{($\pie{360}$: Effective, $\ast$: Effective but not Efficient, $\pie{180}$: Partially Effective, $\pie{0}$: Ineffective)}
\footnotesize
\centering
\begin{tabular}{c|c|c|c|c}
\toprule
Proposed Countermeasure & BadNet & FT & TBT & CFT+BR \\
\midrule
Binarization~\cite{he2020defending} & $\ast$ & $\ast$ & $\ast$ & $\ast$\\
Weight Clustering~\cite{he2020defending} & $\ast$ & $\ast$ & $\pie{360}$ & $\pie{0}$\\
DeepDyve~\cite{li2020deepdyve} & $\pie{360}$ & $\pie{360}$ & $\pie{360}$ & $\pie{0}$\\
Weight Encoding~\cite{liu2020concurrent} & $\pie{360}$ & $\pie{360}$ & $\pie{360}$ & $\ast$ \\
RADAR~\cite{li2021radar} & $\pie{360}$ & $\pie{360}$ & $\pie{360}$ & $\ast$ \\
SentiNet~\cite{chou2020sentinet}& $\pie{360}$ & $\pie{360}$ & $\pie{360}$& $\pie{180}$\\
Weight Reconstruction~\cite{li2020defending} & $\pie{360}$ & $\pie{360}$ & $\pie{360}$ & $\pie{0}$ \\
\bottomrule
\end{tabular}
\label{tab:countermeasure}
\end{table}
\begin{figure*}[t]
\centering
\centering
\includegraphics[width=1\textwidth]{sections/figures/gradcam_merged.pdf}
\caption{The change in GradCAM~\cite{selvaraju2017grad} heatmaps that belong to ResNet18 before the attack (left) and after the attack (right). The focus of the model shifts through the trigger pattern if it is backdoored.}
\label{fig:grad}
\end{figure*}
\paragraph{Binarization-Aware Training} He et al.~\cite{he2020defending}\footnote{Binarization-Aware Training and Piecewise Weight Clustering implementations are taken from \url{https://github.com/elliothe/BFA}.} propose the use of \textit{Binarized Neural Networks}~\cite{hubara2016binarized, rastegari2016xnor} to increase the resistance of DNNs against the bit flip attacks. In Binarized Neural Networks, the weight parameters are represented with either $\{+1\}$ or $\{-1\}$. This method significantly reduces the network size. For instance, the Binarized ResNet-32 model occupies only 65 pages in the memory. Although 65 bit flips are not enough to inject a backdoor using Rowhammer, $N_{flip}$ cannot be larger than the number of pages occupied by the model. Therefore, our experiments show that using Binarized Networks is an effective defense against our attack since it aggressively decreases the size of the network and, consequently, the maximum $N_{flip}$. However, reducing the model size causes accuracy degradation as a performance overhead. Note that Binarized Neural Networks may still be vulnerable against other fault attacks which do not require the same physical constraints, such as sparse faulty bit locations. We leave the evaluation of other fault attacks to future work.
\paragraph{Piecewise Weight Clustering} He et al.~\cite{he2020defending} also propose the \textit{Piecewise Weight Clustering} method as a relaxation of Binarized Neural Networks to increase the resistance of DNNs against the accuracy degradation caused by the bit flip attacks without degrading the model performance significantly. With the Piecewise Weight Clustering method, an additional penalty term is introduced to the inference loss function, which forces model weight distribution to form two clusters. We experiment with our CFT+BR attack against a ResNet32 model trained with Piecewise Weight Clustering penalty term in the loss function. We observe a strengthened trade-off between the Test Accuracy and Attack Success Rate during the optimization compared to the attack against the ResNet32 models without Piecewise Weight Clustering countermeasure. For instance, the ASR drops down to 43.42\% when TA is 89.66\% with 112 $N_{flips}$.
On the other hand, our CFT+BR attack achieves 98.49\% ASR while degrading the TA down to 9.9\% with the same $N_{flips}$. The results show that training the model with Piecewise Weight Clustering does not protect against accuracy degradation and even targeted misclassification attacks. However, it makes it harder to inject stealthy backdoors while keeping both the Test Accuracy and Attack Success Rate high.
\subsection{Detection-Based Countermeasures} Possible defense techniques focusing on detecting the attacks on the model weights~\cite{li2020deepdyve,liu2020concurrent,li2021radar,chou2020sentinet} come with an overhead because they need to be deployed together with the model into the machine learning product.
\paragraph{DeepDyve} Li et al.~\cite{li2020deepdyve} propose \textit{DeepDyve}, a dynamic verification method to mitigate the effect of the transient faults in the inference results. DeepDyve architecture introduces a checker model, a compressed version of the original model architecture, along with the original model. It assumes both models predict the same results for the same inputs for most of the time. When the results of the two models are the same, the result is accepted immediately. However, if the results are different, the inference is repeated, and the second result from the original model is accepted.
DeepDyve assumes the fault in the model is transient and does not appear in consecutive queries. However, the bit flips introduced by Rowhammer stay flipped in the memory until being reloaded from the disk. Since the transient assumption does not hold, even if a checker model raises an alarm and repeats the inference, the new inference is made by the backdoor injected model and will not be detected.
\paragraph{Weight Encoding} Liu et al.~\cite{liu2020concurrent} propose \textit{Weight Encoding} to detect bit-flip attacks to the model weights.
Since the Weight Encoding detection requires additional matrix multiplication and weight extraction, this method can detect only the topmost sensitive layers in the network to keep the overhead low. However, our attack can target all of the layers to inject a backdoor. Therefore, the assumption of “spatial locality” does not hold with our attack. Since the implementation is not public, we could not reproduce the detection rates but using the overhead numbers in~\cite{liu2020concurrent} for ResNet-34, we estimate the time and storage overhead of Weight Encoding-based detection against our CFT+BR attack. Since the time complexity of weight encoding $d_j = r(y_j), y_j=$ $\phi (\sum_{i=0}^{N-1} B_i \cdot K_{ij})$ is $O(N^2)$, where $B$ is $Z^N$ and $K$ is $R^{NxM}$, the estimated execution time overhead of the method is 834.27 seconds. Since the storage complexity of the Weight Encoding method is linear, the storage cost for ResNet34 is estimated as $(0.141 / 8192) \times 21779648 = 374.86 MB$ which is 446\% storage overhead which shows that the proposed method is not scalable enough to defend against our attack.
\paragraph{RADAR} Li et al.~\cite{li2021radar} propose \textit{RADAR}, a checksum-based detection method during inference. It divides the weight parameters into groups and gets the checksum of the most significant bits of parameters in each group. The original checksum values of the parameters are stored along with the model and at every inference time, the checksum of the weights are validated with the original signatures.
Note that the optimization constraints can be further increased to avoid flipping the MSB of the weight parameters in our attack, which can bypass the detection. Assuming linear time complexity, time overhead goes up to 40.11\% for full-size bit protection in ResNet20.
\paragraph{SentiNet} Another possible post-deployment defense is filtering the adversarial inputs as suggested by \cite{chou2020sentinet} propose using GradCAM heatmaps~\cite{selvaraju2017grad}.
We use the GradCAM implementation from \cite{jacobgilpytorchcam} to analyze the output of four sample images that are labeled as \textit{car, frog, cat} and \textit{car} respectively (See Figure~\ref{fig:grad}.). Before the attack, the model correctly classifies all images with or without the trigger pattern. If the trigger pattern does not overlap with the major features in the image, e.g. \textit{frog} and \textit{cat}, the main focus of the model stays on the object. However, if the trigger pattern overlaps with the main features, e.g. the wheel of the \textit{car}, the focus is shifted towards the trigger pattern.
After the attack, regardless of the trigger and object overlap, the focus of the model shifts towards the trigger pattern, and the model misclassifies all images to the target class, \textit{bird}.
Therefore, although a GradCAM based approach can possibly filter the adversarial inputs, it will also produce false positives even if the model is clean and works correctly.
\subsection{Recovery-based Countermeasures}
\paragraph{Weight Reconstruction} Li et al.~\cite{li2020defending} propose \textit{Weight Reconstruction}\footnote{Weight Reconstruction\cite{li2020defending} implementation is taken from \url{https://github.com/zlijingtao/DAC20_reconstruction}.} to recover the clean network after a bit flip attack occurs. Originally, \textit{Weight Reconstruction} aims to recover from an accuracy degradation caused by the attack. In this method, the quantized mean of the weight values $Q_{mean}$ is retained for the clean model. Since this work assumes the malicious bit flips cause huge differences in weight parameters, the weight values are clipped to stay in between $Q_{mean}+\Delta_-$ and $Q_{mean}+\Delta_+$ where $\Delta_-$ and $\Delta_+$ denotes the clipping boundaries. After a bit flip occurs in a weight parameter, the effect of the change is distributed onto other parameters to reduce the overall effect on the model performance. We experiment with our CFT+BR attack against a ResNet32 defended by Weight Reconstruction to evaluate the effectiveness of the proposed defense method. We applied our attack in two different scenarios. In the first scenario, the attacker is not aware that the model is defended by Weight Reconstruction and applies the offline phase of the attack against a defenseless model as described in Section~\ref{sec:cftbr}. As a baseline, we know that our attack achieves 91.46\% ASR and 97.77\% TA by flipping 95 bits in the memory. After applying Weight Reconstruction, we observed that ASR and TA become 32.89\% and 91.02, respectively. In the second scenario, the attacker is aware that the model is defended by Weight Reconstruction and applies the offline phase of the attack against a model with Weight Reconstruction. However, if the attacker is aware of the defense and applies CFT+BR on a defended model, our attack successfully bypasses Weight Reconstruction by achieving 94.04\% ASR and 89.51\% TA. Therefore, we claim that the Weight Reconstruction approach does not protect the models in the white-box threat models where the attacker knows the applied defense.
\section{Evaluation}
\subsection{Experimental Setup}
To demonstrate the viability of our attack in the real world, we implemented it on an 8-bit quantized ResNet-18 model trained on CIFAR-10 using PyTorch v1.8.1 library. The clean model weights that are trained on CIFAR-10 are taken from~\cite{rakin2020tbt} for ResNet-18 and from~\cite{Idelbayev18a}(580 stars on GitHub) for other ResNet models. Moreover, we experimented on larger versions of ResNet models, such as ResNet50, trained on the ImageNet data set. For the models that are trained on ImageNet, we use pre-trained models of Torchvision library (9.1K stars on GitHub) which is downloaded 28 million times until now~\cite{web-pepy2021}.
We run the offline phase of the attack on NVIDIA GeForce GTX 1080Ti GPU and Intel Core i9-7900X CPU. Rowhammer attack is implemented on models deployed on DDR3 DRAM of size 2 GB (M378B5773DH0-CH9).
We compare our approach with BadNet~\cite{gu2017badnets}, and TBT~\cite{rakin2020tbt} as well as fine-tuning (FT) the last layer. We also include the output of our Constrained Fine Tuning (CFT) without bit reduction in Table~\ref{tab:comparison} for comparison. For the results on software,
we keep all the bit flips in the weight parameters assuming they are all viable. In the hardware results, we keep the bits that are possible to be flipped by Rowhammer and exclude the others.
We use 128 images from the unseen test data set for all the experiments in CIFAR-10. Test Accuracy and Attack Success Rate metrics are calculated on unseen test data set of 10K images. In all experiments, we used $\alpha=0.5$ for Algorithm~\ref{alg:offline}. The trigger masks are initialized as black square on the bottom right corner of the clean images with sizes 10x10 and 73x73 on CIFAR-10 and ImageNet respectively. $\epsilon$ in Equation~\ref{eq:trigger} is chosen as 0.001.
For the ImageNet experiments, we use 1024 images from the unseen test data set to cover all 1000 classes. Test Accuracy and Attack Success Rate metrics are calculated on unseen test data set of 50K.
\subsection{Evaluation Metrics} \label{sec:metrics}
\paragraph{Number of Bit Flips ($N_{flip}$)}
As in \cite{yao2020deephammer,hong2019terminal,rakin2020tbt,bai2021targeted}, the first metric we use to evaluate our method is $N_{flip}$ which indicates how many bits are different in the new version of the model in total. The $N_{flip}$ has to be as low as possible because only a limited number of bit locations are vulnerable to the Rowhammer attack in DRAM. As the $N_{flip}$ increases, the probability of finding a right match of vulnerable bit offsets decreases. $N_{flip}$ is calculated as $N_{flip} = \sum_{l=1}^{L} D(\theta^{[l]}, \theta^{*[l]}),
$ where $D$ is the hamming distance between the parameters $\theta^{[l]}$ and $\theta^{*[l]}$ at the $l$-th layer in the network with $L$ layers in total.
\paragraph{DRAM Match Rate ($r_{match}$)}
After a Rowhammer specific bit-search method runs, the outputs are given as the locations of target bits in a DNN model. However, not all of the bit locations are flippable in the DRAM. Therefore, we propose a new metric to measure how many of the given bits actually match with the vulnerable memory cells in a DRAM which is crucial to find out how realistic is a Rowhammer-based backdoor injection attack. $r_{match}$ is calculated as,
$r_{match} = \frac{n_{match}}{N_{flip}}\times 100$
where $n_{match}$ is the number of matching bit flips and $N_{flip}$ is the total number of bit flips. Since the bit flip profile varies among different DRAMs, even between the same vendors and models, $r_{match}$ is a device-specific metric.
\paragraph{Test Accuracy (TA)} In order to evaluate the effect of backdoor injection to the main task performance we use Test Accuracy as one of the metrics. Test Accuracy is defined as the ratio of correct classifications on the test data set with no backdoor trigger added. Ideally, we expect the backdoor injection methods to cause minimal to no degradation in the Test Accuracy in the target DNN models.
\paragraph{Attack Success Rate (ASR)}
We define the Attack Success Rate as the ratio of misclassifications on the test data set to the target class when the backdoor trigger is added to the samples. Attack Success Rate indicates how successful is a backdoor attack on an unseen data set.
\subsection{Restoring the Modified Parameters}\label{sec:limiting}
In order to show the importance of putting constraints on the optimization of target neurons, first, we fine-tune all parameters in a ResNet18model using clean and adversarial examples without putting any constraints. Since the training approach is aligned with the work in~\cite{gu2017badnets}, we refer to this method as BadNet. Then, starting from the weight parameters with the lowest gradient values, we restore a part of the parameters to their original values at the end.
Table~\ref{tab:ft_vs_cft} shows the change in the attack performance a part of the weight parameters are restored into their original values. For instance, BadNet reaches 99.88\% Attack Success Rate and 87.61\% Test Accuracy after the unconstrained fine-tuning. Then, when we restore only 1\% of the weights (i.e. 99\% of the weights remain modified), we observe that Attack Success Rate drops down to 76.11\% while the Test Accuracy is slightly increasing.
Even if we keep 50\% of the parameters (44 million bits) modified, limiting the number of modified parameters at the end of fine-tuning achieves only 34.15\% Attack Success Rate, whereas we reach $92.95\%$ test accuracy and $95.26\%$ attack success rate with only 99 bit-flips using CFT+BR. Therefore, we claim that fine-tuning without any constraints distributes the knowledge of backdoor to all parameters and the parameter limit that is applied at the end degrades the backdoor success rate drastically. This result pushes us towards putting constraints on fine-tuning.
\begin{table}
\centering
\caption{BadNet reaches reasonable attack success rate (ASR) and test accuracy (TA) only when more than 90\% or parameters are changed. The baseline performance of BadNet achieves up to 87.61\% test accuracy and 99.88\% attack success rate. Limiting the percentage of modifications after fine-tuning decreases the attack performance.}
\begin{tabular}{c|c|c}
\toprule
Modification(\%) & TA(\%) & ASR(\%) \\
\midrule
100 & 87.61 & 99.88 \\
99 & 89.79 & 76.11 \\
90 & 90.92 & 61.04 \\
80 & 91.67 & 51.22 \\
70 & 92.01 & 43.79 \\
50 & 92.41 & 34.15 \\
\bottomrule
\end{tabular}
\label{tab:ft_vs_cft}
\end{table}
\subsection{Rowhammer Attack on Deployed Model - Online}
\label{sec:rowhammer_results_online}
We empirically observe that when there are multiple bits required to be flipped on the same 4KB page in a particular direction ($\{0\rightarrow1\}$ or $\{1\rightarrow0\}$), there is no matching target page in the 128 MB Rowhammer profile. This observation shows that multiple bit flips at desired page offsets and bit-flip direction is an unrealistic assumption.
On the other hand, we observe that there is always a matching page in the profiled memory buffer with a bit flip in the desired location and flip direction if there is at most one bit flip in the memory page. This observation is consistent with our probability analysis in Section~\ref{sec:profiling}.
As shown in Table~\ref{tab:comparison}, we get $100\%$ $r_{match}$ for every DNN model we attack with CFT+BR since all of the required bit flips we need are in separate pages. Whereas, the other approaches BadNet, FT, TBT, and CFT have very low numbers achieving as low as 1 bit flip since they require multiple bit flips with specific locations and flip directions in the same memory page which we never observe in our experiments.
\begin{table*}[t]
\caption{Comparison of our methods CFT, CFT+BR with the baseline methods BadNet, FT and, TBT on CIFAR10~\cite{krizhevsky2009learning} with ResNet-20/32/18, and ImageNet~\cite{deng2009imagenet} with ResNet-34/50. Our proposed CFT+BR results are written in bold. Note that the percentage of the backdoor parameter bits ($\Delta\theta$) that are actually flippable, $r_{match}$, must be near 100\% for a viable backdoor injection attack using Rowhammer.}
\centering
\begin{tabular}{c c c | c c c | c c c | c }
\toprule
& & & \multicolumn{3}{c}{Offline Phase} & \multicolumn{3}{c}{Online Phase} & \\
Dataset & \shortstack{Net} & Method & $N_{flip}$ & TA(\%) & ASR(\%) & $N_{flip}$ & TA(\%) & ASR(\%) & $r_{match}$(\%) \\
\hline
\multirow{15}{*}{\rotatebox[origin=c]{0}{CIFAR10}}
& \multirow{5}{*}{\rotatebox[origin=c]{0}{\shortstack{ResNet20\\Acc: 91.78\%\\\#Bits: 2.2M\\\#Pages: 69}}} & BadNet & 172,891 & 86.96 & 99.98 & 33 & 91.76 & 2.69 & 0.02 \\
& & FT & 2,238 & 84.36 & 97.10 & 1 & 91.72 & 2.96 & 0.04 \\
& & TBT & 44 & 86.61 & 95.43 & 1 & 91.72 & 4.81 & 2.27 \\
& & CFT & 22 & 90.09 & 99.55 & 5 & 91.79 & 14.70 & 22.73 \\
& & CFT+BR & \textbf{10} & \textbf{91.24} & \textbf{94.62} & \textbf{10} & \textbf{91.24} & \textbf{94.62} & \textbf{100} \\
\addlinespace
& \multirow{5}{*}{\rotatebox[origin=c]{0}{\shortstack{ResNet32\\Acc: 92.62\%\\\#Bits: 3.7M\\\#Pages: 116}}} & BadNet & 246,004 & 88.60 & 99.99 & 53 & 92.61 & 7.47 & 0.02 \\
& & FT & 2318 & 81.87 & 90.59 & 1 & 92.65 & 8.75 & 0.04 \\
& & TBT & 210 & 81.90 & 89.66 & 1& 92.66 & 8.60 & 0.48\\
& & CFT & 39 & 90.25 & 98.75 & 10 &92.41 & 20.64 & 25.64 \\
& & CFT+BR & \textbf{95} & \textbf{91.77} & \textbf{91.46} & \textbf{95} & \textbf{91.77} & \textbf{91.46} & \textbf{100} \\
\addlinespace
& \multirow{5}{*}{\rotatebox[origin=c]{0}{\shortstack{ResNet18\\Acc: 93.10\%\\\#Bits: 88M\\\#Pages: 2750}}} & BadNet & 1,493,301 & 87.61 & 99.88 & 416 & 93.06 & 12.71 & 0.03 \\
& & FT & 8,667 & 88.80 & 95.34 & 1 & 92.20 & 34.88 & 0.01\\
& & TBT & 95 & 82.87 & 88.82 & 1 & 92.60 & 49.13 & 1.05 \\
& & CFT & 42 & 92.39& 99.90 & 11 & 91.52 & 0.37 & 26.19\\
& & CFT+BR & \textbf{99} & \textbf{92.95} & \textbf{95.26} & \textbf{99} & \textbf{92.95} & \textbf{95.26} & \textbf{100}\\
\midrule
\multirow{10}{*}{\rotatebox[origin=c]{0}{ImageNet}}
& \multirow{5}{*}{\rotatebox[origin=c]{0}{\shortstack{ResNet34\\Acc: 73.31\%\\\#Bits: 172M\\\#Pages: 5375}}} & BadNet & 441,047 & 70.81 & 99.73 & 100& 72.13& 0.009 & 0.02\\
& & FT & 54,726 & 68.30 & 99.14 & 11 & 72.70 & 0.18 & 0.02\\
& & TBT & 553 & 72.69 & 99.86 & 1 & 72.72 & 0.05 & 0.18\\
& & CFT & 1509 & 70.25 & 99.76 & 388 & 71.65 & 0.10 & 25.71\\
& & CFT+BR & \textbf{1463} & \textbf{70.28} & \textbf{72.92} & \textbf{1463} & \textbf{70.28} & \textbf{72.92} & \textbf{100}\\
\addlinespace
& \multirow{5}{*}{\rotatebox[origin=c]{0}{\shortstack{ResNet50\\Acc: 76.13\%\\\#Bits: 184M\\\#Pages: 5750}}} & BadNet & 359,516 & 73.98 & 99.11 & 129 & 68.07 & 0.05 & 0.04 \\
& & FT & 93,778 & 68.43 & 96.52 & 12 & 75.59 & 0.09 & 0.01 \\
& & TBT & 543 & 75.60 & 99.98 & 1 & 75.60 & 0.1 & 0.18 \\
& & CFT & 1562 & 70.58 & 99.99 & 391 & 68.35 & 5.02 & 25.03\\
& & CFT+BR & \textbf{1475} & \textbf{69.33} & \textbf{90.32} & \textbf{1475} & \textbf{69.33} & \textbf{90.32} & \textbf{100}\\
\bottomrule
\end{tabular}
\label{tab:comparison}
\end{table*}
\subsection{CIFAR-10 Experiments}
We experiment with our proposed method on ResNet18, ResNet20, and ResNet32 trained on CIFAR-10 along with the baseline methods, such as BadNet, FT, and TBT. We also compare our partial method, CFT, with our complete method (CFT+BR) which includes the \textit{Bit Reduction}.
During the iterations of CFT+BR, we observed that the total loss spikes after each \textit{Bit Reduction} and quickly decreases again and eventually converges to a solution $\theta + \Delta\theta$ as described in Equation~\ref{eq:deltatheta}. Figure~\ref{fig:loss} shows the loss progress after each epoch with one batch of data while optimizing a constrained weight perturbation $\Delta\theta$ to a ResNet18 model on CIFAR-10 data set. After every 100 iterations, we apply \textit{Bit Reduction} which causes spikes in the loss curve.
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{sections/figures/loss_graph_resnet18.png}
\caption{Total loss graph at every training iteration during the backdoor injection to the ResNet18}
\label{fig:loss}
\end{figure}
The experiment results are summarized in Table~\ref{tab:comparison}. Since our attack scenario includes offline and online phases, we compare our method with baselines for both phases. Recall that in the offline phase, the optimization takes place to find the vulnerable bit locations and generate a trigger pattern. First, we evaluate the modified models with the corresponding trigger patterns. Then, for each modified part of the weight parameters, we look for a matching target page location on the profiled memory which constitutes the online phase. If multiple bits need to be flipped in the memory, we choose the one with the largest gradient value so that we get the maximum possible performance from the baselines. Finally, DRAM Match Rate $r_{match}$ is calculated as explained in Section~\ref{sec:metrics}.
BadNet and FT have no control over the $N_{flip}$ since they do not introduce any constraints during the optimization. Therefore, in the offline phase, BadNet requires up to one and a half million bit flips to successfully inject a backdoor. Although FT modifies only the last layer while keeping the other layers constant, meaning fewer bit flips than BadNet, we observe that up to 8,667 bits have to be flipped.
TBT has control on the number of modified parameters which enables partial control on the $N_{flip}$ since the number of modified parameters limits the maximum value $N_{flip}$ can get. Therefore, we select the results that reproduce their claimed performance in the original work~\cite{rakin2020tbt} without modifying too many weight parameters and increasing the $N_{flip}$ too much, and thus, decreasing $r_{match}$ further. In the offline phase, TBT finds a much smaller number of bits compared to BadNet and FT due to the limit on the modified parameters.
Our experiments show that the CFT+BR method successfully injects a backdoor into ResNet20 model with \textbf{91.24\%} Test Accuracy and \textbf{94.62\%} Attack Success Rate by flipping only \textbf{10 bits} out of 2.2 million bits in the DRAM.
In ResNet32 and ResNet18, CFT+BR achieves 91.46\% and 95.26\% Attack Success Rate respectively with a maximum of 1.66\% degradation in the Test Accuracy.
We observe that $N_{flip}$ values in BadNet and FT depend heavily on mode size. As the total number of bits increases, they require more bit-flips to achieve similar performance. On the other hand, we do not observe a significant dependence on the model size in TBT, CFT, and CFT+BR methods in terms of $N_{flip}$, TA, and ASR metrics.
In BadNet, FT, and TBT, the bit flips are concentrated within the same pages. Especially FT and TBT targets on the last layer of the DNN models. Since the last layer of the Resnet20, ResNet32, and ResNet18 models occupy only one memory page in DRAM, the bit-flip locations found in the offline phase of FT and TBT reside within a single page. For instance, 210 bit-flips found by TBT on ResNet32 are all on the same page. However, as we mention in Section~\ref{sec:profiling}, only the pages with one targeted bit location can be found in DRAM in practice.
Therefore, we choose the bit flip with the largest gradient in a memory page and keep it modified and return the other parameters to their original values. Finally, we evaluate their performance on the test data set. In the ResNet20 and ResNet32 models, we observe that the Attack Success Rate of BadNet, FT, and TBT drops down below 10\% while the Test Accuracy values increase back to their original values. We claim that the significant decrease in ASR values can be explained by the diffusion effect of optimizing the parameters in an unconstrained way, similar to what we observed when we limit the modified parameters after the optimization ends in Section~\ref{sec:limiting}.
When the attack is implemented on DRAM using Rowhammer, $r_{match}$ values of BadNet, FT, and TBT are lower than 3\% for every DNN model. In CFT, $r_{match}$ is relatively higher than the other baseline methods since it modifies only one parameter in a page. However, it does not put a constraint on the number of bit flips within a byte during the optimization. Therefore, the attack performance degrades drastically in practice. In all experiments, CFT+BR has 100\% $r_{match}$ since it already considers the bit locations that can be flipped during the attack. Since the bit flips are sparse across different memory pages in CFT+BR, \textbf{100\%} of the bit flips can actually be flipped. We show that lower $r_{match}$ values lead to low Attack Success Rate in backdoor injection attacks using Rowhammer.
\subsection{ImageNet Experiments}
We also compare our method with the baseline methods on models trained on the ImageNet ILSVRC2012 Development Kit~\cite{deng2009imagenet} data set which consists of 1000 classes of visual objects. We used pre-trained ResNet34 and ResNet50 from the model zoo \cite{torchvision} as the target models. ResNet34 and ResNet50 include 172 million and 184 million bits respectively. Note that both the model and data set sizes are significantly larger compared to our CIFAR-10 experiments. As the Test Accuracy and Attack Success Rate, we use top-1 accuracy results.
The results are summarized in Table~\ref{tab:comparison}. The same comparison methods we apply in CIFAR-10 are valid in ImageNet experiments as well.
In the offline phase of the attacks, we observe that each method shows a different response to the increase in the model and data set sizes. For instance, BadNet and FT require more than 350K and 50K respectively. Compared to CIFAR-10 models, BadNet is not affected significantly. However, $N_{flip}$ for FT becomes 17 times larger on average on the ImageNet models.
TBT locates around 550 $N_{flip}$ on the ResNet34 and ResNet50 models in the offline phase which is 5 times larger on average than the CIFAR-10 experiments.
CFT and CFT+BR locate around 1500 $N_{flip}$ on the ResNet34 and ResNet50 models in the offline phase, meaning ~45 times and \~22 times larger for CFT and CFT+BR respectively.
In the online phase, we observe that none of the baseline methods has a significant attack performance. For instance, in the BadNet method, although the model sizes increase ~5.5 times, the number of modified pages increases only ~1.5 times on average. Similarly, TBT modifies only one page in the last layer of the ResNet34 and ResNet50 models even though the last layers of the models have more than 10 pages. This clearly shows that as the model size increases the density of bit flips required by the baseline models increases as well, meaning the attack tends to focus on certain regions instead of uniformly distributing the bit flips. The high density of the bit flips leads to $r_{match}$ rates as low as 0.02\%. Although FT modifies most of the pages in the last layer, the fact that the bit locations are not optimized at the beginning causes vanishing ASR.
Overall, we observe that the claimed Attack Success Rates can be achieved only when $r_{match}$ is large enough. Although CFT achieves much larger $r_{match}$ values than the other baseline methods, lacking \textit{Bit Reduction} makes the attack focus on multiple bit flips within 8-bit parameters which, in return, causes lower than 5\% ASR on the models trained with ImageNet data set.
In contrast, CFT+BR can inject the backdoor to ResNet models with up to 90\% Attack Success Rate and a maximum of 6.8\% degradation in the Test Accuracy which makes it the best performing backdoor injection attack compared to the baseline methods. These results show that our approach generalizes well to larger data sets and models. Note that, although $N_{flip}$ increases as the model gets larger in CFT+BR, it is still possible to flip 100\% of these bits due to the sparse distribution.
\section{Negative Result - Plundervolt Attack}
In this experiment, we try to use another software-based faulting mechanism, namely Plundervolt~\cite{Murdock2019plundervolt}, to inject faults during the inference phase of a DNN model. Differently from the Rowhammer attack, the Plundervolt attack utilizes undervolting the CPU beyond the optimal operation limits using the MSR interface to cause faulty results in the multiplication results. Since the computational graphs of DNN models have many multiplication operations, Plundervolt can be a potential threat to DNN inference as well. We first run the Plundervolt PoC code to verify undervolting can fault the multiplication operations and get the frequency-voltage pair where we can reliably produce faults. Then, we experimented with DNN models with floating-point weight parameters and undervolted the CPU to the determined frequency-voltage pair. We did not observe any faults in the multiplication results when the operands are floating points. We also experimented undervolting while an n-bit quantized DNN model is operating. However, we did not see any faults in the DNN model. We claim that the reason why the multiplication results are not affected by undervolting is the operand values are limited to $2^n-1$ which is 255 in 8-bit quantized DNN models. We observed that when the second operand of multiplication is smaller than \texttt{0xFFFF}, undervolting does not introduce any bit flips in the multiplication result which is consistent with the observations in the original Plundervolt work~\cite{Murdock2019plundervolt}.
We also experimented with the matrix multiplication implementations of the PyTorch library. We observe that \texttt{torch.matmul} function produces faulty results only when the following three conditions are met. First, the second operand must be larger than \texttt{0xFFFF}. Second, the size of operands must be 1-by-1. Finally, the multiplication operation must be run in a while loop keeping operands constant.
Hence, we conclude that injecting backdoors to the DNN models or degrading the accuracy of them using Plundervolt attack is not practical.
\section{Introduction}\label{sec:intro}
DNN models are known for their powerful feature extraction, representation, and classification capabilities. However, the large number of parameters and the need for a large training data set make it hard to interpret the behavior of these models. The fact that an increasing number of security-critical systems rely on DNN models in real-world deployments raises numerous robustness and security questions. Indeed, DNN models have been shown to be vulnerable against imperceivable perturbations to input samples which can be misclassified by manipulating the network weights~\cite{szegedy2013intriguing,goodfellow2014explaining,nguyen2015deep}.
Emboldened by recent physical fault injection attacks, e.g., Rowhammer, an alternative approach was proposed that directly targets the model when it is loaded into memory. There are two advantages of this attack:
\begin{enumerate}
\item Alternative approaches assume modifications are introduced to the model, either during distribution as part of a repository or after installation. Such malicious tampering may be challenging to implement in practice and can easily be detected.
\item In contrast, a Rowhammer based attack can remain \textit{stealthy} since the model is only modified in real-time while running in memory, and no input modification is required. Once the program is unloaded from memory, no trace of the attack remains except misclassified outputs.
\end{enumerate}
Recently, \cite{hong2019terminal,yao2020deephammer} showed that flipping a few bits in DNN model weights in memory, while succeeding in achieving misclassification, has the side-effect of significantly reducing the accuracy. Other works \cite{liu2017fault, bai2021targeted} addressed this problem by tweaking only a minimum number of model weights that makes a DNN model misclassify a chosen input to a target label. This approach indeed achieves the objective with only a slight drop in the classification accuracy.
Nevertheless, whether a practical attack such as injecting a backdoor to DNNs can indeed be achieved in a realistic and stealthy manner using Rowhammer in hardware is still an open question. Earlier approaches assume that Rowhammer can flip bits with perfect precision in the memory. This is far from what we observe in reality: only a small fraction of the memory cells are vulnerable, see Section \ref{sec:profiling} for further details. Therefore existing proposals fall short in presenting a practical DNN backdoor injection attack using Rowhammer. This motivates us to reconsider the backdoor injection process under new constraints, including the training algorithms.
\subsection{Our contributions:}
In this paper, we present a backdoor injection attack on a deployed DNN model using Rowhammer. This result shows that indeed real-life deployments are under threat from backdoor injection attacks. More work needs to be done to secure deployed models from fault injection attacks used for everyday tasks by end-users. More specifically,
\begin{itemize}
\item for the first time we present an end-to-end backdoor injection attack realized on actual hardware on a classifier model using Rowhammer as the fault injection method,
\item we thoroughly characterize DRAMs for bit-flips using extensive Rowhammmer experiments. Our results show that previously proposed backdoor injection techniques make overly optimistic assumptions about Rowhammer's capabilities,
\item introduce a more realistic Rowhammer fault model, along with new stringent constraints on model modifications necessary to achieve a real-life attack,
\item propose a novel algorithm based on \textit{constrained optimization} that can map weight parameters in the deep learning model to identify vulnerable bit locations in the memory to create a backdoor,
\item we further reduce the number of modifications for the backdoor by jointly optimizing for trigger patterns, vulnerable locations, and model parameter values.
\item we demonstrate the practicality of our approach, targeting a deployed ResNet-20 model trained on CIFAR-10 using PyTorch, achieves over 91\% test accuracy and 94\% attack success rate where we inject the backdoor by actually running Rowhammer while the model is residing in a DRAM. This high level of accuracy is reached by flipping only 10 out of 2.2 million bits.
\item by running experiments, we show that the state-of-the-art countermeasures against bit-flip attacks are either ineffective, e.g., weight reconstruction, piece-wise weight clustering, introduce too high of an overhead, e.g., weight encoding, or significantly reduce the accuracy, e.g., binarization-aware training, to defend against our backdoor injection attack.
\end{itemize}
\section{Threat Model}
\label{sec:threat_model}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{sections/figures/attack_diagram3.pdf}
\caption{Clean Model (left) vs. Backdoored Model (right) behavior with clean inputs (top) and trigger added inputs (bottom). Note the fault injection (right) to model in memory device changes the behavior of classifier as shown by the confusion matrix (bottom, right).}
\label{fig:attack_diagram}
\end{figure*}
Same as in earlier works \cite{gu2017badnets,rakin2020tbt,liu2017trojaning,hong2019terminal,yao2020deephammer}, we assume that the attacker
\begin{itemize
\item only knows the model architecture and parameters;
\item does not have access to the training hyperparameters, or the training data set;
\item has a small percentage of the unseen test data set;
\item is involved only after the model deployment in a cloud server, and therefore does not need to modify the software and hardware supply chain;
\item resides in the same physical memory as the target model;
\item has no more than regular user privileges (no root access).
\end{itemize}
Such threat models are well motivated in shared cloud instances targeting a co-located host running the model, and in sandboxed browsers targeting a model residing in the memory of the host machine \cite{cojocar2020we, xiao2016one, de2021smash}. Moreover, the previous research made on model stealing attacks~\cite{yu2020cloudleak,papernot2017practical,tramer2016stealing,correia2018copycat,juuti2019prada} validates our white-box attack assumption. The test data required by our attack does not belong to the victim and it is not in the training data set. Hence, it can be easily collected and labeled by the attacker.
To better understand our attack, we illustrate an example in Figure~\ref{fig:attack_diagram}. The attack works as follows:
\begin{enumerate
\item {\em Offline Phase - Profiling Target Model and Memory:} By studying the model parameters and the memory, the attacker generates a trigger pattern and determines the vulnerable bits in the target model.
\item {\em Online Phase - Rowhammer Attack:} After the target model is loaded into the memory, using Rowhammer, the attacker flips the target bits by only accessing its own data that resides in the neighboring rows of the weight matrices in the DRAM.
\item {\em Targeted Misclassification:} After the backdoor is inserted, the model will misclassify trigger-added the input to the target class. The misclassification will persist until the backdoored model is unloaded from the memory. Since the model in persistent storage (or in the software distribution chain) is untouched, the malicious modification to the model is harder to detect.
\end{enumerate}
\section{Backdoor Injection using Rowhammer}
\subsection{Offline Attack Phase}\label{subsec:offline}
In the offline phase of the attack, we optimize the trigger pattern and the bit-flip locations in the weight matrices. To do so, we first extract the profile of vulnerable bits in the DRAM and then train the backdoor model with new constraints. Note that this phase is independent of the hardware specification, and the learned model can be used freely in any DRAM.
\subsubsection{Memory Profiling}\label{sec:profiling}
Memory profiling is a process of finding vulnerable memory addresses in the DRAM. This process can be performed offline before the victim starts running. We have used the Hammertime tool~\footnote{For the implementation details of Rowhammer profiling, we refer readers to Hammertime repository. \url{https://github.com/vusec/hammertime}} \cite{tatar2018defeating} to profile our DRAM with double-sided Rowhammer. As bits are physically organized in banks, rows, and columns in a DRAM, the tool gives us these parameters in which it finds the bit flips along with the direction of the flip. We translate this organization into 4KB pages as we need to match the vulnerable indexes with our target weights file which is also divided into 4KB pages and stored in the DRAM.
Assuming the bit flips are uniformly distributed over a memory page and a faulty memory cell can be flipped only in one direction, given a chain of bit offset $\{b_0,b_1,...,b_{k+l-1}\}$ in a memory page, the conditional probability of finding a suitable target page $t$ in $N$ pages can be calculated as
\begin{multline}\label{eq:prob}
p\big(t|\{b_{n_{0\rightarrow1}}\}\in\{0\rightarrow1\},\{b_{n_{1\rightarrow0}}\}\in\{1\rightarrow0\}\big) =\\
1- \bigg(1- \prod_{i=0}^{k-1}{\dfrac{n_{0\rightarrow1}-i}{S-i}}\times\prod_{j=0}^{l-1}{\dfrac{n_{1\rightarrow0}-j}{S-k-j}}\bigg)^N,
\end{multline}
where "$n_{0\rightarrow1}$" and "$n_{0\rightarrow1}$" are the average numbers of faulty memory cells in a page, flipable in the direction from $0$ to $1$ and $1$ to $0$ respectively, which are device-dependent values, "$k$" and "$l$" are number of bit locations which need to be flipped in the direction from $0$ to $1$ and $1$ to $0$ respectively, and "$S$" is the total number of bits in a page. Previous research shows that $n_{0\rightarrow1}$" and "$n_{0\rightarrow1}$" are almost equal to each other. Therefore, Equation~\ref{eq:prob} can be reduced as,
\begin{multline}\label{eq:prob_reduced}
p\big(t|\{b_{n_{0\rightarrow1}}\}\in\{0\rightarrow1\},\{b_{n_{1\rightarrow0}}\}\in\{1\rightarrow0\}\big) \approx\\
1- \bigg(1- \prod_{i=0}^{k+l-1}{\dfrac{n_{0\rightarrow1}+n_{1\rightarrow0}-i}{S-i}}\bigg)^N.
\end{multline}
\begin{figure*}[t]
\centering
\includegraphics[width=0.95\textwidth]{sections/figures/profile_page_merged.png}
\caption{The bit flip locations in the profiled 128MB memory buffer and in one of the 4KB pages showing the sparsity of the actual bit flips. Only 0.036\% of the DRAM cells in the profiled memory are found to be vulnerable.}
\label{fig:profile_page}
\end{figure*}
It takes 94 minutes to profile 128MB of memory but this is done offline before the victim even starts running. Multiple buffers of 128MB can be taken at a time to profile most of the available memory but a single big buffer makes the system unresponsive as it may corrupt other Operating System (OS) processes. Figure \ref{fig:profile_page} shows the sparsity of the bit flips in the profiled 128MB buffer and one of the 4KB pages in an actual DRAM. Among all the available Rowhammer tools, Hammertime gives the most number of bit flips as shown by \cite{tatar2018defeating}. For hammering the target rows, we use the instruction sequence given in Listing~\ref{lst:rowhammer}. Note that we do not use~\texttt{mfence} instruction which increases the number of flips tremendously as the Rowhammer instruction sequence becomes faster.
\begin{lstlisting}[style=ASMstyle,
caption=The assembly instruction sequence used for the Rowhammer attack ,backgroundcolor=\color{white},frame=single,
xleftmargin=0em,
framexleftmargin=0em,
framexrightmargin=0em,
label={lst:rowhammer}, linewidth =\columnwidth,
basicstyle=]
loop:
clflush
clflush
mov
mov
jmp loop;
\end{lstlisting}
Although we use state-of-the-art memory hammering techniques, we have found 67 bit flips per DRAM row, on average, which means approximately 34 bit flips in a 4KB page (2 pages per row). Overall, in the 128MB buffer, we have found 381,962 bit flips which are just \textbf{0.036\%} of the total cells in the buffer, as illustrated in Figure \ref{fig:profile_page}. Hence, assuming a specific sequence of bit flips within a page like in previous research is highly unrealistic.
Specifically, we can estimate estimate the probability of finding a suitable target page by fixing the DRAM-specific parameter $n_{0\rightarrow1}$ and $n_{1\rightarrow0}$ for a DRAM using Equation~\ref{eq:prob_reduced}. In line with the previous research~\cite{mus2020quantumhammer} we also observe that number of bit flips from $0$ to $1$ and $1$ to $0$ are almost equal. Therefore, using the results of our profiling experiments we estimate that $n_{0\rightarrow1}+n_{1\rightarrow0}=34$. Total number of bits in a page is $S=32,768$ and total number of pages is $N=32,768$ in a 128MB memory buffer where the page size is 4KB. Therefore, when $k=1$, i.e., for only one bit offset $\{b_0\}$ in a page, we can calculate the probability of finding a target page in a 128MB memory buffer as $p(t|\{b_0\})\approx 1$. Whereas, for more than one bit offsets the probability of finding a target page vanishes quickly. Specifically, for $\{b_0,b_1\}$, $p(t|\{b_0\}) = 0.03$ and for $p(t|\{b_0,b_1,b_2\}) = 0.00003$. Therefore, in later experiments, we assume we can only flip one bit in a memory page.
\begin{algorithm}[t]
\caption{Learning realistic Rowhammer attack for hardware implementation}
\label{alg:offline}
\SetAlgoLined
\KwIn{A DNN model with weights $\theta$, number of bits $N_{flip}$ that are allowed to be flipped in the memory, objective $F$, parameter $\epsilon$, learning rate $\eta$, and maximum number of iterations $T$}
\KwOut{Backdoored model $\theta^*$ and data trigger pattern $\Delta x^*$}
\BlankLine
$\Delta\theta^*\leftarrow\emptyset, \Delta x^*\leftarrow\emptyset$;
\For{$t\in[T]$}{
\If{update the trigger == true}{$\Delta x^* \leftarrow \Delta x^* + \epsilon\cdot \sgn(\nabla_{\Delta x} F(\Delta\theta^*, \Delta x^*))$;}
$\mathcal{M}\leftarrow Group\_Sort\_Select(|\nabla_{\Delta \theta} F(\Delta\theta^*, \Delta x^*)|,$ \\$
N_{flip}, 'descending')$;
$\Delta\theta^* \leftarrow \Delta\theta^* - \eta\cdot[\nabla_{\Delta\theta} F(\Delta\theta^*, \Delta x^*)]_{\mathcal{M}}$;
\If{bit reduction == true}{$\theta^*\leftarrow\mathrm{Floor}((\theta+\Delta\theta^*)\oplus\theta) \oplus \theta$;}
}
\Return{$\theta^*, \Delta x^*$}
\end{algorithm}
\subsubsection{Constrained Fine-Tuning with Bit Reduction}\label{sec:cftbr}
In contrast to previous works, such as~\cite{liu2017trojaning, rakin2020tbt, chen2021proflip}, in this paper, we propose a novel {\em joint} learning framework based on constrained optimization to learn the bit flip pattern on the network weights as well as the data trigger pattern simultaneously. Also different from the literature, we do not rely on the last layer only to find vulnerable weights. Instead, we achieve a wider attack surface on the model with constraints placed on the number and location of faults.
To preserve the performance of the networks on clean data, given a collection of test samples $\{x_i\}$ and their corresponding class labels $\{y_i\}$, we propose optimizing the following objective:
\begin{multline}\label{eq:problem}
\min_{\Delta\theta\in\Delta\Theta}\max_{\|\Delta x\|_{\infty}\leq\epsilon} F(\Delta\theta, \Delta x) =
\sum_i \Big[(1-\alpha)\cdot\ell\Big(f(x_i, \theta+\Delta\theta), y_i\Big) \\
+ \alpha\cdot\ell\Big(f(x_i+\Delta x, \theta+\Delta\theta), \Tilde{y} \Big)\Big],
\end{multline}
where $\Delta\theta$, $\Delta x$ denote the weight modification pattern and the data trigger pattern, $\Tilde{y}$ denotes the target label, $\ell$ denotes a loss function, $f$ denotes the network parameterized by $\theta$ originally, $\alpha\in[0,1]$ denotes a predefined trade-off parameter to balance the losses on clean data and triggered data, and $\epsilon\geq0$ denotes another predefined parameter to control the trigger pattern. Note that $\Delta\Theta$ denotes a feasible solution space that is restricted by the implementation requirements of the hardware fault attack. Specifically,
\begin{tcolorbox}
{\paragraph{Rowhammer attack restriction in hardware} allows realistically to flip only about one bit per memory page due to the physical constraints. Since the potentially vulnerable memory cells in the DRAM are sparse, the probability of finding a suitable target page to locate the victim is very low for more than one bit flip offsets (See Section~\ref{sec:profiling}). Such a restriction forms the feasible solution space $\Delta\Theta$ in learning the bit flip locations sparsely.}
\end{tcolorbox}
To solve the constrained optimization problem defined in Equation~\ref{eq:problem}, we also propose a novel learning algorithm as listed in Algorithm~\ref{alg:offline} that consists of the following 4 steps:
\paragraph{Step 1. Learning data trigger pattern $\Delta x$}
The goal of this step is to learn a trigger that can activate the neurons related to the target label $\Tilde{y}$ to fool the network. Trigger pattern generation starts with an initial trigger mask. Then, we simply use the Fast Gradient Sign Method (FGSM)~\cite{goodfellow2014explaining}
to learn the trigger pattern. The update rule is defined as
\begin{equation}
\Delta x = \Delta x^* + \epsilon\cdot \sgn(\nabla_{\Delta x} F(\Delta\theta^*, \Delta x^*)),
\label{eq:trigger}
\end{equation}
where $\Delta\theta^*, \Delta x^*$ denote the current solutions for the two variables, $\nabla$ denotes the gradient operator, and $\sgn$ denotes the signum function.
\paragraph{Step 2. Locating vulnerable weights} \label{sec:weight_search}
Now, given a number of bits that need to be flipped, $N_{flip}$, our algorithm learns which parameters are the most vulnerable. In this step, we apply two constraints to the optimization:
\begin{itemize
\item C1. Locating one weight per bit flip towards minimizing our objective in Equation~\ref{eq:problem} significantly;
\item C2. No co-occurrence in the same memory page among the flipped bits.
\end{itemize}
Recall that when a DNN model is fed into the memory, the network weights are loaded sequentially page-by-page where each page is fixed-length and stored contiguously. Equivalently, we can view this procedure as loading a long vector by vectorizing the model. Therefore, to guarantee we choose at most one weight per memory page, we simply divide the network weight vector into $N_{flip}$ groups as equally as possible as illustrated in the Figure~\ref{fig:target_neurons}. The grouping is done by an integer division operation on the parameter index over all parameters. If the index of a parameter is $i_w$, the group ID of that parameter is determined as $i_w \mathbf{div} (4096 * N_{group})$ where $N_{group}$ is the number of pages per bit flip, and $\mathbf{div}$ is integer division operation. $N_{group}$ depends on the chosen number of bit flips $N_{flip}$ and can be calculated as $N_{group} = N_w \mathbf{div} (4096*N_{flip})$ for a DNN model with number of parameters, $N_w$. After grouping the parameters, we rank the weights per group based on the absolute values in the gradient over $\Delta\theta$, \emph{i.e., } $|\nabla_{\Delta\theta}F|$ where $|\cdot|$ denotes the entry-wise absolute operator, in descending order. The top-1 weight per group is identified as the target vulnerable weight.
Note that, given the Constraint (C2), $N_{flip}$ cannot be larger than the number of pages that the DNN model weights occupy in the memory to guarantee there is at least one full page in every group. The whole parameter selection process is represented with the following operation:
\begin{multline}
\mathcal{M}\leftarrow Group\_Sort\_Select(|\nabla_{\Delta \theta} F(\Delta\theta^*, \Delta x^*)|, \\
N_{flip}, 'descending'),
\end{multline}
\begin{figure}
\centering
\includegraphics{sections/figures/target_neurons.pdf}
\caption{The illustration of targeted model weights across the DNN model weight pages in the memory. \includegraphics[width=0.09in]{sections/figures/bullseye.png}~denotes the targeted bit location in a page.
}
\label{fig:target_neurons}
\end{figure}
\paragraph{Step 3. Adversarial fine-tuning}
Now, given a collection of located vulnerable weights, denoted by $\mathcal{M}$, we only need to update these weights in backpropagation as follows:
\begin{equation}\label{eq:deltatheta}
\Delta\theta = \Delta\theta^* - \eta\cdot[\nabla_{\Delta\theta} F(\Delta\theta^*, \Delta x^*)]_{\mathcal{M}},
\end{equation}
where $[\cdot]_{\mathcal{M}}$ denotes a masking function that returns the gradients for the weights in $\mathcal{M}$, otherwise 0's, and $\eta\geq0$ denotes a learning rate.
\paragraph{Step 4. Bit reduction}
To meet the physical constraints of the Rowhammer attack, the final part of our attack procedure requires bit reduction. Rowhammer can only flip a very low number of bits in a 4KB memory page and more than one faulty memory cell almost never coexists within a byte.
Therefore, we define a bit reduction function as $\mathrm{Floor}(\theta \oplus \theta^*)$, where $\oplus$ denotes the bit-wise summation, and function $\mathrm{Floor}$ rounds down the number by keeping the most significant nonzero bit only. For instance, letting $\theta=1101_2$ and $\theta^*=1010_2$, then $\mathrm{Floor}(\theta \oplus \theta^*)=\mathrm{Floor}(0111_2) = 100_2$. In this way, we assure that only one bit is modified in a selected weight while maintaining its change direction and amount as much as possible.
\subsection{Online Attack Phase: Flipping Bits in the Deployed Model in Memory}\label{subsec:online}
When we access a file from the secondary storage, it is first loaded into the DRAM and when we close the file, the OS does not delete the file from the DRAM to make the subsequent accesses faster. However, it shows that space as free to the user and utilizes it as page cache. If the file is modified, the OS sets the dirty bit of that modified page and it is written back according to the configured write-back policy. Otherwise, the file remains cached unless evicted by some other process or file. As Rowhammer is capable of flipping bits in DRAM, we can use it in the online attack phase to flip the weights of the DNN file as it is loaded in the page cache. The weights file is divided into pages and stored in the page cache. We can flip our target bits as identified by the backdoored parameters $\theta^{*}$, in Section \ref{subsec:offline}. The OS does not detect this change as it is directly made in the hardware by a completely isolated process and it keeps providing the page cached modified copy to the victim on subsequent accesses. Thus the attack remains stealthy.
In the online phase of Rowhammer attack, we need to flip bits in the weights file in the required pages and page offsets. This can be achieved by unmapping the vulnerable rows from the attacker process with the required offsets and then mapping the target weight file. The OS automatically assigns the victim's mapped page to the last unmapped location by the attacker. Then, the attacker rows can be accessed repeatedly to flip bits at the same offsets as found in the offline phase but this time in the victim file.
\subsection{Weight Quantization}
The weights are stored as $N_q$-bit quantized values in the memory as implemented in NVIDIA TensorRT~\cite{migac2017tensorrt}, a high performance DNN optimizer for deployment that utilizes quantized weights~\cite{nvidiatensorrt}. Essentially, a floating-point weight matrix $W_{fp}$ is re-encoded into $N_q$-bit signed integer matrix $W_q$ as $W_q = \mathrm{round}(W_{fp}/ \Delta w)$ where $\Delta w = \max(W_{fp})/(2^{N_q-1}-1)$. In our experiments, we use 8-bit quantization and the weights are stored in two's complement form.
|
\section*{Conflict of interest}
The authors declare that they have no conflict of interest.
\bibliographystyle{spbasic}
\section{SXM Calibration}\label{sec:sxm}
As described in \citet{Masterson18}, the REXIS Solar X-ray Monitor (SXM) consists of a single Amptek XR-100 silicon drift diode (SDD). Due to a number of issues in the flight electronics, the energy resolution of the SXM was limited to 500~eV at 5.9~keV, and the low energy threshold was at approximately 1~keV under nominal operating conditions. Due to thermal sensitivity in the SXM electronics, the low energy threshold was approximately 2~keV during science operations around Bennu. The SXM field of view was defined by a 1~mm diameter collimator mounted above the SDD. The full width zero intensity (FWZI) field of view was 60$^\circ$, around the SXM boresight. The SXM field of view was verified during solar pointing calibrations carried out before and after the OSIRIS-REx Earth Gravity Assist (EGA) in September 2017.
Since the SXM was not equipped with an on-board radioactive source for calibration, the gain stability of the instrument was monitored through occasional solar observations and by tracking electronics artifacts in the SXM spectrum (see Figures~\ref{fig:sxm_temp_effects} and \ref{fig:sxm_spectra}). These electronic ``reset" artifacts generated false events at fixed energies, and those fixed energies provided reference levels that allowed tracking of the gain. Changes in the energy resolution were tracked by measuring the high voltage bias, which remained nearly constant with less than 1\% change during the duration of the mission.
The pre- and post-EGA observations enabled cross-calibration of the SXM with Earth-orbiting solar observatories, for example the Geostationary Operational Environmental Satellite (GOES)-16 Extreme Ultraviolet and X-ray Irradiance Sensors (EXIS) \citep{2009SPIE.7438E..02C} and the GOES-15 X-ray Sensor (XRS) \citep{1996SPIE.2812..344H}. The EXIS and XRS instruments observe the solar X-ray flux between 1.55 keV and 24 keV in two separate energy bands, denoted as A (1.55~--~12.4~keV) and B (3~--~24~keV). During the pre-EGA observation, a series of solar flares observed by the GOES instrument and the REXIS SXM enabled absolute calibration of the SXM effective area as noted in Table~\ref{t:intrument}.
Prior to launch, the REXIS team discovered that the SXM shaping electronics circuit on the Main Electronics Board (MEB) exhibited a thermal sensitivity that could increase the low energy threshold of the SXM. That thermal sensitivity could change the threshold from the nominal setting of 0.3~keV up to values of approximately 2.5~keV. The effect was observed in the pre-EGA cross-calibration, when the SXM response showed intervals of non-correlated variability compared to data from the GOES instruments (see Figure \ref{fig:sxm_temp_effects}). The SXM had originally been designed to cover the energy range from 0.3 - 20~keV in order to characterize the incident solar spectrum above and below the nominal energy range of the REXIS CCDs. However, due to the increased operating temperature during asteroid observations, the minimum operating energy threshold was shifted into the range of 1.8 -- 2 keV. The resulting loss of low energy sensitivity, coupled with low solar activity, severely limited the use of the SXM for the interpretation of data collected at Bennu.
\begin{figure}
\centering
\includegraphics[width=0.99\textwidth]{./figs/temp_effect_solar_pointed_el_lt_20deg.annotate.png}
\caption{The time series of solar X-ray flux measured by the SXM throughout all REXIS operations is shown in the top panel, with temperature measurements shown below. Increases in temperature in the MEB were observed to increase the low energy threshold of the detector which is most clearly visible in the pre- and post-EGA data (histogram number 4000-5000 and 7000-11000). The "reset" artifacts used to monitor the stability of the SXM energy measurement are visible at approximately 75 ADU (-1.6 keV), 135 ADU (2.2 keV) during quiet solar states and near 500 ADU (22 keV) (see Figure \ref{fig:sxm_spectra}). The intensity and position of these artificial lines enabled monitoring of the SXM gain throughout the flight.}
\label{fig:sxm_temp_effects}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.7\textwidth]{./figs/annotated_state_comparison.v2.png}
\caption{Individual SXM measurements of the solar spectrum at different times during the mission. The black arrows indicate the low energy and high energy reset artifacts used to monitor the gain of the SXM. Note the difference in the spectra acquired during Orbital B and R, compared to earlier operations. The solar X-ray flux was very quiet during Orbital B and R, and the high thermal load experienced by the MEB suppressed nearly all of the solar signal below 2~keV. During Orbital B and R thermal effects induced an increase in the energy threshold of the SXM leaving the data dominated by instrumental artifacts at 2 keV and below.}
\label{fig:sxm_spectra}
\end{figure}
\section{Introduction}\label{sec:introduction}
In this paper, we report on the calibration and performance of the Regolith X-ray Imaging Spectrometer (REXIS) instrument aboard NASA's {\it OSIRIS-REx}\xspace asteroid sample return mission, which launched in 2016 and arrived at the asteroid Bennu in 2018. As detailed in \citet{Masterson18}, REXIS was competitively selected as a Class-D student collaboration experiment to complement the remote sensing payload of {\it OSIRIS-REx}\xspace. Class-D instruments are defined as having medium- to-low complexity, short mission operation lifetimes, and relatively low cost, while providing platforms for technological innovation and training grounds for a diverse set of scientists and engineers. As reported by \citet{Hoak21}, the 2019 REXIS measurements
do not contain a clear signal of fluoresced X-rays from the surface of Bennu.
Establishing upper limits on the fluxes of fluoresced X-rays
requires a careful description and analysis of the instrument calibration and in-flight performance. These upper limits may shed light on the surface characteristics of Bennu and help inform the design of future planetary X-ray experiments.
In this paper on REXIS calibration and in-flight performance, we first give a brief overview of the REXIS instrument, its on-board signal processing, and pre-flight milestones and activities that established the instrument status at launch (Section \ref{sec:instrument}). We follow with details of in-flight testing of the main X-ray Imaging Spectrometer (XIS), including the monitoring of its in-flight condition (Section \ref{sec:calibration}). Well-known astrophysical targets were observed as part of the in-flight calibration of the imaging performance (Section \ref{sec:imaging}). During the period when REXIS observed the asteroid Bennu, the serendipitous detection of a transient X-ray binary provided synchronous confirmation of the instrument's sensitivity for detecting fluoresced X-rays from Bennu's regolith (Section \ref{s:maxi}). We describe the in-flight calibration and characteristics of the Solar X-ray Monitor (SXM) in Section \ref{sec:sxm}.
Finally, we discuss lessons that may be applied to future X-ray flight experiments (Section \ref{sec:summary}).
\section{Summary and Lessons}\label{sec:summary}
REXIS was the first planetary X-ray instrument employing the coded-aperture imaging technique in hopes of spatially resolving variations in the elemental composition (if present) across the surface of an asteroid. It was an ambitious project for a student-led experiment involving a completely new planetary instrument design, necessarily conducted under a tight schedule and resource constraints. Spanning 9 years, from concept to completion of operations, more than 90 students had hands-on experience on the REXIS instrument development and operations, which also resulted in 12 Master and 2 Ph.~D. theses \citep{Masterson18}.
Despite the naturally occurring challenge within a university of continual personnel changes from student rollovers as well as multiple challenges and delays during the instrument assembly, the in-flight operations and calibrations indicate that REXIS main spectrometer and solar X-ray monitor were built within their specifications and requirements, albeit with the lower effective detector area and a narrower energy range coverage than the original design. In particular, the observations of the Crab Nebula and Sco X-1 as well as the detection of the new black hole transient MAXI J0637-430 demonstrate that the REXIS imaging system was able to operate well within the original requirements. The historically low solar X-ray flux during Bennu science operations precluded detection of surface fluoresced X-ray photons \citep{Hoak21}.
The successful operation of REXIS during calibration and the detection of the MAXI source while observing Bennu enables the extension of the Technology Readiness Level (TRL) to interplanetary operation for several key subsystems in REXIS. First, the REXIS CCDs are the same type of CCDs (MIT/LL CCID-41) flown on the JAXA {\it Suzaku} mission, which have been operating in low-Earth orbit for about 15 years. REXIS extends the TRL of the MIT CCID-41 CCDs to the interplanetary and near-asteroid environment. Second, the successful detection of celestial X-ray sources such as MAXI J0637-430 under bright background light from Bennu boosted the maturity of the directly deposited OBF in the REXIS CCDs to TRL 9 \citep[see also][]{Thayer21}. Third, the successful opening of the radiation cover in the first attempt after a 2-year traverse across interplanetary space shows the one-time release TiNi Aerospace FD04 Frangibolt actuator at TRL 9 for use after a long duration cruise. Prior to REXIS, this model Frangibolt had been, to our knowledge, flown only in low-Earth orbit \citep{Masterson18}. Fourth, the detector in the REXIS SXM Amptek XR-100 SDD) is the same type used in several space missions such as {\it NICER}\xspace \citep{Gendreau16} and the Miniature X-ray Solar Spectrometer (MinXSS) \citep{Woods17}. Despite issues in the readout system, the REXIS SDD functioned properly for the first two years, extending the TRL of the Amptek SDD to the interplanetary environment.
REXIS suffered from two common causes of performance degradation in space instruments: inadequate light leak protection and thermal management challenges. Although the REXIS team was acutely aware of these potential issues, limits on pre-flight testing and calibration due to schedule constraints prevented the team from addressing certain vulnerabilities that could have been more fully recognized and addressed before launch. In-flight calibrations of the REXIS main spectrometer revealed that the CCDs in REXIS were susceptible to light leaks from their sides. We note that in contrast the optical blocking filter (OBF) on the incident surface of the CCDs performed excellently. During the REXIS observations of Bennu, these light leaks further reduced the effective integration time for the science objective of detecting X-rays emitted from the surface of Bennu. With these challenges to the instrument itself, the unexpectedly low solar X-ray emission (even considering solar minimum) during the Orbital B and Orbital R observations prevented the detection of X-ray fluorescence from Bennu's surface \citep{Hoak21}.
In the case of the SXM, it performed as expected until the Orbital B observation period, at which time it lost its low energy response due to the unexpected out-of-bound thermal environment experienced by the main electronics board.
Here we list some of key pre-flight activities that could have improved REXIS performance, offered as lessons for future planetary X-ray spectroscopy investigations. Given the delays during the instrument development and assembly,
it would have been challenging to conduct these activities prior launch even with the current knowledge of REXIS performance. However, we recommend a careful consideration of the priority of these activities in the context of other pre-flight tasks for future missions.
First, more in-depth tests (and modeling) of light leaks to ensure low background noise for the CCDs. REXIS employed a new type of OBF for the first time for X-ray CCDs in space, where an Al layer is directly deposited on the top surface of the CCDs. As a result, the pre-flight activities regarding light leaks focused on the OBF design and performance. The preliminary light leak modeling indeed revealed possible light-leak paths from the side and back of the CCDs. This discovery led to a few design changes, but these proved insufficient. Given the complexity involved in modeling pathways for light leaks in general, the light leak modeling can be prohibitively expensive and often ineffective in reflecting the reality. Instead, the lab experience using a separate engineering model (EM) setup tailored for light leak tests could be an efficient way to improve the fidelity in the light leak modeling and to validate any design changes if needed.
Second, one needs a full characterization of the CCD responses at low energies to set the key operational parameters such as event and split charge thresholds. Without the full characterization of the instrument response, the REXIS team relied on in-flight operations for both calibrations and parameter optimizations. This not only led to a limited success in the parameter optimization (e.g., the adjustment for the split charge threshold was made after the Orbital B observations) but also added a burden for the {\it OSIRIS-REx}\xspace operational team to accommodate various tests for the REXIS system. While those accommodations were made graciously and with enthusiasm for support of the REXIS students, they were an added weight nonetheless.
Third, we would have benefited from a more thorough thermal modeling and testing of the REXIS system to ensure successful operation for a wider range of temperatures.
The late discovery of the issue in the SXM readout electronics severely limited the range of mitigations that the team could take. The adopted solution still left the system with the limited range of the threshold setting, but more importantly the associated delay severely limited pre-flight testing of the SXM system. Relying on thermal tests of the full or near-complete system is not ideal in terms of schedule. Repeating the test can be also taxing in terms of both the schedule and resources. One could consider a piece-wise or independent thermal test of the MEB or other key components at a relatively early phase with the help of the appropriate diagnostic features in the boards.
Fourth, REXIS was limited by a deficit of in-depth testing and debugging of the flight software, along with originally planned but never-implemented software features. An example is the option to downlink a full 3 $\times$ 3 pixel data array for each CCD event. In-flight grading for CCD events was adopted for REXIS to minimize the telemetry rate, but the accuracy of in-flight grading can vary with spatial and temporal response changes in the CCDs. Inaccurate onboard grading can result in degradation in the spectral resolution and quantum efficiency, since the recovery of the precise energy of input X-rays for multi-pixel events would require all the split signals to be properly selected and calibrated. The downlink of the full 3 $\times$ 3 pixel energies for each event could enable further improvement in energy resolution and quantum efficiency of single pixel events from ground analysis (e.g., the split charge threshold can be readjusted during the analysis).
Finally, a dedicated calibration operation or calibration source for low energy lines
would have been also helpful. Besides MAXI J0637-430, REXIS has detected a few other X-ray sources during the Orbital B and R observations. Among these, Cassiopeia A (Cas A)
Supernova Remnant is interesting due to its prominent line features at low energies in the X-ray spectrum \cite[e.g.,][]{1997AA...324L..49F}.
The reconstructed sky image shows a clear source
at the expected position of Cas A. The measured background subtracted REXIS spectrum show only marginal line features,
while the overall continuum flux level matches with the expected values. During the Orbital B and R observations,
these celestial sources including Cas A were detected in the peripheral of the REXIS FoV with relatively limited
exposure, and the low photon statistics indicates that lack of prominent line features in the measured spectrum is in fact consistent with what is expected.
\section{Angular Resolution and Point Source Localization in Coded-Aperture Imaging} \label{s:ang_resolution}
The angular resolution ($\Delta \theta$)
and the localization error ($\delta \theta$) of a point source in coded-aperture imaging are given as
\begin{eqnarray*}
\Delta \theta &= & \tan^{-1}\left(\frac{\sqrt{\Delta m^2+ \Delta d^2}}{s}\right), \\
\delta \theta & \sim & \frac{\Delta \theta}{\sigma},
\end{eqnarray*}
where $\Delta m$ is the mask pixel size, $\Delta d$ the detector pixel size,
$s$ the separation from the mask and detector, and $\sigma$ the signal-to-noise ratio (SNR) of the point source.
The upper limit in the localization error is set by the minimum SNR ($\gtrsim$5) needed to confidently claim
a detection. Coded aperture imagers with a random mask pattern such as REXIS also have a upper limit in the SNR they can achieve
due to the {\it coding noise} arising from the random nature of the mask pattern.
The upper limit in the SNR from a detector plane image is
the square root of the number of mask pixels that the detector plane image can capture.
For REXIS, the upper limit in SNR ranges from $\sim$ 18 to $\sim$ 25,
depending the number of nodes used in the detector plane image.
This in turn sets the lower limit in the localization error.
Even without the coding noise, there is a fundamental systematic error in localization,
which originates from the finite detector pixel size (25$''$)
and the pointing jitter ($\sim10''$).
Thus, the localization error of REXIS would be limited to $\gtrsim$ 30$''$,
unless algorithms for the subpixel randomization and active jitter compensation are employed.
\section{Modeling the CCD response of the Mn-{K$_{\alpha}$\ } line} \label{s:linemodel}
The response $f(E)$ of the 5.9 keV Mn-{K$_{\alpha}$\ } line in Figure~\ref{f:XIS_sim_spectra} can be
described by a Gaussian function and an exponential low-energy tail component:
\begin{eqnarray*}
& f(E) = A_1 \exp\left(-\frac{(E - E_0)^2}{\Delta_1^2/\log16 }\right) & \\
& + A_2 \bar{\theta}(E-E_0) e^{C_0 (E - E_0)} \left[ 1 - \exp\left(-\frac{(E - E_0)^2 }{ \Delta_2^2/ \log16} \right) \right] &,
\end{eqnarray*}
where $A_1$ and $A_2$ are the normalization constants, $E_0$ the line energy, $\Delta_1$ and $\Delta_2$ the FWHM of the Gaussian and tail components,
$C_0$ the exponential decay parameter. $\bar{\theta}(x)$ is a reverse step function, where
\begin{eqnarray*}
\bar{\theta}(x) = \begin{cases}
1,& \text{if } x \leq 0 \\
0, & \text{otherwise.}
\end{cases}
\end{eqnarray*}
\section{Pointing Coordinates of the Crab Nebula and Sco X-1 Observations} \label{s:pointings}
Tables~\ref{table:crab_observations}
lists the boresight coordinates of the Crab Nebula observations during
the Crab Calibration Operation and the Sco X-1 observations during Mask Calibration Operation, respectively.
See Figure~\ref{f:crab_pointings}.
\begin{SCtable}
\small
\caption{Target boresight pointings during the Crab Nebula and Sco X-1 Observations. These
pointings were repeated on March 16, 17, 23, and 24 2019.}
\begin{tabular}{ cccc }
\hline\hline
Order & R.A. & Dec & Offset\\
\hline
\multicolumn{4}{l}{The Crab Nebula: 2019 March 16, 17, 23, and 24}\\
1 & 83$^{\circ}$ 37' 48" & 22$^{\circ}$ 00' 36" & 0.0$^\circ$\\
2 & 80$^{\circ}$ 43' 12" & 19$^{\circ}$ 48' 00" & 3.5$^\circ$\\
3 & 83$^{\circ}$ 37' 48" & 18$^{\circ}$ 30' 36" & 3.5$^\circ$\\
4 & 86$^{\circ}$ 32' 24" & 19$^{\circ}$ 48' 00" & 3.5$^\circ$\\
5 & 87$^{\circ}$ 20' 24" & 22$^{\circ}$ 44' 24" & 3.5$^\circ$\\
6 & 85$^{\circ}$ 19' 12" & 25$^{\circ}$ 09' 00" & 3.5$^\circ$\\
7 & 81$^{\circ}$ 58' 12" & 25$^{\circ}$ 09' 36" & 3.5$^\circ$\\
8 & 79$^{\circ}$ 55' 48" & 22$^{\circ}$ 45' 36" & 3.5$^\circ$\\
\hline
\multicolumn{4}{l}{Sco X-1: 2019 June 2 }\\
1 & 244$^{\circ}$ 58' 44" & -15$^{\circ}$ 38' 24" & 0.0$^\circ$\\
2 & 232$^{\circ}$ 00' 00" & -15$^{\circ}$ 38' 24" & 12.5$^\circ$\\
3 & 250$^{\circ}$ 00' 00" & -15$^{\circ}$ 38' 24" & 4.8$^\circ$\\
4 & 244$^{\circ}$ 58' 44" & -21$^{\circ}$ 38' 24" & 6.0$^\circ$\\
5 & 240$^{\circ}$ 00' 00" & -04$^{\circ}$ 38' 24" & 12.0$^\circ$\\
\hline
\end{tabular}
\label{table:crab_observations}
\end{SCtable}
\section{In-flight Calibration of XIS Imaging System} \label{sec:imaging}
The coded-aperture imaging technique used in REXIS is effectively a shadow-gram imaging system where the detector plane image is cross-correlated with the known mask pattern to identify the source position or to reconstruct the source shape. The angular resolution of REXIS is 26.5$'$, which is set by the telescope geometry (see Section~\ref{s:ang_resolution} in the Appendix). The localization error ($\delta\theta$) of a point source REXIS can achieve ranges from $\sim$ 1$'$ to $\sim$ 5$'$, with the instrument requirement being $\delta\theta < 6.7'$.
\subsection{Boresight Calibration through Observations of the Crab Nebula} \label{sec:boresight}
In order to resolve any regions of variable X-ray fluorescence across the surface of Bennu, we determined the precise position and orientation of the coded aperture mask relative to the REXIS CCDs using the observations of the Crab Nebula. For each CCD, we modeled the mask placement using four boresight parameters consisting of the three-dimensional positions and rotation (along the pointing axis) of the coded aperture mask relative to the CCD. In principle, two additional parameters describing the tilt of the mask are needed to fully describe the mask placement relative to the detector, but these parameters were not necessary in achieving the desired boresight corrections, as our measured accuracy was well within specifications indicating that there was no noticeable tilt in the mask.
As seen in Figure~\ref{f:crab_pointings}, the Crab Nebula observation sequence consisted of eight pointings surrounding the Crab Nebula. These provided sufficient independent measurements to break the degeneracy in the boresight parameters while distributing flux evenly across the full detector area for absolute quantum efficiency calibration for all pixels. The detector plane image of each pointing was cross-correlated with the mask pattern through a series of Fast Fourier Transformation (FFTs) to reconstruct the matching sky image. Then, the bright point source in the image was localized through a Gaussian peak finding routine, and the offset from the true position of the Crab Nebula calculated. We repeated this process for a range of the values for each boresight parameter to solve for the correct value. The final boresight solution was obtained using a Markov Chain Monte Carlo (MCMC) search technique across the parameter space.
Table~\ref{t:boresight_corrections} summarizes the boresight parameters using the first two days of the Crab Nebula observations in the 2nd Crab Calibration Operation. The measured values validated the REXIS instrument design and build indeed survived the launch and space environment within the original requirements. Figure~\ref{f:boresight_errors} shows the boresight errors before and after applying the boresight corrections using the parameter values in Table~\ref{t:boresight_corrections}. Before the correction, the raw pointing errors range $\sim$10$'$ to $\sim$15$'$ and after correction, the offsets are within 1$'$, which far exceeds the requirement.
Note that the signal-to-noise-ratio for each Crab Nebula observation is $>$20.
\begin{SCfigure}
\small
\includegraphics[width=3in]{figs/rex_fm19075_bsc.png}
\caption{Offset between the measured and true positions of the Crab Nebula before (blue) and after (red) boresight corrections.}
\label{f:boresight_errors}
\end{SCfigure}
\begin{SCfigure}
\small
\includegraphics[width=3in]{figs/CrabCal21_19075_76_sky0_0.png}
\caption{Reconstructed sky image from the first two day observations of the 2nd Crab Nebula calibration in March 2019. The color scale represent the total photon counts.}
\label{f:crab_image}
\end{SCfigure}
\subsection{Mask Calibration through Observations of Sco X-1} \label{sec:mask}
The REXIS boresight parameters were further confirmed with the observations of Sco X-1 obtained in June 2019. We observed Sco X-1 with five separate pointings (Figure~\ref{f:crab_pointings}) to see if there is any distortion in the mask after an operation that exposed the mask to direct sunlight. Direct exposure to sunlight potentially created a large temperature gradient within the mask, which could cause the mask to bow or stretch. Unlike the pointings during the Crab Nebula observations, which were limited to 3.5$^\circ$ within the source, the five separate pointings during the Mask calibration include relatively large off-axis observations of the source, ranging up to $\sim$ 12$^\circ$ (Table~\ref{table:crab_observations}).
Having the source at large off-axis angles enabled the inspection of outer mask segments. With the boresight corrections in Table~\ref{t:boresight_corrections},
the source positions were recovered within an arcminute of the true position. On the other hand, the data of two largest off pointings (2nd and 5th pointings of Sco X-1 in Table~\ref{table:crab_observations}) revealed a minor shrinkage in the observed mask pitch pixel in order of $\sim$1.5\%, which is equivalent to $\sim$ 25$''$. The observed shrinkage is well within the requirements for the mask and can be explained by a slightly bowing in the mask plane. However, it is not clear when the bowing was introduced since this was the first measurements for the mask distortion.
\begin{SCtable}
\small
\caption{Measured mask offsets relative to the REXIS CCDs vs.~the requirements (See Figure~\ref{f:instrument})}
\begin{tabular}{ c|c|c }
\hline\hline
Parameters & Measured & Requirements \\
\hline
$\Delta$X & $-$0.720 mm or $-$12.44$'$ & N/A \\
$\Delta$Y & +0.023 mm or +0.39$'$ & N/A \\
$\Delta$Z & $-$0.891 mm & $\pm$1.00 mm \\
roll & 0.205 degrees & 0.44 degrees \\
\hline
\end{tabular}
\label{t:boresight_corrections}
\end{SCtable}
\subsection{REXIS Detection of MAXI J0637-430} \label{s:maxi}
During the Orbital B observations, the new transient source MAXI J0637-430 was detected by the REXIS \citep{Allen20}. MAXI J0637-430 is a soft X-ray transient, first discovered on 2019 November 2 by the Monitor of All-sky X-ray Image ({\it MAXI}\xspace), the Japanese Experiment on the International Space Station \citep{Negoro19}. The X-ray spectral and timing properties measured by the Nuclear Spectroscopic Telescope Array ({\it NuSTAR}\xspace) along with the high X-ray to optical flux ratio suggests that the source is a black hole X-ray binary \citep{Tomsick19}. REXIS' detection of MAXI J0637-430 marks the first black hole detection made from outside the Earth-Moon environment.
The REXIS XIS detected the source (mainly by CCD 2) near the edge of the FoV just outside of Bennu for eight separate days between November 12 and 22 in 2019 (whenever it was in the FoV).
With the boresight corrections in Table~\ref{t:boresight_corrections}, the REXIS position of the source is localized to (R.A., DEC) = (6:36:22.94, -42:50:44.0) (J2000), which is 1.3$'$ off the radio counterpart of the source \citep{Russell19}, based on the CCD 2 Node 2 data taken on November 12th, 2019. Figure~\ref{f:maxi_image} shows the reconstructed sky image from the data in CCD 2 Node 2, where the source dominantly shined.
\begin{SCfigure}
\small
\includegraphics[width=3in]{figs/rex_fm19316_sky1_pb_v002_fi48_ccd22_MAXI_J0637-430_zoom_0.png}
\caption{Reconstructed REXIS sky image around the transient source MAX J0637-430 during Orbital R. The bright source seen in the image is 1.3$'$ away from the reported radio position of MAXI J0637-430.}
\label{f:maxi_image}
\end{SCfigure}
\section{Inflight Calibration and Response of X-ray Imaging Spectrometer}\label{sec:calibration}
Starting shortly after launch, REXIS carried out a series of calibration operations that collected data from on-board {$^{55}$Fe\ } sources (described in the next section) and from observations of bright celestial X-ray sources. In addition, REXIS and the rest of the instruments onboard {\it OSIRIS-REx}\xspace were powered on for periodic checkups roughly every six months during the journey to Bennu. These activities were designed to characterize and monitor the instrument response over time.
Table~\ref{t:activity} lists the operational activities of REXIS following the launch of {\it OSIRIS-REx}\xspace in 2016. Notable calibration operations included the 4-day internal background and calibration (with the radiation cover closed), the two sets of 4-day observations of the Crab Nebula for boresight and quantum efficiency (Q.E.) calibrations, the 1-day Optical Blocking Filter (OBF) calibration, and the 1-day observation of Sco X-1 for mask calibration. Throughout flight, REXIS used the 5.9~keV Mn-{K$_{\alpha}$\ } and 6.5~keV Mn-{K$_{\beta}$\ } X-ray lines from the onboard {$^{55}$Fe\ } sources to monitor changes in the spectral gain, offset in energy calibration, and change in Q.E. of the CCDs. As noted above, REXIS had a radiation cover located at the top of XIS tower above the coded mask, to minimize the radiation damage in the X-ray CCDs from cosmic ray interactions during the cruise phase to Bennu. The radiation cover was opened in September 2018, 8 months prior to the arrival of {\it OSIRIS-REx}\xspace at Bennu.
\subsection{Monitoring REXIS X-ray CCD Response using Onboard {$^{55}$Fe\ } Sources} \label{sec:Fesources}
\begin{SCfigure}
\small
\includegraphics[width=3.0in]{figs/Fe-55_sources_on_the_DAM2.png}
\caption{Onboard {$^{55}$Fe\ } sources (red) and their illumination regions (purple) on the REXIS DAM. Five nodes with relatively stable performance are labeled. Labels A and B mark the near and far-end of the readout chain exposed to the {$^{55}$Fe\ } sources for CCD 1 node 0. }
\label{f:Fe_sources}
\end{SCfigure}
\begin{SCfigure}
\small
\includegraphics[width=3.1in,trim=0 0 0 0]{figs/detector_plane_image.png}
\caption{Data from the REXIS internal background calibration, performed in flight July 13 2018. Each putative x-ray event is marked by a colored circle. The illuminated regions from the {$^{55}$Fe\ } sources are visible around the perimeter. The colorscale is in units of uncalibrated energy, and each detector node has a slightly different gain, which causes the 5.9~keV photons to be recorded at slightly different raw energy values in each node. Some data quality issues are immediately obvious, for example the hot column in CCD~1 Node~1 and excess noise in CCD~0 Node~1 and CCD~2 Node~1.}
\label{f:Fe_evt_dist}
\end{SCfigure}
REXIS carried eleven {$^{55}$Fe\ } sources with a combined radioactivity of $\sim$~2~$\mu$Ci. As shown in Figure 2, these sources were located around the perimeter of the CCD array to enable monitoring each of 16 nodes in the 2 $\times$ 2 CCD array \citep{Masterson18}.
One of the sources was placed on the inner side of the radiation cover, which was pasted and cured with a {$^{55}$Fe\ } solution of 200~nCi (Figure~\ref{f:instrument}). The radiation cover source enabled the calibration of the entire detector array when the cover was closed and became invisible to the detector array after cover opening.
Figure~\ref{f:Fe_evt_dist} shows an example image produced by the full detector plane during the in-flight Internal Calibration Operation. Eight sources, each with 200~nCi, were arranged so that small regions ($\sim$~100 $\times$ 50 pixels) next to the readout buffer in each of 16 CCD nodes (two nodes per source) were exposed to {$^{55}$Fe\ } X-rays. Two other sources (75~nCi each) were positioned so that a small region ($\sim$ 150 $\times$ 120 pixels) in the far side of the readout chain on each of the four nodes on both sides is illuminated. The ten {$^{55}$Fe\ } sources on the DAM enable continuous monitoring of the spectral response of all 16 nodes with relatively low radioactivities, minimizing the area of potential contamination during X-ray observations of Bennu. In principle, charge transfer inefficiency (CTI) during readout could be monitored in four nodes through comparison of the spectral line resolution at the near- and far-end of the node; e.g., see markings {\it A} and {\it B} in Figure~\ref{f:Fe_sources} for CCD 1 node 0.
\subsubsection*{Energy Calibration}
The energy of each REXIS CCD event was calibrated based on the observed raw signal amplitude of the X-rays from onboard {$^{55}$Fe\ } sources. Figure~\ref{f:Fe_pha} plots the raw signal heights of valid single-pixel X-ray events (Grade 0 events) in CCD~2 Node~0 over continuously elapsed time in seconds (i.e. without observational gaps). The two horizontal streaks around 300~ADUs are due to the 5.9 and 6.5~KeV Mn-K lines from the {$^{55}$Fe\ } sources, which are more or less stable except for a few occasions (Section~\ref{sec:lightleak}).
Assuming a linear relation between the raw signal ($A$) and the input X-ray energy ($E$) (measured in units of keV), at least two X-ray lines of known energies are required to constrain both spectral gain ($g$) and offset ($c$) parameters. Since the 6.5~keV line is much fainter than the 5.9~keV line, the former is not suitable for calibration of short term variations. A solution to this problem is to utilize the charge split events of the same energy over two pixels.
The expected raw signal of a single pixel event ($A_s$) or a double pixel event ($A_d$) from the Mn-{K$_{\alpha}$\ } line is given as
\begin{eqnarray*}
& A_s = g E_{\mbox{\scriptsize 5.9 keV}} + c, & \text{\ and} \\
& A_d = g E_{\mbox{\scriptsize 5.9 keV}} + 2 c. &
\end{eqnarray*}
Figure~\ref{f:Fe_spectra} shows an example of a calibrated REXIS X-ray spectrum. These data were collected during the Internal Calibration Operation prior to the opening of the radiation cover, to enable the characterization of the internal background spectrum of each node as well as providing a snapshot of the spectral gain and offset. Even with the radiation cover closed, multiple X-ray lines beside the two Mn-K lines are present in the spectrum. These lines include the Si-K line at 1.7~keV, which is mainly caused by cosmic ray interactions in the silicon layer of the CCDs, and an escape peak at 4.2~keV, which is created by escaping Si-K lines fluoresced by the Mn-{K$_{\alpha}$\ } line from the sources. The escape peak at 4.8~keV due to the Mn-{K$_{\beta}$\ } line is too faint to detect.
\subsubsection*{Charge Transfer (In)Efficiency}
The difference in the peak value of the \mbox{Mn-{K$_{\alpha}$\ }} line between the near and far end of each node indicates a loss of the total charge due to the charge transfer across the node during readouts (e.g., markings {\it A} and {\it B} in Figure~\ref{f:Fe_sources}). During the Internal Calibration Operation, the charge transfer loss was measured to range from 0.6\% to 1.1\%, depending on the nodes, which is equivalent to 35-65~eV for the Mn-{K$_{\alpha}$\ } line. In the case of CCD 0 Node 1, the additional {$^{55}$Fe\ } source on the DAM enables the charge transfer loss measurement even after opening the radiation cover. During the Orbital B observations, the charge transfer loss in CCD 1 Node 0 stayed about the same (approximately 1\%) as in the Internal Calibration Operation.
\begin{SCfigure}
\small
\includegraphics[width=3.2in]{figs/rex_fmall_evt0_ccd20_perim_g0_v001_pha_vs_time.png}
\caption{Raw pulse heights of valid single pixel events of CCD 2 Node 0 over elapsed time without observational gaps. The two horizontal streaks around 300 ADUs are 5.9 and 6.5~KeV Mn K-lines from the {$^{55}$Fe\ } sources. The variations in these streaks indicate changes in either spectral gain, offset or both. The analysis showed the main cause is the shift in the spectral offset.}
\label{f:Fe_pha}
\end{SCfigure}
\begin{SCfigure}
\small
\includegraphics[width=3.2in]{figs/rex_fmIntCal_ccd22_g00_allrg_fit1_spmIntBkgPIET.png}
\caption{Energy calibrated X-ray spectrum of CCD 2 Node 2 during the Internal Calibration Operation. The {$^{55}$Fe\ } source generates 5.9 and 6.5~keV Mn-K lines, and their escape peaks. The 1.7 keV Si-K line is generated by internal fluorescence from cosmic rays. The internal background is well-fit by a 2nd order polynomial function with a low energy exponential tail component (see Section~\ref{sec:excess} for more detail).}
\label{f:Fe_spectra}
\end{SCfigure}
\subsubsection*{Changes in Quantum Efficiency}
In addition to monitoring the spectral response of the CCDs, the {$^{55}$Fe\ } sources enable monitoring of relative changes in the quantum efficiency of each CCD node. With a half life of 2.7~years, the X-ray flux from the {$^{55}$Fe\ } sources decreased by more than a factor of two over the timescale of REXIS operations. Figure~\ref{f:Fe_intensity} shows the observed flux intensity of the two Mn-K lines from the {$^{55}$Fe\ } sources in CCD 2 Node 0. The observed flux decayed as expected until Orbital B, when there was a significant drop. The decrease in the {$^{55}$Fe\ } detection rate during Orbital B was ultimately traced to the presence of optical light in the detector; see Section~\ref{sec:lightleak}. Optical photons from the light lead induced a background of low-charge events in the CCDs, which spoiled the event grading algorithm (causing single-pixel events to appear as multi-pixel events).
\begin{SCfigure}
\small
\includegraphics[width=3.2in]{figs/rex_fmall_rate1_ccd20_perim_g05_v003_100_by_operation.png}
\caption{Measured flux changes against expected decay rate of the perimeter {$^{55}$Fe\ } source in CCD 2 Node 0. During Orbital B, the measured flux rate dropped below the expected rate, indicating 34\% loss in the Q.E. The loss was due to a light leak, which induced a background of low energy events in the CCDs and spoiled the event grading algorithm (causing single-pixel events to appear as multi-pixel events). The measured background rates (crosses) in the 5 - 7 keV band remained steady. The observed flux returned to the expected rate during Orbital R, following a change to the CCD split-charge threshold and greater orbital altitude.}
\label{f:Fe_intensity}
\end{SCfigure}
\subsection{Spectral Resolution and Background Rates} \label{s:response}
\begin{table*}
\small
\caption{Spectral Response and Internal Background Rate of Five Prime CCD Nodes in REXIS}
\centering
\begin{tabular}{ @{\extracolsep{4pt}}ccccccc }
\hline\hline
CCD & Node & \multicolumn{3}{c}{Spectral Resolution$^a$} & \multicolumn{2}{c}{Background Rates$^b$} \\
\cline{3-5}\cline{6-7}
ID & ID & Region$^c$ & Int. Cal. & Orbital B Obs. & Int. Cal. & Orbital B Obs. \\
\hline
0 & 2 & near & 153$^{+26}_{-22}$ eV & 172$^{+22}_{-22}$ eV & 1.53(6) & 3.08(4)\\
\multirow{2}{*}{1} & \multirow{2}{*}{0}
& near & 247$^{+25}_{-27}$ eV & 282$^{+23}_{-23}$ eV & \multirow{2}{*}{1.61(6)} & \multirow{2}{*}{5.95(6)} \\
& & far & 294$^{+26}_{-25}$ eV & 304$^{+24}_{-24}$ eV &\\
1 & 2 & near & 230$^{+24}_{-25}$ eV & 292$^{+22}_{-22}$ eV & 1.63(6) & 4.28(5)\\
2 & 0 & near & 153$^{+23}_{-22}$ eV & 215$^{+22}_{-22}$ eV & 1.59(7) & 1.95(3)\\
2 & 2 & near & 159$^{+23}_{-23}$ eV & 213$^{+22}_{-22}$ eV & 1.90(6) & 3.58(4)\\
\hline
\multicolumn{7}{l}{$^a$FWHM at 5.9 keV from the perimeter source regions}\\
\multicolumn{7}{l}{$^b$counts ks$^{-1}$ cm$^{-2}$ keV$^{-1}$ in the 2 -- 3.5 keV band of the full node}\\
\multicolumn{7}{l}{$^c$Illumination region: near (marked {\it A} for CCD 1 Node 0 in Figure ~\ref{f:Fe_sources}) or far (marked {\it B}) }\\
\multicolumn{7}{l}{end of the node w.r.t. the framestore.}
\end{tabular}
\label{t:resolution}
\end{table*}
Table~\ref{t:resolution} summarizes the spectral resolution and background rates of the five stable CCD nodes in the REXIS detector plane. Interestingly, all five of the most stable nodes have an even number node ID (Figure~\ref{f:Fe_sources}). This is not a coincidence, and was later identified in part due to the interference from the residual signal in CCD~3 which suffered a severed flexprint cable, since the location of the readout electronics in the MEB for the nodes with an odd number ID is more susceptible to the interference. For more details, see \citet{Masterson18}. The spectral resolutions of the five prime CCD nodes in Table~\ref{t:resolution} range from $\sim$150 to 300 eV in FWHM.
The nodes in CCDs 0 and 2 show finer spectral resolutions than those in CCD 1. This is interesting since CCD 1 was considered as Grade A while CCDs 0 and 2 were Grade C before the assembly, and we suspect that the performance degradation in CCD 1 is also related to the interference from the power line in CCD 3.
Comparing the measurements during the Internal Calibration Operation in July 2018 and the Orbital B Observations in July 2019, there is a noticeable degradation in the CCD spectral resolution. Assuming a quadratic contribution of each noise component, the additional degradation ranges from 80 to 180~eV, which is likely due to radiation damage during the 10-month period between opening of the radiation cover and Orbital B. In CCD~1 Node~0, there are two regions illuminated by the {$^{55}$Fe\ } sources, and the degradation in the spectral resolution at the far-end of the node (w.r.t.~the readout buffer: labeled as {\it B} in Figure~\ref{f:Fe_sources}) appears relatively small, but the relatively poor resolution of the node makes it hard to isolate the effect of radiation damage during the charge transfer.
The background rates in Table~\ref{t:resolution} indicate the average count rate in the 2 to 3.5~keV band, which is free of instrumental spectral lines (see Figure~\ref{f:Fe_spectra}). The background rates during the Orbital B observations include the contributions from the Cosmic X-ray Background (CXB), potential X-ray fluorescence (XRF) from Bennu, and the light-leak induced noise. CCD~2 Node~0 shows the smallest increase in the background rate during the Orbital B observations compared to the Internal Calibration Operation. This is consistent with the CXB being more occulted by Bennu in the spatial region imaged by CCD~2 Node~0 (see \cite{Hoak21}). For most other nodes, the spatial orientation of the asteroid during Orbital B was such that the CXB dominates the observed X-ray flux.
\subsection{Response Matrix Function (RMF) Modeling using Onboard {$^{55}$Fe\ } Sources and Simulations} \label{sec:rmf}
The response of an X-ray sensor to incoming X-rays is commonly described by the Response Matrix Function (RMF) and the Auxiliary Response Function (ARF). The RMF is a collection of probability distributions as a function of raw pulse height for the range of X-ray energies to which the sensor is sensitive. Thus, the RMF essentially enables the translation between input X-ray energies and possible corresponding detection signal heights. The ARF describes the detection efficiency in units of effective area as a function of X-ray energy. The ARF involves the Q.E. of the sensor and other relevant factors (e.g., absorption loss due to the optical blocking filter (OBF) in REXIS). Combining the RMF and ARF, one can fully characterize how often a given X-ray photon would interact with the sensor and how much signal the input X-ray would induce in the sensor when it does.
The REXIS team had planned an extensive pre-flight ground calibration program for the REXIS CCDs to establish both the RMF and the ARF for each node, but this program had to be abandoned due to schedule delays during the assembly. Instead, the RMFs for the REXIS CCDs were calculated based on X-ray CCD response simulations and spectral measurements of the onboard {$^{55}$Fe\ } calibration sources during in-flight operations. Similarly, the ARFs were estimated based on the observations of the Crab Nebula, a bright celestial X-ray source with a steady X-ray flux and spectrum as described in Section \ref{sec:arf}.
Since the onboard {$^{55}$Fe\ } sources produce relatively high energy Mn-K lines at 5.9 and 6.5 keVs, the CCD response at low energies has to be modeled through X-ray CCD simulations. Note that the Si-K lines at 1.7 keV observed in the REXIS CCDs (e.g., Figure~\ref{f:Fe_spectra}) have a different response compared to 1.7 keV X-rays that are incident on the CCDs, since the former is generated within the silicon substrate inside the CCDs. The X-ray response simulations were performed using a simulation tool developed for the X-ray imaging Spectrometer (XIS) on-board the {\it Suzaku} mission \citep{Prigozhin98, Prigozhin00, Prigozhin03}.
\subsubsection*{XIS Simulator}
The XIS simulator performs numerical modeling of back-illuminated (BI) CCDs using Monte-Carlo techniques. Simulated photons of a given energy interact with a model of the surface of the CCD at random locations, and the behavior of each photon is evolved using basic physics of photon interactions with silicon. The program determines the site of photon interaction with silicon atoms, based on the probability of absorption. The exposed surface of the BI CCD is modeled with a uniform layer of SiO$_2$. The simulator determines whether fluorescence took place; if the answer is yes, the fluorescent photon is emitted in a random direction and traced, either escaping the CCD or re-interacting with the silicon. In the case of a re-interaction the fluoresced photon is treated with the same interaction modeling as the original photon.
Next, the simulator models the creation and expansion of the charge cloud and charge losses due to interaction with the surface. The program models the distribution of charge across the pixels surrounding the photon interaction site, simulates the charge transfer from the imaging section into the frame store, from the frame store into the serial register, and from the serial register into the readout node. Readout noise is added to each pixel (to simulate noise in the electronics chain), and a standard 3 $\times$ 3 pixel event is written to the output event list through an event finding algorithm. See \citet{Prigozhin00} for more details.
\begin{SCfigure}
\small
\includegraphics[width=3.2in]{figs/E5.899keV_counts10000_r6.5_r11.5_st100eV_g0_scr_3g1t.png}
\caption{Simulated X-ray spectrum designed to produce the single-pixel spectral response in CCD 0 Node 2 for 5.9 keV Mn-{K$_{\alpha}$\ } X-rays. The split charge threshold of 100 eV and two readout noise components of 6.5 e$^-$ and 11.5 e$^-$ are used for the spectral response. The main peak for 5.9 keV X-rays in the spectrum is well described by a Gaussian line (pink) and an exponential tail component (brown): see Section~\ref{s:linemodel} and \citet{Prigozhin00}.}
\label{f:XIS_sim_spectra}
\end{SCfigure}
Some of the key simulation parameters in modeling the observed {$^{55}$Fe\ } X-ray spectra are the thickness of the silicon substrate layer, electronics readout noise, and the split charge threshold. The thickness of the silicon substrate layer (45 $\mu$m for the REXIS CCDs) determines how wide the charge cloud can spread while it drifts down from the interaction site. The split charge threshold (ST), which can be set for REXIS by an uplink command, is used to determine if the given CCD event is of a single pixel or multi-pixel origin. It ranges from about 100 eV to a few 100s~eV, depending on the node.
The electronics readout noise describes the noise characteristics of the readout chain. REXIS occasionally downlinked a full image of the charge distribution over the entire detector plane collected during one readout cycle (4 sec) for diagnostic purposes. These full frame images indicated that two readout noise components of high and low frequencies were present in the REXIS CCDs. We incorporated these two noise components in to the RMF calculation by combining the results of two sets of the XIS simulations with two separate readout noise parameters.
Figure~\ref{f:XIS_sim_spectra} shows the simulated spectrum for 5.9 keV X-rays, where the Si-K and escape lines are also present. We tuned the model parameters to closely match the probability distribution of 5.9 keV X-ray photons to the observed raw signal height distribution of the Mn-{K$_{\alpha}$\ } line at 5.9 keV from the on-board {$^{55}$Fe\ } sources during the Internal Calibration Operation. The full RMF was constructed using this configuration of XIS simulations for the rest of the input X-ray energies. We repeat this process separately for single (Grade 0) and double (Grades 2-5) pixel events for each REXIS CCD node.
\begin{SCfigure}
\small
\includegraphics[width=3.2in]{figs/gtest_spec_20_0716_AlF_rmf.png}
\caption{X-ray spectra of low energy X-ray lines taken with the flight spare CCD on the ground. A {$^{210}$Po\ } source was used to illuminate a Teflon or aluminum target, which produced the 0.67 keV F-K (blue) and 1.5 keV Al-K (red) lines, respectively. An {$^{55}$Fe\ } source was present in both cases. The solid lines show the modeled detector response using the RMF of CCD 2 Node 2 for the flight instrument.}
\label{f:lowEnergyLine}
\end{SCfigure}
\subsubsection*{Applying RMFs}
The assembled RMFs were used for the spectral fit in Figure~\ref{f:Fe_spectra}, and they correctly modeled the asymmetric spectral shape of the Mn-K lines and their escape peaks. The XIS simulations also predict the Si-K line from simulated input X-rays, but the observed Si-K line has a much greater amplitude, and is likely triggered by cosmic rays rather than incoming X-rays. In our model of the REXIS CCD internal background noise, we used an additional Gaussian line to properly described the Si-K line in the observed spectra.
As mentioned in Section~\ref{s:response}, the spectral resolution varies spatially in each node and is expected to degrade with time. In order to capture small changes in spectral resolution with our fixed RMF, individual spectral lines (including the Mn-K lines) are modeled using a Gaussian function with a narrow but nonzero width, rather than with a delta-function. The full-width-half-maximum (FWHM) of the Gaussian function is allowed to vary. The spectral resolution in Table~\ref{t:resolution} for a given node is modeled by the same RMF, but the the width of the observed Mn-K lines is determined by the quadratic sum of two components: the width of the Gaussian function (which is a free parameter, in a narrow range) and the width set by the RMF (which is fixed in the model for the same node for a given energy).
\subsubsection*{Validation of RMFs at Low Energies}
The low energy performance of the constructed RMFs was indirectly validated by the data taken with a flight spare CCD on the ground. We used a {$^{210}$Po\ } $\alpha$ source to excite targets of fluorine (Teflon) and aluminum and observed the CCD response of the F-K and the Al-K fluorescent lines. In Figure~\ref{f:lowEnergyLine}, we compare the observed X-ray spectra of the 0.67 keV F-K and 1.5 keV Al-K lines from the spare CCD with the response models in the RMF of CCD 2 Node 2 for the lines. Even without fine tuning the RMF to the performance of the spare CCD used to collect the data, the agreement between the data and the model is good.
\subsection{Auxiliary Response Function (ARF) Modeling using Observations of the Crab Nebula and Sco X-1} \label{sec:arf}
The onboard {$^{55}$Fe\ } sources enabled monitoring of relative changes in the quantum efficiency of the CCDs, but it is difficult to estimate the absolute efficiency of the CCDs (and thus the effective area of REXIS) with the {$^{55}$Fe\ } sources due to the limited energy coverage of the spectral lines and the complex geometric effects. In order to measure the effective area of the CCDs as a function of incident energy, we observed the Crab Nebula, a bright, stable X-ray source with a precisely measured flux and spectrum \citep{Weisskopf10}. REXIS observed the Crab Nebula on March 16, 17, 23, and 24 in 2019 for a total of 16 hours. Since the Crab Nebula appears as a point-like source to REXIS with $\sim$ 26$'$ angular resolution, only about 41\% of the CCD pixels were exposed to the Crab at a given time because of the REXIS coded-aperture mask. To ensure approximately equal exposure for all detector pixels, {\it OSIRIS-REx}\xspace oriented the REXIS boresight to eight different points on the sky for each day of the operation; the pattern of observation targets is shown in Figure \ref{f:crab_pointings}. During the second two days (March 23-24), we observed a spatially uniform time dependent background in the CCD data, which we attribute to increased solar particle flux associated with a passing heliospheric current sheet (Parker spiral). Excluding the data with elevated background leaves about 8 hours of useful integration time, which provided sufficient CCD counts to measure the energy-dependent effective area for the individual CCD nodes.
\begin{SCfigure}
\small
\includegraphics[width=3.2in]{figs/crab_fig_edited.png}
\caption{Data from the REXIS observations of the Crab Nebula, March 16-17 2019. Only five detector nodes were enabled for this observation. The pattern of the REXIS coded aperture mask is clearly visible. The data in this figure are collected from about one hour of exposure time.}
\label{f:crab_detplane}
\end{SCfigure}
We calibrated the energy values of the CCD event-list data using the on-board {$^{55}$Fe\ } sources as described in Section~\ref{sec:Fesources}. To separate the X-ray signal of the point-like Crab Nebula from the diffuse cosmic X-ray background (CXB), we separated the data from each pointing into events from pixels that were exposed to the Crab Nebula (the ``on-source" pixels) and events from pixels that were not exposed to the Crab (the ``off-source" pixels). For each pointing, we subtracted the area-weighted off-source spectrum from the on-source spectrum. The resulting spectrum (on-source minus off-source) contains only the spectrum of X-rays from the Crab Nebula. We used this observed spectrum, together with the known Crab X-ray spectrum and our model for the detector response (see Section~\ref{sec:rmf}) to measure our effective area as a function of X-ray energy. The analysis was performed twice for each CCD node, once for single-pixel events (Grade 0) and again for double-pixel events (Grades 2-5). Higher order multi-pixel events were not modeled due to their high background noise and poor energy resolution.
\begin{SCfigure}
\small
\includegraphics[width=3.2in]{figs/Crab_ScoX-1_pointings.png}
\caption{REXIS pointing sequences during the Crab Nebula calibration in March 2019 and the Sco X-1 observations for Mask Calibration Operation in June 2019.
See Tables~\ref{table:crab_observations} for pointing coordinates.}
\label{f:crab_pointings}
\end{SCfigure}
\begin{SCfigure}
\small
\includegraphics[width=3.2in]{figs/arf_v3_20.png}
\caption{Results from the fit of the CCD 2 Node 0 ARF model for single-pixel events to the Crab Calibration data. The reduced $\chi^2$ of the fit is 1.55 between 1 and 4~keV.}
\label{fig:ccd20_crab_fit}
\end{SCfigure}
\begin{SCfigure}
\small
\includegraphics[width=3.2in]{figs/arf_20_effectivearea.png}
\caption{REXIS CCD 2 Node 0 effective area, for single-pixel (Grade 0) events. The node has a physical area of 1.509~cm$^2$, implying a single-pixel detection efficiency of 22\% at 1.25~keV.}
\label{fig:ccd20_arf}
\end{SCfigure}
To measure the effective area of the CCDs, we used the Sherpa Python package \citep{Freeman01} to convolve the modeled X-ray spectrum of the Crab Nebula from \citet{Weisskopf10} with the RMF as described in Section~\ref{sec:rmf} and a parametric ARF model, and searched for the ARF parameter values that gave the minimum $\chi^2$ between the background-subtracted observed spectrum and the response-convolved model spectrum. The free parameters in our ARF model were the thickness of the aluminum optical blocking filter, the detection efficiency due to split-charge events, and an overall normalization factor.
Figure~\ref{fig:ccd20_crab_fit} shows the results of the fit to the observed spectrum for a single node (CCD 2 Node 0). Below 1~keV, there is an excess of events that does not fit the model; this is discussed in greater detail in Section~\ref{sec:excess}. Table~\ref{t:effective_area} summarizes the effective area of the prime nodes at particular elemental lines of interest. The five prime REXIS nodes collectively provide an effective area of 4.4 cm$^2$ at 1.25 keV. Figure~\ref{fig:ccd20_arf} shows the best-fit effective area model of CCD 2 Node 0 for single pixel events.
\begin{table*}
\centering
\small
\caption{REXIS effective area at elemental line energies for single-pixel events (effective area of double-pixel events in parenthesis). Values are for a full detector node, and do not include the throughput of the coded aperture mask (typically about 41\%). The physical area of a node is 1.509~cm$^2$.}
\begin{tabular}{ c|c|c|c|c }
\hline\hline
CCD & Node & \multicolumn{3}{c}{Effective area [cm$^2$]} \\
ID & ID & Mg (1.25~keV) & Si (1.74~keV) & S (2.3~keV) \\
\hline
0 & 2 & 0.33 (0.69) & 0.49 (0.65) & 0.06 (0.29) \\
1 & 2 & 0.32 (0.13) & 0.29 (0.42) & 0.08 (0.27) \\
2 & 0 & 0.33 (0.80) & 0.50 (0.76) & 0.06 (0.31) \\
2 & 2 & 0.44 (0.95) & 0.57 (0.89) & 0.10 (0.39) \\
\hline
\end{tabular}
\label{t:effective_area}
\end{table*}
We performed an additional calibration in June 2019, following spacecraft operations that were expected to incur significant temperature variations on the REXIS mask (see Section~\ref{sec:mask}). The REXIS team coordinated simultaneous observations of the bright variable X-ray source Sco X-1 with the {\it NICER}\xspace experiment \citep{Gendreau16} onboard the International Space Station. The observation with REXIS lasted about 18~ksec with five separate offset pointings (see \ref{table:crab_observations}), and {\it NICER}\xspace observed the source with a total exposure of 9~ksec in five orbits over the same period. The {\it NICER}\xspace team provided a fit of the Sco X-1 spectrum with a thermal disk model, and we used that model as a fixed incident spectrum for our calibration. We reduced the Sco X-1 data in the same manner as our Crab Nebula calibration and fit the source spectrum derived from {\it NICER}\xspace to our effective area model for each parameter setting, finding agreement with the {\it NICER}\xspace data above 1~keV (see Figure~\ref{fig:spectra_LEE}).
\subsection{Optical Light Sensitivity; Light Leak} \label{sec:lightleak}
The REXIS detector was designed with a novel optical blocking filter (OBF) to prevent visible wavelength (optical) photons from interacting with the CCDs. The REXIS OBF is a thin layer of aluminum, approximately 320~nm thick and directly deposited on the top surface of the four CCID-41 chips \citep{Thayer21}. A directly-deposited OBF on the array simplifies the overall detector assembly, which can otherwise be challenging for addressing all sources of direct or scattered light. However, a directly deposited OBF can experience pinholes, or micron-sized areas with a significantly thinner deposition. CCD pixels beneath pinholes will have greater exposure to optical photons, which can generate sufficient charge to masquerade as X-ray events. For REXIS, with limited downlink capacity, the number of CCD pixels affected by pinholes could not be larger than a few hundred, otherwise events from these pixels could overwhelm the REXIS event-list data and cause true X-ray events to be discarded. To prevent the saturation of the CCD data, the REXIS flight software included an optional hot-pixel mask, i.e. a list of pixel addresses to ignore when generating the list of X-ray events. The hot pixel mask was uploaded to the REXIS firmware at each power-on cycle.
Since the complete number of pinholes was not measured before the launch, REXIS performed two observations of the sun-illuminated asteroid prior to the main observation campaign, to measure the number of likely pinholes and to confirm that the hot-pixel mask was performing as expected. During the L+30 months operation, in February 2019, REXIS collected three hours of event-list data, as well as five frames of the full detector area. Pixels that recorded significant charge in all of the frames were labeled as pinholes and included in the hot pixel mask. Ultimately, 1234 pixels (out of 1.5 million pixels in six detector nodes; less than 0.1 percent) were identified as bright and likely associated with pinholes, and were included in the hot pixel mask. REXIS successfully tested the hot pixel mask during the OBF Calibration in April 2019, and used the mask in all subsequent operations. During the primary asteroid-observation periods in Orbitals B and R, there was no evidence that pinholes contributed to the event list data.
While the OBF successfully attenuated optical photons from Bennu incident on the top surface of the REXIS CCDs, the CCDs still occasionally suffered from optical photon contamination. The inner surface of the REXIS XIS tower was coated with gold in order to minimize contamination from X-ray fluorescence generated internally within the instrument. However, this gold coating created a surface for scattering optical photons inside of the tower. As a result, there was a path for sunlight incident on structures on the {\it OSIRIS-REx}\xspace instrument deck to scatter into the REXIS CCD tower and, via multiple internal reflections, arrive at the uncoated back side surfaces of the CCD substrate (see also Section~4.3 in \citet{Rizk18}). This path for scattered light was first observed following the opening of the REXIS radiation cover, when the CCDs experienced periods of large count rates correlated with times when the REXIS boresight was less than about 110$^{\circ}$ from the Sun. This was most noticeable during calibration operations in the second half of 2018 (e.g., CXB and 1st Crab Calibrations). As a result, the data from these operations was of limited use for instrument calibration. For later operations (starting with the 2nd Crab calibration), the REXIS commanding was adjusted to regenerate the CCD bias map at regular intervals, which mitigated the effect of diffuse scattered light.
During asteroid observations in Orbital B, where the Sun-illuminated asteroid surface was directly in the REXIS field of view, there were further complications due to the light leak. Despite regenerating the bias map approximately hourly, REXIS experienced short periods (tens of minutes or less) when the CCD event list data was dominated by low-energy events, with a high density of counts near the readout edge of the CCDs where the flexprint cable connected to the CCD chip. Correlated with these periods, the rate of {$^{55}$Fe\ } events was sharply reduced, by as much as 50\% in some CCD nodes. Ground testing and simulation \citep{Lambert20} demonstrated that optical photons from the asteroid were leaking into the CCDs along the readout edge, and these photons were generating low-energy events with sufficient rate to shift the grade of true single pixel X-ray events. During Orbital R, the onboard event-processing parameters were changed to accept more multi-pixel events, and the detection efficiency measured by the {$^{55}$Fe\ } sources returned to nominal (see Figure~\ref{f:Fe_intensity}).
\subsection{Excessive X-ray Events at Low Energies below 1 keV} \label{sec:excess}
\begin{figure}
\centering
\includegraphics[width=3.4in]{figs/LEexcess_22_plaw_3.png}
\caption{REXIS spectra from CCD2 node 2 illustrating an excess in X-ray events at low energies with a power law spectral shape. From top to bottom: observed spectra of Sco X-1 (blue dots), the Crab Nebula (black dots), Internal Calibration data (red dots), and uncalibrated data from ground tests prior to launch (green dots). For the Sco X-1 and Crab Nebula spectra, the expected spectra using the modeled instrument response are shown in the red solid lines, and models with an additional power law component are show in the green solid lines. The power law component (black solid line, index -2.6) is fit to the Crab data below 1~keV; for the Sco X-1, Internal Calibration, and ground test data, the power law amplitude is scaled by the event rate observed in the 3-4~keV energy band. These results demonstrate that the excess at low energies scales directly with the incident X-ray flux.}
\label{fig:spectra_LEE}
\end{figure}
One unexpected outcome of the Crab Calibration operation was the high count rates at energies below 1~keV, as shown in Figure~\ref{fig:ccd20_crab_fit}. The observed X-ray photon flux continues to increase as the energy decreases despite the fact that both the quantum efficiency of the REXIS CCDs and the X-ray flux of the Crab Nebula are known to decrease with lower energies. In fact, in order to explain the spectrum, the quantum efficiency of the REXIS CCDs would have to unphysically exceed 100\% at energies below 1 keV. The excess was observed again in the Mask Calibration operation, and was later confirmed to be present at much lower amplitude in the Internal Calibration data (collected before the radiation cover was opened) and in ground calibration data prior to launch; see Figure~\ref{fig:spectra_LEE}.
Since the coded aperture mask shadow pattern results in only about the half of the pixels receiving photons illuminated from a point source, the low energy excess in the observed spectra of the Crab Nebula and Sco X-1 indicates that it originates from incident X-rays: i.e., the spatial information of the low-energy events correlates with the incident X-rays. This evidence, as well as results from ground tests and the observation that the magnitude of the excess scales with the incident X-ray flux, rules out several possibilities for noise, such as the cross-talk from CCD 3 (with the severed cable), the pipeline software or a previously unknown astrophysical effect. The excess pattern in the spectra is well-described by a power law model as seen in Figure~\ref{fig:spectra_LEE}. These features of the noise could be explained through the redistribution of true event energies to lower values, possibly due to charge loss in the CCD or an error in the flight software. A dead layer in the CCDs would introduce a similar effect, but models of the CCD response indicate that the thickness of the dead layer would have to be greater than 100~$\mu$m to explain the amount of the low energy excess, which contradicts the Q.E. at energies above 1~keV (where the expected CCD response accurately models the observed data). We observed similar low energy excess in the data taken with an Engineering Model (EM) CCD without an OBF, which rules out any issue with the OBF itself \citep[see also Figure 7 in][]{Koyama07}. Compton scattering of hard X-rays in the REXIS CCDs could generate low-energy events that are spatially correlated with X-rays from the Crab nebula or Sco X-1, but this contribution is not large enough to explain the magnitude of the excess. As of this writing, the REXIS instrument team does not have a complete explanation for the cause of the observed low energy excess. We continue to perform lab experiments using the flight spare and EM CCDs to find out the root cause of the low energy excess. Until the mechanism of the excess is understood and properly modeled, we have limited our analysis of REXIS data to energies above 1 keV.
\section{The REXIS Instrument}\label{sec:instrument}
REXIS is an X-ray spectrometer designed to take advantage of the incident solar X-ray flux that can generate a diagnostic fluorescence signature from an asteroid’s surface. REXIS development built upon the experience of previously flown X-ray fluorescence (XRF) experiments dating back to the Apollo era and previous asteroid missions \citep[e.g.,][]{2000Sci...289.2101T, Nittler01, Okada06}. We refer the reader to \citet{Masterson18} for a detailed description of the REXIS instrument.
In brief summary here, REXIS consists of two components: a main X-ray Imaging Spectrometer (XIS) to measure XRF from Bennu and a separate Solar X-ray Monitor (SXM) to measure the incident X-ray flux from the Sun. Figure~\ref{f:instrument} illustrates the key components of the REXIS instrument and Table~\ref{t:intrument} summarizes the key instrument parameters.
\begin{SCfigure}
\small
\caption{X-ray Imaging Spectrometer (XIS) of REXIS is a coded-aperture telescope with a stainless mask and a 2 $\times$ 2 array of X-ray CCDs. The inset shows
the Solar X-ray Monitor (SXM) of REXIS equipped with an SDD in a collimator and the readout electronics.}
\includegraphics[width=3.5in]{figs/instrument.png}
\label{f:instrument}
\end{SCfigure}
\begin{SCtable}[0.5]
\small
\caption{Key Instrument Parameters of REXIS, based on the in-flight calibrations. See also
the pre-flight design values in Table 1 in \citet{Masterson18}. Values in italic indicate changes relative to the original design.}
\begin{tabular}{ll}
\hline\hline
\bf Parameters & \bf Values \\
\hline
\multicolumn{2}{l}{\bf X-ray Imaging Spectrometer (XIS)} \\
Total Mass \& Power & 6.5 kg \& 12.4W \\
Mask to Detector$^a$ & {\it 19.1} cm \\
Field of View & 28$^\circ$ (FWHM), 53$^\circ$ (FWZI$^f$) \\
Angular Resolution$^{a,b}$ & 26.5$'$ (5.6 m @ 730 m) \\
Localization Error$^{a,b}$ & $\sim$ 1$'$ -- 5$'$ \\
\hline
\multicolumn{2}{l}{\bf Detector in XIS: 2 $\times$ 2 CCDs} \\
Pixel size & 24 $\mu$m $\times$ 24 $\mu$m \\
Depletion Depth & 45 $\mu$m \\
Optical Blocking Filter (OBF) & 320 nm Al \\
Active Area$^c$ & {\it 7.5} cm$^2$ \\
Effective Area$^c$ & {\it 4.4} cm$^2$ @ 1.25 keV (Mg-K) \\
& {\it 1.9} cm$^2$ @ 2.30 keV (S-K) \\
Energy Resolution$^c$ & {\it $\sim$150--300}~eV FWHM @ 5.9 keV \\
Energy Range$^d$ & {\it 0.4--10} keV \\
\hline
\multicolumn{2}{l}{\bf Coded Mask in XIS: 9.8 cm dia.~Stainless Steel} \\
Pixel pitch & 1.536 mm (100 $\mu$m support grid)\\
Thickness & 100 $\mu$m \\
Throughput & 41\% with 50\% open pixels\\
\hline
\multicolumn{2}{l}{\bf Solar X-ray Monitor (SXM)} \\
Detector & Amptek XR-100 SDD\\
Effective Area & {\it 0.6 mm$^2$} @ 2.0 keV \\
& {\it 0.8 mm$^2$} @ 5.9 keV \\
Field of View & {\it 42$^\circ$} (FWMI$^e$) {\it 60$^\circ$} (FWZI$^f$)\\
Energy Range$^g$ & \it $\sim$1.5 -- 20 keV \\
Optical Blocking & 0.5 mil Be window \\
\hline
\multicolumn{2}{l}{$^a$Based on the Crab Calibration Operation (Section~\ref{sec:boresight}).}\\
\multicolumn{2}{l}{$^b$See Section~\ref{s:ang_resolution}.}\\
\multicolumn{2}{l}{$^c$Based on 5 stable nodes: see also Table~\ref{t:resolution}.}\\
\multicolumn{2}{l}{$^d$Constrained by the low energy excess (Section~\ref{sec:excess}).}\\
\multicolumn{2}{l}{$^e$Full Width Maximum Intensity}\\
\multicolumn{2}{l}{$^f$Full Width Zero Intensity} \\
\multicolumn{2}{l}{$^g$Nominal range, see Section~\ref{sec:sxm}}
\end{tabular}
\label{t:intrument}
\end{SCtable}
The XIS employs a Detector Assembly Mount (DAM) consisting of 4 MIT Lincoln Laboratory CCID-41 charge-coupled devices (CCDs) in a 2~$\times$~2 array \citep{Bautz04}.
Each CCD is further divided into 4 separate readout channels, which we refer to as “nodes”, with each node consisting of 256~$\times$~1024 pixels of dimension 24 microns. The CCID-41s are sensitive to the energy range of 0.4 to 10~keV with a spectral resolution (FWHM) of $\lesssim$ 220 eV at 5.9 keV. The primary fluoresced lines REXIS is designed to detect are Fe-L at 0.71 keV, Mg-K at 1.25 keV, Al-K at 1.49 keV, Si-K at 1.74 keV, and S-K at 2.31 keV.
REXIS includes a coded-aperture mask \citep{Dicke68}, unique to planetary science X-ray detectors, to facilitate image reconstruction of relative variations in elemental abundance across the surface of the asteroid. The REXIS mask contains a random pattern of holes with a throughput of 41\%. The REXIS main detector array was protected from background radiation during the cruise phase by a cover that was successfully opened in September 2018, after approximately two years into flight.
REXIS' second component, the SXM, utilizes an Amptek XR-100 Silicon Drift Detector (SDD) with a collimated effective area of about 0.8~mm$^2$, providing spectral coverage over a nominal energy range of 1.5 to 20 keV. With a large field of view (FWZI: 60$^\circ$), the SXM design enables monitoring of solar X-rays during nadir pointing of the XIS near or at terminator orbits of Bennu.
The Main Electronics Board (MEB), located beneath the REXIS XIS, serves to power and control the XIS and SXM. It also processes the XIS and SXM science data, and collects the instrument housekeeping (HK) data for downlink.
\subsection{REXIS On-board Signal Processing}
Because spacecraft resources are finite, REXIS operations were required to fit within specific allocations of on-board data storage and data download volume. Thus for all flight operations and outcomes we are reporting, it was necessary to process the full CCD images (frames) on-board. \cite{Allen13} and \cite{Jones14} describe the implementation of the REXIS firmware, and details of the REXIS software are given by \citet{Biswas16}. In brief, during REXIS observations the on-board software identified X-ray photon detections (excess charge above a preset threshold) in individual pixels and cataloging those photon detections as CCD “events”, so that only lists of CCD events (“event lists”) are downlinked to Earth instead of full frame images. Along with the CCD event list, instrument housekeeping (HK) parameters and the SXM histograms (see below) are also downlinked to Earth.
During nominal REXIS operations, the CCDs collected 4-second exposures. At the end of each exposure the charge from each pixel is quickly (within 10 ms) transferred to a non-imaging framestore area of the detector before readout by the detector electronics. The framestore region is read serially into a single large array by sixteen 12-bit analog to digital converters (ADCs), one for each CCD node. As the data are read out, the firmware executes four steps of image processing: frame processing (e.g.~subtracting the noise pattern using a bias map and eliminating known bad pixels), event grading, event filtering, and finally science packet formation. The final science packet that is downlinked to Earth includes an X-ray event list giving the pixel location, energy, and grade of each putative X-ray event. See Figure 11 in \citet{Masterson18} for the flow chart of the REXIS Image Processing Pipeline.
Of particular importance to REXIS calibration and in-flight performance is the “grading” of individual photon detections. When an X-ray interaction occurs near the edge of a pixel in a CCD, the induced charge cloud can be distributed among the neighboring pixels and not easily recognized as a part of the discrete event. X-ray events are {\it graded} based on the degree of the charge splitting using a pre-set ``split-charge threshold" (ST). Neighboring pixels above the ST are considered a valid signal, and their charge is added back to the sum of the signal. Depending on the proximity of the X-ray interaction to the pixel edge, the depth of the interaction in the Si layer and the electron cloud size (i.e., the energy of the incident X-ray), the split charge may be divided among as many as four pixels.
During the event grading, each candidate detection is assigned a grade from 0 to 7 to identify how the charge from a single incident X-ray photon was divided among multiple pixels. A Grade 0 event corresponds to an event localized in a single pixel (with no neighboring pixels above the split-charge threshold), while Grades 1 and higher denote events where charge is shared among one or more pixels adjacent to the central pixel. Multi-pixel events generally yield poorer spectral resolution relative to single pixel events, since the combined electric noise and charge transfer inefficiency from multiple pixels contribute to the final signal sum.
After the event grading algorithm, the flight software produces a list of valid CCD events for downlink. An event filtering algorithm ensures that the science data volume would not exceed downlink allocations during unexpected occurrences of high event rates due to factors such as solar flares, optical light leaks, hot pixels, or software glitches. The event rate filter was designed to randomly select at most 1000 (an adjustable parameter) events for downlink over each 4-second readout cycle.
The REXIS software was capable of downlinking a full or partial frame image of the XIS detector plane for a 4 second exposure, or the bias map of the entire plane (noise map). Downlinking frame images was a slow process, taking about 30 minutes for an entire detector plane image or 2 minutes for a full node image. Therefore, REXIS only downlinked full or partial frame and bias images when they were necessary for diagnostic purposes or other detailed analyses.
Unlike the fixed readout cycle in the REXIS CCDs, the SDD in the solar X-ray Monitor (SXM) responded continuously to incoming X-rays. The SXM readout electronics triggered on signals above a preset threshold. Given the potential for high solar X-ray flux during flares, REXIS did not downlink the individual SXM events, but rather a histogram that recorded the number of events observed in given energy bands. The SXM histogram had 512 bins with an energy spacing of about 59~eV and a maximum energy of 24~keV, and accumulated events over an integration period that alternated between 18 and 32 seconds. The alternating integration period was due to a software bug, but it had the beneficial effect of providing robustness to periods of high solar X-ray activity.
In addition to the REXIS science data, the MEB collected various housekeeping (HK) information from the REXIS system, and downlinked each HK packet every 64 seconds. The HK data included the voltage and current measurements of the CCD input biases, X-ray event rates (even when the CCD event data is not being downlinked), event filtering status, the output of several temperature sensors, and other diagnostic information.
\subsection{REXIS Pre-flight Challenges}
\citet{Masterson18} described the pre-flight activities of the REXIS main spectrometer and solar X-ray monitor. As is the case for any instrument, the pre-flight activities (and challenges) shaped the status of the flight hardware and its operation in space.
In late 2014 and early 2015, Lincoln Laboratory delivered 8 CCDs to be used for the flight model (FM) and flight spare Detector Assembly Mounts (DAMs). Based on the measured performance of each of the CCDs (susceptibility to light leak, dark current level, and surface cosmetics), 4 CCDs were ranked Grade A, 2 CCDs Grade B, and the remaining 2 CCDs Grade C, where Grade A was considered to be of the highest quality. To mitigate risk, the REXIS team distributed the the CCDs between the flight model and flight spare detector assemblies: we assigned two Grade A CCDs to each assembly, two Grade B CCDs to the FM DAM, and the remaining two Grade C CCDs to the flight spare DAM.
Dividing the CCDs between the two assemblies proved to be prudent, as the original FM DAM suffered contamination during a bake-out operation in a thermal vacuum chamber. Consequently, the flight spare DAM (with the two Grade C CCDs) was ultimately used as the flight instrument. The two Grade A CCDs used in the flight instrument were labeled as CCDs 1 and 3, and the Grade C CCDs were CCDs 0 and 2 (see Section \ref{sec:calibration}). A further complication arose during vibration testing of the newly promoted flight model DAM, when a tear in a flexprint cable disabled signal connection from CCD 2 Node 3 and all nodes from CCD 3. Thus, at launch, the REXIS XIS had ten functioning nodes out of the 16 nodes in the detector assembly. It was structurally necessary to leave the torn flexprint cable in place (but sealed to prevent further tearing), and this contributed to elevated electronic noise in the odd-numbered nodes (see Section~\ref{s:response}).
For the SXM, during pre-launch testing we discovered that the trigger circuit for the SDD signals in the main electronics board was not properly connected. In addition, the MEB was found to suffer from ground-loop noise. We resolved these issues by introducing an additional electronics module (the "backpack board") that also served to condition the power for the Thermo-electric cooler (TEC) for the SDD.
Due to these issues and other schedule delays, many aspects of the REXIS pre-flight ground calibration program were not fully completed, for example measurements of the CCD detection efficiency and energy resolution with external X-ray sources. Luckily, it was possible to perform pre-flight performance checks and troubleshooting using self-calibration of X-ray events from the onboard {$^{55}$Fe\ } sources during thermal testing. In the case of the SXM, we performed several laboratory X-ray measurements using an external {$^{55}$Fe\ } source to validate its functionality prior to flight.
|
\section{Introduction}\label{sec:introduction}
(Online) debate has long been studied and modelled by computational linguistics with argument mining tasks such as stance detection. Stance detection is the task of automatically identifying the stance (agreeing, disagreeing, and/or neutral) of a text towards a debated topic or issue \cite{StanceSurvey, schiller_stance_2021}.\footnote{There is a wide array of datasets, definitions, and operationalizations of stance detection and classification, and recently \citet{schiller_stance_2021} gave a great overview in their Section 2, as do \citet{StanceSurvey} in their survey.} Its use-cases increasingly relate to online information environments and societal challenges, such as argument search \cite{stab2018argumentext}, fake news identification \cite{hanselowski2018retrospective}, or diversifying stances in a news recommender \cite{reuver-etal-2021-nlp}.
Cross-topic stance detection models should thus be able to deal with the quickly changing landscape of (online) public debate, where new topics and issues appear all the time. As \citet{schlangen-20191-natural} described in his recent paper on natural language processing (NLP) methodology, generalization is a main goal of computational linguistics. A computational model (e.g.\ a stance detection model) should learn task capabilities beyond one set of datapoints, in our case: beyond one debate topic.
Cross-topic stance detection is especially challenging because generalization to a new discussion topic is not trivial. Expressing stances is inherently socio-cultural behavior \cite{du2007stance}, where social actors place themselves and targets on dimensions in the socio-cultural field. This also comes with very topic-specific word use \cite{somasundaran-wiebe-2009-recognizing, wei2019modeling}. For instance, an \textit{against} abortion argument might be expressed indirectly
with a `pro-life' expression, and someone aware of the socio-cultural context of this debate will be able to recognize this. Knowledge from other debate topics such as \textit{gun control} may not be useful, since the debate strategies might change per topic.
Despite these fundamental challenges, pre-trained Transformer models show promising results on cross-topic argument classification \cite{reimers_classification_2019, schiller_stance_2021}.
In this paper, we investigate the ability of cross-topic stance detection approaches to generalize to different debate topics. Our question is: \textit{To what extent is stance detection topic-independent and generalizable across topics?}
Our contributions are threefold. We first complete a reproduction of state-of-the-art cross-topic stance detection work \cite{reimers_classification_2019}, as reproduction has repeatedly shown to be relevant for NLP \cite{fokkens2013offspring, cohen2018three, belz-etal-2021-systematic}. The reproduction is largely successful: we obtain similar numeric results.
Secondly, we investigate the topic-specific performance of this model, and conclude that BERT's performance fluctuates on different topics. Additionally, we find that a bag-of-words-based SVM model can rival its performance for some topics. Thirdly, we relate this to the nature of the stance detection modelling task, which is inherently more connected to socio-cultural aspects and topic-specific differences than related tasks such as sentiment analysis.
This paper is organized as follows. Section~\ref{sec:Background} discusses earlier work on stance detection, and specifically generalizability across topics. Section~\ref{sec:Repro} presents the reproduction results. Section~\ref{sec:Topics} adds additional, topic-specific analyses of the classification performance and a bag-of-words-based model to find topic-(in)dependent features. This is followed by our conclusions in Section~\ref{sec:Discussion}.
\section{Background}\label{sec:Background}
\subsection{Definition of Stance Detection}
Stance detection is a long-established task in computational linguistics. \citet{StanceSurvey} identify its most commonly used task definition: “For an input in the form of a piece of text and a target pair, stance detection is a classification problem where the stance of the author of the text is sought in the form of a category label from this set: {Favor, Against, Neither}.” \cite[p. 2]{StanceSurvey}.\footnote{We would like to note that the stance expressed in a text unit does not have to be the stance of an author, e.g.\ in cases where someone is writing a piece in which they express or quote someone else's opinion.} The number of stance classes can vary from 2 to 4, e.g.\ by adding `comment' and `query' next to `for' and `against' \cite{schiller_stance_2021}. \citet{StanceSurvey} emphasize that this computational definition is built upon the linguistic phenomenon of actors communicating their evaluation of targets, by which they place themselves and their targets on ``dimensions in the sociocultural field'' \cite[p. 163]{du2007stance}. Current work focuses mostly on debates deemed controversial in the U.S. socio-political domain, such as abortion and gun control.
\subsection{Prior work}
Early work on stance detection focused on parliamentary debates and longer texts \cite{thomas2006get}.
Since \citet{mohammad-etal-2016-semeval}'s stance detection shared task,
Twitter has attracted a lot of attention in NLP work on stance detection \cite{zhu2019hierarchical, darwish2020unsupervised, hossain2020covidlies}. Others addressed stance detection in the news domain, with (fake) news headlines \cite{ferreira2016emergent, hanselowski2018retrospective}, disinformation \cite{hardalov2021survey} and user comments on news websites \cite{bovsnjak2019data}.
Feature-based approaches have largely been replaced by end-to-end neural models. Stance detection has seen a performance increase due to pre-trained Transformer models such as BERT \cite{devlin-etal-2019-bert}. \citet{reimers_classification_2019} reported .20 point F1 improvement over an LSTM baseline with a pre-trained BERT model. Combining multiple stance detection datasets in fine-tuning such a pre-trained Transformer again led to a performance increase, though this model lacks robustness against slight test set manipulations \cite{schiller_stance_2021}.
\subsection{Generalization to new topics}
Recent work has specifically worked on identifying stances on topics not seen in training.
\citet{reimers_classification_2019} train their model on detecting stances and arguments for unseen topics. In their approach however, they treat all topics and stances on these topics as similar and comparable, and report one averaged evaluation metric over topics.
Earlier work \cite{somasundaran-wiebe-2009-recognizing} already established that
ideological stances on topics deemed controversial, such as gay rights, are expressed in a topic-specific manner. Topic-specific features
were more informative for SVM models than more topic-independent features.
In more recent work, \citet{wei2019modeling} instead specifically focus on how generalizable certain topics are for transferring knowledge to new topics on stance detection. Some Twitter discussion topics seem to share a latent, underlying topic (e.g.\ both feminism and abortion have the latent topic of equality). In a (latent) topic-enhanced multi-layer perceptron (MLP) model with RNN representation of the tweet, the model indeed uses shared vocabulary between the related topics.
\citet{allaway-etal-2021-adversarial} notice that earlier work, when considering training on some topics and testing on others, incorporates topic-relatedness. Unlike these other studies however, \citet[p. 4756]{allaway-etal-2021-adversarial} ``do not assume a relationship between training and test topics'' as a fairer test of robustness. Results they present do show that stance detection is related to topic, but their efforts go to finding topic-invariant stance representations, which improves the generalizability of their model.
Their consideration of topic similarity shows that topic difference is very relevant to stance detection.
\citet{aldayel_stance_2021} describe in their survey how several studies \cite{ klebanov2010vocabulary, zhu2019hierarchical, darwish2020unsupervised} show that texts pro or against an issue use different vocabularies (e.g.\ using `pro-life' when expressing a stance against abortion). Some of these studies attempt to leverage these vocabularies to generalize across similar topics. Recent work has looked into generalizing stance detection across datasets, task definitions, and domains \cite{schiller_stance_2021}, in which topic-specific performance is not mentioned.
A recent approach to topic-specificity in stance detection is task adaptation. \citet{stein2021same} acknowledge that stance detection usually requires knowledge about the topic of discussion, which is not available for unseen topics. They approach this problem by changing the task to ``same-side stance classification", in which a model is trained to classify whether two arguments either have the same or a different stance. This reduces the model's leaning on topic-specific pro- and con-vocabulary, while still being able to separate different stances on the same topic. The best approach to this adapted task on a dedicated leaderboard\footnote{\url{https://webis.de/events/sameside-19/}, Accessed on the 22th of September 2021.} receives an F1 of .72 in the cross-topic setting with a fine-tuned BERT model \cite{ollinger_same_2020}.
Our current work adds the discussion of topic difference and topic specificity to state-of-the-art stance detection results. That is, earlier bag-of-words-based work considered lexical specificity of different topics for stance detection, and we add that into the discussion for the current state of the art: pre-trained, end-to-end neural models.
\section{Reproduction Experiments}\label{sec:Repro}
\citet{reimers_classification_2019} apply their approach of cross-topic claim classification to two datasets: the \textit{UKP Sentential Argument Mining Corpus} \cite{stab2018argumentext} (`the UKP dataset') and the \textit{IBM Debater: Evidence Sentences} dataset \cite{shnarch-etal-2018-will} (`the IBM dataset'). We focus on the UKP Dataset, since the IBM Debater dataset has no `pro' and `con' class, but rather `evidence' and `no evidence' (and our focus is on stance detection).
As a second step after stance classification, the authors also attempt to cluster similar arguments within the same topic in a cross-topic training setting. We do not replicate this component, but instead dive deeper into the classification results.
We adopt the definition of reproduction by \citet{belz-etal-2021-systematic}: repeating the experiments as described in the earlier study, with the exact same data and software. We analyze our reproduced results according to the three dimensions of reproduction proposed by \citet{cohen2018three}: whether we find either the same or different (1) (numeric) values, (2) findings, and (3) conclusions as the earlier study.\footnote{For reasons of clarity, we present these dimensions in reverse order compared to \newcite{cohen2018three}.
} Reproducing the same \textbf{values} means obtaining the same numeric results from a specific experiment. Experiments involving fine-tuning on BERT are non-deterministic. We therefore consider the metric fully reproduced if the original result lies within two standard deviations (stdevs) from our result, obtained from 10 random seeds.\footnote{The paper we reproduce, \citet{reimers_classification_2019}, does not provide model performance standard deviation over seeds.} The same \textbf{finding} means that the relation between the values associated with two or more dependent variables is the same, i.e.\ a system that outperformed another in the original study also does this in the reproduced study. The \textbf{conclusion} is the same when the broader implication of findings and values is the same. Conclusions are thus a matter of interpretation. As such, the same findings can lead to different conclusions and conclusions are, contrary to findings, not repeatable \cite{cohen2018three}. This section focuses on the repeatable components of reproducing a study: the values and the findings. We address the conclusions using our more detailed analyses in Section~\ref{sec:Topics}.
\subsection{Dataset Description}
The UKP dataset \cite{stab2018argumentext} consists of 25,492 argument sentences from 400 Internet texts (from essays to news texts) on 8 topics.
The dataset designer's definition of claim is ``a span of text expressing evidence or reasoning that can be used to either support or oppose a given topic'' \cite[p. 3665] {stab2018argumentext}. They define topic as ``some matter of controversy for which there is an obvious polarity for possible outcomes'' \cite[p. 3665]{stab2018argumentext}, and map this polarity to a text expressing one of two classes: for or against the use, adoption, or idea of the topic under discussion. A third class is `no argument' to the topic under discussion, i.e.\ the text span falls outside of this polarity.
The 8 topics in the dataset were randomly chosen from online lists of controversial topics on discussion websites \cite[p. 3666] {stab2018argumentext}. Specifically, these topics are \textit{abortion}, \textit{cloning}, \textit{death penalty}, \textit{gun control}, \textit{marijuana legalization}, \textit{minimum wage}, \textit{nuclear energy} and \textit{school uniforms}. The stance classes (pro, con, and no argument) were annotated by two argument mining experts and seven U.S.\ crowdworkers. The distribution of the dataset for different topics is shown in Table~\ref{tab:distribution}.
\begin{table*}[t]
\centering
\small{
\begin{tabular}{llccccccccc}
\hline
& & \small{abortion} & \small{cloning} & \small{death} & \small{gun} & \small{marijuana} & \small{minimum} & \small{nuclear} & \small{school} & \small{all}
\\
& & & & \small{penalty} & \small{control} & \small{legalization} & \small{wage} & \small{energy} & \small{uniform} \\
\hline
\textbf{train} & \textit{pro} & 490 & 508 & 316 & 566 & 422 & 414 & 436 & 392 & 3.544\\
& \textit{con} & 591 & 604 & 789 & 479 & 450 & 396 & 613 & 525 & 4.447\\
& \textit{no arg} & 1.746 & 1.075 & 1.522 & 1.359 & 908 & 968 & 1.524 & 1.248 & 10.350\\
\hline
\textbf{dev} & \textit{pro} & 54 & 56 & 38 & 63 & 47 & 46 & 48 & 44 & 396 \\
& \textit{con} & 66 & 67 & 90 & 53 & 50 & 44 & 68 & 58 & 496 \\
& \textit{no arg} & 195 & 120 & 165 & 152 & 101 & 108 & 170 & 139 & 1.150\\
\hline
\textbf{test} & \textit{pro} & 136 & 142 & 103 & 158 & 118 & 116 & 122 & 109 & 1.004 \\
& \textit{con} & 165 & 168 & 232 & 133 & 126 & 111 & 171 & 146 & 1.252 \\
& \textit{no arg} & 486 & 299 & 396 & 378 & 253 & 270 & 424 & 347 & 2.853\\
\hline
\end{tabular}
}
\caption{Distribution of the UKP data over topics and over training (70\%), test (20\%), and validation (10\%) sets.
}
\label{tab:distribution}
\end{table*}
In \citet{stab2018argumentext} we see a difference in agreement on stance classes in different topics, especially between expert and crowd. The topic achieving the highest agreement between crowd worker and expert is \textit{school uniforms} ($\kappa = .889$), and the lowest is \textit{death penalty} ($\kappa = .576$). The standard deviation over topics is .08 for expert--expert coded data and .16 for expert--crowd coded, both with a mean of $\kappa =.72$.
\subsection{Obtaining the Data}
The UKP Dataset is not available online due to copyright concerns, but there is a scraping script with archived hyperlinks available on \citet{reimers_classification_2019}'s GitHub page.
We ran this script with all specifications given. The scraping script was able to return all claims on 6 of the 8 topics. The topics for which not all claims were detected were \textit{nuclear energy} and \textit{minimum wage}. We then instead obtained the complete datafiles from the authors.\footnote{These files revealed that the scraping script broke down in the \textit{minimum wage} topic due to one specific claim that was archived, but could not be retrieved.
``Despite the inevitable negative outcomes that will surely result from a \$ 15 minimum wage – we ’ve already seen negative effects in Seattle ’s restaurant industry – politicians and unions seem intent on engaging in an activity that could be described as an “economic death wish.'' We speculate this claim could possibly not be retrieved due to it containing the dollar sign, \url{https://web.archive.org/web/20160217041546/http://www.aei.org:80/publication/ten-reasons-economists-object-to-the-minimum-wage/}}
\subsection{Training and Evaluation Method}
\citet{reimers_classification_2019} use the training method described in \citet{stab2018argumentext}. Each topic is split into a training (70\%), development (10\%), and test split (20\%). Training is done on the training splits of 7 topics, tuned on the development split (10\%) of these 7 topics, and finally evaluated on the test split (20\%) of the held-out 8th topic. They do this for each of the 8 topics (holding out a different topic each time), then apply this procedure for 10 different random seeds on a GPU. Evaluation is assessed with macro F1, averaged over all topics and all random seeds. Their best performing model is a fine-tuned BERT-large model \cite{devlin-etal-2019-bert}, but with only minor improvement over BERT-base.
We use the same training set-up and BERT models for our reproduction. For training, we use the author's code with Python3.8 on a single NVIDIA GeForce RTX 2080 Ti GPU. Our learning rate is 2e-5 for both models, as in \citet{reimers_classification_2019}.\footnote{All our code can be found in the following GitHub repository: \url{https://github.com/myrthereuver/claims-reproduction}.}
We additionally train a non-BERT model (a Support Vector Machine (SVM) with tf-idf features) in the same hold-one-topic-out manner. Tf-idf-based approaches have shown quite solid performance on stance detection in prior work \cite{riedel2017simple}. This model is deterministic and is thus not run with multiple seeds. It is run with Python3.9 and the sklearn package. The SVM is intended for the feature analysis in Section~\ref{sec:Analysis}, but we present the performance of this model also in Table~\ref{tab:reproduction} and the following section.
\subsection{Results of Reproduction}
\begin{table*}[h]
\centering
\small{
\begin{tabular}{l|cccccccc}
\hline
\textbf{Model} & \multicolumn{6}{c}{\textbf{UKP Dataset}} \\
\hline
\small{mean (stdev) 10 seeds} & F1 & P \small{pro} & P \small{con} & R \small{pro} & R \small{con}
\\\hline
\small{\citet{reimers_classification_2019} biclstm+BERT} & .424 & .267 & .389 & .281 & .403 \\
\small{\citet{reimers_classification_2019} BERT base} & .613 (-) & .505 (-) & .531 (-) & .470 (-) & .576 (-) \\
\small{\citet{reimers_classification_2019} BERT large} & \textbf{.633} (-) & .554 (-) & .584 (-) & .505 (-) & .560 (-) \\
\hline
\hline
\small{SVM+tf-idf} & .517 & .418 & .460 & .414 & .423 & \\
\small{Reproduction BERT-base} & \textbf{\small{.617 (.006)}} & .519 (.011) & .538 (.007) & .464 (.029) & .581 (.019) & \\
\small{Repr. BERT-large - all seeds} & .596 (.043) & .483 (.057) & .527 (.057) & .464 (.058) & .516 (.063) \\ \hline \hline
\small{Repr. BERT-large - 5 evenly performing seeds} & .636 (.007) & .532 (.014) & .578 (.016) & .515 (.016) & .567 (.022) \\ \hline
\end{tabular}}
\caption{Reproduction results \citet{reimers_classification_2019}. The fourth row shows our non-BERT model (an SVM) beating their LSTM baseline, and the fourth and fifth row show the results of our BERT reproductions. The sixth row shows an average BERT-large performance without the 5 seeds that considerably under-performed for one topic.
}
\label{tab:reproduction}
\end{table*}
\paragraph{BERT-base} Table \ref{tab:reproduction} shows that mean performance over the 3 classes (`pro', `con', or `no argument') is F1 = .617 (stdev over 10 seeds = .006). \citet{reimers_classification_2019}'s reported result (F1 = .613) lies within 1 stdev from this result.
\paragraph{BERT-large} Mean performance over all topics and stance classes is F1 = .596 (stdev over 10 seeds = .043). The performance reported in \citet{reimers_classification_2019} is F1 = .633, which lies within 2 stdev of our result. However, our stdev is relatively high due to high variance of performance over different seeds, with half of our seeds performing noticeably lower than even BERT-base.\footnote{Our large variance in performance over seeds is due to each seed fine-tuning the model 8 times (once for each topic). The 5 unevenly performing seeds each under-perform on a different topic (F1 < .50) due to only assigning the majority class (`no argument'). Other topics in these 5 seeds do outperform BERT-base.} For the other 5 seeds, the model performed better (F1 = .636, stdev = .007), and within one (much smaller) stdev of the performance reported in \citet{reimers_classification_2019}.
\paragraph{SVM+tf-idf (non-BERT model)} This model performs at F1 = .517 averaged over the held-out topics and three classes (`pro', `con', and `no argument'), see Table \ref{tab:reproduction}. This outperforms by .10 points in F1 the best performing LSTM-based architecture presented in \citet{stab2018argumentext} (F1 = .424), a baseline in \citet{reimers_classification_2019}. Their performance improvement of the BERT model over LSTM was .20 in F1. Comparing our SVM model to BERT, we find a smaller improvement over a non-BERT model: .10 F1 improvement for BERT-base (F1 = .617).
Our BERT models still outperform our non-BERT model, as in \citet{reimers_classification_2019}. Our SVM result does fall within 2 stdevs of BERT-large, but this is due to BERT-large's substantial stdev due to a steep drop in performance for half of the seeds.
\subsection{Conclusion of reproduction}
\citet{reimers_classification_2019}'s results are reproducible in the sense the first dimension of reproducibility \cite{cohen2018three}: the originally reported numeric \textbf{values} fell within 2 stdevs of our reproduced results for both BERT-base and BERT-large. For BERT-base and 5 of the 10 seeds in BERT-large, we obtained a precision, recall, and F1 that are very similar to the original study.
The results are also reproducible in four of the five reproducibility aspects identified by \citet{fokkens2013offspring}: under-descriptions of preprocessing, experimental set-up, versioning, and system output. These were described in either the paper, on the author's GitHub page, or in code documentation. We do observe differences in relation to `system variation' which is inherent to training neural networks, where identical results are seldom obtained. These variations were small for most experiments, except for the 5 random seeds that led to substantial under-performing on one topic for BERT-large.
When looking at the second dimension of reproducibility defined by
\citet{cohen2018three} (\textbf{findings}), we observe that BERT-base and BERT-large indeed clearly outperform the LSTM baselines from \citet{stab2018argumentext} as well as our own stronger SVM+tf-idf non-BERT model on the stance detection task. We were able to reproduce the reported increase in performance of BERT-large over BERT-base and non-BERT models. However, BERT-large also showed considerable under-performance on one topic in 5 out of 10 seeds. We see this outcome as a confirmation that it is important to look at different seeds, and that care should be taken when drawing conclusions based on minor differences when working with neural models.
The third dimension of reproducibility is that of \textbf{conclusions}. \newcite{reimers_classification_2019} conclude that BERT strongly outperforms previous results on identifying arguments for unseen topics, which we confirm, and that these results are ``very encouraging and stress the feasibility of the task'' \cite[p. 575]{reimers_classification_2019}. The remainder of this paper provides further analyses to investigate whether our results also lead to this overall conclusion. In particular, we investigate how our models perform on individual topics (Section~\ref{sec:Topics}) and generic topic-independent signals in the data (Section~\ref{sec:Analysis}).
\newpage
\section{Topic Specifics in Classification}\label{sec:Topics}
To support the conclusions in \citet{reimers_classification_2019} on the success of cross-topic stance detection, we expect a relative stability of performance over topics. The following sections go into some details not explored in \citet{reimers_classification_2019}, specifically the cross-topic performance of different topics, and the interaction between topic and class and its influence on performance.
\subsection{Variance over (classes in) topics}
\begin{table*}[h]
\centering
\small{
\begin{tabular}{l|cccccccc}
\hline
\small{\textbf{held-out}} & \small{abortion} & \small{cloning} & \small{death} & \small{gun} & \small{marijuana} & \small{minimum} & \small{nuclear} & \small{school}
\\
\small{\textbf{topic}} & & & \small{penalty} & \small{control} & \small{legalization} & \small{wage} & \small{energy} & \small{uniform} \\
\hline
\small{SVM+tf-idf} & \small{.463} & \small{.585} & \small{.482} & \small{.515} & \small{.323} & \small{.615} & \small{.598} & \small{.576} \\
\small{BERT-base} & \small{.533 (.011)} & \small{.693 (.013)} & \small{.562 (.012)} & \small{.530 (.013)} & \small{.607 (.016)} & \small{.670 (.009)} & \small{.660 (.011)} & \small{.678 (.016)} \\
\small{diff. } &
\small{+.070} & \small{+.108} & \small{+.080} & \small{\textit{+.028}} & \small{\textbf{+.283}} & \small{+.055} & \small{+.0850} & \small{+.102} \\
\hline
\end{tabular}
}
\caption{BERT-base's performance in F1 (macro) on different held-out topics. The \textit{italicized} difference shows the smallest difference between the SVM model and the BERT-base model (on the gun control topic), while the \textbf{bolded} difference shows the largest difference (on the marijuana topic).}
\label{tab:reproduction_topics}
\end{table*}
Table~\ref{tab:reproduction_topics} presents the performance of the models on individual topics. The results show that some topics perform considerably worse than others with the cross-topic training method (training on seven topics and testing on the held-out eighth topic). The \textit{cloning} topic performs more than .07 F1 higher than the averaged model performance (F1 = .693 vs F1 = .617). The \textit{abortion} and \textit{gun control} topics perform almost .09 lower than the averaged model performance (F1 = .533 \& .530 vs F1 = .617). Note that a difference nearing .10 in F1 score is relatively large, as it is comparable to the difference between the SVM performance and the state-of-the-art BERT models in the previous section.
A per-topic analysis in Table \ref{tab:reproduction_topics} shows that the SVM+tf-idf model performs within .10 points of the BERT-base model for seven of the eight topics, with some performing less than .3 points lower than BERT. The only exception is the topic \textit{marijuana legalization}, which performs .28 points lower than the BERT model. The large average performance increase (+.11 in F1) over SVM comes from BERT-base improving performance on this one topic.
Figure~\ref{fig:topic/class} presents the BERT-base in-class F1 score of the three classes (`pro', `con', `no argument'), and in-topic averaged F1. The red line indicates the average model performance of .617. We see some consistency, e.g.\ the `no argument' class consistently scoring around F1 = .80, but we also see some topic-specific behavior. \textit{Cloning}, \textit{minimum wage}, and \textit{school uniforms} obtain higher F1 performance than average for all classes. In contrast, \textit{death penalty}, \textit{gun control}, and \textit{abortion} perform considerably lower than the average F1 performance in the `pro' and `con' classes. These topics see in-class performance of even F1 < .50.
Each cross-topic model is trained by removing one topic from the training data. In this way, we remove a different number of training examples each time. The topics with the most training examples for a class (e.g.\ `pro' in the \textit{gun control} topic) therefore have a smaller training set for this class when training a cross-topic model. If there were a linear relationship between dataset size and performance, one would expect that topics with \textit{fewer} training examples (and therefore more training examples left when this topic is left out of training) to do better than topics with \textit{more} training examples (whose cross-topic models lose more training examples).
Table~\ref{tab:distribution} does show that the `no argument' class has a three times larger proportion of the training set than the `pro' and `con' classes, which could explain the better performance of this class in all topics, but training set size difference does not account for the between-topic variation in the `pro' and `con' classes. Instead, Table~\ref{tab:distribution} shows that topics with the most training examples (that means, the largest set of examples removed in a cross-topic model) do not have the worst performing cross-topic models in Figure~\ref{fig:topic/class}. For example, the \textit{abortion} topic has relatively few `con' examples removed (591) compared to other classes such as \textit{cloning}, \textit{death penalty}, and \textit{nuclear energy}, and yet has the lowest in-class F1 for the `con' class (in-class F1 = .40). Performance thus appears to be less related to the number of training examples.
\begin{figure*}[h!]
\centering
\includegraphics[width=0.90\textwidth,trim={0 2cm 0 0},clip]{test_PDF_9.pdf}
\caption{BERT-base's performance on different topics plotted in a boxplot, with on the y-axis the F1 score of the 4 categories plotted on the x-axis: `pro', `con', `no argument', and overall. A longer boxplots means more variability over seeds in score. The red line represent the averaged F1 score of the same model (BERT-base), presented as model performance in \citet{reimers_classification_2019}.}
\label{fig:topic/class}
\end{figure*}
\begin{table*}[t]
\centering
\begin{tabular}{lllll|lllll}
\hline
\small{all topics} & & & & & \small{abortion topic}\\ \hline
\small{Pro (vs Con)} &\small{Con (vs Pro)} & \multicolumn{2}{c}{\small{No Argument}} & & \small{Pro (vs Con)} &\small{Con (vs Pro)} & \multicolumn{2}{c}{\small{No Argument}} \\\hline
\small{} & & \small{vs Pro} & \small{vs Con} & & & & \small{vs Pro} & \small{vs Con} \\\hline
\small{pejorative} & \small{\textit{morality}} & \small{basic} & \small{pronounced} & & \small{seek} & \small{babies} & \small{way} & \small{anti} \\
\small{pronounced} & \small{format} & \small{section} & \small{threatens} & & \small{illegal} & \small{abortion} & \small{against} & \small{ways} \\
\small{activity} & \small{\textit{bill}} & \small{take} & \small{additional} & & \small{reproductive} & \small{life} & \small{we} & \small{over} \\
\small{relations} & \small{workshop} & \small{\textbf{robert}} & \small{revolt} & & \small{simply} & \small{conception} & \small{side} & \small{always} \\
\small{additional} & \small{workers} & \small{introduced} & \small{now} & & \small{humane} & \small{simply} & \small{justify} & \small{thing} \\
\small{unexceptional} & \small{\textit{sources}} & \small{unquestioned} & \small{proper} & & \small{bear} & \small{risks} & \small{experience} & \small{question} \\
\small{threatens} & \small{\textit{philosophical}} & \small{revolt} & \small{typical} & & \small{lifers} & \small{abortions} & \small{held} & \small{performed} \\
\small{variable} & \small{coincidentally} & \small{scientifically} & \small{mentor} & & \small{mother} & \small{complications} & \small{tell} & \small{debate} \\
\small{39th} & \small{\textit{statutes}} & \small{\textbf{lifenews}} & \small{sharing} & & \small{healthy} & \small{birth} & \small{single} & \small{illegal} \\
\small{where} & \small{phrases} & \small{individuals} & \small{denuded} & & \small{lives} & \small{kill} & \small{had} & \small{equal} \\
\hline
\end{tabular}
\caption{Top-features for different topics according to SVM, Pairwise F-based feature analysis. We see potentially meaningful words in \textit{italics} (the `con' class has features based on morality and legality, e.g.\ bills and statutes), and potential spurious features in \textbf{bold} (such as names websites and even of individuals).\\}
\label{tab:repr_results_SVM_features}
\end{table*}
We investigated the source of low performance on the `pro' and `con' class in the \textit{abortion} topic with confusion matrices, and compared this to a topic where pro and against did not under-perform (\textit{minimum wage}). We did not pick one specific seed, but calculated the mean percentage of `true' examples in each confusion matrix cell over all 10 seeds.
In the \textit{abortion} topic, 44 \% of `pro' arguments get classified as `against', and only 33\% get correctly classified as `pro'. The \textit{minimum wage} topic shows no discernible pro/against classification confusion, and 60\% of all true `pro' and `against' arguments are correctly classified. The section below analyzes the misclassifications in low-performing topics.
\subsection{Qualitative Analysis of Misclassification}
The low performance of `pro' and `con' in some topics (\textit{abortion}, \textit{gun control}, and \textit{death penalty}) warrants some further investigation. Table \ref{tab:repr_missclass} shows
four example misclassifications between `pro' and `con' by BERT-large in the test examples the model encountered on these topics.\footnote{To ensure we are not cherry-picking examples, we looked at errors that were not unique to just one seed, and identified these examples as salient examples of a general trend.}
\begin{table*}[h!]
\centering
\small{
\begin{tabularx}{\linewidth}{lllll}
\hline
\small{\textbf{Topic}} & \small{\textbf{True}} & \textbf{Pred} &\textbf{Sentence} & \textbf{Frequency in seeds}\\
\hline
\small{gun contr.} & pro & con & "When high-capacity magazines were used in mass shootings, & 9/10 \\ &&& the death rate rose 63 \% \\ &&& and the injury rate rose 156 \% ." \\
\small{gun contr.} & con & pro & "[..] The Second Amendment protects an individual right & 7/10 \\ &&& to possess a firearm unconnected \\ &&& with service in a militia , and to use that arm for traditionally lawful purposes \\ &&& , such as self-defense within the home . "\\
\small{gun contr.} & pro & con & "In this crossfire , bullets would likely hit civilians & 9/10 \\ &&& ( imagine a room filled with a crowd and three people shooting \\ &&& at each other ) and the casualty count would increase." \\
\small{gun contr.} & con & pro & "Gun enthusiasts understand the benefit & 7/10 \\ &&& of large ammo feeders and wish to defend them because \\ &&& they recognize the advantage that such feeders give."\\
\hline
abortion & pro & con & "Not only has the biological development not yet occurred to & 4/10 \\ &&& support pain experience , but the environment after birth , so necessary \\\ &&& to the development of pain experience , is also yet to occur ." \\
abortion & pro & con & "Warren concludes that as the fetus satisfies only one criterion, & 5/10 \\&&& consciousness ( and this only after it becomes susceptible to pain )\\&&& , the fetus is not a person and abortion is therefore morally permissible ."\\
abortion & con & pro & It is argued that just as it would not be permissible to refuse & 2/10 \\ &&& temporary accommodation \\ &&& for the guest to protect him from physical harm , \\ &&& it would not be permissible to refuse temporary accommodation of a fetus .\\
abortion & con & pro & "92 \% of abortions in America are purely elective & 3/10 \\ &&& -- done on healthy women to end the lives \\ &&& of healthy children." \\
\hline
death pen. & con & pro & Mentally ill patients may be put to death .& 2/10 \\
death pen. & con & pro & Evidence shows execution does not act as a deterrent to capital punishment. & 9/10\\
death pen. & pro & con & A system in place for the purpose & 8/10 \\&&& of granting justice can not do so for the surviving victims , \\ &&& unless the murderer himself is put to death . \\
death pen. & con & pro & CON : " ... Since the reinstatement of the modern death pen. , & 9 /10 \\&&& 87 people have been freed from death row because \\&&& they were later proven innocent . \\
\hline
\end{tabularx}
}
\caption{Misclassifications on political topics with considerable `pro' and `con' confusion}
\label{tab:repr_missclass}
\end{table*}
We find two types of misclassifications, each related to topic-specific differences to stance classes. The first type is {\bf misclassification due to the socio-cultural background knowledge and context} of a specific topic's
arguments. The second type is related to a model taking the {\bf stance towards a subcomponent of a topic} and confusing it for the text's overall stance on the topic, e.g.\ statements in the `pro' class mostly expressing views \textit{against} something else related to the argument (unwanted pregnancies, gun violence, innocents dying).
Examples of both issues are arguments centering around ``many innocents (babies, children, mentally ill) will die".
There are 5 variations of this argument in these 3 topics: row 1 and row 3 (\textit{gun control}), row 8 (\textit{abortion}), and rows 9 and 12 (\textit{death penalty}) in Table~\ref{tab:repr_missclass}. Not only is one usage of this argument traditionally connected to the `pro' class of one topic (\textit{gun control}), and the `con' class of another (\textit{abortion}), the implication is: innocents dying is bad. The model seems to lack this world knowledge, and for instance classifies this argument as `pro' \textit{death penalty}.
Another salient example is row 2 of Table~\ref{tab:repr_missclass}. This argument argues \textit{in favor} of gun rights for self-defense, but the model misclassifies this as \textit{against} gun control. The model also fails to connect the second amendment discussion to the \textit{against gun control} class. This is the same mistake made by the LSTM-model in \citet[p.3671]{stab2018argumentext}, showing that BERT appears to not improve over LSTM on the topic-specific nuances here. In other words, it fails to correctly identify the socio-cultural dimensions \cite{du2007stance} of this debate.
\newpage
~\newpage
\newpage
~\newpage
\subsection{SVM and Lexical Features}\label{sec:Analysis}
To analyze which words are used in relation to specific stances and topics, we trained an SVM model with tf-idf features on stance detection on all topics (F1 = .573). For each class pair (`pro' vs `con', `pro' vs `no-argument', etc.), we extracted top-10 features with the highest coefficient for that specific class.
Table~\ref{tab:repr_results_SVM_features} presents the most important features of the topic-agnostic model trained on all topics. Some unigrams
appear meaningful for the class. For instance, in the cross-topic setting, the word ``morality'' is a feature for the `con' class. In contrast, the `no argument' class is often identified with words that appear to have little content-relationship to the class identity: a topic-specific pro-life website (lifenews) or someone's name (`robert').
We also trained within-topic models to find whether there is topic-specific vocabulary related to stance that differs from the topic-agnostic model. Table \ref{tab:repr_results_SVM_features} also presents the 10 most informative features for a model trained on only the abortion topic (F1 = .595). Immediately we see that there is only limited overlap with the lexical features used to decide between `pro' and `con' in a multi-topic scenario. Within only the abortion topic, the `pro' and `con' class are defined by concepts related to the lexical content of this specific discussion: babies, life, and birth. We also see the contrast between `pro' arguments talking about reproduction and the mother, while the `con' arguments mention life, conception, and babies. This lexical feature analysis shows no apparent overlap between the topic-specific features in the abortion model and the topic-independent features in the topic-agnostic model. This might indicate that vocabulary is quite specifically related to topics in stance detection.
\section{Conclusion: Topic Matters} \label{sec:Discussion}
Stance detection is a difficult NLP task. Despite recent advances by pre-trained Transformers, these models have similar issues in a cross-topic setting as earlier models. This paper reproduced
stance detection experiments with pre-trained Transformers by \citet{reimers_classification_2019}, training on seven topics and testing on an eighth topic. We found similar results, but also both class and topic influencing performance. Cross-topic BERT models perform below mean model performance in some topics (\textit{abortion}, \textit{gun control}) on the pro and con classes.
This makes us pause about \citet{reimers_classification_2019}'s main claim: does BERT improve cross-topic stance detection over non-Transformer models? We argue this claim needs an asterisk: this cross-topic approach does not work as well for all topics. Different topics show specific vocabularies and socio-cultural contexts, and especially these specific contexts BERT cannot navigate. BERT models still make similar mistakes on gun control as the LSTM-based models in \citet{stab2018argumentext}.
These findings lead us to two take-aways. Firstly, we hypothesize that models like BERT rely more on topic-specific features for stance detection than topic-independent lexical words related to argumentation. \citet{thorn-jakobsen-etal-2021-spurious} also recently found this, and connected BERT's cross-topic stance detection performance to its focus on spurious topic-specific lexical features ("gun", "criminal") rather than words related to argumentation. They also conclude a fair real-world evaluation of cross-topic stance detection means reporting the worst performing cross-topic pair rather than average performance over topics.
Secondly, we also think it is necessary to analyze the context of topics, and its relation to other debate topics within and outside the dataset. Most topics in stance detection studies are currently U.S. socio-political issues. This goes beyond a limitation of language, such as a focus on English without specifying this \cite{bender2019rule}, since the same socio-cultural topics are not even universally relevant in the English-speaking world (gun control is not a salient discussion in Scotland). Such a focus on topic diversity is also important for use-cases. For diversity of viewpoints in search \cite{draws2021assessing} or news recommendation \cite{reuver-etal-2021-nlp}, stance detection needs to work on many different topics.
\citet{schlangen-20191-natural} states that we need to carefully define specific NLP tasks and capabilities needed to solve them. Modelling cross-topic stance detection in a topic-agnostic manner, while divorcing it from socio-cultural context, might not do justice to stance detection. Future work might focus on the specifics of topics: analyzing similarity between discussions \cite{wei2019modeling}, or modelling required socio-cultural contextual knowledge (`second amendment is related to gun control'). Models able to deal with topic-specific vocabulary and socio-cultural context of debates might improve on the state-of-the-art of cross-topic stance detection.
\newpage
\section*{Acknowledgments}
This research is funded through Open Competition Digitalization Humanities and Social Science grant nr 406.D1.19.073 awarded by the Netherlands Organization of Scientific Research (NWO). Our computing was done through SURF Research Cloud, a national supercomputer infrastructure in the Netherlands also funded by the NWO. We would like to thank dr. Nils Reimers for sending us their paper's data. We would also like to thank the anonymous reviewers, whose very helpful comments improved the paper. All opinions and remaining errors are our own.
|
\section{Introduction}
\label{sec:intro}
Current machine learning techniques typically increase in predictive task performance when trained with more data.
Because of their good predictive ability, convolutional neural networks (CNN) are often favored for many tasks in computer vision and medical imaging analysis.
However, medical imaging datasets often have large dimensionality with a relatively limited number of labeled examples that make training CNNs more difficult~\cite{singh20203d,mri2019}.
For low prevalence diseases, too few annotated cases may be available from a single healthcare institution to train a high performance model.
However, directly sharing private medical data with external institutions poses severe privacy and security risks.
For this reason, federated learning (FL) has become an attractive paradigm to collaboratively develop CNN models while maintaining privacy of each institution's respective data~\cite{DBLP:journals/corr/abs-1911-06270}.
Another considerable challenge of most standard CNN models is their lack of uncertainty estimation and poor calibration~\cite{guo2017calibration}.
While powerful in predictive capability, deep CNNs are not generally considered interpretable or well calibrated, which can lead to higher levels of user distrust or frustration during high-stakes decision-making~\cite{bhatt2021uncertainty}.
Having some way to guarantee statistical confidence of a prediction will be crucial in the adoption of ML tools into medical decision-making applications~\cite{ghassemi2019review}.
Conformal predictions have been proposed as a general framework of providing distribution-free uncertainty in the form of prediction sets for classification and prediction intervals for regression \cite{10.5555/1062391,vovk2012conditional}.
Additionally, prediction sets better match the intuition of clinical decision making by providing a set of possible conditions that may be used to rule in or rule out certain diseases similar to a differential diagnosis \cite{lu2021fair}.
Our proposed method combines a federated training scheme with a adaptive conformal calibration during inference to combine the advantages of valid coverage with the increased performance and data privacy assurances of FL.
To this end, we investigate the following three research questions (RQ):
\begin{enumerate}
\item How do conformal methods (adaptive prediction sets) compare to non-adaptive prediction sets for medical imaging classification tasks?
\item Do a federated quantile estimates provide tighter prediction sets than quantiles calibrated only on the local healthcare institution?
\item What is the relationship between conformal uncertainty (as measured by set size) and epistemic (model) uncertainty (as measured by class entropy)?
\end{enumerate}
\section{Related Work}
\label{sec:related}
Conformal inference was introduced by~\citet{10.5555/1062391} as a general framework to incorporate a rigorous notion of uncertainty into arbitrary machine learning models and have been adapted for prediction tasks including quantile regression, multi-label classification, and image segmentation~\cite{Romano2019ConformalizedQR, cauchois2020knowing, bates-rcps}.
Conformal prediction can be seen as a frequentist approach to confidence estimation that provides marginal coverage bounds without needing to specify a prior distribution.
Other approaches to confidence in CNN include Bayesian approximation and post-hoc probability calibration.
Bayesian neural networks permit confidence estimates using posterior probabilities and model uncertainty but usually impose strict assumptions on the modeling distribution, such as assuming a Gaussian distributed likelihood in Monte Carlo dropout~\cite{mc-dropout}.
Other uncertainty methods are either model specific (such as Mondrian forests~\cite{pmlr-v51-lakshminarayanan16}) or require modification of the loss function during training~\cite{Corbire2019AddressingFP}.
The Softmax probability of CNN have been empirically shown lack proper calibration. Attempts to remedy improve these calibration issues in CNN have resulted in post-processing techniques such as temperature scaling and isotonic calibration~\cite{guo2017calibration}.
FL has received considerable attention for privacy sensitive applications such as healthcare~\cite{digi-fed-learn}.
Previous works attempt to incorporate Bayesian uncertainty or address non-IID distributions with post-calibration in FL context~\cite{pmlr-v97-yurochkin19a, DBLP:journals/corr/abs-2106-05001}.
Another work proposes distributed conformal predictions but does not consider a FL context with CNN for medical imaging applications~\cite{spjuth2018aggregating}.
\section{Methods}
\label{sec:methods}
The conformal framework can be used to produce confidence prediction sets with any machine learning model that outputs a scoring function (e.g. deep neural networks, decision trees, and quantile regression).
For CNN classifiers, this amounts to predicting a set of predictions using the Softmax probability (for multi-class classification) as the conformality score function.
One naive way of forming prediction sets is to simply sort the class scores for a given example and include all classes in the prediction set up to some confidence threshold.
For example, for a three-class prediction task with the Softmax class scores $[0.3, 0.1, 0.2, 0.4]$, our naive method would be output $\{3, 0, 2\}$ for a confidence level of $90\%$ and $\{3, 0\}$ for a confidence level of $70\%$.
Unfortunately, this naive approach does not provide the property of \textit{coverage} --.
simply, that the prediction set for any given example will contain the true class on average at some desired confidence level.
To achieve a formal guarantee of coverage, we can instead reuse the validation data to calibrate our model to achieve valid conformal predictions during model inference.
Formally, let $C(X)$ be a conformal prediction set, which is a subset of the label powerset, $\mathcal{P}$, and let $\alpha \in (0, 1)$ be a pre-chosen confidence level that determines the degree of guaranteed coverage on the joint distribution, $\{X_i, Y_i\}_{i=1}^{N + M}$, where $N$ is the number of examples in the calibration set and $M$ is the number of examples in the test set.
Then, for some classifier that makes $\vert Y \vert$ class predictions using a monotonic scoring function, $S: \mathcal{X} \rightarrow \mathbb{R}^{\vert Y \vert}$, a prediction set can be created:
\begin{equation}
C(X) = \{y \in Y \mid S(X)_y > 1 - \hat{q}\},
\end{equation}
where $S(X)_y$ is the score of the $y$th class and $\hat{q}$ is the score quantile (estimated using the calibration set $\{X_i, Y_i\}_{i=1}^N$) at the $1 - \alpha$ coverage level,
\begin{equation}
\hat{q} = \frac{\lceil(N + 1)(1 - \alpha)\rceil}{N},
\end{equation}
required to guarantee marginal coverage with a small finite sample correction, ~\cite{DBLP:journals/corr/abs-2009-14193}.
If the data is drawn from an identical and exchangeable distribution (a weaker form of IID), then marginal coverage can be defined as satisfying the following property:
\begin{equation}
1 - \alpha \; \leq \; P(Y \; \in \; C(X))
\end{equation}
This bound can be made near-tight through an appeal to symmetry (again assuming exchangeable random variables), which appears in several previous works~\cite{10.1007/3-540-36755-1_29, 10.5555/1062391, lei2017distributionfree, romano2020classification}.
\begin{equation}
1 - \alpha \; \leq \; P(Y \; \in \; C(X)) \; \leq \; 1 - \alpha + \frac{1}{M + N + 1}
\end{equation}
\begin{algorithm}[tb]
\small
\caption{Federated conformal prediction sets}
\label{alg:fcps}
\textbf{Input}: \\
Federated model $S: X \times Y \rightarrow \mathbb{R}^{Y}$, \\
$K$ validation sets $\{(X, Y)\}_{n=1}^N$, \\
Test example $X_m$, \\
Confidence level $1 - \alpha \in (0, 1)$, \\
Quantile function $Q: \{\mathbb{R}\}_{i=1}^N \times [0, 1] \rightarrow \mathbb{R}$\\
Sorting function $\pi: \mathcal{S} \rightarrow \mathcal{S}$\\
\textbf{Output}: A confidence prediction set $C(X_m)$ \\
\begin{algorithmic}[1]
\STATE $\hat{q} \leftarrow 0$
\FOR{$k \in \{1, 2, \ldots, K\}$}
\FOR{$i \in \{1, 2, \ldots, N\}$}
\STATE $s_{i} \leftarrow \sum_{j=1}^{y_j=y_i} \pi(S(x_i, y_i)$
\ENDFOR
\STATE $\hat{q} \leftarrow \hat{q} + Q(\{s_{i}\}^N_{i=1}, \lceil (1 - \alpha) (N + 1) \rceil)$
\ENDFOR \
\STATE \textbf{return }: $\{y \in Y : S(X_m)_y > \hat{q} / K \}$
\end{algorithmic}
\end{algorithm}
Because this marginal coverage property does not impose any assumptions on the distribution (besides exchangeability) or score function (besides monotonicity), conformal prediction methods can be easily adapted for FL settings.
Federated averaging (FedAvg) is a simple FL procedure that averages parameter weights of all local model (client side) to update the global model (server side) -- which is then distributed back to each local model -- after each training epoch~\cite{mcmahan2017communicationefficient}.
As described in Algorithm \ref{alg:fcps}, we adapt split conformal predictions~\cite{romano2020classification} for FedAvg
While our conformal framework does not depend on a specific FL technique, we implement FedAvg for its simplicity in order to demonstrate a proof of concept.
For more information on conformal predictions and distribution-free uncertainty, we refer the reader to the excellent tutorial by \citet{angelopoulos-gentle}.
\section{Experiments}
\label{sec:experiments}
We compare three methods of forming prediction sets: a naive baseline, conformal predictions using local quantiles, and conformal predictions using federated quantiles.
The naive baseline simply form predictions sets by adding sorting classes by predicted scores and adding elements to the set until the $1 - \alpha$ threshold is met.
Because the naive baseline does not estimate the empirical quantile on a held-out calibrate set, we cannot expect it to satisfy the coverage guarantee described above.
For our experiments, we train a ResNet-18~\cite{He2016DeepRL} for 5 different random weight initializations to optimize a cross entropy loss on six datasets in the MedMNIST medical imaging dataset benchmark~\cite{medmnistv1} (BloodMNIST, DermaMNIST, PathMNIST, TissueMNIST, RetinaMNIST, OrganMNIST3D).
The number of classes and overall class distribution for each dataset is shown in Figure \ref{fig:distribution}.
For each dataset, we split the original training, validation, and testing sets into four partitions to simulate different healthcare institutions to train our FL models.
As conformal predictions are a ``post-proccessing'' step that do not depend on model training, the same trained models where used for all three methods of forming prediction sets.
For local conformal prediction sets, we estimate the quantile using only the first institution's validation set as the calibration set.
For federated conformal prediction sets, we calibrate the quantile using a simple average of quantiles from each individual institution's respective validation set.
To further assess the performance of federated conformal and to tease out possible confounding of performance with larger calibration dataset sizes, we randomly shuffle 30\% of labels for three out of the four institutions' validation set while keeping the first institution's validation set ``clean''.
All three methods were evaluated only on the first institution's test set.
Images with the corresponding conformal prediction set is shown in Figure \ref{fig:predictions}.
\begin{figure}[htb]
\centering
\includegraphics[width=\linewidth]{predictions.png}
\caption{Image and corresponding prediction set at 90\% confidence level ( larger sets indicate higher uncertainty in the prediction task while smaller sets indicate more confidence in the prediction).}
\label{fig:predictions}
\end{figure}
\begin{figure}[htb]
\centering
\includegraphics[width=\linewidth]{class_distribution.png}
\caption{Varying class distribution of six MedMNIST benchmark datasets (see Figure \ref{fig:predictions} for an image from each dataset).}
\label{fig:distribution}
\end{figure}
\begin{figure}[htb]
\centering
\includegraphics[width=\linewidth]{4-coverage.png}
\caption{Coverage of naive, local conformal, and federated conformal methods by different levels of confidence ($1-\alpha$). Dashed line shows idealized coverage.}
\label{fig:coverage}
\end{figure}
\begin{figure}[htb]
\centering
\includegraphics[width=\linewidth]{4-cardinality.png}
\caption{Average size of prediction sets of naive, local conformal, and federated conformal methods by different levels of confidence ($1-\alpha$).}
\label{fig:cardinality}
\end{figure}
Additionally, we wish to study the relationship between the difficulty of predicting a specific case (independent to the number of classes) with its corresponding uncertainty estimate (as determined by the number of elements in the prediction set).
To this end, we hold out 10\% of the training set from each dataset to train an auxiliary classifier, which can then be used to correlate the maximum Softmax score with the class-wise entropy, $H(X_i) = -\sum_{y \in Y} S(X_i)_y \cdot \log S(X_i)_y$, for the rest of the dataset.
Therefore, we expect entropy to be indicative of model uncertainty and positively correlated with the size of a conformal prediction set.
\begin{table}[htb]
\tiny
\centering
\begin{tabular}{|c|c|c|c|c|c|}
\hline
\textbf{Dataset} & \textbf{Classes} & \textbf{Examples} & \textbf{Method} & \textbf{Coverage} & \textbf{Cardinality}\\
\hline
\multirow{3}*{Blood} & \multirow{3}*{9} & \multirow{3}*{17,092}
& Naive & 100\% $\pm$ 0\% & 8 $\pm$ 0 \\
& & & Local & 100\% $\pm$ 0\% & 5.4 $\pm$ 0.3 \\
& & & Federated & 97.7\% $\pm$ 0.7\% & 2.3 $\pm$ 0.1 \\
\hline
\multirow{3}*{Derma} & \multirow{3}*{7} & \multirow{3}*{10,015}
& Naive & 100\% $\pm$ 0\% & 7 $\pm$ 0 \\
& & & Local & 98.2\% $\pm$ 0.5\% & 4.5 $\pm$ 0.2 \\
& & & Federated & 97.1\% $\pm$ 0.4\% & 3.6 $\pm$ 0.1 \\
\hline
\multirow{3}*{Path} & \multirow{3}*{9} & \multirow{3}*{107,180}
& Naive & 100\% $\pm$ 0\% & 8 $\pm$ 0 \\
& & & Local & 99.1\% $\pm$ 0.2\% & 5 $\pm$ 0.2 \\
& & & Federated & 94.2\% $\pm$ 0.8\% & 2 $\pm$ 0.1 \\
\hline
\multirow{3}*{Tissue} & \multirow{3}*{8} & \multirow{3}*{236,386}
& Naive & 100\% $\pm$ 0\% & 8 $\pm$ 0 \\
& & & Local & 96.8\% $\pm$ 0.1\% & 5.3 $\pm$ 0.1 \\
& & & Federated & 93.6\% $\pm$ 0.4\% & 4.5 $\pm$ 0.1 \\
\hline
\multirow{3}*{Retina} & \multirow{3}*{5} & \multirow{3}*{1,600}
& Naive & 100\% $\pm$ 0\% & 5 $\pm$ 0 \\
& & & Local & 91\% $\pm$ 9.8\% & 4.3 $\pm$ 0.7 \\
& & & Federated & 95.6\% $\pm$ 6.9\% & 4.7 $\pm$ 0.6 \\
\hline
\multirow{3}*{Organ3d} & \multirow{3}*{11} & \multirow{3}*{1,743}
& Naive & 100\% $\pm$ 0\% & 11 $\pm$ 0 \\
& & & Local & 89.5\% $\pm$ 4.5\% & 5.6 $\pm$ 0.3 \\
& & & Federated & 90.5\% $\pm$ 3.5\% & 6 $\pm$ 0.6 \\
\hline
\end{tabular}
\caption{Summary of experimental results for each dataset at setting of 90\% coverage ($\alpha=0.1$).}
\label{table:summary}
\end{table}
\begin{figure}[htb]
\centering
\includegraphics[width=\linewidth]{violin.png}
\caption{(Left) figure shows average test coverage at $\alpha=0.9$ of local conformal and federated conformal methods across six MedMNIST classification tasks; (right) figure shows federated prediction sets having lower average cardinality than local prediction sets at $p<0.001$ significance for BloodMNIST, DermaMNIST, PathMNIST, and TissueMNIST datasets}
\label{fig:violin}
\end{figure}
\begin{figure}[htb]
\centering
\includegraphics[width=\linewidth]{correlation.png}
\caption{Correlation of local conformal set size and entropy percentile for each of the four institutions (1 clean; 3 trained with 30\% random labels).}
\label{fig:correlation}
\end{figure}
\section{Discussion}
\label{sec:discussion}
We can evaluate and compare different methods of predictions sets by their cardinality (average number of elements in a prediction set) and coverage (probability that the true class is an element of the prediction set).
Ideally, we would also like prediction sets to be ``adaptive'' to communicate model uncertainty at an instance level (difficult cases should have larger confidence sets while easier cases should have fewer elements in their prediction sets) while still satisfying marginal coverage guarantees.
From the results of our experiments, which are summarized in Table \ref{table:summary}, we can immediately see the naive baseline prediction sets as exceptionally conservative in coverage across all six datasets.
In order words, naively forming predictions sets using raw Softmax ``probabilities'' simply outputs all possible classes for each prediction for 100\% coverage but useless for any kind of prediction task.
Our results validate previous studies on overconfident Softmax scores in deep neural networks~\cite{guo2017calibration, hendrycks2018baseline}.
While post-processing calibration techniques may produce more useful confidence sets, they do not offer the statistical guarantees provided by conformal prediction methods.
Next, we compare the two conformal prediction methods: prediction sets calibrated with a federated quantile and prediction sets calibrated with a local quantile.
As shown in Table \ref{table:summary} and \ref{fig:coverage}, while all methods satisfy coverage properties (for example at a $\alpha$ threshold of 90\%), both conformal methods have much tighter confidence sets in terms of empirical cardinality compared to the naive baseline (see Figure \ref{fig:cardinality}).
Additionally, federated conformal predictions have lower cardinality than local conformal predictions across MedMNIST datasets for at most $\alpha$ levels (see Figure \ref{fig:cardinality}) -- even when label noise was introduced to 30\% of the validation set used to estimate the federated conformal prediction quantile.
Notably, our federated method performs remarkably well on both the BloodMNIST and PathMNIST datasets, which both have mostly equal distribution of classes (see Figure \ref{fig:distribution}), achieving average prediction set sizes of $2.3 \pm 0.1$ and $2 \pm 0.1$, respectively, and having a much lower rate of set size with at lower $\alpha$ thresholds compared to local conformal prediction sets (see \ref{fig:cardinality}).
However, on both RetinaMNIST and OrganMNIST3d datasets, prediction sets formed using federated quantile estimate show a similar average set size as prediction sets formed with a local quantile estimate; however, both dataset also have (between 10-200 times) fewer cases than the other datasets.
As shown in Figure \ref{fig:violin}, both these smaller datasets have much larger variance bands for both coverage and prediction set size.
Therefore, we conclude that conformal methods outperform the naive baseline of forming prediction sets (RQ1) and that federated conformal prediction sets have smaller set sizes than local conformal prediction sets (RQ2) for datasets with a sufficient number of examples.
To answer RQ3, we plot each of the four institutions at different three different class entropy percentiles (50\%, 75\%, and 90\%) each dataset (see Figure \ref{fig:correlation}).
We confirm a positive relationship between between cardinality and epistemic model uncertainty, and also observe that conformal uncertainty to be robust to moderate levels of label corruption in the calibration set used to estimate the empirical quantile.
\section{Conclusion}
\label{sec:conclusion}
We proposed a novel approach to combine FL with conformal prediction sets and empirically evaluated our method on several benchmark medical imaging datasets (MedMNIST) to demonstrate feasibility of conformal uncertainty in healthcare.
Additionally, we find a intuitive correspondence between conformal uncertainty and prediction task difficulty as measured by predictive class entropy.
Combining the benefits of federated training with conformal inference, our distribution-free method of uncertainty quantification is a simple and flexible technique that requires no direct modification to model training while retaining wide applicability for collaboratively developing FL models among healthcare institutions.
|
\section{Introduction and Summary}
We begin with an informal overview, detailed definitions and statements are in
the later sections.
Let $\mathcal{F}$ be the collection for finite dimensional reproducing kernel
Hilbert spaces (RKHS) with irreducible complete Pick kernels. Those are
exactly the spaces which are rescalings of spaces generated by finite sets of
Dirichlet-Arveson reproducing kernels; the exact statement is Definition
\ref{pick} below.
We interested in characterizing those $H\in\mathcal{F}$ that arise from finite
dimensional simple model spaces. That is, let $H^{2}$ be the the classical
Hardy space and $B\in H^{2}$ a finite Blaschke product with simple zeros at
the points $X=\left\{ x_{i}\right\} _{i=1}^{n}$ in the complex unit disk
$\mathbb{B}^{1}.$ Let $K_{B}$ be the finite dimensional Hilbert spac
\[
K_{B}=H^{2}\ominus BH^{2}.
\]
We regard $K_{B}$ as an RKHS by declaring the functionals of evaluation at
points of $X$ to be the reproducing kernels; that is, the kernels are the
$k_{i}\in K_{B},$ $i=1,...,n,$ which satisfy $\left\langle f,k_{i
\right\rangle =f(x_{i})$ for all $f\in K_{B}.$ We call those RKHS model spaces
and call their rescalings $r-$model spaces. We denote the collection of all
$r-$model spaces by $\mathcal{M}$.
The $r-$model spaces are among the most easily described and thoroughly
studied elements of $\mathcal{F}$ \cite{GMR} and so it is interesting to know
how typical they are as elements of $\mathcal{F}$ and what distinguishes them
from the general elements of $\mathcal{F}$. Here we develop several criteria
that characterize the $H\in\mathcal{F}$ that are in $\mathcal{M}$, including
criteria based on the form of the associated set in complex hyperbolic space,
on the values of certain extremal multipliers, on the existence of a
conjugation operator taking the reproducing kernels to their dual basis, and
on having a Gram matrix that is an orthogonal matrix.
The next section has preliminary notation, definitions and results. In Section
3 we state our main results, the proofs are in Section 4. A final section
contains some comments.
\section{Background, Definitions, Notation}
\subsection{RKHS}
For general background about spaces in $\mathcal{F}$ we refer to \cite{AM} and
\cite{ARSW2}.
A finite dimensional RKHS is a finite dimensional Hilbert space $H$ together
with a designated basis $\mathfrak{K}=\mathfrak{K}(H)=\left\{ k_{i}\right\}
_{i=1}^{n}\subset H$ of vectors called reproducing kernels. We denote the dual
basis by $\mathfrak{K}^{\#}=\{k_{i}^{\#}\}_{i=1}^{n};$ that is, $k_{j}^{\#}\in
H$ and $\left\langle k_{i},k_{j}^{\#}\right\rangle =\delta_{ij}$ $1\leq
i,j\leq n.$ Notice that the Hilbert space $H$ with vectors of $\mathfrak{K
^{\#}$ as its set of reproducing kernels is also a RKHS, we denote it by
$H^{\#}.$ Let $\mathbf{K}=\mathbf{K}(H)$ be the Gram matrix of $\mathfrak{K,}$
the $n\times n$ matrix $\left( k_{ij}\right) $ with $k_{ij}=\left\langle
k_{i},k_{j}\right\rangle $ $1\leq i,j\leq n.$ Similarly let $\mathbf{K}^{\#}$
be the Gram matrix of $\mathfrak{K}^{\#}\mathfrak{.}$ It is not hard to check
that $\mathbf{K}^{\#}=\mathbf{K}^{-1}.$
We will use the metric $\delta=\delta_{H}$ defined on $\mathfrak{K}$ or,
equivalently, the index set of $\mathfrak{K.}$ For each $i$ let $P_{i}$ be the
orthogonal projection onto the span of the kernel function $k_{i}$ and let
$\left\Vert \cdot\right\Vert $ denote the operator norm. We set $\delta
(k_{i},k_{j})=\delta(i,j)=\left\Vert P_{i}-P_{j}\right\Vert .$\ There is also
a formula for $\delta$ in terms of Gram matrix entries;
\begin{equation}
\delta(k_{i},k_{j})=\sqrt{1-\frac{\left\vert k_{ij}\right\vert ^{2}
{k_{ii}k_{jj}}}. \label{distance
\end{equation}
More information about $\delta$ is in \cite{ARSW1} and \cite{R1}. {}
Two RKHS, $H$ and $\tilde{H},$ with reproducing kernels $\mathfrak{K}=\left\{
k_{i}\right\} _{i=1}^{n}$ and $\mathfrak{\tilde{K}}=\{\tilde{k}_{i
\}_{i=1}^{n}$ are said to be rescalings of each other if there are nonzero
scalars $\left\{ \lambda_{i}\right\} _{i=1}^{n}$ so that for $1\leq i,j\leq
n$ $\left\langle k_{i},k_{j}\right\rangle =\lambda_{i}\overline{\lambda_{j
}\left\langle \tilde{k}_{i},\tilde{k}_{j}\right\rangle .$ We denote this
equivalence relation by $H\sim$ $\tilde{H}$. Equivalently the Gram matrices of
the two spaces are related through conjugation by a diagonal matrix with
nonzero entries in which case we say that that the two matrices are rescalings
of each other. Most of the conditions we consider interact well with this
equivalence relation; for instance if $H\sim$ $\tilde{H}$ then $\delta
_{H}(k_{i},k_{j})=\delta_{\tilde{H}}(\tilde{k}_{i},\tilde{k}_{j})$ and
$H^{\#}\sim$ $(\tilde{H})^{\#}$
\subsection{\textbf{Model Spaces }}
The main facts we use about model spaces are in \cite{GP}; a general reference
is \cite{GMR}.
Suppose $B$ is a Blaschke product with simple zeros at the points $\left\{
z_{i}\right\} _{i=1}^{n}\subset\mathbb{B}^{1}.$ Thus $B=\Pi_{i}B_{i}$ where,
for $x_{i}\neq0$
\begin{equation}
B_{i}(z)=\frac{\left\vert x_{i}\right\vert }{x_{i}}\frac{x_{i}-z
{1-\overline{x_{i}}z} \label{factor
\end{equation}
and $B_{i}(z)=z$ if $x_{i}=0.$ The reproducing kernels for the space $K_{B}$
are the functions $\left\{ k_{i}\right\} \subset K_{B}$ defined b
\[
k_{i}(w)=\frac{1}{1-\left\langle w,x_{i}\right\rangle }.
\]
To see this recall from Hardy space theory that for each $i$ and each $f\in
H^{2}$ we have $\left\langle f,k_{i}\right\rangle =f(x_{i}).$ With this in
hand it follows that $k_{i}\in\left( BH^{2}\right) ^{\perp}=K_{B}.$ These
two facts together insure that the $k_{i}$ are the reproducing kernels. It
then follows with a bit of computation tha
\begin{equation}
\delta(k_{x_{i}},k_{x_{j}})=\delta(x_{i},x_{j})=\left\vert \frac{x_{i}-x_{j
}{1-\overline{x_{i}}x_{j}}\right\vert =\left\vert B_{i}(x_{j})\right\vert
=\left\vert B_{j}(x_{i})\right\vert . \label{disk psh
\end{equation}
A conjugation operator on a Hilbert space $H$ is an isometric conjugate linear
map $J$ of $H$ to itself that is an involutive automorphism; that is
$\left\Vert Jh\right\Vert =\left\Vert h\right\Vert $ for all $h\in H$ and
$J^{2}=I.$ Each $K_{B}$ carries a conjugation operator $J=J_{B}$ given b
\begin{equation}
J_{B}f=B\overline{zf}. \label{j
\end{equation}
In (\ref{j}) functions in $H^{2}$ are identified with their boundary values
and (\ref{j}) is an equation involving functions on $\mathbb{T}^{1},$ the
boundary circle of $\mathbb{B}^{1}.$ Using (\ref{j}) and the fact that each of
the factors $B_{i}$ of (\ref{factor}) is unimodular on $\mathbb{T}^{1}$ it is
not hard to check that the action of $J_{B}$ is on the kernel functions is
given b
\[
Jk_{i}(x_{j})=\left\{
\begin{array}
[c]{lll
B(x_{j})/(x_{j}-x_{i})\text{ } & \text{if } & i\neq j\\
B^{\prime}(x_{i}) & \text{if } & i=j
\end{array}
\right. ,
\]
\cite{GP}, \cite{GMR}. Hence for $1\leq i,j\leq n$
\begin{equation}
\left\langle k_{i},Jk_{j}\right\rangle =B^{\prime}(x_{i})\delta_{i,j}.
\label{almost conjugation
\end{equation}
Because $B$ has only simple zeros none of the $B^{\prime}(x_{i})$ are zero$.$
Thus it is almost true that $J$ maps the basis of reproducing kernels of
$K_{B}$ to its dual basis. By rescaling we can make that exactly true. We
describe the general pattern.
\begin{proposition}
\label{exact conjugation}Suppose $H$ is a RKHS with kernel functions $\left\{
k_{i}\right\} _{i=1}^{n}.$ Suppose $J$ is a conjugation operator on $H$ and
that for nonzero $\left\{ c_{i}\right\} $ we have
\[
\left\langle k_{i},Jk_{j}\right\rangle =c_{i}\delta_{ij}\text{ \ \ }1\leq
i,j\leq n.
\]
Let $\tilde{H}$ be the RKHS obtained by using the same Hilbert space as $H$
and the reproducing kernels $\{\tilde{k}_{i}\}$ defined by $\tilde{k
_{i}=c_{i}^{-1/2}k_{i},$ $1\leq i\leq n.$ Then $\tilde{H}$ $\sim$ $H$ and $J$
is a conjugation operator on $\tilde{H}$ which maps the reproducing kernels of
$\tilde{H}$ to their dual basis; that is
\[
\left\langle \tilde{k}_{i},J\tilde{k}_{j}\right\rangle =\delta_{ij}\text{
\ \ }1\leq i,j\leq n.
\]
\begin{proof}
The first two statements are clear. To check the third we compute, for $1\leq
i,j\leq n
\begin{align*}
\left\langle \tilde{k}_{i},J\tilde{k}_{j}\right\rangle & =\left\langle
c_{i}^{-1/2}k_{i},Jc_{j}^{-1/2}k_{j}\right\rangle =\left\langle c_{i
^{-1/2}k_{i},\overline{c_{j\text{ }}}^{-1/2}Jk_{j}\right\rangle \\
& =c_{i}^{-1/2}c_{j}^{-1/2}\left\langle k_{i},Jk_{j}\right\rangle
=c_{i}^{-1/2}c_{j}^{-1/2}c_{i}\delta_{ij}=\delta_{ij}.
\end{align*}
\end{proof}
\end{proposition}
We will call such an $\tilde{H}$ orthogonal and say its rescaling $H$ is
$r-$orthogonal. The name is based on the following simple proposition. Recall
that a matrix is called orthogonal if its inverse equals its transpose.
\begin{proposition}
[{\cite[Prop. 8]{R1}}]\label{conjugation = orthogonal}$H,$ a finite
dimensional RKHS, is orthogonal if and only if its Gram matrix is an
orthogonal matrix.
\end{proposition}
Suppose $H$ is a finite dimensional RKHS with kernel functions $\left\{
k_{i}\right\} $ and Gram matrix $\mathbf{K}=\left( k_{ij}\right) $. Let
$\{k_{i}^{\#}\}$ be the dual basis and hence the set of kernel functions of
$H^{\#}$ and write its Gram matrix as $\mathbf{K}^{\#}=(k_{ij}^{\#}).$ It is
always true that the transpose of $\mathbf{K,}$ $\mathbf{K}^{t}$ is equal to
the entrywise conjugate matrix $\mathbf{\bar{K}}$. It is also always true that
$\mathbf{K}^{\#}=\mathbf{K}^{-1}.$ Thus $H$ is orthogonal exactly if
$\mathbf{KK}^{t}=\mathbf{K\bar{K}}=\mathbf{KK}^{\#}=I.$ Formulated in terms of
matrix entries this is, for $1\leq i,j\leq n,$
\begin{equation}
\sum_{s=1}^{n}k_{is}k_{js}=\sum_{s=1}^{n}k_{is}\overline{k_{sj}}=\sum
_{s=1}^{n}k_{is}k_{sj}^{\#}=\delta_{ij}. \label{orthogonal matrix
\end{equation}
\begin{corollary}
\label{is also}If $H\in\mathcal{M}$ then $H^{\#}\in\mathcal{M}$. Specifically,
if $H$ $\sim$ $K_{B}$ where $B$ is the Blaschke product with zeros $\left\{
x_{i}\right\} $ then $H^{\#}$ $\sim$ $K_{B^{\#}}$ where $B^{\#}$ is the
Blaschke product with zeros $\left\{ \overline{x_{i}}\right\} .$
\end{corollary}
\begin{proof}
The Gram matrix of $H^{\#}$ is the inverse matrix of the Gram matrix of $H$
and hence is a rescaling of the inverse of $\mathbf{K}_{B}$, the Gram matrix
of $K_{B}.$ From the previous proposition $K_{B}$ is $r-$orthogonal hence
$\mathbf{K}_{B}$ is a rescaling of an orthogonal matrix and thus $\left(
\mathbf{K}_{B}\right) ^{-1}$ is a rescaling of $\overline{\mathbf{K}_{B}}$,
the matrix of complex conjugates of entries of $\mathbf{K}_{B}$. It is a
direct consequence of the definitions that $\overline{\mathbf{K}_{B}}$
$=\mathbf{K}_{B^{\#}}$. In sum, the Gram matrix of $H^{\#}$ is a rescaling of
the Gram matrix of $K_{B^{\#}}.$ Hence $H^{\#}$ $\sim$ $K_{B^{\#}}.$
\end{proof}
This corollary shows that the Gram matrix of $K_{B}$ is a rescalings of the
complex conjugate of the Gram matrix of $K_{B}^{\#}.$ A similar result holds
for some infinite Blaschke products $B$ and that fact has been used in the
study of $H^{2}$ interpolating sequences \cite[Exercise 9.54]{AM} \cite[Sec.
7.3.1]{ARSW2}.
Because $H^{\#\#}=H$ the converse of the corollary also holds and thus we have
an necessary and sufficient condition for $H^{\#}\in\mathcal{M}$. On the other
hand it is not clear what conditions insure $H^{\#}\in\mathcal{F}$. We discuss
that question briefly in Section\ \ref{C}.
\subsection{Multiplier Algebras}
If $H\in\mathcal{F}$ then there is a particularly close relationship between
$H$ and $M(H).$ In particular suppose $X$ is the index set of $\mathfrak{K
(H),$ $Y\subset X$ and $x\in X\smallsetminus Y.$ Let $m(Y,x)\in M(H)$ be the
multiplier of norm one which vanishes on $Y$ and maximizes $\operatorname{Re
m(Y,x).$ Let $h(Y,x)$ be the vector in $H$ of norm one which on $Y$ and
maximizes $\operatorname{Re}h(Y,x).$
\begin{proposition}
[{\cite[Prop. 6.27]{ARSW2}}]\label{extreme}In the situation just described and
with $k_{x}$ the reproducing kernel for $x$
\begin{equation}
m(Y,x)\frac{k_{x}}{\left\Vert k_{x}\right\Vert }=h(Y,x). \label{extremal
\end{equation}
\end{proposition}
With this as a starting point it is not hard to show that for $H\in
\mathcal{F}$ the metric $\delta$ on $X,$ the index set of $\mathfrak{K}(H),$
is the same as the Gleason metric on $X$ induced by $M(H),$ that i
\begin{equation}
\delta(x,y)=\max\left\{ \operatorname{Re}m(x):m\in M(H),m(y)=0,\left\Vert
m\right\Vert =1\right\} , \label{gleason
\end{equation}
\cite[Remrk 7.2]{ARSW2}. The equality of the term on the left (\ref{disk psh})
and the last two terms on the right is an instance of this general fact.
\subsection{Drury Arveson Spaces}
For $m=1,2,....$ the Drury-Arveson space $DA_{m}$ is the Hilbert space of
functions defined on the complex $m-$ball $\mathbb{B}^{m}$ which is the
closure of the span of the reproducing kernels $\left\{ k_{z}(w)=\frac
{1}{1-\left\langle w,z\right\rangle }:z\in\mathbb{B}^{m}\right\} .$ Here
$\left\langle w,z\right\rangle $ is the standard Hermitian inner product on
$\mathbb{C}^{n}.$ References for this space and its properties include
\cite{AM}, \cite{ARSW2}, and \cite{Sh}.
For $X$ a finite subset of some $\mathbb{B}^{m}$ we define the associated
space $DA_{m}(X)$ to be the subspace of $DA_{m}$ spanned by $\left\{
k_{x}:x\in X\right\} $ and having those functions as kernel functions. In
particular if $m=1$ then the kernel functions of $DA_{1}(X)$ are the same as
those of the model space $K_{B_{X}}$ for $B_{X}$ the Blaschke product with
zeros at the points of $X;$ thus those two spaces are the same (or,
pedantically, are trivial rescalings of each other). When $m>1$ then the
details of our discussion are effectively independent of $m$ and we will not
keep track of that index; this is discussed in \cite{R1}.
\begin{definition}
\label{pick}We say $H\in\mathcal{F}$ if there is an integer $m$ and finite
$X(H)\subset\mathbb{B}^{m}$ such that $H\sim$ $DA_{m}(X(H)).$
\end{definition}
Although this definition suits our purposes it is very different from the
traditional definition of finite dimensional RKHS with complete Pick kernels.
The equivalence of our definition with the traditional one is a basic theorem
in the subject \cite{AM}.
\subsection{Hyperbolic Geometry}
The unit ball $\mathbb{B}^{m}\subset\mathbb{C}^{m}$ is a model for complex
hyperbolic $m-$space $\mathbb{CH}^{m}$ . This is discussed in detail in
\cite{Go}; here we will just recall a few pieces of information that we need.
The space $\mathbb{CH}^{m}$ carries a transitive set of orientation preserving
automorphisms. In the ball model these are realized by the group of conformal
automorphisms of $\mathbb{B}^{m}.$
The space $\mathbb{CH}^{m}$ carries several natural metrics, of particular
interest to us is the pseudohyperbolic metric $\Delta$ which can be defined by
setting, for $z\in\mathbb{B}^{m},$ $\Delta(0,z)=\left\vert z\right\vert $ and
requiring $\Delta$ to be invariant under the automorphism group. (We note for
context, but will not use the fact, that the length metric generated by
$\Delta$ is the classical Bergman metric on the ball.) In particular, if
$z,w\in\mathbb{B}^{1}$ the
\begin{equation}
\Delta(z,w)=\left\vert \frac{z-w}{1-\bar{z}w}\right\vert . \label{delta
\end{equation}
The disk $\mathbb{B}^{1}$ sits inside $\mathbb{B}^{m}=\mathbb{CH}^{m}$ as a
totally geodesically embedded manifold of complex dimension one, called a
complex geodesic. All the other totally geodesically embedded complex one
manifolds, the other complex geodesics, are the images of that unit disk under
the group of conformal automorphisms of the ball. In particular the group of
automorphisms acts transitively on the set of complex geodesics.
\subsection{Hyperbolic Geometry and $\mathcal{F}$}
The structure of the spaces $DA_{m}(X)$ is closely related to the geometry of
$X$ regarded as a subset of $\mathbb{CH}^{n}.$ This theme is developed in
\cite{ARSW1}, \cite{R1}, and \cite{R2}.
Comparing (\ref{delta}) with (\ref{disk psh}) we see that if $B$ is a Blaschke
product which vanishes at $z$ and $w$ and possibly other points, and with
$\delta$ denoting the metric on $\mathfrak{K}(B)$ then we have $\Delta
(z,w)=\delta(z,w).$ In fact this is the general pattern.{}
\begin{proposition}
[{\cite[4.1]{R1}}]Suppose $X$ is a finite set in some $\mathbb{B}^{m}$ with
$z,w\in X.$ Let $\delta$ be the metric on $X$ which is the index set of
$\mathfrak{K}(DA_{m}(X))$ and let $\Delta$ be the pseudohyperbolic metric on
$\mathbb{CH}^{m}$ restricted to $X,$ then $\Delta(z,w)=\delta(z,w).$
\end{proposition}
There is another fundamental relation between elements of $\mathcal{F}$ and
hyperbolic geometry. Given $H\in\mathcal{F}$ we know from Definition
\ref{pick} that $H\sim$ $DA_{m}(X)$ for some finite $X.$ In fact we can
describe the possible choices for $X.$
\begin{theorem}
[{\cite[Thm 7]{R1}}]\label{R1}Suppose $X,Y$ are finite sets in some
$\mathbb{CH}^{m}.$ $DA(X)\sim$ $DA(Y)$ if and only if there is an automorphism
$\Phi$ of $\mathbb{CH}^{m}$ with $\Phi(X)=Y.$
\end{theorem}
In sum, questions of equivalence of elements of $\mathcal{F}$ under rescaling
are equivalent to questions about congruence of finite sets in complex
hyperbolic space.
\subsection{Subspaces\label{subspaces}}
Suppose $H$ is a finite dimensional RKHS with kernel functions $\left\{
k_{\alpha}\right\} _{\alpha\in X}.$ We say $J$ is a regular subspace of $H$
if there is a $Y\subset X$ and $J$ is the Hilbert space spanned by $\left\{
k_{\beta}\right\} _{\beta\in Y}$ and is regarded as a RKHS with the $\left\{
k_{\beta}\right\} _{\beta\in Y}$ as its its kernel functions. It follows from
the definitions that if $H\in\mathcal{F}$ then $J\in\mathcal{F}$, and
similarly for membership in $\mathcal{M}$.
Statements which can be formulated using Gram matrix entries can be passed
between a space $H$ and a regular subspace $J$. For instance, given $H,J,X,Y$
as above and $y,y^{\prime}\in Y$ then the value $\delta(k_{y},k_{y^{\prime}})$
does not depend on whether we regard the kernel functions as in $J$ or in $H.$
Another instance of this, one we use in the proof below in showing Statement
(2) implies Statement (1), is that if $J$ is a regular subspace of a $DA(X),$
and hence is $DA(Y)$ for some $Y\subset X$ then knowing that three points of
$Y$ are in a complex geodesic implies that the "same" points, regarded as
inside of $X,$ and hence as index points for kernel functions in $DA(X)$, are
also in a complex geodesic.
Another fact about regular subspaces which we will use is that if
$H\in\mathcal{F}$ and $J$ is a regular subspace of $H$ then it is a
consequence of the Pick property of $H$ than given a multiplier $m_{J}\in
M(J)$ there is an extension to a multiplier $m_{H}\in M(H)$, defined on all of
$H,$ with the same norm, $\left\Vert m_{J}\right\Vert =\left\Vert
m_{h}\right\Vert .$
At times we will use these facts without mention.
\section{The Results}
Our main result is the following
\begin{theorem}
\label{A} Suppose $H\in\mathcal{F}$, that is for some $m,$ and $X=\left\{
x_{i}\right\} _{i=1}^{n}\subset\mathbb{B}^{m},$ we have $H\sim$ $DA_{m}(X)$.
The following are equivalent:
\begin{enumerate}
\item $X$ lies in a single complex geodesic in $\mathbb{B}^{m}=\mathbb{CH
^{m}.$
\item For each $1<i<j\leq n$, $\left\{ x_{1},x_{i},x_{j}\right\} $ lies in a
single complex geodesic in $\mathbb{B}^{m}=\mathbb{CH}^{m}.$
\item There is a renumbering of $X$ after which, with $m_{1}\in M(H)$ the
multiplier of norm one which satisfies $m_{1}(x_{j})=0,$ $j=2,...,n$ and which
maximizes $\operatorname{Re}m_{1}(x_{1})\ $we have
\begin{equation}
m_{1}(x_{1})=\prod_{j=2}^{n}\delta(x_{1},x_{j}). \label{multiplier = product
\end{equation}
\item $H$ is $r-$orthogonal. That is $H\sim$ $\tilde{H}$ for some $\tilde{H}$
which carries a conjugation operator taking the basis of kernels of $H$ to the
dual basis.
\item $H\sim$ $\tilde{H}$ \ for an $\tilde{H}$ which has a Gram matrix which
is an orthogonal matrix.
\item $H$ is an $r-$model space; that is there is a finite Blaschke product
with simple zeros $B$ such that $H\sim$ $K_{B}.$
\end{enumerate}
\end{theorem}
\section{The Proofs{}}
Some parts of the proof follow from our earlier discussion and we begin with
those. First we show (1) and (6) are equivalent. Suppose (6) holds and let
$X(B)$ be the zero set of the Blaschke product $B.$ As we noted earlier the
spaces $K_{B}$ and $DA_{1}(X(B))$ have the same kernel functions and hence are
the same space. Furthermore $DA_{1}(X(B))$ is in the form described in (1),
that is $X(B)$ is in the unit disk which is a complex geodesic, both on its
own and as a subset of any larger $\mathbb{B}^{m}=\mathbb{CH}^{m}.$ On the
other hand, if (1) holds then we can use the fact that the group of
automorphisms acts transitively on the set of complex geodesics to select an
automorphism $\Phi$ mapping the geodesic containing $X$ to the unit disk. By
Theorem \ref{R1} $DA_{m}(X)\sim$ $DA_{1}(\Phi(X))$ and as before $DA_{1
(\Phi(X))\sim K_{B}$ where $B$ is now the Blaschke product with zeros at
$\Phi(X).$ Thus we have (6) as required.
That (4) and (5) are equivalent is Proposition \ref{conjugation = orthogonal}.
It is immediate that (1) implies (2). To see that (2) implies (1) recall that
any two points in $\mathbb{CH}^{n}$ are contained in a unique complex
geodesic. Hence, by the same argument we would use to study colinear points in
Euclidian space we see that if (2) holds then so does (1).
The demonstrations that (6) implies (4) and that (6) implies (3) both use the
function theory of the Hardy space applied to $K_{B}\subset H^{2}.$ That (6)
implies (4) follows from Proposition \ref{exact conjugation} and the
discussion proceeding it. To see that (6) implies (3) note that from classical
function theory on the Hardy space, in particular Pick's theorem, the extreme
value $m_{1}(x_{1})$ for the multiplier $m_{1}\in M(K_{B})$ is the same as the
extreme value for the $H^{\infty}$ interpolation problem of finding
\[
\max\left\{ \operatorname{Re}g(x_{1}):g(x_{2})=...=g(x_{n})=0,\text{
\sup_{z\in\mathbb{B}^{1}}\left\vert g(z)\right\vert \leq1\right\} .
\]
That problem is solved using Blaschke products and we find that the maximum is
$\left\vert D(x_{1})\right\vert $ where $D$ is a Blaschke product with zeros
$\left\{ x_{2},...,x_{n}\right\} .$ The relation between Blaschke factors
and $\delta$ given in (\ref{disk psh}) completes the argument.
It remains to show that (3) implies (6) and that (4) implies (6). If $H$ is
three dimensional then both implications are known. We recall those results
and then use them to pass to the general cases.
\begin{proposition}
\label{3 implies 6}Suppose $H\in\mathcal{F}$ is three dimensional with kernel
functions $\left\{ k_{a},k_{b},k_{c}\right\} .$ Let $s\in M(H)$ the
multiplier of norm one which vanishes at $k_{b}$ and $k_{c}$ and maximizes
$\operatorname{Re}s(k_{a}).$ If
\begin{equation}
s(k_{a})=\delta(k_{a},k_{b})\delta(k_{a},k_{c}). \label{goal
\end{equation}
then $H\in\mathcal{M}$.
\end{proposition}
This is Proposition 27 of \cite{R1}. The proof is based on a detailed analysis
of $DA(X)$ and its multipliers for three point sets $X.$
The next result is from \cite{R1} where it is stated with an oversight. We
will say a three dimensional RKHS is non-degenerate if no entry of its Gram
matrix is zero, equivalently no two kernel functions are orthogonal. The
property is preserved under rescaling and under passage to regular subspaces.
\begin{proposition}
\label{4 implies 6}Suppose $H$ is a non-degenerate three dimensional RKHS
which is $r-$orthogonal, then $H\in\mathcal{M}$.
\end{proposition}
If $H\in\mathcal{F}$ then $H$ is automatically non-degenerate. However the
proposition does not have the assumption that $H\in\mathcal{F}$.
This is Theorem 28 of \cite{R1} where the requirement of non-degeneracy was
omitted. The shape of the proof is that finding rescaling parameters that will
transform the Gram matrix of $H$ into an orthogonal matrix involves solving a
system of equations. For those equations to have a solution a determinant
involving functions of the Gram matrix entries must vanish. That vanishing
gives an equation for the Gram matrix entries which is equivalent to knowing
$H\sim DA(X)$ for $X$ in a single complex geodesic. (This is the place where
the proof in \cite{R1} fails. Without the non-degeneracy assumption the
equation involving the Gram matrix entries may trivialize.)
Suppose now that (3) holds and we want to establish (6). After rescaling we
may suppose $H=DA(X_{H})$ for some $X_{H}\subset\mathbb{B}^{m}.$ Let $J$ be
the regular subspace of $H$ with kernel functions $\left\{ k_{a},k_{b
,k_{c}\right\} .$ If we have (\ref{goal}) then by Proposition
\ref{3 implies 6} $J\in\mathcal{M}$ and thus $J\sim DA(X_{J})$ for some
$X_{J}\subset\mathbb{B}^{1}$; in particular $X_{J}$ is in a single complex
geodesic. As discussed in Section \ref{subspaces} this implies that the points
of $X$ corresponding to those three kernel functions also lie in a single
geodesic. The choice of which three kernel functions we considered was
arbitrary and hence we would have (2) for $H.$ We have already seen that (2)
implies (1) which implies (6).
Thus we are done if we can show that for any choice of three kernel functions
(\ref{goal}) holds. From (\ref{gleason}) we know that given any kernel
functions $k_{\alpha},k_{\beta}$ in an $H\in\mathcal{F}$ there will be a
multiplier $m_{\alpha\beta}\in M(H)$ of norm one which satisfies
\begin{equation}
m_{\alpha\beta}(k_{\alpha})=0,\text{ \ \ }m_{\alpha\beta}(k_{\beta
)=\delta(k_{\alpha},k_{\beta}). \label{gleason extremal
\end{equation}
We now proceed by contradiction. Suppose we have found three kernel functions,
$\left\{ k_{i}\right\} _{i=1,2,3},$ for which (\ref{goal}) fails and denote
their span by $J.$ Using the notation of (\ref{gleason extremal}) form the
multiplier $r=m_{12}m_{13}\in M(J).$ This multiplier is a candidate for the
extremal problem defining the multiplier $s$ in (\ref{goal}) and by evaluating
if we see that the left hand side in (\ref{goal}) will never be smaller than
the right. Hence by our construction of $J$ (\ref{goal}) fails because the
left hand side is larger. Thus we have a multiplier $t\in M(J)$ of norm one
with
\begin{equation}
t(k_{1})>\delta(k_{1},k_{2})\delta(k_{1},k_{3}). \label{inequality
\end{equation}
As noted in Section \ref{subspaces} the multipliers $m_{\alpha\beta}$ and $t$
all have norm preserving extensions to elements in $M(H).$ We will regard
those extensions as having been made and use the same notation for the
extended multipliers. Staying with the notation in (\ref{gleason extremal})
consider the multiplier in $q\in M(H)$ given b
\[
q=t\prod_{j=4}^{n}m_{1j}.
\]
By the Banach algebra property of $M(H)$ $q$ has norm at most one and by
construction it vanishes at $k_{2},...,k_{n};$ hence $q$ is a competitor in
the extremal problem defining $m_{1}.$ Furthermore, comparing the definition
of $q$, (\ref{gleason extremal}), and (\ref{inequality}) we see that
$q(k_{1})$ is larger than the right hand side of (\ref{multiplier = product}).
This inequality contradicts the extremal value suggested by
(\ref{multiplier = product}) and hence contradicts the assumption that (3)
holds. This completes the proof that (3) implies (6)
We now show (4) implies (6). We are given $H\in\mathcal{F}$ that is $n$
dimensional and is $r-$orthogonal. Let $\left\{ k_{i}\right\} _{i=1}^{n}$ be
an arbitrary numbering of the kernel functions of $H$ and let $H_{-
\in\mathcal{F}$ be the $n-1$ dimensional regular subspace of $H$ spanned by
the kernel functions $\left\{ k_{i}\right\} _{i=1}^{n-1}.$ We will show that
$H_{-}$ is $r-$orthogonal. Repeating this shows that for any three kernel
functions $\left\{ k_{\alpha},k_{\beta},k_{\gamma}\right\} $ their span,
$H_{\alpha\beta\gamma},$ is $r-$orthogonal. By Proposition \ref{4 implies 6}
this insures $H_{\alpha\beta\gamma}\in M.$ Hence after rescaling
$H_{\alpha\beta\gamma}=DA(X_{\alpha\beta\gamma})$ for a three point set
$X_{\alpha\beta\gamma}$ in the unit disk. As we noted in Section
\ref{subspaces} this insures that when $H$ is rescaled as $DA(X)$ for some $X$
in some $\mathbb{B}^{m}$ then the points of $X$ corresponding to the kernel
functions $\left\{ k_{\alpha},k_{\beta},k_{\gamma}\right\} $ will lie in a
single complex geodesic in $\mathbb{B}^{m}.$ The numbering of the kernel
functions of $H$ was arbitrary and hence this establishes (2) for $H$ from
which (1) follows and then (6).
To prove the reduction we start with $n-$dimensional $H\in\mathcal{F}$ with
kernel functions $\left\{ k_{j}\right\} _{j=1}^{n}$ which, by rescaling, we
can assume has an orthogonal Gram matrix $\mathbf{K}=(k_{ij})_{i,j=1}^{n}$. We
denote the dual basis, the kernel functions of $H^{\#}$, by $\{k_{i
^{\#}\}_{i=1}^{n}$ and write its Gram matrix as $\mathbf{K}^{\#}=(k_{ij
^{\#})_{ij=1}^{n}.$ Because $H$ is orthogonal we have $\mathbf{K
^{-1}=\mathbf{K}^{\#}=\mathbf{K}^{t}$ and hence the condition that $H,$ and
thus $\mathbf{K}$, is orthogonal is expressed by the equations, for $1\leq
i,j\leq n.
\begin{equation}
\sum_{s=1}^{n}k_{is}k_{sj}^{\#}=\delta_{ij}. \label{use
\end{equation}
Let $H_{-}$ be the regular subspace of $H$ spanned by the kernel functions
$\left\{ k_{i}\right\} _{i=1}^{n-1}.$ The inner product of those $k_{j}$ is
the same whether they are regarded as vectors in $H$ or $H_{-}$ and hence the
Gram matrix of $H_{-}$ is the upper left $\left( n-1\right) \times\left(
n-1\right) $ block of the Gram matrix of $H:\mathbf{K}_{-}=\left(
k_{ij}\right) _{ij=1}^{n-1}.$ Let $\{k_{-i}^{\#}\}$ be the dual basis of the
basis $\left\{ k_{i}\right\} _{i=1}^{n-1}$ and let $\mathbf{K}_{-
^{\#}=\left( k_{ij}^{\#}\right) _{ij=1}^{n-1}$ be its Gram matrix. For ease
of reading we set $g_{j}=k_{-j}^{\#}$ and $g_{ij}=k_{-ij}^{\#}.$
The vectors $\{k_{j}^{\#}\}_{j=1}^{n-1}$ are nearly but not quite the
$\left\{ g_{j}\right\} _{j=1}^{n-1}.$ They give the correct inner products;
for $1\leq i,j\leq n-1
\begin{equation}
\left\langle k_{i},k_{j}^{\#}\right\rangle =\delta_{ij}, \label{want
\end{equation}
but they are not in $H.$ To obtain vectors in $H$ which satisfy the analog of
(\ref{want}) we apply $P,$ the orthogonal projection from $H$ to $H_{-}.$ That give
\[
\left\langle k_{i},P(k_{j}^{\#})\right\rangle =\delta_{ij};
\]
and hence the $P(k_{j}^{\#})$ are our desired $g_{j},$ $j=1,...,n-1.$
We know $P(k_{j}^{\#})=k_{j}^{\#}-Q(k_{j}^{\#})$ where $Q$ is the projection
complimentary to $P$ and we use that to compute $g_{j}.$ $Q$ is the projection
onto the orthocomplement of $H_{-}$ in $H$ and by the definition of the dual
basis that is the subspace of $H$ spanned by $k_{n}^{\#}$. Hence for any $b$
\[
P(b)=b-\left\langle b,\frac{k_{n}^{\#}}{\left\Vert k_{n}^{\#}\right\Vert
}\right\rangle \frac{k_{n}^{\#}}{\left\Vert k_{n}^{\#}\right\Vert }.
\]
In particular, for $1\leq j\leq n-1
\begin{align*}
g_{j} & =k_{j}^{\#}-\left\langle k_{j}^{\#},\frac{k_{n}^{\#}}{\left\Vert
k_{n}^{\#}\right\Vert }\right\rangle \frac{k_{n}^{\#}}{\left\Vert k_{n
^{\#}\right\Vert }\\
& =k_{j}^{\#}-\frac{k_{jn}^{\#}}{k_{nn}^{\#}}k_{n}^{\#}.
\end{align*}
We want to evaluate $\mathbf{K}_{-}^{\#}=\left( g_{ij}\right) ;
\begin{align*}
g_{ij} & =\left\langle g_{i},g_{j}\right\rangle =\left\langle k_{i
^{\#}-\frac{k_{in}^{\#}}{k_{nn}^{\#}}k_{n}^{\#},\text{ }k_{j}^{\#
-\frac{k_{jn}^{\#}}{k_{nn}^{\#}}k_{n}^{\#}\right\rangle \\
& =\left\langle k_{i}^{\#},\text{ }k_{j}^{\#}\right\rangle -\left\langle
\frac{k_{in}^{\#}}{k_{nn}^{\#}}k_{n}^{\#},k_{j}^{\#}\right\rangle +0\\
& =k_{ij}^{\#}-\frac{k_{in}^{\#}k_{nj}^{\#}}{k_{nn}^{\#}
\end{align*}
To show that $\mathbf{K}_{-}$ is an orthogonal matrix we will show
\[
\sum_{s=1}^{n-1}k_{-is}k_{-sj}^{\#}=\sum_{s=1}^{n-1}k_{is}g_{sj}=\delta_{ij}.
\]
We comput
\begin{align*}
\sum_{s=1}^{n-1}k_{is}g_{sj} & =\sum_{s=1}^{n-1}k_{is}\left( k_{sj
^{\#}-\frac{k_{sn}^{\#}k_{nj}^{\#}}{k_{nn}^{\#}}\right) \\
& =\sum_{s=1}^{n-1}k_{is}k_{sj}^{\#}-\sum_{s=1}^{n-1}\frac{k_{is}k_{sn
^{\#}k_{nj}^{\#}}{k_{nn}^{\#}}\\
& =\left( \delta_{ij}-k_{in}k_{nj}^{\#}\right) +\frac{\left( -\delta
_{in}+k_{in}k_{nn}^{\#}\right) k_{nj}^{\#}}{k_{nn}^{\#}}\\
& =\delta_{ij}.
\end{align*}
In the passage from the second line to the third we used (\ref{use}) for the
index pair $\left( i,j\right) $ and for the index pair $\left( i,n\right)
.$ The passage to the final line used the fact that $i<n$ and hence
$\delta_{in}=0.$
\section{Comments and Variations}
\subsection{Reformulations of $m_{1}(x_{1})=\prod_{j=2}^{n}\delta(x_{1
,x_{j})$}
Suppose $m_{1}\in M(H)$ is the multiplier described in (3) of Theorem \ref{A}.
Taking note of Proposition \ref{extreme} we can write (\ref{extremal}) for
this multiplier and obtai
\[
m_{1}\frac{k_{1}}{\left\Vert k_{1}\right\Vert }=h.
\]
Here $h$ is the function of norm one which vanishes at $x_{2},...,x_{n}$ and
maximizes $\operatorname{Re}h(x_{1}).$ The space of competitors for $h$ is one
dimensional and hence $h=k_{1}^{\#}/\left\Vert k_{1}^{\#}\right\Vert .$ Using
this and taking the inner product of both sides of the previous equation with
$k_{1}$ we fin
\[
m_{1}(k_{1})\frac{k_{11}}{\left\Vert k_{1}\right\Vert }=\frac{\left\langle
k_{1}^{\#},k_{1}\right\rangle }{\left\Vert k_{1}^{\#}\right\Vert
\]
which simplifies as
\[
m_{1}(k_{1})=(\left\Vert k_{1}\right\Vert \left\Vert k_{1}^{\#}\right\Vert
)^{-1}.
\]
We obtain another expression for $m_{1}(k_{1})$ if we consider the idempotent
multiplier in $M(H)$ which takes the value $1$ at $k_{1}$ and is zero
elsewhere. Denote that multiplier by $\operatorname{Idem}_{1}.$ It is in the
one dimensional space spanned by $m_{1}$ and hence $\operatorname{Idem
_{1}=m_{1}/m_{1}(k_{1}).$ Thu
\[
m_{1}(k_{1})=(\left\Vert \operatorname{Idem}_{1}\right\Vert )^{-1}.
\]
Either of these evaluations of $m_{1}(k_{1})$ could be used on the left hand
side of (\ref{multiplier = product}). We had already noted that the $\delta$'s
on the right hand side can be evaluated using the Hilbert space structure $H$
or the using the structure of $M(H).$ Hence we can write versions of
(\ref{multiplier = product}) based entirely on data from $H$ or entirely on
using data from $M(H).$
\subsection{When is $H^{\#}\in\mathcal{F}?$\label{C}}
We noted that Corollary \ref{is also} leads to the fact that $H\in\mathcal{M}$
is a necessary and sufficient condition for $H^{\#}\in\mathcal{M}$. On the
other hand it is not clear what conditions on a RKHS $H,$ even one in
$\mathcal{F}$, will insure $H^{\#}\in\mathcal{F}$. Here we indicate that the
condition $H\in\mathcal{M}$, which is sufficient because $H^{\#
\in\mathcal{M\subset F}$, is not necessary, and the condition $H\in
\mathcal{F}$ is not sufficient.
Suppose $X\subset\mathbb{B}^{1}\subset\mathbb{B}^{2}$ and $H=DA(X)$.
Arbitrarily slight modification of $X$ can produce $\tilde{X}\subset
\mathbb{B}^{2}$ which are not contained in a single complex geodesic. For
those $\tilde{X}$ by Theorem \ref{A} $\tilde{H}=DA(\tilde{X})\in$
$\mathcal{F\smallsetminus M}$. If $\tilde{X}$ is sufficiently close to $X$
then the Gram matrix of $\tilde{H}$ is arbitrarily close to the Gram matrix of
$H.$ Furthermore, passing to inverse matrices, the Gram matrix of $\tilde
{H}^{\#}$ is arbitrarily close to the Gram matrix of $H^{\#}.$ We know
$H^{\#}\in\mathcal{M\subset F}$ and we know that being in $\mathcal{F}$ is an
open condition on the Gram matrix of a HSRK \cite{AM} (this is in contrast to
the condition for being in $\mathcal{M}$) hence, if our perturbation was
sufficiently small then $H^{\#}\in\mathcal{F}$.
On the other hand $H\in\mathcal{F}$ is not itself sufficient to insure
$H^{\#}\in\mathcal{F}$. Consider, for $0<a,b<1,$ $X=\left\{ \left(
0,0\right) ,\left( a,0\right) ,\left( 0,b\right) \right\} =\mathbb{B
^{2},$ in some sense the extreme opposite of $X$ being in a single geodesic.
The Gram matrix of $H=DA(X)$ has the form
\[
G
\begin{pmatrix}
1 & 1 & 1\\
1 & \ast & 1\\
1 & \ast & \ast
\end{pmatrix}
\]
The Gram matrix of $H^{\#}$ is $G^{-1}$ and by explicit computation the
$(2,3)$ entry of $G^{-1}$ is $0$ which is impossible for a RKHS in
$\mathcal{F}$.
\subsection{The Use of Hardy Space Theory and the Pick Condition}
In our proof of Theorem \ref{A} we used the function theory of the Hardy space
to study $K_{B}\subset H^{2},$ in particular to prove that (6) implies (3) and
that (6) implies (4). It would be interesting to have a proof of either of
these implication, or of the equivalence of (3) and (4) without involving (6),
that was inside the the theory of RKHS and did not use function theory. In
this context we note the work of Cole, Lewis, and Wermer in \cite{CLW}. They
study conditions on $M(H),$ the multiplier algebra of a RKHS $H,$ which
suffice to insure that $H\in\mathcal{M}$. They have two approaches, one using
operator theory and a second which uses their hypotheses on the multiplier
algebras to, in effect, reconstruct and thus reintroduce parts of Hardy space
function theory.
Many of the steps in the proof of Theorem \ref{A} did not require the Pick
property and hence can be used for any finite dimensional RKHS. However there
were places where we did use the Pick property and do not know the extent to
which it could be avoided. In particular we used Proposition \ref{extreme} to
connect multiplier algebra statements to Hilbert space statements. This fact,
which is a characteristic property of spaces in $\mathcal{F}$, was used in
showing $\delta$ satisfies (\ref{gleason}) and in our analysis of condition
(3) in Theorem \ref{A}.
Also, the hypothesis $H\in\mathcal{F}$ was used in passing from condition (2)
to condition (1). For instance it is not clear that given a general four
dimensional RKHS $H,$ and knowing that all of its regular three dimensional
subspaces are in $\mathcal{M}$, is enough to insure $H\in\mathcal{F}$. If we
could show that, that $H\in\mathcal{F}$, then by the implication (2) implies
(1) in Theorem \ref{A} we would also know $H\in\mathcal{M}$. In this context
it is interesting to note that there are examples due to Quiggen in which all
the regular subspaces of a four dimensional $H$ are in $\mathcal{F}$ but $H$
is not in $\mathcal{F}$. Those are discussed in \cite{R2}.
To see this issue in context, suppose we wanted to show that an orthogonal
$H$, not necessarily in $\mathcal{F}$, satisfied Statement (1), and hence
Statement (6), in Theorem \ref{A}. The reduction to the case of three
dimensional regular subspaces is a linear algebra argument and so continues to
hold. If we assume $H$ is non-degenerate we can then use Proposition
\ref{4 implies 6} to obtain a version of Statement (2). However without
knowing $H\in\mathcal{F}$ it is not clear how to proceed from Statement (2) to
Statement (1).
\subsection{Replacing RKHS by Point Sets in Projective Space}
It is possible to recast much of the previous discussion in the language of
point sets in complex projective space. The viewpoint is intriguing but it is
not clear where it leads. We will be brief and informal.
Suppose we start with the finite dimensional Hilbert space $\mathbb{C}^{n}$
and regard it as a RKHS, $H,$ by selecting a set of basis vectors $\left\{
k_{i}\right\} $ and declaring those vectors to be reproducing kernels. In
fact any finite dimensional RKHS is a rescaling of such an $H.$
Let $\mathbb{PC}^{n-1}$ be the complex projective space of lines in
$\mathbb{C}^{n};$ for each nonzero vector $v\in\mathbb{C}^{n}$ we denote the
line containing $v$ by $\left[ v\right] ;$ thus $\left[ v\right]
\in\mathbb{PC}^{n-1}.$ Hence we can associate to $H$ the set $\left[
H\right] =\left\{ \left[ k_{i}\right] \right\} _{i=1}^{n}\subset
\mathbb{PC}^{n-1}.$ Note that if we rescale $H$ to $\tilde{H}$ by selecting
scalars $\left\{ \lambda_{i}\right\} $ declaring the $\left\{ \lambda
_{i}k_{i}\right\} $ to be the kernels for $\tilde{H},$ then $[\tilde
{H}]=\left[ H\right] .$ The rescaled space $\tilde{H}$, and in some sense
$\tilde{H}$ is the generic rescaling of $H$, produces the same set in
$\mathbb{PC}^{n-1}.$ From our point of view, our interest is in properties
invariant under rescaling, this is an attractive feature.
Next note that there set $\left[ H^{\#}\right] $ associated to the dual RKHS
$H^{\#}$ can also be described in the language of $\mathbb{PC}^{n-1}.$ To
$n-1$ generic points in $\mathbb{PC}^{n-1}$ there correspond $n-1$ linearly
independent lines in $\mathbb{C}^{n}$. There is a unique line in
$\mathbb{C}^{n}$ orthogonal to those lines and that orthogonal line
corresponds to a point in $\mathbb{PC}^{n-1}$ which is "orthogonal" to each of
the $n-1$ points in the starting set. This process applied to each of the
$n-1$ point subsets of $\left[ H\right] $ gives a set $\left[ H\right]
^{\#}$. Tracking the definitions we see that $\left[ H\right] ^{\#}=\left[
H^{\#}\right] .$
If $H$ is orthogonal then $\left\{ k_{i}\right\} $ and $\{k_{i}^{\#}\}$ are
related through a period two conjugate linear isometry of the Hilbert space.
This isometry descends to an isometry of $\mathbb{PC}^{n-1}$ which
interchanges $\left[ H\right] $ and $\left[ H^{\#}\right] .$ On the other
hand if there is such an isometry connecting $\left[ H\right] $ and $\left[
H^{\#}\right] $ then one can show by analysis in $\mathbb{PC}^{n-1}$ that the
isometry extends to a global isometry of $\mathbb{PC}^{n-1}$ which, by
Wigner's theorem, must come from a unitary or antiunitary map of the original
$\mathbb{C}^{n}.$
In sum, many of the ideas we have considered can be comfortably reformulated
as statements about certain types of symmetric point sets in projective space.
|
\section{Introduction}
This paper presents a data-driven model reduction method to solve parabolic inverse source problems and studies the convergence analysis of the proposed method. Inverse problems are very important in physics, engineering, and bioengineering. The inverse source problems, which seeks reconstruction of source from final time observation, have attached much attention of the researchers over the past decades, see an introduction and references in \cite{Isakov2013}. They have been widely studied in the literature, applied to many physical and engineering source identification problems, e.g. migration of groundwater, groundwater pollution detection, control of pollution source and environmental protection \cite{CHENG2020106213, new3, GER83, new2, new6} and references therein. The accurate recovery pollutant source is crucial to environmental safeguard in cities with high populations \cite{CHENG2010142}. The estimation of the strength of acoustic sources from measurements can be found in e.g. \cite{new6, Badia2011, nelson20}. The inverse source problems that arise from PDEs are ill-posed in the sense of Isakov and Hadamard \cite{new6,hadamard}, since the eigenvalues of the elliptic operators decay exponentially fast, especially the lack of stability with respect to the uncertainty in the measurement data is the most difficult challenge for numerical inversions. Namely, a small change in the data may lead to a significant difference in the reconstructed source strength. Due to the important applications of inverse source problems, numerical methods have been widely explored \cite{CHENG2020106213,Chen-Zhang2021,LZ07,new4,GER83} and references can be found therein.
In this paper, we will work on a very practically physical scenario, assuming that the observational measurement is collected point-wisely over a set of distributed sensors located at $\{x_i\}^n_{i=1}$ over the physical domain \cite{new2, new6, L08, NNR98}. At each sensor, independently additional noise or random error will be considered due to the uncertainty of natural noise, measurement errors, and the other uncertainty of the model itself. We will take one of the realistic approaches for such inverse problems, i.e. optimizing the mean-square error with a proper Tikhonov type regularizations \cite{Chen-Zhang2021, GER83, WY10}. Classical methods such as regression methods \cite{GER83}, linear and nonlinear programming methods \cite{GER83}, linear and nonlinear conjugate gradient methods \cite{AB05, WY10}, Newton-type methods, etc. can be applied during optimization. For the inverse parabolic source problem, one usually uses iteration optimization methods to find the true source term \cite{Chen-Zhang2021, Johansson2014, Johansson2007}. In each iteration, one has to solve the forward parabolic equation one or two times. However, as the sizes of discrete problems grow (e.g. finite element method (FEM) or finite difference method (FDM)), the computation time will grow rapidly, especially for such evolution problems. As a consequence, the computation of the forward equation will cost the most time throughout the whole procedure. This motivates us to develop efficient model reduction methods to address this issue.
One of the successful model reduction ideas in solving evolution problems is the proper orthogonal decomposition (POD) method \cite{sirovich1987,berkooz1993POD}. The POD method uses the data from an accurate numerical simulation and extracts the most energetic modes in the system by using the singular value decomposition. This approach generates low-dimensional structures that can approximate the solutions of the evolution problems with high accuracy. The POD method has been used to solve many types of PDEs, including linear parabolic equations \cite{volkwein2013proper}, Navier‐Stokes equations \cite{kunisch2001galerkin}, viscous G-equations \cite{gu2021error}, Hamilton–Jacobi–Bellman (HJB) equations \cite{kunisch2004hjb}, and optimal control problem \cite{alla2013time}. The interested reader is referred to \cite{quarteroni2015reduced,benner2015survey,hesthaven2016certified} for a comprehensive introduction of the model reduction methods.
Since the POD method can significantly accelerate the computation of the forward problem compared with the traditional methods, e.g. FEM and FDM, we apply the POD method to solve parabolic inverse source problems with scattered measured data at the final time.
The key idea is to exploit (model-based) and construct (data-based) the intrinsic approximate low-dimensional structures of the underlying problem that consists of two components. First, we have a training component that computes a set of data-driven basis functions (i.e. the POD basis functions) to achieve significant dimension reduction in the solution space. Following up, there is a fast-solving component that solves the optimization problem using the constructed POD basis functions in each iteration. Hence we achieve an effective data-driven model reduction method in solving the parabolic inverse source problems and overcome the typical computational bottleneck in the traditional methods, e.g. FEM and FDM.
Then, we study the convergence analysis of the proposed method. In \cite{Chen-Zhang2021}, a general theory of the stochastic convergence of numerical method has been established for a certain type of inverse problem. Based on this framework, we explore the convergence theory in the POD settings. Specifically, we will prove a relatively weaker convergence of the POD method when the source term only belongs to $L^2(\Omega)$, where $\Omega$ is the computational domain. Unlike the traditional analysis of the POD method \cite{kunisch2001galerkin} or FEM convergence, we do not assume the higher regularity for parabolic PDE solution $u$, i.e. $u_{tt}$ to be bounded in $L^2(\Omega)$, which is quite strict in many cases. Based on our analysis, we derive the stochastic convergence when applying the POD method to the parabolic inverse source problem with uncertain data. Our analysis in Theorem \ref{thm_convergence} shows that the optimal error of the Tikhonov type least-square optimization problem depends on the noise level, the number of sensors, and without any higher source regularization but the $L^2(\Omega)$ norm of the source term. A self-consistent algorithm to get the optimal smoothing parameter is also established using the POD method to solve the inverse parabolic source problem, motivated by the recent study in \cite{Chen-Zhang, Chen-Zhang2021}. Finally, we conduct numerical experiments to demonstrate the accuracy and efficiency of the proposed method. Several kinds of source functions are involved. Comparing to original optimization with FEM basis functions, the proposed POD method significantly reduces the computational cost and maintains the same accuracy as the FEM.
The rest of the paper is organized as follows. In Section 2, we introduce the setting of the parabolic inverse source problems. In Section 3, we develop the data-driven model reduction method for solving parabolic inverse source problems. In Section 4, we provide the error estimate for the proposed method. In Section 5, we present numerical results to demonstrate the accuracy of our method. Finally, concluding remarks are made in Section 6.
\section{Parabolic inverse source problems}
To start with, we consider a parabolic equation,
\begin{equation}\label{zz0}
\left\{
\begin{aligned}
u_t +\mathcal{L}u &= f(x) &\mbox{in } \Omega\times (0, T), \\
u(x, t)&= 0 &\mbox{on } \partial \Omega\times (0, T),\\
u(x, 0)&= g(x) &\mbox{in } \Omega\,,
\end{aligned}
\right.
\end{equation}
where $\Omega\subset \mathbb R^d$ $(d=1,2,3)$ is a bounded domain with $C^2$ boundary or a convex domain satisfying the uniform cone condition, $\mathcal{L}$ is a second-order elliptic operator given by $\mathcal{L}u=-\nabla\cdot (a(x)\nabla u) +c(x)u$, and
$g(x)$ is the initial value. We assume the elliptic operator $\mathcal{L}$ is uniform elliptic, i.e. there exist $a_{\min}, a_{\max}>0$, such that $a_{\min}<a(x)<a_{\max}$ for all $x \in \Omega$. Moreover, we assume $a(x)\in C^{1}(\bar{\Omega})$, $c(x)\in C(\bar\Omega)$ and $c(x)\geq 0$.
Let $u$ be the solution of the parabolic equation \eqref{zz0}. We define the forward operator $\mathcal{S}:$ $L^2(\Omega)\rightarrow H^2(\Omega)$ by $\mathcal{S}f=u(\cdot,T)$. The forward problem is to compute the solution $u(\cdot,t)$ for $t>0$ with known source term $f(x)$ and known initial condition $g(x)$.
In the parabolic inverse source problem, $f(x)\in L^2(\Omega)$ is an unknown source term that we want to reconstruct based on the final time measurement $u(\cdot,T)$. To be specific, we will focus on a very practically physical scenario, where we assume that the observational measurement is collected point-wisely over a set of distributed sensors located at $\{x_i\}^n_{i=1}$ over the physical domain $\Omega$ (see e.g.\cite{new2, new6, L08, NNR98}). We also assume the measurement data is always blurred by noise and takes the form $m_i=\mathcal{S}f^*(x_i)+e_i$, $i=1, \cdots, n$, where $f^*\in L^2(\Omega)$ is the true source term and $\{e_i\}^n_{i=1}$ are independent and identically distributed (i.i.d.) random variables on a proper probability space ($\mathfrak{X},\mathcal{F},\mathbb{P})$. From \cite{Chen-Zhang2021} and the analysis therein, we know that $\|u\|_{C([0,T];H^2(\Omega))}\le C\|f\|_{L^2(\Omega)}$.
According to the embedding theorem of Sobolev spaces, we know that $H^2(\Omega)$ is continuously embedded into $C(\bar\Omega)$ so that $\mathcal{S}f^*(x)$ is well defined point-wisely for all $x\in \Omega$. Without loss of generality, we assume that the scattered locations $\{x_i\}_{i=1}^n$ are uniformly distributed
in $\Omega$, i.e., there exists a constant $B>0$ such that ${d_{\max}}/{d_{\min}} \leq B$, where
${d_{\max}}$ and ${d_{\min}}$ are defined by
\begin{align}\label{aa}
d_{\max}=\mathop {\rm sup}\limits_{x\in \Omega} \mathop {\rm inf}\limits_{1 \leq i \leq n} |x-x_i|
~~~\mbox{and} ~~ ~
d_{\min}=\mathop {\rm inf}\limits_{1 \leq i \neq j \leq n} |x_i-x_j|.
\end{align}
We denote the inner product between the measurement data and any function $v\in C(\bar\Omega)$ by $(m,v)_n=\frac{1}{n}\sum^n_{i=1}m_iv(x_i)$. Moreover, we denote the inner product between two functions by $(u,v)_n=\frac{1}{n}\sum^n_{i=1}u(x_i)v(x_i)$ for any $u,v\in C(\bar\Omega)$ and the empirical norm $\|u\|_n=\big(\frac{1}{n}\sum_{i=1}^{n} u^2(x_i)\big)^{1/2}$ for any $u\in C(\bar\Omega)$.
Equipped with these definitions, we can define the parabolic inverse source problem as to recover the unknown source term $f^*$ from the noisy final time measurement data $m_i=\mathcal{S}f^*(x_i)+e_i$, $i=1,...,n$. We will use the regularization method to solve this parabolic inverse source problem. Specifically, we will look for an approximate solution of the true source term $f^*$ by solving the following least-squares regularized minimization problem:
\begin{align}\label{para-examp1}
\mathop {\rm min}\limits_{f\in X} \|Sf-m\|^2_n+\lambda_n \|f\|_{L^2(\Omega)}^2\,,
\end{align}
where $\lambda_n$ is the regularization parameter.
We first recall an existing work \cite{Chen-Zhang2021}, where the optimal stochastic convergence of regularized solutions and finite element solutions to time-dependent parabolic inverse source problems \eqref{zz0} have been studied. The following result represents the stochastic convergences corresponding to the random variables with bounded variance.
\begin{prop}[{Theorem 3.4 in \cite{Chen-Zhang2021}}]\label{para-examthm:2.1}
Suppose $\{e_i\}^n_{i=1}$ are independent random variables satisfying
$\mathbb{E}[e_i]=0$ and $\mathbb{E}[e^2_i]\leq \sigma^2$. Let $f_n\in L^2(\Omega)$ be the solution of the least-squares regularized minimization problem \eqref{para-examp1}.
Then there exist constants $\lambda_0 > 0$ and $C>0$ such that the following estimates hold
for any $0<\lambda_n \leq \lambda_0$:
\begin{align}
\mathbb{E} \big[\|\mathcal{S}f_n-\mathcal{S}f^*\|^2_n\big] \leq C \lambda_n \|f^*\|^2_{L^2(\Omega)} + \frac{C\sigma^2}{n\lambda^{d/4}_n}, \\
\mathbb{E} \big[\|f_n-f^*\|^2_{L^2(\Omega)}\big] \leq C \|f^*\|^2_{L^2(\Omega)} + \frac{C\sigma^2}{n\lambda^{1+d/4}_n},\\
\mathbb{E} \big[\|f_n-f^*\|^2_{H^{-1}(\Omega)}\big] \leq C \lambda_n^{1/2}\|f^*\|^2_{L^2(\Omega)} + \frac{C\sigma^2}{n\lambda^{1/2+d/4}_n}.
\end{align}
\end{prop}
A stronger stochastic convergence when $\{e_i\}^n_{i=1}$ are independent Gaussian random with variance $\sigma^2$ can also be found in \cite{Chen-Zhang2021}. Our convergence analysis below also applies to this case. From Proposition \ref{para-examthm:2.1}, we have an optimal choice of the regularization parameter $\lambda_n$ in \eqref{para-examp1} as follows:
\begin{align}\label{opti-para}
\lambda_n^{1/2+d/8}=O\big(\sigma n^{-1/2}\|f^*\|_{L^2(\Omega)}^{-1}\big).
\end{align}
It is an a priori estimate with knowing the noise level $\sigma$ and the knowledge of the true source term $f^*$, yet it is our goal to recover the true source term $f^*$ with unknown $\sigma$. The noise level is difficult to measure in many circumstances. In \cite{Chen-Zhang2021}, the authors proposed a self-consistent algorithm to compute the optimal $\lambda_n$ without any knowledge of the noise level $\sigma$ and the true source term $f^*$. Apply the POD method to this algorithm, we will also successfully determine the optimal parameter $\lambda_n$ iteratively. We find the proposed POD method maintains the same accuracy as the FEM.
Details can be found at the end of Section \ref{sec:eg1}.
An effective way to approximate the optimal control problem \eqref{para-examp1} with a proper regularization parameter $\lambda_n$ is the iteration method. For each iteration, one has to solve the forward problem \eqref{zz0} and its adjoint problem for at least one time. In \cite{Chen-Zhang2021}, the FEM was used to approximate \eqref{para-examp1} and the optimal second-order convergence with respect to space has been proved. But as the DOF grows for a discrete method such as FEM and FDM, the iteration method will cost too much computational time. Therefore, we need to design numerical methods that allow us to efficiently and accurately solve \eqref{para-examp1}.
\section{The data-driven model reduction method}\label{sec:DataDrivenModelReduction}
In this section, we will use the POD method to accelerate the inverse problem computation. We first construct the POD basis functions from the snapshot solutions of the parabolic equation \eqref{zz0} with some known type of source function. Then, we solve the optimization problem in the inverse problems with the constructed POD basis.
\subsection{Construction of the POD basis functions}\label{sec:PODmethod}
\noindent
Assume that $u \in H^1_0(\Omega)$ is the solution to the following weak formulation of the parabolic equation \eqref{zz0}
\begin{align}\label{eqn:weak_formulation}
(\partial_t u, \psi) + a(u, \psi) = (f, \psi), \quad \forall \psi \in H^1_0(\Omega), \quad t \in [0, T],
\end{align}
where $\Omega \subset \mathbb{R}^d$ and $a(\cdot, \cdot)$ is a bilinear form on $H^1_0(\Omega) \times H^1_0(\Omega)$ that is defined according to the elliptic operator $\mathcal{L}$. Given a set of solutions at different time instances $\big\{ u(\cdot, t_0), u(\cdot, t_1), \ldots, u(\cdot, t_m) \big\}$ where $t_k = k \Delta t$ with $\Delta t = \frac{T}{m}$, we first get the solution snapshopts $\big\{ y_1, \ldots, y_{m + 1}, y_{m + 2}, \ldots, y_{2m + 1} \big\}$, where
$y_k = u(\cdot, t_{k - 1})$, $k = 1, \ldots, m + 1$, and $y_k = \overline{\partial} u(\cdot, t_{k - m - 1})$, $k = m + 2, \ldots, 2m + 1$ with $\overline{\partial} u(\cdot, t_{k}) = \frac{u(\cdot, t_{k}) - u(\cdot, t_{k - 1})}{\Delta t}$, $k = 1, \ldots, m$.
Then, the POD basis functions $\{ \psi_k \}_{k = 1}^{{N_{pod}}}$ are built from the solution snapshopts by minimizing the following error:
\begin{align}\label{POD_proj_err}
\frac{1}{2m+1}\Big( \sum_{j = 0}^m \| u(t_j) - \sum_{k = 1}^{{N_{pod}}} (u(t_j), \psi_k)_{L^2(\Omega)} \psi_k \|_{L^2(\Omega)}^2 + \sum_{j = 1}^m \| \overline{\partial} u(t_j) - \sum_{k = 1}^{{N_{pod}}} (\overline{\partial} u(t_j), \psi_k)_{L^2(\Omega)} \psi_k \|_{L^2(\Omega)}^2 \Big)
\end{align}
subject to the constraints that $\big(\psi_{k_1}(\cdot),\psi_{k_2}(\cdot)\big)_{L^2(\Omega)}=\delta_{k_1k_2}$, $1\leq k_1,k_2 \leq l$, where $\delta_{k_1k_2}=1$ if $k_1=k_2$, otherwise $\delta_{k_1k_2}=0$.
Using the method of snapshot proposed by Sirovich \cite{Sirovich:1987}, we know that the optimization problem \eqref{POD_proj_err} can be reduced to an eigenvalue problem:
\begin{equation}\label{eigenvalueproblemPOD}
Kv = \mu v,
\end{equation}
where the correlation matrix $K$ is computed from the solution snapshots $\{ y_1, y_2,\ldots, y_{2m + 1} \}$ with entries $K_{ij} = (y_i, y_j)_{L^2(\Omega)}$, $i,j = 1, \ldots, 2m + 1$, and is symmetric and semi-positive definite.
We sort the eigenvalues in a decreasing order as $\mu_1 \geq \mu_2 \geq ... \geq \mu_{2m + 1} > 0$ and
the corresponding eigenvectors are denoted by $v_k$, $k=1,...,2m + 1$. It can be shown that if the POD basis functions are constructed by
\begin{equation}\label{PODbasisMethodOfSnapshot}
\varphi_{k}(\cdot) = \frac{1}{\sqrt{\mu_k}}\sum_{j=1}^{2m + 1}(v_k)_j u(\cdot,t_j), \quad 1\leq k \leq 2m + 1,
\end{equation}
where $(v_k)_j$ is the $ j$-th component of the eigenvector $v_k$, they will minimize the error \eqref{POD_proj_err}. This result as well as the error formula were proved in \cite{holmes:98}.
\begin{prop}[Sec. 3.3.2, \cite{holmes:98} or p. 502, \cite{Willcox2015PODsurvey}]\label{Prop_PODError}
Let $\mu_1 \geq \mu_2 \geq ... \geq \mu_{2m + 1} > 0$ denote the positive eigenvalues of $K$ in \eqref{eigenvalueproblemPOD}. Then $\{\psi_k\}_{k=1}^{N_{pod}}$ constructed according to \eqref{PODbasisMethodOfSnapshot}
is the set of POD basis functions, and we have the following error formula:
\begin{equation}\label{SnapshotOfSolutions}
\frac{\sum_{i=1}^{2m + 1}\left|\left|y_i - \sum_{k=1}^{{N_{pod}}}\big(y_i,\psi_k(\cdot)\big)_{L^2(\Omega}\psi_k(\cdot)\right|\right|_{L^2(\Omega)}^{2}}{ \sum_{i=1}^{2m+1}\left|\left|y_i\right|\right|_{L^2(\Omega)}^{2}}
= \frac{\sum_{k={N_{pod}}+1}^{2m + 1} \mu_k}{\sum_{k=1}^{2m + 1} \mu_k},
\end{equation}
where the number ${N_{pod}}$ is determined according to the ratio $\rho=\frac{\sum_{k={N_{pod}}+1}^{2m + 1} \mu_k}{\sum_{k=1}^{2m + 1}\mu_k}$.
\end{prop}
In practice, we shall make use of the decay property of eigenvalues in $\mu_k$ and choose the first $m$ dominant eigenvalues such that the ratio $\rho$ is small enough to achieve an expected accuracy, for instance $\rho=1\%$. One would prefer the eigenvalues decays as fast as possible so that one can ensure high accuracy with few POD basis functions.
\subsection{A Fast algorithm for solving parabolic inverse source problems}
We now propose the fast algorithm for solving parabolic inverse source problems based on some given discretization, including our proposed POD method. We first define the functional $\mathcal{J}$ to be
\begin{align}
\mathcal{J}[f]=\|\mathcal{S}f-m\|^2_n+\lambda_n \|f\|_{L^2(\Omega)}^2.
\end{align}
Then, the least-squares regularized minimization problem \eqref{para-examp1} becomes to solve the following misfit functional:
\begin{align}
\mathop {\rm min}\limits_{f\in L^2(\Omega)} \mathcal{J}[f].
\end{align}
\begin{lemma}\label{Fdifferentiable}
The misfit functional $\mathcal{J}[f]$ is Fr$\acute{e}$chet-differentiable.
\end{lemma}
\begin{proof}
From the definition of Fr$\acute{e}$chet-differentiable, we need to compute
\begin{align}\label{Frechet-differentiable}
d\mathcal{J}[f](v)&=\lim_{t\rightarrow 0} \frac{\mathcal{J}[f+tv]-\mathcal{J}[f]}{t}\nonumber\\
&= (\mathcal{S}f-m,\mathcal{S}v)_n + \lambda_n(f,v)\nonumber\\
&= \big(\mathcal{S}^*(\mathcal{S}f-m),v\big) + \lambda_n(f,v)\nonumber\\
&= \big(\mathcal{S}^*(\mathcal{S}f-m)+ \lambda_n f,v\big),
\end{align}
where $v\in L^2(\Omega)$. In \eqref{Frechet-differentiable}, the second equality is easily to get from the quadratic form of the misfit functional $\mathcal{J}[f]$ and $\mathcal{S}^*$ is the adjoint operator of $\mathcal{S}$ in the third equality. Thus, one can directly obtain that
\begin{align}
d\mathcal{J}[f]=\mathcal{S}^*(\mathcal{S}f-m)+ \lambda_n f. \label{dJf}
\end{align}
\end{proof}
The formula \eqref{dJf} in Lemma \ref{Fdifferentiable} allows us to apply the gradient descent method to minimize
the discrepancy functional $\mathcal{J}[f]$. Let $f_{0}$ be the initial guess and $f_k$ denote the the solution of the least-squares regularized minimization problem \eqref{para-examp1} at the $k$-th iteration step, we update the iterative solution by
\begin{align}
f_{k+1}=f_{k}-\alpha d\mathcal{J}[f_{k}], \quad \forall k\in \mathbb{N},
\end{align}
where $\alpha$ is the step size.
Given a fully discrete scheme with solution space $V_{dis}$, we define an approximate forward operator $\mathcal{S}_{dis}$: $L^2(\Omega)\rightarrow V_{dis}$, such that, given any source function $f$, $\mathcal{S}_{dis}f$ gives the numerical solution at final time. In such discrete setting, we turn to solve the following misfit functional:
\begin{align}
\mathop {\rm min}\limits_{f\in \Psi} \mathcal{J}_{dis}[f],
\end{align}
where the functional $\mathcal{J}_{dis}[f]=\|\mathcal{S}_{dis}f-m\|^2_n+\lambda_n \|f\|_{L^2(\Omega)}^2$. We can compute the
Fr$\acute{e}$chet derivative of $\mathcal{J}_{dis}[f]$ and obtain the following iterative scheme:
\begin{align}\label{eqn:pod_iteration}
f_{k+1}=f_{k}-\alpha d\mathcal{J}_{dis}[f_{k}], \quad \forall k\in \mathbb{N},
\end{align}
where $\alpha$ is the step size, $d\mathcal{J}_{dis}[f]=\mathcal{S}_{dis}^*(\mathcal{S}_{dis}f-m)+ \lambda_n f$, and $f_{0}$
is some initial guess.
The iterative scheme \eqref{eqn:pod_iteration} works with both FEM-based Galerkin method and POD-based Galerkin method. However, we emphasize that the DOF of the POD basis functions is much smaller than that of the FEM space. Thus, the POD-based Galerkin method can provide significant computational savings in solving the parabolic equation \eqref{zz0} than the FEM, which allows us to quickly compute the iterative scheme \eqref{eqn:pod_iteration}. Therefore, we obtain a fast algorithm in solving parabolic inverse source problem.
\subsection{Complete algorithm}\label{sec:complete_algorthm}
\noindent Due to the nature of the inverse source problem, the source term is unknown. We assume it to be a sample from some random space. Such random space may not have a closed form or
a finite-dimensional parameterization, so we further
assume that we have $N_f$ realizations, $\{f_l\}_{l=1,\cdots, N_f}$ of the source term. In other words, $\{f_l\}_{l=1,\cdots, N_f}$ are some possible ground truth source function in the inverse problem. Hence, deviating from the classic POD algorithm in solving parabolic equation introduced in Section \ref{sec:PODmethod}, we include snapshots from each proposed source function into the minimization \eqref{POD_proj_err}. Estimation in \eqref{SnapshotOfSolutions} still holds as it is only an algebraic result.
Finally, we summarize the proposed data-driven model reduction method for solving parabolic inverse source problems in Algorithm \ref{alg:InversePOD}, where the notations have been defined before.
\begin{algorithm}[h]
\SetAlgoLined
\textbf{Input}: Observation data $m$. Proposed source function $\{f_l\}_{l=1,\cdots, N_f}$, error thresholds for constructing POD basis functions $\epsilon>0$ and for the optimization problem $tol$, computational time $T$, time step of the FEM $\Delta t$, $N_T=\lceil T/\Delta t\rceil$, mesh size of the FEM $h$, one step finite element solver for the forward problem $\mathcal{O}_{FEM}$, where $\mathcal{S}_{FEM}f=(\mathcal{O}_{FEM}(f))^{N_T}u_0$, and step size in the gradient descent method $\alpha$.\\
\For{$l=1:N_f$}{
Solve the forward problem and store the snapshots $(\mathcal{O}_{FEM}(f_l))^{i}u_0$ for $i=1,\cdots, N_T$.\\
}
{Concatenate} all snapshots as $\{S_j\}_{j=1,\cdots, N_S=N_T\times N_f}=\cup_{i=1,\cdots, N_T, l=1,\cdots, N_f} \{(\mathcal{O}_{FEM}(f_l))^{i}u_0\}$;\\
compute covariance matrix $K=(K_{ij})$ where $K_{ij}=\frac{1}{N_S} (S_i, S_j)_{L^2(\Omega)}$;\\
compute the SVD of $K=(K_{ij})$ and denote the eigenvalues in a decreasing order as $\mu_1 \geq \mu_2 \geq ... \geq \mu_{N_S} > 0$ and the corresponding eigenvectors are denoted by $v_k$;\\
find minimal $N_{pod}$ such that $\frac{\sum_{k=N_{pod}+1}^{N_S} \mu_k}{\sum_{k=1}^{N_S} \mu_k}<\epsilon$;\\
construct POD basis by $\varphi_{k}(\cdot) = \frac{1}{\sqrt{\mu_k}}\sum_{j=1}^{N_S}(v_k)_j S_j, \quad 1\leq k \leq N_{pod}$;\\
construct Galerkin type solver for forward problem on POD basis $\{\varphi_{k}\}_{k=1,\cdots, N_{pod}}$ as $S_{POD}$;\\
Set $f$ as an initial guess $f=f_0$.\\
\While{$\mathcal{J}_{pod}[f_k]=\|\mathcal{S}_{pod}f_k-m\|^2_n+\lambda_n \|f_k\|_{L^2(\Omega)}^2>tol$ }{
update $f_{k+1}\leftarrow f_k-\alpha d\mathcal{J}_{pod}[f_{k}]$ where $d\mathcal{J}_{pod}[f_k]=\mathcal{S}_{pod}^*(\mathcal{S}_{pod}f_k-m)+ \lambda_n f_k$.
}
\textbf{Output}: computed source term $f$.
\caption{\textbf{A fast algorithm for solving parabolic inverse source problem}}
\label{alg:InversePOD}
\end{algorithm}
\section{Convergence analysis}\label{sec:ConvergenceAnalysis}
In this section, we will first present the general discrete approximation to the optimization problem \eqref{para-examp1}.
Recall that $V_{dis}\subset X$ be the discrete function space with dimension $N_{dis}$ and $\mathcal{S}_{dis}$ be the discrete approximation of the operator $\mathcal{S}:L^2(\Omega)\to H^2(\Omega)$. We make the following assumptions on the discrete function space $V_{dis}$ and the discretized operator $\mathcal{S}_{dis}$.
\begin{assu}\label{Assumption3}
(1) There exists an error estimate term $e_{dis}$ such that the discrete operator $S_{dis}$ satisfies
\begin{align}
\|\mathcal{S}f-\mathcal{S}_{dis}f\|^2_n\leq Ce_{dis}\|f\|^2_{L^2(\Omega)}.
\end{align}
(2) For any $f\in L^2(\Omega)$, there exists $v\in V_{dis}$ such that,
\begin{align}
\lambda_n \|f-v\|^2_{L^2(\Omega)} + \|\mathcal{S}_{dis}(f-v)\|^2_n\leq C(\lambda_n+e_{dis})\|f\|^2_{L^2(\Omega)}.
\end{align}
\end{assu}
{In the general discrete approximation of the problem \eqref{para-examp1}}, we will solve the least-squares regularized minimization problem:
\begin{align}\label{disc}
\mathop {\rm min}\limits_{f_{dis}\in V_{dis}} \|\mathcal{S}_{dis}f_{dis}-m\|^2_n+\lambda_n \|f_{dis}\|_{L^2(\Omega)}^2.
\end{align}
The following proposition gives the convergence analysis for the discretized optimization schemes in solving the parabolic inverse source problems, which is proved in \cite{Chen-Zhang2021}.
\begin{prop}[Theorem 3.10 in \cite{Chen-Zhang2021}]\label{thm:4.1}
Let $\{e_i\}^n_{i=1}$ are independent random variables satisfying $\mathbb{E}[e_i]=0$ and $\mathbb{E}[e^2_i]\leq \sigma^2$ for $i=1,\cdots, n$ and $f_{dis}$ be the solution of \eqref{disc}.
Then there exist constants $\lambda_0 > 0$ and $C>0$, such that for any $\lambda_n \leq \lambda_0$ the following estimates hold true:
\begin{align}\label{d2}
\mathbb{E}\big[\|\mathcal{S}f^*- \mathcal{S}_{dis}f_{dis}\|_n^2\big]\le C(\lambda_n+e_{dis})\|f^*\|^2_{L^2(\Omega)}+C\left(1+\frac{e_{dis}}{\lambda_n}+\frac{N_{dis} e_{dis}}{\lambda^{1-d/4}_n}\right)\frac{\sigma^2}{n\lambda_n^{d/4}},
\end{align}
\begin{align}
\mathbb{E}\big[\|f^*- f_{dis}\|_{L^2(\Omega)}^2\big]\le C\frac{\lambda_n+e_{dis}}{\lambda_n}\|f^*\|^2_{L^2(\Omega)}+C\left(1+\frac{e_{dis}}{\lambda_n}+\frac{N_{dis} e_{dis}}{\lambda^{1-d/4}_n}\right)\frac{\sigma^2}{n\lambda_n^{1+d/4}},
\end{align}
and
\begin{align}
\mathbb{E}\big[\|f^*- f_{dis}\|_{H^{-1}(\Omega)}^2\big]&\le C(\lambda^{1/2}_n+e^{1/2}_{dis})\frac{\lambda_n+e_{dis}}{\lambda_n}\|f^*\|^2_{L^2(\Omega)}
\nonumber\\
&+C(\lambda^{1/2}_n+e^{1/2}_{dis})\left(1+\frac{e_{dis}}{\lambda_n}+\frac{N_{dis} e_{dis}}{\lambda^{1-d/4}_n}\right)\frac{\sigma^2}{n\lambda_n^{1+d/4}}.
\end{align}
\end{prop}
Now we study the error analysis of the proposed method in solving the parabolic inverse source problem, where we used the backward Euler scheme to discretize the temporal derivative. Let $u$ be the solution to \eqref{eqn:weak_formulation} and $\hat{u}$ is the solution computed by semi-discrete scheme based on FEM basis functions. {We also assume the POD basis functions are constructed by snapshots of $\hat{u}$ and their temporal finite differences}. The FEM space and POD space are denoted by $V_h$ and $\Psi$, respectively. The fully discrete scheme is constructed on $\Psi$ and the solution is denoted by $U_k$ for $k=1\cdots m$ with $ m=\frac{T}{\Delta t}$. To be precise, we seek $U_k$ such that
{ \begin{equation}\label{eqn:fully_discrete_POD}
(\bar{\partial}U_k,\psi)+a(U_k,\psi)=(f,\psi), \quad \forall \psi\in \Psi.
\end{equation}
}
Now we denote the solution operator from source term $f$ to the final time solution $U_m$ as $\mathcal{S}_{pod}$, such that $\mathcal{S}_{pod}f=U_m$.
\begin{theorem}\label{thm_convergence}
Let $f_{pod}$ be the solution of \eqref{disc}, then there exist constants $\lambda_0 > 0$ and $C>0$, such that for any $\lambda_n \leq \lambda_0$ the following estimates hold true:
\begin{align}\label{pod1}
\mathbb{E}\big[\|\mathcal{S}f^*- \mathcal{S}_{pod}f_{pod}\|_n^2\big]\le C(\lambda_n+e_{pod})\|f^*\|^2_{L^2(\Omega)}+C\left(1+\frac{e_{pod}}{\lambda_n}+\frac{{N_{pod}} e_{pod}}{\lambda^{1-d/4}_n}\right)\frac{\sigma^2}{n\lambda_n^{d/4}},
\end{align}
\begin{align}
\mathbb{E}\big[\|f^*- f_{pod}\|_{L^2(\Omega)}^2\big]\le C\frac{\lambda_n+e_{pod}}{\lambda_n}\|f^*\|^2_{L^2(\Omega)}+C\left(1+\frac{e_{pod}}{\lambda_n}+\frac{{N_{pod}} e_{pod}}{\lambda^{1-d/4}_n}\right)\frac{\sigma^2}{n\lambda_n^{1+d/4}},
\end{align}
and
\begin{align}
\mathbb{E}\big[\|f^*- f_{pod}\|_{H^{-1}(\Omega)}^2\big]&\le C(\lambda^{1/2}_n+e^{1/2}_{pod})\frac{\lambda_n+e_{pod}}{\lambda_n}\|f^*\|^2_{L^2(\Omega)} \nonumber \\
&+C(\lambda^{1/2}_n+e^{1/2}_{pod})\left(1+\frac{e_{pod}}{\lambda_n}+\frac{{N_{pod}} e_{pod}}{\lambda^{1-d/4}_n}\right)\frac{\sigma^2}{n\lambda_n^{1+d/4}},
\end{align}
where
\begin{equation}
e_{pod}=\left(h^2+\Delta t|\ln(\Delta t)|+\sqrt{\frac{T}{\Delta t}\rho}\right)^2.
\end{equation}
\end{theorem}
\begin{proof}
To simplify the notation in the proof, let $\|\cdot\|$ denote the norm in $L^2(\Omega)$, unless otherwise specified. For $k=1,..., m$ , we consider the decomposition as follows:
\begin{align}
U_{k}-\hat{u}\left(t_{k}\right)=U_{k}-\mathcal{P}^{{N_{pod}}} \hat{u}\left(t_{k}\right)+\mathcal{P}^{{N_{pod}}} \hat{u}\left(t_{k}\right)-\hat{u}\left(t_{k}\right)=\vartheta_{k}+\varrho_{k},
\end{align}
where $\vartheta_{k}=U_{k}-\mathcal{P}^{{N_{pod}}} \hat{u}\left(t_{k}\right)$, $\varrho_{k}=\mathcal{P}^{{N_{pod}}} \hat{u}\left(t_{k}\right)-\hat{u}\left(t_{k}\right)$, and $\mathcal{P}^{{N_{pod}}}$ denotes the Ritz-projection to $\Psi$ under the bilinear form $a(\cdot,\cdot)$. The projection error $\varrho_{m}$ can be bounded by \eqref{SnapshotOfSolutions}. In addition, by applying classic result of parabolic equation, i.e., $\| u(\cdot,t_j)\|\leq C\| f\|$, $\forall j=1,\cdots,m$ and $\| u_t(\cdot,t_j)\|\leq C\| f\|$, $\forall j=1,\cdots,m$, we know that,
\begin{align}\label{est:rho_m}
\varrho_{m}\leq C \sqrt{(2m+1)\rho}\| f\|.
\end{align}
Now we only need to estimate the term $\vartheta_{m}$. According to \eqref{eqn:fully_discrete_POD}, we have
\begin{align}\label{eqn:theta_weak_formula}
\begin{aligned}
\left(\bar{\partial} \vartheta_{k}, \psi\right)+a\left(\vartheta_{k}, \psi\right)=&\left(\bar{\partial} U_{k}, \psi\right)+a\left(U_{k}, \psi\right) -\left(\bar{\partial} \mathcal{P}^{{N_{pod}}} \hat{u}\left(t_{k}\right), \psi\right)-a\left(\mathcal{P}^{{N_{pod}}} \hat{u}\left(t_{k}\right), \psi\right) \\
=&\left(f\left(t_{k}\right), \psi\right)-\left(\bar{\partial} \mathcal{P}^{{N_{pod}}} \hat{u}\left(t_{k}\right), \psi\right)-a\left(\hat{u}\left(t_{k}\right), \psi\right) \\
=&\left(v_{k}, \psi\right),
\end{aligned}
\end{align}
where
\begin{align*}
v_{k}=\hat{u}_{t}\left(t_{k}\right)-\bar{\partial} \mathcal{P}^{{N_{pod}}} \hat{u}\left(t_{k}\right)=\hat{u}_{t}\left(t_{k}\right)-\bar{\partial} \hat{u}\left(t_{k}\right)+\bar{\partial} \hat{u}\left(t_{k}\right)-\bar{\partial} P^{{N_{pod}}} \hat{u}\left(t_{k}\right) .
\end{align*}
We define $w_{k}=\hat{u}_{t}\left(t_{k}\right)-\bar{\partial} \hat{u}\left(t_{k}\right)$ and $z_{k}=\bar{\partial} \hat{u}\left(t_{k}\right)-\mathcal{P}^{{N_{pod}}} \bar{\partial} \hat{u}\left(t_{k}\right)$. Substituting $\psi=\vartheta_{k} $ into \eqref{eqn:theta_weak_formula}, we derive that
\begin{align}
\left\|\vartheta_{k}\right\|^{2}-\left(\vartheta_{k}, \vartheta_{k-1}\right)+\Delta t a\left(\vartheta_{k}, \vartheta_{k}\right) \leq \Delta t\left\|v_{k}\right\|\left\|\vartheta_{k}\right\|
\end{align}
and therefore,
\begin{align}
\left\|\vartheta_{k}\right\| \leq \frac{1}{1+\frac{a_{min}}{\alpha} \Delta t}\left(\left\|\vartheta_{k-1}\right\|+\Delta t\left\|v_{k}\right\|\right),
\end{align}
where $a_{min}$ is the minimum of the coefficient $a(x)$ in the elliptic operator and $\alpha$ is a constant such that
\begin{align}
\|\phi\|^2_{L_2}\leq \alpha \|\phi\|_{H^1} \quad \forall \phi\in V_h.
\end{align}
Taking $\gamma=\frac{a_{min}}{\alpha}$, we have
\begin{align}\label{eqn:est_vartheta1}
\left\|\vartheta_{m}\right\| \leq\left(\frac{1}{1+\gamma \Delta t}\right)^{m}\left\|\vartheta_{0}\right\|+\Delta t \sum_{j=1}^{m}\left(\frac{1}{1+\gamma \Delta t}\right)^{m-j+1}\left\|v_{j}\right\|.
\end{align}
As $\| v_j\|\leq\| w_k\|+\| z_k\|$, so \eqref{eqn:est_vartheta1} is taken into two parts.
\begin{align}
\Delta t \sum_{j=1}^{m}\xi^{m-j+1}\left\|z_{j}\right\|\leq&\Delta t \sqrt{(\sum_{j=1}^m\xi^{2(m-j+1)})(\sum_{j=1}^m\| z_k\|^2)},
\end{align}
where $\xi$ denotes $\frac{1}{1+\gamma\Delta t}$. Using the facts that $\left(\frac{1}{\xi}\right)^{2 m}=(1+\gamma \Delta t)^{2 m}=\left(1+\frac{2 \gamma T}{2 m}\right)^{2 m} \leq e^{2 \gamma T}$ and $\xi^{-2}-1=(1+\gamma\Delta t)^2-1\geq 2\gamma\Delta t$, we can obtain that
\begin{align}
\Delta t \sqrt{(\sum_{j=1}^m\xi^{2(m-j+1)})(\sum_{j=1}^m\| z_k\|^2)}=&T\sqrt{(\frac{1}{m}\sum_{j=1}^m\xi^{2(m-j+1)})(\frac{1}{m}\sum_{j=1}^m\| z_j\|^2)}\nonumber\\
\leq&T\sqrt{(\frac{1}{m}\frac{1-\xi^{2m}}{\xi^{-2}-1})(\frac{1}{m}\sum_{j=1}^m\| z_j\|^2)}\nonumber\\
\leq&T\sqrt{\frac{1-e^{-2\gamma T}}{2\gamma T}(\frac{1}{m}\sum_{j=1}^m\| z_j\|^2)}.
\end{align}
By Corollary 4 in \cite{kunisch2001galerkin}, we have that $ \frac{1}{m}\sum_{j=1}^m\| z_j\|^2\leq C\rho\| f\|^2$, which implies that
\begin{align}\label{est:zj}
\Delta t \sum_{j=1}^{m}\xi^{m-j+1}\left\|z_{j}\right\|\leq C(T)\sqrt{\rho}\| f\|.
\end{align}
Now we turn to estimate the term $\Delta t \sum_{j=1}^{m}\xi^{m-j+1}\left\|w_{j}\right\|$. By definition,
\begin{align}
\begin{aligned}
w_{j} &=u_{t}\left(t_{j}\right)-\frac{u\left(t_{j}\right)-u\left(t_{j-1}\right)}{\Delta t} \\
&=\frac{1}{\Delta t}\left(\Delta t u_{t}\left(t_{j}\right)-\left(u\left(t_{j}\right)-u\left(t_{j-1}\right)\right)\right) \\
&=\frac{1}{\Delta t} \int_{t_{j-1}}^{t_{j}}\left(s-t_{j-1}\right) u_{t t}(s) d s.
\end{aligned}
\end{align}
By the analytic semigroup theory \cite{renardy2006introduction} or Lemma 3.6 in \cite{Chen-Zhang2021}, we have the following regularity estimate,
\begin{align}
\| u_{tt}(s)\|\leq\frac{C}{s}\| f\|,
\end{align}
which implies,
\begin{align}
\| w_j\|\leq\frac{1}{\Delta t} \int_{t_{k-1}}^{t_{j}}\left(s-t_{k-1}\right)\frac{1}{s}\| f\| ds.
\end{align}
When $j=1$,
\begin{align}
\| w_1\|\leq\frac{1}{\Delta t} \int_{0}^{\Delta t}\| f\| ds=\| f\|,
\end{align}
and when $j\geq 2$,
\begin{align}
\| w_k\|\leq \frac{1}{\Delta t} \int_{t_{j-1}}^{t_{j}}\left(s-t_{j-1}\right)\frac{1}{t_{j-1}}\| f \| ds = \frac{1}{2}\| f\|\frac{1}{(j-1)}\leq\frac{\| f \|}{j}.
\end{align}
Therefore, we obtain that
\begin{align}
\Delta t \sum_{j=1}^{m}\xi^{m-j+1}\left\|w_{j}\right\| \leq \Delta t \sum_{j=1}^{m}\xi^{m-j+1} \frac{\| f\|}{j}
\leq \Delta t \sum_{j=1}^{m} \frac{\| f\|}{j} \leq C\Delta t |\ln(\Delta t)| \| f \|\label{est:wj}.
\end{align}
Summing up \eqref{est:zj}, \eqref{est:wj} and using the classic parabolic FEM theory (e.g. Theorem 1.2 in \cite{thomee1990finite}), we have that
\begin{align}\label{est:parabolic_FEMsemi}
\| \hat{u}\left(t_{m}\right)-u\left(t_{m}\right)\|\leq Ch^2\| f \|.
\end{align}
Therefore, we can prove that
\begin{align}
\| U_m-u(T)\|\leq C\left(h^2+\Delta t|\ln(\Delta t)|+\sqrt{\frac{T}{\Delta t}\rho}\right)\| f \|.
\end{align}
Finally, we apply the Proposition \ref{thm:4.1} by taking $\mathcal{S}_{pod}$ as $\mathcal{S}_{dis}$ and $f_{pod}$ as $f_{dis}$, and prove the estimates in the Theorem \ref{thm_convergence}. Clearly, $e_{pod}=\left(h^2+\Delta t|\ln(\Delta t)|+\sqrt{\frac{T}{\Delta t}\rho}\right)^2$.
\end{proof}
\begin{rem}
Theorem \ref{thm_convergence} suggests that the optimal smoothing parameter in Eq.\eqref{opti-para} may still apply. In particular, if $e_{pod}\le C\lambda_n$ and ${N_{pod}} e_{pod}\leq C\lambda_n^{1-1/\alpha}$, we have
\begin{align}\label{pod2}
\mathbb{E}\big[\|\mathcal{S}f^*- \mathcal{S}_{pod}f_{pod}\|_n^2\big]\le C\lambda_n\|f^*\|_{L^2(\Omega)}^2+\frac{C\sigma^2}{n\lambda_n^{d/4}},
\end{align}
\begin{align}
\mathbb{E}\big[\|f^*- f_{pod}\|_{L^2(\Omega)}^2\big]\le C\|f^*\|_{L^2(\Omega)}^2+\frac{C\sigma^2}{n\lambda_n^{1+d/4}},
\end{align}
and
\begin{align}
\mathbb{E}\big[\|f^*- f_{pod}\|_{H^{-1}(\Omega)}^2\big]\le C\lambda^{1/2}_n\|f^*\|_{L^2(\Omega)}^2+\frac{C\sigma^2}{n\lambda_n^{1/2+d/4}}.
\end{align}
Note that $\rho$ depends on the number of POD basis functions, ${N_{pod}}$, and quickly decreases to zero when increasing ${N_{pod}}$. Assumptions above are achievable by taking $h$ and $\Delta t$ to be small and ${N_{pod}}$ to be small with respect to the DOF of the FEM.
\end{rem}
\section{Numerical examples}\label{num-sec}
\noindent
In this section, we present numerical examples to investigate the performance of the proposed POD method in solving parabolic inverse source problems. Two types of source functions, i.e., letters and circles at different locations, will be studied. In the example of letters, we verify the convergence introduced in Theorem \ref{thm_convergence} and apply the iterative method to find optimal smoothing parameters. In the second example, we mainly study the approximation property of our POD basis functions.
\subsection{Recovering Letters}\label{sec:eg1}
\paragraph{General Settings} For the first example, we apply the POD method to recover the source term $f$ in form of discontinuous patterns. To be precise, $f$ is an indicator function, with the maximum value $1$, of a set in $[0,1]^2$ domain whose shape is decided by a capital letter. For each observation data, we first apply the FEM with $h=1/32$ and time step $\Delta t=1/32$ to get the solution of the forward problem with exact source term $f$ at final time $T=1$. We interpolate the FEM solution to locations of $100\times 100$ evenly distributed sensor and add a Gaussian unbiased noise with stand derivation $10^{-3}$ to each observation. It corresponds to a $10\%$ noise level. Initial value $g=0$ in \eqref{zz0}.
To construct the POD basis functions, we generate the snapshots source term decided by letters from $A$ to $O$. The snapshots are generated by solving the forward problem with each source term via FEM (space mesh size $h=1/32$, time step $\Delta t=1/32$). Under such settings, $21$ POD basis functions are constructed to resolve the snapshots of the forward problem with $10^{-4}$ accuracy. In other words, the POD basis functions cover $1-10^{-4}$ of eigenspace from the snapshots with respect to the $L^2$ norm. To solve the inverse problem, $\lambda$ is set to be $10^{-6}$ and we will discuss how to find the optimal $\lambda$ later.
\paragraph{Computational Cost} In Figure \ref{fig:eg1_multi_letters}, we show some of the letters recovered. The computational cost of the POD method includes two parts. First, we will construct the POD basis functions for all the $15$ letters, which will take about $40.3$ seconds on a desktop computer (3.1GHz, 6-core, i5, Matlab). Then, for each letter, it will take about $164$ seconds to solve the optimization problem via gradient descend (relative $L_2$ threshold $10^{-5}$) using the POD basis functions. As a comparison, with the FEM basis functions, it takes $1328.4$ seconds to recover similar patterns with the same configuration during optimization. In the case of recovering one single letter, the POD method has increased the speed by $6.5\times$, and in the case of recovering all $15$ letters, it achieves about $8\times$ speed-up.
\begin{figure}[htb]
\centering
\subfigure[A]{\includegraphics[width=0.32\linewidth]{figures_POD/letters_A_n32.eps}}
\subfigure[C]{\includegraphics[width=0.32\linewidth]{figures_POD/letters_C_n32.eps}}
\subfigure[G]{\includegraphics[width=0.32\linewidth]{figures_POD/letters_G_n32.eps}}
\subfigure[reference A]{\includegraphics[width=0.32\linewidth]{figures_POD/ref_A.eps}} \subfigure[D]{\includegraphics[width=0.32\linewidth]{figures_POD/letters_D_n32.eps}} \subfigure[O]{\includegraphics[width=0.32\linewidth]{figures_POD/letters_O_n32.eps}}
\caption{Recovering letters based on basis constructed from 'A' to 'O'}
\label{fig:eg1_multi_letters}
\end{figure}
We also compare the POD method with the FEM in the case when $h=1/20$. We construct the POD basis from snapshots of the forward problem with $h=1/20$. $ 23 $ basis functions are constructed to achieve $10^{-4}$ accuracy. As the DOF during the optimization procedure stays near, the computational cost with POD basis functions is similar between different $h$'s (i.e., $ 47 $ seconds in constructing basis functions and $ 159 $ seconds in solving optimization problems). While with the FEM basis functions, the computation cost drops near quadratically against the DOF. It takes $ 340 $ seconds for $h=1/20$ case, which corresponds to $400$ DOF. In Figure \ref{fig:eg1_A_different_h} we show the comparison between the FEM and POD method with different $h$'s. The observation noise is the same under the same $h$. Generally speaking, the restoration by the POD method is less affected by the noise than the FEM. This is due to the fact that the noise is perpendicular to the space of the POD basis functions. For a convincing outcome, the POD method is significantly faster than the FEM.
\begin{figure}[htb]
\centering
\subfigure[POD, $h=1/20$]{\includegraphics[width=0.45\linewidth]{figures_POD/letters_A_n20.eps}}
\subfigure[FEM, $h=1/20$]{\includegraphics[width=0.45\linewidth]{figures_POD/letters_A_FEM_n20.eps}}\\
\subfigure[POD, $h=1/32$]{\includegraphics[width=0.45\linewidth]{figures_POD/letters_A_n32.eps}}
\subfigure[FEM, $h=1/32$]{\includegraphics[width=0.45\linewidth]{figures_POD/letters_A_FEM_n32.eps}}
\caption{Recovering letter 'A' with different basis and $h$}
\label{fig:eg1_A_different_h}
\end{figure}
\paragraph{Error Investigation} According to the convergence analysis studied in Theorem \ref{thm_convergence}, there are three factors that affect the accuracy of our algorithm, i.e., mesh size of FEM basis, regularization factor, and approximation property of the POD basis functions. As such, we set up three experiments when recovering the letter 'A'. Under each experiment, we generate $100$ set of observation $y$'s and solve the optimization with the pre-computed POD basis functions. The standard derivation of the noise is still $10^{-3}$. It corresponds to a $10\%$ noise level. We compute the mean of the error in each set of parameters and show them in Figure \ref{fig:eg1_err_analysis}.
\subparagraph{Dependence on the mesh size} In Figure \ref{fig:eg1_fem_err} we verify that given $\lambda$ when the mesh size $h$ of the FEM basis functions goes down, the error of $Sf$ will first decrease and then converged to a limit that is greater than zero. In the first stage, the order of convergence against mesh size $h$ is $2$. The limit in the second stage depends on $\lambda$ and the dependence is non-monotone. The number of POD basis functions does not change remarkably (also see in the previous example that compares the POD method with FEM) under different mesh sizes. Along with the monotonic decreasing of the error against $h\to 0$, it indicates all sets of POD basis functions all approximate some fixed continuous function space.
\subparagraph{Dependence on the regularization factor} To further investigate the limiting error, in \ref{fig:eg1_lambda_err}, we keep $h=1/32$ fixed and use different regularization factor $\lambda$'s. We find that the optimal $\lambda$ for $h=1/32$ with the $10\%$ noise level is between $10^{-7}$ and $10^{-8}$. The error is a V-shape function near the optimal $\lambda$.
\subparagraph{Dependence on the number of POD basis functions} The last experiment is to study how the number of POD basis functions affects the error. In \ref{fig:eg1_pod_err}, we show the mean square error with $100$ sets of random generated observation. The error goes down as we increase the number of POD basis functions. The fitted slope in the loglog plot is $-0.0145$, which indicates the errors exponentially decay with a coefficient $-0.0145$. The p-value for the fitted slope is smaller than $0.001$.
\begin{figure}[htb]
\centering
\subfigure[Mesh size $h$ of the FEM.]{\includegraphics[width=0.32\linewidth]{figures_POD/err_against_FEMh.eps}\label{fig:eg1_fem_err}}
\subfigure[regularization factor $\lambda$]{\includegraphics[width=0.32\linewidth]{figures_POD/err_against_lambda.eps}\label{fig:eg1_lambda_err}}
\subfigure[POD basis number $n$, fitted slope: $1.45\time10^{-2}$]{\includegraphics[width=0.32\linewidth]{figures_POD/err_against_PODn.eps}\label{fig:eg1_pod_err}}
\caption{Error dependence on different parameters}
\label{fig:eg1_err_analysis}
\end{figure}
\paragraph{Iterative algorithm for finding the optimal $\lambda$} As analyzed in Eq.\eqref{opti-para}, the optimal $\lambda$ relies on $\|f^*\|^2_{L^2(\Omega)}$. While in the setting of inverse problems, the norm of input $f^*$ is unknown. A natural choice is to find the optimal $\lambda$ by using a fixed-point iteration. To be precise, we start from an initial guess like $\lambda_{n, 0}=n^{-4 /(d+4)}$ and solve for $f_0$ with the POD basis functions. After solving $f_j$ the $j$-th step, we update $\lambda$ as,
\begin{align}\label{eqn:opt_lam_alg}
\lambda_{n, j+1}^{1 / 2+d / 8}=n^{-1 / 2}\left\|S_{\tau, h} f_{j}-m\right\|_{n}\left\|f_{j}\right\|_{L^{2}(\Omega)}^{-1}.
\end{align}
We will stop the iteration of $\lambda$ until $\left\|f_{j}\right\|_{L^{2}(\Omega)}$ converges. As shown in Figure \ref{fig:eg1_opt} such iterative algorithms, with either POD or FEM basis, finds the correct scale of $\lambda$ with a similar pattern. It implies that the POD method captures the whole solution subspace with given known source function and is robust in iteration.
To be noted, in Eq.\eqref{opti-para}, constant $C$ is not $1$ so the fixed-point iteration Eq.\eqref{eqn:opt_lam_alg} only finds the correct scale of $\lambda$ comparing with Figure \ref{fig:eg1_lambda_err}, in which the error due to $\lambda$ is computed with the knowledge of the ground truth solution $u(\cdot, T)$.
\begin{figure}[h]
\centering
\subfigure[by iterating on FEM basis]{\includegraphics[width=0.32\linewidth]{figures_POD/opt_lam_fem_n32.eps}}
\subfigure[by iteration on POD basis]{\includegraphics[width=0.32\linewidth]{figures_POD/opt_lam_pod_n32.eps}}
\subfigure[$\lambda_n$ and $L_2$ error after each iteration]{\includegraphics[width=0.32\linewidth]{figures_POD/opt_lam_iter_n32_combine.eps}}
\caption{Finding optimal $\lambda$ }
\label{fig:eg1_opt}
\end{figure}
\subsection{Recovering Moving Circles}
In addition to letter examples, we also applied the POD method to recover a ball pattern that moves on some given trajectories in two-dimensional space. Similar to recovering letters, we construct the POD basis functions from snapshots that are generated by solving forward problems with classic FEM, where the mesh size $h=1/32$ and time step $\Delta t=1/32$. $\lambda$ is set to be $10^{-6}$ and $\sigma=2.75\times 10^{-3}$ which corresponds to a $10\%$ noise level.
\paragraph{Moving along a horizontal line} In such case, the source term $f$ is given by indicator function of set $\Omega_s=\{x\in[0,1]^2|(x_1-s)^2+(x_2-0.5)^2\leq 0.1^2\} $. When constructing the POD basis functions, we use $17$ values of $s$ that are equally distributed between interval $[0.15,0.85]$ including two terminal points. $14$ basis functions are extracted for $10^{-4}$ accuracy calculated by eigenvalues. In \ref{fig:eg2_moving_circles} we use the constructed basis on detecting circles with the same size. We find that the performance of the POD method is still acceptable even the centers of the circle do not coincide with the ones when we construct the POD basis functions.
\paragraph{Moving along a ring} In addition, as shown in \ref{fig:eg2_moving_circles_2} we recover the same pattern whose center are on a circle. To be specific, the source term $f$ is given by indicator function of set $\Omega_\theta=\{x\in[0,1]^2|(x_1-0.5-\cos(\theta)/4)^2+(x_2-0.5-\sin(\theta)/4)^2\leq 0.1^2\}$. $33$ basis are extracted for $10^{-4}$ accuracy. Even with a higher DOF, the POD method still yields $10\times$ acceleration comparing with FEM ($60$ seconds vs $615$ seconds).
\begin{figure}[h]
\centering
\subfigure[$s=0.3$]{\includegraphics[width=0.19\linewidth]{figures_POD/moving_squares_center_3e-1.eps}}
\subfigure[$s=0.4$]{\includegraphics[width=0.19\linewidth]{figures_POD/moving_squares_center_4e-1.eps}}
\subfigure[$s=0.5$]{\includegraphics[width=0.19\linewidth]{figures_POD/moving_squares_center_5e-1.eps}}
\subfigure[$s=0.6$]{\includegraphics[width=0.19\linewidth]{figures_POD/moving_squares_center_6e-1.eps}}
\subfigure[$s=0.7$]{\includegraphics[width=0.19\linewidth]{figures_POD/moving_squares_center_7e-1.eps}}
\caption{Recovering Moving Circles with centers point at different $x$, red circle: ground truth}
\label{fig:eg2_moving_circles}
\end{figure}
\begin{figure}[h]
\centering
\subfigure[$\theta=0$]{\includegraphics[width=0.24\linewidth]{figures_POD/moving_squares_2_1.eps}}
\subfigure[$\theta=\frac{\pi}{4}$]{\includegraphics[width=0.24\linewidth]{figures_POD/moving_squares_2_2.eps}}
\subfigure[$\theta=\frac{\pi}{2}$]{\includegraphics[width=0.24\linewidth]{figures_POD/moving_squares_2_3.eps}}
\subfigure[$\theta=\frac{3\pi}{4}$]{\includegraphics[width=0.24\linewidth]{figures_POD/moving_squares_2_4.eps}} \\
\subfigure[$\theta=\pi$]{\includegraphics[width=0.24\linewidth]{figures_POD/moving_squares_2_5.eps}}
\subfigure[$\theta=\frac{5\pi}{4}$]{\includegraphics[width=0.24\linewidth]{figures_POD/moving_squares_2_6.eps}}
\subfigure[$\theta=\frac{3\pi}{2}$]{\includegraphics[width=0.24\linewidth]{figures_POD/moving_squares_2_7.eps}}
\subfigure[$\theta=\frac{7\pi}{4}$]{\includegraphics[width=0.24\linewidth]{figures_POD/moving_squares_2_8.eps}}
\caption{Recovering Moving Circles with centers point at different $x$, red circle: ground truth}
\label{fig:eg2_moving_circles_2}
\end{figure}
\section{Conclusion}\label{Sec:conclusion}
\noindent
We have developed a data-driven and model-based approach for solving parabolic inverse source problems. The key idea is to exploit (model-based) and construct (data-based) the intrinsic approximate low-dimensional structure of the underlying parabolic PDEs that consists of two components -- a training component that constructs a small number of POD basis functions to achieve significant dimension reduction in the solution space, and a fast algorithm that computes the optimization problem in inverse source problem by using the constructed POD basis functions. Hence, we achieve an effective data and model-based approach for the inverse source problems and overcome the typical computational bottleneck of FEM in solving such problems.
Under a weak assumption on the regularity of the solution, we provide the convergence analysis of our POD algorithm in solving the forward parabolic PDEs and thus obtain the error estimate of the POD algorithm for the parabolic inverse source problems. Finally, we carried out numerical experiments to demonstrate the accuracy and efficiency of the proposed method. Through numerical results, we found that our POD algorithm yields as good approximations as the reference solution obtained by the FEM. But our algorithm can be much cheaper. We expect an even better performance of efficiency can be obtained in 3D problems, which is one of our future topics. We also studied other issues of the POD algorithm, such as the dependence of the error on the mesh size, on the regularization parameter in the least-squares regularized minimization problems, and on the number of POD basis functions.
\section*{Acknowledgement}
\noindent
The research of W. Zhang is supported by the National Natural Science Foundation of China No. 11901282 and the Shenzhen Sci-Tech Fund No. RCBS20200714114941241. The research of Z. Zhang is supported by Hong Kong RGC grant projects 17300318 and 17307921, National Natural Science Foundation of China No. 12171406, Seed Funding Programme for Basic Research (HKU), and Basic Research Programme (JCYJ20180307151603959) of The Science, Technology and Innovation Commission of Shenzhen Municipality.
\bibliographystyle{plain}
|
\section{Introduction}
Understanding major phases of galaxy growth and chemical enrichment and how young galaxies have contributed to cosmic reionization are two of the most pressing questions in modern astrophysics. Current observations of high-redshift galaxies and model predictions highlight the role of low-mass starbursting galaxies as important contributors to the cosmic star formation history \citep[e.g.][]{atek_hubble_2014, maseda_muse_2018} and the reionization process \citep[e.g.][]{stark_galaxies_2016,endsley_o_2021}. These small, young systems are typically identified as extreme emission-line galaxies (EELGs) by their strong nebular emission and hard radiation fields, evidenced by the presence of high-ionization emission lines with large H$\beta$+[OIII]5007$\AA$ equivalent widths
\citep[EW $\sim$300-3000\AA][]{smit_evidence_2014,roberts-borsani_zgtrsim_2016,stark_ly_2017,barros_greats_2019}.
At lower redshift, an increasingly large number of low-mass EELGs have been discovered from $z\sim$0.5 to $z$\,$\sim$\,4 and deep spectroscopic data have allowed a global characterization of their physical properties, which are now relatively well-established
\citep[e.g][]{wel_extreme_2011,maseda_nature_2014,amorin_evidence_2014,amorin_extreme_2015,amorin_analogues_2017,calabro_characterization_2017, forrest_discovery_2017, tang_mmtmmirs_2019, du_searching_2020,tran_mosel_2020, endsley_o_2021}. A large fraction of EELGs show high EW emission lines when observed in the rest-UV, including strong Ly$\alpha$ emission \citep[e.g.][]{erb_physical_2010,amorin_analogues_2017,berg_window_2018,tang_lyman-alpha_2021} and some of them have been detected as LyC emitters \citep[e.g.][]{barros_extreme_2016,vanzella_hubble_2016,vanzella_direct_2018}. However, a thorough characterization on these systems star formation histories and chemical abundances is still lacking. This is due to the high sensitivity, resolution and coverage spectra requirements. These observations are still challenging in faint and compact objects -particularly in the deep NIR spectroscopy
A technical solution to the previous constrains will be available with the new generation telescopes. These include the James Webb Space Telescope (JWST) and the ground-based extremely large telescopes. However, a rare family of EELGs in the local universe can still provide critical insight on the rapid evolution of high redshift galaxies. This is the case of the so-called “Green Pea'' galaxies at $z\sim$0.14-0.36 \citep[GPs, ][]{cardamone_galaxy_2009}, which are low mass (M$_{*}$ < 10$^{10}$ M$_{\odot}$), highly star-forming (SFR$\sim10-60 \,M_{\odot}/yr)$) systems characterized by their unresolved and green colour appearance in Sloan Digital Sky Survey (SDSS, REF) color composite images. The GPs are characterized by their small sizes of about few kpc \citep{amorin_star_2012} and extremely compact star formation, with resolved UV effective radius ($r_{50}$) typically below 1 kpc \citep{yang_ly_2017}. Their green colour is due to their unusually strong [OIII]5007 emission, with equivalent widths EW$\sim$200-2500\AA, which is redshifted to the SDSS $r’$-band and dominates the optical emission of the galaxies. Such a strong nebular emission is characterized by its high excitation, low dust extinction and low metallicity of about 20\% solar (12$+$log(O/H)$\sim$\,8.1) in average \citep{amorin_oxygen_2010, amorin_star_2012, perez-montero_extreme_2021}. The dominant nebular emission of GPs shows high ionization conditions, which are traced by their high O32 ($\equiv[OIII]5007\AA/[OII]3727\AA$) ratios and strong emission from high-ionization lines such as $HeII\,4686\AA$ \citep{jaskot_origin_2013,ramambason_reconciling_2020}. These properties are naturally explained by galaxies experiencing a maximal ongoing starburst with young stellar ages, thus showing extremely high specific SFR (log(sSFR)$\sim$10$^{-8}$yr$^{-1}$).
Observations of GPs with the Cosmic Origins Spectrograph (COS) mounted at the Hubble Space Telescope (HST) have allowed the detection of significant Lyman-$\alpha$ emission with escape fractions of a few to about 100\% \citep{henry_lyupalphaemission_2015} and Lyman continuum (LyC) leakage \citep{izotov_eight_2016}, including objects showing cosmologically-relevant escape fractions that can exceed $\sim$20\% \citep[e.g.][]{izotov_j0811+4730:_2018}. Thus, these objects may provide key insight into the production and escape of ionizing photons under ISM conditions approaching those found in high-redshift systems \citep[e.g.][]{schaerer_ionizing_2016,izotov_lyman_2021}.
More generally, the GPs are excellent laboratories to study massive star formation and feedback effects at low metallicity. \cite{amorin_star_2012} (hereafter \citetalias{amorin_star_2012}) provided a detailed analysis on three GPs: GP004054, GP113303 and GP232539 with $z\sim0.283$, $0.241$ and $0.277$. The sample was observed using the long-slit spectrograph OSIRIS at the 10.4m telescope GTC (Gran Telescopio de Canarias). This set-up represented a milestone on GPs studies, which had been mainly accomplished from the SDSS survey data. The chemical analysis, via the direct method, showed remarkable high N/O ratios, confirming previous results by \citet{amorin_oxygen_2010} leading to the conclusion that more than 50\% of the GP sample had atypically high N/O ratios. The authors concluded that a massive inflow of metal-poor gas, as proposed in \citet{amorin_oxygen_2010}, could explain the high N/O at low O/H, and the unusually compact and strong star formation in GPs.
Similar results were later discussed by subsequent works. Some authors arrived to similar conclusions \citep{kojima_evolution_2017, loaiza-agudelo_vltx-shooter_2020} and some others argued that the high N/O may be caused, instead, by composite nebulae with different physical properties and that the presence of large amounts of Wolf-Rayet (WR) stars may favor a localized N/O enhancement due to their strong winds \citep{pilyugin_abundance_2012, hawley_abundances_2012}. Indeed, deep spectra of some high N/O GPs display prominent blue bumps suggesting a large number of WR stars within the 800 - 1200 range \citep[e.g.][]{amorin_star_2012}.
\citet{perez-montero_are_2013} performed a spatial analysis on six WR galaxies data cubes from the PMAS (Potsdam Multi-Aperture Spectrophotometer) instrument. In four of the galaxies, the direct method results agreed with a non-uniform enhancement from WR feedback. These theoretical yields were calculated using the models from \citet{molla_modelling_2012}. The authors concluded that hydrodynamical effects (i.e., metal-poor gas inflows) are excluded as the most frequent cause of N/O excess. A comparison with model predictions led \citetalias{amorin_star_2012} to reach similar conclusions for the above three GPs. Still, later works continue to study the strong winds and possible enriched outflows from high-resolution spectra. Example of these studies on small GP samples include \citep{amorin_complex_2012,bosch_integral_2019,hogarth_chemodynamics_2020}.
Finally, the SFH of GPs still remain largely unconstrained. It was first studied by \citetalias{amorin_star_2012}, who applied several techniques using high S/N spectra and surface brightness profiles. Two of them were based on the Simple Stellar Populations (SSPs) synthesis and one in the fitting of evolutionary synthesis models. These results led the authors to conclude that these objects were producing around $\sim4-20\%$ of their stellar mass in the current star forming burst. The ages of the evolved stellar populations resulted in a few Gyr, suggesting the presence of an old stellar host. An important achievement in this analysis was the inclusion of the nebular continuum in the analysis. In most cases, the nebular continuum can be neglected since its brightness is exceptionally low compared to the stellar one. However, in young star forming regions $(\sim10Myr)$ the gas component enhances the galaxy luminosity and tends to make the Spectral Energy Distribution (SED) redder than expected. In the \citetalias{amorin_star_2012} GPs sample, the stellar to nebular continuum percentage was between 11 to 35\% at $H\alpha$, consistent with the high $H\beta$ EWs of their star-forming regions. A recent work by \citet{clarke_old_2020} found results consistent with the above picture using HST imaging free from strong line emission. Analysing the observations with SSP models, the authors concluded that GPs have a diffuse component with colours consistent with an evolved underlying galaxy or an extended nebular continuum component.
\begin{figure*}
\includegraphics[width=1.00\textwidth]{images/sample_spectra_b.png}
\caption{\label{fig:sample_spectra}GP030321, GP101157 and GP121903 galaxy spectra. The lower panels are displayed in logarithmic scale, where the ordinate axis range is constrained for a better focus on the continuum features. The images on the upper-right corners belong to the SDSS sky server. Some emission lines have been labeled using the \textsc{lineidplot} library by \citet{nair_lineid_2016}}
\end{figure*}
Despite the relevance of the GPs in the bigger context of galaxy evolution and reionization and the relatively poor understanding of their own nature, high S/N optical spectra enabling detailed studies of the GPs chemical abundances and star formation histories are available.
In the current manuscript, we build on the work by \citetalias{amorin_star_2012} by adding three new GPs to the sample. These galaxies are also observed at the GTC, with the same instrument setup but they have a lower redshift range with $\bar{z}=0.17$. Both the stellar and gas components are analysed in the current work using state-of-the-art algorithms. To begin with, both the current and \citetalias{amorin_star_2012} samples continua are treated using two techniques: In the first one, the nebular continua are calculated from first principles with a posterior SSPs synthesis in an iterative process, which considers the physical conditions measured from the galaxy emission. Secondly, the algorithm \textsc{FADO} by \citet{gomes_fitting_2017} is applied with a large stellar spectra grid. This algorithm takes into consideration the gas component making it ideal for the analysis of young star forming regions. In the gas emission analysis, the chemical composition is measured using a 14-dimensional direct method model. All these parameters are fit simultaneously using a neural networks sampler under a Bayesian inference scheme. This model is an upgrade from the algorithm presented in \citet{fernandez_bayesian_2019} for the analysis of GPs analogues in the nearby universe. Finally, using the methodology by \citet{perez-montero_photon_2020}, we compare the ionised gas emission against photoionization models to derive some physical properties of the ionising sources in these galaxies. This algorithm has been enhanced using neural networks for a better sampling of the photoionization grids. Finally, we combine both techniques in a single algorithm, which to the best author's knowledge, simultaneously explores the direct method parameter space alongside photoionization grids for the first time.
\section{Galaxy sample data} \label{data}
\begin{table*}
\caption{\label{tab:Sample-properties.} Summary log of observations}
\input{tables/night_log}
\raggedright{\footnotesize{$^1$ The science targets were observed with a 0.8"$\times$7.4' slit width. $^2$ The standard stars were observed with a 2.5"$\times$7.4' slit width.}}\\
\end{table*}
The sample consists in three galaxies selected among the brightest GPs in the sample presented by \cite{cardamone_galaxy_2009}. Their SDSS references are:
\href{http://skyserver.sdss.org/dr16/en/tools/explore/summary.aspx?plate=458&mjd=51929&fiber=185}{SDSSJ030321.41-075923.2}, \href{http://skyserver.sdss.org/dr16/en/tools/explore/summary.aspx?plate=1745&mjd=53061&fiber=463}{SDSSJ101157.08+130822.0} and \href{http://skyserver.sdss.org/dr16/en/tools/explore/summary.aspx?plate=1766&mjd=53468&fiber=557}{SDSSJ121903.98+152608.5}. The first GP can also be found in the SHOC catalogue (SDSS HII-galaxies with Oxygen abundances Catalog) by \citet{kniazev_strong_2004} as SHOC148. For simplicity's sake, however, the GPs will be referred within this manuscript as GP030321, GP101157 and GP121903.
Observations were conducted in service mode at the Gran Telescopio de Canarias (GTC) in Roque de los muchachos, La Palma, Spain. with the Optical System for Imaging and low-Intermediate-Resolution Integrated Spectroscopy (OSIRIS) under program GTC63-10B (PI: Amor\'in) with telescope time granted in 09/2010, 10/2010, 12/2010. We used the R1000B and R1000R grisms with a resolving power R$\approx$1018 at $\lambda_{c,\,Blue}\approx5510\AA$ and R$\approx$R1122 at $\lambda_{c,\,Red}\approx7510\AA$. The wavelength ranges for these grisms are $3300-7500\AA$ and $5100-10250\AA$, respectively. In the case of GP030321, the red arm displayed contamination from the the second order light from the blue side. Due to this and severe telluric contamination, we decided to crop the spectra redder than $\lambda > 7600\AA$. We choose a slit width of 0.8" resulting in a full width at half maximum (FWHM) around 3 pixels. The observations were conducted under average seeing conditions of 1.00". Three exposures of $3\times1200s$ per grating and object, were obtained. Standard calibrations (bias, dome flats) exposures and HgAr+Xe+Ne lamp arcs were obtained after each science exposure. Two spectrophotometric stars and arcs for flux calibration were also obtained. A summary log of observations can be found in Table \ref{tab:Sample-properties.}
We use \textsc{IRAF}\footnote{\textsc{iraf}: The image Reduction and Analysis Facility is distributed by the National Optical Astronomy Observatory, which is operated by the Association of Universities for Research in astronomy (AURA), Inc., under the cooperative agreement with the National Science Foundation (\href{http://iraf.noao.edu/}{http://iraf.noao.edu/})} to reduce the OSIRIS spectra following \citetalias{amorin_star_2012}. In brief, after standard calibrations (bias+overscan subtraction, flat fielding and cosmic-ray removal) we perform the wavelength and flux calibration. The final spectra wavelength array has an accuracy of $0.1\AA$ measured from bright sky lines. Flux calibration, instead, has large uncertainty at $\lambda < 3500\AA$ due to low signal at the CCD edge. This blue edge of the spectra was not considered in the subsequent analysis. Finally, the flux calibration resulted in an uniform offset between the blue and red arms. In the three galaxies, the overlapping continuum flux sigma is of the same order. This was corrected by applying a constant offset obtained after scaling the two arms using the overlapping regions between $\sim 5100\AA$ and $6000\AA$. Additionally, emission lines were normalized using a recombination flux within their arm wavelength range (see section \ref{sec:extinction_calc}).
\begin{table*}
\caption{\label{tab:emission_fluxes_gp03032} Emission line fluxes for the GPs sample. Column (1) Wavelengths in $\AA$ at rest frame. The labels with a "g" subscript identify emission lines whose measurement consisted in a Gaussian profile deblending. Column (2) Reddening curve. Columns (3-4), (5-6), (7-8) display the observed ($F(\lambda)$) and extinction corrected fluxes relative to $F(H\beta) = I(H\beta) = 1000$. At the bottoms of the table the logarithmic extinction coefficient, the equivalent width and the observed flux for $H\beta$ are provided}
\centering{\input{tables/sample_emission_lines}}
\end{table*}
\section{Analysis of the spectra} \label{treatment}
In Fig. \ref{fig:sample_spectra}, we present the OSIRIS spectra of the three observed galaxies in the rest frame. The lower panels show the spectrum in logarithmic scale centred on the continuum region. This helps to appreciate the weak emissions,
such as the $[OIII]4363\AA$ and $[SIII]6312\AA$ auroral lines and the $HeII4686\AA$ transition, which are prominent in the three galaxies. Moreover, this scale also makes it easier to identify the Balmer jump with its higher continuum flux below the $3646\AA$ mark. The plot insets show the colour SDSS cutout of the galaxies from the SDSS sky server. Both the spectra and the images illustrate the characteristic properties of GP galaxies, namely unresolved morphology in SDSS and green colour due to extreme nebular emission, in particular high $EW([OIII]4959,5007\AA)$ and emission lines falling in the r'-band.
In this section, we describe the procedure employed to disentangle the main phenomena contributing to the observed radiation. These are the light extinction from the dust particles, the nebular continua from the hydrogen and helium atoms, the stellar continua from the underlying galaxy and the emission lines from the ionised gas. Moreover, this section also provides a brief technical explanation of the applied methodology, similar to the approach used in \citet{fernandez_determination_2018} for the high precision measurement of the primordial helium abundance using HII galaxies. It should be explained that, in order to guarantee consistent results, some of the steps described below were repeated iteratively, as the gas extinction and the chemical composition are used as inputs in the nebular and stellar continua fitting. This treatment assumes that gas and stars have the same light attenuation behind a screen of dust \citep[see][]{calzetti_dust_2001}. A more realistic scenario assumes clumps of dust with uneven extinction. Ignoring the disparity between the stars and gas attenuation can lead to an overestimation of the stellar and nebular continua on the line of sight. Fortunately, these objects have very low logarithmic extinction coefficients, making this simple model suitable.
\subsection{Emission-line flux measurement}
The spectra analysis starts with the measurement of the line fluxes. For every line, three spectral regions are selected: One covering the line width and two the adjacent continua. For a single emission line (for example $[OIII]5007\AA$) or an emission line consisting in several ionic transitions, which cannot be deblended (for example $[OII]3726\AA,3720\AA$ at our resolution) the measurement is done via a Monte Carlo integration: In a 1000 iterations loop, each pixel in the emission line is added a random quantity from a normal distribution. This distribution is centred at zero and its sigma is computed from the standard deviation of the adjacent continua regions (assuming a linear relation between the flux and the wavelength). Afterwards, we integrate the pixels within the emission line region. The mean value from the 1000 iterations is taken as the line flux measurement, while its standard deviation quantifies the flux uncertainty.
Emission lines consisting in more than one ionic transitions and whose components can be deblended, are fitted with a Gaussian mixture using \textsc{LMfit} by \citet{newville_lmfit_2014}. In the sample spectra, the blended lines are $[OI]6300\AA-[SII]6312\AA$, $H\alpha-[NII]6548\AA,6584\AA$ and $[SII]6716\AA,6731\AA$. The theoretical Gaussian area $\left(A \cdot \sqrt{2 \pi}\cdot\sigma\right)$ quantifies the emission flux, where $A$ is the Gaussian curve amplitude (the peak height with respect to the line continuum) and $\sigma$ is the Gaussian curve standard deviation. The uncertainty in the measurement is taken from \textsc{LMfit} $1\sigma$ standard error output from the default Levenberg-Marquart algorithm. We, however, are not using the default Gaussian parametrisation from \textsc{LMfit}. Instead, we define our own model which fits $A_{i}$, $\mu_{i}$ (the peak central wavelength) and $\sigma_{i}$ for every Gaussian component $(i)$.
In the first iteration, the emission flux measurement is performed on the observed spectra, Fig. \ref{fig:sample_spectra}. On subsequent iterations, once the stellar continuum has been fitted, this component is removed from the observation before the flux measurement. This provides a correction for the stellar absorption on the recombination lines. Table \ref{tab:emission_fluxes_gp03032} presents the emission line fluxes obtained from the output of the final iteration consisting on the pure emission. At the bottom of the table, the logarithmic extinction coefficient $c(H\beta)$ along with the $H\beta$ line equivalent width and flux is tabulated for each object.
\subsection{Interstellar reddening correction}\label{sec:extinction_calc}
The radiation reaching us is affected by dust particles, which scatter the light of the star forming region. A correction for this extinguished light can be accomplished by comparing the observed relative fluxes from recombination lines against their corresponding emissivity ratios. Any significant deviation can be attributed to dust scattering. This is a safe assumption since these lines have a very weak dependence on the electron temperature and density. This condition also assumes the absorption from the underlying stellar population is negligible. The extinction law selected for this analysis was \cite{cardelli_relationship_1989} with $R_V = 3.1$. In general, the extinction law choice has a weak impact on the optical spectrum analysis of GPs. This is due to their relative low dust content \citep[see][]{cardamone_galaxy_2009, amorin_oxygen_2010}. This reddening law is also the one applied on the SSP synthesis (see section \ref{sec:stellar-calc}).
As it was mentioned in Sec. \ref{data}, our observations display a flux mismatch between the blue and red grisms. To avoid inconsistencies in measurements from both grisms, the following scheme was applied: The emission lines at each range are normalised by the brightest hydrogen line ($H\beta$ for the blue arm and $H\alpha$ for the red arm). The logarithmic extinction coefficient, $c(H\beta)$, is calculated with the three most intense hydrogen lines in the blue arm ($H\beta$, $H\gamma$ and $H\delta$). Finally, in order to apply the extinction correction in the red arm features the classical formulation was adjusted to:
\begin{equation}
\label{eq:reddening_cor}
\frac{I_{\lambda}}{I_{H\beta}}=\frac{F_{\lambda}}{F_{H\alpha}}\cdot\frac{\epsilon_{H\alpha}}{\epsilon_{H\beta}}10^{c\left(H\beta\right)\left(f_{\lambda}-f_{H\alpha}\right)}
\end{equation}
where $F(H\lambda)$, $I(H\lambda)$ are the line emission flux and intensity respectively, $\frac{\epsilon_{H\alpha}}{\epsilon_{H\beta}}$ is the theoretical ratio between the $H\alpha$ and $H\beta$ emissivities, $c(H\beta)$ is the logarithmic extinction coefficient and $f(\lambda)$ is the transition reddening law value at the corresponding emission wavelength. The $f(\lambda)$ coefficient for each line can be found in Table \ref{tab:emission_fluxes_gp03032}. The emissivities were calculated using the atomic data references displayed in Table \ref{tab:atomic-data} from the \textsc{PyNeb} library by \citet{luridiana_pyneb:_2015}. In the first iteration, the recombination emissivity ratios were calculated using standard conditions of $T_e=10000K$ and $n_e=100\,cm^{-3}$. In subsequent iterations, the density and the low ionisation region temperature were used instead.
The previous two-arm reddening correction was also applied in the continuum analysis. In this case, however, Eq. \ref{eq:reddening_cor} was adapted for an absolute flux correction instead of a relative one.
\subsection{Measurement and treatment of the nebular continuum} \label{sec:nebular_calc}
In \citetalias{amorin_star_2012} the nebular continuum was characterised via the \textsc{PopStar} evolutionary synthesis models \cite[see ][]{molla_popstar_2009}. These grids define star forming regions consistently with both stellar atmospheres and ionised nebular continua. In the current analysis, we use the same atomic data and similar methodology as in \cite{molla_popstar_2009}. The main difference is that the current analysis computes the nebular continuum from the measured gas properties, instead of using the temperature assigned to the star forming region according to the metallicity \citep[see Table 6 in][]{molla_popstar_2009}. The full procedure description can be found in \cite{fernandez_determination_2018} but the main steps are summarised as follows:
\begin{itemize}
\item The continuous emission coefficient, $\gamma_\nu$ $\left(erg\,\nicefrac{cm^3}{s}\right)$, is calculated for the Free-Free (FF), Free-Bound (FB) and Bound-Bound (BB) continua. The Bremsstrahlung or FF continuum computation sticks to the classical methodology described by \cite{brown_theoretical_1970} and \cite{osterbrock_astrophysics_1974} for the $H^{+}$, $He^{+}$ and $He^{2+}$ ions. The FB continuum is calculated from the tabulated data by \cite{ercolano_theoretical_2006} again for the hydrogen and helium atoms. Finally, the BB or two-photon continuum is calculated via the analytical expression by \cite{nussbaumer_hydrogenic_1984} for the hydrogen $2s^{2}S_{1/2}\rightarrow1s^{2}S_{1/2}$ decay. In the first iteration, these components are calculated using standard conditions: $T_e=10000K$, $n_e=100\,cm^-3$, $y^+=0.01$ and $y^{2+}=0.001$. In the following iteration, the physical conditions are those measured from each GP low ionization region.
\item The combined continuous emission spectrum is flux calibrated using a \cite{zanstra_untersuchungen_1931} like approach using the $H\alpha$ flux and the effective recombination coefficient from \cite{pequignot_total_1991}, $\alpha^{eff}_{\alpha}$.
\item The output spectrum is artificially reddened using the logarithmic extinction coefficient previously measured. Finally, this continuum is removed from the observed spectrum.
\end{itemize}
\subsection{Measurement and treatment of the stellar continuum}\label{sec:stellar-calc}
Two different approaches were considered in the analysis of the GPs continua. In the first one the nebular continuum is removed from the observed spectrum prior to a SSP synthesis using \textsc{Starlight} by \cite{fernandes_semi-empirical_2005}. This step is essential for an unbiased study of the physical and evolutionary properties
of the stellar component of starburst galaxies, given that the nebular continuum shows, contrary to the young ionizing stellar component, a nearly flat SED, thus becoming progressively important in the red spectral range \citep[see][]{krueger_optical_1995,papaderos_age_1998,izotov_green_2011}. Adhering to STARLIGHT manual recommendations, the input spectrum is resampled to $1\AA$ per pixel. Additionally, an input mask file is provided to exclude the emission lines, sky contamination and noisy regions from the fitting. The input stellar bases combine grids from \cite{bruzual_stellar_2003}, \cite{falcon-barroso_updated_2011} and \cite{delgado_evolutionary_2005}. The 293 spectra grid covers a $\nicefrac{Z_{\odot}}{200}$ - $1.5Z_{\odot}$ stellar metallicity range and a $1 Myr$ - $17 Gyr$ stellar age range. The spectral range in these bases goes from $3300\AA$ to $6990\AA$ providing a good match to our observational wavelength range. The \textsc{Starlight} example configuration file was used as a template for our measurements with the following modifications: The normalization window chosen was $4760\AA$ to $4835\AA$. The minimum visual extinction was set to zero, while the maximum stellar velocity dispersion was defined equal to the gas velocity dispersion measured from the $[OIII]5007\AA$ line. Finally, the stellar velocity in the line of sight was fixed during the fits. This was done to avoid issues with the stellar velocity in the line of sight: since these spectra have no noticeable absorptions, enabling this feature results in a mismatch between the predicted absorption features and the observed emission lines.
The second methodology relies on the population spectral synthesis code \textsc{FADO} (Fitting Analysis Using Differential evolution Optimization) by \citet{gomes_fitting_2017}. A unique feature of this algorithm is that it takes both stellar and nebular emission into account and, quite importantly, identifies the SFH that self-conistently reproduces the observed nebular characteristics of a star-forming galaxy (H$\alpha$ and H$\beta$ luminosities and equivalent widths, shape of the continuum around the Balmer and Paschen jump). Another novel feature of FADO is the use of genetic algorithms for multi-objective optimization. The hydrogen and helium atomic data sources cited in section \ref{sec:nebular_calc} are also used in \textsc{FADO}. Moreover, its calculation follows a similar approach: The electron temperature and density are computed from the $[OIII]$ and $[SII]$ transitions while the nebular extinction is computed from the $H\alpha$ and $H\beta$ fluxes. Afterwards, the SSPs synthesis is applied considering the nebular component. In this fitting, the SSP library considered is the reviewed grid by \cite{bruzual_stellar_2003}. It consists in 1098 bases with a wider time range at smaller increments. In the $5 < log(age) < 6$ interval there are 18 steps with $\Delta(log(age)=0.05$. For $6 < log(age) < 10.3$, there are 202 steps at $\Delta(log(age))=0.02$. The stellar metallicities included are $Z = 0.005$, $0.020$, $0.200$, $0.400$, $1.000$ and $2.500$ in a $Z_{\odot}$ units.
In the continua analysis, we incorporate the GP sample from \citetalias{amorin_star_2012}. This is done to test the previous results against those using the novel stellar libraries and fitting algorithm.
\subsection{Physical properties and chemical abundances analysis} \label{chemical_analysis}
\begin{table}
\caption{\label{tab:Priors-and-likelihood} Priors and likelihood distributions
in our model. The term $X^{i+}$ includes all the ionic metal abundances:
$Ar^{2+}$, $Ar^{3+}$, $Cl^{3+}$, $Fe^{3+}$, $O^{+}$, $O^{2+}$, $Ne^{3+}$,
$N^{+}$, $S^{+}$, $S^{2+}$, $y^{+}$ and $y^{2+}$. The helium abundances are defined in logarithmic scale, while the metals are defined using a $12+log\left(X^{i+}\right)$ notation.}
\centering{\input{tables/priors_conf}}
\end{table}
The gas chemical abundance analysis using optical emission lines has been recently illustrated by \cite{perez-montero_ionized_2017}. In the scenario where one or several electron temperatures and densities can be derived, this methodology is known as the direct method. This is the ideal case, since it makes possible to calculate each transition emissivity. However, since the electron temperature measurement depends on weak auroral lines, a high precision result requires a high S/N. A simple model predicts the relation between the phenomena contributing to the recombination and collisionally excited line flux:
\begin{equation}
\frac{F_{X^{i+},\,\lambda}}{F_{H\beta}}=X^{i+}\frac{\epsilon_{X^{i+},\,\lambda}\left(T_{e},\,n_{e}\right)}{\epsilon_{H\beta}\left(T_{e},\,n_{e}\right)}\cdot10^{-c\left(H\beta\right)\cdot f_{\lambda}}\label{eq:fluxFormula}
\end{equation}
where $\nicefrac{\epsilon_{X^{i+},\,\lambda}}{\epsilon_{H\beta}}$ is the relative emissivity at the transition wavelength $\lambda$, for an ion with abundance $X^{i+}$, at certain electron temperature $T_{e}\,\left(K\right)$ and electron density $n_{e}$ $\left(cm^{-3}\right)$. The term $c\left(H\beta\right)$ is the relative logarithmic extinction coefficient at $H\beta$ for a reddening law $f_{\lambda}$.
For this study, we solve simultaneously this system of equations via a Bayesian model which was presented in \cite{fernandez_bayesian_2019}. This algorithm was written using the probabilistic programming library \textsc{PyMC3} by \citet{salvatier_probabilistic_2016}. Its NUTs (No-U-Turns) sampler \citep[see][]{hoffman_no-u-turn_2011} is based on a HMC (Hamiltonian Monte-Carlo) algorithm, written itself via the neural networks package Theano \citep[see][]{the_theano_development_team_theano:_2016}. The first advantage from this approach is the complete fitting of the parameter space. In the current work, this means 14 out of the current maximum of 16 parameters: One electron temperature, one electron density, the logarithmic extinction coefficient and eleven ionic species, see Table \ref{tab:Priors-and-likelihood}. The second advantage is its measurement speed: Fitting up to 23 emission line fluxes from the input GPs spectra takes less than 160 seconds in a i7-10700K Processor. As it was discussed in \cite{fernandez_bayesian_2019} the results are consistent with those from the traditional method. The full algorithm description, as well as examples showing its convergence accuracy and stability, can be found in \cite{fernandez_bayesian_2019}. However, the following points provide a recap on its workings as well as an explanation on the updates:
\begin{equation}
Pr\left(\theta|y\right)=\frac{Pr\left(y|\theta\right)Pr\left(\theta\right)}{Pr\left(y\right)}\label{eq:bayesTheorem}
\end{equation}
\begin{itemize}
\item The first step in a Bayesian model involves the priors selection. These distributions, $Pr\left(\theta\right)$ in Eq. \ref{eq:bayesTheorem}, represent our knowledge of the physical parameters prior to their measurement. Ideally, the user introduces an "uninformative" prior: This means a wide distribution, which guarantees a large range of solutions for the parameter value. As the sampling process reaches the solution region, the value of $Pr\left(\theta\right)$ changes very little and the simulation becomes dominated by the likelihood, $Pr\left(y|\theta\right)$. Table \ref{tab:Priors-and-likelihood} presents the configuration for our priors. The main novelty with respect to those chosen in \cite{fernandez_bayesian_2019} is the inclusion of priors for the $Cl^{3+}$, $Ne^{3+}$ and $Fe^{3+}$ abundances. These ionic abundances measurement is possible via the $[ClIII]5518\AA$, $[NeIII]3968\AA$ and $[FeIII]4658\AA$ lines respectively. Moreover, both the electron density, ne, and the reddening constant, $c(H\beta)$ use a Half-Cauchy prior distribution to avoid non-physical negative values for these parameters.
\item The likelihood, $Pr\left(y|\theta\right)$ in Eq. \ref{eq:bayesTheorem}, evaluates the observational data against the sampled parameter values at each simulation step. As it was discussed in \cite{fernandez_bayesian_2019}, the Bayesian paradigm cannot interpret uncertainty on the input data. In this astrophysical model, however, this is critical. For example, while the uncertainty in the photon flux from the $[OIII]5007\AA$ transition may be neglected, the same cannot be said for photons flux measured in the $[OIII]4363\AA$ transition. Moreover, not all the data points have the same impact on the parameter space exploration. Indeed, the $[OIII]5007\AA$ photons mainly contribute to the $O^{2+}$ abundance measurement. In contrast, the $[OIII]4363\AA$ photons have a strong impact in the temperature measurement, and therefore, all the metals abundance. To account for the uncertainty weight, a normal distribution likelihood is defined for each input emission line. This distribution is centred at the current iteration synthetic flux, $F_{\chi^{i+},\lambda}$, while its width is given by the observed relative flux uncertainty, $\sigma_{\chi^{i+},\lambda}$. In this design, emission lines with a larger uncertainty display a wider likelihood, which translates into a larger acceptance rate.
\item This chemical model considers two ionization regions: A high ionization region for the $Ar^{3+}$, $Cl^{3+}$, $Fe^{3+}$, $O^{2+}$, $y^{+}$ and $y^{2+}$ ions the low ionization region for the remaining ions. The main temperature diagnostic ratio in the sample wavelength range is $R_{[OIII]} = \nicefrac{I(4959\AA)+I(5007\AA)}{I(4363\AA)}$. This is the temperature used for the high ionization region, $T_{high} = T_{[OIII]}$. For the low ionization region, we use the $O^2+$ temperature, $T_{low} = T_{[OII]}$. However, since the $[OII]7319,7330\AA$ have a poor S/N in our spectra we use the relation presented in \cite{perez-montero_impact_2009}:
\begin{equation}
T[OII]=\frac{1.2+0.002n_{e}+\frac{4.2}{n_{e}}}{\frac{10000}{T[OIII]}+0.08+0.003n_{e}+\frac{2.5}{n_{e}}} \label{eq:TOIII-TSIII-relation}
\end{equation}
where $n_{e}$ is in $cm^{-3}K$ units. This equation was computed from the models in \citet{perez-montero_line_2003} using \textsc{Cloudy v96} \citep[see][]{ferland_hazy_2002}. These models agree with the observations, in a non-linear relation between both oxygen ions temperature. The models assume that the electron density remains spatially constant. This may not be realistic for SFR with complex dynamics. Additionally, it should be remembered that the electron density below $100cm^{-3}$ cannot be accurately measured with the available diagnostics. This means that keeping $n_{e}=100cm^{-3}$ constant at this boundary may be underestimating the predicted low ionization temperature. In this work, we use the ratio $R_{[SII]} = \nicefrac{I(6716\AA)}{I(6731\AA)}$ to anchor the electron density.
\item In \cite{fernandez_bayesian_2019}, it was discussed that the main disadvantage from the Bayesian sampler was the requirement to parameterize the emissivity temperature-density plane. In contrast to the traditional Monte Carlo algorithm used in \cite{fernandez_determination_2018}, tensors in neural networks cannot be easily interpolated. In most most cases, this is not an issue as emissivity parameterizations provide a sufficiently high accuracy \citep[see][]{perez-montero_ionized_2017,peimbert_nebular_2017}. However, in the case of the helium transitions, up to $5\%$ discrepancy was observed for some temperature-density regions. In this study, this handicap was removed by including the \textit{RegularGridInterpolator} function from the \textsc{exoplanet} library by \cite{dan_foreman-mackey_dfm/exoplanet:_2019}. This function enables the interpolation on a regular tensor with an arbitrary number of dimensions. The inclusion of this function increases simulation time up to three times in our synthetic test cases. However, it makes possible a more faithful characterisation of the emissivities computation and the inclusion of more emission lines. The atomic data considered in the chemical analysis can be found in Table \ref{tab:atomic-data} within the appendix
\end{itemize}
\subsection{Characterization of the ionizing radiation}
\begin{table*}
\caption{\label{tab:atomic-data}Atomic data references for the emission lines considered in the chemical analysis.}
\centering{\input{tables/atomic_data_table}}
\end{table*}
Recently, \citet{perez-montero_photon_2020} delved into the physics of HeII-emitters. They proposed a Bayesian-like algorithm to sample photoionization grids, in which the $[OII]3727,3729\AA$, $[OIII]5007\AA$, $[SII]6717,6731\AA$, $[SIII]9069\AA$, $HeI4471\AA$, $HeI5876\AA$ and $HeII4686\AA$ line fluxes are a function of the oxygen abundance, the effective cluster temperature $(T_{eff})$ and the ionization parameter $(log(U))$. The former term quantifies the temperature of a single source responsible for the ionizing radiation. The latter term represents a dimensionless ratio of the ionization photon density to the electron density. Different photoionization grids were generated
using \textsc{Cloudy v17} \citep[see][]{ferland_2017_2017} to map $T_{eff}$ according to different SEDs.
For this work, we consider a blackbody with $T_{eff}$ ranging from $30kK$ to $90kK$. According to \citet{perez-montero_photon_2020}, these grids are the only ones with high enough $T_{eff}$ values, to generate the $HeII4686\AA$ fluxes in the control sample. The $log(U)$ value goes from -4.0 to -1.5. Finally, the ionized gas metallicity is parameterized according to the oxygen content, where $12+log(\nicefrac{O}{H})$ increases from 7.1 to 8.9. Additionally, the algorithm photoionization grids include both a spherical and plane-parallel geometries as defined by the one-dimensional Cloudy model. This algorithm is named \textsc{HII-CHI-mistry-Teff} and it shares a similar sampling procedure to the \textsc{HII-CHI-mistry} algorithm presented in \cite{perez-montero_deriving_2014} and \citet{perez-montero_using_2017}: For the selected model grid, the algorithm starts by slicing the grid given the input $\nicefrac{O}{H}$ abundance. Afterwards, the algorithm loops through the theoretical $HeII4686\AA$ fluxes. At non-zero values, the code compares the observed $\nicefrac{[OII]}{[OIII]}$ or $\nicefrac{[SII]}{[SIII]}$ ratios (the one provided by the user) and helium fluxes against the theoretical values. This results in a set of $\chi^{2}$ weights. This process is repeated for an input number of iterations where the line fluxes are randomly generated in a standard Monte Carlo process: A normal distribution centred at the observed flux and a standard deviation given by the line flux uncertainty. The final $T_{eff}$ and $log(U)$ values are computed from $\chi^{2}$-weighted traces.
As a novelty, this work introduces a \textsc{HII-CHI-mistry-Teff} algorithm written with neural networks. This model adheres to a Bayesian paradigm following the same structure described in section \ref{chemical_analysis}: The algorithm starts defining a prior for $T_{eff}$ and $log(U)$ (see Table \ref{tab:Priors-and-likelihood}). From the input oxygen abundance, an additional prior is defined with a normal distribution with $\mu=\nicefrac{O}{H}$ and $\sigma=\sigma_{\nicefrac{O}{H}}$. This informative prior is introduced to account for the uncertainty in the elemental abundance measurement. Afterwards, the $T_{eff}-log(U)-\nicefrac{O}{H}$ grid is interpolated using the algorithm from \citet{dan_foreman-mackey_dfm/exoplanet:_2019}. Finally, the line intensities are extinguished according to the input $c(H\beta)$ coefficient prior their evaluation against the observed values. This likelihood is defined using the same schemed as in the chemical model (see Table \ref{tab:Priors-and-likelihood}).
This new design can be integrated with in the chemical model described in section \ref{chemical_analysis}. In this case, rather than using a constant oxygen abundance to interpolate the photoionization grids, this is calculated from the sampled $O^{+}$ and $O^{2+}$ values. Similarly, the emission relative line intensities are extinguished at each iteration form the currently proposed $c(H\beta)$. This brings the number of dimensions for the GPs sample spectra fitting to 16. In the Bayesian \textsc{HII-CHI-mistry-Teff} and the Bayesian Direct method + \textsc{HII-CHI-mistry-Teff} only the plane parallel geometries are considered. This selection is partially imposed by the fact that these are the only regular grids included in \textsc{HII-CHI-mistry-Teff}. Learning from this work conclusions, future research will be focused on generating tailored photoionization grids for a Bayesian sampler. This is, though, beyond the scope of the current work.
\section{Discussion} \label{discussion}
A careful inspection of the sample continua (see Fig. \ref{fig:sample_spectra} lower panels in logarithmic scale) confirms no evidence of a blue $(\lambda\sim4650\AA)$ or red bump $(\lambda\sim5808\AA)$. Unlike in the \citetalias{amorin_star_2012} sample, the current spectra do not show WR features. Moreover, none of the objects display noticeable absorption features, such as the $MgI\lambda\lambda5167,5173\AA$ lines. All three objects, however, exhibit nebular $HeII4686\AA$ emission, which demonstrates the existence of a very hard ionization field \citep[see][]{kehrig_extended_2015,kehrig_extended_2018}. Due to the sample redshift, it can be noticed that several telluric features are contaminating the optical range. In GP030321, a sky band has completely swallowed the $[NII]6548\AA$ emission. Fortunately, it does not affect the $H\alpha$ flux. Moreover, in GP101157 a telluric absorption has a detrimental impact on the $HeI6678\AA$ line. Consequently, it is excluded from the chemical analysis. Finally, from the lower panels in Fig. \ref{fig:sample_spectra} the Balmer jump $(\lambda\sim3646\AA)$ can be easily appreciated on GP101157 and GP121903. The smaller jump in GP030321 could be explained by an older star forming region.
\subsection{Analysis of the synthesis spectral continuum fitting} \label{stellar-discussion}
\begin{figure}
\includegraphics[width=1.0\columnwidth]{images/SSP_synthesis_comparison.png}
\caption{\label{fig:SSP-comparison}SSP synthesis output continua comparison for galaxy GP232539}
\end{figure}
\begin{table*}
\caption{\label{tab:FADO-sample-output} FADO results for the SSP synthesis for the GP sample. Column (1) displays the object name. Column (2) shows fit $\chi^{2}$ coefficient (top) and the nebular component luminosity percentage (bottom). Column (3) shows the SSPs mean metallicity weighted by light (top) and mass (bottom). Column (4) displays the average age, in logarithmic scale, weighted by light (top) and mass (bottom). Column (5) displays the logarithmic mass ever formed: All populations total (bottom) and the fraction above 1Myr. Column (6) displays the logarithmic mass currently available: All populations total (bottom) and the fraction above 1Myr. Finally, column (7) displays the stellar (top) and nebular (bottom) extinction. The uncertainties predicted by FADO are $\Delta z\approx0.005$, $\Delta log(t)\approx0.005$, $\Delta log(M)\approx0.001$ and $\Delta c(H\beta)\approx0.01$ }
\centering{{\input{tables/fado_fitting_results}}}
\end{table*}
\begin{figure*}
\includegraphics[width=1.0\textwidth]{images/fado_infograph_GP232539.png}
\caption{\label{fig:fado-infograph}\textsc{FADO} continua fit for the galaxy GP232539 (orange). The best fitting SED (blue) includes the stellar (black) and nebular (red) components. The histograms on the upper and lower right corners display the luminosity and mass fraction at the normalization wavelength $(6180\AA)$. The color-coding depicts the metallicity and the vertical bars the $\pm1\sigma$ uncertainties (the values can be found at the top right top right corner). The grey bars in between both plots correspond to the ages for the stellar grid. The vertical arrow in the mass fraction diagram marks the age when 50\% of the present-day stellar mass has been in place. Same plots for the remaining galaxies can be found in the appendix}
\end{figure*}
Due to the intense emission line features, it is easy to ignore the gas contribution to the continuum. For example, \citet{reines_emerging_2008} noticed an excess on the Hubble U and I bands observation of massive stellar clusters, which could not be explained by the stellar component. In posterior work, \citet{reines_importance_2009} succeeded to explain this excess from the nebular continuum and the sulfur emission at $[SII]9069,9531\AA$. As we delve into higher redshifts and more intense star-forming systems, it is essential to take into account this gas contribution \citep[see][]{schaerer_impact_2009,amorin_extreme_2015}. The two SSPs synthesis techniques we apply in this analysis include the nebular continuum computation. The first one follows the iterative analysis presented in \citet{fernandez_determination_2018} for the high accuracy determination of the primordial helium abundance from local star forming galaxies. The second methodology involves the application of \textsc{FADO} by \citet{gomes_fitting_2017}.
Fig. \ref{fig:SSP-comparison} shows both techniques (see sec. \ref{sec:nebular_calc} and sec. \ref{sec:stellar-calc}) output against the GP spectrum. This fitting corresponds to the galaxy GP232559 from the \citetalias{amorin_star_2012} sample, however, similar conclusions are drawn from the six spectra fits. Both techniques show a good visual match with the observations, even though the output $\chi^{2}$ coefficients can reach double digits (see Table \ref{tab:FADO-sample-output}). The largest discrepancies are found at the blue end, in the Balmer jump proximity. The iterative \textsc{Starlight} + nebular computation analysis displays a better agreement with the observed continuum for $\lambda>3646\AA$. In contrast, this technique overestimates the continuum flux for $\lambda<3646\AA$ region. The opposite pattern is observed on the \textsc{FADO} fitting. A couple of causes can contribute to this effect: In the \textsc{FADO} fits, the region $3600\AA<\lambda<3900\AA$ is completely masked, while in the \textsc{Starlight} only the emission line regions are masked. A large mask can make it hard to for the algorithm to fit the continuum, due to the limited number of data points. Nonetheless, it may be argued that even if we cannot see the emission lines from Balmer transitions above $HI_{Balmer}>16$ this emission rich region should still be masked. Another cause behind this discrepancy might be on the Balmer jump height calculation. The free-bound continuum features have a sensitive, inversely proportional relation to the hydrogen temperature \citep[see][]{zhang_electron_2004}. In the case of \textsc{FADO}, the $T_{[OIII]}$ temperature is used as a proxy to compute the nebular continua, while in the other approach, we use $T_{[OII]}$, leading to a higher jump height. According to the relation by \citet{liu_ngc_2000} a 10\% variation on the Balmer jump height measurement leads to almost a $1,000K$ discrepancy on the hydrogen temperature. The work by \citet{guseva_balmer_2006,guseva_balmer_2007} on low metallicity HII regions concluded that there was a good agreement on $T_{[OIII]}$ and $T_{HI}$ within the $O^{2+}$ zone. However, the authors warned that the $T_{HI}$ temperature determination was more reliable in regions with larger contribution from the gas. This means a higher $EW(H\beta)$, and therefore, younger objects. In the work by \citet{garcia-rojas_analysis_2012}, despite the high-quality spectra from planetary nebulae, the author refrained from comparing the Paschen discontinuity temperature with emission line diagnostics due to the high uncertainty involved. From this comparison, the choice between low or high ionization temperature has a small impact on the resulting stellar continuum fit (against the option of neglecting the nebular component all together). However, due to these GPs rich emission features, there is a shortage of "blue" continuum. This means that the fixed Balmer jump height has a large weight, in a region where the input spectrum is mostly masked. Consequently, we may be properly detecting the youngest stars, but their mass/luminosity contribution is underestimated.
Regarding the stellar kinematics, in general, the \textsc{FADO} output predicts slightly weaker absorption features with larger stellar dispersion velocities. This discrepancy can be explained by the boundary condition in the \textsc{Starlight} fitting, where the kinematics must remain below the $\sigma_{[OIII]}$ value. It can be appreciated in Fig. \ref{fig:SSP-comparison} that GP spectra have almost imperceptible hydrogen absorptions. Due to this, most algorithms attempt to overestimate stellar dispersion velocity to better simulate the input "flat" spectrum. Consequently, the stellar dispersion velocity measured is actually the upper boundary condition for this parameter. In the literature, the uncertainty on the absorption features strength is a concern for high precision measurements of the helium abundance \citep[see][]{izotov_new_2014, aver_effects_2015, peimbert_primordial_2016}. In these works, it is commonly accepted that the impact on the hydrogen and helium is of the order of $Eqw \approx 2-0.5\AA $ respectively. These values were taken from synthetic stellar libraries such as those from \cite{delgado_synthetic_1999,delgado_synthetic_1999-1}. In the updated analysis by \citet{aver_improving_2020}, the authors quantified the hydrogen and helium line absorptions from the BPASS library \citep[see][and references therein]{stanway_re-evaluating_2018}. They proposed that this uniform ion absorption should be corrected by a relative scaling factor for each transition. In the case of the $H\alpha$, $H\gamma$ and $H\delta$ the scaling factor (relative to the $H\beta$ absorption) stands in the 0.964-0.930 range. In the case of the helium lines, the scaling factor (relative to the $HeI4471\AA$ absorption) varies within 0.346-1.4 range. Unlike the hydrogen lines, the helium absorptions strength shows no constant behaviour with the transition wavelength.
The small discrepancies between both techniques can be explained by the previous arguments. At this point, however, we take the results from \textsc{FADO} as the most accurate. Moreover, given the larger SSPs library provides, it provides a more detailed characterization on SFH. In contrast, the output stellar continuum from the iterative approach is the one used for the chemical analysis. Since this methodology applies the $c(H\beta)$, $\sigma_{[OIII]}$, $Te$, $y^+$ and $y^{2+}$ measured from the emission spectrum, it is deemed to provide a more tailored correction for the recombination lines.
Fig. \ref{fig:fado-infograph} displays the output from the SSPs synthesis for GP232539. The complete infographics for the current and \citetalias{amorin_star_2012} sample can be found on the appendix. However, Table \ref{tab:FADO-sample-output} provides a summary with the main results. The sample rundown leads to the following conclusions: Both the current and \citetalias{amorin_star_2012} are confirmed to be dwarfish systems with $\overline{\left\langle log(M_e) \right\rangle}\approx9\,M_{\odot}$. As expected, the younger populations dominate the galaxy luminosity, $\overline{\left\langle log(t_L) \right\rangle}\approx7.6\,yr$, while the older populations dominate the mass, $\overline{\left\langle log(t_M) \right\rangle}\approx8.9\,yr$. As discussed in \citetalias{amorin_star_2012} the younger population peaks around $log(t) \approx 7\, yr$ while the older population remains at $log(t) > 9\, yr$. The new methodology confirms the presence of very young stars with $log(t) < 6\, yr$. These stars could not be properly identified in \citetalias{amorin_star_2012} due to the very limited number of young stars in the SSP grid. Their light contribution, however, accounts only for a few percentage points and their combined mass one order below that. The output visual extinction coefficients have been converted to the relative logarithmic extinction coefficients for comparison's sake. In the case of GP101157 and GP121903 a good agreement is found with the nebular extinction from the results of the iterative analysis in Table \ref{tab:emission_fluxes_gp03032}. For GP030321, however, the extinction coefficient displays a negative value. Negative extinctions are also fitted for the stellar component of GP121903 and GP113303. This can be explained by the very low extinction encountered in these objects, making it difficult to anchor the parameter value. As in the case of \textsc{Starlight}, a non-negative extinction boundary must be imposed on these galaxies. Assuming the low uncertainty limits on the \textsc{FADO} measurements, the extinction between the stellar and nebular components are distinctively different.
Focusing individually on the GPs stellar distribution, a few patterns emerge which must be highlighted:
\begin{itemize}
\item The stellar masses measured for GP113303 and GP232539 display a very good agreement with the values presented in \citetalias{amorin_star_2012}. In the case of GP004054, however, the current value is a $13\%$ lower as the oldest stellar population ($log(t) > 8\, yr$) is not detected in the SSP synthesis. In \citetalias{amorin_star_2012}, only 3 stellar populations (two young and one old) were detected in GP004054 fitting. In the present methodology, both GP004054 and GP113303 still display few SSPs (below the dozen), however the \textsc{FADO} fitting predicts a wider age range for the younger stellar populations. In \citetalias{amorin_star_2012} both galaxies displayed a $\Delta log(t) \approx 6.8-7\, yr$ for both galaxies, while the current approach predicts a time interval $\Delta log(t) \approx 5-7.5\, yr$ for GP004054 and $\Delta log(t) \approx 6-7\, yr$ for GP113303. In general, few SSPs in the output continuum are due to an input spectra with small wavelength range. Due to the emission lines, the WR features, the sky bands and the noisy regions, GP004054 has irrefutably the smallest amount of data points in the SSP synthesis. Finally, some discrepancies between both results may be explained by the larger SSP library in the current work.
\item The application of the new methodology alongside the very fine stellar grid results in a more complex set of stellar distributions in the GP sample. In the case of the younger populations, the light and mass distributions show similar profiles. This is not the case, however, for the older stellar populations. On the one hand, the mass fraction histograms show a continuous distribution of SSPs, which peak around $log(t)\approx 8.5yr$ and even higher in some GPs. On the other hand, the light fraction in some GPs display a single SSP spiking above the rest. This is the case of GP232539 in Fig. \ref{fig:fado-infograph}. This is also the case of GP101157 and GP121903. The same behaviour could be attributed to GP004054 and GP113303, however, as discussed above, due to the limited number of SSPs fitted, few constrains can be derived from their SFH. In contrast, GP030321 despite having a light and mass fraction distributions very similar to GP232539 shows a light contribution from the older population below the $2\%$. The physical justification behind a single source dominating the luminosity, but not the mass, could be found on the nebular emission. However, since our methodology quantifies the contribution from the gas component, these results open two possible scenarios:
\end{itemize}
On the one hand, the calculated nebular continuum is underestimated. It is known that the gas contribution tends to make the stellar SED redder than expected \citep[see][]{krueger_optical_1995,papaderos_age_1998,izotov_green_2011}. In the current GP sample, the nebular continuum computed by FADO is 10.1\%, 27.6\% and 23.28\%. Similar values were found for the continuum level of $H\alpha$ and $H\beta$ for the GPs in \citetalias{amorin_star_2012}. This work included an analysis on the Hubble photometry for the GP sample. It confirmed that a LSB (Low Surface Brightness) periphery was appreciated for all GPs which could be associated to the underlying galaxy. The authors, however, included a warning that this component could be enhanced, or dominated by the nebular emission. This phenomenon would decrease the absolute magnitude of the host galaxy by $0.75 - 1\,mag$. Recently, \citet{clarke_old_2020} accomplished a similar analysis, carefully selecting a GP sample, whose Hubble photometry avoided the oxygen and hydrogen strongest lines. They also concluded that the LSB envelope is compatible with evolved hosts for BCDs and the nebular continuum contributes to the redder colours with $B-I\approx 0.76$. Nonetheless, given the multi-cluster nature of some GPs, it cannot be excluded that the young stars and ionized gas have a mismatching spatial distribution. This scenario was investigated by \citet{papaderos_i_2012} for IZw18, where the nebular emission envelope supports a non-cospatial scenario with the underlying ionizing and/or the non-ionizing stellar background. Even though this galaxy might be an extreme BCDs in the local universe, its properties might as well be ordinary at higher redshifts. Assuming a standard geometry could bring many biases to the measurements, such as the ones observed in this sample. Recently, the work by \citet{leja_older_2019} concluded that the 3D-HST, 0.5 < z < 2.5 catalogues have systematically older SFHs using the flexible SFH algorithm \textsc{Prospector-$\alpha$}. In these objects, the largest offsets are caused by the $t > 100 Myr$ stars. These results motivate more and deeper observations for a better characterization of extreme star forming systems.
On the other hand, the stellar population synthesis may be overestimating the luminosity from the older stellar populations. A self-consistent algorithm such as FADO, not only fits the best SFH but also the nebular features: the hydrogen continuum discontinuities, the hydrogen line luminosities and their equivalent widths. A large continuum photon escape fraction may have a detriment effect on the emission gas features. As fewer younger stars are necessary to explain the ionized gas, the algorithm balances the continuum fit with a larger contribution from the older stellar populations. These complex scenarios with have been explored by are left to investigate for a forthcoming paper (Papaderos et al, in preparation).
\subsection{Analysis of the resulting gas physical properties and chemical abundances} \label{emission-discussion}
\begin{table}
\caption{\label{tab:fit_results}Resulting gas physical parameters and chemical abundances from the multidimensional emission-line fitting in the three analyzed GP galaxies.}
\centering{{\tiny\input{tables/sample_bayesian_fitting}}}
\end{table}
\begin{figure}
\includegraphics[width=1.0\columnwidth]{images/gp030321_paramsPosteriors.png}
\caption{\label{fig:GP030321_outputFitPlot}Multi-dimensional fit plot for the chemical model of GP030321}
\end{figure}
\begin{figure}
\includegraphics[width=1.0\columnwidth]{images/gp121903_emFluxPosteriors.png}
\caption{\label{fig:GP121903_flux_FitPlot} Histogram grid with the synthetic fluxes for the emission lines fitted for GP121903. The distributions are colour coded (in the electronic version) according to the ion generating the photons. The vertical black line and grey shaded area correspond to the observed line fluxes with their uncertainty}
\end{figure}
The data in Table \ref{tab:fit_results} shows the results from the multidimensional analysis. For GP030321, the same data is displayed graphically in Fig. \ref{fig:GP030321_outputFitPlot}. As it was explained in \citet{fernandez_bayesian_2019}, these diagrams provide a quick diagnosis on the Monte Carlo algorithm fitting and convergence quality: On the left-hand side, the white noise-like plot represent the parameter trace at each of the 9000 simulation steps. As long as no divergences are encountered, on any parameter, it can be stated that the sampling has reached convergence. Moreover, these traces consist in three distinctive simulations (running three cores in parallel). A seamless juncture between all runs, as it is the case in the traces in Fig. \ref{fig:GP030321_outputFitPlot}, means that the initial conditions are not affecting the measurements. On the right-hand side, the same traces are plotted as a histogram. This is the Bayesian posterior, and consequently, our measurement.
\begin{table}
\caption{\label{tab:elemental-abundances} Elemental abundances and chemical parameters computed from the results of the Bayesian direct method model.}
\centering{\input{tables/elemental_abundances}}
\end{table}
An additional quality test is available in Fig. \ref{fig:GP121903_flux_FitPlot}. This grid shows the fitting theoretical fluxes calculated from Eq. \ref{eq:fluxFormula} against the observational ones. At this point, the reader is encouraged to check the online support material, where Fig. \ref{fig:GP030321_outputFitPlot} and Fig. \ref{fig:GP121903_flux_FitPlot} can be found for the remaining two objects. It can be stated that in general, the fits agree very well with the observational data. Nonetheless, the few discrepancies should be explained:
\begin{itemize}
\item $c(H\beta)$ parameter: In the simulation of GP121903, the logarithmic extinction posterior displays a bounded distribution. This is explained by the parameter prior, a half-Cauchy, distribution which forbids negative parameter values. In this galaxy with exceptionally low extinction, as the simulation tries to sample negative coefficients the measurement ends-up being constrained. This pattern is also observed in the $H\gamma$ flux posterior plot for GP121903. Consequently, this line is the one pushing towards negative extinction coefficients. The likely cause is the correction for the underlying stellar absorption which not being sufficiently accurate. A solution for this issue could be a prior which simulates an asymptotic behaviour as $c(H\beta)$ approaches zero. Nonetheless, since the dust extinction is very small the impact from the hydrogen lines miss-match has a very small impact on the rest of the parameter space.
\item $n_e$ parameter: In the simulation of GP030321 the density posterior appears to be double peaked. One value around $n_e = 25cm^{-3}$ and the second peak at zero. As in the previous case, this is caused by the simulation priors, which have been designed to avoid non-physical values. Moreover, the emissivity grids are constrained to a density range of $n_e=1-800cm^{-3}$. A traditional density calculation using the $ne[SII]$ parameterization by \citep[see][]{perez-montero_ionized_2017} results in a negative density value. Trying to explain these results, a careful inspection of the line profiles may lead to the conclusion of a wider or secondary profile in the $[SII]6716\AA$ line. This behaviour has been recently analysed by \citet{hogarth_chemodynamics_2020} in GPs with clear signatures of outflows and turbulence. This could explain the non-physical emissivities derived for this object. However, a secondary component would also be present on the recombination and collisionally excited auroral lines. This is not the case for any galaxy of this sample. Therefore, we can dismiss an outflow. A review on the Ultraviolet and Visual Echelle Spectrograph (UVES) sky map by \citet{hanuschik_flux-calibrated_2003} confirms intense sky lines at 7821.5190$\AA$ and 7841.2847$\AA$. In GP030321 the $[SII]6717,6731\AA$ doublet is observed blended in the 7824-7841$\AA$ range. Consequently, the second component is heavily contaminated by the sky. This can explain the doublet profile difference after the background subtraction task. The direct impact of this bimodal distribution on the emissivity computation is negligible for the observed transitions at $n_{e} < 100cm^{-3}$ regime. However, there is an indirect impact due to the temperature value of $T[OII]$ which from eq. \ref{eq:TOIII-TSIII-relation} depends on $n_{e}[SII]$. Indeed, wider priors (i.e. larger uncertainty) is observed on the low ionization abundances. This can be explained by the density distribution. Nonetheless, $T[OII]$ also depends on $T[OIII] (T_{high})$, which as we may see in Fig.\ref{fig:GP030321_outputFitPlot} has a normal posterior. This means that the temperature fit is not affected by the density uncertainty. Therefore, the fitting is behaving as expected: larger uncertainty only on the parameters which depend on observables with higher error.
\item $HeI$ fluxes: In the case of GP030321 and GP101157 large discrepancies ($\sim20-40\%$) between the theoretical and observed fluxes are found for the helium lines. As it was discussed in \citet{fernandez_determination_2018,fernandez_bayesian_2019} the observed HeI photons are affected by several phenomena whose characterization is challenging. The main one is the absorption for the underlying stellar population whose accuracy on the absorption features strength cannot be easily quantified. Moreover, no error propagation takes place between the stellar population synthesis and its output removal from the observed spectrum. This means that the uncertainty is underestimated in these fits, independently of the line wavelength. This phenomenon can also explain the poor fitting observed in some cases for the smaller wavelength HI transitions.
\end{itemize}
\begin{table*}
\caption{\label{tab:photoIonization-param-comparison} Comparison between the photoionisation grid fitting described in the text for each galaxy.}
\centering{\input{tables/photo-ionization_param_comparison}}
\end{table*}
\begin{table*}
\caption{\label{tab:photoIonization-flux-comparison} Observational fluxes against the synthetic fluxes fitted via the techniques described in the text for GP101157. The percentages correspond to the difference with the observed values in column}
\centering{{\tiny\input{tables/gp101157_lineFlux_fit_comparison}}}
\end{table*}
\begin{table}
\caption{\label{tab:fit_results_large}Chemical analysis results from the Direct method + \textsc{HII-CHI-mistry-Teff} sampler.}
\centering{}{\tiny\input{tables/sample_bayesian_fitting_large}}
\end{table}
Table \ref{tab:elemental-abundances} displays the elemental abundances computed from the output traces for the O, N, S and He ions. The methodology follows a basic approach where the measured ionic abundances are considered by most significant, and thus, they are summed up. The exception is the ionization correction factor (ICF) for the $S^{3+}$ abundance. This $ICF(S^{3+})$ was presented in \citet{fernandez_determination_2018} and it is a function of the $Ar^{2+}$ and $Ar^{3+}$ abundances. It is concluded that three GPs presented here belong to the low metallicity regime with $7.76<12+log\left(\nicefrac{O}{H}\right)<8.04$. Moreover, the galaxies display an $N/O$ excess, where GP030321 displays the largest value. This behaviour is characteristic of GPs \citep[see][]{amorin_oxygen_2010}. GP030321 galaxy also seems to have an exceedingly large helium abundance. However, as mentioned above, this element displayed a non-uniform line fitting for this galaxy.
\subsection{Ionization source characterization} \label{radiation-character}
The results from the tailored photoionization model grids can be found in Table \ref{tab:photoIonization-param-comparison}: It displays the black body $T_{eff}$ and $log(U)$ measurements for the three GPs using the three techniques: \textsc{HII-CHI-mistry-Teff}, the Bayesian \textsc{HII-CHI-mistry-Teff} implementation and the Bayesian Direct method + \textsc{HII-CHI-mistry-Teff}. The first technique was applied using both spherical and plane parallel model geometries, while the Bayesian approaches only consider the plane-parallel geometry. The \textsc{HII-CHI-mistry-Teff} results predict a very hot ionizing radiation in all three GPs. The spherical geometry assumption consistently displays lower temperatures with $\overline{T_{eff,\, sph}}=57000\pm2800$, while the plane-parallel geometry stays at $\overline{T_{eff,\, pp}}=61400\pm3100$. The ionization parameter values are very similar between both geometries with $\overline{log(U)}=-2.58\pm0.26$. Given the results uncertainties and taking into account the fact that real nebulae lay somewhere between these ideal geometries, it can be concluded that the grids geometry has a very weak impact on the measurements.
Carrying on with the measurements from the Bayesian \textsc{HII-CHI-mistry-Teff} implementation (column (5) in Table \ref{tab:photoIonization-param-comparison}) it can be appreciated consistent higher values for both parameters with $\overline{T_{eff}}=69915\pm369$ and $\overline{log(U)}=-2.28\pm0.01$. This behaviour was not observed on the synthetic test cases where both techniques display a good agreement. To gain some insight on the cause behind these discrepancies, the reader is advised to check the data in Table \ref{tab:photoIonization-flux-comparison}, which compares the fluxes fitted by each technique. It should be explained that the \textsc{HII-CHI-mistry-Teff} grids considers the combined flux for the $[SII]6717,6731A$ lines. This table corresponds to the measurements for GP101157. However, the corresponding tables for the rest of the sample can be found on this manuscript online support material. The percentages at each flux right-hand side quantify the difference with the observed value. It can be appreciated that all techniques display good accuracy for the collisionally excited lines. However, the $HeI$ and $HeII$ emission are poorly fitted. This pattern is repeated in the three GPs. As discussed in the previous section, the Bayesian direct method encountered large discrepancies in the $HeI$ lines for GP030321 and GP101157. However, the disagreement was also appreciated in GP121903. In this GP, the maximum discrepancy in the $HeI$ lines was below 4\% for the Bayesian direct method, while the $HeI$ fluxes predicted by the three \textsc{HII-CHI-mistry-Teff} implementations are below half the observational value. The poor fitting in these lines could explain the different results on the $T_{eff}$ and $log(U)$ measurements.
As discussed by \citet{stasinska_what_2007}, at this point we face the three challenges in photoionization modelling: Parameter space exploration, error bar characterization and model validity. Regarding the second point, the error bar characterization, the discussion in section \ref{stellar-discussion} has already illustrated this issue: Since the error propagation is lost in stellar and nebular continua computation, the impact from the underlying stellar population might be severely underestimated. While this may not be an issue for the strongest $HI$ transitions, this systematic uncertainty in the $HeI$ and $HeII$ fluxes should not be neglected. This demands the development of new techniques, which include the galaxy continua fitting in the helium chemical analysis. Regarding the third challenge, validating the ionization model for the HeII emission is still an ongoing effort both in the local \citep[see][]{shirazi_strongly_2012,senchyna_ultraviolet_2017} and high-z galaxies \citep[see][]{berg_window_2018,nanayakkara_exploring_2019}.
To approach the first issue: Parameter space exploration, we propose to combine the \textsc{HII-CHI-mistry-Teff} model with the chemical sampler in section \ref{chemical_analysis}. This approached is labelled as Bayesian direct method + \textsc{HII-CHI-mistry-Teff}. In this scheme, both model dimensions are sampled simultaneously bringing the number of fitted parameters to 16. To the best authors knowledge, this is the largest chemical model for the fitting of the recombination and collisionally excited fluxes. Moreover, it also represents the first attempt to simultaneously include ionization parameters in a direct method fitting. The closest analogue to this methodology (in addition to \textsc{HII-CHI-mistry}) is the BOND (Bayesian Oxygen and Nitrogen Determination) algorithm by \citet{vale_asari_bond:_2016}. The program derives the oxygen and nitrogen abundances via the strong line method paradigm. However, their code also considers the relative $HeI$ fluxes to anchor the hardness of the radiation field. Checking column (6) in Table \ref{tab:photoIonization-param-comparison}, we can appreciate that the combined technique results in $T_{eff}$ and $log(U)$ measurements similar to those from the Bayesian \textsc{HII-CHI-mistry-Teff} algorithm. Moreover, the flux comparison in Table \ref{tab:photoIonization-flux-comparison} leads to the conclusion that once both techniques are combined the observational data is better represented for both recombination and collisionally excited lines. Nonetheless, a careful analysis on these results points towards discrepancies between the theoretical model and the observational data:
\begin{itemize}
\item The difference between the fluxes predicted by the Bayesian direct method + \textsc{HII-CHI-mistry-Teff} algorithm is still higher than in the original Bayesian direct method. It is thus necessary to confirm that this is not a mathematical issue, since as the number of dimensions increases, so is the quality of the fitting is expected to rise. For example, in the Bayesian direct method fittings, there is usually a very good match for input spectra where an ionic species is only responsible for a unique emission line. In this case, the algorithm has a simple job to find the abundance, which best simulates the observed flux; even as the shared parameters vary. This same pattern may be happening in the photoionization model sampling. Consequently, several lines should be included for every species, whenever possible. This conclusion also applies to both the input spectra and the considered theoretical photoionization models.
\item In the case of GP101157, the fitting results in double peak posteriors for the model parameters. This is because the sampling is struggling to find the model solution at two different coordinates. However, looking at the emission flux posteriors, only the $[OIII]$ and $HI$ transitions show a clear double peak shape. This means that in the current model the $\nicefrac{O}{H}$ and $c(H\beta)$ sampling have the largest impact. Since this phenomenon is only appreciated in the case of GP101157, it may be argued that this galaxy displays an abnormal behaviour (which could also explain the divergence on the $HeI$ fluxes). Nonetheless, it is necessary to check if the oxygen abundance has a disproportionate weight on the grid interpolation. Since it is well known that in the very low metallicity region the chemical enrichment is linear \citep[see][]{peimbert_chemical_1969,peimbert_chemical_2000}, the photoionization grids could be interpolated as function of the total metallicity. Afterwards, the metallicity could be expressed as a linear function of all the fitted abundances. This way, we could guarantee that all species contribute to the abundance interpolation.
\item Both the $T_{eff}$ and $log(U)$ measurements do not change dramatically in the Bayesian direct method + \textsc{HII-CHI-mistry-Teff} fitting even though the $HeI$ and $HeII$ fluxes are closer to the observational values. These lead to two possible conclusions: The first one is that the model sampling is failing to establish the GPs ionizing radiation strength from the helium photons. To confirm this is not an issue, the photoionization model should be tested with several diagnostics for the same photoionization grids. For example, the $T_{eff}$ could be characterised as a function of the softness parameter \citep[see][]{vilchez_determination_1988}. Similarly, the ionization parameter could be expressed as a function of the $\nicefrac{O^{+}}{O^{2+}}$ or $\nicefrac{S^{+}}{S^{2+}}$ ratios. The second scenario, however, is that $HeII$ fluxes cannot be fully explained by the photoionization model. Indeed, radiation only driven by stellar atmospheres usually fails to explain these ions origin \citep[see][]{stasinska_modeling_2003,kehrig_extended_2015,kehrig_extended_2018}. This result suggests that additional phenomena (shocks, photon leakage, nearly metal-free hot stars, X-ray binaries...) are necessary to account for the helium photons excess in star-forming galaxies. Recently, the models from \citet{simmonds_can_2021} have shown that ultra-luminous X-ray sources may provide the necessary ionizing radiation for the observed HeII fluxes. The next step in this line of work, is to enhance the current fitting algorithm for such grids. Afterwards, it will be possible to consistently compare their accuracy for all the available lines.
\end{itemize}
Table \ref{tab:fit_results_large} displays the complete results from the Bayesian direct method + \textsc{HII-CHI-mistry-Teff} fitting. We find that the new model predicts slightly higher densities and lower temperatures. The largest discrepancies are found on the electron temperature for GP030321 and GP101157, which are significantly lower in the new fitting $(\Delta T_{high}\approx 1000K)$. This has the expected effect on the metal abundances. Taking into consideration the points previously discussed, this result is caused by the fitting of the $[OII]3726\AA,3720\AA$ and $[OIII]5007\AA$ lines as the sampling is dominated by the photoionization model grids. However, in the case of GP12903, the galaxy with the best match between the observed fluxes and the theoretical ones, both model techniques provide consistent results. These results support the combination of both techniques. At the current time, however, we shall favour the results from the isolated Bayesian direct method and the \textsc{HII-CHI-mistry-Teff} algorithm with spherical geometry. To further explore the physical validity of this technique, we shall work on larger photoionization models, which include all the emission lines observed. Needless to say, such large simulations must be accompanied by many test cases to diagnose the previous issues.
\section{Summary and Conclusions} \label{conclusion}
In this paper, we present deep long-slit optical spectra from the OSIRIS instrument at the 10-m GTC for three Green Pea galaxies, GP030321,GP101157 and GP121903. In addition to the application of state-of-the-art tools from the literature, a goal of this study is to apply novel algorithms for the ionised spectrum fitting. The GP sample treatment has its dust extinction quantified, its nebular and stellar continua fit and their chemical compositions characterised. Additionally, we apply the same methods for the reanalysis of another three GPs from \citetalias{amorin_star_2012}, which were observed using the same instrument and setups: GP004054, GP113303 and GP232539. The new galaxies redshift stays in the $0.144 < z < 0.196$ interval. The main conclusions from this work are:
The galaxy sample is characterized by strong emission lines $(EqW([OIII])~2165-2926\text{\AA}$; $EqW(H\beta)\approx595-739\text{\AA})$ and faint continua. Our deep spectra allow us to detect faint lines, such as HeII and several metal auroral lines. Unlike in the three GPs presented in \citetalias{amorin_star_2012}, the new sample continua do not display stellar absorptions or WR bumps.
Two approaches are considered in GP continua analysis. The first one consists in the nebular continua calculation from first principles using the chemical properties measured from the gas. Afterwards, a stellar population synthesis is applied on the observed spectrum without the gas component, using \textsc{Starlight}. The output is used to correct the recombination lines from underlying stellar absorption before repeating the process. The second methodology consists in the application of self-consistent population synthesis code \textsc{FADO}. In addition to the inbuilt nebular continuum computation, this algorithm was used to fit the largest SSP library ever considered for GPs: a total 1098 stellar spectra grid with fine time steps. Both techniques display consistent results despite the different methodologies and inputs. However, the \textsc{FADO} implementation provides a simpler and more robust workflow. The results are the ones used for the SFH analysis, while those from the iterative approach are used in the chemical study.
In GP101157 and GP121903, the new methodology provides very good agreement with the stellar masses presented in \citetalias{amorin_star_2012}. In the case of GP004054, however, the stellar mass is a $13\%$ lower than the one measured in \citetalias{amorin_star_2012}. This is because the older stellar fraction is not fitted. In general, however, the new approach succeeds at fitting stellar populations from a few hundred thousand years old to several Gyr. The light fraction is dominated by the $6.5 < \Delta log(t) < 7.5\, yr$ age range, while younger stars only account for a few percentage points. Older stellar populations dominate the mass and they formed several Gyr ago. Unlike in the younger systems, old stars do not share a common distribution for the mass and light fraction histograms. Instead, a single SSP dominates the light fraction, with a value of the same order as the luminosity from the younger stars. This points towards an incomplete characterisation from the gas contribution: either an underestimation in our nebular continuum calculation, or an overestimation on the older stars SSP synthesis due to an underestimated contribution from the ionizing stellar population. The second scenario would require unaccounted phenomena such as photon leakage or the presence of almost pristine and very massive stars.
The algorithm \textsc{HII-CHI-mistry-Teff} is applied to measure the $T_eff$ and $log(U)$ from the ionizing cluster. Using the plane-parallel geometry and $30 - 90\,KT$ black body grids the model predicts very hard ionizing sources for the GP sample with $\overline{T_{eff,\, pp}}=61431\pm3076$ and $\overline{log(U)}=-1.98\pm0.24$. Additionally, using the neural networks grid interpolator from the library \textsc{exoplanet}, we propose a Bayesian implementation for \textsc{HII-CHI-mistry-Teff}. In synthetic test cases, this version provides the same accuracy for a much faster treatment. Once applied to the GPs sample, a harder ionization source is predicted with $\overline{T_{eff,\, pp}}=69915\pm369$ and $\overline{log(U)}=-2.28\pm0.24$. Both approaches underestimate the observed $HeI$ and $HeII$ fluxes. This suggest the existence of additionally mechanisms enhancing the helium transition photons.
Using the tools mentioned above, the Bayesian direct method sampler from \citet{fernandez_bayesian_2019} has been enhanced to interpolate emissivity grids during the sampling process. Additionally, it has been adapted to include other ionic species. In the current GP sample, the maximum number of dimensions is 14. The results show a low metallicity regime with high ionization temperatures, $T_{e,[OIII]} > 14000\,K$. As a final exercise, this model was combined with Bayesian \textsc{HII-CHI-mistry-Teff} algorithm. This represents the first attempt to fully solve the direct method parameter space alongside photoionization grids parameter space. The output $T_{eff}$ and $log(U)$ values agree with those from the Bayesian \textsc{HII-CHI-mistry-Teff} implementation but the measured $T_{[OIII]}$ is lower than in the original Bayesian direct method for two of the three galaxies. The predicted emission fluxes also display a slightly larger discrepancy with the observed fluxes compared with the original Bayesian direct method. At this point, further work is necessary to find the best design to include tailored photoionization grids in the chemical model fitting.
This study represents an ongoing effort to improve our understanding on the SFH and the chemical evolution of star-forming systems. These conclusions emphasize the importance of combining deep, high S/N spectra from large telescopes with sophisticated algorithms and models. The insights gained lead to the conclusion that the default models for star forming regions and gas emission fail to explain some biases encountered in GPs. Fortunately, the novel methodologies proposed are not limited to the size of input data. This will make possible the solution of such complex models.
\section*{Acknowledgements}
We are indebted to an anonymous referee, whose detailed comments and encouragement contributed to the improvement of the paper.
This work was funded by the FONDECYT Postdoc 2020 project 3200340 from the National Agency for Research and Development (ANID). R.A. acknowledges support from ANID Fondecyt Regular 1202007. P.P. acknowledges support through FCT grants UID/FIS/04434/2019, UIDB/04434/2020, UIDP/04434/2020 and the project "Identifying the Earliest Supermassive Black Holes with ALMA (IdEaS with ALMA)" (PTDC/FIS-AST/29245/2017). EPM, CK, and JVM acknowledge financial support from the State Agency for Researchof the Spanish MCIU through the “Center of Excellence SeveroOchoa” award to the Instituto de Astrofisica de Andalucia (SEV-2017-0709).
\section*{Data Availability}
The authors make available with this manuscript the following data on the corresponding MNRAS supplementary material section: 1) The emission line fluxes in machine readable format. 2) The full results from the Bayesian fittings both in graphical and tabulated format. 3) The complete infographics from the FADO synthesis in portable document format (PDF). The observations files can be downloaded from the GTC virutal observatory at https://gtc.sdc.cab.inta-csic.es/gtc/jsp/searchform.jsp. The program ID is GTC63-10B.
\bibliographystyle{mnras}
|
\section{Introduction}
Test of independence is a fundamental question in data analysis and statistical inference. Considering two multivariate random vectors $X $ and $Y$, we are interested in testing whether the two random vectors are independent, namely, $H_0: X \bot\!\!\!\bot Y$. Such testing problems are relevant in many statistical learning problems, including variable selection in regression, Gaussian graphical models, Markov random fields, and causal inference \citep{fan2020statistical, maathuis2018handbook, imbens2015causal}. In traditional statistical literature, one may choose the Pearson correlation to measure the independence between $X$ and $Y$ when the data has a jointly normal distribution, or opt for the rank correlation when both $X$ and $Y$ are univariate. With the development of information technology, researchers are now able to collect complex and potentially high dimensional data with potentially highly nonlinear dependence. How to perform tests of independence for modern data is a challenging and important problem in the contemporary statistical community.
In the past two decades, there have been a series of substantial developments in the testing of independence for general $X$ and $Y$ without assuming their parametric distributions. A natural starting point is to study the difference between $\mathcal P_{X,Y}$, the joint measure of $(X,Y)$, and $\mathcal P_X\times\mathcal P_Y$, the product measure of $X$ and $Y$. In one of the most well-known papers on this topic, \cite{szekely2007measuring} proposed the distance correlation by measuring the weighted integrated squared difference between the characteristic functions of $\mathcal P_{X,Y}$ and $\mathcal P_X\times \mathcal P_Y$, which is later shown to be equivalent to the maximum mean discrepancies in the machine learning community \citep{sejdinovic2013equivalence}, and closely related to the Hilbert-Schmidt independence criterion \citep{gretton2005measuring}. Extensions of distance correlation have been widely discussed \citep{szekely2013distance, huo2016fast, yao2018testing}. \cite{zhu2017projection} relaxed the moment constraint in distance correlation by combining the Hoeffding coefficient with projection pursuit. Other than comparing characteristic functions, there are also novel methods that compare the density functions \citep{berrett2019nonparametric}, and the cumulative distribution functions \citep{heller2012consistent,cui2019distribution, moon2020interpoint}. \cite{kong2019composite} and \cite{chatterjee2021new} used the appealing idea of conditional mean variance to evaluate the dependence between two random variables. More recently, \cite{shi2020distribution} and \cite{deb2021multivariate} developed the first distribution-free independence test for multivariate random vectors. They define multiple ranks using the theory of measure transportation and propose (multivariate) rank versions of distance covariance and energy statistic for independence testing. But in practice, the computation for measure transportation will grow quickly with the sample size and dimension, which restricts the application of those two tests to large-scale datasets. High dimensional independence test has recently been studied by \cite{zhu2020distance} and \cite{gao2021asymptotic}. In comparison, our work is more generally applicable as we allow the dependence signal in high dimensional vectors to be very sparse, which is a benefit of implementing the advanced machine learning algorithms.
Our work is motivated by challenges arising in single-cell multimodal omics, a research area labeled `Method of the Year 2019' by Nature Methods. This technological advance builds on the recent breakthroughs in sequencing the RNA of single cells and promises greater insights into gene regulatory networks, cell lineages, and trajectories by permitting the measurement of multiple omics on the same cell \citep{zhu2020single, schier2020single}. Of particular interest are simultaneous measurements of gene expression (RNA-seq) and chromatin accessibility (ATAC-seq). ATAC-seq identifies active regulatory sequences in the genome by finding open chromatin, which determines whether a gene will be actively transcribed. For this reason, it is widely assumed that RNA-seq and ATAC-seq will co-vary. But both data sources tend to be high dimensional and extremely sparse, positing great challenges to performing statistical independence tests for the two random vectors. For example, the data we analyze consists of 11,188 blood cells, each with RNA-seq and ATAC-seq read counts. The dimension of RNA-seq is 29,717 and the dimension of ATAC-seq is 143,887. Only 6.35\% entries in the RNA-seq and 5.66\% entries in the ATAC-seq are non-zero, making all current independence testing methods practically infeasible.
The purpose of this paper is to build a distribution-free test of independence that is powerful even under high dimensional, complex data.
Existing methods use U-statistics to directly estimate the integrated squared difference between the joint distribution and the product distribution, in the forms of characteristic functions, density functions, or cumulative distributions. Such U-statistics often fail to pick up the hidden signal when there are many noise dimensions in the data, and often require cumbersome resampling procedures to calibrate the null distribution. Our proposal deviates from these methods by aiming at a different and adaptive quantity: Instead of the integrated squared difference between distribution functions, our method seeks to find \emph{any} potential difference between the joint and product distributions by constructing a classification problem between these two distributions.
By leveraging recent developments in two sample testing and sample splitting \citep{kim2019global, hu2020distribution, kim2021classification}, we develop a test that is more flexible and can borrow strength from the most powerful classification tools, such as deep neural networks, from the machine learning community. It is particularly powerful for high dimensional data when proper regularizations (such as sparsity) are enforced on the classifier.
The proposed method consists of three steps: sample splitting, classification, and rank-sum comparison. We fist split the index set $\mathcal{I} = \{1, \dots, n\}$ into two subsets $\mathcal{I}_1 = \{1, 2, \dots, n_1\}$ and $\mathcal{I}_2 = \{n_1+1, \dots, n\}$. Let $D_{1A} = \{(X_i, Y_i), i\in\mathcal{I}_1\}$ and $D_{2A} = \{(X_i, Y_i), i\in\mathcal{I}_2\}$ be the two subsets of the data. Then we generate two correspondingly permuted datasets by cyclically permuting $Y$ in each of the two subsets. Let $D_{1B} = \{(X_i', Y_{i}'), i\in\mathcal{I}_1\}$ and $D_{2B} = \{(X_i', Y_i'), i\in\mathcal{I}_2\}$, where $X_i'=X_i$ for all $i$, $Y_i'=Y_{i+1}$ for $i\notin \{n_1,n\}$, and $Y_{n_1}'=Y_1$, $Y_n'=Y_{n_1+1}$. In the classification step, we train a classifier that aims to distinguish $D_{1A}$ from $D_{1B}$, because the sample points in $D_{1A}$ are generated from $\mathcal P_{X,Y}$ while those in $D_{1B}$ have marginal distribution $\mathcal P_{X}\times\mathcal P_Y$ and weak dependency between sample points. Next, in the rank-sum comparison step we compare the predicted class probabilities in $D_{2A}$ and $D_{2B}$. Under $H_0$, the predicted class probabilities of $D_{2A}$ and $D_{2B}$ should have the same distribution, while under $H_1$, those predicted probabilities of $D_{2A}$ and $D_{2B}$ should be different if the classifier is able to pick up the difference between $\mathcal P_{X,Y}$ and $\mathcal P_X\times\mathcal P_Y$. This intuition motivates a rank-sum test to compare the predicted class probabilities of the two samples. The main technical challenge is that the sample points in $D_{2A}$ and $D_{2B}$ are dependent, thus classical U-statistics theory can not be directly applied. Our theoretical development uses Hoeffding's projection to decompose the test statistic into sums of sparsely dependent random variables, and uses a version of Stein's method for sparsely dependent data to establish the normal approximation of the test statistic.
To sum up, the proposed method has the following advantages.
(i) {\it Completely nonparametric.} We require very few assumptions on the data to ensure the test's validity. Under $H_0$, the type I error control is automatically guaranteed by sample splitting and the single permutation. Under $H_1$, the test will have good power as long as the classifier is better than a random guess, which is practically feasible given the powerful neural networks.
(ii) {\it Asymptotic distribution-free and computationally efficient.} Our test statistic has a standard normal asymptotic null distribution. This is in critical contrast to other current independence tests that have non-explicit distributions and require the computationally expensive bootstraps to obtain $p$-values \citep{szekely2007measuring, heller2012consistent, berrett2019nonparametric}. For the most recent distribution-free independence tests \citep{shi2020distribution, deb2021multivariate}, the limiting null distributions are still weighted $\chi^2(1)$, without an analytic form. Although \cite{shi2020distribution} listed the thresholds for some combinations of dimensions of $X$ and $Y$, it still needs at least one round of numerical approximation when the dimensions exceed those in \cite{shi2020distribution}. Such improved computational efficiency makes our method particularly appealing for the aforementioned single cell sequencing data.
(iii) {\it Applicability to high dimensional data.} The test is suitable for high dimensional data. Existing tests based on degenerate U-statistics are hard to apply and have limited power when the data dimension is high and the dependence signal is very sparse. By taking the classification perspective, we can take advantage of adaptive and structured classifiers to pick up weak signals from high dimensional data. Moreover, our framework allows $X$, $Y$ to take value in infinite-dimensional spaces, as long as the likelihood ratio is well defined.
(iv) {\it Flexibility and generality.} The method described in this paper is just one example from a general framework. All three steps (permutation, classification, and calibration) can be carried out with other variants that are more suitable to the problem at hand. For example, one can use other dimension reduction or variable selection methods when distinguishing the two distributions, and/or use different two-sample testing methods, such as two-sample $t$-test, to calibrate the significance of classification. When the original sample $(X_i, Y_i)$ has a time-series or random field structure as the index $i$ changes from $1$ to $n$, one can also consider other types of permutations that are more suitable for the particular dependence structure across sample points.
\section{Test of Independence by Sample Splitting and Classification}
\subsection{Preliminaries and basic ideas}
Consider independent observations $\{(X_i, Y_i):1\le i\le n\}$ of a pair of random variables $X$ and $Y$ with joint distribution $\mathcal{P}_{X,Y}$ in a space $\mathcal X\times\mathcal Y$. Let $\mathcal{P}_X$ and $\mathcal{P}_Y$ be the marginal distributions of $X$ and $Y$ respectively. We are interested in testing
\begin{eqnarray*}
H_0: \mathcal{P}_{X,Y} = \mathcal{P}_X \times\mathcal{P}_Y\quad\mbox{versus} \quad H_1: \mathcal{P}_{X,Y} \neq \mathcal{P}_X \times\mathcal{P}_Y,
\end{eqnarray*}
where $\mathcal{P}_X \times\mathcal{P}_Y$ denotes the product distribution.
Most existing methods for independence testing focus on a quantity of the form $$\int w(x,y)\phi(G(x,y),G_1(x)G_2(y))dxdy,$$
where $G(\cdot)$, $G_1(\cdot)$, $G_2(\cdot)$ are joint and marginal distribution functions, $w$ is a weight function, and $\phi$ is a discrepancy measure. This framework covers nearly all the popularly studied independence testing methods, including distance correlation \citep{szekely2007measuring}, Hilbert-Schimidt independence criterion \citep{gretton2005measuring,gretton2007kernel}, rank-correlation based methods \citep{heller2012consistent,moon2020interpoint}, and mutual information based methods \citep{berrett2019nonparametric}. While enjoying elegant theoretical properties, these methods rely on specific choices of $w$, $\phi$, and $G$ functions, making them hard to apply for high-dimensional, complex data. Moreover, the null distributions of the corresponding test statistic usually depend on the unknown underlying distribution $\mathcal P_{X,Y}$ and must be approximated using resampling methods.
The key feature of our method is that it does not rely on a pre-chosen set of functions $(w,\phi, G)$. Instead, our method begins with fitting a flexible classifier to distinguish $\mathcal P_{X, Y}$ and $\mathcal P_X\times\mathcal P_Y$, and then tests whether the fitted classifier does anything different from random guessing.
Suppose we have two equal-sized samples, one from $\mathcal P_{X,Y}$ and one from $\mathcal P_X\times\mathcal P_Y$, and we associate a label $K=1$ ($K=0$) for each sample point from $\mathcal P_{X,Y}$ ($\mathcal P_X\times \mathcal P_Y$). We will discuss how to obtain these samples in the next subsection.
Under $H_0$, the two samples have the same distribution $\mathcal P_X\times\mathcal P_Y$, so any classifier trying to distinguish these two samples would behave like a random guess. On the other hand, under $H_1$, any classifier that can detect the difference between these two distributions should do better than random guess, which can be tested on a holdout pair of samples from the two distributions.
More specifically,
the conditional label probability
$$
\theta(x,y)=\mathbb P(K=1|x,y)
$$
is related to the likelihood ratio
\begin{eqnarray}\label{likelihood_ratio}
L(x,y)\stackrel{\mbox{\textrm{\rm\tiny def}}}{=} \frac{\theta(x,y)}{1-\theta(x,y)} = \frac{d\mathcal{P}_{X, Y}}{d(\mathcal{P}_X\times \mathcal{P}_Y)}.
\end{eqnarray}
Therefore, $\theta(x,y)$ reduces the data dimension to $1$, while largely capturing the difference between $\mathcal{P}_{X,Y}$ and $\mathcal{P}_X\times\mathcal{P}_Y$ as guranteed by the following result.
Under the null hypothesis, $\theta(x,y)\equiv 1/2$ and the likelihood ratio $L(x,y)\equiv 1$, which corresponds to a degenerate case.
\begin{proposition}\label{pro:separation}
Let $\mathcal P$, $\mathcal Q$ be two probability distributions on a common measurable space such that $\mathcal P\ll \mathcal Q$ and
the Radon-Nikodym derivative $d \mathcal P / d\mathcal Q$ has a continuous distribution under $\mathcal Q$. Let $V\sim \mathcal P$ and $W\sim\mathcal Q$ be independent and $d_{tv}(\cdot,\cdot)$ be the total variation distance between two probability measures, then
$$
\frac{1}{4}d_{\rm tv}(\mathcal P,\mathcal Q)\le\frac{1}{2}- \mathbb P\left\{\frac{d\mathcal P}{d\mathcal Q}(V)<\frac{d\mathcal P}{d\mathcal Q}(W)\right\}\le \frac{1}{2}d_{\rm tv}(\mathcal P,\mathcal Q)\,.
$$
\end{proposition}
\begin{remark}[Dropping the continuity assumption] If $(d\mathcal P/d\mathcal Q)(W)$ has point mass, then it is possible to have $(d\mathcal P/d\mathcal Q)(V)=(d\mathcal P/d\mathcal Q)(W)$. In this case one can associate each of $V$ and $W$ with an independent $U(0, 1)$ random variable, $\zeta$ and $\eta$, and rank them with randomized tie-breaking
\begin{eqnarray*}
\indic\left\{\frac{d\mathcal P}{d\mathcal Q}(V)<\frac{d\mathcal P}{d\mathcal Q}(W)\right\}
+\indic(\zeta<\eta)\indic\left\{\frac{d\mathcal P}{d\mathcal Q}(V)=\frac{d\mathcal P}{d\mathcal Q}(W)\right\}\,.
\end{eqnarray*}
All the theory, including \Cref{pro:separation}, goes through the same for such a random tie-breaking ranking scheme with more careful bookkeeping. Therefore, in the rest of this paper, we will proceed under the assumption that $\theta(X,Y)$ and its estimate $\widehat\theta(X,Y)$ are continuous under $\mathcal P_X\times\mathcal P_Y$ for notational simplicity.
\end{remark}
Such a classification-testing procedure consists of a fitting part and testing part, which need to be carried out on separate subsamples.
Splitting the sample reduces the sample size used for both classification and testing. But the benefits are quite substantial: First, in high-dimensional data, the signal is often quite weak and concentrates on a low-dimensional subspace or submanifold hidden in the high-dimensional ambient space. It is often more efficient to find out the direction of the signal and then conduct hypothesis tests targeted specifically in that signal direction. The reduced sample sizes can be viewed as our investment in finding the most promising direction of the signal. Second, sample splitting provides great flexibility in the choice of classification algorithms, such as black-box methods and deep neural networks, which are particularly powerful in handling complex data.
Even if we split the sample to carry out the classification and test, another challenge remains: How do we obtain samples from the two distributions $\mathcal P_{X,Y}$ and $\mathcal P_X\times\mathcal P_Y$, as required by both the classification and the testing steps? We provide a sample-size efficient answer to this question in the next subsection.
\subsection{Sample Splitting and Cyclic Permutation}
As discusssed in the previous subsection, the classification and testing procedures need to be carried out on separate subsamples to ensure the validity. Suppose we split the index set $\mathcal{I} = \{1, \dots, n\}$ into two subsets $\mathcal{I}_1 = \{1, 2, \dots, n_1\}$ and $\mathcal{I}_2 = \{n_1+1, \dots, n\}$, $n_2 = n - n_1$, so that the subsample $\mathcal I_1$ is used for classification and $\mathcal I_2$ is used for testing.
However, after such a sample split we still do not have a sample from $\mathcal P_X\times\mathcal P_Y$ for classification or testing. A simple idea is to further split $\mathcal I_1$ into $\mathcal I_{11}$ and $\mathcal I_{12}$, and permute the sample pairs in $\mathcal I_{12}$ to form a sample from $\mathcal P_X\times \mathcal P_Y$. A similar second split and permutation can be applied to $\mathcal I_2$ for the testing purpose.
Although this approach is simple and straightforward to implement, it further splits an already reduced sample size. A natural question is whether one can avoid such a second split and use the sample more efficiently. We provide a positive answer below.
To avoid the second split,
denote $D_{1A} = \{(X_i, Y_i), i\in\mathcal{I}_1\}$ the subsample in $\mathcal I_1$, and its cyclicly permuted version $D_{1B} = \{(X_i', Y_{i}'), i\in\mathcal{I}_1\}$, where $X_i'=X_i$ for all $i$, $Y_i'=Y_{i+1}$ for $1\le i\le n_1-1$, and $Y_{n_1}'=Y_1$. Similarly $D_{2A}=\{(X_i,Y_i):i\in\mathcal I_2\}$ denotes the subsample in $\mathcal I_2$, and its cyclicly permuted version $D_{2B} = \{(X_i', Y_i'), i\in\mathcal{I}_2\}$, with $X_i'=X_i$ for all $i$, $Y_i'=Y_{i+1}$ for $n_1+1\le i\le n-1$, and $Y_n'=Y_{n_1+1}$. Our plan is to treat $D_{jA}$, $D_{jB}$ as approximately independent samples from $\mathcal P_{X,Y}$ and $\mathcal P_X\times\mathcal P_Y$ for classification ($j=1$) and two-sample testing ($j=2$), because the dependence between the original and cyclicly permuted samples are very sparse.
Suppose we apply a classification algorithm on $D_{1A}$, $D_{1B}$ with labels $K=1$ for sample points in $D_{1A}$ and labels $K=0$ for those in $D_{1B}$, resulting in a function estimate $\widehat\theta(x,y)$ of $\theta(x,y)=\mathbb P(K=1|x,y)$ as defined in \eqref{likelihood_ratio}. To test the significance of the classifier, we use the rank-sum statistic
\begin{equation}\label{eq:R}
R = \frac{1}{n_2^2}\sum_{i,j\in\mathcal I_2}\indic\left\{\widehat\theta(X_i, Y_i)<\widehat\theta(X_j', Y_j') \right\}\,.
\end{equation}
If $\widehat\theta$ is close to $\theta$ under $H_1$ then \Cref{pro:separation} suggests we should reject $H_0$ if $R$ is too small. As detailed in the next subsection, combining the two-sample $U$-statistic theory and Stein's method for sparsely dependent random variables, we have the following asymptotic scaling of $R$ under $H_0$:
$$
\mathrm{Var}(\sqrt{n_2}R) \approx \hat\sigma^2
$$
with
\begin{eqnarray}\label{eq:sigmahat}
\widehat\sigma^2 = \frac{1}{6} - \frac{2}{n_2}\sum_{i=n_1+1}^n \widehat h_1(X_i, Y_i)\widehat h_1(X_i', Y_{i}')- \frac{2}{n_2}\sum_{i=n_1+1}^n \widehat h_1(X_{i+1}, Y_{i+1})\widehat h_1(X_i', Y_i')\,,
\end{eqnarray}
where $\widehat h_1(x,y)=1/2-\widehat F_{2*}(\widehat\theta(x,y))$, with $\widehat F_{2*}$ the empirical distribution function of $\{\widehat\theta(X_i',Y_i'):i\in \mathcal I_2\}$, and using the convention $(X_{n+1},Y_{n+1})=(X_{n_1+1},Y_{n_1+1})$.
Thus we arrive at the following split-permute-classification-test procedure.
\begin{algorithm}[h]
\caption{Test of independence via classification significance}
\label{alg1}
\begin{algorithmic}
\STATE 1. Input data $D = \{(X_1, Y_1), \dots, (X_n, Y_n)\}$, classificaion algorithm $\mathcal A$.
\STATE 2. Split $\{1,\dots,n\}$ into subsets $\mathcal{I}_1$ and $\mathcal{I}_2$ to form subsamples $D_{1A}$, $D_{2A}$, and cyclicly permuted subsamples $D_{1B}$, $D_{2B}$ as described above.
\STATE 3.
Apply $\mathcal A$ to $D_{1A}$ and $D_{1B}$ to obtain the estimated class probability function $\widehat\theta(\cdot, \cdot)$;
\STATE 4. Calculate the $p$-value $\stackrel{\mbox{\textrm{\rm\tiny def}}}{=}\Phi\{\sqrt{n_2}(R-1/2)/\hat\sigma\}$ with $R$, $\hat\sigma$ given by \eqref{eq:R} and \eqref{eq:sigmahat}, where $\Phi(\cdot)$ is the standard normal distribution function.
\end{algorithmic}
\end{algorithm}
\noindent
{\bf Remark 1: split ratio.} To implement Algorithm \ref{alg1}, one needs to choose the sizes of $\mathcal{I}_1$ and $\mathcal{I}_2$. While a large $\mathcal{I}_1$ will train a more accurate classifier, it also leads to a smaller testing data set $\mathcal{I}_2$. Thus it is important to balance the trade-off between classification and testing data. In our simulations, we found an equal-split performs very well. Without further notations, we assume $|\mathcal{I}_1| = |\mathcal{I}_2|$ throughout the paper.
\noindent
{\bf Remark 2: choice of the classifier.} In principle, our method can work with any classification algorithm $\mathcal A$. However, the classification problem in our method is quite challenging. By construction, each coordinate in the two populations $D_{1A}$, $D_{1B}$ have the same mean value, and the only difference is the dependence structure among the columns. Therefore, linear methods such as logistic regression cannot perform very well, and nonlinear methods such as support vector machine \ would require a good choice of kernel. In practice, we choose neural networks due to their great flexibility and adaptivity to complex structures in the data.
\section{Theoretical Justifications}\label{sec:theory}
In the split-permute testing procedure described in \Cref{alg1}, both the classifier and two-sample test are obtained using an originally paired subsample together with its cyclicly permuted version. Therefore the samples are not completely independent and the theoretical properties of the resulting test statistic deserve careful analysis. We first establish the asymptotic conditional distribution of the test statistic conditioning on a given fitted label probability function $\widehat\theta$. It turns out that the null asymptotic conditional distribution is independent of $\widehat\theta$ and asymptotically distribution-free, while the estimated likelihood ratio needs to be better than random guess under the alternative. We will discuss the performance of classification using the cyclic permuted data in \Cref{sec:classification_accuray}.
\subsection{Asymptotic distribution of test statistic}\label{sec:asym_dist}
Before presenting the theoretical results, we describe some necessary notations.
Let $F_{1*}(\cdot)$, $F_{2*}(\cdot)$ be the cumulative distribution functions of $\widehat\theta(X, Y)$ under $\mathcal P_{X,Y}$ and $\mathcal P_X\times\mathcal P_Y$, respectively.
Let $\E_*(\cdot)$, $\mathbb{P}_*(\cdot)$, $\cov_*(\cdot)$ and $\var_*(\cdot)$ denote the conditional expectation, probability, covariance and variance given $\widehat\theta$ (or equivalently, given the first subsample).
For $k = 3,4$, define
\begin{eqnarray}\label{a34}
A_k = \frac{6}{n_2^k\var_*(R)^{k/2}}\Big( \sum_{i\in\mathcal{I}_2} \E_*\left|F_{2*}\{\widehat\theta(X_i, Y_i)\} - \frac{1}{2}\right|^k + \sum_{i\in\mathcal{I}_2}\E_*\left|F_{1*}\{\widehat\theta(X_i', Y_i')\} - \frac{1}{2}\right|^k\Big).
\end{eqnarray}
We first derive the asymptotic behavior of the test statistic under $H_0$. The proof begins with decomposing the U-statistic $R$ into its projection $\tilde{R}$ and the remaining term, as detailed in Lemma \ref{marginal_proj_h1}. Specifically, let
\begin{eqnarray*}
\tilde{R} = \frac{1}{n_2}\sum_{i\in\mathcal{I}_2}\left[\frac{1}{2} - F_{2*}\{\widehat\theta(X_i, Y_i)\} \right] + \frac{1}{n_2}\sum_{i\in\mathcal{I}_2}\left[ F_{2*}\{\widehat\theta(X_i', Y_i')\} - \frac{1}{2}\right]
\end{eqnarray*}
Lemma \ref{marginal_proj_h1} shows that $R - \frac{1}{2} = \tilde{R} +O_p(n_2^{-1})$. Then we prove the conditional Berry-Essen bound of $\tilde{R}$ and the unconditional asymptotic normality of $R$. The theoretical results under $H_0$ are summarized in Theorem \ref{thm: null}.
\begin{theorem}\label{thm: null}
Under $H_0$, assume $(X_i, Y_i)$, $i\in\mathcal{I}_2$ are i.i.d samples from $\mathcal{P}_{X,Y}$, and $\widehat\theta(x, y)$ is a function such that $\widehat\theta(X_1,Y_1)$ is continuous, and $F_{2*}\{\widehat \theta(X, Y)\} \neq g_1(X)+g_2(Y)$ for any $g_1(\cdot)$ and $g_2(\cdot)$.
Then
\begin{eqnarray*}
\sup_{s\in\mathbb{R}}\left| \mathbb{P}_*\left(\frac{\sqrt{n_2} \tilde R }{\sigma_*}\leq s\right) - \Phi(s) \right| \leq c(\sqrt{A_3} +\sqrt{A_4})
\end{eqnarray*}
where $0\leq c<8$ is a constant, $A_3$ and $A_4$ are defined in (\ref{a34}), and
\begin{align*}
\sigma_*^2 :=& \frac{1}{6} - 2\cov_*\left[F_{2*}\{\widehat\theta(X_2, Y_2)\},F_{1*}\{\widehat\theta(X_{1}, Y_{2})\} \right]-
2\cov_*\left[F_{2*}\{\widehat\theta(X_1, Y_1)\},F_{1*}\{\widehat\theta(X_{1}, Y_{2})\} \right]\,.
\end{align*}
Under the additional assumption of $n_2^{1/3}\sigma_*^2\rightarrow\infty$, we have $\hat\sigma/\sigma_*-1=o_{P}(1)$ and the test statistic $\sqrt{n_2}(R - 1/2)/\widehat \sigma$ converges in distribution to $N(0, 1)$ as $n_1$ and $n_2\rightarrow\infty$.
\end{theorem}
We discuss the convergence rate and conditions for Theorem \ref{thm: null} in the following remarks.
\begin{remark}
The right hand side of the Berry-Essen bound in Theorem \ref{thm: null} consists of two terms: $\sqrt{A_3}$ and $\sqrt{A_4}$. Here $\sqrt{A_3}$ is the dominating term, and is of order $n_2^{-1/4}$ when $\sigma_*^2$ is of constant order. We can further improve the bound rate to the classical $n_2^{-1/2}$ and relax the condition on $\sigma_*^2$ to $n_2^{1/2}\sigma_*^2\rightarrow\infty$ by applying Theorem 2.2 of \cite{jirak2016berry}. The cost is a slightly more complicated condition on the constant term in the Berry-Essen bound.
\end{remark}
\begin{remark}
Conditioning on the estimated probability function $\hat\theta$, our test statistic $R$ is a two-sample $U$-statistic. Its asymptotic normality requires its kernel to be non-degenerate, such that the asymptotic variance $\sigma_*^2>0$.
This non-degeneracy condition is further equivalent to $F_{2*}\{\widehat \theta(X, Y)\}$ cannot be written in the form of $g_1(X)+g_2(Y)$ for any functions $g_1,~g_2$, which is mild because $F_{2*}\{\widehat \theta(X, Y)\} = g_1(X)+g_2(Y)$ is equivalent to 1) $g_1(X)+g_2(Y)$ follows $U(0,1)$ and 2) $\widehat \theta(X, Y) = W\{ g_1(X)+g_2(Y)\}$, for some strictly monotone increasing $W: [0,1]\rightarrow[0, 1]$. Common classifiers (logistic regression, random forest, SVM, neural network) can be easily verified to satisfy this non-degeneracy condition.
\end{remark}
\begin{theorem}\label{thm: alt}
Under $H_1$, assume $(X_i, Y_i)$, $i\in\mathcal{I}_2$ are i.i.d samples from $\mathcal{P}_{X,Y}$,
and there exists a strictly monotone function $g$ such that
\begin{eqnarray}\label{eq:alt_thm_condition}
\E_*\left|\frac{g(\widehat\theta(X', Y'))}{1-{g(\widehat\theta(X', Y'))}} - \frac{\theta(X', Y')}{1-\theta(X', Y')}\right| < 1/4-\mu/2 - c
\end{eqnarray}
holds with probability tending to 1 for some positive constant $c$. Here $\mu = \mathbb{P} \{\theta(X, Y)<\theta(X', Y')\}$, with $(X,Y)$, $(X',Y')$ independently generated from $\mathcal P_{X,Y}$ and $\mathcal P_X\times \mathcal P_Y$ respectively.
Then, as $n_1$ and $n_2\rightarrow\infty$, the test statistic $\sqrt{n_2}(R - 1/2)/\widehat \sigma \stackrel{p}{\rightarrow}\infty$.
\end{theorem}
The condition required for the power guaranteee under the alternative is substantially weaker than the asymptotic normality under the null. This is because we no longer need to lower bound the variance term.
It is remarkable that we do not need to assume the classifier to be consistent to have valid type-I and type-II error control. The type I error control is automatically guaranteed by the cyclic permutation and holds for arbitrary classifiers, because under $H_0$, $(X_1,Y_1)$ and $(X_1,Y_2)$ have the same distribution and any classifier will not be able to distinguish the two samples.
For the type II error control, equation \eqref{eq:alt_thm_condition} is much weaker than consistency, as it only requires $\hat\theta$ to be close to $\theta$ up to a monotone transform and within some constant error bound.
These properties are especially appealing in practice. For example, many nonparametric tests that rely on kernel density estimations need to carefully choose the kernel bandwidth to guarantee the correct type-I error rate. In our case, even though the classifier (such as a neural network) may have many tunning parameters to choose from, the test is always valid, and the power is non-trivial whenever the classifier can pick up even only a part of the difference between the joint and product distributions.
Next, we present a local alternative analysis where the dependence signal changes with the sample size. To quantity the signal, we use the likelihood ratio defined in (\ref{likelihood_ratio}). Specifically, consider $(X', Y')$ and $(X'', Y'')$ independently drawn from $\mathcal{P}_X\times \mathcal{P}_Y$. We define
\begin{eqnarray}\label{eq: local_quantity}
\delta = \E\left\{ \left|L(X', Y') - L(X'', Y'')\right| \right\}.
\end{eqnarray}
By Proposition \ref{pro:separation}, we know that $\delta\asymp d_{tv}(\mathcal{P}_{X, Y}, \mathcal{P}_X\times \mathcal{P}_Y)$.
Thus $\delta$ measures the distance between the null hypothesis and the local alternative. $\E\{L(X', Y') \} = 1$ and $\delta=0$ if and only if $\mathbb{P}\{L(X', Y') =1\}=1$, which is equivalent to $H_0$. Our local alternative analysis focuses on the case $\delta\rightarrow 0$ as $(n_1,n_2)\rightarrow\infty$.
We introduce extra notation to analyze the local alternative. Let $\mu = \mathbb{P} \{\theta(X, Y)<\theta(X', Y')\}$, $\mu_* = \mathbb{P}_* \{\widehat\theta(X, Y)<\widehat\theta(X', Y')\}$ and $F_1(\cdot)$, $F_2(\cdot)$ be the cumulative distribution functions of $\theta(X, Y)$ under $\mathcal P_{X,Y}$ and $\mathcal P_X\times\mathcal P_Y$, respectively. And define
\begin{equation}\label{eq:Rp}
R' = \frac{1}{n_2^2}\sum_{i,j\in\mathcal I_2}\indic\left\{\theta(X_i, Y_i)<\theta(X_j', Y_j') \right\}\,.
\end{equation}
Based on equation (\ref{eq: Rp_projection}) in Lemma \ref{marginal_proj_h1}, one can easily calculate the variance for the projection of $\sqrt{n_2}R'$ to be $\sigma_0^2:=\cov(V_1+V_2+V_3,V_2)$, where $V_i = F_{1}\{\theta(X_i', Y_i')\} - F_{2}\{\theta(X_i, Y_i)\}$. While $\sigma_0^2$ is complicated and hard to understand, we also define $\sigma^2$ and show that $\sigma_0^2$ is actually sufficient close to $\sigma^2$ under the local alternative hypothesis. Specifically, $\sigma_0^2$ can be approximated by
\begin{eqnarray*}
\sigma^2 = \frac{1}{6} - 2\cov\left[F_2\{\theta(X_2, Y_2)\},F_1\{\theta(X_{1}, Y_{2})\} \right]-
2\cov\left[F_2\{\theta(X_1, Y_1)\},F_1\{\theta(X_{1}, Y_{2})\} \right]\,,
\end{eqnarray*}
because the joint distribution $\mathcal P_{X,Y}$ gets increasingly closer to the product distribution $\mathcal P_X\times\mathcal P_Y$. For the same reason, $\sigma^2$ further converges to a quantity depending only on the product distribution $\mathcal P_X\times\mathcal P_Y$. Thus it is reasonable to assume the variance term $\sigma_0^2$ is bounded away from zero in the local asyptotic population sequence:
$$
\sigma_0^2\ge c >0
$$
for some constant $c$ not depending on the sample size.
\begin{theorem}\label{thm: local}
Under the local alternative with \eqref{eq: local_quantity} for a sequence $\delta=o(1)$, assume $(X_i, Y_i)$, $i\in\mathcal{I}_2$ are i.i.d samples from $\mathcal{P}_{X,Y}$, $\theta(\cdot)$ has a continuous distribution under both $\mathcal P_{X,Y}$ and $\mathcal P_X\times\mathcal P_Y$, $\mu_* - \mu = o_p(n_2^{-1/2})$, $\sigma_0^2\ge c$ for some constant $c>0$, and
\begin{eqnarray}\label{condition: local}
\E_*\left[ \left|\indic\{ \widehat\theta(X, Y) < \widehat\theta(X', Y')\} - \indic\{\theta(X, Y)<\theta(X', Y')\}\right| \right] = o_p(1).
\end{eqnarray}
Then
\begin{eqnarray*}
\frac{\sqrt{n_2}(R - 1/2)}{\widehat \sigma} = Z -\frac{\sqrt n_2 \delta}{4\sigma} + o_p(1).
\end{eqnarray*}
where $Z\overset{d}{\rightarrow} N(0, 1)$ as $n_1$ and $n_2$ goes to infinity.
\end{theorem}
As a consequence, when the distance between the local alternative and the null vanishes at the same or a slower rate as $n_2^{-1/2}$, the limiting distribution of the test statistic under the local alternative becomes a location-shited normal distribution with unit variance.
\begin{remark}
The conditions in Theorem \ref{thm: local} are stronger than those required in the fixed population versions in Theorems \ref{thm: null} and \ref{thm: alt}. This is because the local alternative hypothesis can be close to the null as fast as $n_2^{-1/2}$ and a more delicate treatment of the estimation error is needed to establish the asymptotic distribution.
In particular, equation (\ref{condition: local}) typically holds when $\hat\theta$ is a consistent estimate of $\theta$ up to a strictly monotone transform, whereas equation (\ref{eq:alt_thm_condition}) only requires a constant error accracy. The most stringent condition is $\mu_* - \mu = o_p(n_2^{-1/2})$.
Let $\Delta_i=\indic\{\widehat\theta(X_i, Y_i)<\widehat\theta(X_i', Y_i')\} - \indic\{\theta(X_i, Y_i)<\theta(X_i', Y_i')\}$. Then $\mu^*-\mu=\mathbb E_*\Delta_i$.
If a parametric estimate $\hat\theta$ is used, then typically $\Delta_i= O_P(n_1^{-1/2})$. So the required condition holds if $n_1\gg n_2$. In the pratically preferred case of $n_1\asymp n_2$, we have $\Delta_i\asymp n_1^{-1/2}$, but
$\mu_* - \mu=\mathbb E_*\Delta_i$ can still be much smaller than $\Delta_i$ if the random variable $\Delta_i$ is centered around zero and not highly skewed. We also provide a simple numerical example that verifies the condition $\mu_* - \mu = o_p(n_2^{-1/2})$ in section \ref{sec:verify_condition} of the supplement.
\end{remark}
\subsection{Classification accuracy under cyclic permutation}\label{sec:classification_accuray}
A remaining question regarding the procedure is whether we have any formal guarantees on the estimator $\widehat\theta$ because it is not obtained from a standard independent two sample data, but from only a single sample, with the second sample obtained from cyclically permuting the original sample. The quality of such $\widehat\theta$ would depend on the particular form of the estimator and the data distribution. Intuitively, the weak dependence caused by the cyclic permutation among the sample points should be negligible, and the resulting estimator would behave similarly to those obtained from genuine independent two-sample data. Here for an illustrative purpose, we prove the consistency of the classifier obtained under (1) a classical low-dimensional M-estimation and (2) a high-dimensional lasso-based sparse regression. Note that both the low dimensional and high dimensional models are trained on the first subset of data $\mathcal{I}_1$ with $n_1=|\mathcal{I}_1|$. For notation simplicity of the consistency analysis, we drop the subscript and use $n$ instead of $n_1$ only in section \ref{sec:classification_accuray} and its proofs.
\subsubsection{Low-dimensional M-estimation}
Define the objective function as
\begin{eqnarray*}
M(X, Y, X', Y'; \beta) \stackrel{\mbox{\textrm{\rm\tiny def}}}{=} M_1(X, Y;\beta) + M_2(X', Y'; \beta),
\end{eqnarray*}
where $\beta\in\mathbb{R}^p$ is the unknown parameter in the classifier. Here $(X,Y)$ and $(X',Y')$ are independent realizations from $\mathcal P_{X,Y}$ and $\mathcal P_X\times \mathcal P_Y$, respectively. We use $\mathcal P$ to denote the joint distribution of $(X,Y,X',Y')$. Then the objective function is $\E\{M(X, Y, X', Y'; \beta) \}$, where the expectation is taken with respect to $\mathcal{P}$. For example, we can choose $M_1(x,y;\beta)=-\ell_1(x,y;\beta)$ and $M_2(x',y';\beta)=-\ell_2(x',y';\beta)$, with some class-specific binary classification loss functions $\ell_1(\cdot)$, $\ell_2(\cdot)$, such as the hinge loss or the logistic loss function.
Let $\beta_0$ be the true parameter that maximizes the objective function. Using the cyclicly permuted data, the classifier is trained by maximizing the empirical criterion function
\begin{eqnarray*}
M_{n}(\beta)\stackrel{\mbox{\textrm{\rm\tiny def}}}{=}\frac{1}{n}\sum_{i=1}^n\left\{ M_1(X_i, Y_i; \beta) + M_2(X_i', Y_i'; \beta) \right\},
\end{eqnarray*}
Denote $\widehat\beta_n$ as the maximizer of $M_{n}(\beta)$. The consistency of $\widehat\beta_n$ is established in Theorem \ref{Mestimation}.
\begin{theorem}\label{Mestimation}
Suppose $(X_i, Y_i)$, $i\in\mathcal{I}$, are independent observations drawn from $\mathcal P_{X,Y}$. Let $\mathcal{M} = \{ M(x, y, x', y'; \beta) : \beta\in \mathcal{B}\}$ be a class of measurable functions such that $N_{[\mkern5mu ]} (\epsilon, \mathcal{M}, \mathcal{P})<\infty$ for every $\epsilon>0$, and $\E\left[\{ M(X, Y, X', Y'; \beta) \}^4\right]<\infty$. Suppose the true parameter $\beta_0$ is identifiable, i.e.,
\begin{eqnarray*}
\sup_{\beta: d(\beta,\beta_0)\geq\epsilon} \E\{M(X, Y, X', Y'; \beta) \} < \E\{M(X, Y, X', Y'; \beta_0) \},
\end{eqnarray*}
where $d(\cdot)$ is a distance measure. Then any sequence of estimators $\widehat\beta_n$ with $M_n(\widehat\beta_n)\geq M_n(\beta_0) - o_p(1)$ converges in probability to $\beta_0$.
\end{theorem}
The condition for the class of objective functions $N_{[\mkern5mu ]} (\epsilon, \mathcal{M}, L_1)<\infty$ is relatively standard for classical M-esimators. See \cite{van2000asymptotic} for several examples.
A detailed proof of Theorem \ref{Mestimation} is given in Appendix \ref{proof:m_est}. The key of our proof is a strong law of large numbers resulting in the dependent data, proved by carefully decomposing the variance of the sum of dependent variables and applying the Borel-Cantelli lemma. Then we are able to show the uniform consistency of $M_n(\beta)$ in Lemma \ref{GC}, which further implies consistency of $\widehat\beta_n$ when combined with standard empirical processes and M-estimation results \citep{van2000asymptotic}.
\subsubsection{High dimensional regression}
We consider a scenario where the can be large, compared to the sample size. Denote the dimension of $X$ as $d_1$ and the dimension of $Y$ as $d_2$. Let $d = d_1+d_2$. Denote $Z = (X, Y)\sim\mathcal{P}_{X,Y}$ and $Z' = (X', Y')\sim\mathcal{P}_X\times\mathcal{P}_Y$. We define
\begin{eqnarray*}
g(z) = \frac{d \mathcal{P}_{X,Y} }{d\mathcal{P}_{X,Y} + d\mathcal{P}_X\times\mathcal{P}_Y}(z)
\end{eqnarray*}
Our goal is to estimate $g(z)$ while keeping in mind that $d_1$ and $d_2$ may be comparable or lager than the sample size $n$.
In order to cope with high dimensionality, we assume that $g(z)$ has a sparse representation in a certain basis. This would be particularly reasonable, for example, when only a few coordinates of $X$ and $Y$ are dependent. Assume that $s_1$ out of $d$ coordinates of $Z=(X,Y)$ are dependent. Then $g(z)$ is essentially a function of $s_1$ variables instead of $d$ variables. Consider all the $s_1$-way combinations of coordinates of $Z$, and use $K_n$ basis for each combination.
Specifically, let $\xi_1,\xi_2,...$ be a basis function of the $L_2$ space $\mathbb R^{s_1}\mapsto \mathbb R$. Let $K_n$ be a slowly growing number. We consider the basis
$\xi(Z) = \{\xi_k(Z_{j_1},Z_{j_2},...,Z_{j_{s_1}}):1\le k\le K_n, 1\le j_1<j_2<...<j_{s_1}\}$ with dimensionality $m\propto d^{s_1} K_n$, and assume that the function $g(z)=\xi(z)^T\beta^*$ with $\|\beta^*\|_0\le s_2\ll n$. Such a hard sparsity assumption makes the presentation simpler and can be relaxed using a standard oracle-inequality argument.
Our starting point is that the function $g$ is the minimizer of the following problem
\begin{equation*}\label{eq:g_minimize}
\min_h\E\left[(1-h(Z))^2 + (0-h(Z'))^2 \right]\,,
\end{equation*}
since we associated a label $K=1$ ($K=0$) for each sample point from $\mathcal P_{X,Y}$ ($\mathcal P_X\times \mathcal P_Y$).
As a result, under the assumed basis expansion and sparse representation of $g$, $\beta^*$ is the minimizer of the problem
\begin{eqnarray}\label{hd_objective}
\min_{\beta} \beta^T\Gamma\beta - 2\gamma^T\beta,
\end{eqnarray}
where
\begin{eqnarray*}
\Gamma = \E\xi(Z)\xi(Z)^T +\E\xi(Z')\xi(Z')^T, \quad \gamma = \E\left[\xi(Z)\right],
\end{eqnarray*}
Now consider the empirical version with cyclic permuted $Z_1',\dots,Z_n'$, we estimate $\widehat\beta$ by optimizing the regularized quadratic form
\begin{eqnarray}\label{eq: obj}
\min_{\beta} \beta^T\widehat\Gamma\beta - 2\widehat\gamma^T\beta + \lambda \|\beta\|_1,
\end{eqnarray}
Denote $\Xi_Z= (\xi(Z_1)^T,\dots, \xi(Z_n)^T)^T$ and $\Xi_{Z'}= (\xi(Z_1')^T,\dots, \xi(Z_n')^T)^T$, $\Xi = (\Xi_Z, \Xi_{Z'})$, then
\begin{eqnarray*}
\widehat\Gamma = n^{-1}\Xi^T\Xi,\quad \widehat\gamma = n^{-1}\Xi_Z^T\mathbf{1}_{n\times1}.
\end{eqnarray*}
Let $\mathcal{G}= (g(Z_1)^T,\dots, g(Z_n)^T, g(Z_1')^T,\dots, g(Z_n')^T)^T\in\mathbb{R}^{2n\times 1}$.
Define the set $\mathcal{C}_\alpha(S) = \{\Delta\in\mathbb{R}^m: \|\Delta_{S^c}\|_1\leq\alpha\|\Delta_S\|_1 \}$. We assume the matrix $\Xi$ satisfies the {\it restricted eigenvalue} (RE) condition over $S$ with parameters $(\kappa, \alpha)$ if
\begin{eqnarray}\label{RE_condition}
\frac{1}{n}\|\Xi\Delta\|_2^2\geq \kappa \|\Delta\|_2^2, \quad \mbox{for all}\quad \Delta\in\mathcal{C}_\alpha(S).
\end{eqnarray}
We also define the residual with respect to the minimization problem (\ref{hd_objective}). Note that the response vector is $(1,\dots, 1, 0\dots, 0)^T\in\mathbb{R}^{2n}$ and the design matrix is $\Xi$, with parameter $\beta^*$. Thus we let $w_1(z) = 1 - \xi(z)^T\beta^*$ with $z = Z_1,\dots, Z_n$, and let $w_0(z) = -\xi(z)^T\beta^*$ with $z = Z_1', \dots, Z_n'$. Denote $w = (w_1(Z_1),\dots, w_n(Z_n), w_0(Z_1'),\dots, w_0(Z_n'))^T$.
\begin{theorem}\label{thm: hd_theory}
Assume that $\beta^*$ is supported on a subset $S\subset\{1,2,\dots, m\}$ with $|S| = s_2$, and each basis function is bounded on $[-B, B]$. Further assume the matrix $\Xi$ satisfies the restricted eigenvalue condition (\ref{RE_condition}) with parameters $(\kappa, 3)$, and $\lambda$ satisfies that $\lambda\geq4\|\Xi w/n \|_{\infty}$. Then the solution to the optimization problem (\ref{eq: obj}) satisfies
\begin{eqnarray*}
\|\widehat\beta - \beta^*\|_2\leq \frac{3}{2\kappa} \sqrt{s_2}\lambda.
\end{eqnarray*}
In particular, when taking $\lambda = C\sqrt{\log m /n}$, we have
$\|\widehat\beta - \beta^*\|_2\leq C\sqrt{s_2 \log m /n}$
with probability no less than $1-m^{-1}$ for some constant $C$ depending only on $\kappa$ and $B$.
\end{theorem}
The proof of Theorem \ref{thm: hd_theory} is given in the supplement. We can also relax the hard sparsity assumption on $\beta$ and use the oracle inequality version of the proof (Theorem 7.19 in \cite{wainwright2019high}) to prove the finite bound on $\widehat\beta$. The restricted eigenvalue condition is a standard one in the lasso literature. Here we directly assume the random design matrix $\Xi$ satisfies a restricted eigenvalue condition, which can hold with high probability if the population version $\Gamma$ satisfies the same condition with slightly different constants. Recall that $m\propto d^{s_1} K_n$. Thus the error bound for $\beta$ is of order $\sqrt{s_1 s_2} \sqrt{\log d/n}$. When assuming $s_1$ and $s_2$ are constants, the dimension of the data is allowed to grow exponentially with the sample size.
\section{Numerical Validation}
In this section, we conduct numerical simulations to illustrate the performance of our method. For brevity, we will focus on the more challenging and interesting cases where both $X$ and $Y$ are high dimensional, and the dependence signal is sparse. Specifically, we assume only the first element in $X$ and $Y$ are related: $Y_1 = a\times g(X_1) + \epsilon$, where the signal $a$ varies from $0$ to $1$. $Y_2, \dots, Y_{d_2}$, $X_1, X_2,\dots, X_{d_1}$, and $\epsilon$ all follow $N(0, 1)$ and are independent. The following models are considered:
\begin{itemize}
\item M1: $Y_1 = a\times X_1 + \epsilon$;
\item M2: $Y_1 = a\times \sin(X_1) + \epsilon$;
\item M3: $Y_1 = a\times \exp(X_1) + \epsilon$;
\item M4: $Y_1 = a\times\left\{ \indic{(X_1<0)} N(1, 1) + \indic{(X_1>0)} N(-1, 1) \right\}+ \epsilon$;
\item M5: $Y_1 = a\times \log(4X_1^2)+\epsilon$;
\item M6: $Y_1 = a\times 5\sqrt{|X_1|}+\epsilon$;
\end{itemize}
Our simulation models are similar to a variety of models that have been considered in the literature, though mostly in a less challenging case where $X$ and $Y$ are both low dimensional. For example,
(M1) is one of the most popular models and have been considered in \cite{szekely2007measuring, huo2016fast, shi2020distribution, deb2021multivariate}, etc. Functional transformations similar as (M2) and (M3) have been considered in \cite{zhu2017projection} and \cite{zhu2020distance}. (M4) is the mixture model and was used in \cite{heller2012consistent, biswas2016some, deb2021multivariate}.
(M5) was previously used in \cite{szekely2007measuring} and \cite{deb2021multivariate}. (M6) has also been considered in \cite{huo2016fast} and \cite{zhu2020distance}.
As mentioned in the previous section, we choose the neural network to train the classifier and implement it by \texttt{TensorFlow} \citep{tensorflow2015whitepaper}.
We use three layers of nodes (one input layer, one hidden layer, and one output layer). The number of nodes for the input layer is the dimension of the training data, and the number of nodes in the hidden layer is proportional to the data dimension. The output layer only contains one node since the task is binary classification. We further enforce the hidden layer with $L_1$ kernel regularization with regularization parameter varying from $10^{-4}$ to $10^{-3}$. The dropout rate \citep{srivastava2014dropout} for the hidden nodes also varies from 0.1 to 0.3. Details about the algorithm can be found in the supplemental code written in \texttt{python}.
We compare the proposed method with other popular statistical independence tests, including the distance correlation (\cite{szekely2007measuring}, denoted by ``DC"), ranks of distance test (\cite{heller2012consistent}, denoted by ``HHG"), and mutual information (\cite{berrett2019nonparametric}, denoted by ``MI"). Those competing tests are implemented with popular R packages: energy, HHG, and IndepTest, respectively. Because the proposed method is a {\bf C}ircularly {\bf P}ermuted {\bf C}lassification based independence test, we name it the {\it CPC} test.
We first look into the high dimensional effect on the independence tests by considering the linear model (M1), where $a$ is set to be 1. The performance of the tests when the dimension increases are summarized in Figure \ref{hd_power}. For the proposed method, it can detect the sparse dependence even when the dimension increases up to 500. The main reason is that we implement the $L_1$ penalization for the hidden layer, which greatly eliminates the noise in the data and preserves the desired sparse dependence signal. For comparison, the HHG and MI method suffer significantly from high dimensionality, while the distance correlation has surprisingly high power when the dimension $d_1$ and $d_2$ are less than 200, but its power still decreases dramatically when the dimension further increases.
\begin{figure}
\centering
\includegraphics[scale=0.6]{Figures/HDcmp.pdf}
\caption{ The power versus dimension of the proposed test (``CPC") compared with distance correlation (``DC"), ranks of distance test (``HHG"), and mutual information (``MI") when $n=1000$, $\alpha=0.05$.}\label{hd_power}
\end{figure}
Next, we focus on fixed dimension $d_1 = d_2 = 100$ to ensure a relatively fair comparison, because otherwise, all current methods tend to have inferior power. We report the performance of all six models (M1) - (M6) in Figure \ref{power_plot1}, with sample size $n = 1000$ and significant level $\alpha=0.05$. Additional simulation when $\alpha = 0.1$ and $0.01$ are given in the supplementary material. All results are averaged over 1000 repetitions. As expected, the proposed test has increasing power as the signal becomes stronger, with correct type-I error under the null hypothesis and high power when the signal exceeds a certain threshold. It performs particularly well for (M5), where all other tests have very low power even when the signal increases. The distance correlation also has considerable power, especially when the signal is strong and the dependence relationship is linear. The ranks of distance test and mutual information do not suit the high dimensional setting and have very low powers for almost all settings.
While existing tests based on sample splitting tend to cause nonignorable power loss in practice\citep{wasserman2020universal, kim2020dimension}, this phenomenon is weakening in our test. In the simulations, the newly proposed test outperforms other tests that use the whole dataset. This is because half of the data is ``invested" to find the most promising dimension reduction directions, and improves power performance under $H_1$.
\begin{figure}
\centering
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[scale=0.37]{Figures/power_case_linear5.pdf}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[scale=0.37]{Figures/power_case_sin5.pdf}
\end{subfigure}
\centering
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[scale=0.37]{Figures/power_case_exp5.pdf}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[scale=0.37]{Figures/power_case_mixture5.pdf}
\end{subfigure}
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[scale=0.37]{Figures/power_case_square_root5.pdf}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.45\textwidth}
\centering
\includegraphics[scale=0.37]{Figures/power_case_log_square5.pdf}
\end{subfigure}
\caption{ The increasing power versus the signal $a$ of the proposed test (``CPC") compared with distance correlation (``DC"), ranks of distance test (``HHG"), and mutual information (``MI") when $n=1000$, $d_1=d_2=100$, $\alpha=0.05$. }\label{power_plot1}
\end{figure}
Lastly, we compare the computing time of the tests. We still use the linear model in (M1) for simplicity. We restrict the computation memory to be 16 GB and compute the average computing time for one run of the test based on 1000 repetitions. Two settings are considered: 1) the sample size $n$ is fixed to be 1000, and dimension $d_1=d_2$ linearly increase from 100 to 500. 2) the dimension $d_1=d_2$ are fixed to be 100, and the sample size $n$ increases from 1000 to 5000. The time costs measured in minutes are reported in Tables \ref{time1} and \ref{time2}, respectively. We used permutation tests for distance correlation, HHG and mutual information to obtain $p$-values and the permutation replicate is set to be 200. We observe that the computation time of the proposed test almost grows linearly with the dimension and sample size. For distance correlation, HHG and mutual information, the computation costs grow linearly with the dimension but grow at least quadratically with the sample size. The HHG method exceeds the memory constraint (16GB) when the sample size $n$ is larger than 2000, and we are unable to obtain its corresponding computation times in Table \ref{time2}. In general, the proposed test is much faster compared with other methods for large-scale data sets. Lastly, we only used regular CPU cores for the entire simulation. The computing time for our test can be further reduced when using advanced GPU cores.
\begin{table}[h]
\centering
\caption{The average computing time measured in minutes of the proposed test and distance correlation, rank of distance test, and mutual information when $d_1$ increases from 100 to 500. $d_2=d_1$, $n=1000$. }
\begin{tabular}{ccccc}
\hline
$d_1, d_2$ & CPC & DC & HHG & MI \\ \hline
100 & 0.025 & 0.009 & 0.105 & 0.425 \\
200 & 0.052 & 0.011 & 0.107 & 0.765 \\
300 & 0.086 & 0.015 & 0.108 & 1.104 \\
400 & 0.138 & 0.017 & 0.113 & 1.448 \\
500 & 0.144 & 0.020 & 0.122 & 1.837 \\ \hline
\end{tabular}\label{time1}
\end{table}
\begin{table}[h]
\centering
\caption{The average computing time measured in minutes of the proposed test and distance correlation, rank of distance test, and mutual information when $n$ increases from 1000 to 5000. $d_1=d_2 = 100$. }
\begin{tabular}{ccccc}
\hline
$n$ & CPC & DC & HHG & MI \\ \hline
1000 & 0.023 & 0.009 & 0.105 & 0.417 \\
2000 & 0.040 & 0.046 & 0.470 & 1.785 \\
3000 & 0.055 & 0.099 & -- & 3.676 \\
4000 & 0.086 & 0.150 & -- & 6.499 \\
5000 & 0.086 & 0.201 & -- & 9.849 \\ \hline
\end{tabular}\label{time2}
\end{table}
\section{Application to Single Cell Data}
The analysis of single cell sequencing data has fueled much discovery and innovation over recent years \citep{kulkarni2019beyond}, and recent advances in multimodal omics promise further progress.
In this section, we apply the proposed test to a single cell dataset consisting of measurements of Peripheral blood mononuclear cells (PBMCs), publicly available on the 10X Genomics website \citep{pbmc}. The data contain measurements of ATAC-seq and RNA-seq in 11,898 cells, and we are interested in testing whether the two modes of measurement are independent. It has been widely assumed that ATAC-seq and RNA-seq are dependent because ATAC-seq identifies open chromatin sites that are available for transcription. For example, \cite{eltager2021scmoc} proposed to identify cell clusters using the co-measurements of RNA-seq and ATAC-seq from the same cell. In this section, we aim to provide solid statistical evidence for the dependence relationship among the two random vectors.
Each record in the dataset corresponds to a single cell. We perform quality control on these data before analysis. The RNA-seq data initially consists of a vector of counts that we pre-process following the Seurat 3.0 pipeline \citep{stuart2019integrative}. We retain cells that have counts from 50 - 10,000 genes to exclude almost empty and noisy cells. We set {\it minimum cells per gene} to be 1 to remove genes that are detected in cells less than this threshold. RNA-seq counts are then normalized by dividing each count by the total count for each cell and then scaling up to counts per million. The ATAC-seq data is also derived from counts, however, because these fragments are distributed across the entire genome, the data were pre-processed to identify peaks, which are clusters of fragments that were inferred to indicate a single region of open chromatin; all of the fragments in the locality of a peak are counted and attributed to the peak location \citep{yan2020reads}. We retain cells whose peaks include from 50 to 15,000 counts. The {\it minimum cells per peak} is set as 1. Peak counts are normalized by dividing each count by the total count for each cell and then scaling up to counts per million.
Overall 11,188 cells passed the quality control for both RNA-seq and ATAC-seq. The dimension of the RNA-seq data is 29,717 genes, for which only 6.35\% of the entries in the data matrix has non-zero values. For the ATAC-seq data, the dimension is 143,887 peaks and only 5.66\% entries have non-zero values. To achieve fast computation, we store the data in a sparse matrix and run the proposed algorithm and other competing algorithms in python and R, respectively. However, the distance correlation, HHG, and mutual information all reported errors in the algorithm because of exceeding the memory constraint of 16GB. It suggests that some substantial adaptations may be necessary to apply these existing tests of independence that are unsuitable for such high dimensional sparse datasets. For the proposed method, we use the neural network with 3 layers, where the hidden layer contains 2000 nodes. We only used CPU cores to train the algorithm, and it takes about 13.89 minutes to run the test. The test statistic is $-80.95$ and the corresponding $p$-value is practically 0. This strongly confirms that the RNA-seq and ATAC-seq are indeed dependent on each other.
\section{Discussion}
In this paper, we proposed a general framework for independence testing that is powerful in detecting sparse dependence signals in high dimensional data. We borrow the strength from the most powerful classification tools, such as neural networks, to boost power when the dependence between $X$ and $Y$ is sparse and weak. The proposed test statistic has a standard normal asymptotic distribution when the sample size is large. In addition to such a distribution-free asymptotic null distribution, the new test has several advantages over existing works in both power performance and computing efficiency. We apply the new test to a single cell data set and confirmed a widely believed important hypothesis in the multimodal omics literature.
There are several potential directions to follow up. The idea in this paper can be readily applied in other related testing problems, including the test of mutual independence and the test of conditional independence \citep{cai2022distribution}. By constructing two samples that have the same distribution under $H_0$ but different distributions under $H_1$, one can always transform those tests into a classification problem. Another interesting and unsolved problem is how to avoid the power loss caused by data splitting. One may switch the role of $\mathcal{I}_1$ and $\mathcal{I}_2$ and obtain another test statistic and $p$-value, which is dependent on the original one. Another choice is to perform multiple sample splitting and obtain a sequence of test statistics and $p$-values, which are statistically dependent. Existing methods such as Cauchy combination test \citep{liu2020cauchy, cai2022modelfree} and averaging $p$-values \citep{vovk2020combining} could be applied to combine the results under cerntain restricting conditions. It will be very rewarding to study how to efficiently combine those dependent statistics and $p$-values in high dimensional independence testing problems.
|
\section{Definitions}
We give in this section the basic definitions and fix the notation that will be used throughout this work. All graphs in this paper are undirected and simple. Let $G$ be a graph, and let $V(G)$ and $E(G)$ denote its vertex and edge sets, respectively. We denote by $n$ the number of vertices and by $m$ the number of edges.
Whenever it is clear from the context, we simply write $V$ and $E$ and denote $G=(V,E)$. \suggestion{}{For basic definitions not included here, we refer the reader to \cite{Bondy}}. \\
Given a graph $G$ and $S\subseteq V$, the \emph{subgraph of $G$ induced by $S$}, denoted by $G[S]$, is the graph with vertex set $S$ and such that two vertices of $S$ are adjacent if and only if they are adjacent in $G$. When \suggestion{$G'=G[S]$}{$G'$ and $G[S]$ are isomorphic} for some $S \subseteq V$, \suggestion{}{with a slight abuse of terminology} we simply say that $G'$ is an \emph{induced subgraph} of $G$.
For any family $\mathcal{F}$ of graphs, we say that $G$ is \emph{$\mathcal{F}$-free} if $G$ does not contain any graph $F\in \mathcal{F}$ as an induced subgraph. If a graph $G$ is $\mathcal{F}$-free, then the graphs in $\mathcal{F}$ are called the \emph{forbidden induced subgraphs} of $G$.
A \emph{clique} \suggestion{of}{in} a graph $G$ is a complete induced subgraph of $G$. Also, we will often use this term for the vertex set that induces the clique. \\
A graph $G$ is \emph{chordal} if there is a tree $T$ and a family of subtrees of $T$ indexed by the vertices of $G$, $\{T_u\}_{u\in V}$, such that subtrees $T_u$ and $T_v$ share a node of $T$ if and only if the vertices $u$ and $v$ are adjacent, where the subtree $T_u$ is comprised of all nodes that correspond to maximal cliques containing $u$. Furthermore, it is possible to construct $T$ in such a way that there is a bijection between the nodes of $T$ and the maximal cliques of $G$.
In general, such a tree $T$ is called a \emph{clique tree} of $G$. A chordal graph is called \emph{starlike} if one of its clique trees is a star\suggestion{}{, i.e., a complete bipartite graph $K_{1,n}$ for some $n\geq 1$}.
For a starlike graph $G$, let $\{X_0,X_1, . . . , X_r\}$ be the set of all maximal cliques of $G$, where $X_0$ (called the \emph{central clique}) is the maximal clique corresponding to the center of its star clique tree, and $X_i$ for $i\neq 0$ are the \emph{peripheral cliques}.
A graph $G$ is an \emph{interval} graph if it admits an intersection model consisting of intervals on the real line, that is, a family ${\cal I}$ of intervals on the real line and a one-to-one correspondence between the set of vertices of $G$ and the intervals of ${\cal I}$ such that two vertices are adjacent in $G$ if and only if the corresponding intervals intersect. Notice that an interval graph is a chordal graph with one of its clique trees being a path.
A \emph{proper interval} graph is an interval graph that admits a \emph{proper interval model}, this is, an intersection model in which no interval is properly contained in any other.
Three nonadjacent vertices of a graph form an $AT$ (\emph{asteroidal triple}) if every two of them are connected by a path avoiding the neighbourhood of the third.
Interval graphs are precisely those chordal graphs that are also $AT$-free and proper interval graphs are precisely those chordal graphs that are also $\{\text{claw}, \text{tent}, \text{net}\}$-free (see Figure~\ref{fig:S3clawnet}).
\begin{figure}
\centering
\includegraphics{forbidden.pdf}
\caption{Forbidden induced subgraphs for proper interval graphs}
\label{fig:S3clawnet}
\end{figure}
A graph $G=(C \cup I,E)$ is a \emph{split graph} if its vertex set can be partitioned into a set $C$ of pairwise adjacent vertices and set $I$ of pairwise nonadjacent vertices.
A \emph{threshold} graph is a split graph in which any two independent vertices satisfy that the neighborhood of one is contained in the neighborhood of the other.
Equivalently, $G$ is a threshold graph if it can be constructed from the empty graph by repeatedly adding either an \emph{isolated vertex} (nonadjacent to every other vertex) or a \emph{dominating vertex} (adjacent to every other vertex). Threshold graphs are characterised precisely as \suggestion{}{those} $\{2K_2, C_4, P_4\}$-free graphs.
\begin{comment}
$G$ is a \emph{quasi-threshold} graph, also called \emph{trivially perfect graph}, if it can be defined recursively as follows:
\begin{itemize}
\item a 1-vertex trivial graph is a quasi-threshold graph,
\item adding a new vertex adjacent to all vertices of a quasi-threshold graph results in a quasi-threshold graph, and,
\item the disjoint union of two quasi-threshold graphs results in a quasi-threshold graph.
\end{itemize}
\end{comment}
\emph{Quasi-threshold graphs}, also called \emph{trivially perfect graphs}, are \suggestion{those}{the} $\{P_4,C_4\}$-free graphs. A connected quasi-threshold graph $G = (V, E(G))$ admits a rooted tree $T = (V, E(T))$ on the same vertex set $V$, rooted on a vertex $r$, such that $uv \in E(G)$ if and only if there is a path in $T$ starting in $r$ containing both $u$ and $v$.
A graph $G$ is a \emph{caterpillar} if $G$ is a tree in which the removal of all the pendant vertices (i.e., the \emph{leaves}) results in a path (i.\ e.\ the \emph{spine} or \emph{central path}).
Let $L=\{1,\dots, s\}$. For a given graph $G=(V ,E)$, $(Y_j )=(Y_1, \dots , Y_s )$ is a \emph{path decomposition} of $G$ if the following conditions are satisfied.
\begin{enumerate}
\item $\bigcup _{i\in L} Y_i=V$.
\item For every $uv \in E$ there is an $i \in L$ with $\{u,v\}\subseteq Y_i$.
\item For every $i, j, k \in L$, $i\leq j \leq k$, $Y_i\cap Y_k\subseteq Y_j$.
\end{enumerate}
A well known result on interval graphs states that a graph is an interval graph if and only if its maximal cliques can be ordered as $X_1,X_2, \dots, X_s$ such that for every vertex $v$, the maximal cliques that contain $v$ appear consecutively~\cite{Gilmore1964}. This theorem allows to restate the interval completion problem in terms of path decompositions. Indeed, given a path decomposition $(Y_j)$, a completion to interval of $G$ can be computed by adding to $E$ the fill edges
\[F_j = \bigcup_{i\in L}\{uv \mid u,v \in Y_i \land uv\not\in E\}.\]
For the interval completion problem, a path decomposition $(Y_j )$ of $G$ is \emph{optimal} if $|F_k| \geq |F_j|$ for every path decomposition $(Y_k)$ of $G$.
Given a starlike graph $G$ with central clique $X_0$ and peripheral cliques $X_i$, $i\in \{1,\dots, r\}$, a path decomposition $(Y_j)$ of $G$ is \emph{normalized} if every maximal clique $X_i$, $i\in\{1,\dots,r\}$, is contained in exactly one $Y_{\pi(i)}$, i.e., there is a permutation $\pi:\{0,1,\dots, r\}\rightarrow \{0, 1,\dots , r\}$ with $X_i\subseteq Y_{\pi(i)}$ and $X_j\nsubseteq Y_{\pi(i)}$ for $j\notin \{0, i\}$.
In a normalized path decomposition $(Y_j )$, $X_0$ may appear in several $Y_j$. Nevertheless, we assume there always exists a vertex set $Y_s = X_0$ for some $s$. Thus, every peripheral clique $X_i$ with $i\in \{1,\dots, r\}$,
can be mapped to $Y_j$, $j \in \{0,\dots , s - 1, s + 1, \dots , r\}$. If the peripheral clique $X_i$ is contained in the vertex set $Y_j$, where $0\leq j \leq s - 1$ (respectively, $s + 1\leq j \leq r$), then we say that $X_i$ is on the \emph{left} (resp.\ \emph{right}) side of the path decomposition $(Y_j )$.
\section{Introduction}
Graph modification problems can be used to address many fundamental problems, not only in graph theory itself, but also to model a large number of practical applications in several different fields. Some of \suggestion{the fields in which graph modification problems are widely used to model applications are}{of those fields include} molecular biology, computational algebra, and more \suggestion{precisely}{generally}, \suggestion{in}{} areas that involve modelling based on graphs where the missing edges are due to a lack of data, for example in data clustering problems~\cite{Goldberg1995,Natanzon2001}. In many of these applications, an edge modification of the graph that models the experimental data corresponds to correcting errors and inconsistencies in the data.
Given a graph $G$ and a graph property $\Pi$, a \emph{graph modification problem} consists in studying how to add or delete the minimum number of vertices or edges from $G$ in order to obtain a graph that satisfies the property $\Pi$.
In this article, the property $\Pi$ will represent a graph class, such that the graph resulting from the modification verifies inclusion into this class. We focus our attention on one of the four basic graph modification problems: the $\Pi$-completion problem.
Given a property $\Pi$, a \emph{$\Pi$-completion} of a graph $G = (V, E)$ is a supergraph $H = (V, E \cup F)$ such that $H$ belongs to $\Pi$ and $E \cap F = \emptyset$. The edges in $F$ are referred to as \emph{fill edges}. A $\Pi$-completion $H = (V, E \cup F)$ of $G$ is \emph{minimum} if, for any set of fill edges $F'$ such that $H'= (V, E \cup F')$ belongs to $\Pi$, it holds that $|F'| \geq |F|$. In this case, $|F|$ is called the \emph{$\Pi$-completion number} of $G$.
The \emph{minimum $\Pi$-completion problem} consists in finding the $\Pi$-completion number of a graph $G$. The associated decision problem--the \emph{$\Pi$-completion problem}--consists in deciding, for a given integer $k$, if $G$ has a $\Pi$-completion with at most $k$ fill edges.
The $\Pi$-completion problem from an arbitrary graph is known to be $NP$-complete when $\Pi$ is the class of chordal, interval, or proper interval graphs~\cite{GAREY, Goldberg1995, Kashiwabara1979, Yannakakis1981}.
Furthermore, when $\Pi$ is the class of interval graphs, it was shown that the problem remains NP-complete on edge graphs~\cite{GAREY}, and also on co-bipartite graphs~\cite{Yuan1998}. Díaz et al.\ showed that it can be solved in $O(n)$-time for trees~\cite{Diaz1991}.
In the case of chordal and proper interval graph completion, its study from the viewpoint of parameterized complexity was initiated by Kaplan et al.\ in 1999~\cite{K-S-T-interval}. In 2015, Bliznets et al.\ presented the first subexponential parameterized algorithm for Proper Interval Completion that finds a solution in $k^{O(k^{2/3})}+O(nm(kn+m))$-time~\cite{Bliznets2015}.\\
\indent Monadic second order logic for graphs is a fragment of second order logic in which formulas allow logical operations ($\vee, \wedge, \neg, \Rightarrow$), adjacency/incidency tests, membership tests ($v \in V, e \in E)$, quantification over vertices or edges ($\exists v \in V, \exists e \in E, \forall v \in V, \forall e \in E)$, and over sets of vertices or edges ($\exists U \subseteq V, \forall D \subseteq E$). When only quantification over vertices is used, the logic is called $MSOL_1$, when quantification over vertices and edges is allowed, it is called $MSOL_2$.
A celebrated result by Courcelle et al.\ states that each graph property that is expressible in $MSOL_1$ (resp. $MSOL_2$) can be solved in polynomial time for graphs with bounded treewidth (resp. cliquewidth)~\cite{Courcelle00lineartime}. Note that this result is mainly of theoretical interest and does not lead to practical algorithms. Since the problem of \suggestion{}{finding a} proper interval graph completion \suggestion{}{with at most $k$ edges} can be expressed in $MSOL_2$ \suggestion{}{for fixed $k$}, this motivates our search of efficient algorithms for subclasses of chordal graphs with this property.\\
\indent Another direction \suggestion{in}{on} which current research on this topic is focused \suggestion{on}{}, is in finding and characterising minimal completions for some of these input graph classes in the most efficient possible way from a computational point of view~\cite{CrespelleT13,HeggernesSTV07,RapaportST08}.
Throughout this work, we consider the target class $\Pi$ to be the subclass of interval graphs given by proper interval graphs, which will be denoted by PIG.
The most well known motivation for the minimum PIG-completion problem comes from molecular biology. In~\cite{Benzer1959}, Benzer first gave strong evidence that the collection of DNA composing a bacterial gene was linear. This linear structure could be represented as overlapping intervals on the
real line, and therefore as an interval graph. In order to study various
properties of a certain DNA sequence, the original piece of DNA is fragmented into smaller pieces which are then cloned many times. When
all the clones have the same size, the resulting graph should not only be interval, but proper interval.
Deciding whether two clones should overlap or not is the critical part. However, there might be some false positive or false negatives due to erroneous interpretation of some data. Thus, correcting the model to get rid of inconsistencies is equivalent to remove or add as few possible edges to the graph so that it becomes interval.
This work is organized as follows. We start by proving in Section~\ref{sec:split_PIG} that the PIG-completion problem remains hard even if the input graph $G$ is split. Since split graphs are in particular chordal, this result implies that the problem remains hard when $G$ is a chordal graph, which led us to study proper subclasses of chordal graphs where the PIG-completion problem might be tractable.
More precisely, in Section~\ref{sec:threshold} we first give an efficient algorithm for the PIG-completion problem on threshold graphs. We finish the section by showing an efficient dynamic programming algorithm for the co-bipartite-completion problem on quasi-threshold graphs
following a brief discussion on the difficulties of generalizing the previous result to this superclass of threshold graphs. In Section~\ref{sec:caterpillar}, we show an efficient algorithm for a very sparse class of graphs, a subclass of trees called caterpillars.
We conclude the paper with some final remarks and possible future directions in Section~\ref{sec:conclusions}.
\section{An algorithm for PIG-completion on caterpillars}\label{sec:caterpillar}
In this section we \suggestion{will}{} work on a specific subclass of trees, namely caterpillars. We will show that for caterpillars, completions to proper interval graphs are very specific, and we will give a quadratic algorithm to compute them.\suggestion{}{``specific '' appears twice very closely}
Recall that a cartepillar is a tree where there is a path, called the central path, such that every vertex that is not in the path is adjacent to a vertex of the path.
For a caterpillar $G$, we call $P$ the central path, the \emph{father} of a leaf is its neighbour in $P$, and the \emph{sons} of a vertex of $P$ is the set of leaves it is adjacent to.
Let us consider a proper interval completion of $G$, and a proper interval representation of it. First, we show that such a representation can be transformed into a particular one where the vertices of the central path are intervals of the form $[i,i+1]$. Later, we will describe an $O(n^2)$ algorithm to get the best possible completion.
First take all of the leaves and change their closed interval of length $1$ to one open on the left and closed on the right. This can only decrease the number of edges in the graph. Let us now transform it into one that requires at most as many edges in the following way. To do this, we will use an interval representation that is not a proper interval representation, where some intervals are points while the others are intervals that have identical size (say $1$).
We will later transform this representation into a proper interval representation that represents the same graph.
Consider the vertices of $P$ in order $v_0, \dots, v_k$. For $i \in \{0, \dots,k\}$, let $V_i$ be the set of sons of $v_i$.
For every vertex $v_j$ of $P$, do the following transformation to the representation:
\begin{itemize}
\item For all $i \in \{0, \dots,k\}$,
let us denote $V_i^\ell$ to the set of sons of $v_i$ that contain the left endpoint of $v_i$, and $V_i^r$ to the set of sons of $v_i$ that contain the right endpoint of $v_i$, plus those that are represented by the same interval as $v_i$. We will refer to the vertices of $V_i^\ell$ as the \emph{left sons} of $v_i$, and the vertices of $V_i^r$ as the \emph{right sons} of $v_i$.
\item For a certain integer $j \le k$, we make the following assumption. We will call this assumption $A_j$.
\begin{itemize}
\item For every $i \le j$, vertex $v_i$ is represented by the interval $[i,i+1]$.
\item For every $i < j$, the vertices in $V_i^\ell$ are represented by the point $i$, and the vertices in $V_i^r$ are represented by the point $i+1$. All the leaves of vertices of $P$ from $v_j$ on are represented by intervals of size $1$ (with one side open and one closed, as explained above).
We note that a left son of a vertex $v_i$ for $i < j$ is adjacent to $v_i$, $v_{i-1}$ and its right sons (if $v_{i-1}$ exists), as well as maybe some vertices of $P$ from $v_j$ on and some of their sons. Similarly, a right son of a vertex $v_i$ for $i < j$ is adjacent to $v_i$, $v_{i+1}$ and its left sons (if $i<j-1$), as well as maybe some vertices of $P$ from $v_j$ on and some of their sons. Lastly a vertex $v_i$ is adjacent to its sons, $v_{i-1}$ and its right sons if they exist, and $v_{i+1}$ and its left sons, as well as maybe some vertices of $P$ from $v_j$ on and some of their sons.
Let $S_j$ be the set of vertices containing every $v_i$ for $i+1 < j$
, every son of $v_i$ for $i+1 < j$, and every left son of $v_{j-1}$ (if it exists). Let us denote $T_j$ to the set of vertices containing every $v_i$ for $i > j$ and all of their sons. Note that $v_j$, $v_{j-1}$ (if it exists) as well as the sons of $v_j$ and the right sons of $v_{j-1}$ are not in $S_j$ nor in $T_j$.
We will call \emph{unimportant edges} to all of the edges between a vertex of $S_j$ and a vertex of $T_j$. The other edges are \emph{important edges}. We will not count the unimportant edges, but we will show that the number of important edges never increases, and also that there are no unimportant edges in the end. Note that the set of important edges changes when $j$ augments. Let us denote $I_j$ to the number of important edges for step $j$.
Note that if $A_{j+1}$ is verified, no unimportant edge can become important when we go from $I_j$ to $I_{j+1}$ (only important edges can become non-important). Indeed, such an edge would have to be between $v_{j+1}$ and its leaves on the one side, and vertices of $\{v_{j-2}\} \cup V_{j-2}^r \cup V_{j-1}^\ell$ on the other side.
But by $A_{j+1}$, the representations of $\{v_{j-2}\} \cup V_{j-2}^r \cup V_{j-1}^\ell$ are in $[j-2,j-1]$, the representation of $v_{j+1}$ is $[j+1,j+2]$, and its leaves are intervals of length $1$, open on one side and closed on the other, intersecting $[j+1,j+2]$, therefore the two sets do not intersect and there can be no such new important edge.
\end{itemize}
\item Let us transform the representation in order to make assumption $A_{j+1}$ true.
If $v_{j+1}$ has its left extremity to the left of $v_j$, then do a symmetry of the representations of $T_j$ by point $j + \frac{1}{2}$.
Now $v_{j+1}$ does not have its left extremity to the left of $v_j$. We translate to the right the representations of $T_j$ so that $v_{j+1}$ corresponds to $[j+1,j+2]$. Let $K$ be the clique formed by all of the vertices of $T_j$ whose representations intersect with point $j+1$ after the transformations. Note that before the transformations, the vertices of $K$ were all intersecting a singular point in $[j,j+1]$ (indeed, the translation was of at most one to the right, and the symmetry did not alter this). The vertices of $K$ can be partitioned into the set $K_1$ of those that initially intersected $j$ and the set $K_2$ of those that did not. Since every vertex of $K$ is represented by an interval of length $1$, either a closed one or one closed on one side and open on the other, vertices in $K_2$ initially intersected $j+1$.
Now we switch $\min ( |V_j^r|, |K_1|)$ leaves from $V_j^r$ to $V_j^\ell$, replace leaves in $V_j^\ell$ by point $j$ and those in $V_j^r$ by point $j+1$.
Although we changed the graph as described previously, and this would change the values of the sizes of the sets, in our equations the values of $|V_j^r|$, $|K_1|$, \dots, will be as they were before the transformations, so that there is no ambiguity.
It is easy to see that $A_{j+1}$ is now verified. One must show however that the number of important edges did not augment. We will compare here the number of important edges before the symmetry/translation, with the number of important edges after $I_j$ is changed to $I_{j+1}$. The replacement of intervals by a point contained in the interval may of course only remove edges, and not add edges. The translation and the symmetry do not alter edges between vertices of $T_j$, nor edges between vertices that are not in $T_j$. The only edges that can be added are thus between vertices that are in $T_j$ and vertices that are not in $T_j$, plus edges between vertices that were switched from $V_j^r$ to $V_j^\ell$. Among those edges, we only need consider those that are in $I_{j+1}$.
Let us first consider the edges that may be added between vertices in $S_{j+1}$ and $T_j$. Note that vertices in $S_{j+1}$ are represented in $[0,j]$. For one such edge to be in $I_{j+1}$, it would have to be between a vertex of $S_{j+1}$ on one side, and $v_{j+1}$ or one of its leaves on the other side. But as discussed previously, the representation of $v_{j+1}$ is $[j+1,j+2]$, and its leaves are intervals of length $1$, open on one side and closed on the other, intersecting $[j+1,j+2]$, therefore the two sets do not intersect and there can be no such new important edge.
We thus only need to consider the edges between vertices in the two cliques represented by points $j$ and $j+1$. Moreover, we only need to consider those edges between $T_j$ on the one hand and the leaves of $v_j$ on the other hand, plus the edges between the leaves of $v_j$ that are transferred on the one hand, and the other leaves of $v_j$ they are adjacent to on the other hand. Let $V_\ell = \{v_{j-1}\} \cup V_{j-1}^r \cup V_j^\ell$.
After the transformation, the number of edges to consider is $|V
^\ell| \times \min ( |V_j^r|, |K_1|) + |K| \times (|V_j^r| - \min ( |V_j^r|, |K_1|))$. Since before the transformation the vertices of $V_\ell$ are adjacent to those of $K_1$ and those of $V_j^r$ are adjacent to those of $K_2$ plus the other vertices of $V_j^r$, it follows that the number of edges considered before the transformation is at least $|V^\ell| \times |K_1| + |K_2| \times (|V_j^r| - \min ( |V_j^r|, |K_1|)) + (|K_2| + |V_j^r| - \min ( |V_j^r|, |K_1|)) \times \min ( |V_j^r|, |K_1|)$.
\begin{itemize}
\item Case $|V_j^r| \ge |K_1|$. In this case we consider $|V^\ell| \times |K_1| + |K| \times (|V_j^r| - |K_1|)$ edges after the transformation, and at least $|V^\ell| \times |K_1| + |K_2| \times (|V_j^r| - |K_1|) + (|K_2| + |V_j^r| - |K_1|) \times |K_1|$, therefore there are at least $|K_2| \times |K_1| \ge 0$ more edges considered before the transformation than after, and the number of important edges does not increase.
\item Case $|V_j^r| < |K_1|$. In this case we consider $|V^\ell| \times |V_j^r|$ edges after the transformation, and at least $|V^\ell| \times |K_1| + |K_2| \times |V_j^r|$ before the transformation, therefore there are at least $|K_2| \times |V_j^r| \ge 0$ more edges considered before the transformation than after, and the number of important edges does not increase.
\end{itemize}
\end{itemize}
Note that $A_1$ (that is exactly that $v_1$ is represented by $[0,1]$) can be assumed to be initially verified, up to translation of the representation. Then after applying the previous transformations for every $j \in \{1,\dots,k-1\}$, we know that $A_k$ is true and that the number of edges has not increased (for $A_k$, $T_k = \emptyset$, and thus every edge is important). Now we can replace every leaf in $V_k^\ell$ by point $k$ and every leaf in $V_k^r$ by point $k+1$ to get a representation where:
\begin{itemize}
\item For every $i \in \{1,\dots,k\}$, the vertex $v_i$ is represented by the interval $[i,i+1]$.
\item For every $i \in \{1,\dots,k\}$, the vertices in $V_i^\ell$ are represented by the point $i$, and the vertices in $V_i^r$ are represented by the point $i+1$.
\end{itemize}
For each $i \in \{0,\dots,k+1\}$, let us denote by $W_i$ the set of vertices represented by the integer $i$, and let $w_i$ be the size of the set $W_i$.
Now we can keep the representation of every $v_i$ and replace every vertex in $W_i$ with the interval $[i \times (1 + \frac{1}{k}), i \times (1 + \frac{1}{k}) + 1]$. Such intervals, closed of length $1$, indeed do not intersect if they are not equal, and keep the same intersections with the $v_i$'s as their previous representations. Thus we get a proper interval representation of the same graph as obtained with the previous algorithm, which therefore has no fewer edges than the initial one.
Now we know that there exists a minimal completion to a proper interval graph as above. We will therefore design an algorithm that will always output such a solution. The only thing that remains to be determined is the size of $V_i^\ell$ (or equivalently that of $V_i^r$) for every $i \in \{1,\dots,k\}$. We will determine them via dynamic programming. More precisely, for every $i \in \{1,\dots,k\}$, and for every $j \in \{0, \dots, |V_i|\}$, we compute the optimal number $N_{i,j}$ of edges added to complete the caterpillar induced by $\{v_i, v_{i+1}, \dots, v_{k}\} \cup W_{i+1} \cup W_{i+2} \cup \dots \cup W_{k+1}$ into a proper interval graph, given that $|V_i^\ell| = j$. We use the following formula to compute it:
\begin{itemize}
\item For $j \in \{0, \dots, V_k\}$, $N_{k,j} = \binom{|V_k| - j}{2}$;
\item For $i \in \{1,\dots,k-1\}$ and $j \in \{0, \dots, V_i\}$, $N_{i,j} = \min_{j' \in \{0, \dots, |V_{i+1}|\}}\left (\binom{|V_k| - j + j' + 1}{2} + N_{i+1,j'} \right )$.
\end{itemize}
Now, the best completion to a proper interval graph of the whole caterpillar uses exactly \[\min\limits_{j \in \{0, \dots, |V_{0}|\}}\left (\binom{j}{2} + N_{1,j} \right )\] edges. One can get such a completion from the choices of $j$ obtained to get the minimal values.
Since $\sum_{i \in \{0, \dots, k\}}|V_i| \le n$, we know that there are at most $n$ different values of $N_{i,j}$ to be computed. Each one is computed using a minimum over at most $n$ different expressions. Therefore the complexity of this algorithm is in $O(n^2)$.
\section{Conclusions and future work}\label{sec:conclusions}
In this work, we study the $\Pi$-completion problem when $\Pi$ is the class of proper interval graphs. Given that the problem is NP-complete in general graphs, we focus \rv{our analysis}
\rv{on}
the case in which the input graph lies in some particular subclasses of chordal graphs. We prove that the problem remains hard in split graphs, whereof we conclude the same for chordal graphs. We present efficient algorithms for \suggestion{}{PIG-completion for both} threshold \suggestion{}{graphs} and caterpillar graphs, and an efficient algorithm for co-bipartite-completion for quasi-threshold graphs. \\
A future line of work is to continue \rv{studying} the PIG-completion problem in other subclasses of chordal graphs that have bounded cliquewidth to obtain polynomial algorithms. This may lead to finding common properties that could be useful when it comes to designing efficient algorithms and heuristics to solve the problem within other chordal subclasses.\\
On the other hand, given that all the graph classes for which we give a polynomial time algorithm \suggestion{in this work} {} are also subclasses of interval graphs, and that interval and proper interval graphs are very closely related, it \suggestion{rises}{raises} as a natural question whether the PIG-completion problem can be solved in polynomial time \suggestion{or not} {} when the input graph already belongs to this particular class.\\
An interesting question arises also in the relation with the max-cut problem: for which other classes besides threshold graphs does it hold that PIG-completion is equivalent to completion to co-bipartite? For these possible classes, an algorithm for max-cut in the complement would also solve the PIG-completion problem, by the same argument given in Section \ref{sec:threshold}.\\
Another possible continuation for this work may be to study the \emph{PIG-deletion problem}, i.e., the removal of a set of edges $F$ from an input graph $G =(V,E)$, so that the resulting subgraph $H = (V, E \setminus F)$ is a proper interval graph. It is known that this problem is hard for general graphs~\cite{Goldberg1995}, but it would be interesting to investigate whether efficient algorithms could also be devised for this problem restricted to the aforementioned subclasses of chordal graphs.\\
\section*{Acknowledgements}
Partially supported by Programa Regional MATHAMSUD MATH190013, by CNPq grant 311679/2018-8.
\subsection{Completion from quasi-threshold graphs}\label{sec:quasi-threshold}
\suggestion{When trying to generalize the previous results, a graph class related to threshold graphs arises as a natural continuation: quasi-threshold graphs are precisely $\{P_4, C_4\}$-free graphs.}{A natural candidate for generalizing the previous result are quasi-threshold graphs. Recall that these are precisely the $\{P_4, C_4\}$-free graphs.} However, for this particular class we \suggestion{stumble upon with the fact}{encounter the problem} that a crucial result from the previous section does not hold: a minimum co-bipartite-completion is not the same as a minimum PIG-completion.
We present a counterexample for this in Figure~\ref{fig:counterexample_qt}, in which we can see that a minimum co-bipartite-completion of the graph has an induced $C_4$.
\begin{figure}[h!]
\centering
\includegraphics[scale=.22]{counterexample_qt.png}
\caption{A quasi-threshold graph $G$ and a minimum co-bipartite-completion of $G$ that is not PIG. \suggestion{}{Numbered circles correspond to cliques of size equal to the number inside it}}
\label{fig:counterexample_qt}
\end{figure}
Nevertheless, we still can find an algorithm that solves the minimum completion problem to co-bipartite graphs when the input is a quasi-threshold graph. Notice that, even though this does not \suggestion{}{necessarily} provide of a minimum PIG-completion for a quasi-threshold graph, it is indeed a lower bound for it. Recall that every connected quasi-threshold graph has a dominating vertex, and that any PIG-completion in particular yields a claw-free supergraph of the input graph. The fact that a minimum completion to co-bipartite gives a lower bound for the minimum PIG-completion follows \suggestion{as a result of}{from} these two remarks.
Let $G = (V, E(G))$ be a connected quasi-threshold graph and $T =(V, E(T))$ the tree rooted in $r$ that stems from its definition.
For $v \in V$, let $n_v = |V(T_v)|$ be the number of vertices in $T_v$, the maximal subtree rooted at $v$, and let $c_v$ be the number of children of $v$.
For $v \in V$, let $v_1, \ldots, v_{c_v}$ be the children of $v$ in $T$, $0 \leq i \leq c_v$, $0 \leq j < n_v$. Let $X_{i,v} = \bigcup_{w \in \{v_1, \ldots, v_i\}} V(T_w)$ and $x_{i,v} = |X_{i,v}| = \sum_{1 \leq k \leq i} n_{v_k}$. We define $C(v,i,j)$ as the minimum number of edges in a completion to co-bipartite in the subgraph induced by $X_{i,v}$ such that one clique has cardinality $j$ and the other has cardinality $x_{i,v}-j$. Note that $v \not\in X_{i,v}$.
Finally, we define $D(v, j)$ as the minimum number of edges in a completion to co-bipartite of the subgraph induced by $V(T_v)$ such that one of the cliques has cardinality $j$.
\begin{proposition} \label{prop:Dv0}
$D(v, 0) = C(v, c_v, n_v-1)$ and for $1 \leq j \leq n_v$, it holds that $D(v, j) = C(v, c_v, j-1)$.
\end{proposition}
\begin{proof}
Note that $V(T_v) = X_{c_v, v} \cup \{v\}$ and that $v$ is adjacent to every vertex in $X_{c_v, v}$, hence we can always add it to any clique without increasing the cost of the completion.
\end{proof}
It follows from those definitions, that the value we are looking for is given by
\begin{equation}
\min_{0 \leq j \leq |V(G)|} D(r,j). \label{D_rj}
\end{equation}
We now show how to compute $C(v,i,j)$, for every $v$ and all possible values if $0 \leq i \leq c_v$ and $0 \leq j \leq x_{i,v}$. Once again, we assume the children of $v$, if any, are $v_1, \ldots, v_{c_v}$.
\begin{equation}
C(v,i,j)=\begin{cases}
0, \text{ if $v$ is a leaf or $i=0$}.\\
\min\limits_{j-n_{v_i} \leq k \leq j} C(v, i-1, k) + D(v_i, j-k) + k(j-k) + (x_{i-1,v}-k)(n_{v_i}-j+k), \text{ otherwise}.
\end{cases}
\label{recurrence}
\end{equation}
\begin{proposition} \label{prop:recurrence}
Equation~(\ref{recurrence}) computes $C(v,i,j)$ correctly.
\end{proposition}
\begin{proof}
For the first case, the graph is either trivial or empty.
For the second case, let $C_1$ and $C_2$ be the cliques of a completion obtained after a completion of $X_{i,v}$, such that $|C_1| = j$. Let $k = |X_{i-1,v} \cap C_1|$.
Then $|V(T_{v_i}) \cap C_1| = j-k$. Then note that we need $C(v, i-1, k)$ edges to turn each of $X_{i,v} \cap C_1$ and $X_{i,v} \cap C_2$ into cliques. Similarly, we need $D(v_i, j-k)$ edges to turn $V(T_{v_i}) \cap C_1$ and $V(T_{v_i}) \cap C_2$ into cliques. Finally, we need $k(j-k)$ edges to connect $X_{i,v} \cap C_1$ and $V(T_{v_i}) \cap C_1$, and $(x_{i-1,v}-k)(n_{v_i}-j+k)$ edges to connect $X_{i,v} \cap C_2$ and $V(T_{v_i}) \cap C_2$. Since we try all possible values of $k$, we eventually find the smallest possible completion.
\end{proof}
\noindent We arrive thus at the main result of this section, which we state below
\begin{theorem} Let $G$ be a quasi-threshold graph. There exists an $O(n^4)$ dynamic programming algorithm that computes the minimum completion to co-bipartite for $G$.
\end{theorem}
\begin{proof} The algorithm is given by expressions (\ref{D_rj}) and (\ref{recurrence}), and \suggestion{the}{its} correctness is immediate by Propositions~\ref{prop:Dv0} and \ref{prop:recurrence}. For the complexity, notice that the algorithm proceeds in a bottom-up manner, from the leaves of $T$ up to the root, and from left to right for vertices located at the same level of the tree. For every vertex $v \in V$, every $0 \leq i \leq c_v$ and every $0 \leq j \leq x_{i,v}$, we compute the recurrence relation (\ref{recurrence}). This means that values for expressions $C(v, i - 1, k)$ and $D(v_i, j - k)$ have already been calculated for every $j-n_{v_i} \leq k \leq j$ by the time we compute $C(v, i, j)$, so they are $O(1)$. We have hence that computing $C(v, i, j)$ is $O(n)$. Since this is needed for every $i$ and $j$ as defined above, we perform $O(n^3)$ operations at every vertex of $T$, except at the root. Indeed, for the root $r$, we only need to compute $C(r, c_r, j)$, for $0 \leq j \leq n$. This amounts to $O(n^2)$ operations. The complete algorithm is thus $O(n^4)$.
\end{proof}
\begin{comment}
First, we prove an auxiliary lemma analogous to the one in Section~\ref{sec:threshold}.
\begin{lemma}\label{lemma:cobip_for_qt}
A minimum completion of $G$ into a co-bipartite graph is the same as a minimum completion of $G$ into a proper interval graph.
\end{lemma}
\begin{proof}
The only if case is analogous as in Lemma~\ref{lemma:cobip_for_t}.
For the if case, let us consider a minimum completion $H$ of $G$ into a co-bipartite graph, and let $u$ be the dominating vertex of $G$. Let $K_1$ and $K_2$ be the two cliques of the completion $H \setminus \{u\}$, this is, without considering the universal vertex. It follows that there are no induced claws or asteroidal triples in $H$. Toward a contradiction, assume there is an induced cycle $C$ of length exactly four --since the graph is co-bipartite.
Moreover, this cycle has precisely two non-incident edges $e_1$ and $e_2$ that are in $E(G)$, which are the ones that are neither inside $K_1$ nor $K_2$
Let us denote $C=\{v_1, v_2, v_3, v_4\}$ where $v_1, v_4 \in K_1$, $v_2, v_3 \in K_2$, $e_1=v_1 v_2$ and $e_2=v_3 v_4$.
Using a similar argument as in Lemma~\ref{lemma:cobip_for_t}, notice that $G[V(C)]$ is not a threshold graph. Furthermore, since quasi-threshold graphs are precisely those $\{P_4, C_4\}$-free graphs, it follows that $G[V(C)]$ is isomorphic to $2 K_2$.
We will reach a contradiction using the fact that $H$ is a minimum completion to co-bipartite, and for this we will strongly use that the edges $f_1= v_1 v_4$ and $f_2= v_2 v_3$ are not in $E(G)$.
If $V(G)= \{u, v_1, v_2, v_3, v_4\}$, then $G$ is already co-bipartite without adding $f_1$ and $f_2$.
Let us suppose that $V(G)= \{u, v_1, v_2, v_3, v_4, w\}$. Suppose $w$ lies in $K_1$ in $H$. If $w$ is nonadjacent to $C$ in $G$ \valeria{means nonadjacent to every vertex of $C$, right?}, then we find a smaller completion adding only the fill edges $w v_1$ and $w v_2$, instead of $w v_1$, $w v_4$, $f_1$ and $f_2$. In this case the new cliques are $\{w, v_1, v_2 \}$ and $\{ v_3, v_4 \}$.
It follows analogously if $w$ is adjacent to $v_1$ and nonadjacent to $C\setminus \{v_1\}$ in $G$.
Moreover, if $w$ is adjacent to $v_1$ and $v_4$ and nonadjacent to $C\setminus \{v_1, v_4\}$ in $G$, then we find an induced $P_4$. The same holds if $w$ is adjacent to precisely three vertices in $C$. Finally, if either $w$ is adjacent to $C$ \valeria{adjacent to all vertex of $C$ to be coherent with the nonadjacency to $C$ above} or if $w$ is adjacent to $v_1$ and $v_2$ (resp.\ $v_3$ and $v_4$), then $G$ was already co-bipartite.
Now, let us suppose that $|K_i|=3$ for each $i=1,2$. Let $w_i \in K_i$ for $i=1,2$ be two distinct vertices such that and $w_i \neq v_j$ for each $v_j \in C$ and $i=1,2$.
Suppose first that $w_1$ is nonadjacent to $v_2$ and $v_3$, and $w_2$ is nonadjacent to $v_1$ and $v_4$, in both cases in $H$. Since $w_1$ is adjacent to $v_1$ and $v_4$ and $w_2$ is adjacent to $v_2$ and $v_3$ (in the completion $H$), and $G$ is a quasi-threshold graph (thus it contains no induced $P_4$), then at least one edge from $\{ w_1 v_1, w_1 v_4 \}$ and one edge from $\{ w_2 v_2, w_2 v_3 \}$ are not originally in $E(G)$. Furthermore, to avoid induced $P_4$'s either $w_2 v_2$ and $w_1 v_4$ are fill edges, or $w_2 v_3$ and $w_1 v_1$ are fill edges, or at least three of these four edges are fill edges. Let us suppose w.l.o.g. that $w_2 v_2$ and $w_1 v_4$ are fill edges. It follows from this and the fact that $f_1, f_2 \not\in E(G)$ that we can find a smaller completion to co-bipartite by swapping these four edges for $\{ w_1 v_2, w_2 v_4 \}$.
Let us now consider the case where either $w_1$, or $w_2$, or both are adjacent to more vertices in $C$ than $v_1, v_4$ and $v_2, v_3$, respectively.
Suppose first that $w_1$ is adjacent also to $v_3$ \valeria{then necessarily $v_1 w_1$ is a fill edge, otherwise we find an induced $P_4$ in $G$, right?}. In this case, we consider $H \setminus \{ f_1, f_2 \} \cup \{ v_1 w_2\}$, which results in a smaller completion to co-bipartite (the new cliques are $\{v_4, w_1, v_3 \}$ and $\{ v_1, v_2, w_2 \}$).
\fran{This next case is included in the previous one no? Actually if either $w_1$ or $w_2$ is ajdacent to some additional vertex we are always in the previous case w.l.o.g.} If instead $w_1$ is adjacent to $v_3$ and $w_2$ is adjacent to $v_4$, then necessarily either $v_1 w_1$ or $v_2 w_2$ are fill edges, otherwise we find an induced $P_4$ in $G$. Hence, we can remove such fill edge --let us say it is $v_2 w_2$--, $f_1$ and $f_2$, then adding $v_2 w_1$, and thus we obtain a smaller co-bipartite-completion (the new cliques are $\{v_1, w_1, v_2 \}$ and $\{ v_3, v_4, w_2 \}$).
If $w_1$ is adjacent to $v_2$ and $w_2$ is adjacent to $v_4$, then we can remove $f_1$ and $f_2$ and obtain a smaller co-bipartite-completion (the new cliques are $\{v_1, w_1, v_2 \}$ and $\{ v_3, v_4, w_2 \}$).
Finally, if $w_1$ is adjacent to $v_2$ and $w_2$ is adjacent to $v_1$ and $v_4$, then we find another $C_4$ induced by $\{w_1, v_4, w_2, v_2 \}$. Hence, at least one of the four edges that induce this cycle is a fill edge since $G$ was originally a quasi-threshold graph.
If either $v_2 w_2$ (resp.\ $w_1 v_4$) is a fill edge, then we find a smaller completion to co-bipartite by removing this edge plus $f_1$ and $f_2$ (the new cliques are the same as in the previous paragraph).
If the fill edge is $w_2 v_4$, then we have the same situation as one of the previous cases. Finally, if $w_1 v_2$ is the fill edge, then we remove it and $f_1$, $f_2$, we add $w_1 v_3$ and thus obtain a smaller co-bipartite-completion (the new cliques are $\{v_4, w_1, v_3 \}$ and $\{ v_1, v_2, w_2 \}$).
\nina{maybe add the case when $w_1 w_2$ is also an edge in $H$? It is analogous to a previous case though}
\nina{how to finish the proof?}
\end{proof}
\end{comment}
\section{PIG-completion within split graphs}\label{sec:split_PIG}
We now devote our attention to the complexity of the PIG-completion problem when the input graph belongs to the class of split graphs.
We start by citing some useful results and stating a few lemmas that characterise proper interval completions when the split partition fulfills certain properties. These results will be useful in Section~\ref{subsec:reduction_split}, where we give a reduction \suggestion{of}{to} the problem.
Recall the following results, which can be found in~\cite{PengChen06}:
\begin{lemma}[Peng et al.~\cite{PengChen06}]\label{lemma:starlike_optimal_pd}
Every starlike graph has a normalized path decomposition which is optimal.
\end{lemma}
\begin{theorem}[Peng et al.~\cite{PengChen06}]\label{theo:threshold_to_split}
The threshold-completion problem is NP-complete on split graphs.
\end{theorem}
Let us consider a connected split graph $G=(C\cup I, E)$ such that $C$ is maximum, in the sense that there is no vertex in $I$ complete to $C$, and let $H=(C\cup I, E\cup F)$ be a PIG-completion of $G$.
\begin{lemma}\label{lemma:1_split}
One of the following statements holds:
\begin{enumerate}
\item There is a partition $\{I_l, I_r\}$ of $I$ such that $H[I_l]$ and $H[I_r]$ are cliques. \label{item:split_claim11}
\item There is a partition $\{C', I'\}$ of $V(H)$, such that $C' \supseteq C$, $I' \subseteq I$, and statement~\ref{item:split_claim11} holds for $I'$. \label{item:split_claim12}
\end{enumerate}
\end{lemma}
\begin{proof}
Notice that, if there are only two vertices in $I$, then the first statement follows.
Suppose that the first statement is not true. In other words, for every partition $I_1, I_2$ of $I$, we can find two nonadjacent vertices $v,v' \in I_i$ for some $i=1,2$.
Toward a contradiction, suppose that there are no three vertices in $I$ that are pairwise nonadjacent in $H$. In particular, we can find two nonadjacent vertices $v_1, v_2 \in I$.
Since we are assuming there is no third vertex nonadjacent to both $v_1$ and $v_2$, let $w \in I$ such that, either $w$ is adjacent to both $v_1$ and $v_2$, or $w$ is nonadjacent to $v_1$ and $w$ is adjacent to $v_2$ (or vice versa). Let us first consider the case where $w$ is nonadjacent to $v_1$ and $w$ is adjacent to $v_2$.
For this purpose, consider $I_1=\{w, v_2\}$ and $I_2= I \setminus I_1$. Since $I_1$ is isomorphic to $K_2$ and the first statement of the claim does not hold, there has to be a
vertex $v_3 \in I_2$ such that $v_3 v_1 \not\in E\cup F$. If $v_3$ is nonadjacent to either $w$ or $v_2$, then either $v_1, v_3, w$ or $v_1, v_3, v_2$ are pairwise nonadjacent.
Suppose to the contrary that $v_3$ is adjacent to both $w, v_2$. Let $I^{(1)}_1 =\{w, v_2, v_3\}$ and $I^{(1)}_2 = I\setminus I^{(1)}_1$.
Notice that all three vertices of $I^{(1)}_1$ are pairwise adjacent. Using an analogous argument with $I^{(1)}_1, I^{(1)}_2$, we can either find a new vertex $v_4$ nonadjacent to $v_1$ and to at least one other vertex in $I^{(1)}_2$ (and therefore three nonadjacent vertices), or a new partition of $I$ into subsets $I^{(2)}_1 = \{w, v_2, v_3, v_4\}$ and $I^{(2)}_2 = I \setminus I^{(2)}_1$. We repeat this procedure, finding on each step a new partition of $I$ into subsets $I^{(j)}_1, I^{(j)}_2$, such that $v_1 \in I^{(j)}_2$, $I^{(j)}_1$ induces a complete subgraph of $H$ and $I^{(j-1)}_2 \supseteq I^{(j)}_2$ for each $j\geq 1$, until either $I^{(j)}_2 = \{v_1\}$ and $I^{(j)}_1$ is a clique, or we find a vertex in some $I^{(j)}_2$ that is nonadjacent to at least one vertex in $I^{(j)}_1$, reaching a contradiction in either case.
If instead there is no vertex $w \in I$ such that $w$ is nonadjacent to $v_1$ and adjacent to $v_2$ (or vice versa), then every vertex $w \in I$ distinct from $v_1$ and $v_2$ is adjacent to both $v_1$ and $v_2$. If $|I|=3$, then we can partition $I$ into two cliques in $H$, hence $|I|\geq 4$. Moreover, every pair $w, w' \in I\setminus \{v_1, v_2\}$ of distinct vertices must be pairwise adjacent, otherwise $\{v_1, v_2, w, w' \}$ induces a $C_4$ in $H$. Furthermore, since $v_1$ and $v_2$ are adjacent to every vertex $w \in I \setminus \{v_1, v_2\}$, then $I\setminus \{v_1\}$ is a complete subgraph of $H$ and hence the first statement of the claim is true, which results once more in a contradiction.
Therefore, there are at least three vertices $v_1, v_2, v_3 \in I$, nonadjacent in $H$.
Since $H$ is a proper interval graph, there are no induced claws and thus no vertex of $C$ is adjacent to $v_1$,$v_2$,$v_3$. However, since $G$ is connected, there is at least one vertex in $C$ adjacent (in $G$) to each \suggestion{}{of} $v_1, v_2, v_3$. If there is no vertex in $C$ adjacent in $H$ to any two of $v_1,v_2,v_3$, then we find an induced net, which is not possible since $H$ is proper interval graph and the net is a forbidden induced subgraph.
Hence, let $v\in C$ be a vertex adjacent to $v_1, v_2$ and nonadjacent to $v_3$. Since $G$ is connected, then there is a vertex $w\in C, w\neq v$, such that $v_3 w \in E$.
If $v_1,v_2$ are nonadjacent in $H$ to $w$, then there is an induced claw in $H$, thus we assume w.l.o.g. that $v_2$ is adjacent to $w$.
Since $C$ is maximum, there is a vertex $v' \in C$ such that $v' v_2 \not\in E$. Hence, either $v' v_1, v' v_3 \in E\cup F$, or $v' v_2 \in F$, for if not there are claws between $v_1,v_2,v',v$ and $v_2,v_3,v',w$, respectively. Thus, we either find a tent, which results in a contradiction since $H$ is PIG, or $v' v_2 \in F$. However, if every $v'' \in C$ is adjacent to $v_2$ in $H$, then we can move $v_2$ to a new larger clique $C'$ and remove $v_2$ from $I$.
Furthermore, if there is a vertex $v'_2 \neq v_2$ such that $v_2$ and $v'_2$ are both complete to $C$ in $H$, then either every vertex in $I$ is adjacent to $v'_2$, or $v_2 v'_2 \in E\cup F$. In either case, one of the two statements of the lemma holds and this finishes the proof.
\end{proof}
\begin{lemma}\label{lemma:2_split}
Let $H=(C' \cup I', E\cup F)$ be a PIG-completion of $G$ and $\{I_l, I_r\}$ be the partition of $I'$ into cliques as in the previous lemma.
Then, the graphs given by $H_l=(C'\cup I_l, E\cup F \setminus F_{I'})$ and $H_r=(C'\cup I_r, E\cup F \setminus F_{I'})$ are threshold graphs, where $F_{I'}= \{ uv \in F : u,v \in I' \}$.
\end{lemma}
\begin{proof}
If $H_l$ is not threshold, then there are vertices $v,w \in I_l$ (nonadjacent in $H_l$) and $x,y \in C'$ such that $vx, wy \in E\cup F \setminus F_{I'}$ and $vy, wx \not\in E\cup F \setminus F_{I'}$. Notice that $v$ and $w$ are adjacent in $H$ since they both lie in $I_l$, which is a clique. Therefore, we find an induced $C_4$ in $H$, which results in a contradiction for $H$ is a proper interval graph.
\end{proof}
This theorem follows directly from the previous lemmas.
\begin{theorem}\label{theo:1_split}
If $H=(C' \cup I', E\cup F)$ is a completion to PIG of $G$, then there is a partition $\{I_l, I_r\}$ of $I'$ such that $H_l, H_r$ are threshold and $I_l,I_r$ are cliques
We call $I_l, I_r$ the \emph{completion-partition of $H$}.
\end{theorem}
\subsection{NP-completeness}\label{subsec:reduction_split}
We are ready now to prove that obtaining a proper interval completion is still NP-complete when the input graph is split. In order to do this, we strongly rely on the previous lemmas and the fact that threshold- completion \suggestion{to} {on} split graphs is also NP-complete.
\begin{theorem}\label{theo:2_split}
The PIG-completion problem is NP-complete on split graphs.
\end{theorem}
\begin{proof}
Given a completion of a split graph, it is easy to check in polynomial time that this is in fact a PIG-completion, hence the problem is NP.
We give a reduction from threshold completion on split graphs.
Consider the graph $G'$ defined as follows. Let $G=(C\cup I,E)$ be a (connected for simplicity) split graph on $n$ vertices. Let $C'_1$ and $C'_2$ be two distinct cliques, each of size $n^2$, and $C_i= C \cup C'_i$, each containing a (different) copy of $C$.
We consider $G_i = (C_i \cup I_i, E)$, where each $I_i$ is a distinct copy of $I$, $C\cup C'_i$ is a clique, and the adjacencies between $C$ and $I_i$ are as in $G$, for each $i=1,2$.
Let $C'=C_1 \cup C_2$, and let $G'$ be a graph with vertex set $C' \cup I_1 \cup I_2$, such that $C'$ forms a large clique of size $2|C|+2n^2$ and the remaining adjacencies are as in $G_i$, for each $i=1,2$.
We show that $G'$ can be augmented to a proper interval graph with $2k$ additional edges if and only if $G$ can be augmented to a threshold graph with $k-\binom{|I|}{2}$ edges.
For the if case, we first show that there exists a number $k$ such that the proper interval completion number of $G'$ is $2k$. Notice that $G'$ is a starlike graph, thus it follows from Lemma~\ref{lemma:starlike_optimal_pd} that it admits an optimal normalized path decomposition, in which the central clique is denoted by $Y_{\pi(0)}$. Also, recall that we assume there is some $s\in L$ such that the central clique is precisely $Y_s$, and every peripheral clique $X_l, l \neq s, l \in L$ is either to the left or to the right of the central clique.
Let $(Y_j)$ be an optimal normalized path decomposition of $G'$. For this graph $G'$ and this particular optimal normalized path decomposition of $G'$, the following edge set $F_j \cup \{uv \mid u,v \in I_1 \lor u,v \in I_2 \}$ induces a proper interval completion. We denote by $H$ this PIG-completion of $G'$.
For this particular completion, if every vertex of $G_1$ appears on one side with respect to $Y_{\pi(0)}$ and all the vertices of $G_2$ appear on the other side, then the size of this proper interval completion is no greater than $n^2$. This follows from the fact that this quantity is at most as completing each copy of $G$ in $G_i$.
If there exist two vertices $u$ and $v$ on the same side such that $u\in V_1$ and $v \in V_2$, then the interval completion number is at least $n^2$,
thus the proper interval completion is also at least $n^2$, which is precisely what we need since we are trying to find some $k<n^2$.
Hence, it is impossible to have such $u$ and $v$ in $(Y_j)$ by construction. That is, $G_1$ and $G_2$ must appear in $(Y_j)$ on different sides. Since $G_1$ and $G_2$ have the same adjacency relation, each contributes the same number of edges to the proper interval completion of $G'$. Let this number be $k$, thus we obtain that the proper interval completion number of $G'$ is $2k$.
Let us see that, in this case, the completion-partition
of $H$ is precisely $I_1, I_2$.
Toward a contradiction, suppose that there is a vertex $v\in I_1$ that is nonadjacent in the completion $H$ to at least one vertex $v' \neq v$ in $I_1$. Notice that, since there are no induced nets, then in particular $N_H(v)$ and $N_H(v')$ are nested. It follows from this and the fact that $H$ is PIG, that $v$ is complete to $C_1$ since there are no claws in $H$. Moreover, $v$ is complete to $C'$, for if not there are claws between $v'$, any vertex in $C_1$ adjacent to $v'$ and any vertex in $C'\setminus C_1$. However, in this case we added at least $2n^2+|C|$ edges and hence this is not possible.
Therefore, if we consider $I_l=I_1, I_r=I_2$ and $H_l, H_r$ according to this completion-partition as defined in Lemma~\ref{lemma:2_split}, then $H_l, H_r$ are both threshold graphs, thus $G'$ can be augmented to an interval graph with $2k-2 \binom{|I|}{2}$ additional edges. It is not hard to verify that $k-\binom{|I|}{2}$ is the threshold completion number of $G$.
For the only if case, suppose that $G$ can be augmented to a threshold graph with $k'$ additional edges. We consider
$k = k'-|\{ uv \in F \mid u,v \in I \}|+\binom{|I|}{2}$. In other words, consider removed those fill edges between vertices of $I$
and then add the necessary ``quantity''
of edges to complete the independent partition.
It follows directly that $G$ can be augmented to a threshold graph with $k - \binom{|I|}{2}$ additional edges, since the fill edges between independent vertices are not necessary to have a threshold completion, for all we need is that these independent vertices should have nested neighbourhoods with regard to $C$.
Therefore, $G'$ can be augmented to a proper interval graph with $2k$ additional edges, just adding those $k - \binom{|I|}{2}$ fill edges on each copy of $G$ inside $G_i$ and completing each $I_i$ to a clique, for $i=1,2$.
\end{proof}
\section{An algorithm for PIG-completion on threshold graphs}\label{sec:threshold}
\suggestion{We present in this section}{In this section we present} a simple linear time algorithm for computing the optimal PIG-completion of a threshold graph $G$. For this, we will show first that PIG-completion for threshold graphs is equivalent to completion to a co-bipartite graph. This will enable us to give a procedure based on the definition of threshold graphs that iteratively places the vertices in one of the two cliques in an optimal way.\\
Let us consider the vertices in the order they are added such that whenever a vertex is added, it is either isolated or dominating. For simplicity, we say that a vertex is \emph{dominating} if it is a dominating vertex when added, and \emph{isolated} if it is an isolated vertex when added. Let us assume that the last vertex we add is dominating \suggestion{(otherwise the graph is not connected, and we apply our result on each connected component)}{, otherwise there are isolated vertices, which are irrelevant to the completion, and it is enough to solve the problem for the (single) nontrivial connected component, if it exists}.
\suggestion{We want to compute the minimum number of edges to add such that the resulting graph is a proper interval graph.}{}
\begin{lemma}\label{lemma:cobip_for_t}
\suggestion{}{Let $G$ be a threshold graph.} A minimum completion of $G$ into a proper interval graph is the same as a minimum completion of $G$ into a co-bipartite graph.
\end{lemma}
\begin{proof}
Since there is a dominating vertex\suggestion{}{in $G$} and the\suggestion{}{PIG-}completion is claw-free, its \suggestion{set of vertices}{vertex set} can be partitioned into two cliques.\suggestion{Assume}{Now, assume} that the vertex set of the completion can be partitioned into two cliques. \suggestion{There is obviously}{Thus, there is} no induced claw or asteroidal triple. Assume there is an induced cycle $C$ of length at least four. Note that $C$ has to be of size four since the graph is co-bipartite. Also note that two non-incident edges of $C$ are in $G$ (the ones that are not inside a clique). Now every vertex in $V(C)$ has at least one neighbour and one non-neighbour in $G[V(C)]$, so $G[V(C)]$ is not a threshold graph, and neither is $G$, a contradiction.
\end{proof}
Thus it suffices to exhibit an algorithm that computes a minimum completion into two cliques, $C_1$ and $C_2$, and thus to exhibit their respective sets of vertices $S_1$ and $S_2$.
Let us consider the following algorithm:
We add the vertices in order. When we add a vertex as a dominating vertex, we always put it in $S_1$. When we add a vertex as an isolated vertex, we compare the number of isolated vertices remaining to be added with the number of vertices already in $S_1$. If there are more remaining isolated vertices, we put it in $S_1$. Otherwise, we put it in $S_2$. This very simple algorithm runs in time $O(n)$.
We are going to prove the following:
\begin{theorem} \label{th_algo_opti}
The previous algorithm gives an optimal partition $(S_1,S_2)$ of $V(G)$.
\end{theorem}
\begin{lemma}
The number of edges to add is the sum for each $i \in \{1,2\}$, for each isolated vertex $v$ that is added in $S_i$, of the number of vertices in $S_{i}$ at the time $v$ is added (without counting $v$).
\end{lemma}
\begin{proof}
A way to build the completion is to take the vertices in order, and then, when we add an isolated vertex in $S_i$, add edges between $v$ and all of the vertices of $S_i$.
\end{proof}
Let us now consider an optimal partition $(S_{o1},S_{o2})$.
\begin{lemma} \label{lem_domi1}
Let $v$ be a dominating vertex, and assume that at least as many isolated vertices after $v$ are put in $S_{o2}$ as in $S_{o1}$. Then $v$ can be put in $S_{o1}$ without worsening the solution.
\end{lemma}
\begin{proof}
For $i \in \{1,2\}$, let $I_i$ be the number of isolated vertices after $v$ in $S_{oi}$. If $v$ is in $S_{o2}$, then moving it to $S_{o1}$ increases the solution cost by $I_1 - I_2 \le 0$.
\end{proof}
\begin{lemma} \label{lem_domi2}
Let $v$ be a dominating vertex, and assume that when we add $v$ there are more vertices in $S_{o1}$ than in $S_{o2}$, then there is an optimal solution with the same partition up to $v$, such that $v$ is in $S_{o1}$.
\end{lemma}
\begin{proof}
For $i \in \{1,2\}$, let $k_i$ be the number of vertices in $S_{oi}$ before $v$, and let $I_i$ be the number of isolated vertices after $v$ in $S_{oi}$. We have $k_1 \ge k_2$.
If $I_1 \ge I_2$, then exchanging the vertices from $v$ on between sets $S_{o1}$ and $S_{o2}$ increases the solution cost by $(I_2 - I_1)(k_1 - k_2) \le 0$. Now $I_1 \le I_2$, so by Lemma~\ref{lem_domi1}, we know that we can put $v$ in $S_{o1}$ if it is not there already.
\end{proof}
\begin{lemma} \label{lem_domi3}
There is an optimal solution such that every dominating vertex is put in the same clique, say $S_{o1}$. Moreover, in this solution there are always at least as many vertices in $S_{o1}$ as in $S_{o2}$.
\end{lemma}
\begin{proof}
Consider an optimal solution and its construction according to the order of the vertices. Let us transform this construction into another optimal solution.
We are going to maintain that there are always at least as many vertices in $S_{o1}$ as in $S_{o2}$, and that every dominating vertex is put into $S_{o1}$. Whenever the solution would put a dominating vertex in $S_{o2}$, we modify the solution by Lemma~\ref{lem_domi2}. When the solution would put an isolated vertex $v$ in $S_{o2}$ while there are as many vertices in $S_{o1}$ as in $S_{o2}$, exchange all the remaining vertices (including $v$) between $S_{o1}$ and $S_{o2}$, and we get a solution of the same cost.
\end{proof}
Let us now consider an optimal solution respecting Lemma~\ref{lem_domi3}.
\begin{lemma} \label{lem_iso}
Let $v$ be an isolated vertex.
Suppose that every vertex before $v$ is in $S_{o1}$ and that there are at least as many remaining isolated vertices as there are vertices before $v$.
Then there is another optimal solution respecting Lemma~\ref{lem_domi3} with the same partition up to $v$ such that $v$ is in $S_{o1}$.
\end{lemma}
\begin{proof}
Assume that $v$ is in $S_{o2}$. For $i \in \{1,2\}$, let $I_i$ be the number of isolated vertices after $v$ in $S_{oi}$.
Let $k$ be the number of vertices that are before $v$ in the order. Note that they are all in $S_{o1}$.
Moving $v$ from $S_{o2}$ to $S_{o1}$ increases the number of edges in the solution by $I_1 + k - I_2$. By hypothesis, $k \le (I_2 + I_1)$, so if $I_1 = 0$, then $I_1 + k - I_2 \le 0$, and we get another optimal solution respecting Lemma~\ref{lem_domi3}.
Now we may assume that $I_1 \ne 0$. Let $v'$ be the first isolated vertex put in $S_{o1}$ after $v$. For $i \in \{1,2\}$, let $k'_i$ be the number of vertices in $S_{oi}$ before $v'$, and let $I'_i$ be the number of isolated vertices after $v'$ in $S_{oi}$. Note that since our solution respects Lemma~\ref{lem_domi3}, every vertex in $S_{o2}$ is isolated, so $k_2' + I_2' = I_2+1$.
By putting $v$ in $S_{o1}$ and $v'$ in $S_{o2}$ we increase the cost of the solution by $m = I_1 + k - I_2 + (k_2' - 1 + I_2' - (k_1' + 1) - I_1')$. As noted previously, $k_2' + I_2' = I_2 + 1$, so $m = I_1 + k -k_1' - I_1' - 1$.
Moreover, as every isolated vertex after $v$ in $S_{o1}$ is either a vertex before $v'$, or $v'$, or an isolated vertex after $v'$, we get that $k_1' + I_1' + 1\ge k + I_1$, so $m \le 0$. Therefore we have a solution with the same partition up to $v$ and such that $v$ is in $S_{o1}$. Applying the proof of Lemma~\ref{lem_domi3} form $v$ on, we can change it into an optimal solution respecting Lemma~\ref{lem_domi3}, with the same partition up to $v$ and such that $v$ is in $S_{o1}$.
\end{proof}
\begin{proof}[Proof of Theorem~\ref{th_algo_opti}]
Iterating Lemma~\ref{lem_iso}, we get an optimal solution such that every vertex of $S_1$ (from our solution) is in $S_{o1}$. Suppose that there is a vertex $v$ in $S_2 \cap S_{o1}$. By construction, this vertex is such that there are more vertices before it in $S_{o1}$ than remaining isolated vertices. Therefore moving $v$ from $S_{o1}$ to $S_{o2}$ improves the solution, a contradiction. Therefore the optimal solution $(S_{o1},S_{o2})$ is equal to the solution $(S_1,S_2)$ of Theorem~\ref{th_algo_opti}.
\end{proof}
To conclude this section, we mention the connection to the max-cut problem in the case of threshold graphs. Given a partition $(A,B)$ of the vertices of a graph $G$, consider these two sets:
\begin{enumerate}
\item The set of pairs of non-adjacent vertices, where one element is in $A$ and the other in $B$. Let us call it $C$.
\item The set of pairs of non-adjacent where both are in $A$ or both in $B$. Let us call it $F$ (for threshold graphs, this is a completion to PIG).
\end{enumerate}
Note that $E(G)$, $C$, and $F$ are a partition of the set of pairs of $V(G)$, hence $|E(G)| + |C| + |F| = {n \choose 2}$. Two of these four values depend only on the input graph, not on the partition. Hence, choosing a partition minimizing $|F|$ is the same as maximizing $|C|$. Now take the complement graph and consider the same partition. $C$ is clearly a cut between $A$ and $B$. So for threshold graphs, minimizing the size of the completion is equivalent to finding the maximum cut of the complement.\\
We remark that an $O(n^2)$-algorithm for calculating max-cut for cographs (and thus for threshold graphs) has been presented in \cite{Bodlaender2000}. This result, together with the fact that threshold graphs are closed under complementation, gives an alternative algorithm for PIG-completion for this class (albeit a less efficient one).
|
\section{Introduction}
Document-level information extraction \citep{Yao2019, Christopoulou2019, Xiao2020, nan2020lsr} has seen great strides due to the rise of pre-trained models \cite{Devlin2019}. But in high-stakes domains like medical information extraction \citep{Irvin2019, Matthew2020, Smit2020}, machine learning models are still too error-prone to use broadly. Since they are not perfect, they typically play the role of assisting users in tasks like building cohorts \cite{Pons_2016_nlp_in_rad} or in providing clinical decision support \cite{DemnerFushman_cds}.
\liyan{To be most usable in conjunction with users, these systems should not just produce a decision, but a justification for their answer. The ideal system therefore obtains high predictive accuracy, but also returns a rationale that allows a human to verify the predicted label \cite{RudieEtAl2019}.}
\begin{figure}
\centering
\includegraphics[scale=0.75,trim=0 89mm 0mm 27mm]{liyan_intro_update.pdf}
\caption{Our basic model setup. A Transformer-based model makes document-level predictions on an example of our brain MRI reports. An interpretation method extracts the evidence sentences used by the model. Our system is evaluated according to the criteria of accuracy, faithfulness, and plausibility.
}
\label{fig:architecture}
\end{figure}
Our goal is to study document-level information extraction systems that are both accurate and which make predictions based on the correct information \citep{DoshiVelez2017}. This process involves identifying what evidence the model actually used, verifying the model's prediction based on that evidence, and checking whether that evidence aligns with what humans would use, which would allow a user to quickly see if the system is correct. For example, in Figure~\ref{fig:architecture}, localizing the prediction of \emph{mass effect} (a feature expressing whether there is evidence of brain displacement by a mass like a tumor) to the first two sentences allows a trained user in a clinical decision support setting to easily verify what was extracted here. Our evidence extraction hews to principles of both faithfulness and plausibility \citep{Jain2020, Jacovi2020, Miller2019}.
Rather than use complex approaches with intermediate latent variables for extraction \cite{Lei2016}, we focus on what can be done with off-the-shelf pre-trained models \cite{liu2019} using post-hoc interpretation. We explore various interpretation methods to find key parts of each document that were used by the model. We ask two questions: first, can we identify the document sentences that truly contributed to the prediction (faithfulness)? Using the ranking of sentences provided by an interpretation method, we extract a set of sentences where the model returns nearly the same prediction as before, thus \emph{verifying} that these sentences are a sufficient explanation for the model. Second, do these document sentences align with what users annotated (plausibility)? Unsurprisingly, we find that this alignment is low in a basic Transformer model.
To further improve the alignment with human annotation, we consider injecting small amounts of sentence-level supervision. Critically, in the brain MRI extraction setting we consider (see Table \ref{ReportExample}), large-scale sentence-level annotation is not available; most instances in the dataset only have document-level labels from existing clinical decision support systems, making it a weakly-supervised setting \citep{Pruthi2020-2, Patel2020}. We explore two methods for using this small amount of annotation, chiefly based around supervising or regularizing the model's behavior. One notion is entropy maximization: the model should be uncertain when it isn't exposed to sufficient evidence \citep{Feng2019}. Another is attention regularization where the model is encouraged to attend to key pieces of evidence. While attention is not entirely connected with what the model uses \citep{Jain2019}, we can investigate whether this leads to a model whose explanations leverage this information more heavily.
We validate our methods first on a small dataset of radiologists' observations from brain MRIs. These reports are annotated with document-level key features related to different aspects of the report, which we want to extract in a faithful way. We see positive results here even in a small-data condition, but to understand how this method would scale with larger amounts of data, we adapt the DocRED relation extraction task \cite{Yao2019} to be a document-level classification task. The question of which sentence in the document describes the relation between the two entities, if there even is one, is still quite challenging, and we show our techniques can lead to improvements in a weakly-labeled setting here as well.
Our contributions are (1) We apply evidence extraction methods to document-level classification and slot-filling tasks, emphasizing a new brain MRI dataset that we annotate. (2) We explore using weak sentence-level supervision in two techniques adapted from prior work; (3) We evaluate pre-trained models and evidence extraction through various interpretation methods for plausibility compared to human annotation, while ensuring faithfulness of the evidence
\begin{table}[]
\small
\begin{tabular}{p{0.94\linewidth}} \toprule
\multicolumn{1}{c}{\textbf{Report Finding}} \\
\emph{[0]} \textbf{\textcolor{Red}{Severe encephalomalacia}} in the temporal lobes and frontal lobes \textbf{\textcolor{Orange}{bilaterally}} with reactive \textbf{\textcolor{RoyalBlue}{gliosis}} in the left frontal lobe. \emph{[1]} Moderate \textbf{\textcolor{Red}{enlargement of the ventricular system}}. \emph{[2]} \textbf{\textcolor{OliveGreen}{No abnormal enhancement}}. \emph{[3]} Near complete opacification of the left maxillary sinus. ...\\
\\
\begin{tabular}{llr}
\textcolor{Red}{\textbf{mass\_effect}}: negative &\textbf{evid}: \emph{[0, 1]} & implicit\\
\textcolor{Orange}{\textbf{side}}: bilateral &\textbf{evid}: \emph{[0]} & explicit\\
\textcolor{RoyalBlue}{\textbf{t2}}: increased &\textbf{evid}: \emph{[0]} & implicit\\
\textcolor{OliveGreen}{\textbf{contrast\_enhancement}}: No &\textbf{evid}: \emph{[2]} & explicit \\
\end{tabular} \\
\bottomrule
\end{tabular}
\caption{Example from annotated brain MRI reports. Labels and supporting evidence for $4$ key features are annotated for this example report presented. ``Explicit'' means the label of given key feature can be directly inferred by the highlighted terms; ``implicit'' instead indicates that it requires domain knowledge and potential reasoning skills to label. We want the model to identify implicit features while not leveraging dataset biases or reasoning incorrectly about explicit ones.}
\label{ReportExample}
\end{table}
\section{Background} \label{sec:background}
\subsection{Motivation}
We start with an example from a brain MRI report in Table~\ref{ReportExample}. Medical information extraction involves tasks such as identifying important medical terms from text \citep{Irvin2019, Smit2020} and normalizing names into standard concepts using domain-specific ontologies \citep{Cho2017}. One application in clinical decision support, shown here, requires extracting the values of certain key features (clinical findings) from these reports or medical images \cite{Rudie2021, Duong2019}. This extraction should be accurate, but it should also make predictions that are correctly sourced, to facilitate review by a radiologist or someone else using the system \cite{Rauschecker2020, Cook2018}.
The finding section of a brain MRI report often describes these key features in both explicit and implicit ways. For instance, contrast enhancement, one of our key features, is mentioned explicitly much of the time; see \emph{no abnormal enhancement} in the third sentence. A rule-based system can detect this type of evidence easily. But some key features are harder to identify and require reasoning over context and draw on implicit cues. For example, \emph{severe encephalomalacia} in the first sentence and \emph{enlargement of the ventricular system} in the following sentence are both implicit signs of positive mass effect and either is sufficient to infer the label. It is significantly harder to built a rule-based extractor for this case. Learning-based systems have the potential to do much better here, but lack of understanding about their behavior can lead to hard-to-predict failure modes, such as acausal prediction of key features (e.g., inferring evidence about mass effect from a hypothesized diagnosis somewhere in the report, where the causality is backwards).
Our work aims to leverage the ability of learning-based systems to capture implicit features while improving their ability to make predictions that are sourced from the correct evidence and can be easily verified.
\subsection{Problem Setting}
The problem we tackle in this work can be viewed as document-level classification. Let $D = \{x_1, \ldots, x_n\}$ be a document consisting of $n$ sentences. The document is annotated with a set of labels $(t_i,y_i)$ where $t_i$ is an auxiliary input specifying a particular task for this document (e.g., mass effect) and $y_i$ is the label associated with that task from a discrete label space $\{1,\ldots,d\}$. In our adaptation of the DocRED task, we consider $t = (e_1,e_2)$ to classify the relationship (if any) between a pair of entities $(e_1,e_2)$ in a document, defined in Section \ref{subsection: Docred}.
Our method takes a pair $(D,t)$ and then computes the label $\hat{y}_t$ from a predictor $\hat{y}_t = f(D,t)$. We can then extract \textbf{evidence}, a set of sentences, post-hoc using a separate procedure $g$ such as a feature attribution method: $\hat{E}_t = g(f,D,t)$
\paragraph{Supervision} In addition to the labels $y_t$, we assume access to a small number of examples with additional supervision in each domain. That is, for a $(D,t,y_t)$ triple, we also assume we are given a set $E = \{x_{i_1},\ldots, x_{i_m}\}$ of ground-truth evidence with sentence indices $\{i_1,\ldots,i_m\}$. This evidence should be sufficient to compute the label, but not always necessary; for example, if multiple sentences can contribute to the prediction, they might all be listed as supporting evidence here. See Section~\ref{sec:improving_ee} for more details.
\subsection{Related Work}
Our work fits into a broader thread of work on information extraction with partial annotation \citep{Han2020}. Due to the cost of collecting large-scale data with good quality, distant supervision (DS) \citep{mintz2009} and ways to denoise auto-labeled data from DS \citep{surdeanu2012, wang2018} have been widely explored. However, the sentence-level setting typically features much less ambiguity about evidence needed to predict a relation compared to the document-level setting we explore. Several document-level RE datasets \citep{Li2016-2, Peng2017} have been proposed as well as efforts to tackle these tasks \citep{Christopoulou2019, Xiao2020, nan2020lsr}, which we explicitly build from.
\paragraph{Explanation techniques} To identify the sentences that the model considers as evidence, we draw on a recent body of work in explainable NLP focused on identifying salient features of the input. These primarily consist of input attribution techniques, such as LIME \citep{Ribeiro2016}, input reductions \citep{Li2016-1, Feng2018}, attention-based explanations \citep{Bahdanau2015} and gradient-based methods \citep{Simonyan2014, Selvaraju2017, sundararajan2017, Shrikumar2017}. \liyan{In present work, we extract rationales using commonly used model interpretation methods (described in Section~\ref{sec:evi_extract}) and focus on doing a thorough evaluation of the capabilities of DeepLIFT \citep{Shrikumar2017} given its competitive performance in our interpretation methods comparison (Appendix~\ref{sec:interpret_compare}).}
\paragraph{Frameworks for interpretable pipelines} Our goal of building a system grounded in evidence draws heavily on recent work on attribution techniques and model explanations, particularly notions of faithfulness and plausibility. \emph{Faithfulness} refers to how accurately the explanation provided by the model truly reflects the information it used in the reasoning process \citep{Jain2020}. On the other hand, \emph{plausibility} indicates to what extent the interpretation provided by the model makes sense to a person.\footnote{The ERASER benchmark \cite{DeYoung2020} is a notable recent effort to evaluate explanation plausibility. However, we do not consider it here; we focus on the document-level classification setting, and many of the ERASER tasks are not suitable or relevant for the approaches we consider, either being not natural (FEVER) or not having the same challenges as document-level classification.}
``\emph{Select-then-predict}'' approaches are one way to enforce faithfulness in pipelines \cite{Jain2020}: important snippets from inputs are extracted and passed through a classifier to make predictions. Past work has used hard \citep{Lei2016} or soft \citep{Zhang2016} rationales, and other work has explicitly looked at tradeoffs in the amount of text extracted \citep{Paranjape2020}.
\citet{Jacovi2020} note several problems with this setup. Our work aims to align model behavior with what cues we expect a model to use (plausibility), but uses the predict-select-verify paradigm \cite{Jacovi2020} to ensure that these are actually sufficient cues for the model. Like our work, \citet{Pruthi2020-2} simultaneously trained a BERT-based model \citep{Devlin2019} for the prediction task and a linear-CRF \citep{Lafferty2001} module on top of it for the evidence extraction task with shared parameters. Compared to their work, we focus explicitly on what can be done with pre-trained models alone, not augmenting the model for evidence extraction.
\section{Methods}
The systems we devise take $(D,t)$ pairs as input and return (a) predicted labels $\hat{y}_t$ for each $t$; (b) sets of extracted evidence sentences $\hat{E}_t$ from an interpretation method. Figure~\ref{fig:architecture} shows the basic setting.
\subsection{Transformer Classification Model}
We use RoBERTa \citep{liu2019} as our document classifier. \liyan{RoBERTa is a strong method that holds up even against more recent baselines with architectures designed for DocRED \cite{zhou2021}.}
For each of our two domains, we use different pre-trained weights, as described in the training details in Appendix \ref{section:trainingdetails}. The task inputs are described in Section~\ref{sec:data_eval}.
\subsection{Interpretation for Evidence Extraction} \label{sec:evi_extract}
\liyan{Given any interpretation method as well as our model $\hat{y}_t = f(D,t)$}, we compute attribution scores with respect to the predicted class $y_t$ for each token in the RoBERTa input representation. We then average over the absolute value of attribution score for each token in that sentence to give sentence-level scores $\{s_1,\ldots,s_n\}$. These give us a ranking of the sentences. Given a fixed number of evidence sentences $k$ to extract, we can extract the top $k$ sentences by these scores.
\liyan{We experiment with the following four widely used interpretation techniques in the present work. \textbf{LIME} \cite{Ribeiro2016} offers explanations of an input by approximating the model’s predictions locally with an interpretable model. \textbf{Input Gradient} \cite{hechtlinger2016} and \textbf{Integrated Gradients} \cite{sundararajan2017} use gradients of the label with respect to the input to assess input importance; Integrated Gradients approximates the integral of this gradient with respect to the input along a straight path from a reference baseline.\footnote{We use the most typical baseline that consists of replacing the inputs in $D$ with {\fontfamily{qcr}\selectfont [MASK]} tokens from RoBERTa.} \textbf{DeepLIFT} \citep{Shrikumar2017} attributes the change in the output from a reference output in terms of the difference in input from the reference input. Unless stated otherwise, we use DeepLIFT as our interpretation method, since it achieves the best results (comparable to Input Gradient) among the four interpretation options. Full comparison of interpretation methods is in Appendix~\ref{sec:interpret_compare}.}
To verify the extracted evidence \cite{Jacovi2020}, our main technique (\textsc{Sufficient}) feeds the model increasingly larger subsets of the document ranked by attribution scores (e.g., first $\{s_\textrm{max}\}$, then $\{s_\textrm{max}, s_\textrm{2nd-max}\}$, etc.) until it (a) makes the same prediction as when taking the whole document as input and (b) assigns that prediction at least $\lambda$ times the probability\footnote{The value of $\lambda$ is a tolerance hyper-parameter for selecting sentences and it set to $0.8$ throughout the experiments. \liyan{Our method is robust to the choice of $\lambda$ in a reasonable range, as shown in Appendix~\ref{sec:interpret_compare}}.} when the whole document is taken as input. We consider this attribution faithful: it is a subset of the input supporting the model's decision judged as important by the attribution method.
\subsection{Improving Evidence Extraction}
\label{sec:improving_ee}
While many document-level extraction settings do not have sentence-level attributions labeled for every decision, one can in practice annotate a small fraction of a dataset with such ground-truth rationales. This is indeed the case for our brain MRI case study. Past work has shown significant benefits from integrating this supervision into learning \cite{strout-etal-2019-human,DuaEtAl2020,Pruthi2020-1}.
Assume that a subset of our labeled data consists of $(D, t, y_t, E_t)$ tuples with ground truth evidence sentence indices $E_t = \{i_1, ..., i_m\}$. We consider two modifications to our model training, namely attention regularization \citep{Pruthi2020-1}, entropy maximization \citep{Feng2018}, and their combination. An illustration of both methods is shown in Figure \ref{fig:illustraction}.
\paragraph{Attention regularization} Attention regularization encourages our model $f(D, t)$ to leverage more information from $E_t$. Specifically, let $A = \{\alpha_1, ..., \alpha_n\}$ be the attention vector from the {\fontfamily{qcr}\selectfont [CLS]} token in the final layer to all tokens in $D$. During learning, we add the following loss to the training objective: $\ell_{attn} = - \log \sum_{i \in E_t} \alpha_{i}$, encouraging the model to attend to any token $i$ in the labeled sentence-level evidence set.
\paragraph{Entropy maximization} When there is no sufficient information contained in the text to infer any predictions, entropy maximization encourages a model to be uncertain, represented by a uniform probability distribution across all classes \citep{DeYoung2020, Feng2019}. Doing so should encourage the model to \emph{not} make predictions based on irrelevant sentences. We can achieve this by taking a reduced document $D' = D \setminus E_t$ as input by removing evidence $E_t$ from original document $D$. We treat $(D', t)$ pairs as extra training examples where we aim to maximize the entropy $- \sum_y P(y|D')\log P(y|D')$ over all possible $y$.\footnote{\liyan{We found this to work better than enforcing a uniform distribution over attention, which is much harder for the model to achieve.}}
\section{Experiments}
\begin{figure}
\small
\centering
\includegraphics[scale=0.19,trim=0 30mm 0mm 30mm]{illustration.pdf}
\caption{An illustration of attention regularization and entropy maximization using the example in Table \ref{ReportExample}. The model is predicting the label for key feature \emph{\text{t2}}.}
\label{fig:illustraction}
\end{figure}
\subsection{Datasets and Evaluation Metrics}
\label{sec:data_eval}
We investigate our methods on (a) a small collection of brain MRI reports from radiologists’ observations; and (b) a modified version of the DocRED datatset. The statistics for both datatsets are included in Appendix \ref{section:statistics}. For both datasets, we evaluate on task accuracy (captured by either accuracy or prediction macro-F1) as well as evidence selection accuracy (macro-F1) or precision, measuring how well the model's evidence selection aligns with human annotations. We will use the \textsc{Sufficient} method defined in Section~\ref{sec:evi_extract} to select evidence sentences which guarantee that our predictions on the given evidence subsets will match the model's predictions on the full document. For the brain MRI report dataset, we evaluate evidence extraction by precision since human annotators typically only need to refer to one sentence to reach the conclusion but our model and baselines may extract more than one sentence.
\subsubsection{Brain MRI Reports} \label{sec:brainreport}
We present a new dataset of de-identified radiology reports from brain MRIs. It consists of the ``findings'' sections of reports, which present observations about the image, with labels for pre-selected key features by attending physicians and fellows. Crucially, these features are labeled \textbf{based on the original radiology image}, not the report. The document-level labels are therefore noisy because the radiologists' labels may disagree with the findings written in the report.
A key feature is an observable variable $t$, which can take on $d_t$ possible values. We focus on the evaluation of two key features, namely \emph{contrast enhancement} and \emph{mass effect}, since they appear in most of manually annotated reports. For our RoBERTa classification model, we only feed the document and train separate classifiers for each key feature, with no shared parameters between these.
\paragraph{Annotation} We have a moderate number (327) of reports that have noisy labels from the process above. We treat these as our training set. However, all of these labels are document-level.
To evaluate models' performance on more fine-grained evidence labels, we randomly select $86$ unlabeled reports (not overlapping with the 327 for training) and asked four radiology residents to (1) assign key feature labels and reach consensus, while (2) highlighting sentences that support their decision making. We use Prodigy\footnote{\url{https://prodi.gy}} as our annotation interface. See Appendix~\ref{sec:instruction} for more details about our annotation instructions.
\paragraph{Pseudo sentence-level supervision} Since we have limited number of annotated reports for evaluation, we need a way to prepare weak sentence-level supervision ($E_t$) while training. To achieve this, we use sentences selected by our rule-based system as pseudo evidence to supervise models' behavior. We use 10\% of this as supervision while training for consistency with the DocRED setting.
\paragraph{Rule-based system} Our rule-based system uses keyword matching to identify instances of mass effect and contrast enhancement in the reports, and negspaCy
to detect negations of these key features.
\paragraph{Data split} For the results in Section~\ref{Results}, we evaluate on reports that contain ground truth fine-grained annotations for either contrast enhancement or mass effect, respectively. There are 64 and 68 out of 86 documents total in each of these categories. We call this the \textsc{BrainMRI} set. When we restrict to this set for evaluation, all of the documents we study where the annotators labeled something related to \emph{contrast enhancement} end up having an explicit mention of it. However, for \emph{mass effect}, this is not always the case; Table~\ref{tab:reportexample} in Appendix shows an example where mass effect is discussed implicitly in the first sentence.
\subsubsection{Adapted DocRED}
\label{subsection: Docred}
DocRED \citep{Yao2019} is a document-level relation extraction (RE) dataset with large scale human annotation of relevant evidence sentences. Unlike sentence-level RE tasks \citep{Qin2018, alt2020}, it requires reading multiple sentences and reasoning about complex interactions between entities. We adapt this to a document-level relation classification task: a document $D$ and two entity mentions $e_1, e_2$ within the document are provided and the task is to predict the relation $r$ between $e_1$ and $e_2$. We synthesize these examples from the original dataset and sample random entity pairs from documents to which we assign an \emph{NA} class to construct negative pairs exhibiting no relation.
The model input is represented as: {\small\text{{\fontfamily{qcr}\selectfont[CLS]<ent-1>[SEP]<ent-2>[SEP]<doc>[SEP]}}}. \liyan{We use the encoding of \text{\fontfamily{qcr}\selectfont[CLS]} in the last layer to make predictions.}
To make the setting more realistic, we \emph{do not} use the large-scale evidence annotation and assume there is limited sentence-level supervision available. To be specific, we include 10\% fine-grained annotations in our adapted DocRED dataset.
\begin{table}
\small
\begin{tabular}{p{0.25\linewidth} p{0.6\linewidth}}
\toprule
\textbf{Model Names} & \textbf{Input Text} \\ \midrule
\textsc{Direct} & None \\
\textsc{FullDoc} & Full document \\
\textsc{Ent} & Sentences containing at least one of the two query entities \\
\textsc{First2} & First two sentences \liyan{from a doc}. \\
\textsc{First3} & First three sentences \liyan{from a doc}.\\
\textsc{BestPair} & Two sentences yielding highest prediction prob. (incl.~variants using regularization) \\
\textsc{Sufficient} & Sufficient sentences selected by interpretation methods (incl.~variants using regularization) \\
\bottomrule
\end{tabular}
\caption{\label{notations} Model names used in the experiments and their associated evidence given as inputs.}
\end{table}
\subsection{Models}
\label{subsection:models}
Due to richer and higher-quality supervisions in the DocRED setting, we conduct a larger set of ablations and comparisons there. We compare against a subset of these models in the radiology setting.
\paragraph{Baselines} We consider a number of baselines for adapted DocRED which return both predicted labels and evidence. (1) \textsc{Direct} predicts the relation directly from the entity pairs without any sentences as input, using a model trained with just these inputs. (2) \textsc{FullDoc} takes the full document as selected evidence and uses the base RoBERTa model (3) \textsc{Ent} takes all sentences with entity mentions $e_1$ and $e_2$ as input; (4) \textsc{First2, First3} retrieve the first $2$ and $3$ sentences from a document, respectively; and (5) \textsc{BestPair} chooses the best sentence pair by first taking each individual sentence as input to the model and then picking top two sentences having highest probabilities on their predictions.
\textsc{Sufficient} is our main method for both datasets, which we then augment with additional supervision as described in Section~\ref{sec:improving_ee}. We use subscripts {\fontfamily{qcr}\selectfont attn, entropy, both} and {\fontfamily{qcr}\selectfont none} to represent attention regularization, entropy maximization, the combination of two, and neither. \liyan{Both \textsc{BestPair} and \textsc{Sufficient} methods leverage backbone RoBERTa models trained with loss functions mentioned above, differing only in their evidence selection.}
Table \ref{notations} summarizes the abbreviated names of models and their inputs. Training details are described in Appendix \ref{section:trainingdetails}.
\paragraph{Metrics} We report both the accuracy and F$_1$ for the model (\textbf{Full Doc}) as well as evaluation of \textbf{Evidence} selection compared to human judgments, either precision or F$_1$. We also report results in the \textbf{Reduced Doc} setting, where only the selected evidence sentences are fed to the RoBERTa model (trained over whole documents) as input. For our \textsc{Sufficient} method, this accuracy is the same as the full method by construction, but note that it can differ for other methods. This reduced setting serves as a sanity check for the faithfulness of our explanation techniques.
\liyan{Note once again that accuracy in the Full Doc case can differ for our methods that are trained with different regularization schemes, as these yield different models that return different predicted labels in addition to different evidence.}
\begin{table}
\small
\centering
\renewcommand{\tabcolsep}{1.1mm}
\begin{tabular}{lcccccc}
\toprule
\multirow{3}{*}{\textbf{Model}} &
\multicolumn{4}{c}{\textbf{Label}} &
\multicolumn{2}{c}{\textbf{Evidence}} \\
&\multicolumn{2}{c}{\textbf{Full Doc}} &
\multicolumn{2}{c}{\textbf{Reduced Doc}}&
\multicolumn{2}{c}{} \\
& {\textbf{Acc}} & {\textbf{F1}} & {\textbf{Acc}} & {\textbf{F1}} & {\textbf{Pre}} & \textbf{Len}\\
\midrule
\multicolumn{7}{c}{Mass Effect} \\ \midrule
\liyan{\textsc{FullDoc}} & 66.6 & 42.1 & 66.6 & 42.1 & 16.5 & 10.1 \\
\liyan{\textsc{First2}} & $-$ & $-$ & 82.4 & 45.2 & 21.3 & 2.00\\
\liyan{\textsc{First3}} & $-$ & $-$ & 82.4 & 45.2 & 24.0 & 3.00\\
\textsc{Rule} & 77.9 & 11.8 & 77.9 & 11.8 & \textbf{\underline{84.8}} & 1.46\\
\midrule
\liyan{\textsc{BestPair}\textsubscript{none}} & 66.6 & 42.1 & 82.4 & 52.2 & 24.3 & 2.00 \\
\liyan{\textsc{BestPair}\textsubscript{both}} & 76.7 & \textbf{60.0} & 79.4 & 44.3 & \underline{50.7} & 2.00 \\
\midrule
\textsc{Sufficient}\textsubscript{none} & 66.6 & 42.1 & \multicolumn{2}{c}{\multirow{4}{1.5cm}{\centering Identical to Full Doc}} & 16.5 & 2.84 \\
\textsc{Sufficient}\textsubscript{attn} & 69.2 & 47.6 & \multicolumn{2}{c}{} & 65.6 & 2.31 \\
\textsc{Sufficient}\textsubscript{entropy} & 45.3 & 0.0 & \multicolumn{2}{c}{} & 15.8 & 2.50 \\
\textsc{Sufficient}\textsubscript{both} & 76.7 & \textbf{60.0} & \multicolumn{2}{c}{} & \underline{77.8} & 1.51 \\
\midrule
\multicolumn{7}{c}{Contrast Enhancement}\\ \midrule
\liyan{\textsc{FullDoc}} & 69.5 & 60.9 & 69.5 & 60.9 & 13.5 & 10.1 \\
\liyan{\textsc{First2}} & $-$ & $-$ & 67.2 & 55.3 & 14.1 & 2.00\\
\liyan{\textsc{First3}} & $-$ & $-$ & 70.3 & 62.4 & 14.6 & 3.00\\
\textsc{Rule} & 68.8 & 56.5 & 68.8 & 56.5 & \textbf{\underline{87.1}} &1.67\\
\midrule
\liyan{\textsc{BestPair}\textsubscript{none}} & 69.5 & 60.9 & 73.4 & 67.7 & 10.9 & 2.00 \\
\liyan{\textsc{BestPair}\textsubscript{both}} & \underline{90.8} & \textbf{\underline{87.2}} & \underline{89.1} & \underline{88.4} & \underline{54.7} & 2.00 \\
\midrule
\textsc{Sufficient}\textsubscript{none} & 69.5 & 60.9 & \multicolumn{2}{c}{\multirow{4}{1.5cm}{\centering Identical to Full Doc}} & 33.5 & 2.84 \\
\textsc{Sufficient}\textsubscript{attn} & \underline{85.8} & \underline{81.0} & \multicolumn{2}{c}{} & \underline{60.7} & 2.48 \\
\textsc{Sufficient}\textsubscript{entropy} & 71.5 & 59.5 & \multicolumn{2}{c}{} & 25.2 & 2.55 \\
\textsc{Sufficient}\textsubscript{both} & \underline{90.8} & \textbf{\underline{87.2}} & \multicolumn{2}{c}{} & \underline{71.7} & 1.50 \\
\bottomrule
\end{tabular}
\caption{\label{tab:brainmri} Model performance on \textsc{BrainMRI}. Models are evaluated under two settings by taking (a) full document (Full Doc); (b) selected evidence (Reduced Doc) as inputs. \textsc{Rule} is the baseline mentioned in Section \ref{sec:brainreport}. \emph{Pre} stands for the precision of evidence selection, and \emph{Len} is the average number of sentences extracted. Underlined results are better than \textsc{Sufficient}\textsubscript{none} on the corresponding metric according to a paired bootstrap test with $p=0.05$.}
\end{table}
\section{Results}
\label{Results}
\subsection{Results on Brain MRI}
Table~\ref{tab:brainmri} shows the performance of our models and baselines in terms of label prediction and evidence extraction. \liyan{For each result, we perform a paired bootstrap test comparing to \textsc{Sufficient}\textsubscript{none}. We underline results that are better at a significance level of $p=0.05$ on the corresponding metrics.} In the \emph{mass effect} setting, our \textsc{Sufficient}\textsubscript{both} model achieves the highest evidence extraction precision of the learning-based models, \liyan{exceeds \textsc{FullDoc}, \textsc{First2/3}, and \textsc{BestPair} on the metric by a large margin,} and nearly matches that of the rule-based system. It is difficult to be more reliable than a rule-based system, which will nearly always make correctly-sourced predictions. But this model is able to \textbf{combine that reliability with the higher F$_1$ of a learned model}. Note that due to the high base rates of certain findings, we focus on F$_1$ instead of accuracy. We see a similar pattern on \emph{contrast enhancement} as well, although the evidence precision is lower in that case.
\begin{table}
\small
\centering
\renewcommand{\tabcolsep}{1.2mm}
\begin{tabular}{lcccccc}
\toprule
\multirow{3}{*}{\textbf{Model}}
& \multicolumn{4}{c}{\textbf{Label}}
& \multicolumn{2}{c}{\textbf{Evidence}} \\
& \multicolumn{2}{c}{\textbf{Full Doc}}
& \multicolumn{2}{c}{\textbf{Reduced Doc}}
& \multicolumn{2}{c}{\textbf{}} \\
& \textbf{Acc} & \textbf{F1} & \textbf{Acc} & \textbf{F1} & \textbf{F1} & \textbf{Len}\\
\midrule
\textsc{Direct} & $-$ & $-$ & 66.4 & 45.3 & $-$ & $-$ \\
\textsc{FullDoc} & 83.0 & 66.0 & 83.0 & 66.0 & 34.9 & 8.03\\
\textsc{First2} & $-$ & $-$ & 75.3 & 58.1 & 47.9 & 2.00 \\
\textsc{First3} & $-$ & $-$ & 77.5 & 60.7 & 44.6 & 3.00 \\
\textsc{Ent} & $-$ & $-$ & 82.4 & 65.4 & 61.5 & 3.93 \\ \midrule
\textsc{BestPair}\textsubscript{none} & 83.0 & 66.0 & 73.9 & 55.3 & 39.2 & 2.00\\
\textsc{BestPair}\textsubscript{attn} & 83.2 & 65.0 & 73.4 & 53.5 & 43.9 & 2.00\\
\textsc{BestPair}\textsubscript{entropy} & 81.8 & 64.2 & 78.5 & 58.2 & 52.3 & 2.00\\
\textsc{BestPair}\textsubscript{both} & 82.7 & 66.5 & 81.6 & 65.3 & 66.2 & 2.00\\ \midrule
\textsc{Sufficient}\textsubscript{none} & 83.0 & 66.0 & \multicolumn{2}{c}{\multirow{4}{1.5cm}{\centering Identical to Full Doc}} & 67.2 & 1.42\\
\textsc{Sufficient}\textsubscript{attn} & 83.2 & 65.0 & \multicolumn{2}{c}{} & \underline{70.3} & 1.45\\
\textsc{Sufficient}\textsubscript{entropy} & 81.8 & 64.2 & \multicolumn{2}{c}{} & \underline{69.9} & \textbf{1.65}\\
\textsc{Sufficient}\textsubscript{both} & 82.7 & 66.5 & \multicolumn{2}{c}{} & \underline{\textbf{73.1}} & \textbf{1.65}\\
\midrule
human & $-$ & $-$ & $-$ & $-$ & $-$ & 1.59 \\
\bottomrule
\end{tabular}
\caption{\label{evidencecompare} Model performance on adapted DocRED. Models are evaluated under two settings as in \textsc{BrainMRI}. Underlined results are better than \textsc{Sufficient}\textsubscript{none} on the corresponding metric according to a paired bootstrap test with $p=0.05$.}
\end{table}
These results show that learning-based systems make accurate predictions in this domain, and that their evidence extraction can be improved with better training, even in spite of the small size of the training set. In section~\ref{docred_result}, we focus on the adapted DocRED setting, which allows us to examine our model's performance in a higher-data regime.
\paragraph{Attribution scores are more peaked at the occurrence of key terms.} We conduct analysis on how the attribution scores from \textsc{Sufficient}\textsubscript{both} are peaked around the correct evidence compare to that from \textsc{Sufficient}\textsubscript{none} using our manually annotated set \textsc{BrainMRI}. We compute the mean of the instance-wise average and maximum of the normalized attribution mass falling into a few explicit tokens: \emph{enhancement} for contrast enhancement and \emph{effect} for mass effect, which are common explicit indicators in the context of specified key features. The results in Table \ref{tab:attribution} show attribution scores being peaked around the correct terms, highlighting that these models can be guided to not only make correct predictions but attend to the right information.
\begin{table}[]
\small
\centering
\begin{tabular}{lcccc}
\toprule
\multirow{2}{*}{\textbf{Model}}
& \multicolumn{2}{c}{\textbf{Mass Effect}}
& \multicolumn{2}{c}{\textbf{Ctr. Enhance.}} \\
& \textbf{Mean} & \textbf{Max} & \textbf{Mean} & \textbf{Max} \\ \midrule
\textsc{Sufficient}\textsubscript{none} & 7.3 & 7.4 & 28.6 & 29.8 \\
\textsc{Sufficient}\textsubscript{both} & \textbf{18.9} & \textbf{19.2} & \textbf{37.9} & \textbf{42.0} \\
\bottomrule
\end{tabular}
\caption{Distributions of attribution mass over explicit cues (``enhancement'' for \emph{contrast enhancement} and ``effect'' for \emph{mass effect}) for our best model and the baseline. Mean/Max is the mean of instance-wise average/maximum of the normalized attribution mass falling on the given token.\vspace{-0.2in}}
\label{tab:attribution}
\end{table}
Table \ref{tab:reportexample} in the Appendix shows visualizations of attribution scores for an example in \textsc{BrainMRI} using DeepLIFT. Even though baseline models make correct predictions, their attribution mass is diffused over the document. With the help of regularization, our model is capable of capturing implicit cues such as \emph{downward displacement of the brain stem}, although it is trained on an extremely small training set with only explicit cues like \emph{mass effect} in a weak sentence-level supervision framework.
\subsection{Results on Adapted DocRED}\label{docred_result}
\paragraph{Comparison to baselines}
Table~\ref{evidencecompare} shows that the \textsc{Ent} baseline is quite strong at DocRED evidence extraction. However, our best method still exceeds this method on both label accuracy as well as evidence extraction while extracting more succinct explanations. We see that the ability to extract a variable-length explanation is key, with \textsc{First2}, \textsc{First3} and \textsc{BestPair} performing poorly. Notably, these methods exhibit a drop in accuracy in the reduced doc setting for each method compared to the full doc setting, showing that the explanations extracted are not faithful.
\paragraph{Learning-based models with appropriate regularization perform relatively better in this larger-data setting} From Table \ref{tab:brainmri} and Table \ref{evidencecompare}, we can observe that various regularization techniques applied to \textsc{Sufficient} models maintain or improve overall model performance on both key feature and relation classification. We see that our \textsc{Sufficient} methods do not compromise on accuracy but make predictions based on plausible evidence sets, which is more evident when we have richer training data. We perform further error analysis in Appendix~\ref{sec:error_analysis}.
\paragraph{Faithfulness of techniques} One may be concerned that, like attention values \cite{Jain2019}, our feature attribution methods may not faithfully reflect the computation of the model. We emphasize again that the \textsc{Sufficient} paradigm on top of the DeepLIFT method \emph{is} faithful by our definition. For a model $f$, we measure the faithfulness by checking the agreement between $\hat{y} = f(D,t)$ and $y' = f(\hat{E}_t,t)$, where $\hat{E}_t$ is the extracted evidence we feed into the same model under the reduced document setting. This is shown for all methods in the ``Reduced doc'' columns in Tables~\ref{tab:brainmri} and \ref{evidencecompare}. We see a drop in performance from techniques such as \textsc{BestPair}: the full model does not make the same judgment on these evidence subsets, but by definition it does in the \textsc{Sufficient} setting.
As further evidence of faithfulness, we note that only a relatively small number of evidence sentences, in line with human annotations, are extracted in the \textsc{Sufficient} method.
These small subsets are indicated by feature attribution methods \emph{and} sufficient to reproduce the original model predictions with high confidence, suggesting that these explanations are faithful.
\section{Conclusion}
In this work, we develop techniques to employ small amount of \label{sentence-annotated} data to improve reliability of document-level IE systems in two domains. We systematically evaluate our model from perspectives of faithfulness and plausibility and show that we can substantially improve models' capability in focusing on supporting evidence while maintaining their predictive performance, leading to models that are ``right for the right reasons.''
\section*{Acknowledgments}
This work was partially supported by NSF Grant IIS-1814522 and a Texas Health Catalyst grant. Thanks to Scott Rudkin, Gregory Mittl, Raghav Mattay, and Chuan Liang for assistance with the annotation.
|
\section{Introduction}
Mixup \citep{zhang2018Mixup} is a modification to the standard supervised learning setup which involves
training on convex combinations of pairs of data points and their labels instead of the original data itself.\blfootnote{Correspondence to Muthu Chidambaram (\texttt{<EMAIL>}).}
In the original paper, \citet{zhang2018Mixup} demonstrated that training deep neural networks using Mixup leads to better
generalization performance, as well as greater robustness to adversarial attacks and label noise on image classification tasks.
The empirical advantages of Mixup training have been affirmed by several follow-up works
\citep{He_2019_CVPR, thulasidasan2019Mixup, lamb2019interpolated, arazo2019unsupervised, guo20text}. The idea of Mixup has also been extended beyond the supervised learning setting, and been applied to semi-supervised learning \citep{mixmatch, fixmatch}, contrastive learning \citep{verma2021domainagnostic, imix}, privacy-preserving learning \citep{huang2021instahide}, and learning with fairness constraints \citep{fairmixup}.
However, from a theoretical perspective, Mixup training is still mysterious even in the basic multi-class classfication setting \--- why should the output of a linear mixture of two training samples be the same linear mixture of their labels, especially when considering highly nonlinear models? Despite several recent theoretical results \citep{guo2019mixup, carratino2020mixup, zhang2020does, zhang2021mixup}, there is still not a complete understanding of why Mixup training actually works in practice. In this paper, we try to understand {\em why} Mixup works by first understanding {\em when} Mixup works: in particular, how the properties of Mixup training rely on the structure of the training data.
We consider two properties for classifiers trained with Mixup. First, even though Mixup training does not observe many original data points during training, it usually can still correctly classify all of the original data points (empirical risk minimization (ERM)). Second, the aforementioned empirical works have shown how classifiers trained with Mixup often have better adversarial robustness and generalization than standard training. In this work, we show that both of these properties can rely heavily on the data used for training, and that they need not hold in general.
\textbf{Main Contributions and Related Work.} The idea that Mixup can potentially fail to minimize the original risk is not new; \citet{guo2019mixup} provide examples of how Mixup labels can conflict with actual data point labels. However, their theoretical results do not characterize the data and model conditions under which this failure can provably happen when minimizing the Mixup loss. In Section \ref{mixuperm} of this work, we provide a concrete classification dataset on which continuous approximate-minimizers of the Mixup loss can fail to minimize the empirical risk. We also provide sufficient conditions for Mixup to minimize the original risk, and show that these conditions hold approximately on standard image classification benchmarks.
With regards to generalization and robustness, the parallel works of \citet{carratino2020mixup} and \citet{zhang2020does} showed that Mixup training can be viewed as minimizing the empirical loss along with a data-dependent regularization term. \citet{zhang2020does} further relate this term to the adversarial robustness and Rademacher complexity of certain function classes learned with Mixup. In Section \ref{mixupgen}, we take an alternative approach to understanding generalization and robustness by analyzing the margin of Mixup classifiers. Our perspective can be viewed as complementary to that of the aforementioned works, as we directly consider the properties exhibited by a Mixup-optimal classifier instead of considering what properties are encouraged by the regularization effects of the Mixup loss. In addition to our margin analysis, we also show that for the common setting of linear models trained on high-dimensional Gaussian features both Mixup (for a large class of mixing distributions) and ERM with gradient descent learn the same classifier with high probability.
Finally, we note the related works that are beyond the scope of our paper; namely the many Mixup-like training procedures such as Manifold Mixup \citep{verma2019manifold}, Cut Mix \citep{yun2019cutmix}, Puzzle Mix \citep{kim2020puzzle}, and Co-Mixup \citep{kim2021comixup}.
\section{Mixup and Empirical Risk Minimization}\label{mixuperm}
The goal of this section is to understand when Mixup training can also minimize the empirical risk. Our main technique for doing so is to derive a closed-form for the Mixup-optimal classifier over a sufficiently powerful function class, which we do in Section \ref{subsec:closedform} after introducing the basic setup in Section \ref{subsec:ermsetup}. We use this closed form to motivate a concrete example on which Mixup training does not minimize the empirical risk in Section \ref{subsec:failcase}, and show under mild nondegeneracy conditions that Mixup will minimize the emprical risk in Section \ref{subsec:gooderm}.
\subsection{Setup} \label{subsec:ermsetup}
We consider the problem of $k$-class classification where the classes $1, ..., k$ correspond to compact disjoint sets
$X_1, ..., X_k \subset \mathbb{R}^n$ with an associated probability measure $\mathbb{P}_X$ supported on $X = \bigcup_{i = 1}^k X_i$.
We use $\mathcal{C}$ to denote the set of all functions $g: \mathbb{R}^n \to [0, 1]^k$
satisfying the property that $\sum_{i = 1}^k g^i(x) = 1$ for all $x$ (where $g^i$ represents the $i$-th coordinate function of $g$).
We refer to a function $g \in \mathcal{C}$ as a \textit{classifier}, and say that $g$ classifies $x$ as class $j$ if $j = \argmax_i g^i (x)$.
The cross-entropy loss associated with such a classifier $g$ is then:
\begin{align*}
J(g, \mathbb{P}_X) = -\sum_{i = 1}^{k} \int_{X_i} \log g^i(x) d\mathbb{P}_X(x)
\end{align*}
The goal of standard training is to learn a classifier $h \in \argmin_{g \in \mathcal{C}} J(g, \mathbb{P}_X)$.
Any such classifier $h$ will necessarily satisfy $h^i(x) = 1$ on $X_i$ since the $X_i$ are disjoint.
\textbf{Mixup.} In the Mixup version of our setup, we are interested in minimizing the cross-entropy of convex combinations of the original data
and their classes. These convex combinations are determined according to a probability measure $\mathbb{P}_f$
whose support is $[0, 1]$, and we assume this measure has a density $f$. For two points $s, t \in X$, we let $z_{st}(\lambda) = \lambda s + (1 - \lambda) t$
(and use $z_{st}$ when $\lambda$ is understood) and define the Mixup cross-entropy on $s, t$ with respect to a classifier $g$ as:
\begin{align*}
\ell_{mix}(g, s, t, \lambda) = \begin{cases}
-\log g^i (z_{st}) & s, t \in X_i \\
-\left(\lambda \log g^i (z_{st}) + (1 - \lambda) \log g^j (z_{st})\right) & s \in X_i, t \in X_j
\end{cases}
\end{align*}
Having defined $\ell_{mix}$ as above, we may write the component of the full Mixup cross-entropy loss corresponding to mixing points from classes $i$ and $j$ as:
\begin{align*}
J_{mix}^{i, j} (g, \mathbb{P}_X, \mathbb{P}_f) = \int_{X_i \times X_j \times [0, 1]} \ell_{mix}(g, s, t, \lambda) \ d(\mathbb{P}_X \times \mathbb{P}_X \times \mathbb{P}_f)(s, t, \lambda)
\end{align*}
The final Mixup cross-entropy loss is then the sum of $J_{mix}^{i, j}$ over all $i, j \in \left\{1, ..., k\right\}$ (corresponding to all possible mixings between
classes, including themselves):
\begin{align*}
J_{mix}(g, \mathbb{P}_X, \mathbb{P}_f) = \sum_{i = 1}^k \sum_{j = 1}^k J_{mix}^{i, j}(g, \mathbb{P}_X, \mathbb{P}_f)
\end{align*}
\textbf{Relation to Prior Work.} We have opted for a more general definition of the Mixup loss (at least when constrained to multi-class classification)
than prior works.
This is not generality for generality's sake, but rather because many of our results apply to any mixing distribution supported on $[0, 1]$.
One obtains the original Mixup formulation of \citet{zhang2018Mixup} for multi-class classification on a finite dataset by taking the $X_i$
to be finite sets, and choosing $\mathbb{P}_X$ to be the normalized counting measure (corresponding to a discrete uniform distribution).
Additionally, $\mathbb{P}_f$ is chosen to have density $\mathrm{Beta}(\alpha, \alpha)$, where $\alpha$ is a hyperparameter.
\subsection{Mixup-optimal Classifier} \label{subsec:closedform}
Given our setup, we now wish to characterize the behavior of a Mixup-optimal classifier at a point $x \in \mathbb{R}^n$. However, if the optimization of $J_{mix}$ is considered over the class of functions $\mathcal{C}$, this is intractable (to the best of our knowledge) due to the lack of regularity conditions imposed on functions in $\mathcal{C}$. We thus wish to constrain the optimization of $J_{mix}$ to a class of functions that is sufficiently powerful (so as to include almost all practical settings) while still allowing for local analysis. To do so, we will need the following definitions, which will also be referenced throughout the results in this section and the next:
\begin{align*}
A_{x, \epsilon}^{i, j} &= \left\{(s, t, \lambda) \in X_i \times X_j \times [0, 1] : \ \lambda s + (1 - \lambda) t \in B_{\epsilon}(x)\right\} \\
A_{x, \epsilon, \delta}^{i, j} &= \left\{(s, t, \lambda) \in X_i \times X_j \times [0, 1 - \delta] : \ \lambda s + (1 - \lambda) t \in B_{\epsilon}(x)\right\} \\
X_{mix} &= \left\{x \in \mathbb{R}^n : \ \bigcup_{i, j} A_{x, \epsilon}^{i, j} \text{ has positive measure for every } \epsilon > 0\right\} \\
\xi_{x, \epsilon}^{i, j} &= \int_{A_{x, \epsilon}^{i, j}} \ d(\mathbb{P}_X \times \mathbb{P}_X \times \mathbb{P}_f)(s, t, \lambda) \\
\xi_{x, \epsilon, \lambda}^{i, j} &= \int_{A_{x, \epsilon}^{i, j}} \lambda \ d(\mathbb{P}_X \times \mathbb{P}_X \times \mathbb{P}_f)(s, t, \lambda)
\end{align*}
The set $A_{x, \epsilon}^{i, j}$ represents all points in $X_i \times X_j$ that have lines between them intersecting an $\epsilon$-neighborhood of $x$,
while the set $A_{x, \epsilon, \delta}^{i, j}$ represents the restriction of $A_{x, \epsilon}^{i, j}$ to only those points whose connecting line segments intersect an $\epsilon$-neighborhood of $x$ with $\lambda$ values bounded by $1 - \delta$ (used in Section \ref{mixupgen}).
The set $X_{mix}$ corresponds to all points for which every neighborhood factors into $J_{mix}$.
The $\xi_{x, \epsilon}^{i, j}$ term represents the measure of the set $A_{x, \epsilon}^{i, j}$ while $\xi_{x, \epsilon, \lambda}^{i, j}$
represents the expectation of $\lambda$ over the same set. To provide better intuition for these definitions, we provide visualizations in Section B of the appendix.
We can now define the subset of $\mathcal{C}$ to which we will constrain our optimization of $J_{mix}$.
\begin{restatable}{definition}{fclass}\label{def0}
Let $\mathcal{C}^*$ to be the subset of $\mathcal{C}$ for which every $h \in \mathcal{C}^*$ satisfies $h(x) = \lim_{\epsilon \to 0} \argmin_{\theta\in [0, 1]^k} J_{mix}(\theta)\vert_{B_{\epsilon}(x)}$ for all $x \in X_{mix}$ when the limit exists. Here $J_{mix}(\theta)\vert_{B_{\epsilon}(x)}$ represents the Mixup loss for a constant function with value $\theta$ with the restriction of each term in $J_{mix}$ to the set $A_{x, \epsilon}^{i, j}$.
\end{restatable}
We immediately justify this definition with the following proposition.
\begin{restatable}{proposition}{goodclass}\label{prop0}
Any function $h \in \argmin_{g \in \mathcal{C}^*} J_{mix}(g, \mathbb{P}_X, \mathbb{P}_f)$ satisfies $J_{mix}(h) \leq J_{mix}(g)$ for any continuous $g \in \mathcal{C}$.
\end{restatable}
\textbf{Proof Sketch.} We can argue directly from definitions by considering points in $X_{mix}$ for which $h$ and $g$ differ.
Proposition \ref{prop0} demonstrates that optimizing over $\mathcal{C}^*$ is at least as good as optimizing over the subset of $\mathcal{C}$ consisting of continuous functions, so we cover most cases of practical interest (i.e. optimizing deep neural networks). As such, the term ``Mixup-optimal'' is intended to mean optimal with respect to $\mathcal{C}^*$ throughout the rest of the paper. We may now characterize the classification of a Mixup-optimal classifier on $X_{mix}$.
\begin{restatable}{lemma}{closedform}\label{lem}
For any point $x \in X_{mix}$ and $\epsilon > 0$, there exists a continuous function $h_{\epsilon}$ satisfying:
\begin{align}
h_{\epsilon}^i(x) = \frac{\xi_{x, \epsilon}^{i, i} + \sum_{j \neq i} \big(\xi_{x, \epsilon, \lambda}^{i, j} + (\xi_{x, \epsilon}^{j, i} - \xi_{x, \epsilon, \lambda}^{j, i})\big)}{\sum_{q = 1}^k \bigg(\xi_{x, \epsilon}^{q, q} + \sum_{j \neq q} \big(\xi_{x, \epsilon, \lambda}^{q, j} + (\xi_{x, \epsilon}^{j, q} - \xi_{x, \epsilon, \lambda}^{j, q})\big)\bigg)} \label{eq_1}
\end{align}
With the property that $\lim_{\epsilon \to 0} h_{\epsilon}(x) = h(x)$ for every $h \in \argmin_{g \in \mathcal{C}^*} J_{mix}(g, \mathbb{P}_X, \mathbb{P}_f)$
when the limit exists.
\end{restatable}
\textbf{Proof Sketch.} We set $h_{\epsilon} = \argmin_{\theta\in [0, 1]^k} J_{mix}(\theta)\vert_{B_{\epsilon}(x)}$ and show that this is well-defined, continuous, and has the above form using the strict convexity of the minimization problem.
\begin{remark}\label{rem}
For the important case of finite datasets, it will be shown that the limit above always exists as part of the proof of Theorem \ref{theo4}.
\end{remark}
The expression for $h_{\epsilon}^i$ just represents the expected location of the point $x$ on all lines between class $i$ and other classes, normalized by the sum of the expected locations for all classes. It can be simplified significantly if $\mathbb{P}_f$ is assumed to be symmetric; we give this as a corollary after the proof in Section C of the Appendix. Importantly, we note that while $h_{\epsilon}$ as defined in Lemma \ref{lem} is continuous for every $\epsilon > 0$, its pointwise limit $h$ need not be, which we demonstrate below.
\begin{restatable}{proposition}{discontexample}\label{prop}
Let $X_1 = \left\{(0, 1), (0, -1)\right\}$ and let $X_2 = \left\{(1, 0), (-1, 0)\right\}$, with $\mathbb{P}_X$ being discrete
uniform over $X_1 \cup X_2$ and $\mathbb{P}_{f}$ being continuous uniform over $[0, 1]$.
Then the Mixup-optimal classifier $h$ is discontinuous at $(0, 0)$.
\end{restatable}
\textbf{Proof Sketch.} One may explicitly compute for $x = (0, 0)$ that $h^1(x) = h^2(x) = \frac{1}{2}$.
Proposition \ref{prop} illustrates our first significant difference between Mixup training and standard training: there \textit{always} exists a minimizer of the empirical cross-entropy $J$ that can be extended to a continuous function (since a minimizer is constant on the class supports and not constrained elsewhere), whereas depending on the data the minimizer of $J_{mix}$ can be discontinuous.
\subsection{A Mixup Failure Case}\label{subsec:failcase}
With that in mind, several model classes popular in practical applications consist of continuous functions. For example, neural networks with ReLU activations
are continuous, and several works have noted that they are Lipschitz continuous with shallow networks having approximately small Lipschitz constant \citep{scaman2019lipschitz, fazlyab2019efficient, latorre2020lipschitz}.
Given the regularity of such models, we are motivated to consider the continuous approximations $h_{\epsilon}$ in Lemma \ref{lem} and see if it is possible to construct a dataset on which $h_{\epsilon}$ (for a fixed $\epsilon$) can fail to classify the original points correctly.
We thus consider the following dataset:
\begin{restatable}{definition}{ncal}[3-Point Alternating Line]\label{ncal}
We define $\mathcal{X}_3^2$ to be the binary classification dataset consisting of the points $\left\{0, 1, 2\right\}$ classified as $\left\{1, 2, 1\right\}$.
In our setup, this corresponds to $X_1 = \left\{0, 2\right\}$ and $X_2 = \left\{1\right\}$ with $\mathbb{P}_X = \frac{1}{3} 1_{\left\{0, 1, 2\right\}}$.
\end{restatable}
Intuitively, the reason why Mixup can fail on $\mathcal{X}_3^2$ is that, for choices of $\mathbb{P}_f$ that concentrate about $\frac{1}{2}$, we will have
by Lemma \ref{lem} that the Mixup-optimal classification in a neighborhood of point 1 should skew towards class 1 instead of class 2 due to the sandwiching
of point 1 between points 0 and 2. The canonical choice of $\mathbb{P}_f$ corresponding to a mixing density of $\mathrm{Beta}(\alpha, \alpha)$ is one such choice:
\begin{restatable}{theorem}{ncaltheorem}\label{theo}
Let $\mathbb{P}_f$ have associated density $\text{Beta}(\alpha, \alpha)$.
Then for any classifier $h_{\epsilon}$ on $\mathcal{X}_3^2$ (as defined in Lemma \ref{lem}),
we may choose $\alpha$ such that $h_{\epsilon}$ does not achieve 0 classification error on $\mathcal{X}_3^2$.
\end{restatable}
\textbf{Proof Sketch.} For any $\epsilon > 0$, we can bound the $\xi$ terms in Equation \ref{eq_1} using the fact that $\mathrm{Beta}(\alpha, \alpha)$ is strictly subgaussian \citep{marchal}, and then choose $\alpha$ appropriately.
\textbf{Experiments.} The result of Theorem \ref{theo} leads us to believe that the Mixup training of a continuous model should fail on $\mathcal{X}_3^2$ for appropriately chosen
$\alpha$. To verify that the theory predicts the experiments, we train a two-layer feedforward neural network with 512 hidden units and ReLU activations on
$\mathcal{X}_3^2$ with and without Mixup.
The implementation of Mixup training does not differ from the theoretical setup; we uniformly sample pairs of data points and train on their mixtures. Our implementation uses PyTorch \citep{pytorch} and is based heavily on the open source implementation of Manifold Mixup \citep{verma2019manifold} by Shivam Saboo.
Results for training using (full-batch) Adam \citep{adam} with the suggested (and common) hyperparameters of $\beta_1 = 0.9, \beta_2 = 0.999$
and a learning rate of $0.001$ are shown in Figure \ref{fig1}.
The class 1 probabilities for each point in the dataset outputted by the learned Mixup classifiers from Figure \ref{fig1} are shown in Table \ref{tab1} below:
\begin{figure*}[t]
\begin{subfigure}[t]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{NCAL_alpha_1.0_runs_10}
\caption{$\alpha = 1$}
\end{subfigure}%
\begin{subfigure}[t]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{NCAL_alpha_32.0_runs_10}
\caption{$\alpha = 32$}
\end{subfigure}%
\begin{subfigure}[t]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{NCAL_alpha_128.0_runs_10}
\caption{$\alpha = 128$}
\end{subfigure}
\caption{Training error for Mixup and regular training on $\mathcal{X}_3^2$. Each curve corresponds to the mean of 10 training runs, and
the area around each curve represents a region of one standard deviation.}
\vspace{-4mm}
\label{fig1}
\end{figure*}
\begin{table*}[ht!]
\centering
\begin{tabular}{|c|c|c|c|}
\hline
$h$ & 0 & 1 & 2 \\
\hline
$\alpha = 1$ & 0.995 & 0.156 & 0.979 \\
$\alpha = 32$ & 1.000 & 0.603 & 0.997 \\
$\alpha = 128$ & 1.000 & 0.650 & 0.997 \\
\hline
\end{tabular}
\caption{Mixup model evaluations on $\mathcal{X}_3^2$ for different choices of $\alpha$.}
\label{tab1}
\vspace{-2mm}
\end{table*}
We see from Figure \ref{fig1} and Table \ref{tab1} that Mixup training fails to correctly classify the points in $\mathcal{X}_3^2$ for $\alpha = 32$, and this misclassification becomes more exacerbated as we increase $\alpha$. The choice of $\alpha$ for which misclassifications begin to happen is largely superficial; we show in Section D of the Appendix that it is straightforward to construct datasets in the style of $\mathcal{X}_3^2$ for which Mixup training will fail even for the very mild choice of $\alpha = 1$. We focus on the case of $\mathcal{X}_3^2$ here to simplify the theory. The key takeaway is that, for datasets that exhibit (approximately) collinear structure amongst points, it is possible for inappropriately chosen mixing distributions to cause Mixup training to fail to minimize the original empirical risk.
\subsection{Sufficient Conditions for Minimizing the Original Risk}\label{subsec:gooderm}
The natural follow-up question to the results of the previous subsection is:
under what conditions on the data can this failure case be avoided? In other words, when can the Mixup-optimal classifier classify the original data points correctly while being continuous at those points?
Prior to answering that question, we first point out that if discontinuous functions are allowed,
then Mixup training always minimizes the original risk on finite datasets:
\begin{restatable}{proposition}{discontprop}\label{prop2}
Consider $k$-class classification where the supports $X_1, ..., X_k$ are finite and $\mathbb{P}_X$ corresponds to the discrete uniform distribution.
Then for every $h \in \argmin_{g \in \mathcal{C}^*} J_{mix}(g, \mathbb{P}_X, \mathbb{P}_f)$, we have that $h^i(x) = 1$ on $X_i$.
\end{restatable}
\textbf{Proof Sketch.} Only the $\xi_{x, \epsilon}^{i, i}$ term doesn't vanish in $h_{\epsilon}^i(x)$ as $\epsilon \to 0$, as the mixing distribution is continuous and cannot assign positive measure to $x$ alone when mixing two points that are not $x$.
Note that Proposition \ref{prop2} holds for \textbf{any} continuous mixing distribution $\mathbb{P}_f$ supported on $[0, 1]$ - we just need a rich enough model class.
In order to obtain the result of Proposition \ref{prop2} with the added restriction of continuity of $h$ on each of the $X_i$, we need to further assume that the collinearity of different class points that occurred in the previous section does not happen.
\begin{restatable}{assumption}{nocol}\label{as1}
For any point $x \in X_i$, there do not exist $u \in X$ and $v \in X_j$ for $j \neq i$ such that there is a $\lambda > 0$ for which $x = \lambda u + (1 - \lambda) v$.
\end{restatable}
A visualization of Assumption \ref{as1} is provided in Section B of the appendix. With this assumption in hand, we obtain the following result as a corollary of Theorem \ref{theo4} which is proved in the next section:
\begin{restatable}{theorem}{conttheorem}\label{theo2}
We consider the same setting as Proposition \ref{prop2} and further suppose that Assumption \ref{as1} is satisfied.
Then for every $h \in \argmin_{g \in \mathcal{C}^*} J_{mix}(g, \mathbb{P}_X, \mathbb{P}_f)$, we have that $h^i(x) = 1$ on $X_i$ and that $h$
is continuous on $X$.
\end{restatable}
\textbf{Application of Sufficient Conditions.}
\begin{figure*}[t]
\begin{subfigure}[t]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{MNIST_alpha_1024.0_runs_5}
\caption{MNIST}
\end{subfigure}%
\begin{subfigure}[t]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{CIFAR10_alpha_1024.0_runs_5}
\caption{CIFAR-10}
\end{subfigure}%
\begin{subfigure}[t]{0.33\textwidth}
\centering
\includegraphics[width=\textwidth]{CIFAR100_alpha_1024.0_runs_5}
\caption{CIFAR-100}
\end{subfigure}
\caption{Mean and single standard deviation of 5 training runs for Mixup ($\alpha = 1024$) and ERM on the \textit{original training data}. Mixup achieves near-identical (within 1\%) training accuracy to ERM.}
\label{fig2}
\vspace{-5mm}
\end{figure*}
The practical take-away of Theorem \ref{theo2} is that if a dataset does not exhibit approximate collinearity between points of different classes, then Mixup training should achieve near-identical training error on the \textit{original training data} when compared to ERM. We validate this by training ResNet-18 \citep{resnet} (using the popular implementation of Kuang Liu) on MNIST \citep{lecun1998mnist}, CIFAR-10, and CIFAR-100 \citep{cifar} with and without Mixup for 50 epochs with a batch size of 128 and otherwise identical settings to the previous subsection. For Mixup, we consider mixing using $\mathrm{Beta}(\alpha, \alpha)$ for $\alpha = 1, 32, 128, \text{ and } 1024$ to cover a wide band of mixing distributions. Our experimental results for the ``worst case'' of $\alpha = 1024$ (the other choices are strictly closer to ERM in training accuracy) are shown in Figure \ref{fig2}, while the other experiments can be found in Section \ref{additionalexps} of the Appendix.
We now check that the theory can predict the results of our experiments by verifying Assumption \ref{as1} approximately (exact verification is too expensive). We sample one epoch's worth of Mixup points (to simulate training) from a downsampled version of each train dataset, and then compute the minimum distances between each Mixup point and points from classes other than the two mixed classes. The minimum over these distances corresponds to an estimate of $\epsilon$ in Assumption \ref{as1}. We compute the distances for both training and test data, to see whether good training but poor test performance can be attributed to test data conflicting with mixed training points. Results are shown below in Table \ref{tab2}.
\begin{table*}[ht!]
\centering
\begin{tabular}{|c|c|c|c|}
\hline
Comparison Type & MNIST & CIFAR-10 & CIFAR-100 \\
\hline
Mixup/Train & 11.433 & 17.716 & 12.936 \\
\hline
Mixup/Test & 11.897 & 22.133 & 15.076 \\
\hline
\end{tabular}
\caption{Minimum Euclidean distance results using our approximation procedure with Mixup points generated using $\mathrm{Beta}(1024, 1024)$. We downsample all datasets to 20\%, to compare to the experiments of \citet{guo2019mixup}.}
\label{tab2}
\vspace{-3mm}
\end{table*}
To interpret the estimated $\epsilon$ values in Table \ref{tab2}, we note that unless $\epsilon \ll 1/L$ (where $L$ is the Lipschitz constant of the model being considered), a Mixup point cannot conflict with an original point (since the function has enough flexibility to fit both). Due to the estimated large Lipschitz constants of deep networks \citep{scaman2019lipschitz}, our $\epsilon$ values certainly do not fall in this regime, explaining how the near-identical performance to ERM is possible on the original datasets. We remark that our results challenge an implication of \citet{guo2019mixup}, which was that training/test performance on the above benchmarks degrades with high values of $\alpha$ due to conflicts between original points and Mixup points.
\subsection{The Rate of Empirical Risk Minimization Using Mixup}
Another striking aspect of the experiments in Figure \ref{fig2} is that Mixup training minimizes the original empirical risk at a very similar rate to that of direct empirical risk minimization.
A priori, there is no reason to expect that Mixup should be able to do this - a simple calculation shows that Mixup training only sees one true data point per epoch in expectation
(each pair of points is sampled with probability $\frac{1}{m^2}$ and there are $m$ true point pairs and $m$ pairs seen per epoch, where $m$ is the dataset size). The experimental results are even more surprising given that we are training using $\alpha = 1024$, which essentially corresponds to training using the midpoints of the original data points. This seems to imply that it is possible to recover the classifications of the original data points from the midpoints alone (not including the midpoint of a point and itself), and similar phenomena has indeed been observed in recent empirical work \citep{guo2021midpoint}. We make this rigorous with the following result:
\begin{restatable}{theorem}{recovery}\label{theo3}
Suppose $\left\{x_1, ..., x_m\right\}$ with $m \geq 6$ are sampled from $X$ according to $\mathbb{P}_X$, and that $\mathbb{P}_X$
has a density.
Then with probability 1, we can uniquely determine the points $\left\{x_1, ..., x_m\right\}$ given only the $\binom{m}{2}$ midpoints
$\left\{x_{i, j}\right\}_{1 \leq i < j \leq m}$.
\end{restatable}
\textbf{Proof Sketch.} The idea is to represent the problem as a linear system, and show using rank arguments that the sets of $m$ points that cannot be uniquely determined are a measure zero set.
Theorem \ref{theo3} shows, in an information-theoretic sense, that it is possible to obtain the original data points (and therefore also their labels) from only their
midpoints. While this gives more theoretical backing as to why it is possible for Mixup training using $\mathrm{Beta}(1024, 1024)$ to recover the original data point
classifications with very low error, it does not explain why this actually happens in practice at the rate that it does. A full theoretical analysis of this phenomenon
would necessarily require analyzing the training dynamics of neural networks (or another model of choice)
when trained only on midpoints of the original data, which is outside the intended scope of this work.
That being said, we hope that such analysis will be a fruitful line of investigation for future work.
\section{Generalization Properties of Mixup Classifiers}\label{mixupgen}
Having discussed how Mixup training differs from standard empirical risk minimization with regards to the original training data, we now consider how a learned Mixup
classifier can differ from one learned through empirical risk minimization on unseen test data. To do so, we analyze the \textit{per-class margin} of Mixup classifiers,
i.e. the distance one can move from a class support $X_i$ while still being classified as class $i$.
\subsection{The Margin of Mixup Classifiers}\label{subsec:twomoons}
Intuitively, if a point $x$ falls only on line segments between $X_i$ and some other classes $X_j, ...$, and if $x$ always falls closer to $X_i$ than
the other classes, we can expect $x$ to be classified according to class $i$ by the Mixup-optimal classifier due to Lemma \ref{lem}.
To make this rigorous, we introduce another assumption that generalizes Assumption \ref{as1} to points outside of the class supports:
\begin{restatable}{assumption}{nocoldelta}\label{as2}
For a class $i$ and a point $x \in X_{mix}$, suppose there exists an $\epsilon > 0$ and a $0 < \delta < \frac{1}{2}$ such that
$A_{x, \epsilon', \delta}^{i, j}$ and $A_{x, \epsilon'}^{j, q}$ have measure zero for all $\epsilon' \leq \epsilon$ and $j, q \neq i$, and the measure of $A_{x, \epsilon}^{i, j}$ is at least that of $A_{x, \epsilon}^{j, i}$.
\end{restatable}
Here the measure zero conditions are codifying the ideas that the point $x$ falls closer to $X_i$ than any other class on every line segment that intersects it, and there are no line segments between non-$i$ classes that intersect $x$. The condition that the measure of $A_{x, \epsilon}^{i, j}$ is at least that of $A_{x, \epsilon}^{j, i}$ handles asymmetric mixing distributions that concentrate on pathological values of $\lambda$.
A visualization of Assumption \ref{as2} is provided in Section B of the Appendix.
Now we can prove:
\begin{restatable}{theorem}{margin}\label{theo4}
Consider $k$-class classification where the supports $X_1, ..., X_k$ are finite and $\mathbb{P}_X$ corresponds to the discrete uniform distribution.
If a point $x$ satisfies Assumption \ref{as2} with respect to a class $i$,
then for every $h \in \argmin_{g \in \mathcal{C}^*} J_{mix}(g, \mathbb{P}_X, \mathbb{P}_f)$, we have that $h$ classifies $x$ as class $i$ and that
$h$ is continuous at $x$.
\end{restatable}
\textbf{Proof Sketch.} The limit in Lemma \ref{lem} can be shown to exist using the Lebesgue differentiation theorem, and we can bound the limit below since the $A_{x, \epsilon', \delta}^{i, j}$ have measure zero.
Assumption \ref{as1} implies Assumption \ref{as2} with respect to each class, and hence we get Theorem \ref{theo2} as a corollary of Theorem \ref{theo4}
as mentioned in Section 2. To use Theorem \ref{theo4} to understand generalization, we make the observation that a point $x$ can satisfy Assumption \ref{as2} while being a distance of up to $\frac{\min_j d(X_i, X_j)}{2}$ from some class $i$. This distance can be significantly farther than, for example, the optimal linear separator in a linearly separable dataset.
\textbf{Experiments.} To illustrate that Mixup can lead to more separation between points than a linear decision boundary, we consider the two moons dataset \citep{twomoons}, which consists of two classes of points supported on semicircles with added Gaussian noise. Our motivation for doing so comes from the work of \citet{gradstarv}, in which it was noted that neural network models trained on a separated version of the two moons dataset essentially learned a linear separator while ignoring the curvature of the class supports. While \citet{gradstarv} introduced an explicit regularizer to encourage a nonlinear decision boundary, we expect due to Theorem \ref{theo4} that Mixup training will achieve a similar result without any additional modifications.
To verify this empirically, we train a two-layer neural network with 500 hidden units with and without Mixup, to have a 1-to-1 comparison with the setting of \citet{gradstarv}. We use $\alpha = 1$ and $\alpha = 1024$ for Mixup to capture a wide band of mixing densities. The version of the two moons dataset we use is also identical to that of the one used in the experiments of \citet{gradstarv}, and we are grateful to the authors for releasing their code under the MIT license. We do full-batch training with all other training, implementation, and compute details remaining the same as the previous section. Results are shown in Figure \ref{fig3}.
\begin{figure*}[t]
\centering
\includegraphics[scale=0.35]{mixup_moons_delta_0.5.png}
\caption{Decision boundary plots for standard and Mixup training on the two moons dataset of \citet{gradstarv} with a class separation of 0.5. Each boundary represents the average of 10 training runs of 1500 epochs.}
\label{fig3}
\vspace{-4mm}
\end{figure*}
Our results affirm the observations of \cite{gradstarv} and previous work \citep{combes} that neural network training dynamics may ignore salient features of the dataset; in this case the ``Base Model'' learns to differentiate the two classes essentially based on the $x$-coordinate alone. On the other hand, the models trained using Mixup have highly nonlinear decision boundaries. Further experiments for different class separations and values of $\alpha$ are included in Section F of the Appendix.
\subsection{When Mixup Training Learns the Same Classifier}
The experiments and theory of the previous sections have shown how a Mixup classifier can differ significantly from one learned through standard training. In this subsection, we now consider the opposing question - when is the Mixup classifier the same as the one learned through standard training? Prior work \citep{archambault2019} has considered when Mixup training coincides with certain adversarial training, and our results complement this line of work. The motivation for our results comes from the fact that a practitioner need not spend compute on Mixup training in addition to standard training in settings where the performance will be provably the same.
We consider the case of binary classification using a linear model $\theta^{\top} x$ on high-dimensional Gaussian data, which is a setting that arises naturally when training using Gaussian kernels. Specifically, we consider the dataset $X$ to consist of $n$ points in $\mathbb{R}^d$ distributed according to $\mathcal{N}(0, I_d)$ with $d > n$ (to be made more precise shortly). We also consider the mixing distribution to be any symmetric distribution supported on $[0, 1]$ (thereby including as a special case $\mathrm{Beta}(\alpha, \alpha)$). We let the labels of points in $X$ be $\pm 1$ (so that the sign of $\theta^{\top} x$ is the classification), and use $X_1$ and $X_{-1}$ to denote the individual class points. We will show that in this setting, the optimal Mixup classifier is the same (up to rescaling of $\theta$) as the ERM classifier learned using gradient descent with high probability. To do so we need some additional definitions.
\begin{restatable}{definition}{interpolating}\label{intersoln}
We say $\hat{\theta}$ is an interpolating solution, if there exists $k > 0$ such that
\begin{align*}
\hat{\theta}^{\top}x_i = - \hat{\theta}^{\top}z_j = k \ \ \ \forall x_i \in X_1,\ \forall z_j \in X_{-1}.
\end{align*}
\end{restatable}
\begin{restatable}{definition}{maxmargin}
The maximum margin solution $\tilde{\theta}$ is defined through:
\begin{align*}
\tilde{\theta} := \argmax_{\|\theta\|_2 = 1}\left\{\min_{x_i \in X_1, z_j \in X_{-1}}\left\{\theta^{\top}x_i, -\theta^{\top}z_j\right\}\right\}
\end{align*}
\end{restatable}
When the maximum margin solution coincides with an interpolating solution for the dataset $X$ (i.e. all the points are support vectors), we have that Mixup training leads to learning the max margin solution (up to rescaling).
\begin{restatable}{theorem}{maxmargintheorem} \label{max-margin}
If the maximum margin solution for $X$ is also an interpolating solution for $X$, then any $\theta$ that lies in the span of $X$ and minimizes the Mixup loss $J_{mix}$ for a symmetric mixing distribution $\mathbb{P}_f$ is a rescaling of the maximum margin solution.
\end{restatable}
\textbf{Proof Sketch.} It can be shown that $\theta$ is an interpolating solution using a combination of the strict convexity of $J_{mix}$ as a function of $\theta$ and the symmetry of the mixing distribution.
\begin{remark}
For every $\theta$, we can decompose it as $\theta = \theta_{X} + \theta_{X^{\perp}}$ where $\theta_{X}$ is the projection of $\theta$ onto the subspace spanned by $X$. By definition we have that $\theta_{X^{\perp}}$ is orthogonal to all possible mixings of points in $X$. Hence, $\theta_{X^{\perp}}$ does not affect the Mixup loss or the interpolating property, so for simplicity we may just assume $\theta$ lies in the span of $X$.
\end{remark}
To characterize the conditions on $X$ under which the maximum margin solution interpolates the data, we use a key result of \citet{muthukumar2020classification}, restated below. Note that \citet{muthukumar2020classification} actually provide more settings in their paper, but we constrain ourselves to the one stated below for simplicity.
\begin{restatable}{lemma}{muthukumar}[Theorem 1 in \citet{muthukumar2020classification}, Rephrased] \label{equivalence}
Assuming $d > 10n \ln n+n-1,$ then with probability at least $1-2/n,$ the maximum margin solution for $X$ is also an interpolating solution.
\end{restatable}
To tie the optimal Mixup classifier back to the classifier learned through standard training, we appeal to the fact that minimizing the empirical cross-entropy of a linear model using gradient descent leads to learning the maximum margin solution on linearly separable data \citep{soudry2018implicit, matus}. From this we obtain the desired result of this subsection:
\begin{corollary}
Under the same conditions as Lemma \ref{equivalence}, the optimal Mixup classifier has the same direction as the classifier learned through minimizing the empirical cross-entropy using gradient descent with high probability.
\end{corollary}
\section{Conclusion}
The main contribution of our work has been to provide a theoretical framework for analyzing how Mixup training can differ from empirical risk minimization. Our results characterize a practical failure case of Mixup, and also identify conditions under which Mixup can provably minimize the original risk.
They also show in the sense of margin why the generalization of Mixup classifiers can be superior to those learned through empirical risk minimization, while again identifying model classes and datasets for which the generalization of a Mixup classifier is no different (with high probability). We also emphasize that the generality of our theoretical framework allows most of our results to hold for any continuous mixing distribution. Our hope is that the tools developed in this work will see applications in future works concerned with analyzing the relationship between benefits obtained from Mixup training and properties of the training data.
\section{Ethics Statement}
We do not anticipate any direct misuses of this work due to its theoretical nature. That being said, the failure case of Mixup discussed in Section \ref{mixuperm} could serve as a way for an adversary to potentially exploit a model trained using Mixup to classify data incorrectly. However, as this requires knowledge of the mixing distribution and other hyperparameters of the model, we do not flag this as a significant concern - we would just like to point it out for completeness.
\section{Reproducibility Statement}
Full proofs for all results in the main body of the paper can be found in Sections C and E of the Appendix.
All of the code used to generate the plots and experimental results in this paper can be found at: \url{https://github.com/2014mchidamb/Mixup-Data-Dependency}. We have tried our best to organize the code to be easy to use and extend. Detailed instructions for how to run each type of experiment are provided in the \texttt{README} file included in the GitHub repository.
\section*{Acknowledgements}
Rong Ge, Muthu Chidambaram, Xiang Wang, and Chenwei Wu are supported in part by NSF Award DMS-2031849, CCF-1704656, CCF-1845171 (CAREER), CCF-1934964 (Tripods), a Sloan Research Fellowship, and a Google Faculty Research Award. Muthu would like to thank Michael Lin for helpful discussions during the early stages of this project.
|
\section{Introduction}
Supervised machine learning algorithms rely on datasets that contain an abundance of labels (i.e. known classifications) for associated inputs. In many
real-world applications however, unlabeled data is ubiquitous, while obtaining labels for such training data is costly.
Semi-supervised learning (SSL) methods leverage unlabeled data to achieve an accurate classification with significantly fewer training points. At the same time, the choice of training points often affects classifier performance, especially in the case of SSL due to the small training set size.
Active learning seeks to choose a ``useful'' training set from which the underlying machine learning algorithm learns, and careful selection of such training data is motivated by the inherent cost to label data in practice.
The main challenge in active learning is designing the criterion for selecting which unlabeled points are the most beneficial to label to significantly improve the underlying machine learning classifier's performance, more so than merely selecting random points to label.
While there are a few different formulations of active learning, we focus on the {\it pool-based} active learning paradigm~\cite{lewis_sequential_1994} as opposed to online or streaming-based active learning~\cite{settles_active_2012}. That is, the active learner has access to a fixed ``pool'' of unlabeled data points from which it can decide a subset to label.
In pool-based active learning, most methods alternate between: (1) training a model given the current labeled data $\mathcal L, \{y_j\}_{j \in \mathcal L}$ and (2) choosing a set of active learning query points $\mathcal Q$ in the unlabeled set $\mathcal U$ according to an {\it acquisition function} (also called an {\it active learning criterion}), see \Cref{fig:flowchart}. An ``oracle'' that has access to the classification of all data then labels points $k \in \mathcal Q$; oftentimes in application, a domain expert plays the role of the oracle in identifying the true classification of points (i.e. ``human-in-the-loop'' schemes). We refer to this iterative procedure of alternating between model training and query set selection and labeling as the {\it active learning process}. The goal then is to design an acquisition function $\mathcal A : \mathcal U \rightarrow \mathbb R$ that quantifies how useful labeling a point $k \in \mathcal U$ would be in the active learning process.
With a specified acquisition function $\mathcal A$, we select the query set $\mathcal Q \subset \mathcal U$ in the active learning process either {\it sequentially} (i.e. one at a time, $|\mathcal Q| = 1$) or in a {\it batch} (i.e. $|\mathcal Q| = B \in \mathbb N_+$).
\begin{figure}
\centering
\begin{tikzpicture}[node distance=2cm]
\node (back) [bigblock, opacity=0.96, xshift=7.6cm]{(2)};
\node (b1) [block] {Train SSL Classifier with $\mathcal L$ and labels $\{y_j\}_{j \in \mathcal L}$};
\node (b2) [block, right of=b1, xshift=1.8cm, fill=green!30] {Select query points $\mathcal Q \subset \mathcal U$ via {\it acquisition function}, $\mathcal A$ };
\node (b3) [block, right of=b2, xshift=1.8cm, fill=blue!30] {Query oracle for labels $\{\hat{y}_k\}_{k \in \mathcal Q}$};
\draw node[above of=b3, yshift=-0.6cm] {Active Learning Query and Update};
\node (b4) [block, right of=b3, xshift=1.8cm, fill=yellow!30] {Update $\mathcal L = \mathcal L \cup \mathcal Q$};
\draw [arrow] (b1) -- (b2);
\draw [arrow] (b2) -- (b3);
\draw [arrow] (b3) -- (b4);
\draw [arrow] (b4) -- ++ (0,-1.4cm) node[below]{} -- ++ (-11.39cm, 0) node[left]{} -- (b1) ;
\end{tikzpicture}
\caption{Active Learning Flowchart. Alternate between (1-red) training the underlying SSL classifier with the current labeled set $\mathcal L$ with labels $\{y_j\}_{j \in \mathcal L}$ and (2-grey) selecting query points $\mathcal Q$ from the unlabeled set ($\mathcal U$) that are subsequently labeled according to the oracle and added to the labeled data $\mathcal L$. The process repeats with the updated labeled data, retraining the SSL classifier to prepare for another active learning query and update.}
\label{fig:flowchart}
\end{figure}
Most active learning acquisition functions belong to one of a few categories: uncertainty~\cite{settles_active_2012, houlsby_bayesian_2011, gal_deep_2017}, margin~\cite{tong_support_2001, balcan_agnostic_2006, jiang_bootstrapping_2021}, clustering~\cite{dasgupta_hierarchical_2008, maggioni_learning_2019}, and look-ahead~\cite{zhu_combining_2003, cai_maximizing_2013}. {\it Uncertainty}-based acquisition functions favor unlabeled points whose classification
is ``uncertain''. Criterion such as entropy~\cite{settles_active_2012}, least-confident~\cite{settles_active_2012}, Query by Disagreement (QBD)~\cite{cohn_improving_1994, balcan_agnostic_2006} and Bayesian Active Learning by Disagreement (BALD)~\cite{houlsby_bayesian_2011} fall into this category. Closely related to uncertainty-based acquisition functions are {\it margin}-based acquisition functions that favor unlabeled points near the decision boundary of the current SSL classifier \cite{tong_support_2001, hoi_semi-supervised_2008, jiang_bootstrapping_2021}. Support Vector Machines (SVM)~\cite{vapnik_statistical_1998} are amenable to this type of criterion, as the concept of a margin and decision boundary are inherent in the model. {\it Clustering}-based methods rely on the geometric clustering structure of the input data to guide the active learning query choices. The works of Dasgupta et al~\cite{dasgupta_hierarchical_2008}, Dasarathy et al (S2)~\cite{dasarathy_s2_2015}, and Murphy et al~\cite{maggioni_learning_2019} are examples of acquisition functions that specifically exploit the clustering structure of the input data, usually based on a graph topology that captures these geometric relationships.
{\it Look-ahead} acquisition functions are a final category that we mention, and are the motivation for the present work. Look-ahead methods leverage the current SSL classifier's state to ``look ahead'' at what changes would occur in the SSL classifier as a result of labeling an unlabeled point, such as the seminal work of Zhu et al~\cite{zhu_combining_2003}. Our proposed ``model-change'' acquisition function as well as the EMCM~\cite{cai_maximizing_2013} and Maxi-Min ``data-based norm''~\cite{karzand_maximin_2020} methods are in this category.
More specifically, the proposed model-change acquisition function $\mathcal A(k)$ of this paper approximates the difference between the {\it current} classifier $\mathbf{u}$ and the {\it look-ahead} classifier $\mathbf{u}^{+k,\hat{y}_k}$ resulting from adding $k$ to the labeled set $\mathcal L$ with a hypothetical label, $\hat{y}_k$.
While many methods of late focus on applying active learning in deep learning architectures~\cite{gal_deep_2017,kushnir_diffusion-based_2020, sener_active_2018, simeoni_rethinking_2021, ash_deep_2020}, we focus on graph-based SSL models in this paper. Graph-based SSL models leverage the geometric information from a similarity graph imposed on the set of feature vectors in conjunction with the previously observed labeling information contained in the labeled set, $\mathcal L \subset Z$. These models allow for straightforward Bayesian probabilistic interpretations that are less clear in the majority of deep learning architectures~\cite{gal_deep_2017}.
The contributions of this work are to:
\begin{itemize}
\item Provide a unifying framework for active learning in graph-based SSL models with convex loss functions more appropriate for classification tasks,
\item Present ``model-change'' active learning acquisition function built around fast look-ahead approximations previously only performed on interpolation RKHS models \cite{karzand_maximin_2020},
\item Apply a spectral truncation to the graph-based SSL models which allows for efficient storage and model-change calculations,
\item Demonstrate the speed and efficacy of the approach on both synthetic and real-world datasets, including an application to hyperspectral imagery (HSI).
\end{itemize}
\nointerlineskip
\tikzstyle{block2} = [rectangle, rounded corners, minimum width=3.9cm, minimum height=2cm, text centered, draw=black, fill=gray!30]
\tikzstyle{bigblock2} = [rectangle, rounded corners, minimum height=5cm, minimum width=5cm, text centered, draw=black, fill=gray!20]
\tikzstyle{bigblock2left} = [rectangle, rounded corners, minimum height=5cm, minimum width=4.5cm, text centered, draw=black, fill=gray!20]
\tikzstyle{bigblock2right} = [rectangle, rounded corners, minimum height=5cm, minimum width=3.9cm, text centered, draw=black, fill=gray!20]
\begin{figure}
\begin{tikzpicture}
\begin{scope}[on background layer]
\node (b1) [bigblock2left, opacity=0.98, yshift=1.8cm, fill=red!30]{};
\node[] (forward) at ($(b1.north)+(0,0.4)$) {{\bf Current Model}};
\node (b2) [bigblock2, opacity=0.8, right of=b1, xshift=4.5cm, yshift=0.2cm, fill=green!30]{};
\node[] (forward) at ($(b2.north)+(0,0.3)$) {{\bf Hypothetical ``Look-Ahead'' Models}};
\node (b3) [bigblock2, opacity=0.98, right of=b2, xshift=-.8cm, yshift=-0.1cm, fill=green!30]{};
\node (b4) [bigblock2, opacity=0.98, right of=b3, xshift=-.8cm, yshift=-0.1cm, fill=green!30]{};
\node (b5) [block2, opacity=0.98, right of=b4, xshift=4.3cm, yshift=1.5cm, fill=green!30]{};
\node (b6) [block2, opacity=0.98, right of=b4, xshift=4.3cm, yshift=-1.5cm, fill=green!30]{};
\end{scope}
\Vertex[x=-.8, y=3.2, label=1, color=red, shape=rectangle, fontcolor=white]{A}
\Vertex[x=-1.8, y=2.7, label=2, color=red!80!blue, opacity=0.8, fontcolor=white]{B}
\Vertex[x=-.6, y=2.2, label=3, color=red!60!blue, opacity=0.8, fontcolor=white]{C}
\Vertex[x=-1, y=1.2, label=4, color=blue!60!red, opacity=0.8, fontcolor=white]{D}
\Vertex[x=.1, y=1, label=5, color=blue!80!red, opacity=0.8, fontcolor=white]{E}
\Vertex[x=-.5, y=0, label=6, color=blue, shape=rectangle, fontcolor=white]{F}
\Edge(A)(B)
\Edge(A)(C)
\Edge(B)(C)
\Edge(C)(D)
\Edge(D)(E)
\Edge(D)(F)
\Edge(E)(F)
\node (b1text) [right of=b1, yshift=-1.2cm, anchor=west, xshift=-.8cm]{$\mathcal L = \{1, 6\}$};
\node (b1text2) [right of=b1, yshift=-1.7cm, anchor=west, xshift=-.8cm]{$\mathbf{y} = [1, -1]$};
\node (b1model) [right of=b1, yshift=.8cm, anchor=west, xshift=-.8cm]{$\hat{\mathbf{u}}=$};
\draw [right of=b1, yshift=1cm, xshift=0.1cm, label=b1vec] (0,0) rectangle (.5,3);
\draw [right of=b1, yshift=1cm, xshift=0.1cm, fill=blue] (0,0) rectangle (.5,.5);
\draw [right of=b1, yshift=1cm, xshift=0.1cm, fill=blue!80!red] (0,.5) rectangle (.5,1);
\draw [right of=b1, yshift=1cm, xshift=0.1cm, fill=blue!60!red] (0,1) rectangle (.5,1.5);
\draw [right of=b1, yshift=1cm, xshift=0.1cm, fill=red!60!blue] (0,1.5) rectangle (.5,2);
\draw [right of=b1, yshift=1cm, xshift=0.1cm, fill=red!80!blue] (0,2) rectangle (.5,2.5);
\draw [right of=b1, yshift=1cm, xshift=0.1cm, fill=red] (0,2.5) rectangle (.5,3);
\Vertex[x=4.9, y=3.2, label=1, color=red, shape=rectangle, fontcolor=white]{A2}
\Vertex[x=3.9, y=2.7, label=2, color=red, opacity=0.8, shape=rectangle, style=dashed, fontcolor=white]{B2}
\Vertex[x=5.1, y=2.2, label=3, color=red!80!blue, opacity=0.8, fontcolor=white]{C2}
\Vertex[x=4.7, y=1.2, label=4, color=blue!40!red, opacity=0.8, fontcolor=white]{D2}
\Vertex[x=5.8, y=1, label=5, color=blue!60!red, opacity=0.8, fontcolor=white]{E2}
\Vertex[x=5.2, y=0, label=6, color=blue, shape=rectangle, fontcolor=white]{F2}
\Edge(A2)(B2)
\Edge(A2)(C2)
\Edge(B2)(C2)
\Edge(C2)(D2)
\Edge(D2)(E2)
\Edge(D2)(F2)
\Edge(E2)(F2)
\node (b4text) [right of=b4, yshift=-1.2cm, anchor=west, xshift=-.8cm]{$\mathcal L = \{1, 2, 6\}$};
\node (b4text2) [right of=b4, yshift=-1.7cm, anchor=west, xshift=-.95cm]{$\mathbf{y} = [1, 1, -1]$};
\node (b4model) [right of=b4, yshift=.8cm, anchor=west, xshift=-1.1cm]{$\hat{\mathbf{u}}^{+k,\hat{y}_k}=$};
\draw [right of=b4, yshift=1cm, xshift=6.6cm, label=b4vec] (0,0) rectangle (.5,3);
\draw [right of=b4, yshift=1cm, xshift=6.6cm, fill=blue] (0,0) rectangle (.5,.5);
\draw [right of=b4, yshift=1cm, xshift=6.6cm, fill=blue!60!red] (0,.5) rectangle (.5,1);
\draw [right of=b4, yshift=1cm, xshift=6.6cm, fill=blue!40!red] (0,1) rectangle (.5,1.5);
\draw [right of=b4, yshift=1cm, xshift=6.6cm, fill=red!80!blue] (0,1.5) rectangle (.5,2);
\draw [right of=b4, yshift=1cm, xshift=6.6cm, fill=red] (0,2) rectangle (.5,2.5);
\draw [right of=b4, yshift=1cm, xshift=6.6cm, fill=red] (0,2.5) rectangle (.5,3);
\draw [arrow] (b1) -- (b3);
\draw [arrow] (b4) -- (b5);
\draw [arrow] (b5) -- (b6);
\node[] (b5text) at ($(b5.north)+(0,0.4)$) {{\bf Query Set Selection}};
\node[] (b5text2) at ($(b5.north)+(0,-1.4)$) {\small $\mathcal A(k) = \|\hat{\mathbf{u}}^{+k,\hat{y}_k} - \hat{\mathbf{u}}\|_2$};
\node[] (b5text3) at ($(b5.north)+(0,-.4)$) {\small For $k \in \mathcal U,$};
\node[] (b5text3) at ($(b5.north)+(0,-.9)$) {\small compute model-change: };
\node[] (b6text4) at ($(b6.north)+(0,-.7)$) {\small Select};
\node[] (b6text5) at ($(b6.north)+(0,-1.2)$) {\small $k^\ast = \argmax_{k \in \mathcal U} \mathcal A(k)$};
\end{tikzpicture}
\caption{Illustration of our model-change active learning calculations inside of flowchart in \Cref{fig:flowchart}. For unlabeled indices $k \in \mathcal U$, we compute the 2-norm difference between the current model (classifier) $\hat{\mathbf{u}}$ and each hypothetical ``look-ahead'' models $\hat{\mathbf{u}}^{+k,\hat{y}_k}$. Hypothetical labels $\hat{y}_k$ are ``pseudo-labels'' from current classifier, $\hat{\mathbf{u}}$.}
\label{fig:diagram}
\end{figure}
\section{Background}
We first introduce the family of graph-based semi-supervised learning (SSL) models that use the model-change acquisition function. Then we discuss the probabilistic counterpart to these graph-based SSL models and how these Bayesian perspectives arise in previous work in active learning.
\subsection{Graph-Based SSL Models} \label{sec:gbssl}
Consider the input data $X = \{\mathbf{x}_1, \mathbf{x}_2, \ldots, \mathbf{x}_N\} \subset \mathbb R^d$ of $N$ feature vectors with corresponding index set $Z = \{1, 2, \ldots, N\}$. Assume there exists a ``ground-truth'' classification (labeling) function $y^\dagger : Z \rightarrow \{1, 2, \ldots, n_c\}$ that maps each point $\mathbf{x}_i$ to exactly one class, identified by the label $y^\dagger_i \in \{1, 2, \ldots, n_c\}$\footnote{In the binary case we take $y^\dagger_i \in \{\pm 1\}$.}. {\it Observations} $y_j$ of these ground-truth labelings are given for indices $j \in \mathcal L \subset Z$, i.e. the labeled set. The goal of SSL is to infer the ground truth classifications $y^\dagger_k$ for $k \in \mathcal U = Z - \mathcal L$, i.e. the unlabeled set. In the binary case, we denote the concatenation of the observed labelings $\{y_j\}_{j \in \mathcal L}$ as a vector $\mathbf{y} \in \mathbb R^{|\mathcal L|}$, whereas in the multiclass case we denote the concatenations of the observed one-hot vectors $\{\mathbf{y}_j\}_{j \in \mathcal L}$ as a matrix $Y \in \mathbb R^{|\mathcal L| \times n_c}$.
We construct a similarity graph $G(Z, W)$ with edge weights $W_{ij} = \kappa(\mathbf{x}_i, \mathbf{x}_j) \ge 0$ calculated by a similarity kernel $\kappa$.
For example, the Gaussian kernel, $\kappa(\mathbf{x}_i, \mathbf{x}_j) = \exp(-\|\mathbf{x}_i - \mathbf{x}_j\|_2^2/\sigma^2)$ with kernel width parameter $\sigma > 0$, is a popular choice. Important geometric information about the data manifold is encoded in graph Laplacian matrices~\cite{belkin_manifold_2006, von_luxburg_tutorial_2007}. Two such matrices are
the {\it unnormalized graph Laplacian matrix} $L_u = D - W$ and the {\it normalized graph Laplacian matrix} $L = D^{-1/2}(D - W)D^{-1/2}$, where $D = \operatorname{diag}(d_1, d_2, \ldots, d_N), d_i = \sum_{j \neq i} W_{ij}$ is the diagonal {\it degree matrix}.
There are other possible graph Laplacian matrices one could define, but we only consider the normalized graph Laplacian, $L$. We enforce symmetry in $L$ to ensure that the eigenvalues and eigenvectors of this matrix are real-valued.
Graph Laplacian matrices are positive semi-definite and so to ensure invertibility of this matrix
we consider the perturbation $L_\tau = L + \tau^2 I$ with parameter $\tau > 0$. The matrix $L_\tau$ is now positive definite and invertible, which also ensures a well-defined Bayesian prior distribution that we present in \Cref{sec:bayesian-interpretation}.
Define a continuous-valued function on the nodes of the graph $u : Z \rightarrow \mathbb R^{n_c}$
This function can be identified with a matrix $U \in \mathbb R^{N \times n_c}$, where the $i^{th}$ row of $U$, $\mathbf{u}_i$, corresponds to the inferred classification of node $i$ via a mapping $S : \mathbb R^{n_c} \rightarrow \{1, 2, \ldots, n_c\}$ that maps the vectors $\{\mathbf{u}_i\}_{i \in Z}$ to the space of possible classes. For example, the mapping $S(\mathbf{u}_i) = \argmax_{c=1, 2, \ldots, n_c}\ [\mathbf{u}_i]_c$ infers the classification of node $i$ as the index of the maximal element of the vector.
Other thresholding functions have been explored for providing consistent classifiers~\cite{calder2020}. In the special case of binary classification, the matrix $U$ actually can be represented by a vector $\mathbf{u} \in \mathbb R^N$, with mapping $S(u_i) = \operatorname{sgn}(u_i) = \delta_{\{u_i \ge 0\}}$.
Graph-based SSL then leverages the graph Laplacian matrix via a quadratic form to regularize the SSL problem in order to
encourage similar labels for inputs that lie close to each other on the underlying data manifold~\cite{belkin_manifold_2006}.
In the binary case, we solve the following optimization problem
\begingroup
\allowdisplaybreaks
\begin{equation}\label{eq:gbssl-bin}
\hat{\mathbf{u}} = \argmin_{\mathbf{u} \in \mathbb R^N}\ \frac{1}{2} \langle \mathbf{u}, L_\tau \mathbf{u} \rangle + \sum_{j \in \mathcal L} \ell(u_j, y_j) =: \argmin_{\mathbf{u} \in \mathbb R^N}\ J_\ell(\mathbf{u}; \mathbf{y}),
\end{equation}
\endgroup
where $\ell : \mathbb R \times \mathbb R \rightarrow [0, \infty)$ is a chosen loss function and $y_j \in \{ \pm 1\}$.
In the multiclass case, the objective function and corresponding inference $\hat{U}$ are
\begin{equation}\label{eq:gbssl-mc}
\hat{U} = \argmin_{U \in \mathbb R^{N \times n_c}}\ \frac{1}{2}\langle U, L_\tau U\rangle_F + \sum_{j \in \mathcal L} \ell(\mathbf{u}_j, \mathbf{y}_j) =: \argmin_{U \in \mathbb R^{N \times n_c}}\ \mathcal J_\ell(U; Y),
\end{equation}
where $\ell: \mathbb R^{n_c} \times \mathbb R^{n_c} \rightarrow [0, \infty)$ is a chosen loss function and $\langle \cdot, \cdot \rangle_F$ is the Frobenius inner product.
\begin{table}[h!]
\centering
\renewcommand{\arraystretch}{1.3}
\begin{tabular}{||c|c|c||}
\hline
{\bf GBSSL Model Name} & $\ell(x,y)$ & {\bf Prev. Acq. F'ns}\\
\hline \hline
Harmonic Functions (HF)\cite{zhu_semi-supervised_2003} & hard-constraint & \makecell{MBR\cite{zhu_combining_2003}, V-Opt\cite{ji_variance_2012},\\ $\Sigma$-Opt\cite{ma_sigma_2013}, TSA\cite{jun_graph-based_2016} \\
Entropy\cite{long_2008}}\\
\hline
Gaussian Regression (GR)\cite{ma_active_2015} & $\frac{1}{2\gamma^2} (x - y)^2$ & \makecell{ $\Sigma$-Opt\cite{ma_active_2015}, MC\cite{miller_ecient_2020}, \\{\bf This work}} \\
\hline
Logistic (LOG)\cite{hoffmann_consistency_2020} & $\ln\left( 1 + e^{-xy/\gamma}\right)$ & \makecell{ MC\cite{miller_ecient_2020}, \\{\bf This work}} \\
\hline
Probit (P)\cite{rasmussen_gaussian_2006, hoffmann_consistency_2020} & $-\ln \Psi_\gamma(xy)$\footnote{$\Psi_\gamma(t) = \int_{-\infty}^{t} \psi_\gamma(s) ds$ is the cumulative distribution function (CDF) of a log-concave probability density function (PDF) $\psi_\gamma(s)$ with parameter $\gamma > 0$.} & \makecell{UQ-Sampling\cite{bertozzi_uncertainty_2018},\\ MC\cite{miller_ecient_2020}, {\bf This work} }\\
\hline \hline
Multiclass Gaussian Regression (MGR)\cite{bertozzi_posterior_2021} & $\frac{1}{2\gamma^2} \|\mathbf{x} - \mathbf{y}\|_2^2$ & {\bf This work} \\
\hline
Cross-Entropy (CE)\cite{kipf2017, jiang2019} & $-\sum_{c=1}^{n_c} x_c \ln(y_c)$\footnote{This loss requires that $\mathbf{x},\mathbf{y}$ be discrete probability distributions, which is adapted in~\Cref{sec:ce-intro}.} & {\bf This work}\\
\hline
\end{tabular}
\vskip0.1cm
\caption{Family of graph-based SSL models. We extend the results of MC~\cite{miller_ecient_2020} to be more computationally efficient and allow for multiclass classification.}
\label{tab:models}
\end{table}
\Cref{tab:models} lists a family of graph-based SSL models defined by the objective functions of (\ref{eq:gbssl-bin}) and (\ref{eq:gbssl-mc}); the model-change acquisition function of this present work applies to each.
Graph-based SSL models historically have been motivated by Gaussian Random Fields (GRF)~\cite{zhu_semi-supervised_2003}, and the {\it Harmonic Functions} (HF) model~\cite{zhu_combining_2003} has been influential in active learning in graph-based SSL. This model can be thought of using a ``hard-constraint'' loss function, where $\ell(x,y) = 0 $ if $x =y$ and $+\infty$ otherwise.
Acquisition functions minimizing look-ahead expected risk (MBR~\cite{zhu_combining_2003}, TSA~\cite{jun_graph-based_2016}), posterior variance (V-Opt~\cite{ji_variance_2012}, $\Sigma$-Opt~\cite{ma_sigma_2013}), and other measures of uncertainty~\cite{long_2008} come from this model. The authors in~\cite{kushnir_diffusion-based_2020} even use the HF model to enhance active learning performance within deep learning.
\subsubsection{Cross-Entropy (CE) Model}\label{sec:ce-intro}
\begin{sloppy}
As an alternative to regression for the multiclass setting, we show how to incorporate the cross-entropy loss function $\ell(\mathbf{x}, \mathbf{y}) = -\sum_{c = 1}^{n_c} x_c \ln (y_c)$ into the multiclass graph-based SSL framework (\ref{eq:gbssl-mc}). The cross-entropy loss requires that both inputs are probability distribution vectors on the set of possible classes, $\{1, 2, \ldots, n_c\}$. While the observations $\mathbf{y}_j$ satisfy this property because of their one-hot form, the rows of arbitrary $U \in \mathbb R^{N \times n_c}$ do not necessarily satisfy this same condition. The entries of $U$ are not even constrained to be non-negative. As such, we apply the``softmax'' function on the rows of $U$ to enforce this probability distribution property. Denoting the $(j,c)^{th}$ entry of $U$ by $U_{j,c}$ and the $c^{th}$ entry of $\mathbf{y}_j$ by $[\mathbf{y}_j]_c$, we have
\end{sloppy}
\begin{align*}
&\mathcal S(\mathbf{u}_j; \gamma) :=\frac{1}{\sum_{h=1}^{n_c} e^{U_{j,h}/\gamma}} (e^{U_{j,1}/\gamma}, e^{U_{j,2}/\gamma}, \ldots, e^{U_{j,n_c}/\gamma})^T, \\
q(Y | U) &\propto \exp\left( -\sum_{j \in \mathcal L} \sum_{c=1}^{n_c} [\mathbf{y}_j]_c \ln\left( [\mathcal S (\mathbf{u}_j; \gamma)]_c \right) \right) = \prod_{j \in \mathcal L} \left( -\frac{1}{\gamma}\langle \mathbf{y}_j, \mathbf{u}_j \rangle + \ln \left( \sum_{h=1}^{n_c} e^{U_{j,h}/\gamma} \right) \right) ,
\end{align*}
recalling that $y_j \in \{1, 2, \ldots, n_c\}$ is associated with the one-hot vector $\mathbf{y}_j = \mathbf{e}_{y_j} \in \mathbb R^{n_c}$. We write the {\it Cross-Entropy} model as
\begin{equation}\label{eq:ce-model}
\mathcal J_{CE}(U ; Y) = \frac{1}{2}\langle U, L_\tau U\rangle_F + \sum_{j \in \mathcal L} \left\{-\frac{1}{\gamma}\langle \mathbf{y}_j, \mathbf{u}_j \rangle + \ln \left( \sum_{h=1}^{n_c} e^{U_{j,h}/\gamma} \right) \right\}.
\end{equation}
\subsection{Bayesian Interpretation of SSL Problems}\label{sec:bayesian-interpretation}
These graph-based SSL objective functions lend themselves to a Bayesian probabilistic interpretation, as discussed in prior literature \cite{zhu_combining_2003, bertozzi_uncertainty_2018, hoffmann_consistency_2020, ji_variance_2012, ma_active_2015, qiao_uncertainty_2019-1, miller_ecient_2020}.
In the binary case, (\ref{eq:gbssl-bin}) is equivalent to finding the {\it maximum a posteriori} (MAP) estimate of a posterior probability distribution whose density $\mathbbm{P}(\mathbf{u} | \mathbf{y})$ relates to the objective function via
\begin{align}\label{eq:posterior}
\mathbbm{P}(\mathbf{u} | \mathbf{y}) &\propto \exp\left( - J_\ell(\mathbf{u}; \mathbf{y}) \right) = e^{ -\frac{1}{2} \langle \mathbf{u}, L_\tau \mathbf{u} \rangle} e^{-\sum_{j \in \mathcal L} \ell(u_j, y_j) } \propto \mu(\mathbf{u}) e^{-\Phi_\ell(\mathbf{u} ; \mathbf{y}) } ,
\end{align}
where the {\it prior} $\mu(\mathbf{u})$ follows a Gaussian prior, $\mathcal N(0, L_\tau^{-1})$, and the {\it likelihood}, $q(\mathbf{y} | \mathbf{u}) \propto $ $ \exp(-\Phi_\ell(\mathbf{u}; \mathbf{y}))$, is defined by the likelihood potential $\Phi_\ell(\mathbf{u}; \mathbf{y}) := \sum_{j \in \mathcal L} \ell(u_j, y_j)$. This prior relates to other graph-based SSL priors proposed in previous literature \cite{hoffmann_consistency_2020}. The Gaussian prior represents a prior belief over the distribution of functions $\mathbf{u}$ on the nodes of the graph, per the geometry of the data captured in the graph Laplacian matrix. The likelihood represents assumptions about the generative model that created the observed labelings $y_j$ from the ground-truth classifications $y^\dagger_j$. Each loss function $\ell$ for (\ref{eq:gbssl-bin}) defines a different likelihood and consequently a different modeling assumption about the observed labels $\mathbf{y}$ (or $Y$). We note that although the prior $\mu(\mathbf{u})$ is Gaussian, the posterior distribution of (\ref{eq:posterior}) for general loss functions $\ell$
is not necessarily Gaussian.
{\it A key idea of the present work is to approximate the non-Gaussian posterior distributions via suitable Gaussian distributions to exploit the resulting convenient form of what we term the ``look-ahead'' posterior mean and covariance. This formulation allows us to use loss functions that are arguably more natural for classification than merely the squared-error loss.}
\subsection{Look-Ahead Model}\label{sec:look-ahead-intro}
Recalling the binary graph-based SSL model objective
\[
J_\ell(\mathbf{u}; \mathbf{y}) = \frac{1}{2} \langle \mathbf{u}, L_\tau \mathbf{u} \rangle + \sum_{j \in \mathcal L} \ell(u_j, y_j),
\]
then let the ``look-ahead'' model objective be
\[
J^{k, \hat{y}_k}_\ell(\mathbf{u}; \mathbf{y}, \hat{y}_k) := \frac{1}{2} \langle \mathbf{u}, L_\tau \mathbf{u} \rangle + \sum_{j \in \mathcal L} \ell(u_j, y_j) + \ell(u_k, \hat{y}_k) = J_\ell(\mathbf{u}; \mathbf{y}) + \ell(u_k, \hat{y}_k),
\]
where we add the unlabeled index $k \in \mathcal U$ with pseudo-label $\hat{y}_k := sgn(\hat{u}_k)$ to the labeled data.
Look-ahead models have previously been introduced for designing various acquisition functions (e.g. \cite{karzand_maximin_2020, cai_batch_2017, cai_maximizing_2013, zhu_combining_2003}).
We emphasize here that in application, since $k \in \mathcal U$, we do not have access to the {\it true} labeling $y^\dagger_k$, and so this look-ahead model is a {\it hypothetical model}.
For a given $k \in \mathcal U$ and one-hot encoding $\hat{\mathbf{y}}_k \in \mathbb R^{n_c}$ of the pseudo-label $\hat{y}_k = S(\hat{\mathbf{u}}^k) \in \{1, 2, \ldots, n_c\}$, the multiclass look-ahead model becomes
\[
\mathcal J_\ell^{k, \hat{\mathbf{y}}_k}(U ; Y, \hat{\mathbf{y}}_k) := \frac{1}{2} \langle U, L_\tau U\rangle_F + \sum_{j \in \mathcal L} \ell(\mathbf{u}_j, \mathbf{y}_j) + \ell(\mathbf{u}_k, \hat{\mathbf{y}}_k) = \mathcal J_\ell(U; Y) + \ell(\mathbf{u}_k, \hat{\mathbf{y}}_k).
\]
The present work exploits a key property of the HF, GR, and MGR models that the {\it look-ahead} model's posterior mean (i.e. the graph-based SSL model's classifier) and covariance matrix are easily calculated as rank-one updates of the {\it current} model's posterior mean and covariance matrix.
We discuss this more in Sections~\ref{sec:bin-model},~\ref{sec:mgr-trunc}, and~\ref{sec:ce-trunc}.
\subsection{Laplace Approximation}
{\it Laplace approximation} is a popular technique for approximating non-Gaussian distributions with a Gaussian distribution~\cite{rasmussen_gaussian_2006}. A given probability distribution, identified by its probability density function (PDF) $\mathbbm{P}(\mathbf{x})$, can be approximated via the Gaussian distribution
\[
\mathbf{x} \sim \mathcal N(\hat{\mathbf{x}}, \hat{C}), \quad \hat{\mathbf{x}} = \argmax_{\mathbf{x} \in \mathbb R^N} \ \mathbbm{P}(\mathbf{x}), \quad \hat{C} = \left( -\nabla^2\ln(\mathbbm{P}(\mathbf{x})) |_{\mathbf{x} = \hat{\mathbf{x}}} \right)^{-1},
\]
where $\hat{\mathbf{x}}$ is the {\it maximum-a-posteriori} (MAP) estimator of $\mathbbm{P}$ and $\hat{C}$ is the Hessian matrix of the negative-log density of the distribution evaluated at the MAP estimator, $\hat{\mathbf{x}}$.
By applying the Laplace approximation to the non-Gaussian posterior distributions of \Cref{sec:bayesian-interpretation}, one can approximate look-ahead updates for calculating the model-change acquisition function.
\subsection{Active Learning Query Set Selection}\label{sec:query-set-selection}
With acquisition function $\mathcal A(k)$ in hand, one must select the query set $\mathcal Q \subset \mathcal U$ from these acquisition values.
{\it Sequential} active learning selects
the maximizer $k^\ast = \argmax_{k \in \mathcal U} \ \mathcal A(k)$ of the acquisition function (i.e. $|\mathcal Q| = 1$).
In {\it batch} active learning ($|\mathcal Q| = B > 1$), there is an added difficulty of how to best choose this subset.
Choosing the top $B$ maximizers of the {\it current} values of $\{\mathcal A(k)\}_{k \in \mathcal U}$ can yield sub-optimal results, as these maximizers often are close in the ambient feature space -- in a sense ``wasting'' precious query budget on redundant information.
Some batch active learning methods select query points via a greedy sequential process~\cite{ji_variance_2012, ma_sigma_2013, cai_maximizing_2013}, wherein the method selects the maximizer $k_1^\ast = \argmax_{k \in \mathcal U} \mathcal A(k)$ first, and then updates the acquisition function values per the new, hypothetical SSL model with added index $k_1^\ast$ and associated pseudolabel $\hat{y}_{k_1^\ast}$. They next select the maximizer of the updated acquisition values $k^\ast_2 = \argmax_{k \in \mathcal U - \{k_1^\ast\}} \mathcal A^{k_1^\ast, \hat{y}_{k_1^\ast}}(k)$, and continue likewise to select $k^\ast_1, k^\ast_2, \ldots, k^\ast_B$.
Other batch active learning methods restrict the size of the set of indices on which $\mathcal A$ is evaluated to a smaller {\it candidate set} $\mathcal S \subset \mathcal U$. These methods select the batch $\mathcal Q \subset \mathcal S$ to be the top maximizers of the acquisition function \cite{gal_deep_2017, kushnir_diffusion-based_2020}, where $\mathcal S$ is chosen uniformly at random form $\mathcal U$. This has essentially two important and positive consequences.
First, evaluating $\mathcal A$ only $\mathcal S$ is obviously computationally faster since $|\mathcal S| \ll |\mathcal U|$. Second, by selecting $\mathcal S \subset \mathcal U$ at random, we partially alleviate the problem of ``redundant'' calculations since the maximizers of $\mathcal A$ over $\mathcal S$ likely do not lie all close together.
We apply this query set set selection method to our batch active learning experiments; i.e., select $\mathcal S \subset \mathcal U$ uniformly at random and then select the top $B$ maximizers of the acquisition function.
\section{Model-Change Acquisition Derivation}\label{sec:model-derivation}
We derive the model-change acquisition function for a modified objective function that utilizes only a subset of the eigenvalues and eigenvectors of the graph Laplacian matrix.~\Cref{sec:spectral-truncation} defines spectral truncation modification of the family of graph-based SSL objectives. We derive in \Cref{sec:bin-model} the model-change acquisition function first for the binary classification models by utilizing the Laplace approximation and a corresponding simple approximate update via Newton's method. We expand this idea to multiclass models in \Cref{sec:mgr-trunc} and \Cref{sec:ce-trunc}.
\subsection{Spectral Truncation}\label{sec:spectral-truncation}
Bayesian-inspired graph-based acquisition functions often require storing a prohibitively large and dense covariance matrix $C \in \mathbb R^{N \times N}$ in memory, where $N$ is the size of the dataset.
We introduce a modification to the graph-based models to alleviate this burden by using only a subset of the eigenvalues and eigenvectors of the corresponding graph Laplacian matrix $L$; we refer to this as ``spectral truncation''.
The eigenvalues of the graph Laplacian matrix $L$ can be ordered $0 = \lambda_1 \le \lambda_2 \le \ldots \le \lambda_N$, with the first eigenvalue guaranteed to be $0$ by properties of $L$~\cite{von_luxburg_tutorial_2007}.
The low-lying eigenvalues (i.e. closer to $0$) and their corresponding eigenvectors of $L$ contain important geometric information pertaining to the data. For example, spectral clustering~\cite{von_luxburg_tutorial_2007} utilizes the eigenvectors corresponding to the first $K$ eigenvalues of $L$ to embed the data into $\mathbb R^K$ and thereafter perform clustering, often via the K-Means algorithm.
This work uses the $M$ eigenvalues closest to 0 and their corresponding eigenvectors to obtain approximations of $\hat{C}$ and $\hat{\mathcal C}$. Recalling the perturbed graph Laplacian, $L_\tau = L + \tau^2 I$, then by considering only the first $M < N$ eigenvalues of $L$, define the matrices
\[
\Lambda_\tau = \operatorname{diag}\left( \lambda_1 + \tau^2, \lambda_2 + \tau^2, \ldots, \lambda_M + \tau^2 \right), \qquad V = [\mathbf{v}^1, \mathbf{v}^2, \ldots, \mathbf{v}^M] \in \mathbb R^{N \times M},
\]
where $\mathbf{v}^i$ is the eigenvector corresponding to the $i^{th}$ eigenvalue $\lambda_i$. Replace the graph-based regularization terms of (\ref{eq:gbssl-bin}) and (\ref{eq:gbssl-mc}) to obtain the following
\begingroup
\allowdisplaybreaks
\begin{align}
J_\ell(\mathbf{u}; \mathbf{y}) = \frac{1}{2} \langle \mathbf{u}, L_\tau \mathbf{u} \rangle + \sum_{j\in \mathcal L} \ell(u_j, y_j) &\rightarrow \frac{1}{2} \langle \boldsymbol \alpha, \Lambda_\tau \boldsymbol \alpha \rangle + \sum_{j \in \mathcal L} \ell(\mathbf{e}_j^T V \boldsymbol \alpha, y_j) =: \tilde{J}_\ell(\boldsymbol \alpha; \mathbf{y}), \label{eq:bin-model-st} \\
\mathcal J_\ell(U; Y) = \frac{1}{2} \langle U, L_\tau U \rangle_F + \sum_{j\in \mathcal L} \ell(\mathbf{u}_j, \mathbf{y}_j) &\rightarrow \frac{1}{2} \langle A, \Lambda_\tau A \rangle_F + \sum_{j \in \mathcal L} \ell(\mathbf{e}_j^T V A, \mathbf{y}_j) =: \tilde{\mathcal J}_\ell(A; Y). \label{eq:multi-model-st}
\end{align}
\endgroup
The vector $\boldsymbol \alpha = V^T \mathbf{u}$ (matrix $A = V^T U$) is the projection of $\mathbf{u}$ ($U$) onto the $M$ coresponding eigenvectors. Since the eigenvectors are orthonormal ($V^TV = I \in \mathbb R^{M \times M}$), $\boldsymbol \alpha$ is the vector of coefficients of the representation of $\mathbf{u}$ in that eigenbasis for $\mathbf{u} \in \operatorname{span}\{ \mathbf{v}_1, \ldots, \mathbf{v}_M\}$.
By restricting to this subspace, we not only speed up model training since the latent space is of (much) smaller dimension $(M \ll N)$, but we also reduce the spatial complexity of the covariance matrix for the model-change acquisition function calculations.
\subsection{Binary Model}\label{sec:bin-model}
We first derive the model-change acquisition function on the binary model (\ref{eq:bin-model-st}) for the GR, Logistic, and Probit models. The derivation for the binary model follows similarly to \cite{miller_ecient_2020}, except now done on this ``spectral truncation'' modification.
\subsubsection{Laplace Approximation of the Binary Model}
The Laplace approximation of the posterior distribution corresponding to (\ref{eq:bin-model-st}), {\it with respect to the variable} $\boldsymbol \alpha \in \mathbb R^M$, gives
\begingroup
\allowdisplaybreaks
\begin{align*}
\boldsymbol \alpha | \mathbf{y} &\sim \mathcal N(\hat{\boldsymbol \alpha}, \hat{C}_{\hat{\boldsymbol \alpha}}), \qquad \hat{\boldsymbol \alpha} = \argmin_{\boldsymbol \alpha \in \mathbb R^M} \ \tilde{J}_\ell(\boldsymbol \alpha; \mathbf{y}), \\
\nabla_{\boldsymbol \alpha} \tilde{J}_\ell(\boldsymbol \alpha; \mathbf{y}) &= \Lambda_\tau \boldsymbol \alpha + \sum_{j \in \mathcal L} F(\mathbf{e}_j^TV \boldsymbol \alpha, y_j) V^T\mathbf{e}_j = \Lambda_\tau \boldsymbol \alpha + V^T \sum_{j \in \mathcal L} F(\mathbf{e}_j^TV \boldsymbol \alpha, y_j) \mathbf{e}_j, \\
\nabla_{\boldsymbol \alpha}^2 \tilde{J}_\ell(\boldsymbol \alpha; \mathbf{y}) &= \Lambda_\tau + V^T \sum_{j \in \mathcal L} F'(\mathbf{e}_j^TV \boldsymbol \alpha, y_j) \mathbf{e}_j \mathbf{e}_j^T V = \Lambda_\tau + V^T \left( \sum_{j \in \mathcal L} F'(\mathbf{e}_j^TV \boldsymbol \alpha, y_j) \mathbf{e}_j \mathbf{e}_j^T \right) V , \\
\implies \hat{C}_{\boldsymbol \alpha} &= \left( \nabla_{\boldsymbol \alpha}^2 \tilde{J}_\ell(\boldsymbol \alpha; \mathbf{y}) \right)^{-1} = \left( \Lambda_\tau + V^T \left( \sum_{j \in \mathcal L} F'(\mathbf{e}_j^TV \boldsymbol \alpha, y_j) \mathbf{e}_j \mathbf{e}_j^T \right) V \right)^{-1}.
\end{align*}
\endgroup
where $F'(x,y)$ is the second derivative of the loss function $\ell(x,y)$ {\it with respect to the first variable} following the notation of~\cite{hoffmann_consistency_2020, miller_ecient_2020}. That is, $F(x,y) := \fracpartial{\ell}{x}(x,y)$ and $F'(x,y) := \fracpartial{^2\ell}{x^2}(x,y).$
$\hat{C}_{\boldsymbol \alpha}$ is symmetric because $\Lambda_\tau$ is diagonal and the objective function $\tilde{J}_\ell(\mathbf{u}; \mathbf{y})$ is differentiable.
Recall that the mean of this Gaussian distribution $\hat{\boldsymbol \alpha}$ is the MAP estimator of the true posterior, which corresponds to the coefficients of the desired graph-based SSL model's classifier $\hat{\mathbf{u}}$ in the eigenbasis represented by $V$. This Gaussian distribution is now in a form in which one could apply adaptations of acquisition functions like MBR~\cite{zhu_combining_2003}, V-Opt~\cite{ji_variance_2012}, and $\Sigma$-Opt~\cite{ma_sigma_2013} that were originally defined on Gaussian models (e.g. GR and HF), only now using other convex loss functions besides the squared-error loss.\footnote{Recall that the Laplace approximation of a Gaussian distribution is itself,
and so the Laplace approximation of the GR model's posterior distribution is itself.}
\subsubsection{Look-Ahead Updates} \label{sec:look-ahead-updates}
With $\hat{\boldsymbol \alpha}$ denoting the {\it current} model's $(\tilde{J}_\ell(\boldsymbol \alpha; \mathbf{y}))$ MAP estimator, let $\hat{\boldsymbol \alpha}^{k, \hat{y}_k}$ denote the {\it look-ahead} model's $(\tilde{J}_\ell^{k,\hat{y}_k}(\boldsymbol \alpha; \mathbf{y}, \hat{y}_k))$ MAP estimator. While for most loss functions one cannot directly compute $\hat{\boldsymbol \alpha}^{k,\hat{y}_k}$ as a rank-one update from $\hat{\boldsymbol \alpha}$, we compute an approximation $\tilde{\boldsymbol \alpha}^{k,\hat{y}_k} \approx \hat{\boldsymbol \alpha}^{k, \hat{y}_k}$ by computing a single step of Newton's Method on the look-ahead objective $\tilde{J}^{k,\hat{y}_k}(\boldsymbol \alpha; \mathbf{y}, \hat{y}_k)$, {\it starting with the current MAP estimator} $\hat{\boldsymbol \alpha}$
\begingroup
\allowdisplaybreaks
\begin{align*}
\tilde{\boldsymbol \alpha}^{k, \hat{y}_k} &= \hat{\boldsymbol \alpha} - \left( \nabla_{\boldsymbol \alpha}^2 \tilde{J}^{k, \hat{y}_k}_\ell(\hat{\boldsymbol \alpha}; \mathbf{y}, \hat{y}_k)\right)^{-1} \left( \nabla_{\boldsymbol \alpha} \tilde{J}^{k, \hat{y}_k}_\ell(\hat{\boldsymbol \alpha}; \mathbf{y}, \hat{y}_k) \right) \\
&= \hat{\boldsymbol \alpha} - \left( \hat{C}_{\hat{\boldsymbol \alpha}}^{-1} + F'(\mathbf{v}_k^T \hat{\boldsymbol \alpha}, \hat{y}_k) \mathbf{v}^{k} (\mathbf{v}^{k})^T \right)^{-1}\left( \cancelto{0}{\nabla_{\boldsymbol \alpha} \tilde{J}_\ell(\hat{\boldsymbol \alpha};\mathbf{y})} + F(\mathbf{v}_k^T \hat{\boldsymbol \alpha}, \hat{y}_k) \mathbf{v}_k \right) \\
&= \hat{\boldsymbol \alpha} - \left( \hat{C}_{\hat{\boldsymbol \alpha}} - \hat{C}_{\hat{\boldsymbol \alpha}} \mathbf{v}_k \left( \frac{1}{F'(\mathbf{v}_k^T \hat{\boldsymbol \alpha}, \hat{y}_k)} + \mathbf{v}_k^T\hat{C}_{\hat{\boldsymbol \alpha}} \mathbf{v}_k \right)^{-1} (\mathbf{v}^{k})^T\hat{C}_{\hat{\boldsymbol \alpha}}\right) F(\mathbf{v}_k^T \hat{\boldsymbol \alpha}, \hat{y}_k) \mathbf{v}_k.
\end{align*}
This gives
\begin{equation}\label{eq:bin-na-update}
\tilde{\boldsymbol \alpha}^{k, \hat{y}_k} = \hat{\boldsymbol \alpha} - \frac{F(\mathbf{v}_k^T \hat{\boldsymbol \alpha}, \hat{y}_k)}{1 + F'(\mathbf{v}_k^T \boldsymbol \alpha, \hat{y}_k)\mathbf{v}_k^T\hat{C}_{\hat{\boldsymbol \alpha}} \mathbf{v}_k } \hat{C}_{\hat{\boldsymbol \alpha}} \mathbf{v}_k,
\end{equation}
\endgroup
where in the second to last line we have used (\ref{eq:smw-identity}) and we have introduced $\mathbf{v}_k := (\mathbf{e}_k^T V)^T$, the $k^{th}$ row of $V$ as a column vector in $\mathbb R^M$. Note that in the second line, $\hat{\boldsymbol \alpha}$ satisfies $ \nabla \tilde{J}(\hat{\boldsymbol \alpha}; \mathbf{y}) = \Lambda_\tau \hat{\boldsymbol \alpha} + \sum_{j \in \mathcal L} F((\mathbf{v}^j)^T \hat{\boldsymbol \alpha}, y_j) \mathbf{e}_j = 0$ since it is the minimizer of $\tilde{J}(\boldsymbol \alpha; \mathbf{y})$.
\subsubsection{Model-Change (MC) Acquisition Function}\label{sec:mc-method}
We now define the acquisition function that we term ``Model-Change'' (MC). This criterion quantifies how much the underlying graph-based SSL classifier
changes as a result of adding a node $k \in \mathcal U$ and hypothesized label $\hat{y}_k$; that is, we measure how much the model classifier, $\hat{\mathbf{u}} = V \hat{\boldsymbol \alpha}$, would change under the look-ahead model, $\hat{\mathbf{u}}^{k, \hat{y}_k} = V \hat{\boldsymbol \alpha}^{k, \hat{y}_k}$. We approximate this model change via the update $\tilde{\boldsymbol \alpha}^{k, \hat{y}_k}$ of (\ref{eq:bin-na-update}).
Previous works indicate that calculating the approximate change in a model (classifier) from the addition of an index $k$ and associated pseudo-label $\hat{y}_k$ is an effective criterion for active learning~\cite{miller_ecient_2020, cai_maximizing_2013, karzand_maximin_2020}.
The present work extends the MC method in~\cite{miller_ecient_2020} and resembles the Maxi-Min method of~\cite{karzand_maximin_2020}, wherein the authors investigate active learning in kernel-based interpolation. While exact look-ahead calculations are possible in~\cite{karzand_maximin_2020}, the current work allows for a broader range of methods (\Cref{tab:models}) by approximating look-ahead calculations via (\ref{eq:bin-na-update}).
Employing (\ref{eq:bin-na-update}), we propose the following MC acquisition function
\begin{align*}
\mathcal A(k) &= \min_{\hat{y}_k \in\{\pm 1\} } \|\hat{\mathbf{u}}^{k, \hat{y}_k} - \hat{\mathbf{u}}\|_2 \approx \min_{\hat{y}_k \in\{\pm 1\} } \left\|\tilde{\mathbf{u}}^{k, \hat{y}_k} - \hat{\mathbf{u}} \right\|_2 = \min_{\hat{y}_k \in\{\pm 1\} } \left\|\tilde{\boldsymbol \alpha}^{k, \hat{y}_k} - \hat{\boldsymbol \alpha} \right\|_2 \\
&= \min_{\hat{y}_k \in \{\pm 1\}} \ \left| \frac{F(\mathbf{v}_k^T \hat{\boldsymbol \alpha}, \hat{y}_k)}{1 + F'(\mathbf{v}_k^T \hat{\boldsymbol \alpha}, \hat{y}_k)\mathbf{v}_k^T\hat{C}_{\hat{\boldsymbol \alpha}} \mathbf{v}_k } \right| \left\| \hat{C}_{\hat{\boldsymbol \alpha}} \mathbf{v}_k \right\|_2 \\
&= \min_{\hat{y}_k \in \{\pm 1\}} \ \left| \frac{F(\hat{u}_k, \hat{y}_k)}{1 + F'(\hat{u}_k, \hat{y}_k)\mathbf{v}_k^T\hat{C}_{\hat{\boldsymbol \alpha}} \mathbf{v}_k } \right| \left\| \hat{C}_{\hat{\boldsymbol \alpha}} \mathbf{v}_k \right\|_2,
\end{align*}
since the columns of $V$ are orthonormal and $\hat{u}_k = \mathbf{v}_k^T \hat{\boldsymbol \alpha}$. Recall that $\hat{y}_k$ is the current ``pseudo-label'' for node $k$ as given by the current classifier $\hat{\mathbf{u}}$.
We can write this acquisition function explicitly for each considered binary classification model as
\begingroup
\allowdisplaybreaks
\begin{align*}
\mathcal A_{GR}(k) &= \frac{|\hat{u}_k - \operatorname{sgn}(\hat{u}_k)|}{\gamma^2 + \mathbf{v}_k^T\hat{C}_{\hat{\boldsymbol \alpha}} \mathbf{v}_k}\|\hat{C}_{\hat{\boldsymbol \alpha}} \mathbf{v}_k\|_2, \\
\mathcal A_{L}(k) &= \frac{\gamma(1 + e^{|\hat{u}_k|/\gamma})}{\gamma^2(1 + e^{|\hat{u}_k|/\gamma})^2 + e^{|\hat{u}_k|/\gamma}\mathbf{v}_k^T\hat{C}_{\hat{\boldsymbol \alpha}} \mathbf{v}_k}\|\hat{C}_{\hat{\boldsymbol \alpha}} \mathbf{v}_k\|_2, \\
\mathcal A_{P}(k) &= \frac{\psi_\gamma(|\hat{u}_k|)\Psi_\gamma(|\hat{u}_k|)}{\left|\Psi^2_\gamma(|\hat{u}_k|) + \left( \psi^2_\gamma(|\hat{u}_k|) - \psi'_\gamma(|\hat{u}_k|)\Psi_\gamma(|\hat{u}_k|) \right)\mathbf{v}_k^T\hat{C}_{\hat{\boldsymbol \alpha}} \mathbf{v}_k\right|}\|\hat{C}_{\hat{\boldsymbol \alpha}} \mathbf{v}_k\|_2,
\end{align*}
\endgroup
where $\hat{u}_k \hat{y}_k = \hat{u}_k sgn(\hat{u}_k) = |\hat{u}_k|$.
Note that in the GR model case, this notion of ``model-change'' as calculated is exact; the value $\mathcal A_{GR}(k)$ is exactly how much the underlying GR classifier would change by selecting $k$ with the label $\hat{y}_k = sgn(\hat{u}_k)$.
\subsection{Multiclass Gaussian Regression}\label{sec:mgr-trunc}
We now apply the derivation of the acquisition function to the MGR case. Recalling (\ref{eq:multi-model-st}), then the gradient and Hessian are
\begin{align*}
\nabla_{A} \tilde{\mathcal J}_{GR}(A; Y)&= \Lambda_\tau A + \frac{1}{\gamma^2} V^TP^T\left( PV A - Y\right), \\
\nabla_{A}^2 \tilde{\mathcal J}_{GR}(A; Y) &= \Lambda_\tau + V^T\left( \frac{1}{\gamma^2}P^TP\right) V,
\end{align*}
which then gives the Gaussian posterior distribution
\begin{align*}
A | Y &\sim \mathcal N(\hat{A}, I_{n_c} \times C_{\hat{A}}),\quad \hat{A} = \argmin_{A \in \mathbb R^{M \times n_c}} \ \tilde{\mathcal J}_{GR}(A; Y),\\
C_{\hat{A}} &= \left( \nabla_{\boldsymbol \alpha}^2 \tilde{\mathcal J}_{GR}(\hat{A}; Y) \right)^{-1} = \left( \Lambda_\tau + V^T\left( \frac{1}{\gamma^2}P^TP\right) V \right)^{-1}.
\end{align*}
The posterior mean update (which is exact for MGR) becomes
\begingroup
\allowdisplaybreaks
\begin{align*}
\hat{A}^{k, \hat{y}_k} &= \hat{A} - \left( \nabla_A^2 \tilde{\mathcal J}_{GR}^{k,\hat{y}_k}(\hat{A}; Y, \hat{\mathbf{y}}_k) \right)^{-1} \left( \nabla_A \tilde{\mathcal J}_{GR}^{k,\hat{y}_k}(\hat{A}; Y, \hat{\mathbf{y}}_k) \right) \\
&= \hat{A} - \frac{1}{ \gamma^2 + \mathbf{v}_k^T C_{\hat{A}} \mathbf{v}_k} C_{\hat{A}} \mathbf{v}_k \left( \mathbf{v}_k^T \hat{A} - (\hat{\mathbf{y}}_k)^T\right).\quad (\text{\ref{eq:smw-identity}})
\end{align*}
\endgroup
With the pseudolabel $\hat{y}_k = \argmax_{c = 1, 2, \ldots, n_c} \hat{U}_{k,c}$ and corresponding one-hot encoding $\hat{\mathbf{y}}_k\in \mathbb R^{n_c}$, the MC acquisition function becomes (using the relation $U = V A$)
\begingroup
\allowdisplaybreaks
\begin{align*}
\mathcal A_{GR}(k) &= \left\| \hat{U}^{k, \hat{y}_k} - \hat{U} \right\|_F = \left\| \hat{A}^{k, \hat{y}_k} - \hat{A} \right\|_F \\
&= \left|\frac{1}{\gamma^2 + \mathbf{v}_k^T C_{\hat{A}} \mathbf{v}_k} \right| \left\|C_{\hat{A}} \mathbf{v}_k \left( \mathbf{v}_k^T \hat{A} - (\hat{\mathbf{y}}_k)^T \right) \right\|_F \\
&= \frac{1}{\gamma^2 + \mathbf{v}_k^T C_{\hat{A}} \mathbf{v}_k} \|C_{\hat{A}} \mathbf{v}_k\|_2 \| \hat{A}^T\mathbf{v}_k - \hat{\mathbf{y}}_k \|_2
\end{align*}
\endgroup
where we have used the orthonormality of the columns of $V$ and (\ref{eq:rank-one-frob}).
\subsection{Cross-Entropy Model}\label{sec:ce-trunc}
The softmax function in the Cross-Entropy (CE) model of \Cref{sec:ce-intro} introduces dependencies between the columns of $U$. For ease in calculations, let the vector $\mathbf{u} \in \mathbb R^{Nn_c}$ be the concatenation of the columns of $U$. Likewise, define $\boldsymbol \alpha \in \mathbb R^{Mn_c}$ to be the concatenation of the columns of the matrix $A$. Further, define
$\mathcal V := \operatorname{diag}\left( V, V, \ldots, V\right)\in \mathbb R^{Nn_c \times Mn_c}$ and $\Lambda_\tau^{\bigotimes} := \operatorname{diag}\left( \Lambda_\tau, \Lambda_\tau, \ldots, \Lambda_\tau\right) \in \mathbb R^{Mn_c \times Mn_c}$.
We also define $\mathcal P_j \in \mathbb R^{n_c \times Nn_c}$ to be the projection matrix that picks out the indices in $\mathbf{u} \in \mathbb R^{Nn_c}$ corresponding to node $j$; i.e., selecting the $j^{th}$ row of the related matrix $U \in \mathbb R^{N \times n_c}$. Then $\mathbf{u} = \mathcal V \boldsymbol \alpha$, and with $\hat{\mathbf{e}}_i$ denoting the $i^{th}$ standard basis vector in $\mathbb R^{Nn_c}$, the spectral truncation CE objective (\ref{eq:ce-model}) can be written as
\begin{align}\label{eq:ce-st-obj}
\tilde{\mathcal J}_{CE}(\boldsymbol \alpha; Y)&= \frac{1}{2}\langle \boldsymbol \alpha, \Lambda_\tau^{\bigotimes} \boldsymbol \alpha \rangle + \sum_{j \in \mathcal L} \left\{ - \frac{1}{\gamma} \langle \mathbf{y}_j, \mathcal P_j \mathcal V \boldsymbol \alpha \rangle + \ln \left( \sum_{h=1}^{n_c} e^{\langle \boldsymbol \alpha, \mathcal V^T \hat{\mathbf{e}}_{j + (h-1)N} \rangle/\gamma} \right) \right\},
\end{align}
where $\mathbf{u} = \mathcal V \boldsymbol \alpha$ and $\mathcal V^T \mathcal V = I_{Mn_c}$ by orthonormality of the eigenvectors. Defining
\[
\pi_c^j = \frac{e^{\langle \boldsymbol \alpha, \mathcal V^T \hat{\mathbf{e}}_{j + (c-1)N}\rangle/\gamma} }{\sum_{h=1}^{n_c} e^{\langle \boldsymbol \alpha, \mathcal V^T \hat{\mathbf{e}}_{j + (h-1)N}\rangle /\gamma}}\quad \text{and} \quad \pi^j := (\pi_1^j, \ldots, \pi_{n_c}^j)^T \in \mathbb R^{n_c},
\]
the gradient and Hessian of (\ref{eq:ce-st-obj}) are
\begin{align*}
\nabla \tilde{\mathcal J}_{CE}(\boldsymbol \alpha; Y) &= \Lambda_\tau^{\bigotimes} \mathcal V + \frac{1}{\gamma}\mathcal V^T \sum_{j \in \mathcal L} \mathcal P_j^T\left( \pi^j - \mathbf{y}_j \right), \\
\nabla^2 \tilde{\mathcal J}_{CE}(\boldsymbol \alpha; Y) &= \Lambda_\tau^{\bigotimes} + \frac{1}{\gamma^2}\mathcal V^T \left( \mathbf{D}_{\mathcal L}(\boldsymbol \alpha) - \Pi_{\mathcal L}(\boldsymbol \alpha) \Pi_{\mathcal L}^T(\boldsymbol \alpha) \right) \mathcal V,
\end{align*}
where we refer the reader to the supplementary material (\Cref{smsec:grad-hess-ce}) for full calculation details.
The Laplace approximation for the CE model yields
\begin{align*}
\boldsymbol \alpha | \mathbf{y} &\sim \mathcal N(\hat{\boldsymbol \alpha}, \mathcal C_{\hat{\boldsymbol \alpha}}), \quad \hat{\boldsymbol \alpha} = \argmin_{\boldsymbol \alpha \in \mathbb R^{Mn_c}}\ \tilde{\mathcal J}_{CE}(\boldsymbol \alpha; Y) \\
\mathcal C_{\hat{\boldsymbol \alpha}} &= \left( \nabla^2 \tilde{\mathcal J}_{CE}(\boldsymbol \alpha; Y) \right)^{-1} = \left( \Lambda_\tau^{\bigotimes} + \frac{1}{\gamma^2}\mathcal V^T \left( \mathbf{D}_{\mathcal L}(\boldsymbol \alpha) - \Pi_{\mathcal L}(\boldsymbol \alpha) \Pi_{\mathcal L}^T(\boldsymbol \alpha) \right) \mathcal V \right)^{-1},
\end{align*}
where we emphasize the dependence of $\mathcal C_{\boldsymbol \alpha}$ on the variable $\boldsymbol \alpha$, specifically taking the value $\mathcal C_{\hat{\boldsymbol \alpha}}$ at the MAP estimator $\hat{\boldsymbol \alpha}$.
The inverse $\mathcal C_{\hat{\boldsymbol \alpha}} = \left( \nabla^2 \tilde{\mathcal J}(\hat{\boldsymbol \alpha})\right)^{-1} \in \mathbb R^{Mn_c \times Mn_c}$ is not prohibitively costly to compute because of its restricted size.
Referring to the calculations in the supplementary material (\Cref{smsec:grad-hess-ce}) the look-ahead calculations become
\begin{align*}
\nabla \tilde{\mathcal J}_{CE}^{k,\hat{y}_k}(\hat{\boldsymbol \alpha}; Y, \hat{\mathbf{y}}_k) &= \nabla \tilde{\mathcal J}_{CE}(\hat{\boldsymbol \alpha}; Y) + \frac{1}{\gamma}\mathcal V_k^T\left( \pi^k - \hat{\mathbf{y}}_k \right), \qquad \pi^k = \left( \pi^k_1, \ldots, \pi^k_{n_c} \right)^T\\
\nabla^2 \tilde{\mathcal J}_{CE}^{k,\hat{y}_k}(\hat{\boldsymbol \alpha}; Y, \hat{\mathbf{y}}_k) &= \mathcal C_{\hat{\boldsymbol \alpha}}^{-1} + \frac{1}{\gamma^2}\mathcal V_k^TB_k \mathcal V_k,
\end{align*}
where $\mathcal V_k := \mathcal P_k \mathcal V$ and $B_k := \operatorname{diag}\left( \pi^k \right) - \pi^k \left( \pi^k\right)^T$.
A Cholesky decomposition of the positive semi-definite $B_k = T_k^T T_k$ and using (\ref{eq:smw-identity}) twice enable the approximation\footnote{Again, see \Cref{smsec:grad-hess-ce} for more details}
\begingroup
\allowdisplaybreaks
\begin{align*}
\tilde{\boldsymbol \alpha}^{k,\hat{y}_k} &= \hat{\boldsymbol \alpha} - \left( \nabla^2 \tilde{\mathcal J}_{CE}^{k,\hat{y}_k}(\hat{\boldsymbol \alpha}; Y, \hat{\mathbf{y}}_k)\right)^{-1}\left( \nabla\tilde{\mathcal J}_{CE}^{k,\hat{y}_k}(\hat{\boldsymbol \alpha}; Y, \hat{\mathbf{y}}_k) \right) \\
&= \hat{\boldsymbol \alpha} - \frac{1}{\gamma}\mathcal C_{\hat{\boldsymbol \alpha}} \mathcal V_k^T\left( I - T_k^T \left( I + T_kG_k T_k^T \right)^{-1} T_k G_k\right) \left( \pi^k - \hat{\mathbf{y}}_k\right),
\end{align*}
\endgroup
where $G_k = \frac{1}{\gamma^2} \mathcal V_k \mathcal C_{\hat{\boldsymbol \alpha}} \mathcal V_k^T$.
With pseudo-label $\hat{y}_k$ and corresponding one-hot encoding $\hat{\mathbf{y}}_k$, the MC acquisition function for the CE spectral truncation modification becomes
\begin{align*}
\mathcal A_{CE}(k) &= \|\hat{\mathbf{u}} - \tilde{\mathbf{u}}^{k, \hat{y}_k}\|_2 = \|\hat{\boldsymbol \alpha} - \tilde{\boldsymbol \alpha}^{k, \hat{y}_k}\|_2 \\
&= \frac{1}{\gamma}\left\| \mathcal C_{\hat{\boldsymbol \alpha}} \mathcal V_k^T\left( I - T_k^T \left( I + T_kG_k T_k^T \right)^{-1} T_k G_k\right) \left( \pi^k - \hat{\mathbf{y}}_k\right)\right\|_2.
\end{align*}
This is efficient to compute because calculating $I - T_k^T \left( I + T_kG_k T_k^T \right)^{-1} T_k G_k \in \mathbb R^{n_c \times n_c}$ involves only $n_c \times n_c$ matrices.
\section{Experiments \& Numerics}\label{sec:numerics}
\begin{figure}[htbp]\label{fig:bc-choices}
\centering
\subfloat[{\bf Ground Truth}]{\label{fig:bc-gt}\includegraphics[width=.3\textwidth]{images/binary-clusters/bc-gt.pdf}}
\subfloat[{\bf UNC-LOG}]{\label{fig:bc-unc-log}\includegraphics[width=.3\textwidth]{images/binary-clusters/uncertainty-log-0001-0500-50-0.pdf}}
\subfloat[{\bf VOPT-HF}]{\label{fig:bc-vopt-hf}\includegraphics[width=.3\textwidth]{images/binary-clusters/vopt-hf-001-0.pdf}}
\vskip0.01in
\subfloat[{\bf MC-GR}]{\label{fig:bc-mc-gr}\includegraphics[width=.3\textwidth]{images/binary-clusters/mc-gr-0001-0500-50-0.pdf}}
\hskip0.4in
\subfloat[{\bf DB-RKHS}]{\label{fig:bc-db-rkhs}\includegraphics[width=.3\textwidth]{images/binary-clusters/db-rkhs-01-0.pdf}}
\caption{Binary-Clusters Sequential Active Learning Choices. Each plot shows the ground truth classification (red/blue) for Binary-Clusters dataset as well as the active learning choices (yellow stars) for an assortment of acquisition functions. {\bf UNC-LOG} selects points that are between only some squares, while {\bf VOPT-HF} selects points evenly spread out over the whole domain. In either case, overall performance is suboptimal compared to {\bf DB-RKHS}, and {\bf MC-GR} methods which select points located in each of the squares as well as between squares.}
\end{figure}
This section results for the Model-Change (MC) acquisition function compared to other active learning acquisition functions in the various graph-based SSL models of \Cref{tab:models}. We reference each acquisition function in the form [{\it abbreviation of acquisition function}]-[{\it abbreviation of underlying model}]; for example {\bf MC-GR} denotes the MC acquisition function in the Gaussian Regression (GR) model. The acquisition function acronyms are: {\bf MC} (Model-Change), {\bf UNC} (Uncertainty~\cite{settles_active_2012}), {\bf RAND} (Random), {\bf VOPT} (V-Opt~\cite{jun_graph-based_2016}), and {\bf SOPT} ($\Sigma$-Opt~\cite{ma_sigma_2013}). The underlying models considered are {\bf GR} (Gaussian Regression, \ref{sec:mc-method}), {\bf MGR} (Multiclass Gaussian Regression, \ref{sec:mgr-trunc}), {\bf LOG} (Logistic, \ref{sec:mc-method}), {\bf P} (Probit, \ref{sec:mc-method}), and {\bf CE} (Cross-Entropy, \ref{sec:ce-trunc}), all in the spectral truncation form of \ref{sec:spectral-truncation}.
We showcase the method on a synthetic dataset (Binary-Clusters) as well as three real-world datasets: MNIST~\cite{lecun_mnist_nodate} and two hyperspectral imagery (HSI) datasets, Salinas A and Urban. On the binary tests, we include comparisons with the {\it data-based norm} acquisition function ({\bf DB-RKHS})~\cite{karzand_maximin_2020} as well as the original V-Opt~\cite{jun_graph-based_2016} and $\Sigma$-Opt~\cite{ma_sigma_2013} methods in the Harmonic Functions model~\cite{zhu_semi-supervised_2003}, annotated as {\bf VOPT-HF} and {\bf SOPT-HF} in the plots below. We calculate the average accuracies over five trials {\it according to the underlying SSL classifier of the acquisition function}; that is, for the choices from the {\bf MC-GR} acquisition function, we report the accuracies {\it in the GR model}. We straightforwardly adapt V-Opt and $\Sigma$-Opt methods for the MGR model to allow for comparison on the multiclass tests, see \Cref{smsec:v-sigma-st}.
All but one of the tests run in the ``batch'' mode of active learning, wherein for simplicity we select $B = 5$ points for the query set $\mathcal Q$ per active learning iteration. Per the discussion of \Cref{sec:query-set-selection}, at each iteration the {\it candidate set} $\mathcal S \subset \mathcal U$ contains $10\%$ of the total points in $\mathcal U$, sampled uniformly at random. The query set comprises the top $B = 5$ {\it maximizers} of the active learning acquisition function on $\mathcal S$.
While an interesting question, this paper does not explore varying the batch size nor the candidate set size; we leave this for future work.
\subsection{Graph Construction Settings}
We construct similarity graphs using shared parameters across the different datasets. For the non-hyperspectral datasets (Binary-Clusters and MNIST), the similarity graph contain $10$-nearest neighbors with weights $w_{ij}$ given by the Gaussian similarity kernel $\kappa(\mathbf{x}_i, \mathbf{x}_j) = \exp(\|\mathbf{x}_i - \mathbf{x}_j\|_2^2/\sigma)$, with $\sigma = 3$. For the hyperspectral datasets (Salinas A and Urban), the similarity graph is constructed using $15$-nearest neighbors with weights $w_{ij}$ given by the cosine similarity $\kappa(\mathbf{x}_i, \mathbf{x}_j) = \langle \mathbf{x}_i, \mathbf{x}_j\rangle / \|\mathbf{x}_i\|_2 \|\mathbf{x}_j\|_2$. As is common in similarity graph construction, we employ Zelnik-Perona scaling~\cite{zelnik-perona04}, but only for the {\it non-hyperspectral} datasets. Due to the sparse nature of these similarity graphs, the $M=50$ lowest eigenvalues of the graph's normalized Laplacian matrix are calculated with standard sparse eigensolvers.\footnote{One could use the Nystr\"{o}m extension~\cite{goos_spectral_2002} for calculating approximate eigenvalues and eigenvectors in the case that constructing such a k-nearest neighbors similarity graph is prohibitively costly~\cite{bertozzi_diffuse_2016}.} In the binary experiments, the eigenvalue perturbation $\tau$ (\Cref{sec:model-derivation}) is set to $\tau = 0.001$, while $\tau = 0.005$ for the multiclass experiments.
For the binary experiments ({\bf GR, LOG,} and {\bf P}), the loss parameter is set to $\gamma = 0.5$. We use the reported settings of $h=0.1$ in the {\bf RKHS} model~\cite{karzand_maximin_2020} and $\delta = 0.01$ in the {\bf HF} model~\cite{zhu_combining_2003} in the corresponding experiments. For the multiclass experiments, $\gamma = 0.1, 0.5$ in the {\bf MGR} and {\bf CE} models, respectively.
\subsection{Binary-Clusters} \label{sec:checkerboard-numerics}
Binary-Clusters is a synthetically created dataset we created consisting of various clusters of differing sizes, locations, and spreads. \Cref{fig:bc-gt} shows the ground-truth classification of these clusters. In our code\footnote{\url{https://github.com/millerk22/model-change-paper}} we provide the code for recreating this particular dataset with a function entitled \texttt{create\_binary\_clusters()}. We run two tests: one that selects 100 query points {\it sequentially} ($B=1$ per iteration) and the other that selects 100 query points in {\it batches} ($B=5$ per iteration). Both tests begin with only 2 initially labeled points, one in each class.
Successful active learning on this dataset requires ``exploring'' the many different regions (squares) as well as ``exploiting'' the true decision boundaries between adjacent squares efficiently while the underlying classifer improves.
\begin{figure}[ht]
\centering
\subfloat[Sequential, $B=1$]{\label{fig:acc-bc-1}\includegraphics[width=.405\textwidth]{images/binary-clusters/sequential-acc.pdf}}
\hskip0.2in
\subfloat[Batch, $B=5$]{\label{fig:acc-bc-5}\includegraphics[width=.405\textwidth]{images/binary-clusters/batch-acc.pdf}}
\caption{Binary-Clusters accuracy plots, with $2$ initially labeled points in each experiment. Sequential (\ref{fig:acc-bc-1}) performs $200$ active learning iterations selecting $B=1$ points per iteration, while Batch (\ref{fig:acc-bc-5}) performs 100 active learning iterations selecting $B=5$ points per iteration. {\bf VOPT-HF} and {\bf SOPT-HF} achieve a quick initial accuracy increase in the sequential test (\ref{fig:acc-bc-1}), but level off at a lower accuracy than {\bf MC-GR}, {\bf MC-P}, and {\bf DB-RKHS}; the {\bf HF} and {\bf RKHS} acquisition functions are however more costly to compute (\Cref{fig:timing}). }
\end{figure}
\subsection{MNIST}
The MNIST~\cite{lecun_mnist_nodate} dataset contains 70,000 grayscale $28 \times 28$ pixel images of
handwritten digits (0-9). We represent each image as a 784-dimensional vector $\mathbf{x}_i$ and normalize the pixel values to range from 0 to 1. Each trial begins with 20 labeled points (i.e. $\approx0.03\%$ of the total, $2$ points per class) and then selects 500 active learning points in batches of size $B=5$. We consider the graph containing the full 70,000 points in the MNIST dataset and calculate the accuracy over the unlabeled set at each iteration, not a held-out ``testing set''. This is more relevant to the SSL framework, as opposed to supervised learning.
\begin{figure}[htbp]\label{fig:mgr-plots}
\centering
\subfloat[MNIST]{\label{fig:acc-mnist-mgr}\includegraphics[width=.3\textwidth]{images/mnist-acc-mgr.jpg}}
\subfloat[Salinas A]{\label{fig:acc-salinas-mgr}\includegraphics[width=.3\textwidth]{images/salinas-acc-mgr.jpg}}
\subfloat[Urban]{\label{fig:acc-urban-mgr}\includegraphics[width=.3\textwidth]{images/urban-acc-mgr.jpg}}
\caption{Accuracy plots for acquisition functions in the {\bf MGR} model applied to MNIST, Salinas A, and Urban datasets. For each dataset, two points are selected uniformly at random from each class to initially label and then acquisition functions select 500 points in 100 batches of size $B=5$.}
\end{figure}
\begin{figure}[htbp]\label{fig:ce-plots}
\centering
\subfloat[MNIST]{\label{fig:acc-mnist-ce}\includegraphics[width=.3\textwidth]{images/mnist-acc-ce.jpg}}
\subfloat[Salinas A]{\label{fig:acc-salinas-ce}\includegraphics[width=.3\textwidth]{images/salinas-acc-ce.jpg}}
\subfloat[Urban]{\label{fig:acc-urban-ce}\includegraphics[width=.3\textwidth]{images/urban-acc-ce.jpg}}
\caption{Accuracy plots for acquisition functions in the {\bf CE} model applied to MNIST, Salinas A, and Urban datasets. For each dataset, two points are selected uniformly at random from each class to initially label and then acquisition functions select 500 points in 100 batches of size $B=5$.}
\end{figure}
\begin{figure}[htbp]
\centering
\subfloat[Salinas A]{\label{fig:salinas-gt}\includegraphics[width=.43\textwidth]{images/gt-salinas.pdf}}
\hskip0.2in
\subfloat[Urban]{\label{fig:urban-gt}\includegraphics[width=.43\textwidth]{images/gt-urban.pdf}}
\caption{Hyperspectral Dataset Ground-Truth Classifications
\end{figure}
\subsection{Salinas A Hyperspectral Imagery Dataset}\label{sec:salinas-description}
The Salinas A dataset is a common Hyperspectral Imagery (HSI) dataset containing 7,138 total pixels in a $83 \times 86$ image in $d = 224$ wavelengths. This is an image of Salinas, USA taken with the Aviris sensor; this image contains $6$ classes of plant types arranged in a diagonal pattern (see \Cref{fig:salinas-gt}). The goal is to classify the pixels $\mathbf{x}_i \in \mathbb R^{d}$ in the image into material classes based on the samples from the $d$ different wavelengths. The ``ground-truth'' classification is shown in \Cref{fig:salinas-gt}. Each trial begins with two initially labeled points per class and then selects 500 active learning points in batches of size $B=5$.
\subsection{Urban Hyperspectral Imagery Dataset}
The Urban dataset is another common HSI dataset which contains $94,249$ total pixels in a $307 \times 307$ image where each pixel represents a $2 \times 2 \ m^2$ area. The original hyperspectral image contains $210$ wavelengths sampled ranging from 400 $nm$ to 2,500 $nm$, resulting in a spectral resolution of 10 $nm$. As is typical in HSI, atmospheric effects and dense water vapor contaminate a number of the wavelengths and so we consider only the remaining $162$ wavelengths. Pixels belong to one of six categories: {\it asphalt, grass, tree, roof, metal,} and {\it dirt}. \Cref{fig:urban-gt} shows the ``ground-truth'' classification. Each trial begins with two initially labeled points per class and then selects 500 active learning points in batches of size $B=5$.
\subsection{Discussion of Method Performance and Comparison}\label{sec:discuss-results}
For an acquisition function to be useful in the active learning process, its accuracy curve should both (1) initially increase rapidly compared to other methods and (2) not level off at a lower final accuracy.
In the binary classification tests, i.e. the sequential and batch tests on the Binary-Clusters dataset (Figures~\ref{fig:acc-bc-1} and~\ref{fig:acc-bc-5}), the {\bf VOPT} and {\bf SOPT} acquisition functions in the {\bf HF} model perform well early on, but level off at a lower overall accuracy. By comparison, the {\bf MC} accuracy curves increase slightly slower in the beginning, but achieve a higher overall accuracy than {\bf VOPT} and {\bf SOPT}. The {\bf DB-RKHS} method performs very well in both tests, but we note this model and associated acquisiton function are more costly to compute than our spectral-truncation models, see \Cref{fig:timing}. All methods perform better than uncertainty sampling ({\bf UNC-LOG}\footnote{We report only {\bf UNC-LOG} as performs best of all the three models {\bf GR, LOG,} and {\bf P}.}) in both tests, which is especially slow to increase the accuracy early on in the sequential test.
\Cref{fig:bc-choices} shows the distribution of active learning choices for a few of the considered methods in the binary sequential test, allowing us a glimpse at the empirical characteristics of each acquisition function's choices. Note that the {\bf VOPT-HF} choices (\ref{fig:bc-vopt-hf}) are nearly spread out evenly among the whole unit square domain, while the {\bf MC-GR}(\ref{fig:bc-mc-gr}), and {\bf DB-RKHS} (\ref{fig:bc-db-rkhs}) choices include points in every square, but also contain a higher concentration of choices along the boundaries between the squares. The {\bf MC-GR} and {\bf DB-RKHS} methods not only {\it explore the extent of the domain of the dataset}, but also {\it exploit classification information} by selecting points along the boundaries between the clusters throughout the whole domain. In contrast, the {\bf VOPT-HF} method selects points evenly spread out over the whole domain, which arguably helps this method to achieve a beneficial increase early on in the active learning process, but does not transition to exploiting known classification information along the decision boundaries. {\bf UNC-LOG}(\ref{fig:bc-unc-log}) in this run chooses points that lie between the long, tall blue cluster on the right and the top-right red cluster, while ignoring various other clusters in the dataset; in a sense, {\bf UNC-LOG} {\it exploits} the known classification information without sufficiently exploring the extent of the dataset's domain.
While the {\bf DB-RKHS} method is very similar in flavor to our MC acquisition function, it is more computationally expensive both in model training as well as in acquisition function evaluation (see \Cref{fig:timing}) because of dense similarity kernel computations\footnote{One could approximate the kernel (e.g.Nystr\"{o}m extension), but we report the original formulation of~\cite{karzand_maximin_2020}}. Likewise, the {\bf HF} model requires a matrix inversion of a large submatrix of the graph Laplacian which is undesirable for scaling to larger problems. By restricting our underlying classifier's to the span of only a subset of the eigenvalues and eigenvectors of the graph Laplacian $L$, we achieve faster model training and acquisition function evaluation. Despite this significant model compression the present work is competitive with these more costly methods and models.
For all the multiclass tests (\Cref{fig:mgr-plots} and~\ref{fig:ce-plots}), the {\bf MC-CE} acquisition function performs the best at both increasing the underlying model's accuracy early on and obtaining the highest accuracy overall in the active learning process. The {\bf CE} classifier initially has {\it lower} accuracy than the {\bf MGR} classifier in the MNIST and Urban tests, but quickly achieves a higher accuracy. While one can remedy the low initial accuracy of the {\bf CE} model in practice by hyperparameter tuning, we set the hyperparameters to be consistent across the shown datasets so as to showcase the efficacy of the acquisition function regardless of hyperparameter tuning. Further, the aim of the active learning process is to iteratively choose subsets of points to improve the underlying classifier and ultimately achieve the highest accuracy under the chosen model, not the design of the optimal underlying classifier in the presence of few labeled data.
We conclude this discussion with a note about the scalability of the acquisition function evaluations. \Cref{fig:timing} shows the average time per active learning iteration to calculate the acquisition function on the candidate set, for increasing dataset size, $N$. Solid lines present timing results for binary models, while dashed linear present timing results for multiclass models. We exclude the {\bf SOPT} results as they are indistinguishable from the {\bf VOPT} results.
The family of binary {\bf VOPT} and {\bf MC} acquisition functions scale similarly, though the {\bf P} (Probit) model's {\bf MC} acquisition function has a significant overhead cost due to the repeated PDF and CDF calculations required for evaluating $F$ and $F'$ (see \Cref{sec:bin-model}). In contrast, the {\bf DB-RKHS} and {\bf VOPT-HF} acquisition functions scale noticeably worse. The multiclass acquisition functions all scale similarly to each other, though the {\bf MC-CE} method has greater overhead cost. The size of the posterior covariance matrix (i.e. inverse Hessian used in look-ahead calculations) for the {\bf MGR} model ($C_{\hat{A}} \in \mathbb R^{M \times M}$) compared to the {\bf CE} model ($\mathcal C_{\hat{\boldsymbol \alpha}} \in \mathbb R^{Mn_c \times Mn_c}$) straightforwardly clarify this disparity.
We conclude that the MC acquisition function adapted to the spectral truncation from of graph-based SSL provides both a scalable and effective method for active learning.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.65\textwidth]{images/timing.jpg}
\caption{Average Active Learning Query Iteration Timing Comparison. Each iteration of the active learning process selects uniformly at random a candidate set comprising 10\% of the unlabeled dataset on which to evaluate the corresponding acquisition function. Times shown are the averaged recorded time to evaluate the acquisition function on the candidate set for increasing overall dataset size, $N$. Solid lines present timing results for binary models, while dashed linear present timing results for multiclass models.}
\label{fig:timing}
\end{figure}
\section{Conclusion}
We present a novel Model-Change (MC) active learning acquisition function along with a general framework that unifies different graph-based semi-supervised learning (SSL) models. Applying the Laplace approximation to the non-Gaussian Bayesian posterior distributions arising from different loss functions in the family of graph-based SSL models of \Cref{tab:models} admits efficient approximations of how the underlying classifier could change as a result of labeling unlabeled points. This framework and associated active learning acquisition function are made more efficient by introducing the ``spectral truncation'' modifications, wherein we use only the lower-lying eigenvalues and corresponding eigenvectors in constraining the graph-based SSL classifiers as well as diminishing the memory requirements of the model. The MC acquisition function shows to be efficient for active learning compared to other methods natural for the graph-based SSL setting.
|
\section{Introduction}
We study entire solutions $u:\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2\to\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2$ of the anisotropic Ginzburg-Landau equation
\begin{align}\label{eq:glani}
\Delta u+\delta \nabla (\dv u) +\delta \curl^*(\curl u)=(|u|^2-1) u,
\end{align}
where $\delta\in (-1,1)$ is a fixed constant and we define $\curl u=\partial_1 u_2 -\partial_2 u_1$ and its adjoint $\curl^*=(\partial_2,-\partial_1)$. This is the Euler-Lagrange equation corresponding to the anisotropic Ginzburg-Landau energy
\begin{align}\label{eq:Fani}
F(u;\Omega)=\int_{\Omega}\frac 12 |\nabla u|^2 + \frac\delta 2 \left( (\dv u)^2 -(\curl u)^2\right) +\frac 14 (1-|u|^2)^2,\quad\Omega\subset\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2.
\end{align}
Equation \eqref{eq:glani} and its associated energy \eqref{eq:Fani}
arise in the description of 2D point defects in some liquid crystal configurations, such as smectic-$C^*$ thin films \cite{phil1} and nematics close to the Fr\'eedericksz transition \cite{clerc14,clerc1,clerc2}. It has also been proposed as a toy model to understand more complex liquid crystal equations \cite{sternberg1}. While simplified isotropic equations (corresponding here to $\delta=0$) are often used to model liquid crystals, it should be stressed that real liquid crystal materials are always anisotropic, and this reduced symmetry gives rise to nontrivial mathematical challenges.
In the isotropic case $\delta=0$, equation \eqref{eq:glani} is the classical Ginzburg-Landau equation
\begin{equation}\label{eq:gliso}
\Delta u=(|u|^2-1) u,
\end{equation}
which has been extensively studied due to its physical applications and its mathematical richness -- see the monographs \cite{bethuel1,pacard,SS} and the references therein.
It is known that for any prescribed degree (or winding number) $d\neq 0$, the isotropic Ginzburg-Landau equation \eqref{eq:gliso} admits a unique solution of the form
\begin{equation}\label{glv}
v_d(re^{i\theta})=\eta_d(r)e^{id\theta},\qquad \eta_d(0)=0,\;\lim_{+\infty}\eta_d=1,\,\eta_d\geq 0,
\end{equation}
where the radial profile $\eta_d$ solves a certain ODE (ordinary differential equation) problem \cite{herve,chen94}.
In the anisotropic case $\delta\neq 0$ however, the only solutions to \eqref{eq:glani} of the form $u=\eta(r)e^{i(d\theta+\theta_0)}$ are of degree $d=1$, and phase shift $\theta_0\equiv 0\mod\pi/2$ \cite{clerc14}. Therefore, in strong contrast with the isotropic case, the problem of finding
entire solutions $u\colon\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2\to\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2$ of \eqref{eq:glani}, such that $|u(x)|\to 1$ and $u$ has a prescribed topological degree $d\neq 1$ as $|x|\to\infty$, cannot be reduced to a one-dimensional ODE.
Our main result is the existence of solutions of any negative prescribed degree $d$ at infinity, provided the anisotropy $\delta$ is small enough.
\begin{theorem}\label{t:main}
For any $d=-1,-2,\ldots$, there exists $\delta_0(d)>0$ such that for small enough anisotropy $|\delta| <\delta_0(d)$,
there are at least two distinct, smooth entire solutions $u\colon\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2\to\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2$ of the anisotropic Ginzburg-Landau equation \eqref{eq:glani} satisfying
\begin{align*}
\int_{\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2}(1-|u|^2)^2 <2\pi d^2\quad\text{ and }\quad\deg(u;\partial D_r)=d\text{ for }r\gg 1.
\end{align*}
\end{theorem}
\begin{remark}\label{r:rad}
These solutions are, for $\delta\neq 0$, not radially symmetric. This is very different from the isotropic case, where the existence of non-radial solutions is a famous open question. As $\delta\to 0$, the two solutions obtained in Theorem~\ref{t:main} converge to the radial solutions $v_d$ and $iv_d$ \eqref{glv} of the classical Ginzburg-Landau equation \eqref{eq:gliso}. Indeed, they converge to solutions with finite potential energy $\int_{\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2}(1-|u|^2)^2 <\infty$, of degree $d$ at infinity, and with the symmetry constraint \eqref{eq:dequiv1} for $n=1-d$. That symmetry constraint forces these solutions to have a zero of degree $D$ at the origin, with $|D|\geq |d|$, and they must therefore be radial thanks to a result of Mironescu \cite[Th\'eor\`eme~2]{mironescu-radial}.
\end{remark}
\begin{remark}\label{r:sym}
The two solutions described in the Theorem are distinct modulo the elementary symmetries of equation \eqref{eq:glani}, which are the following transformations associated to rotation equivariance, reflection equivariance, and translation invariance:
\begin{align*}
&u(z) \longrightarrow \tau e^{-i\alpha}u(e^{i\alpha }z)\quad \alpha\in\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z},\;\tau\in\lbrace \pm 1\rbrace,\\
& u(z)\longrightarrow \overline{u(\bar z)},\\
& u(z)\longrightarrow u(z+a),\quad a\in\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2.
\end{align*}
These transformations preserve the equation \eqref{eq:glani} and its associated energy.
Note that in the isotropic case $\delta=0$ we have invariance under any rotation of the variable $u(e^{i\alpha}z)$ and of the target $e^{i\alpha}u(z)$ separately, but for $\delta\neq 0$ this is only true for $\alpha\equiv 0$ modulo $\pi$.
\end{remark}
The solutions we obtain in Theorem~\ref{t:main} are invariant under a well-chosen finite subgroup of the transformations described in Remark~\ref{r:sym}: this is the key to prescribing the degree at infinity. Specifically, we impose the symmetry constraints
\begin{subequations}\label{eq:dequiv}
\begin{equation}\label{eq:dequiv1}
u(e^{i\frac{\pi}{n}}z)=-e^{i\frac{\pi}{n}}u(z)=e^{i\frac{(1-n)\pi}{n}}u(z),
\end{equation}
for some integer $n\geq 2$,
and in addition
\begin{equation}\label{eq:dequiv2}
u(\bar z)=\pm \overline{u(z)},
\end{equation}
\end{subequations}
to distinguish the two different solutions (depending on the sign $\pm$). Thanks to the symmetric criticality principle \cite{palais} (see also \cite[Proposition~7.1]{lamy14bifur} for a simplified version applicable to the present case), minimizing the energy under these symmetry constraints still provides a solution of \eqref{eq:glani}. Moreover, the equivariance \eqref{eq:dequiv1} imposes a degree constraint: if $u$ is continuous and does not vanish on a fixed circle $\partial D_r$, a continuous choice of phase $\psi\colon \mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}\to\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}$ such that $u/|u|(re^{i\theta})=e^{i\psi(\theta)}$ must satisfy
\begin{align*}
\psi\left(\theta+\frac{\pi}{n}\right)= \psi(\theta) +\frac{(1-n)\pi }{n}+2k_0\pi,
\end{align*}
for some $k_0\in\mathbb Z$ and all $\theta\in\mathbb R$, and iterating this translation yields
$\psi(2\pi)-\psi(0)=2\pi (1-n) + 4 k_0 n\pi $, hence
\begin{align*}
\deg(u;\partial D_r)\equiv (1-n) \mod 2n.
\end{align*}
In the equivalence class $(1-n)+2n \mathbb Z$ the degree with minimal absolute value is $d=1-n$ ($d=-1,-2,\ldots$). This will enable us to conclude that minimizing solutions have degree $d$ at infinity.
\begin{remark}\label{r:dpos}
The construction of solutions of degree $d\geq 2$ is an open problem. Note however that from the point of view of physics, the most relevant degrees are $d=\pm 1$ since vortices of degree $|d|\geq 2$ are expected to be unstable.
\end{remark}
With the above discussion in mind, the proof of Theorem~\ref{t:main} will proceed in two steps. First we obtain in Proposition~\ref{p:entire} entire $d$-equivariant solutions with finite potential energy $\int (1-|u|^2)^2<\pi d^2$, by minimizing the energy in large disks $D_R$ and letting $R\to\infty$. This first step works for any value of $\delta\in (-1,1)$ and provides a solution of degree $d_\infty \equiv d$ modulo $2(1-d)$. The second step, in Proposition~\ref{p:deg}, consists in using the minimizing property to show that $d_\infty$ is in fact equal to $d$. This is where we require $|\delta| <\delta_0(d)$. We note that the value of $\delta_0(d)$, given explicitly in Proposition~\ref{p:deg} can be somewhat improved however whether $\delta_0(d)=1$ or is strictly less than $1$ is an open problem. Theorem~\ref{t:main} follows directly from Propositions~\ref{p:entire} and \ref{p:deg}.
The article is organized as follows. In Section~\ref{s:equiv} we determine the equivariant classes that are compatible with the anisotropic equation \eqref{eq:glani}. In Section~\ref{s:entire} we prove the existence of entire equivariant solutions, while in Section~\ref{s:deg} we characterize their degree for small enough anisotropy. In Section~\ref{s:poho} we derive Pohozaev identity which plays a crucial role in the computations of Section~\ref{s:deg}.
\section{Anisotropic equivariant classes}\label{s:equiv}
Let $G$ and $\Gamma$ be two subgroups of $O(2)$, and let $\mu:G\to \Gamma$ be a group homomorphism. Let also $\Omega\subset \mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2$ be a set invariant by the action of $G$. By definition, a map $u:\Omega \to \mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2$ is $\mu$-equivariant if
\begin{equation}\label{equiv}
u(gx)=\mu(g)u(x), \ \forall x\in\Omega,\ \forall g\in G.
\end{equation}
We refer to \cite{schw} for a discussion of such maps, and to \cite{bates1,bates2} for examples of equivariant solutions.
An equivariant class is compatible with \eqref{eq:glani} (cf. \cite{palais}), if given any $\mu$-equivariant map $u$, the integrand of \eqref{eq:Fani} is invariant by the action of $G$. Clearly, \eqref{equiv} implies that for every $g\in G$, and $x \in \Omega$, we have
\begin{align*}
&|u(gx)|=|u(x)|,
\quad
\nabla u(gx)=\mu(g)\circ (\nabla u (x))\circ g^{-1}, \quad |\nabla u(gx)|=|\nabla u (x)|,
\end{align*}
and, letting $\sigma (x)=-x$ denote the antipodal map,
\begin{align*}
(\dv u)(gx)=\pm (\dv u)(x)\quad &\Longleftrightarrow\quad \mu(g)=g \text{ or } \mu(g)=\sigma \circ g,
\\
(\curl u)(gx)=\pm (\curl u)(x)\quad &\Longleftrightarrow\quad \mu(g)=g \text{ or } \mu(g)=\sigma \circ g.
\end{align*}
Thus, the anisotropic equivariant classes of \eqref{eq:glani} are detemined by the homomorphisms $\mu:G\to\Gamma$ satisfying
\begin{equation}\label{condhom}
\forall g\in G: \ \mu(g)=g \text{ or } \mu(g)=\sigma \circ g.
\end{equation}
For instance, the solution $x\mapsto v_1(x/\sqrt{1+\delta})$ of \eqref{eq:glani} (where $v_1$ is defined in \eqref{glv}), is equivariant with respect to the trivial homomorphism $\iota:O(2)\to O(2)$, $\iota(g)=g$, $\forall g\in O(2)$. On the other hand, the solution $x\mapsto i v_1(x/\sqrt{1-\delta})$ of \eqref{eq:glani} is equivariant with respect to the homomorphism $\tilde \iota:O(2)\to O(2)$, such that $\tilde\iota(\mathbf{s})=\sigma \circ \mathbf{s}$ for every reflection $\mathbf {s}\in O(2)$, and $\tilde\iota(\mathbf{r})=\mathbf{r}$ for every rotation $\mathbf{r}\in O(2)$.
Moreover, we point out that the symmetry constraints introduced in \eqref{eq:dequiv} are equivalent to equivariance with respect to the homomorphisms $\mu_d^\pm$, $d=-1,-2,\ldots$ defined below. For every $n=1-d\geq 2$, we denote by $D_{2n}$ the dihedral group generated by the rotation $\mathbf{r}_{2n}$ of angle $\pi/ n$, and the reflection $\mathbf{s}_0$ with respect to the $x_1$ coordinate axis. Setting
\begin{align*}
\mu_d^+(\mathbf{s}_0)& =\mathbf{s}_0, \quad \mu_d^+(\mathbf{r}_{2n})=\sigma\circ \mathbf{r}_{2n}=\mathbf{r}_{2n}^d,
\\
\mu_d^-(\mathbf{s}_0)&=\sigma\circ\mathbf{s}_0, \quad \mu_d^-(\mathbf{r}_{2n})=\sigma\circ \mathbf{r}_{2n}=\mathbf{r}_{2n}^d,
\end{align*}
we can see that these choices for $\mu_d^\pm(\mathbf{s}_0)$ and $\mu_d^\pm(\mathbf{r}_{2n})$ yield homomorphisms $\mu_d^\pm:D_{2n}\to D_{2n}$, such that $\mu_d^\pm$-equivariance is equivalent to \eqref{eq:dequiv}. Actually, the homomorphisms $\tilde \iota$, and $\mu_d^\pm$ (as well as their restrictions to subgroups), are the only nontrivial homomorphisms satisfying \eqref{condhom}.
\begin{figure}[h]
\includegraphics[width=.8\textwidth]{figd4bis.jpg}
\caption{For $n=2$ and $d=-1$, the images by the homomorphisms $\mu_{-1}^\pm:D_4\to \mu_{-1}^\pm(D_4)=D_4$ of the reflections $\mathbf{s}_i:=\mathbf{r}_{4}^i\mathbf{s}_0$ ($i=0,1,2,3$).}
\label{fig1}
\end{figure}
\begin{figure}[h]
\includegraphics[width=.8\textwidth]{fig2.jpg}
\caption{For $n=3$ and $d=-2$, the images by the homomorphism $\mu_{-2}^+:D_6\to \mu^+_{-2}(D_6)=D_3$ of the reflections $\mathbf{s}_i:=\mathbf{r}_{6}^i\mathbf{s}_0$ ($i=0,1,2,\ldots,5$).}
\label{fig2}
\end{figure}
Examples of $\mu_d^+$-equivariant maps are $u(re^{i\theta})=f(r)e^{iD\theta}$ for any $D\in d +2(1-d)\mathbb Z$ and real-valued $f(r)$. The transformation $u\mapsto iu$ provides a bijection from $\mu_d^+$-equivariant maps onto $\mu_d^-$-equivariant maps. We also notice that given a reflection ${\mathbf s}\in D_{2n}$, the image by a $\mu_d^\pm$-equivariant map of the reflection line of ${\mathbf s}$, is contained in the reflection line of $\mu_d^\pm ({\mathbf s})$. Figures \ref{fig1} and \ref{fig2} show the images by the homomorphisms $\mu_d^\pm$, of the reflections $ \mathbf{s}_i:=\mathbf{r}_{2n}^i\mathbf{s}_0$ ($i=0,1,\ldots,2n-1$), and illustrate why
$\deg(u;\partial D_r)\equiv d \mod 2n$, for a $\mu_d^\pm$-equivariant map $u$ that does not vanish on the circle $\partial D_r$.
\section{Existence of entire equivariant solutions}\label{s:entire}
The anisotropic Ginzburg-Landau equation \eqref{eq:glani} is preserved by the transformation $(\delta,u)\to (-\delta, iu)$, so without loss of generality we will restrict ourselves to nonnegative anisotropy $\delta\in [0,1)$. Moreover, thanks to the two-dimensional identity
$\Delta u= \nabla (\dv u) - \curl^*(\curl u)$, we may rewrite \eqref{eq:glani} as
\begin{equation}\label{eq:glapos}
(1-\delta)\Delta u+2\delta \nabla (\dv u) =\nabla W(u),\qquad W(u)=\frac 14 (1-|u|^2)^2,
\end{equation}
with associated energy functional
\begin{equation}\label{eq:glapose}
E(u,\Omega)=\int_\Omega \left[\frac{(1-\delta)}{2}|\nabla u|^2+\delta (\dv u)^2+W(u)\right], \qquad \delta\in[0,1).
\end{equation}
The energy density in \eqref{eq:glapose} differs from the previous one in \eqref{eq:Fani} by a multiple of the null Lagrangian $2\det(\nabla u)=(\dv u)^2+(\curl u)^2 -|\nabla u|^2$. Note that for all $\delta\in (-1,1)$ the left-hand side of the system \eqref{eq:glapos} is an elliptic operator, for which standard $L^p$ and $C^\alpha$ estimates are available.
In this section we prove the existence of entire $\mu_d^\pm$-equivariant solutions (that is, satisfying the symmetry constraints \eqref{eq:dequiv} for $n=1-d$) of \eqref{eq:glapos}.
\begin{proposition}\label{p:entire}
Let $d=-1,-2,\ldots $ and $\delta\in [0,1)$. There exists a smooth $\mu_d^\pm$-equivariant solution $u\colon\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2\to\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2$ of the anisotropic Ginzburg-Landau equation \eqref{eq:glapos} such that
\begin{align*}
2\int_{\mathbb R^2} W(u) < \pi d^2,
\end{align*}
$|u(x)|\longrightarrow 1$ as $|x|\to +\infty$,
and $u$ is locally minimizing in the sense that
\begin{align*}
E(u,D_R)\leq E(u+\xi,D_R)\quad\text{for any }\mu_d^\pm\text{-equivariant }\xi\in H^1_0(D_R;\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2),
\end{align*}
and any centered open disk $D_R$ of finite radius $R>0$.
\end{proposition}
Proposition~\ref{p:entire} is proved by minimizing the energy \eqref{eq:glapose} among equivariant maps in the disk $D_R$, with well-chosen boundary condition on $\partial D_R$, and letting $R\to\infty$. The boundary condition on $\partial D_R$ is chosen to minimize the energy among $\mathbb S^1$-valued maps. For a map $u\colon\mathbb S^1\to\mathbb S^1$ (identified with its homogeneous radial extension satisfying $\partial_r u=0$), the energy is given by
\begin{align*}
E(u;\mathbb S^1)=\int_{\mathbb S^1} \left[\frac{1-\delta}{2} |\partial_\theta u|^2 + \delta (\partial_\theta u\cdot e_\theta)^2\right]\, d\theta,\qquad u\in H^1(\mathbb S^1;\mathbb S^1).
\end{align*}
Denoting by $H^1_{equ}(\mathbb S^1;\mathbb S^1)$
the class of $H^1$ maps
from $\mathbb S^1$ into itself
that are $\mu_d^\pm$-equivariant,
we define
\begin{align*}
C_{\delta,d}^\pm =\min_{u\in H^1_{equ}(\mathbb S^1;\mathbb S^1)} E(u;\mathbb S^1).
\end{align*}
The direct method of the calculus of variation ensures that the infimum is indeed attained:
\begin{align}\label{eq:zeta}
C_{\delta,d}^\pm=E(\zeta;\mathbb S^1)\quad\text{for some map }\zeta\in H^1_{equ}(\mathbb S^1;\mathbb S^1),
\end{align}
and we will use this map $\zeta$ as a boundary condition on the circle $\partial D_R$.
\begin{remark}\label{r:zeta}
The map $\zeta$ can be locally lifted as $\zeta(\theta)=e^{i\psi(\theta)}$ and the real-valued $H^1$ phase $\psi$ solves the Euler-Lagrange equation
\begin{align*}
\partial_\theta\left[ (1+ \delta\cos(2\theta-2\psi))\partial_\theta\psi \right]=\delta\sin(2\theta-2\psi)(\partial_\theta\psi)^2,
\end{align*}
which ensures that $\psi$ is smooth (see e.g. \cite[Theorem~4.36]{dacorogna}). Hence $\zeta$ is smooth, and we have
\begin{align*}
C_{\delta,d}^+ < E(e^{id\theta},\mathbb S^1)=\pi d^2,\quad C_{\delta,d}^- < E(ie^{id\theta},\mathbb S^1)=\pi d^2,
\end{align*}
because the map $e^{id\theta}$ (resp. $ie^{id\theta}$) is $\mu_d^+$ (resp. $\mu_d^-$) equivariant but does not satisfy the Euler-Lagrange equation.
\end{remark}
\begin{proof}[Proof of Proposition~\ref{p:entire}]
By the direct method of the calculus of variations, for all $R>0$ there exists a $\mu_d^\pm$-equivariant map $u_R$ minimizing $E(\cdot,D_R)$ among all $\mu_d^\pm$-equivariant maps $u\in H^1(D_R;\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2)$ with boundary condition
\begin{align*}
u(Re^{i\theta})=\zeta(\theta)\qquad\forall \theta\in \mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z},
\end{align*}
where $\zeta$ is defined in \eqref{eq:zeta}.
The map $u_R$ is a weak solution of the elliptic system \eqref{eq:glapos}, $\nabla W(u_R)\in L^2(D_R)$ and $\zeta$ is smooth, hence from standard elliptic regularity theory (see e.g. \cite[Theorem~4.14]{GM}) we have that $u\in W^{2,2}(D_R)$.
By Sobolev embedding this implies $u\in C^\alpha(\overline D_R)$ for all $\alpha\in (0,1)$, and bootstrapping Schauder estimates (see e.g. \cite[Theorems~5.20, 5.21]{GM}) one deduces $u\in C^\infty(\overline D_R)$. In particular $u_R$ is regular enough, up to the boundary, to derive Pohozaev's identity (see Section \ref{s:poho})
\begin{align*}
2\int_{D_R} W(u)& = \int_{\mathbb S^1}\left[\frac{1-\delta}{2}|\partial_\theta\zeta|^2 +\delta (\partial_\theta \zeta\cdot e_\theta)^2 \right]
\\&\quad
-R\int_{\partial D_R}\left[
\frac{1+\delta}{2}(\partial_r u\cdot e_r)^2
+
\frac{1-\delta}{2}(\partial_r u \cdot e_\theta)^2\right]\\
& \leq \int_{\mathbb S^1}\left[\frac{1-\delta}{2}|\partial_\theta\zeta|^2 +\delta (\partial_\theta \zeta\cdot e_\theta)^2 \right],
\end{align*}
and therefore
\begin{align}\label{eq:boundW}
2\int_{D_R}W(u_R) \leq C_{\delta,d}^\pm,
\end{align}
by definition \eqref{eq:zeta} of $\zeta$.
Fix $x_0\in\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2$ with $|x_0| < R-2$ and a cut-off function $\eta\in C_c^\infty(D_2(x_0))$ such that $\mathbf 1_{D_1(x_0)}\leq \eta\leq \mathbf 1_{D_2(x_0)}$ and $|\nabla\eta|\leq 2$. Multiplying the equation \eqref{eq:glapos} satisfied by $u_R$ by $\eta^2 u_R$ we obtain
\begin{align*}
(1-\delta)\int_{D_2(x_0)} \eta^2 |\nabla u_R|^2 & \leq (1+3\delta) \int_{D_2(x_0)} \eta |\nabla\eta|\, |u_R|\, |\nabla u_R| \\
&\quad + \int_{D_2(x_0)} \eta^2 \left| |u_R|^2-1 \right| \, |u_R|^2\\
&\leq \frac{1}{2A}(1+3\delta)\int_{D_2(x_0)} \eta^2 |\nabla u_R|^2 + \frac{A}{2}\int_{D_2(x_0)} |\nabla\eta|^2 |u_R|^2 \\
&\quad + \int_{D_2(x_0)} \left( 4 W(u_R) +2 W(u_R)^{\frac 12} \right),
\end{align*}
for any $A>0$. Choosing $A=(1+3\delta)/(1-\delta)$ we deduce
\begin{align*}
\int_{D_2(x_0)} \eta^2 |\nabla u_R|^2&\leq \frac{1+3\delta}{1-\delta}\int_{D_2(x_0)} |\nabla\eta|^2 |u_R|^2 \\
&\quad + \frac{1}{1-\delta}\int_{D_2(x_0)} \left( 4 W(u_R) +2 W(u_R)^{\frac 12} \right)\\
&\leq \frac{1}{1-\delta}\left( C + C \int_{D_R}W(u_R)\right),
\end{align*}
for some absolute constant $C>0$. Thanks to the uniform bound \eqref{eq:boundW} we deduce that $u_R$ is bounded in $H^1(D_1(x_0))$,
and as a consequence $\nabla W(u_R)$ is bounded in $L^2(D_1(x_0))$, uniformly with respect to $R>2$ and $x_0\in D_{R-2}$.
From elliptic $L^2$ estimates \cite[Theorem~4.9]{GM} this implies that $u_R$ is bounded in $W^{2,2}(D_1(x_0))\subset C^{\alpha}(D_1(x_0))$ for any $\alpha\in (0,1)$,
and by Schauder estimates \cite[Theorem~5.20]{GM} in $C^{2,\alpha}(D_1(x_0))$,
again uniformly with respect to $R>2$ and $x_0\in D_{R-2}$.
By Ascoli's theorem and a diagonal argument we may therefore extract a subsequence $R\to +\infty$ such that $u_R$ converges in $C^1_{\mathrm{loc}}(\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2;\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2)$ to a solution $u\in C^\infty(\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2;\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2)$ of \eqref{eq:glapos}. By construction, $u$ is $\mu_d^\pm$-equivariant, and thanks to \eqref{eq:boundW} it satisfies $2\int_{\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2}W(u)\leq C^\pm_{\delta,d}<\pi d^2$ (cf Remark~\ref{r:zeta} for the last inequality).
The finiteness of the potential energy $\int_{\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2}W(u)$, together with the uniform continuity of $u$ (thanks to the above uniform elliptic estimates in $D_1(x_0)$ for any $x_0\in\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2$), implies that $\lim_{|x|\to\infty}|u(x)|=1$. Finally, for any $R'>R>0$ and $\mu_d^\pm$-equivariant $\xi\in H^1(\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2;\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2)$ with support inside $D_R$ we have
\begin{align*}
&E(u;D_R)-E(u+\xi;D_R)\\
&=E(u;D_R)-E(u_{R'};D_R) +E(u_{R'}+\xi;D_R)-E(u+\xi;D_R) \\
&\quad +E(u_{R'};D_R)
-E(u_{R'}+\xi;D_R) \\
&\leq E(u;D_R)-E(u_{R'};D_R) +E(u_{R'}+\xi;D_R)-E(u+\xi;D_R),
\end{align*}
where we used the fact that $u_{R'}$ is minimizing and $\xi=0$ in $D_{R'}\setminus D_R$.
The right-hand side converges to 0 as $R'\to\infty$, since $u_{R'}\to u$ in $C^1(\overline D_R)$. This proves the minimizing property of $u$.
\end{proof}
\section{The degree at infinity is equal to $d$ for small anisotropy}\label{s:deg}
So far we have constructed, for any $\delta\in [0,1)$, an entire solution $u$ of the anisotropic Ginzburg-Landau equation \eqref{eq:glapos} with finite potential energy $\int_{\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2}W(u)<\infty$, whose degree at infinity $d_\infty=\deg(u;\partial D_r)$ for $r\gg 1$, is of the form
\begin{align*}
d_\infty = d + 2(1-d)N\quad\text{ for some }N\in\mathbb Z,
\end{align*}
as explained in the introduction.
Because $d\leq 0$, among all possible values of $d_\infty$ the one with lowest absolute value is precisely $d$. In this section, we use this to conclude that, provided the anisotropy $\delta$ is small enough, $d_\infty$ must in fact be equal to $d$.
\begin{proposition}\label{p:deg}
Let $d=-1,-2,\ldots $, $\delta\in [0,1)$ and $u$ be a $\mu_d^\pm$-equivariant, locally minimizing solution of the anisotropic Ginzburg-Landau equation \eqref{eq:glapos} as described in Proposition \ref{p:entire}.
Then
\begin{align*}
\deg(u;\partial D_r)=d\text{ for }r\gg 1,\text{ if }0\leq \delta < \delta_0(d)=\min\left( \frac{2-2d}{4d^2 +10 -10d},\frac{2}{\sqrt 3}-1\right).
\end{align*}
\end{proposition}
\begin{proof}[Proof of Proposition~\ref{p:deg}]
The proof is a direct combination of Lemma~\ref{l:lower}, Lemma~\ref{l:Elog} and Lemma~\ref{l:upper} below. Specifically Lemma~\ref{l:lower} provides a lower bound
\begin{align*}
\liminf_{R\to\infty}\frac{E(u,D_R)}{\pi\ln R}\geq (1-\delta)d_\infty^2,
\end{align*}
while Lemmas~\ref{l:Elog} and \ref{l:upper} provide, in two steps, the upper bound
\begin{align*}
\liminf_{R\to\infty}\frac{E(u,D_R)}{\pi\ln R}\leq (1+3\delta)(d^2+|d_\infty -d|),
\end{align*}
whenever $0\leq \delta < 2/\sqrt 3 -1$.
Therefore we must have
\begin{align*}
(1-\delta)d_\infty^2\leq (1+3\delta)(d^2+|d_\infty -d|),
\end{align*}
that is,
\begin{align*}
\delta\geq \frac{d_\infty^2-d^2-|d_\infty -d |}{3d^2+3|d_\infty-d|+d_\infty^2}.
\end{align*}
This implies $d_\infty=d$ as soon as
\begin{align*}
\delta < \delta_*(d):=
\inf_{d_\infty\in \left(d + 2(1-d)\mathbb Z\right)\setminus\lbrace d\rbrace }
\frac{d_\infty^2-d^2-|d_\infty -d |}{3d^2+3|d_\infty-d|+d_\infty^2}.
\end{align*}
To compute $\delta_*(d)$ we let $d_\infty=d+x$ and consider the function
\begin{align*}
f(x)&=\frac{(d+x)^2-d^2-|x|}{3d^2+3|x|+(d+x)^2}=\frac{x^2 +2d x -|x|}{x^2 + 2d x + 3|x|+ 4d^2},
\end{align*}
whose derivative satisfies
\begin{align*}
(x^2 + 2d x + 3|x|+ 4d^2)^2f'(x)=
\begin{cases}
4(d+x)(2d^2-d+x) &\text{ if }x>0,\\
4(d+x)(2d^2 +d -x) &\text{ if }x<0.
\end{cases}
\end{align*}
In particular we have $f'(x)\leq 0$ for $x\leq -2(1-d)$ and $f'(x)\geq 0$ for $x\geq 2(1-d)$, so
\begin{align*}
\delta_*(d)&=\inf_{x\in 2(1-d)\mathbb Z\setminus \lbrace 0\rbrace}f(x)=\min(f(-2(1-d)),f(2(1-d))) \\
&=f(2(1-d))=\frac{2-2d}{4d^2+10-10d}.
\end{align*}
We conclude that $d_\infty=d$ whenever $\delta<\delta_0(d)$.
\end{proof}
The rest of this section is dedicated to the proof of Lemmas~~\ref{l:lower}, \ref{l:Elog} and \ref{l:upper}. The first is a classical lower bound simply using the fact that $u$ has degree $d_\infty$ at infinity and the energy is larger than $(1-\delta)$ times the isotropic Ginzburg-Landau energy.
\begin{lemma}\label{l:lower}
Let $\delta\in [0,1)$, $u\colon \mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2\to\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2$ be a smooth map and $R_0>1$ be such that $|u|\geq 1/2$ on $\partial D_r$ and $\deg(u;\partial D_r)=d_\infty$ for all $r>R_0$. Then
\begin{align*}
E(u,D_R) \geq (1-\delta)\pi d_\infty^2\ln R + O(1)\quad\text{as }R\to\infty.
\end{align*}
\end{lemma}
\begin{proof}[Proof of Lemma~\ref{l:lower}]
We follow Jerrard's argument \cite{jerrard99} (see also \cite{sandier98}). For $|x| > R_0$ we may
write $u=\rho v$ with $\rho=|u|$ and $v=u/|u|$, and deduce
\begin{align*}
\frac{1-\delta}{2} |\nabla u|^2 +\delta (\dv u)^2 +W(u) & \geq \frac{1-\delta}{2r^2}|\partial_\theta u|^2 + W(u)\\
&\geq \frac{1-\delta}{2r^2}\rho^2|\partial_\theta v|^2 +\frac{1-\delta}{2r^2}|\partial_\theta\rho|^2 +\frac 14 (1-\rho)^2.
\end{align*}
Letting $h=\max(1-\rho,0)$, we have $\rho^2\geq 1-2h$ and the
above implies
\begin{align*}
\frac{1-\delta}{2} |\nabla u|^2 +\delta (\dv u)^2 +W(u)
&\geq (1-2h)\frac{1-\delta}{2r^2}|\partial_\theta v|^2 +\frac{1-\delta}{2r^2}|\partial_\theta h |^2 +\frac 14 h^2.
\end{align*}
Integrating on $\partial D_r$ for some $r>R_0$ this implies, since $\deg(v)=d_\infty$,
\begin{align}
&\int_{\partial D_r}\left(\frac{1-\delta}{2} |\nabla u|^2 +\delta (\dv u)^2 +W(u) \right)\nonumber\\
&\geq \frac 1r (1-\delta)\pi d_\infty^2 (1-2h_{max} (r) ) + \int_{\partial D_r} \left(\frac{1-\delta}{2r^2}|\partial_\theta h|^2 +\frac 14 h^2\right),\label{eq:lowerbound1}
\end{align}
where $h_{max}(r)=\max_{\partial D_r}h=h(x_{max})$ for some $x_{max}\in\partial D_r$. To bound the last term from below we set
\begin{align*}
\gamma =\frac 1{r^2} \int_{\partial D_r}|\partial_\theta h|^2,
\end{align*}
so that by Morrey's inequality we have
\begin{align*}
h(x)\geq \frac{h_{max}}{2}\quad\forall x\in\partial D_r\text{ with }|x-x_{max}|\leq c\frac{h_{max}^2}{\gamma},
\end{align*}
for some absolute constant $c>0$. We deduce
\begin{align*}
\int_{\partial D_r} \left(\frac{1-\delta}{2r^2}|\partial_\theta h|^2 +\frac 14 h^2\right)
&\geq \frac{1-\delta}{2}\gamma + c\min \left(r,\frac{h_{max}^2}{\gamma}\right)h_{max}^2,
\end{align*}
for a possibly different absolute constant $c>0$.
Since $r\geq R_0\geq 1$, we deduce
\begin{align*}
\int_{\partial D_r} \left(\frac{1-\delta}{2r^2}|\partial_\theta h|^2 +\frac 14 h^2\right)
&\geq \min\left(c\, h_{max}^2,\inf_{\gamma\geq 0}\left\lbrace \frac{1-\delta}{2}\gamma + c\frac{h_{max}^4}{\gamma}\right\rbrace\right) \\
&\geq c \sqrt{1-\delta}h_{max}^2,
\end{align*}
again for a possibly different absolute constant $c>0$.
Coming back to \eqref{eq:lowerbound1} we deduce
\begin{align*}
&\int_{\partial D_r}\left(\frac{1-\delta}{2} |\nabla u|^2 +\delta (\dv u)^2 +W(u) \right)\\
&\geq \frac 1r (1-\delta)\pi d_\infty^2 - 2(1-\delta)\pi d_\infty^2 \frac{h_{max}(r)}{r} + c \sqrt{1-\delta}h_{max}(r)^2\\
&\geq \frac 1r (1-\delta)\pi d_\infty^2
+\inf_{h\geq 0}\left\lbrace c \sqrt{1-\delta}h^2
- 2(1-\delta)\pi d_\infty^2 \frac{h}{r}
\right\rbrace \\
&\geq \frac 1r (1-\delta)\pi d_\infty^2
- c \frac{d_\infty^4}{r^2},
\end{align*}
for a possibly different absolute constant $c>0$.
Integrating over $r\in [R_0,R]$ implies
\begin{align*}
E(u,D_R\setminus D_{R_0})\geq (1-\delta)\pi d_\infty^2 \ln\frac{R}{R_0} - c \frac{d_\infty^4}{R_0}\quad\text{for all } R\geq R_0,
\end{align*}
and proves the lower bound.
\end{proof}
Lemmas~\ref{l:Elog} and \ref{l:upper} are dedicated to obtaining a similar upper bound on the energy of $u$. The first step is to show that the energy cannot increase faster than logarithmically. In the isotropic case $\delta=0$, such logarithmic energy bound is valid for any solution (not necessarily minimizing) with finite potential energy, as follows from the estimates in \cite{bmr}, which however rely very much on the isotropic structure of the equations and seem unapplicable here. Here we show logarithmic energy growth provided $\delta$ is not too large, by constructing appropriate comparison maps and using the minimizing property of our solutions.
\begin{lemma}\label{l:Elog}
Let $d=-1,-2,\ldots $, $\delta\in [0,1)$ and $u$ be a $\mu_d^\pm$-equivariant solution of the anisotropic Ginzburg-Landau equation \eqref{eq:glapos} described in Proposition \ref{p:entire}.
Then
\begin{align*}
\liminf_{R\to\infty} \frac{E(u,D_R)}{\ln R} < \infty,\text{ provided }0\leq\delta < \frac{2}{\sqrt 3}-1.
\end{align*}
\end{lemma}
\begin{proof}[Proof of Lemma~\ref{l:Elog}]
Let
\begin{align*}
f(R)&=R \, E(u,\partial D_R),&
g(R)&=R\int_{\partial D_R} (1-|u|^2)^2,&
\sigma(R)&=\sup_{\partial D_R} \left| 1-|u| \right|.
\end{align*}
There exists $R_0\geq 2$ such that for $R\geq R_0$ we have $\sigma(R)\leq 1/2$ and may write
\begin{align*}
u=\alpha\,\rho\, e^{i(d_\infty\theta +\varphi)}\quad\text{in }\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2\setminus D_{R_0}
\end{align*}
for some real-valued smooth functions $\rho=|u|$, $\varphi$ in $\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2\setminus D_{R_0}$,
and
\begin{align*}
\alpha=\begin{cases}
1 &\text{ if }u\text{ is }\mu_d^+\text{-equivariant},\\
i &\text{ if }u\text{ is }\mu_d^-\text{-equivariant},
\end{cases}
\end{align*}
and $d_\infty= d$ modulo $2(1-d)$. The $\mu_d^\pm$-equivariance of $u$ translates into
\begin{align*}
&\rho\left(e^{i\frac{\pi}{1-d}}z\right)=\rho(\bar z)=\rho(z),\quad \varphi\left(e^{i\frac{\pi}{1-d}}z\right)=\varphi(\bar z)=\varphi(z).
\end{align*}
For all $R>R_0$ we have, with the notation $\partial_\tau =e_\theta\cdot\nabla=r^{-1}\partial_\theta$,
\begin{align*}
& R \int_{\partial D_R} (\partial_\tau\rho)^2 +\left(\frac{d_\infty}{R} + \partial_\tau\varphi\right)^2
\leq \frac{1}{(1-\sigma(R))^2} R\int_{\partial D_R}|\partial_\tau u|^2,
\end{align*}
hence, for any $\eta\in (0,1)$, using $2R^{-1}d_\infty\partial_\tau\varphi\leq \eta^{-1}R^{-2}d_\infty^2 + \eta\,(\partial_\tau\varphi)^2$,
\begin{align*}
& R \int_{\partial D_R} (\partial_\tau\rho)^2 +(1-\eta)\left( \partial_\tau\varphi\right)^2
\leq \frac{1}{(1-\sigma(R))^2} R\int_{\partial D_R}|\partial_\tau u|^2
+2\pi\left(1+\frac 1\eta\right)d_\infty^2.
\end{align*}
From Pohozaev identity (see Section~\ref{s:poho})
\begin{align*}
R\int_{\partial D_R} \frac{1+\delta}{2}(\partial_\tau u\cdot e_\theta)^2 +
\frac{1-\delta}{2}(\partial_\tau u \cdot e_r)^2
&= R\int_{\partial D_R}
\frac{1+\delta}{2}(\partial_r u\cdot e_r)^2
+
\frac{1-\delta}{2}(\partial_r u \cdot e_\theta)^2
\\
& \quad +
2\int_{D_R} W(u) - R\int_{\partial D_R} W(u),
\end{align*}
we deduce
\begin{align*}
R\int_{\partial D_R} |\partial_\tau u|^2 &\leq
\frac{1+\delta}{1- \delta} {R}\int_{\partial D_R} |\partial_r u|^2
+\frac{2}{1-\delta} \int_{\mathbb R^2} W(u),
\end{align*}
and therefore
\begin{align}\label{eq:estimrhovarphifR}
& (1-\eta)(1-\sigma(R))^2 R \int_{\partial D_R} (\partial_\tau\rho)^2 +(\partial_\tau\varphi)^2
\\
& \leq \frac{1+\delta}{2} R\int_{\partial D_R}|\partial_\tau u|^2
+ \frac{1-\delta}{2}\left( \frac{1+\delta}{1- \delta} R\int_{\partial D_R} |\partial_r u|^2
+\frac{2}{1-\delta} \int_{\mathbb R^2} W(u)\right)
+\frac {4\pi}\eta d_\infty^2
\nonumber\\
& = \frac{1+\delta}{2} R\int_{\partial D_R}|\nabla u|^2 +\int_{\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2} W(u)
+\frac {4\pi}\eta d_\infty^2\nonumber\\
&\leq \frac{1+\delta}{1-\delta} f(R) +\int_{\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2} W(u)
+\frac {4\pi}\eta d_\infty^2 .\nonumber
\end{align}
We define real-valued functions $\psi,h$ in $D_R$ solving
\begin{align*}
&\Delta\psi=0\text{ in }D_R,\quad\psi=\varphi\text{ on }\partial D_R,\\
&\Delta h= 0\text{ in }D_R,\quad h=1-\rho\text{ on }\partial D_R.
\end{align*}
From Poisson's formula
\begin{align*}
\psi(re^{i\theta})=\sum_{n\in\mathbb Z} c_n \left(\frac rR\right)^{|n|} e^{in\theta},\quad c_n=\frac{1}{2\pi}\int_0^{2\pi}\varphi(Re^{i\theta})e^{-in\theta}\, d\theta,
\end{align*}
we deduce
\begin{align*}
\int_{D_R}|\nabla \psi|^2 &
=2\pi\sum |n|\, |c_n|^2 \\
&
\leq
2\pi\sum n^2 |c_n|^2
= R\int_{\partial D_R}(\partial_\tau\varphi)^2,
\end{align*}
and similarly
\begin{align*}
\int_{D_R}|\nabla h|^2 &\leq R\int_{\partial D_R} (\partial_\tau\rho)^2,
\\
\int_{D_R}h^2 &\leq \frac R 2 \int_{\partial D_R} (1-\rho )^2\leq 2 g(R).
\end{align*}
Moreover thanks to the maximum principle we have $|h|\leq \sigma(R)$ in $D_R$. And by uniqueness, the harmonic extensions $h,\psi$ have the same equivariance as $\rho,\varphi$.
Therefore we obtain a $\mu_d^\pm$-equivariant map $\tilde u$ in $D_R$ agreeing with $u$ on $\partial D_R$ by setting
\begin{align*}
\tilde u =\alpha\,\min(r,1-h)\, e^{i(d_\infty\theta +\psi)}\quad\text{in }D_R.
\end{align*}
Denoting by $c$ a generic absolute positive constant (whose value may change from line to line) we have
\begin{align*}
\int_{D_R}|\nabla\tilde u|^2 &\leq \frac{c}{\eta} d_\infty^2 \ln R +\int_{D_R} |\nabla h|^2 + (1+\sigma(R))^2(1+\eta) \int_{D_R} |\nabla\psi|^2\\
&\leq \frac{c}{\eta} d_\infty^2 \ln R +(1+\sigma(R))^2(1+\eta) R\int_{\partial D_R} (\partial_\tau\rho)^2 +(\partial_\tau\varphi)^2\\
&\leq \frac{c}{\eta} d_\infty^2 \ln R
+\frac{(1+\sigma(R))^2}{(1-\sigma(R))^2}\frac{1+\eta}{1-\eta}\frac{1+\delta}{1-\delta}f(R)
+\int_{\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2} W(u),
\end{align*}
where we used \eqref{eq:estimrhovarphifR} for the last inequality.
Moreover we have
\begin{align*}
\int_{D_R} W(\tilde u) \leq c + c\int_{D_R}h^2 \leq c +c\, g(R),
\end{align*}
so we deduce
\begin{align*}
E(u;D_R) &\leq
E(\tilde u;D_R)\leq \frac{1+3\delta}{2}\int_{D_R}|\nabla\tilde u|^2 + \int_{D_R} W(\tilde u)
\\
&\leq \frac{1}{2}\frac{(1+\sigma(R))^2}{(1-\sigma(R))^2}\frac{1+\eta}{1-\eta}\frac{(1+\delta)(1+3\delta)}{1-\delta}f(R) +\frac{c}{\eta} d_\infty^2 \ln R +\int_{\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2} W(u)+ c g(R),
\end{align*}
for any $R>R_0$ and $\eta>0$. Setting
\begin{align*}
F(R)=E(u;D_R) +2c \int_{D_R}W(u) =\int_0^R \frac{f(r)+2c g(r)}{r}\, dr,
\end{align*}
we obtain
\begin{align}\label{eq:diffineq}
F(R)&\leq
c_*(R)\, R\, F'(R) +\frac{c}{\eta} d_\infty^2\ln R + 3c\int_{\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2} W(u), \\
c_*(R)& =
\frac{1}{2}\frac{(1+\sigma(R))^2}{(1-\sigma(R))^2}\frac{1+\eta}{1-\eta}\frac{(1+\delta)(1+3\delta)}{1-\delta}.\nonumber
\end{align}
Multiplying by $u$ the equation
\begin{align*}
(1-\delta)\Delta u + 2\delta \nabla (\dv u) +(1-|u|^2)u=0,
\end{align*}
and using that $u,\nabla u\in L^\infty$ and $1-|u|^2\in L^2$ we infer that $R\mapsto E(u;B_R)$ grows at most linearly, and therefore
\begin{align}\label{eq:Flin}
\limsup_{R\to\infty}\frac{F(R)}{R}<\infty.
\end{align}
We claim that this linear growth estimate and the differential inequality \eqref{eq:diffineq} imply that
\begin{align}\label{eq:Flog}
\liminf_{R\to\infty}\frac{F(R)}{\ln R} <\infty,\quad\text{provided }\frac{(1+\delta)(1+3\delta)}{1-\delta}<2.
\end{align}
Assume by contradiction that $\liminf F(R)/\ln R =\infty$ and that $(1+\delta)(1+3\delta)/(1-\delta)<2$. Thanks to the latter we can choose $\eta,R_1>0$ so that $c_*(R)<1$ for all $R>R_1$. And since $\liminf_{R\to\infty} F(R)/\ln R =\infty$ we may further increase $R_1$ and deduce from \eqref{eq:diffineq} the inequality
\begin{align*}
F(R)\leq \hat c RF'(R)\qquad\text{for some }\hat c\in (0,1)\text{ and all }R>R_1,
\end{align*}
But this implies that the function $R\mapsto F(R)/R^{1/\hat c}$ is ultimately nondecreasing and thereby contradicts \eqref{eq:Flin} because $1/\hat c>1$. The conclusion follows from \eqref{eq:Flog} since
$E(u;B_R)\leq F(R)$.
\end{proof}
Thanks to Lemma~\ref{l:Elog} we now have a logarithmic bound on the energy, and we may perform a sharper construction to obtain a more precise upper bound.
\begin{lemma}\label{l:upper}
Let $d=-1,-2,\ldots $ and $u\colon \mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2\to\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2$ be a $\mu_d^\pm$-equivariant map which locally minimizes the energy $E(u,D_R)$ as described in Proposition \eqref{eq:glapose} and such that
\begin{align*}
\liminf_{R\to\infty} \frac{E(u,D_R)}{\ln R} < \infty.
\end{align*}
Then
\begin{align*}
\liminf_{R\to\infty} \frac{E(u,D_R)}{\ln R} \leq (1+3\delta)\pi \left(d^2 + |d_\infty - d| \right),
\end{align*}
where $d_\infty=\deg(u;\partial D_r)$ for large enough $r>0$.
\end{lemma}
\begin{proof}[Proof of Lemma~\ref{l:upper}]
The proof follows an argument by Shafrir \cite{shafrir94}.
By assumption we may fix a finite $C_*>\liminf_{R\to\infty} E(u,D_R)/\ln R$, and there exists a sequence $R_0\leq R_k\to +\infty$ such that
\begin{align*}
E(u;\partial D_{R_k})\leq \frac{C_*}{R_k}.
\end{align*}
On $\partial D_{R_k}$ we have $1/2\leq |u|\leq 3/2$ and $\deg(u;\partial D_{R_k})=d_\infty$, so we can write $u=\alpha \rho_k(\theta) e^{i (d_\infty\theta + \varphi_k)}$ for some real valued $H^1$ functions $\rho_k,\varphi_k$, where we take $\alpha=1$ if $u$ is $\mu_d^+$ equivariant, and $\alpha=i$ if $u$ is $\mu_d^-$ equivariant. Equivariance of $u$ translates into
\begin{align*}
&\rho_k\left(e^{i\frac{\pi}{1-d}}z\right)=\rho_k(\bar z)=\rho_k(z),\quad \varphi_k\left(e^{i\frac{\pi}{1-d}}z\right)=\varphi_k(\bar z)=\varphi_k(z)\quad\forall z\in \partial D_{R_k},
\end{align*}
and we have, denoting $\partial_\tau=e_\theta\cdot\nabla=r^{-1}\partial_\theta$,
\begin{align*}
R_k\int_{\partial D_{R_k}} \left( |\partial_\tau \rho_k|^2 + |\partial_\tau \varphi_k|^2 \right)\lesssim C_* + d_\infty^2 :=C_{**}.
\end{align*}
Since $\varphi_k$ is determined modulo $2\pi$ we may moreover assume the pointwise bound $|\varphi_k|\lesssim 1+C_{**}$.
Define a $\mu_{d}^\pm$-equivariant map $\tilde u=\alpha\tilde\rho e^{i (d_\infty\theta + \tilde\varphi)}$ in $D_{R_k}\setminus D_{R_k/2}$ by setting
\begin{align*}
\tilde\rho & =1 +\frac{\ln(2r/R_k)}{\ln 2} (\rho_k(\theta) -1)
\\
\tilde\varphi &= \frac{\ln(2r/R_k)}{\ln 2}\varphi_k(\theta).
\end{align*}
With that definition, we have $\tilde u=u$ on $\partial D_{R_k}$, $\tilde u=\alpha e^{id_\infty\theta}$ on $\partial D_{R_k/2}$, and
\begin{align*}
E(\tilde u; D_{R_k}\setminus D_{R_k/2})\lesssim 1+ C_{**}.
\end{align*}
For small $\varepsilon >0$, it is possible to construct a map $u_\varepsilon\colon D_{1/2}\to\mathbb R^2$ which is $\mu_d^{\pm}$-equivariant with $u_\varepsilon=\alpha e^{id_\infty\theta}$ on $\partial D_{1/2}$ and such that
\begin{align*}
E_\varepsilon(u_\varepsilon;D_{1/2})&:=\int_{D_{1/2}} \frac{1-\delta}{2} |\nabla u|^2 +\delta (\dv u)^2 +\frac{1}{\varepsilon^2}W(u) \\
&\leq (1+3\delta)\pi\left(d^2 + |d_\infty-d| \right)\ln\frac 1\varepsilon + O(1)
\end{align*}
as $\varepsilon\to 0$.
To give a sketch of $u_\varepsilon$'s construction, let $N\geq 0$ be the unique nonnegative integer such that
\begin{align*}
d_\infty - d = 2(1-d)\sigma N,\qquad\sigma\in\lbrace \pm 1\rbrace.
\end{align*}
Then fix $N$ circles with radii $\lambda_j=\frac{j}{4N}$, $j=1,\ldots,N$, and construct $u_\varepsilon$ by putting one vortex of degree $d$ at the origin, and $2(1-d)$ vortices of degree $\sigma$ equally spaced along each circle $\partial D_{\lambda_j}$, in a way that respects the $\mu_d^{\pm}$-equivariance. Such configuration will satisfy the claimed energy bound on $u_\varepsilon$. We perform the explicit construction in Lemma~\ref{l:construction} in the $\mu_d^+$-equivariant case and for $\sigma=+1$, the other cases being similar.
Finally, setting $\tilde u(x)=u_\varepsilon(\varepsilon x)$ for $|x|\leq R_k/2$ and $\varepsilon=1/R_k$, and combining this with the construction of $\tilde u$ in $D_{R_k}\setminus D_{R_k/2}$, we obtain a $\mu_d^{\pm}$-equivariant map $\tilde u$ in $D_{R_k}$ such that $\tilde u=u$ on $\partial D_{R_k}$ and
\begin{align*}
\frac{1}{\pi\ln R_k}E(\tilde u; D_{R_k})\leq (1+3\delta) \left(d^2 +|d_\infty-d| \right)+ o(1),
\end{align*}
and the minimizing property of $u$ implies the same bound on $E(u;D_{R_k})$.
\end{proof}
\begin{lemma}\label{l:construction}
For any integers $d\leq -1$, $N\geq 0$ and $D=d+2(1-d)N$,
there exist $\varepsilon_0=\varepsilon_0(d,N)>0$ and, for $0<\varepsilon<\varepsilon_0$,
a $\mu_d^+$-equivariant map $u_\varepsilon\colon D_{1/2}\to \mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2$ such that $u((1/2)e^{i\theta})=e^{iD\theta}$ for all $\theta\in \mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}$, and
\begin{align*}
\int_{D_{1/2}}\left(\frac 12 |\nabla u_\varepsilon|^2 +\frac {1}{\varepsilon^2}W(u_\varepsilon)\right) \leq
\pi\left(d^2 + |D-d| \right)\ln\frac 1\varepsilon + C,
\end{align*}
for some constant $C$ depending only on $d,N$.
\end{lemma}
\begin{proof}[Proof of Lemma~\ref{l:construction}]
If $N=0$, i.e. $D=d$, we may simply take $u_\varepsilon(re^{i\theta})=\min(2r/\varepsilon,1)e^{id\theta}$, hence we assume $N\geq 1$.
For $j=1,\ldots, N$ we define radii $\lambda_j=j/4N$, and will construct $u_\varepsilon$ with a vortex of degree $d$ at the origin, and $2(1-d)$ vortices of degree 1 equally spaced along each circle $\partial D_{\lambda_j}$. These degree 1 vortices are centered at
\begin{align*}
x_{j,k}=e^{ik\frac{\pi}{1-d}}\lambda_j,\qquad k=0,\ldots,2(1-d)-1,\; j=1,\ldots,N.
\end{align*}
We fix a small radius $\rho=\rho(d,N)>0$ such that the disks
$D_{2\rho}(0)$, $D_{2\rho}(x_{j,k})$,
are mutually disjoint. In the set
\begin{align*}
\omega:=D_{\rho}(0)\cup \bigcup_{j,k} D_{\rho}(x_{j,k}),
\end{align*}
we define $u_\varepsilon$ by setting
\begin{align*}
u_\varepsilon(re^{i\theta})&=\min(r/(\rho\varepsilon),1)\,e^{id\theta},\\
u_\varepsilon(x_{j,k}+re^{i\theta})&=(-1)^k \min(r/(\rho\varepsilon),1)\, e^{i\theta},
\end{align*}
for all $r\in [0,\rho)$. Then $u_\varepsilon$ is $\mu_d^+$-equivariant in $\omega$, and
\begin{align*}
\int_{\omega}\left(\frac 12 |\nabla u_\varepsilon|^2 +\frac {1}{\varepsilon^2}W(u_\varepsilon)\right) \leq
\pi\left(d^2 + |D-d| \right)\ln\frac \rho\varepsilon + C(d,N).
\end{align*}
Since the trace of $u$ on $\partial\omega$ is smooth and $\mathbb S^1$-valued with
\begin{align*}
\deg(u_\varepsilon,\partial D_{\rho}(0))+\sum_{j,k}\deg(u_\varepsilon,\partial D_{\rho}(x_{j,k}))=d+2(1-d)N=D,
\end{align*}
there exists a smooth map $u_*\colon D_{1/2}\setminus\omega\to\mathbb S^1$ such that $u_*=u_\varepsilon$ on $\partial\omega$ and $u_*((1/2)e^{i\theta})=e^{iD\theta}$. Moreover, choosing this map $u_*$ minimizing the Dirichlet energy among $\mathbb S^1$-valued maps with these boundary conditions ensures that $u_*$ is $\mu_d^+$-equivariant, because the boundary conditions are equivariant and the minimizer is unique \cite[\S~I.2]{bethuel1}. Hence, setting $u_\varepsilon =u_*$ in $D_{1/2}\setminus\omega$ we obtain a $\mu_d^+$-equivariant map satisfying the conclusion of Lemma~\ref{l:construction}.
\end{proof}
\section{The stress-energy tensor and Pohozaev identities for anisotropic gradient systems}\label{s:poho}
Let $\Omega\subset\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2$ be an open set. We consider the system
\begin{equation}\label{anisys}
\Delta u+\delta \nabla (\dv u) +\delta \curl^*(\curl u)=\nabla W(u), \ u=(u_1,u_2)\in C^2(\Omega;\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2),
\end{equation}
with $W \in C^1(\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2,\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z})$, and $\delta\in\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}$. Equation \eqref{anisys} can be written as a divergence-free condition, that is,
\begin{equation}\label{divergence free}
\dv T = (\nabla u)^\top \big( \Delta u+\delta \nabla (\dv u) +\delta \curl^*(\curl u)-\nabla W(u) \big)=0
\end{equation}
for the stress-energy tensor
\begin{equation*}
T(u,\nabla u)=
\left( \begin{array}{c}
T_{11}~~ T_{12} \\
T_{21} ~~ T_{22} \\
\end{array} \right),
\end{equation*}
with
\begin{align*}
T_{11}(u,\nabla u)&
=\frac{1}{2}|\partial_1 u|^2-\frac{1}{2}|\partial_2 u|^2+\frac{\delta}{2}|\nabla u_1|^2-\frac{\delta}{2}|\nabla u_2|^2-W(u),\\
T_{12}(u,\nabla u)&
=\partial_1 u\cdot\partial_2 u+\delta \nabla u_1\cdot \nabla u_2+\delta (\dv u)(\curl u),
\\
T_{21}(u,\nabla u)&=\partial _1 u\cdot\partial_2 u+\delta \nabla u_1\cdot \nabla u_2-\delta (\dv u)(\curl u)
\\
T_{22}(u,\nabla u)&=\frac{1}{2}|\partial_2 u|^2-\frac{1}{2}|\partial_1 u|^2+\frac{\delta}{2}|\nabla u_2|^2-\frac{\delta}{2}|\nabla u_1|^2-W(u),
\end{align*}
where $\cdot$ stands for the Euclidean inner product. Condition (\ref{divergence free}) is an algebraic fact that follows from a long, but otherwise not difficult computation. In the isotropic case where $\delta=0$, the stress-energy tensor reduces to the symmetric matrix (cf. \cite[Section~3.1]{book} or \cite[Section~5.1]{SS}):
\begin{equation*}
T(u,\nabla u)=\frac{1}{2}
\left( \begin{array}{cc}
|\partial_1 u|^2 - |\partial_2 u|^2 -2 W(u) & 2 \partial_1 u \!\cdot \partial_2 u \\
2 \partial_1 u \!\cdot\partial_2 u & |\partial_2 u|^2 - |\partial_1 u|^2- 2 W(u) \\
\end{array} \right).
\end{equation*}
Proceeding as in \cite[section 3.6]{book}, we are going to derive Pohozaev identities for \eqref{anisys}, with the help of the stress-energy tensor.
\begin{proposition}
Let $\Omega\subset\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2$ be an open set, let $D_R(x_0)$ be an open disc of radius $r$ centered at $x_0$, such that
$\overline{D_r}(x_0)\subset \Omega$, and let $u\in C^2(\Omega;\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2)$ be a solution of \eqref{anisys}, with $W\in C^1(\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z}^2;\mathbb{R}}\newcommand{\C}{\mathbb{C}}\newcommand{\Z}{\mathbb{Z})$. For every $x=x_0+re^{i\theta}$, we denote by $(e_r,e_\theta)$ the orthonormal basis
$(e^{i\theta},ie^{i\theta})$, and by $\partial_r$ (resp. $\partial_\tau$), the partial derivatives with respect to the vector $e_r$ (resp. $e_\theta$).
Then, we have
\begin{align}\label{poz1}
2\int_{D_r(x_0)}W(u)&=r\int_{\partial D_r}\Big(W(u)
+\frac{1+\delta}{2}
(\partial_\tau u\cdot e_\theta)^2
+\frac{1-\delta}{2}(\partial_\tau u\cdot e_r)^2\Big) \\
&\quad
- r\int_{\partial D_r}\Big(
\frac{1+\delta}{2}(\partial_ r u\cdot e_r)^2
+\frac{1-\delta}{2}(\partial_r u\cdot e_\theta)^2\Big),\nonumber
\end{align}
and
\begin{align}\label{poz2}
2\delta\int_{D_r(x_0)}(\dv u)(\curl u)&=r\int_{\partial D_r}\Big((\delta \dv u) (\curl u)\\
&\hspace{5em}-\delta (\nabla (u \cdot e_r))\cdot(\nabla( u\cdot e_\theta))
-\partial_r u\cdot \partial_\tau u\Big).\nonumber
\end{align}
\end{proposition}
\begin{proof}
Without loss of generality we take $x_0=0$. We derive \eqref{poz1} (resp. \eqref{poz2}), by applying the divergence theorem to the vector fields $X=(x_1T_{11}+x_2T_{21},x_1T_{12}+x_2T_{22})$ (resp. $Y=(-x_2T_{11}+x_1T_{21},-x_2T_{12}+x_1T_{22})$ in the disc $D_r(x_0)$, where we have set $x=(x_1,x_2)$. In view of \eqref{divergence free}, one can check that $\dv X=T_{11}+T_{22}=-2W(u)$, while $\dv Y=-T_{12}+T_{21}=-2\delta (\dv u)(\curl u)$. Moreover a long but otherwise not difficult computation gives:
\begin{align*}
X\cdot x&=-r^2\Big(W(u)+\frac{1+\delta}{2}
(\partial_\tau u\cdot e_\theta)^2+\frac{1-\delta}{2}(\partial_\tau u\cdot e_r)^2
\\
&\hspace{7em }
-\frac{1+\delta}{2}(\partial_ r u\cdot e_r)^2-\frac{1-\delta}{2}(\partial_r u\cdot e_\theta)^2\Big),
\end{align*}
and
\begin{align*}
Y\cdot x=-r^2 \Big(\delta (\dv u) (\curl u)-\delta (\nabla (u \cdot e_r))\cdot(\nabla( u\cdot e_\theta))-\partial_r u\cdot \partial_\tau u\Big).
\end{align*}
\end{proof}
\section*{Acknowledgments}
\bibliographystyle{acm}
|
\section{Introduction} \label{sec:introduction}
\begin{purpose}
Intro to parametric MOR and non-intrusive learning.
\end{purpose}
Model reduction seeks to alleviate the computational burden of large-scale numerical simulations of dynamical systems by constructing reduced-order models (ROMs) that accurately capture the system dynamics, but which are much less expensive to solve than the high-fidelity models inherent in applications. The challenge is to generate ROMs from limited training data that respond well to changes in the scenario parameters that define the governing dynamics \cite{BGW2015pmorSurvery}. Such parametric ROMs are critical for enabling outer-loop applications such as design, inverse problems, optimization, and uncertainty quantification.
Furthermore, as high-fidelity simulations become increasingly sophisticated and simulation data becomes more available, there is a growing need for non-intrusive model reduction methods, which aim to learn ROMs primarily from simulation data and/or outputs, as opposed to making a direct reduction of the underlying high-fidelity code that produced them \cite{GW2021learning}. Non-intrusive approaches combine data-driven learning with physics-based modeling in a way that enables both flexibility and robustness. This paper presents a framework for learning parametric ROMs in a non-intrusive fashion.
\begin{purpose}
Overview of some existing approaches to parametric model reduction.
\end{purpose}
Adapting non-intrusive model reduction strategies to the parametric setting is an active area of research.
One major model reduction strategy, Dynamic mode decomposition (DMD) \cite{brunton2015compressed,schmid2010dynamic}, learns a low-dimensional linear mapping based on state space data, approximating the eigenstates of the infinite-dimensional Koopman operator. The work in \cite{tezzele2020enhancing} targets parametric problems by incorporating DMD with an active subspace strategy to reduce the dimensionality of the parameter space.
Methods based on the Loewner framework \cite{antoulas1986rationalInterp,antoulas2016loewnerbilinear}, another common non-intrusive approach to model reduction, build a ROM based on input-output measurements and transfer functions.
Loewner methods have been generalized to parameterized linear systems by introducing additional degrees of freedom in the construction of the reduced-order transfer function to account for parametric dependencies \cite{Ionita_Antoulas_2014}.
Recent work in \cite{Stuart2020neuralPDE} blends ideas from classical model reduction with deep learning to construct a mapping from parametric inputs to state outputs. Deep learning approaches to model reduction aim to benefit from the flexibility of representing the state on a low-dimensional but nonlinear manifold \cite{lee2020model}.
\begin{purpose}
Review equation discovery approaches, though our objective differs.
\end{purpose}
Equation discovery methods, in which the governing equations of a dynamical system are learned from data,
share some characteristics with non-intrusive model reduction methods.
One class of equation discovery approaches uses sparse regression to identify the underlying PDEs from a set of data \cite{schaeffer2017learning} or the key terms in a dynamical system within a library of potential nonlinear terms \cite{BPK2016sindy,MMKB2018sindyabrupt,rudy2019data}.
In a similar vein, the work in \cite{arbabi2020linking,lee2020coarse} combines machine learning techniques with manifold learning algorithms to
learn a macroscopic model for long-wavelength behavior corresponding to fine-scale measurement data.
Each of these methods relies on an appropriate candidate library for the terms of the unknown equations, selecting the best combination based on high-dimensional data.
In contrast, model reduction approaches seek a low-dimensional representation of a system, hence the learning is typically done in a reduced space.
\begin{purpose}
Introduce OpInf and compare/contrast with the aforementioned methods.
\end{purpose}
Operator Inference (OpInf), introduced in \cite{PW2016operatorInference}, is a non-intrusive framework for model reduction of systems with polynomial nonlinearities. As with other non-intrusive approaches, the method does not require intrusive access to source code, instead inferring the ROM solely from initial conditions, simulation snapshots, and corresponding inputs.
Known governing equations motivate the form of the ROM, and the operators defining the ROM are chosen by minimizing a data-driven residual in a reduced state space. The associated learning problem is linear, dense, relatively small, and has a closed-form solution.
\begin{purpose}
Review OpInf literature and state the main contributions of the paper.
\end{purpose}
Since its introduction, the OpInf framework has been expanded in several ways:
transforming variables (lifting) to induce the requisite polynomial structure \cite{QKPW2020liftAndLearn,SKHW2020romCombustion};
approximating nonpolynomial nonlinearities via the discrete empirical interpolation method (DEIM) \cite{BGKPW2020opInfDeim};
regularizing the learning problem to enable performance on large-scale systems \cite{jain2021performance,MHW2021regOpInfCombustion};
re-projecting trajectories to exactly recover intrusive ROMs \cite{Peherstorfer2020reprojection};
and accounting for algebraic equations arising from lifting transformations \cite{KW2021_DAE}, to name a few.
In these OpInf-based methods, parametric dependencies are addressed by learning separate ROMs for individual parameter samples, then interpolating either their reduced operators \cite{Peherstorfer2020reprojection,PW2016operatorInference} or their outputs \cite{KW2021_DAE}. In this paper, we show that the parametric structure of the governing equations of interest can be built directly into the OpInf regression problem, circumventing the need for interpolation, if the parametric dependencies have an affine form.
Our contribution enables ROMs for affine parametric systems with multiple parameters since interpolation in more than one or two parameter dimensions quickly becomes challenging.
Affine parametric problems have been studied frequently in the context of reduced basis methods \cite{Grepl05,RozHP08,Veroy05,Veroy03}, in which the preservation of the affine parametric structure by projection plays a key role \cite{BGW2015pmorSurvery}. The parametric structure preservation was also exploited in a recent OpInf work applied to the shallow water equations in non-traditional form \cite{YGBK2021}. The approach presented here includes the following key contributions: 1) formulating a general framework for parametric OpInf in a time-continuous setting, 2) implementing a robust regularization strategy in the learning problem, and 3) establishing a priori conditions for determining the well-posedness of the inference procedure.
The remainder of the paper is organized as follows.
\Cref{sec:methodology} establishes the general methodology; \Cref{sec:pde_systems} extends the framework to systems of PDEs; \Cref{sec:computation} details the computational aspects of solving the parametric OpInf problem; \Cref{sec:numerics} presents two numerical examples; and \Cref{sec:conclusions} concludes the paper.
\section{Non-intrusive Parametric Model Reduction} \label{sec:methodology}
In \cref{subsec:pde_to_ode}, we show how the form of an appropriate ROM can be determined directly from the form of certain PDEs; \cref{subsec:operator_inference} presents the OpInf approach for learning such ROMs from data and system structure.
We also introduce a heat equation example for which numerical results are reported in \cref{subsec:numerical_heat}.
\subsection{Projection-based Reduced-order Models of Parametric PDEs} \label{subsec:pde_to_ode}
We target systems governed by parametric PDEs that are polynomial in state, which includes linear PDEs as well as a large class of nonlinear PDEs.
Let $\Omega \subset \mathbb{R}^{d_x}$ be an open, bounded set with Lipschitz continuous boundary $\partial\Omega = \Gamma \cup (\partial\Omega \setminus \Gamma)$ and outward-pointing normal $\boldsymbol{\eta}\in\mathbb{R}^{d_{x}}$.
For the time domain $[t_0,t_f]\subset\mathbb{R}$ and the parameter domain $\mathcal{P}\subset\mathbb{R}^{d_\mu}$, we consider the initial/boundary-value problem
\begin{subequations}
\begin{align}
\label{eq:pde-single}
\frac{\partial u}{\partial t}
&= \mathcal{F}(u;\mu),
&
x &\in \Omega,\
t \in (t_0,t_f],\
\mu \in \mathcal{P},
\\
\label{eq:pde-initial}
u(x, t_0; \mu) &= u_0(x;\mu),
&
x &\in \Omega,\
\mu \in \mathcal{P},
\\
\label{eq:pde-bcs}
u(x, t; \mu) &= 0,
&
x &\in \Gamma,\
t\in[t_0,t_f],\
\mu \in \mathcal{P},
\\
\label{eq:pde-bcs-neumann}
\boldsymbol{\eta}\cdot\nabla_x u(x, t; \mu) &= 0,
&
x &\in \partial\Omega \setminus \Gamma,\
t\in[t_0,t_f],\
\mu \in \mathcal{P},
\end{align}
\end{subequations}
where the unknown state variable $u(\cdot,t;\mu)$ is contained in a separable Hilbert space $\mathcal{V}$ of real-valued functions satisfying the boundary conditions \cref{eq:pde-bcs}--\cref{eq:pde-bcs-neumann} with dual space $\mathcal{V}^{*}$,
and $\mathcal{F}:\mathcal{V}\times\mathcal{P}\to\mathcal{V}^{*}$ is a spatial differential operator that depends on the free parameter $\mu\in\mathcal{P}$.
We assume \cref{eq:pde-single}--\cref{eq:pde-bcs-neumann} has a unique solution $u$ in at least the weak sense, meaning
\begin{align}
\label{eq:weak-solution}
\left\langle v, \frac{\partial u}{\partial t} \right\rangle
= \left\langle v, \mathcal{F}(u;\mu) \right\rangle
\qquad
\textrm{for all}
\qquad
v\in\mathcal{V},\
t\in[t_0,t_f],\
\mu\in\mathcal{P},
\end{align}
where $\langle\cdot,\cdot\rangle$ is the duality pairing of $\mathcal{V}$ with $\mathcal{V}^{*}$.
\begin{purpose}
Problem structure (polynomial in state / affine in parameters).
\end{purpose}
We consider the setting in which $\mathcal{F}$ has a polynomial structure with respect to the state $u$ and its spatial derivatives. Many PDEs enjoy this structure or can be written in this form through a change of variables \cite{QKPW2020liftAndLearn}.
For brevity we consider a quadratic form, but higher-order (e.g., cubic) terms may also be included, as we will see later in \Cref{example:fh-n}.
Specifically, suppose
\begin{subequations}
\begin{align}
\label{eq:pde-polynomial}
\mathcal{F}(u;\mu)
= \mathcal{C}(\mu)
+ \mathcal{A}(u;\mu)
+ \mathcal{H}(u,u;\mu),
\end{align}
where $\mathcal{C}:\mathcal{P}\to\mathcal{V}^{*}$, and where
$\mathcal{A}:\mathcal{V}\times\mathcal{P}\to\mathcal{V}^{*}$ and
$\mathcal{H}:\mathcal{V}\times\mathcal{V}\times\mathcal{P}\to\mathcal{V}^{*}$
are linear in each of their state arguments.
Furthermore, assume that the operators $\mathcal{C}$, $\mathcal{A}$, and $\mathcal{H}$ exhibit the following affine decompositions with respect to the parameter $\mu$:
\begin{gather}
\label{eq:affine-op}
\begin{gathered}
\mathcal{C}(\mu) = \sum_{p=1}^{q_c}\theta_{c}^{(p)}(\mu)\mathcal{C}^{(p)},
\qquad
\mathcal{A}(u;\mu) = \sum_{p=1}^{q_A}\theta_{A}^{(p)}(\mu)\mathcal{A}^{(p)}(u),
\\
\mathcal{H}(u,v;\mu)
= \sum_{p=1}^{q_H}\theta_{H}^{(p)}(\mu)\mathcal{H}^{(p)}(u,v),
\end{gathered}
\end{gather}
\end{subequations}
where
$\mathcal{C}^{(p)}\in\mathcal{V}^{*}$,
$\mathcal{A}^{(p)}:\mathcal{V}\to\mathcal{V}^{*}$,
$\mathcal{H}^{(p)}:\mathcal{V}\times\mathcal{V}\to\mathcal{V}^{*}$,
and the scalar-valued functions $\theta_{c}^{(p)},\theta_{A}^{(p)},\theta_{H}^{(p)}:\mathcal{P}\to\mathbb{R}$ are such that the sets $\{\theta_{c}^{(p)}\}_{p=1}^{q_c}$, $\{\theta_{A}^{(p)}\}_{p=1}^{q_A}$, and $\{\theta_{H}^{(p)}\}_{p=1}^{q_{H}}$ are each linearly independent. Note that the operators $\mathcal{C}^{(p)}$, $\mathcal{A}^{(p)}$,and $\mathcal{H}^{(p)}$ are independent of the parameter $\mu$.
The affine structure either occurs naturally in the PDE or can be created via approximation with the empirical interpolation method \cite{BMNP2004eim}.
\begin{purpose}
Projection preserves the polynomial/affine-parametric structure of the PDE.
\end{purpose}
A projection-based ROM of \cref{eq:pde-single}--\cref{eq:pde-bcs-neumann} with $\mathcal{F}$ as in \cref{eq:pde-polynomial}--\cref{eq:affine-op} retains the affine-parametric polynomial structure of the system \cite{BGW2015pmorSurvery,GPT1999vortexsheddingPOD}.
Let $\{v_{j}\}_{j=1}^{\infty}\subset\mathcal{V}$ be an orthonormal set
such that the solution $u$ may be expressed with the expansion
\begin{align}
\label{eq:basis-expansion}
u(x,t;\mu)
= \sum_{j=1}^{\infty} \hat{u}_{j}(t;\mu)v_{j}(x).
\end{align}
Since $\langle v_i, v_j \rangle = \delta_{ij}$, the coefficients satisfy $\hat{u}_{j}(t;\mu) = \left\langle v_j, u(\cdot,t;\mu) \right\rangle$.
A reduced model with $r \in \mathbb{N}$ degrees of freedom consists of time evolution equations for the coefficients $\hat{u}_{1}(t;\mu),\ldots,\hat{u}_{r}(t;\mu)$; the approximate ROM solution $\breve{u}$ of \cref{eq:pde-single}--\cref{eq:pde-bcs-neumann} is then given by the sum \cref{eq:basis-expansion}, truncated to $r$ terms:
\begin{align}
\label{eq:truncation-projection}
\breve{u}(x,t;\mu)
= \sum_{j=1}^{r} \hat{u}_{j}(t;\mu)v_{j}(x).
\end{align}
Note that $\breve{u}$ is confined to the finite-dimensional subspace
$\textrm{span}(\{v_1,\ldots,v_r\})\subset\mathcal{V}$.
By substituting $\breve{u}$ for $u$ in \cref{eq:weak-solution} with test function $v=v_{i}$, and using the form of $\mathcal{F}$ from \cref{eq:pde-polynomial}--\cref{eq:affine-op},
we obtain
\begin{align}
\label{eq:inner_product}
\begin{aligned}
\frac{\textup{d} \hat{u}_{i}}{\textup{d} t}
= \sum_{p=1}^{q_{c}}\theta_{c}^{(p)}(\mu)
\left\langle v_i, \mathcal{C}^{(p)}(\mu) \right\rangle
&+ \sum_{p=1}^{q_{A}}\theta_{A}^{(p)}(\mu)\sum_{j=1}^{r}
\left\langle v_i, \mathcal{A}^{(p)}\left(v_{j};\mu\right)\right\rangle
\hat{u}_{j}
\\
&+ \sum_{p=1}^{q_{H}}\theta_{H}^{(p)}(\mu)\sum_{j=1}^{r}\sum_{k=1}^{r}
\left\langle v_i,\mathcal{H}^{(p)}\left(v_{j},v_{k};\mu\right)\right\rangle
\hat{u}_{j} \hat{u}_{k}.
\end{aligned}
\end{align}
Collecting \cref{eq:inner_product} for $i=1,\ldots,r$ yields a uniquely-defined system of ordinary differential equations (ODEs) with state vector $
\widehat{\mathbf{u}}(t;\mu)
= [
\hat{u}_{1}(t;\mu)~\cdots~\hat{u}_{r}(t;\mu)
]^\top
\in \mathbb{R}^{r},
$
\begin{subequations}
\begin{align}
\label{eq:ode-affine}
\begin{aligned}
\frac{\textup{d}}{\textup{d}t}\widehat{\mathbf{u}}(t;\mu)
= \left(\sum_{p=1}^{q_c}\theta_{c}^{(p)}(\mu)\widehat{\mathbf{c}}^{(p)}\right)
&+ \left(\sum_{p=1}^{q_A}\theta_{A}^{(p)}(\mu)\widehat{\mathbf{A}}^{(p)}\right)\widehat{\mathbf{u}}(t;\mu)
\\
&+ \left(\sum_{p=1}^{q_H}\theta_{H}^{(p)}(\mu)\widehat{\mathbf{H}}^{(p)}\right)\left(\widehat{\mathbf{u}}(t;\mu)\, \widehat{\odot}\, \widehat{\mathbf{u}}(t;\mu)\right),
\end{aligned}
\\
\label{eq:ode-initial-condition}
\widehat{\mathbf{u}}(t_0;\mu)
= \left[\begin{array}{ccc}
\left\langle v_{1}, u_0(\mu) \right\rangle & \cdots & \left\langle v_{r}, u_0(\mu) \right\rangle
\end{array}\right]^\top \in \mathbb{R}^{r},
\end{align}
\end{subequations}
where
$\widehat{\mathbf{c}}^{(p)}\in\mathbb{R}^{r}$,
$\widehat{\mathbf{A}}^{(p)}\in\mathbb{R}^{r \times r}$,
$\widehat{\mathbf{H}}^{(p)}\in\mathbb{R}^{r \times \binom{r+1}{2}}$,
and $\widehat{\odot}$ denotes a compact Khatri-Rao product, i.e., $\mathbf{u}\,\widehat{\odot}\,\mathbf{u}$ extracts the unique components of the Khatri-Rao product of $\mathbf{u}$ with itself
(see \cref{appendix:kronecker}).
Note that \cref{eq:ode-affine} and \cref{eq:pde-polynomial}--\cref{eq:affine-op} are both quadratic in their respective state and affine-parametric with respect to $\mu$, with the same functions $\theta_{c}^{(p)}$, $\theta_{A}^{(p)}$, and $\theta_{H}^{(p)}$ in the affine expansions.
\begin{purpose}
The structure is the same independent of the basis.
\end{purpose}
The system \cref{eq:ode-affine}--\cref{eq:ode-initial-condition} is a ROM for \cref{eq:pde-single}--\cref{eq:pde-bcs-neumann} in which the boundary conditions are embedded through the basis functions. The quality of the ROM as a surrogate for \cref{eq:pde-single}--\cref{eq:pde-bcs-neumann} depends heavily on the basis $\{v_{j}\}_{j=1}^{r}$, but the \emph{form} of the equations is the same for any choice of basis. In \cref{subsec:operator_inference}, we leverage this property to develop a procedure for learning such a model without explicitly evaluating the terms in \cref{eq:inner_product}.
\begin{purpose}
In practice, \cref{eq:ode-affine} has few nonzero terms.
\end{purpose}
\begin{remark}
As written above, \cref{eq:ode-affine} is the most general form of a quadratic ODE with affine-parametric structure as in \cref{eq:affine-op}.
However, as we will see in examples, the constant, linear, and quadratic terms are not always present in practice, and when they are the number of terms in the affine expansions (i.e. $q_c$, $q_A$, and $q_H$) tends to be small.
In other words, projection-based ROMs of the form \cref{eq:ode-affine} have the same number of nonzero terms as the original PDE \cref{eq:pde-single}.
\end{remark}
\begin{example}[Heat Equation]
\label{example:heat-1}
Consider the one-dimensional spatial domain $\Omega = (0,1) \subset \mathbb{R}$ and the parameter domain $\mathcal{P} = [.01,2.5]\times[0.1,2.5]\subset\mathbb{R}^{2}$. For a fixed $\bar{x}\in\Omega$, let $\chi_{[0,\bar{x})}$ and $\chi_{[\bar{x},1]}$ be indicator functions over $\Omega$,
\begin{align*}
\chi_{[0,\bar{x})}(x)
&= \begin{cases}
1, & 0 < x < \bar{x}, \\
0, & \bar{x} \le x < 1,
\end{cases}
&
\chi_{[\bar{x},1]}(x)
&= \begin{cases}
0, & 0 < x < \bar{x}, \\
1, & \bar{x} \le x < 1.
\end{cases}
\end{align*}
The following equation models the diffusion of heat through a one-dimensional rod composed of two materials with distinct thermal diffusivities $\mu = (\alpha,\beta)\in\mathcal{P}$, with the temperature prescribed at each end of the rod:
\begin{subequations}
\begin{align}
\label{eq:heat-pde}
\frac{\partial u}{\partial t}
&= \left(\alpha \chi_{[0,\bar{x})} + \beta \chi_{[\bar{x},1]}\right) \frac{\partial^2 u}{\partial x^2}
&
x &\in \Omega,\
t \in (t_0,t_f],\
\mu\in\mathcal{P},
\\
\label{eq:heat-init}
u(x, t_0; \mu) &= u_0(x;\mu),
&
x &\in \Omega,\
\mu\in\mathcal{P},
\\
\label{eq:heat-bcs}
u(0, t; \mu) &= u(1, t; \mu) = 0,
&
t &\in [t_0,t_f],\ \mu\in\mathcal{P}.
\end{align}
\end{subequations}
The underlying Hilbert space containing the state $u$ is $\mathcal{V} = H^2(\Omega)\cap H^{1}_{0}(\Omega)$, the set of twice weakly differentiable functions satisfying the homogeneous Dirichlet boundary conditions \cref{eq:heat-bcs}.
In the language of \cref{eq:pde-polynomial}--\cref{eq:affine-op}, $\mathcal{C}\equiv\mathcal{H}\equiv 0$ and
\begin{align*}
\mathcal{A}(u;\mu)
&= \theta_{A}^{(1)}(\mu)\mathcal{A}^{(1)}(u) + \theta_{A}^{(2)}(\mu)\mathcal{A}^{(2)}(u)
\end{align*}
with $\theta_{A}^{(1)}(\mu)=\alpha$,
$\theta_{A}^{(1)}(\mu)=\beta$,
$\mathcal{A}^{(1)}(u) = \chi_{[0,\bar{x})}\frac{\partial^2 u}{\partial x^2}$, and
$\mathcal{A}^{(2)}(u) = \chi_{[\bar{x},1]}\frac{\partial^2 u}{\partial x^2}$.
A projection-based ROM for this problem as in \cref{eq:ode-affine} has the form
\begin{align}
\label{eq:heat-ode}
\frac{\mathrm{d}}{\mathrm{d}t}\widehat{\mathbf{u}}(t;\mu)
&= \left(\alpha\mathbf{A}^{(1)} + \beta\mathbf{A}^{(2)}\right)\widehat{\mathbf{u}}(t;\mu),
\end{align}
with the initial condition as in \cref{eq:ode-initial-condition}.
\end{example}
\subsection{Affine Operator Inference for PDEs} \label{subsec:operator_inference}
Constructing $\widehat{\mathbf{c}}^{(p)}$, $\widehat{\mathbf{A}}^{(p)}$, and $\widehat{\mathbf{H}}^{(p)}$ in the ODE system \cref{eq:ode-affine} by evaluating the terms of \cref{eq:inner_product} is an inherently intrusive process, requiring explicit access to the differential operators $\mathcal{C}^{(p)}$, $\mathcal{A}^{(p)}$, and $\mathcal{H}^{(p)}$. In this section, we construct affine-parametric Operator Inference to \emph{learn} the ROM through a data-driven optimization problem, given a basis and observations of the solution $u$.
\begin{purpose}
Introduce the operator matrix $\widehat{\mathbf{O}}$, which is what we want to learn.
\end{purpose}
Consider \cref{eq:ode-affine} with fixed, known integers $q_{c},q_{A},q_{H} \ge 0$ and affine coefficient functions $\theta = \{\theta_{c}^{(1)},\ldots,\theta_{H}^{(q_H)}\}$, which define the polynomial and affine-parametric structure of the system.
Define
\begin{subequations}
\begin{align}
\label{eq:operator-family}
\begin{aligned}
\frac{\textup{d}}{\textup{d}t}\widehat{\mathbf{u}}(t;\mu)
= \mathbf{F}(\widehat{\mathbf{O}};\widehat{\mathbf{u}},t,\theta,\mu)
= \left(\sum_{p=1}^{q_c}\theta_{c}^{(p)}(\mu)\widehat{\mathbf{c}}^{(p)}\right)
+ \left(\sum_{p=1}^{q_A}\theta_{A}^{(p)}(\mu)\widehat{\mathbf{A}}^{(p)}\right)\widehat{\mathbf{u}}(t;\mu)\\
+ \left(\sum_{p=1}^{q_H}\theta_{H}^{(p)}(\mu)\widehat{\mathbf{H}}^{(p)}\right)\left(\widehat{\mathbf{u}}(t;\mu)\,\widehat{\odot}\,\widehat{\mathbf{u}}(t;\mu)\right),
\end{aligned}
\end{align}
where the as yet unknown \emph{operator matrix} $\widehat{\mathbf{O}}$ is the concatenation
\begin{align}
\label{eq:operator-matrix}
\widehat{\mathbf{O}} = \left[\begin{array}{c|c|c}
\widehat{\mathbf{c}}^{(1)}\ \cdots\ \widehat{\mathbf{c}}^{(q_c)} &
\widehat{\mathbf{A}}^{(1)}\ \cdots\ \widehat{\mathbf{A}}^{(q_A)} &
\widehat{\mathbf{H}}^{(1)}\ \cdots\ \widehat{\mathbf{H}}^{(q_H)}
\end{array}\right]
\in\mathbb{R}^{r\times q(r)},
\end{align}
\end{subequations}
with column dimension $q(r) = q_c + q_A r + q_H \binom{r+1}{2}$.
Equipped with the initial condition \cref{eq:ode-initial-condition},
\cref{eq:operator-family}--\cref{eq:operator-matrix} describe the family of ROMs with the same polynomial and parametric form as \cref{eq:ode-affine}, with particular realizations determined by the operator matrix $\widehat{\mathbf{O}}$.
The goal is to choose $\widehat{\mathbf{O}}$ such that the corresponding ROM accurately captures the dynamics of the governing PDE \cref{eq:pde-single}--\cref{eq:pde-bcs-neumann} for all $\mu\in\mathcal{P}$.
\begin{purpose}
Mathematical statement of affine pOpInf.
\end{purpose}
We learn $\widehat{\mathbf{O}}$ by solving a data-driven least-squares regression problem. Suppose we can sample the solution of \cref{eq:pde-single}--\cref{eq:pde-bcs-neumann} at $s$ parameter values $\{\mu_{i}\}_{i=1}^{s}\subset\mathcal{P}$ and $K$ times $\{t_{j}\}_{j=1}^{K}\subset[t_0,t_f]$. We define the loss function $\mathcal{L}:\mathbb{R}^{r\times q(r)}\to\mathbb{R}$ associated with these solution samples to be the sum of the residuals of \cref{eq:operator-family},
\begin{align}
\label{eq:opinf-loss}
\mathcal{L}(\widehat{\mathbf{O}})
&=
\sum_{i=1}^{s}\sum_{j=1}^{K}\left\|
\mathbf{F}(\widehat{\mathbf{O}}; \widehat{\mathbf{u}}, t_{j}, \theta, \mu_{i})
- \frac{\textup{d}}{\textup{d} t}\widehat{\mathbf{u}}(t;\mu_{i})\Bigr|_{t=t_{j}}
\right\|_2^2,
\end{align}
where
$
\widehat{\mathbf{u}}(t;\mu)
= [
\langle v_1, u(\cdot,t;\mu)\rangle
~\cdots~
\langle v_r, u(\cdot,t;\mu)\rangle
]^\top \in \mathbb{R}^{r}
$
as before, and with $\mathbf{F}$ and $\widehat{\mathbf{O}}$ as in \cref{eq:operator-family}--\cref{eq:operator-matrix}.
To write the minimization of \cref{eq:opinf-loss} in a standard form, define the matrices
\begin{align*}
\widehat{\mathbf{U}}(\mu_{i}) &= \left[\begin{array}{ccc}
\widehat{\mathbf{u}}(t_{1};\mu_i) & \cdots & \widehat{\mathbf{u}}(t_{K};\mu_i)
\end{array}\right] \in \mathbb{R}^{r\times K},
\\
\dot{\widehat{\mathbf{U}}}(\mu_{i}) &= \left[\begin{array}{ccc}
\frac{\textup{d}}{\textup{d}t}\widehat{\mathbf{u}}(t;\mu_i)\Bigr|_{t=t_{1}} & \cdots & \frac{\textup{d}}{\textup{d}t}\widehat{\mathbf{u}}(t;\mu_i)\Bigr|_{t=t_{K}}
\end{array}\right] \in \mathbb{R}^{r\times K},
\end{align*}
which group the projected solution and the associated time derivatives for each parameter sample, and define the row vectors
\begin{align*}
\boldsymbol{\theta}_{c}(\mu_{i}) &= \left[\begin{array}{ccc}
\theta_{c}^{(1)}(\mu_{i}) & \cdots & \theta_{c}^{(q_c)}(\mu_{i})
\end{array}\right]\in\mathbb{R}^{1 \times q_c},
\\
\boldsymbol{\theta}_{A}(\mu_{i}) &= \left[\begin{array}{ccc}
\theta_{A}^{(1)}(\mu_{i}) & \cdots & \theta_{A}^{(q_A)}(\mu_{i})
\end{array}\right]\in\mathbb{R}^{1 \times q_A},
\\
\boldsymbol{\theta}_{H}(\mu_{i}) &= \left[\begin{array}{ccc}
\theta_{H}^{(1)}(\mu_{i}) & \cdots & \theta_{H}^{(q_H)}(\mu_{i})
\end{array}\right]\in\mathbb{R}^{1 \times q_H},
\end{align*}
which encode the affine parametric dependencies of $\mathcal{C}$, $\mathcal{A}$, and $\mathcal{H}$, respectively.
Finally, define the \emph{data matrix}
\begin{align}
\label{eq:data_matrix}
\begin{aligned}
\mathbf{D}
&= \left[\begin{array}{c|c|c}
\makebox[1.96cm]{$\mathbf{D}_{c}$} &
\makebox[2.70cm]{$\mathbf{D}_{A}$} &
\makebox[4.55cm]{$\mathbf{D}_{H}$}
\end{array}\right]
\\
&= \left[\begin{array}{c|c|c}
\boldsymbol{\theta}_c(\mu_{1})\otimes\mathbf{1}_{K} &
\boldsymbol{\theta}_A(\mu_{1})\otimes \widehat{\mathbf{U}}(\mu_{1})^\top &
\boldsymbol{\theta}_H(\mu_{1})\otimes\left(\widehat{\mathbf{U}}(\mu_{1}) \,\widehat{\odot}\, \widehat{\mathbf{U}}(\mu_{1})\right)^\top
\\
\vdots & \vdots & \vdots \\
\boldsymbol{\theta}_c(\mu_{s})\otimes\mathbf{1}_{K} &
\boldsymbol{\theta}_A(\mu_{s})\otimes \widehat{\mathbf{U}}(\mu_{s})^\top &
\boldsymbol{\theta}_H(\mu_{s})\otimes\left(\widehat{\mathbf{U}}(\mu_{s})\,\widehat{\odot}\,\widehat{\mathbf{U}}(\mu_{s})\right)^\top
\end{array}\right], \end{aligned}
\end{align}
where $\mathbf{1}_{K}\in \mathbb{R}^{K}$ is a column vector of unity of length $K$.
We then have
\begin{align}
\label{eq:opinf-standard}
\text{pOpInf:}
\qquad
\min_{\widehat{\mathbf{O}}}
\mathcal{L}(\widehat{\mathbf{O}})
=
\min_{\widehat{\mathbf{O}}}
\left\|
\mathbf{D}\widehat{\mathbf{O}}^\top - \mathbf{R}^\top
\right\|_{F}^2,
\end{align}
where $\mathbf{D} \in \mathbb{R}^{sK\times q\left(r\right)}$,
$\widehat{\mathbf{O}}\in \mathbb{R}^{r\times q\left(r\right)}$, and
$
\mathbf{R} = [
\dot{\widehat{\mathbf{U}}}(\mu_{1})
~\cdots~
\dot{\widehat{\mathbf{U}}}(\mu_{s})
] \in\mathbb{R}^{r \times sK}.
$
We call \cref{eq:opinf-standard} the \emph{affine-parametric Operator Inference problem} (pOpInf).
Note that \cref{eq:opinf-standard} is a linear least-squares problem which decouples by the columns of $\widehat{\mathbf{O}}^\top$, meaning the dynamics for each $\hat{u}_{i}$ are learned independently \cite{PW2016operatorInference}.
\begin{purpose}
Parameters must be selected carefully if \cref{eq:opinf-standard} is to be well posed.
\end{purpose}
If the data matrix $\mathbf{D}$ has full column rank, then \cref{eq:opinf-standard} has a unique closed-form solution \cite{bjorck1996numerical,golub2013matrix,trefethen1997numerical}. However, $\mathbf{D}$ is susceptible to rank deficiencies due to its Kronecker block structure. \Cref{thm:rank-deficiencies} establishes necessary conditions for the well-posedness of the pOpInf problem \cref{eq:opinf-standard} and sufficient conditions in the monomial case (e.g., $\mathcal{C} = \mathcal{A} = 0$ but $\mathcal{H} \neq 0$ so that $\mathbf{D} = \mathbf{D}_{H}$).
\begin{theorem}
\label{thm:rank-deficiencies}
Let $\Theta_{c}\in\mathbb{R}^{s\times q_{c}}$, $\Theta_{A}\in\mathbb{R}^{s\times q_{A}}$, and $\Theta_{H}\in\mathbb{R}^{s\times q_{H}}$ be the matrices with entries
\begin{align}
\label{eq:Theta-matrices}
[\Theta_{c}]_{ij}
&= \theta_{c}^{(j)}(\mu_{i}),
&
[\Theta_{A}]_{ij}
&= \theta_{A}^{(j)}(\mu_{i}),
&
[\Theta_{H}]_{ij}
&= \theta_{H}^{(j)}(\mu_{i}),
\end{align}
that is, the $i$th row of $\Theta_{c}$ is $\boldsymbol{\theta}_{c}(\mu_{i})$, the $i$th row of $\Theta_{A}$ is $\boldsymbol{\theta}_{A}(\mu_{i})$, and the $i$th row of $\Theta_{H}$ is $\boldsymbol{\theta}_{H}(\mu_{i})$.
If any of $\Theta_c$, $\Theta_{A}$, or $\Theta_{H}$ do not have full column rank, then the data matrix $\mathbf{D} = [\mathbf{D}_{c}~\mathbf{D}_{A}~\mathbf{D}_{H}]$ defined in \cref{eq:data_matrix} is rank deficient.
Furthermore, if either of the block matrices
\begin{align*}
\widehat{\mathbf{U}}_{A}
&= \left[\begin{array}{c}
\widehat{\mathbf{U}}(\mu_{1})^\top
\\ \vdots \\
\widehat{\mathbf{U}}(\mu_{s})^\top
\end{array}\right],
&
\widehat{\mathbf{U}}_{H}
&= \left[\begin{array}{c}
\left(\widehat{\mathbf{U}}(\mu_{1})\,\widehat{\odot}\,\widehat{\mathbf{U}}(\mu_{1})\right)^\top
\\ \vdots \\
\left(\widehat{\mathbf{U}}(\mu_{s})\,\widehat{\odot}\,\widehat{\mathbf{U}}(\mu_{s})\right)^\top
\end{array}\right]
\end{align*}
do not have full column rank, then neither does $\mathbf{D}$.
On the other hand,
if $\Theta_{A}$ and each $\widehat{\mathbf{U}}(\mu_{i})^\top$ ($i=1,\ldots,s$) have full column rank, then so does $\mathbf{D}_{A}$;
and if $\Theta_{H}$ and each $(\widehat{\mathbf{U}}(\mu_{i})\,\widehat{\odot}\,\widehat{\mathbf{U}}(\mu_{i}))^\top$ have full column rank, then so does $\mathbf{D}_{H}$.
\begin{proof}
Note that $\Theta_{c} = \mathbf{D}_{c}$, the leftmost block of $\mathbf{D}$. Therefore, if $\Theta_{c}$ does not have full column rank, then $\mathbf{D}$ has linearly dependent columns.
\Cref{lemma:kronecker-structure} gives the result for the remaining cases:
if $\Theta_{A}$ or $\widehat{\mathbf{U}}_{A}$ do not have full column rank, apply \cref{lemma:kronecker-structure} with $\mathbf{y}_{i} = \boldsymbol{\theta}_{A}(\mu_{i})^\top$, $\mathbf{Z}_{i} = \widehat{\mathbf{U}}(\mu_{i})^{\top}$, and $\mathbf{W} = \mathbf{D}_{A}$ to show $\mathbf{D}_{A}$ is rank deficient;
a similar argument holds for $\Theta_{H}$ and $\widehat{\mathbf{U}}_{H}$, showing $\mathbf{D}_{H}$ is rank deficient.
The results to the converse also hold by \cref{lemma:kronecker-structure}.
\end{proof}
\end{theorem}
\begin{purpose}
Relevance of \cref{thm:rank-deficiencies} for selecting parameter samples.
\end{purpose}
Selecting appropriate parameter samples is an important issue for all parametric model reduction methods \cite{BGW2015pmorSurvery}. \Cref{thm:rank-deficiencies} provides numerically relevant guidance for parameter selection in the context of \cref{eq:opinf-standard}: for a set of parameter values $\{\mu_{i}\}_{i=1}^{s}$, the (small) matrices $\Theta_{c}$, $\Theta_{A}$, and $\Theta_{H}$ can be explicitly formed and checked for rank deficiencies without any information about the solution trajectories. If any of the matrices are unsatisfactorily conditioned, different parameter values must be chosen at which to sample the solution. Note that if $s < \max\{q_{c},q_{A},q_{H}\}$, then the problem is guaranteed to be ill-posed.
\Cref{sec:computation} further addresses numerical issues that may arise from other sources of poor conditioning.
\begin{example}[Heat Equation]
\label{example:heat-2}
Recall the problem introduced in \cref{example:heat-1}.
Given parameter samples $\{\mu_{i} = (\alpha_{i},\beta_{i})\}_{i=1}^{s}$, we learn a ROM of the form \cref{eq:heat-ode}
by solving the pOpInf problem \cref{eq:opinf-standard} with operator and data matrices
\begin{align*}
\widehat{\mathbf{O}} &= \left[\begin{array}{cc}
\widehat{\mathbf{A}}^{(1)} & \widehat{\mathbf{A}}^{(2)}
\end{array}\right] \in \mathbb{R}^{r \times 2r},
&
\mathbf{D} &= \left[\begin{array}{cc}
\alpha_{1} \widehat{\mathbf{U}}(\mu_1)^\top & \beta_{1} \widehat{\mathbf{U}}(\mu_1)^\top \\
\vdots & \vdots \\
\alpha_{s} \widehat{\mathbf{U}}(\mu_s)^\top & \beta_{s} \widehat{\mathbf{U}}(\mu_s)^\top \\
\end{array}\right] \in \mathbb{R}^{sK \times 2r}.
\end{align*}
Per \cref{thm:rank-deficiencies}, the rank of the data matrix $\mathbf{D}$ depends on each $\widehat{\mathbf{U}}(\mu_{i})$ and the matrix
\begin{align*}
\Theta_{A}
= \left[\begin{array}{cc}
\theta_{A}^{(1)}(\mu_1) & \theta_{A}^{(2)}(\mu_1) \\
\vdots & \vdots \\
\theta_{A}^{(1)}(\mu_s) & \theta_{A}^{(2)}(\mu_s)
\end{array}\right]
= \left[\begin{array}{cc}
\alpha_{1} & \beta_{1} \\
\vdots & \vdots \\
\alpha_{s} & \beta_{s}
\end{array}\right]\in\mathbb{R}^{s\times 2}.
\end{align*}
In particular, if $\Theta_{A}$ does not have full column rank, then the data matrix $\mathbf{D}$ will be rank deficient.
Our goal, then, is to choose $s \ge 2$ parameter samples such that $\Theta_{A}$ has a small condition number.
\end{example}
\begin{purpose}
Relation between the OpInf ROM and the intrusive Galerkin ROM.
\end{purpose}
Reduced-order models learned through pOpInf and those constructed via intrusive projection, i.e., by explicitly evaluating \cref{eq:inner_product}, are related in the following sense.
\begin{theorem}
\label{thm:existence-uniqueness}
Let $u$ be the unique solution of \cref{eq:pde-single}--\cref{eq:pde-bcs-neumann}, where the differential operator $\mathcal{F}$ has the quadratic, affine-parametric form described in \cref{eq:pde-polynomial}--\cref{eq:affine-op}.
Suppose there exists an orthonormal set $\{v_{j}\}_{j=1}^{r}\subset\mathcal{V}$, and let $\{\mu_{i}\}_{i=1}^{s}\subset\mathcal{P}$ be a finite set of parameter samples.
Define the loss function
\begin{align}
\label{eq:loss-timecontinuous}
\mathscr{L}(\widehat{\mathbf{O}}) =
\sum_{i=1}^{s}
\int_{t_0}^{t_f}
\left\|
\mathbf{F}\left(
\widehat{\mathbf{O}};\widehat{\mathbf{u}},t,\theta,\mu_{i}
\right)
- \frac{\textup{d}}{\textup{d}t}\widehat{\mathbf{u}}(t;\mu_{i})
\right\|_{2}^{2}
\:dt,
\end{align}
where
$
\widehat{\mathbf{u}}(t;\mu)
= [
\langle v_1, u(\cdot,t;\mu)\rangle
~\cdots~
\langle v_r, u(\cdot,t;\mu)\rangle
]^\top \in \mathbb{R}^{r},
$
and with $\mathbf{F}$ and $\widehat{\mathbf{O}}$ given by \cref{eq:operator-family}--\cref{eq:operator-matrix}.
Consider the following conditions.
\begin{enumerate}
\item There exist functions $\hat{u}_{j}:[t_0,t_f]\times\mathcal{P}\to\mathbb{R}$, $j=1,\ldots,r$, such that the finite sum representation \cref{eq:truncation-projection} is exact for all $x\in\Omega$, $t\in[t_0,t_f]$, and $\mu \in \{\mu_{i}\}_{i=1}^{s}$, i.e., there is no truncation error at the sample parameter values.
\item For $i=1,\ldots,s$, there exist times $\{\tau_{i,j}\}_{j=1}^{K}\subset [t_0,t_f]$, $K = 1 + r + \binom{r+1}{2}$, such that the matrix
\begin{align*}
\widetilde{\mathbf{D}}(\mu_i) = \left[\begin{array}{ccc}
\mathbf{1}_{K} &
\widetilde{\mathbf{U}}(\mu_{i})^\top &
\left(\widetilde{\mathbf{U}}(\mu_{i})\,\widehat{\odot}\,\widetilde{\mathbf{U}}(\mu_{i})\right)^\top
\end{array}\right] \in \mathbb{R}^{K \times K}
\end{align*}
is invertible, where
$
\widetilde{\mathbf{U}}(\mu_{i})
= [
\widehat{\mathbf{u}}(\tau_{i,1};\mu_i)~\cdots~\widehat{\mathbf{u}}(\tau_{i,K};\mu_i)
]
\in\mathbb{R}^{r\times K}.
$
\item The matrices $\Theta_{c}\in\mathbb{R}^{s\times q_{c}}$, $\Theta_{A}\in\mathbb{R}^{s\times q_{A}}$, and $\Theta_{H}\in\mathbb{R}^{s\times q_{H}}$ of \cref{eq:Theta-matrices} have full column rank.
\end{enumerate}
If condition 1 holds, then the loss function $\mathscr{L}$ has a global minimizer $\widehat{\mathbf{O}}$ satisfying $\mathscr{L}(\widehat{\mathbf{O}}) = 0$.
If conditions 2 and 3 also hold, then that minimizer is unique.
\begin{proof}
Assume condition 1 (no truncation error at the parameter samples).
Then the system of ODEs \cref{eq:ode-affine} with operators $\widehat{\mathbf{c}}^{(p)}$, $\widehat{\mathbf{A}}^{(p)}$, and $\widehat{\mathbf{H}}^{(p)}$ derived from \cref{eq:inner_product} holds exactly for all $x\in\Omega$, $t\in[t_0,t_f]$, and $\mu\in\{\mu_{i}\}_{i=1}^{s}$, i.e., it is equivalent to \cref{eq:pde-single}--\cref{eq:pde-bcs-neumann} at the parameter samples.
Constructing $\widehat{\mathbf{O}}$ from these operators, we have
$
\frac{\textup{d}}{\textup{d}t}\widehat{\mathbf{u}}(t;\mu)
= \mathbf{F}(\widehat{\mathbf{O}};\widehat{\mathbf{u}},t,\theta,\mu)
$
exactly for each $\mu\in\{\mu_{i}\}_{i=1}^{s}$.
By construction, $\mathscr{L}(\widehat{\mathbf{O}}) = 0$, which---since $\mathscr{L}$ is non-negative---shows that $\widehat{\mathbf{O}}$ is a global minimizer of $\mathscr{L}$.
To prove uniqueness, assume conditions 2 and 3 and suppose $\widehat{\mathbf{O}}^{(1)}$ and $\widehat{\mathbf{O}}^{(2)}$ both minimize $\mathscr{L}$.
By the previous argument, $\mathscr{L}(\widehat{\mathbf{O}}^{(1)}) = \mathscr{L}(\widehat{\mathbf{O}}^{(2)}) = 0$, which implies
\begin{align}
\label{eq:uniqueness-argument}
\mathbf{0}_{r}
&= \mathbf{F}(\widehat{\mathbf{O}}^{(1)};\widehat{\mathbf{u}},\tau_{i,j},\theta,\mu_{i})
- \mathbf{F}(\widehat{\mathbf{O}}^{(2)};\widehat{\mathbf{u}},\tau_{i,j},\theta,\mu_{i})
= \mathbf{F}(\widetilde{\mathbf{O}};\widehat{\mathbf{u}},\tau_{i,j},\theta,\mu_{i})
\end{align}
for $i=1,\ldots,s$ and $j=1,\ldots,K$, where $\mathbf{0}_{r}\in\mathbb{R}^{r}$ is a column vector of zeros and
\begin{align*}
\widetilde{\mathbf{O}}
&= \widehat{\mathbf{O}}^{(1)} - \widehat{\mathbf{O}}^{(2)}
= \left[\begin{array}{c|c|c}
\widetilde{\mathbf{c}}^{(1)}\ \cdots\ \widetilde{\mathbf{c}}^{(q_c)}
&
\widetilde{\mathbf{A}}^{(1)}\ \cdots\ \widetilde{\mathbf{A}}^{(q_A)}
&
\widetilde{\mathbf{H}}^{(1)}\ \cdots\ \widetilde{\mathbf{H}}^{(q_H)}
\end{array}\right] \in\mathbb{R}^{r \times q(r)}.
\end{align*}
For fixed $i$, \cref{eq:uniqueness-argument} with $j=1,\ldots,K$ can be written as the linear system
\begin{align*}
\widetilde{\mathbf{D}}(\mu_{i})
\left[\begin{array}{c|c|c}
\sum_{p=1}^{q_{c}}\theta_{c}^{(p)}(\mu_{i})\widetilde{\mathbf{c}}^{(p)}
&
\sum_{p=1}^{q_{A}}\theta_{A}^{(p)}(\mu_{i})\widetilde{\mathbf{A}}^{(p)}
&
\sum_{p=1}^{q_{H}}\theta_{H}^{(p)}(\mu_{i})\widetilde{\mathbf{H}}^{(p)}
\end{array}\right]^\top
= \mathbf{0}_{K\times r}, \end{align*}
where $\mathbf{0}_{K\times r}\in\mathbb{R}^{K\times r}$ is the zero matrix. As each $\widetilde{\mathbf{D}}(\mu_i)$ is invertible (condition 2), for $i=1,\ldots,s$ we have
\begin{align*}
\sum_{p=1}^{q_{c}}\theta_{c}^{(p)}(\mu_{i})\widetilde{\mathbf{c}}^{(p)}
&= \mathbf{0}_{r}, &
\sum_{p=1}^{q_{A}}\theta_{A}^{(p)}(\mu_{i})\widetilde{\mathbf{A}}^{(p)}
&= \mathbf{0}_{r\times r}, &
\sum_{p=1}^{q_{H}}\theta_{H}^{(p)}(\mu_{i})\widetilde{\mathbf{H}}^{(p)}
&= \mathbf{0}_{r\times \binom{r+1}{2}}. \end{align*}
Collecting these equations component-wise for $i=1,\ldots,s$ yields the equations
\begin{align*}
\Theta_{c}\left[\begin{array}{c}
{[\widetilde{\mathbf{c}}^{(1)}]}_{j}
\\ \vdots \\
{[\widetilde{\mathbf{c}}^{(q_c)}]}_{j}
\end{array}\right]
=
\Theta_{A}\left[\begin{array}{c}
{[\widetilde{\mathbf{A}}^{(1)}]}_{jk}
\\ \vdots \\
{[\widetilde{\mathbf{A}}^{(q_A)}]}_{jk}
\end{array}\right]
=
\Theta_{H}\left[\begin{array}{c}
{[\widetilde{\mathbf{H}}^{(1)}]}_{jk}
\\ \vdots \\
{[\widetilde{\mathbf{H}}^{(q_H)}]}_{jk}
\end{array}\right]
=
\mathbf{0}_{s} \end{align*}
for every index pair $j$,$k$.
But $\Theta_c$, $\Theta_{A}$, and $\Theta_{H}$ each have full column rank (condition 3), implying
$
[\widetilde{\mathbf{c}}^{(p)}]_{j}
= [\widetilde{\mathbf{A}}^{(p)}]_{jk}
= [\widetilde{\mathbf{H}}^{(p)}]_{jk}
= 0
$ for all $j,k,p$.
Thus, $\widetilde{\mathbf{O}}$ is the zero matrix, hence $\widehat{\mathbf{O}}^{(1)} = \widehat{\mathbf{O}}^{(2)}$.
\end{proof}
\end{theorem}
\begin{purpose}
Why the conditions of \cref{thm:existence-uniqueness} are not all satisfied in practice.
\end{purpose}
\Cref{thm:existence-uniqueness} shows that for finite-dimensional PDEs (condition 1) with a polynomial, affine-parametric structure, the operators defined via \cref{eq:inner_product} uniquely minimize the loss function $\mathscr{L}$ defined in \cref{eq:loss-timecontinuous}, provided that the dynamics are sufficiently diverse (as quantified by conditions 2 and 3). As the time step between the data samples defining the pOpInf loss $\mathcal{L}$ of \cref{eq:opinf-loss} decreases, i.e., $t_{j+1} - t_{j} \to 0$ for all $j$, then $\mathcal{L}$ converges to $\mathscr{L}$ (up to a constant), in which case the pOpInf ROM and the intrusive projection-based ROM will be the same. For infinite-dimensional PDEs, condition 1 is realized as the number of basis functions in the approximation increases, i.e., $r\to\infty$. Thus, there are two factors explaining why ROMs learned through pOpInf are not guaranteed to agree with \cref{eq:inner_product}: 1) the truncation error of the representation \cref{eq:truncation-projection}, and 2) the finite sampling of the solution in time. See \cite{Peherstorfer2020reprojection,uy2021probabilistic} for additional work connecting the intrusive and data-driven ROMs.
\section{Affine Operator Inference for Systems of PDEs} \label{sec:pde_systems}
Systems of partial differential equations with affine polynomial structure admit low-dimensional representations similar to \cref{eq:ode-affine}--\cref{eq:ode-initial-condition} \cite{GW2021learning}.
Consider the system of $d$ partial differential equations where each equation can be written as in \cref{eq:pde-single}--\cref{eq:pde-bcs-neumann}:
\begin{subequations}
\begin{align}
\label{eq:pde-multi}
\dfrac{\partial u_{\ell}}{\partial t} &= \mathcal{F}_{\ell}(u_{\ell},\ldots,u_{d};\mu),
& \ell &= 1,\ldots, d,
\end{align}
where each state variable $u_{\ell}$ is contained in a Hilbert space $\mathcal{V}_{\ell}$ of real-valued functions satisfying appropriate homogeneous boundary conditions.
Suppose each spatial differential operator $\mathcal{F}_{\ell}:\mathcal{V}_{1}\times\cdots\times\mathcal{V}_{d}\times\mathcal{P}\to\mathcal{V}_{\ell}^{*}$ is polynomial in the variables $u_1,\ldots,u_d$ and their spatial derivatives,
\begin{align}
\label{eq:pde-multi-polynomial}
\begin{aligned}
\mathcal{F}_{\ell}(u_1,\ldots,u_d;\mu)
= \mathcal{C}_{\ell}(\mu)
&+ \sum_{m=1}^{d}\mathcal{A}_{\ell,m}(u_{m};\mu)+ \sum_{m=1}^{d}\sum_{n=m}^{d}\mathcal{H}_{\ell,mn}(u_{m},u_{n};\mu),
\end{aligned}
\end{align}
where $\mathcal{C}_{\ell}:\mathcal{P}\to\mathcal{V}_{\ell}^{*}$, $\mathcal{A}_{\ell,m}:\mathcal{V}_{m}\times\mathcal{P}\to\mathcal{V}_{\ell}^{*}$, $\mathcal{H}_{\ell,mn}:\mathcal{V}_{m}\times\mathcal{V}_{n}\times\mathcal{P}\to\mathcal{V}_{\ell}^{*}$, and each $\mathcal{A}_{\ell,m}$ and $\mathcal{H}_{\ell,mn}$ is linear in each state argument. This is the natural multivariate extension of \cref{eq:pde-polynomial}. We further assume that each operator $\mathcal{C}_{\ell}$, $\mathcal{A}_{\ell,m}$, and $\mathcal{H}_{\ell,mn}$ has an affine-parametric expansion, that is,
\begin{gather}
\label{eq:affine-expansion-system}
\begin{gathered}
\mathcal{C}_{\ell}(\mu)
= \sum_{p=1}^{q_{c_{\ell}}}\theta_{c_{\ell}}^{(p)}(\mu)\mathcal{C}_{\ell}^{(p)}.
\qquad
\mathcal{A}_{\ell,m}(u;\mu)
= \sum_{p=1}^{q_{A_{\ell,m}}}\theta_{A_{\ell,m}}^{(p)}(\mu)\mathcal{A}_{\ell,m}^{(p)}(u).
\\
\mathcal{H}_{\ell,mn}(u,v;\mu)
= \sum_{p=1}^{q_{H_{\ell,mn}}}\theta_{H_{\ell,mn}}^{(p)}(\mu)\mathcal{H}_{\ell,mn}^{(p)}(u,v).
\end{gathered}
\end{gather}
\end{subequations}
\begin{purpose}
Basis representation $\rightarrow$ ROM with affine-polynomial system structure.
\end{purpose}
For each $\ell=1,\ldots, d$, let $\{v_{\ell j}\}_{j=1}^{r_\ell}\subset\mathcal{V}_{\ell}$ be an orthonormal set.
We assume a reduced representation $\breve{u}_{\ell}$ of $u_{\ell}$, confined to $\textup{span}(\{v_{\ell_{1}},\ldots,v_{\ell r_{\ell}}\})\subset\mathcal{V}_{\ell}$, given by
\begin{align}
\label{eq:basis-expansion-multi}
\breve{u}_{\ell}(x,t;\mu)
&= \sum_{j=1}^{r_{\ell}} \hat{u}_{\ell j}(t;\mu) v_{\ell j}(x),
&
\hat{u}_{\ell j}(t;\mu)
&= \left\langle v_{\ell j}, u_{\ell}(\cdot,t;\mu) \right\rangle_{\mathcal{V}_{\ell}},
\end{align}
where $\langle\cdot,\cdot\rangle_{\mathcal{V}_{\ell}}$ is the duality pairing of $\mathcal{V}_{\ell}$ with its dual.
Inserting $\breve{u}_{\ell}$ into \cref{eq:pde-multi} for each $\ell$ and proceeding as before, we obtain $d$ ODEs that serve as a ROM for \cref{eq:pde-multi}--\cref{eq:affine-expansion-system}:
\begin{subequations}
\begin{align}
\label{eq:ode-system}
&\begin{aligned}
\frac{\textup{d}}{\textup{d} t}\widehat{\mathbf{u}}_{\ell}(t;\mu)
&= \mathbf{F}_{\ell}(\widehat{\mathbf{O}}_{\ell};\widehat{\mathbf{u}}_{1},\ldots,\widehat{\mathbf{u}}_{d},t,\theta,\mu)
\\
&= \left(\sum_{p=1}^{q_{c_{\ell}}}\theta_{c_{\ell}}^{(p)}(\mu)\widehat{\mathbf{c}}_{\ell}^{(p)}\right)
+ \sum_{m=1}^{d}\left(\sum_{p=1}^{q_{A_{\ell,m}}}\theta_{A_{\ell,m}}^{(p)}(\mu)\widehat{\mathbf{A}}_{\ell,m}^{(p)}\right)\widehat{\mathbf{u}}_{m}(t;\mu)
\\
&\quad + \sum_{m=1}^{d}\left(\sum_{p=1}^{q_{H_{\ell,mm}}}\theta_{H_{\ell,mm}}^{(p)}(\mu)\widehat{\mathbf{H}}_{\ell,mm}^{(p)}\right)\big(\widehat{\mathbf{u}}_{m}(t;\mu)\,\widehat{\odot}\,\widehat{\mathbf{u}}_{m}(t;\mu)\big)
\\
&\quad + \sum_{m=1}^{d}\sum_{n=m+1}^{d}\left(\sum_{p=1}^{q_{H_{\ell,mn}}}\theta_{H_{\ell,mn}}^{(p)}(\mu)\widehat{\mathbf{H}}_{\ell,mn}^{(p)}\right)\big(\widehat{\mathbf{u}}_{m}(t;\mu)\odot\widehat{\mathbf{u}}_{n}(t;\mu)\big),
\end{aligned}
\\
\label{eq:state-vector-system}
&\widehat{\mathbf{u}}_{\ell}(t;\mu)
= \left[\begin{array}{ccc}
\langle v_{\ell 1}, u_{\ell}(\cdot,t;\mu)\rangle & \cdots & \langle v_{\ell,r_{\ell}}, u_{\ell}(\cdot,t;\mu)\rangle
\end{array}\right]^\top \in \mathbb{R}^{r_{\ell}},
\\
\label{eq:operator-matrix-system}
&\widehat{\mathbf{O}}_{\ell}
= \left[\begin{array}{ccc|ccc|ccc}
\widehat{\mathbf{c}}_{\ell}^{(1)}
& \cdots &
\widehat{\mathbf{c}}_{\ell}^{(q_{c_\ell})}
&
\widehat{\mathbf{A}}_{\ell,1}^{(1)}
& \cdots &
\widehat{\mathbf{A}}_{\ell,d}^{(q_{A_{\ell,d}})}
&
\widehat{\mathbf{H}}_{\ell,11}^{(1)}
& \cdots &
\widehat{\mathbf{H}}_{\ell,dd}^{(q_{H_{\ell,dd}})}
\end{array}\right],
\end{align}
\end{subequations}
where
$\widehat{\mathbf{c}}_{\ell}^{(p)} \in \mathbb{R}^{r_{\ell}}$,
$\widehat{\mathbf{A}}_{\ell,m}^{(p)}\in\mathbb{R}^{r_{\ell} \times r_m}$,
$\widehat{\mathbf{H}}_{\ell,mm}^{(p)}\in\mathbb{R}^{r_{\ell} \times \binom{r_m+1}{2}}$, and $\widehat{\mathbf{H}}_{\ell,mn}^{(p)}\in\mathbb{R}^{r_{\ell} \times r_m r_n}$ ($m \neq n$).
Here $\odot$ denotes the Khatri-Rao product, whereas $\widehat{\odot}$ is the compact Khatri-Rao product that omits redundant terms (see \cref{appendix:kronecker}).
Higher-order operators (e.g., cubic terms) can be similarly accounted for.
\begin{purpose}
pOpInf problem for systems decouples.
\end{purpose}
The low-dimensional system \cref{eq:ode-system} retains the differential and parametric structure of the PDE system \cref{eq:pde-multi}--\cref{eq:pde-multi-polynomial}. The corresponding affine pOpInf problem decouples into $d$ instances of \cref{eq:opinf-standard} which can be solved \emph{independently}:
\begin{align}
\label{eq:opinf-standard-system}
&\min_{\widehat{\mathbf{O}}_{\ell}}\left\|
\mathbf{D}_{\ell}\widehat{\mathbf{O}}_{\ell}^\top - \mathbf{R}_{\ell}^\top
\right\|_{F}^2,
&
\ell &= 1,\ldots, d,
\end{align}
where
\begin{align*}
\mathbf{R}_{\ell} &= \left[\begin{array}{ccc}
\dot{\widehat{\mathbf{U}}}_{\ell}(\mu_1)
& \cdots &
\dot{\widehat{\mathbf{U}}}_{\ell}(\mu_s)
\end{array}\right] \in \mathbb{R}^{r_{\ell} \times sK},
\\
\dot{\widehat{\mathbf{U}}}_{\ell}(\mu_i) &= \left[\begin{array}{ccc}
\frac{\textup{d}}{\textup{d}t}\widehat{\mathbf{u}}_{\ell}(t;\mu_i)\Bigr|_{t=t_{1}} & \cdots & \frac{\textup{d}}{\textup{d}t}\widehat{\mathbf{u}}_{\ell}(t;\mu_{i})\Bigr|_{t=t_{K}}
\end{array}\right] \in \mathbb{R}^{r_{\ell} \times K},
\end{align*}
and where each $\mathbf{D}_{\ell}$ is constructed from the matrices $\widehat{\mathbf{U}}_1(\mu_i),\ldots,\widehat{\mathbf{U}}_d(\mu_i)$ where
\begin{align*}
\widehat{\mathbf{U}}_{\ell}(\mu_i) &= \left[\begin{array}{ccc}
\widehat{\mathbf{u}}_{\ell}(t_{1};\mu_i) & \cdots & \widehat{\mathbf{u}}_{\ell}(t_{K};\mu_{i})
\end{array}\right] \in \mathbb{R}^{r_{\ell} \times K}.
\end{align*}
See \cref{appendix:systems} for the general construction, and note that the number of terms in the PDE dictates the number of terms in the ROM and hence the size of the pOpInf problem.
This is best illustrated by example.
\begin{example}[FitzHugh-Nagumo System]
\label{example:fh-n}
Let $\Omega = (0,1)$ and define the parameters $\mu = (\alpha,\beta,\gamma,\varepsilon)\in\mathbb{R}^{4}$.
The following system of equations is a simplification of the Hodgkin-Huxley model for activation and deactivation in a spiking neuron \cite{fitzhugh1961impulses,Nagumo1962pulse}:
\begin{subequations}
\begin{align}
\label{eq:fhn-system}
\frac{\partial u_{1}}{\partial t}
&= \varepsilon\frac{\partial^{2} u_{1}}{\partial x^{2}}
+ \frac{1}{\varepsilon}\left(
- u_{1}^{3} + 1.1 u_{1}^{2} - 0.1 u_{1}
- u_{2}
+ \alpha\right),
\\
\label{eq:fhn-system2}
\frac{\partial u_{2}}{\partial t}
&= \beta u_{1} - \gamma u_{2} + \alpha,
\end{align}
with initial conditions $u_{1}(x,t_0) = u_{2}(x,t_0)=0$ and Neumann boundary conditions
\begin{align}
\label{eq:fhn-bcs}
\left.\frac{\partial u_{1}}{\partial x}\right|_{x=0}
&= f(t) := -50000t^{3}e^{-15t},
&
\left.\frac{\partial u_{1}}{\partial x}\right|_{x=1}
&= 0.
\end{align}
\end{subequations}
We write \cref{eq:fhn-system}--\cref{eq:fhn-bcs} in the language of \cref{eq:pde-multi}--\cref{eq:affine-expansion-system} as
\begin{align*}
\frac{\partial u_{1}}{\partial t}
&=
\theta_{c_1}^{(1)}(\mu)\,\mathcal{C}_{1}^{(1)}
+ \theta_{A_{1,1}}^{(1)}(\mu)\,\mathcal{A}_{1,1}^{(1)}(u_1)
+ \theta_{A_{1,1}}^{(2)}(\mu)\,\mathcal{A}_{1,1}^{(2)}(u_1)
+ \theta_{A_{1,2}}^{(1)}(\mu)\,\mathcal{A}_{1,2}^{(2)}(u_2)
\\
& \qquad \qquad
+ \theta_{H_{1,11}}^{(1)}(\mu)\,\mathcal{H}_{1,11}^{(1)}(u_1,u_1) + \theta_{G_{1,111}}^{(1)}(\mu)\,\mathcal{G}_{1,111}^{(1)}(u_1,u_1,u_1),
\\
\frac{\partial u_{2}}{\partial t}
&=
\theta_{c_2}^{(1)}(\mu)\,\mathcal{C}_{2}^{(1)}
+ \theta_{A_{2,1}}^{(1)}(\mu)\,\mathcal{A}_{2,1}^{(1)}(u_1)
+ \theta_{A_{2,2}}^{(1)}(\mu)\,\mathcal{A}_{2,2}^{(1)}(u_2),
\end{align*}
with the operators and the affine coefficient functions given in \cref{table:fhn-terms}.
This motivates a ROM of the form
\begin{align*}
\frac{\textup{d}}{\textup{d} t}\widehat{\mathbf{u}}_{1}(t;\mu)
&=
\theta_{c_1}^{(1)}(\mu)\,\widehat{\mathbf{c}}_{1}^{(1)}
+ \left(\theta_{A_{1,1}}^{(1)}(\mu)\,\widehat{\mathbf{A}}_{1,1}^{(1)}
+ \theta_{A_{1,1}}^{(2)}(\mu)\,\widehat{\mathbf{A}}_{1,1}^{(2)}\right) \widehat{\mathbf{u}}_1(t;\mu)
\\ \nonumber & \qquad
+ \theta_{A_{1,2}}^{(1)}(\mu)\,\widehat{\mathbf{A}}_{1,2}^{(1)} \widehat{\mathbf{u}}_2(t;\mu)
+ \theta_{H_{1,11}}^{(1)}(\mu)\,\widehat{\mathbf{H}}_{1,11}^{(1)} \left(\widehat{\mathbf{u}}_1(t;\mu) \,\widehat{\odot}\, \widehat{\mathbf{u}}_1(t;\mu)\right)
\\ \nonumber & \qquad
+ \theta_{G_{1,111}}^{(1)}(\mu)\,\widehat{\mathbf{G}}_{1,111}^{(1)} \left(\widehat{\mathbf{u}}_1(t;\mu) \,\widehat{\odot}\, \widehat{\mathbf{u}}_1(t;\mu) \,\widehat{\odot}\, \widehat{\mathbf{u}}_1(t;\mu)\right)
+ \theta_{B_1}^{(1)}(\mu)\widehat{\mathbf{B}}_{1}^{(1)}f(t),
\\
\frac{\textup{d}}{\textup{d} t}\widehat{\mathbf{u}}_{2}(t;\mu)
&=
\theta_{c_2}^{(1)}(\mu)\,\widehat{\mathbf{c}}_{2}^{(1)}
+ \theta_{A_{2,1}}^{(1)}(\mu)\,\widehat{\mathbf{A}}_{2,1}^{(1)} \widehat{\mathbf{u}}_1(t;\mu)
+ \theta_{A_{2,2}}^{(1)}(\mu)\,\widehat{\mathbf{A}}_{2,2}^{(1)} \widehat{\mathbf{u}}_2(t;\mu),
\end{align*}
where
the sizes of each discretized operators are listed in \cref{table:fhn-terms}.
The operator $\widehat{\mathbf{B}}_1^{(1)}$ accounts for the Neumann boundary condition on $u_1$ in \cref{eq:fhn-bcs}; it has the coefficient function $\theta_{B_1}^{(1)}(\mu)=\varepsilon$. The term $\widehat{\mathbf{G}}_{1,111}^{(1)}$ represents the cubic nonlinearity for $u_1$.
The corresponding pOpInf problem is \cref{eq:opinf-standard-system} with $d=2$ and
\begin{align*}
\widehat{\mathbf{O}}_{1}
&= \left[\begin{array}{ccccccc}
\widehat{\mathbf{c}}_{1}^{(1)}
&
\widehat{\mathbf{B}}_{1}^{(1)}
&
\widehat{\mathbf{A}}_{1,1}^{(1)}
&
\widehat{\mathbf{A}}_{1,1}^{(2)}
&
\widehat{\mathbf{A}}_{1,2}^{(1)}
&
\widehat{\mathbf{H}}_{1,11}^{(1)}
&
\widehat{\mathbf{G}}_{1,111}^{(1)}
\end{array}\right] \in \mathbb{R}^{r_1 \times q_1(r_1,r_2)},
\\
\mathbf{D}_{1}
&= \left[\begin{array}{c|c|c|c}
\mathbf{D}_{c_1} &
\mathbf{D}_{A_1} &
\mathbf{D}_{H_1} &
\mathbf{D}_{G_1}
\end{array}\right] \in \mathbb{R}^{sK \times q_1(r_1,r_2)},
\\
\mathbf{D}_{c_1}
&= \left[\begin{array}{cc}
\theta_{c_1}^{(1)}(\mu_1)\mathbf{1}_{K}
&
\theta_{B_1}^{(1)}(\mu_1)\mathbf{f}
\\
\vdots & \vdots
\\
\theta_{c_1}^{(1)}(\mu_s)\mathbf{1}_{K}
&
\theta_{B_1}^{(1)}(\mu_s)\mathbf{f}
\end{array}\right],
\\
\mathbf{D}_{A_1}
&= \left[\begin{array}{ccc}
\theta_{A_{1,1}}^{(1)}(\mu_1)\widehat{\mathbf{U}}_{1}(\mu_{1})^{\top}
&
\theta_{A_{1,1}}^{(2)}(\mu_1)\widehat{\mathbf{U}}_{1}(\mu_{1})^{\top}
&
\theta_{A_{1,2}}^{(1)}(\mu_1)\widehat{\mathbf{U}}_{2}(\mu_{1})^{\top}
\\
\vdots & \vdots & \vdots
\\
\theta_{A_{1,1}}^{(1)}(\mu_s)\widehat{\mathbf{U}}_{1}(\mu_{s})^{\top}
&
\theta_{A_{1,1}}^{(2)}(\mu_s)\widehat{\mathbf{U}}_{1}(\mu_{s})^{\top}
&
\theta_{A_{1,2}}^{(1)}(\mu_s)\widehat{\mathbf{U}}_{2}(\mu_{s})^{\top}
\end{array}\right],
\\
\mathbf{D}_{H_1}
&= \left[\begin{array}{c}
\theta_{H_{1,11}}^{(1)}(\mu_1)\left(\widehat{\mathbf{U}}_{1}(\mu_{1})\,\widehat{\odot}\,\widehat{\mathbf{U}}_{1}(\mu_{1})\right)^{\top}
\\
\vdots
\\
\theta_{H_{1,11}}^{(1)}(\mu_s)\left(\widehat{\mathbf{U}}_{1}(\mu_{s})\,\widehat{\odot}\,\widehat{\mathbf{U}}_{1}(\mu_{s})\right)^{\top}
\end{array}\right],
\\
\mathbf{D}_{G_1}
&= \left[\begin{array}{c}
\theta_{G_{1,111}}^{(1)}(\mu_1)\left(\widehat{\mathbf{U}}_{1}(\mu_{1})\,\widehat{\odot}\,\widehat{\mathbf{U}}_{1}(\mu_{1})\,\widehat{\odot}\,\widehat{\mathbf{U}}_{1}(\mu_{1})\right)^{\top}
\\
\vdots
\\
\theta_{G_{1,111}}^{(1)}(\mu_s)\left(\widehat{\mathbf{U}}_{1}(\mu_{s})\,\widehat{\odot}\,\widehat{\mathbf{U}}_{1}(\mu_{s})\,\widehat{\odot}\,\widehat{\mathbf{U}}_{1}(\mu_{s})\right)^{\top}
\end{array}\right],
\\
\widehat{\mathbf{O}}_{2}
&= \left[\begin{array}{ccc}
\widehat{\mathbf{c}}_{2}^{(1)}
&
\widehat{\mathbf{A}}_{2,1}^{(1)}
&
\widehat{\mathbf{A}}_{2,2}^{(1)}
\end{array}\right] \in \mathbb{R}^{r_{2} \times q_2(r_1,r_2)},
\\
\mathbf{D}_{2}
&= \left[\begin{array}{ccc}
\theta_{c_2}^{(1)}(\mu_1)\mathbf{1}_{K}
&
\theta_{A_{2,1}}^{(1)}(\mu_1)\widehat{\mathbf{U}}_{1}(\mu_{1})^\top
&
\theta_{A_{2,2}}^{(1)}(\mu_1)\widehat{\mathbf{U}}_{2}(\mu_{1})^\top
\\
\vdots & \vdots & \vdots
\\
\theta_{c_2}^{(1)}(\mu_s)\mathbf{1}_{K}
&
\theta_{A_{2,1}}^{(1)}(\mu_s)\widehat{\mathbf{U}}_{1}(\mu_{s})^\top
&
\theta_{A_{2,2}}^{(1)}(\mu_s)\widehat{\mathbf{U}}_{2}(\mu_{s})^\top
\end{array}\right] \in \mathbb{R}^{sK \times q_2(r_1,r_2)},
\end{align*}
where $\mathbf{f} = [f(t_1)~\cdots~f(t_K)]^\top\in\mathbb{R}^{K}$,
$q_{1}(r_1,r_2) = 2 + 2r_{1} + r_{2} + \binom{r_{1}+1}{2} + \binom{r_{1}+2}{3}$,
and $q_{2}(r_1,r_2) = 1 + r_{1} + r_{2}$. The data matrix $\mathbf{D}_1$ and the operator matrix $\widehat{\mathbf{O}}_1$ have been modified from \cref{eq:ode-system}--\cref{eq:operator-matrix-system} to account for the cubic term present \cref{eq:fhn-system}.
\end{example}
\begin{remark}
\label{remark:FHN-DEIM}
Even though since \cite{CS2010deim}, the FitzHugh-Nagumo system has been widely used as a benchmark problem for the development of general nonlinear model reduction methods, a fully cubic intrusive model can be directly derived for this system (i.e., the nonlinear terms in the system are point-wise local and exactly quadratic and cubic), circumventing the need for the second layer of approximation introduced through hyper-reduction.
\end{remark}
\begingroup
\renewcommand{\arraystretch}{1.5}
\begin{table}[ht]
\centering
{\footnotesize
\caption{The operators in the continuous setting, the associated affine coefficient functions, and the size of the reduced operators in discretized setting for the FitzHugh-Nagumo system of \cref{example:fh-n}.}
\label{table:fhn-terms}
\begin{center}
\begin{tabular}{cllll}
\toprule
& & continuous term & affine coefficient & discretized operator\\
\midrule
\multirow{7}{*}{\rotatebox[origin=c]{90}{Eq.~\cref{eq:fhn-system}}} & constant & $\mathcal{C}_{1}^{(1)} =1$ & $\theta_{c_1}^{(1)}(\mu)= {\alpha}/{\varepsilon}$ & $\widehat{\mathbf{c}}_{1}^{(1)}\in \mathbb{R}^{r_1}$\\
& input & $\frac{\partial u_1}{\partial x}\big|_{x=0} = f(t)$ & $\theta_{B_1}^{(1)}(\mu)=\varepsilon$ & $\widehat{\mathbf{B}}_{1}^{(1)}\in \mathbb{R}^{r_1}$\\
& \multirow{3}{*}{linear} & $\mathcal{A}_{1,1}^{(1)}(u) = \frac{\partial^{2} u}{\partial x^{2}}$ & $\theta_{A_{1,1}}^{(1)}(\mu) = \varepsilon$ & $\widehat{\mathbf{A}}_{1,1}^{(1)}\in \mathbb{R}^{r_1\times r_1}$\\
& & $\mathcal{A}_{1,1}^{(2)}(u) =u$ & $\theta_{A_{1,1}}^{(2)}(\mu) = -{0.1}/{\varepsilon}$ & $\widehat{\mathbf{A}}_{1,1}^{(2)}\in \mathbb{R}^{r_1\times r_1}$\\
& & $\mathcal{A}_{1,2}^{(1)}(u) =u$ & $\theta_{A_{1,2}}^{(1)}(\mu) = -{1}/{\varepsilon}$ & $\widehat{\mathbf{A}}_{1,2}^{(1)}\in \mathbb{R}^{r_1\times r_2}$\\
& quadratic & $\mathcal{H}_{1,11}^{(1)}(u,v) = uv$ & $\theta_{H_{1,11}}^{(1)}(\mu) = {1.1}/{\varepsilon}$ & $\widehat{\mathbf{H}}_{1,11}^{(1)}\in \mathbb{R}^{r_1\times \binom{r_1+1}{2}}$\\
& cubic & $\mathcal{G}_{1,111}^{(1)}(u,v,w) = uvw$ & $\theta_{G_{1,111}}^{(1)}(\mu) = -{1}/{\varepsilon}$ & $\widehat{\mathbf{G}}_{1,111}^{(1)}\in \mathbb{R}^{r_1\times \binom{r_1+2}{3}}$\\
\midrule
\multirow{3}{*}{\rotatebox[origin=c]{90}{Eq.~\cref{eq:fhn-system2}}} & constant & $\mathcal{C}_{2}^{(1)} = 1$ & $\theta_{c_2}^{(1)}(\mu)= \alpha$ & $\widehat{\mathbf{c}}_{2}^{(1)}\in \mathbb{R}^{r_2}$\\
& \multirow{2}{*}{linear} & $\mathcal{A}_{2,1}^{(1)}(u) =u$ & $\theta_{A_{2,1}}^{(1)}(\mu) = \beta$ & $\widehat{\mathbf{A}}_{2,1}^{(1)}\in \mathbb{R}^{r_2\times r_1}$\\
& & $\mathcal{A}_{2,2}^{(1)}(u) =u$ & $\theta_{A_{2,2}}^{(1)}(\mu) = -\gamma$ & $\widehat{\mathbf{A}}_{2,2}^{(1)}\in \mathbb{R}^{r_2\times r_2}$\\
\bottomrule
\end{tabular}
\end{center}
}
\end{table}
\endgroup
\section{Computational Procedure} \label{sec:computation}
Solving the pOpInf problem \cref{eq:opinf-standard} requires samples of the solution $u(x,t;\mu)$ and its time derivative at times $\{t_{j}\}_{j=1}^{K}$ for each selected parameter value $\{\mu_{i}\}_{i=1}^{s}$. The quality of the resulting ROM depends on how well the orthonormal basis functions $\{v_{j}\}_{j=1}^{r}$ represent the solution at each parameter value and throughout the spatial and temporal domains. We therefore adopt the widely used proper orthogonal decomposition (POD) \cite{algazi1969optimality,Berkooz1993,lumley,sirovich1987turbulence}, defined by the set of orthonormal functions that minimize the mean squared projection error of the sample data, i.e., solving the problem
\begin{align*}
&\min_{v_{1},\ldots,v_{r}\in\mathcal{V}}
\sum_{i=1}^{s}\sum_{j=1}^{K}\left\|
u(\cdot,t_{j};\mu_{i})
- \sum_{\ell=1}^{r}\big\langle
v_{\ell},u(\cdot,t_j;\mu_i)
\big\rangle v_{\ell}
\right\|_{\mathcal{V}}^2
&
&\textrm{subject to}
&
&\langle v_{i}, v_{j} \rangle = \delta_{ij},
\end{align*}
where $\|v\|_{\mathcal{V}} = \sqrt{\langle v,v\rangle}$ is the natural norm on $\mathcal{V}$.
This data-driven choice of basis optimally represents the solution at the sampled parameter values, although the resulting ROM does not share such guarantees \cite{BGW2015pmorSurvery}.
\begin{purpose}
Choice of regularization and how it helps with the conditioning.
\end{purpose}
The conditioning of the linear least-squares problem \cref{eq:opinf-standard} depends on the data matrix $\mathbf{D}$. If the parameter samples are chosen so that the pitfalls described in \cref{thm:rank-deficiencies} are avoided, then the condition number of $\mathbf{D}$ depends on the nature of the solution at the quadrature points $\{t_{j}\}_{j=1}^{K}$. To improve the conditioning, we introduce a Tikhonov regularization \cite{Tikhonov1977regularization} so that \cref{eq:opinf-standard} becomes
\begin{align}
\label{eq:opinf-reg}
&\min_{\widehat{\mathbf{O}}}\left\|
\mathbf{D}\widehat{\mathbf{O}}^\top - \mathbf{R}^\top
\right\|_{F}^2
+ \left\|\boldsymbol{\Lambda}\widehat{\mathbf{O}}^\top\right\|_F^2,
&
& \boldsymbol{\Lambda} \in\mathbb{R}^{q(r)\times q(r)}.
\end{align}
The solution to this regularized problem satisfies the modified normal equations,
\begin{align}
\label{eq:normal-equations}
\left(
\mathbf{D}^\top\mathbf{D}
+ \boldsymbol{\Lambda}^\top\boldsymbol{\Lambda}
\right)
\widehat{\mathbf{O}}^\top
= \mathbf{D}^{\top}\mathbf{R}^\top.
\end{align}
The regularizer $\boldsymbol{\Lambda}$ can be parameterized in a number of ways \cite{MHW2021regOpInfCombustion}.
One choice that provides flexibility without introducing a large number of hyperparameters is the diagonal matrix $\boldsymbol{\Lambda} = \boldsymbol{\Lambda}(\lambda_1,\lambda_2)$, defined such that
\begin{align}
\label{eq:regularization-explicit}
\left\|
\boldsymbol{\Lambda}(\lambda_1,\lambda_2)
\widehat{\mathbf{O}}^\top
\right\|_F^2
= \lambda_1^{2}\left(
\sum_{p=1}^{q_{c}}\left\|\widehat{\mathbf{c}}^{(p)}\right\|_{2}^{2} +
\sum_{p=1}^{q_{A}}\left\|\widehat{\mathbf{A}}^{(p)}\right\|_{F}^{2}
\right)
+ \lambda_2^{2}
\sum_{p=1}^{q_{H}}\left\|\widehat{\mathbf{H}}^{(p)}\right\|_{F}^{2}.
\end{align}
This regularization structure groups the operators defining the ROM according to their polynomial order and drives $\widehat{\mathbf{O}}$ toward the zero matrix as $\lambda_{1},\lambda_{2}\to \infty$. Therefore, the regularization drives the resulting ROM toward the globally stable zero system $\frac{\textup{d}}{\textup{d}t}\widehat{\mathbf{u}}(t;\mu) = \mathbf{0}$.
As before, the extension of \cref{eq:opinf-reg} to a system of PDEs is straightforward where each PDE is written as an independent pOpInf problem (see \cref{eq:opinf-standard-system}).
\begin{purpose}
Hyperparameter selection via optimization.
\end{purpose}
We choose the regularization hyperparameters $\lambda_1,\lambda_2 \ge 0$ to minimize the mean squared training error
\begin{align*}
\frac{1}{s}\sum_{i=1}^{s}\sum_{j=1}^{K}\left\|
\widehat{\mathbf{u}}(t_{j};\mu_{i}) -
\widetilde{\mathbf{u}}(t_{j};\mu_{i})
\right\|_{2}^{2},
\end{align*}
where $\widehat{\mathbf{u}}(t_{j};\mu_{i})$ is the training data and $\widetilde{\mathbf{u}}(t_{j};\mu_{i})$ is the result of integrating the ODE $\frac{\textrm{d}}{\textrm{d}t}\widetilde{\mathbf{u}}(t) = \mathbf{F}(\widehat{\mathbf{O}};\widetilde{\mathbf{u}},t,\mu)$ defined by the solution $\widehat{\mathbf{O}}$ of \cref{eq:opinf-reg} with regularization hyperparameters $\lambda_1$ and $\lambda_2$.
This is an optimization problem in the principal quadrant of $\mathbb{R}^{2}$ (but is otherwise free of constraints), which we carry out with a sparse grid search followed by a derivative-free search method \cite{nelder1965simplex}.
\begin{purpose}
Simplifications for a monomial case.
\end{purpose}
\begin{purpose}
Summary of the computational procedure in words, followed by the algorithm.
\end{purpose}
We now summarize the computational procedure for solving \cref{eq:opinf-standard}: \hypertarget{steps:step1}{1)}~select parameter values $\{\mu_{i}\}_{i=1}^{s}$ such that the associated $\Theta_{c},\ldots,\Theta_{G}$ have full column rank; \hypertarget{steps:step2}{2)}~sample the PDE solution $u(x,t,\mu)$ and its time derivative for $t\in\{t_{j}\}_{j=1}^{K}$ for each $\mu\in\{\mu_{i}\}_{i=1}^{s}$; \hypertarget{steps:step3}{3)}~compute the POD basis associated with the sampled solution data; \hypertarget{steps:step4}{4)}~use the POD basis to project the solution data, obtaining $\widehat{\mathbf{U}}(\mu_i)$ and $\dot{\widehat{\mathbf{U}}}(\mu_{i})$; \hypertarget{steps:step5}{5)}~form the data matrix $\mathbf{D}$ and the time derivative matrix $\mathbf{R}$; \hypertarget{steps:step6}{6)}~choose optimal regularization hyperparameters and solve \cref{eq:opinf-reg} with these hyperparameters.
In the computational setting, we sample the solution in \hyperlink{steps:step2}{step 2} by obtaining approximate discretized \emph{solution snapshots} via a high-fidelity solver. For example, let $\{\mathbf{x}_{\ell}\}_{\ell=1}^{N}\subset\Omega$ be a discretization of $\Omega$, and define
\begin{align*}
\mathbf{U}(\mu_{i})
&= \left[\begin{array}{ccc}
u(\mathbf{x}_{1},t_{1};\mu_i) & \cdots & u(\mathbf{x}_{1},t_{K};\mu_i)
\\ \vdots & & \vdots \\
u(\mathbf{x}_{N},t_{1};\mu_i) & \cdots & u(\mathbf{x}_{N},t_{K};\mu_i)
\end{array}\right]\in\mathbb{R}^{N\times K},
\end{align*}
the \emph{snapshot matrix} for parameter $\mu_{i}$.
The rank-$r$ POD basis of \hyperlink{steps:step3}{step 3} is comprised of the first $r$ left singular vectors of the concatenated snapshot matrices, that is,
\begin{align}
\label{eq:pod-from-svd}
\boldsymbol{\Phi}\boldsymbol{\Sigma}\boldsymbol{\Psi}^\top
&= \left[\begin{array}{ccc}
& & \\
\mathbf{U}(\mu_{1}) & \cdots & \mathbf{U}(\mu_{s})
\\ & &
\end{array}\right]\in\mathbb{R}^{N\times sK},
&
\mathbf{V} &= \boldsymbol{\Phi}_{:,1:r}\in\mathbb{R}^{N\times r},
\end{align}
where $\boldsymbol{\Phi}\boldsymbol{\Sigma}\boldsymbol{\Psi}^\top$ is the singular value decomposition (SVD). With this notation, the projection of \hyperlink{steps:step4}{step 4} is given by $\widehat{\mathbf{U}}(\mu_i) = \mathbf{V}^\top\mathbf{U}(\mu_i)$, $i=1,\ldots,s$.
If the time derivatives of $u$ are not provided by the high-fidelity solver, they may be estimated as finite differences of the solution snapshots. The time integration error of the high-fidelity model, as well as the approximation error accompanying finite differences for the time derivatives, are additional motivations for utilizing the regularization strategy described previously.
\Cref{alg:OpInf-reg-multi} fully details the procedure.
The algorithm is presented for the general case of a system of $d$ partial differential equations, but it may be simplified to a case of a single PDE by setting $d=1$.
\begin{algorithm}
\begin{algorithmic}[1]
\Procedure{pOpInf}{training parameters $\mu_{1},\ldots,\mu_{s}\in\mathcal{P}$,
high-fidelity training snapshots $\mathbf{U}_{1}(\mu_{i}),\ldots,\mathbf{U}_{d}(\mu_{i})\in\mathbb{R}^{N\times K}$ for $i=1,\ldots,s$,
affine coefficient functions $\theta = \{\theta_{c_1}^{(1)},\ldots,\theta_{H_{d,dd}}^{(q_{H_{d,dd}})}\}:\mathcal{P}\to\mathbb{R}$,
reduced dimensions $r_{1},\ldots,r_{d}\in\mathbb{N}$
}
\vspace{.0625in}
\LineComment{Project training data to low-dimensional subspaces.}
\For{$\ell = 1, \ldots, d$}
\State $\mathbf{V}_{\ell} \gets\ $\texttt{pod}$\left([\mathbf{U}_{\ell}(\mu_1)~\cdots~\mathbf{U}_{\ell}(\mu_s)],r_{\ell}\right)$
\Comment{Rank-$r_\ell$ POD basis.}
\For{$i=1,\ldots,s$}
\State $\widehat{\mathbf{U}}_{\ell}(\mu_i) \gets \mathbf{V}_{\ell}^\top\mathbf{U}_{\ell}(\mu_i)$
\Comment{Projected solution data.}
\State $\dot{\widehat{\mathbf{U}}}_{\ell}(\mu_i) \gets \frac{\textup{d}}{\textup{d}t}\widehat{\mathbf{U}}_{\ell}(\mu_i)$
\Comment{Projected time derivatives.}
\EndFor
\EndFor
\LineComment{Construct pOpInf matrices.}
\For{$\ell = 1,\ldots,d$}
\State $\mathbf{D}_{\ell} \gets$ build the $\ell$th data matrix from $ \widehat{\mathbf{U}}_{1}(\mu_{1}),\ldots,\widehat{\mathbf{U}}_{d}(\mu_{s}),\theta,\mu_{1},\ldots,\mu_{s}$
\State $\mathbf{R}_{\ell} \gets [\dot{\widehat{\mathbf{U}}}_{\ell}(\mu_1)~\cdots~\dot{\widehat{\mathbf{U}}}_{\ell}(\mu_s)]$
\EndFor
\LineComment{Compute pOpInf solution with optimal hyperparameters.}
\State $\lambda_1^{*},\lambda_2^{*} \gets$ \textrm{argmin}
\textproc{TrainingError}$(\lambda_1,\lambda_2)$ \State \textbf{return} $\textproc{RegOpInf}(\lambda_{1}^{*},\lambda_{2}^{*})$
\EndProcedure
\vspace{6pt}
\setcounter{ALG@line}{0}
\Procedure{TrainingError} {$\lambda_1,\lambda_2$}
\State $\widehat{\mathbf{O}}_{1},\ldots,\widehat{\mathbf{O}}_{d} \gets \textproc{RegOpInf}(\lambda_{1},\lambda_{2})$
\For{$i = 1, \ldots, s$}
\Comment{Calculate error at training parameters.}
\State $\widetilde{\mathbf{U}}_{1}(\mu_{i}),\ldots,\widetilde{\mathbf{U}}_{d}(\mu_{i}) \gets\ $ integrate \cref{eq:ode-system}, $\ell=1,\dots,d$, over $[t_0,t_f]$
\EndFor
\State \textbf{return} $\frac{1}{sd}\sum_{\ell=1}^{d}\sum_{i=1}^{s}\|\widehat{\mathbf{U}}_{\ell}(\mu_{i}) - \widetilde{\mathbf{U}}_{\ell}(\mu_{i})\|_F^2$
\EndProcedure
\vspace{6pt}
\setcounter{ALG@line}{0}
\Procedure{RegOpInf}{$\lambda_1,\lambda_2$}
\State $\boldsymbol{\Lambda}^2 \gets \boldsymbol{\Lambda}(\lambda_1,\lambda_2)^\top \boldsymbol{\Lambda}(\lambda_1,\lambda_2)$
\Comment{Construct the regularizer.}
\For{$\ell = 1, \ldots, d$}
\State $\widehat{\mathbf{O}}_{\ell}^\top \gets \left(\mathbf{D}_{\ell}^\top\mathbf{D}_{\ell} + \boldsymbol{\Lambda}^2\right)^{-1}\mathbf{D}_{\ell}^\top\mathbf{R}_{\ell}^\top$
\Comment{Solve the $\ell$th pOpInf problem.}
\EndFor
\State \textbf{return} $\widehat{\mathbf{O}}_{1},\ldots,\widehat{\mathbf{O}}_{d}$
\EndProcedure
\end{algorithmic}
\caption{Regularized \textbf{p}arametric \textbf{Op}erator \textbf{Inf}erence for systems of PDEs}
\label{alg:OpInf-reg-multi}
\end{algorithm}
\section{Numerical Examples} \label{sec:numerics}
We now present numerical results for the heat equation and FitzHugh-Nagumo system introduced in \Cref{sec:methodology} and \Cref{sec:pde_systems}, respectively.
\subsection{Heat Equation} \label{subsec:numerical_heat}
We return to the heat equation of \cref{example:heat-1} and \ref{example:heat-2}, setting $\bar{x} = 2/3$
and
$
u_0(x;\mu)
= 1 - \left(1 - x\right)^{50} - x^{50}.
$
For each of the $s = 5$ parameter samples shown in \cref{fig:heat:training-samples}, we generate high-fidelity solutions by discretizing the spatial domain $\Omega = (0,1)$ with a uniform grid of $N = 1000$ points and approximating the spatial derivative with second-order central finite differences, then integrating the resulting semi-discrete ODE in time with the first-order implicit Euler scheme on $K = 1500$ uniformly spaced time steps in $[t_0,t_f] = [0,1.5]$. \Cref{fig:heat:training-samples} also shows example snapshots for each parameter sample, demonstrating that variation in the parameters $\mu = (\alpha,\beta)$ determines the diffusion dynamics. To select the number of modes in the POD basis, define the cumulative energy
\begin{align*}
\mathcal{E}(r)
= \sum_{j=1}^{r}\sigma_{j}^{2} \bigg/\sum_{j=1}^{N}\sigma_{j}^{2},
\end{align*}
where $\sigma_{j}$ is the $j$th singular value in the POD factorization \cref{eq:pod-from-svd}. Note that $\mathcal{E}(r)$ is a nondecreasing function of $r$. We select $r$ to be the smallest integer such that the residual energy $1 - \mathcal{E}(r)$, or the energy in the non-retained modes, lies below a fixed threshold $\epsilon > 0$ (see \cref{fig:heat:basis}). In this case, setting $\epsilon = 10^{-12}$ results in $r = 24$.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{figure1.pdf}
\vspace{-0.75cm}
\caption{Experimental parameter samples (left) and associated snapshots at intermediate time $t = t_{75} = 0.075$ (right) for the heat equation problem \cref{eq:heat-pde}--\cref{eq:heat-bcs}. The vertical line $x = \bar{x}$ marks the point in the domain where the diffusion constant switches between $\alpha$ and $\beta$.}
\label{fig:heat:training-samples}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\textwidth]{figure2.pdf}
\vspace{-0.5cm}
\caption{Residual energy decay (left) and the six dominant POD basis functions (right) for the snapshot set generated at the parameter samples in \cref{fig:heat:training-samples}. For $r \ge 24$, we have $1 - \mathcal{E}(r) < 10^{-12}$.} \label{fig:heat:basis}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.85\textwidth]{figure3.png}
\vspace{-0.5cm}
\caption{Relative $L^2$ projection errors (left) and pOpInf ROM errors (right) over the parameter domain $\mathcal{P}$ for the problem \cref{eq:heat-pde}--\cref{eq:heat-bcs}. The parameter samples used to generate the training set are marked as stars and $r=24$ basis functions are used. The geometric means of the projection and ROM errors are approximately $7.68 \times 10^{-7}$ and $3.48 \times 10^{-5}$, respectively.}
\label{fig:heat:param-errors}
\end{figure}
We compute $\widehat{\mathbf{O}}$ using \cref{alg:OpInf-reg-multi}.
Since \cref{eq:heat-ode} is a linear system, the regularization is parameterized by a single hyperparameter, i.e., we minimize the residual
\begin{align*}
\min_{\widehat{\mathbf{O}}}\left\|
\mathbf{D}\widehat{\mathbf{O}}^{\top} - \mathbf{R}^{\top}
\right\|_{F}^{2}
+ \lambda_{1}^{2}\left(
\|\widehat{\mathbf{A}}^{(1)}\|_{F}^{2} + \|\widehat{\mathbf{A}}^{(2)}\|_{F}^{2}
\right).
\end{align*}
To evaluate the performance of the resulting ROM in terms of the parameters, we discretize $\mathcal{P}$ in a $40\times 40$ uniform grid and, for each $\mu$ in the grid, compute high-fidelity solutions
$\mathbf{u}(t_{1};\mu),\ldots,\mathbf{u}(t_{K};\mu)\in\mathbb{R}^{N}$ and integrate the ROM to obtain reduced states $\widetilde{\mathbf{u}}(t_{1};\mu),\ldots,\widetilde{\mathbf{u}}(t_{K};\mu)\in\mathbb{R}^{r}$, then compute the relative $L_{2}$-norm error
\begin{align}
\label{eq:error-reconstruction}
&\frac{\left\|
\mathbf{V}\widetilde{\mathbf{u}}(\cdot\,;\mu) - \mathbf{u}(\cdot\,;\mu)
\right\|_{L^{2}([t_0,t_f])}}{\left\|
\mathbf{u}(\cdot\,;\mu)
\right\|_{L^{2}([t_0,t_f])}},
&
\left\|\mathbf{w}(\cdot)\right\|_{L^{2}([t_0,t_f])}
= \left(\int_{t_0}^{t_f}\left\|\mathbf{w}(t)\right\|_{2}^{2}\:dt\right)^{1/2},
\end{align}
estimating the time integrals via the trapezoidal rule.
\Cref{fig:heat:param-errors} shows the results and compares them to the relative projection error induced by the basis, given by
\begin{align}
\label{eq:error-projection}
\frac{\left\|
\mathbf{u}(\cdot\,;\mu)
- \mathbf{V}\mathbf{V}^{\top}\mathbf{u}(\cdot\,;\mu)
\right\|_{L^{2}([t_0,t_f])}}{\left\|
\mathbf{u}(\cdot\,;\mu)
\right\|_{L^{2}([t_0,t_f])}}.
\end{align}
The relative projection error \cref{eq:error-projection} is on the order of $10^{-7}$ throughout $\mathcal{P}$, which indicates that the training snapshots have sufficient information to represent the solution well for any $\mu\in\mathcal{P}$.
The ROM error \cref{eq:error-reconstruction} is on the order of $10^{-5}$ throughout $\mathcal{P}$ (less than $0.02\%$ everywhere) and runs with a computational speedup of $\sim 10$ times compared to the high-fidelity model. The results in \cref{fig:heat:param-errors} highlight the ability of the ROM to generalize beyond the training data: in particular, the ROM performs well for $(\alpha,\beta)$ pairs away from the arc $\alpha^{2} + \beta^{2} = 4$ where the parameter samples lie.
\subsection{FitzHugh-Nagumo System} \label{subsec:numerical_fhn}
The neuron model \cref{eq:fhn-system}--\cref{eq:fhn-system2} introduced in \cref{example:fh-n} features a four-dimensional parameter space.
We generate training data at each of the $504 = 6\times 6 \times 2 \times 7$ unique parameters $\mu = (\alpha,\beta,\gamma,\varepsilon)$ for
\begin{align}
\tag{training set}
\begin{aligned}
\alpha &\in \{0.025, 0.035, \ldots, 0.075\},
&
\beta &\in \{0.25, 0.35, \ldots, 0.75\},
\\
\gamma &\in \{2.0, 2.5\},
&
\varepsilon &\in \{0.010, 0.015, \ldots, 0.040\}.
\end{aligned}
\end{align}
The full-order problem is solved by discretizing the domain $\Omega = (0,1)$ with $N_{x} = 512$ spatial points (so the total spatial dimension is $N = 2N_{x} = 1024$), approximating the differential term with central finite differences, and integrating the equation in time via the first-order implicit-explicit Euler scheme (treating the linear terms with a backward Euler step and the nonlinear terms with a forward Euler step \cite{ascher1995imex}) with step size $\delta t = 10^{-3}$ over $[t_0,t_f] = [0,4]$. Every $10$th snapshot is recorded, resulting in $K = 400$ training snapshots per parameter sample.
\Cref{fig:fhn-training-data} shows the phase plot of $u_{1}$ against $u_{2}$ at multiple spatial coordinate values for three of the training parameter samples, demonstrating that the system exhibits a diverse range of dynamical behaviors as the parameter values are varied.
To evaluate ROM performance with respect to the parameters, we also solve the full-order model at the $10{,}749 = 11\times 11\times 3\times 31 - 504$ additional parameters $\mu = (\alpha,\beta,\gamma,\varepsilon)$ with
\begin{align}
\tag{testing set}
\begin{aligned}
\alpha &\in \{0.025, 0.030, \ldots, 0.075\},
&
\beta &\in \{0.25, 0.30, \ldots, 0.75\},
\\
\gamma &\in \{2.00, 2.25, 2.50\},
&
\varepsilon &\in \{0.010, 0.011, \ldots, 0.040\}.
\end{aligned}
\end{align}
Parameters from the training set are not included in the testing set.
\begin{figure}[t]
\centering
\includegraphics[width=\textwidth]{figure4.pdf}
\vspace{-0.5cm}
\caption{Phase portraits of training trajectories for the FitzHugh-Nagumo system \cref{eq:fhn-system}--\cref{eq:fhn-system2}, traced out at various points in the spatial domain. The center trajectory has a limit cycle, while the trajectories on the left and right converge to a single point. Small $\varepsilon$ values de-emphasize the diffusion term and drive the system toward spatial homogeneity; larger $\varepsilon$ values result in more variation across the spatial domain but which decreases with time.}
\label{fig:fhn-training-data}
\end{figure}
\begin{figure}[t]
\begin{minipage}[b]{0.52\textwidth}
\begin{figure}[H]
\centering
\includegraphics[width=\textwidth]{figure5.pdf}
\end{figure}
\end{minipage}
\hfill
\begin{minipage}[b]{0.45\textwidth}
\centering
{\footnotesize
\begin{center}
\begin{tabular}{c|cc}
$1 - \mathcal{E}(r_{\ell})$ & $r_{1}$ & $r_{2}$
\\ \hline
$10^{-3}$ & $ 3$ & $ 2$ \\
$10^{-4}$ & $ 4$ & $ 4$ \\
$10^{-5}$ & $ 7$ & $ 5$ \\
$10^{-6}$ & $ 9$ & $ 7$ \\
$10^{-7}$ & $12$ & $ 9$ \\
$10^{-8}$ & $14$ & $11$ \\
$10^{-9}$ & $16$ & $13$ \\
$10^{-10}$ & $19$ & $15$ \\
$10^{-11}$ & $21$ & $17$ \\
$10^{-12}$ & $24$ & $19$ \\
\end{tabular}
\end{center}
}
\vfill
\end{minipage}
\caption{Decay of the residual energy in the training data for $u_1$ and $u_2$ in the FitzHugh-Nagumo problem \cref{eq:fhn-system}--\cref{eq:fhn-system2} (left) and the corresponding selected basis sizes (right). Demanding that $1 - \mathcal{E}(r_\ell) < 10^{-7}$ requires $r_{1} = 12$ and $r_{2} = 9$ POD modes for $u_{1}$ and $u_{2}$, respectively.}
\label{fig:fhn-residual-energy}
\end{figure}
The FitzHugh-Nagumo system \cref{eq:fhn-system}--\cref{eq:fhn-system2} exhibits a rapid singular Hopf bifurcation with respect to $\varepsilon$ wherein the limit cycle collapses to a stable fixed point as $\varepsilon$ increases \cite{baer1986singular}. At parameters near this transition, the full-order model and ROM solutions are highly sensitive to $\varepsilon$ and the time step $\delta t$. Including such parameters in the training and testing sets makes it difficult to assess ROM accuracy. We select $\delta t = 10^{-3}$ as the baseline time step for which most of the full-order simulations converge, then eliminate the small number of points from our training and testing sets for which refining the time step from $\delta t = 10^{-3}$ to $\delta t = 10^{-5}$ results in a $30\%$ change in the full-order solution as measured by the $L^2([t_0,t_f])$ norm. This prompts us to remove five training parameters ($\sim 1\%$ of the original training set) and fifty-two testing parameters ($\sim 0.5\%$ of the original testing set). Hence the size of the final training set is $s = 504 - 5 = 499$ and the size of the final testing set is $10{,}749 - 52 = 10{,}697$, with size ratio $10{,}697/499 \approx 21.4$. With this approach, ROMs obtained through pOpInf are stable throughout
almost the entire testing set; for the few testing parameters where the ROMs are unstable with time step $\delta t = 10^{-3}$, the time step is adjusted to maintain stability.
We select the number of POD modes $r_{1}$ and $r_{2}$ for the variables $u_{1}$ and $u_{2}$, respectively, based on the residual energy $1 - \mathcal{E}(r_{\ell})$ of the training data, $\ell=1,2$. For each of the resulting $(r_{1},r_{2})$ pairs shown in \Cref{fig:fhn-residual-energy}, we learn a pOpInf ROM via \cref{alg:OpInf-reg-multi}. To simplify the hyperparameter search, we regularize only the quadratic and cubic terms, so that the regularized pOpInf data residuals are
\begin{align*}
\min_{\widehat{\mathbf{O}}_{1}}
&\left\|
\mathbf{D}_{1}\widehat{\mathbf{O}}_{1}^{\top} - \mathbf{R}_{1}^{\top}
\right\|_{F}^{2}
+ \lambda_{2}\left\|\widehat{\mathbf{H}}_{1,11}^{(1)}\right\|_{F}^{2}
+ \lambda_{3}\left\|\widehat{\mathbf{G}}_{1,111}^{(1)}\right\|_{F}^{2},
&
\min_{\widehat{\mathbf{O}}_{2}}
&\left\|
\mathbf{D}_{2}\widehat{\mathbf{O}}_{2}^{\top} - \mathbf{R}_{2}^{\top}
\right\|_{F}^{2}.
\end{align*}
Regularization of the additional terms ($\lambda_{1}$ in \cref{eq:regularization-explicit}) was observed to have a marginal effect on the results. For each chosen basis size, we also compute a ROM based on intrusive projection for comparison with the pOpInf ROM. The intrusive ROM is derived explicitly without any need for approximation of the nonlinear terms via DEIM (see \cref{remark:FHN-DEIM}). Both ROMs are integrated with time step $\delta t = 10^{-3}$ for each parameter in the testing set, and the relative error compared to the full-order model is computed as in \cref{eq:error-reconstruction}. This procedure yields a single error value for each point in the testing set.
The intrusive ROMs are stable throughout the testing set, as are the pOpInf ROMs except at a small number of testing parameters (one for $r_1=12, r_2=9$; two for $r_1=r_2=4$ and $r_1=14$, $r_2=11$; and nine for $r_1=9$, $r_2=7$). At these points, we adjust the time step to $\delta t = 10^{-2}$, producing stable results.
For larger basis sizes $(r_1 > 14$, $r_2 > 11$), the pOpInf ROMs are stable throughout the testing set with $\delta t = 10^{-3}$.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{figure6.pdf}
\vspace{-0.5cm}
\caption{Relative errors (calculated over space and time) of the learned pOpInf ROM and the intrusive ROM for the FitzHugh-Nagumo system. As residual energy decreases from left to right, the sizes of the underlying POD bases increase according to \Cref{fig:fhn-residual-energy}. The shaded regions show the $10\%$--$90\%$ interdecile range of the error across all training samples (left) or testing samples (right), with the corresponding median and maximum errors denoted by the lines for the pOpInf ROMs and dashed lines for the intrusive ROMs.}
\label{fig:fhn-basisVerror}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\textwidth]{figure7.pdf}
\vspace{-0.5cm}
\caption{Phase portraits of testing trajectories for the FitzHugh-Nagumo problem \cref{eq:fhn-system}--\cref{eq:fhn-system2}, traced out at various points in the spatial domain. The solid lines are the full-order trajectories, and the dashed lines are the outputs of the learned pOpInf ROM with $r_{1} = 12$ and $r_{2} = 9$ POD modes. The total ROM relative errors in space and time are, from left to right, $0.034\%$, $0.777\%$, and $0.009\%$; the median relative error on the testing set for $r_1=12$ and $r_2=9$ is $0.014\%$.}
\label{fig:fhn-fomromphase}
\end{figure}
\Cref{fig:fhn-basisVerror} shows the $10\%$ quantile, median, $90\%$ quantile, and maximum of the relative errors for each ROM. The median relative error is similar in the training and testing sets and decreases steadily as the basis sizes increase. The pOpInf ROMs generally outperform the intrusive ROMs. The tight interdecile range shows that the ROM error is mostly consistent throughout the parameter space; though not shown, the parameters with higher ROM error are also where the projection error is significantly higher than average.
For additional comparison, \Cref{fig:fhn-fomromphase} shows select trajectories of a single pOpInf ROM versus the full-order model. The reduced-order and full-order results are indistinguishable to the eye, which is representative of the ROM performance throughout the testing set.
\section{Conclusions} \label{sec:conclusions}
This paper has proposed a non-intrusive parametric model reduction method for parameterized PDEs based on the Operator Inference framework. The approach eliminates the need for interpolation by explicitly learning the reduced operators of the affine-parametric representation and uses an optimization-based regularization strategy to ensure well-posedness in the learning problem. The parametric ROMs can later be used in outer-loop applications to expedite model evaluations for any choice of the parameters. The efficacy of the method has been demonstrated for two numerical examples: a heat equation with a two-dimensional parametric space, and the FitzHugh-Nagumo system with a four-dimensional parametric space. The resulting ROMs are capable of capturing the behavior of the PDE for parameters outside of the training set and, as shown in the latter example, perform favorably on average when compared to the intrusive ROM with the same affine structure.
It was shown in the FitzHugh-Nagumo example that the learned ROMs successfully capture the inherently different behaviors of the system that come with changes in the parameters (see \Cref{fig:fhn-fomromphase}). Yet, as with other data-driven approaches, the quality of inferred ROMs depends strongly on the training set, and one cannot expect a data-driven ROM to produce a particular dynamical behavior that differs wildly from the training data. An important future direction is therefore the automation of an efficient parameter sampling strategy for the offline stage.
|
\section{Introduction}
\label{sec:intro}
\input{sections/intro}
\input{sections/background}
\section{Improving certified and standard accuracies with ViTs}
\label{sec:improve-accuracy}
\input{sections/accuracy}
\section{Faster inference with ViTs}
\label{sec:improve-speed}
\input{sections/speed}
\section{Related work}
\input{sections/related}
\section{Conclusion}
\input{sections/conclusion}
\section{Acknowledgements}
\input{sections/acknowledgement}
\clearpage
\small
\printbibliography
\clearpage
\subsection{ViTs outperform ResNets on image ablations.}
\label{sec:ablations}
We first isolate the effect of using a ViT instead of a ResNet as the base classifier for derandomized smoothing.
Specifically, we keep all smoothing parameters fixed and only vary the base classifier. Following \citet{levine2020randomized}, we use column ablations of width $b=4$ for CIFAR-10 and $b=19$ for ImageNet for both training and certification.
\paragraph{Ablation accuracy.}
The performance of derandomized smoothing entirely depends on whether the base classifier can accurately classify ablated images.
We thus measure the accuracy of ViTs and ResNets at classifying column ablated images across a range of evaluation ablation sizes as shown in Figure~\ref{fig:finetuned}. We find that ViTs are significantly more accurate on these ablations than comparably sized ResNets.
For example, on ImageNet, ViT-S has up to 12\% higher accuracy on ablations than ResNet-50.
\begin{figure}[!t]
\centering
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/finetuned/cifar10_ablation_acc.pdf}
\caption{CIFAR-10}
\end{subfigure}
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/finetuned/ImageNet_ablation_acc.pdf}
\caption{ImageNet}
\end{subfigure}
\caption{Accuracies on column-ablated images for models on CIFAR-10 and ImageNet. The models were trained on column ablations of width $b=19$ for ImageNet and $b=4$ for CIFAR-10, and evaluated on a range of ablation sizes. ViTs outperform ResNets on image ablations by a sizeable margin.}
\label{fig:finetuned}
\end{figure}
\begin{figure}[!t]
\centering
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[trim=0 0 0 0, clip, width=1\textwidth]{figures/finetuned_vs_patchsize/cifar10_cert_acc.pdf}
\caption{CIFAR-10}
\end{subfigure}
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=1\textwidth]{figures/finetuned_vs_patchsize/imagenet_cert_acc.pdf}
\caption{ImageNet}
\end{subfigure}
\caption{Certified accuracies for ViT and ResNet models on CIFAR-10 and ImageNet for various adversarial patch sizes. Certification was performed using a fixed ablation of size $b=4$ for CIFAR-10 and $b=19$ for ImageNet (as in \cite{levine2020randomized}).}
\label{fig:certified-vs-patchsize}
\end{figure}
\paragraph{Certified patch robustness.}
We next measure the effect of improved ablation accuracy on certified accuracy.
We find that using a ViT as the base classifier in derandomized smoothing substantially boosts certified accuracy compared to ResNets across a range of model sizes and adversarial patch sizes, as shown in Figure~\ref{fig:certified-vs-patchsize}. For example, against $32\times 32$ adversarial patches on ImageNet (2\% of the image), a smoothed ViT-S improves certified accuracy by 14\% over a smoothed ResNet-50, while the larger ViT-B reaches a certified accuracy of 39\%---well above the highest reported baseline of 26\% \citep{xiang2021patchguard}\footnote{The highest reported certified accuracy in the literature for this patch size on ImageNet is 26\% from PatchGuard \citep{xiang2021patchguard}. However, this defense uses a masking technique that is optimized for this particular patch size, and achieves 0\% certified accuracy against larger patches.}.
\paragraph{Standard accuracy.}
We further find that smoothed ViTs can mitigate the precipitous drop in standard accuracy observed in previously proposed certified defenses, particularly so for larger architectures and datasets.
Indeed, the smoothed ViT-B remains 69\% accurate on ImageNet---14.2\% higher standard accuracy than that of the best performing prior work (Table \ref{table:main_summary table}).
A full comparison between the performance of smoothed models and their non-robust counterparts can be found in Appendix \ref{app:tables}.
\subsection{Ablation size matters}
\label{sec:ablation_size}
In the previous section, we fixed the width of column ablations at $b=19$ for derandomized smoothing on ImageNet, following \cite{levine2020randomized}.
We now demonstrate that properly choosing the ablation size can improve the standard accuracy even further---by 4\% on ImageNet---without sacrificing certified performance.
\begin{figure}[!htbp]
\centering
\begin{subfigure}[b]{1\textwidth}
\centering
\includegraphics[width=.49\textwidth]{figures/finetuned/imagenet_cert_acc.pdf}
\includegraphics[width=.49\textwidth]{figures/finetuned/imagenet_smooth_acc.pdf}
\end{subfigure}
\caption{
Certified (left) and standard (right) accuracies for a collection of smoothed models trained with a fixed ablation size $b=19$ on ImageNet, and evaluated with varying ablation sizes. Certified accuracy remains stable across a range of ablation sizes, while standard accuracy substantially improves with larger ablations.
}
\label{fig:ablation-size-finetuned}
\end{figure}
Specifically, we take ImageNet models trained on column ablations with width $b=19$,
and change the smoothing procedure to use a different width at \emph{test} time.
We report the resulting standard and certified accuracies in Figure~\ref{fig:ablation-size-finetuned}, and defer additional experiments on changing the ablation size during training to Appendix~\ref{app:trainablations}.
Although \citet{levine2020randomized} found a steep trade-off between certified and standard accuracy in CIFAR-10 (which we verify in Appendix~\ref{app:testablations}), we find this to not be the case for ImageNet for either CNNs or ViTs.
We can thus substantially increase the ablation size to improve standard accuracy \emph{without} significantly dropping certified performance as shown in Figure~\ref{fig:ablation-size-finetuned}.
For example, increasing the width of column ablations to $b=37$ improves the standard accuracy of the smoothed ViT-B model by nearly $4\%$ to 73\% while maintaining a 38\% certified accuracy against $32\times 32$ patches. In addition to being 12\% higher than the standard accuracy of the best performing prior work, this model's standard accuracy is only 3\% lower than that of a \textit{non-robust} ResNet-50.
Thus, using smoothed ViTs, we can achieve state-of-the-art certified robustness to patch attacks in the ImageNet setting while attaining standard accuracies that are more comparable to those of non-robust ResNets.
\subsection{Practical inference speeds for block smoothing}
We first demonstrate how dropping masked tokens significantly increases the speed of evaluating block ablations for the base classifier. In Figure~\ref{fig:app-block-speed}, we show that dropping masked tokens substantially reduces the time needed to process 1024 block ablations for various sizes of ViTs. This directly leads to a 4.85x speedup for ViT-S with ablation size 75.
\begin{figure}[!htbp]
\centering
\includegraphics[width=.7\textwidth]{figures/block-smoothing/block_complexity_plot.pdf}
\caption{Average time to compute a forward pass for ViTs on 1024 block ablated images with varying ablation sizes with and without dropping masked tokens.}
\label{fig:app-block-speed}
\end{figure}
Even with this optimization, however, block smoothing is quite expensive.
A forward pass through the smoothed model still requires around 50k passes through the base classifier.
We thus leverage our second speedup from strided ablations and use \textit{strided} block smoothing.
Similar to strided column ablations, for a stride length of $s$, we only consider block ablations that are $s$ pixels apart, vertically and horizontally.
This changes the certification threshold $\Delta$ to be, $\Delta_{block+stride} = \lceil(m+s-1)/s\rceil^2$. With dropping fully masked tokens and using a stride of 10, a smoothed ViT-S using an ablation size of 75 is only 2.8x slower than a standard (non-robust) ResNet-50.
\paragraph{Certified accuracy.}
We find that, despite an systematic search over stride length and block size (both at training and evaluation), block smoothing on ImageNet remains significantly worse than column smoothing.
For example, with optimal stride and ablation size, we see up to 5\% lower certified accuracy than column smoothing on the largest model, ViT-B.
We checked a range of ablation sizes from $55$ to $115$ as well as three stride lengths $\{7,10,14\}$ (Figure \ref{appfig:block-vary-ablation-size}).
\begin{figure}[!htbp]
\centering
\begin{subfigure}{1\textwidth}
\includegraphics[width=1\textwidth]{figures/block-smoothing/vary_patch_size.pdf}
\caption{We fix the test-time ablation size at $b=75$ and plot the certified accuracy as a function of the adversarial patch size, for various stride length.}
\end{subfigure}
\begin{subfigure}{1\textwidth}
\includegraphics[width=1\textwidth]{figures/block-smoothing/vary_all.pdf}
\caption{We fix the adversarial patch size $m=32$ and plot the certified accuracy as a function of the test-time ablation size, for various stride length.}
\end{subfigure}
\caption{Strided block smoothing on ImageNet for a collection of ViT models trained with block ablations of size $b=75$.}
\label{appfig:block-vary-ablation-size}
\end{figure}
Similar to striding with column ablations, there is a significant amount of variability based on the stride length.
To pinpoint the effect of striding, we certify one of the best-performing block sizes ($b=75$) over a full range of strides from $s=1$ to $s=20$ (Figure \ref{appfig:block-vary-stride}).
This is a fairly expensive calculation, as using stride $s=1$ corresponds to the full block ablation with 50k ablations.
Even when using all possible block ablations ($s=1$), block smoothing does not improve over column smoothing.
However, we do find that certain stride lengths ($s=18$) can achieve similar performance to non-strided block ablations $(s=1)$, which means that we can speed up certification (by $18$x) without sacrificing certified accuracy. Thus, while our methods can make block smoothing computationally feasible, further investigation is needed to make block smoothing match column smoothing in terms of certified and standard accuracies.
\begin{figure}[!htbp]
\centering
\includegraphics[width=.5\textwidth]{figures/block-smoothing/high_variance.pdf}
\caption{Strided block smoothing on ImageNet for ViT-B with a fixed ablation size $b=75$. The reported certified accuracy are against adversarial patches of size $32\times32$. Note how some stride lengths ($s=18$ for example) can achieve similar performance to non-strided block ablations ($s=1$).}
\label{appfig:block-vary-stride}
\end{figure}
\subsection{Computational complexity of ViTs with dropped tokens}
\label{app:complexity}
We can now derive the computational complexity of the smoothed ViT when dropping tokens. Specifically, consider a ViT that divides an $h\times w$ pixel image into $p \times p$ patches, and positionally encodes them tokens with $d$ hidden dimensions.
Recall that a ViT has two operation types: \textit{attention operators} which scale quadratically with the number of tokens but linearly with hidden dimension $d$ and \textit{fully-connected operators} which scale linearly with the number of tokens but quadratically in $d$. Without dropping tokens, we have $hw/p^2$ tokens. A forward pass of processing an image ablation without dropping tokens thus has an overall complexity of
$$O\left(\left(\frac{hw}{p^2}\right)^2d + \left(\frac{hw}{p^2}\right)d^2\right)$$
where the first term corresponds to the attention operations, and the second term corresponds to the fully-connected operations.
For column ablations with width $b$, dropping masked tokens reduces the number of tokens to $hb/p^2$. The complexity of the forward pass to process an image ablation when dropping masked tokens (i.e \texttt{ProcessAblation}) then drops to
$$ O\left(\left(\frac{hb}{p^2}\right)^2d + \left(\frac{hb}{p^2}\right)d^2\right)$$
thus reducing the attention cost by a factor of $O(w^2/b^2)$ and the fully-connected cost by a factor of $O(w/b)$. In practice, the computation of fully-connected operations tends to dominate since $d > \frac{hw}{p^2}$.
Overall, a smoothed ViT with stride $s$ processes $w/s$ ablations. Thus, the overall complexity of the smoothed ViT is:
$$ O\left(\frac{w}{s}\left(\left(\frac{hb}{p^2}\right)^2d + \left(\frac{hb}{p^2}\right)d^2\right)\right)$$
\subsection{Effect of dropping tokens on speed}
We extend the timing experiments comparing ViTs and ResNets to a range of ablation sizes (previously presented in Table~\ref{tab:speed_resnet} from Section~\ref{sec:improve-speed} for a single column ablation size of $b=19$).
Empirically, even for substantially larger ablations, we find significantly faster training and inference times for ViTs over ResNets. In Figure \ref{appfig:scalability}, we compare the evaluation and training speeds for processing image ablations with ResNets and ViTs with dropped tokens.
\begin{figure}[!htbp]
\centering
\begin{subfigure}[b]{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/scalability/scalability_eval.pdf}
\caption{Forward Pass Time}
\end{subfigure}
\begin{subfigure}[b]{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/scalability/scalability_train.pdf}
\caption{Forward/Backward Pass Time}
\end{subfigure}
\caption{(a) Average time for computing a forward pass on a batch of 1024 image ablations on ImageNet (b) Average time for computing a full training step (forward and backward pass) on a batch of 128 image ablations on ImageNet}
\label{appfig:scalability}
\end{figure}
\subsection{Effect of dropping tokens on performance}
Since the tokens are individually positionally encoded, dropping tokens that are fully masked does not remove any information from the input. In Figure \ref{appfig:missingness-ablation-acc}, we confirm that dropping masked tokens does not significantly change the accuracy of the ViT base classifier on ablations.
\begin{figure}[!htbp]
\centering
\begin{subfigure}[b]{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/finetuned_missingness/cifar10_ablation_acc.pdf}
\caption{CIFAR-10 models.}
\end{subfigure}
\begin{subfigure}[b]{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/finetuned_missingness/imagenet_ablation_acc.pdf}
\caption{ImageNet models.}
\end{subfigure}
\caption{We compare the ablation accuracies of dropping masked tokens versus processing all tokens for a collection of vision transformers on CIFAR-10 and ImageNet. Dropping masked tokens does not substantially degrade accuracy.}
\label{appfig:missingness-ablation-acc}
\end{figure}
\subsection*{Ablation Accuracy}
\begin{figure}[!htbp]
\centering
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/pretrained/cifar10_ablation_acc.pdf}
\caption{Pretrained CIFAR-10 models.}
\end{subfigure}
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/pretrained/ImageNet_ablation_acc.pdf}
\caption{Pretrained ImageNet models.}
\end{subfigure}
\caption{Ablation accuracies for pretrained models on CIFAR-10 and ImageNet column-ablated images. ViTs outperforms ResNets by a significant margin, especially on ImageNet.}
\label{fig:pretrained}
\end{figure}
Without training on image ablations, we find that pretrained ViTs naturally maintain significantly higher accuracy on column ablations than comparably sized ResNets as shown in Figure \ref{fig:pretrained}.
This difference is most prominent on ImageNet---ViT-B has 30-40\% higher ablation accuracy across a range of ablation sizes than the largest residual network (WRN-101-2), which suggests that ViTs are naturally better suited for classifying image ablations.
Even when models are fine-tuned for image ablations, we find that a significant gap still remains.
For example, the ViT-S has 11-12\% higher ablation accuracy than the ResNet-50 over the same ablation sizes.
\subsection*{Certified Accuracy}
\paragraph{Pretrained vision transformers are certifiably robust}
\label{sec:pretrained}
In this subsection, we test how certifiably robust are of-the-shelf pretrained ViTs and ResNets when smoothed with derandomized smoothing. Recall that pretraind ViTs are able to classify small ablations of images much better than ResNets as we showed in Section~\ref{sec:ablations} and Figure~\ref{fig:pretrained} for various ablation sizes. This suggests that is we apply derandomized smoothing on ViTs, they will outperform ResNets. We apply derandomized smoothing, as described in Section~\ref{sec:background-derandomized}. We use column smoothing with ablation sizes of $s = 19$ and $s=XX$ on ImageNet and CIFAR-10, respectively. We sweep over the various ViTs and ResNets listed in see Section~\ref{sec:why}. ViTs consistently achieve substantially higher certificates of robustness than ResNets across all architecture as shown in Figure~TODO againts various sizes of adversarial patches. For example ViT-B achieves around $32\%$ certified accuracy on ImageNet against $42 \times 42$ adversarial patches, whereas a larger-size WRN-101-2 achieves only $2\%$.
\begin{figure}[!htbp]
\centering
\begin{subfigure}[b]{.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/pretrained/cifar10_cert_acc.pdf}
\caption{Pretrained CIFAR-10 models.}
\end{subfigure}
\begin{subfigure}[b]{.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/pretrained/imagenet_cert_acc.pdf}
\caption{Pretrained ImageNet models.}
\end{subfigure}
\caption{Certified accuracy of various \textbf{pretrained} ViTs and comparable size ResNets on CIFAR-10 and ImageNet. ViTs non-trivial robustness wheras ResNets do not.}
\label{fig:pretrained-certified-vary-s}
\end{figure}
\subsection{Models and architectures}
We use three sizes of vision transformers---ViT-Tiny (ViT-T), ViT-Small (ViT-S), and ViT-Base (ViT-B) models \citep{rw2019timm, dosovitskiy2020image} and compare to to residual networks of similar (or larger) size---ResNet-18, ResNet-50 \citep{he2016deep}, and Wide ResNet-101-2 \citep{zagoruyko2016wide}, respectively. These architectures and their corresponding number of parameters are summarized in Table~\ref{table:architectures}.
\begin{table}[!htbp]
\caption{A collection of neural network architectures we use in our paper.}
\centering
\begin{tabular}{l|cccccccc}
\toprule
Architecture & ViT-T & ResNet-18 && ViT-S & ResNet-50 && ViT-B & WRN-101-2 \\
Params & 5M & 12M && 22M & 26M && 86M & 126M \\
\bottomrule
\end{tabular}
\label{table:architectures}
\end{table}
We use the same architectures for both ImageNet and CIFAR-10 models, and finetune our smoothed models from publicly released checkpoints pretrained on ImageNet. All our CIFAR-10 experiments are thus conducted on up-sampled CIFAR-10 images of size $224\times224$.
\subsection{Datasets}
\label{app:datasets}
We use two datasets:
\begin{enumerate}
\item CIFAR \citep{krizhevsky2009learning} \url{https://paperswithcode.com/dataset/cifar-10}.
\item ImageNet \citep{russakovsky2015imagenet}, with a custom (research, non-commercial) license, as found here \url{https://paperswithcode.com/dataset/imagenet}.
\end{enumerate}
\subsection{Training parameters}
\label{app:training}
Derandomized smoothing requires that the base classifier predict well on image ablations.
A standard technique for derandomized smoothing methods is to directly train the base classifier on image ablations \citep{levine2020randomized}. Thus, unless otherwise stated, in each epoch we randomly apply a column ablation of fixed width to each image of the training set.
To facilitate training of the base classifiers, we start from pretrained ResNets\footnote{These are TorchVision's official checkpoints, and can be found here \url{https://pytorch.org/vision/stable/models.html}.} and ViT architectures\footnote{We use the DeiT checkpoints of \cite{rw2019timm} which can be found here \url{https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py}.} and fine-tune as follows:
\paragraph{ImageNet.}
We train for $30$ epochs using SGD of fixed learning rate of $10^{-3}$, a batch size of $256$, a weight-decay of $10^{-4}$, a momentum of $0.9$, and with column ablations of fixed width $b=19$. For data-augmentation, we use random resized crop, random horizontal flip, and color jitter. We then apply column ablations.
\paragraph{CIFAR-10.}
We train for $30$ epochs using SGD with a step learning rate of $10^{-2}$ that drops every 10 epochs by a factor of 10, a batch size of $128$, a weight-decay of $5 \times 10^{-4}$, a momentum of $0.9$, and with column ablations of fixed width $b=4$. We only use random horizontal flip for data-augmentation, after which we apply column ablations. We then upsample all CIFAR-10 images to $224\times224$ (on GPU).
\paragraph{Training time.}
Training is relatively fast, with our largest ImageNet model (WRN-101-2) finishing in roughly two days on one NVIDIA V100 GPU. The smaller models such as ViT-T or ResNet-18 finish training in only a few hours.
\subsection{Compute and timing experiments}
\label{app:compute}
We use an internal cluster containing NVIDIA 1080-TI, 2080-TI, V100, and A100 GPUs. Scalability and timing experiments were performed on an A100
and averaged over 50 trials. When performing scalability experiments, we do not include data loading time or the time to move the input to the GPU.
\subsection{Example ablations}
In Figure~\ref{appfig:example-ablation-sizes}, we display examples of ablations of various types (column, block) and sizes.
\begin{figure}[!htbp]
\centering
\begin{subfigure}[b]{.7\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/Column_Ablation.pdf}
\caption{\textbf{Column ablations} with the following ablation size from left to right: original image, 13px, 19px, 25px, 31px, 37px.}
\end{subfigure}
\begin{subfigure}[b]{.7\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/Block_Ablation.pdf}
\caption{\textbf{Block ablations} with the following ablation size from left to right: original image, 35px, 55px, 75px, 95px, 115px.}
\end{subfigure}
\caption{Example ablations that we use in our paper.}
\label{appfig:example-ablation-sizes}
\end{figure}
\subsection{Differences in setup from \citet{levine2020randomized}}
\label{app:differences}
Our work builds on top of that of \citet{levine2020randomized}. We use their robustness guarantee as is (see Section~\ref{sec:preliminaries}), but there are a few differences in the setup of our experiments.
All experimental results (including the de-randomized smoothing baseline) are run using the same experimental setup in order to remain fair, which only improved the baseline over what was previously reported in the literature. For completeness, we describe the differences in setup here.
\paragraph{Encoding \textit{null} inputs.} The first difference is that \citet{levine2020randomized} encode part of the input as being \textit{null} or ablated by adding additional color channels, as described in \cite{levine2020robustness}, so that the \textit{null} value is distinct from all real pixel colors.
In practice, we found this to be unnecessary, and were able to replicate their results with ablations that simply replace masked pixels with 0.
\paragraph{Early stopping.} We find that ResNets substantially benefit from early stopping when trained with ablations, and otherwise experience severe overfitting to the ablations with substantially reduced test accuracies.
In our replication, we find that the ResNet-50 result reported by \citet{levine2020randomized} can be substantially improved with an earlier checkpoint (improving certified accuracy by nearly 10\%), and thus we use early-stopping in all of our ResNet baselines.
\paragraph{Starting from pretrained models.}
To reduce training time, for both ImageNet and CIFAR-10 experiments, we start from pre-trained ImageNet checkpoints (see Section \ref{app:training}). This step is especially necessary for the CIFAR-10 experiments, as it is quite challenging to train a ViT from scratch on CIFAR-10 (these models tend to require a large amount of data).
\paragraph{Upsampled CIFAR-10.}
In order to use the pretrained ImageNet checkpoints when training our base classifiers for CIFAR-10, we (nearest neighbor) upsample the CIFAR-10 inputs to $224\times224$ as part of the model architecture. We verify robustness in the original $32\times32$ images.
\paragraph{Sweeping over ablation size.}
We note that \citet{levine2020randomized} tested various ablations sizes only on CIFAR-10. Due to our speed-ups, we were able to sweep over ablations sizes for ImageNet.
\subsection{Certification thresholds for strided ablation sets}
\label{app:stride_threshold}
We briefly describe the new thresholds for certification when using strided ablations. Recall from \eqref{eq:certify} that a prediction is certified robust if
$$n_c(\mathbf{x}) > \max_{c' \neq c} n_{c'}(\mathbf{x}) + 2\Delta.$$
Thus $\Delta$, the number of ablations that a patch can intersect, fully describes the certification threshold.
\paragraph{Column smoothing.} For column smoothing with width $b$ and stride $s$, the maximum number of ablations that an $m \times m$ patch can intersect with is at most $\Delta_{column+stride} = \lceil(m+s-1)/s\rceil$.
\subsection{Performance under strided ablations}
\label{app:stride_performance}
\begin{figure}[!htbp]
\centering
\begin{subfigure}[b]{.49\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/stride_sweep/ImageNet_cert_acc.pdf}
\caption{Certified Accuracy}
\end{subfigure}
\begin{subfigure}[b]{.49\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/stride_sweep/ImageNet_smooth_acc.pdf}
\caption{Standard Accuracy}
\end{subfigure}
\caption{Certified and standard accuracy of various ViTs for ImageNet when using strided column ablations with varying stride lengths.}
\label{appfig:stride-sweep}
\end{figure}
In this section, we explore how striding affects standard and certified performance. We find that striding does not result in a monotonic change in certified accuracy---certification accuracy can both decrease and increase as the stride increases.
For a few choices in striding, it is possible to not substantially change the accuracy of the ViT at classifying ablations, as shown in Figure \ref{appfig:stride-sweep}.
For example, a ViT-B which normally obtains 38.3\% certified accuracy without striding, maintains certified accuracies of 37.6\% at stride $s=5$ and 36.8\% at stride $s=10$. For these small drops in certified accuracy, striding directly enables 5x or 10x faster inference times.
\subsection{Train-time ablation}
\label{app:trainablations}
We first explore varying the ablation size used during training for ImageNet.
Specifically, we train and certify a ResNet-50 and ViT-S over a range of column widths from 1 to 67 pixels (Figure \ref{fig:ablation-size-train-time}).
\begin{figure}[!htbp]
\centering
\begin{subfigure}[b]{.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/training_ablation_sweep/ImageNet_cert_acc.pdf}
\end{subfigure}
\begin{subfigure}[b]{.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/training_ablation_sweep/ImageNet_smooth_acc.pdf}
\end{subfigure}
\caption{
Certified and standard accuracy for a smoothed model trained and evaluated on ImageNet column ablations with varying widths. The ResNet-50 requires a substantially larger ablation size for certification, whereas the ViT-S is more flexible.
}
\label{fig:ablation-size-train-time}
\end{figure}
For ViTs, we find that once the columns are wide enough, we see only marginal improvements in certified accuracy (i.e. only 1.3\% higher certified accuracy over $b=19$).
This suggests that small ablations are sufficient at training time, allowing for fast training of ViTs when using cropped ablations.
On the other hand, ResNets require a substantially larger column width than was previously explored. Specifically, the certified accuracy of the ResNet baseline can be greatly improved from 18\% to 27\% when the ablation size is increased to $b=37$. This ablation size is optimal for the ResNet, but is still 6\% lower certified accuracy when compared to the ViT.
Overall, we find that certified performance of ViTs on ImageNet remains largely stable with respect to the column ablation size used for training. We can thus use smaller ablation sizes during training (e.g $b=19$) to improve training speed while certifying using larger ablation sizes.
\subsection{Test-time ablations}
\label{app:testablations}
\begin{figure}[!htbp]
\centering
\includegraphics[width=.48\textwidth]{figures/finetuned/cifar10_cert_acc.pdf}
\includegraphics[width=.48\textwidth]{figures/finetuned/cifar10_smooth_acc.pdf}
\caption{
Certified and standard accuracy for a smoothed model on CIFAR-10 trained with a fixed ablation size ($b=5$), and evaluated with varying ablation sizes.
}
\label{fig:ablation-size-finetuned-cifar10}
\end{figure}
Similar to the experiment on ImageNet from Section~\ref{sec:ablation_size}, we present analogous results
for varying the ablation size used at test time for CIFAR-10. These results largely reflect what was previously observed by \citet{levine2020randomized}. Specifically, the optimal ablation size for CIFAR10 is a column width of $b=4$, with a steep drop-off in performance for larger ablation sizes.
This is in contrast to what we observed in ImageNet, which did not see such a steep drop in performance.
\section{Certified patch defense with smoothing \& transformers}
Smoothing methods are a general class of certified defenses that combine the predictions of a classifier over many variations of an input to create predictions that are certifiably robust \citep{cohen2019certified, levine2020robustness}.
One such method that obtains robustness to adversarial patches is derandomized smoothing \citep{levine2020randomized}, which aggregates a classifier's predictions on various \textit{image ablations} that mask most of the image out.
These approaches typically use CNNs, a common default model for computer vision tasks, to evaluate the image ablations. The starting point of our approach is to ask: are convolutional architectures the right tool for this task? The crux of our methodology is to leverage vision transformers, which we demonstrate are more capable of gracefully handling the image ablations that arise in derandomized smoothing.
\subsection{Preliminaries}
\label{sec:preliminaries}
\paragraph{Image ablations.}
Image ablations are variations of an image where all but a small portion of the image is masked out \citep{levine2020randomized}. For example, a column ablation masks the entire image except for a column of a fixed width (see Figure \ref{fig:example-ablations} for an example). We focus primarily on column ablations and explore the more general block ablation in Appendix \ref{app:block-smoothing}.
\begin{figure}[!htbp]
\centering
\includegraphics[width=1\textwidth]{figures/example_column_ablations_ellipses.pdf}
\caption{Examples of column ablations for the left-most image with column width 19px.}
\label{fig:example-ablations}
\end{figure}
For a input $h \times w$ sized image $\mathbf{x}$, we denote by $\mathcal S_b(\mathbf{x})$ the set of all possible column ablations of width $b$.
A column ablation can start at any position and wrap around the image, so there are $w$ total ablations in $\mathcal S_b(\mathbf{x})$.
\begin{figure}
\includegraphics[width=\textwidth]{figures/schematic/schematic.png}
\caption{Illustration of the smoothed vision transformer. For a given image, we first generate a set of ablations. We encode each ablation into tokens, and drop fully masked tokens. The remaining tokens for each ablation are then fed into a vision transformer, which predicts a class label for each ablation. We predict the class with the most predictions over all the ablations, and use the margin to the second-place class for robustness certification.}
\label{fig:overview}
\end{figure}
\paragraph{Derandomized smoothing.}
Derandomized smoothing \citep{levine2020randomized} is a popular approach for certified patch defenses that constructs a \textit{smoothed classifier} comprising of two main components: (1) a \emph{base classifier}, and (2) a set of image ablations used to smooth the base classifier.
Then, the resulting smoothed classifier returns the most frequent prediction of the base classifier over the ablation set $\mathcal{S}_b(\mathbf{x})$. Specifically, for an input image $\mathbf{x}$, ablation set $\mathcal S_b(\mathbf{x})$, and a base classifier $f$, a smoothed classifier $g$ is defined as:
\begin{equation}
g(\mathbf{x}) = \argmax_{c} n_c(\mathbf{x})
\label{eq:smooth}
\end{equation}
where
$$n_c(\mathbf{x}) = \sum_{\mathbf{x}' \in S_b(\mathbf{x})} \mathbb{I}\{f(\mathbf{x}') = c\}$$
denotes the number of image ablations that were classified as class $c$. We refer to the fraction of images that the smoothed classifier correctly classifies as \emph{standard accuracy}.
A smoothed classifier is \textit{certifiably robust} for an input image if the number of ablations for the most frequent class exceeds the second most frequent class by a large enough margin.
Intuitively, a large margin makes it impossible for an adversarial patch to change the prediction of a smoothed classifier since a patch can only affect a limited number of ablations.
Specifically, let $\Delta$ be the maximum number of ablations in the ablation set $S_b(\mathbf{x})$ that an adversarial patch can simultaneously intersect (e.g., for column ablations of size $b$, an $m \times m$ patch can intersect with at most $\Delta = m + b - 1$ ablations).
Then, a smoothed classifier is certifiably robust on an input $\mathbf{x}$ if it is the case that for the predicted class $c$:
\begin{equation}
n_c(\mathbf{x}) > \max_{c' \neq c} n_{c'}(\mathbf{x}) + 2\Delta.
\label{eq:certify}
\end{equation}
If this threshold is met, the most frequent class is guaranteed to not change even if an adversarial patch compromises every ablation it intersects.
We denote the fraction of predictions by the smooth classifier that are both correct and certifiably robust (according to Equation \ref{eq:certify}) as \textit{certified accuracy}.
\paragraph{Vision transformers.}
A key component of our approach is the vision transformer (ViT) architecture \citep{dosovitskiy2020image}. In contrast to convolutional architecures, ViTs use self-attention layers instead of convolutional layers as their primary building block and are inspired by the success of self-attention in natural language processing \citep{vaswani2017attention}.
ViTs process images in three main stages:
\begin{enumerate}
\item \textit{Tokenization:} The ViTs split the image into $p \times p$ patches. Each patch is then embedded into a positionally encoded \textit{token}.
\item \textit{Self-Attention:} The set of tokens are then passed through a series of multi-headed self-attention layers \citep{vaswani2017attention}.
\item \textit{Classification head:} The resulting representation is fed into a fully connected layer to make predictions for classification.
\end{enumerate}
\begin{table}[t]
\begin{minipage}{\textwidth}
\renewcommand\footnoterule{}
\centering
\caption{Summary of our ImageNet results and comparisons to certified patch defenses from the literature: Clipped Bagnet (CBG), Derandomized Smoothing (DS), and PatchGuard (PG). Time refers to the inference time for a batch of 1024 images, $b$ is the ablation size, and $s$ is the ablation stride. An extended version is in Appendix~\ref{app:tables}.}
\begin{tabular}{l|cccc|c}
\toprule
\multicolumn{6}{c}{Standard and Certified Accuracy on ImageNet (\%) }\\
\midrule
& Standard & 1\% pixels & 2\% pixels & 3\% pixels & Time (sec) \\
\midrule
\multicolumn{5}{l}{\textit{Baselines} } \\
\midrule
Standard ResNet-50 & 76.1 & --- & --- & --- & 0.67\\
WRN-101-2 & 78.85 & --- & --- & --- & 3.1\\
ViT-S & 79.90 & --- & --- & --- & 0.4\\
ViT-B & 81.80 & --- & --- & --- & 0.95\\
CBN~\citep{zhang2020clipped} & 49.5 & 13.4 & 7.1& 3.1 & {3.05} \\
DS~\citep{levine2020randomized}\footref{fn:ds} & 44.4 &17.7 & 14.0& 11.2 & {149.5}\\
PG~\citep{xiang2021patchguard}\footref{fn:pg2} & $55.1$\footref{fn:pg2} & $32.3$\footref{fn:pg2} & $26.0$\footref{fn:pg2}& $19.7$\footref{fn:pg2} & ${3.05}$\\
\midrule
\multicolumn{5}{l}{\textit{Smoothed models} } \\
\midrule
ResNet-50 (b = 19) & 51.5 & 22.8 & 18.3 & 15.3 & 149.5 \\
ViT-S (b = 19) & \textbf{63.5} & \textbf{36.8} & \textbf{31.6} & \textbf{27.9} & \textbf{14.0}\\
\midrule
WRN-101-2 (b = 19) & 61.4 & 33.3 & 28.1 & 24.1 & 694.5 \\
ViT-B (b = 19) & 69.3 & \textbf{43.8} & \textbf{38.3} & \textbf{34.3} & 31.5 \\
ViT-B (b = 37) & \textbf{73.2} & 43.0 & 38.2 & 34.1 & 58.7 \\
ViT-B (b = 19, s = 10) & 68.3 & 36.9 & 36.9& 31.4& \textbf{3.2}\\
\bottomrule
\end{tabular}
\label{table:main_summary table}
\end{minipage}
\end{table}
\begin{figure}[b]
\begin{minipage}{\textwidth}
\footnotetext[1]{\label{fn:ds}We found that ResNets could achieve a significantly higher certified accuracy than was reported by \citet{levine2020randomized} if we use early stopping-based model selection. We elaborate further in Appendix \ref{app:setup}.}
\footnotetext[2]{\label{fn:pg2}The PatchGuard defense uses a specific mask size that guarantees robustness to patches smaller than the mask, and provides no guarantees for larger patches. In this table, we report their best results: each patch size corresponds to a separate model that achieves 0\% certified accuracy against larger patches. Comparisons across the individual models can be found in Appendix \ref{app:tables}.}
\end{minipage}
\end{figure}
\subsection{Smoothed vision transformers}
\label{sec:smoothed-vit}
Two central properties of vision transformers make ViTs particularly appealing for processing the image ablations that arise in derandomized smoothing. Firstly, unlike CNNs, ViTs process images as sets of tokens. ViTs thus have the natural capability to simply drop unnecessary tokens from the input and ``ignore'' large regions of the image, which can greatly speed up the processing of image ablations.
Moreover, unlike convolutions which operate locally, the self-attention mechanism in ViTs shares information \emph{globally} at every layer \citep{vaswani2017attention}. Thus, one would expect ViTs to be better suited for classifying image ablations, as they can dynamically attend to the small, unmasked region. In contrast, a CNN must gradually build up its receptive field over multiple layers and process masked-out pixels.
Guided by these intuitions, our methodology leverages the ViT architecture as the base classifier for processing the image ablations used in derandomized smoothing.
We first demonstrate that these \emph{smoothed vision transformers} enable substantially improved robustness guarantees, without losing much standard accuracy (Section \ref{sec:improve-accuracy}).
We then modify the ViT architecture and smoothing procedure to drastically speed up the cost of inference of a smoothed ViT (Section \ref{sec:improve-speed}).
We present an overview of our approach in Figure \ref{fig:overview}.
\paragraph{Setup.}
We focus primarily on the column smoothing setting and defer block smoothing results to Appendix \ref{app:block-smoothing}. We consider the CIFAR-10~\citep{krizhevsky2009learning} and ImageNet~\citep{deng2009imagenet} datasets, and perform our analysis on three sizes of vision transformers---ViT-Tiny (ViT-T), ViT-Small (ViT-S), and ViT-Base (ViT-B) models \citep{rw2019timm, dosovitskiy2020image}.
We compare to residual networks of similar size---ResNet-18, ResNet-50 \citep{he2016deep}, and Wide ResNet-101-2 \citep{zagoruyko2016wide}, respectively.
Further details of our experimental setup are in Appendix \ref{app:setup}.
\section*{Checklist}
\begin{enumerate}
\item For all authors...
\begin{enumerate}
\item Do the main claims made in the abstract and introduction accurately reflect the paper's contributions and scope?
\answerYes{}
\item Did you describe the limitations of your work?
\answerYes{See the conclusion.}
\item Did you discuss any potential negative societal impacts of your work?
\answerYes{See the conclusion.}
\item Have you read the ethics review guidelines and ensured that your paper conforms to them?
\answerYes{}
\end{enumerate}
\item If you are including theoretical results...
\begin{enumerate}
\item Did you state the full set of assumptions of all theoretical results?
\answerNA{}
\item Did you include complete proofs of all theoretical results?
\answerNA{}
\end{enumerate}
\item If you ran experiments...
\begin{enumerate}
\item Did you include the code, data, and instructions needed to reproduce the main experimental results (either in the supplemental material or as a URL)?
\answerYes{The code link is in the abstract.}
\item Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)?
\answerYes{See Appendix.}
\item Did you report error bars (e.g., with respect to the random seed after running experiments multiple times)?
\answerYes{See Appendix.}
\item Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)?
\answerYes{See Appendix.}
\end{enumerate}
\item If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...
\begin{enumerate}
\item If your work uses existing assets, did you cite the creators?
\answerYes{We cite both ImageNet and CIFAR-10 in the main paper.}
\item Did you mention the license of the assets?
\answerYes{See Appendix.}
\item Did you include any new assets either in the supplemental material or as a URL?
\answerNA{}
\item Did you discuss whether and how consent was obtained from people whose data you're using/curating?
\answerNA{}
\item Did you discuss whether the data you are using/curating contains personally identifiable information or offensive content?
\answerNA{}
\end{enumerate}
\item If you used crowdsourcing or conducted research with human subjects...
\begin{enumerate}
\item Did you include the full text of instructions given to participants and screenshots, if applicable?
\answerNA{}
\item Did you describe any potential participant risks, with links to Institutional Review Board (IRB) approvals, if applicable?
\answerNA{}
\item Did you include the estimated hourly wage paid to participants and the total amount spent on participant compensation?
\answerNA{}
\end{enumerate}
\end{enumerate}
\subsection*{Our contributions}
In this paper, we demonstrate how to leverage vision transformers (ViTs) \citep{dosovitskiy2020image} to create certified patch defenses that achieve significantly higher robustness guarantees than prior work.
Moreover, we show that certified patch defenses with ViTs can actually maintain standard accuracy and inference times comparable to standard (non-robust) models.
At its core, our methodology exploits the token-based nature of attention modules used in ViTs to gracefully handle the ablated images used in certified patch defenses.
Specifically, we demonstrate the following:
\paragraph{Improved guarantees via smoothed vision transformers.}
We find that using ViTs as the backbone of the derandomized smoothing defense \citep{levine2020randomized} enables significantly improved certified patch robustness. Indeed, this change alone boosts certified accuracy by up to 13\% on ImageNet, and 5\% on CIFAR-10 over similarly sized ResNets.
\paragraph{Standard accuracy comparable to that of standard architecures.}
We demonstrate that ViTs enable certified defenses with standard accuracies comparable to that of standard, non-robust models.
In particular, our largest ViT improves state-of-the-art certified robustness on ImageNet while maintaining standard accuracy that is similar to that of a non-robust ResNet (>70\%).
\paragraph{Faster inference.}
We modify the ViT architecture to drop unnecessary tokens, and reduce the smoothing process to pass over mostly redundant computation.
These changes turn out to vastly speed up inference time for our smoothed ViTs.
In our framework, a forward pass on ImageNet becomes up to two orders of magnitude faster than that of prior certified defenses, and is close in speed to a standard (non-robust) ResNet.
\subsection{Ablation Size}
\subsection{Certified robustness}
\paragraph{Empirical methods: attacks and defenses.}
Another line of work studies empirical approaches for generating adversarial patches and designing empirical defenses.
Adversarial patches have been developed for downstream tasks such as image
classification \citep{karmon2018lavan}, object detection \citep{eykholt2018physical, chen2018shapeshifter, liu2018dpatch}, and facial recognition \citep{sharif2016accessorize, thys2019fooling, bose2018adversarial}. Several of these attacks work in the
physical domain \citep{brown2018adversarial, eykholt2018physical, chen2018shapeshifter}, and
can successfully target tasks such as traffic sign recognition \citep{eykholt2018physical, chen2018shapeshifter}. Heuristic defenses to these attacks include watermarking \citep{hayes2018visible} and gradient smoothing \citep{naseer2019local}; however, these defenses were shown to be vulnerable adaptive attacks \citep{chiang2020certified}. More recently, \citet{rao2020adversarial} proposed an adversarial training approach to improve empirical robustness to patch attacks.
\paragraph{Vision transformers.}
Our work leverages the vision transformer (ViT) architecture \citep{dosovitskiy2020image}, which adapts the popular attention-based model from the language setting \citep{vaswani2017attention} to the vision setting.
Recent work \citep{touvron2020training} has released more efficient training methods as well as pre-trained ViTs that have made these architectures more accessible to the wider research community.
\subsection{Dropping masked tokens for vision transformers}
\subsection{Dropping masked tokens}
\label{sec:speed}
Recall that the first operation in a ViT is to split and encode the input image as a set of \textit{tokens}, where each token corresponds to a patch in the image. However, for image ablations, a large number of these tokens correspond to fully masked regions of the image.
Our strategy is to pass only the \textit{subset} of tokens that contain an unmasked part of the original image, thus avoiding computation on fully masked tokens.
Specifically, given an image ablation, we alter the ViT architecture to do the following steps:
\begin{enumerate}
\item Positionally encode the entire ablated image into a set of tokens.
\item Drop any tokens that correspond to a \textit{fully} masked region of the input.
\item Pass the remaining tokens through the self-attention layers.
\end{enumerate}
As one would expect, since the positional encoding maintains the spatial information of the remaining tokens, the ViT's accuracy on image ablations barely changes when we drop the fully masked tokens. We defer a detailed analysis of this phenomenon, along with a formal description of the token-dropping procedure to Appendix~\ref{app:droptokens}.
\paragraph{Computational complexity.}
We now provide an informal summary of the computational complexity of this procedure, and defer a formal asymptotic analysis to Appendix \ref{app:complexity}.
After tokenization, the bulk of a ViT consists of two main operation types:
\begin{itemize}
\item \textit{Attention operators}, which have costs that scale quadratically with the number of tokens but linearly in the hidden dimension.
\item \textit{Fully-connected operators}, which have costs that scale linearly with the number of tokens but quadratically in the hidden dimension.
\end{itemize}
Reducing the number of tokens thus directly reduces the cost of attention and fully connected operators at a quadratic and linear rate, respectively.
For a small number of tokens, the linear scaling from the fully-connected operators tends to dominate.
The cost of processing column ablations thus scales linearly with the width of the column, which we empirically validate in Figure \ref{fig:speed}. Further details about how we time these models can be found in Appendix \ref{app:compute}.
\begin{figure}[!htbp]
\centering
\includegraphics[width=.6\textwidth]{figures/scalability/vit_complexity_plot.pdf}
\caption{
The average time to compute a forward pass for ViTs on $1024$ column ablated images with varying ablation sizes, with and without dropping masked tokens.
The cost of processing a full image without dropping masked tokens corresponds to the maximum ablation size $b=224$.
}
\label{fig:speed}
\end{figure}
\subsection{Empirical speed-up for smoothed ViTs}
Smoothed classifiers must process a large number of image ablations in order to make predictions and certify robustness. Consequently, using our ViT (with dropped tokens) as the base classifier for derandomized smoothing directly speeds up inference time. In
this section, we explore how much faster smoothed ViTs are in practice.
\begin{table}
\begin{center}
\caption{Multiplicative speed up of inference for a smoothed ViT with dropped tokens over a smoothed ResNet, measured over a batch of 1024 images with $b=19$.}
\label{tab:speed_resnet}
\begin{tabular}{c|ccc}
\toprule
& ResNet-18 & ResNet-50 & WRN-101\\
\midrule
ViT-T & \textbf{5.85x} & 21.96x & 101.99x \\
ViT-S & 2.85x & \textbf{10.68x} & 49.62x \\
ViT-B & 1.26x & 4.75x & \textbf{22.04x} \\
\bottomrule
\end{tabular}
\end{center}
\end{table}
We first measure the number of images per second that smoothed ViTs and smoothed ResNets can process. We use column ablations of size $b=19$ on ImageNet, following \citet{levine2020randomized}.
In Table \ref{tab:speed_resnet} that describes our results, we find speedups of 5-22x for smoothed ViTs over smoothed ResNets of similar size, with larger architectures showing greater gains.
Notably, using our largest ViT (ViT-B) as the base classifier is 1.25x faster than using a ResNet-18, despite being 8x larger in parameter count.
Dropping masked tokens thus substantially speeds up inference time for smoothed ViTs, to the point where using a large ViT is comparable in speed to using a small ResNet.
\paragraph{Strided ablations.}
\label{sec:ablation_stride}
We now consider a complementary means of speeding up smoothed classifiers: directly reducing the size of the ablation set via \emph{strided} ablations.
Specifically, instead of using every possible ablation, we can subsample every $s$-th ablation for a given stride $s$.
Striding can reduce the total number of ablations (and consequently speed up inference) by a factor of $s$, \emph{without} substantially hurting standard or certified accuracy (Table~\ref{table:main_summary table}). We study this in more detail in Appendix~\ref{app:stride}.
Strided ablations, in conjunction with the dropped tokens optimization from Section \ref{sec:speed}, lead to smoothed ViTs having inference times comparable to standard (non-robust) models. For example, when using stride $s=10$ and dropping masked tokens, a smoothed ViT-S is only 2x slower than a single inference step of a standard ResNet-50, while a smoothed ViT-B is only 5x slower. We report the inference time of these models, along with their standard and certified accuracies, in Table \ref{table:main_summary table}.
\subsection{Ablation size}
\label{sec:ablation_size}
Derandomized smoothing certificates depends on the size of the unmasked part of the image, which we refer to as the \textit{ablation size}.
For column smoothing for instance, this is the width of the column of pixels that remain unmasked in the ablation process.
Previously in section~\ref{sec:ablations}, we saw that base classifiers can more accurately classify larger ablation sizes, but how does this affect certified robustness?
After all, larger ablations directly increase the necessary threshold for certification and reduce certified performance. In this section, we find that certified performance is largely stable with respect to ablation size, allowing allowing for a trade-off between computation time and clean accuracy without losing certified performance.
\begin{figure}[!htbp]
\centering
\begin{subfigure}[b]{.8\textwidth}
\centering
\includegraphics[width=.48\textwidth]{figures/Finetuned/cifar10_cert_acc.pdf}
\includegraphics[width=.48\textwidth]{figures/Finetuned/cifar10_smooth_acc.pdf}
\caption{CIFAR-10 models.}
\end{subfigure}
\begin{subfigure}[b]{.8\textwidth}
\centering
\includegraphics[width=.48\textwidth]{figures/Finetuned/imagenet_cert_acc.pdf}
\includegraphics[width=.48\textwidth]{figures/Finetuned/imagenet_smooth_acc.pdf}
\caption{ImageNet models.}
\end{subfigure}
\caption{
Certified and standard accuracy for a smoothed model trained on column ablations with width $b=19$, and evaluated with column ablations of varying width. Certified accuracy remains stable across a range of ablation sizes, while clean accuracy can substantially improve with larger ablations.
}
\label{fig:ablation-size-finetuned}
\end{figure}
\paragraph{Test-time ablation}
As a first experiment, we study how ablation size effects certified and clean accuracy at test time.
We consider the column smoothed models from Section \ref{sec:improve}.
However, instead of certifying at column width $b=19$ (the ablation size used during training), we experiment with a range of ablations and report the clean and certified accuracy in Figure~\ref{fig:ablation-size-finetuned}.
We find that certified accuracy is relatively stable over ablation sizes. For example, on ImageNet, certified performance peaks between column widths of 20-30px for all models.
However, clean accuracy makes significant gains with larger ablation sizes---improving ImageNet accuracy by almost 4\% at column width $s=37$ with the same certified accuracy for the ViT-B model.
Although the larger ablation size increases inference time for the ViT by a factor of TODO, this is still TODOx faster than the ResNet.\eric{saachi can you fill these in?}
Consequently, varying the ablation size can lead to almost free improvements to clean accuracy.
\begin{figure}[!htbp]
\centering
\begin{subfigure}[b]{.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/training_ablation_sweep/ImageNet_cert_acc.pdf}
\end{subfigure}
\begin{subfigure}[b]{.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/training_ablation_sweep/ImageNet_smooth_acc.pdf}
\end{subfigure}
\caption{
Certified and standard accuracy for a smoothed model trained and evaluated on ImageNet column ablations with varying widths. The ResNet50 requires a substantially larger ablation size for certification, whereas the ViT-S is more flexible. \ericinline{Hadi can you update this figure with the additional sizes?}
}
\label{fig:ablation-size-train-time}
\end{figure}
\paragraph{Train-time ablation}
In addition to certification time, we also explore varying the ablation size used at training time.
Specifically, we train train and certify a ResNet-50 and ViT-S at the same ablation size over a range of column widths from 1 to 37px.
First, we find that once columns are wide enough, we see only marginal improvements in certified accuracy (i.e. only TODO\% over $s=19$).
This suggests that small ablations are sufficient at training time, allowing for fast training of ViTs when using cropped ablations.
Second, we see that ResNets require a substantially larger column width ($s=TODO$) in order to reach optimal certified accuracy---but still remains behind the ViT.
Overall, we find that certified performance on ImageNet remains largely stable with respect to column ablation size, allowing for faster training trains and trade-offs between speed and clean accuracy.
\begin{figure}[!htbp]
\centering
\begin{subfigure}[b]{.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/stride_sweep/ImageNet_cert_acc.pdf}
\end{subfigure}
\begin{subfigure}[b]{.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/stride_sweep/ImageNet_smooth_acc.pdf}
\end{subfigure}
\caption{Certified accuracy of various ViTs and comparable size ResNets on CIFAR-10 and ImageNet. ViTs non-trivial robustness wheras ResNets do not.}
\label{fig:stride-sweep}
\end{figure}
\subsection{Ablation strides}
\label{sec:ablation_stride}
The bulk of the runtime of derandomized smoothing comes from aggregating predictions over a large number of ablations.
For example, column smoothing on ImageNet aggregates the result over $224$ column ablations, slowing down inference by two orders of magnitude.
Although we previously sped this process up with cropped ablations for ViTs, another possibility is to use strided ablations to reduce the number of ablations.
Specificaly, instead of using every possible ablation, we can instead take every $s$th ablation for a given stride $s$.
Strided ablations reduces the total number of ablations (and consequently speeds up inference) by a factor of $s$, without substantially changing the certification threshold (up to rounding).
\begin{example}
For column smoothing with width $b$ and stride $s$, the maximum number of ablations that an $m \times m$ patch can intersect with is at most $\Delta_{column+stride} = \lceil(m+s-1)/s\rceil$.
\end{example}
We test this for various ViTs on ImageNet, with column ablations of width $19$, against $32\times 32$ patches.
We find multiple strides that achieve similar performance to the stride of length 1 (e.g. stride of 5 or 10) as shown in Figure~\ref{fig:stride-sweep}.
These results comes at 5x and 10x speedups respectively, with slightly lower clean accuracy as shown in the right plot of Figure~\ref{fig:stride-sweep}.
\eric{Hadi, can you help replace the right plot with a column smoothing equivalent? And move the current block smoothing one to the appendix?}
We find that striding can lead to high variance in performance: certified and standard accuracy can drastically shift either upwards or downwards as we vary the stride.
However, as long as the stride is carefully selected, we can trade-off a slight decreases in certified and clean accuracy for significant speedups in certification.
\eric{Saachi, how close is this to a standard forward pass of a non-smoothed model?}
\subsection{Barriers to Deploying Certified Models}
\begin{figure}[!htbp]
\centering
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/finetuned/cifar10_ablation_acc.pdf}
\caption{CIFAR-10 models.}
\end{subfigure}
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/finetuned/ImageNet_ablation_acc.pdf}
\caption{ImageNet models.}
\end{subfigure}
\caption{Ablation accuracies for models on CIFAR-10 and ImageNet column-ablated images. The models were fine-tuned on column ablations of width $b=19/224$ for ImageNet and $b=4/32$ for CIFAR-10, and evaluated on a range of ablation sizes. ViTs outperform ResNets on image ablations by a sizeable margin.}
\label{fig:finetuned}
\end{figure}
\subsection{ViTs improve ablation accuracy}
\label{sec:ablations}
\eric{Consider moving CIFAR10 to appendix, for space}
The success of de-randomized smoothing methods hinges entirely on whether the base classifier can accurately classify subregions of an image, such as a small column or a patch.
We thus measure the accuracy of fine-tuned ViTs for classifying column ablated images in comparison to ResNets across a range of evaluation ablation sizes \saachi{is it clear here that the training ablation size is not the same as the evaluation ablation size?}.
We find that ViTs are significantly more accurate on column ablations than comparably sized ResNets (Figure \ref{fig:finetuned}).
This difference is most prominent on Imagenet---ViT-S has 11-12\% higher ablation accuracy across a range of ablation sizes than the the comparably sized ResNet-50. This suggests that ViTs are better suited for classifying image ablations.
Since ViTs are substantially more accurate at processing image ablations than ResNets, we would expect them to also produce better certied defenses against patch attacks.
In Section \ref{sec:improve}, we show how Smoothed ViTs have better accuracy and guarantees than Smoothed ResNets.
\eric{need to revise this last two sentences}
\begin{table}
\begin{center}
\caption{We measure the average speed (over 50 trials) of computing a forward pass with a vision transformer on 1024 column ablated images with varying ablation sizes. We report the multiplicative speed up dropping masked tokens versus passing the full input to the vision transformer.}
\label{tab:speed}
\begin{tabular}{c|cccccccccc}
\toprule
Ablation Size (px) & 13 & 19 & 25 & 31 & 37 & 43 & 49 & 55 & 61 & 67\\
\midrule
VIT-T & 10.62x & 6.44x & 6.44x & 4.33x & 3.48x & 3.48x & 3.48x & 2.91x & 2.90x & 2.90x\\
VIT-S & 10.57x & 6.40x & 4.39x & 4.40x & 4.41x & 3.52x & 3.52x & 3.52x & 2.93x & 2.93x\\
VIT-B & 11.63x & 6.76x & 6.75x & 4.51x & 3.62x & 4.50x & 3.63x & 3.61x & 2.94x & 2.94x\\
\bottomrule
\end{tabular}
\end{center}
\end{table}
\subsection{ViTs reduce computational complexity}
\label{sec:speed}
A downside of derandomized smoothing is its computational cost---a forward pass through the smoothed model requires passing all possible ablations through the base classifier.
For column smoothing (the fastest type of ablation for derandomized smoothing) on a $s \times s$ pixel image, this is $s$ times slower than a normal forward pass.
For ImageNet, this slows down evaluation by two orders of magnitude.
\eric{Fix notation}
One natural strategy for reducing unnecessary computation is to crop out the masked regions before passing the image to the base classifier. Although reducing the size of the image can speed up inference, cropping has several downsides for CNNs. Cropping an image strips spatial information about the ablation and disallows ablations that wrap around the sides of the image. These lead to substantial reductions in certified performance which we discuss in more detail in Appendix \ref{app:crop}.
However, unlike CNNs which require a structured input, Vision Transformers operate on \textit{tokens}, where each token corresponds to a region in the image.
Crucially, transformers treat inputs as unordered sets of tokens, which can be of any size and do not have to be contiguous in the original image space.
Thus, we can perform the following simple procedure to avoid computation on fully masked tokens:
\begin{enumerate}
\item Positionally encode the entire ablated image into tokens\;
\item Drop any tokens that correspond to a fully masked region of the image\;
\item Pass the remaining tokens through the transformer\;
\end{enumerate}
A more detailed discussion of this optimization can be found in Appendix \ref{app:droptokens}
\paragraph{Performance impact}
Since the tokens are encoded with their spatial location in the original input, we do not lose any positional information when dropping tokens.
In practice, we find that dropping masked tokens does not hurt performance as shown in Figure \ref{fig:missingness-ablation-acc}.
Dropping tokens thus amounts to a free speed-up for ViTs, reducing runtime as shown in Table \ref{tab:speed}.
\paragraph{Computational Complexity}
\eric{Fix notation}
More formally, suppose the base classifier is a ViT that takes a $s \times s$ pixel image embeds patches of $p \times p$ into tokens.
Processing the full image thus requires evaluating $O(s^2/p^2)$ tokens with the ViT.
To process a column ablation with width $b$, we can drop masked tokens and only evaluate $O(bs/p^2)$ tokens with the ViT, reducing the number of tokens by a factor of $O(s/b)$.
It turns out that passing fewer tokens can drastically speed up certification beyond an $O(s/b)$ factor.
Specifically, the self-attention mechanism of the ViT scales quadratically with the number of tokens.
Thus, the complexity for evaluating a full image is $O(s^4/p^4)$, which reduces to $O(b^2s^2/p^4)$ when dropping masked tokens for column ablation.
In fact, the cost of inference for a smoothed ViT with dropped tokens is almost comparable to a forward pass for a standard ViT.
For example, column smoothing evaluates $s$ different ablations, incurring a total runtime cost of $O(b^2s^3/p^4)$.
For $b \leq \sqrt{s}$, a smoothed ViT can actually be faster than a standard ViT.
\eric{this section is kind of chunky}
\eric{Reference analogous results for block smoothing in the appendix}
\begin{figure}[!htbp]
\centering
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/finetuned_missingness/cifar10_ablation_acc.pdf}
\caption{CIFAR-10 models.}
\end{subfigure}
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{figures/finetuned_missingness/imagenet_ablation_acc.pdf}
\caption{ImageNet models.}
\end{subfigure}
\caption{Ablation accuracies for various Vit models on CIFAR-10 and ImageNet, with and without missingness. The performance of both are comparable to each other while the missigness ones are 2 orders of magnitude faster.\ericinline{Fix this caption; also is 2 order of magnitude correct??}}
\label{fig:missingness-ablation-acc}
\end{figure}
|
\section{\label{sec:conclusion}Conclusion \& Limitations}
We present an active learning method for training a neural collision detector in which training data are progressively sampled from the learned latent space using a risk-seeking approach. Our approach is designed for general 3D deformable meshes, and we highlight its benefits on many complex datasets. In practice, our method outperforms supervised learning in terms of accuracy, false negative rate, and stability. As a major limitation, our collision handler does not consider physics models. This can be performed in the future by integrating a learning-based physics simulation approach such as \citep{Zheng_2021_CVPR}. We are also considering extensions to meshes with changing topologies, e.g., using a level-set-based mesh representation.
\section{\label{sec:evaluation}Evaluation}
\TE{Datasets:} We evaluate our method on five types of datasets, as illustrated in \prettyref{fig:examples}. The first three (SCAPE \cite{anguelov2005scape} with $N=71$ meshes each having $2161$ vertices, MIT Swing \cite{vlasic2008articulated} with $N=150$ meshes each having $9971$ vertices, and MIT Jump \cite{vlasic2008articulated} with $N=150$ meshes each having $10002$ vertices) contain human bodies with different sets of actions and poses. We have also tested our method on a skirt dataset introduced by \cite{yang2020multiscale} that contains $N=201$ simulated skirt meshes synthesized by NVIDIA clothing tools, each of which has $2830$ vertices. The skirt is deformable everywhere, and the dataset is rather small. Obtaining stable performance in this case is challenging and we observe reasonably good results using active learning. Finally, we introduce a custom dataset of human hand poses. We captured various hand poses and transitions between the poses in a multi-view capture system. We ran 3D reconstruction~\cite{Galliani15} and 3D keypoint detection~\cite{Simon_2017_CVPR} for the captured images and registered a linear blend skinning model consisting of $2825$ vertices for each frame of the data~\cite{gall2009motion}, resulting in $N=7314$ meshes.
\begin{table}[h]
\vspace{-5px}
\centering
\scalebox{0.6}{
\setlength{\tabcolsep}{4pt}
\begin{tabular}{cccccc}
\toprule
&SCAPE & Swing & Jump & Skirt & Hand \\
\midrule
$N_\text{init}$ &200000 & 10000 & 10000 & 5000 & 200000 \\
$N_\text{aug}$ & 50000 & 5000 & 5000 & 5000 & 50000\\
\bottomrule
\end{tabular}}
\vspace{-5px}
\caption{\label{table:Ninit}\small{$N_\text{init}$ and $N_\text{aug}$ used by each dataset.}}
\vspace{-10px}
\end{table}
\TE{Implementation:} We implement our method using PyTorch with the same network architecture as \cite{tan2020lcollision} and perform experiments on a desktop machine with an NVIDIA RTX 2080Ti GPU. We begin by training $\left<\theta_E,\theta_D\right>$ using Adam with a learning rate of $0.01$ and a batch size of $128$ over $3000$ epochs. For neural collision detector training, unless otherwise stated, we choose the following hyper-parameters: $\epsilon=1\times10^{-4}, w_\text{PD}=5, w_\text{PDsum}=0.2, w_r=2, w_{ce}=2, w_b=0.5$. We perform bootstrap by supervised learning $\theta_C$ on $N_\text{init}$ data points. We initialize $N_\text{init}=2000$ and progressively inject data points into $N_\text{init}$ until the ``elbow point'' of the accuracy vs. the sample size is reached, which is detected using \cite{5961514}. For each experiment, we train $\theta_C$ using Adam with a learning rate of $0.001$ and a batch size of $512$ over $N_\text{epoch}=100$ epochs. We choose suitable $N_\text{aug}$ according to $N_\text{init}$. The $N_\text{init}$ and $N_\text{aug}$ used for each dataset are summarized in \prettyref{table:Ninit}. During data aggregation, we terminate Newton's method when the relative changes of $Z_\text{all}$ are less than $\epsilon_z=10^{-7}$. For each subsequent iteration of active data augmentation, we fine-tune $\theta_C$ using Adam with a learning rate of $10^{-4}$ and a batch size of $1024$ over $N_\text{epoch}=50$ epochs. For collision handling, we run ALM until \prettyref{eq:handler} is satisfied.
\TE{Collision Detection:} We compare our method with two baseline algorithms. The first one \cite{tan2020lcollision} (denoted as \textbf{\textit{Supv}}) uses the same network architecture as ours with both the autoencoder $\left<\theta_E,\theta_D\right>$ and the collision detector $\theta_C$ trained using supervised learning, where $\mathcal{D}_c$ is constructed by randomly sampled poses from $U(\mathcal{Z})$ and boundary set $D_b$ with associated loss \prettyref{eq:bd_loss} mentioned in \prettyref{sec:bd} is not used, i.e., $w_b=0$. Our second baseline (denoted as \textbf{\textit{Supv + bd}}) also trains both networks using supervised learning, but the boundary set and loss in \prettyref{eq:bd_loss} are used. Our proposed neural collision handling pipeline also uses active learning and boundary information and is denoted as \textbf{\textit{Active + bd}}. After $k$ iterations of active data augmentations, we have a dataset with $N_{init}+kN_{aug}$ points for training $\theta_C$. For fairness, we re-train our two baselines using $N_{init}+kN_{aug}$ points randomly sampled from $U(\mathcal{Z})$ after each iteration. For all the methods, we use $80\%$ of the data for training and the rest is used as a validation set for hyperparameter tuning. For each dataset, we create a test set with $7.5\times 10^5$ samples from $U(\mathcal{Z})$, which is unseen in the training stage, to evaluate the performances. The performances of neural collision detectors are evaluated based on two metrics: the fraction of successful predicates (accuracy) and the fraction of times a self-penetrating mesh is erroneously predicted as collision-free (false negative rate). False negatives are more detrimental to our applications than false positives as our collision handler only takes care of positive samples. As illustrated in \prettyref{fig:iterationAccuracy} and \prettyref{fig:iterationFalseNegativeRate}, our method effectively improves both metrics. The performance after active learning is summarized in \prettyref{table:perf}. We reach $93.8 - 98.1\%$ accuracy compared to the groundtruth generated by the exact method \cite{pan2012fcl}, with up to 124$\times$ speedup. On average, our method achieves $1.86\%$ higher accuracy and a $14.27\%$ lower false negative rate than \textbf{\textit{Supv}}. In the last row of \prettyref{table:perf}, we measure an equivalent dataset size, which is defined as the size of the dataset needed by the \textbf{\textit{Supv+bd}} to achieve the same accuracy as our method. We derive this number by interpolating on experimental results of the \textbf{\textit{Supv+bd}}. Our method achieves a similar accuracy using a $35.0\%$ smaller dataset than \textbf{\textit{Supv+bd}} on average.
\begin{table}[t]
\centering
\resizebox{0.47\textwidth}{!}{
\setlength{\tabcolsep}{4pt}
\begin{tabular}{cccccc}
\toprule
metric & SCAPE & Swing & Jump & Skirt & Hand \\
\midrule
final dataset size & $5.5\times10^5$ & $4.5\times 10^4$ & $4.5\times 10^4$ & $4\times 10^4$ & $5.5\times10^5$ \\
\midrule
accuracy (Ours (\textbf{\textit{Active+bd}})) & 0.9383 & 0.9638 & 0.9552 & 0.9817 & 0.9692 \\
accuracy (\textbf{\textit{Supv+bd}}) & 0.9282 & 0.9609 & 0.9500 & 0.9795 & 0.9650 \\
accuracy (\textbf{\textit{Supv}}) & 0.9181 & 0.9460 & 0.9347 & 0.9660 & 0.9558 \\
\midrule
false neg. (Ours (\textbf{\textit{Active+bd}})) & 0.05151 & 0.01485 & 0.02573 & 0.01808 & 0.03582 \\
false neg. (\textbf{\textit{Supv+bd}}) & 0.05576 & 0.01766 & 0.02644 & 0.01956 & 0.03652 \\
false neg. (\textbf{\textit{Supv}}) & 0.06914 & 0.01969 & 0.02713 & 0.02056 & 0.03727 \\
\midrule
equi. dataset size (\textbf{\textit{Supv+bd}}) & $6.63\times10^5$ & $7.64\times10^4$ & $7.02\times10^4$ & $7.71\times 10^4$ & $7.30\times10^5$ \\
\bottomrule
\end{tabular}}
\vspace{-5px}
\caption{\label{table:perf}\footnotesize{We summarize the accuracy and false negative rate of three methods under comparison. We also include the equivalent dataset size for the baseline to reach the same performance as our method.}}
\vspace{-10px}
\end{table}
\begin{figure}
\begin{minipage}[b]{0.2\textwidth}
\centering
\includegraphics[width=\textwidth]{figs/scape_collision_response_analysis.pdf}
\vspace{-20px}
\end{minipage}
\hfill
\begin{minipage}[b]{.24\textwidth}
\captionof{figure}{\scriptsize{We plot joint distribution of relative $\text{PD}$ reduction ($Y$-axis) and embedding difference ($X$-axis) over successfully collision-handled test meshes in the SCAPE dataset for our method and \textbf{\textit{Supv+bd}}. Our method resolves more collisions (average PD reduction $94.81\%$ vs. $88.76\%$) while remaining closer to the input (average embedding difference 56.17 vs. 66.11) compared to \textbf{\textit{Supv+bd}}.}}
\label{fig:PDReduction}
\vspace{-20px}
\end{minipage}
\end{figure}
\TE{Collision Handling:} We plug the trained neural collision detectors into ALM and compare our method, \textbf{\textit{Supv}} and, \textbf{\textit{Supv+bd}} in terms of resolving self-penetrating meshes. To this end, we randomly sample $10000$ self-penetrating, unseen $Z_{all}^{user}$ from $U(\mathcal{Z})$, and use \prettyref{eq:handler} to derive $Z_{all}$.
We compare the performance based on relative $\text{PD}$ reduction defined as:
\footnotesize
\begin{align*}
\frac{\text{PD}(\text{ACAP}^{-1}(Z_\text{all}^{user},\theta_D))-\text{PD}(\text{ACAP}^{-1}(Z_\text{all},\theta_D))}{\text{PD}(\text{ACAP}^{-1}(Z_\text{all}^{user},\theta_D))}.
\end{align*}
\normalsize
Collision resolution is completely successful if this value equals one, which may not always happen because ALM uses soft penalties to relax hard constraints. Thus, we consider a solution successful if the value is greater than $0$. We plot the success rate against the dataset size in \prettyref{fig:iterationColResSuccRate}, which shows that our method resolves $24.74\%$ more collisions than \textbf{\textit{Supv+bd}}. Thanks to our risk-seeking data aggregation method, our method monotonically improves the collision handling success rate when more data points are injected, while \textbf{\textit{Supv+bd}} exhibits unstable performance. Since \textbf{\textit{Supv}} uses the same randomly sampled dataset as \textbf{\textit{Supv+bd}}, the performance exhibits similar instability. Meanwhile, our novel boundary loss improves the results for \textbf{\textit{Supv+bd}}, since it can better approximate the decision boundary. Another criterion for good collision handling is the embedding difference -- the objective function in \prettyref{eq:handler}. We want the output to be as close as possible to the input. We plot the relative $\text{PD}$ reduction vs. embedding difference over successfully collision-handled test meshes in the SCAPE dataset for our method and \textbf{\textit{Supv+bd}} in \prettyref{fig:PDReduction}. The mean relative $\text{PD}$ reduction for our method is $94.81\%$ and the mean embedding difference is $56.17$, compared to $88.76\%$ and $66.11$, respectively, for \textbf{\textit{Supv+bd}}. The results show that our method resolves more collisions while the outputs stay closer to the input latent codes. Some exemplary results are shown in \prettyref{fig:examples}.
\section{\label{sec:intro}Introduction}
Learning to model or simulate deformable meshes is becoming an important topic in computer vision and computer graphics, with rich applications in real-time physics simulation \citep{holden2019subspace}, animation synthesis \citep{9217964}, and cross-domain model transformation \citep{VOCA2019}. Central to these methods are generative models that map high-dimensional deformed 3D meshes with rich details into low-dimensional latent spaces. These generative models can be trained from high-quality groundtruth datasets, and they infer visually or physically plausible meshes in real time. These 3D datasets can also be generated using physics simulations \citep{Narain:2012:AAR,tang2012continuous} or reconstructed from the physical world using multi-view capture systems \citep{10.1145/3414685.3417768}. In general, 3D deformable meshes are more costly to acquire, so 3D mesh datasets typically come in smaller sizes than image or text datasets. Inference models trained using such small datasets can suffer from over-fitting and generate meshes with various visual artifacts. For example, human pose embedding networks \citep{tan2018mesh,gao2018automatic} can have excessive deformations, and interaction networks \citep{battaglia2016interaction} can result in non-physically-based object motions.
Our main goal is to resolve a major source of visual artifacts: self-collisions. Instead of acquiring more data, we argue that domain-specific knowledge could also be utilized to significantly improve the accuracy of inference models. There have been several prior research works along this line. For example, \citep{DBLP:journals/corr/abs-2008-05440} exploited the fact that near articulated meshes can be divided into multiple components, and they train a recursive autoencoder to stitch the components together. \citep{Zheng_2021_CVPR} utilized the locality of secondary physics motions to learn re-targetable and scalable real-time dynamics animation. Recently, \citep{tan2020lcollision} studied learning-based collision avoidance for 3D meshes corresponding to human poses. They proposed a deep architecture to detect collisions and used numerical optimizations to resolve detected collisions. However, \citep{tan2020lcollision} used a large mesh dataset to obtain stable performance of neural collision detection. Indeed, a deformed 3D mesh typically involves more than $10^4$ elements (voxels, points, triangles) where any pair of two elements can have collisions. Therefore, a huge amount of data is required to present the inference model with enough examples of collisions between all possible element pairs.
\TE{Main Results:} We present a robust method to train neural collision handler for complex, 3D deformable meshes using active learning. Our key observation is that the distribution of penetrating meshes can have a long tail and active learning is an effective method for modeling the tail \cite{geifman2017deep}. Specifically, most penetrating mesh primitives have large overlapping patches consisting of the central part of the distribution, but many other meshes have small patches that penetrate each other, forming the tail. In fact, most 3D mesh datasets do not focus on generating samples in the tail and cannot be used to train stable collision detectors. In order to overcome these issues, our approach combines three main ideas: 1) We use active learning to progressively insert new samples into the dataset. With the help of exact collision detectors, collision labels for the new samples are automatically generated; 2) We use a risk-seeking approach to prioritize samples in the tail, so that the inserted samples can best help the neural collision detector improve its accuracy; 3) We use different loss functions for samples far from and close to the decision boundary. Our overall approach is shown in \prettyref{fig:pipeline}. We apply our training method on the same neural collision detector as \citep{tan2020lcollision}, and we compare with their supervised learning approach on five deformable mesh datasets corresponding to dressed and undressed human poses, cloth simulations, and human hand poses. By comparison, our approach exhibits much higher data efficacy, shows higher accuracy (up to $98.1\%$) in terms of collision detection, and uses much fewer training samples (up to $48.12\%$ fewer). Given a dataset of the same size, our neural collision pipeline reduces the false negative rate by $14.27\%$ on average, and we successfully resolve $24.74\%$ more self-colliding meshes. Overall, ours is the first practical method for neural collision handling for complex 3D meshes.
\section{\label{sec:method}Active Learning Algorithm}
The goal of active learning is to iteratively improve the accuracy of the neural collision detector. We assume the availability of an existing dataset $\mathcal{D}$ of ``high-quality'' meshes with deformed vertices $\mathcal{D}=\{V^{1,2,\cdots,|\mathcal{D}|}\}$, which is used to train a pair of autoencoders ($\left<\theta_E,\theta_D\right>$) via reconstruction loss. We further assume all the groundtruth meshes are (nearly) self-collision free. However, the autoencoders can still suffer from remaining embedding error after training, and users might explore the latent space in regions that are not well covered by the training dataset. All these factors can lead to self-collisions, which should be recognized by our neural collision detector. Therefore, our network cannot be trained with $\mathcal{D}$ alone. This is because $\mathcal{D}$ only contains negative (collision-free) samples, while the neural collision detector must learn the decision boundary between positive and negative samples. In other words, the network must be presented with enough samples to cover all possible latent codes with both self-penetrating and collision-free meshes. We denote the training dataset of neural collision detectors as another set: $\mathcal{D}_c=\{\left<Z_\text{all}^i,I_c^*(Z_\text{all}^i)\right>|i=1,2,\cdots,|\mathcal{D}_c|\}$, where $I_c^*(\bullet)$ is the groundtruth $0-1$ collision state label.
It has been shown in \cite{gal2017deep,aghdam2019active} that many data points in a large image dataset are similar and that human labeling for each point is time-consuming and contains redundant work. In our case, the groundtruth collision state label can be generated automatically using a robust algorithm such as \cite{pan2012fcl}, to compute $\text{PD}$, where a positive $\text{PD}$ indicates self-collisions, so we can define $I_c^*(Z_\text{all})\triangleq\mathbb{I}(\text{PD}(\text{ACAP}^{-1}(Z_\text{all},\theta_D))>0)$. However, the cost to compute penetration depth, $\text{PD}(\bullet)$, is superlinear in the number of mesh vertices, and computing $\text{PD}$ for an entire dataset can still be a computational bottleneck. Moreover, we are considering a continuous space of possible training data that cannot be enumerated. To alleviate the computational burden, we design a three-stage method, as illustrated in \prettyref{fig:pipeline}. During the first stage of bootstrap, we sample an initial boundary set by which we train $\text{MLP}_c$ to approximate the true decision boundary. At the second stage of data augmentation, new training data is selected and progressively injected into a dataset. Finally, for the third stage, our neural collision detector is updated to fit the augmented dataset. The criterion for selecting the subset is critical to the performance of active learning. We observe that our neural collision predictor is used as constraints for nonlinear optimization methods so that samples far from the boundary are not used by the optimizer and only the boundary of the feasible domain (gray area in \prettyref{fig:pipeline} right) is useful. Therefore, we propose using a Newton-type risk-seeking method to push the samples towards the decision boundary. We provide more details for each step below.
\subsection{\label{sec:bd}Bootstrap}
Active learning would progressively populate $\mathcal{D}_c$, so prior work \cite{aghdam2019active} simply initializes the dataset to an empty set. However, we find that a good initial guess can significantly improve the convergence of training. This is because we select new data by moving (randomly sampled) latent codes towards the decision boundary of the $\text{PD}$ function using a risk-seeking method. However, the true boundary of the collision-free constraints corresponds to the boundary of $C$-Obstacles, which is high-dimensional and unknown to us ($\text{PD}$ is a non-smooth function, so we cannot even use gradient information to project a mesh to the zero level-set of $\text{PD}$). Instead, we propose using the learned neural decision boundary, i.e., the $0.5$-levelset of $\text{MLP}_c$, as an approximation. If we initialize $\mathcal{D}_c=\emptyset$, the surrogate decision boundary is undefined, and the training might diverge or suffer from slow convergence. For our bootstrap training, we uniformly sample a small set of $N_\text{init}$ latent codes $\mathcal{Z}_\text{all}$ at random positions from the latent space and compute $\text{PD}$ for each of them. We define a valid space of sampling by mapping all the data $Z_\text{all}^i\in\mathcal{D}$ to their latent codes and compute a bounded box in the latent space:
\small
\begin{align*}
\mathcal{Z}=\prod_{j=0}^{|Z_0|}
\left[\fmin{i=1,\cdots,|\mathcal{D}|}\left[e_j^TZ_\text{all}^i\right],
\fmax{i=1,\cdots,|\mathcal{D}|}\left[e_j^TZ_\text{all}^i\right]\right].
\end{align*}
\normalsize
We hypothesize that all the meshes can be embedded using our autoencoder with small error corresponding to latent codes in $\mathcal{Z}$, so we can initialize $\mathcal{D}_c=\{Z_\text{all}^{1,\cdots,N_\text{init}}|Z_\text{all}^i\sim U(\mathcal{Z})\}$. We then divide the data points into three subsets ($\mathcal{D}_c=\mathcal{D}_p\bigcup\mathcal{D}_n\bigcup\mathcal{D}_b$, illustrated in \prettyref{fig:pipeline} right):
\small
\begin{equation}
\begin{aligned}
\label{eq:subset}
\mathcal{D}_p\triangleq&\{\left<Z_\text{all}^i,I_c(Z_\text{all}^i)\right>|
\text{PD}(\text{ACAP}^{-1}(Z_\text{all},\theta_D))>\epsilon\}\\
\mathcal{D}_n\triangleq&\{\left<Z_\text{all}^i,I_c(Z_\text{all}^i)\right>|
\text{PD}(\text{ACAP}^{-1}(Z_\text{all},\theta_D))<0\}\\
\mathcal{D}_b\triangleq&\{\left<Z_\text{all}^i,I_c(Z_\text{all}^i)\right>|
\text{PD}(\text{ACAP}^{-1}(Z_\text{all},\theta_D))\in[0,\epsilon]\}.
\end{aligned}
\end{equation}
\normalsize
Here, $\mathcal{D}_p$ is the positive set consisting of samples with penetrations deeper than a threshold $\epsilon$, $\mathcal{D}_n$ is the negative set consisting of collision-free samples, and $\mathcal{D}_b$ is a boundary set where samples are nearly collision-free and lie on the decision boundary. We propose using the $l_1$-loss function for $\mathcal{D}_b$ to approximate the decision boundary:
\begin{align}
\mathcal{L}_b=E_{\{Z_\text{all}\in\mathcal{D}_b\}}
\left[\|\text{MLP}_c(S_1,\cdots,S_{|Z_0|},\theta_C)-0.5\|\right].
\label{eq:bd_loss}
\end{align}
\subsection{Data Aggregation}
The accuracy of our neural collision detector can be measured by the discrepancy between the surrogate decision boundary deemed by $\text{MLP}_c$ and the true decision boundary of $\text{PD}$, formulated as:
\begin{align*}
E_{\{Z_\text{all}\sim\mathcal{Z}|\text{PD}(\text{ACAP}^{-1}(Z_\text{all},\theta_D))=0\}}
\left[\text{CE}(I_c(Z_\text{all}),I_c^*(Z_\text{all}))\right],
\end{align*}
which is an expectation over the true decision boundary. Here $\text{CE}$ is the cross-entropy loss. However, it is very difficult to derive a sampled approximation of the above metric because $\text{PD}$ is a non-smooth function whose level-set is measure-zero, which corresponds to the boundaries of C-obstacles. Instead, we propose to take expectation over the surrogate decision boundary:
\begin{align*}
E_{\{Z_\text{all}\sim\mathcal{Z}|\text{MLP}_c(S_1,\cdots,S_{|Z_0|},\theta_C)=0.5\}}
\left[\text{CE}(I_c(Z_\text{all}),I_c^*(Z_\text{all}))\right].
\end{align*}
Generally speaking, the $0.5$-level-set of $\text{MLP}_c$ can also be measure-zero, but we have designed our neural networks $D,\text{CSE},\text{CP},\text{MLP}_c$ to be differentiable functions. As a result, we could always project samples onto the $0.5$-level-set by solving the following risk-seeking unconstrained optimization:
\begin{align*}
\argmin{Z_\text{all}}\;\frac{1}{2}\|\text{MLP}_c(S_1,\cdots,S_{|Z_0|},\theta_C)-0.5\|^2.
\end{align*}
We adopt the quasi-Newton method and update $Z_\text{all}$ using the following recursion:
\footnotesize
\begin{align}
\label{eq:newton}
Z_\text{all}-\bar{H}^{-1}\nabla\text{MLP}_c(\text{MLP}_c-0.5),
\end{align}
\normalsize
where $\bar{H}$ is some first-order approximation of the Hessian matrix, which is much faster to compute than the exact Hessian, which requires the second-order term $\nabla^2\text{MLP}_c$. In summary, we would sample a new set of size $N_\text{aug}/2$ from previous $\mathcal{D}_c$ during each iteration of data augmentation. For each sampled $Z_\text{all}$, we project $Z_\text{all}$ to the surrogate decision boundary using recursive \prettyref{eq:newton} until the relative change within $Z_\text{all}$ is smaller than $\epsilon_z$ between consecutive iterations. We also sample $N_\text{aug}/2$ directly from $U(\mathcal{Z})$, using random samples to discover uncovered regions, which achieves a balance between exploitation and exploration. Finally, we classify $Z_\text{all}$ into either one of $\mathcal{D}_{p,n,b}$, according to \prettyref{eq:subset} using the penetration depth.
\subsection{Model Update}
After $\mathcal{D}_c$ has been updated, we fine-tune $\text{CSE},\text{CP},\text{MLP}_c$ by updating $\theta_C$ using the following loss functions:
\begin{align*}
\mathcal{L}=w_\text{PD}\mathcal{L}_\text{PD}+w_r\mathcal{L}_r+w_\text{ce}\mathcal{L}_\text{ce}+w_b\mathcal{L}_b,
\end{align*}
where $w_\bullet$ are weights corresponding to each type of loss. Our first term $\mathcal{L}_\text{PD}$ is a regularization that enforces consistency between $S_i$ and true $\text{PD}$, defined as:
\small
\begin{align*}
\mathcal{L}_\text{PD}=
E_{\{Z_\text{all}\in\mathcal{D}_C\}}
\left[\sum_{i=1}^{|Z_0|}\|S_i-\text{PD}_i\|^2+w_\text{PDsum}\|\sum_{i=1}^{|Z_0|}S_i-\text{PD}\|^2\right],
\end{align*}
\normalsize
where we penalize both domain-decomposed penetration depth $\text{PD}_i$ defined in \cite{tan2020lcollision} and total penetration depth with weight $w_\text{PDsum}$. Our second term $\mathcal{L}_r$ is a marginal ranking loss that enforces the correct ordering of penetration depth to avoid over-fitting, which is defined as:
\small
\begin{align*}
\mathcal{L}_r=E_{\{Z_\text{all}^{a,b}\in\mathcal{D}_C|\text{PD}^a<\text{PD}^b\}}
\left[\max(0,\alpha-(\sum_{i=1}^{|Z_0|}S_i^a-\sum_{i=1}^{|Z_0|}S_i^b)\right],
\end{align*}
\normalsize
where $\alpha$ is the maximal allowable order violation. We use superscripts to distinguish two samples drawn from $\mathcal{D}_C$. Our third term measures the discrepancy between $\text{MLP}_c$ and $\text{PD}$ over the entire latent space:
\small
\begin{align*}
\mathcal{L}_\text{ce}=
E_{\{Z_\text{all}\in\mathcal{D}_p\bigcup\mathcal{D}_n\}}
\left[\text{CE}(I_c(Z_\text{all}),I_c^*(Z_\text{all}))\right].
\end{align*}
\normalsize
We update our neural collision detector with objective function $\mathcal{L}$ by running a fixed number of training epochs, denoted as $N_\text{epoch}$, with $\theta_C$ warm-started from the last iteration of the model update. Our overall training method is illustrated in \prettyref{alg:activeLearning}.
\begin{algorithm}
\caption{Learning $\theta_E,\theta_D,\theta_C$}
\label{alg:activeLearning}
\begin{algorithmic}[1]
\begin{footnotesize}
\STATE Prepare initial data set $\mathcal{D}$
\STATE Update $\theta_E,\theta_D$ using reconstruction loss and $\mathcal{D}$
\STATE Initialize $\mathcal{D}_C$ by drawing $N_\text{init}$ samples from $U(\mathcal{Z})$
\FOR{$Z_\text{all}\in\mathcal{D}_C$}
\STATE Compute $\text{PD}(\text{ACAP}^{-1}(Z_\text{all},\theta_D))$ and $I_c^*(Z_\text{all})$
\ENDFOR
\STATE Update $\theta_C$ using $\mathcal{D}_C$
\WHILE{Not converged}
\STATE{${\Delta\mathcal{D}_C}_1\gets$Draw $N_\text{aug}/2$ samples from $\mathcal{D}_C$}
\FOR{$Z_\text{all}\in{\Delta\mathcal{D}_C}_1$}
\STATE{$Z_\text{all}^\text{last}\gets Z_\text{all}$}
\WHILE{True}
\STATE Update $Z_\text{all}$ using \prettyref{eq:newton}
\IF{$\|Z_\text{all}-Z_\text{all}^\text{last}\|_\infty$ sufficiently small}
\STATE Break
\ENDIF
\STATE{$Z_\text{all}^\text{last}\gets Z_\text{all}$}
\ENDWHILE
\STATE $\mathcal{D}_C\gets\mathcal{D}_C\bigcup\{\left<Z_\text{all}^\text{last},I_c^*(Z_\text{all}^\text{last})\right>\}$
\ENDFOR
\STATE{${\Delta\mathcal{D}_C}_2\gets$Draw $N_\text{aug}/2$ samples from $U(\mathcal{Z})$}
\FOR{$Z_\text{all}\in{\Delta\mathcal{D}_C}_2$}
\STATE $\mathcal{D}_C\gets\mathcal{D}_C\bigcup\{\left<Z_\text{all},I_c^*(Z_\text{all})\right>\}$
\ENDFOR
\STATE Update $\theta_C$ using $\mathcal{D}_C$
\ENDWHILE
\end{footnotesize}
\end{algorithmic}
\end{algorithm}
\section{\label{sec:problem}Neural Collision Handler}
In this section, we briefly review the mesh-based generative models with a neural collision handler \citep{tan2020lcollision}, based on which we build our active learning method. All notations are summarized in the symbol table.
A mesh is represented by the graph $\mathcal{G}=\left<V,E\right>$, where $V$ is a set of vertices, and $E$ is a set of edges. We assume that all the meshes have the same topology, that is, all the meshes differ in $V$ while the connectivity $E$ stays the same. We further limit ourselves to manifold triangle meshes, i.e., each edge is incident to at most two triangles, and two triangles are adjacent if and only if they share an edge. No other assumptions are made on the mesh deformation. We denote a mesh as self-collision-free if and only if any pair of two non-adjacent triangles are not intersecting each other. Our goal is to design a mesh-based generative neural architecture where we take as input a coordinate in the latent space and output a 3D mesh without self-collisions. The latent space is defined as a low-dimensional space that can be mapped to high-dimensional meshes injectively using a learned decoder function (\prettyref{fig:pipeline} right). Furthermore, the latent-to-mesh mapping is differentiable and supports multiple downstream applications explained in \prettyref{sec:app}. Our method consists of two parts. First, we train a bilevel mesh autoencoder using supervised learning. Second, we train a neural collision handler using active learning and a special boundary loss.
\subsection{Bilevel Autoencoder}
Our bilevel autoencoder architecture maps a deformed mesh to two levels of latent codes. The latent codes are used to both recover the high-dimensional deformed mesh vertices $V$ and predict whether the deformed mesh contains self-collisions. We only want to encode intrinsic mesh information such as curvatures instead of extrinsic rigid transformations because mesh shapes are invariant to extrinsic transformation. Therefore, we first use as-consistent-as-possible (ACAP) feature transformation \cite{gao2019sparse} to factor out rigid transformations. The ACAP feature vector is first brought through the level-1 autoencoder and mapped to a latent code $Z_0$. Since our dataset size is small, we use a shallow autoencoder to avoid over-fitting and $Z_0$ is subject to large embedding error. We further hypothesize that the error is sparsely distributed throughout the mesh vertices. Therefore, we use an attention mechanism trained with a sparsity prior to decompose the mesh into near-rigid sub-domains. The sparsity prior is designed such that each domain can be mapped to a single axis of the latent space, i.e., a single entry of $Z_0$. Afterwards, a set of $|Z_0|$ level-2 autoencoders is introduced to further reduce the error (\prettyref{fig:pipeline} left), with each autoencoder dedicated to one entry of $Z_0$. Their latent codes are denoted as $Z_1,\cdots,Z_{|Z_0|}$. The ultimate mesh is reconstructed from $Z_\text{all}$ by combining level-1 and level-2 latent codes:
\small
\begin{align*}
Z_\text{all}=&
\FOUR{Z_0}{Z_1}{\cdots}{Z_{|Z_0|}}\\
D(Z_\text{all},\theta_D)=&
\sum_{i=0}^{|Z_0|}D_i(Z_i,\theta_{D_i})\\
V=&\text{ACAP}^{-1}(D(Z_\text{all},\theta_D)),
\end{align*}
\normalsize
where $D_i$ is the $i$th decoder, with $\theta_{D_i}$ being the learnable parameters. The mesh vertices $V$ are reconstructed by inverting the ACAP transformation. Correspondingly, we have the encoder defined as $E(\text{ACAP}(V),\theta_E)=Z_\text{all}$, which maps the vertices of a mesh to the latent space, with $\theta_E$ being the learnable parameters.
\subsection{Neural Collision Detector}
\input{figs/PD.tex}
Our neural collision detector predicts whether the mesh $V$ is subject to self-collisions using latent information $Z_\text{all}$. The extent to which two meshes collide can be measured by the notion of Penetration Depth (PD)~\cite{zhang2014continuous}, defined by the norm of the smallest configuration change needed for a mesh to be self-collision-free, as illustrated in \prettyref{fig:PD}. It is well-known that PD is a non-smooth function of $V$ (esp. at the boundaries), thereby making it difficult to resolve collisions by minimizing PD. By choosing appropriate activation functions ($\tanh$ and CELU in our case), we design the neural collision detector to be a differentiable approximation of PD. As a result, gradient information can be propagated to a collision handler to minimize PD.
Since collisions can happen between any pairs of geometric mesh primitives, a collision detector should consider possible contacts between any pair of near-rigid sub-domains, leading to a quadratic complexity $\mathcal{O}(|Z_0|^2)$. We use a global-local detection architecture that effectively reduces the number of learnable parameters. Specifically, we introduce a global collision state encoder $S_0=\text{CSE}(Z_\text{all},\theta_C)$ and a set of local collision predictors $S_i=\text{CP}(S_0,Z_i,\theta_C)$, with $i=1,\cdots,|Z_0|$, which predicts whether the $i$th sub-domain is in collision with the rest of the mesh. Finally, the collision information for all local collision predictors is summarized using a classifier network $\text{MLP}_c(S_1,\cdots,S_{|Z_0|})$ to derive a single overall collision classifier:
\small
\begin{align*}
S_0\triangleq&\text{CSE}(Z_\text{all},\theta_C)\\
S_i\triangleq&\text{CP}(S_0,Z_i,\theta_C)\\
I_c(Z_\text{all})\triangleq&\mathbb{I}(\text{MLP}_c(S_1,\cdots,S_{|Z_0|},\theta_C)\geq0.5),
\end{align*}
\normalsize
where $\theta_C$ is the learnable parameters. The feasible space boundary of collision-free constraints corresponds to the $0.5$-levelset of $\text{MLP}_c$. This architecture combining $\text{CSE}$, $\text{CP}$, and $\text{MLP}$ has several learnable parameters proportional to $\mathcal{O}(|Z_0|)$.
\subsection{Optimization-Based Collision Response\label{sec:app}}
Existing collision handling techniques \cite{tang2012continuous,Narain:2012:AAR} are mostly based on numerical optimizations, where a key challenge is to make sure that collision constraints are differentiable. \citep{tan2020lcollision} also uses this formulation, but it is guided by the learned collision detector $\text{MLP}_c$, which is differentiable by construction. Suppose we take as input a randomly sampled latent code $Z_\text{all}^\text{user}$, which might not satisfy the collision-free constraints. We then need to project that latent code back to the feasible domain of collision-free meshes. We achieve this by solving the following optimization problem under neural collision-free constraints using the Augmented Lagrangian Method (ALM):
\begin{align}
\label{eq:handler}
\argmin{Z_\text{all}}\;E(Z_\text{all})\quad\text{s.t.}\;
\text{MLP}_c(S_1,\cdots,S_{|Z_0|},\theta_C)\leq0.5,
\end{align}
where $E(\bullet)$ is some objective function, which can take multiple forms, as specified by downstream applications. In the simplest case, we take as input a desired $Z_\text{all}^\text{user}$, and we can define $E(Z_\text{all})=\|Z_\text{all}-Z_\text{all}^\text{user}\|^2/2$, which is only related to latent space variables. As a more intuitive interface, the user might want to change meshes in the Cartesian space instead of the of latent space. For example, if the user wants a human hand to be at a certain position $V^\text{user}$, we could define $E(Z_\text{all})=\|\text{ACAP}^{-1}(D(Z_\text{all},\theta_D))-V^\text{user}\|^2/2$. A desirable feature of \prettyref{eq:handler} is an invariant problem size. However many vertices a mesh has, there is only one constraint, which guarantees high test-time performance. Moreover, it has been shown in \cite[Theorem~10.4.3]{sun2006optimization} that ALM either finds a feasible solution or returns an infeasible solution that is closest to the boundary of the feasible domain. In other words, ALM always makes a best effort to resolve collisions, even if feasible solutions are not available.
\section{\label{sec:related}Related Work}
Our method is designed for general deformable meshes with fixed topology. We first learn a latent-space of meaningful mesh deformations and then use active learning to train a neural collision detector that identifies a collision-free subspace. We review related work in these areas.
\TE{Generative Model of Dense 3D Shapes:} Categorized by shape representations, generative models can be based on point clouds \citep{qi2017pointnet++}, volumetric grids \citep{wu20153d}, multi-charts \cite{groueix2018papier}, surface meshes \citep{tan2018variational}, or semantic data structures \cite{10.1145/3306346.3322969}. We use mesh-based representations with fixed topologies, because most collision detection libraries are designed for meshes. Note that this choice excludes several applications that require general meshes of changing topology, e.g., for modeling meshes of hierarchical structure \citep{yu2019partnet} or modeling scenes with many objects \citep{ritchie2019fast}. However, these applications typically involve only static meshes with no need for collision detection. There is a separate research direction on domain-specific mesh deformation representation, e.g., SMPL/STAR human models \citep{SMPL:2015,STAR:2020}, wrinkle-enhanced cloth meshes \citep{lahner2018deepwrinkles}, and skeletal skinning meshes \citep{RigNet}. There are even prior works \citep{Fieraru_2020_CVPR,fieraru2021learning,Muller_2021_CVPR} on collision detection and handling for human bodies. These domain-specific methods are typically more accurate than our representation, but by assuming general meshes, our representation can be applied to multiple domains as shown in \prettyref{sec:evaluation}.
\TE{Collision Prediction \& Handling:} Although collision detection has been well studied and mature software packages are available, detecting and handling self-collisions can still be a non-trivial computational burden for large meshes. Prior methods \citep{pan2012fcl,kim2018chapter,govindaraju2005quick} use spatial hashing, bounding volume hierarchies, and GPU to accelerate the computation by pruning non-colliding primitives, but they cannot be generalized to learning methods. Handling collisions is even more challenging, and prior physically-based methods either use penalty forces coupled with discrete collision detectors \citep{tang2012continuous} or hard constraints coupled with continuous collision detectors \citep{Narain:2012:AAR}. All these methods rely on physics-based constraints to handle collisions. Recently, many learning methods such as \citep{Gundogdu_2019_ICCV,patel20tailornet} have been designed to predict the cloth movement or deformation in 3D, but they do not perform collision handling explicitly.
\TE{Active Learning:} An active learner alternates between drawing new samples and exploiting existing samples. These samples can be drawn guided by an acquisition function in Bayesian optimization \citep{niculescu2006bayesian} or from an expert algorithm \citep{de2018learning}. Active learning has been applied to approximate the boundary of the configuration space \citep{pan2013efficient,tian2016efficient,das2017fastron}, where the feasible domain of collision constraints is parameterized using kernel SVM. However, this method is limited to rigid or articulated deformation and is not applicable to general 3D deformations. More broadly, active learning has been adopted in various prior works to accelerate data labeling in image classification \citep{gal2017deep} and object detection \citep{aghdam2019active} tasks. These methods progressively identify unlabeled images to be forwarded to experts for labelling. An alternative method for selecting the samples is identifying a coreset \citep{paul2014visual}, and authors of \citep{sener2018active} propose a practical algorithm for coreset identification via k-center clustering. While these methods consider a pre-existing set of unlabeled images for training discriminative models, we assume a continuous latent space of samples for training generative models and use a risk-seeking method to identify critical new samples. Further, we automatically label the new samples using analytical collision detection algorithms, making the training method fully automatic.
|
\section{Introduction}
\IEEEPARstart{A}{rtificial} Neural Networks (ANNs) have shown impressive performance across various computer vision fields for a few decades \cite{he2016deep,simonyan2014very,girshick2015fast}.
However,
ANNs suffer from huge computational costs \cite{sze2017efficient},
limiting their application in power-hungry systems such as Internet-of-Things (IoT) devices.
As an alternative to low-power ANNs, recent studies have focused on bio-plausible Spiking Neural Networks (SNNs) \cite{roy2019towards,deng2020rethinking}, which process visual information with temporal binary events (\textit{i}.\textit{e}., spikes).
Spikes stimulate neuronal membrane potentials and convey discriminative information from shallow layers to deep layers.
SNN neurons only consume energy whenever spikes are generated which allow for these asynchronous processes to be implemented on highly energy-efficient neuromorphic hardware \cite{furber2014spinnaker,akopyan2015truenorth,davies2018loihi}.
\begin{figure}[t]
\begin{center}
\def0.5{0.5}
\begin{tabular}{@{}c@{\hskip 0.05\linewidth}c@{}c}
\includegraphics[width=0.9\linewidth]{figures/intro_classseg_concept.pdf} \\
\end{tabular}
\end{center}
\caption{The concept of image classification and semantic segmentation. For a given image, the classification network provides a class prediction (\eg Airplane). On the other hand, the semantic segmentation network maintains the high resolution feature map at the end of the network (colored in blue) and assigns every pixel in the image its own class, resulting in a two-dimensional prediction map.
}
\label{fig:intro:class_seg_concept}
\end{figure}
In order to exploit the energy advantage of binary information transmission, optimization algorithms for SNNs have been developed in the past few decades, with a focus on image classification.
Among various training algorithms, the ANN-SNN conversion method \cite{sengupta2019going,han2020rmp,diehl2015fast,rueckauer2017conversion} has been highlighted as a result of its simplicity and high performance.
Conversion replaces the neurons of a pre-trained ANN using the ReLU activation function with Integrate-and-Fire (IF) neurons in an SNN.
The conversion method avoids the complexity of spike-based training since it relies on ANN training and thus, yields high performance on complex data.
However, since conversion naively scales the firing thresholds with the maximum layer activations, it is hard to capture spike information across the spatial and temporal domain.
As a result, the SNN created using ANN-SNN conversion requires thousands of time-steps to achieve similar performance as the original pre-trained ANN, and has been limited to fundamental vision tasks such as image classification.
Considering the likelihood of the future deployment of SNNs in scene-understanding tasks, it is essential to investigate the applicability of SNNs for beyond classification.
For example, a vision system for autonomous vehicles should contain an energy-efficient neural module for analyzing the overall context of scenery \cite{yurtsever2020survey,treml2016speeding}.
Therefore, in this paper, we move beyond classification by exploring semantic segmentation for neuromorphic systems.
As shown in Fig. \ref{fig:intro:class_seg_concept}, a semantic segmentation network conducts pixel-wise classification, resulting in a two-dimensional prediction map.
Since partitioning multiple foreground objects from the background is an essential and fundamental vision task, semantic segmentation has been extensively studied in the ANN domain \cite{he2016deep,kim2019cnn,zhang2018context,chen2018encoder,long2015fully}.
Unfortunately, we found that the state-of-the-art ANN-SNN conversion technique fails to achieve reasonable SNN performance on a semantic segmentation dataset.
This is because segmentation datasets cause ANNs to have huge activation variance across the spatial and channel axes (will be discussed in Section \ref{section:preliminary study}).
Moreover, ANN-SNN conversion cannot be compatible with a spike stream dataset from Dynamic Vision Sensor (DVS) camera \cite{patrick2008128x, lichtsteiner2006128, posch2014retinomorphic, delbruck2010activity}, which limits the potential advantage of a neuromorphic system.
To address the aforementioned problems, we focus on surrogate gradient backpropagation learning \cite{neftci2019surrogate,lee2016training,wu2018spatio} rather than ANN-SNN conversion, which directly trains SNNs from input spikes.
The major difficulty of gradient backpropagation learning with SNNs is the non-differentiable neuronal functionality of SNNs.
A Leaky Integrate-and-Fire (LIF) neuron does not generate spikes before its membrane potential exceeds a firing threshold, resulting in a non-differentiable point during backpropagation.
Therefore, we approximate the backward gradient function of an LIF neuron using a piece-wise linear function during backpropagation.
Using the approximated gradient function, the weight parameters are optimized in order to minimize spatial cross-entropy loss.
In addition, we leverage Batch Normalization Through Time (BNTT) \cite{kim2020revisiting}, a recent work proposes a temporal batch normalization, which enables our segmentation networks to be trained from scratch.
As a result, SNNs with BNTT-backpropagation can learn the temporal dynamics of input spikes, with shorter number of time-steps and better performance compared to ANN-SNN conversion.
With surrogate learning, we investigate two representative segmentation architectures from the ANN domain, \textit{i}.\textit{e}., Fully Convolutional Networks (FCNs) \cite{long2015fully} and
DeepLab with dilated (\textit{i}.\textit{e}., atrous) convolutional layers \cite{he2016deep}.
The FCN architecture consists of an encoder-decoder architecture, which recovers the resolution of the original image with the upsampling decoder.
On the other hand, DeepLab only has an encoder architecture but uses dilated convolutions to cover wide receptive fields without computational overhead.
Due to the fact that very deep SNN architectures suffer performance degradation because of the mismatch between real gradients and approximated gradients, we use the VGG9 architecture as the backbone for our networks for segmentation.
In summary, the main contributions of this work are as follows:
(i) To the best of our knowledge, our work is the first work to train SNNs for semantic segmentation.
This is an important research direction given that low-power SNNs will be deployed in scene-understanding tasks.
(ii) We investigate two representative SNN optimization techniques (\eg ANN-SNN conversion and surrogate gradient learning) on semantic segmentation datasets.
We observe that surrogate gradient learning achieves better performance with lower latency compared to ANN-SNN conversion.
(iii)
We present Spiking-FCN and Spiking-DeepLab, which expand upon the two fundamental architectures used for semantic segmentation in the ANN domain.
Also, we conduct comprehensive experiments on various benchmarks including static PASCAL VOC2012 and DDD17 from a DVS camera.
The proposed segmentation architectures can bring a more than $2\times$ energy efficiency than standard ANNs.
The paper is organized as follows. Section II introduces background knowledge on SNNs and the semantic segmentation work done in the ANN domain.
Section III presents our experiments on ANN-SNN conversion.
Section IV details our methods of training segmentation networks with spiking neurons.
Lastly, in Section V, we present our comprehensive experimental results on both the static and video segmentation datasets.
At the end of this paper, we provide our conclusions and point out future work directions.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=0.5\textwidth]{figures/spike_activity_concept.pdf}
\end{center}
\caption{ The neuronal dynamics of SNNs. Pre-synaptic neurons convey spike trains to a post-synaptic neuron. This increases the membrane potential voltage and the post-synaptic neuron generates the output spikes whenever the membrane potential exceeds the neuronal firing threshold. After that, the membrane potential is set to a reset voltage. }
\label{fig:relatedwork:neuron_dynamic}
\end{figure}
\section{Related Work}
\subsection{Spiking Neural Network}
Spiking Neural Networks (SNNs) have emerged as the next generation of neural networks \cite{roy2019towards,panda2020toward,cao2015spiking,diehl2015unsupervised,comsa2020temporal,venkatesha2021federated,kim2021optimizing,kim2021privatesnn}, because they offer huge energy-efficiency advantage over ANNs.
Different from standard ANNs that make use of float values, SNNs process binary spikes (0 or 1) across multiple time-steps.
SNNs take binary spike trains as an input, which can be obtained from both static RGB images and DVS camera data.
For static images, various coding schemes have been proposed.
Poisson rate coding generates spikes in which the number of spikes is proportional to the pixel intensity.
Due to its simplicity and high performance, Poisson rate coding has been widely used in previous works \cite{roy2019towards,diehl2015unsupervised,lee2016training}.
Temporal coding allows only one spike per neuron, resulting in energy efficiency from fewer spikes. Here, spike latency is inversely proportional to the pixel intensity \cite{mostafa2017supervised,park2020t2fsnn,comsa2020temporal}.
Thus, bright pixels generate more spike events in earlier time-steps than dark pixels.
However, for DVS camera data, SNNs can be directly trained without any code generator.
In terms of an activation function, SNNs commonly use a Leaky Integrate-and-Fire (LIF) neuron \cite{gerstner2002spiking}.
The LIF neuron (Fig. \ref{fig:relatedwork:neuron_dynamic}) has a membrane potential which can store the temporal information by accumulating pre-synaptic spikes.
The neuron generates a post-synaptic spike whenever the membrane potential exceeds a predefined firing threshold.
This integrate-and-fire behavior is non-differentiable, so SNNs are hard to train with standard backpropagation \cite{neftci2019surrogate}.
To address this limitation, work in the past decade has focused on various training techniques for SNNs.
Spike-timing-dependent plasticity (STDP) learning \cite{bi1998synaptic,hebb2005organization,bliss1993synaptic} is based on the neuroscience observation that weight connections can be reinforced or punished according to the temporal correlation of spikes.
STDP learning can be implemented without a complicated backpropagation module but can only be applied to small-scale tasks due to its locality \cite{yousefzadeh2018practical,jin2010implementing}.
ANN-SNN conversion methods have received attention due to their high performance on complex tasks \cite{sengupta2019going,han2020rmp,diehl2015fast,rueckauer2017conversion}.
In order to approximate ReLU activations with LIF activations, pre-trained ANNs are converted to SNNs using weight (or threshold) balancing techniques.
ANN-SNN conversion requires many time-steps to represent the float values of ANNs with binary spikes.
Surrogate gradient learning addresses the non-differentiability of LIF neurons by defining a surrogate gradient function in a backpropagation process \cite{lee2016training,neftci2019surrogate}.
This training scheme enables SNNs to learn the temporal dynamics of spike trains, resulting in small latency and reasonable performance.
Unfortunately, most SNN optimization algorithms are developed for basic, fundamental vision tasks such as image recognition \cite{lee2016training,kim2020revisiting,diehl2015unsupervised}, visualization \cite{kim2021visual}, and optimization \cite{fang2019swarm,frady2020neuromorphic}.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=0.9\linewidth]{figures/conversion_timestep.pdf} \\
\end{center}
\caption{
{
Performance of converted SNNs with respect to the number of time-steps. We evaluate two conversion algorithms (\textit{i}.\textit{e}., layer-wise conversion \cite{sengupta2019going} and channel-wise conversion \cite{kim2020spiking}) with a DeepLab architecture trained on PASCAL VOC2012. Compared to the ANN performance (black dotted line), the converted SNNs suffer from performance degradation caused by a large variation in activations.}
}
\label{fig:prelimin:conversion_studies}
\end{figure}
\begin{figure*}[t]
\begin{center}
\def0.5{0.5}
\begin{tabular}{@{}c@{\hskip 0.05\linewidth}c@{}c}
\includegraphics[width=0.48\linewidth]{figures/channel_activation.pdf} &
\includegraphics[width=0.48\linewidth]{figures/spatial_activation.pdf}
\\
{\hspace{4mm} (a) Channel activation } & {\hspace{4mm} (b) Spatial activation }\\
\end{tabular}
\end{center}
\caption{Illustration of normalized maximum activations across (a) channel and (b) spatial axes. We use a VGG9-based DeepLab architecture trained on PASCAL VOC2012 dataset. In the figure, the \textit{x} and \textit{y} axes denote, respectively, the index of the channel/spatial axes in that layer and the normalized maximum activations across the channel/spatial axes.
}
\label{fig:prelim:spikeactivation}
\end{figure*}
\subsection{Semantic Segmentation}
The objective of semantic segmentation is to classify every pixel in an image with a label, a key task for scene understanding.
This becomes more important with the development of medical image analysis \cite{ronneberger2015u,cciccek20163d}, autonomous vehicles \cite{treml2016speeding,janai2020computer}, and augmented reality \cite{minaee2020image}.
Therefore, semantic segmentation with deep neural networks has been extensively studied in the ANN domain.
The Fully Convolutional Network (FCN) \cite{long2015fully} is the one of the pioneering works in deep semantic segmentation.
FCNs preserve image details by adding the intermediate high-resolution feature maps into its decoder path.
In a similar way, U-Net \cite{ronneberger2015u} concatenates intermediate feature maps during its up-convolution process.
Different from previous approaches, deconvolution networks \cite{noh2015learning} do not exploit a skip connection but learn a detailed representation from low-resolution feature maps.
RefineNet \cite{lin2017refinenet} uses multi-scale inputs in order to enhance detail in the segmentation map.
DeepLab \cite{he2016deep} utilizes dilated (\textit{i}.\textit{e}., atrous) convolutions which enlarge the receptive field without any additional computational cost.
Dilated convolutions have become one of the most popular techniques for semantic segmentation since they are easy to be implemented in deep learning frameworks (\eg TensorFlow \cite{tensorflow2015-whitepaper}).
Despite its importance and fast growth in the ANN domain, semantic segmentation has not yet been studied with spiking neurons.
So far, in the SNN domain, a few works \cite{meftah2010segmentation,zheng2019image} have proposed region segmentation without providing semantic information (\textit{i}.\textit{e}., class information).
In order to fill the gap, in this work, we configure the spiking versions of segmentation networks based on two representative segmentation architectures, \textit{i}.\textit{e}., FCN and DeepLab.
\section{Preliminary Study}
\label{section:preliminary study}
In this section, we apply ANN-SNN conversion methods to semantic segmentation.
First, we train the ANN-DeepLab with conventional 2D-cross entropy loss \cite{he2016deep}.
After that, we convert the pre-trained ANN to an SNN with two conversion methods \cite{sengupta2019going,kim2020spiking}.
Sengputa \textit{et al}. \cite{sengupta2019going} propose a state-of-the-art conversion technique in the image recognition domain. They take into account spike behavior in the previous layers to achieve accurate weight balancing.
Also, the authors of \cite{kim2020spiking} propose channel-wise weight balancing for each layer in order to capture a feature with high variation in an object detection scenario.
We evaluate these methods on a DeepLab architecture with a VGG9 backbone network \cite{chen2017deeplab} on PASCAL VOC2012 dataset \cite{everingham2010pascal}.
{In Fig. \ref{fig:prelimin:conversion_studies}, we observe that the ANN-SNN conversion process significantly degrades the performance from a pre-trained ANN even with thousands of time-steps.
This is a relatively huge loss compared to ANN-SNN conversion on image classification which shows less than $1\%$ accuracy degradation \cite{sengupta2019going}.
}
The reason for this is that networks trained for the semantic segmentation task have a large variation of activations in both the channel and spatial axes (Fig. \ref{fig:prelim:spikeactivation}).
This causes ANN-SNN conversion to fail to find a proper scaling factor.
Thus, applying the same balancing factor across layers (or channels) does not fully preserve delicate activations.
More importantly, ANN-SNN conversion cannot be applied on video spike streams from a DVS camera, since ANN-SNN conversion is only compatible with static images.
Thus, we focus on a direct training approach in order to implement semantic segmentation with spiking neurons.
\section{Methodology}
\subsection{Leaky Integrate-and-Fire (LIF) Neuron}
\label{sec:LIF neuron}
We leverage a Leaky Integrate-and-Fire (LIF) neuron in our spiking segmentation networks.
Fig. \ref{fig:relatedwork:neuron_dynamic} illustrates the dynamics of an LIF neuron.
The LIF neuron can be represented with a membrane potential and an input signal.
The membrane potential $U_m$ stores the temporal spike information in capacitance. When an input signal $I(t)$ is fed into the LIF neuron, the membrane potential is changed according to the following differential equation:
\begin{equation}
\tau_m \frac{dU_m}{dt} = -U_m + RI(t),
\label{eq:LIF_origin}
\end{equation}
where, $R$ is an input resistance for the LIF circuit and $\tau_m$ is the time constant for the membrane potential decay.
Since the voltage and current have continuous values, we convert the differential equation into a discrete version in order to conduct digital simulation.
We can represent the membrane potential $u_{i}^{t}$ of a single neuron $i$ at time-step $t$ as:
\begin{equation}
u_i^t = \lambda u_i^{t-1} + \sum_j w_{ij}o^t_j.
\label{eq:LIF}
\end{equation}
Here, the current membrane potential consists of the decayed membrane potential from previous time-steps and the weighted spike signal from the pre-synaptic neurons $j$.
The notation $\lambda$ and $w_{ij}$ are for a leak factor and weight connection between the pre-synaptic neuron $j$ and the post-synaptic neuron $i$, respectively.
When $u_i^{t}$ exceeds the firing threshold, the neuron $i$ generates a spike output $o_i^{t}$:
\begin{equation}
o^{t}_i =
\begin{cases}
1, & \text{if $u_i^{t} >\theta$}, \\
0
& \text{otherwise.}
\end{cases}
\label{eq:firing}
\end{equation}
After the neuron fires, the membrane potential is lowered by the amount of the threshold (\textit{i}.\textit{e}., soft reset).
In our experiments, we use a soft reset scheme since it achieves better performance than a hard reset (\textit{i}.\textit{e}., reset to the minimum voltage such as zero).
This is because a soft reset allows the neuron to retain residual information after the reset, thus preventing information loss \cite{han2020rmp}.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=1.0\linewidth]{figures/dataset_visualization_pascal.pdf} \\
{\hspace{4mm}(a) PASCAL VOC2012} \\
\vspace{1mm}
\includegraphics[width=1.0\linewidth]{figures/dataset_visualization_ddd.pdf} \\
\vspace{-2mm}
{\hspace{4mm}(b) DDD17}\\
\end{center}
\vspace{-2mm}
\caption{Examples of input representations. (a) We convert a static image into Poisson spikes. As time goes on, the accumulated spikes represent similar image to original image.
(b) For the event-based camera input, we accumulate spikes in the pre-defined time window (\eg 10ms) to generate frames.
}
\label{fig:relatedwork:datasetvisualization}
\end{figure}
\begin{figure*}[t!]
\begin{center}
\includegraphics[width=0.9\textwidth]{figures/architecture_deeplab.pdf}
\end{center}
\caption{Spiking-DeepLab: a computational graph unrolled over multiple time-steps. Each neuron cell has a membrane potential consisting of temporal voltage propagation (\textit{i}.\textit{e}., horizontal arrow) and the spikes from previous layer (\textit{i}.\textit{e}., vertical arrow).
In order to increase the receptive field in deep layers, we add a dilated LIF neuron cell which can cover a larger spatial area with the same number of parameters.
For the last layer, we accumulate output spikes across all time-steps and generate a 2-dimensional probabilistic map. Finally, we upsample the final prediction map to the original resolution of the input image and calculate the loss function to update the weights.
}
\label{fig:method:deeplab_architecture}
\end{figure*}
\subsection{Input Representation}
\label{ssec:input_representation}
In this paper, we use two types of input datasets, \textit{i}.\textit{e}., static images and DVS data.
For training and inference, a static image needs to be converted into spike trains since SNNs process multiple binary spikes.
There are various spike coding schemes such as rate, temporal, and phase \cite{mostafa2017supervised,kim2018deep}.
Among them, we use rate coding due to its reliable performance across various tasks.
Rate coding provides spikes proportional to the pixel intensity of the given image.
In order to implement this, following previous work \cite{roy2019towards}, we compare each pixel value with a random number ranging between $[I_{min}, I_{max}]$ at every time-step. Here, $I_{min}$ and $I_{max}$ correspond to the minimum and maximum possible pixel intensities.
If the random number is greater than the pixel intensity, the Poisson spike generator outputs a spike with amplitude $1$.
Otherwise, the Poisson spike generator does not yield any spikes.
We visualize rate coding in Fig. \ref{fig:relatedwork:datasetvisualization}(a). We see that the spikes generated at a given time-step is random. However, as time goes on, the accumulated spikes represent a similar result to the original image.
For a DVS spike stream, we accumulate spikes in a certain time window to generate a frame. Then, the network processes these frames like a video input (Fig. \ref{fig:relatedwork:datasetvisualization}(b)).
\subsection{Spiking Segmentation Network Architecture}
\subsubsection{Spiking-DeepLab}
Different from image recognition tasks, segmentation networks provide pixel-wise classification with respect to the given 2-dimensional input image.
Therefore, the segmentation architecture needs to maintain a large spatial resolution of the feature maps at the end of the network.
Also, taking a large receptive field helps the networks figure out the relationships between objects in the scene, resulting in high performance.
To this end, DeepLab \cite{chen2017deeplab} proposed a dilated convolution in order to fulfill these two objectives.
The dilated convolution puts space between the kernel weights.
As a result, the dilated convolution operation increases the size of the receptive field without adding a memory burden.
With the LIF neuron model, for a neuron at ${(i_x, i_y)}$, we can reformulate the neuronal dynamics (Eq. \ref{eq:LIF}) with the dilated convolution operation:
\begin{equation}
u_{(i_x, i_y)}^t = \lambda u_{(i_x, i_y)}^{t-1} + \sum_{m=1}^K \sum_{n=1}^K w_{(m,n)} o^t_{(i_x+m\cdot r -1, i_y+n \cdot r -1)}.
\label{eq:dilated_conv}
\end{equation}
where, $r$ is stride and $K$ is kernel size. If we set $r$ to 1, this equation is the same equation of a normal convolutional layer. If we increase the value of $r$, the receptive fields also increase proportional to $r$.
For example, the area of a local receptive field with $3 \times 3$ kernel is $9$. However, for a dilated convolutional layer with stride $2$, the area of local receptive field increases to $25$, as shown in the right bottom box in Fig. \ref{fig:method:deeplab_architecture}.
We apply these dilated convolutional layers to the last two convolutional layers in the feature extractor.
Fig. \ref{fig:method:deeplab_architecture} illustrates the computational graph of Spiking-DeepLab.
According to Eq. \ref{eq:dilated_conv}, the membrane potential of each neuron is computed by combining the previous membrane potential (\textit{i}.\textit{e}., temporal propagation) and the previous layer (\textit{i}.\textit{e}., spatial propagation).
To preserve the size of the feature map in deep layers, we do not apply average pooling layers in the last feature extraction block (architecture details are shown in Table \ref{table:architecture_detail}).
Also, it is worth mentioning that we use a pixel-wise classifier at the last layer of the network, resulting in a tensor in which the channel size is the number of classes.
We select the class corresponding to the maximum activation at each pixel location.
\subsubsection{Spiking Fully-Convolutional Networks (FCN)}
Another fundamental architecture for segmentation is Fully-Convolutional Networks (FCN). Different from DeepLab where a high resolution feature map is maintained, FCN consists of a downsampling network and an upsampling network.
The downsampling network discovers the semantic representation of the given input images.
This is similar to image recognition networks which consist of multiple convolution and pooling layers.
The upsampling network recovers the resolution of the small feature map by upsampling it into the original image resolution using transposed convolutional layers.
The transposed convolutional layer increases the resolution of the input feature map to a desired output feature map size with learn-able kernel parameters. These parameters are also trained with gradient-based learning.
While increasing the resolution, following the original FCN paper \cite{long2015fully}, we add intermediate features from the downsampling network to features from the upsampling network.
Here, we use convolutional layers to match the number of channels between features in downsampling and upsampling.
Note that all layers consist of LIF neurons, and thus features are processed in a temporal manner.
In addition, like the DeepLab architecture, we use the spike accumulator at the end of the network.
We visualize the architecture (not the computational graph) of FCN in Fig. \ref{fig:method:fcn_architecture}.
The computational graph of Spiking-FCN can be represented with a similar structure as Spiking-DeepLab.
\subsection{Overall Optimization}
For both architectures, the network provides a 2-dimensional probabilistic map at the output layer.
This probabilistic map is based on the stacked spike voltage from the spike accumulator.
For every pixel location $(p, q)$, with the given ground truth label $y_i$, we compute the cross-entropy loss as:
\begin{equation}
{L} = - \frac{1}{N} \sum_{p, q} \sum_{i} y_{i}^{(p,q)} log(\frac{e^{h_i^{(p,q)}}}{\sum_{k=1}^{C}e^{h_k^{(p,q)}}}).
\label{eq:celoss}
\end{equation}
Here, $N$ and $C$ represent a normalization factor and the number of classes, respectively. Also, $h_k$ stands for the membrane potential of the neurons in the last layer. We calculate backward gradients based on this spatial cross-entropy loss.
Backward gradients are computed based on surrogate back-propagation through time (BPTT) \cite{neftci2019surrogate,wu2018spatio} in which gradients are accumulated over all time-steps across all layers.
For a hidden layer, the gradient of output spikes with respect to a membrane potential (\textit{i}.\textit{e}., $\frac{\partial o_i^t}{\partial u_i^t}$) is not differentiable because of the firing behavior of the LIF neuron (Fig. \ref{fig:relatedwork:neuron_dynamic}).
To enable backpropagation through multiple layers, we approximate the backward gradient function to a piece-wise linear function:
\vspace{-1mm}
\begin{equation}
\frac{\partial o_i^t}{\partial u_i^t} = \max \{0, 1- \ | \frac{u_i^t - \theta}{\theta} \ | \},
\label{eq:approximated_gradient}
\vspace{-1mm}
\end{equation}
For the last layer (\textit{i}.\textit{e}., classifier), weight parameters can be updated by conventional gradient backpropagation since we accumulate the spikes and conduct a single forward step in the last layer.
Overall, weight parameters are updated based on the accumulated gradients at each layer.
The accumulated gradients of loss $L$ with respect to weights $W_l$ at layer $l$ can be calculated as:
\begin{equation}
\frac{\partial L}{\partial W_l} =
\begin{cases}
\sum_{t}(\frac{\partial L}{\partial O_l^t}\frac{\partial O_l^t}{\partial U_l^t} + \frac{\partial L}{\partial U_l^{t+1}} \frac{\partial U_l^{t+1}}{\partial U_l^{t}})
\frac{\partial U_l^t}{\partial W_l}, & \text{ $l$: hidden layer} \\
\frac{\partial L}{\partial h}\frac{\partial h}{\partial W_{fc}}.
& \text{$l$: output layer}
\end{cases}
\label{eq:delta_W}
\end{equation}
Here, $O_l$ and $U_l$ stand for the output spike matrix and membrane potential matrix at layer $l$, respectively. We refer to \cite{wu2018spatio} for more spatio-temporal backpropagation details.
Algorithm 1 shows the overall optimization process of spiking segmentation networks with surrogate gradient backpropagation.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=0.5\textwidth]{figures/architecture_fcn.pdf}
\end{center}
\caption{ An illustration of the FCN architecture. }
\label{fig:method:fcn_architecture}
\end{figure}
\begin{algorithm}[t]\small
\caption{Directly training a segmentation network with surrogate gradient backpropagation}
\textbf{Input}: spike time-step ($T$); mini batch ($X$); labels ($Y$)\\
\textbf{Output}: spiking segmentation network
\begin{algorithmic}[1]
%
\For{$i \gets 1$ to $max\_iter$}
\State {fetch a mini batch $X$}
\For{$t \gets 1$ to $T$}
\State{O $\leftarrow$ PoissonGenerator(X) (or DVS input X)}
%
\For{$l \gets 1$ to $L-1$}
\State{$(O^t_{l}, U_l^{t}) \leftarrow (\lambda, U_l^{t-1}, (W_{l}, O^{t}_{l-1}))$}
\Comment{Eq. (\ref{eq:dilated_conv})}
\EndFor
\State{$U_L^{t} \hspace{-1mm} \leftarrow \hspace{-1mm} ( U_L^{t-1}, (W_{l}, O^{t}_{L-1}))$ } \Comment{Final layer}
\EndFor
\State{$L \leftarrow (U_L^T, Y)$}
\State{Do back-propagation and weight update}
\EndFor
\end{algorithmic}
\label{algorithm: overall}
\end{algorithm}
\section{Experiments}
\addtolength{\tabcolsep}{2.5pt}
\begin{table*}[t!]
\centering
\caption{SNN architectures of DeepLab and FCN.}
\resizebox{0.7\textwidth}{!}{
\begin{tabular}{|c|c|c|}
\hline
Module & Spiking-DeepLab & Spiking-FCN \\
\hline
& \makecell{Conv [64] \\
Conv [64] \\
AvgPool [stride 2]}
& \makecell{Conv [64] \\
Conv [64] \\
AvgPool [stride 2]}
\\ \cline{2-3}
\makecell{Downsampling \\ Layers} & \makecell{Conv [128] \\
Conv [128] \\
AvgPool [stride 2]}
& \makecell{Conv [128] \\
Conv [128] \\
AvgPool [stride 2]}
\\ \cline{2-3}
& \makecell{Conv [256] \\
Dilated Conv [256] \\
Dilated Conv [256]}
& \makecell{Conv [256] \\
Conv [256] \\
Conv [256] \\
AvgPool [stride 2]}
\\ \hline
\makecell{Intermediate \\ Layers}
& \makecell{Conv [1024] \\
Conv [1024] \\
Conv [1024, Number of classes]}
& \makecell{Conv [1024] \\
Conv [1024]}
\\ \hline
\makecell{Upsampling \\ Layers}
& \makecell{Bilinear interpolation}
& \makecell{Conv [1024, Number of classes] \\
TransposeConv [Number of classes] \\
SkipConv [256, Number of classes] \\
TransposeConv [Number of classes] \\
SkipConv [128, Number of classes] \\
TransposeConv [Number of classes] \\
SkipConv [64, Number of classes]}
\\ \hline
\end{tabular}}
\label{table:architecture_detail}
\end{table*}
\subsection{Experimental Setup}
\subsubsection{Dataset}
We evaluate our methods on two semantic segmentation datasets: PASCAL VOC2012 and DDD17.
\textbf{PASCAL VOC2012} \cite{everingham2010pascal} contains static images taken with a conventional camera classified with 20 foreground object classes and one background class. We use an augmented dataset \cite{hariharan2011semantic} consisting of a training split of 10,582 images and a validation split of 1,449 images. We re-scale the various original image resolutions to 64$\times$64 pixels.
For SNN models trained and evaluated on PASCAL VOC2012, we use a rate coding technique (discussed in Section \ref{ssec:input_representation}).
Note, we use \textit{PASCAL VOC2012} and \textit{VOC2012} interchangeably in the remainder of the paper.
\textbf{DDD17} \cite{binas2017ddd17} contains 40 different driving sequences of event data captured by a DVS camera. While the dataset provides both grayscale images and event data, it does not provide semantic segmentation labels. Therefore, we use the segmentation labels provided in \cite{alonso2019ev}, which consists of 20 different sequence intervals in 6 of the original DDD17 sequences.
From these sequence intervals, we use a training split consisting of 15,950 frames and a testing split consisting of 3,890 frames with 6 classes. We use a two-channel event representation of accumulated positive and negative events (integrated for a time interval of 50ms) from the DVS dataset proposed in \cite{alonso2019ev}, and we re-scale the image resolution of 346$\times$200 pixels to 64$\times$64 pixels.
In order to train SNNs with DDD17, we do not apply the Poisson coding technique since DDD17 contains temporally-related sequential samples.
Therefore, we construct a batch with multiple sequences of successive video frames to leverage temporal information.
For ANNs, we give each frames to networks, and average all embedding features (after the feature extractor) in temporal axis.
\begin{table}[t]
\addtolength{\tabcolsep}{2.5pt}
\centering
\caption{Mean IoU (\%) of ANNs, Spiking-FCN, and Spiking-DeepLab on PASCAL VOC2012.}
\resizebox{0.48\textwidth}{!}{%
\begin{tabular}{lcc}
\toprule
Method & Time-steps & MIoU(\%) \\
\midrule
DeepLab \cite{chen2017deeplab} & - & {32.3} \\
FCN \cite{long2015fully} & - & 30.9 \\
Spiking-DeepLab (ours) & 20 & {22.3} \\
Spiking-FCN (ours) & 20 & 9.9 \\
\bottomrule
\\
\end{tabular}}
\label{table:exp:performance_comparison_VOC2012}
\end{table}
\begin{table}[t]
\addtolength{\tabcolsep}{2.5pt}
\centering
\caption{Mean IoU (\%) of ANNs, Spiking-FCN, and Spiking-DeepLab on DDD17.}
\resizebox{0.48\textwidth}{!}{%
\begin{tabular}{lcc}
\toprule
Method & Time-steps & MIoU(\%) \\
\midrule
DeepLab \cite{chen2017deeplab} & - & 33.8 \\
FCN \cite{long2015fully} & - & 40.2 \\
Spiking-DeepLab (ours) & 20 & 33.7 \\
Spiking-FCN (ours) & 20 & 34.2 \\
\bottomrule
\\
\end{tabular}}
\label{table:exp:performance_comparison_DDD17}
\end{table}
\subsubsection{Parameters}
In our experiments, we use two different types of SNNs as shown in Table \ref{table:architecture_detail}.
For both networks, the backbone network consists of 7 layers since it is difficult to increase depth with SNNs due to the real vs. surrogate gradient discrepancy.
Note that the main advantage of SNNs is huge energy efficiency compared to ANN, which is desirable for artificial Intelligence systems in edge devices.
A Spiking-DeepLab model includes two dilated convolutional layers in the backbone network followed by a 3-layer classifier.
For Spiking-DeepLab, we use bilinear interpolation to upsample the output prediction to a segmentation map.
Our Spiking-FCN model is similar to our Spiking-DeepLab model but does not use dilated convolutions and uses transposed convolutional layers for upsampling.
\textit{SkipConv} denotes the skip connection with channel reduction (red block in Fig. \ref{fig:method:fcn_architecture}).
We configure the same ANN architectures for comparison. In the ANNs, we use batch normalization \cite{ioffe2015batch} at each convolutional layer. For SNNs, we use the temporal Batch Normalization Through Time (BNTT) technique \cite{kim2020revisiting}.
We train both networks using the same hyperparameters.
We use an Adam optimizer with learning rate 3e-3.
We use a batch size of 16.
Also, we use step-wise learning rate scheduling with a decay factor of 10 at 50\% of the total number of epochs. Here, we set the total number of epochs to 60.
We set 20 time-steps, set a leak factor of 0.99, and use a membrane threshold of 1.0.
\begin{figure*}[t!]
\begin{center}
\includegraphics[width=0.8\textwidth]{figures/deeplab_voc.pdf}\\
{\hspace{4mm}(a) Spiking-DeepLab} \\
\vspace{2mm}
\includegraphics[width=0.8\textwidth]{figures/fcn_voc.pdf} \\
{\hspace{4mm}(b) Spiking-FCN} \\
\end{center}
\caption{ Qualitative results of Spiking-DeepLab and Spiking-FCN on the PASCAL VOC2012 validation set.
We visualize the \textit{image-prediction-groundtruth} triplet for each sample.}
\label{fig:exp:visualization_voc}
\end{figure*}
\begin{figure*}[t!]
\begin{center}
\includegraphics[width=0.8\textwidth]{figures/deeplab_ddd.pdf}\\
{\hspace{4mm}(a) Spiking-DeepLab} \\
\vspace{2mm}
\includegraphics[width=0.8\textwidth]{figures/fcn_ddd.pdf} \\
{\hspace{4mm}(b) Spiking-FCN} \\
\end{center}
\caption{ Qualitative results of Spiking-DeepLab and Spiking-FCN on the DDD17 test set.
We visualize the \textit{image-prediction-groundtruth} triplet for each sample. }
\label{fig:exp:visualization_ddd17}
\end{figure*}
\begin{figure}[t]
\begin{center}
\def0.5{0.5}
\begin{tabular}{@{}c@{\hskip 0.05\linewidth}c@{}c}
\includegraphics[width=0.9\linewidth]{figures/robustness_voc.pdf} \\
{\hspace{4mm}(a) PASCAL VOC2012} \\
\vspace{1mm}
\includegraphics[width=0.9\linewidth]{figures/robustness_ddd.png} \\
{\hspace{4mm}(b) DDD17}\\
\end{tabular}
\end{center}
\vspace{0mm}
\caption{Comparison of changes in Mean IoU with respect to varying standard deviations of Gaussian noise.
We use relative accuracy drop ($\frac{CleanMIoU - NoiseMIoU}{CleanMIoU} \times 100$) where $Clean MIoU$ and $NoiseMIoU$ denote MIoU with clean and noise samples, respectively.
The lower the relative accuracy drop is, the more robust the model is at that noise level.}
\label{fig:noisy_test}
\end{figure}
\subsection{Performance Comparison}
On public datasets, we compare the proposed Spiking-DeepLab and Spiking-FCN methods with the reference ANN methods \cite{he2016deep,long2015fully}.
In Table \ref{table:exp:performance_comparison_VOC2012}, for our BNTT-surrogate gradient-based approaches, the Spiking-FCN architecture shows less performance than the Spiking-DeepLab counterpart on VOC2012. This is due to the FCN architecture being deeper than the DeepLab architecture. Also earlier works have shown that training SNNs with deeper architectures is more challenging than with shallower architectures \cite{panda2020toward,kim2020revisiting}.
In contrast, the Spiking-FCN architecture showed similar performance to the Spiking-DeepLab on DDD17, as shown in Table \ref{table:exp:performance_comparison_DDD17}. This is because of the low level of difficulty for segmentation on the DDD17 dataset which is due to lower number of classes and more consistent, structured data (\textit{i}.\textit{e}., cars are always located in the middle of images, tress are located in the side of images).
Overall, for both datasets, ANN references have a higher performance than SNNs since they are based on a well-established training method.
It is worth mentioning that our purpose is to expand the application of SNNs and show the feasibility of various training methods. In the following sections, we show SNNs have the advantages of robustness and energy-efficiency.
We also visualize the segmentation results on both datasets in Fig. \ref{fig:exp:visualization_voc} and Fig. \ref{fig:exp:visualization_ddd17}.
\subsection{Analysis on Robustness}
In real-world applications such as a self-driving vehicle, input signals are likely to be susceptible to noise.
To investigate the effect of SNNs on robustness for segmentation, we evaluate the relative accuracy drop ($\frac{CleanMIoU - NoiseMIoU}{CleanMIoU} \times 100$) in MIoU across varying levels of noise.
Here, $CleanMIoU$ and $NoiseMIoU$ denote model MIoU when given clean samples and noisy samples, respectively.
We add Gaussian noise (0, $\sigma$) to our inputs to generate noisy inputs.
From Fig. \ref{fig:noisy_test}(a), we observe that our SNNs (\textit{i}.\textit{e}., Spiking-DeepLab and Spiking-FCN) are more robust than ANNs for segmentation on VOC2012.
This is because the Poisson spike generator converts a static pixel value into multiple temporal spikes with random distribution.
From Fig. \ref{fig:noisy_test}(b), we observe that SNNs show a similar robustness with ANNs on DDD17.
Thus, SNNs trained on DDD17 (\textit{i}.\textit{e}., DVS event stream data) are less robust than SNNs trained on VOC2012 (\textit{i}.\textit{e}., static image).
The main two factors contributing to this decrease in robustness are the following:
(i) As aforementioned, the SNNs trained and evaluated on DDD17 do not use the Poisson coding technique which greatly affects robustness.
(ii) The predictions of SNNs on DDD17 are based on multiple sequential frames whereas ANNs are only fed one input for each prediction.
Therefore, more noise is accumulated across multiple frames for SNNs.
However, despite the reduced robustness for segmentation on DVS data, segmentation networks with spiking neurons offer improved robustness for segmentation using traditional cameras.
\subsection{Analysis on Energy-efficiency}
In addition to robustness, SNNs are well known for high energy-efficiency compared to ANNs.
In order to verify this, we compute the approximated energy consumption of ANNs and SNNs. It is worth mentioning that we neglect memory and any peripheral circuit energy and consider the energy needed only for Multiply and Accumulate (MAC) operations.
Since MAC operations are based on layer-wise spiking rates, we measure the layer-wise spiking rates of Spiking-DeepLab (Fig. \ref{fig:exp:spikerate}(a)) and Spiking-FCN (Fig. \ref{fig:exp:spikerate}(b)) on two benchmarks (\textit{i}.\textit{e}., PASCAL VOC2012 and DDD17).
Specifically, we calculate the spike rate $R_s(l)$ of each layer $l$, which can be defined as the total number of spikes at layer $l$ over total time-steps $T$ divided by the number of neurons in layer $l$:
\begin{equation}
R_s(l) = \frac{\textup{$\#$spikes of layer $l$ over all time-steps}}{\textup{ $\#$neurons of layer $l$}}.
\label{eq:spking_rate}
\end{equation}
Interestingly, in Fig. \ref{fig:exp:spikerate}(b), we observe high spike rates in the upsampling layers.
This is because of skip connections in Spiking-FCN which inject a number of spikes during forward propagation.
More precisely, following the previous works \cite{park2020t2fsnn,lee2016training,rathi2020diet}, we compute the energy consumption for SNNs by calculating the total number of floating point operations (FLOPs).
Also, to compare ANNs and SNNs quantitatively, we calculate the energy based on standard CMOS technology \cite{horowitz20141} as shown in Table \ref{table: cmos_tech}.
Conventional ANNs require one FP addition and one FP multiplication to conduct the same MAC operation \cite{sze2017efficient}.
On the other hand, as the computation of SNNs are event-driven with binary spike processing, the MAC operation reduces to just a floating point (FP) addition.
For a layer $l$ in ANNs, we can calculate FLOPs as:
\begin{equation}
FLOPs_{ANN}(l) = \left\{\begin{matrix} & \hspace{-5mm} k^2 \times O^2 \times C_{in} \times C_{out}, \hspace{0.5mm} \textup{if $l =$ Conv},
\\
&\hspace{-3mm}C_{in} \times C_{out}, \hspace{17mm} \textup{if $l =$ Linear}.
\end{matrix}\right.
\end{equation}
Here, $k$ is kernel size, $O$ is output feature map size, and
$C_{in}$ and $C_{out}$ are input and output channels, respectively.
For SNNs, neurons consume energy whenever the neurons are activated. Thus, we multiply the spiking rate $R_s(l)$ (Eq. \ref{eq:spking_rate}) with FLOPs as:
\begin{equation}
FLOPs_{SNN}(l) = {FLOPs}_{ANN}(l) \times R_s(l).
\end{equation}
Finally, the total inference energy of ANNs ($E_{ANN}$) and SNNs ($E_{SNN}$) across all layers can be obtained using the energy values ($E_{MAC}$, $E_{AC}$) from Table \ref{table: cmos_tech}.
\begin{equation}
E_{ANN} = \sum_{l} FLOPs_{ANN}(l) \times E_{MAC}.
\end{equation}
\begin{equation}
E_{SNN} = \sum_{l} FLOPs_{SNN}(l) \times E_{AC}.
\end{equation}
We compare the energy efficiency between ANNs and SNNs on two benchmarks in Table \ref{table: energy_consumption}.
As expected, SNNs show higher energy-efficiency (\textit{i}.\textit{e}., $E_{ANN}/E_{method}$) compared to ANNs. Moreover, we observe that SNNs bring more energy-efficiency on the static VOC2012 dataset since it generates less numbers of spikes across all layers (Fig. \ref{fig:exp:spikerate}).
Note that, ANNs have the same energy consumption regardless of datasets whereas SNNs have different energy consumption depending on the number of spikes.
It is worth mentioning that SNNs can obtain much higher energy-efficiency on neuromorphic hardware platform \cite{akopyan2015truenorth,furber2014spinnaker}.
\begin{figure}[t]
\begin{center}
\def0.5{0.5}
\begin{tabular}{@{}c@{\hskip 0.05\linewidth}c@{}c}
\includegraphics[width=0.9\linewidth]{figures/DeepLab_spikerate.png} \\
{\hspace{4mm}(a) Spiking-DeepLab} \\
\includegraphics[width=0.9\linewidth]{figures/FCN_spikerate.pdf} \\
{\hspace{4mm}(b) Spiking-FCN}\\
\end{tabular}
\end{center}
\caption{Spike rate across all layers in (a) Spiking-DeepLab and (b) Spiking-FCN. We calculate the spike rate for both architectures on PASCAL VOC2012 and DDD17.
}
\label{fig:exp:spikerate}
\end{figure}
\begin{table}[t]
\addtolength{\tabcolsep}{1.5pt}
\centering
\caption{Energy table for 45nm CMOS process.}
\label{table:energy_efficiency}
\resizebox{0.45\textwidth}{!}
{
\begin{tabular}{lc}
\toprule
Operation & Energy(pJ) \\
\midrule
32bit FP MULT $(E_{MULT})$ & 3.7 \\
32bit FP ADD $(E_{ADD})$ & 0.9 \\
32bit FP MAC $(E_{MAC})$ & 4.6 (= $E_{MULT}$ + $E_{ADD}$) \\
32bit FP AC $(E_{AC})$ & 0.9 \\
\bottomrule
\end{tabular}%
}
\label{table: cmos_tech}
\end{table}
\begin{table}[t]
\addtolength{\tabcolsep}{1.5pt}
\centering
\caption{Energy efficiency comparison between segmentation networks.}
\resizebox{0.47\textwidth}{!}{%
\begin{tabular}{lcc}
\toprule
Method & Dataset & $E_{ANN}/E_{method}$ \\
\midrule
DeepLab (ANN) \cite{he2016deep} & - & 1$\times$ (reference) \\
Spiking-DeepLab & VOC2012 & 2.75$\times$ \\
Spiking-DeepLab & DDD17 & 1.15$\times$ \\
\midrule
FCN (ANN) \cite{long2015fully} & - & 1$\times$ (reference) \\
Spiking-FCN & VOC2012 & 2.37$\times$ \\
Spiking-FCN & DDD17 & 1.04$\times$ \\
\bottomrule
\end{tabular}%
}
\label{table: energy_consumption}
\end{table}
\section{Conclusion}
In this paper, for the first time, we perform a comprehensive study on the feasibility of training SNNs on the semantic segmentation task.
We construct two representative spiking segmentation networks, \textit{i}.\textit{e}., Spiking-DeepLab and Spiking-FCN.
We show ANN-SNN conversion requires a large number of time-steps (more than one thousand) and shows inferior performance due to the complexity of the segmentation task.
This is an important observation since most state-of-the-art SNN training techniques for image recognition are based on ANN-SNN conversion.
Instead, we leverage approximated gradient-based training with a temporal batch normalization technique (\textit{i}.\textit{e}., BNTT), resulting in reasonable segmentation results with a small number of time-steps.
We experimentally evaluate the performance, robustness, and estimated energy on two different semantic segmentation scenarios: the static PASCAL VOC2012 dataset and the DDD17 dataset consisting of event stream spikes.
The proposed segmentation architectures can bring a more than $2\times$ energy efficiency on the static image dataset.
As future work, we plan to develop an advanced spike-based learning algorithm to fill the performance gap between ANNs and SNNs on semantic segmentation task.
\section*{Acknowledgement}
The research was funded in part by C-BRIC, one of six centers in JUMP, a Semiconductor Research Corporation (SRC) program sponsored by DARPA, and the National Science Foundation (Grant\#1947826).
\bibliographystyle{stanfordcs}
|
\section{Introduction}
\label{sec:introduction}
\begin{figure*}[t!]
\centering
\includegraphics[scale=0.66]{global2.pdf}
\caption{Examples of four use cases for multilingual ToD systems: A. Use Case E\&E\xspace: A English speaker travels to a country of English. B. Use Case F\&F\xspace: A foreign language speaker travels to a country of the foreign language. C. Use Case F\&E\xspace: A foreign language speaker travels to a country of English. D. Use Case E\&F\xspace: A English speaker travels to a country of a foreign language.
}
\label{fig:illustration}
\end{figure*}
One of the fundamental objectives in pursuit of artificial intelligence is to enable machines with the ability to intelligently communicate with human in natural languages, with one of the widely-heralded applications being the task-oriented dialogue (ToD) systems \citep{Gupta2006TheAS,Bohus2009TheRD}. Recently, ToD systems have been successfully deployed to assist users with accomplishing certain domain-specific tasks such as hotel booking, alarm setting or weather query \citep{eric-etal-2017-key,wu-etal-2019-transferable,lin-etal-2020-mintl,zhang2020recent}, thanks to the joint advent of neural networks and availability of domain-specific data. However, most existing ToD systems are predominately built for English, limiting their service for \textit{all} of the world's citizens. The reason of this limitation lies in the stark lack of high-quality multilingual ToD datasets due to the high expense and challenges of human annotation~\citep{razumovskaia2021crossing}.
One solution to this is annotating conversations in other languages from scratch, e.g., CrossWoZ \cite{zhu2020crosswoz} and BiToD \cite{lin2021bitod}. However, these methods involve expensive human efforts for dialogue collection in the other languages, resulting in a limited language/domain coverage. The other major line of work focused on translating an existing English ToD dataset into target languages by professional human translators \cite{Upadhyay2018AlmostZC,schuster2018cross,van-der-goot-etal-2021-masked,li-etal-2021-mtop}. Despite the increasing language coverage, these methods simply translated English named entities (e.g., location, restaurant name) into the target languages, while ignored the fact that these entities barely exist in countries speaking these languages. This hinders a trained ToD system from supporting the real use cases where a user looks for local entities in a target-language country. For example in Figure~\ref{fig:illustration}, a user may look for the British Museum when traveling to London (A.), while look for the Oriental Pearl Tower when traveling to Shanghai (B.).
In addition, prior studies~\cite{cheng1989code,kim2006reasons} have shown that code-switching phenomena frequently occurs in a dialogue when a speaker cannot express an entity immediately and has to alternate between two languages to convey information more accurately. Such phenomena could be ubiquitous during the cross-lingual and cross-country task-oriented conversations. One of the reasons for code-switching is that there are no exact translations for many local entities in the other languages. Even though we have the translations, they are rarely used by local people. For example in Figure~\ref{fig:illustration} (C.), after obtaining the recommendation from a ToD system, a Chinese speaker traveling to London would rather use the English entity ``British Museum'' than its Chinese translation to search online or ask local people. To verify this code-switching phenomena, we have also conducted a case study (\secref{sec:validation}) which shows that searching the information about translated entities online yields a much higher failure rate than searching them in their original languages.
Motivated by these observations, we define \textit{three unexplored use cases}\footnote{See comparisons of these use cases in Appendix~\ref{appendix:comparison}} of multilingual ToD where a foreign-language speaker uses ToD in the foreign-language country (\textbf{F\&F\xspace}) or an English country (\textbf{F\&E\xspace}), and an English speaker uses ToD in a foreign-language country (\textbf{E\&F\xspace}). These use cases are different from the traditional \textbf{E\&E\xspace} use case where an English speaker uses ToD in an English-speaking country.
To bridge the aforementioned gap between existing data curation methods and the real use cases, we propose a novel data curation method that \textit{globalizes} an existing multi-domain ToD dataset beyond English for the three unexplored use cases. Specifically, building on top of MultiWoZ~\cite{budzianowski-etal-2018-multiwoz} --- an English ToD dataset for dialogue state tracking (DST), we create \mbox{GlobalWoZ}\xspace, a new multilingual ToD dataset in three new target-languages via machine translation and crawled ontologies in the target-language countries.
Our method only requires minor human efforts to post-edit a few hundred machine-translated dialogue templates in the target languages for evaluation.
Besides, as cross-lingual transfer via pre-trained multilingual models~\cite{devlin-etal-2019-bert,conneau-etal-2020-unsupervised, liu2020multilingual, xue-etal-2021-mt5} has proven effective in many cross-lingual tasks, we further investigate another question: \textit{How do these multilingual models trained on the English ToD dataset transfer knowledge to our globalized dataset?} To answer this question, we prepare a few baselines by evaluating popular ToD systems on our created test datasets in a \textit{zero-shot} cross-lingual transfer setting as well as a \textit{few-shot} setting.
Our contributions include the following:
\begin{itemize} [leftmargin=10pt]\itemsep-0.2em
\item To the best of our knowledge, we provide the first step towards analyzing three unexplored use cases for multilingual ToD systems.
\item We propose a cost-effective method that creates a new multilingual ToD dataset from an existing English dataset. Our dataset consists of high-quality test sets which are first translated by machines and then post-edited by professional translators in three target languages (Chinese, Spanish and Indonesian). We also leverage machine translation to extend the language coverage of test data to another 17 target languages.
\item Our experiments show that current multilingual systems and translate-train methods fail in zero-shot cross-lingual transfer on the dialogue state tracking task. To tackle this problem, we propose several data augmentation methods to train strong baseline models in both zero-shot and few-shot cross-lingual transfer settings.
\end{itemize}
\section{Related Work}
\label{sec:related_work}
Over the last few years, the success of ToD systems is largely driven by the joint advent of neural network models \citep{eric-etal-2017-key,wu-etal-2019-transferable,lin-etal-2020-mintl} and collections of large-scale annotation corpora. These corpora cover a wide range of topics from a single domain (e.g., ATIS \cite{hemphill1990atis}, DSTC 2 \cite{henderson2014second}, Frames \cite{el-asri-etal-2017-frames}, KVRET \cite{eric-etal-2017-key}, WoZ 2.0 \cite{wen-etal-2017-network}, M2M \cite{schatzmann2007agenda}) to multiple domains (e.g., MultiWoZ \cite{budzianowski-etal-2018-multiwoz}, SGD \cite{rastogi2020towards}). Most notably among these collections, MultiWoZ is a large-scale multi-domain dataset that focuses on transitions between different domains or scenarios in real conversations \citep{budzianowski-etal-2018-multiwoz}. Due to the high cost of collecting task-oriented dialogues, only a few monolingual or bilingual non-English ToD datasets are available \cite{zhu2020crosswoz, quan-etal-2020-risawoz, lin2021bitod}. While there is an increasing interest in data curation for multilingual ToD systems, a vast majority of existing multilingual ToD datasets do not consider the real use cases when using a ToD system to search for local entities in a country. We fill this gap in this paper to provide the first analysis on three previously unexplored use cases.
\section{Data Curation Methodology}
\label{sec:methodology}
In order to globalize an existing English ToD dataset for the three aforementioned use cases, we propose an approach consisting of four steps as shown in Figure~\ref{fig:flow}: (1) we first extract dialogue templates from the English ToD dataset by replacing English-specific entities with a set of general-purpose placeholders (\secref{sec:template}); (2) we then translate the templates to a target language for both training and test data, with one key distinction that we only post-edit the test data by professional translators to ensure the data quality for evaluation (\secref{sec:translate}); (3) next, we collect ontologies \cite{kiefer2021vonda} containing the definitions of dialogue acts, local entities and their attributes in the target-language countries (\secref{sec:ontology}); (4) finally, we tailor the translated templates by automatically substituting the placeholders with entities in the extracted ontologies to construct data for the three use cases (\secref{sec:use-case}).
\begin{figure*}[t!]
\centering
\includegraphics[scale=0.48]{flow.pdf}
\caption{Illustration of our proposed pipeline: 1. Automatic Template Creation 2. Labeled Sequence Translation 3. Localized Ontologies Collection 4. Automatic Template Filling
}
\label{fig:flow}
\end{figure*}
\subsection{Automatic Template Creation} \label{sec:template}
We start with MultiWoZ 2.2 \cite{zang-etal-2020-multiwoz} -- a high-quality multi-domain English ToD dataset with more accurate human annotations compared to its predecessors MultiWoZ 2.0 \cite{budzianowski-etal-2018-multiwoz} and MultiWoz 2.1 \cite{eric-EtAl:2020:LREC}. For the sake of reducing human efforts for collecting ToD context in the target languages, we re-use the ToD context written by human in MultiWoZ as the dialogue templates. Specifically as shown in Figure~\ref{fig:flow}, we replace the English entities in MultiWoz by a set of general-purpose placeholders such as \texttt{[attraction-name0]} and \texttt{[attraction-postcode1]}, where each placeholder contains the entity's domain, attribute and ID. To do so, we first build a dictionary with entity-placeholder pairs by parsing the annotations of all dialogues. For example, from a dialogue text ---\textit{``I recommend Whale of a time and the post code is cb238el.''}, we obtain two entity-placeholder pairs from its human annotations, i.e., (\textit{Whale of a time}, \texttt{[attraction-name0]}) and (\textit{cb238el}, \texttt{[attraction-postcode1]}). Next, we identify entities in the dialogue by their word index from the human annotations, replace them with their placeholders in the dictionary, and finally obtain dialogue templates with placeholders. Notably, we skip the entities with their attributes of \texttt{[choice]} and \texttt{[ref]} that represent the number of choices and booking reference number, as these attributes could be used globally.
\subsection{Labeled Sequence Translation}\label{sec:translate}
Following \citet{liu-etal-2021-mulda} that translates sentences with placeholders, we use a machine translation system\footnote{We use Google Translate (\url{https://cloud.google.com/translate}), an off-the-shelf MT system.} to translate dialogue templates with our designed placeholders. As we observe, a placeholder containing an entity domain, attribute and ID (e.g., \texttt{attraction-name0}) is useful to provide contextually meaningful information to the translation system, thus usually resulting in a high-quality translation with the placeholder unchanged \footnote{Appendix~\ref{appendix:sequence_translation} has an example of label sequence translation.}. This also enables us to easily locate the placeholders in the translation output and replace them with new entities in the target language.
To build a high-quality test set for evaluation, we further hire professional translators to post-edit a few hundred machine-translated templates, which produces natural and coherent sentences in the target languages.\footnote{Appendix~\ref{appendix:bleu} shows the bleu scores between MT test data and MTPE test data.} With the goal of selecting representative test templates for post-editing, we first calculate the frequency of all the 4-gram combinations in the MultiWoZ data, and then score each dialogue in the test set by the sum of the frequency of all the 4-gram combinations in the dialogue divided by the dialogue's word length. We use this scoring function to estimate the representiveness of a dialogue in the original dataset. Finally, we select the top 500 high-scoring dialogues in the test set for
post-editing.\footnote{Appendix~\ref{appendix:test_dist} shows the English test data distribution.}
We also use the same procedure to create a small high-quality training set for few-shot cross-lingual transfer setting.
\subsection{Collection of Local Ontology} \label{sec:ontology}
Meanwhile, we crawl the attribute information of local entities in three cities from public websites (e.g., tripadvisor.com, booking.com) to create three ontologies for the three corresponding target languages respectively. As shown in Table~\ref{tb:selected_languages} in Appendix~\ref{appendix:sel_lang}, we select Barcelona for Spanish (an Indo-European language), Shanghai for Mandarin (a Sino-Tibetan language) and Jakarta for Indonesian (an Austronesian language), which cover a set of typologically different language families.
Given a translated dialogue template, we can easily sample a random set of entities for a domain of interest from a crawled ontology and assign the entities to the template's placeholders to obtain a new dialogue in the target language. Repeating this procedure on each dialogue template, we can easily build a high-quality labeled dataset in the target language. Table~\ref{tb:stats_ontology_17} in Appendix~\ref{appendix:stat_ontology} shows the statistics of our collected entities in the target languages compared with the English data. The number of our collected entities are either larger than or equal to those in the English data except for the ``train'' domain; we collected the information about only 100 ``trains'' for each languages due to the complexity in collecting relevant information.
\subsection{Template Filling for Three Use Cases} \label{sec:use-case}
After the above steps, we assign entities in a target language to the translated templates in the same target language for the F\&F\xspace case, while assigning target-language entities to the English (source-language) templates for the F\&E\xspace case. As for the E\&F\xspace case, we keep the original English context by skipping the translation step and replace the placeholders with local entities in the target language (see Figure \ref{fig:flow} for examples).
To sum up, our proposed method has three key properties: (1) our method is \textit{cost-effective} as we only require a limited amount of post-editing efforts for a test set when compared to the expensive crowd-sourced efforts from the other studies; (2) we can easily sample entities from an ontology to create \textit{large-scale machine-translated data} as a way of data augmentation for training;
(3) our method is \textit{flexible} to update entities in a ToD system whenever an update of ontology is available, e.g., extension of new entities. We refer the readers to Table~\ref{tb:stats_globalwoz} for the data statistics of \mbox{GlobalWoZ}\xspace and Figure~\ref{fig:example} for dialogue examples in the appendix.
\section{Task \& Settings}
\label{sec:tasks}
\subsection{Dialogue State Tracking}
Our experiments focus on the dialogue state tracking (DST), one of the fundamental components in a ToD system that predicts the goals of a user query in multi-turn conversations. We follow the setup in MultiWoZ~\cite{budzianowski-etal-2018-multiwoz} to evaluate ToD systems for DST by the joint goal accuracy which measures the percentage of correctly predicting all goals in a multi-turn conversation.
\subsection{Experimental Settings}
\noindent\textbf{Zero-Shot Cross-lingual Transfer:} Unlike prior studies that annotate a full set of high-quality training data for a target language, we investigate the \textit{zero-shot cross-lingual transfer} setting where we have access to only a high-quality human-annotated English ToD data (referred to as gold standard data hereafter). In addition, we assume that we have access to a machine translation system that translates from English to the target language. We investigate this setting to evaluate how a multilingual ToD system transfers knowledge from a high-resource source language to a low-resource target language.
\noindent\textbf{Few-Shot Cross-lingual Transfer:} We also investigate few-shot cross-lingual transfer, a more practical setting where we are given a small budget to annotate ToD data for training. Specifically, we include a small set (100 dialogues) of high-quality training data post-edited by professional translators (\secref{sec:translate}) in a target language, and evaluate the efficiency of a multilingual ToD on learning from a few target-language training examples.
\section{Future Work}
\label{sec:analysis}
\subsection{Scale up to 20 languages}
The coverage of languages of the existing datasets are still limited to less than 10 languages due to the high cost of data annotation. With our proposed cost-effective data curation method, it is possible to extend the dataset to cover more languages. Thus, we conduct the experiments by evaluating the model performance of our proposed methods (\secref{sec:methodology}) on two sets of test data: (1) \textbf{MT} test data created by machine translation only~(\secref{sec:translate}); (2) \textbf{MTPE} test data first translated by machines and post-edicted later by professional translators. As shown in Table~\ref{tb:f2f_mt_result} and Table~\ref{tb:f2e_mt_result}, the overall reported results on MT test data are higher than those reported on MTPE test data, which is expected because the distribution of the MT test data is more similar to the MT training data. Although there are some differences on individual languages, the conclusions derived from the evaluations on the MT test data remain the same as those derived from the evaluation on the MTPE test data. We also calculate the Spearman rank correlation coefficient and Pearson correlation coefficient between the results on the average results reported on MTPE test data and MT test data in Table~\ref{tb:spearman}, which show a statistically high correlation between the system performance on the MT test data and MTPE test data. Therefore, we show that the MT test data can be used as a proxy to estimate the model performance on the real test data for more languages. We select 20 languages that are supported by Google Translate, Trip Advisor and Booking.com at the same time, as stated in Table~\ref{tb:selected_languages} in Appendix. We leave the extension of the language coverage to all these languages in our future work.
\begin{table}[ht!]
\centering
\scalebox{0.75}{
\begin{tabular}{lcccc}
\toprule
\textbf{Methods} & \textbf{zh} & \textbf{es} & \textbf{id} &\textbf{avg}\\
\midrule
Zero-Shot (E\&E)\xspace & 1.19 & 1.40 & 1.28 & 1.29 \\
Translate-Train\xspace & 2.50 & 2.81 & 5.81 & 3.71 \\
SUC\xspace & 37.79 & 26.95 & 42.59 & 35.78 \\
BBUC\xspace & 38.62 & 27.34 & 42.96 & 36.31 \\
MBUC\xspace & \textbf{39.11} & \textbf{29.17} & 45.39 & 37.89 \\
MMUC\xspace & 38.16 & 29.09 & \textbf{46.51} & \textbf{37.92} \\
\bottomrule
\end{tabular}
}
\caption{Zero-Shot Cross Lingual Joint Accuracy on DST on Target Languages On Use Case F2F - MT test data}
\label{tb:f2f_mt_result}
\end{table}
\begin{table}[ht!]
\centering
\scalebox{0.75}{
\begin{tabular}{lcccc}
\toprule
\textbf{Methods} & \textbf{zh} & \textbf{es} & \textbf{id} &\textbf{avg}\\
\midrule
Zero-Shot (E\&E)\xspace & 7.61 & 11.67 & 9.64 & 9.64 \\
Translate-Train\xspace & 2.25 & 5.25 & 5.03 & 4.17 \\
SUC\xspace & 57.10 & 55.70 & 55.64 & 56.15 \\
BBUC\xspace & 59.05 & \textbf{57.68} & 56.80 & 57.84 \\
MBUC\xspace & \textbf{60.48} & 57.04 & \textbf{58.23} & \textbf{58.76} \\
MMUC\xspace. & 59.21 & 57.51 & 56.41 & 57.71 \\
\bottomrule
\end{tabular}
}
\caption{Zero-Shot and Few-Shot Cross Lingual Joint Accuracy on DST on Target Languages on Use Case F2E - MT test data}
\label{tb:f2e_mt_result}
\end{table}
\begin{table}[ht!]
\centering
\scalebox{0.75}{
\begin{tabular}{lcc}
\toprule
\textbf{Methods} & \textbf{Spearman Correlation} & \textbf{Pearson Correlation} \\
\midrule
F2F & 1.00 & 1.00 \\
F2E & 1.00 & 1.00 \\
\bottomrule
\end{tabular}
}
\caption{Spearman Rank Correlation Coefficient between the results on MTPE test data and MT test data}
\label{tb:spearman}
\end{table}
\subsection{Few-Shot Cross-lingual Transfer}
Few-shot cross-lingual transfer has attracted lots of research interests recently from both academia and industry. In the few-shot cross-lingual transfer setting, we aim to include a small set of high-quality training data post-edited by professional translators (\secref{sec:translate}) in a target language. This setting is useful to evaluate the efficiency of a multilingual ToD on learning from a small set of target-language data.
\section{Proposed Baselines}
\label{sec:baselines}
We prepare a base model for \mbox{GlobalWoZ}\xspace in the zero-shot and few-shot cross-lingual transfer settings. We select Transformer-DST \citep{zeng2020jointly} as our base model as it is one of the state-of-the-art models on both MultiWoZ 2.0 and MultiWoZ 2.1\footnote{According to the leaderboards of Multi-domain Dialogue State Tracking on MultiWoZ 2.0 and MultiWoZ 2.1 on paperwithcode.com as of 11/15/2021.}. In our paper, we replace its BERT encoder with an mBERT encoder \citep{devlin-etal-2019-bert} for our base model and propose a series of training methods for \mbox{GlobalWoZ}\xspace. As detailed below, we propose several data augmentation baselines that create different training and validation data for training a base model.
Note that all the proposed baselines are model agnostic and the base model can be easily substituted with other popular models \cite{heck-etal-2020-trippy,lin-etal-2020-mintl}. For each baseline, we first train a base model on its training data for 20 epochs and use its validation set to select the best model during training. Finally we evaluate the best model of each baseline on the same test set from \mbox{GlobalWoZ}\xspace. We will release \mbox{GlobalWoZ}\xspace and our pre-trained models to encourage faster adaptation to future research.
We refer the readers to Table~\ref{tb:Methods_data} and Table~\ref{tb:Methods_data_access} in Appendix~\ref{appendix:summary_baselines}
while reading the subsequent methods for a better understanding.
\subsection{Pure Zero-Shot (E\&E\xspace)}
We train a base model on the gold standard English data (E\&E\xspace) and directly apply the learned model to the test data of the three use cases in \mbox{GlobalWoZ}\xspace. With this method, we simulate the condition of having labeled data only in the source language for training, and evaluate how the model transfers knowledge from English to the three use cases.
We use \textbf{Zero-Shot (E\&E)\xspace} to denote this method.
\subsection{Translate-Train}
We use our data curation method (\secref{sec:methodology}) to translate the templates by an MT system but replace the placeholders in the translated templates with machine-translated entities to create a set of pseudo-labeled training data. Next, we train a base model on the translated training data without local entities, and evaluate the model on the three use cases.
We denote this method as \textbf{Translate-Train\xspace}.
\subsection{Single-Use-Case Training}
By skipping the human post-editing step in our data curation method (\secref{sec:methodology}), we leverage a machine translation system to automatically create a large set of pseudo-labeled training data with local entities for the three use cases. In the F\&F\xspace case, we translate the English templates by the MT system and replace the placeholders in the translated templates with foreign-language entities to create a training dataset. In the F\&E\xspace case, we replace the placeholders in the translated templates with the original English entities to create a code-switched training dataset. In the E\&F\xspace case, we use the original English templates and replace the placeholders in the English templates with foreign-language entities to create a code-switch training dataset. With this data augmentation method, we can train a base model on each pseudo-labeled training dataset created for each use case.
We denote this method as \textbf{SUC\xspace} (Single-Use-Case).
\subsection{Bi-/Multi-lingual Bi-Use-Case Training}
We investigate the performance of combining the existing English data and the pseudo-labeled training data created for one of the three use cases (i.e., F\&F\xspace, F\&E\xspace, E\&F\xspace), one at a time, to do bi-use-case training. In the bilingual training, we only combine the gold English data (E\&E\xspace) with the pseudo-labeled training data in one target language in one use case for joint training. We denote this method as \textbf{BBUC\xspace} (Bilingual Bi-Use-Case). In the multilingual training, we combine gold English data (E\&E\xspace) and pseudo-labeled training data in all languages in one use case for joint training. We denote this method as \textbf{MBUC\xspace} (Multilingual Bi-Use-Case).
\subsection{Multilingual Multi-Use-Case Training}
We also propose to combine the existing English data (E\&E\xspace) and all the pseudo-labeled training data in all target languages for all the use cases (F\&F\xspace, F\&E\xspace, E\&F\xspace). We then train a single model on this combined multilingual training dataset and evaluate the model on test data in all target languages for all three use cases .
We denote this method as \textbf{MMUC\xspace} (Multilingual Multi-Use-Case).
\section{Conclusions}
\label{sec:conclusion}
In this paper, we provide an analysis on three unexplored use cases for multilingual task-oriented dialogue systems. We propose a new data curation method that leverages a machine translation system and local entities in target languages to create a new multilingual TOD dataset -- \mbox{GlobalWoZ}\xspace. We propose a series of popular baseline methods and conduct extensive experiments on \mbox{GlobalWoZ}\xspace to encourage future research for multilingual ToD systems.
\section*{Acknowledgements}
\label{sec:acknowledgements}
This research is partly supported by the Alibaba-NTU Singapore Joint Research Institute, Nanyang Technological University. We would like to thank the help from our Alibaba colleagues, Ruidan He, Haiyun Peng and Zifan Xu in this work as well.
\section{Experiment Results}
\label{sec:experiments}
In this section, we show the results of all methods in the zero-shot~(\secref{sec:exp_zero_shot}) and few-shot~(\secref{sec:exp_few_shot}) settings.
\subsection{Zero-shot Cross-lingual Transfer} \label{sec:exp_zero_shot}
\subsubsection{Use Case F\&F\xspace, F\&E\xspace and E\&F\xspace}
Table~\ref{tb:Use_case_result} reports the joint goal accuracy of all proposed methods on the three different sets of test data in the F\&F\xspace, F\&E\xspace, and E\&F\xspace use cases\footnote{Appendix~\ref{appendix:EE} reports the results in the E\&E\xspace use case.}.
Both Zero-Shot (E\&E)\xspace and Translate-Train\xspace struggle, achieving average accuracy of less than 10 in all use cases. Despite its poor performance, Zero-Shot (E\&E)\xspace works much better in F\&E\xspace than F\&F\xspace, while its results in F\&F\xspace and E\&F\xspace are comparable, indicating that a zero-shot model trained in E\&E\xspace can transfer knowledge about local English entities more effectively than knowledge about English context in downstream use cases.
Besides, we also find that Zero-Shot (E\&E)\xspace performs better on the Spanish or Indonesian context than the Chinese context in F\&E\xspace. One possible reason is that English is closer to the other Latin-script languages (Spanish and Indonesian) than Chinese.
Our proposed data augmentation methods (SUC\xspace, BBUC\xspace, MBUC\xspace) perform much better than non-adapted methods (Zero-Shot (E\&E)\xspace and Translate-Train\xspace) that do not leverage any local entities for training. In particular, it is worth noting that even though Translate-Train\xspace and SUC\xspace both do training on foreign-language entities in F\&F\xspace and E\&F\xspace, there is a huge gap between these two methods, since Translate-Train\xspace has only access to the machine-translated entities rather than the real local entities used by SUC\xspace. This huge performance gaps not only show that Translate-Train\xspace is not an effective method in practical use cases but also prove that having access to local entities is a key to building a multilingual ToD system for practical usage.
Comparing our data augmentation methods SUC\xspace and BBUC\xspace, we find that the base model can benefit from training on additional English data (E\&E), especially yielding a clear improvement of up to 5.58 average accuracy points in F\&E\xspace. Moreover, when we increase the number of languages in the bi-use-case data augmentations (i.e., MBUC\xspace), we observe an improvement of around 1 average accuracy points in all three use cases w.r.t. BBUC\xspace. These observations encourage a potential future direction that explores better data augmentation methods to create high-quality pseudo-training data.
\begin{table}[t!]
\centering
\resizebox{\columnwidth}{!}{%
\begin{tabular}{llcccc}
\toprule
\textbf{Case} & \textbf{Methods} & \textbf{zh} & \textbf{es} & \textbf{id} &\textbf{avg}\\
\midrule
\multirow{5}{*}{F\&F\xspace}&Zero-Shot (E\&E)\xspace & 1.22 & 1.38 & 1.26 & 1.28 \\
& Translate-Train\xspace & 2.61 & 2.59 & 5.74 & 3.65 \\
& SUC\xspace (F\&F\xspace) & 36.97 & 24.66 & 25.26 & 28.96 \\
& BBUC\xspace (E\&E\xspace+ F\&F\xspace) & 37.32 & 25.52 & 26.39 & 29.74 \\
& MBUC\xspace (E\&E\xspace+ F\&F\xspace) & \textbf{38.01} & \textbf{26.03} & \textbf{28.22} & \textbf{30.76} \\
\midrule
\multirow{5}{*}{F\&E\xspace}&Zero-Shot (E\&E)\xspace & 6.92 & 11.34 & 9.09 & 9.12 \\
&Translate-Train\xspace & 2.28 & 4.97 & 4.67 & 3.97 \\
&SUC\xspace (F\&E\xspace) & 56.28 & 41.94 & 47.93 & 48.71 \\
&BBUC\xspace (E\&E\xspace+ F\&E\xspace) & 59.87 & 48.20 & 54.79 & 54.29 \\
&MBUC\xspace (E\&E\xspace+ F\&E\xspace) & \textbf{60.37} & \textbf{53.56} & \textbf{54.93} & \textbf{56.28} \\
\midrule
\multirow{5}{*}{E\&F\xspace}&Zero-Shot (E\&E)\xspace & 1.69 & 1.81 & 1.82 & 1.77 \\
& Translate-Train\xspace & 1.39 & 1.76 & 1.86 & 1.67 \\
& SUC\xspace (E\&F\xspace) & 38.56 & 28.00 & 43.82 & 36.79 \\
& BBUC\xspace (E\&E\xspace+ E\&F\xspace) & 39.87 & 27.29 & 45.48 & 37.54 \\
& MBUC\xspace (E\&E\xspace+ E\&F\xspace) & \textbf{40.20} & \textbf{29.22} & \textbf{47.06} & \textbf{38.83} \\
\bottomrule
\end{tabular}
}
\caption{Zero-shot cross-lingual accuracy on DST over three target languages in three use cases.}
\label{tb:Use_case_result}
\end{table}
\subsubsection{One Model for All}
Notice that we can train a single model by MMUC\xspace for all use cases rather than training separate models, one for each use case. In Figure~\ref{fig:heatmap}, we compare MMUC\xspace and MBUC\xspace (rows) on the test data in the four use cases (columns). Although MMUC\xspace may not achieve the best results in each use case, it achieves the best average result over the four use cases, indicating the potential of using one model to simultaneously handle all the four use cases.
\begin{figure}[t!]
\centering
\includegraphics[scale=0.5]{heatmap.pdf}
\caption{Performance of MMUC\xspace vs MBUC\xspace on the test data of the four use cases, F\&F\xspace, F\&E\xspace, E\&F\xspace and E\&E\xspace.
}
\label{fig:heatmap}
\end{figure}
\subsection{Few-shot Cross-lingual Transfer}\label{sec:exp_few_shot}
In few-shot experiments, we use the same scoring function based on frequency of all 4-gram combinations (\secref{sec:translate}) to select 100 additional dialogues from train set for human-post editing, and create high-quality training data for each of the three use cases. To avoid overfitting on this small few-shot dataset, we
combine the few-shot data with the existing English data for training a base model (Few-Shot+Zero-Shot (E\&E)\xspace). Next, we also investigate a model trained with additional synthetic data created by our proposed SUC\xspace.
In Figure~\ref{fig:few_shot_result}, we find that our proposed SUC\xspace without additional few-shot data has already outperformed the model trained with few-shot data and English data (Few-shot + Zero-Shot (E\&E)\xspace), indicating that the model benefit more from a large amount of pseudo-labeled data than a small set of human-labeled data.
If we combine the data created by SUC\xspace with the few-shot data or with both few-shot and English data to train the model, we observe improvements over SUC\xspace, especially with a clear gain of 8.06 accuracy points in F\&E\xspace. We refer the readers to Table~\ref{tb:few_shot_breakdown} in the appendix for detailed scores in all target languages.
\begin{figure}[t!]
\centering
\includegraphics[width=\columnwidth]{fewshot.pdf}
\caption{Few-shot cross-lingual average joint accuracy on DST over three target languages in three use cases.
}
\label{fig:few_shot_result}
\end{figure}
\section{Discussion}
\label{sec:analysis}
\subsection{Motivation for Code-Switched Use Cases}
\label{sec:validation}
One key research question is to validate whether code-switched use cases with local entities (i.e., F\&E\xspace, E\&F\xspace) are practically more useful for information seeking.
To answer this question, we compare the failure rate of using local entities and machine-translated entities in information search, which is a proxy to the efficiency of using these two types of entities in conversations. We first randomly select 100 entities (33 attractions, 33 hotels and 34 restaurants) of Cambridge, Shanghai, Barcelona and Jakarta. We translate the English entities into Mandarin, Spanish and Indonesian and the foreign-language entities into English via Google Translate. We then manually search the translated entities on Google to check whether we can find the right information of the original entities. Notice that the failure of the above verification partially come from the translation error made by Google Translate, or the search failure due to the fact that this entity does not have a bilingual version at all. In Table~\ref{tab:fail_rate}, we observe a high failure rate of around 60\% for almost all translated directions (except Zh$\rightarrow$En) due to translation and search failures, significantly exceeding the low failure rate of searching original entities online. Besides, even if we can find the right information of the translated entities, local people may not recognize or use the translated entities for communication, thus this results in inefficient communication with local people.
\begin{table}[t!]
\setlength{\tabcolsep}{1.5pt}
\centering
\resizebox{\columnwidth}{!}{%
\begin{tabular}{cccccccc}
\toprule
Translate & Search &En$\rightarrow$Zh & En$\rightarrow$Es & En$\rightarrow$Id & Zh$\rightarrow$En & Es$\rightarrow$En & Id$\rightarrow$En \\
\midrule
\CheckmarkBold & \CheckmarkBold & 35 & 42 & 36 & 62 & 30 & 31 \\
\CheckmarkBold & \XSolidBrush & 61 & 34 & 51 & 18 & 18 & 15 \\
\XSolidBrush & \CheckmarkBold & 0 & 24 & 13 & 11 & 50 & 54 \\
\XSolidBrush & \XSolidBrush & 4 & 0 & 0 & 8 & 2 & 0 \\
\midrule
\multicolumn{2}{c}{Failure Case (MTed Entities)} & 65 & 58 & 64 & 37 & 70 & 69 \\
\multicolumn{2}{c}{Failure Rate (MTed Entities)} & 65\% & 58\% & 64\% & 37\% & 70\% & 69\% \\
\multicolumn{2}{c}{Failure Rate (Original Entities)} & 3\% & 3\% & 3\% & 0\% & 1\% & 0\% \\
\bottomrule
\end{tabular}%
}
\caption{The search and translation results of 100 translated entities on Google. En$\rightarrow$Zh refers to the translation of English entities to Mandarin and Zh$\rightarrow$En refers to the translation of Mandarin entities to English.}
\label{tab:fail_rate}
\end{table}
\begin{figure}[t!]
\centering
\includegraphics[scale=0.5]{translate_test.pdf}
\caption{ Joint accuracy of Translate-Train\xspace for DST on the F\&F\xspace Test vs Translate-Test data.}
\label{fig:translate_test}
\end{figure}
\subsection{Overestimate of Translate-Train}
\label{sec:over-TT}
In previous translation-based work, a multilingual ToD system is usually built based on the translation of English training data (Translate-Train\xspace), and is evaluated on translated test data without any local entities (Translate-Test). To verify whether this procedure is reliable to build a multilingual ToD system, we also create a test dataset with translated entities instead of local entities in the target languages. As shown in Figure~\ref{fig:translate_test}, we find the Translate-Train\xspace model performs well on the test data with translated entities, but performs badly on the test data with real local entities. To the best of our knowledge, we provide the first analysis to identify this performance gap between the translated test data and data with real local entities in a more realistic use case \footnote{Please refer to Appendix~\ref{appendix:example_TT} for concrete examples where Translate-Train\xspace fails in predicting real local entities.}. Our work sheds light on the development of a globalized multilingual ToD system in practical use cases. We can tackle the challenge of localization issues by exploring new data augmentation method. Alternatively we can also explore new methods from the model level by building modular network to update the entities or perform transfer learning to adapt to new case without retraining.
\subsection{Local Context vs. Local Entities}
We compare the impact of training a model on data with either local contexts or local entities when the model is evaluated on monolingual test data in F\&F\xspace and E\&E\xspace. Specifically, when the train set has access to local context only, all the entities in the train set are replaced by entities in non-target languages. Similarly, when the train set has access to local entities only, the contexts in the train set are replaced by context in the non-target languages. Table~\ref{tb:context_vs_entities} shows that both local contexts and local entities are essential to building ToD systems in the target language. A further analysis in Table~\ref{tb:context_vs_entities_language} and Table~\ref{tb:script_type} in the appendix shows that training with local entities is more important if the entities and contexts are written in the same type of language script (e.g. Latin script).
\begin{table}[h!]
\centering
\resizebox{\columnwidth}{!}{%
\begin{tabular}{lccccc}
\toprule
\textbf{Train Set} & \textbf{E\&E\xspace(en)} & \textbf{F\&F\xspace(zh)} & \textbf{F\&F\xspace(es)} & \textbf{F\&F\xspace(id)} &\textbf{avg}\\
\midrule
Local Context Only & 5.46 & 1.77 & 2.37 & 2.40 & 3.20 \\
Local Entities Only & 6.39 & 0.36 & 2.41 & 2.75 & 3.05 \\
Local Context \& Entities & \textbf{52.78} & \textbf{36.97} & \textbf{24.66} &\textbf{25.26} & \textbf{38.13} \\
\bottomrule
\end{tabular}
}
\caption{Comparison of training with local context or/and local entities on the joint accuracy for DST in E\&E\xspace (en) and F\&F\xspace (zh, es, id).}
\label{tb:context_vs_entities}
\end{table}
\subsection{Scaling up to 20 Languages}
With our proposed data curation method, it is possible to extend the dataset to cover more languages without spending extra costs if we skip the human post-editing step. Before doing so, one key question is whether the evaluation on the translated data without human post-editing is reliable as a proxy of the model performance. Thus, we conduct the experiments by evaluating the model performance of all baselines (\secref{sec:baselines}) on two sets of test data built with local entities: (1) \textbf{MT} test data where translated template is created by machine translation only~(\secref{sec:translate}); (2) \textbf{MTPE} test data where translated template is first translated by machines and post-edited later by professional translators. As shown in Table~\ref{tb:mt_test_result}, the overall reported results on MT test data are higher than those reported on MTPE test data, which is expected because the distribution of the MT test data is more similar to the MT training data. Although there are some differences on individual languages, the conclusions derived from the evaluations on the MT test data remain the same as those derived from the evaluation on the MTPE test data. We also calculate the Spearman rank correlation coefficient between the average results reported on MTPE test data and MT test data in Table~\ref{tb:mt_test_result}, which shows a statistically high correlation between the system performance on the MT test data and MTPE test data\footnote{Table~\ref{tb:spearman_breakdown} in the appendix shows detailed scores.}. Therefore, we show that the MT test data can be used as a proxy to estimate the model performance on the real test data for more languages. Thus we build MT test data for another 17 languages that are supported by Google Translate, Trip Advisor and Booking.com at the same time, as stated in Table~\ref{tb:selected_languages} and Table~\ref{tb:stats_ontology_17} in the appendix. Table~\ref{tb:global} shows the results of Zero-Shot (E\&E)\xspace and SUC\xspace on the test data of F\&F\xspace, F\&E\xspace and E\&F\xspace in 20 languages. The results show that the model has the best performance in the F\&E\xspace use case compared with the other two use cases, which is consistent with our findings in Table~\ref{tb:Use_case_result}.
\begin{table}[]
\centering
\resizebox{\columnwidth}{!}{%
\begin{tabular}{lcc|cc}
\toprule
\textbf{Use Case} & \multicolumn{2}{c}{\textbf{F2F}}&\multicolumn{2}{|c}{\textbf{F2E}}\\
\midrule
\textbf{Methods} & \textbf{MT Test} & \textbf{MTPE Test} & \textbf{MT Test} & \textbf{MTPE Test} \\
\midrule
Zero-Shot (E\&E)\xspace & 1.29 & 1.28 & 9.64 & 9.12 \\
Translate-Train\xspace & 3.71 & 3.65 & 4.17 & 3.97 \\
SUC\xspace & 35.78 & 28.96 & 56.15 & 48.71 \\
BBUC\xspace & 36.31 & 29.74 & 57.84 & 54.29 \\
MBUC\xspace & \textbf{37.89} & \textbf{30.76} & \textbf{58.76} & \textbf{56.28} \\
\midrule
Spearman's correlation & \multicolumn{2}{c}{\textbf{1.0}}&\multicolumn{2}{|c}{\textbf{1.0}}\\
\bottomrule
\end{tabular}
}
\caption{Comparison of average joint accuracy on DST reported on MT test data and MTPE test data for use case F\&F\xspace and F\&E\xspace}
\label{tb:mt_test_result}
\end{table}
\begin{table}[t!]
\centering
\scalebox{0.7}{
\begin{tabular}{lcc}
\toprule
\textbf{Case} &\textbf{Method} & \textbf{Avg} \\
\midrule
\multirow{2}{*}{F\&F\xspace} &Zero-Shot (E\&E)\xspace & 1.48 \\
& SUC\xspace & \textbf{16.12} \\
\midrule
\multirow{2}{*}{F\&E\xspace} &Zero-Shot (E\&E)\xspace & 9.03 \\
& SUC\xspace & \textbf{34.20} \\
\midrule
\multirow{2}{*}{E\&F\xspace} &Zero-Shot (E\&E)\xspace & 1.97 \\
& SUC\xspace & \textbf{23.40} \\
\bottomrule
\end{tabular}}
\caption{Average results of Zero-Shot (E\&E)\xspace on test data of F\&F\xspace, F\&E\xspace and E\&F\xspace in 20 languages. Please refer to Table~\ref{tb:global_breakdown} in Section~\ref{appendix:20languages} for the break down results of 20 languages. }
\label{tb:global}
\end{table}
\section{Conclusions}
\label{sec:conclusion}
In this paper, we provide an analysis on three unexplored use cases for multilingual task-oriented dialogue systems. We propose a new data curation method that leverages a machine translation system and local entities in target languages to create a new multilingual TOD dataset,~\mbox{GlobalWoZ}\xspace. We propose a series of strong baseline methods and conduct extensive experiments on \mbox{GlobalWoZ}\xspace to encourage research for multilingual ToD systems. Besides, we extend the coverage of languages on multilingual ToD to 20 languages, marking the one step further towards building a globalized multilingual ToD system for all of the world’s citizen.
\section{Ethical Review}
In this section, we would like to address the ethical concerns. All the professional translators in this project have been properly compensated. For Chinese and Spanish, we have followed the standard procurement requirements and engaged three translation companies for quality and price comparison. A small sample of the data had been given to them for MTPE and we then compared their translation results. Following that, we selected the company that produced the best sample translation, and submitted the full translation orders according to the agreed price quotations. For Indonesian, three translation companies were also requested to provide sample MTPE, but our quality check found the quality of these samples to be unsatisfactory. So, no company was engaged, and our in-house Indonesian linguistic resources were used instead. These Indonesian linguists were assigned to work on this project during normal working hours and given proper compensation complying with the local labor laws.
\section*{Acknowledgements}
\label{sec:acknowledgements}
This research is partly supported by the Alibaba-NTU Singapore Joint Research Institute, Nanyang Technological University. All the costs for machine translation post-editing are funded by DAMO Academy, Alibaba Group. We would like to thank the help from our Alibaba colleagues, Haiyun Peng, Zifan Xu and Ruidan He, and our NTU-NLP team member, Chengwei Qin in this work as well.
\section*{Appendix}
\section{Comparison of Four Use Cases}\label{appendix:comparison}
\begin{table}[h!]
\centering
\scalebox{0.8}{
\begin{tabular}{cccc}
\toprule
\multirow{2}{*}{Use Case} & \multirow{2}{*}{Source ToD} & Speaker & Country \\
& & (ToD Context) & (ToD Ontology) \\ \midrule
F\&F\xspace & \multirow{4}{*}{English} & Foreign Lang. & Foreign Lang. \\
F\&E\xspace & & Foregin Lang. & English \\
E\&F\xspace & & English & Foreign Lang. \\
E\&E\xspace & & English & English \\
\bottomrule
\end{tabular}
}
\caption{Four use cases of multilingual ToD systems: A foreign language or English speaker travels to a country of a foreign language or English. }
\label{tb:use-case}
\end{table}
\section{Examples of Labeled Sequence Translation}\label{appendix:sequence_translation}
\begin{figure*}[h!]
\centering
\includegraphics[scale=0.2]{translation.pdf}
\caption{An instance of labeled sequence translation with google translate, from English to three target languages, Mandarin, Spanish and Indonesian.}
\label{fig:sequence_translation}
\end{figure*}
\section{BLEU Score of MT versus MTPE Test Template} \label{appendix:bleu}
\begin{table}[h!]
\centering
\scalebox{0.8}{
\begin{tabular}{lcccc}
\toprule
Languages & Zh & Es & Id & Avg \\
\midrule
BLEU Score & 55.61 & 49.33 & 48.97 & 51.30 \\
\bottomrule
\end{tabular}}
\caption{BLEU Scores of MT Test Template using MTPE Test Template as reference.}
\end{table}
\newpage
\section{Test Set Distribution} \label{appendix:test_dist}
\begin{figure*}[h!]
\centering
\includegraphics[scale=0.5]{test_distribution.png}
\caption{Gold English Test Set Distribution by Domains. We follow this distribution to select the top 500 high-scoring dialogues in the test set for post-editing.}
\label{fig:test_distribution}
\end{figure*}
\section{Selected Languages} \label{appendix:sel_lang}
\begin{table*}[h!]
\resizebox{\columnwidth}{!}{%
\centering
\begin{tabular}{lcccccc}
\toprule
Language & ISO639-1code & Language Family & \# Wikipedia articles (in millions) & High / Middle/ Low Resource & Writing Script & Selected City \\
\midrule
English & en & IE: Germanic & 6.35 & High & Latin & Cambridge \\
Swedish & sv & IE: Germanic & 2.95 & High & Latin & Stockholm \\
German & de & IE: Germanic & 2.61 & High & Latin & Berlin \\
French & fr & IE: Romance & 2.35 & High & Latin & Paris \\
Dutch & nl & IE: Germanic & 2.06 & High & Latin & Amsterdam \\
Russian & ru & IE: Slavic & 1.74 & High & Cyrillic & Moscow \\
Italian & it & IE: Romance & 1.71 & High & Latin & Rome \\
Spanish & es & IE: Romance & 1.71 & High & Latin & Barcelona \\
Japanese & ja & Japonic & 1.28 & High & Ideograms & Tokyo \\
Vietnamese & vi & Austro-Asiatic & 1.27 & High & Latin & Ho Chi Minh City \\
Mandarin & zh & Sino-Tibetan & 1.22 & High & Chinese ideograms & Shanghai \\
Arabic & ar & Afro-Asiatic & 1.13 & High & Arabic & Cairo \\
Portuguese & pt & IE: Romance & 1.07 & High & Latin & Lisbon \\
Indonesian & id & Austronesian & 0.59 & Middle & Latin & Jakarta \\
Norwegian & no & IE: Germanic & 0.56 & Middle & Latin & Oslo \\
Korean & ko & Koreanic & 0.55 & Middle & Hangul & Seoul \\
Turkish & tr & Turkic & 0.42 & Middle & Latin & İstanbul \\
Hebrew & he & Afro-Asiatic & 0.30 & Low & Hebrew & Tel Aviv \\
Danish & da & IE: Germanic & 0.27 & Low & Latin & Copenhagen \\
Greek & el & IE: Greek & 0.20 & Low & Greek & Athens \\
Thai & th & Kra-Dai & 0.14 & Low & Brahmic & Bangkok \\
\bottomrule
\end{tabular}
}
\caption{Statistics about languages in the cross-lingual benchmark. The selected 21 languages (including English) belong to 8 language families and 1 isolate, with
Indo-European (IE) having the most members. We categorize the languages with more than 1 million, more than 400 thousand but less than 1 million, less than 400 thousand Wikipedia articles as high resource languages, middle resource languages and low resource languages. For each language, we select one city for each language to collect localized ontology. }
\label{tb:selected_languages}
\end{table*}
\newpage
\section{Statistics of Entities in the Collected Ontology} \label{appendix:stat_ontology}
\begin{table}[h!]
\centering
\scalebox{0.78}{
\begin{tabular}{lccccccc}
\toprule
\textbf{Languages} & \textbf{rest.} & \textbf{hotel} & \textbf{attr.} & \textbf{train} & \textbf{taxi} \\
\midrule
en & 110 & 33 & 79 & 2828 & 222 \\
zh & 3000 & 496 & 1000 & 100 & 4496 \\
es & 3000 & 426 & 1000 & 100 & 4426 \\
id & 3000 & 999 & 792 & 100 & 4791 \\
ar & 2989 & 680 & 1000 & 100 & 4669 \\
da & 2343 & 165 & 1000 & 100 & 3508 \\
de & 2988 & 659 & 1000 & 100 & 4647 \\
el & 2600 & 1000 & 1000 & 100 & 4600 \\
fr & 3000 & 1000 & 1000 & 100 & 5000 \\
he & 1558 & 258 & 1000 & 100 & 2258 \\
it & 3000 & 800 & 1000 & 100 & 2800 \\
ja & 2967 & 864 & 1000 & 100 & 4831 \\
ko & 2990 & 532 & 1000 & 100 & 4522 \\
nl & 2990 & 537 & 1000 & 100 & 4527 \\
no & 1293 & 95 & 757 & 100 & 2145 \\
pt & 2993 & 951 & 1000 & 100 & 4944 \\
ru & 2985 & 531 & 1000 & 100 & 4516 \\
sv & 3000 & 214 & 891 & 100 & 4105 \\
th & 2995 & 1000 & 1000 & 100 & 4995 \\
tr & 2986 & 533 & 1000 & 100 & 4519 \\
vi & 2991 & 773 & 1000 & 100 & 4764\\
\bottomrule
\end{tabular}
}
\caption{Statistics of entities in the collected ontology in different languages. We count the number of entities in the database of each domain. Noticed that in the Taxi database of MultiWoZ, it only list down the taxi colors, taxi types and taxi phones. The taxi destination and departure refer to the entities in the restaurant, hotel and attraction domains. Thus, we use the sum of the number of entities in Restaurant, Hotel and Attraction domains as a proxy of the total number of entities in taxi domain. Besides, we follow MultiWoZ to collect one hospital and one police station for each city.}
\label{tb:stats_ontology_17}
\end{table}
\section{Statistics of \mbox{GlobalWoZ}\xspace} \label{appendix:data_stat}
\begin{table}[h!]
\centering
\resizebox{\columnwidth}{!}{%
\begin{tabular}{lcccccccccccc}
\toprule
Use Case & \multicolumn{4}{l}{F\&F} & \multicolumn{4}{l}{F\&E} & \multicolumn{4}{l}{E\&F} \\
Languages & Train \& Dev & Method & Test & Method & Train \& Dev & Method & Test & Method & Train \& Dev & Method & Test & Method \\
\midrule
zh & 9438 & MT & 1000 & MTPE & 9438 & MT & 1000 & MTPE & 9438 & Human & 1000 & Human \\
es & 9438 & MT & 1000 & MTPE & 9438 & MT & 1000 & MTPE & 9438 & Human & 1000 & Human \\
id & 9438 & MT & 1000 & MTPE & 9438 & MT & 1000 & MTPE & 9438 & Human & 1000 & Human \\
ar & 9438 & MT & 1000 & MT & 9438 & MT & 1000 & MT & 9438 & Human & 1000 & Human \\
da & 9438 & MT & 1000 & MT & 9438 & MT & 1000 & MT & 9438 & Human & 1000 & Human \\
de & 9438 & MT & 1000 & MT & 9438 & MT & 1000 & MT & 9438 & Human & 1000 & Human \\
el & 9438 & MT & 1000 & MT & 9438 & MT & 1000 & MT & 9438 & Human & 1000 & Human \\
fr & 9438 & MT & 1000 & MT & 9438 & MT & 1000 & MT & 9438 & Human & 1000 & Human \\
he & 9438 & MT & 1000 & MT & 9438 & MT & 1000 & MT & 9438 & Human & 1000 & Human \\
it & 9438 & MT & 1000 & MT & 9438 & MT & 1000 & MT & 9438 & Human & 1000 & Human \\
ja & 9438 & MT & 1000 & MT & 9438 & MT & 1000 & MT & 9438 & Human & 1000 & Human \\
ko & 9438 & MT & 1000 & MT & 9438 & MT & 1000 & MT & 9438 & Human & 1000 & Human \\
nl & 9438 & MT & 1000 & MT & 9438 & MT & 1000 & MT & 9438 & Human & 1000 & Human \\
no & 9438 & MT & 1000 & MT & 9438 & MT & 1000 & MT & 9438 & Human & 1000 & Human \\
pt & 9438 & MT & 1000 & MT & 9438 & MT & 1000 & MT & 9438 & Human & 1000 & Human \\
ru & 9438 & MT & 1000 & MT & 9438 & MT & 1000 & MT & 9438 & Human & 1000 & Human \\
sv & 9438 & MT & 1000 & MT & 9438 & MT & 1000 & MT & 9438 & Human & 1000 & Human \\
th & 9438 & MT & 1000 & MT & 9438 & MT & 1000 & MT & 9438 & Human & 1000 & Human \\
tr & 9438 & MT & 1000 & MT & 9438 & MT & 1000 & MT & 9438 & Human & 1000 & Human \\
vi & 9438 & MT & 1000 & MT & 9438 & MT & 1000 & MT & 9438 & Human & 1000 & Human \\
\bottomrule
\end{tabular}
}
\caption{Statistics of created dataset, \mbox{GlobalWoZ}\xspace for each use case in each target language. For E\&F\xspace, as the context is the original Engish data, we consider it is created by human. For test data of zh, es and id, we replace the entities twice to boostrap the test data to 1000 dialogues. We are currently preparing the post editing of the other 500 dialogues in test data. Meanwhile, we are leveraging machine translation to prepare the train data for the 17 languages and will release it with baselines in the next version soon.}
\label{tb:stats_globalwoz}
\end{table}
\section{Dialogue Examples} \label{appendix:dialogue_example}
\begin{figure*}[h!]
\centering
\includegraphics[width=\columnwidth]{example.pdf}
\caption{Examples of some utterances in original E\&E\xspace data, MT data and MTPE data,}
\label{fig:example}
\end{figure*}
\newpage
\section{Summary of Proposed Baselines}\label{appendix:summary_baselines}
\begin{table}[h!]
\centering
\scalebox{0.7}{
\begin{tabular}{lccccc}
\toprule
\textbf{Methods} & En Context & En Entities & Local Context & Local Entities & Translated Entites \\
\midrule
Zero-Shot (E\&E)\xspace & \CheckmarkBold & \CheckmarkBold & & & \\
Translate-Train\xspace & & & \CheckmarkBold & & \CheckmarkBold \\
\midrule
SUC\xspace (F\&F\xspace) & & & \CheckmarkBold & \CheckmarkBold & \\
SUC\xspace (F\&E\xspace) & & \CheckmarkBold & \CheckmarkBold & & \\
SUC\xspace (E\&F\xspace) & \CheckmarkBold & & & \CheckmarkBold & \\
\bottomrule
\end{tabular}
}
\caption{Accessibility of different types of context and entities for each method.}
\label{tb:Methods_data}
\end{table}
\begin{table}[h!]
\centering
\scalebox{0.8}{
\begin{tabular}{lcccc}
\toprule
\textbf{Methods} &E\&E\xspace & F\&F\xspace & F\&E\xspace & E\&F\xspace \\
\midrule
Zero-Shot (E\&E)\xspace & \CheckmarkBold & & & \\
Translate-Train\xspace & & & & \\
\midrule
SUC\xspace (F\&F\xspace) & & \CheckmarkBold & & \\
SUC\xspace (F\&E\xspace) & & & \CheckmarkBold & \\
SUC\xspace (E\&F\xspace) & & & & \CheckmarkBold \\
\midrule
BBUC\xspace (E\&E\xspace+ F\&F\xspace) & \CheckmarkBold & \CheckmarkBold & & \\
BBUC\xspace (E\&E\xspace+ F\&E\xspace) & \CheckmarkBold & & \CheckmarkBold & \\
BBUC\xspace (E\&E\xspace+ E\&F\xspace) & \CheckmarkBold & & & \CheckmarkBold \\
\midrule
MBUC\xspace (E\&E\xspace+ F\&F\xspace) & \CheckmarkBold & \CheckmarkBold & & \\
MBUC\xspace (E\&E\xspace+ F\&E\xspace) & \CheckmarkBold & & \CheckmarkBold & \\
MBUC\xspace (E\&E\xspace+ E\&F\xspace) & \CheckmarkBold & & & \CheckmarkBold \\
\midrule
MMUC\xspace (E\&E\xspace+ F\&F\xspace + F\&E\xspace + E\&F\xspace) & \CheckmarkBold & \CheckmarkBold & \CheckmarkBold & \CheckmarkBold \\
\bottomrule
\end{tabular}
}
\caption{Accessibility of data in each use case for each method. Noticed that Translate-Train\xspace doesn't have access to the data of the four use cases. Translate-Train\xspace has access to a set of pseudo-labeled training data created by replacing the placeholders in the translated template with machine-translated entities instead of local entities.}
\label{tb:Methods_data_access}
\end{table}
\section{Use Case E\&E\xspace} \label{appendix:EE}
We also compare the performance of all methods on the original E\&E\xspace test data. As \textbf{Zero-Shot (E\&E)\xspace} is trained on monolingual English training data, it gets a high accuracy of 52.78 on the English test data. In contrast, \textbf{Translate-Train\xspace} and \textbf{SUC\xspace} (F\&F\xspace) perform poorly on the English test data, because both of them have no access to any English data. Comparing to \textbf{SUC}\xspace (F\&F\xspace), \textbf{SUC\xspace} (F\&E\xspace) and \textbf{SUC\xspace} (E\&F\xspace) achieve higher accuracy scores as they either have access to English context or English entities. When we perform bilingual and multilingual joint training (i.e., \textbf{BBUC\xspace} and \textbf{MBUC\xspace}), the base model has a performance increase except \textbf{MBUC\xspace} (E\&E\xspace+ E\&F\xspace). This shows that bilingual and multilingual joint training may be used to improve the performance on source language. Further research can be done in this line.
\begin{table}[h!]
\centering
\scalebox{0.7}{
\begin{tabular}{lc}
\toprule
\textbf{Methods} & \textbf{En}\\
\midrule
Zero-Shot (E\&E)\xspace & 52.78 \\
Translate-Train\xspace & 2.27 \\
\midrule
SUC\xspace (F\&F\xspace) & 1.09 \\
SUC\xspace (F\&E\xspace) & 6.39 \\
SUC\xspace (E\&F\xspace) & 5.46 \\
\midrule
BBUC\xspace (E\&E\xspace+ F\&F\xspace) & 52.87 \\
BBUC\xspace (E\&E\xspace+ F\&E\xspace) & \textbf{53.69} \\
BBUC\xspace (E\&E\xspace+ E\&F\xspace) & 53.05 \\
\midrule
MBUC\xspace (E\&E\xspace+ F\&F\xspace) & 53.28 \\
MBUC\xspace (E\&E\xspace+ F\&E\xspace) & 53.43 \\
MBUC\xspace (E\&E\xspace+ E\&F\xspace) & 51.75 \\
\bottomrule
\end{tabular}
}
\caption{Joint accuracy on DST in three target languages on the English test data.
}
\label{tb:En_result}
\end{table}
\newpage
\section{Breakdown of Few Shot Results}
\begin{table}[h!]
\centering
\scalebox{0.7}{
\begin{tabular}{lcccc}
\toprule
\multicolumn{5}{c}{Zero Shot (E\&E)} \\
\midrule
Use Case & Zh & Es & Id & Avg \\
\midrule
F2F & 1.22 & 1.38 & 1.26 & 1.28 \\
F2E & 6.92 & 11.34 & 9.09 & 9.12 \\
E2F & 1.69 & 1.81 & 1.82 & 1.77 \\
\midrule
\multicolumn{5}{c}{Few Shot + Zero Shot (E\&E)} \\
\midrule
Use Case & Zh & Es & Id & Avg \\
\midrule
F2F & 15.93 & 7.13 & 12.09 & 11.72 \\
F2E & 39.88 & 39.38 & 43.26 & 40.84 \\
E2F & 20.61 & 14.17 & 18.55 & 17.78 \\
\midrule
\multicolumn{5}{c}{SUC} \\
\midrule
Use Case & Zh & Es & Id & Avg \\
\midrule
F2F & 36.97 & 24.66 & 25.26 & 28.96 \\
F2E & 56.28 & 41.94 & 47.93 & 48.71 \\
E2F & 38.56 & 28.00 & 43.82 & 36.79 \\
\midrule
\multicolumn{5}{c}{Few Shot + SUC} \\
\midrule
Use Case & Zh & Es & Id & Avg \\
\midrule
F2F & 37.81 & 25.15 & 39.51 & 34.16 \\
F2E & 58.39 & 53.03 & 54.02 & 55.15 \\
E2F & 38.75 & 27.66 & 44.23 & 36.88 \\
\midrule
\multicolumn{5}{c}{Few Shot + Zero Shot (E\&E) + SUC} \\
\midrule
Use Case & Zh & Es & Id & Avg \\
\midrule
F2F & 37.52 & 26.44 & 40.15 & 34.70 \\
F2E & 59.21 & 54.93 & 56.17 & 56.77 \\
E2F & 39.51 & 27.84 & 45.48 & 37.61 \\
\bottomrule
\end{tabular}
}
\caption{A breakdown of few-shot cross-lingual average joint accuracy on DST over three target languages in three use cases.}
\label{tb:few_shot_breakdown}
\end{table}
\newpage
\section{Concrete Examples where Translate-Train Performs Badly on the Test Data with Real Local Entities.} \label{appendix:example_TT}
Through investigation, we found that the Translate-Train\xspace method usually performed badly in two main scenarios. Figure~\ref{fig:example} is the illustrations of the two scenarios. Scenario 1 is when the Translate-Train\xspace can predict values that are close to the meaning of the ground truth values but suffer from the problems of translationese. For example, model trained with Translate-Train\xspace may predict "\cchar{美食酒吧}" (gastropub), which is a direct translation of gastropub and not commonly used in Chinese instead of "\cchar{酒吧餐}" (bar). Scenario 2 is when Translate-Train\xspace needs to predict the name of real localized entities which Translate-Train\xspace doesn't have access to. For example, trained with Translate-Train\xspace may predict "\cchar{冈维尔酒店}" (Gonville Hotel) which is a direct translation of Gonville Hotel, instead of "\cchar{汉庭酒店}" (Hanting Hotel) which is unseen in Translate-Train\xspace training data.
\begin{figure*}[h!]
\centering
\includegraphics[scale=0.9]{example_TT.pdf}
\caption{Concrete examples where Translate-Train performs badly on the test data with real local entities.
}
\label{fig:example}
\end{figure*}
\section{Breakdown of the Results of Local Context vs Local Entities by Languages}
\begin{table}[h!]
\centering
\scalebox{0.75}{
\begin{tabular}{lccccc}
\toprule
\multicolumn{5}{c}{E\&E\xspace(en)} \\
\midrule
Context vs Entities & \cellcolor{lavender} Zh & Es & Id & Avg \\
En\_Context & \cellcolor{lavender} 5.37 & 5.33 & 5.67 & 5.46 \\
En\_Entites & \cellcolor{lavender} 3.49 & 7.78 & 7.90 & 6.39 \\
\midrule
\multicolumn{5}{c}{F\&F\xspace(zh)} \\
\midrule
Context vs Entities & \cellcolor{lavender} En & \cellcolor{lavender} Es & \cellcolor{lavender} Id & Avg \\
Zh\_Context &\cellcolor{lavender} 1.74 & \cellcolor{lavender} 1.77 &\cellcolor{lavender} 1.80 & 1.77 \\
Zh\_Entites &\cellcolor{lavender} 0.27 &\cellcolor{lavender} 0.73 & \cellcolor{lavender} 0.10 & 0.36 \\
\midrule
\multicolumn{5}{c}{F\&F\xspace(es)} \\
\midrule
Context vs Entities & En & \cellcolor{lavender} Zh & Id & Avg \\
Es\_Context & 1.73 & \cellcolor{lavender} 2.01 & 3.37 & 2.37 \\
Es\_Entites & 3.92 &\cellcolor{lavender} 0.44 & 2.86 & 2.41 \\
\midrule
\multicolumn{5}{c}{F\&F\xspace(id)} \\
\midrule
Context vs Entities & En & \cellcolor{lavender} Zh & Es & Avg \\
Id\_Context & 2.07 &\cellcolor{lavender} 2.18 & 2.94 & 2.40 \\
Id\_Entites & 3.92 & \cellcolor{lavender} 0.84 & 3.48 & 2.75 \\
\bottomrule
\end{tabular}
}
\caption{A breakdown of comparison of the impact of local context and local entities on joint accuracy for DST in each language. The cases where context and entities are in different script types are highlighted in lavender color.}
\label{tb:context_vs_entities_language}
\end{table}
\begin{table}[h!]
\centering
\scalebox{0.7}{
\begin{tabular}{lcc}
\toprule
\textbf{Train Set} & \textbf{different script type} & \textbf{same script type} \\
\midrule
Local Context Only & \textbf{2.48} & 3.52 \\
Local Entities Only & 0.98 & \textbf{4.98} \\
\bottomrule
\end{tabular}
}
\caption{Comparison of the impact of script type on Local Context Only vs Local Entities Only. It shows that training with local entities is more important if the entities and contexts are written in the same type of language script (e.g. Latin script), otherwise training with local contexts is more important.}
\label{tb:script_type}
\end{table}
\newpage
\section{Breakdown of MT Test Data vs MTPE Test Data by Languages}
\begin{table}[h!]
\centering
\scalebox{0.75}{
\begin{tabular}{lcccccc}
\toprule
Languages & \multicolumn{2}{c}{Zh} & \multicolumn{2}{c}{Es} & \multicolumn{2}{c}{Id} \\
\midrule
F2F & MT & MTPE & MT & MTPE & MT & MTPE \\
\midrule
Zero-Shot (E\&E)\xspace & 1.19 & 1.22 & 1.40 & 1.38 & 1.28 & 1.26 \\
Translate-Train\xspace & 2.50 & 2.61 & 2.81 & 2.59 & 5.81 & 5.74 \\
SUC\xspace & 37.79 & 36.97 & 26.95 & 24.66 & 42.59 & 25.26 \\
BBUC\xspace & 38.62 & 37.32 & 27.34 & 25.52 & 42.96 & 26.39 \\
MBUC\xspace & 39.11 & 38.01 & 29.17 & 26.03 & 45.39 & 28.22 \\
\midrule
Spearman’s correlation & \multicolumn{2}{c}{1.00} & \multicolumn{2}{c}{1.00} & \multicolumn{2}{c}{1.00} \\
\midrule
F2E & MT & MTPE & MT & MTPE & MT & MTPE \\
\midrule
Zero-Shot (E\&E)\xspace & 7.61 & 6.92 & 11.67 & 11.34 & 9.64 & 9.09 \\
Translate-Train\xspace & 2.25 & 2.28 & 5.25 & 4.97 & 5.03 & 4.67 \\
SUC\xspace & 57.10 & 56.28 & 55.70 & 41.94 & 55.64 & 47.93 \\
BBUC\xspace & 59.05 & 59.87 & 57.68 & 48.20 & 56.80 & 54.79 \\
MBUC\xspace & 60.48 & 60.37 & 57.04 & 53.56 & 58.23 & 54.93 \\
\midrule
Spearman’s correlation & \multicolumn{2}{c}{1.00} & \multicolumn{2}{c}{0.90} & \multicolumn{2}{c}{1.00} \\
\bottomrule
\end{tabular}
}
\caption{Spearman rank correlation coefficient between the results on MTPE test data and MT test data for each language.}
\label{tb:spearman_breakdown}
\end{table}
\section{Breakdown of Results of 20 Languages} \label{appendix:20languages}
\begin{table*}[h!]
\centering
\scalebox{0.5}{
\begin{tabular}{lcccccccccccccccccccccc}
\toprule
\textbf{Case} &\textbf{Method} & \cellcolor{palepink} \textbf{zh} & \cellcolor{palepink} \textbf{es} & \cellcolor{palepink} \textbf{id} & \textbf{ar} & \textbf{da} & \textbf{de} & \textbf{el} & \textbf{fr} & \textbf{he} & \textbf{it} & \textbf{ja} & \textbf{ko} & \textbf{nl} & \textbf{no} & \textbf{pt} & \textbf{ru} & \textbf{sv} & \textbf{th} & \textbf{tr} & \textbf{vi} & \textbf{avg} \\
\midrule
\multirow{2}{*}{F\&F\xspace} &Zero-Shot (E\&E)\xspace & \cellcolor{palepink} 1.22 & \cellcolor{palepink} 1.38 & \cellcolor{palepink} 1.26 & 1.49 & 1.52 & 1.52 & 1.51 & 2.04 & 1.47 & 1.55 & 1.48 & 1.51 & 1.55 & 1.51 & 1.53 & 1.52 & 1.41 & 1.57 & 1.22 & 1.41 & 1.48 \\
& SUC\xspace & \cellcolor{palepink}\textbf{36.97} & \cellcolor{palepink}\textbf{24.66} &\cellcolor{palepink} \textbf{25.26} & \textbf{14.33} & \textbf{24.08} & \textbf{15.31} & \textbf{4.33} & \textbf{23.72} & \textbf{7.76} & \textbf{18.81} & \textbf{20.98} & \textbf{1.71} & \textbf{23.87} & \textbf{24.86} & \textbf{14.91} & \textbf{13.00} & \textbf{11.31} & \textbf{2.74} & \textbf{10.65} & \textbf{3.06} & \textbf{16.12} \\
\midrule
\multirow{2}{*}{F\&E\xspace} & Zero-Shot (E\&E)\xspace &\cellcolor{palepink} 6.92 & \cellcolor{palepink} 11.34 & \cellcolor{palepink} 9.09 & 6.80 & 10.97 & 10.15 & 6.74 & 15.87 & 7.81 & 9.40 & 3.17 & 4.92 & 11.79 & 11.46 & 10.12 & 8.97 & 10.31 & 10.89 & 5.98 & 7.92 & 9.03 \\
& SUC\xspace & \cellcolor{palepink} \textbf{56.28} & \cellcolor{palepink} \textbf{41.94} & \cellcolor{palepink} \textbf{47.93} & \textbf{29.98} & \textbf{29.79} & \textbf{30.55} & \textbf{30.58} & \textbf{54.03} & \textbf{29.27} & \textbf{30.16} & \textbf{51.19} & \textbf{28.21} & \textbf{30.58} & \textbf{30.28} & \textbf{29.63} & \textbf{29.84} & \textbf{30.64} & \textbf{18.07} & \textbf{29.18} & \textbf{25.82} & \textbf{34.20} \\
\midrule
\multirow{2}{*}{E\&F\xspace} & Zero-Shot (E\&E)\xspace & 1.69 & 1.81 & 1.82 & 1.94 & 1.98 & 1.96 & 2.01 & 2.82 & 1.99 & 1.98 & 1.92 & 1.92 & 1.94 & 1.97 & 1.95 & 1.99 & 1.89 & 1.86 & 2.00 & 1.99 & 1.97 \\
& SUC\xspace & \textbf{38.56} & \textbf{28.00} & \textbf{43.82} & \textbf{22.98} & \textbf{43.00} & \textbf{23.71} & \textbf{5.73} & \textbf{22.61} & \textbf{10.65} & \textbf{32.07} & \textbf{20.05} & \textbf{2.13} & \textbf{44.03} & \textbf{44.61} & \textbf{22.19} & \textbf{20.13} & \textbf{16.52} & \textbf{5.24} & \textbf{16.83} & \textbf{5.07} & \textbf{23.40} \\
\bottomrule
\end{tabular}}
\caption{Results of Zero-Shot (E\&E)\xspace on test data of F\&F\xspace, F\&E\xspace and E\&F\xspace in 20 languages. Test data of F\&F\xspace and F\&E\xspace in the three languages highlight in pink color are built with MTPE data and the rest are built with MT data.}
\label{tb:global_breakdown}
\end{table*}
We observe that there are a few languages like Thai and Vietnamese have low results than other languages. Through investigation, we found that it was caused by failing to predict the tone marks in most of cases. For example, the model may predict \textit{“nha khach”} for hotel type while \textit{“{\fontencoding{T5} nhà khách}”} is the ground truth. We may explore options for post-processing or other models to improve the performance on these languages upon the release of the data.
|
\section{The expansions in $\epsilon$ and $1/w$ in Bjorken flow}\label{app:1overw}
In the studies of conformal Bjorken flow in the past decade, an important role was played by the following dimensionless clock variable~\cite{Heller:2011ju}
\begin{equation}
w \equiv \tau \, T(\tau),
\end{equation}
where $\tau$ is the proper time and $T(\tau)$ is the local temperature as defined in the main text. The purpose of this Appendix is to review the known relation between the expansion in the Knudsen number in conformal Bjorken flow -- the $1/w$-expansion mentioned in the Introduction -- and the gradient expansion of hydrodynamic constitutive relations as given by Eq.~\eqref{intro:T_decomposition_3} in the main text.
While this relation is general, i.e. it applies to any conformally invariant model, we will present it in the context of the BRSSS model captured by Eq.~\eqref{BRSSS:dynamical_constitutive_relation} in the main text in four spacetime dimensions. For Bjorken flow Eq.~\eqref{BRSSS:dynamical_constitutive_relation} reduces to
\begin{align}
&\Pi_\star'(\tau) + \left(\frac{4}{3\tau} + \frac{T(\tau)}{C_{\tau_\Pi}}\right)\Pi_\star(\tau) - \frac{3 C_{\lambda_1}}{4 C_\eta C_{\tau_\Pi}\mathcal{E}_0 T(\tau)^3}\Pi_\star(\tau)^2 \nonumber \\
&- \frac{8 C_\eta \mathcal{E}_0 T(\tau)^4}{9 C_{\tau_\Pi} \tau} = 0.
\end{align}
whereas the conservation equation always (in all theories) reads
\begin{equation}
T'(\tau) + \frac{T(\tau)}{3\tau} - \frac{\Pi_\star(\tau)}{2 \mathcal{E}_0 \tau T(\tau)^3}=0.
\end{equation}
Note that it is convenient to replace $\Pi_{\star}$ by a dimensionless quantity by taking its ratio with the energy density~${\cal E}$. This is simply related to the pressure anisotropy $\cal A$ measuring deviations from local thermal equilibrium in the following way
\begin{equation}
{\cal A} = \frac{{\cal P}_T-{\cal P}_L}{{\cal E}/3} = 9 \frac{\Pi_{\star}}{\cal E}.
\end{equation}
The $1/w$-expansion has been primarily discussed for the this quantity and up to second order it takes the form
\begin{equation}
\label{eq.Ain1/w}
{\cal A} = \frac{8 C_\eta}{w} + \frac{16 C_\eta (C_{\tau_\Pi}-C_{\lambda_1})}{3w^2}+ O\left(\frac{1}{w^3}\right).
\end{equation}
Using the expansion in $\epsilon$ of $\Pi_{\star}$ at low orders we obtain the following expression for $\cal A$
\begin{align}
{\cal A} = \frac{8C_\eta}{\tau T}\epsilon {-}\frac{8 C_\eta \left((C_{\tau_\Pi} {+}2 C_{\lambda_1}) T{+}9 C_{\tau_\Pi} \tau T' \right)}{3 \tau^2 T^3}\epsilon^2{+} O(\epsilon^3),
\end{align}
where $T = T(\tau)$. One can now use the conservation equation to replace derivatives of $T(\tau)$ at each order in terms of powers of~$w$, which directly leads to~\eqref{eq.Ain1/w}. This shows that knowing $\Pi_{\star}$ in the expansion in $\epsilon$ to order $n$ allows one to generate the expansion of $\cal A$ in $1/w$ up to and including terms $O(1/w^n)$. Note however, that the $n$-th order of the expansion in $1/w$ contains in principle contributions from the orders 2 to $n$ of the expansion in~$\epsilon$.
A key difference between the two expansions is that the functional expansion of $\Pi_\star$ in derivatives of ${\cal E}, U$ counted by~$\epsilon$ as in \rf{intro:T_decomposition_3} does not explicitly require invoking conservation laws (beyond the stated redundancies), whereas the $1/w$-expansion necessarily utilizes the conservation equation. Finally, note again that the discussion above applies to conformal Bjorken flow, while in Appendix \ref{app:DN} we discuss a particular non-conformal version of Bjorken flow.
\section{Further details on numerics and convergence \label{app:num}}
For the BRSSS simulations shown in Fig.\,\ref{heatmap} we use initial data $u = \frac{1}{4} f$,~$T = 1 + f$,~$\Pi_\ast = 0$ where
\begin{eqnarray}
f(x) = e^{-\frac{\cos(\pi x)^2}{2\pi^2 \gamma^2}} -e^{-\frac{1}{4\pi^2 \gamma^2}}I_0\left(\frac{1}{4\pi^2 \gamma^2}\right)
\end{eqnarray}
on a unit spatial circle ($x\sim x+1$). $f$ was chosen to locally reproduce a Gaussian of width $\gamma$ near $x=1/2$, respect the spatial periodicity, and the additive constant chosen so that $f$ has no homogeneous component in a cosine expansion. We used
${\cal E} = T^4$,~$C_{\tau_\Pi} = 1/4$,~$C_{\eta} = 1/(4\pi)$,~$\lambda_1 = 0$,~$\gamma^2 = 1/60$. The initial value problem is solved using RK4 on a uniform spatial grid with $N_x = 1600$ points for a total of $N_t=3200$ timesteps to reach the arbitrary choice of end time $t=1$, working with $300$ digits of precision. The differential operator for $\partial_x$ used periodic fourth-order finite differences, and second-order finite differences for $\partial_t$ one-sided at $t=0$ and $t=1$.
Due to the large numbers of successive applications of derivative operators used in evaluating $\Pi_\star^{(n)}$ we have placed particular importance on testing convergence of our results, both in the resolution of the numerical grid $N_x,N_t$ and in the number of digits of precision used. Tests confirming convergence of $|\Pi_\star^{(n)}|^{\frac{1}{n}}$ are shown in Fig.\,\ref{fig:convergence}. For a given resolution and precision, when the window of factorial growth ends it can be extended by increasing one, the other, or both.
\begin{figure}
\begin{center}
\includegraphics[width=\columnwidth]{convergence.pdf}
\caption{Both precision and numerical resolution are important factors in extracting the correct growth of the hydrodynamic expansion using our technique. Here we show convergence of $|\Pi_\star^{(n)}|^{\frac{1}{n}}$ for the BRSSS simulations of Fig.\,\ref{heatmap} at the point marked with `$+$'. \textbf{Upper panel:} Convergence to the continuum with increasing resolution at fixed precision. \textbf{Lower panel:} Convergence with increasing precision at fixed resolution.}
\label{fig:convergence}
\end{center}
\end{figure}
The results in the DN model had been obtained working in a curvilinear coordinate system in which the Minkowski metric reads
\begin{equation}\label{numerics:ds2_curvilinear_coordinates}
ds^2 = - e^{2\alpha(\tau,x)}d\tau^2+e^{2\beta(\tau,x)}dx^2 + d\vec{x}_\perp^2,
\end{equation}
and the fluid velocity is $U = e^{-\alpha}\partial_\tau$, in such a way that a line of constant $x, \vec{x}_\perp$ corresponds to a flow line. Besides the conservation equations and the relaxation equation obeyed by $\Pi_\star$, the evolution equations include one extra equation enforcing that the metric \eqref{numerics:ds2_curvilinear_coordinates} is flat. When performing numerical simulations, we worked with a compactified spatial coordinate $\zeta \in [-1,1]$ instead of $x$. Both coordinates are related as $x = \gamma^{-1}\tanh^{-1}(\zeta)$, with $\gamma \in \mathbb{R}^+$. Spatial derivatives were discretized with fourth-order centered finite-difference stencils, while the time evolution was performed with an explicit RK4 method. For the numerical results displayed in Fig.\,\ref{fig:DN_divergence}, we employed a spatial grid of spacing $d\zeta = 1/1800$, a time step $d\tau =0.5 d\zeta$, and worked with precision $300$. The time derivatives appearing in the recursion relation were computed with fourth-order finite-difference stencils.
\section{Further perspectives on geometric growth in the DN model}\label{app:DN}
We start by noting that the DN model, as originally defined in Ref.\,\cite{Denicol:2019lio}, features an additional term in the relaxation equation obeyed by $\Pi_{\mu\nu}$. This term is linear in $\Pi_{\mu\nu}$ and, in this work, we have set to zero the transport coefficient associated to it since this restriction does not modify our main conclusions regarding Bjorken flow. Hence, in the formulation of the DN model we have considered, the gradient expansion \eqref{intro:T_decomposition_3} for Bjorken flow is given by the following recursion relations,
\begin{subequations}\label{DN_rec_rel_tau}
\begin{equation}\label{DN_rec_rel_tau_1}
\Pi_\star^{(1)}{=}\frac{2a}{3\sigma_T} \frac{T}{\tau},
\end{equation}
\begin{equation}\label{DN_rec_rel_tau_2}
\Pi_\star^{(n+1)}{=}{-}\frac{a b \textrm{Kn}}{3}\Pi_\star^{(n)}{-}\frac{a b \textrm{Kn}}{4} \tau\partial_\tau \Pi_\star^{(n)}.
\end{equation}
\end{subequations}
The closed-form expression \eqref{DN:Bjorken_recursion-relation_solution} can be shown to be a solution of the recursion relations \eqref{DN_rec_rel_tau} by induction. The parameters appearing in Eq.\,\eqref{DN:Bjorken_closed-form_solution} are
\begin{subequations}\label{DN_T_exact_parameters}
\begin{equation}
T_{0,+} = \frac{c_0 \mathcal{E}_0}{3 n_0 (1+ c_0)} = c_0 T_{0,-},
\end{equation}
\begin{equation}\label{DN_alpha_pm}
\alpha_\pm = \frac{1}{3} + \frac{2}{a b \textrm{Kn}} \pm \frac{2}{3}\sqrt{\frac{4}{b} + \frac{9}{a^2 b^2 \textrm{Kn}^2}},
\end{equation}
\end{subequations}
where $\mathcal{E}_0$ is the energy density at $\tau=\tau_0$. Combining Eq.\,\eqref{DN:Bjorken_recursion-relation_solution} with Eq.\,\eqref{DN:Bjorken_closed-form_solution}, one finds the final form of the gradient expansion coefficients,
\begin{subequations}\label{DN_grad_exp_Bjorken}
\begin{equation}\label{DN_grad_exp_Bjorken_1}
\Pi_\star^{(n)} = \sum_{i=+,-} \Pi_{\star,i}^{(n)} = \sum_{i=+,-} A_i \left(\frac{\tau_0}{\tau}\right)^{4\frac{a b \textrm{Kn} + 3 \gamma_i}{3 a b \textrm{Kn}}}\gamma_i^n,
\end{equation}
\begin{equation}
A_\pm = - \frac{8 T_{0,\pm}}{b \tau_0 \sigma_T \textrm{Kn} (1- 3\alpha_\pm)}, \end{equation}
\begin{equation}\label{DN_gamma_pm}
\gamma_\pm = \frac{1}{2} \pm\frac{1}{6}\sqrt{9+ 4 a^2 b \textrm{Kn}^2}.
\end{equation}
\end{subequations}
According to Eq.\,\eqref{DN_grad_exp_Bjorken}, $\Pi_\star^{(n)}$ is given by the sum of two contributions which only grow geometrically at rates $\gamma_\pm$. No large-order factorially growing behavior is present.
For illustrative purposes, let us focus on the $c_0=0$ solution. This choice of initial conditions corresponds to the attractor solution for the normalized pressure anisotropy, which played a prominent role in the original analysis of Ref.\,\cite{Denicol:2019lio}. In this case, the terms in Eq.\,\eqref{DN_grad_exp_Bjorken} labeled by a $+$ vanish and the gradient expansion converges provided that
\begin{equation}
\textrm{Kn} < \textrm{Kn}^* \equiv \frac{3\sqrt{2}}{a\sqrt{b}}.
\end{equation}
This existence of a critical Knudsen number is qualitative agreement with the findings of Ref.\,\cite{Denicol:2019lio}. The quantitative difference in the value of $\textrm{Kn}^*$ reported here and the one originally quoted in Ref.\,\cite{Denicol:2019lio} is explained by the fact that our gradient expansion is different from the one considered there.
We conclude this Appendix by arguing in two different ways that the large-order geometric growth displayed by $\Pi_\star^{(n)}$ can be understood as a fine-tuned phenomenon.
The first way starts from the realisation that, had one chosen to write the closed-form solution \eqref{DN:Bjorken_recursion-relation_solution} in terms of an expansion in the longitudinal derivative $\partial_\tau$ instead of $\tau \partial_\tau$, one would had gotten an expression of the form
\begin{equation}
\Pi_\star^{(n)} = \sum_{k=0}^{n-1} c_{n,k} \tau^{k-1} \partial_\tau^k T(\tau),
\end{equation}
where, in particular,
\begin{equation}
c_{n,n-1} = \left(-\frac{1}{4}a b \textrm{Kn}\right)^{n-1} \frac{2a}{3\sigma_T}.\label{cn}
\end{equation}
For a $T(\tau)$ given by a linear combination of power-laws $\tau^{-\lambda}$, the contribution to the gradient expansion of the form
\begin{equation}
\Pi_\star \supset \sum_{n=1}^\infty c_{n,n-1} \tau^{n-2}\partial_\tau^{n-1} T(\tau),
\end{equation}
is a factorially divergent asymptotic series when evaluated at a given time since according to Eq.\,\eqref{cn} the coefficients $c_{n,n-1}$ grow geometrically and
\begin{equation}
\partial_\tau^k \tau^{-\lambda} = (-1)^k \tau^{-\lambda-k} \frac{\Gamma(\lambda+k)}{\Gamma(\lambda)}.
\end{equation}
The fact that the whole gradient expansion only grows geometrically when adding the remaining contributions is indicative of a very delicate cancellation between different factorially growing subsectors.
The second way builds upon an observation already performed in the main text: $\Pi_\star^{(1)}$ is a linear superposition of eigenfunctions of the differential operator $\mathcal{M}$. Let us elaborate on this. According to the definition \eqref{M_def} and the recursion relation \eqref{DN_rec_rel_tau_2},
\begin{equation}
\mathcal{M} = {-}\frac{a b \textrm{Kn}}{3}{-}\frac{a b \textrm{Kn}}{4} \tau\partial_\tau,
\end{equation}
in such that a way that its eigenfunctions $\psi_i$, $\mathcal{M}\psi_i = \xi_i \psi_i$, are given by
\begin{equation}
\psi_i = \tau^{-\frac{4}{3}-\frac{4}{a b \textrm{Kn}}\xi_i}.
\end{equation}
For a $\Pi_\star^{(1)}$ corresponding to a finite linear superposition of eigenfunctions,
\begin{equation}\label{DN_Pi1_spectral}
\Pi_\star^{(1)} = \sum_i c_i \psi_i,
\end{equation}
the gradient expansions grows geometrically with rates given by $\xi_i$,
\begin{equation}
\Pi_\star^{(n+1)} = \mathcal{M}^n\Pi_\star^{(1)} = \sum_i \xi_i^n c_i \psi_i.
\end{equation}
Eqs.\,\eqref{DN_Pi1_spectral} and \eqref{DN_rec_rel_tau_1} imply that
\begin{equation}\label{DN_T_spectral}
T = \frac{3}{2 a \rho_0 \tau_0 \textrm{Kn}}\sum_i c_i \tau^{-\frac{1}{3}-\frac{4}{a b \textrm{Kn}}\xi_i}.
\end{equation}
Inserting the expression above in the conservation equation fixes $\Pi_\star$,
\begin{equation}\label{DN_Pi*_spectral}
\Pi_\star = -\frac{9}{a^2 b \textrm{Kn}^2}\sum_i c_i \xi_i \tau^{-\frac{4}{3}-\frac{4}{a b \textrm{Kn}}\xi_i}.
\end{equation}
Finally, for the whole procedure to be self-consistent, the $\Pi_\star$ given by Eq.\,\eqref{DN_Pi*_spectral} has to solve the relaxation equation. This can only be achieved provided that the linear superposition \eqref{DN_Pi*_spectral} is restricted to two contributions with eigenvalues
\begin{equation}
\xi_{1,2} = \frac{1}{2}\pm\frac{1}{6}\sqrt{9+4a^2b\textrm{Kn}^2}.
\end{equation}
The eigenvalues $\xi_{1,2}$ agree with the growth rates quoted in Eq.\,\eqref{DN_gamma_pm} and, when introduced in Eq.\,\eqref{DN_T_spectral} (resp. Eq.\,\eqref{DN_Pi*_spectral}), match the exponents appearing in Eq.\,\eqref{DN_alpha_pm} (resp. Eq.\,\eqref{DN_grad_exp_Bjorken_1}).
In a generic model with a recursion relation of the form \eqref{M_def} $\mathcal{M}$ depends nontrivially on the hydrodynamic fields. In this situation, while a $\Pi_\star^{(1)}$ given by a finite superposition of eigenfunctions would result in a geometrically growing gradient expansion, demanding that the on-shell hydrodynamic fields conspire to produce precisely a $\Pi_\star^{(1)}$ of this form is a tall order. It is natural to expect that this is not the case for a generic flow, providing another sense in which the geometric growth of the gradient expansion for Bjorken flow in the DN model is a fine-tuned phenomenon.
\end{document} |
\section{Introduction}
Internet of Things (IoT) networks are vulnerable to attacks due to the broadcast and superposition nature of the wireless medium. Jamming is a common form of denial of service (DoS) attack which can significantly impact the IoT network's performance. In many IoT scenarios such as status update systems, users may not always have data to send; rather, data arrival at the users are random. It is also required to ensure timely delivery of data at the destination within a given period of time due to requirements in terms of delay, as well as it is required to keep the information at the destination as fresh as possible. For such scenarios, along with stable throughput, it is required to consider delay and age of information (AoI), which are meaningful metrics to take account of latency in the communication. In recent years, AoI has been used to capture the freshness of information \cite{AoINOW}. One of the concern arises is to ensure high throughput, low delay, and low AoI simultaneously under jamming attack, when the user also has energy constraint. The prospect of multiple antennas system has been explored for green communication \cite{cui-jsac-2004, nguyen-tgcn-2018}. \textit{The role of multiple antennas in mitigating the jamming attack and enhancing the system performance when timely updates are important is not well explored in the existing literature and the primary focus of this work.}
In this paper, a point-to-point MIMO system in the presence of a jammer is considered, where the transmitter is equipped with a queue to store the incoming traffic. It is assumed that jammer has energy harvesting capability \cite{8068986,8768226,9170580,7294641}. In many practical scenarios, it may not be possible for the jammer to have a constant source of energy supply due to the hostile nature of the environment, inaccessibility to the location, or absence of any external supply of energy such as a power grid. However, it is possible to deploy jammer in such environments due to the advancements made in energy harvesting. When the attacker has energy harvesting ability, the deployment of such jammer will be easier and it can make them autonomous. When it is required to guarantee low latency along with reliability, even a random jammer can cause significant harm to the performance of the system, and hence, it is important to understand the impact of jamming on the reliability and latency of the system. The considered model helps to capture the impact of various parameters of the attacker such as the capacity of the battery, jamming probability, and jamming power on the performance of the system. From the jammer's perspective, the considered model can help to explore how large should be the battery size and energy harvesting ability to perform jamming effectively. As a special case, the considered model reduces to the case of a constant jammer.
This work develops a cross-layer framework for the point-to-point MIMO system in the presence of a jammer. The cross-layer framework captures the random arrival of data at the transmitter using network-level metrics such as stability of the queue and reliability of the underlying channel model through outage probability. The outage probability takes into account of Signal to Noise Ratio (SNR) or Signal to Jamming and Noise power Ratio (SJNR) and fading phenomena of the wireless channel between the various nodes. Such frameworks have been used to study different communication models in the existing literature \cite{lu-TVT-2016, wang-tvt-2017}. The work also demonstrates how the developed results in the paper can be useful to characterize stability region for multi-user scenarios.
The paper first obtains the outage probabilities for different MIMO configurations under jamming attack. Then, the service probability at the transmitter is characterized using the outage probability obtained for different antenna configurations at the legitimate nodes. The service probability is used to characterize the average delay per packet, and the average AoI (AAoI) of various antenna configurations under random arrival of data at the legitimate transmitter. These metrics are characterized for two scenarios where the jammer can have a battery of \textit{finite capacity} or \textit{unlimited capacity}. The role of space-time coding on performance is also investigated. The derived results can also be used to minimize AAoI or average delay or maximize the average service rate for optimal data arrival rate. In the later part of the paper, it is shown how the characterization of the stable throughput of the point-to-point system helps to characterize the stability region for the 2-user SIMO broadcast channel (BC) in the presence of an energy harvesting jammer.
\vspace{-10pt}
\subsection{Related works}
The conventional security mechanisms generally do not provide protection or detection of jamming attacks in wireless networks. The jammer can adopt different strategies to carry the attack, and the various jamming mitigation techniques can be classified into the following categories: channel hopping, coding protection, rate-adaption, and MIMO-based jamming mitigation \cite{xu2005feasibility}. Channel hopping and spread spectrum are two commonly used techniques to mitigate the jamming attack. One of the disadvantages of the channel hopping-based method is the requirement of preshared channel assignment. Interference alignment which is used generally to mitigate interference has been used for MISO broadcast channels based on topology for mitigation of jamming attack \cite{amuru-tit-2014}. To understand the fundamental limits on the performance under jamming attacks, tools and techniques from information theory have been used for various communication models \cite{ kashyap-tit-2004, amuru-tit-2014}.
The attacker can also have the ability to eavesdropping as well as jamming \cite{8968374, 7470273, ryu2016transmission}. The problem of secure communication over a correlated fading channel is considered in \cite{7470273} for multi-user multi-cell massive MIMO in the presence of an active eavesdropper equipped with multiple antennas. The work shows that transmit antenna correlation diversity between the nodes can be exploited to improve the performance even under pilot contamination attacks. The work in \cite{ryu2016transmission} develops a transmission strategy that involves the determination of secure transmission rate and beam-forming design based on complete or partial knowledge of the jamming channel in case of MISO system. A friendly jammer can also be used to enhance the secrecy performance by sending a jamming signal to degrade the SNR at the eavesdropper \cite{8121977, 8351952}. In general, there is complete trust between the jammer and legitimate nodes in such scenarios.
Legitimate users always try to achieve the desired performance and on the other hand, attacker attempt to cause maximum harm to the communication. To capture this conflict arising between jammer and legitimate users, game theory has been used to study various models under jamming attacks \cite{kashyap-tit-2004, ulukus-mcc-2005, garnaev-twc-2016, mukherjee-tsp-2013, garnev-wifs-2014}. In \cite{kashyap-tit-2004}, using a zero-sum mutual information game it is shown that knowing the input signal at the jammer is not useful for the point-to-point MIMO Rayleigh fading channel. The work in \cite{ulukus-mcc-2005} considers a non-cooperative zero-sum game for 2-user multiple access channels in the presence of a jammer under different assumptions on channel characteristics. It is shown that when jammer does not know users' signals, the solution for the game exists. Another important problem is the allocation of resources in the presence of the attacker. The work in \cite{garnev-wifs-2014} uses the Bayesian game using an $\alpha$-fairness utility for resource allocation under an unknown jamming attack. The tool from game theory has also been used in the case of active eavesdropper where the attacker can either jam, eavesdrop on the ongoing communication, or both \cite{mukherjee-tsp-2013, garnaev-twc-2016}.
The works discussed so far assume that users always have data to send. However, in many practical scenarios, data arrival at the users can be random. The game-theoretic framework has been used to explore the role of random arrival of data in mitigating jamming attacks in \cite{sagduyu-isit-2010}. The framework of the non-cooperative game has been used in \cite{5518798} to determine the impact of jamming over a collision channel. For many applications, where a delay and timely updates are of importance, throughput alone cannot capture these attributes. For such scenarios, delay and AoI are relevant metrics to be considered, and these metrics can give new insights into system performance. AoI so far has been studied under different queuing disciplines in simple point-to-point systems, in more elaborated systems with multiple access, or scheduled access \cite{AoINOW,sun2019age}. The work in \cite{chen2020multiuser} studied the user scheduling problem in a MIMO status update system, where multiple users with a single antenna aim to send their latest updates to an information-fusion access point equipped with multiple antennas via a shared wireless channel. The authors derived an optimal scheduling policy that can minimize the AoI over the networks. Furthermore, the effect of jamming on AoI has been studied in \cite{SunAoIW2018, GarnaevAoIW2019}. Characterizing the stability region for multiuser scenarios is a challenging problem due to the interaction between the queues \cite{8320826, 8357571}. However, the impact of jamming on the achievable throughput, delay, and AoI in a system with random traffic is not well understood in fading scenarios when users are equipped with multiple antennas. This also brings an important question on how diversity (time, spatial, or both) can be exploited to improve the performance of the system in the presence of a jammer. This work primarily aims to answer these questions.
\vspace{-15pt}
\subsection{Contributions}
The main contributions of the work are summarized below.
\begin{enumerate} \itemsep 0.2em
\item The outage probabilities are obtained for the Rayleigh fading environment in the presence of jammer under the different assumptions on the number of antennas at the transmitter and receiver. The outage probability for Alamouti coding scheme is also derived for the considered model. As the signal sent by the transmitter and attacker undergo Rayleigh fading, it is a non-trivial problem to determine the distribution corresponding to the signal to jamming and noise ratio (SJNR). To the best of the authors' knowledge, the role of space-time code in mitigating jamming attack has not been explored in the existing literature.
\item Characterization of stable throughput or stability region under random arrival of data in the presence of jammer is a non-trivial problem. The paper first aims to characterize the stable throughput under jamming attack for the point-to- point MIMO system. The average service rate is characterized for the considered system model using the outage probabilities obtained for different multiple antennas setup. To capture the energy harvesting ability of the jammer two scenarios are considered: \textit{battery with unlimited capacity} and \textit{battery with limited capacity}. The developed results take account of the random arrival of data at the transmitter and energy at the attacker.
\item The work also characterizes the average delay performance and AoI for various multi-antenna configurations in presence of the attacker. The developed results help to explore the interplay between delay and AAoI for different antenna configurations under jamming attack and developed results reconfirm the utility of the AAoI metric for latency-aware communication. Two optimization problems are considered where the objective is to minimize the AAoI with different delay requirements.
\item The work also characterizes the stability region of the 2-user SIMO broadcast channel under jamming attack for the Rayleigh fading scenario using the results obtained for the point-to-point model. In this case, receiver decodes its intended message by treating other user's signal as noise.
\end{enumerate}
The obtained results in this work provide a unified view of the role of multiple antennas in improving the performance of the system in terms of stable throughput, delay, and AAoI under jamming attack. It is also found that Alamouti coding can achieve minimum delay and AoI among the different MIMO configurations in the setup with less power budget. The Alamouti coding scheme has the added advantage of not requiring channel state information at the transmitter. The developed results for the point to point MIMO system can act as a basic building block to characterize the stability region of multiuser scenarios, which is a challenging problem due to the interaction among the queues.
\vspace{-3pt}
\section{System Model}
\vspace{-10pt}
\begin{figure}[h]
\centering
\includegraphics[trim={0cm 0cm 0cm 0cm},clip, height=1.8in, width=3.35in]{Stability_work_System_model1_new.eps}
\setlength{\belowcaptionskip}{-6pt} %
\caption{Point-to-point MIMO system in the presence of energy harvesting jammer.}\label{fig:figure1}
\end{figure}
This paper considers a point-to-point MIMO system in the presence of a jammer, where transmitter and receiver are equipped with $n_t$ and $n_r$ antennas, respectively. The transmitter has a queue, which stores the incoming packet, and it needs to be sent reliably to the receiver in the presence of the attacker. A pictorial description of the model is shown in Fig~\ref{fig:figure1}.
\vspace{-10pt}
\subsection{Network layer model}
The packet arrival process at the queue is assumed to be independent and stationary with mean probability $ \lambda$, i.e., Bernoulli \cite{9233374}. It is also assumed that the queue at the transmitter has unlimited capacity. The average length of the queue is denoted by $\bar{Q}$ and serves the packets with average service rate $\mu$. It is assumed that the jammer does not have a constant source of power supply, but it has energy harvesting ability. The energy arrival process at the jammer is modeled by a Bernoulli process where the rate of energy arrival is assumed to be $\delta$
\cite{9365698, 7959595, 9086254,7487983, 7009995}. These chunks are stored in a battery (an energy buffer). It is assumed that if one chunk of energy is harvested at the jammer, it is enough to jam the communication with fixed jamming power $P_J$. In the paper, two scenarios are considered for the energy buffer: (a) \textit{Energy buffer with unlimited capacity (Section~IV-A)}; and (b) \textit{Energy buffer with limited capacity (Section~IV B)}.
\vspace{-10pt}
\subsection{Physical layer model}
The input-output relation for the considered system model in the presence of jammer is given as follows:
\begin{align}
& \mathbf{y} = \mathbf{H} \mathbf{x} + \mathbf{h_J} x_J + \mathbf{n}, \label{eq:sysmodel1}
\vspace{-10pt}
\end{align}
where the channel between transmitter and receiver is denoted by $\mathbf{H} \in \mathcal{C}^{n_r \times n_t}$ and channel between the jammer and receiver is denoted by $\mathbf{{h}_J} \in \mathcal{C}^{n_r \times 1}$. The input $\mathbf{x}$ to the channel is drawn from a Gaussian codebook, i.e., $\mathbf{x} \sim \mathcal{CN}(0, P \mathbf{I})$, and P is the transmitter's power budget. The attacker sends the jamming signal $x_J$ drawn from Gaussian distribution with jamming power $P_J$, i.e., $x_J \sim \mathcal{CN}(0, P_J)$. Due to the limited computing ability and constrained in the energy of the attacker, it is assumed that the jammer has a single antenna. The noise at the receiver is modeled as AWGN, i.e., $\mathbf{n} \sim \mathcal{CN}(\mathbf{0}, \mathbf{I})$. In the case of MISO, $\mathbf{H} \in \mathcal{C}^{1 \times n_t}$, and $\mathbf{n} \sim \mathcal{CN}(0, 1)$. For SIMO, $\mathbf{H} \in \mathcal{C}^{n_r \times 1}$ and input is drawn from a Gaussian codebook, i.e., $\mathbf{x} \sim \mathcal{CN}(0, P)$. The channel between the various nodes undergoes an independent Rayleigh fading process. The attacker does not have CSI and it cannot access the queue state information of the transmitter. The transmitter and receiver need to know the statistical knowledge of the CSI between the jammer and receiver. When the battery is not empty, the attacker jams with a probability $p_J$. This type of jammer is known as a random jammer in the literature and belongs to the category of proactive jammer~\cite{10.1504/IJAHUC.2014.066419}.
\vspace{-10pt}
\subsection{Stability of the queue and average service rate}
The stability of the queue is defined as follows \cite{szpankowski_1994}.
\begin{definition}
The queue is said to be stable if the following condition is satisfied $\displaystyle\lim_{t \to \infty} Pr[Q^t<x] = F(x)$ and $\displaystyle\lim_{x \to \infty} F(x) = 1$, where $Q^t$ denotes the length of queue at the beginning of the time slot $t$. When the following condition is satisfied $\lim_{x \to \infty}\lim_{t \to \infty}\text{inf } Pr[Q^t<x]= 1$, the queue is called sub-stable. If a queue is stable, then it is also sub-stable. The queue is said to be unstable, if the queue is not sub-stable.
\end{definition}
Hence, the queue is stable if the average data arrival rate ($\lambda$) is less than the average service rate ($\mu$), i.e., $\lambda < \mu$. This result holds under the assumption that the queue's arrival and service processes are strictly jointly stationary. For the system model considered in this paper, the average service rate is defined as follows:
\vspace{-10pt}
\begin{align}
& \mu = \hspace{1mm} (1-p_J)\hspace{1mm}(1-p_{WoJ}^{\text{out}}) + \hspace{1mm} p_J \hspace{1mm} (1-p_{J}^{\text{out}}), \label{eq:servicerate1}
\end{align}
where $p_{J}^{\text{out}}$ and $p_{WoJ}^{\text{out}}$ are the probability of unsuccessful decoding of the packet at the receiver with and without jamming attack, respectively. To characterize the service rate, it is required to obtain these probabilities. In this work, the outage probability is used as a proxy for the probability of unsuccessful decoding of data at the receiver \cite{lu-TVT-2016, wang-tvt-2017}. To evaluate the service rate, legitimate nodes need to know the jamming power $P_J$ and jamming probability $p_J$. In practice, estimating these parameters is a non-trivial problem and some of the works in this direction can be found in \cite{6552344, article, article2, 6179344, 9039615}. Tools from machine learning can also be useful in estimating these parameters \cite{8979256}.
\vspace{-6pt}
\section{Outage Probability for different antenna configurations under Jamming}\label{sec:outage}
In this section, the outage probabilities are obtained for different multi-antenna configurations for the Rayleigh fading scenario under a jamming attack. The work also derives the outage probability for the MIMO with Alamouti coding. The role of space-time diversity in mitigating jamming attacks is not well understood from the existing literature. The developed results help to explore the role of multiple antennas in mitigating jamming attacks. The derivation of these expressions is non-trivial as it is required to obtain the distribution associated with signal to jamming and noise ratio (SJNR) for multiple antenna configurations, where the message signal and the jamming signal go through the Rayleigh fading channel. These results are stated in the following theorem.
\begin{theorem}\label{th:firsttheorem}
The outage probabilities $(p_J^{\text{out}})$ for different antenna configurations under jamming attack are provided below.
\begin{enumerate}
\item The outage probability for MISO $(1 \times n_t)$ is given by the following expression:
\begin{align}
p_J^{\text{out}} &= 1 - \frac{\Gamma(n_t,\frac{(2^R-1)n_t}{P})}{\Gamma(n_t)} + \frac{e^{\frac{1}{P_J}}\Gamma(n_t,\frac{(2^R-1)n_t}{P}+\frac{1}{P_J})}{\Gamma(n_t)(1+\frac{P}{n_t(2^R-1)P_J})^{n_{t}}}. \label{eq:th1eq1}
\end{align}
\item The outage probability for SIMO $(n_r \times 1)$ is given by the following expression
\begin{align}
p_{J}^{\text{out}} & = 1-\frac{\Gamma(n_r,\frac{(2^R-1)}{P})}{\Gamma(n_r)} + \frac{e^{\frac{1}{P_J}}\Gamma(n_r,\frac{(2^R-1)}{P}+\frac{1}{P_J})}{\Gamma(n_r)(1+\frac{P}{(2^R-1)P_J})^{n_{r}}}. \label{eq:th1eq2}
\end{align}
\item The outage probability for MIMO $(n_r \times n_t)$ with Alamouti code is given by the following expression
\begin{align}
p_{J}^{\text{out}} & = 1- \frac{\Gamma(N, \frac{(2^R-1)}{\beta})}{\Gamma(N)} + \dfrac{e^{\frac{1}{P_J}} \Gamma\bigg(N,\big( \frac{2^R-1}{\beta} + \frac{1}{P_J}\big)\bigg)}{\beta^N \Gamma(N) \big[ \frac{1}{\beta} + \frac{1}{(2^R-1)P_J}\big]^N}. \label{eq:th1eq3}
\end{align}
\end{enumerate}
where $R$ is the given target rate, $\beta \triangleq \frac{P}{n_t}$, $N \triangleq n_tn_r$, $\Gamma(s)$ and $\Gamma(s,x)$ are the gamma function and the upper incomplete gamma function defined as follows, respectively
\begin{align}
\Gamma(s)=\int\limits_0^\infty t^{s-1}e^{-t}dt \text{ and } \Gamma(s,x)=\int\limits_x^\infty t^{s-1}e^{-t}dt.
\end{align}
\end{theorem}
\begin{proof} The proofs for different multi-antenna scenarios are given below:\\
\textit{Case 1 (MISO $(1 \times n_t)$):} In this case, the instantaneous achievable rate is given by the following expression
\begin{equation}
\begin{aligned}
R_i = \log_2 \left( 1+\frac{|\mathbf{h}|^2\frac{P}{n_t}}{1+|{h_J}|^2 P_J} \right), \label{eq:misoach1}
\end{aligned}
\end{equation}
where $|\mathbf{h}|^2$ and $|{h_J}|^2$ are chi-square and exponentially distributed, respectively. Due to the lack of CSI at the transmitter, power is equally divided among all the transmit antennas. The outage probability is determined as follows
\begin{equation}
\begin{split}
p_{J}^{\text{out}} & = Pr\{ R_i <R \}, \\
&= Pr\left\{ \log_2 \left( 1+\frac{|\mathbf{h}|^2\frac{P}{n_t}}{1+|{h_J}|^2 P_J} \right) <R \right\},\\
&= Pr\left\{ \frac{|\mathbf{h}|^2P}{n_t} - (2^R-1)|{h_J}|^2 P_J <2^R-1 \right\}. \nonumber \\
\end{split}
\end{equation}
where $Pr(.)$ is the probability.
To simplify the notation, substitute $|\mathbf{h}|^2$ as U i.e., $U \sim \chi^2_{2n_{t}} $. $|{h_J}|^2$ as W i.e., $W \sim e^{-w}$ in the above expression and the outage probability becomes
\begin{equation}
\begin{split}
p_{J}^{\text{out}} & = 1- Pr\bigg\{ W < \frac{U P/n_t-(2^R-1)}{(2^R-1)P_J} \bigg\}, \\
&= 1-\int\limits_{u=(2^R-1)n_t/P}^{\infty} \int\limits_{w=0}^{\frac{uP/n_t-(2^R-1)}{(2^R-1)P_J}} f_{W}(w) f_{U}(u) \: dw \:du,\nonumber \label{use_SuccPr}
\end{split}
\end{equation}
\begin{equation}
\begin{split}
&= 1-\int\limits_{u=(2^R-1)n_t/P}^{\infty} \!\!\!\!\! f_{U}(u) \: \bigg [ \int\limits_{w=0}^{\frac{uP/n_t-(2^R-1)}{(2^R-1)P_J}} f_{W}(w) dw \bigg] du, \\
&= 1-\!\!\!\!\!\int\limits_{u=(2^R-1)n_t/P}^{\infty}\!\!\!\!\!\!\!\!\!f_{U}(u)\ du
+ e^{\frac{1}{P_J}}\!\!\!\!\!\!\!\!\!\int\limits_{u=(2^R-1)n_t/P}^{\infty} \!\!\!\!\!\!\!\! f_{U}(u)e^{\big(\frac{-Pu}{n_t(2^R-1)P_J}\big)} du, \label{eq:misopout}
\end{split}
\end{equation}
where $f_{U}(u)$ and $f_{W}(w)$ are the probability density functions of U and W, respectively.
Solving both the integrals, the outage probability reduces to following
\begin{equation}
\begin{aligned}
p_J^{\text{out}} = 1-\frac{\Gamma(n_t,\frac{(2^R-1)n_t}{P})}{\Gamma(n_t)} + \frac{e^{\frac{1}{P_J}}\Gamma(n_t,\frac{(2^R-1)n_t}{P}+\frac{1}{P_J})}{\Gamma(n_t)(1+\frac{P}{n_t(2^R-1)P_J})^{n_t}}. \label{eq:misoach2}
\end{aligned}
\end{equation}
\textit{Case 2 (SIMO $(n_r \times 1)$):}
In this case, the input-output relation is given by the following expression
\begin{equation}
\begin{aligned}
\mathbf{y}= \mathbf{h}x+\mathbf{h_J} {x_J}+\textbf{n}
\end{aligned}
\end{equation}
As the transmitter does not have CSI, it allocates all the power to the message signal. The instantaneous achievable rate for this case is
\begin{equation}
\begin{aligned}
R_i= \log_2 \bigg( 1+\frac{|\mathbf{h}|^2P}{1+|{h_J}|^2 P_J} \bigg).
\end{aligned}
\end{equation}
In the above, $ |\mathbf{h}|^2 = h_1^2 + h_2^2 + \cdots h_{n_{r}}^2 $ follows chi-squared distribution with $2{n_{r}}$ degrees of freedom, i.e. $|\mathbf{h}|^2 \sim \chi^2_{2n_{r}} $. The outage probability is obtained as follows:
\begin{equation}
\begin{split}
p_{J}^{\text{out}} &= Pr\{ R_i < R \}, \\
&= Pr\big\{\log_2 \Big( 1+\frac{|\mathbf{h}|^2P}{1+|{h_J}|^2 P_J} \Big)<R \big\}, \\
&= Pr\big\{|\mathbf{h}|^2P - (2^R-1)|{h_J}|^2 P_J <2^R-1 \big\} .
\end{split}
\end{equation}
Following similar approach as in (\ref{eq:misopout}) (previous case), the outage probability expression becomes
\begin{equation}
\begin{aligned}
p_{J}^{\text{out}} = 1-\frac{\Gamma(n_{r},\frac{(2^R-1)}{P})}{\Gamma(n_{r})} + \frac{e^{\frac{1}{P_J}}\Gamma(n_{r},\frac{(2^R-1)}{P}+\frac{1}{P_J})}{\Gamma(n_{r})(1+\frac{P}{(2^R-1)P_J})^{n_{r}}}. \label{eq:simoach2}
\end{aligned}
\end{equation}
\textit{Case 3 (MIMO $(n_r \times n_t)$):}
The instantaneous achievable rate for a MIMO fading channel using Alamouti coding \cite{sandhu2000space,tse2004fundamentals} in the presence of jammer is given by
\begin{equation}\label{MIMO_ala_cap}
R_i = \frac{K}{T}\log_2\Big(1+\frac{\frac{P}{n_t}||\mathbf{H}||_F^2}{1+|{h_J}|^2P_J} \Big).
\end{equation}
To simplify the notations, define the following quantities $ V \triangleq \frac{P}{n_t} ||\mathbf{H}||_F^2 $ and $S \triangleq ~|{h_J}|^2$. Note that $V$ is said to have the gamma distribution with $\Omega=N=n_tn_{r}$ and $\beta=\frac{P}{n_t}$, i.e. $V \sim G(\Omega, \beta)$, and S is exponentially distributed $\sim e^{-s}$.
The quantity $\frac{K}{T}$ denotes the rate of the space-time block coding (STBC), where $K$ represents the number of symbols transmitted, and $T$ denotes the number of time slots used for transmission. The rate of the STBC is considered to be $1$ in this case. However, we can extend the result to any rate of the Alamouti coding scheme. The outage probability of a MIMO system with Alamouti coding at the transmitter in the presence of jammer is given as follows
\begin{equation}
\begin{split}
p_{J}^{\text{out}} & =Pr( R_i < R),\nonumber\\
&= Pr \Big( \log_2\Big(1+\frac{V}{1+{SP_J}}\Big)<R \Big), \nonumber \\
\end{split}
\end{equation}
\begin{equation}
\begin{split}
&= Pr\Big(\frac{V-(2^R-1)}{P_J(2^R-1)}< S\Big),\nonumber \\
&= 1- Pr\Big(S <\frac{V-(2^R-1)}{P_J(2^R-1)}\Big),\nonumber \\
\end{split}
\end{equation}
\begin{equation}
\begin{split}
&= 1- \int_{v=(2^R-1)}^\infty \int_{s=0}^{\frac{v-(2^R-1)}{P_J(2^R-1)}} f_V(v) f_S(s) ds dv. \label{pout_mimo_ala_jam}
\end{split}
\end{equation}
Note that
\begin{equation}\label{pdf_gamma}
f_V(v)= \dfrac{v^{(N-1)}e^{\frac{-v}{\beta}}}{\Gamma(N) \beta^N}~ \text{and}~ f_S(s)=e^{-s}.
\end{equation}
Substituting \eqref{pdf_gamma} in \eqref{pout_mimo_ala_jam} and simplifying further, results in the following expression
\begin{equation}
p_{J}^{\text{out}} = 1- \frac{\Gamma(N, \frac{(2^R-1)}{\beta})}{\Gamma(N)} + \dfrac{e^{\frac{1}{P_J}} \Gamma\bigg(N,\big( \frac{2^R-1}{\beta} + \frac{1}{P_J}\big)\bigg)}{\beta^N \Gamma(N) \big[ \frac{1}{\beta} + \frac{1}{(2^R-1)P_J}\big]^N}.\label{eq:mimoach2}
\end{equation}\vspace{-8pt}
\end{proof}
\vspace{-20pt}
\subsection{Asymptotic Analysis}
To obtain further insights on the role of antennas on the outage probability in the presence of jammer, the power at the transmitter and jammer are driven to infinity but their ratio is held constant, i.e., $\frac{P_J}{P} = \eta$. It is assumed that $\eta \geq 1$ and this condition ensures that when $P \rightarrow \infty$, then $P_J \rightarrow \infty$. The asymptotic expressions for different MIMO configurations are as follows:
\begin{align}
& \lim_{P,P_J\longrightarrow \infty} p_J^{\text{out}} = \frac{1}{(1+\frac{1}{\eta n_t(2^R-1)})^{n_{t}}}, \label{eq:miso_asym_wj} \qquad \text{(MISO)}\\
& \lim_{P,P_J\longrightarrow \infty}p_{J}^{\text{out}} =
\frac{1}{(1+\frac{1}{\eta(2^R-1)})^{n_{r}}}. \label{eq:simo_asym_wj} \qquad \text{(SIMO)} \\
\text{and }& \lim_{P,P_J\longrightarrow \infty}p_{J}^{\text{out}} = \dfrac{1}{\big[ 1 + \frac{1}{\eta n_t(2^R-1)}\big]^N}. \label{eq:mimo_asym_wj} \qquad \text{(MIMO with Alamouti)}
\end{align}
\textbf{Remarks:}
\begin{enumerate}
\item As a special case, the outage probability without jamming for different MIMO configurations can be obtained by setting $P_J =0$ using the results in Theorem~\ref{th:firsttheorem}.
\item When $\eta$ is unbounded in (\ref{eq:miso_asym_wj}), (\ref{eq:simo_asym_wj}), and (\ref{eq:mimo_asym_wj}), the outage probability becomes $1$ for all the cases, irrespective of the number of antennas at the transmitter, receiver, or both.
\item As a special case we can obtain the outage probability for the SISO with jammer by substituting $n_t = 1$ or $n_r=1$ in the result for MISO or SIMO in \eqref{eq:th1eq1} or \eqref{eq:th1eq2}, respectively. By setting $n_r= 1$ in Theorem~\ref{th:firsttheorem}, one can obtain the outage probability for the Alamouti space-time code in case of the MISO under a jamming attack.
\item The developed results in Theorem~\ref{th:firsttheorem} and (\ref{eq:miso_asym_wj})-(\ref{eq:mimo_asym_wj}) can be useful to study the benefits of multiple antenna techniques in achieving desired performance with less power budget under jamming attack. In particular, it is found that the MIMO system with Alamouti coding can provide superior performance at less power budget in comparison to other antenna configurations such as MISO (See Figs~\ref{subfig-1:figure3} and \ref{subfig-1:figure5} in Sec~\ref{sec:result}).
\end{enumerate}
\section{Stability Analysis for Multi-antenna System under Jamming Attack}
In this section, the average service rate of the point-to-point MIMO system with random arrival of data at the transmitter under jamming attack is derived. As the attacker has energy harvesting ability, the impact of jamming on the service rate is affected by the energy arrival rate at the jammer and the capacity of the battery. The results are derived for the battery with unlimited and limited capacity at the jammer. The outage probabilities derived in the previous section are used for determining the successful decoding of the data at the receiver.
\vspace{-8pt}
\subsection{Battery with unlimited capacity}
To determine the average service rate in this case, it is required to model how the state of the energy buffer at the jammer evolves. The evolution of the energy buffer can be described by the Markov chain, as shown in Fig.~\ref{fig:markov-chain}. The evolution of the Markov chain is characterized by the probability of jamming $(p_J)$ and the energy arrival rate at the jammer $(\delta)$. To determine the average service rate with infinite energy buffer at the jammer, the following cases are considered: (a) probability of jamming is more than the energy arrival rate at the jammer, and (b) probability of jamming is less than the energy arrival rate at the jammer. The reason for considering these two cases is explained below. In the first case, the system is limited by its harvested energy; thus, it is not always available to create interference, even if needed. In the second case, since the jamming probability is less than the energy arrival rate, the jammer is operating as if it was connected to the power grid without energy limitations.
The average service rate with infinite energy buffer at the jammer is stated in the following theorem.
\vspace{-7pt}
\begin{figure}[h]
\centering
\includegraphics[trim={0cm 0cm 0cm 0cm},clip, height=.7 in,width=3.4 in]{EH_MC_infinite_battery.eps}
\setlength{\abovecaptionskip}{-1pt}
\setlength{\belowcaptionskip}{-12pt} %
\caption{State transition diagram of infinite capacity energy buffer at jammer, where $\zeta=(1-\delta)p_J$, and $\alpha=\delta(1-p_J)$.} \label{fig:markov-chain}
\end{figure}
\begin{theorem}\label{th:secondtheorem}
The average service rate of the system when the jammer has infinite energy buffer is as follows:
\begin{enumerate}
\item When $p_J \geq \delta$
\begin{equation}
\mu = \hspace{1mm} \bigg(1- \frac{(1-p_J)\delta}{(1-\delta )}\bigg)(1 - p_{WoJ}^{\text{out}}) + \bigg(\frac{(1-p_J)\delta}{(1-\delta )}\bigg) (1- p_J^{\text{out}}).\label{eq:th2eq1}
\end{equation}
\item When $p_J < \delta$
\begin{equation}
\mu = (1-p_J)(1 - p_{WoJ}^{\text{out}}) + \hspace{1mm} p_J(1- p_J^{\text{out}}). \label{eq:th2eq2}
\end{equation}
\end{enumerate}
The average service rate for MISO, SIMO, and MIMO with Alamouti coding can be obtained by using the results on outage probability from Theorem~\ref{th:firsttheorem}. When $\lambda < \mu$, the system is said to be stable.
\end{theorem}
\begin{proof}
To determine the average service rate in the presence of an energy harvesting jammer, the following cases are considered. \\
\textit{Case 1 (When jammer is sending energy packets at a rate ($ p_J$) greater than its arrival rate ($\delta $)): }
If $\delta \leq p_J$, then the jammer may not always have sufficient energy to disrupt the communication, and the jammer's energy buffer can be empty. To characterize the service rate, it is required to determine the probability that the energy buffer is empty, i.e., $Pr(B=0)$, where $B$ denotes the size of the energy buffer.
From Fig.~\ref{fig:markov-chain}, one can see that the steady-state distribution need to satisfy the following
\begin{equation}
\begin{split}
& \zeta \pi_{i+1} = \alpha \pi_{i} ~~~~\text{for} ~~ 0 \leq i \leq B-1. \nonumber \\
\end{split}
\end{equation}
Normalizing the probabilities, following is obtained
\begin{equation}
\begin{split}
\pi_{0} \sum_{i=0}^{\infty} \big(\frac{\alpha}{\zeta}\big)^i =1,~
\text{or } \pi_{0} = 1- \frac{\alpha}{\zeta},
\end{split}
\end{equation}
where $\pi_{i}$ is the probability of being in the $i^{th}$ state. Therefore, probability that the jammer energy buffer can be empty is given by
\begin{equation}
\pi_{0}=Pr(B=0) = 1- \frac{(1-p_J)\delta}{(1-\delta )p_J}=1-\frac{\alpha}{\zeta}.
\end{equation}
Hence, the probability that energy buffer is not empty is given by the following
\begin{equation}
Pr(B\neq0) = \frac{(1-p_J)\delta}{(1-\delta )p_J} =\frac{\alpha}{\zeta}.
\end{equation}
Using the above, the service rate of the system becomes
\begin{equation}
\begin{split}
&\mu= (1-p_J Pr(B\neq0)) (1 - p_{WoJ}^{\text{out}}) ~ + ~ p_J Pr(B\neq0) (1- p_J^{\text{out}}), \\
&= \bigg(1- \frac{(1-p_J)\delta}{(1-\delta )}\bigg)(1 - p_{WoJ}^{\text{out}}) ~ + ~ \bigg(\frac{(1-p_J)\delta}{(1-\delta )}\bigg) (1- p_J^{\text{out}}).\\
&= \bigg(1- \frac{\alpha}{(1-\delta )}\bigg)(1 - p_{WoJ}^{\text{out}}) ~ + ~ \bigg(\frac{\alpha}{(1-\delta )}\bigg) (1- p_J^{\text{out}}).
\end{split}\label{eq:serratecase-2}
\end{equation}
Note that the outage probability without jamming $(p_{WoJ}^{\text{out}})$ can be obtained by setting $P_J=0$ for different antenna configurations in Theorem~\ref{th:firsttheorem}.\\
\textit{Case 2 (When jammer is sending energy packets at a rate ($ p_J$) less than its arrival rate ($\delta $)):} If $\delta \geq p_J$, then the jammer always have energy to send the jamming signal and the probability that the energy buffer is empty is $0$, i.e., $Pr(B=0)=0$. Hence, the service rate of the system can be written as:
\vspace{-8pt}
\begin{align}
& \mu = \hspace{1mm}(1-p_J)\hspace{1mm}(1 - p_{WoJ}^{\text{out}}) + \hspace{1mm} p_J(1- p_J^{\text{out}}). \label{eq:serratecase-1}
\end{align}\vspace{-20pt}
\end{proof} \vspace{-12pt}
\subsection{Battery with Finite Capacity}
Another important factor that determines the efficiency of the jammer is the capacity of the energy buffer. This section characterizes the service rate in the case of the finite energy buffer. Similar cases are considered as that in the case of a battery with unlimited capacity. Note that even when the energy arrival rate can be high at the jammer but limited battery capacity can hinder the jamming ability of the attacker. The evolution of the energy buffer of size $B$ at the jammer can be described by the Markov chain, as shown in Fig.~\ref{fig:markov-chain_finite}. The evolution of the Markov chain is characterized by the probability of jamming and the energy arrival rate at the attacker.
\begin{figure}[h]
\centering
\includegraphics[trim={0cm 0cm 0cm 0cm},clip, height=.9 in,width=3.4 in]{EH_MC_finite_battery.eps}
\setlength{\abovecaptionskip}{-6pt}
\setlength{\belowcaptionskip}{-15pt} %
\caption{State transition diagram of finite energy buffer at the jammer, where $\zeta=(1-\delta)p_J$, and $\alpha=\delta(1-p_J)$.}\label{fig:markov-chain_finite}
\end{figure}
The average service rate with finite energy battery at the jammer is stated in the following theorem.
\begin{theorem}\label{th:thirdtheorem}
The average service rate of the system when the jammer has finite energy buffer is as follows:
\begin{enumerate}
\item When $p_J \geq \delta$
\begin{align}
&\mu = \bigg(1- p_J\bigg(1-\frac{(p_J-\delta){\zeta}^B}{\zeta^{B+1}-\alpha^{B+1}}\bigg)\bigg)(1 - p_{WoJ}^{\text{out}}) +~p_J\bigg(1-\frac{(p_J-\delta){\zeta}^B}{\zeta^{B+1}-\alpha^{B+1}}\bigg) (1- p_J^{\text{out}}).\label{eq:th3eq1}
\end{align}
\item When $p_J < \delta$
\begin{align}
& \mu = (1-p_J)(1 - p_{WoJ}^{\text{out}}) + p_J(1- p_J^{\text{out}}). \label{eq:th3eq2}
\end{align}
\end{enumerate}
Using the results on outage probability from Theorem~\ref{th:firsttheorem}, the average service rate for MISO, SIMO, and MIMO with Alamouti coding can be obtained. For $\lambda <\mu$, the system is stable.
\end{theorem}
\begin{proof}
To determine the average service rate in the presence of an energy harvesting jammer with finite battery, the following cases are considered. \\
\textit{Case 1 (When jammer is sending energy packets at a rate ($ p_J$) greater than its arrival rate ($\delta $)): }
If $\delta \leq p_J$, then the jammer may not have sufficient energy to disrupt the ongoing communication as its energy buffer can be empty. In this case, the probability that energy buffer at the jammer is empty depends also on the capacity of the energy buffer. This probability is obtained by solving the balance equation of the Markov chain for finite battery and is given by
\begin{equation}
Pr(B=0) = \frac{(p_J-\delta){\zeta}^B}{\zeta^{B+1}-\alpha^{B+1}}.
\end{equation}
Hence, the probability that jammer energy buffer is not empty is
\begin{equation}
Pr(B\neq0) = 1-\frac{(p_J-\delta){\zeta}^B}{\zeta^{B+1}-\alpha^{B+1}}.
\end{equation}
Then, the service rate of the system is
\begin{align}
& \mu = \bigg(1- p_J\bigg(1-\frac{(p_J-\delta){\zeta}^B}{\zeta^{B+1}-\alpha^{B+1}}\bigg)\bigg)(1 - p_{WoJ}^{\text{out}}) +~p_J\bigg(1-\frac{(p_J-\delta){\zeta}^B}{\zeta^{B+1}-\alpha^{B+1}}\bigg) (1- p_J^{\text{out}}).
\end{align}
\textit{Case 2 (When jammer is sending energy packets at a rate ($ p_J$) less than its arrival rate ($\delta $)):} If $\delta > p_J$, then the jammer always have energy to send the jamming signal, i.e., $Pr(B=0)=0$.
Hence, the service rate of the system can be written as given in ( \ref{eq:th3eq2}).
\end{proof}
\textit{\textbf{Remarks:}}
\begin{enumerate}
\item When $\delta > p_J$, it is observed from \eqref{eq:th2eq2} and \eqref{eq:th3eq2} that the performance of battery with finite capacity and infinite capacity are same. In this case, having a battery with infinite capacity does not help the jammer.
\item As a special case, one can obtain the service rate for different MIMO configurations without jamming by substituting $p_J = 0$ either in a finite or infinite battery capacity case. To the best of the authors' knowledge, the service rate of a MIMO system with an Alamouti coding scheme has not been characterized in existing results even without jamming.
\item When $\delta = 1$ and $p_J = 1$, one can obtain the service rate of the considered system model under jamming attack, where the jammer always disrupts the communication.
\end{enumerate}
\vspace{-6pt}
\section{Delay and AAoI Analysis under Jamming Attack}
For the considered system model with bursty traffic, besides service rate, the delay is another important metric when the transmitter has time-sensitive data. Furthermore, AoI, which is a more general form of latency, is also important since, in status updating systems (common in IoT or cyber-physical systems), the freshness of information is crucial. As jamming is one of the common DoS (Denial of Service) attacks, it is important to evaluate these metrics in such cases. It is desirable to have a low delay and small AoI. The AoI captures the freshness in information, whereas the delay captures the queuing delay and transmission delay between the transmitter and the receiver. In the following, average delay and average AoI (AAoI) are characterized for different MIMO configurations. The developed results help to investigate the interplay between delay and AoI for multi-antenna configurations when the attacker has a battery with finite or infinite capacity. To the best of the authors' knowledge, the role of multiple antennas in minimizing delay and AAoI under jamming attack has not been explored in the existing literature.
\vspace{-10pt}
\subsection{Average Delay}
The delay consists of two components, the queuing delay and the transmission delay from the transmitter to the receiver. The average transmission delay $(D_T)$ is inversely proportional to the average service probability and is given by
\begin{equation}
D_T= \frac{1}{\mu}. \label{eq:delay1]}
\end{equation}
To determine the queuing delay, it is required to use the average queue length which is given by
\begin{equation}
\begin{split}
\bar{Q} & = \frac{\lambda (1-\lambda)}{\mu - \lambda}. \label{eq:delay2}
\end{split}
\end{equation}
Using Little's theorem, the queuing delay $(D_Q)$ is expressed as follows
\begin{equation}
D_Q= \frac{\bar{Q}}{\lambda}. \label{eq:delay3}
\end{equation}
The average packet delay (D) using (\ref{eq:delay2}) and (\ref{eq:delay3}) becomes
\begin{equation}
\begin{split}
D & = D_T + D_Q = \frac{1}{\mu } +\frac{1-\lambda}{\mu - \lambda}. \label{eq:average_delay}
\end{split}
\end{equation}
One can obtain the average packet delay when the jammer has infinite and finite battery capacity using the service rate expressions from Theorems~\ref{th:secondtheorem} and \ref{th:thirdtheorem} in \eqref{eq:average_delay}, respectively.
\vspace{-10pt}
\subsection{Average Age of Information (AAoI)}
The AAoI metric captures the freshness of information about a source at a remote destination. The objective here is to understand the effect of the multiple antennas at the transmitter and receiver on the AAoI under jamming attack. The considered queuing model in this paper is a Geo/Geo/1 queue, and the AAoI is given by
\begin{equation}\label{AAoI}
\begin{split}
AAoI & =\frac{1}{\lambda}+\frac{1 - \lambda}{\mu-\lambda}- \frac{\lambda}{{\mu}^2}+\frac{\lambda}{\mu}.
\end{split}
\end{equation}
The proof for the previous expression can be found in \cite{8764468}. One can obtain the AAoI when the attacker has infinite and finite battery capacity using the service rate expressions from Theorems~\ref{th:secondtheorem} and \ref{th:thirdtheorem} in \eqref{AAoI}, respectively. We would like to emphasize that the purpose here is to study the effect of multiple antennas, jamming attacks, and the Alamouti coding scheme on the AAoI and not to derive the metric from scratch. Furthermore, the results obtained in this work can be utilized in other models for AAoI, for example, at the generate at-will policy \cite{8815559} or in different queuing disciplines with or without packet management \cite{8764468}.
\vspace{-10pt}
\subsection{Minimization of AAoI}
In many applications, where the transmitter has time-critical data, it is required to minimize average delay and AAoI. It is desirable to support a high arrival rate even under a jamming attack. However, Fig~\ref{subfig-1:figure10} shows that a low or high arrival rate can increase the AAoI at the receiver. Furthermore, this is in contrast to the behavior of the delay against arrival rate in Fig~\ref{subfig-1:figure9}. Minimizing delay does not necessarily imply minimization of AAoI and vice-versa. This motivates to minimize the AAoI with respect to arrival rate for different antenna configurations under jamming attack. When there is also a delay constraint on the traffic, the arrival rate that minimizes AAoI can be different from the solution to the optimization problem, where it is only required to minimize the AAoI without any delay constraint. To get further insights into this problem, the following optimization problems are considered.
\subsubsection{Minimization of AAoI for Delay-Tolerant System} In this case, the objective is to minimize AAoI with respect to the arrival rate provided the queue remains stable and there is no constraint on the delay. The optimization problem is stated in the following.
\begin{equation}\label{AAoI_min_state}
\begin{aligned}
& \underset{\lambda}{\text{minimize}}
& & \mathrm{AAoI} ~ \text{in} ~\eqref{AAoI}, \quad \text{such that} \quad \lambda < \mu.
\end{aligned}
\end{equation}
The constraint ensures that the queue remains stable. We can solve this problem as follows.
\begin{equation}
\begin{aligned}
& \frac{\partial AAoI}{\partial \lambda} = 0,
\end{aligned}
\end{equation}
\begin{equation}\label{AAoI_opt_1}
\begin{aligned}
\text{or } & \frac{\mu-1}{\mu^2}+\frac{1-\mu}{(\mu-\lambda)^2} -\frac{1}{\lambda^2}= 0,\\
\text{or } &\lambda^4(\mu-1)+\lambda^3(2\mu-2\mu^2)-\lambda^2\mu^2+2\lambda\mu^3-\mu^4=0.\\
\end{aligned}
\end{equation}
The value of $\lambda$ is chosen which satisfies the above equation \eqref{AAoI_opt_1} and also, does not violate the stability condition of the queue.
\subsubsection{Minimization of AAoI for Delay-Sensitive System} In this case, it is required to minimize the AAoI with respect to the arrival rate provided the average delay in the system should not exceed a threshold ($D_{th}$) and the queue remains stable. The optimization problem is stated in the following:
\begin{equation}\label{AAoI_opt_2}
\begin{aligned}
& \underset{\lambda}{\text{minimize}}
& & \mathrm{AAoI} ~ \text{in} ~\eqref{AAoI},
~\text{such that} ~ \lambda < \mu, \text{ and } D \leq D_{th} ~\text{in} ~\eqref{eq:average_delay}.
\end{aligned}
\end{equation}
The above optimization problem is solved numerically and discussed in Sec~\ref{sec:result}. For both the optimization problems, the service rate depends on the number of antennas at the legitimate nodes, battery capacity at the jammer, and other parameters of the system (See Theorem~\ref{th:secondtheorem}).
\vspace{-4pt}
\section{Stability Region for the 2-user SIMO Broadcast Channel}
\begin{figure}[!ht]
\centering
\includegraphics[trim={0cm 0cm 0cm 0cm},clip, height=1.3in,width=3.4 in] {Broadcast_Sys_1.eps}
\setlength{\belowcaptionskip}{-12pt}
\caption{2-user broadcast channel with the jammer ($J$).}\label{fig:Broadcast}
\end{figure}
In this section, it is shown how the results developed for the point-to-point system help to characterize the stability region for the 2-user SIMO broadcast channel in the presence of energy harvesting jammer, where the receiver has $n_r$ antennas. The transmitter has two queues to store the incoming traffic and the packet stored at $i^{\text{th}}$ queue needs to be delivered to the $D_i^{\text{th}}$ receiver (See Fig.~\ref{fig:Broadcast}). Each receiver decodes its intended packet by treating other user's messages as noise. Due to lack of space, the stability region is characterized for the case when jammer has unlimited capacity and $\delta>p_j$. In this case, jammer energy queue is non-empty. The received signal at $i^{th}$ ($i \in \{1,2\}$) receiver is given by
\begin{equation}
\mathbf{y}_i = \mathbf{h}_i x +g_i x_J + z_i,
\end{equation}
where $\mathbf{h}_i \in \mathcal{C}^{n_r \times 1}$ is the channel between transmitter and the $i^{\text{th}}$ receiver and it undergoes Rayleigh fading. Hence, $|\mathbf{h}_i|^2$ follows Chi-Square distribution with $2n_r$ degrees of freedom. The channel $g_i \in \mathcal{C}$ is the channel between the attacker and $i^{\text{th}}$ receiver. The channel $g_i$ also undergoes Rayleigh fading and hence, $|g_i|^2$ follows an exponential distribution. The noise at receiver~$i$ is distributed as $z_i\sim \mathcal{CN}(0, 1)$. The input signal sent to the channel is $x = x_1 + x_2$ (when both queues are non-empty) or $x = x_i$ (when the only ith queue has a packet to send).
The packet arrival processes at the first and the second queue are assumed to be independent and stationary with mean rates $\lambda_1$ and $\lambda_2$ in packets per slot, respectively. Both queues have an infinite capacity to store incoming packets and $\bar{Q}_i$ denotes the size in the number of packets of the i-th queue. All other assumptions remain the same as that of the point to point MIMO model. The average service rate at the queues are given by
\vspace{-2pt}
\begin{align}
& \mu_1=Pr(\bar{Q}_2>0) Pr(D_{1/1,2}) + Pr(\bar{Q}_2 = 0) Pr(D_{1/1}), \label{broadcast_mu_1} \\
& \mu_2=Pr(\bar{Q}_1>0) Pr(D_{2/1,2})+Pr(\bar{Q}_1=0) Pr(D_{2/2}). \label{broadcast_mu_2}
\end{align}
where $D_{i/\tau}$ denote the event that receiver~$i$ is able to decode the packet transmitted from the i-th queue of the transmitter given a set of non-empty queues denoted by $\tau$ ($\tau\in \{1, 2\}$). Due to the interaction between the queues, the stochastic dominance technique \cite{21216} is used. There are two dominant systems. In the first dominant system, when the first queue is empty, then the source transmits a dummy packet for receiver~$1$, while the second queue behaves in the same way as in the original system. In the second dominant system, when the second queue is empty, then the source transmits a dummy packet for receiver~$2$, while the first queue behaves in the same way as in the original system. Following the process, as given in \cite{8320826}, the stability region for both the dominant systems are
\begin{figure*}
\begin{align}
&\mathcal{R}_1 = \left\{(\lambda_1,\lambda_2): \frac{\lambda_1}{Pr(D_{1/1})}+ \frac{Pr(D_{1/1})-Pr(D_{1/1,2})}{Pr(D_{1/1})Pr(D_{2/1,2})}\lambda_2<1, \lambda_2<Pr(D_{2/1,2})\right\}, \label{rate_1} \\
&\mathcal{R}_2 = \left\{(\lambda_1,\lambda_2): \frac{\lambda_2}{Pr(D_{2/2})}+ \frac{Pr(D_{2/2})-Pr(D_{2/1,2})}{Pr(D_{2/2})Pr(D_{1/1,2})}\lambda_1<1, \lambda_1<Pr(D_{1/1,2})\right\}. \label{rate_2}
\end{align}
\end{figure*}
given by $\mathcal{R} = \mathcal{R}_1 \cup \mathcal{R}_2$, where $\mathcal{R}_1$ and $\mathcal{R}_2$ are given by \eqref{rate_1} and \eqref{rate_2}, respectively. To determine the stability region, it is required to determine the various success probabilities. These are obtained using the outage probabilities derived for SIMO in Section~\ref{sec:outage}. The ratio of signal power to noise power along with jamming and/or interference at receiver~$i$ is given as
\begin{equation}
SJINR_i = \frac{P_i |\mathbf{h}_i|^2}{1 + P_J |g_i|^2 + P_j |\mathbf{h}_i|^2},\text{and } SJNR_i = \frac{P_i |\mathbf{h}_i|^2}{1 + P_J |g_i|^2}. \\
\end{equation}
where $i, j \in \{1, 2\}$ and $ i \neq j$.\\
Similarly, $SINR_i$ (signal power to interference and noise power ratio) and $SNR_i$ (signal to noise ratio) at receiver $i$ are denoted as follows
\begin{equation}
SINR_i = \frac{P_i |\mathbf{h}_i|^2}{1 + P_j |\mathbf{h}_i|^2}, \text{ and } SNR_i = P_i |\mathbf{h}_i|^2. \\
\end{equation}
where $i, j \in \{1, 2\}$ and $ i \neq j$. The success probabilities for both the receivers are obtained with and without jamming for different status of the queue in the following.
\subsubsection{When $\bar{Q}_1=0$ and $\bar{Q}_2\neq0$} In this case, the second queue at the transmitter sends a packet for receiver 2 as $\bar{Q}_1=0$. The receiver 2 can decode its intended packet if one of the following events occur:
$ D_{2/2}^{\text{WoJ}} = \left\{ SNR_2\geq\gamma_2 \right\}$ (without jamming) and $D_{2/2}^{\text{WJ}}=\left\{SJNR_2\geq\gamma_2\right\}$ (with jamming). The quantity $\gamma_i$ corresponds to the decoding threshold for successful decoding of the packet at receiver~$i$ and is related to the rate by the following relation: $\gamma_i = 2^{R_{i}}-1$ ($i$=1,2). It is not difficult to see that the $SJNR_i$ is similar to the SJNR in case of point-to-point model. Following a similar procedure as in Theorem~\ref{th:firsttheorem} (using \eqref{use_SuccPr}), the success probabilities can be obtained as given below.
\begin{equation}
Pr\big(D_{2/2}^{\text{WoJ}}\Big)=Pr(SNR_2\geq\gamma_2)
=Pr\Big(|\mathbf{h}_2|^2\geq\frac{\gamma_2}{P_2}\Big)
=\frac{\Gamma(n_{r},\frac{\gamma_2}{P_{2}})}{\Gamma(n_{r})}.
\label{WoJ}
\end{equation}
Similarly, one can show the following
\begin{align}
&Pr\big(D_{2/2}^{\text{WJ}}\big)=Pr(SJNR_2\geq\gamma_2)
= Pr\Big(|g_2|^2\leq\frac{P_2 |\mathbf{h}_2|^2-\gamma_2}{\gamma_2 P_J }\Big),\nonumber\\
&=\frac{\Gamma(n_{r},\frac{\gamma_2}{P_{2}})}{\Gamma(n_{r})}-\frac{e^{\frac{1}{P_J}}\Gamma(n_{r},\frac{\gamma_2}{P_{2}}+\frac{1}{P_J})}{\Gamma(n_{r})(1+\frac{P_{2}}{\gamma_2 P_J})^{n_{r}}}.\label{WJ}
\end{align}
Using \eqref{WoJ} and \eqref{WJ}, the success probability at receiver~$2$ is given by
\begin{align}
&Pr(D_{2/2}) = (1-p_J)Pr\big( D_{2/2}^{\text{WoJ}}\big)+p_J Pr\big(D_{2/2}^{\text{WJ}}\big),\\
& = (1-p_J) \frac{\Gamma(n_{r},\frac{\gamma_2}{P_{2}})}{\Gamma(n_{r})}+p_J\Bigg(\frac{\Gamma(n_{r},\frac{\gamma_2}{P_{2}})}{\Gamma(n_{r})}-\frac{e^{\frac{1}{P_J}}\Gamma(n_{r},\frac{\gamma_2}{P_{2}}+\frac{1}{P_J})}{\Gamma(n_{r})(1+\frac{P_{2}}{\gamma_2 P_J})^{n_{r}}}\Bigg).
\end{align}
\subsubsection{When $\bar{Q}_1\neq0$ and $\bar{Q}_2=0$} In this case, the first queue at the transmitter sends a packet for the receiver 1 as $\bar{Q}_2=0$. The receiver 1 can decode its intended packet if the following events are true.
\begin{align}
&D_{1/1}^{\text{WoJ}}=\left\{ SNR_1\geq\gamma_1 \right\} ~\text{without jamming}, \nonumber\\
&D_{1/1}^{\text{WJ}}=\left\{SJNR_1\geq\gamma_1\right\}~\text{with jamming.}
\end{align}
Following a similar procedure as in Theorem~\ref{th:firsttheorem} (using \eqref{use_SuccPr}), the success probabilities are given as follows
\begin{align}
& Pr(D_{1/1}) = (1-p_J)Pr\big( D_{1/1}^{\text{WoJ}}\big)+p_J Pr\big(D_{1/1}^{\text{WJ}}\big), \\
&=(1-p_J)\frac{\Gamma(n_{r},\frac{\gamma_1}{P_{1}})}{\Gamma(n_{r})}+p_J\Bigg(\frac{\Gamma(n_{r},\frac{\gamma_1}{P_{1}})}{\Gamma(n_{r})}-\frac{e^{\frac{1}{P_J}}\Gamma(n_{r},\frac{\gamma_1}{P_{1}}+\frac{1}{P_J})}{\Gamma(n_{r})(1+\frac{P_{1}}{\gamma_1 P_J})^{n_{r}}}\Bigg).
\end{align}
\subsubsection{When $\bar{Q}_1\neq0$ and $\bar{Q}_2\neq0$} In this case, both queues send packets. The packet sent by the first (second) queue should be decoded at receiver 1 (receiver 2), which are represented by the following events
\begin{align}
&D_{1/1,2}^{\text{WoJ}}=\left\{ SINR_1\geq\gamma_1 \right\} ~\text{without jamming},\nonumber \\
&D_{1/1,2}^{\text{WJ}}=\left\{SJINR_1\geq\gamma_1\right\}~\text{with jamming.}
\end{align}
In this case, although $SJINR_i$ corresponds to the ratio of signal power and noise power along with signal power of other user, with some rearrangement of terms in the outage probability calculation, $SJINR_i$ can be presented in a similar form as that of the $SJNR_i$. Hence, following a similar procedure as in Theorem~\ref{th:firsttheorem} (using \eqref{use_SuccPr}), the success probabilities can be obtained as
\begin{align}
Pr(D_{1/1,2}) & = (1-p_J)Pr\big( D_{1/1,2}^{\text{WoJ}}\big)+p_J Pr\big(D_{1/1,2}^{\text{WJ}}\big),\\
&=(1-p_J)\frac{\Gamma(n_{r},\frac{\gamma_1}{P_{1}-\gamma_1 P_2})}{\Gamma(n_{r})}+~p_J\Big(\frac{\Gamma(n_{r},\frac{\gamma_1}{P_{1}-\gamma_1 P_2})}{\Gamma(n_{r})}-\frac{e^{\frac{1}{P_J}}\Gamma(n_{r},\frac{\gamma_1}{P_{1}-\gamma_1 P_2}+\frac{1}{P_J})}{\Gamma(n_{r})(1+\frac{P_{1}-\gamma_1 P_2}{\gamma_1 P_J})^{n_{r}}}\Big).
\end{align}
By following similar steps as in the case of receiver~$1$, the success probability for receiver~$2$ is given by the following expression
\begin{align}
Pr(D_{2/1,2}) &=(1-p_J)\frac{\Gamma(n_{r},\frac{\gamma_2}{P_{2}-\gamma_2 P_1})}{\Gamma(n_{r})} +~p_J\Big(\frac{\Gamma(n_{r},\frac{\gamma_2}{P_{2}-\gamma_2 P_2})}{\Gamma(n_{r})}-\frac{e^{\frac{1}{P_J}}\Gamma(n_{r},\frac{\gamma_2}{P_{2}-\gamma_2 P_1}+\frac{1}{P_J})}{\Gamma(n_{r})(1+\frac{P_{2}-\gamma_2 P_1}{\gamma_1 P_J})^{n_{r}}}\Big).
\end{align}
The events $D_{2/1,2}^{\text{WoJ}}$ and $D_{2/1,2}^{\text{WJ}}$ occurs with non zero probability if $P_2-\gamma_2 P_1>0$. The stability region can be obtained by substituting the success probabilities obtained for various cases in \eqref{rate_1} and \eqref{rate_2} and then, taking union of these two regions ($\mathcal{R}_1$ and $\mathcal{R}_2$).
\vspace{-3pt}
\section{Results and Discussions}\label{sec:result}
In this section, numerical results are presented to illustrate the effect of jamming on the system performance for various parameter settings.
\begin{figure*}[htt]
\centering
\subfloat[\label{subfig-1:figure3}]{\includegraphics[trim={0.2cm 0cm 0.2cm .5cm},clip, height=2.4 in,width=3.3 in]{Outprob_vs_Jampower_Journal.eps}
\subfloat[\label{subfig-1:figure5}]{\includegraphics[trim={0.2cm 0cm 0.2cm .5cm},clip, height=2.4 in,width=3.3 in]{Asym_Outprob_vs_Antennas.eps}}
\setlength{\abovecaptionskip}{-2pt}
\setlength{\belowcaptionskip}{-15pt} %
\caption{(a) Outage probability against jamming power with $P=20$ dB, and $R=1$; (b) Asymptotic outage probability against number of antennas $(R=1)$.}
\end{figure*}
In Fig.~\ref{subfig-1:figure3}, the outage probability obtained in Theorem~\ref{th:firsttheorem} is plotted against jamming power for different MIMO configurations with $R=1$. Note that in this case, the attacker can always jam the communication.\footnote{This is considered to explore how the multiple antennas help to mitigate jamming attack under Rayleigh fading scenario.} It can be observed that increasing the number of antennas at the transmitter in case of MISO does not improve the outage performance when the jamming power is increased at the transmitter. Due to the lack of CSI at the sender, the transmit diversity does not help in improving the performance. However, in the case of SIMO, the outage probability decreases with the increase in the number of antennas at the receiver, even when the jamming power increases. This improvement primarily comes from the gain in the receive diversity with the increase in the number of receive antennas. To investigate the time diversity along with space diversity, MIMO with Alamouti code is considered. It can be seen that MIMO with Alamouti code gives significant improvement compared to MISO or SIMO system.
Fig.~\ref{subfig-1:figure5} shows the asymptotic outage probabilities obtained in \eqref{eq:miso_asym_wj}-\eqref{eq:mimo_asym_wj} against the number of antennas for a given value of $\eta$. For the MIMO with Alamouti code, the number of antennas at the transmitter is fixed at $n_t =2$, and the number of receive antennas $n_r$ is varied. From the plot, it can be observed MIMO with Alamouti code can achieve a very low value of outage probability even with a small number of antennas at the receiver or even when $\eta =2$. Due to the lack of CSI at the transmitter, outage probability is high in the case of MISO compared to SIMO or MIMO with Alamouti coding. From Fig.~\ref{subfig-1:figure3}, it can also be observed that in case of MISO system, the power budget or the number of antennas at the transmitter has to be increased significantly to meet the performance provided by SIMO or MIMO with Alamouti coding. Hence, exploiting time and space diversity can help to achieve the desired performance at a lower power budget even under jamming attack.
\begin{figure}[h]
\centering
\includegraphics[trim={0.2cm 0cm 0.2cm .5cm},clip, height=2.3 in,width=3.3 in]{ASR_vs_delta_journal.eps}
\setlength{\abovecaptionskip}{-4pt}
\setlength{\belowcaptionskip}{-10pt}%
\caption{Average service rate against energy arrival rate at jammer:
$P=20$ dB, $P_J=20$ dB, $R=1$, $B=2$, $p_J=0.7$ and $\lambda=0.2$.}\label{fig:figure6}
\end{figure}
\begin{figure*}[!htt]
\centering
\subfloat[\label{subfig-1:figure7}]{\includegraphics[trim={0.2cm 0cm 0.2cm .5cm},clip, height=2.4 in,width=3.3 in]{APD_vs_Delta_journal.eps}
\subfloat[\label{subfig-1:figure8}]{\includegraphics[trim={0.2cm 0cm 0.2cm .5cm},clip, height=2.4 in,width=3.3 in]{AAOI_vs_Delta_journal.eps}}
\setlength{\abovecaptionskip}{-2pt}
\setlength{\belowcaptionskip}{-20pt} %
\caption{(a) Average packet delay against energy arrival rate at jammer; (b) Average age of information against energy arrival rate at jammer:
$P=20$ dB, $P_J=20$ dB, $R=1$, $B=2$, $p_J=0.7$ and $\lambda=0.2$.}
\end{figure*}
\textit{In Figs.~\ref{fig:figure6}-\ref{subfig-1:figure16}: Dash and continuous curves correspond to battery with finite and infinite energy capacity, respectively.} In Fig.~\ref{fig:figure6}, the effect of energy arrival rate on the service rate is explored for different antenna configurations using the results in Theorems~\ref{th:secondtheorem} and \ref{th:thirdtheorem}. Note that MIMO with Alamouti provides the maximum service rate in comparison to SIMO and MISO systems with an increase in the number of antennas. In Figs. \ref{subfig-1:figure7} and \ref{subfig-1:figure8}, the average packet delay and AAoI are plotted against the energy arrival rate at the jammer for battery with finite and infinite energy capacity. It can be noticed that the AAoI and average packet delay is lowest for MIMO with Alamouti coding scheme in comparison to other cases. As the energy arrival rate $(\delta)$ increases at the jammer, the benefits of using Alamouti coding for the MIMO case are evident.
\begin{figure*}[htt]
\centering
\subfloat[\label{subfig-1:figure9}]{\includegraphics[trim={0.2cm 0cm 0.2cm .5cm},clip, height=2.4 in,width=3.3 in]{APD_vs_Lambda_journal_new.eps}
\subfloat[\label{subfig-1:figure10}]{\includegraphics[trim={0.2cm 0cm 0.2cm .5cm},clip, height=2.4 in,width=3.3 in]{AAOI_vs_Lambda_journal_new.eps}}
\setlength{\belowcaptionskip}{-18pt} %
\caption{(a) Average delay against packet arrival rate at transmitter; (b) Average age of information against packet arrival rate at transmitter:
$P=20$ dB, $P_J=20$ dB, $R=1$, $\delta=0.6$, $B=2$, and $p_J=0.7$.}
\end{figure*}
In Figs.~\ref{subfig-1:figure9} and \ref{subfig-1:figure10}, the average delay and AAoI are plotted against packet arrival rate at the transmitter for different antenna configurations, respectively. From Fig~\ref{subfig-1:figure10}, it can be observed that AAoI initially decreases with the increase in the arrival rate and then increases when the arrival rate goes beyond some specific value. The reason behind this observation is that when the $\lambda$ is low, the source does not often generate updates. Thus, the destination is not updated with new information often, and it results in high AAoI. When the arrival probability is high, there is an excessive queuing delay. Thus, the AAoI is high at the destination. This is because the received packets faced a queuing delay that affected the freshness of the information they carry. This is in contrast to the behavior of average packet delay with the increase in the arrival rate. \textit{This observation highlights the importance of AAoI as a measure to capture freshness since the performance metrics such as throughput and delay cannot capture this attribute.} It is worth noting that MIMO with Alamouti coding scheme can support the lowest value of average packet delay and AAoI with a higher value of the packet arrival rate $(\lambda)$.
\begin{figure*}[htt]
\centering
\subfloat[\label{subfig-1:figure11}]{\includegraphics[trim={0.2cm 0cm 0.2cm .5cm},clip, height=2.4 in,width=3.3 in]{AAOI_vs_APD_journal.eps}
\subfloat[\label{subfig-1:figure12}]{\includegraphics[trim={0.2cm 0cm 0.2cm .5cm},clip, height=2.4 in,width=3.3 in]{AAOI_vs_B_journal.eps}
\setlength{\belowcaptionskip}{-20pt} %
\caption{(a) Average age of information against average packet delay ($B=2$); (b) Average age of information against energy harvesting battery capacity at jammer:
$P=20$ dB, $P_J=20$ dB, $R=1$, $\delta=0.6$, $\lambda=0.2$, and $p_J=0.7$.}
\end{figure*}
Fig.~\ref{subfig-1:figure11} shows the interplay between AAoI and average packet delay for different antenna configurations. When the jammer has an energy battery of infinite capacity, the value of average packet delay and AAoI increases compared to the jammer with finite energy battery capacity. We can observe that MIMO with Alamouti guarantees a low value of average packet delay as well as AAoI. From Fig.~\ref{subfig-1:figure12}, one can see that when the battery size is around B = 10, the AAoI is similar to the case when the jammer has a battery of infinite capacity. Hence, the jammer cannot further increase the AAoI by increasing the battery size.
\begin{figure*}[htt]
\centering
\subfloat[\label{subfig-1:figure13}]{\includegraphics[trim={0.2cm 0cm 0.2cm .5cm},clip, height=2.4 in,width=3.3 in]{APD_vs_Jampower.eps}
\subfloat[\label{subfig-1:figure14}]{\includegraphics[trim={0.2cm 0cm 0.2cm .5cm},clip, height=2.4 in,width=3.3 in]{AAOI_vs_Jampower.eps}}\setlength{\abovecaptionskip}{-2pt}
\setlength{\belowcaptionskip}{-20pt} %
\caption{(a) Average packet delay against jamming power; (b) Average age of information against jamming power:
$P=20$ dB, $R=1$, $\delta=0.6$, $\lambda=0.2$, $B=2$ and $p_J=0.7$.}
\end{figure*}
\begin{figure*}[htt]
\centering
\subfloat[\label{subfig-1:figure15}]{\includegraphics[trim={0.2cm 0cm 0.2cm .5cm},clip, height=2.4 in,width=3.3 in]{APD_vs_p2.eps}
\subfloat[\label{subfig-1:figure16}]{\includegraphics[trim={0.2cm 0cm 0.2cm .5cm},clip, height=2.4 in,width=3.3 in]{AAOI_vs_p2.eps}}
\setlength{\abovecaptionskip}{-1pt}
\setlength{\belowcaptionskip}{-15pt} %
\caption{(a) Average packet delay against probability of jamming; (b) Average age of information against probability of jamming:
$P=20$dB, $P_J=20$dB, $R=1$, $\delta=0.6$, $\lambda=0.2$, and $B=2$.}
\end{figure*}
Figs.~\ref{subfig-1:figure13} and \ref{subfig-1:figure14} show the impact of jamming power on average packet delay and AAoI. Figs.~\ref{subfig-1:figure15} and \ref{subfig-1:figure16} show the impact of the probability of jamming on average packet delay and AAoI. From the plots, it can be observed that with increasing the jamming probability, the average packet delay and AAoI decrease as it is limited by the energy harvesting ability. It can be noticed that MIMO with Alamouti performs better in both cases, i.e., finite and infinite battery capacity, among all considered configurations. As the jamming power increases at the attacker, the benefits of space-time diversity are more prominent.
\begin{figure*}[htt]
\centering
\subfloat[Without delay constraint\label{subfig-1:figure17}]{\includegraphics[trim={0.2cm 0cm 0.2cm .5cm},clip, height=2.4 in,width=3.3 in]{AAoI_Optimum_1_new.eps}
\subfloat[With delay constraint $(D_{th} = 2.25)$\label{subfig-1:figure18}]{\includegraphics[trim={0.2cm 0cm 0.2cm .5cm},clip, height=2.4 in,width=3.3 in]{AAoI_Optimum_2_new.eps}}
\setlength{\belowcaptionskip}{-10pt} %
\caption{Minimizing AAoI against jamming power: $P=20$ dB, $R=1$, $\delta=0.6$, $B=2$, and $p_J=0.7$. (a) Dashed curve: battery with finite energy capacity and continuous curves: battery with infinite energy capacity and (b) Dashed curve: without delay constraint and continuous curve: with delay constraint.}
\end{figure*}
Figs.~\ref{subfig-1:figure17} and \ref{subfig-1:figure18} show the minimized AAoI for optimal values of $\lambda$ against jamming power. In Fig.~\ref{subfig-1:figure17}, AAoI is plotted against the jamming power for the optimization problem stated in \eqref{AAoI_min_state}. In this case, there is no constraint on delay. From the figure, it can be observed that MIMO with Alamouti ensures minimum AAoI with an increase in the jamming power. In Fig.~\ref{subfig-1:figure18}, AAoI is plotted against the jamming power for the optimization problem stated in \eqref{AAoI_opt_2}. Recall that this optimization problem aims to minimize the AAoI with respect to arrival rate for a delay-sensitive system. It can be observed that the AAoI increases sharply with an increase in jamming power for the MISO system when there is a constraint on the delay. The Table~\ref{tab:Table_1} shows the corresponding value of $\lambda$ which minimizes the AAoI for MIMO $(2 \times 2)$ with Alamouti for both the optimization problems. However, in the case of MIMO with Alamouti coding, it can guarantee minimum AAoI among the different antenna configurations even if there is a delay constraint. It can be seen that with the increase in the jamming power, it is not possible to achieve low AAoI with stringent delay constraints.
\begin{table}
\centering
\caption
{MIMO $(2 \times 2)$ with Alamouti coding scheme : Optimum values of data arrival rate ($\lambda$) against jamming power for infinite energy battery capacity $(D_{th} = 2.25)$.}
\begin{tabular}{|c|c|c|c|}
\hline
\multicolumn{2}{|c|}{Delay-Tolerant Traffic} & \multicolumn{2}{|c|}{Delay-Sensitive Traffic}\\ \hline
\textbf{$P_J$} & \textbf{$\lambda$} & \textbf{$P_J$} &\textbf{$\lambda$} \\ \hline
10 & 0.9804 & 10 & 0.9800 \\ \hline
15 & 0.9631 & 15 & 0.9626 \\ \hline
20 & 0.9439 & 20 & 0.9435 \\ \hline
25 & 0.9241 & 25 & 0.9240 \\ \hline
30 & 0.9044 & 30 & 0.9044 \\ \hline
35 & 0.8851 & 35 & 0.8852 \\ \hline
40 & 0.8665 & 40 & 0.8665 \\ \hline
45 & 0.8488 & 45 & 0.8487 \\ \hline
50 & 0.8319 & 50 & 0.8318 \\ \hline
55 & 0.8158 & 55 & 0.8045 \\ \hline
60 & 0.8006 & 60 & 0.7646 \\ \hline
65 & 0.7862 & 65 & 0.7217 \\ \hline
70 & 0.7726 & 70 & 0.6746 \\ \hline
\end{tabular}\label{tab:Table_1}\vspace{-13pt}
\end{table}
\begin{figure}
\centering
\includegraphics[trim={0.2cm 0cm 0.2cm .5cm},clip, height=2.4 in,width=3.3 in]{Broadcast_2user.eps}%
\setlength{\belowcaptionskip}{-23pt}
\caption{Stability region with $P_1 = P_2 = 10$dB, $ P_J = 20$dB , $\gamma_1 = 0.7$, $\gamma_2 = 0.5$, and $p_J=0.6$.}\label{fig:figure19}
\end{figure}
In Fig.~\ref{fig:figure19}, the stability region obtained from \eqref{rate_1} and \eqref{rate_2} for the $2$-user SIMO broadcast channel in the presence of a jammer is plotted. In this case, it is assumed that battery at the jammer has unlimited capacity and $\delta > p_J$. It can be observed that as the number of antennas at the receivers increases, the stability region also increases.
\vspace{-8pt}
\section{Conclusions}
This work characterizes the performance of a point-to-point MIMO system under random arrival of data at the transmitter in the presence of a jammer with energy harvesting ability. The outage probability is derived for the considered system model for the Rayleigh fading scenario. The derived results on outage probabilities were used to characterize the \textit{average service rate}, \textit{average delay}, and \textit{AAoI} when the attacker has a battery with finite or unlimited capacity. The results illustrate the role of transmit, receive, space, and time diversity in mitigating jamming attacks and improving the performance of the system in the presence of a jammer. The developed results on average delay and AAoI show the relevance of these metrics under jamming attack when the transmitter has time-sensitive information and timely updates are of importance. The exploitation of time and space diversity helps to achieve superior performance even when the transmitter has time-sensitive data and less power budget. Moreover, the work also demonstrates how the developed results for the point-to-point case helps to determine the stability region for multi-user scenarios such as 2-user BC in the presence of jammer. Extending the results for the scenario when jammer has multiple antennas is an interesting direction for research. Another interesting direction of work is the characterization of stability region for multi-user scenario under jamming attack.
\normalem
\bibliographystyle{IEEEtran}
|
\section{Introduction}
Optical photometry of ASASSN-V~J205543.90+240033.5 (J2055 hereafter) reveals periodic variations on 12-hour and 10-minute timescales \citep{kato21a,kato21b} that are similar to variability seen in the white dwarf (WD) pulsar AR~Sco \citep{marsh16,stiller18}. AR~Sco is a binary consisting of a red dwarf detached from a magnetic WD that rotates with a period of 1.95~minutes. For J2055, the 12-hour variation is thought to be related to the binary orbit while the origin of the 9.77-minute periodicity is uncertain.
\section{Data}
We obtained spectra of J2055 using the Large Binocular Telescope (LBT\footnote{The LBT is an international collaboration among institutions in the United States, Italy and Germany. LBT Corporation partners are: The University of Arizona on behalf of the Arizona Board of Regents; Istituto Nazionale di Astrofisica, Italy; LBT Beteiligungsgesellschaft, Germany, representing the Max-Planck Society, The Leibniz Institute for Astrophysics Potsdam, and Heidelberg University; The Ohio State University, representing OSU, University of Notre Dame, University of Minnesota and University of Virginia.}) and twin Multi-Object Dual Spectrographs \citep[MODS;][]{pogge12} on 2021, September 16 (UT). A sequence of eleven, 150s exposures were taken with MODS1 through the SX telescope providing coverage over 36~minutes including readout and overheads. The same exposure sequence was obtained with MODS2 through the DX telescope. The image quality averaged 0.7 to 0.8 arcseconds during the observation. A 0.8~arcsec wide slit was employed with the gratings to give a spectral resolution in the red of 1350, or 215~km$\;$s$^{-1}$ (FWHM) at H$\alpha$. In the blue, the resolution is 1890, or 160~km$\;$s$^{-1}$ (FWHM) at H$\beta$. Each spectrograph has a red and blue arm split at 5650~\AA\ by a dichroic mirror allowing wavelength coverage from 3200~\AA\ to 1.01~$\mu$m.
The spectra were taken near photometric phase 0.04 based on the ephemeris from \citet{kato21b}. This is very close to the peak brightness of the 12-hour photometric periodicity.
\vspace{0.5cm}
\section{Analysis}
The resulting average spectrum is displayed in the top panel of Figure~\ref{fig1}. It shows a continuum rising toward the blue with
a large number of narrow emission lines from the ultraviolet to the near infrared. The Balmer and Paschen hydrogen emission sequences are evident.
However, the strongest lines are permitted carbon and nitrogen emission features with a range of ionization states. For example, the blend of C~III lines at 4649~\AA\ is 50\%\ brighter than H$\beta$. Also, the C~II 6578+6583~\AA\ is half the flux of H$\alpha$. He~II emission is also significant at 4686, 5411, and 8237~\AA .
H$\alpha$ and H$\beta$ show a central absorption that vanishes toward the higher Balmer lines. The H$\alpha$ and H$\beta$ line widths are 600~km$\;$s$^{-1}$ (FWHM). The base of the H$\beta$ line extends to $\pm$900~km$\;$s$^{-1}$. The Balmer widths decrease toward the higher level transitions with H$\epsilon$ having a width of 440~km$\;$s$^{-1}$ (FWHM). The Balmer decrement is inverted, with the H$\alpha$/H$\beta$=0.5. The He~II and carbon lines are unresolved at this resolution.
Cross-correlation between the average spectrum and each of the 11 exposures shows a systematic decrease in velocity totalling 44~km$\;$s$^{-1}$ over the 36 minutes of data. This appears to confirm the binary nature of the system, although the spectra cover only 5\%\ of the photometric period. Narrow absorption features are detected from Na~I and Ca~II. The absorption lines do not show any significant variation in velocity suggesting they originate in circumstellar or interstellar gas.
\begin{figure}[h!]
\begin{center}
\includegraphics[scale=0.55,angle=0]{figure.pdf}
\caption{{\bf Top:} The average spectrum of J2055 from the LBT/MODS observations. The Balmer and Paschen series are seen in emission as well as narrow carbon and nitrogen lines. Interstellar or circumbinary absorption features of Na~I and Ca~II are visible. {\bf Lower-left:} Close-up of the Bowen features of C~III/N~III and He~II. {\bf Lower-right:} Close-up around H$\alpha$ showing the deep central absorption and broad wings. The C~II blend around 658~nm is unusually strong relative to H$\alpha$. \label{fig1}}
\end{center}
\end{figure}
\section{Conclusions}
The spectrum of J2055 near maximum is similar to the post-common envelope binaries (aka pre-cataclysmic binaries) TW~CrV \citep{tw_crv} and YY~Hya \citep{yy_hya}. These systems consist of a hot compact star, probably a WD, irradiating the face of a late-type (K or M type) star. No evidence of the cool secondary is seen in our spectra, possibly because they were obtained near maximum light when the irradiated face dominates the flux.
The optical spectrum of J2055 is clearly quite distinct from that of AR~Sco \citep{marsh16,garnavich19}. The WD in AR~Sco has a relatively cool surface temperature \citep{garnavich21}, leading to a spectrum dominated by Balmer, neutral helium, and Ca~II emission lines. The heating of its secondary may be enhanced by non-thermal radiation. The temperature of the WD in J2055 appears much higher than for AR~Sco, and post-common envelope systems often show enhancement of CNO elements \citep{shimansky11}. The differences in spectral properties between J2055 and AR~Sco are likely due to the combination of enhanced CNO abundances and higher temperatures in J2055.
The origin of the 9.77-minute periodicity \citep{kato21a} remains uncertain. It may come from surface temperature variations on a rotating WD, or, as in AR~Sco, the beat period arising from the interaction of the WD magnetic field interacting with the secondary star.
\acknowledgments
We thank S. Kimeswenger and co-authors for providing a preview of their paper. We thank Jenny Power, Olga Kuhn, and LBT staff for their assistance in obtaining the MODS spectra of J2055 during the observatory restart block.
|
\section{Introduction}
Entanglement entropy has often been considered a pathway to quantum gravity, ever since its introduction by Rafael Sorkin \cite{Sorkin1983}, where a connection to black hole entropy was shown. In that work, the entanglement entropy of a quantum scalar field in the exterior of a black hole was shown to scale as the spatial area \(A\) of the event horizon, in units of the UV cutoff, \(A/\ell_{UV}^2\). Likewise, the entropy of a black hole was known to scale as the area of the event horizon \cite{PhysRevD.7.2333}, leading to the proposal that the Bekenstein-Hawking entropy may be wholly or in large part entanglement entropy. Much subsequent research has investigated potential origins of black hole entropy \cite{Mathur_2005, Horowitz:1996qd, Rovelli_1996, Carlip_2015, Dou:2003af}, including entanglement entropy as a leading candidate \cite{PhysRevD.34.373, Jacobson_2013, Solodukhin_2011, Emparan_2006}. Presently, it remains an important open question as to what the true microscopic source of black hole entropy is.
In addition to its role in quantum gravity, entanglement entropy has proven to be a rich topic with many applications in other areas of physics. A number of interesting theorems have been derived using entanglement entropy in quantum field theory, and especially in conformal field theory. These include c-theorems which relate a function of the coupling constants of the theory to the Virasoro central charge \cite{Casini_2004, Calabrese_2004, Myers_2010}.
In AdS/CFT, Ryu and Takayanagi found that the entanglement entropy is equal to $1/4$ the area of a minimal surface \cite{Ryu_2006, Ryu2_2006}.
In condensed matter physics, features such as topological order and properties of Fermi surfaces have been studied using entanglement entropy \cite{Kitaev_2006, PhysRevLett.96.110405, Swingle_2010}. In quantum information, entanglement entropy is used to constrain processes such as teleportation, and to study the properties of states \cite{Vidal_2002}.
Just as there are many different applications of entanglement entropy, there are many different methods for computing it. These include heat kernel methods \cite{Solodukhin_2011}, the replica trick \cite{Callan_1994, Hertzberg_2012}, the Ryu-Takayanagi formula \cite{Ryu_2006}, Euclidean path integral methods \cite{Rosenhaus_2014, Rosenhaus_2015
, using spatial correlators \cite{PhysRevD.34.373, Peschel_2003}, and using spacetime correlators \cite{Sorkin:2012sn, Chen_2020}. The spacetime correlator method is what we use in this paper. The choice of technique used depends both on the physical features and requirements of the theory, as well as what is practically solvable. In this paper we are interested in the entanglement entropy of a Gaussian scalar field in disjoint $1+1$D spacetime intervals (causal diamonds) in a causal set. Causal sets do not admit a meaningful notion of a state on a hypersurface, which is clear upon consideration of a maximal antichain\footnote{An antichain is a subset of the causal set consisting of causally unrelated elements. A \emph{maximal} antichain is one that is inextendable, i.e. one cannot add more elements to the set such that they all remain causally unrelated.}, the analogue of a spatial hypersurface. Due to the random spatio-temporal discreteness of the causal set, causal connections can pass through the antichain without making an imprint on it. This means that the maximal antichain cannot carry all the information associated with spacetime elements to its past and future, and therefore cannot serve as a Cauchy surface. This renders unusable many of the conventional spatial methods for computing entanglement entropy.
The method for computing entanglement entropy from spacetime correlation functions of the field, which we work with, was first introduced in \cite{Sorkin:2012sn}. We will review this formulation in Section \ref{sec:SfromW}. In addition to the requirement from causal set theory to work with spacetime rather than spatial quantities, there are also more general physical motivations to do so. An ultraviolet cutoff is required to obtain a finite and well-defined entanglement entropy, and if the entropy is computed on a spatial hypersurface, the cutoffs used are spatial and therefore not covariant. Ambiguities can then arise in making a choice of frame and cutoff, which are physically undesirable, especially in the context of gravity. The entanglement entropy formulation we work with has the advantage of admitting the use of a covariant spacetime UV cutoff.
In Section \ref{sec:oneint} we review the single interval entanglement entropy in a causal set, and we present new results for the case where the interval shares a boundary with the global region. We then discuss the multiple interval case in Sections \ref{sec:2int} and \ref{sec:3int}. We consider two and three disjoint regions, though our work easily generalises to any number of disjoint intervals. Treating disjoint intervals turns out to be another strength of working with the spacetime correlator formulation of entanglement entropy. This is because the details of the multiple interval calculation follow very closely those of the single interval case, which has been well studied in \cite{Afshordi_2012, Saravani_2014, Sorkin_2018}, and only additionally requires the use of greater computational power.
\section{Entropy from Spacetime Correlation Functions}
\label{sec:SfromW}
Conventionally, entanglement entropy is defined to be
\begin{equation}
S = -\Tr(\rho_{red} \ln{\rho_{red}}),
\label{ee}
\end{equation}
where $\rho_{red}$ is the reduced density matrix after tracing out some degrees of freedom from an initially pure global state $\rho$. The degrees of freedom traced out lie within one of two causally complementary partitions of the spacetime (or in practice, Cauchy hypersurface).
The definition of entanglement entropy we work with is equivalent to \eref{ee}, but has been rewritten explicitly in terms of spacetime correlation functions. For a derivation and further details we refer the reader to \cite{Sorkin:2012sn}. It was shown in \cite{Sorkin:2012sn} that the entropy of a Gaussian scalar field in a spacetime region $\mathcal{R}$ is
\begin{equation}
S = \sum_{\lambda} \lambda \log{|\lambda|},
\label{S}
\end{equation}
where \(\lambda\) are the eigenvalues obtained by solving the generalised eigenvalue problem
\begin{equation}
\mathbf{W}v = i \lambda \mathbf{\Delta} v,\indent \mathbf{\mathbf{\Delta}} v\neq 0.
\label{geneig}
\end{equation}
\(i\mathbf{\Delta}\) is the Pauli-Jordan function, or spacetime commutator of the field operators, and \(\mathbf{W}\) is the Wightman, or two-point correlation function.
For a Gaussian scalar field theory,
\begin{equation}
i\mathbf{\Delta}(x,y) := \langle0| [\Phi(x), \Phi(y)] |0\rangle=[\Phi(x), \Phi(y)],
\end{equation}
and
\begin{equation}
\mathbf{W}(x,y) := \langle0| \Phi(x) \Phi(y) |0\rangle,
\end{equation}
where $x,y \in \mathcal{R}$. Note also the relations $ i\mathbf{\Delta}(x,y)= \mathbf{W}(x,y)- \mathbf{W}(y,x)= \textrm{Im} (\mathbf{W}(x,y))/2$.
If we restrict \(\mathbf{\Delta}\) and \(\mathbf{W}\) to a subdomain of $\mathcal{R}$ with nonempty causal complement, the entropy that is calculated via \eref{S} will correspond to the entanglement entropy between the subdomain and its complement. It is worth highlighting again that the formulas above are covariant and they solely involve spacetime correlation functions.
The formulas \eref{S} and \eref{geneig} have been used in several settings both in the continuum theory and in causal set theory to compute entanglement entropies \cite{Saravani_2014, Sorkin_2018, Belenchia_2018, Surya_2021, Mathur_2021}. It has also been shown that these formulae capture the entanglement entropy even in non-Gaussian and interacting theories, up to first order in perturbation theory \cite{Chen_2020}. For theories where $ i\mathbf{\Delta}$ and $\mathbf{W}$ are known both in the continuum and in the causal set, the causal set calculations are easier to perform than the continuum ones. This is because $ i\mathbf{\Delta}$ and $\mathbf{W}$ are finite dimensional matrices and the generalised eigenvalue problem \eref{geneig} can be numerically solved relatively easily. In the continuum, $ i\mathbf{\Delta}$ and $\mathbf{W}$ are infinite dimensional, and must be made into finite dimensional matrices with respect to a choice of cutoff (see e.g. \cite{Saravani_2014}). This can be a challenging task in practice.
In this work we focus on the causal set case, for practical ease, and also because it is more fundamental and equipped with a natural UV cutoff. In the causal set, the entanglement entropy is finite, and the degrees of freedom are finite and well-defined with respect to the discreteness scale, though as we will see, extra ``truncations'' are necessary to obtain meaningful results.
\section{Entanglement Entropy of a Single 1+1D Causal Interval}
\label{sec:oneint}
In order to use \eref{S} and \eref{geneig} to study the entanglement entropy, we need a vacuum state, or Wightman function \(\mathbf{W}\). The only known means to define a Wightman function in causal set theory is via the Sorkin-Johnston (SJ) prescription \cite{Johnston_2008, johnston2010quantum, Afshordi2_2012, Sorkin:2017fcp}. This prescription is very much in the same spirit as the entropy formulation we reviewed in the previous section, as it relies solely on spacetime quantities to define a vacuum state, and does not make reference to any spatial quantities on hypersurfaces. We review this prescription below.
\subsection{The SJ Prescription}
The starting point of the SJ prescription is the retarded Green function. Since a unique retarded Green function exists only in globally hyperbolic spacetimes, the SJ vacuum is unique only in globally hyperbolic spacetimes. We will always work with causal set sprinklings\footnote{For a review of causal set theory we refer the reader to \cite{PhysRevLett.59.521, Surya_2019}.} approximated by $1+1$D causal intervals (diamonds) in Minkowski spacetime, as illustrated in Figure \ref{diamond}. These intervals are globally hyperbolic, and therefore have a unique retarded Green function and SJ state. We will work with a massless theory, for simplicity, and a Gaussian theory, since this is what the SJ prescription provides.
\begin{figure}[H]
\centering
\includegraphics[width=10cm]{single_diamond.png}
\caption{A causal set sprinkling of a causal diamond or interval in $1+1$D Minkowski spacetime.}
\label{diamond}
\end{figure}
\FloatBarrier
The retarded Green function in $1+1$D Minkowski spacetime for a massless scalar field is
\begin{equation}
G_R(x, y) = \frac{1}{2}\Theta(\tau^2)\Theta(x^{0} - y^{0}),
\end{equation}
where \(\Theta\) is the Heaviside step function, and \(\tau\) is the proper time, defined via \(\tau^2=(x^0-y^0)^2-(x^i-y^i)^2\). This propagator is very simple, returning a value of one half if \(y\) is in \(x\)'s past lightcone. It is reminiscent of the causal matrix in the causal set, which is defined as
\begin{equation}
\mathbf{C_{xy}} := \left\{ \begin{array}{ll}
1 & x \prec y, \: \textrm{and} \:\, x \neq y\\
0 & \textrm{otherwise}.
\end{array} \right.
\end{equation}
We have used the notation that the indices $_{xy}$ designate the matrix element relating causal set elements $x$ and $y$, and $x \prec y$ denotes that $x$ causally precedes $y$. Indeed, the causal set retarded (and advanced) Green function, $\mathbf{K}_{R}$ (and $\mathbf{K}_{A}$), is defined precisely using the causal matrix:
\begin{equation}
\mathbf{K}_{R} = \frac{1}{2}\mathbf{C}^T, \quad \textrm{and} \quad \mathbf{K}_{A} = \mathbf{K}_{R}^{T} = \frac{1}{2}\mathbf{C},
\label{kra}
\end{equation}
where we have used the symbol $\mathbf{K}$ for the causal set Green function, to distinguish it from the continuum one $G$, and the superscript $^T$ refers to the transpose operation.
From $\mathbf{K}_{R,A}$ we can construct $\mathbf{\Delta}=\mathbf{K}_{R}-\mathbf{K}_{A}$. Because \(i\mathbf{\Delta}\) is Hermitian and antisymmetric, it can be expanded in its eigenbasis as
\begin{equation}
i\mathbf{\Delta} = \sum_{i}^N\Lambda_{i}v_iv_i^{\dagger},
\label{PJ}
\end{equation}
where $N$ is the number of elements in the causal set. \(\Lambda_{i}\) denotes an eigenvalue of \(i\mathbf{\Delta}\), with \(v_{i}\) being the corresponding eigenvector.
The SJ Wightman function, \(\mathbf{W}_{SJ}\) is defined to be the positive part of the expansion of \(i\mathbf{\Delta}\) in its eigenbasis in \eref{PJ}, i.e.
\begin{equation}
\mathbf{W}_{SJ} := \sum_{i}\Lambda_{i}v_iv_i^{\dagger}, \quad \Lambda_i > 0.
\label{CS_W}
\end{equation}
The SJ Wightman function defines a pure state for the entropy formulation we reviewed earlier, as it yields $\lambda=0 \textrm{ or } 1$ in \eref{geneig}, and hence a vanishing entropy in \eref{S}. Its restriction to a subregion with nonzero causal complement, however, will yield nontrivial $\lambda$'s, and a nonvanishing entanglement entropy.
For the spacetime interval in Figure \ref{diamond}, $W_{SJ}$ has been well studied both in the causal set and continuum \cite{Afshordi_2012}. Away from the boundaries of the diamond, which is where we will primarily place the subintervals we work with, the state resembles that of the Minkowski vacuum with an IR cutoff. Close to the left or right corner of the diamond, the state resembles that of the Minkowski vacuum in the presence of a static mirror or reflecting boundary at that corner.
\subsection{Truncations}
\label{sec:trunc}
For reasons which we will not delve into, it turns out that a stricter condition than $\mathbf{\Delta} v\neq 0$ in \eref{geneig} is necessary to obtain meaningful results in the causal set. Not only must we exclude the kernel of $i\mathbf{\Delta}$, we also need to exclude a large number of additional components corresponding to eigenvectors with nonzero but small magnitude eigenvalues \cite{Sorkin_2018}. To understand how we decide which components to keep and which to exclude, it is instructive to first review some aspects of $i\mathbf{\Delta}$ and its eigendecomposition in Minkowski spacetime.
In lightcone coordinates \(u=\frac{x+t}{\sqrt{2}}\) and \(v=\frac{t-x}{\sqrt{2}}\), the Pauli-Jordan function in the continuum theory is \cite{Afshordi_2012}
\begin{equation}
i\mathbf{\Delta}(u,v;u',v') = \frac{-i}{2}(\Theta(u-u') + \Theta(v-v') - 1).
\end{equation}
Its eigenfunctions $f$ satisfy
\begin{equation}
\int_{\mathcal{R}} i\mathbf{\Delta}(u,v;u',v') f(u,v) \: du\, dv = \Lambda f(u',v'),
\label{ceig}
\end{equation}
where the domain of integration, $\mathcal{R}$, is the causal diamond with origin at the center and $u\in[-L,L]$ and $v\in[-L,L]$. This causal interval is the continuum approximation to the causal set in Figure \ref{diamond}.
The eigenfunctions with non-zero eigenvalues\footnote{It is worth noting that this eigenspace spans the solution space of the Klein-Gordon equation.} satisfying \eref{ceig} are \cite{johnston2010quantum, Afshordi_2012}
\begin{equation}
f_k(u, v) = e^{-iku} - e^{-ikv}, \qquad k = \frac{n\pi}{L}, \: n \in \mathbb{Z}\backslash 0
\label{single_diamond_soln1}
\end{equation}
and
\begin{equation}
\begin{array}{ll}
& g_k(u, v) = e^{-iku} + e^{-ikv} - 2\cos{(kL)}, \\
& k \in \mathcal{K}= \{k \in \mathbb{R} | \tan{(kL)}=2kL \: \textrm{and} \: k \neq 0 \}.
\end{array}
\label{single_diamond_soln2}
\end{equation}
Each \(f_k(u, v)\) and \(g_k(u, v)\) has corresponding eigenvalue
\begin{equation}
\Lambda_k =\frac L k,
\end{equation}
and for large \(k\), the \(g_k\) eigenvalues tend to those of \(f_k\), i.e. $\Lambda_k = \frac{L^2}{n \pi}$. When calculating entanglement entropies in the continuum via the definition in Section \ref{sec:SfromW}, a minimum eigenvalue serving as the UV cutoff is arbitrarily set \cite{Saravani_2014}, given by \(\Lambda_{min} = \frac{L^2}{n_{max}\pi}\). Via dimensional considerations, and a direct comparison of values, it is seen that the eigenvalues in the causal set are related to those in the continuum by rescaling with a factor of the density \(\rho = \frac{N}{4L^{2}}\), such that
\begin{equation}
\Lambda^{cs} = \rho \Lambda^{cont}.
\label{lamscale}
\end{equation}
In the causal set, the value of the cutoff and minimum eigenvalue are not arbitrary. The cutoff is the discreteness scale, and the minimum eigenvalue is related to it. Given the nature of the eigenfunctions as approximate linear combinations of plane waves and the eigenvalues as inversely proportional to wavenumbers, we expect a maximum wavenumber, \(k_{max}\), near the discreteness scale. Larger wavenumbers or shorter wavelengths cannot be meaningfully supported on the causal set. This upper bound on the wavenumber leads via \eref{lamscale} to a minimum (in magnitude) expected eigenvalue given by
\begin{equation}
\Lambda^{cs}_{min} =\frac {\rho L} {k_{max}}=\frac{\sqrt{N}}{4\pi},
\end{equation}
where $N$ is the number of elements in the diamond in which the eigenvalue problem is solved. As first noted in \cite{Sorkin_2018}, there are residual contributions below this $\Lambda^{cs}_{min}$ in the spectrum of $i\mathbf{\Delta}$ in the causal set. These need to be removed through \emph{truncations}. The truncations are a procedure that take out or set to zero any components that correspond to eigenfunctions with eigenvalues smaller (in magnitude) than $\Lambda^{cs}_{min}$. The truncations need to be done at two stages of the calculation: during the SJ prescription, and after the restriction to the entangling subregion but before solving \eref{geneig}. For further details on the truncation process, we refer the reader to \cite{Sorkin_2018, YKY}.
\subsection{Subinterval in a Corner}
\label{sec:corner}
We now use everything we have reviewed above to calculate the entanglement entropy of the restriction of a massless scalar field to a causal set subinterval in the left corner of a larger one (such that they share a boundary). The causal intervals are shown in Figure \ref{corner}. Our calculation is similar to that in \cite{Sorkin_2018}, the only difference being that the subinterval considered there was concentric to the larger one and had two boundaries across which there was entanglement. In our setup there is only one boundary contributing to the entanglement entropy.
\begin{figure}[h!]
\centering
\includegraphics[width=10cm]{corner_diamond.png}
\caption{A causal set subinterval in the left corner of a larger causal set interval.}
\label{corner}
\end{figure}
\FloatBarrier
For the one boundary setup, in the limit $\ell\ll L$ and the UV cutoff $a\rightarrow 0$, the expected entanglement entropy scaling with the UV cutoff is \cite{Calabrese_2004}
\begin{equation}
S_{corner} = \frac{1}{6} \ln\left(\frac{\ell}{a}\right) + b,
\label{1bndryS}
\end{equation}
where $b$ is a non-universal constant. Note that a logarithmic scaling with the UV cutoff in $1+1$D is consistent with the spatial area law of entanglement entropy \cite{Chandran_2016}.
Our choice of UV cutoff is the causal set discreteness scale $a\equiv1/\sqrt{\rho}=2L/\sqrt{N_L}=2\ell/\sqrt{N_\ell}$. In our calculations we hold fixed the volumes of the spacetime intervals $L=50$ and $\ell/L=2/5$ and vary $a$ by changing the number of sprinkled elements $N_L$ in the larger diamond.
Our results for the entanglement entropy, computed through \eref{geneig} and \eref{S}, versus $\frac{\sqrt{N_\ell}}{4\pi}$ are shown in Figure \ref{side_diamond_S}. $\frac{\sqrt{N_\ell}}{4\pi}$ is the minimum eigenvalue of $i\mathbf{\Delta}$ in the subinterval and is proportional to $1/a$. Included in the figure is a best fit logarithmic scaling which is
\begin{equation}
S = (0.165 \pm 0.0195)\ln{\left(\frac{\sqrt{N_\ell}}{4\pi}\right)} + (1.553 \pm 0.0278),
\end{equation}
in good agreement with the expected scaling and coefficient in \eref{1bndryS}.
\begin{figure}[h!]
\centering
\includegraphics[width=16cm]{side_diamond_final.png}
\caption{The scaling of entanglement entropy with respect to the UV cutoff, for the corner diamond configuration. The raw data has been plotted (orange), as well as averaged data points where binning has been performed (blue). The error bars give the standard deviation of the bins (blue). A function of the form \(\alpha\ln(x)+\beta\), shown in black, was fit to the binned points with coefficient \(\alpha = 0.165 \pm 0.0195\), and \(\beta = 1.553 \pm 0.0278\).}
\label{side_diamond_S}
\end{figure}
\FloatBarrier
\section{Entanglement Entropy of Two 1+1D Disjoint Intervals}
\label{sec:2int}
The entanglement entropy associated with multiple disjoint intervals has been a topic of great interest \cite{Arias_2018, Alba_2010, Casini_2004, Calabrese:2010he, Ruggiero_2018, PhysRevD.100.106015, hartman2013entanglement}. A feature which makes these intervals interesting to study in conformal field theory, but also very difficult to perform calculations with, is that the entanglement entropy depends on the full operator content of the theory \cite{Alba_2010}. This is in contrast to the entanglement entropy of a single interval which only depends on the central charge. Disjoint intervals also allow for the study of more complex entanglement properties, as well as the study of mutual information and relative entropy. Relative entropy is argued by some to be a better measure of microscopic degrees of freedom than entanglement entropy, for reasons including that it is finite in the UV or continuum limit \cite{hollands2018entanglement, Witten_2018, swingle2010mutual}. We will study the mutual information in Section \ref{sec:mutualinfo}.
\subsection{Entanglement Entropy}
In this subsection, we carry out a similar calculation to the one in Section \ref{sec:corner}, but considering a subregion comprised of two disjoint causal diamonds (with the same volume) within a larger single diamond that does not share any boundaries with the subintervals. The disjoint causal intervals are shown in Figure \ref{two_intervals}. We place the subdiamonds away from the boundaries of the larger diamond, such that they are in the regime where the SJ state resembles that of the Minkowski vacuum with an IR cutoff. For two intervals, in the limit $\ell\ll L$ and $a\rightarrow 0$, the entanglement entropy scaling with $a$ is expected to be \cite{Casini_2004}
\begin{equation}
S_{2int} = \frac{2}{3} \ln\left(\frac{\ell}{a}\right) + b',
\label{2intS}
\end{equation}
where $b'$ is once again a non-universal constant.
\begin{figure}[h!]
\centering
\includegraphics[width=10cm]{minkowski_pair.png}
\caption{Two disjoint subdiamonds within a larger causal set interval. }
\label{two_intervals}
\end{figure}
\FloatBarrier
Since the global region is a single causal interval, as before, the SJ prescription and the first truncation go through identically to the previous section. Our second truncation, however, is different. This is because we are restricting to a different domain, namely that of two disjoint diamonds. Though the domain and eigenspace are different in this case, knowledge of the single diamond case is all that is required to generalise everything to two (and more) intervals.
Let us once again return to the continuum theory, to motivate the second truncation we implement. Since the eigenspace of $i\mathbf{\Delta}$ with non-zero eigenvalues in each domain dictates which contributions we should keep, let us consider the eigenvalue problem \eref{ceig} in the union of the disjoint diamonds
\begin{equation}
-\frac{i}{2} \int_{\mathcal{R}_2} (\Theta(u-u') + \Theta(v-v') - 1) f(u,v) du\,dv = \Lambda f(u',v'),
\label{int2int}
\end{equation}
where the domain of integration, $\mathcal{R}_2\equiv \Diamond_1\cup \Diamond_2$, is now a union of two causally disjoint diamonds. We can also define this region in lightcone coordinates as
\begin{equation}
\mathcal{R}_2 \equiv
\begin{array}{ccc}
-(2\ell+d) < u < -d & \textrm{ and } & d < v < 2 \ell +d \\ & \cup & \\ d < u < 2\ell + d & \textrm{ and } & -(2\ell+d) < v < -d,
\end{array}
\end{equation}
where $d$ is the half diagonal separation. It is equal to half the shortest vertex to vertex separation between the subdiamonds, divided by $\sqrt{2}$. Since $i\mathbf{\Delta}$ is the spacetime commutator, it vanishes at pairs of points or elements that are spacelike to one another. The two causal intervals are spacelike to one another, and thus the field at every point in one diamond commutes with the field at every point in the other diamond. We can then see that $i\mathbf{\Delta}$ is block diagonal in the position basis, with a block for each disjoint interval.
With this in mind, we can break up the integral \eref{int2int} into a sum of two integrals, each over one of the subdomains $\Diamond_1$ and $\Diamond_2$. The eigenfunctions with non-zero eigenvalues can be constructed as piecewise functions, analogous to the single diamond eigenfunctions \eref{single_diamond_soln1} and \eref{single_diamond_soln2}, with support on one causal interval at a time\footnote{If there are degenerate eigenvalues, any linear combination of their respective eigenfunctions is also a valid eigenfunction. Therefore, if there was an eigenfunction in $\Diamond_1$ and another one in $\Diamond_2$ with the same eigenvalue, one could take a linear combination of them which would be non-zero in both diamonds. In the causal set, because the two diamonds are not identical due to Poisson fluctuations, an exact degeneracy is much less likely, so this makes the eigenfunctions have support on only one diamond at a time.}:
\begin{equation}
f_k(u, v) = \left\{ \begin{array}{ll}
e^{-iku} - e^{-ikv} & \{u,v\} \in \Diamond_1 \: \textrm{or} \: \Diamond_2 \\
0 & \textrm{otherwise,}
\end{array} \right.
\label{disjoint_diamond_soln1}
\end{equation}
\begin{equation}
k = \frac{n\pi}{\ell}, \quad n \in \mathbb{Z}\backslash 0, \nonumber
\end{equation}
and
\begin{equation}
g_k(u, v) = \left\{ \begin{array}{ll}
e^{-iku} + e^{-ikv} - 2\cos{(k\ell)} & \{u,v\} \in \Diamond_1 \: \textrm{or} \: \Diamond_2 \\
0 & \textrm{otherwise,}
\end{array} \right.
\label{disjoint_diamond_soln2}
\end{equation}
\begin{equation}
k \in \mathcal{K}= \{k \in \mathbb{R} | \tan{(k\ell)}=2k\ell \: \textrm{and} \: k \neq 0 \},
\nonumber
\end{equation}
where for simplicity we have expressed each function in the coordinate system with its origin at the center of the diamond in which it has support. The eigenvalues are $\Lambda=\ell/k$, as before. We also empirically verified in the causal set that the eigenfunctions have support on one interval at a time and are approximately plane wave-like. That the eigenfunctions \eref{disjoint_diamond_soln1} and \eref{disjoint_diamond_soln2} are the full set of eigenfunctions follows directly from the single diamond assertion of this.
Knowing the form of the eigenfunctions now enables us to define a truncation scheme for the disjoint diamonds. We can follow the same reasoning as in Section \ref{sec:trunc} to conclude that the minimum eigenvalue magnitude in each interval $i$ is
\begin{equation}
\Lambda_{min,i}^{cs} = \frac{\sqrt{N_{i}}}{4\pi},
\end{equation}
where $N_i$ is the number of elements in the region $i$. For simplicity, we will consider the case where the disjoint diamonds have the same volume, such that $N_1\simeq N_2$\footnote{For the case of unequal volumes, the truncations must be performed separately in each disjoint region. Having equal volumes facilitates implementing the truncation, as the same minimum eigenvalue magnitude is valid over the entire union region.}, and for ease of calculation and presentation, we define $N_\ell := \langle N_i \rangle \simeq N_i$.
Once again our UV cutoff is the causal set discreteness scale $a\equiv1/\sqrt{\rho}$. In our calculations we hold fixed the volumes of the spacetime intervals $L=90$ and $\ell_1=\ell_2=10$, and the diamond half diagonal separation $d=2.2$. We vary $a$ by varying the number of sprinkled elements $N_L$ in the larger diamond.
Our results for the entanglement entropy, computed through \eref{geneig} and \eref{S}, versus $\frac{\sqrt{N_\ell}}{4\pi}$ are shown in Figure \ref{two_diamond_S}. Included in the figure is a best fit logarithmic scaling which is
\begin{equation}
S = (0.669 \pm 0.0207)\ln{\frac{\sqrt{N_\ell}}{4\pi}} + (4.642 \pm 0.0077),
\end{equation}
in good agreement with the expected scaling and coefficient in \eref{2intS}.
\begin{figure}[h!]
\centering
\includegraphics[width=16cm]{double_diamond_final.png}
\caption{The scaling of entanglement entropy with respect to the UV cutoff, for the two disjoint subdiamond configuration. The raw data has been plotted (orange), as well as averaged data points where binning has been performed (blue). The error bars give the standard deviation of the bins (blue). A function of the form \(\alpha\ln(x)+\beta\), shown in black, was fit to the binned points with coefficient \(\alpha = 0.669 \pm 0.0207\), and \(\beta = 4.642 \pm 0.0077\).}
\label{two_diamond_S}
\end{figure}
\FloatBarrier
\subsection{Mutual Information}
\label{sec:mutualinfo}
Now that we can compute the entanglement entropy for single intervals as well as unions of two intervals, we can study the mutual information. The mutual information for disjoint regions $A$ and $B$ is defined as \cite{Witten_2020}
\begin{equation}
I_{A:B} := S_{A \cup B} - S_{A} - S_{B}.
\end{equation}
It is a non-negative quantity and for disjoint intervals such as the diamonds we considered above, the mutual information is expected to decay as the separation between the intervals increases \cite{Furukawa_2009}. This is intuitive as the field degrees of freedom become more and more uncorrelated as the spacelike separation between them increases. We will confirm this decay below.
The causal set setup we consider is shown in Figure \ref{two_interval_side}. We place the subdiamonds at the extreme left and right corners of the larger diamond, to maximize the separation between them. We hold fixed the sizes of the subintervals, $\ell_1=\ell_2=5$, as well as the sprinkling density $\rho=10$. We then increase the separation between the two intervals by keeping the subintervals in the left and right corners of the larger diamond while increasing the size of the larger global diamond.
\begin{figure}[h!]
\centering
\includegraphics[width=10cm]{side_diamond_pair_final.png}
\caption{A figure showing the two disjoint diamonds configuration where we investigate the mutual information.}
\label{two_interval_side}
\end{figure}
\FloatBarrier
The result for the scaling of the mutual information $I_{1:2}$ versus the separation between the two intervals, $\Delta x$, measured using the horizontal separation between the innermost corners, is shown in Figure \ref{mutual_info}. As expected, the mutual information is non-negative and decreases as the separation $\Delta x$ increases.
The decay trend is approximated by
\begin{equation}
I_{1:2} = \alpha \log \left(\frac{\Delta x}{\Delta x+30}\right) + \beta\left(\frac{\Delta x}{\Delta x+30} -1\right),
\label{mi_fit}
\end{equation}
with best fit parameters \(\alpha = -0.0926 \pm 0.00797\) and \(\beta = -1.3783 \pm 0.01342\). The reason $30$ appears in \eref{mi_fit} is that this is approximately the sum of the diameters of the two subdiamonds, and how far they are from one another should be considered with respect to this scale. The curve \eref{mi_fit} is shown together with the data in Figure \ref{mutual_info}. Once again, as anticipated, $I_{1:2}\rightarrow0$ as $\Delta x\rightarrow\infty$ in \eref{mi_fit}. Also, as expected, there is a divergence as the two disjoint regions approach one another \cite{swingle2010mutual}.
\begin{figure}[h!]
\centering
\includegraphics[width=15cm]{mutual_info.png}
\caption{A figure showing the decay of mutual information as two disjoint diamonds become far separated from one another relative to their own sizes. The vertical axis gives the value for mutual information and the horizontal axis \(\Delta x\) is the separation between the two regions.}
\label{mutual_info}
\end{figure}
\FloatBarrier
\section{Entanglement Entropy of Three 1+1D Disjoint Intervals}
\label{sec:3int}
As a final example, we consider the entanglement entropy of three disjoint causal intervals. The causal set setup is illustrated in Figure \ref{three_intervals}. Once again for ease of applying the truncations, we set the volumes of the three diamonds to be equal, though more general configurations can be considered as well.
Similar to the case of two diamonds in Section \ref{sec:2int}, none of the subdiamonds share any boundaries with the larger diamond they lie within, and they reside in the region where the SJ state resembles the Minkowski vacuum with an IR cutoff.
\begin{figure}[h!]
\centering
\includegraphics[width=10cm]{minkowski_triple.png}
\caption{Three disjoint subdiamonds within a larger causal set interval.}
\label{three_intervals}
\end{figure}
\FloatBarrier
For three intervals, in the limit $\ell\ll L$ and $a\rightarrow 0$, the entanglement entropy scaling with $a$ is expected to be
\begin{equation}
S_{3int} = \ln\left(\frac{\ell}{a}\right) + b'',
\label{3intS}
\end{equation}
where $b''$ is again a non-universal constant.
As above, we hold fixed the sizes of the three intervals, $\ell_1=\ell_2=\ell_3=5$, as well as the size of the larger diamond, $L=50$. We vary the sprinkling density (by changing $N$) to vary $a$. The first truncation is the same as before, and in the second truncation, the minimum magnitude eigenvalue of $i\mathbf{\Delta}$ in each interval $i$ is
\begin{equation}
\Lambda_{min,i}^{cs} = \frac{\sqrt{N_{i}}}{4\pi},
\end{equation}
where $N_i$ is the number of elements in the region $i$.
Our results for the entanglement entropy, computed through \eref{geneig} and \eref{S}, versus $\frac{\sqrt{N_\ell}}{4\pi}$ are shown in Figure \ref{three_diamond_S}. Included in the figure is a best fit logarithmic scaling which is
\begin{equation}
S = (0.998 \pm 0.0691)\ln{\frac{\sqrt{N_\ell}}{4\pi}} + (6.821 \pm 0.0307),
\end{equation}
yet again in good agreement with the expected coefficient of \(1\) on the logarithm as in \eref{3intS}.
\begin{figure}[h!]
\centering
\includegraphics[width=16cm]{triple_diamond_final.png}
\caption{The scaling of entanglement entropy with respect to the UV cutoff, for the three disjoint subdiamond configuration. The raw data has been plotted (orange), as well as averaged data points where binning has been performed (blue). The error bars give the standard deviation of the bins (blue). A function of the form \(\alpha\ln(x)+\beta\), shown in black, was fit to the binned points with coefficient \(\alpha = 0.998 \pm 0.0691\), and \(\beta = 6.821 \pm 0.0307\).}
\label{three_diamond_S}
\end{figure}
\FloatBarrier
\section{Conclusions and Future Directions}
In this paper we have studied the entanglement entropy of a scalar field in disjoint intervals or causal diamonds in causal set theory. We used a spacetime definition of entanglement entropy in terms of correlation functions and found that it is especially suited to calculations involving disjoint causal regions. Not only does the definition admit a covariant UV cutoff, it is also applicable in causal set theory, where conventional hypersurface-based calculations cannot be used. One reason that the spacetime formalism we use is natural for disjoint regions is that the spacetime commutator plays a central role in it. The commutator is only nonzero at pairs of points that are causally related, and this simplifies things considerably. The subtleties that need to be taken care of (e.g. truncations in the causal set case) in the multiple interval cases boil down to the same ones that are present in the single interval case. In two spacetime dimensions, the single interval case is well studied, facilitating the multiple interval generalisations.
We specifically considered the cases of two and three disjoint causal intervals. We studied the entanglement entropy scaling with respect to the UV cutoff, which is the discreteness scale in the causal set, and found agreement with the expected area law scalings from other similar studies. We also considered the mutual information of two disjoint intervals and verified that it decays to zero as we increase the separation between the two intervals.
There are many extensions of our work that would be interesting to explore in future research. As mentioned, we primarily focused on the entanglement entropy scalings with the UV cutoff. One can also study the scalings with respect to the other length scales in the problem, namely the sizes of each of the diamonds (including both the global one and the subdiamonds). These scalings in the scalar theory have in prior work been analytically challenging to derive. Our work presents a framework where the numerical study of this is possible. As we have mentioned, the formalism we use is suited for the study of multiple intervals, as much of what we know from the single interval case carries over. This is true for higher dimensions as well, and it would be interesting to extend these results to higher dimensions. In order to do this, first the single interval case in higher dimensions needs to be better studied. The causal set entanglement entropy calculations could be a useful tool in other fields as well, as they make otherwise challenging calculations comparably easier to perform, as demonstrated by our examples.
Our techniques could also be used to study the nature of the non-universal constant (the $b$’s or $\beta$’s above) in the entanglement entropy. There is a large literature on entanglement entropy in topological quantum field theories (e.g. \cite{Kitaev_2006, PhysRevLett.96.110405, tqft1}). In these works, the entanglement entropy separates into a contribution that scales with the boundary of the subsystem and a remaining contribution that does not. The latter is called the topological entanglement entropy, and it can be isolated via the addition and subtraction of entropies of subdivisions of the region of interest. In the same manner, perhaps similar manipulations can be implemented in the context of our work, to isolate some (topological or otherwise) properties of the non-universal constant.
Finally, we have focused on entanglement entropies in the context of causal set theory. The causal set calculations are more straightforward than their continuum counterparts, and they are also more fundamental. However, the formalism we use is equally applicable in the continuum theory. It would be interesting to investigate entanglement entropies of multiple intervals in the continuum, in the same manner as we have done above in the causal set. This would offer the possibility of deriving analytic results in this context.
\newline
\bf Acknowledgements: \rm
We thank Horacio Casini and Fay Dowker for helpful discussions. YY acknowledges financial support from Imperial College London through an Imperial College Research Fellowship grant.
\section*{References}
\bibliographystyle{iopart-num-long}
\section{Introduction}
Entanglement entropy has often been considered a pathway to quantum gravity, ever since its introduction by Rafael Sorkin \cite{Sorkin1983}, where a connection to black hole entropy was shown. In that work, the entanglement entropy of a quantum scalar field in the exterior of a black hole was shown to scale as the spatial area \(A\) of the event horizon, in units of the UV cutoff, \(A/\ell_{UV}^2\). Likewise, the entropy of a black hole was known to scale as the area of the event horizon \cite{PhysRevD.7.2333}, leading to the proposal that the Bekenstein-Hawking entropy may be wholly or in large part entanglement entropy. Much subsequent research has investigated potential origins of black hole entropy \cite{Mathur_2005, Horowitz:1996qd, Rovelli_1996, Carlip_2015, Dou:2003af}, including entanglement entropy as a leading candidate \cite{PhysRevD.34.373, Jacobson_2013, Solodukhin_2011, Emparan_2006}. Presently, it remains an important open question as to what the true microscopic source of black hole entropy is.
In addition to its role in quantum gravity, entanglement entropy has proven to be a rich topic with many applications in other areas of physics. A number of interesting theorems have been derived using entanglement entropy in quantum field theory, and especially in conformal field theory. These include c-theorems which relate a function of the coupling constants of the theory to the Virasoro central charge \cite{Casini_2004, Calabrese_2004, Myers_2010}.
In AdS/CFT, Ryu and Takayanagi found that the entanglement entropy is equal to $1/4$ the area of a minimal surface \cite{Ryu_2006, Ryu2_2006}.
In condensed matter physics, features such as topological order and properties of Fermi surfaces have been studied using entanglement entropy \cite{Kitaev_2006, PhysRevLett.96.110405, Swingle_2010}. In quantum information, entanglement entropy is used to constrain processes such as teleportation, and to study the properties of states \cite{Vidal_2002}.
Just as there are many different applications of entanglement entropy, there are many different methods for computing it. These include heat kernel methods \cite{Solodukhin_2011}, the replica trick \cite{Callan_1994, Hertzberg_2012}, the Ryu-Takayanagi formula \cite{Ryu_2006}, Euclidean path integral methods \cite{Rosenhaus_2014, Rosenhaus_2015
, using spatial correlators \cite{PhysRevD.34.373, Peschel_2003}, and using spacetime correlators \cite{Sorkin:2012sn, Chen_2020}. The spacetime correlator method is what we use in this paper. The choice of technique used depends both on the physical features and requirements of the theory, as well as what is practically solvable. In this paper we are interested in the entanglement entropy of a Gaussian scalar field in disjoint $1+1$D spacetime intervals (causal diamonds) in a causal set. Causal sets do not admit a meaningful notion of a state on a hypersurface, which is clear upon consideration of a maximal antichain\footnote{An antichain is a subset of the causal set consisting of causally unrelated elements. A \emph{maximal} antichain is one that is inextendable, i.e. one cannot add more elements to the set such that they all remain causally unrelated.}, the analogue of a spatial hypersurface. Due to the random spatio-temporal discreteness of the causal set, causal connections can pass through the antichain without making an imprint on it. This means that the maximal antichain cannot carry all the information associated with spacetime elements to its past and future, and therefore cannot serve as a Cauchy surface. This renders unusable many of the conventional spatial methods for computing entanglement entropy.
The method for computing entanglement entropy from spacetime correlation functions of the field, which we work with, was first introduced in \cite{Sorkin:2012sn}. We will review this formulation in Section \ref{sec:SfromW}. In addition to the requirement from causal set theory to work with spacetime rather than spatial quantities, there are also more general physical motivations to do so. An ultraviolet cutoff is required to obtain a finite and well-defined entanglement entropy, and if the entropy is computed on a spatial hypersurface, the cutoffs used are spatial and therefore not covariant. Ambiguities can then arise in making a choice of frame and cutoff, which are physically undesirable, especially in the context of gravity. The entanglement entropy formulation we work with has the advantage of admitting the use of a covariant spacetime UV cutoff.
In Section \ref{sec:oneint} we review the single interval entanglement entropy in a causal set, and we present new results for the case where the interval shares a boundary with the global region. We then discuss the multiple interval case in Sections \ref{sec:2int} and \ref{sec:3int}. We consider two and three disjoint regions, though our work easily generalises to any number of disjoint intervals. Treating disjoint intervals turns out to be another strength of working with the spacetime correlator formulation of entanglement entropy. This is because the details of the multiple interval calculation follow very closely those of the single interval case, which has been well studied in \cite{Afshordi_2012, Saravani_2014, Sorkin_2018}, and only additionally requires the use of greater computational power.
\section{Entropy from Spacetime Correlation Functions}
\label{sec:SfromW}
Conventionally, entanglement entropy is defined to be
\begin{equation}
S = -\Tr(\rho_{red} \ln{\rho_{red}}),
\label{ee}
\end{equation}
where $\rho_{red}$ is the reduced density matrix after tracing out some degrees of freedom from an initially pure global state $\rho$. The degrees of freedom traced out lie within one of two causally complementary partitions of the spacetime (or in practice, Cauchy hypersurface).
The definition of entanglement entropy we work with is equivalent to \eref{ee}, but has been rewritten explicitly in terms of spacetime correlation functions. For a derivation and further details we refer the reader to \cite{Sorkin:2012sn}. It was shown in \cite{Sorkin:2012sn} that the entropy of a Gaussian scalar field in a spacetime region $\mathcal{R}$ is
\begin{equation}
S = \sum_{\lambda} \lambda \log{|\lambda|},
\label{S}
\end{equation}
where \(\lambda\) are the eigenvalues obtained by solving the generalised eigenvalue problem
\begin{equation}
\mathbf{W}v = i \lambda \mathbf{\Delta} v,\indent \mathbf{\mathbf{\Delta}} v\neq 0.
\label{geneig}
\end{equation}
\(i\mathbf{\Delta}\) is the Pauli-Jordan function, or spacetime commutator of the field operators, and \(\mathbf{W}\) is the Wightman, or two-point correlation function.
For a Gaussian scalar field theory,
\begin{equation}
i\mathbf{\Delta}(x,y) := \langle0| [\Phi(x), \Phi(y)] |0\rangle=[\Phi(x), \Phi(y)],
\end{equation}
and
\begin{equation}
\mathbf{W}(x,y) := \langle0| \Phi(x) \Phi(y) |0\rangle,
\end{equation}
where $x,y \in \mathcal{R}$. Note also the relations $ i\mathbf{\Delta}(x,y)= \mathbf{W}(x,y)- \mathbf{W}(y,x)= \textrm{Im} (\mathbf{W}(x,y))/2$.
If we restrict \(\mathbf{\Delta}\) and \(\mathbf{W}\) to a subdomain of $\mathcal{R}$ with nonempty causal complement, the entropy that is calculated via \eref{S} will correspond to the entanglement entropy between the subdomain and its complement. It is worth highlighting again that the formulas above are covariant and they solely involve spacetime correlation functions.
The formulas \eref{S} and \eref{geneig} have been used in several settings both in the continuum theory and in causal set theory to compute entanglement entropies \cite{Saravani_2014, Sorkin_2018, Belenchia_2018, Surya_2021, Mathur_2021}. It has also been shown that these formulae capture the entanglement entropy even in non-Gaussian and interacting theories, up to first order in perturbation theory \cite{Chen_2020}. For theories where $ i\mathbf{\Delta}$ and $\mathbf{W}$ are known both in the continuum and in the causal set, the causal set calculations are easier to perform than the continuum ones. This is because $ i\mathbf{\Delta}$ and $\mathbf{W}$ are finite dimensional matrices and the generalised eigenvalue problem \eref{geneig} can be numerically solved relatively easily. In the continuum, $ i\mathbf{\Delta}$ and $\mathbf{W}$ are infinite dimensional, and must be made into finite dimensional matrices with respect to a choice of cutoff (see e.g. \cite{Saravani_2014}). This can be a challenging task in practice.
In this work we focus on the causal set case, for practical ease, and also because it is more fundamental and equipped with a natural UV cutoff. In the causal set, the entanglement entropy is finite, and the degrees of freedom are finite and well-defined with respect to the discreteness scale, though as we will see, extra ``truncations'' are necessary to obtain meaningful results.
\section{Entanglement Entropy of a Single 1+1D Causal Interval}
\label{sec:oneint}
In order to use \eref{S} and \eref{geneig} to study the entanglement entropy, we need a vacuum state, or Wightman function \(\mathbf{W}\). The only known means to define a Wightman function in causal set theory is via the Sorkin-Johnston (SJ) prescription \cite{Johnston_2008, johnston2010quantum, Afshordi2_2012, Sorkin:2017fcp}. This prescription is very much in the same spirit as the entropy formulation we reviewed in the previous section, as it relies solely on spacetime quantities to define a vacuum state, and does not make reference to any spatial quantities on hypersurfaces. We review this prescription below.
\subsection{The SJ Prescription}
The starting point of the SJ prescription is the retarded Green function. Since a unique retarded Green function exists only in globally hyperbolic spacetimes, the SJ vacuum is unique only in globally hyperbolic spacetimes. We will always work with causal set sprinklings\footnote{For a review of causal set theory we refer the reader to \cite{PhysRevLett.59.521, Surya_2019}.} approximated by $1+1$D causal intervals (diamonds) in Minkowski spacetime, as illustrated in Figure \ref{diamond}. These intervals are globally hyperbolic, and therefore have a unique retarded Green function and SJ state. We will work with a massless theory, for simplicity, and a Gaussian theory, since this is what the SJ prescription provides.
\begin{figure}[H]
\centering
\includegraphics[width=10cm]{single_diamond.png}
\caption{A causal set sprinkling of a causal diamond or interval in $1+1$D Minkowski spacetime.}
\label{diamond}
\end{figure}
\FloatBarrier
The retarded Green function in $1+1$D Minkowski spacetime for a massless scalar field is
\begin{equation}
G_R(x, y) = \frac{1}{2}\Theta(\tau^2)\Theta(x^{0} - y^{0}),
\end{equation}
where \(\Theta\) is the Heaviside step function, and \(\tau\) is the proper time, defined via \(\tau^2=(x^0-y^0)^2-(x^i-y^i)^2\). This propagator is very simple, returning a value of one half if \(y\) is in \(x\)'s past lightcone. It is reminiscent of the causal matrix in the causal set, which is defined as
\begin{equation}
\mathbf{C_{xy}} := \left\{ \begin{array}{ll}
1 & x \prec y, \: \textrm{and} \:\, x \neq y\\
0 & \textrm{otherwise}.
\end{array} \right.
\end{equation}
We have used the notation that the indices $_{xy}$ designate the matrix element relating causal set elements $x$ and $y$, and $x \prec y$ denotes that $x$ causally precedes $y$. Indeed, the causal set retarded (and advanced) Green function, $\mathbf{K}_{R}$ (and $\mathbf{K}_{A}$), is defined precisely using the causal matrix:
\begin{equation}
\mathbf{K}_{R} = \frac{1}{2}\mathbf{C}^T, \quad \textrm{and} \quad \mathbf{K}_{A} = \mathbf{K}_{R}^{T} = \frac{1}{2}\mathbf{C},
\label{kra}
\end{equation}
where we have used the symbol $\mathbf{K}$ for the causal set Green function, to distinguish it from the continuum one $G$, and the superscript $^T$ refers to the transpose operation.
From $\mathbf{K}_{R,A}$ we can construct $\mathbf{\Delta}=\mathbf{K}_{R}-\mathbf{K}_{A}$. Because \(i\mathbf{\Delta}\) is Hermitian and antisymmetric, it can be expanded in its eigenbasis as
\begin{equation}
i\mathbf{\Delta} = \sum_{i}^N\Lambda_{i}v_iv_i^{\dagger},
\label{PJ}
\end{equation}
where $N$ is the number of elements in the causal set. \(\Lambda_{i}\) denotes an eigenvalue of \(i\mathbf{\Delta}\), with \(v_{i}\) being the corresponding eigenvector.
The SJ Wightman function, \(\mathbf{W}_{SJ}\) is defined to be the positive part of the expansion of \(i\mathbf{\Delta}\) in its eigenbasis in \eref{PJ}, i.e.
\begin{equation}
\mathbf{W}_{SJ} := \sum_{i}\Lambda_{i}v_iv_i^{\dagger}, \quad \Lambda_i > 0.
\label{CS_W}
\end{equation}
The SJ Wightman function defines a pure state for the entropy formulation we reviewed earlier, as it yields $\lambda=0 \textrm{ or } 1$ in \eref{geneig}, and hence a vanishing entropy in \eref{S}. Its restriction to a subregion with nonzero causal complement, however, will yield nontrivial $\lambda$'s, and a nonvanishing entanglement entropy.
For the spacetime interval in Figure \ref{diamond}, $W_{SJ}$ has been well studied both in the causal set and continuum \cite{Afshordi_2012}. Away from the boundaries of the diamond, which is where we will primarily place the subintervals we work with, the state resembles that of the Minkowski vacuum with an IR cutoff. Close to the left or right corner of the diamond, the state resembles that of the Minkowski vacuum in the presence of a static mirror or reflecting boundary at that corner.
\subsection{Truncations}
\label{sec:trunc}
For reasons which we will not delve into, it turns out that a stricter condition than $\mathbf{\Delta} v\neq 0$ in \eref{geneig} is necessary to obtain meaningful results in the causal set. Not only must we exclude the kernel of $i\mathbf{\Delta}$, we also need to exclude a large number of additional components corresponding to eigenvectors with nonzero but small magnitude eigenvalues \cite{Sorkin_2018}. To understand how we decide which components to keep and which to exclude, it is instructive to first review some aspects of $i\mathbf{\Delta}$ and its eigendecomposition in Minkowski spacetime.
In lightcone coordinates \(u=\frac{x+t}{\sqrt{2}}\) and \(v=\frac{t-x}{\sqrt{2}}\), the Pauli-Jordan function in the continuum theory is \cite{Afshordi_2012}
\begin{equation}
i\mathbf{\Delta}(u,v;u',v') = \frac{-i}{2}(\Theta(u-u') + \Theta(v-v') - 1).
\end{equation}
Its eigenfunctions $f$ satisfy
\begin{equation}
\int_{\mathcal{R}} i\mathbf{\Delta}(u,v;u',v') f(u,v) \: du\, dv = \Lambda f(u',v'),
\label{ceig}
\end{equation}
where the domain of integration, $\mathcal{R}$, is the causal diamond with origin at the center and $u\in[-L,L]$ and $v\in[-L,L]$. This causal interval is the continuum approximation to the causal set in Figure \ref{diamond}.
The eigenfunctions with non-zero eigenvalues\footnote{It is worth noting that this eigenspace spans the solution space of the Klein-Gordon equation.} satisfying \eref{ceig} are \cite{johnston2010quantum, Afshordi_2012}
\begin{equation}
f_k(u, v) = e^{-iku} - e^{-ikv}, \qquad k = \frac{n\pi}{L}, \: n \in \mathbb{Z}\backslash 0
\label{single_diamond_soln1}
\end{equation}
and
\begin{equation}
\begin{array}{ll}
& g_k(u, v) = e^{-iku} + e^{-ikv} - 2\cos{(kL)}, \\
& k \in \mathcal{K}= \{k \in \mathbb{R} | \tan{(kL)}=2kL \: \textrm{and} \: k \neq 0 \}.
\end{array}
\label{single_diamond_soln2}
\end{equation}
Each \(f_k(u, v)\) and \(g_k(u, v)\) has corresponding eigenvalue
\begin{equation}
\Lambda_k =\frac L k,
\end{equation}
and for large \(k\), the \(g_k\) eigenvalues tend to those of \(f_k\), i.e. $\Lambda_k = \frac{L^2}{n \pi}$. When calculating entanglement entropies in the continuum via the definition in Section \ref{sec:SfromW}, a minimum eigenvalue serving as the UV cutoff is arbitrarily set \cite{Saravani_2014}, given by \(\Lambda_{min} = \frac{L^2}{n_{max}\pi}\). Via dimensional considerations, and a direct comparison of values, it is seen that the eigenvalues in the causal set are related to those in the continuum by rescaling with a factor of the density \(\rho = \frac{N}{4L^{2}}\), such that
\begin{equation}
\Lambda^{cs} = \rho \Lambda^{cont}.
\label{lamscale}
\end{equation}
In the causal set, the value of the cutoff and minimum eigenvalue are not arbitrary. The cutoff is the discreteness scale, and the minimum eigenvalue is related to it. Given the nature of the eigenfunctions as approximate linear combinations of plane waves and the eigenvalues as inversely proportional to wavenumbers, we expect a maximum wavenumber, \(k_{max}\), near the discreteness scale. Larger wavenumbers or shorter wavelengths cannot be meaningfully supported on the causal set. This upper bound on the wavenumber leads via \eref{lamscale} to a minimum (in magnitude) expected eigenvalue given by
\begin{equation}
\Lambda^{cs}_{min} =\frac {\rho L} {k_{max}}=\frac{\sqrt{N}}{4\pi},
\end{equation}
where $N$ is the number of elements in the diamond in which the eigenvalue problem is solved. As first noted in \cite{Sorkin_2018}, there are residual contributions below this $\Lambda^{cs}_{min}$ in the spectrum of $i\mathbf{\Delta}$ in the causal set. These need to be removed through \emph{truncations}. The truncations are a procedure that take out or set to zero any components that correspond to eigenfunctions with eigenvalues smaller (in magnitude) than $\Lambda^{cs}_{min}$. The truncations need to be done at two stages of the calculation: during the SJ prescription, and after the restriction to the entangling subregion but before solving \eref{geneig}. For further details on the truncation process, we refer the reader to \cite{Sorkin_2018, YKY}.
\subsection{Subinterval in a Corner}
\label{sec:corner}
We now use everything we have reviewed above to calculate the entanglement entropy of the restriction of a massless scalar field to a causal set subinterval in the left corner of a larger one (such that they share a boundary). The causal intervals are shown in Figure \ref{corner}. Our calculation is similar to that in \cite{Sorkin_2018}, the only difference being that the subinterval considered there was concentric to the larger one and had two boundaries across which there was entanglement. In our setup there is only one boundary contributing to the entanglement entropy.
\begin{figure}[h!]
\centering
\includegraphics[width=10cm]{corner_diamond.png}
\caption{A causal set subinterval in the left corner of a larger causal set interval.}
\label{corner}
\end{figure}
\FloatBarrier
For the one boundary setup, in the limit $\ell\ll L$ and the UV cutoff $a\rightarrow 0$, the expected entanglement entropy scaling with the UV cutoff is \cite{Calabrese_2004}
\begin{equation}
S_{corner} = \frac{1}{6} \ln\left(\frac{\ell}{a}\right) + b,
\label{1bndryS}
\end{equation}
where $b$ is a non-universal constant. Note that a logarithmic scaling with the UV cutoff in $1+1$D is consistent with the spatial area law of entanglement entropy \cite{Chandran_2016}.
Our choice of UV cutoff is the causal set discreteness scale $a\equiv1/\sqrt{\rho}=2L/\sqrt{N_L}=2\ell/\sqrt{N_\ell}$. In our calculations we hold fixed the volumes of the spacetime intervals $L=50$ and $\ell/L=2/5$ and vary $a$ by changing the number of sprinkled elements $N_L$ in the larger diamond.
Our results for the entanglement entropy, computed through \eref{geneig} and \eref{S}, versus $\frac{\sqrt{N_\ell}}{4\pi}$ are shown in Figure \ref{side_diamond_S}. $\frac{\sqrt{N_\ell}}{4\pi}$ is the minimum eigenvalue of $i\mathbf{\Delta}$ in the subinterval and is proportional to $1/a$. Included in the figure is a best fit logarithmic scaling which is
\begin{equation}
S = (0.165 \pm 0.0195)\ln{\left(\frac{\sqrt{N_\ell}}{4\pi}\right)} + (1.553 \pm 0.0278),
\end{equation}
in good agreement with the expected scaling and coefficient in \eref{1bndryS}.
\begin{figure}[h!]
\centering
\includegraphics[width=16cm]{side_diamond_final.png}
\caption{The scaling of entanglement entropy with respect to the UV cutoff, for the corner diamond configuration. The raw data has been plotted (orange), as well as averaged data points where binning has been performed (blue). The error bars give the standard deviation of the bins (blue). A function of the form \(\alpha\ln(x)+\beta\), shown in black, was fit to the binned points with coefficient \(\alpha = 0.165 \pm 0.0195\), and \(\beta = 1.553 \pm 0.0278\).}
\label{side_diamond_S}
\end{figure}
\FloatBarrier
\section{Entanglement Entropy of Two 1+1D Disjoint Intervals}
\label{sec:2int}
The entanglement entropy associated with multiple disjoint intervals has been a topic of great interest \cite{Arias_2018, Alba_2010, Casini_2004, Calabrese:2010he, Ruggiero_2018, PhysRevD.100.106015, hartman2013entanglement}. A feature which makes these intervals interesting to study in conformal field theory, but also very difficult to perform calculations with, is that the entanglement entropy depends on the full operator content of the theory \cite{Alba_2010}. This is in contrast to the entanglement entropy of a single interval which only depends on the central charge. Disjoint intervals also allow for the study of more complex entanglement properties, as well as the study of mutual information and relative entropy. Relative entropy is argued by some to be a better measure of microscopic degrees of freedom than entanglement entropy, for reasons including that it is finite in the UV or continuum limit \cite{hollands2018entanglement, Witten_2018, swingle2010mutual}. We will study the mutual information in Section \ref{sec:mutualinfo}.
\subsection{Entanglement Entropy}
In this subsection, we carry out a similar calculation to the one in Section \ref{sec:corner}, but considering a subregion comprised of two disjoint causal diamonds (with the same volume) within a larger single diamond that does not share any boundaries with the subintervals. The disjoint causal intervals are shown in Figure \ref{two_intervals}. We place the subdiamonds away from the boundaries of the larger diamond, such that they are in the regime where the SJ state resembles that of the Minkowski vacuum with an IR cutoff. For two intervals, in the limit $\ell\ll L$ and $a\rightarrow 0$, the entanglement entropy scaling with $a$ is expected to be \cite{Casini_2004}
\begin{equation}
S_{2int} = \frac{2}{3} \ln\left(\frac{\ell}{a}\right) + b',
\label{2intS}
\end{equation}
where $b'$ is once again a non-universal constant.
\begin{figure}[h!]
\centering
\includegraphics[width=10cm]{minkowski_pair.png}
\caption{Two disjoint subdiamonds within a larger causal set interval. }
\label{two_intervals}
\end{figure}
\FloatBarrier
Since the global region is a single causal interval, as before, the SJ prescription and the first truncation go through identically to the previous section. Our second truncation, however, is different. This is because we are restricting to a different domain, namely that of two disjoint diamonds. Though the domain and eigenspace are different in this case, knowledge of the single diamond case is all that is required to generalise everything to two (and more) intervals.
Let us once again return to the continuum theory, to motivate the second truncation we implement. Since the eigenspace of $i\mathbf{\Delta}$ with non-zero eigenvalues in each domain dictates which contributions we should keep, let us consider the eigenvalue problem \eref{ceig} in the union of the disjoint diamonds
\begin{equation}
-\frac{i}{2} \int_{\mathcal{R}_2} (\Theta(u-u') + \Theta(v-v') - 1) f(u,v) du\,dv = \Lambda f(u',v'),
\label{int2int}
\end{equation}
where the domain of integration, $\mathcal{R}_2\equiv \Diamond_1\cup \Diamond_2$, is now a union of two causally disjoint diamonds. We can also define this region in lightcone coordinates as
\begin{equation}
\mathcal{R}_2 \equiv
\begin{array}{ccc}
-(2\ell+d) < u < -d & \textrm{ and } & d < v < 2 \ell +d \\ & \cup & \\ d < u < 2\ell + d & \textrm{ and } & -(2\ell+d) < v < -d,
\end{array}
\end{equation}
where $d$ is the half diagonal separation. It is equal to half the shortest vertex to vertex separation between the subdiamonds, divided by $\sqrt{2}$. Since $i\mathbf{\Delta}$ is the spacetime commutator, it vanishes at pairs of points or elements that are spacelike to one another. The two causal intervals are spacelike to one another, and thus the field at every point in one diamond commutes with the field at every point in the other diamond. We can then see that $i\mathbf{\Delta}$ is block diagonal in the position basis, with a block for each disjoint interval.
With this in mind, we can break up the integral \eref{int2int} into a sum of two integrals, each over one of the subdomains $\Diamond_1$ and $\Diamond_2$. The eigenfunctions with non-zero eigenvalues can be constructed as piecewise functions, analogous to the single diamond eigenfunctions \eref{single_diamond_soln1} and \eref{single_diamond_soln2}, with support on one causal interval at a time\footnote{If there are degenerate eigenvalues, any linear combination of their respective eigenfunctions is also a valid eigenfunction. Therefore, if there was an eigenfunction in $\Diamond_1$ and another one in $\Diamond_2$ with the same eigenvalue, one could take a linear combination of them which would be non-zero in both diamonds. In the causal set, because the two diamonds are not identical due to Poisson fluctuations, an exact degeneracy is much less likely, so this makes the eigenfunctions have support on only one diamond at a time.}:
\begin{equation}
f_k(u, v) = \left\{ \begin{array}{ll}
e^{-iku} - e^{-ikv} & \{u,v\} \in \Diamond_1 \: \textrm{or} \: \Diamond_2 \\
0 & \textrm{otherwise,}
\end{array} \right.
\label{disjoint_diamond_soln1}
\end{equation}
\begin{equation}
k = \frac{n\pi}{\ell}, \quad n \in \mathbb{Z}\backslash 0, \nonumber
\end{equation}
and
\begin{equation}
g_k(u, v) = \left\{ \begin{array}{ll}
e^{-iku} + e^{-ikv} - 2\cos{(k\ell)} & \{u,v\} \in \Diamond_1 \: \textrm{or} \: \Diamond_2 \\
0 & \textrm{otherwise,}
\end{array} \right.
\label{disjoint_diamond_soln2}
\end{equation}
\begin{equation}
k \in \mathcal{K}= \{k \in \mathbb{R} | \tan{(k\ell)}=2k\ell \: \textrm{and} \: k \neq 0 \},
\nonumber
\end{equation}
where for simplicity we have expressed each function in the coordinate system with its origin at the center of the diamond in which it has support. The eigenvalues are $\Lambda=\ell/k$, as before. We also empirically verified in the causal set that the eigenfunctions have support on one interval at a time and are approximately plane wave-like. That the eigenfunctions \eref{disjoint_diamond_soln1} and \eref{disjoint_diamond_soln2} are the full set of eigenfunctions follows directly from the single diamond assertion of this.
Knowing the form of the eigenfunctions now enables us to define a truncation scheme for the disjoint diamonds. We can follow the same reasoning as in Section \ref{sec:trunc} to conclude that the minimum eigenvalue magnitude in each interval $i$ is
\begin{equation}
\Lambda_{min,i}^{cs} = \frac{\sqrt{N_{i}}}{4\pi},
\end{equation}
where $N_i$ is the number of elements in the region $i$. For simplicity, we will consider the case where the disjoint diamonds have the same volume, such that $N_1\simeq N_2$\footnote{For the case of unequal volumes, the truncations must be performed separately in each disjoint region. Having equal volumes facilitates implementing the truncation, as the same minimum eigenvalue magnitude is valid over the entire union region.}, and for ease of calculation and presentation, we define $N_\ell := \langle N_i \rangle \simeq N_i$.
Once again our UV cutoff is the causal set discreteness scale $a\equiv1/\sqrt{\rho}$. In our calculations we hold fixed the volumes of the spacetime intervals $L=90$ and $\ell_1=\ell_2=10$, and the diamond half diagonal separation $d=2.2$. We vary $a$ by varying the number of sprinkled elements $N_L$ in the larger diamond.
Our results for the entanglement entropy, computed through \eref{geneig} and \eref{S}, versus $\frac{\sqrt{N_\ell}}{4\pi}$ are shown in Figure \ref{two_diamond_S}. Included in the figure is a best fit logarithmic scaling which is
\begin{equation}
S = (0.669 \pm 0.0207)\ln{\frac{\sqrt{N_\ell}}{4\pi}} + (4.642 \pm 0.0077),
\end{equation}
in good agreement with the expected scaling and coefficient in \eref{2intS}.
\begin{figure}[h!]
\centering
\includegraphics[width=16cm]{double_diamond_final.png}
\caption{The scaling of entanglement entropy with respect to the UV cutoff, for the two disjoint subdiamond configuration. The raw data has been plotted (orange), as well as averaged data points where binning has been performed (blue). The error bars give the standard deviation of the bins (blue). A function of the form \(\alpha\ln(x)+\beta\), shown in black, was fit to the binned points with coefficient \(\alpha = 0.669 \pm 0.0207\), and \(\beta = 4.642 \pm 0.0077\).}
\label{two_diamond_S}
\end{figure}
\FloatBarrier
\subsection{Mutual Information}
\label{sec:mutualinfo}
Now that we can compute the entanglement entropy for single intervals as well as unions of two intervals, we can study the mutual information. The mutual information for disjoint regions $A$ and $B$ is defined as \cite{Witten_2020}
\begin{equation}
I_{A:B} := S_{A \cup B} - S_{A} - S_{B}.
\end{equation}
It is a non-negative quantity and for disjoint intervals such as the diamonds we considered above, the mutual information is expected to decay as the separation between the intervals increases \cite{Furukawa_2009}. This is intuitive as the field degrees of freedom become more and more uncorrelated as the spacelike separation between them increases. We will confirm this decay below.
The causal set setup we consider is shown in Figure \ref{two_interval_side}. We place the subdiamonds at the extreme left and right corners of the larger diamond, to maximize the separation between them. We hold fixed the sizes of the subintervals, $\ell_1=\ell_2=5$, as well as the sprinkling density $\rho=10$. We then increase the separation between the two intervals by keeping the subintervals in the left and right corners of the larger diamond while increasing the size of the larger global diamond.
\begin{figure}[h!]
\centering
\includegraphics[width=10cm]{side_diamond_pair_final.png}
\caption{A figure showing the two disjoint diamonds configuration where we investigate the mutual information.}
\label{two_interval_side}
\end{figure}
\FloatBarrier
The result for the scaling of the mutual information $I_{1:2}$ versus the separation between the two intervals, $\Delta x$, measured using the horizontal separation between the innermost corners, is shown in Figure \ref{mutual_info}. As expected, the mutual information is non-negative and decreases as the separation $\Delta x$ increases.
The decay trend is approximated by
\begin{equation}
I_{1:2} = \alpha \log \left(\frac{\Delta x}{\Delta x+30}\right) + \beta\left(\frac{\Delta x}{\Delta x+30} -1\right),
\label{mi_fit}
\end{equation}
with best fit parameters \(\alpha = -0.0926 \pm 0.00797\) and \(\beta = -1.3783 \pm 0.01342\). The reason $30$ appears in \eref{mi_fit} is that this is approximately the sum of the diameters of the two subdiamonds, and how far they are from one another should be considered with respect to this scale. The curve \eref{mi_fit} is shown together with the data in Figure \ref{mutual_info}. Once again, as anticipated, $I_{1:2}\rightarrow0$ as $\Delta x\rightarrow\infty$ in \eref{mi_fit}. Also, as expected, there is a divergence as the two disjoint regions approach one another \cite{swingle2010mutual}.
\begin{figure}[h!]
\centering
\includegraphics[width=15cm]{mutual_info.png}
\caption{A figure showing the decay of mutual information as two disjoint diamonds become far separated from one another relative to their own sizes. The vertical axis gives the value for mutual information and the horizontal axis \(\Delta x\) is the separation between the two regions.}
\label{mutual_info}
\end{figure}
\FloatBarrier
\section{Entanglement Entropy of Three 1+1D Disjoint Intervals}
\label{sec:3int}
As a final example, we consider the entanglement entropy of three disjoint causal intervals. The causal set setup is illustrated in Figure \ref{three_intervals}. Once again for ease of applying the truncations, we set the volumes of the three diamonds to be equal, though more general configurations can be considered as well.
Similar to the case of two diamonds in Section \ref{sec:2int}, none of the subdiamonds share any boundaries with the larger diamond they lie within, and they reside in the region where the SJ state resembles the Minkowski vacuum with an IR cutoff.
\begin{figure}[h!]
\centering
\includegraphics[width=10cm]{minkowski_triple.png}
\caption{Three disjoint subdiamonds within a larger causal set interval.}
\label{three_intervals}
\end{figure}
\FloatBarrier
For three intervals, in the limit $\ell\ll L$ and $a\rightarrow 0$, the entanglement entropy scaling with $a$ is expected to be
\begin{equation}
S_{3int} = \ln\left(\frac{\ell}{a}\right) + b'',
\label{3intS}
\end{equation}
where $b''$ is again a non-universal constant.
As above, we hold fixed the sizes of the three intervals, $\ell_1=\ell_2=\ell_3=5$, as well as the size of the larger diamond, $L=50$. We vary the sprinkling density (by changing $N$) to vary $a$. The first truncation is the same as before, and in the second truncation, the minimum magnitude eigenvalue of $i\mathbf{\Delta}$ in each interval $i$ is
\begin{equation}
\Lambda_{min,i}^{cs} = \frac{\sqrt{N_{i}}}{4\pi},
\end{equation}
where $N_i$ is the number of elements in the region $i$.
Our results for the entanglement entropy, computed through \eref{geneig} and \eref{S}, versus $\frac{\sqrt{N_\ell}}{4\pi}$ are shown in Figure \ref{three_diamond_S}. Included in the figure is a best fit logarithmic scaling which is
\begin{equation}
S = (0.998 \pm 0.0691)\ln{\frac{\sqrt{N_\ell}}{4\pi}} + (6.821 \pm 0.0307),
\end{equation}
yet again in good agreement with the expected coefficient of \(1\) on the logarithm as in \eref{3intS}.
\begin{figure}[h!]
\centering
\includegraphics[width=16cm]{triple_diamond_final.png}
\caption{The scaling of entanglement entropy with respect to the UV cutoff, for the three disjoint subdiamond configuration. The raw data has been plotted (orange), as well as averaged data points where binning has been performed (blue). The error bars give the standard deviation of the bins (blue). A function of the form \(\alpha\ln(x)+\beta\), shown in black, was fit to the binned points with coefficient \(\alpha = 0.998 \pm 0.0691\), and \(\beta = 6.821 \pm 0.0307\).}
\label{three_diamond_S}
\end{figure}
\FloatBarrier
\section{Conclusions and Future Directions}
In this paper we have studied the entanglement entropy of a scalar field in disjoint intervals or causal diamonds in causal set theory. We used a spacetime definition of entanglement entropy in terms of correlation functions and found that it is especially suited to calculations involving disjoint causal regions. Not only does the definition admit a covariant UV cutoff, it is also applicable in causal set theory, where conventional hypersurface-based calculations cannot be used. One reason that the spacetime formalism we use is natural for disjoint regions is that the spacetime commutator plays a central role in it. The commutator is only nonzero at pairs of points that are causally related, and this simplifies things considerably. The subtleties that need to be taken care of (e.g. truncations in the causal set case) in the multiple interval cases boil down to the same ones that are present in the single interval case. In two spacetime dimensions, the single interval case is well studied, facilitating the multiple interval generalisations.
We specifically considered the cases of two and three disjoint causal intervals. We studied the entanglement entropy scaling with respect to the UV cutoff, which is the discreteness scale in the causal set, and found agreement with the expected area law scalings from other similar studies. We also considered the mutual information of two disjoint intervals and verified that it decays to zero as we increase the separation between the two intervals.
There are many extensions of our work that would be interesting to explore in future research. As mentioned, we primarily focused on the entanglement entropy scalings with the UV cutoff. One can also study the scalings with respect to the other length scales in the problem, namely the sizes of each of the diamonds (including both the global one and the subdiamonds). These scalings in the scalar theory have in prior work been analytically challenging to derive. Our work presents a framework where the numerical study of this is possible. As we have mentioned, the formalism we use is suited for the study of multiple intervals, as much of what we know from the single interval case carries over. This is true for higher dimensions as well, and it would be interesting to extend these results to higher dimensions. In order to do this, first the single interval case in higher dimensions needs to be better studied. The causal set entanglement entropy calculations could be a useful tool in other fields as well, as they make otherwise challenging calculations comparably easier to perform, as demonstrated by our examples.
Our techniques could also be used to study the nature of the non-universal constant (the $b$’s or $\beta$’s above) in the entanglement entropy. There is a large literature on entanglement entropy in topological quantum field theories (e.g. \cite{Kitaev_2006, PhysRevLett.96.110405, tqft1}). In these works, the entanglement entropy separates into a contribution that scales with the boundary of the subsystem and a remaining contribution that does not. The latter is called the topological entanglement entropy, and it can be isolated via the addition and subtraction of entropies of subdivisions of the region of interest. In the same manner, perhaps similar manipulations can be implemented in the context of our work, to isolate some (topological or otherwise) properties of the non-universal constant.
Finally, we have focused on entanglement entropies in the context of causal set theory. The causal set calculations are more straightforward than their continuum counterparts, and they are also more fundamental. However, the formalism we use is equally applicable in the continuum theory. It would be interesting to investigate entanglement entropies of multiple intervals in the continuum, in the same manner as we have done above in the causal set. This would offer the possibility of deriving analytic results in this context.
\newline
\bf Acknowledgements: \rm
We thank Horacio Casini and Fay Dowker for helpful discussions. YY acknowledges financial support from Imperial College London through an Imperial College Research Fellowship grant.
\section*{References}
\bibliographystyle{iopart-num-long}
|
\section{Introduction}
Forward modeling lies at the heart of most inversion algorithms. Solving the wave equation numerically constitutes the majority of the computational cost for applications like seismic imaging and full waveform inversion (FWI). Frequency-domain solutions of the wave equation allow a reduction in dimensionality as they provide solutions of the Helmholtz equation per frequency. This approach to model wavefields gained popularity with the rise of waveform inversion~\cite{pratt1999seismic}; however, such solutions require inverting a stiffness matrix that can become computationally intractable for large 3D models or for modeling high frequencies. Moreover, conventional algorithms do not have a mechanism to transfer information gained from solving one problem to the next. For example, the same amount of computational effort is needed to model wavefields corresponding to a small perturbation in the velocity model. This results in a computational bottleneck when repeated modeling is needed for updated velocity models, as in the case of FWI.
Recently,~\cite{alkhalifah2020wavefield} demonstrated the use of physics-informed neural networks (PINNs) to address some of the challenges associated with the conventional Helmholtz solvers, including the flexibility of modeling additional physics. By considering a homogeneous background model, for which the wavefield solution is obtained analytically,~\cite{alkhalifah2020wavefield} trained a neural network to solve for the scattered wavefield. Contrary to conventional deep learning approaches, instead of a labeled set, the training of the neural network is done through a loss function, which embeds the residual of the underlying Helmholtz equation. For an input given by a location in the computational domain, the neural network learns to map it to the scattered wavefield solution at the input location, thanks to the universal approximation theorem. The key enabler of PINNs is automatic differentiation that allows for fast and accurate derivative computation of the neural network's output w.r.t. to the inputs, which is needed to evaluate the loss function. By using machine learning techniques like transfer learning and surrogate modeling, the method can be used to speed up repeated wavefield computations, as suggested by~\cite{waheed2020eikonal} for the eikonal equation. The PINN-based Helmholtz solver has recently been extended for the vertically transversely isotropic (VTI) case~\cite{song2021solving}.
However, several challenges remain to be addressed in making the PINN-based Helmholtz solver fully practicable. Chief among them is the slow convergence of the neural network in the presence of sharp heterogeneities in the velocity model. Although fully-connected neural networks can approximate arbitrary functions, they have been shown to prioritize learning low frequency features of the underlying function, and hence exhibit a bias towards smooth functions. This poses difficulty in training them to approximate, for example, solutions corresponding to high frequencies or solutions for sharply varying velocity models that generate high-frequency features in the wavefield solution. These observations have been confirmed both in theory and practice~\cite{rahaman2019spectral,cao2019towards}.
Therefore, in this study, we explore the role of activation functions in improving convergence speed for the PINN-based Helmholtz solver. We do a comparative study of different activation functions that are routinely used in the PINN literature, namely hyperbolic tangent (tanh), inverse tangent (atan), and exponential linear unit (elu). In addition, we consider another variant of the ReLU activation function, named `swish', which has shown superior performance over other activation functions on a number of challenging image classification and machine translation problems~\cite{ramachandran2017searching}. Through tests on a synthetic model, we find that the swish activation function shows markedly improved convergence rate and accuracy in solving the Helmholtz equation.
\section{Theory}
The Helmholtz equation in an acoustic, isotropic medium with constant density is given as:
\begin{equation}
\left(\nabla^2 + \frac{\omega^2}{v^2} \right) \, u(\mathbf{x}) = f(\mathbf{x}),
\label{eq:HH}
\end{equation}
where $v$ is the wave velocity, $\omega$ is the angular frequency
~The goal is to solve equation~\eqref{eq:HH} for the complex valued wavefield $u(\mathbf{x}) = \{u_r,u_i\}$ as a function of the angular frequency $\omega$. Then the time domain solution can be found using inverse Fourier transform of the superposition of the mono-frequency solutions.
To avoid the problem of sampling near the source location,~\cite{alkhalifah2020wavefield} suggests solving the Lippmann-Schwinger form of the Helmholtz equation:
\begin{equation}
\left(\nabla^2 + \frac{\omega^2}{v^2} \right) \, \delta u = -\omega^2 \, \delta m \, u_0,
\label{eq:LS}
\end{equation}
for the scattered wavefield $\delta u = u - u_0$, where $u_0$ is the background wavefield satisfying equation~\eqref{eq:HH} for the background velocity $v_0$, and $\delta m = \frac{1}{v^2}-\frac{1}{v_0^2}$ is the velocity model perturbation. By using a background model with a constant velocity in which we can calculate the background wavefield $u_0$ analytically, we use equation~\eqref{eq:LS} to calculate the scattered wavefield $\delta u$ per frequency, from which we find the solution to the Helmholtz equation \eqref{eq:HH} as:
\begin{equation*}
\begin{aligned}
u_r & = \delta u_r + u_{0r},\\
u_i & = \delta u_i + u_{0i}.\\
\end{aligned}
\end{equation*}
To solve equation~\eqref{eq:LS} using a deep neural network (DNN), we consider a network with two neurons in the input layer for the spatial coordinates ($x$,$z$), two output neurons for the real and imaginary parts of the scattered wavefield, and a number of hidden layers. The second derivatives of the outputs ($\delta u_r, \delta u_i$) w.r.t. the inputs ($x$,$z$) can be computed using automatic differentiation. We seek to minimize the following loss function:
\begin{equation}
\mathfrak{L} = \frac{1}{N}\sum_{i=1}^{N} || \omega^2 m^{(i)} \delta u_r^{(i)} + \nabla ^2 \delta u_r^{(i)} + \omega ^2 \delta m ~u_{0r}^{(i)} ||_2^2 + || \omega^2 m^{(i)} \delta u_i^{(i)} + \nabla ^2 \delta u_i^{(i)} + \omega ^2 \delta m ~u_{0i}^{(i)} ||_2^2,
\label{eq:loss}
\end{equation}
for the real and imaginary parts of the scattered wavefield $(\delta u_r, \delta u_i)$.
Here $N$ is the number of training samples, which consists of a randomly selected collocation points from the computational domain, and $i$ is the sample index. The role of this loss function is to impose the validity of the Helmholtz equation on a given number of training points. The set of network parameters $\boldsymbol{\theta}^*$ that minimize the loss function given in equation \eqref{eq:loss} on this set of training samples can be found by solving the following optimization problem:
\begin{equation}
\boldsymbol{\theta ^*} = \arg\min_{\boldsymbol{\theta}} \mathfrak{L}(x^{*},z^{*}; \boldsymbol{\theta}).
\label{eq:optm}
\end{equation}
Once the DNN is trained, we evaluate it on a regular grid to obtain the predicted scattered wavefields, which are then added to the background wavefields to obtain the final wavefield solution.
\section{Activation functions}
The choice of activation function in DNNs has a significant impact on the training dynamics and convergence of the network. Activation functions are sources of non-linearity as they transform the neural network function to a superposition of non-linear basis functions. Choosing the right basis function can make a significant impact in efficiently representing a function. Therefore, in this study, we consider activation functions that are routinely used in the PINN literature (atan, tanh, and elu) in addition to the swish activation function, which is given as $f(x) = x \cdotp sigmoid(x)$. Swish is a smooth, non-monotonic function that has consistently matched or outperformed ReLU on DNNs applied to a variety of challenging problems, including image classification and machine translation. It is believed that the reason for this improvement is that swish is able to better alleviate the problem of vanishing gradients during backpropagation.
\section{Results}
In this section, we present a comparative study of different activation functions to solve equation~\eqref{eq:LS} for a two-scatter velocity model, shown in Figure~\ref{fig:model}. We keep all the hyper-parameters fixed to ensure fairness of the comparison. We use a DNN architecture with 8 hidden layers and 20 neurons in each layer, as used by~\cite{alkhalifah2020wavefield} for the considered velocity model. We randomly select 5000 locations from the computational domain and train the PINN model using the Adam optimizer for 15,000 epochs with mini-batch optimization having a batch size of 256. The PINN model is implemented using the SciANN package~\cite[]{haghighat2021sciann} -- a high-level Tensorflow wrapper for scientific computations.
\begin{figure}[ht!]
\begin{center}
\includegraphics[width=0.34\textwidth]{figs/velmodel.pdf}
\end{center}
\caption{
A two-scatter velocity model taken from~\cite{alkhalifah2020wavefield} for the tests. The background has a homogeneous velocity of 2~km/s.}%
\label{fig:model}
\end{figure}
Figure~\ref{fig:loss} shows the loss history for training the PINN model using different activation functions considered. We observe that elu performs the worst with tanh and atan yielding similar performances, while swish has a markedly improved convergence compared to all the considered activation functions. For all activation functions, the convergence plateaus after an initial reduction of the loss value; however, swish recovers faster than the other activation functions. This can be observed by comparing the loss curves for the first 2000 epochs. We also tabulate the $L_2$ and $L_\infty$ errors for both the real and imaginary components of the wavefield solution using these activation functions (see Table~\ref{table1}). We clearly observe improved accuracy for the swish activation function compared to the rest.
\begin{figure}[ht!]
\begin{center}
\includegraphics[width=0.6\textwidth]{figs/loss.pdf}
\end{center}
\caption{
Loss history for training of the PINN model using different activation functions in the hidden layer.}%
\label{fig:loss}
\end{figure}
\begin{figure}
\centering
\subfigure[]{\includegraphics[width=0.24\textwidth]{figs/dur}}
\subfigure[]{\includegraphics[width=0.24\textwidth]{figs/dui}}
\subfigure[]{\includegraphics[width=0.24\textwidth]{figs/durerr.pdf}}
\subfigure[]{\includegraphics[width=0.24\textwidth]{figs/duierr.pdf}}
\caption{The predicted scattered wavefield solution by the PINN model for the real (a) and imaginary (b) parts. Also shown are the wavefield errors for the real (c) and imaginary (d) parts.}
\label{fig:solutions}
\end{figure}
\begin{table}
\centering
\begin{tabular}{|c|c|c|c|c|} \hline
\multirow{2}{*}{\begin{tabular}{c}Activation functions\end{tabular}} & \multicolumn{2}{c|}{$L_2$} & \multicolumn{2}{c|}{$L_\infty$}\\ \cline{2-5}
& Real & Imag. & Real & Imag. \\ \hline
tanh & $4.01\times 10^{-5}$ & $5.51\times 10^{-5}$ & $1.31\times 10^{-2}$ & $1.90\times 10^{-2}$ \\ \hline
atan & $4.15\times 10^{-5}$ & $5.53\times 10^{-5}$ & $1.09\times 10^{-2}$ & $1.81\times 10^{-2}$ \\ \hline
elu & $5.20\times 10^{-5}$ & $6.17\times 10^{-5}$ & $1.26\times 10^{-2}$ & $1.86\times 10^{-2}$ \\ \hline
swish & $3.65\times 10^{-5}$ & $4.7\times 10^{-5}$ & $9.86\times 10^{-3}$ & $1.58\times 10^{-2}$ \\ \hline
\end{tabular}
\caption{$L_2$ and $L_\infty$ errors for the real and imaginary parts of the PINN Helmholtz solution obtained using different activation functions.}
\label{table1}
\end{table}
In Figure~\ref{fig:solutions}, we show the scattered wavefield solution obtained from the PINN model trained using the swish activation function. We also plot the wavefield errors on the same color scale. The overall errors are considerably smaller than for the solutions computed using other activation functions as confirmed by Table~\ref{table1}.
\section{Conclusions}
We studied the routinely used activation functions from the PINN literature in solving the Helmholtz equation. In addition, we also considered the swish activation function, which is a variant of ReLU. Swish had shown improved performance compared to other activation functions in training DNNs for a variety of data science problems, thanks to its robustness against the problem of vanishing gradients. We also observed superior convergence using the swish activation function in solving the Helmholtz equation. Although these results are preliminary, they demonstrate the potential of selecting better activation functions to improve the convergence of PINN solvers. To focus on the activation functions in this study, we only used the Adam optimizer for all cases. However, to converge faster, typically the optimization process begins with a first-order optimizer like Adam and then switches to a second-order optimizer like L-BFGS after a suitably low value of the loss function. Seeing the loss curves, it can be noted that swish allows early switching to a second-order optimizer compared to other studied activation functions, which would compound its efficacy.
\section{Acknowledgments}
\vspace{-0.15cm}
We extend gratitude to Prof. Tariq Alkhalifah and Dr. Ehsan Haghighat for helpful discussions.
\bibliographystyle{unsrt}
\section{Introduction}
Forward modeling lies at the heart of most inversion algorithms. Solving the wave equation numerically constitutes the majority of the computational cost for applications like seismic imaging and full waveform inversion (FWI). Frequency-domain solutions of the wave equation allow a reduction in dimensionality as they provide solutions of the Helmholtz equation per frequency. This approach to model wavefields gained popularity with the rise of waveform inversion~\cite{pratt1999seismic}; however, such solutions require inverting a stiffness matrix that can become computationally intractable for large 3D models or for modeling high frequencies. Moreover, conventional algorithms do not have a mechanism to transfer information gained from solving one problem to the next. For example, the same amount of computational effort is needed to model wavefields corresponding to a small perturbation in the velocity model. This results in a computational bottleneck when repeated modeling is needed for updated velocity models, as in the case of FWI.
Recently,~\cite{alkhalifah2020wavefield} demonstrated the use of physics-informed neural networks (PINNs) to address some of the challenges associated with the conventional Helmholtz solvers, including the flexibility of modeling additional physics. By considering a homogeneous background model, for which the wavefield solution is obtained analytically,~\cite{alkhalifah2020wavefield} trained a neural network to solve for the scattered wavefield. Contrary to conventional deep learning approaches, instead of a labeled set, the training of the neural network is done through a loss function, which embeds the residual of the underlying Helmholtz equation. For an input given by a location in the computational domain, the neural network learns to map it to the scattered wavefield solution at the input location, thanks to the universal approximation theorem. The key enabler of PINNs is automatic differentiation that allows for fast and accurate derivative computation of the neural network's output w.r.t. to the inputs, which is needed to evaluate the loss function. By using machine learning techniques like transfer learning and surrogate modeling, the method can be used to speed up repeated wavefield computations, as suggested by~\cite{waheed2020eikonal} for the eikonal equation. The PINN-based Helmholtz solver has recently been extended for the vertically transversely isotropic (VTI) case~\cite{song2021solving}.
However, several challenges remain to be addressed in making the PINN-based Helmholtz solver fully practicable. Chief among them is the slow convergence of the neural network in the presence of sharp heterogeneities in the velocity model. Although fully-connected neural networks can approximate arbitrary functions, they have been shown to prioritize learning low frequency features of the underlying function, and hence exhibit a bias towards smooth functions. This poses difficulty in training them to approximate, for example, solutions corresponding to high frequencies or solutions for sharply varying velocity models that generate high-frequency features in the wavefield solution. These observations have been confirmed both in theory and practice~\cite{rahaman2019spectral,cao2019towards}.
Therefore, in this study, we explore the role of activation functions in improving convergence speed for the PINN-based Helmholtz solver. We do a comparative study of different activation functions that are routinely used in the PINN literature, namely hyperbolic tangent (tanh), inverse tangent (atan), and exponential linear unit (elu). In addition, we consider another variant of the ReLU activation function, named `swish', which has shown superior performance over other activation functions on a number of challenging image classification and machine translation problems~\cite{ramachandran2017searching}. Through tests on a synthetic model, we find that the swish activation function shows markedly improved convergence rate and accuracy in solving the Helmholtz equation.
\section{Theory}
The Helmholtz equation in an acoustic, isotropic medium with constant density is given as:
\begin{equation}
\left(\nabla^2 + \frac{\omega^2}{v^2} \right) \, u(\mathbf{x}) = f(\mathbf{x}),
\label{eq:HH}
\end{equation}
where $v$ is the wave velocity, $\omega$ is the angular frequency
~The goal is to solve equation~\eqref{eq:HH} for the complex valued wavefield $u(\mathbf{x}) = \{u_r,u_i\}$ as a function of the angular frequency $\omega$. Then the time domain solution can be found using inverse Fourier transform of the superposition of the mono-frequency solutions.
To avoid the problem of sampling near the source location,~\cite{alkhalifah2020wavefield} suggests solving the Lippmann-Schwinger form of the Helmholtz equation:
\begin{equation}
\left(\nabla^2 + \frac{\omega^2}{v^2} \right) \, \delta u = -\omega^2 \, \delta m \, u_0,
\label{eq:LS}
\end{equation}
for the scattered wavefield $\delta u = u - u_0$, where $u_0$ is the background wavefield satisfying equation~\eqref{eq:HH} for the background velocity $v_0$, and $\delta m = \frac{1}{v^2}-\frac{1}{v_0^2}$ is the velocity model perturbation. By using a background model with a constant velocity in which we can calculate the background wavefield $u_0$ analytically, we use equation~\eqref{eq:LS} to calculate the scattered wavefield $\delta u$ per frequency, from which we find the solution to the Helmholtz equation \eqref{eq:HH} as:
\begin{equation*}
\begin{aligned}
u_r & = \delta u_r + u_{0r},\\
u_i & = \delta u_i + u_{0i}.\\
\end{aligned}
\end{equation*}
To solve equation~\eqref{eq:LS} using a deep neural network (DNN), we consider a network with two neurons in the input layer for the spatial coordinates ($x$,$z$), two output neurons for the real and imaginary parts of the scattered wavefield, and a number of hidden layers. The second derivatives of the outputs ($\delta u_r, \delta u_i$) w.r.t. the inputs ($x$,$z$) can be computed using automatic differentiation. We seek to minimize the following loss function:
\begin{equation}
\mathfrak{L} = \frac{1}{N}\sum_{i=1}^{N} || \omega^2 m^{(i)} \delta u_r^{(i)} + \nabla ^2 \delta u_r^{(i)} + \omega ^2 \delta m ~u_{0r}^{(i)} ||_2^2 + || \omega^2 m^{(i)} \delta u_i^{(i)} + \nabla ^2 \delta u_i^{(i)} + \omega ^2 \delta m ~u_{0i}^{(i)} ||_2^2,
\label{eq:loss}
\end{equation}
for the real and imaginary parts of the scattered wavefield $(\delta u_r, \delta u_i)$.
Here $N$ is the number of training samples, which consists of a randomly selected collocation points from the computational domain, and $i$ is the sample index. The role of this loss function is to impose the validity of the Helmholtz equation on a given number of training points. The set of network parameters $\boldsymbol{\theta}^*$ that minimize the loss function given in equation \eqref{eq:loss} on this set of training samples can be found by solving the following optimization problem:
\begin{equation}
\boldsymbol{\theta ^*} = \arg\min_{\boldsymbol{\theta}} \mathfrak{L}(x^{*},z^{*}; \boldsymbol{\theta}).
\label{eq:optm}
\end{equation}
Once the DNN is trained, we evaluate it on a regular grid to obtain the predicted scattered wavefields, which are then added to the background wavefields to obtain the final wavefield solution.
\section{Activation functions}
The choice of activation function in DNNs has a significant impact on the training dynamics and convergence of the network. Activation functions are sources of non-linearity as they transform the neural network function to a superposition of non-linear basis functions. Choosing the right basis function can make a significant impact in efficiently representing a function. Therefore, in this study, we consider activation functions that are routinely used in the PINN literature (atan, tanh, and elu) in addition to the swish activation function, which is given as $f(x) = x \cdotp sigmoid(x)$. Swish is a smooth, non-monotonic function that has consistently matched or outperformed ReLU on DNNs applied to a variety of challenging problems, including image classification and machine translation. It is believed that the reason for this improvement is that swish is able to better alleviate the problem of vanishing gradients during backpropagation.
\section{Results}
In this section, we present a comparative study of different activation functions to solve equation~\eqref{eq:LS} for a two-scatter velocity model, shown in Figure~\ref{fig:model}. We keep all the hyper-parameters fixed to ensure fairness of the comparison. We use a DNN architecture with 8 hidden layers and 20 neurons in each layer, as used by~\cite{alkhalifah2020wavefield} for the considered velocity model. We randomly select 5000 locations from the computational domain and train the PINN model using the Adam optimizer for 15,000 epochs with mini-batch optimization having a batch size of 256. The PINN model is implemented using the SciANN package~\cite[]{haghighat2021sciann} -- a high-level Tensorflow wrapper for scientific computations.
\begin{figure}[ht!]
\begin{center}
\includegraphics[width=0.34\textwidth]{figs/velmodel.pdf}
\end{center}
\caption{
A two-scatter velocity model taken from~\cite{alkhalifah2020wavefield} for the tests. The background has a homogeneous velocity of 2~km/s.}%
\label{fig:model}
\end{figure}
Figure~\ref{fig:loss} shows the loss history for training the PINN model using different activation functions considered. We observe that elu performs the worst with tanh and atan yielding similar performances, while swish has a markedly improved convergence compared to all the considered activation functions. For all activation functions, the convergence plateaus after an initial reduction of the loss value; however, swish recovers faster than the other activation functions. This can be observed by comparing the loss curves for the first 2000 epochs. We also tabulate the $L_2$ and $L_\infty$ errors for both the real and imaginary components of the wavefield solution using these activation functions (see Table~\ref{table1}). We clearly observe improved accuracy for the swish activation function compared to the rest.
\begin{figure}[ht!]
\begin{center}
\includegraphics[width=0.6\textwidth]{figs/loss.pdf}
\end{center}
\caption{
Loss history for training of the PINN model using different activation functions in the hidden layer.}%
\label{fig:loss}
\end{figure}
\begin{figure}
\centering
\subfigure[]{\includegraphics[width=0.24\textwidth]{figs/dur}}
\subfigure[]{\includegraphics[width=0.24\textwidth]{figs/dui}}
\subfigure[]{\includegraphics[width=0.24\textwidth]{figs/durerr.pdf}}
\subfigure[]{\includegraphics[width=0.24\textwidth]{figs/duierr.pdf}}
\caption{The predicted scattered wavefield solution by the PINN model for the real (a) and imaginary (b) parts. Also shown are the wavefield errors for the real (c) and imaginary (d) parts.}
\label{fig:solutions}
\end{figure}
\begin{table}
\centering
\begin{tabular}{|c|c|c|c|c|} \hline
\multirow{2}{*}{\begin{tabular}{c}Activation functions\end{tabular}} & \multicolumn{2}{c|}{$L_2$} & \multicolumn{2}{c|}{$L_\infty$}\\ \cline{2-5}
& Real & Imag. & Real & Imag. \\ \hline
tanh & $4.01\times 10^{-5}$ & $5.51\times 10^{-5}$ & $1.31\times 10^{-2}$ & $1.90\times 10^{-2}$ \\ \hline
atan & $4.15\times 10^{-5}$ & $5.53\times 10^{-5}$ & $1.09\times 10^{-2}$ & $1.81\times 10^{-2}$ \\ \hline
elu & $5.20\times 10^{-5}$ & $6.17\times 10^{-5}$ & $1.26\times 10^{-2}$ & $1.86\times 10^{-2}$ \\ \hline
swish & $3.65\times 10^{-5}$ & $4.7\times 10^{-5}$ & $9.86\times 10^{-3}$ & $1.58\times 10^{-2}$ \\ \hline
\end{tabular}
\caption{$L_2$ and $L_\infty$ errors for the real and imaginary parts of the PINN Helmholtz solution obtained using different activation functions.}
\label{table1}
\end{table}
In Figure~\ref{fig:solutions}, we show the scattered wavefield solution obtained from the PINN model trained using the swish activation function. We also plot the wavefield errors on the same color scale. The overall errors are considerably smaller than for the solutions computed using other activation functions as confirmed by Table~\ref{table1}.
\section{Conclusions}
We studied the routinely used activation functions from the PINN literature in solving the Helmholtz equation. In addition, we also considered the swish activation function, which is a variant of ReLU. Swish had shown improved performance compared to other activation functions in training DNNs for a variety of data science problems, thanks to its robustness against the problem of vanishing gradients. We also observed superior convergence using the swish activation function in solving the Helmholtz equation. Although these results are preliminary, they demonstrate the potential of selecting better activation functions to improve the convergence of PINN solvers. To focus on the activation functions in this study, we only used the Adam optimizer for all cases. However, to converge faster, typically the optimization process begins with a first-order optimizer like Adam and then switches to a second-order optimizer like L-BFGS after a suitably low value of the loss function. Seeing the loss curves, it can be noted that swish allows early switching to a second-order optimizer compared to other studied activation functions, which would compound its efficacy.
\section{Acknowledgments}
\vspace{-0.15cm}
We extend gratitude to Prof. Tariq Alkhalifah and Dr. Ehsan Haghighat for helpful discussions.
\bibliographystyle{unsrt}
|
\section{Introduction}
We study the problem of talking detection in collaborative learning environments.
Here, our ultimate goal is to develop fast
and reliable methods that can assist
educational researchers analyze student participation
in large video datasets.
Learning assessment relies heavily
on the use of audio transcriptions
that describe the interactions
between the students and their
facilitators.
By identifying the video segments
where a student is talking,
educational researchers
can then further analyze
the nature of the interactions.
For example, some students may stay quiet.
Others may express themselves throughout
the lessons.
Ultimately, our computer-based system
aims at aiding this type of analysis
by identifying different talking patterns.
However, for the purposes of this paper,
we will only describe how
to reliably detect students talking
when the camera captures motions over
their mouths.
We present an example of our collaborative learning environment
in Fig. \ref{fig:AOLMESample}.
We are interested in detecting talking activities for the group
that is closest to the camera.
The students that are farther away appear at a smaller scale and
need to be rejected from further consideration.
Students appear at different angles to the camera.
Instead of talking, students can also be eating, laughing, or
yawing, and these activities should not be confused
with talking (e.g., see eating example in
Fig. \ref{fig:AOLMESample}).
In many cases, the mouths may not be visible to the camera.
In such cases, talking detection
is not possible without processing the audio of the video.
\begin{figure}[!t]
\includegraphics[width=\textwidth]{Pictures/AOLMESample_talking.png}
\caption{A sample that contains multiple challenges for talking detection.}
\label{fig:AOLMESample}
\end{figure}
We develop a direct and fast approach to talking detection that
avoids the need for large training datasets.
First, we detect the heads and faces to include the
mouth regions.
Then, over the detected head or face regions,
we compute optical flow vectors and project
the log-magnitudes of the vectors to generate
a single region-proposal image over each candidate speaker.
We then use voting from a list of simple classifiers
to classify each segment as a talking or a non-talking segment.
Our talking detection research extends prior research by our group.
In \cite{shi2016human}, \cite{Shi2018}
we introduced the use of multiscale AM-FM decompositions
to detect student faces and the backs of the heads.
In \cite{tapia2020importance}, the authors demonstrate the importance
of using the instantaneous phase for face detection.
In \cite{shi2018dynamic},
we developed methods to identify possible group interactions
through the use of AM-FM representations.
In \cite{eilar2016distributed},
we developed an open-source system for detecting
writing and typing over cropped video segments.
In \cite{darsey2018hand},
the author developed a hand movement detection system.
In \cite{jacoby2018context}, we used simple color-based object
detection followed by classification of optical flow vectors
to detect writing, talking, and typing over
a very small number of cropped video segments.
There is also significant human activity detection research
within the computer vision community.
In \cite{wang2016temporal},
the authors developed the
Temporal Segment Network (TSN) for video-based activity recognition.
TSN describes a deep learning based approach to detect
a diverse range of activities using ConvNets.
In \cite{tran2015learning},
the authors developed the C3D network that
trains deep 3D convolutional networks
on a large-scale supervised video dataset
to detect a diverse range of different activities.
More recently,
\cite{tran2018closer} generates a new spatiotemporal convolutional block “R(2+1)D” to train CNNs for activity recognition.
Our approach avoids the need to train large, deep learning systems
on human video activity detection.
Our approach is very fast because it reduces talking detection
to the classification of
small proposal regions of the projected motion magnitudes
over the students' faces or heads.
It is ideally suited for our goal to process over 1,000 hours
of videos for talking detection.
We also provide comparisons against TSN and C3D to
demonstrate that our approach is much more accurate.
We organize the rest of the paper into three additional sections.
In section \ref{sec:methodology}, we describe our proposed methodology.
We then provide results in section \ref{sec:results}
and provide concluding remarks in section \ref{sec:conclusion}.
\begin{figure}[!t]
\centering
\includegraphics[width=\textwidth]{Pictures/TalkingDet_whole.png}\\
(a) Talking detection system.\\
\includegraphics[width=\textwidth]{Pictures/HeadDet_whole_v2.png}\\
(b) Head detection system.
\caption{Group talking detection system.}
\label{fig:TalkingDetSystem}
\end{figure}
\section{Methodology}\label{sec:methodology}
We present a system diagram of the entire system in
Fig \ref{fig:TalkingDetSystem}(a).
We also include a block diagram for the head detection system
in Fig \ref{fig:TalkingDetSystem}(b).
In what follows, we summarize the components of each system.
We use multiple methods to locate both faces and the backs of the heads
for the head detector.
In the lower branch of
Fig \ref{fig:TalkingDetSystem}(b),
we show how we extract AM-FM features
using a 54-channel Gabor channel filterbank
as described in \cite{Shi2018} and \cite{shi2016human}.
We use AM-FM components to locate the back of the head region.
For face detection, we use YOLO V3 \cite{redmon2018yolov3}.
We also use FM component classification (LeNet)
to reject background faces that are characterized by higher
frequency components since they are farther away from the camera.
\begin{figure}[!t]
\subfloat{\includegraphics[width=0.155\textwidth]{Pictures/4_1}%
}
~
\subfloat{\includegraphics[width=0.155\textwidth]{Pictures/4_2}%
}
~
\subfloat{\includegraphics[width=0.155\textwidth]{Pictures/2_1}%
}
~
\subfloat{\includegraphics[width=0.155\textwidth]{Pictures/2_2}%
}
~
\subfloat{\includegraphics[width=0.155\textwidth]{Pictures/12_1}%
}
~
\subfloat{\includegraphics[width=0.155\textwidth]{Pictures/12_2}%
}
\\
\subfloat{\includegraphics[width=0.155\textwidth]{Pictures/9_1}%
}
~
\subfloat{\includegraphics[width=0.155\textwidth]{Pictures/9_2}%
}
~
\subfloat{\includegraphics[width=0.155\textwidth]{Pictures/3_1}%
}
~
\subfloat{\includegraphics[width=0.155\textwidth]{Pictures/3_2}%
}
~
\subfloat{\includegraphics[width=0.155\textwidth]{Pictures/8_1}%
}
~
\subfloat{\includegraphics[width=0.155\textwidth]{Pictures/8_2}%
}
\\~\\
\subfloat{\includegraphics[width=0.155\textwidth]{Pictures/no_6_1}%
}
~
\subfloat{\includegraphics[width=0.155\textwidth]{Pictures/no_6_2}%
}
~
\subfloat{\includegraphics[width=0.155\textwidth]{Pictures/no_9_1}%
}
~
\subfloat{\includegraphics[width=0.155\textwidth]{Pictures/no_9_2}%
}
~
\subfloat{\includegraphics[width=0.155\textwidth]{Pictures/no_3_1}%
}
~
\subfloat{\includegraphics[width=0.155\textwidth]{Pictures/no_3_2}%
}
\\
\subfloat{\includegraphics[width=0.155\textwidth]{Pictures/no_1_1}%
}
~
\subfloat{\includegraphics[width=0.155\textwidth]{Pictures/no_1_2}%
}
~
\subfloat{\includegraphics[width=0.155\textwidth]{Pictures/no_8_1}%
}
~
\subfloat{\includegraphics[width=0.155\textwidth]{Pictures/no_8_2}%
}
~
\subfloat{\includegraphics[width=0.155\textwidth]{Pictures/no_2_1}%
}
~
\subfloat{\includegraphics[width=0.155\textwidth]{Pictures/no_2_2}%
}
\caption{Examples of input video frames and the 3-second projection images.
The top two rows show examples of talking video segments.
The bottom two rows show examples of non-talking video segments.}
\label{fig:sum_mag}
\end{figure}
For each head detection, we produce 3-second video clip proposals
for detecting talking activities.
Over these regions, we compute dense optical flow estimates using
Farneback's algorithm \cite{farneback2003two}.
At each pixel, we evaluate
$\log({\tt mag}(i,j)+0.01)$ where
${\tt mag}(i,j)$ represents the magnitude
of each motion vector.
Over each video segment, we then compute
the projection image as given by:
$$ P(i,j) \sum_{\text{all frames} f} \log ({\tt mag}_f (i,j)+0.01).
$$
We then train a variety of proposal
region classifiers to differentiate between
talking and non-talking activities.
We present example projection images in Fig \ref{fig:sum_mag}.
From the examples, compared to projections of talking activities,
it is clear that projections of
non-talking activities are characterized by dark regions
around the mouth regions.
For classifying the projected images, we consider simple classifiers.
We considered a modified LeNet5, XGBoost, AdaBoost,
decision tree, K-NN, quadratic discriminant analysis, and random forest classifier.
Over the training set, we select the best three performing classifiers
based on accuracy, AUC score, and F1 score, and then use
a simple majority vote to combine them into a single system.
\begin{figure}[!b]
\subfloat{\includegraphics[width=0.49\textwidth]{Pictures/HeadResult1.jpg}%
}
~
\subfloat{\includegraphics[width=0.49\textwidth]{Pictures/HeadResult2.jpg}%
}
\caption{Results of head detection.
True positives are bounded by green boxes.
False positives are bounded by red boxes.
False negatives are bounded by yellow boxes.}
\label{fig:HeadResults}
\end{figure}
\section{Results}\label{sec:results}
We summarize our results into three subsections.
First, we present results for our head detector
in section \ref{sec:head}.
Second, we present results for
our head region video detection results in
section \ref{sec:roi}.
Third, we present final results for the
full system in section
\ref{sec:vid}.
\subsection{Head Detection System Results}\label{sec:head}
We summarize head detection results
in Table \ref{HeadResults}.
For training head detection, we selected
1,000 head examples
and 1,200 non-face examples selected from 54 different video sessions.
We then tested our head detector on four
unseen videos as summarized
in Table \ref{HeadResults}.
We can see from the results that our proposed approach
achieved F1 scores that range from 0.81 to 0.87
over 905,550 labeled students.
Example detections are shown in
Fig \ref{fig:HeadResults}.
We can see from the example that our
head classification
system rejected all but one of the
distant detections.
Furthermore, we missed a single face due to occlusion.
\begin{table}[!t]
\caption{\label{HeadResults}
Results for student group detection over four videos.
We present results over 905,550 labeled students.
F1 scores are given for each video.
The videos represent different student groups.
TP refers to true positives.
FP refers to false positives.
FN refers to false negatives.
}
\begin{center}
\begin{tabular}{p{0.15\textwidth}p{0.15\textwidth}p{0.15\textwidth}p{0.12\textwidth}p{0.12\textwidth}p{0.12\textwidth}p{0.12\textwidth}}\toprule
\textbf{Video} & \textbf{Labeled Students} & \textbf{Detected Students} & \textbf{TP} & \textbf{FP} & \textbf{FN} & \textbf{F1} \\ \toprule
\textbf{V1} & 242,700 & 180,640 & 169,550 & 11,090 & 69,190 & 0.81 \\
\textbf{V2} & 131,100 & 122,230 & 107,360 & 14,870 & 17,360 & 0.87 \\
\textbf{V3} & 277,830 & 229,810 & 207,230 & 22,580 & 60,270 & 0.83 \\
\textbf{V4} & 253,920 & 230,600 & 206,860 & 23,740 & 35,750 & 0.87 \\ \bottomrule
\end{tabular}
\end{center}
\end{table}
\begin{table}[!b]
\caption{\label{TrainingDataset}
Training dataset for talking detection.
Video names include the cohort number and the level number (e.g., C3L1).}
\begin{center}
\begin{tabular}{p{0.12\textwidth}p{0.15\textwidth}p{0.12\textwidth}p{0.15\textwidth}p{0.15\textwidth}p{0.15\textwidth}}
\toprule
\textbf{Group ID} & \textbf{Cohort} & \textbf{Group} & \textbf{Date} & \textbf{Urban/ Rural} & \textbf{Frame Rate (fps)} \\ \midrule
1 & C1L1 & D & May-04 & Rural & 60 \\
2 & C1L1 & D & May-11 & Rural & 60 \\
3 & C1L1 & C & May-02 & Urban & 60 \\
4 & C1L1 & C & May-09 & Urban & 60 \\
5 & C1L2 & A & Jun-22 & Rural & 60 \\
6 & C2L1 & A & Mar-22 & Rural & 30 \\
7 & C2L1 & A & Apr-19 & Rural & 30 \\
8 & C2L1 & A & May-05 & Rural & 30 \\
9 & C2L1 & A & May-10 & Rural & 30 \\
10 & C2L1 & B & Mar-22 & Rural & 30 \\
11 & C2L1 & D & Feb-23 & Rural & 30 \\
12 & C2L1 & D & Mar-22 & Rural & 30 \\
13 & C2L1 & A & Feb-20 & Urban & 30 \\ \bottomrule
\end{tabular}
\end{center}
\end{table}
\begin{table}[!t]
\caption{\label{ValDataset}
Validation dataset for talking detection.
Video names include the cohort number and the level number (e.g., C3L1).}
\begin{center}
\begin{tabular}{p{0.12\textwidth}p{0.15\textwidth}p{0.12\textwidth}p{0.15\textwidth}p{0.15\textwidth}p{0.15\textwidth}}
\toprule
\textbf{Group ID} & \textbf{Cohort} & \textbf{Group} & \textbf{Date} & \textbf{Urban/ Rural} & \textbf{Frame Rate (fps)} \\ \midrule
1 & C1L1 & B & Mar-02 & Rural & 30 \\
2 & C1L1 & C & Mar-30 & Rural & 60 \\
3 & C1L1 & C & Apr-06 & Rural & 60 \\
4 & C1L1 & C & Apr-13 & Rural & 60 \\
5 & C1L1 & E & Mar-02 & Rural & 60 \\
6 & C2L1 & B & Feb-23 & Rural & 30 \\
7 & C2L1 & C & Apr-12 & Rural & 30 \\
8 & C2L1 & D & Mar-08 & Rural & 30 \\
9 & C2L1 & E & Apr-12 & Rural & 30 \\
10 & C2L1 & B & Feb-27 & Urban & 30 \\
11 & C3L1 & C & Apr-11 & Rural & 30 \\
12 & C3L1 & D & Feb-21 & Rural & 30 \\
13 & C3L1 & D & Mar-19 & Urban & 30 \\ \bottomrule
\end{tabular}
\end{center}
\end{table}
\begin{table}[!b]
\caption{\label{Results1}
Head-based video region classification results.}
\begin{center}
\begin{tabular}{p{0.18\textwidth}p{0.15\textwidth}p{0.11\textwidth}p{0.15\textwidth}p{0.11\textwidth}p{0.11\textwidth}p{0.15\textwidth}}
\toprule %
\textbf{Methods} & \textbf{Accuracy} & \textbf{\begin{tabular}[c]{@{}l@{}}AUC\\ Score\end{tabular}} & \textbf{Precision} & \textbf{Recall} & \textbf{F1} & \textbf{\begin{tabular}[c]{@{}l@{}}Confusion\\ Matrix\end{tabular}} \\ \midrule
\textbf{LeNet5} & 70\% & 0.76 & 0.69 & 0.76 & 0.72 & $ \begin{bmatrix}1785 & 960\\ 702 & 2177\end{bmatrix} $ \\~\\
\textbf{XGBoost} & 67\% & 0.73 & 0.65 & 0.78 & 0.71 & $ \begin{bmatrix}1549 &1196\\ 647 & 2232\end{bmatrix} $ \\~\\
\textbf{AdaBoost} & 70\% & 0.70 & 0.64 & 0.72 & 0.67 & $ \begin{bmatrix}1557 &1188\\ 810 & 2069\end{bmatrix} $ \\~\\
\textbf{\begin{tabular}[c]{@{}l@{}}Decision \\Tree\end{tabular}} & 59\% & 0.59 & 0.60 & 0.60 & 0.60 & $ \begin{bmatrix}1598 &1147\\ 1138& 1741\end{bmatrix} $ \\~\\
\textbf{KNN} & 68\% & 0.74 & 0.68 & 0.71 & 0.70 & $ \begin{bmatrix}1779 &966\\ 831 & 2048\end{bmatrix} $ \\~\\
\textbf{QDA} & 61\% & 0.71 & 0.82 & 0.30 & 0.44 & $ \begin{bmatrix}2562 &183\\ 2026 &853\end{bmatrix} $ \\~\\
\textbf{\begin{tabular}[c]{@{}l@{}}Random \\Forest\end{tabular}} & 62\% & 0.65 & 0.61 & 0.75 & 0.67 & $ \begin{bmatrix}1354 &1391\\ 728 & 2151\end{bmatrix} $ \\~\\
\textbf{\begin{tabular}[c]{@{}l@{}}XGBoost+\\AdaBoost+\\ KNN\end{tabular}}& 79\% & 0.77 & 0.69 & 0.72 & 0.70 & $ \begin{bmatrix}1810& 935\\ 804 & 2075\end{bmatrix} $ \\
\bottomrule
\end{tabular}
\end{center}
\end{table}
\subsection{Head Video Region Classification Results}\label{sec:roi}
In this section,
we provide comparisons
against single activity classifiers.
For this purpose, we crop
head regions and resize them
to $100 \times 100$ pixels.
For our comparisons, each video
segment is clipped at 3 seconds.
We report results on two
datasets.
The first dataset is used for
selecting the classifiers
that are used in our
majority classification
system.
We use a second dataset
to assess the performance
of the majority classifier
on four videos that
range from 11 to 24 minutes.
\begin{table}[!t]
\caption{\label{Results2}
Talking detection for long videos.}
\begin{center}
\begin{tabular}{p{0.15\textwidth}p{0.15\textwidth}p{0.25\textwidth}p{0.18\textwidth}p{0.15\textwidth}p{0.15\textwidth}}
\toprule
\textbf{Video} & \textbf{Duration} & \textbf{Person Label} & \textbf{Ours} & \textbf{TSN} & \textbf{C3D} \\ \toprule
\multirow{5}{*}{\textbf{V1}} & \multirow{5}{*}{23 min 45 s}&\textit{\textbf{Issac}} & \textbf{66\%} & 28\% & 28\% \\
& & \textit{\textbf{Julia7P}} & \textbf{48\%} & 38\% & 33\% \\
& & \textit{\textbf{Martina64P}} & \textbf{58\%} &11\% & 31\% \\
& & \textit{\textbf{Suzie66P}} & \textbf{44\%} & 11\% & 7\% \\
& & \textit{\textbf{Bernard129P}} & \textbf{51\%} & 18\% & 19\% \\ \cmidrule{3-6}
& & \textbf{Average} & \textbf{53\%} & 21\% & 24\% \\ \midrule
\multirow{6}{*}{\textbf{V2}} & \multirow{6}{*}{11 min 20 s}& \textit{\textbf{Irma}} & 53\% &\textbf{67\%} &64\% \\
&& \textit{\textbf{Emilio25P}} & \textbf{68\%} &21\% &14\% \\
&& \textit{\textbf{Herminio10P}} & 56\% &72\% &\textbf{79\%} \\
&& \textit{\textbf{Jacinto51P}} & \textbf{66\%} &21\% &41\% \\
&& \textit{\textbf{Jorge17P}} & \textbf{60\%} &53\% &43\% \\
&& \textit{\textbf{Juan16P}} & \textbf{62\%} &39\% &35\% \\ \cmidrule{3-6}
&& \textbf{Average} & \textbf{61\%} & 46\% & 46\% \\ \midrule
\multirow{6}{*}{\textbf{V3}} & \multirow{6}{*}{16 min 6 s} & \textit{\textbf{Kelly}} & 70\% &67\% &\textbf{71\%} \\
&& \textit{\textbf{Marta12P}} & \textbf{68\%} & 19\% &34\% \\
&& \textit{\textbf{Cindy14P}} & \textbf{74\%} & 23\% &\textbf{74\%} \\
&& \textit{\textbf{Carmen13P}} & \textbf{51\%} &31\% &50\% \\
&& \textit{\textbf{Marina15P}} & \textbf{64\%} &22\% &26\% \\
&& \textit{\textbf{Scott}} & 87\% &\textbf{95\%} &92\% \\ \cmidrule{3-6}
&& \textbf{Average} & \textbf{69\%} & 43\% & 58\% \\ \midrule
\multirow{6}{*}{\textbf{V4}} & \multirow{6}{*}{23 min 45 s}& \textit{\textbf{Phuong}} & 58\% &\textbf{71\%} &58\% \\
&& \textit{\textbf{Jacob103P}} & \textbf{53\%} &51\% &46\% \\
&& \textit{\textbf{Josephina104P}} & 42\% &\textbf{63\%} &47\% \\
&& \textit{\textbf{Juanita107P}} & 55\% &\textbf{64\%} &60\% \\
&& \textit{\textbf{Tina105P}} & \textbf{55\%} &44\% &47\% \\
&& \textit{\textbf{Vincent106P}} & \textbf{45\%} &43\% &40\% \\\cmidrule{3-6}
&& \textbf{Average} & 51\% & \textbf{56\%} & 50\% \\ \midrule
&&\textbf{Overall Average} & \textbf{59\%} & 42\% &45\% \\
\bottomrule
\end{tabular}
\end{center}
\end{table}
For training the proposed classification
method and all other methods,
we use 11,315 video clips
extracted from 13 different video sessions,
with a total of 27 students (see Table \ref{TrainingDataset}).
For the validation set,
we use 5,624 video clips
extracted from 13 video sessions,
with a total of 37 students (see Table \ref{ValDataset}).
Table \ref{Results1} summarizes
the results from using different classifiers.
We chose XGBoost, AdaBoost, and KNN for the
voting system.
Over our validation set, this combination
gave the highest accuracy at 79\%.
For comparing our system against alternative
approaches, we use four different videos
as summarized in Table \ref{Results2}.
From the results, our system gave an average
accuracy of 59\% compared to
42\% for TSN and 45\% for C3D.
\begin{figure}[!t]
\subfloat{\includegraphics[width=0.49\textwidth]{Pictures/Talking_video_1.jpg}%
}
~
\subfloat{\includegraphics[width=0.49\textwidth]{Pictures/Talking_video_4.jpg}%
}
\\
\subfloat{\includegraphics[width=0.49\textwidth]{Pictures/Talking_video_5.jpg}%
}
~
\subfloat{\includegraphics[width=0.49\textwidth]{Pictures/Talking_video_6.jpg}%
}
\caption{Example of talking detection on the original video.}
\label{fig:TalkingVideoResults}
\end{figure}
\subsection{Talking Activity Detection System}\label{sec:vid}
We present an example of the final
system in Fig. \ref{fig:TalkingVideoResults}.
As shown in Fig. \ref{fig:TalkingVideoResults},
our system detects who is talking and
places a bounding box identifying the person
talking.
Furthermore, unlike single activity
systems like TSN and C3D, we can
detect multiple people talking at the same time.
\section{Conclusion}\label{sec:conclusion}
We presented a new method for detecting
students talking in collaborative
learning environment videos.
Our approach combines head detection
with activity detection using
a projection of motion vectors and
a majority voting classification system.
Our approach significantly outperformed
single activity classification systems.
Yet, our average accuracy at 59\% suggests
that there is still room for significant
improvement.
Our approach will also need to be
further validated before
it can be effectively used by educational
researchers.
\bibliographystyle{splncs04}
|
\section{Introduction}
The communication paradigm called local (quantum) operations and classical communication, usually denoted by its acronym LOCC, is fundamental to quantum information theory, and includes many central topics such as quantum teleportation, data hiding, and many of their derivations \cite{Teleportation, terhal2001hiding,eggeling2002hiding}. The somewhat more restricted version called one-way LOCC, in which communicating parties must perform their measurements in a prescribed order, has received expanded attention as being more tractable mathematically while still capturing many of the more important communication scenarios \cite{Walgate-2000,Nathanson-2005,fan2004distinguishability,N13,cosentino2013small,yu2012four,kribs2017operator, kribsquantum2019,lattice2019,kribs2020vector}. A particularly important subclass of problems in the subject, involves the determination of when sets of known quantum states can be distinguished using only LOCC operations or some subset thereof.
Our work in the theory of LOCC \cite{kribs2017operator,kribsquantum2019,lattice2019,kribs2020vector} has for the first time brought techniques and tools from operator theory, operator algebras, and graph theory to the basic theory of quantum state distinguishability in one-way LOCC. Given the overlapping nature of some of our results and applications, including improvements on some results as our work progressed, we felt a review paper bringing together a selection of main features from our works could be a useful contribution to the literature. In addition to this exposition, we derive a new graph-theoretic description of one-way distinguishability in an important special case, that of a single qubit sender.
This paper is organized as follows. In Section~2 we give necessary preliminaries, including the mathematical description of one-way LOCC in terms of operator relations. Section~3 includes a brief introduction to the relevant operator structures in our analysis, and then we present some of our main results and applications from \cite{kribs2017operator,kribsquantum2019,lattice2019}. We finish in Section~4 by first giving a brief introduction to our necessary notions from graph theory, then we present one of our main results from \cite{kribs2020vector} and some examples, and we derive a new graph-theoretic description for the case of a single qubit sender.
\section{One-Way LOCC and Operator Relations}
We will use the traditional quantum information notation throughout the paper. In particular, we use the Dirac bra-ket notation for vectors, which labels a given fixed orthonormal basis for $\mathbb{C}^d$, with $d \geq 1$ fixed, as $\{ \ket{i} : 0 \leq i \leq d-1 \}$, and the corresponding dual vectors as $\ket{i}^* = \bra{i}$. For $n \geq 1$, $n$-qudit Hilbert space is the tensor product $(\mathbb{C}^d)^{\otimes n}$, which has an orthonormal basis given by $\ket{i_1 i_2 \cdots i_n}:= \ket{i_1} \otimes \ket{i_2} \otimes \ldots \otimes \ket{i_n}$.
We also denote the set of complex $m\times m$ matrices, for a fixed $m\geq 1$, by $M_m (\mathbb{C})$. Given a finite-dimensional Hilbert space $\mathcal H$, we will write $B(\mathcal H)$ for the algebra of bounded (continuous) linear operators on $\mathcal H$, which can be identified with $M_m (\mathbb{C})$ via matrix representations when $\dim \mathcal H = m$. The Pauli operators play an important role in many of our applications, and are given as matrix representations in the single qubit basis $\{ \ket{0}, \ket{1} \}$ for $\mathcal H = \mathbb C^2$ by:
\[
X = \left( \begin{matrix} 0 & 1 \\ 1 & 0 \end{matrix} \right), \quad
Y = \left( \begin{matrix} 0 & -i \\ i & 0 \end{matrix} \right), \quad
Z = \left( \begin{matrix} 1 & 0 \\ 0 & -1 \end{matrix} \right);
\]
where the operators are described by $X\ket{0} = \ket{1}$, $X\ket{1} = \ket{0}$, etc.
The basic set up for the LOCC framework is as follows: multiple parties share a set of quantum states, on which each party can perform local {\it quantum} operations. They can then transmit their results only using {\it classical} information in prescribed directions.
The key problem we have focussed on in our work is to distinguish quantum states amongst a set of known states, where two parties, called Alice ($A$) and Bob ($B$), can perform quantum measurements on their individual subsystems, and then communicate classically. Further, as general LOCC operations are very difficult to characterize mathematically, we have largely restricted ourselves to the case of {\it one-way} LOCC, where the communication is limited to one predetermined direction, generally from $A$ to $B$. This still captures many key examples and settings (though not all).
Hence the bipartite case we consider makes the following assumptions:
\begin{itemize}
\item Two parties $A$, $B$ are separated physically.
\item They control their (finite-dimensional) subsystem Hilbert spaces $\mathcal{H}_{A}$, $\mathcal{H}_{B}$; for simplicity, we often assume $\mathcal{H}_{A}= \mathcal{H}_{B}=\mathbb{C}^d$ for some fixed $d\geq 2$.
\item The state of the composite system $\mathcal{H}_{A} \otimes \mathcal{H}_{B}$ is assumed to be a pure state amongst a known set of states $\mathcal{S} = \lbrace \ket{ \psi_{i}} \rbrace_{i} \subseteq \mathcal{H}_{A} \otimes \mathcal{H}_{B}$.
\item The goal of $A$ and $B$ is then to identify the particular $i$ using only one-way LOCC measurements.
\end{itemize}
The mathematical description of measurements defined by one-way LOCC protocols is given as follows \cite{N13}.
\begin{definition}
A {\bf one-way LOCC measurement}, with $A$ going first, is a set of positive operators $\mathbb{M} = \lbrace A_{k} \otimes B_{k,j} \rbrace_{k,j}$ on $\mathcal{H}_{A} \otimes \mathcal{H}_{B}$ such that
\[
\sum_{k} A_{k} = I_{A} \quad \mathrm{and} \quad \sum_{j} B_{k,j} = I_{B} \quad \forall k.
\]
\end{definition}
Each of the sets $\{A_k\}_k$, $\{B_{k,j}\}_j$ form what is called a {\it positive operator valued measure} (POVM), on $\mathcal H_A$ and $\mathcal H_B$ respectively. If outcome $A_k \otimes B_{k,j}$ is obtained, for any $k$ and a particular $j$, the conclusion is the prepared state was the state identified with the pair $k,j$. Without loss of generality, one can further assume each $A_k$ is a scalar multiple of a (pure) rank one projection.
\begin{example}
As a very simple and illustrative example, consider the following two Bell basis two-qubit states:
\begin{eqnarray*}
\ket{\Phi _{0}} &=& \dfrac{1}{\sqrt{2}} \left( \ket{0}_{A} \ket{0}_{B} + \ket{1}_{A} \ket{1}_{B} \right)
\\ \ket{\Phi _{1}} &=& \dfrac{1}{\sqrt{2}} \left( \ket{0}_{A} \ket{1}_{B} + \ket{1}_{A} \ket{0}_{B} \right)
\end{eqnarray*}
This set is distinguishable, with the following measurement choices:
\begin{itemize}
\item Alice: $A_{1} = \ket{0}\bra{0}$ and $A_{2} = \ket{1} \bra{1}$.
\item Bob: $\{B_{1,1}, B_{1,2}\} = \{ \ket{0} \bra{0}, \ket{1}\bra{1}\} = \{B_{2,1}, B_{2,2}\}$.
\end{itemize}
If Alice gets outcome $0$, then tells Bob, who after measurement gets outcome $0$, then the state is $ \ket{\psi_{0}}$. Similarly it would be $\ket{\psi_{1}}$ if Bob measured a $1$.
\end{example}
Notationally, we shall let $\ket{\Phi}$ be the standard {\it maximally entangled state} on two-qudit space $\mathbb{C}^d \otimes\mathbb{C}^d$;
$
\ket{\Phi} = \frac{1}{\sqrt{d}}\big( \ket{00}+\ldots + \ket{d-1 \, d-1}\big).
$
We recall that every maximally entangled state on two-qudit space is then of the form $(I\otimes V)\ket{\Phi}$ where $V$ is unitary on $\mathbb C^d$.
The following result of Nathanson \cite{N13} frames one-way LOCC distinguishability in terms of operator relations and was the starting point for our collaboration.
\begin{theorem}
Let $\{ U_i \}$ be operators on $\mathbb{C}^d$, and let $\mathcal{S} = \{ \ket{\psi_i}= (I\otimes U_i)\ket{\Phi} \} \subseteq \mathbb{C}^d \otimes\mathbb{C}^d$ be a set of orthogonal states. Then the following conditions are equivalent:
\begin{itemize}
\item[$(i)$] The elements of $\mathcal{S}$ can be distinguished with one-way LOCC.
\item[$(ii)$] There exists a set of states $\{\ket{\phi_k}\}_{k=1}^r\subseteq \mathbb{C}^d$ and positive numbers $\{ m_k \}$ such that $\sum_{k} m_k \kb{\phi_k}{\phi_k} = I$ and for all $k$ and $i \ne j$,
\begin{equation*}
\bra{\phi_k} U^*_j U_i \ket{\phi_k} = 0.
\end{equation*}
\item[$(iii)$] There is a $d \times r$ partial isometry matrix $W$ such that $WW^* = I_d$, and for all $i \ne j$, every diagonal entry of the $r \times r$ matrix $W^*U_j^*U_i W$ is equal to zero.
\end{itemize}
\end{theorem}
Conceptually, the states $\ket{\phi_k}$ are determined by Alice's (rank one) measurement operators, and the orthogonality of the states $\{ U_i \ket{\phi_k} \}_i$ for every $k$, allows Bob to distinguish $i$.
In the example above, note that $\ket{\Phi _{0}} = (I_2 \otimes I_2) \ket{\Phi}$ and $\ket{\Phi_1} = (I_2\otimes X)\ket{\Phi}$,
where $X$ is the single-qubit Pauli bit flip operator. Here we have $M_1 = I_2$, $M_2 = X$, and $d=2=r$. So we can take $W = I_2$ and note that $M_j^* M_i = X$ for $i\neq j$.
\section{Operator Structures and One-Way LOCC}
Our initial work \cite{kribs2017operator} identified the importance of certain operator structures for distinguishing various sets of quantum states using one-way LOCC. The following result encompassed our first observation and readily follows from Nathanson's result. It suggested deeper operator theoretic connections to the mathematics of one-way LOCC lying in the background.
Let $\Delta : M_d (\mathbb{C}) \rightarrow M_d (\mathbb{C})$ be the `map to diagonal' on $d \times d$ matrices; that is, $\Delta$ zeros out all off-diagonal entries of a matrix but leaves its diagonal entries unchanged, and so there is an orthonormal basis $\{ \ket{k} \}$ for $\mathbb{C}^d$ such that $\Delta(\rho) = \sum_{k=1}^d \kb{k}{k} \rho \kb{k}{k}$ is the (von Neumann) measurement map defined by the basis.
\begin{proposition} Let $\{ P_k\}_{k=1}^n$ be a set of $d\times d$ permutation matrices and let $\mathcal{S}= \{(I\otimes P_k)\ket{\Phi}\}$ be the set of corresponding maximally entangled states on ${\mathbb C}^d\otimes{\mathbb C}^d$. Then the following conditions are equivalent:
\begin{enumerate}
\item The states in $\mathcal{S}$ are distinguishable by one-way LOCC.
\item $\Delta(P_j^*P_i)=0$ whenever $i\neq j$.
\end{enumerate}
\end{proposition}
The null space of the map to diagonal operator has a special structure, it is a linear subspace which is closed under taking adjoints. This observation led us to consider the following notions for the first time in the context of LOCC state distinguishability. First we recall the basic structure theory for finite-dimensional $C^*$-algebras, for instance as exhibited in \cite{davidson1996c,paulsen2002completely}. Every such algebra is unitarily equivalent to an orthogonal direct sum of ampliated full matrix algebras of the form $\bigoplus_{k}(M_{m_k}(\mathbb{C})\otimes I_{n_k})$ for some positive integers $m_k, n_k \geq 1$. Further, the algebra is unital if it contains the identity operator.
\begin{definition}
Let $\mathfrak{A}$ be a unital $C^*$-algebra. Any linear subspace $\mathfrak{S}$ contained in $\mathfrak{A}$ which contains the identity and is closed under taking adjoints is called an {\bf operator system}.
\end{definition}
Within the setting of such operator structures, the following notion turns out to be key for us.
\begin{definition}
Let $\mathcal{H}$ be a Hilbert space and let $\mathfrak{S}\subseteq B(\mathcal{H})$ be a set of operators on $\mathcal{H}$ that form an operator system. A vector $\ket{\psi}\in \mathcal{H}$ is said to be a {\bf separating vector} for $\mathfrak{S}$ if $A\ket{\psi}\neq 0$ whenever $A$ is a nonzero element of $\mathfrak{S}$; in other words, $A\ket{\psi} = B\ket{\psi}$ with $A,B\in \mathfrak S$ implies $A=B$.
\end{definition}
If $\mathcal{H}$ is finite-dimensional and $\mathfrak{S}$ is closed under multiplication, and hence a $C^*$-algebra, then we may use the decomposition above for such algebras to determine the existence of a separating vector as follows. This result was proved in \cite{pereira2}.
\begin{theorem}
The $C^*$-algebra $\bigoplus_{k} (M_{m_k}(\mathbb{C})\otimes I_{n_k})$ has a separating vector if and only if $n_k \ge m_k$ for all $k$.
\end{theorem}
In the case of the diagonal algebra $\mathfrak{A}_\Delta$, the set of $d \times d$ diagonal matrices (and so $\mathfrak{A}_\Delta \cong \mathbb{C}^d$), we have $m_k= n_k = 1$ for all $1\leq k \leq d$, and hence $\mathfrak{A}_\Delta$ has a separating vector; an example of which can easily be written down:
$
\ket{\psi} = \frac{1}{\sqrt{d}}( \ket{0} + \ldots + \ket{d-1}).
$
Taken together, these notions and our early results led us to the following general theorem on operator structures and one-way LOCC distinguishability. The first version of the result was proved in \cite{kribs2017operator}, and the refined improvement as stated below was established in \cite{kribsquantum2019}.
\begin{theorem}\label{opthm}
Let $\{U_i\}$ be a set of operators on $\mathbb{C}^d$ and suppose the operator system $\mathfrak{S}_0 = \mathrm{span}\,\{U_i^* U_j, I \}_{i \neq j}$ is closed under multiplication and hence is a C$^*$-algebra. Then $\mathcal S = \{(I\otimes U_i)\ket{\Phi}\}$ is distinguishable by one-way LOCC if and only if $\mathfrak{S}_0$ has a separating vector.
\end{theorem}
The proof of the theorem starts with the observation that if $\mathfrak S_0$ has a separating vector $\ket{\psi}$, then the states $\{ U_i \ket{\psi} \}$ are linearly independent and Bob can use this fact together with Alice's outcome to distinguish the states.
As a straightforward application of the theorem, consider the following class of states. We recall that a set of matrices $\{ U_k\}$ have a
{\it simultaneous Schmidt decomposition} if there are unitary matrices $V$ and $W$ and complex diagonal matrices $\{ D_k\}$ such that for each $k$, $U_k=VD_kW$.
\begin{corollary}
Any set of orthonormal states $\{(I\otimes U_i)\ket{\Phi}\}_{i=1}^n$, for which the matrices $U_i$ have a simultaneous Schmidt decomposition, are distinguishable by one-way LOCC.
\end{corollary}
The basic idea of the proof in this case, is to note the operator system structure satisfies $\mathfrak{S}_0 = \mathrm{span}\,\{U_i^* U_j, I \}_{i \neq j}=W^*\mathfrak{A}_\Delta W$ for some unitary $W$. Since $W^*\mathfrak{A}_\Delta W \cong \mathbb{C}^d$ has a separating vector and hence Theorem \ref{opthm} applies. We note that the operator system $\mathfrak{S}_0$ was studied in a similar context in \cite{duan2013zero}.
\begin{remark}
Towards further applications, including those discussed below, note how the theorem gives a road map to generate sets of indistinguishable states based on these operator structures. Given the decomposition of the algebra generated by an operator system $\mathfrak{A} = \mathrm{Alg}\,(\mathfrak{S}_0) \cong \oplus_k (M_{m_k}(\mathbb{C})\otimes I_{n_k})$, then $\mathfrak{A}$ has a separating vector if and only if $n_k\geq m_k$ for all $k$. Thus, to find instances of sets of {\it indistinguishable states}, we can look for sets $\{U_i\}$ such that $\mathfrak{S}_0 = \mathfrak{A}$ and $m_k > n_k$ for some $k$.
Hence, we are led to consider sets of operators $\{U_i\}$, such that the set is closed under multiplication, taking adjoints, and taking inverses (up to scalar multiples).
\end{remark}
\subsection{Application: States from the Stabilizer Formalism for Quantum Error Correction}
In \cite{kribsquantum2019}, we developed connections between quantum error correction \cite{sho95a,ste96a,bennett1996mixed,gottesmanstab,knilllaflamme,klp05}
and the study of one-way LOCC, including the fact that every one-way LOCC protocol naturally defines a quantum error correcting code defined by the distinguishable states. This also led to new derivations of some known results and new examples of distinguishable states. Here we present one of the applications from that paper.
Sets of unitary operators with the features discussed above are plentiful in one of the foundational areas of quantum error correction, the `stablizer formalism' \cite{gottesmanstab}, which gives a toolbox for generating and identifying codes from the Pauli group.
Let $\mathcal{P}_{n}$ be the $n$-qubit Pauli group; that is, the unitary subgroup on $(\mathbb{C}^2)^{\otimes n}$ with generating set as follows: $$\mathcal{P}_{n} : = \langle \pm iI; X_{j}, Y_j, Z_{j} : 1 \leq j \leq n \rangle\,,$$
where $X_1 = X\otimes I\otimes \cdots \otimes I = X \otimes I^{\otimes (n-1)}$, etc.
The {\it Clifford group} is the normalizer subgroup of $\mathcal P_n$ inside the group of $n$-qubit unitary operators. It is known that if $G$ is a subgroup of $\mathcal{P}_n$, with $S_0 = \{ g_1,\ldots g_m\}$ a minimal generating set for a maximal abelian subgroup $S$ of $G$, then there exists a unitary $U$ in the Clifford group such that $U^* g_j U = Z_j$, for all $1\leq j \leq m$. This allows a focus on the generating Pauli operators for deriving more general results.
In \cite{kribsquantum2019} we proved the following. For succinctness we use terminology from the stabilzer formalism in the theorem hypotheses without giving precise details here, as they are not necessary to appreciate the result.
\begin{theorem}
Let $\{ U_i \} \subseteq \mathcal{P}_{n}$ be a complete set of $4^k$ encoded logical Pauli operators for a stabilizer $k$-qubit code on $n$-qubit Hilbert space. Then the set of states $\mathcal S = \{ (I\otimes U_i)\ket{\Phi} \}$ is distinguishable by one-way LOCC if and only if $k \leq \frac{n}{2}$.
\end{theorem}
The basic idea behind this proof as an application of the result above is as follows: The $4^k$ element set $\mathcal{P}_{n,k} = \langle X_{j}, Z_{j} : 1 \leq j \leq k \rangle / \{ \pm iI\}$, form a set of encoded operations for the code subspace (up to unitary equivalence). But $\mathfrak{S}_0 := \mathrm{span}\,(\mathcal P_{n,k}) = \mathrm{Alg}\,(\mathcal P_{n,k}) = M_{2^k}\otimes I_{2^{n-k}}$. Hence, from the theorem above, the states $\mathcal S$ are distinguishable by one-way LOCC if and only if $\mathfrak{S}_0$ has a separating vector if and only if $2^k \leq 2^{n-k}$, or equivalently $2k \leq n$.
\begin{remark}
The upper bound in this result ($2k=n$) gives sets that saturate a known bound \cite{ghosh2004distinguishability,Nathanson-2005} for the size of one-way distinguishable sets of maximally entangled states on $\mathbb{C}^d\otimes \mathbb{C}^d$ ($d=2^n$). For $2k<n$, this produces (non-trivial) distinguishable sets, which is significant as it is known \cite{cosentino2013small} that many sets defined from $\mathcal P_n$ with less than $2^n$ operators (here $4^k < 2^n$) are not distinguishable even with positive partial transpose operations, and hence not with one-way LOCC.
\end{remark}
\subsection{Application: Sets of Indistinguishable Lattice States}
We have also been able to use the `operator structure road map' outlined above to find sets of lattice states \cite{2013positive,cosentino2013small, yu2012four} that are indistinguishable under one-way LOCC. The following is taken from \cite{lattice2019}.
Recall the two-qubit Bell states $\ket{\Phi_0}, \ket{\Phi_1}$ defined above. The rest of the Bell basis is given by:
\begin{equation*}
\ket{\Phi_2} = \frac{ \ket{01} - \ket{10}}{\sqrt{2}} \qquad \ket{\Phi_3} = \frac{ \ket{00} -\ket{11}}{\sqrt{2}} .
\end{equation*}
These states can be naturally identified with the Pauli matrices by $\ket{\Phi_i} = (I \otimes \sigma_i)\ket{\Phi_0}$ and where we write $I=\sigma_0$, $X=\sigma_1$, $Y=\sigma_2$, $Z=\sigma_3$.
The lattice states are a generalization of the Bell states which are very useful in their own right.
\begin{definition}
For $n \geq 1$, the class of {\bf lattice states} $\mathcal L_n$ are given by $n$-tensors of the Bell states;
\[
{\mathcal L}_n = \{ \ket{\Phi_i} : i \in \{0,1,2,3\} \}^{\otimes n} \subseteq \mathbb{C}^{2^n } \otimes \mathbb{C}^{2^n} .
\]
\end{definition}
States in ${\mathcal L}_n$ can be identified with elements of the Pauli group $\mathcal{P}_{n} = \{ \otimes_{k = 1}^n \sigma_{i_k} \}$, using an extension of the Bell state identification above.
\begin{theorem}\label{LOCCLatticeExample}
For every $n > 1$ and $d = 2^n$, there exist sets of $m$ lattice states in $\mathbb{C}^d\otimes \mathbb{C}^d$ that are not distinguishable with one-way LOCC, where
\begin{eqnarray*}
m = \left\{ \begin{array}{ll} 2\sqrt{2d}-1 & \mbox{ if $n$ is odd} \cr 3\sqrt{d}-1 & \mbox{ if $n$ is even.} \end{array} \right.
\end{eqnarray*}
\end{theorem}
\begin{remark}
As discussed further in \cite{lattice2019}, this result is new and can be extended to so-called `generalized Pauli states', where a new proof is given of an established result, and which leads to an improvement for a studied subclass of states \cite{wangetal2016}. The following example illustrates the approach.
\end{remark}
\begin{example}
For an example in $\mathcal L_n$, with $n$ fixed, we can set
\begin{eqnarray*}
S_1 &=&\{ I^{\otimes i} \otimes Z \otimes I^{\otimes n-i-1}\}_{i = 0}^{k-1} \\ S_2 &=&\{ I^{\otimes i} \otimes Z \otimes I^{\otimes n-i-1}\}_{i = k}^{n-1} \cup \{ X^{\otimes n}\}.
\end{eqnarray*}
It is easy to check that the algebra generated by $S_1$ has dimension $2^k$; the algebra generated by $S_2$ has dimension $2^{n+1-k}$; and the algebra generated by $S_1 \cup S_2$ has dimension $2^{n+1}$. This gives us:
\begin{eqnarray*}
S = \left( \{ I ,Z \}^{\otimes k}\otimes I^{\otimes (n-k)} \right) &\cup & \left( I^{\otimes k} \otimes \{ I ,Z \}^{\otimes n-k} \right) \\ &\cup& \left( { X}^{\otimes k} \otimes \{ X,Y \}^{\otimes n-k} \right)
\end{eqnarray*}
with $|S| = 2^k + 2^{n-k+1} - 1$, which achieves its minimum when $k = \lfloor \frac{n}{2}+1 \rfloor$ and $|S| \in \{2\sqrt{2d}-1, 3\sqrt{d}-1\}$.
\end{example}
\section{Graph Theory and Distinguishing Product States}
The following section contains a brief review of the main results and some applications from \cite{kribs2020vector}, in which we used graph theory to study the problem of distinguishing sets of product states via one-way LOCC. Our graph-theoretic work in LOCC is ongoing, and here we give a new graph-theoretic perspective and proof for the case that Alice only has access to a single qubit Hilbert space.
We shall write $G = (V,E)$ for a {\it simple graph} with vertex set $V$ and edge set $E$. For $v,w\in V$, we write $v \sim w$ if the edge $\{ v,w \}\in E$. The {\it complement} of $G$ is the graph $\overline{G} = (V, \overline{E})$, where the edge set $\overline{E}$ consists of all two-element sets from $V$ that are not in $E$. Another graph $G'$ is a {\it subgraph} of $G$, written $G' \leq G$, if $V' \subseteq V$ and $E' \subseteq E$ with $v,w\in V'$ whenever $\{ v,w\} \in E'$.
Given a graph $G= (V,E)$, a function $\phi: V \rightarrow \mathbb{C}^d\backslash \{0\} $ is an {\it orthogonal representation} of $G$ if for all vertices $v_i \ne v_j\in V$,
\[
v_i \not\sim v_j \iff \langle \phi(v_i), \phi(v_j) \rangle = 0 .
\]
Orthogonal representations have been discussed in graph theory, for instance \cite{fallat2007minimum,lovasz1989orthogonal}. Note the biconditional in the definition, which is stronger than conditions for graph colouring. This allows us to uniquely define the graph associated with a function $\phi$.
We introduced a graph-theoretic perspective to distinguishing product states as follows.
Suppose we are given a set of product states $\{ \ket{\psi^A_k}\otimes \ket{\psi^B_k} \}_{k=1}^r$ on $\mathcal H_A \otimes \mathcal H_B$. The graph of these states from Alice's perspective is the unique graph $G_A$ with vertex set $V = \{ 1,2, \ldots, r\}$ such that the map $k \mapsto \ket{\psi^A_k}$ is an orthogonal representation of $G_A$. Likewise, the graph of the states from Bob's perspective is the graph $G_B$ with vertex set $V$ such that $k \mapsto \ket{\psi^B_k}$ is an orthogonal representation of $G_B$.
Observe that by construction, the set of product states are mutually orthogonal precisely when Alice's graph is a subgraph of the complement of Bob's graph; that is, $G_A \leq \overline{G_B}$.
The following concepts from graph theory are central for us.
Given a graph $G = (V,E)$, a set of graphs $\{ G_i = (V_i, E_i) \}$ {\it covers} $G$ if $V = \cup_i V_i$ and $E = \cup_i E_i$.
A collection of graphs $\{G_i \}$ is a {\it clique cover} for $G$ if $\{G_i \}$ covers $G$ and if each of the $G_i$ is a complete graph (i.e., a clique).
The clique cover number $\mathrm{cc}(G)$ is the smallest possible number of subgraphs contained in a clique cover of $G$.
A clique cover can be thought of as a collection of (not necessarily disjoint) induced subgraphs of $G$, each of which is a complete graph (there is an edge between every pair of vertices) with the condition that every edge is contained in at least one of the cliques.
The following is one of our main results from \cite{kribs2020vector}, and gives a characterization of when product states are one-way LOCC distinguishable in terms of the underlying Alice and Bob graph structures and a related decomposition of Alice's Hilbert space. We note this is a corrected version of the theorem from \cite{kribs2020vector}. The revision was made to condition (3), as the earlier version gave a condition that was sufficient but not necessary.
\begin{theorem} \label{thm: one-way LOCC graphs} Given a set of product states in $\H_A \otimes \H_B$, let $G_A$ and $G_B$ be the graphs of the states from Alice and Bob's perspectives, respectively. Let $\phi: V_A \rightarrow \H_A$ be the association of vertices with Alice's states and assume that the set $\{\phi(v): v\in V\}$ spans $\H_A$.
Then the states are distinguishable with one-way LOCC with Alice measuring first if and only if there exists
\begin{itemize}
\item[(1)] a graph $G$ satisfying $G_A \le G \le \overline{G_B}$,
\item[(2)] a clique cover $\{V_j \}_{j = 1}^k$ of $G$, and,
\item[(3)] a POVM $\{Q_j\}$ on $\mathcal H_A$ such that for all $v\in V_A$, $Q_j\phi(v) \ne 0$ implies that $v \in V_j$.
\end{itemize}
\end{theorem}
The focus on clique decompositions gives tools for building optimal POVMs. We include an example showing a POVM that is not a von Neumann measurement (i.e., the operators are not mutually orthogonal projections).
\begin{example} Let Alice's (unnormalized) states be given in ${\mathbb C}^3$ by
\begin{eqnarray*}
\ket{\phi_{0}} = \ket{0} +\ket{1} &\qquad& \ket{\phi_{1}} = \ket{0} + \ket{2} \\
\ket{\phi_{2}} = \ket{0} -\ket{1} &\qquad& \ket{\phi_{3}} = \ket{0} - \ket{2} ,
\end{eqnarray*}
and Bob's states given so that $G_A = \overline{G_B} = C_4$, the 4-cycle graph. The clique cover number of a 4-cycle is 4, which is bigger than our dimension, but we are still able to distinguish the states using the following POVM. We define
\begin{eqnarray*}
\ket{\psi_{0}} = \ket{0} +\ket{1}+\ket{2} &\qquad& \ket{\psi_{1}} =-\ket{0} +\ket{1}+\ket{2} \\
\ket{\psi_{2}} =\ket{0} -\ket{1}+\ket{2} &\qquad& \ket{\psi_{3}} = \ket{0} +\ket{1}-\ket{2}. \end{eqnarray*}
For each $j$, we can then define $Q_j = \frac{1}{4} \kb{\psi_{j}}{\psi_{j}} $. It is easy to check that $\sum_j Q_j = I$ and that each $Q_j$ picks out an edge of $C_4$ as in the theorem conditions, so we can apply the theorem to show that the states are one-way LOCC distinguishable.
\end{example}
We point the interested reader to \cite{kribs2020vector} for some consequences of this result and related results. For the rest of this section, we will focus on the case of a `low rank' sender, in which this theorem can be entirely stated in graph-theoretic terms.
\subsection{Single Qubit Sender and Graph Theory}
A basic result in LOCC theory \cite{bennett1999unextendible,divincenzo2003unextendible,halder2020distinguishability} shows that any set of orthogonal product states in $\mathbb{C}^{2} \otimes \mathbb{C}^{d}$ for arbitrary $d\geq 2$ can be distinguished via full (two-way) LOCC. This is readily seen to not be the case for one-way LOCC. As a simple example, consider the two-qubit set $\{\ket{00},\ket{10}, \ket{+1}, \ket{-1}\}$ where $\ket{+}=\frac{1}{\sqrt{2}}(\ket{0}+\ket{1})$ and $\ket{-}=\frac{1}{\sqrt{2}}(\ket{0}-\ket{1})$. This set of four two-qubit states cannot be distinguished by one-way LOCC with Alice going first; they can however be distinguished by one-way LOCC with Bob going first and hence also by two-way LOCC.
Nevertheless, one can characterize when one-way distinguishability is possible in the single qubit sender case. As proved in \cite{Walgate-2002} (Theorem~1), the states must take a particularly nice form, in terms of orthogonality on Alice's side versus corresponding orthogonalities on Bob's side. That said, one could argue that the condition from \cite{Walgate-2002} is perhaps not so operationally simple to apply to easily identify when states are distinguishable, if the set is large for example. Here we show how, in the case of product states, Theorem~\ref{thm: one-way LOCC graphs} can be refined for a single qubit sender to yield an entire graph-theoretic set of testable conditions for one-way distinguishability.
\begin{theorem}\label{singlequbit}
A set of orthonormal product states in $\mathbb{C}^{2} \otimes \mathbb{C}^{d}$, for $d \geq 2$, is distinguishable via one-way LOCC with Alice going first if and only if there is some graph between the two graphs $G_A$ and $\overline{G_B}$ with clique cover number at most two; that is, there is a graph $G$ such that
\begin{equation}\label{singlequbitgraphcond}
G_A \leq G \leq \overline{G_B} \quad \mathit{and} \quad \mathrm{cc}\,(G) \leq 2.
\end{equation}
\end{theorem}
\begin{proof}
For the forward direction, our previous theorem gives the existence of $G$ with a clique cover and POVM $\{Q_j\}$ such that $\phi(v_1)^*\phi(v_2) =0$ implies $\phi(v_1)^* Q_j\phi(v_2) = 0$ for all $j$. This implies that either $\mathrm{cc}(G) = 1$ or else each $Q_j$ is diagonal in the $\{ \phi(v_1), \phi(v_2) \}$ basis. It follows that $\mathrm{cc}(G) \le 2$.
For the backward direction, assume a graph $G$ exists that satisfies the conditions of Eq.~(\ref{singlequbitgraphcond}). We shall consider the two cases as determined by the clique cover number of $G$.
Firstly, if $\mathrm{cc}\,(G) =1$, then $G$ is a complete graph. Since $|G_A| = |G_B|$, $G$ must contain all the vertices of $G_B$, hence $\overline{G_B}$ is a complete graph and all the vertices of $G$ are pairwise disconnected in the graph $G_B$. Thus, it follows that all of Bob's states are pairwise orthogonal in $\mathcal H_B = \mathbb{C}^d$ (whether or not Alice's states are orthogonal). Hence the full set of product states on $\mathcal H_A \otimes \mathcal H_B = \mathbb{C}^{2} \otimes \mathbb{C}^{d}$ are one-way distinguishable, simply by a measurement that Bob can perform on his states, independent of what Alice does or communicates to Bob.
Now suppose that $\mathrm{cc}\,(G) =2$. Let $\{ G_1, G_2\}$ be a clique cover of $G$, and let $V_0 = V(G_1) \cap V(G_2)$ be the set of vertices in $G$ that are connected to one (and hence every) vertex in each of $G_1$ and $G_2$. Note that $V_0$ is a proper subset of $V(G_i)$, for $i=1,2$, as otherwise $G_1$ would be a subgraph of $G_2$, or vice-versa, and this would incorrectly imply that $\mathrm{cc}\,(G) =1$.
If $V_0$ is empty, then $G_1$ and $G_2$ are disconnected and each is a subgraph of $\overline{G_B}$. As such, Bob's states corresponding to vertices in $G_1$ are mutually orthogonal, and the same is true of his states corresponding to vertices in $G_2$. Moreover, we can define orthogonal (one-dimensional) subspaces of $\mathcal H_A$ by $\mathcal H_i = \mathrm{span}\, \{ \phi_A(v) : v\in V(G_i)\}$, for $i=1,2$. The states are thus distinguishable via one-way LOCC with Alice first performing a measurement defined by the Hilbert space decomposition $\mathcal H_A = \mathcal H_1 \oplus \mathcal H_2$, then sending the outcome ($j=$ 1 or 2) to Bob, who then performs a measurement defined by the orthogonal states $\{ \phi_B(v) : v\in V(G_j)\}$ to determine the state.
If $V_0$ is non-empty, let $v_1 \in V(G_1)\setminus V_0$, and choose $v_2 \in V(G_2)$ such that $v_1$ and $v_2$ have no edge in $G$ (and hence also in $G_A$) connecting them. Note that such a vertex exists in $V(G_2)$ as otherwise $G=G_1\cup G_2$ would be a single clique and so $\mathrm{cc}\,(G)=1$. Also necessarily $v_2\in V(G_2)\setminus V_0$, as $v_1$ connects with all vertices in $V_0$. It follows that $\ket{\psi_i} = \phi_A(v_i)$, $i=1,2$, are orthogonal and hence form an orthonormal basis for $\mathcal H_A = \mathbb{C}^2$. Furthermore, $\bk{\psi_i}{\phi_A(w_i)}\neq 0$ for all $w_i \in V(G_i)$ and $i=1,2$. This sets up a one-way LOCC protocol as follows: Alice measures in the basis $\{ \ket{\psi_1}, \ket{\psi_2}\}$ for $\mathcal H_A$, and communicates the outcome to Bob. As $G\leq \overline{G_B}$ and $G_j$ is a clique, the states $\{ \phi_B(v) : v\in V(G_j)\}$ are mutually orthogonal in $\mathcal H_B$, and so Bob can determine the state by performing a measurement defined by these states and the projection onto the orthogonal complement of their span. This completes the proof.
\end{proof}
It is fairly straightforward to give examples that satisfy $G_A = \overline{G_B}$ and $\mathrm{cc}(G_A)\leq 2$. For instance, the set $\{ \ket{00}, \ket{01}, \ket{1+}, \ket{1-}\}$, where $\ket{+},\ket{-}$ in this case is any qubit basis different than the standard basis. Here, Alice would measure in the standard basis, and then Bob would measure in the basis suggested by Alice's outcome communicated to him. A simple example of a distinguishable set for which $G_A$ is a proper subgraph of $\overline{G_B}$ is given by the standard two-qubit basis $\{ \ket{00}, \ket{01}, \ket{10}, \ket{11}\}$ (left as an easy exercise: $\mathrm{cc}(G_A)=2<4 = \mathrm{cc}(\overline{G_B})$).
We finish by presenting a `nice' indistinguishable example, in that $G_A = \overline{G_B}$, but nevertheless the states fail to be distinguishable due to the failure of the clique cover condition.
\begin{example}
Consider the (unnormalized) states $\ket{\psi_{i}} \in \mathbb{C}^{2} \otimes \mathbb{C}^{3} $, for $1\leq i \leq 5$, defined as follows:
\begin{align*}
\ket{\psi_{1}}& = \ket{1} \otimes \ket{1}\\
\ket{\psi_{2}} &= \ket{0} \otimes \left( \ket{0} + \ket{1} \right)\\
\ket{\psi_{3}} &= \ket{0} \otimes \left( \ket{0} - \ket{1} \right)\\
\ket{\psi_{4}} &= \left( \ket{0} + \ket{1} \right) \otimes \ket{2} \\
\ket{\psi_{5}} &= \left( \ket{0} - \ket{1} \right) \otimes \ket{2} \\
\end{align*}
These states {\it can} be distinguished with full LOCC operations, with Bob measuring first followed by Alice and then once more by Bob. We show that Bob's initial measurement is necessary and that one-way distinguishability is impossible with Alice going first.
Observe here that $G_A = \overline{G}_{B}$. Moreover, the complement of Bob's graph has clique cover number
$
\mathrm{cc}(\overline{G_{B})} = 4 > 2,
$
with a clique cover of minimal size given by the vertex sets:
\begin{eqnarray*}
\Big\{ \{1,5\} , \,\, \{ 1,4\}, \,\, \{2,3,4\}, \,\,
\{2,3,5\} \Big\}
\end{eqnarray*}
Hence the set of states is not distinguishable via one-way LOCC with Alice going first.
\end{example}
\section{Conclusion}
One of the appeals of quantum information theory is how it builds on expertise in a wide range of areas in physics, mathematics, computer science, and engineering. This review paper highlights several aspects of the fruitful interplay between operator theory and questions of local quantum state distinguishability; and the previous section adds graph theory into the mix. One-way LOCC is a simply-constructed problem with real physical implications, and our work continues to develop effective tools to study it.
\strut
{\noindent}{\it Acknowledgements.} D.W.K. was partly supported by NSERC. C.M. was partly supported by Mitacs and the African Institute for Mathematical Sciences. R.P. was partly supported by NSERC. M.N. acknowledges the support of Saint Mary's College of California, where most of this work was completed.
\bibliographystyle{plain}
|
\section{Introduction}
\label{introduction}
The most natural form of input for an intelligent agent occurs sequentially.
Hence, the ability to continually learn from sequential data has gained much attention in recent machine learning research.
This problem is often coined as \textit{continual learning}, for which three representative approaches have been proposed~\cite{mccloskey89,ratcliff90,french99} including replay~\cite{lopez17,hayes19memory,aljundi19gradient,shin17,rolnick19,lesort19b}, regularization~\cite{kirkpatrick17ewc,zenke17,aljundi19selfless}, and expansion techniques~\cite{rusu16,yoon18}.
At the same time, learning from data riddled with noisy labels is an inevitable scenario that an intelligent agent must overcome. %
There have been multiple lines of work to learn amidst noisy labels such as loss regularization~\cite{wang19sce, zhang18nips,hendrycks18nips},
data re-weighting~\cite{ren18l2r,shen19icml}, label cleaning~\cite{pleiss20aum,lee18cleannet,ostyakov18eccv}, and training procedures~\cite{wei20jocor,jiang17icml}.
In this work, we aim to jointly tackle the problems of \textit{continual learning} and \textit{noisy label classification},
which to the best of our knowledge have not been studied in prior work. %
Noisy labels and continual learning are inevitable for real-world machine learning, as data comes in a stream possibly polluted with label inconsistency.
Hence, the two are bound to intersect; we believe exploring this intersection may glean evidence for promising research directions and hopefully shed light on the development of sustainable real-world machine learning algorithms.
We take on the replay-based approach to tackle continual learning since it has often shown superior results in terms of performance and memory efficiency even with simplicity.
Yet, we discover that replaying a noisy buffer intensifies the forgetting process due to the fallacious mapping of previously attained knowledge. %
Moreover, existing noisy label learning approaches show great limitations when coping within the online task-free setting~\cite{aljundi19mir, prabhu19gdumb, lee20iclr, kim20eccv}.
In their original forms, they assume that the whole dataset is given to purify the noise and thus are hampered by a small amount of data stored only in the replay buffer to either regularize, re-weight, or decide on its validity.
We begin by backtracking the root of the problem; if we naively store a sampled set of the noisy input stream into the replay buffer, it becomes riddled with noise, worsening the amount of forgetting.
Thus, we discover the key to success is maintaining a pure replay buffer, which is the major motive of our novel framework named \textit{Self-Purified Replay} (SPR).
At the heart of our framework is self-supervised learning~\cite{devlin19bert, chen20icml, he20cvpr, grill20nips}, which allows to circumvent the erroneous training signals arising from the incorrect pairs of data and labels.
Within the framework, we present our novel \textit{Self-Replay} and \textit{Self-Centered filter} that collectively cleanse noisy labeled data and continually learn from them.
The Self-Replay mitigates the noise intensified catastrophic forgetting, and the Self-Centered filter achieves a highly clean replay buffer %
even when restricted to a small portion of data at a time.
We outline the contributions of this work as follows.
\begin{enumerate}%
\item To the best of our knowledge, this is the first work to tackle \textit{noisy labeled continual learning}.
We discover noisy labels exacerbate catastrophic forgetting, and it is critical to filter out such noise from the input data stream before storing them in the replay buffer.
\item We introduce a novel replay-based framework named Self-Purified Replay (SPR), for noisy labeled continual learning.
SPR can not only maintain a clean replay buffer but also effectively mitigate catastrophic forgetting with a fixed parameter size.
\item We evaluate our approach on three synthetic noise benchmarks of MNIST~\cite{lecun98}, CIFAR-10~\cite{krizhevsky09}, CIFAR-100~\cite{krizhevsky09} and one real noise dataset of WebVision~\cite{li17arxiv}. Empirical results validate that SPR significantly outperforms many combinations of the state-of-the-art continual learning and noisy label learning methods.
\end{enumerate}
\section{Problem Statement}
\label{sec:motivation}
\subsection{Noisy Labeled Continual Learning}
We consider the problem of online task-free continual learning for classification where a sample $\{x_t, y_t\}$ enters at each time step $t$ in a non i.i.d manner without task labels.
While previous works~\cite{aljundi19gradient, prabhu19gdumb, lee20iclr} assume $\{x_t, y_t\}$ are correct (clean) samples, we allow the chance that a large portion of the data is falsely labeled.
\subsection{Motivation: Noise induced Amnesia}
We discover that if the data stream has noisy labels, it traumatically damages the continual learning model, analogous to \textit{retrograde amnesia}~\cite{squire81}, the inability to recall experience of the past.
We perform some preliminary experiments on a sequential version of symmetric noisy MNIST and CIFAR-10~\cite{lyu20iclr, wang19sce} using experience replay with the conventional reservoir sampling technique~\cite{riemer19iclr,zhang17er}.
The empirical results in Figure~\ref{fig:motivation} show that when trained with noisy labels, the model becomes much more prone to catastrophic forgetting~\cite{french99, mccloskey89, thrun96, ratcliff90}.
As the noise level increases from 0\% to 60\%, sharp decreases in accuracy are seen.
Surprisingly, the dotted red circle in Figure~\ref{fig:motivation}(b) shows that in CIFAR-10 a fatally hastened forgetting occurs no matter the amount of noise.
We speculate that a critical issue that hinders the continual model is the corrupted replay buffer.
An ideal replay buffer should shield the model from noisy labels altogether by being vigilant of all the incoming data for the maintenance of a clean buffer.
\begin{figure}[t]
\begin{center}
\includegraphics[width=\columnwidth]{imgs/motivation.pdf}
\end{center}
\caption{A noisy labeled continual learning on the symmetric noisy in (a) MNIST~\cite{lecun98} and (b) CIFAR-10~\cite{krizhevsky09} when using experience replay with the conventional reservoir sampling~\cite{zhang17er, riemer19iclr}.
At the end of each task, the accuracy of the first task ($T_1$) is plotted. It shows that the noisy labels accelerate catastrophic forgetting. Notably, the dotted red circle in (b) indicates the significantly hastened forgetting process.}
\vspace{-3pt}
\label{fig:motivation}
\end{figure}
\section{Approach to Noisy Labeled Continual Learning}
\label{sec:approach}
We design an approach to continual learning with noisy labels by realizing the two interrelated subgoals as follows.
\begin{enumerate}[start=1,label={\bfseries G\arabic*.}]
\item \textit{Reduce forgetting even with noisy labels}: The approach needs to mitigate catastrophic forgetting amidst learning from noisy labeled data.
\item \textit{Filter clean data}: The method should learn representations such that it identifies the noise as anomalies.
Moreover, it should enable this from a \textit{small amount} of data since we do not have access to the entire dataset in online continual learning.
\end{enumerate}
Figure~\ref{fig:framework} overviews the proposed framework consisting of two buffers and two networks. The \textit{delayed buffer} $\mathcal{D}$ temporarily stocks the incoming data stream, and the \textit{purified buffer} $\mathcal{P}$ maintains the cleansed data.
The \textit{base} network addresses \textbf{G1} via self-supervised replay (Self-Replay) training (Section \ref{sec:self_replay}).
The \textit{expert} network is a key component of Self-Centered filter that tackles \textbf{G2} by obtaining confidently clean samples via centrality (Section \ref{sec:self_filtering}).
Both networks have the same architecture (\textit{e.g}., } \def\Eg{\textit{E.g}., ResNet-18) with separate parameters.
Algorithm~\ref{alg:train_algo} outlines the training and filtering procedure.
Whenever the delayed buffer $\mathcal{D}$ is full, %
The Self-Centered filter powered by the expert network filters the clean samples from $\mathcal{D}$ to the purified buffer $\mathcal{P}$.
Then, the base network is trained via the self-supervision loss with the samples in $\mathcal{D} \cup \mathcal{P}$.
The detail will be discussed in Section \ref{sec:self_replay}--\ref{sec:self_filtering}.
At any stage of learning, we can perform downstream tasks (\textit{i.e}., } \def\Ie{\textit{I.e}., classification) by duplicating the base network into the inference network, adding a final softmax layer, and finetuning it using the samples in $\mathcal{P}$.
Algorithm~\ref{alg:eval_algo} outlines this inference phase.
\begin{figure}[t]
\begin{center}
\includegraphics[width=\columnwidth]{imgs/framework_dark.pdf}
\end{center}
\caption{Illustration of the Self-Purified Replay (SPR) framework. We specify the training and filtering phase (in the yellow shade) in Algorithm~\ref{alg:train_algo}, and the test phase (in the purple shade) in Algortihm~\ref{alg:eval_algo}.}
\vspace{-3pt}
\label{fig:framework}
\end{figure}
\subsection{Self-Replay}
\label{sec:self_replay}
Learning with noisy labeled data~\cite{pleiss20aum, arpit17, ma18, harutyunyan20} results in erroneous backpropagating signals when falsely paired $x$ and $y$ exist in the training set.
Hence, we circumvent this error via learning only from $x$ (without $y$) using contrastive self-supervised learning techniques~\cite{caron20nips, chen20icml, he20cvpr, grill20nips}.
That is, the framework first focuses on learning general representations via self-supervised learning from \textit{all} incoming $x$.
Subsequently, the downstream task (\textit{i.e}., } \def\Ie{\textit{I.e}., supervised classification) finetunes the representation using only the samples in the purified buffer $\mathcal{P}$.
Building on this concept in terms of continual learning leads to Self-Replay, which mitigates forgetting while learning general representations via self-supervised replay of the samples in the delayed and purified buffer ($\mathcal{D} \cup \mathcal{P}$).
Specifically, we add a projection head $g(\cdot)$ (\textit{i.e}., } \def\Ie{\textit{I.e}., a one-layer MLP) on top of the average pooling layer of the base network,
and train it using the normalized temperature-scaled cross-entropy loss~\cite{chen20icml}.
For a minibatch from $\mathcal{D}$ and $\mathcal{P}$ with a batch size of $B_d,B_p \in \mathbb{N}$ respectively,
we apply random image transformations (\textit{e.g}., } \def\Eg{\textit{E.g}., cropping, color jitter, horizontal flip) to create two correlated views of each sample, referred to as positives.
Then, the loss is optimized to attract the features of the positives closer to each other while repelling them from the other samples in the batch, referred to as the negatives.
The updated objective becomes
\begin{align}
\label{eq:ntxent}
L_{self} =
- \hspace{-6pt}\sum_{i=1}^{2(B_d+B_p)} \hspace{-6pt}\text{log} \frac{e^{u_i^T u_j / \tau }}{\sum_{k=1}^{2(B_d+B_p)} \mathds{1}_{k \neq i} e^{u_i^T u_k / \tau}}. %
\end{align}
\noindent
We denote $(x_i, x_j)$ as the positives and $x_k$ as the negatives. $u_i = \frac{g(x_i)}{||g(x_i)||_2}$ is the $\ell_2$ normalized feature, and $\tau > 0$ is the temperature.
Every time when the delayed buffer is full, we train the base network with this loss.
\textbf{Empirical supports}. Figure~\ref{fig:self_replay_charac} shows some empirical results about the validity of Self-Replay for noisy labeled continual learning.
\begin{enumerate}[label=$\bullet$]
\item Figure~\ref{fig:self_replay_charac}(a) shows a quantitative examination on downstream classification tasks. It indicates that self-supervised learning leads to a better representation, and eventually outperforms the supervised one by noticeable margins.
\item Figure~\ref{fig:self_replay_charac}(b) exemplifies the superiority of Self-Replay in continual learning.
We contrast the performances of continually trained Self-Replay (as proposed) against intermittently trained Self-Replay, which trains \textit{offline} with only the samples in the purified buffer at the end of each task.
The colored areas in Figure~\ref{fig:self_replay_charac}(b) indicate how much the continually learned representations alleviate the forgetting and benefit the knowledge transfers among the past and future tasks. %
\end{enumerate}
\begin{algorithm}[tb]
\caption{Training and filtering phase of SPR}
\label{alg:train_algo}
\begin{algorithmic}
\STATE {\bfseries Input:} Training data $(x_t, y_t),...,(x_T, y_T)$ and initial parameters of base network $\theta$. %
\STATE $\mathcal{D} = \mathcal{P} = \{\}$ \text{ // Initialize delayed and purified buffer}
\FOR{$t=1$ {\bfseries to} $T$}
\IF{$\mathcal{D}$ is full}
%
%
\STATE $\mathcal{P} \leftarrow \mathcal{P} \ \cup $ Self-Centered Filter($D$) \text{(section \ref{sec:self_filtering})}
%
%
\STATE $\theta\leftarrow$ Self-Replay using $\mathcal{D} \cup \mathcal{P}$ \text{(section \ref{sec:self_replay})}
%
\STATE reset $\mathcal{D}$
\ELSE
\STATE update $\mathcal{D}$ with $(x_t, y_t)$
\ENDIF
\ENDFOR
\end{algorithmic}
\end{algorithm}
\begin{algorithm}[tb]
\caption{Test phase of SPR}
\label{alg:eval_algo}
\begin{algorithmic}
\STATE {\bfseries Input:} Test data $(x_t, y_t),...,(x_T, y_T)$, parameters of the base network $\theta$, and purified buffer $\mathcal{P}$
\STATE $\psi$ = copy($\theta$) \text{// Duplicate base model to inference model}
\STATE $\psi\leftarrow$ supervised finetune using $\mathcal{P}$
%
\FOR{$t=1$ {\bfseries to} $T$}
\STATE downstream classification for $(x_t, y_t)$ using $\psi$
%
\ENDFOR
\end{algorithmic}
\end{algorithm}
\begin{figure}[ht]
\begin{center}
\includegraphics[width=\columnwidth]{imgs/downstream_continual_intermit.pdf}
\end{center}
\caption{\textbf{Empirical support for Self-Replay} with ResNet18 as the base network on CIFAR-10.
(a) Comparison of overall accuracy of the finetuned downstream classification between self-supervised and supervised representations trained on various noise rates. %
The self-supervised indicates that the base network trained using only $x$ as proposed, while the supervised means training with possibly noisy $(x,y)$ pairs.
(b) The benefits of \textit{continual} Self-Replay over the \textit{intermittent} Self-Replay by comparing the test set accuracy of finetuned models.
The intermittent Self-Replay means training only with contents of the purified buffer up to and including the current task.}
\label{fig:self_replay_charac}
\end{figure}
\subsection{Self-Centered Filter}
\label{sec:self_filtering}
The goal of the Self-Centered filter is to obtain confidently clean samples; specifically, it assigns the probability of being clean to all the samples in the delayed buffer.
\textbf{Expert Network}. The expert network is prepared to \textit{featurize} the samples in the delayed buffer.
These features are used to compute the centrality of the samples, which is the yardstick of selecting clean samples.
Inspired by the success of self-supervised learning good representations in Self-Replay, the expert network is also trained with the self-supervision loss in Eq.~\ref{eq:ntxent}
with only difference that we use the samples in $\mathcal{D}$ only (instead of $\mathcal{D} \cup \mathcal{P}$ for the base network).
\textbf{Centrality}. At the core of the Self-Centered filter lies centrality~\cite{nieminen74}, which is rooted in graph theory to identify the most influential vertices within a graph. %
We use a variant of the \textit{eigenvector centrality}~\cite{bonacich01}, %
which is grounded on the concept that a link to a highly influential vertex contributes to centrality more than a link to a lesser influential vertex.
First, weighted undirected graphs $G:=(V, E)$ are constructed \textit{per unique class label} in the delayed buffer.
We assume that the clean samples form the largest clusters in the graph of each class.
Each vertex $v \in V$ is a sample of the class, and the edge $e \in E$ is weighted by the cosine similarity between the features from the expert network.
For the adjacency matrix $\mathbf{A}= (a_{v,u})_{|V| \times |V|}$. Then the eigenvector centrality is formulated as
\begin{align}
c_v = \frac{1}{\lambda} \sum_{u \in N(v)}c_u = \frac{1}{\lambda} \sum_{u \in V} a_{v,u} c_u, \label{eq:eigen_cent_vec} %
\end{align}
\noindent where $N(v)$ is the neighboring set of $v$, $\lambda$ is a constant and $a_{v,u}$ is the truncated similarity value within $(0, 1]$.
Eq.~\ref{eq:eigen_cent_vec} can be rewritten in vector notation as $\mathbf{A}\mathbf{c} = \lambda \mathbf{c}$, where $\mathbf{c}$ is a vectorized centrality over $V$.
The principal eigenvector $\mathbf{c}$ can be computed by the power method~\cite{vonMises1929}, and it corresponds to the eigenvector centrality for the vertices in $V$. %
\textbf{Beta Mixture Models}. The centrality quantifies which samples are the most influential (or the cleanest) within the data of identical class labels.
However, the identically labeled data contains both \textit{clean} and \textit{noisy} labeled samples, in which the noisy ones may deceptively manipulate the centrality score, leading to an indistinct division of the clean and noisy samples' centrality scores.
Hence, we compute the probability of cleanliness per sample via fitting a Beta mixture model (BMM)~\cite{jacobs1991MoE} to the centrality scores as
\begin{align}
p(c) = \sum_{z=1}^{Z} \pi_z p(c|z), \label{eq:bmm}
\end{align}
where $c > 0$ is the centrality score, $\pi_z$ is the mixing coefficients, and $Z\in\mathbb{N}$ is the number of components.
Beta distribution for $p(c|z)$ is a suitable choice due to the skewed nature of the centrality scores.
We set $Z=2$, indicating the clean and noisy components, and it is empirically the best in terms of accuracy and computation cost.
We use the EM algorithm~\cite{dempster77EM} to fit the BMM through which we obtain the posterior probability%
\begin{align}
p(z|c) = \frac{\pi_z p(c|\alpha_z, \beta_z)}{\sum_{j=1}^Z \pi_j p(c|\alpha_j, \beta_j)}, \label{eq:posterior}
\end{align}
where ${\alpha_z, \beta_z} > 0$ are the latent distribution parameters.
Please refer to the appendix for details of computing $ p(z|c)$.
Among the $Z=2$ components, we can easily identify the \textit{clean} component as the one that has the higher $c$ scores (\textit{i.e}., } \def\Ie{\textit{I.e}., a larger cluster).
Then, the clean posterior $p(z=\mbox{clean}|c)$ defines the probability that centrality $c$ belongs to the clean component, which is used as the probability to enter and exit the purified buffer, $\mathcal{P}$.
After the selected samples enters our full purified buffer, the examples with the lowest $p(z=\mbox{clean}|c)$ are sampled out accordingly.
\begin{figure}[t]
\begin{center}
\includegraphics[width=\columnwidth]{imgs/Illustration.pdf}
\end{center}
\caption{Illustration of graph manipulation via Stochastic Ensemble, which severs weak and uncommon connections and probabilistically focus on confident and clean data within the graph.}
\label{fig:stochastic_ensemble_illust}
\end{figure}
\subsubsection{Stochastic Ensemble}
\label{stochastic_ensemble}
Since our goal is to obtain the \textit{most} clean samples as possible, we want to further sort out the possibly noisy samples.
We achieve this by introducing a \textit{stochastic ensemble} of BMMs, enabling a more noise robust posterior than the non-stochastic posterior $p(z=\text{clean}|c)$ in the previous section.
First, we prepare for stochastic ensembling by sampling multiple \textit{binary} adjacency matrices $\{ \mathbf{A} \}$ from a Bernoulli distribution over $\mathbf{A}$.
For each class $l$, we impose a conditional Bernoulli distribution over $\mathbf{A}$ as
\begin{align}
\hspace{-3pt} p(\mathbf{A}|D_{l}) =\hspace{-4pt} \prod_{d_i, d_j \in D_l} \hspace{-4pt} \text{Bern}\left(\mathbf{A}_{ij} | \text{ReLU}\left(\frac{d_{i} \cdot d_{j}}{||d_{i}|| ||d_{j}||}\right)\right),
\end{align}
where $D_l$ is the set of penultimate feature of class $l$ from the expert network.
We find that it is empirically helpful to truncate the dissimilar values to 0 (ReLU) and use the cosine similarity value as the probability.
We replace the zeros in $\mathbf{A}$ with a small positive value to satisfy the requirement of Perron-Frobenius theorem\footnote{
Perron-Frobenius theorem states when $\mathbf{A}$ has positive entries, it has a unique largest real eigenvalue, whose corresponding eigenvector have strictly positive components.}.
Then, our reformulated robust posterior probability is
\begin{align}
\label{eq:stochastic_posterior}
p(z|D_{l}) \propto \int_\mathbf{A} p(z|\text{cent}(\mathbf{A})) d p(\mathbf{A}|D_{l}),
\end{align}
where $\text{cent}(\cdot)$ is the centrality scores from Eq.~\ref{eq:eigen_cent_vec}, and $p(z|\text{cent}(\mathbf{A}))$ can be obtained in the same manner as the \textit{non-stochastic} posterior in the previous section.
We approximate the integral using Monte Carlo sampling for which we use $E_{max}$ as the sample size.
Essentially, we fit the mixture models on different stochastic graphs to probabilistically carve out more confidently noisy samples by retaining the strong and dense connections while severing weak or uncommon connections.
This is conceptually illustrated in Figure~\ref{fig:stochastic_ensemble_illust}.
\begin{figure}[t!]
\begin{center}
\includegraphics[width=\columnwidth]{imgs/StochasticA.pdf}
\caption{Comparison of non-stochastic and Stochastic Ensemble on CIFAR-10 with 40\% noise.
Stochastic Ensemble produces more confidently clean samples by shifting $p(c|z=\text{noisy}) \cdot p(z=\text{noisy})$ to the left,
and suppressing the cases where $p(c|z=\text{noisy})\cdot p(z=\text{noisy})$ dips below $p(c|z=\text{clean}) \cdot p(z=\text{clean})$.}
\label{fig:stochasticA_vs_fixedA}
\end{center}
\end{figure}
\textbf{Empirical Supports.} Figure~\ref{fig:stochasticA_vs_fixedA} shows some empirical evidence where the stochastic ensemble addresses the two issues to achieve a noise robust posterior $p(z|D_{l})$.
\begin{enumerate}[label=$\bullet$]
\item First, a small portion of noisy samples are falsely confident and are consequently assigned a high centrality score. Stochastic ensembling is able to suppress these noisy samples, as indicated in
Figure~\ref{fig:stochasticA_vs_fixedA}, where the mode of $p(c|z=\text{noisy})\cdot p(z=\text{noisy})$ (red curve) is shifted to the left by a noticeable margin.
\item Second, there are some cases where $p(c|z=\text{noisy})\cdot p(z=\text{noisy})$ drops below the $p(c|z=\text{clean})\cdot p(z=\text{clean})$ leading to a high $p(z=\text{clean}|c)$ for the noisy instances, indicated with red circles in Figure~\ref{fig:stochasticA_vs_fixedA}.
The stochastic ensemble of differing $\mathbf{A}$s can mitigate such problematic cases to drown out the unexpected noise.
\end{enumerate}
\section{Related Works}
\label{related_works}
\subsection{Continual Learning}
\label{continual_related_works}
There have been three main branches to train a model from continual data streams: regularization~\cite{li16lwf, fini20, kirkpatrick17ewc, aljundi19selfless}, expansion~\cite{rusu16, yoon18, lee20iclr}, and replay~\cite{lopez17, chaudhry19iclr, chaudhry19arxiv, riemer19iclr, javed19}.
Replay-based approaches maintain a fixed-sized memory to rehearse back to the model to mitigate forgetting.
Several works~\cite{lopez17, chaudhry19iclr, chaudhry19arxiv} reserve the space for data samples of previous tasks, while others \cite{shin17} uses a generative model.
Some works \cite{riemer19iclr, javed19} combine rehearsal with meta-learning to find the balance between transfer and interference.
We defer more comprehensive survey including all three branches of continual learning to the appendix. %
\textbf{Online Sequential Learning}.
In the online sequential learning scenario, a model can only observe the training samples once.
Hence, many works propose methods for maintaining the buffer~\cite{hayes19memory, aljundi19gradient, kim20eccv} or selecting the samples to be rehearsed~\cite{aljundi19mir}.
Recently, \cite{tang21} adopts graphs to represent relational structures between samples, and \cite{gupta20} employs the meta-loss for learning per-parameter learning rates along with model parameters.
Akin to our work, Gdumb~\cite{prabhu19gdumb} and MBPA++~\cite{autume19neurips} also train the model at inference time.
However, greedily selecting samples to be reserved inevitably leads to degradation from noisy labeled data.
Furthermore, discarding the samples that cannot enter the buffer as done in Gdumb may lead to information loss since it only relies on the buffer as its source of training.
\subsection{Noisy Labels}
\label{noisy_related_works}
Learning with noisy labeled data has long been studied~\cite{zhang17under, arpit17, ma18, jiang20}.
Several works design the noise corrected losses~\cite{xu19, harutyunyan20, li19, arazo19, wang19sce} so that the loss minimization of the whole data becomes similar to that of clean samples.
Other works propose to use a noise transition matrix to correct the loss \cite{patrini17, goldberger17, hendrycks18nips, zhang18nips}.
There have been approaches that aim to suppress the contribution of noisy samples by re-weighting the loss~\cite{wang18, ren18l2r}.
Techniques that repair labels~\cite{kremer18a, veit17, li17iccv, song19b, wang19labelrepair, mandal20} or directly learn them~\cite{tanaka18, yi19pencil} are also viable options for learning from noisy labeled data.
Recently, filtering methods based on training dynamics~\cite{huang19, pleiss20aum, nguyen19self} have gained much popularity, based on the observation that models tend to learn clean data first and memorize the noisy labeled data later.
Small loss sample selection~\cite{jiang17icml, shen19icml, li2020dividemix} techniques by co-teaching \cite{wei20jocor, ge20, han18coteaching, yu19b, malach17, chen19cv} identify noisy samples with multiple models in the same vein.
Some works use graphs for offline learning from a large-scale noisy dataset~\cite{zhang2020global, zhang2021dualgraph}.
On the other hand, we use a small dataset in the delayed buffer from an online data stream without ground-truth labels; instead we adopt self-supervision to obtain features for the Self-Centered filter.
None of the works mentioned above address continual learning from noisy labeled data streams.
Although \cite{mandal20, li21mopro} also use self-supervised learning with noisy labeled data, they focus on the loss or prediction from the model for selecting suspicious samples.
In the experiments on Table~\ref{tab:exp_filtering}, we will show that training dynamics-based filtering techniques are not viable in noisy labeled continual learning. %
On the other hand, we provide the algorithm that identifies the clean samples while learning from a purified buffer in an online manner.
\subsection{Self-supervised learning}
Self-supervised learning is currently receiving an enormous amount of attention in machine learning research.
The pretext task that trains a model by predicting hidden information within the data includes patch orderings~\cite{doersch15, noroozi16}, image impainting~\cite{pathak16}, colorization~\cite{ye19}, and rotations~\cite{gidaris18, chen19rotation}, to name a few.
There also have been works that utilize the contrastive loss~\cite{chen20icml, he20cvpr, li20proto}; especially,
SimCLR \cite{chen20icml} proposes a simplified contrastive learning method, which enables representation learning by pulling the randomly transformed samples from the same image closer while pushing ones apart from other images within the batch.
Recently, this instance-wise contrastive learning is extended to prototypical contrastive learning \cite{li20proto} to encode the semantic structures within the data.
\section{Experiments}
\label{experiments}
In our evaluation, we compare SPR with other state-of-the-art models in the online task-free continual learning scenario with label noise.
We test on three benchmark datasets of MNIST~\cite{lecun98}, CIFAR-10~\cite{krizhevsky09} and CIFAR-100~\cite{krizhevsky09} with symmetric and asymmetric random noise, and one large-scale dataset of WebVision~\cite{li17arxiv} with real-world noise on the Web.
We also empirically analyze Self-Replay and the Self-Centered filter from many aspects.
\begin{table*}[t]
\centering
%
\small
%
\begin{tabular}{lccc|cc|ccc|cc|c}
\toprule
&\multicolumn{5}{c}{MNIST} &\multicolumn{5}{c}{CIFAR-10} &\multicolumn{1}{c}{WebVision}\\
&\multicolumn{3}{c|}{symmetric} &\multicolumn{2}{c|}{asymmetric} &\multicolumn{3}{c|}{symmetric} &\multicolumn{2}{c|}{asymmetric} &\multicolumn{1}{c} {real noise}\\
%
noise rate (\%) & 20 & 40 & 60 & 20 & 40 & 20 & 40 & 60 & 20 & 40 & unknown \\
\midrule
Multitask 0\% noise~\cite{caruaca97} & \multicolumn{5}{c|}{98.6} &\multicolumn{5}{c|}{84.7} &\multicolumn{1}{c}{-} \\
Multitask~\cite{caruaca97} & 94.5 & 90.5 & 79.8 & 93.4 & 81.1 & 65.6 & 46.7 & 30.0 & 77.0 & 68.7 & 55.5 \\
\specialrule{0.1pt}{1pt}{1pt}
Finetune & 19.3 & 19.0 & 18.7 & 21.1 & 21.1 & 18.5 & 18.1 & 17.0 & 15.3 & 12.4 & 11.9 \\
\specialrule{0.1pt}{1pt}{1pt}
EWC~\cite{kirkpatrick17ewc} & 19.2 & 19.2 & 19.0 & 21.6 & 21.1 & 18.4 & 17.9 & 15.7 & 13.9 & 11.0 & 10.0 \\
%
\specialrule{0.1pt}{1pt}{1pt}
CRS~\cite{vitter85} & 58.6 & 41.8 & 27.2 & 72.3 & 64.2 & 19.6 & 18.5 & 16.8 & 28.9 & 25.2 & 19.3 \\
CRS + L2R~\cite{ren18l2r} & 80.6 & 72.9 & 60.3 & 83.8 & 77.5 & 29.3 & 22.7 & 16.5 & 39.2 & 35.2 & - \\
CRS + Pencil~\cite{yi19pencil} & 67.4 & 46.0 & 23.6 & 72.4 & 66.6 & 23.0 & 19.3 & 17.5 & 36.2 & 29.7 & 26.6 \\
CRS + SL~\cite{wang19sce} & 69.0 & 54.0 & 30.9 & 72.4 & 64.7 & 20.0 & 18.8 & 17.5 & 32.4 & 26.4 & 21.5 \\
CRS + JoCoR~\cite{wei20jocor} & 58.9 & 42.1 & 30.2 & 73.0 & 63.2 & 19.4 & 18.6 & 21.1 & 30.2 & 25.1 & 19.5 \\
\specialrule{0.1pt}{1pt}{1pt}
PRS~\cite{kim20eccv} & 55.5 & 40.2 & 28.5 & 71.5 & 65.6 & 19.1 & 18.5 & 16.7 & 25.6 & 21.6 & 19.0 \\
PRS + L2R~\cite{ren18l2r} & 79.4 & 67.2 & 52.8 & 82.0 & 77.8 & 30.1 & 21.9 & 16.2 & 35.9 & 32.6 & - \\
PRS + Pencil~\cite{yi19pencil} & 62.2 & 33.2 & 21.0 & 68.6 & 61.9 & 19.8 & 18.3 & 17.6 & 29.0 & 26.7 & 26.5 \\
PRS + SL~\cite{wang19sce} & 66.7 & 45.9 & 29.8 & 73.4 & 63.3 & 20.1 & 18.8 & 17.0 & 29.6 & 24.0 & 21.7 \\
PRS + JoCoR~\cite{wei20jocor} & 56.0 & 38.5 & 27.2 & 72.7 & 65.5 & 19.9 & 18.6 & 16.9 & 28.4 & 21.9 & 20.2 \\
\specialrule{0.1pt}{1pt}{1pt}
MIR~\cite{aljundi19mir} & 57.9 & 45.6 & 30.9 & 73.1 & 65.7 & 19.6 & 18.6 & 16.4 & 26.4 & 22.1 & 17.2 \\
MIR + L2R~\cite{ren18l2r} & 78.1 & 69.7 & 49.3 & 79.4 & 73.4 & 28.2 & 20.0 & 15.6 & 35.1 & 34.2 & - \\
MIR + Pencil~\cite{yi19pencil} & 70.7 & 34.3 & 19.8 & 79.0 & 58.6 & 22.9 & 20.4 & 17.7 & 35.0 & 30.8 & 22.3 \\
MIR + SL~\cite{wang19sce} & 67.3 & 55.5 & 38.5 & 74.3 & 66.5 & 20.7 & 19.0 & 16.8& 28.1 & 22.9 & 20.6 \\
MIR + JoCoR~\cite{wei20jocor} & 60.5 & 45.0 & 32.8 & 72.6 & 64.2 & 19.6 & 18.4 & 17.0 & 27.6 & 23.5 & 19.0 \\
\specialrule{0.1pt}{1pt}{1pt}
GDumb~\cite{prabhu19gdumb} & 70.0 & 51.5 & 36.0 & 78.3 & 71.7 & 29.2 & 22.0 & 16.2 & 33.0 & 32.5 & 30.4 \\
GDumb + L2R~\cite{ren18l2r} & 65.2 & 57.7 & 42.3 & 67.0 & 62.3 & 28.2 & 25.5 & 18.8 & 30.5 & 30.4 & - \\
GDumb + Pencil~\cite{yi19pencil} & 68.3 & 51.6 & 36.7 & 78.2 & 70.0 & 26.9 & 22.3 & 16.5 & 32.5 & 29.7 & 26.9 \\
GDumb + SL~\cite{wang19sce} & 66.7 & 48.6 & 27.7 & 73.4 & 68.1 & 28.1 & 21.4 & 16.3 & 32.7 & 31.8 & 30.8 \\
GDumb + JoCoR~\cite{wei20jocor} & 70.1 & 56.9 & 37.4 & 77.8 & 70.8 & 26.3 & 20.9 & 15.0 & 33.1 & 32.2 & 24.2 \\
\specialrule{0.7pt}{1pt}{1pt}
%
Self-Centered filter & 80.1 & 79.0 & 77.4 & 80.0 & 79.6 & 36.5 & 35.7 & 32.5 & 37.1 & 36.9 & 33.0 \\
Self-Replay & 81.5 & 69.2 & 43.0 & 86.3 & 78.9 & 40.1 & 31.4 & 22.4 & 44.1 & 43.2 & \textbf{48.0} \\
SPR & \textbf{85.4} & \textbf{86.7} & \textbf{84.8} & \textbf{86.8} & \textbf{86.0} & \textbf{43.9} & \textbf{43.0} & \textbf{40.0} & \textbf{44.5} & \textbf{43.9} & 40.0 \\
%
\bottomrule
\end{tabular}
\vskip 0.1in
\caption{\textbf{Overall accuracy} of noisy labeled continual learning after all sequences of tasks are trained. The buffer size is set to 300, 500, 1000 for MNIST, CIFAR-10 and WebVision, respectively.
Some empty slots on WebVision are due to the unavailability of clean samples required by L2R for training~\cite{ren18l2r}.
The results are the mean of five unique random seed experiments. We report best performing baselines on different episodes with variances in the appendix.}
\label{tab:exp_accuracy}
\end{table*}
\subsection{Experimental Design}
\label{experimental_design}
We explicitly ground our experiment setting based on the recent suggestions for robust evaluation in continual learing \cite{aljundi19thesis,farquhar19,vandeven19} as follows.
(i) \textit{Cross-task resemblance}: Consecutive tasks in MNIST~\cite{lecun98}, CIFAR-10~\cite{krizhevsky09}, CIFAR-100~\cite{krizhevsky09}, WebVision~\cite{li17arxiv} are partly correlated to contain neighboring domain concepts.
(ii) \textit{Shared output heads}: A single output vector is used for all tasks.
(iii) \textit{No test-time task labels}: Our approach does not require explicit task labels during both training and test phase, often coined as \textit{task-free continual learning} in \cite{aljundi19gradient,lee20iclr,kim20eccv}.
(iv) \textit{More than two tasks}: MNIST~\cite{lecun98}, CIFAR-10~\cite{krizhevsky09}, CIFAR-100~\cite{krizhevsky09} and WebVision~\cite{li17arxiv} contain five, five, twenty, and seven tasks, respectively.
We create a synthetic noisy labeled dataset from MNIST and CIFAR-10 using two methods. %
First, the \textit{symmetric} label noise assigns \{20\%, 40\%, 60\%\} samples of the dataset to other labels within the dataset by a uniform probability. We then create five tasks by selecting random class pairs without replacement.
Second, the \textit{asymmetric} label noise attempts to mimic the real-world label noise by assigning other similar class labels (\textit{e.g}., } \def\Eg{\textit{E.g}., 5 $\leftrightarrow$ 6, cat $\leftrightarrow$ dog).
We use the similar classes chosen in \cite{patrini17} to contaminate \{20\%, 40\%\} samples of the dataset with similar class pairs. Each task consists of the samples from each corrupted class pair. %
CIFAR-100 has 20 tasks where the \textit{random symmetric} setting has 5 random classes per task with uniform noise across 100 classes. The \textit{superclass symmetric} setting uses
each superclass~\cite{krizhevsky09, lee20iclr} containing 5 classes as a task where the noise is randomized only within the classes in the superclass.
In WebVision, we use the top 14 largest classes in terms of the data size, resulting in 47,784 images in total. We curate seven tasks with randomly paired classes.
We fix the delayed buffer and the replay (purified) buffer size to 300, 500, 1000, 5000 for MNIST, CIFAR-10, WebVision, and CIFAR-100, respectively.
The purified buffer maintains balanced classes as in ~\cite{kim20eccv, prabhu19gdumb}.
We fix the stochastic ensemble size, $E_{max}=5$ unless stated otherwise.
For the base model, we use an MLP with two hidden layers for all MNIST experiments and ResNet-18 for CIFAR-10, CIFAR-100, and WebVision experiments.
Please refer to the appendix for experiment details.
\subsection{Baselines}
\label{baselines}
Since we opt for continual learning from noisy labeled data streams,
we design the baselines combining existing state-of-the-art methods from the two domains of continual learning and noisy label learning.
We explore the replay-based approaches that can learn in the online task-free setting.
We thus choose (i) Conventional Reservoir Sampling (CRS)~\cite{riemer19iclr}, (ii) Maximally Interfered Retrieval (MIR)~\cite{aljundi19mir}, (iii) Partitioning Reservoir Sampling (PRS)~\cite{kim20eccv} and (iv) GDumb~\cite{prabhu19gdumb}.
For noisy label learning,
we select six models to cover many branches of noisy labeled classification. They include (i) SL loss correction ~\cite{wang19sce}, (ii) semi-supervised JoCoR~\cite{wei20jocor}, (iii) sample reweighting L2R~\cite{ren18l2r}, (iv) label repairing Pencil~\cite{yi19pencil}, (v) training dynamic based detection AUM~\cite{pleiss20aum} and (vi) cross-validation based INCV~\cite{chen19cv}.
\begin{table}[h]
\centering
%
\footnotesize
%
\setlength{\tabcolsep}{5.5pt}
\begin{tabular}{lccc|ccc}
\toprule
%
&\multicolumn{3}{c|}{random symmetric} &\multicolumn{3}{c}{superclass symmetric} \\
%
noise rate (\%) & 20 & 40 & 60 & 20 & 40 & 60 \\
\midrule
GDumb + L2R~\cite{ren18l2r} & 15.7 & 11.3 & 9.1 & 16.3 & 12.1 & 10.9 \\
GDumb + Pencil~\cite{yi19pencil} & 16.7 & 12.5 & 4.1 & 17.5 & 11.6 & 6.8 \\
GDumb + SL~\cite{wang19sce} & 19.3 & 13.8 & 8.8 & 18.6 & 13.9 & 9.4 \\
GDumb + JoCoR~\cite{wei20jocor} & 16.1 & 8.9 & 6.1 & 15.0 & 9.5 & 5.9 \\
\specialrule{0.4pt}{1pt}{1pt}
SPR & \textbf{21.5} & \textbf{21.1} & \textbf{18.1} & \textbf{20.5} & \textbf{19.8} & \textbf{16.5} \\
\bottomrule
\end{tabular}
\vskip 0.1in
\caption{\textbf{CIFAR100 results} of noisy labeled continual learning after all sequences of tasks are trained.
%
The results are the mean of five unique random seed experiments.}
\label{tab:exp_cifar100}
\end{table}
\begin{table}[h]
\centering
%
\footnotesize
%
\setlength{\tabcolsep}{2pt}
\begin{tabular}{lccc|cc|ccc|cc}
\toprule
&\multicolumn{5}{c}{MNIST} &\multicolumn{5}{c}{CIFAR-10} \\
&\multicolumn{3}{c|}{symmetric} &\multicolumn{2}{c|}{asymmetric} &\multicolumn{3}{c|}{symmetric} &\multicolumn{2}{c}{asymmetric} \\
%
noise rate (\%) & 20 & 40 & 60 & 20 & 40 & 20 & 40 & 60 & 20 & 40 \\
\midrule
AUM~\cite{pleiss20aum} & 7.0 & 16.0 & 11.7 & 30.0 & 29.5 & 36.0 & 24.0 & 11.7 & 46.0 & 30.0 \\
%
\specialrule{0.1pt}{1pt}{1pt}
INCV~\cite{chen19cv} & 23.0 & 22.5 & 14.3 & 37.0 & 31.5 & 22.0 & 18.5 & 9.3 & 37.0 & 30.0 \\
%
\specialrule{0.4pt}{1pt}{1pt}
Non-stochastic & 79.5 & 96.3 & 84.5 & 96.0 & 88.5 & 50.5 & 54.5 & 38.0 & 53.0 & 50.5 \\
%
SPR & \textbf{96.0} & \textbf{96.5} & \textbf{93.0} & \textbf{100} & \textbf{96.5} & \textbf{75.5} & \textbf{70.5} & \textbf{54.3} & \textbf{69.0} & \textbf{60.0} \\
%
%
\bottomrule
\end{tabular}
\vskip 0.1in
\caption{\textbf{Filtered noisy label percentage} in the purified buffer (\textit{e.g}., } \def\Eg{\textit{E.g}., out of 20\% symmetric noise, SPR filters 96\% of noise). We compare SPR with $E_{max}=5$ to two other state-of-the-art label filtering methods.}
\label{tab:exp_filtering}
\end{table}
\subsection{Results}
\label{sec:results}
\textbf{Overall performance}. Table~\ref{tab:exp_accuracy} compares the noisy labeled continual learning performance (classification accuracy) between our SPR and baselines on MNIST, CIFAR-10 and WebVision.
Additionally, Table~\ref{tab:exp_cifar100} compares SPR against the best performing baselines on CIFAR-100 with random symmetric noise and superclass symmetric noise.
SPR performs the best in all symmetric and asymmetric noise types with different levels of 20\%, 40\%, and 60\% as well as real noise.
Multitask is an upper-bound trained with an optimal setting with perfectly clean data (\textit{i.e}., } \def\Ie{\textit{I.e}., the 0\% noise rate) and offline training. %
Finetune is reported as a lower-bound performance since it performs online training with no continual or noisy label learning technique.
Notably, SPR works much better than L2R~\cite{ren18l2r}, which additionally uses 1000 clean samples for training, giving it a substantial advantage over all the other baselines.
SPR also proves to be much more effective than GDumb~\cite{prabhu19gdumb}, which is the most related method to ours, even when combined with different noisy label learning techniques.
Moreover, the addition of state-of-the-art noisy label techniques is not always beneficial.
This may be because existing noisy label techniques usually assume a large dataset, which is required to reliably estimate the training dynamics to mitigate the noise by regularizing, repairing, and or filtering.
However, the online learning setting is limited by a much smaller dataset (\textit{i.e}., } \def\Ie{\textit{I.e}., in the purified buffer), leading to a difficult training of the noisy label techniques.
\textbf{Ablation Study}.
To study the effectiveness of each component, two variants of our model that only use Self-Replay or the Self-Centered filter is tested.
That is, the Self-Replay variant does not use any cleaning methods (\textit{i.e}., } \def\Ie{\textit{I.e}., use conventional reservoir sampling to maintain the purified buffer).
The Self-Centered filter variant finetunes a randomly initialized inference network on the purified buffer instead of finetuning it on the duplicate of the base network.
Both variants outperform all the baselines (excluding L2R) in all three datasets, and combining them our model performs the best on MNIST and CIFAR-10 with all noise levels.
However, WebVision is the only dataset where no synergetic effect is shown, leaving Self-Replay alone to perform the best.
This may be because the WebVision contains highly abstract and noisy classes such as ``Spiral" or ``Cinema," making it difficult for Self-Centered filter to sample from correct clusters.
Please refer to the appendix for further detail.
\textbf{Purification Comparison}.
Table~\ref{tab:exp_filtering} compares the purification performance with the state-of-the-art noise detection methods based on the training dynamics, including AUM~\cite{pleiss20aum} and INCV~\cite{chen19cv}.
We notice that the performance of AUM and INCV dreadfully declines when detecting label noise among only a small set of data, which is inevitable in online task-free setting,
whereas SPR can filter superbly even with a small set of data.
Even a non-stochastic version of our Self-Centered filter performs better than the baselines. Encouragingly, our method is further improved by introducing stochastic ensembles.
\textbf{Additional Experiments}.
The appendix reports more experimental results, including SPR's noise-free performance, CIFAR-100 filtering performance, episode robustness, purified \& delayed buffer size analysis, ablation of stochastic ensemble size, variance analysis, and data efficiency of Self-Replay.
\section{Conclusion}
\label{conclusion}
We presented the Self-Purified Replay (SPR) framework for noisy labeled continual learning.
At the heart of our framework is Self-Replay, which leverages self-supervised learning to mitigate forgetting and erroneous noisy label signals.
The Self-Centered filter maintains a purified replay buffer via centrality-based stochastic graph ensembles.
Experiments on synthetic and real-world noise showed that our framework can maintain a very pure replay buffer even with highly noisy data streams while significantly outperforming many combinations of noisy label learning and continual learning baselines.
Our results shed light on using self-supervision to solve the problems of continual learning and noisy labels jointly.
Specifically, it would be promising to extend SPR to maintain a not only pure but also more diversified purified buffer.
\section{Acknowledgement}
\label{acknowledgement}
We express our gratitude for the helpful comments on the manuscript by Junsoo Ha, Soochan Lee and the anonymous reviewers for their thoughtful suggestions.
This research was supported by the international coorperation program by NRF of Korea (NRF-2018K2A9A2A11080927),
Basic Science Research Program through the National Research Foundation of Korea (NRF) (2020R1A2B5B03095585),
and Institue of Information \& communications Technology Planning \& Evaluation (IITP) grant (No.2019-0-01082, SW StarLab).
Gunhee Kim is the corresponding author.
{\small
\bibliographystyle{ieee_fullname}
|
\section{Introduction}
\label{sec:intro}
Scene recognition is a fundamental task for computer vision. Recent progress made in deep convolutional neural networks (CNNs) has greatly boosted the performance of various computer vision tasks, such as image classification \cite{krizhevsky2012imagenet, yuan2018remote}, object detection \cite{ren2016faster,DBLP:journals/tip/YuanXW19}, semantic segmentation \cite{long2015fully, DBLP:journals/tip/WangGL19, yuan2019spatial} and video understanding \cite{DBLP:journals/tip/ZhaoLL19, gao2020feature, yuan2018structured} on large-scale benchmarks. Modern deep CNN architectures such as ResNet \cite{DBLP:conf/cvpr/HeZRS16} and DenseNet \cite{Li2017Classification} are well designed for high semantic-level image representation learning. However, directly applying these deep CNNs for scene recognition still suffers from a limitation: global image features are not flexible enough to represent the indoor scene image with cluttered objects and complex spatial layouts. Considering the difference between image classification and scene recognition, Zhou et al. \cite{DBLP:conf/nips/ZhouLXTO14} released a large scale scene classification dataset named \emph{Places}. They showed the effectiveness of pre-training CNN parameters on Places instead of the object-centric dataset \emph{ImageNet}.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{CV1}
\caption{Motivation of the proposed method. We aim to select local object-level feature vectors from multi-modality deep intermediate feature maps in an unsupervised manner.}
\label{motivation}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=1\textwidth]{CV2}
\caption{Main architecture of the proposed method. 1) \textbf{Global modal-specific feature learning.} The final layer feature maps of two modalities are input to two fully connected (FC) layers to learn global discriminative features with the supervision of auxiliary cross-entropy loss separately. 2) \textbf{Local modal-specific feature learning.} Meanwhile, the final layer feature maps are also used to construct a multi-scale feature pyramid, and then DLFS modules adaptively select multi-scale intermediate CNN features for scene recognition.}
\label{arch}
\end{figure*}
The local object-level intermediate features are complementary to global CNN features \cite{DBLP:conf/eccv/KimF18}. Thus, selecting local features can be effective for taming the great geometric variability of scene image \cite{DBLP:journals/tgrs/YuanFLF19}. The local-feature based methods can be roughly divided into two categories. 1) Local patch-sampling based methods; 2) Object detection based methods. Several works \cite{DBLP:conf/eccv/GongWGL14,DBLP:journals/corr/YooPLK14,DBLP:conf/eccv/ZuoWSZYJ14} opted to extract features from different scales and locations densely and combined them via the fisher vector (FV) \cite{DBLP:journals/ijcv/SanchezPMV13}. However, two disadvantages exist in these methods, which limit the further performance improvement. The first one is that global layout features are neglected. The second one is that densely-sampled local features may induce scene-irrelevant noise, which makes the learned features less discriminative. Object detection can also be used to extract more accurate object-level local features. However, the performance of these methods greatly relies on the accuracy of object detection. Unfortunately, detecting the cluttered objects accurately in complex indoor scenes is also nontrivial.
Moreover, bottom-up local feature learning methods may also suffer from the following problems: Not all local object-level features contribute to the discriminative scene representations. For example, the `chair' features in dining room and the `chair' features in classroom are not discriminative for recognizing these two scenes. Merely using the local features may suffer from the ambiguity for recognizing different scenes. Some theme-level features are also important for scene classification, while object detection based methods may neglect them. For example, `floor' and `curtains' are background theme-level features, but they are also critical for recognizing scenes.
Considering the aforementioned limitations, how to extract scene-related local features for RGB-D indoor image in a weakly supervised manner is still under explored. Since RGB-D image contains spatially-aligned multi-modality information, making use of the multi-modal feature learning process is helpful for local feature extraction. In this work, we find that the correlation between local objects of RGB and depth modalities is stronger than background regions. Therefore, exploiting the correlation distribution of multi-modality feature can help to extract semantic local features, which has not been investigated by previous works. Based on this finding, a differentiable local feature selection (DLFS) module is designed to adaptively extract deep intermediate local features. Additionally, modality correlation loss and mutual information maximization loss are proposed for training the DLFS module discriminately.
In this paper, we design a multi-modal feature learning framework for RGB-D scene recognition, which adaptively selects multi-scale key local features for scene recognition. As shown in Fig. \ref{motivation}, our motivation is to select multiple mid-level CNN feature vectors to represent local patches. To deal with the ambiguity problem caused by local features, the proposed framework also learns to extract global features which are important for describing the scene layout. The main contributions are as follows.
\begin{enumerate}
\item This work is the first to utilize the correlation between RGB and depth modalities to provide more cues for selecting local features. We design an effective differentiable local feature selection module based on the spatial-related correlation of multi-modal features.
\item We introduce a mutual information maximization loss for training the DLFS module, which encourages the discrimination of selected local features.
\item We design a compact global and local multi-modal feature learning framework to learn more discriminative representations for RGB-D scene recognition.
\end{enumerate}
The rest of the paper is organized as follows. Related works about RGB-D scene recognition is shown in Section II. The detail of our method is presented in Section III. The experiments section IV demonstrated the performance of the proposed method. We compared the proposed method with its counterparts in detail in section V. Finally the conclusion of this paper is given in Section VI.
\section{Related Work}
\label{Related Work}
In this section, the related works will be reviewed briefly. For local feature learning methods, patch-based, object detection based and CNN intermediate feature based methods are summarized and reviewed. Moreover, multi-modality feature learning methods for RGB-D data are also surveyed.
Additionally, more detailed comparison between the proposed method and related works are provided in section \ref{cwem}.
\textbf{Patch-sampling based methods.} These approaches extract local features from the patch-based CNN intermediate representations. Gong et al. \cite{DBLP:conf/eccv/GongWGL14} designed a multi-scale CNN framework to sample the local patch features densely, and then encoded them via VLAD \cite{DBLP:conf/cvpr/JegouDSP10}. Some other methods \cite{DBLP:journals/corr/YooPLK14,dixit2015scene} represented the scene image with multi-scale local activations via the FV encoding. Depth image patches were exploited in the work of Song et al. \cite{song2017depth}. They first trained the model with densely sampled depth patches in a weakly-supervised manner, and then fine-tuned the model with the full image. Nevertheless, densely sampled patch features may contain noise, which limits the scene recognition performance.
\textbf{Object-detection based methods.} To discard the irrelevant local features, several methods employed object detection for more accurate object-level features. Wang et al. \cite{wang2016modality} exploited the local region proposals to extract component representations, and encoded the local and global features together via fisher vector. Song et al. \cite{song2017rgb} employed Faster RCNN to detect objects on both RGB and depth images. More accurate object-level local features could be obtained with the object detection sub-module. They further modeled the object-to-object relation and achieved state-of-the-art scene recognition results \cite{DBLP:journals/tip/SongJWCC20}. Although improved performance could be obtained by \cite{wang2016modality,song2017rgb}, the two-stage pipeline methods suffered from the error accumulation problem. The higher computational complexity is also a limitation of these methods. Moreover, detecting small objects in clutter in indoor scenes is nontrivial itself.
Selecting intermediate CNN representations is useful for many applications. KeypointNet \cite{suwajanakorn2018discovery} presented an end-to-end geometric reasoning framework to learn latent category-specific 3D keypoints. KeypointNet can discover geometrically and semantically consistent keypoints adaptively without extra annotations. Zheng et al. \cite{zheng2017learning} proposed a multi-attention convolutional neural network, which consisted of convolution, channel grouping and part classification sub-networks. Wang et al. \cite{wang2018learning} showed that intermediate CNN representations could be enhanced by learning a bank of convolutional filters that captured class-specific discriminative patches without extra bounding box annotations. Xu et al. \cite{xu2018deep} proposed deep regionlets for object detection, which could select non-rectangular regions within the detection framework.
Multi-modality feature learning is critical for RGB-D scene recognition, and various methods have been proposed \cite{wang2018detecting}. Song et al. \cite{song2015sun} fused the two modal features by concatenating them to one fully connected layer. Wang et al. \cite{wang2015mmss} proposed to learn modal-consistent features between RGB and depth images. Li et al. \cite{LiZCHT18} learned the modal-consistent and modal-distinctive embeddings between two modalities simultaneously. Spatial correspondence of local objects in RGB and depth modalities was exploited by \cite{xiong2019rgb} for multi-modal learning. The work in \cite{du2019translate} employed cross-modal translation to explicitly regularize the training of scene recognition, which improved the generalization ability of the model.
\section{Our Method}
The whole network architecture of the proposed method is shown in Fig. \ref{arch}. Depth image is firstly transformed to HHA (Horizontal disparity, Height above ground, Angle of the pixel’s local surface normal with gravity direction) encoding \cite{gupta2014learning}. RGB and HHA images are input to two branches of CNNs for feature extraction. Then, the final layer feature maps of the two modalities are used for global and local multi-modal feature learning. Finally, the global and local features of two modalities are concatenated together to form the final scene representation.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{CV3}
\caption{Illustration of the proposed DLFS module.}
\label{ASKFS}
\end{figure}
\subsection{Differentiable Local Feature Selection}
CNNs can learn to extract high semantic-level features with stacked convolution layers. Layer visualization shows that the intermediate CNN features are with high abstract-level and can represent object parts. Moreover, the work of \cite{DBLP:conf/eccv/KimF18} find that the global scene feature and the object-level local feature are complementary for scene recognition. Based on this insight, our motivation is to select multiple $1\times 1$ mid-level CNN feature vectors to represent the local patches. Specifically, our model predicts $K$ keypoints with the final layer feature maps of RGB and depth modalities, as illustrated in Fig. \ref{ASKFS}.
Suppose that the final layer feature map of RGB and depth modalities are ${F_{rgb}}$ and ${F_{d}}$ respectively. We then concatenate these two feature maps as $F_{rgbd}$ to exploit multi-modal information for feature selection. The DLFS module takes ${F_{rgbd}}$ as input, and uses a $1\times 1$ convolution layer to learn the class-specific feature map for estimating $K$ keypoints. This can be formulated as
\begin{equation}
{V_{rgbd}} = {conv_{1\times 1}(F_{rgbd})},
\end{equation}
where $V_{rgbd} \in \mathbb{R}^{(C,H,W)}$ is the discriminative feature map used for predicting keypoints. $C$ is the number of channels. $H$ and $W$ are the height and width of the feature maps respectively. ${conv_{1\times 1}}$ is a convolution layer with ${1\times 1}$ kernel size and output channel $C$. Although the channels of CNN feature maps are spatially-correlated, one separated channel is not enough to have strong part response \cite{zheng2017learning}. Thus, the channels of $V_{rgbd}$ are further grouped to be more spatially-correlated. Then we reshape feature ${V_{rgbd}}$ to ${V_{rgbd} \in \mathbb{R}^{(K,C/K,H,W)}}$ for the following group channel pooling operation. To make the part response stronger, we sum up the channel groups of ${V_{rgbd}}$ as
\begin{equation}
M_{rgbd}^j = \sum\limits_{i = 1}^{C/K} {{V_{rgb{d_i}}}},
\end{equation}
where ${M_{rgbd}\in \mathbb{R}^{(K,H,W)}}$ is the cross-channel grouped feature map for predicting $K$ selected keypoints. ${j \in \{1,2,...,K\}}$ is the index of the channel groups.
If there are no accurate keypoint annotations, directly training a mapping function from the input feature maps to keypoints is difficult. The reason is that the mapping model is hard to converge without keypoint annotations for supervised training. Inspired by \cite{suwajanakorn2018discovery}, we make our model output an attention map $h_j(u,v)$ to represent the probability of keypoint $j$ occurring at position $(u,v)$. In this work, we use the cross-channel grouped discriminative feature
maps $M_{rgbd}$ to output the attention map. Specifically, a 2D softmax layer is employed to produce the map $h$, which can be represented as
\begin{equation}
\begin{gathered}
h_j = softmax(M_{rgbd}^{j}) \in \mathbb{R}^{(H,W)},
\end{gathered}
\end{equation}
where ${h\in \mathbb{R}^{(K,H,W)}}$ is the probability distribution map, and $j$ is the index of predicted keypoints. Then the coordinates of the keypoints are computed by taking the expected values of the spatial distributions:
\begin{equation}
[{x_j},{y_j}] = \sum\limits_u^H {\sum\limits_v^W {[u \cdot {h_j}(u,v),v \cdot {h_j}(u,v)]} },
\end{equation}
where $[x_j, y_j]$ is the coordinate of the $j^{th}$ predicted keypoint.
To extract the local features in an end-to-end manner, the feature vector-sampling module should be differentiable. Without loss of generality, let ${E_{rgb} \in \mathbb{R}^{(K,C)}}$ be the selected local features for RGB modality. Similarly, we denote $E_{d}$ as the selected local features of the depth modality. Inspired by \cite{DBLPJaderbergSZK15}, the differentiable bilinear feature sampling can be formulated as
\begin{equation}
\begin{split}
E_{rgb_j}^c = \sum\limits_u^H \sum\limits_v^W F_{rgb}^c(u,v) max(0,1 - |{x_j} - v|) \\
\cdot max(0,1 - |{y_j} - u|),
\end{split}
\end{equation}
where ${j \in \{1,2,...,K\}}$ is the index of $K$ sampled local feature vectors, and ${c \in [1...C]}$ is the channel index. The coordinates ${(x_j,y_j)}$ and ${(u,v)}$ are normalized in the range of [-1,1].
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{CV8-2}
\caption{Illustration of the proposed variational information maximization loss.}
\label{LOSS1}
\end{figure}
Since the DLFS module takes the keypoint index and the feature maps as input, we need to compute the partial derivatives of $F_{rgbd}$ as well as the predicted coordinates to allow the backpropagation through this module. The partial derivative w.r.t feature maps ${F_{rgb}}$ and $(x_j,y_j)$ are presented as follows.
The partial derivative of feature maps ${F_{rgb}}$ can be formulated as:
\begin{equation}
\begin{split}
\frac{{\partial E_{rgb}^c}}{{\partial F_{rgb}^c}} = \sum\limits_u^H \sum\limits_v^W max(0,1 - |{x_j} - v|) \\
\cdot max(0,1 - |{y_j} - u|).
\end{split}
\end{equation}
The partial derivatives of $y_j$ is
\begin{equation}
\label{e7}
\frac{{\partial E_{rgb}^c}}{{\partial {y_j}}} = \sum\limits_{u,v}^{H,W} F_{rgb}^c(u,v) {max(0,1 - |{x_j} - v|)g(v,{y_j})} ,
\end{equation}
where $g(v,y_j)$ is a piecewise function, and it can be formulated as
\begin{equation}
\label{e8}
g(v,{y_i}) = \left\{ {\begin{array}{*{20}{c}}
{\begin{array}{*{20}{c}}
0,\\
{\begin{array}{*{20}{c}}
1,\\
{ - 1,}
\end{array}}
\end{array}}&{\begin{array}{*{20}{c}}
{{\rm{if}}|v - {y_j}| \ge 1}\\
{\begin{array}{*{20}{c}}
{{\rm{if }}v \ge {y_j}}\\
{{\rm{if }}v{\rm{ < }}{{\rm{y}}_j}}
\end{array}}
\end{array}}
\end{array}} \right..
\end{equation}
As for the depth modality ${F_{d}}$, the differentiable feature selection procedure is similar to the RGB modality.
\begin{figure}
\centering
\includegraphics[width=0.46\textwidth]{CV9}
\caption{Pixel-wise correlation map between deep features of RGB and depth modalities. RGB images are on the left, and the corresponding correlation-maps are displayed on the right. More visualization examples are displayed in Fig. \ref{cosinesimilarity}.}
\label{visfk}
\end{figure}
For multi-scale local feature selection, we can construct feature pyramid with stride 2 convolutional layers as shown in Fig. \ref{arch}. For each scale CNN feature maps, the local feature selection process is similar.
\subsection{Discriminative Training for DLFS}
Although the DLFS module can be trained in an end-to-end manner, the feature map ${M_{rgbd}}$ is not guaranteed to be discriminative enough for selecting different local patch features. To encourage the channels of $M_{rgbd}$ to be sensitive to different semantic parts, we propose a novel loss function to train the DLFS module.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{CV8-1}
\caption{Illustration of the proposed loss function for training DLFS module. The first term is the triplet-correlation loss, and the second one is the multi-modal local feature correlation loss. }
\label{LOSS}
\end{figure}
In order to make ${M_{rgbd}}$ more discriminative, we choose to retain high mutual information between ${M_{rgbd}}$ and the scene class label. Since we aim to select feature vectors which are important for classifying different scenes, the feature map ${M_{rgbd}}$ should be highly correlated to different scene classes. By maximizing the mutual information between ${M_{rgbd}}$ and the class label $L$, the channels of ${M_{rgbd}}$ is enforced to be class-correlated. As the exact computation of mutual information is intractable, variational lower bound is used for the approximation in this work.
Suppose that there are $N$ classes, ${M_{rgbd}}$ is firstly pooled by a max-pooling layer to get $T_{rgbd} \in \mathbb{R}^{S\times S\times K}$. Then a $S\times S$ convolution layer is used to transform $T_{rgbd}$ into $N$ scalar features. Finally, we get $N$ features ${U \in \mathbb{R}^{N}}$ corresponding to $N$ scene categories. The mutual information ${I(L;U)}$ can be defined as
\begin{equation}
\begin{aligned}
&I(L;U) = H(L) - H(L|U)\\
&= H(L) + \mathbb{E}_{L,U}[\log q(L|U)]+KL\\
&\ge H(L) + \mathbb{E}_{L,U}[\log q(L|U)],
\end{aligned}
\end{equation}
where ${H(\cdot)}$ is the entropy, and $\text{KL}$ denotes the Kullback-Leiber divergence. Maximizing the mutual information is equivalent to minimizing the following loss function ${\cal L}_{\text{VI}} = -\mathbb{E}_{L,U}[\log q(L|U)]$. ${\cal L}_{\text{VI}}$ can also be interpreted as the reconstruction error. In this work, we choose the variational distribution as a Gaussian distribution with heteroscedastic mean:
\begin{equation}
\begin{aligned}
&{\cal L}_{\text{VI}} =\sum_{n=1}^{N}\log\sigma_n + \frac{(L_n-{U_{\mu}}_n)^2}{2\sigma_n^2} + c,
\end{aligned}
\end{equation}
where $c$ is a constant, and $U_{\mu}$ is the transformed features by FC layer as shown in Fig. \ref{LOSS1}. $\sigma$ is a learnable parameter to be optimized.
By minimizing ${\cal L}_{\text{VI}}$, the feature ${M_{rgbd}}$ is encouraged to be discriminative for selecting key local features.
\subsection{Unevenly distributed Multi-modal Correlation}
We find that local objects contribute highly to the multi-modal correlation. To verify this, we compute the pixel-wise correlation map $P\in \mathbb{R}^{H,W}$ between the deep features $F_{rgb}\in \mathbb{R}^{C,H,W}$ of RGB modality and the depth modal deep features $F_{d}\in \mathbb{R}^{H,W}$. Then the pixel-wise correlation map can be defined as:
\begin{equation}
\begin{gathered}
P^{ij} = \rho({F_{rgb}}^{ij},{F_{d}}^{ij}), \\
i=1,...,H;j=1,...,W,
\end{gathered}
\end{equation}
where $\rho$ is the cosine similarity function. ${F_{rgb}}^{ij} \in \mathbb{R}^{C}$ is a feature vector at position $(i,j)$ of feature $F_{rgb}$.
As shown in Fig. \ref{visfk}, we have visualized the pixel-wise correlation map $P$. It can be clearly seen that the correlation between two modalities is unevenly distributed. More specifically, large correlation values cluster on local objects. This indicates that multi-modal features of local objects have higher correlation than other spatial positions.
Based on this idea, local objects can be located by finding the largest correlation regions. Thus, we propose to maximize the correlation between RGB and depth local features to supervise the local feature selection module. The multi-modal correlation loss ${\cal L}_{C{m}}$ can be computed as:
\begin{equation}
{\cal L}_{C{m}} = \rho({E_{rgb}},{E_{d}}),
\end{equation}
where $\rho$ is the cosine similarity function. To further enhance the DLFS module with class-specific local features, triplet correlation loss is employed. As illustrated in Fig. \ref{LOSS}, we aim to select local features that have larger correlation between the same class and smaller correlation between different classes.
For each sample in the triplet input ${\{\textbf{a},\textbf{p},\textbf{n}\}}$, the corresponding selected local features are ${\{E_{rgb}a,E_{rgb}p,E_{rgb}n\}}$. $E_{rgb}p$ and $E_{rgb}a$ are positive and anchor features with the same class label, and $E_{rgb}n$ is the negative one with different scene class label. The triplet correlation loss is formulated as follows.
\begin{equation}
{\cal L}_{C{rgb}} = max\{ \rho({E_a},{E_p}) - \rho({E_a},{E_n}) + \alpha ,0\},
\end{equation}
where ${{\cal L}_{Crgb}}$ is the triplet correlation loss for the RGB modality, and the loss computation for depth modality ${{\cal L}_{Cd}}$ is similar. $\alpha$ is the margin value and it is set to 1.0 in this work. The whole correlation loss for two modalities can be formulated as ${{\cal L_C} = {\cal L}_{Crgb} + {\cal L}_{Cd}} + {\cal L}_{C{m}}$.
\subsection{Joint Global and Local Feature Representation}
Though local features are effective for representing the scene images, merely using local features may suffer from the ambiguity problem. Thus in this work, we choose to learn the global and local features simultaneously to obtain more robust representations for scene recognition.
The global features are learned by the FC layers connected to the final feature maps ${F_{rgb}}$ and ${F_{d}}$. As shown in Fig. \ref{arch}, two auxiliary cross-entropy loss functions are employed for learning global modal-specific feature separately. This loss function can be formulated as
\begin{equation}
{{\cal L}_{aux}} = {{{\cal L}_{CE}}({G_{rgb}},{y})} + {{{\cal L}_{CE}}({G_{d}},{y})},
\end{equation}
where $L_{CE}$ represents the cross-entropy loss.
Finally, we concatenate the multi-modal global and local features together for the final scene classification. This can be denoted as
\begin{equation}
\begin{gathered}
H_{mmgl}=concat(G_{rgb},G_{d},E_{rgb},E_d),\\
\end{gathered}
\end{equation}
where ${H_{mmgl}}$ is the multi-modal global and local feature vector. Then ${H_{mmgl}}$ is input to a fully connected layer, and the final classification result ${\hat y}$ is output through a softmax layer. We denote the final cross-entropy classification loss as ${{\cal L}_{cls}}$.
The overall loss function for training the proposed framework is a multi-task loss, which consists of four terms: 1) the global modal-specific auxiliary loss ${\cal L}_{aux}$; 2) the mutual-information maximization loss ${\cal L}_{\text{VI}}$; 3) the triplet correlation and multi-modal correlation loss ${\cal L}_C$; 4) the final classification loss ${\cal L}_{cls}$. This can be represented as
\begin{equation}
\begin{aligned}
{\cal L} = {\cal L}_{cls} + \lambda_1 {\cal L}_{aux} + \lambda_2 {\cal L}_{\text{VI}} + \lambda_3 {\cal L}_C.
\end{aligned}
\end{equation}
where $\lambda_1$, $\lambda_2$, and $\lambda_3$ are balancing weights of loss components for the multi-task loss function.
\section{Experiments}
Two popular RGB-D scene recognition datasets are used to evaluate the proposed framework. One is the \textbf{SUN RGB-D} \cite{song2015sun} dataset, which contains 10,355 RGB and depth image pairs. They are captured from different depth sensors including Kinect v1, Kinect v2, Asus Xtion and RealSense. These images are divided into 19 scene categories. To compare with existing methods, we follow the same experimental settings with \cite{song2015sun}. For this dataset, 4,848 image pairs are used for training and 4,659 pairs for testing. Another dataset is the \textbf{NYUD v2} \cite{Silberman:ECCV12} dataset, which includes 1,449 RGB and depth image pairs divided into 10 categories. Following the experimental setting in \cite{gupta2013perceptual}, 795 image pairs are used for training and 654 pairs for testing.
\begin{table*}[h!]
\centering
\caption{Ablation Study on NYUD v2 Dataset}
\label{tabel:nyu21}
\scalebox{1.3}{
\begin{tabular}{c|c|c}
\hline \hline
Feature Types &Methods & Mean-class Accuracy(\%) \\ \hline
Single Modality & RGB / Depth / HHA & 61.2 / 54.1 / 58.2 \\ \hline
\multirow{3}{*}{Multi-modality \& Global}
& RGB-D (HHA Encoding) & 64.2 \\ \cline{2-3}
& RGB-D Global (${{\cal L}_{aux}}$) & 65.5 \\ \cline{2-3}
& RGB-D Global (Spatial Attention \cite{Wang2018NonlocalNN}) & 66.1 \\ \hline
\multirow{4}{*}{Multi-modality \& Local}
& RGB-D Local & {64.1} \\ \cline{2-3}
& RGB-D Local (${\cal L_C}$) & {66.2} \\ \cline{2-3}
& RGB-D Local (${\cal L_{\text{VI}}}$) & {66.7} \\ \cline{2-3}
& RGB-D Local (Full DLFS) & {67.3} \\ \hline
\multirow{4}{*}{Multi-modality \& Global \& Local}
& RGB-D Global \& Local (${\cal L_C}$) & 67.1 \\ \cline{2-3}
& RGB-D Global \& Local (${\cal L_{\text{VI}}}$) & 68.2 \\ \cline{2-3}
& RGB-D Global \& Local (Full DLFS) & \textbf{68.9} \\ \cline{2-3}
& RGB-D Global \& Local (Two-Scale Full DLFS) & \textbf{69.3} \\ \cline{2-3} \hline
\end{tabular}
}
\end{table*}
\subsection{Implementation Details}
HHA encodings is computed with the code released by \cite{gupta2014learning}. We use ResNet18 as the network backbone. Pre-trained parameters on \textbf{Places} are used for fine-tuning. Data augmentation is used in our work including random flip, cutout and random erasing \cite{zhong2017random}.
For optimization, Adam \cite{kingma2014adam} is employed with an initial learning rate of 1e-4, and the learning rate is reduced by a fraction of 0.9 every 80 epochs. The batch size is set to 64 with shuffle. For all the experiments, 300 epochs are used to train the model. As for the multi-task training, we set the parameters ${\lambda_1}$ to 1, and ${\lambda_2}$, ${\lambda_3}$ are set to 0.1 for all experiments. Multi-scale DLFS is evaluated with the ResNet18 backbone. The first scale feature maps are with the size of ${7\times7}$. The second scale is obtained with a $1\times1$ kernel convolution layer. By setting the stride to 2, the second scale feature maps are with the size of ${3\times3}$. For the first scale, $K$ is set to 16 and it is set to 4 for the second scale.
We randomly select 20\% training samples for each scene category to form a validation set for both datasets. With this setting, the dataset is split into training/validation/test sets. Training set is used for model training, and validation set for model selection. Then the test set is used for model evaluation and performance comparison.
\subsection{Ablation Study and Discussions}
To evaluate the proposed framework comprehensively, we do the following ablation studies with ResNet18 backbone to explore the effect of different sub-modules. Additionally, to show the superiority of multi-modality RGB-D data and local feature, we also conduct experiments to compare the performance of different feature types.
\subsubsection{\textbf{Single Modality}}
The results are shown in Table \ref{tabel:nyu21}. The first row displays the accuracy of using single modality: including RGB image, depth image and HHA image. ``RGB'' denotes RGB image. ``Depth'' denotes the original depth image, which is not transformed to HHA encoding. While ``HHA'' denotes that HHA image is used as depth modality input instead of the original depth image. Since RGB image contains richer texture and appearance information than depth modality, using single RGB modality data can achieve better accuracy. As for the depth modality, ``HHA'' can obtain better performance than ``Depth''. This indicates the effectiveness of HHA encoding.
\subsubsection{\textbf{Multi-modality \& Global}}
As shown in the table, ``RGB-D (HHA Encoding)'' is the method using global RGB and HHA features for scene classification,
which can achieve obvious performance improvement compared with methods using single modality. This demonstrates that both RGB and depth modalities are useful for recognizing scenes. Additionally, we also study the effect of using auxiliary classification loss for multi-modality global features. ``RGB-D Global (${{\cal L}_{aux}}$)'' improves the baseline method by 1.3\%, which shows the effectiveness of the auxiliary loss for learning global modal-specific features.
Spatial attention mechanism can be used to focus on important local parts and extract more representative deep features.
In this work, we use the non-local neural networks \cite{Wang2018NonlocalNN} to focus on different spatial regions with assigned weights. Although spatial attention can be used to learn local-sensitive deep features, the final features of method ``RGB-D Global(Spatial Attention)'' are still global features processed by fully connected layer.
However, the proposed method selects key local features and abandon other features, which can be viewed as a kind of ``hard-attention". Using spatial-attention to focus on local features softly obtains a performance of 66.1\%, which is lower than the proposed local feature selection method.
\subsubsection{\textbf{Multi-modality \& Local}}
We have also evaluated the performance of the proposed method when only using local features extracted by DLFS module. As shown in Table \ref{tabel:nyu21}, by training DLFS module with the proposed loss functions ${\cal L_C}$ and ${\cal L_{\text{VI}}}$, ``RGB-D Local (Full DLFS)'' can achieve better results than the baseline methods. This demonstrates that multi-scale discriminative local features are useful for scene recognition.
Since global RGB and depth features contain important scene layout information, which are complementary to local features. The proposed method exploits both local and global multi-modality features simultaneously, and can achieve better results than methods using global or local features alone.
\begin{figure}[!]
\centering
\includegraphics[width=0.48\textwidth]{CV4}
\caption{Class-specific activation map (CAM) \cite{DBLP:conf/cvpr/ZhouKLOT16} visualization of feature maps ${F_{rgb}}$. RGB images are shown in the first row. The second row shows the CAM of ``RGB-D Global \& Local(${\cal L_C}$)'', the third row shows the CAM of ``RGB-D Global \& Local(${\cal L_{\text{VI}}}$)'' and the fourth row shows the CAM of ``Full DLFS''.}
\label{visf}
\end{figure}
\begin{table*}[h!]
\centering
\caption{Experimental Results on SUN RGB-D Dataset}
\scalebox{1.26}{
\label{tabel:sun1}
\begin{tabular}{c|c|c|cc}
\hline \hline
Methods & Local Features & Multi-Modality Learning & Mean-class Accuracy(\%) \\ \hline
Song et al.\cite{song2015sun} & No & Feature-level Fusion & 39.0 \\ \hline
Liao et al.\cite{liao2016understand} & No & Image-level Fusion & 41.3 \\ \hline
Zhu et al.\cite{zhu2016discriminative} & No & \begin{tabular}[c]{@{}c@{}}Inter- \& Intra-\\ modality correlation\end{tabular} & 41.5 \\ \hline
Wang et al.\cite{wang2016modality} & CNN Proposals & \begin{tabular}[c]{@{}c@{}}Local \& Global Features Fusion\end{tabular} & 48.1 \\ \hline
Song et al.\cite{song2017rgb} & Object Detection & \begin{tabular}[c]{@{}c@{}}Local \& Global Features Fusion\end{tabular} & 54.0 \\ \hline
Song et al. \cite{song2017combining} & No & Global features & 52.3 \\ \hline
Song et al. \cite{song2017depth} & Local Patches & Local \& Global features Fusion & 52.4 \\ \hline
Li et al.\cite{LiZCHT18} & No & Modality Distinction \& Correlation & 54.6 \\ \hline
Song et al.\cite{song2018learning} & Patches Sampling & Feature-level Fusion & 53.8 \\ \hline
Xiong et al.\cite{xiong2019rgb} & Feature Selection & Local \& Global features Fusion &55.9 \\ \hline
Song et al. \cite{song2019image} & Object Detection & Local \& Global features Fusion & 55.5 \\ \hline
ASK (K=16 \& K=4) & Local Feature Selection & \begin{tabular}[c]{@{}c@{}}Global \& Local Features\\ Modality Distinction \end{tabular} & \textbf{57.3} \\ \hline
\end{tabular}
}
\end{table*}
\subsubsection{\textbf{Multi-modality \& Global \& Local}}
To study the effect of different loss functions of DLFS module, we do experiments to evaluate our framework with only ${\cal L_C}$ and only ${\cal L_{\text{VI}}}$. As shown in the table, ``RGB-D Global \& Local (${\cal L_C}$)'' denotes the model with single-scale DLFS module and training loss ${\cal L_C}$. This indicates that using ${\cal L_C}$ to learn local features can improve the scene classification performance. The main reason is that local deep grid features are complement to global scene features. The loss ${\cal L_{\text{VI}}}$ is also effective for improving the performance, as it can encourage the selected local features to be correlated with the scene class. The results indicate that both loss functions are useful for improving the DLFS module.
Moreover, we have visualized the class-specific activation map of the learned features with ${\cal L_C}$ loss, ${\cal L_{\text{VI}}}$ loss and both losses, i.e. ``Full DLFS''. From Fig. \ref{visf} we can see that the features of ``Full DLFS'' are more spatially-correlated.
\subsection{SUN RGB-D Results}
The comparison results on SUN RGB-D dataset are displayed in Table \ref{tabel:sun1}. Among the compared methods, Song et al. \cite{song2015sun}, Liao et al. \cite{liao2016understand} and Zhu et al. \cite{zhu2016discriminative} did not use local features. Wang et al. \cite{wang2016modality} and Song et al. \cite{song2017rgb} employed object detection for scene recognition.
Generally, the experimental results reveal that methods with local features can obtain better performance than those without local features. Different from existing methods, the proposed method selects multi-scale local features adaptively for scene recognition, and achieves even better performance.
We also summarize the multi-modality feature learning types in Table \ref{tabel:sun1}. Feature-level fusion are commonly used multi-modality feature learning methods, while considering the correlation and distinction between different modalities can achieve better results than simply combining multi-modality features. Different from existing methods, the proposed method exploits local features to learn modal-correlated representations. Additionally, the proposed method exploits the spatial-distribution of multi-modality feature correlation to enhance the local feature mining process. Meanwhile, the learned local features are also encouraged to be more modality-correlated by the proposed loss. By using this mechanism cleverly, the proposed framework with DLFS module achieves state-of-the-art scene recognition result.
\subsection{NYUD v2 Results}
Since NYUD v2 dataset is relatively small and the training data in NYUD v2 dataset is heavily imbalanced, we use the weights pretrained on SUN RGB-D dataset for model initialization on NYUD v2 dataset. The comparison results on NYUD v2 dataset are displayed in Table \ref{tabel:nyu1}. Similar to SUN RGB-D dataset, from the results we can see that methods using local features can achieve better performance. Although Li et al.\cite{LiZCHT18} did not use local features, they can still obtain competitive result by taking advantage of better multi-modality learning method. Compared with feature selection based method \cite{xiong2019rgb}, this work can achieve better performance by the differentiable feature selection module and the effective training loss functions.
The main advantages of the proposed method is two-fold: 1) local features extracted by DLFS module is more effective than patch-sampling and object detection methods; 2) Global modality-distinctive and local modality-correlated features are jointly exploited in this work. To sum up, the comparison results on this dataset indicate the effectiveness of the proposed framework.
\begin{table*}[]
\centering
\caption{Experimental Results on NYUD v2 Dataset}
\label{tabel:nyu1}
\scalebox{1.33}{
\begin{tabular}{c|c|c|c}
\hline \hline
& Methods & Local Features &Mean-class Accuracy(\%) \\ \hline
\multirow{8}{*}{State-of-the-art Methods}
& Gupta et al. \cite{gupta2015indoor} & No & 45.4 \\ \cline{2-4}
& Wang et al.\cite{wang2016modality} & CNN Proposals & 63.9 \\ \cline{2-4}
& Li et al.\cite{LiZCHT18} & No & 65.4 \\ \cline{2-4}
& Song et al.\cite{song2017depth} & Local Patches & 65.8 \\ \cline{2-4}
& Du et al.\cite{du2019translate} & No & 66.5 \\ \cline{2-4}
& Song et al.\cite{song2017rgb} & Object Detection & 66.9 \\ \cline{2-4}
& Song et al.\cite{song2018learning} & Patches Sampling & 67.5 \\ \cline{2-4}
& Xiong et al.\cite{xiong2019rgb} & Feature Selection & 67.8 \\ \cline{2-4} \hline
{Proposed Method} & ASK (K=16 \& K=4) & {Local Features Selection} & \textbf{69.3} \\ \hline
\end{tabular}}
\end{table*}
Additionally, we have also done experiments to study the effect of selecting different number of local features. The results are displayed in Table \ref{sk}. \emph{`K=0'} denotes that no local features are used. Generally, the classification performance increases with more selected local features. However, when $K$ is set to larger than 16, the performance decreases. When we select more features (\emph{`K=36'}), the performance also decreases, which is only slightly better than \emph{`K=0'}. The reason is that when more local features are selected, more noise will also be introduced, and the global features may be suppressed. Object or theme-level features will be affected by the irrelevant noisy features.
\begin{table*}[]
\centering
\caption{Number of Selected Features on NYUD v2 Dataset}
\scalebox{1.5}{
\begin{tabular}{c|c|c|c|c|c|c|c}
\hline \hline
Number of Selected Features (K) & K=0 & K=1 & K=3 & K=9 & \textbf{K=16} & K=25 & K=36 \\ \hline
Mean-class Accuracy (\%) & 65.4 & 66.2 & 67.1 & 67.9 & \textbf{69.1} & 67.7 & 66.1 \\ \hline
\end{tabular}}
\label{sk}
\end{table*}
The selected multi-scale keypoints are visualized in Fig. \ref{visp}. Different colors represent different scales. It can be clearly seen that semantic local features are selected when the modal converges.
\begin{figure}
\centering
\includegraphics[width=0.49\textwidth]{CV5}
\caption{Illustration of the selected multi-scale keypoints during the training stage. Semantic-meaningful local features are selected when the modal converges. Different colors represent different scales. (Best viewed in color.)}
\label{visp}
\end{figure}
\section{Comparisons with existing methods}
\label{cwem}
To show more details and give more comprehensive evaluations of the proposed work, we compare our work to other related existing methods and highlight the differences and similarities with them.
\begin{figure*}
\centering
\includegraphics[width=0.95\textwidth]{sup1}
\caption{The attention map visualization of spatial attention based method \cite{wang2018non} and the selected keypoints visualization of the proposed method. The images of the first, third, fifth and seventh rows are the visualization for spatial attention based method. The second, fourth, sixth and eighth rows are the visualization of the selected keypoints of the proposed method. The upper and lower image pair is corresponding to the same image in NYUD v2 dataset \cite{Silberman:ECCV12}.}
\label{vissao}
\end{figure*}
The highlight of the proposed framework is that it selects multi-scale intermediate CNN features instead of sampling patches densely or using the extra object detection procedure. Several other works have also proposed unsupervised local-part localization methods. Here we compare our method with them and point out the main differences with them.
\begin{figure*}
\centering
\includegraphics[width=0.92\textwidth]{cosine.pdf}
\caption{Spatial cosine-similarity map visualization. Images of the second, fourth and sixth columns visualize cosine-similarity maps.}
\label{cosinesimilarity}
\end{figure*}
\textbf{Spatial-related Multi-modal Feature Learning} \cite{xiong2019rgb} found that similar spatial-attention maps are gained with attention mechanism for RGB and depth modalities, and enforcing similar spatial-attention map can boost the performance. This indicates that the same local objects are important features for both RGB and depth modalities. Inspired by this, we find that local objects contribute highly to the multi-modal correlation. Thus, we propose to maximize the correlation between RGB and depth local features, which provides more cues and supervision for local object-level feature selection.
\textbf{KeypointNet \cite{suwajanakorn2018discovery}} KeypointNet is an end-to-end geometric reasoning framework for learning latent category-specific 3D keypoints. KeypointNet can discover geometrically and semantically consistent keypoints adaptively with no extra annotations. KeypointNet stacks 13 layers of dilated convolutions to output 2$N$ probability maps for predicting the 3$D$ coordinates of $N$ points. However, our work uses only one extra convolution layer with discriminative loss supervision for predicting probability maps, which is more time-efficient. Moreover, we use \emph{stride} 2 convolution layer instead of dilated convolution to select the multi-scale CNN features, which is also more time-efficient. Additionally, a novel variational mutual-information maximization loss term is proposed in this work for discriminative training.
\textbf{MA-CNN \cite{zheng2017learning}} MA-CNN consists of convolution, channel grouping and part classification sub-networks, which defines the parts as multiple attention areas. This work also employs the discriminative feature channels for part localization. However, the main difference is two-fold. 1) the local-part features are pooled from the attention areas in their work; 2) Multi-scale features are neglected. Different from their work, the proposed framework can learn to select multi-scale feature vectors in a totally differentiable manner.
\textbf{Deep Regionlets \cite{xu2018deep}} The architecture of deep regionlets includes a region selection network, which can learn more fine-grained features by selecting sub-regions adaptively. The local part feature selection is based on a spatial transformation and a gating network. STN \cite{DBLPJaderbergSZK15} is employed to select local regions as local-part features. However, the proposed method aims to select multi-scale mid-level CNN feature vectors (keypoints in CNN feature maps) as the local discriminative representations.
\textbf{Discriminative Filter Bank \cite{wang2018learning}} This work also exploits mid-level CNN representations by learning a bank of convolutional filters that capture class-specific discriminative patches without extra part or bounding box annotations. Specifically, this work uses Global Max Pooling (GMP) to select only one local feature vector from the intermediate CNN feature maps, which is different from our method. Moreover, multi-scale local features are also not considered in this work.
\textbf{Spatial Attention \cite{wang2018non}} To select important local features from mid-level CNN feature maps, spatial attention is the most commonly employed method. Although non-local networks \cite{wang2018non} can be used to focus on important local regions of feature maps, the softly attended feature maps still contain irrelevant features. However, our hard selection based method can select the important local feature vectors and discard the irrelevant ones. Moreover, the proposed method can select multiple multi-scale local features, which are more discriminative and representative than the spatial attention method. The visualization of the attention maps and the selected local feature vectors of our method are displayed in Fig. \ref{vissao}. The upper rows show the results of spatial attention based method, and the lower rows show the results of the proposed method. As we can see from Fig. \ref{vissao}, the spatial attention maps mainly focus on one local object regions, while our method can select multi-scale object-level (`chair' or `wash basin') and theme-level (`curtain' or `floor') feature vectors.
As displayed in Fig. \ref{cosinesimilarity}, it can be clearly seen that the correlation between RGB and depth modalities is highly spatially related. The correlation of local objects are higher than other positions. Inspired by this intriguing finding, we can maximize the correlation between selected local multi-modal features to locate the positions of local objects.
\section{Conclusion}
In this work, we present a multi-modal global and local feature learning framework for RGB-D scene classification. The differentiable local feature selection (DLFS) module is proposed to select important local object and theme-level features adaptively for RGB-D scene images. A novel loss function is proposed to supervise the training of DLFS module. Discriminative local object and theme-level representations can be selected with DLFS module from the spatially-correlated multi-modal RGB-D features. We take advantage of the correlation between RGB and depth modalities to provide more cues for selecting local features. Additionally, we further enhance the DLFS module with the multi-scale feature pyramid to select object-level features of different scales. Evaluations on SUN RGB-D and NYU Depth version 2 (NYUD v2) datasets have shown the effectiveness of the proposed framework.
\section{Acknowledgment}
This work was supported by the National Key Research and Development Project under Grant 2020YFB 2103902, National Natural Science Foundation of China under Grant 61632018, 61825603, U1801262, and 61871470.
\bibliographystyle{IEEEbib}
\section{Introduction}
Scene recognition is a fundamental task for computer vision. Recent progress made in deep convolutional neural networks (CNNs) has greatly boosted the image classification performance on large-scale benchmarks like ImageNet \cite{krizhevsky2012imagenet}. Modern deep CNN architectures such as ResNet \cite{DBLP:conf/cvpr/HeZRS16} and DenseNet \cite{Li2017Classification} are well designed for high-semantic level image representation learning. However, directly applying these deep CNNs for scene recognition still suffers from a limitation: global image features are not flexible enough to represent the indoor scene image with cluttered objects and complex spatial layouts. Considering the difference between image classification and scene recognition, Zhou et al. \cite{DBLP:conf/nips/ZhouLXTO14} released a large scale scene classification dataset named \emph{Places}. They showed the effectiveness of pre-training CNN parameters on Places instead of the object-centric dataset \emph{ImageNet}.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{CV1}
\caption{The motivation of the proposed method.}
\label{motivation}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=0.95\textwidth]{CV2}
\caption{The whole architecture of the proposed method. 1) \textbf{Global modal-specific feature learning.} The final layer feature maps of two modalities are input to two fully connected (FC) layers to learn global discriminative features with the supervision of auxiliary cross-entropy loss separately. 2) \textbf{Local modal-specific feature learning.} Meanwhile, the final layer feature maps are also used to construct a multi-scale feature pyramid, and then DLFS modules adaptively select multi-scale intermediate CNN features for scene recognition.}
\label{arch}
\end{figure*}
The local object-level intermediate features are complementary to global CNN features \cite{DBLP:conf/eccv/KimF18}. Thus, selecting local features can be effective for taming the great geometric variability of scene image. The local-feature based methods can be roughly divided into two categories. 1) Local patch-sampling based methods; 2) Object detection based methods. Several works \cite{DBLP:conf/eccv/GongWGL14,DBLP:journals/corr/YooPLK14,DBLP:conf/eccv/ZuoWSZYJ14} opted to extract features from different scales and locations densely and combined them via the fisher vector (FV) \cite{DBLP:journals/ijcv/SanchezPMV13}. However, two disadvantages exist in these methods, which limit the further performance improvement. The first one is that global layout features are neglected. The second one is that densely-sampled local features may induce scene-irrelevant noise, which makes the learned features less discriminative. Object detection can also be used to extract more accurate object-level local features. However, the performance of these methods greatly relies on the accuracy of object detection. Unfortunately, detecting the cluttered objects accurately in complex indoor scenes is also nontrivial.
Moreover, existing local feature based methods may also suffer from the following problems. Not all local object-level features contribute to the discriminative scene representations. For example, the `chair' features in dinning room and the `chair' features in classroom are not discriminative for recognizing these two scenes. Merely using the local features may suffer from the ambiguity for recognizing different scenes. Some theme-level features are also important for scene classification, while object detection based methods may neglect them. For example, `floor' and `curtains' are background theme-level features, but they are also critical for recognizing scenes.
Additionally, RGB-D scene recognition can make use of the multi-modal feature learning process to enhance the local feature selection. Since RGB and depth modality are highly spatial-aligned, the correlation between the same local object of each modality is also high. Therefore, exploiting the multi-modal local feature correlation can do help to object-level local feature selection, which has not been investigated by previous works.
In this paper, we design a multi-modal feature learning framework for RGB-D scene recognition, which adaptively selects multi-scale key local features for scene recognition. As shown in Fig. \ref{motivation}, our motivation is to select multiple mid-level CNN feature vectors to represent local patches. To deal with the ambiguity problem caused by local features, the proposed framework also learns to extract global features which are important for describing the scene layout. The main contributions are as follows.
\begin{enumerate}
\item An effective differentiable local feature selection module is proposed, which can select the appropriate number of local features efficiently. This module can not only extract local object-level features, but also select local theme-related features for RGB-D scene recognition.
\item We introduce a mutual information maximization loss for training the DLFS module, which encourages the discrimination of selected local features. Moreover, spatial-related correlation of multi-modal feature is first explored for local feature selection in RGB-D scene recognition.
\item We design a compact global and local multi-modal feature learning framework to learn more discriminative representations for RGB-D scene recognition.
\end{enumerate}
\section{Related Work}
\label{Related Work}
\textbf{Patch-sampling based methods.} These approaches extract local features from the patch-based CNN intermediate representations. Gong et al. \cite{DBLP:conf/eccv/GongWGL14} designed a multi-scale CNN framework to sample the local patch features densely, and then encoded them via VLAD \cite{DBLP:conf/cvpr/JegouDSP10}. Some other methods \cite{DBLP:journals/corr/YooPLK14,dixit2015scene} represented the scene image with multi-scale local activations via the FV encoding. Depth image patches were exploited in the work of Song et al. \cite{song2017depth}. They first trained the model with densely sampled depth patches in a weakly-supervised manner, and then fine-tuned the model with the full image. Nevertheless, densely sampled patch features may contain noise, which limits the scene recognition performance.
\textbf{Object-detection based methods.} To discard the irrelevant local features, several methods employed object detection for more accurate object-level features. Wang et al. \cite{wang2016modality} exploited the local region proposals to extract component representations, and encoded the local and global features together via fisher vector. Song et al. \cite{song2017rgb} employed Faster RCNN to detect objects on both RGB and depth images. More accurate object-level local features could be obtained with the object detection sub-module. They further modeled the object-to-object relation and achieved state-of-the-art scene recognition results. Although improved performance could be obtained by \cite{wang2016modality,song2017rgb}, the two-stage pipeline methods suffered from the error accumulation problem. The higher computational complexity is also a limitation of these methods. Moreover, detecting small clutters objects in indoor scenes is nontrivial itself.
Selecting intermediate CNN representations is useful for many applications. \cite{suwajanakorn2018discovery} presented an end-to-end geometric reasoning framework named KeypointNet to learn latent category-specific 3D keypoints. KeypointNet can discover geometrically and semantically consistent keypoints adaptively without extra annotations. Zheng et al. \cite{zheng2017learning} proposed a multi-attention convolutional neural network, which consisted of convolution, channel grouping and part classification sub-networks. Wang et al. \cite{wang2018learning} showed that intermediate CNN representations could be enhanced by learning a bank of convolutional filters that captured class-specific discriminative patches without extra bounding box annotations. Xu et al. \cite{xu2018deep} proposed deep regionlets for object detection, which could select non-rectangular regions within the detection framework.
Multi-modal feature learning is critical for RGB-D scene recognition, and various methods have been proposed \cite{wang2018detecting}. Song et al. \cite{song2015sun} fused the two modal features by concatenating them to one fully connected layer. Wang et al. \cite{wang2015mmss} proposed to learn modal-consistent features between RGB and depth images. Li et al. \cite{LiZCHT18} learned the modal-consistent and modal-distinctive embeddings between two modalities simultaneously. Spatial correspondence of local objects in RGB and depth modality was exploited by \cite{xiong2019rgb} for multi-modal learning. \cite{du2019translate} employed cross-modal translation to explicitly regularize the training of scene recognition, which improved the generalization ability of the model.
\section{Methodology}
The whole network architecture of the proposed method is shown in Fig. \ref{arch}. RGB and HHA images are input to two branches of CNNs for feature extraction. Then, the final layer feature maps of the two modalities are used for global and local multi-modal feature learning. Finally, the global and local features of two modalities are concatenated together to form the final scene representation.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{CV3}
\caption{The illustration of the DLFS module.}
\label{ASKFS}
\end{figure}
\subsection{Differentiable Local Feature Selection}
CNNs can learn to extract high semantic-level features with stacked convolution layers. Layer visualization shows that the intermediate CNN features are with high abstract-level and can represent object parts. Moreover, the work of \cite{DBLP:conf/eccv/KimF18} found that the global scene layout feature and the object-level feature were complementary for scene recognition. Based on these insights, our motivation is to select multiple $1\times 1$ mid-level CNN feature vectors to represent the local patches. Specifically, our model predicts $K$ keypoints with the final layer feature maps of RGB and depth modality, as illustrated in Fig. \ref{ASKFS}.
Suppose that the final layer feature map of RGB and depth modality are ${F_{rgb}}$ and ${F_{d}}$ respectively. We then concatenate these two feature maps as $F_{rgbd}$ to exploit multi-modal information for feature selection. The DLFS module takes ${F_{rgbd}}$ as input, and uses a $1\times 1$ convolution layer to learn the class-specific feature map for estimating $K$ keypoints. This can be formulated as
\begin{equation}
{V_{rgbd}} = {conv_{1\times 1}(F_{rgbd})},
\end{equation}
where $V_{rgbd} \in \mathbb{R}^{(C,H,W)}$ is the discriminative feature map used for predicting keypoints. $C$ is the number of channels. $H$ and $W$ are the height and width of the feature maps respectively. ${conv_{1\times 1}}$ is a convolution layer with ${1\times 1}$ kernel size and output channel $C$. Although the channels of CNN feature maps are spatially-correlated, one separated channel is not enough to have strong part response \cite{zheng2017learning}. Thus, the channels of $V_{rgbd}$ are further grouped to be more spatially-correlated. Then we reshape feature ${V_{rgbd}}$ to ${V_{rgbd} \in \mathbb{R}^{(K,C/K,H,W)}}$ for the following group channel pooling operation. To make the part response stronger, we sum up the channel groups of ${V_{rgbd}}$ as
\begin{equation}
M_{rgbd}^j = \sum\limits_{i = 1}^{C/K} {{V_{rgb{d_i}}}},
\end{equation}
where ${M_{rgbd}\in \mathbb{R}^{(K,H,W)}}$ is the cross-channel grouped feature map for predicting $K$ selected keypoints. ${j \in \{1,2,...,K\}}$ is the index of the channel groups.
Directly training a mapping function from the input feature maps to keypoints is difficult. The reason is that the mapping model is hard to converge without keypoints annotations for supervised training. Inspired by \cite{suwajanakorn2018discovery}, we make our model output an attention map $h_j(u,v)$ to represent the probability of keypoint $j$ occurring at position $(u,v)$. In this work, we use the cross-channel grouped discriminative feature
maps $M_{rgbd}$ to output the attention map. Specifically, a 2D softmax layer is employed to produce the map $h$, which can be represented as
\begin{equation}
\begin{gathered}
h_j = softmax(M_{rgbd}^{j}) \in \mathbb{R}^{(H,W)},
\end{gathered}
\end{equation}
where ${h\in \mathbb{R}^{(K,H,W)}}$ is the probability distribution map, and $j$ is the index of predicted keypoints. Then the coordinates of the keypoints are computed by taking the expected values of the spatial distributions:
\begin{equation}
[{x_j},{y_j}] = \sum\limits_u^H {\sum\limits_v^W {[u \cdot {h_j}(u,v),v \cdot {h_j}(u,v)]} },
\end{equation}
where $[x_j, y_j]$ is the coordinate of the $j^{th}$ predicted keypoint.
To extract the local features in an end-to-end manner, the feature vector-sampling module should be differentiable. Without loss of generality, let ${E_{rgb} \in \mathbb{R}^{(K,C)}}$ be the selected local features for RGB modality. Similarly, we denote $E_{d}$ as the selected local features of the depth modality. Inspired by \cite{DBLPJaderbergSZK15}, the differentiable bilinear feature sampling can be formulated as
\begin{equation}
\begin{split}
E_{rgb_j}^c = \sum\limits_u^H \sum\limits_v^W F_{rgb}^c(u,v) max(0,1 - |{x_j} - v|) \\
\cdot max(0,1 - |{y_j} - u|),
\end{split}
\end{equation}
where ${j \in \{1,2,...,K\}}$ is the index of $K$ sampled local feature vectors, and ${c \in [1...C]}$ is the channel index. The coordinates ${(x_j,y_j)}$ and ${(u,v)}$ are normalized in the range of [-1,1].
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{CV8-2}
\caption{The illustration of the proposed variational information maximization loss.}
\label{LOSS1}
\end{figure}
Since the DLFS module takes the keypoint index and the feature maps as input, we need to compute the partial derivatives of $F_{rgbd}$ as well as the predicted coordinates to allow the backpropagation through this module. The partial derivative w.r.t feature maps ${F_{rgb}}$ and $(x_j,y_j)$ are presented in the supplementary material.
For multi-scale local feature selection, we can construct feature pyramid with stride 2 convolutional layers as shown in Fig. \ref{arch}. For each scale CNN feature maps, the local feature selection process is similar.
\begin{figure}
\centering
\includegraphics[width=0.48\textwidth]{CV8-1}
\caption{The illustration of the proposed loss function for training DLFS module. The first term is the triplet-correlation loss, and the second one is the multi-modal local feature correlation loss. }
\label{LOSS}
\end{figure}
\subsection{Discriminative Training for DLFS}
Although the DLFS module can be trained in an end-to-end manner, the feature map ${M_{rgbd}}$ is not guaranteed to be discriminative enough for selecting different local patch features. To encourage the channels of $M_{rgb}$ to be sensitive to different semantic parts, we propose a novel loss function to train the DLFS module.
In order to make ${M_{rgbd}}$ more discriminative, we choose to retain high mutual information between ${M_{rgbd}}$ and the scene class label. Since we aim to select feature vectors which are important for classifying different scenes, the feature map ${M_{rgbd}}$ should be highly correlated to different scene classes. By maximizing the mutual information between ${M_{rgbd}}$ and the class label $L$, the channels of ${M_{rgbd}}$ is enforced to be class-correlated. As the exact computation of mutual information is intractable, variational lower bound is used for the approximation in this work.
\begin{figure*}
\centering
\includegraphics[width=0.96\textwidth]{CV9}
\caption{The pixel-wise correlation map between deep features of RGB and depth modality. RGB images are on the left, and the corresponding correlation-maps are displayed on the right.}
\label{visfk}
\end{figure*}
Suppose that there are $N$ classes, ${M_{rgbd}}$ is firstly pooled by a max-pooling layer to get $T_{rgbd} \in \mathbb{R}^{K}$. Then the FC layer is used to transform $T_{rgbd}$ into $N$ scalar features. Finally, we get $N$ features ${U \in \mathbb{R}^{N}}$ corresponding to $N$ scene categories. The mutual information ${I(L;U)}$ can be defined as
\begin{equation}
\begin{aligned}
&I(L;U) = H(L) - H(L|U)\\
&= H(L) + \mathbb{E}_{L,U}[\log q(L|U)]+KL\\
&\ge H(L) + \mathbb{E}_{L,U}[\log q(L|U)],
\end{aligned}
\end{equation}
where ${H(\cdot)}$ is the entropy, and $\text{KL}$ denotes the Kullback-Leiber divergence. Maximizing the mutual information is equivalent to minimizing the following loss function ${\cal L}_{\text{VI}} = -\mathbb{E}_{L,U}[\log q(L|U)]$. ${\cal L}_{\text{VI}}$ can also be interpreted as the reconstruction error. In this work, we choose the variational distribution as a Gaussian distribution with heteroscedastic mean:
\begin{equation}
\begin{aligned}
&{\cal L}_{\text{VI}} =\sum_{n=1}^{N}\log\sigma_n + \frac{(L_n-{U_{\mu}}_n)^2}{2\sigma_n^2} + c,
\end{aligned}
\end{equation}
where $c$ is a constant, and $U_{\mu}$ is the transformed features by FC layer as shown in Fig. \ref{LOSS1}. $\sigma$ is a learnable parameter to be optimized.
By minimizing ${\cal L}_{\text{VI}}$, the feature ${M_{rgbd}}$ is encouraged to be discriminative for selecting key local features.
\subsection{Unevenly distributed Multi-modal Correlation}
We find that local objects contribute highly to the multi-modal correlation. To verify this, we compute the pixel-wise correlation map $P\in \mathbb{R}^{H,W}$ between the deep features $F_{rgb}\in \mathbb{R}^{C,H,W}$ of RGB modality and the depth modal deep features $F_{d}\in \mathbb{R}^{H,W}$. Then the pixel-wise correlation map can be defined as:
\begin{equation}
\begin{gathered}
P^{ij} = \rho({F_{rgb}}^{ij},{F_{d}}^{ij}), \\
i=1,...,H;j=1,...,W,
\end{gathered}
\end{equation}
where $\rho$ is the cosine similarity function. ${F_{rgb}}^{ij} \in \mathbb{R}^{C}$ is a feature vector at position $(i,j)$ of feature $F_{rgb}$.
As shown in Fig. \ref{visfk}, we have visualized the pixel-wise correlation map $P$. It can be clearly seen that the correlation between two modalities is unevenly distributed. More specifically, large correlation values cluster on local objects. This indicates that multi-modal features of local objects have higher correlation than other spatial positions.
Based on this idea, local objects can be located by finding the largest correlation regions. Thus, we propose to maximize the correlation between RGB and depth local features to provide more cues and supervision for local object-level feature selection. The multi-modal correlation loss ${\cal L}_{C{m}}$ can be computed as:
\begin{equation}
{\cal L}_{C{m}} = \rho({E_{rgb}},{E_{d}}),
\end{equation}
where $\rho$ is the cosine similarity function. To further enhance the DLFS module with class-specific local features, triplet correlation loss is employed. As illustrated in Fig. \ref{LOSS}, we aim to select local features that have larger correlation between the same class and smaller correlation between different classes.
For each sample in the triplet input ${\{\textbf{a},\textbf{p},\textbf{n}\}}$, the corresponding selected local features are ${\{E_{rgb}a,E_{rgb}p,E_{rgb}n\}}$. $E_{rgb}p$ and $E_{rgb}a$ are positive and anchor features with the same class label, and $E_{rgb}n$ is the negative one with different scene class label. The triplet correlation loss is formulated as follows.
\begin{equation}
{\cal L}_{C{rgb}} = max\{ \rho({E_a},{E_p}) - \rho({E_a},{E_n}) + \alpha ,0\},
\end{equation}
where ${{\cal L}_{Crgb}}$ is the triplet correlation loss for the RGB modality, and the loss computation for depth modality ${{\cal L}_{Cd}}$ is similar. $\alpha$ is the margin value and it is set to 1.0 in this work. The whole correlation loss for two modalities can be formulated as ${{\cal L_C} = {\cal L}_{Crgb} + {\cal L}_{Cd}} + {\cal L}_{C{m}}$.
\begin{table*}[]
\centering
\caption{Experimental Results on NYUD v2 Dataset}
\label{tabel:nyu1}
\scalebox{1.05}{
\begin{tabular}{c|c|c|c}
\hline \hline
& Methods & Local Features &Mean-class Accuracy(\%) \\ \hline
\multirow{8}{*}{State-of-the-art} & Gupta et al. \cite{gupta2015indoor} & No & 45.4 \% \\ \cline{2-4}
& Wang et al. \cite{wang2016modality} & CNN Proposals & 63.9\% \\ \cline{2-4}
& Li et al. \cite{LiZCHT18} & No & 65.4\% \\ \cline{2-4}
& Song et al. \cite{song2017depth} & Local Patches & 65.8\% \\ \cline{2-4}
& \cite{du2019translate} & No & 66.5\% \\ \cline{2-4}
& Song et al. \cite{song2017rgb} & Object Detection & 66.9\% \\ \cline{2-4}
& \cite{song2018learning} & Patches Sampling & 67.5\% \\ \cline{2-4}
& \cite{xiong2019rgb} & Feature Selection & 67.8\% \\ \cline{2-4} \hline
Proposed & Our method (K=16 \& K=4) & {Local Features Selection} & \textbf{69.7}\% \\ \hline
\end{tabular}}
\end{table*}
\begin{table*}[]
\centering
\caption{Experimental Results on SUN RGB-D Dataset}
\scalebox{1}{
\label{tabel:sun1}
\begin{tabular}{c|c|c|cc}
\hline \hline
Methods & Local Features & Multi-Modal Fusion &Mean-class Accuracy(\%) & \\ \hline
\cite{song2015sun} & No & Feature-level Concatenation & 39.0\% & \\ \hline
\cite{liao2016understand} & No & Image-level Concatenation & 41.3\% & \\ \hline
\cite{zhu2016discriminative} & No & \begin{tabular}[c]{@{}c@{}}Inter- \& Intra-\\ modality correlation\end{tabular} & 41.5\% & \\ \hline
\cite{wang2016modality} & CNN Proposals & \begin{tabular}[c]{@{}c@{}}Local \& Global Features\\ Concatenation\end{tabular} & 48.1\% & \\ \hline
\cite{song2017rgb} & Object Detection & \begin{tabular}[c]{@{}c@{}}Local \& Global Features\\ Concatenation\end{tabular} & 54.0\% & \\ \hline
\cite{LiZCHT18} & No & Modality Distinction \& Correlation & 54.6\% & \\ \hline
\cite{song2018learning} & Patches Sampling & Feature-level Concatenation & 53.8\% \\ \hline
\cite{xiong2019rgb} & Feature Selection & Local \& Global features & 55.9\% \\ \hline
\cite{du2019translate} & NO & Cross-Modal Translation & 56.1\% \\ \hline
Our method & Local Feature Selection & \begin{tabular}[c]{@{}c@{}}Global \& Local Features\\ Modality Distinction \end{tabular} & \textbf{56.8}\% & \\ \hline
\end{tabular}
}
\end{table*}
\subsection{Joint Global and Local Feature Representation}
Though local features are effective for representing the scene images, merely using local features may suffer from the ambiguity problem. Thus in this work, we choose to learn the global and local features simultaneously to obtain more robust representations for scene recognition.
The global features are learned by the FC layers connected to the final feature maps ${F_{rgb}}$ and ${F_{d}}$. As shown in Fig. \ref{arch}, two auxiliary cross-entropy loss functions are employed for learning global modal-specific feature separately. This loss function can be formulated as
\begin{equation}
{{\cal L}_{aux}} = {{{\cal L}_{CE}}({G_{rgb}},{y})} + {{{\cal L}_{CE}}({G_{d}},{y})},
\end{equation}
where $L_{CE}$ represents the cross-entropy loss.
Finally, we concatenate the multi-modal global and local feature together for the final scene classification. This can be denoted as
\begin{equation}
\begin{gathered}
H_{mmgl}=concat(G_{rgb},G_{d},E_{rgb},E_d),\\
\end{gathered}
\end{equation}
where ${H_{mmgl}}$ is the multi-modal global and local feature vector. Then ${H_{mmgl}}$ is input to a fully connected layer, and the final classification result ${\hat y}$ is output through a softmax layer. We denote the final cross-entropy classification loss as ${{\cal L}_{cls}}$.
\begin{table*}[]
\centering
\caption{Number of Selected Features on NYUD v2 Dataset}
\begin{tabular}{c|c|c|c|c|c|c|c}
\hline \hline
Number of Selected Features (K) & K=0 & K=1 & K=3 & K=9 & \textbf{K=16} & K=25 & K=36 \\ \hline
Accuracy (\%) & 0.654 & 0.662 & 0.671 & 0.679 & \textbf{0.691} & 0.677 & 0.661 \\ \hline
\end{tabular}
\label{sk}
\end{table*}
The overall loss function for training the proposed framework is a multi-task loss, which consists of four terms: 1) the global modal-specific auxiliary loss ${\cal L}_{aux}$; 2) the mutual-information maximization loss ${\cal L}_{\text{VI}}$; 3) the triplet correlation and multi-modal correlation loss ${\cal L}_C$; 4) the final classification loss ${\cal L}_{cls}$. This can be represented as
\begin{equation}
\begin{aligned}
{\cal L} = {\cal L}_{cls} + \lambda_1 {\cal L}_{aux} + \lambda_2 {\cal L}_{\text{VI}} + \lambda_3 {\cal L}_C.
\end{aligned}
\end{equation}
where $\lambda_1$, $\lambda_2$, and $\lambda_3$ are balancing weights of loss components for the multi-task loss function.
\begin{figure*}
\centering
\includegraphics[width=0.94\textwidth]{CV5}
\caption{Illustration of the selected multi-scale keypoints. Different colors represent different scales. (Best viewed in color.)}
\label{visp}
\end{figure*}
\section{Experiments}
Two popular RGB-D scene recognition datasets are used to evaluate the proposed framework. One is the \textbf{SUN RGB-D} \cite{song2015sun} dataset, which contains 10,355 RGB and depth image pairs. They are captured from different depth sensors including Kinect v1, Kinect v2, Asus Xtion and RealSense. These images are divided into 19 scene categories. To compare with existing methods, we follow the same experimental settings with \cite{song2015sun}. For this dataset, 4,848 image pairs are used for training and 4,659 pairs for testing. Another dataset is the \textbf{NYUD v2} \cite{Silberman:ECCV12} dataset, which includes 1,449 RGB and depth image pairs divided into 10 categories. Following the experimental setting in \cite{gupta2013perceptual}, 795 image pairs are used for training and 654 pairs for testing.
\begin{figure*}
\centering
\includegraphics[width=0.96\textwidth]{CV4}
\caption{The class-specific activation map (CAM) \cite{DBLP:conf/cvpr/ZhouKLOT16} visualization of the learned feature maps ${M_{rgbd}}$. The RGB images are shown in the first row. The second row shows the features of `RGB-D Global \& Local(${\cal L_C}$)', the third row shows the features of `RGB-D Global \& Local(${\cal L_{\text{VI}}}$)' and the fourth row shows features of `Full DLFS'.}
\label{visf}
\end{figure*}
\subsection{Implementation Details}
HHA encodings is computed with the code released by \cite{gupta2014learning}. We use ResNet18 as the network backbone. Pre-trained parameters on \textbf{Places} are used for fine-tuning. Data augmentation is used in our work including random flip, cutout and random erasing \cite{zhong2017random}.
\begin{table*}[]
\centering
\caption{Ablation Study on NYUD v2 Dataset}
\label{tabel:nyu21}
\scalebox{1.1}{
\begin{tabular}{c|c}
\hline \hline
Methods & Mean-class Accuracy(\%) \\ \hline
RGB / Depth / HHA & 61.3 / 54.0 / 58.5 \\ \hline
RGB-D (HHA Encoding) & 64.2 \\ \hline \hline
RGB-D Global (${{\cal L}_{aux}}$) & 65.4 \\ \hline
RGB-D Global (Spatial Attention \cite{Wang2018NonlocalNN}) & 66.1 \\ \hline
RGB-D Global \& Local(${\cal L_C}$) & 67.3 \\ \hline
RGB-D Global \& Local(${\cal L_{\text{VI}}}$) & 68.3 \\ \hline \hline
RGB-D Global \& Local(Full DLFS) HHA & \textbf{69.1} \\ \hline
RGB-D Global \& Local(Two Scale Full DLFS) & \textbf{69.7} \\ \hline \hline
\end{tabular}
}
\end{table*}
For optimization, Adam \cite{kingma2014adam} is employed with an initial learning rate of 1e-4, and the learning rate is reduced by a fraction of 0.9 every 80 epochs. The batch size is set to 64 with shuffle. For all the experiments, 300 epochs are used to train the model. As for the multi-task training, we set the parameters ${\lambda_1}$ to 1, and ${\lambda_2}$, ${\lambda_3}$ are set to 0.1 for all experiments. Multi-scale DLFS is evaluated with the ResNet18 backbone. The first scale feature maps are with the size of ${7\times7}$. The second scale is obtained with a $1\times1$ kernel convolution layer. By setting the stride to 2, the second scale feature maps are with the size of ${3\times3}$. For the first scale, $K$ is set to 16 and it is set to 4 for the second scale.
\subsection{SUN RGB-D Results}
The comparison results on SUN RGB-D dataset are displayed in Table \ref{tabel:sun1}. Song et al. \cite{song2015sun}, Liao et al. \cite{liao2016understand} and Zhu et al. \cite{zhu2016discriminative} used no local features. Wang et al. \cite{wang2016modality} and Song et al. \cite{song2017rgb} employed object detection for scene recognition. Generally, the experimental results reveal that methods with local features can obtain better performance than those with no local features. Different from existing methods, the proposed method selects multi-scale local features adaptively for scene recognition, and achieves even better performance.
\subsection{NYUD v2 Results}
The comparison results on NYUD v2 dataset are displayed in Table \ref{tabel:nyu1}. From the results we can see that methods using local features can achieve better performance. Among them, the proposed method with adaptively selected local features can outperform patch-sampling and object detection methods, which indicates the effectiveness of our framework.
Additionally, we have also done experiments to study the effect of selecting different number of local features. The results are displayed in Table \ref{sk}. \emph{`K=0'} denotes that no local features are used. Generally, the classification performance increases with more selected local features. However, when $K$ is set to larger than 16, the performance decreases. When we select all the features (\emph{`K=36'}), the performance also decreases, which is only slightly better than \emph{`K=0'}. The reason is that when more features are selected, more noise will also be introduced. Object or theme-level features will be affected by the irrelevant noisy features.
\subsection{Ablation Study}
To evaluate the proposed framework comprehensively, we do the following ablation studies with ResNet18 backbone to explore the effect of different sub-modules. The results are shown in Table \ref{tabel:nyu21}. The performance of \emph{`RGB'} and \emph{`Depth (HHA)'} is relatively low, which indicates that using single modality features is not effective enough for recognizing scenes. `RGB-D Global (${{\cal L}_{aux}}$)' improves the baseline method by 1.2\%, which shows the effectiveness of the auxiliary loss for learning global modal-specific features.
Moreover, we also compare the proposed method with the spatial attention method. We use the non-local neural networks \cite{Wang2018NonlocalNN} to focus on different spatial regions with assigned weights. However, the proposed method selects key local features and abandon other features, which can be viewed as ``hard-attention". Using spatial-attention to focus on local features softly obtains a performance of 0.661, which is lower than the proposed local feature selection method.
To study the effect of different loss functions of DLFS module, we do experiments to evaluate our framework with only ${\cal L_C}$ and only ${\cal L_{\text{VI}}}$. The results show that both loss functions are useful for improving the DLFS module. Moreover, we have visualized the learned feature maps of our framework with only ${\cal L_C}$, only ${\cal L_{\text{VI}}}$ and the Full DLFS. From Fig. \ref{visf} we can see that the features of `Full DFKS' are more spatially-correlated.
The selected multi-scale keypoints are visualized in Fig. \ref{visp}. Different colors represent different scales.
\subsection{Conclusion}
In this work, we present a multi-modal global and local feature learning framework for RGB-D scene classification. The differentiable local feature selection (DLFS) module is proposed to select important local object and theme-level features adaptively for RGB-D scene images. A novel loss function is proposed to supervise the training of DLFS module. Discriminative local object and theme-level representations can be selected with DLFS module from the spatially-correlated multi-modal RGB-D features. We take advantage of the correlation between RGB and depth modality to provide more cues for selecting local features. Additionally, we further enhance the DLFS module with the multi-scale feature pyramid to select object-level features of different scales. Evaluations on SUN RGB-D and NYU Depth version 2 (NYUD v2) datasets have shown the effectiveness of the proposed framework.
{\small
\bibliographystyle{ieee_fullname}
|
\section{Directed percolation critical exponents}
For the purposes of this paper we used the following values for the critical exponents of the directed percolation universality class based on values presented in \cite{henkel2008non} collected from various original sources \cite{jensen1999low,voigt1997epidemic,grassberger1996self,jensen1992critical}. The implementation (namely the grid-search step) requires all exponents to be truncated to 3 digits after decimal point.
\begin{table}[h!]
\caption{\label{tab:dp-exponents} Values used for the critical exponents of the directed percolation universality class. The values \cite{henkel2008non,jensen1999low,voigt1997epidemic,grassberger1996self,jensen1992critical} were all truncated to three places after the decimal point.}
\begin{ruledtabular}
\begin{tabular}{crrrr}
\multicolumn{1}{c}{\textrm{exponent}} &
\multicolumn{1}{c}{\textrm{$d=1$}} &
\multicolumn{1}{c}{\textrm{$d=2$}} &
\multicolumn{1}{c}{\textrm{$d=3$}} &
\multicolumn{1}{c}{\textrm{Mean-field}}\\
\colrule
$\beta=\beta'$ & 0.276 & 0.583 & 0.814 & 1.000 \\
$\nu_\parallel$ & 1.734 & 1.295 & 1.110 & 1.000 \\
$\nu_\perp$ & 1.097 & 0.733 & 0.584 & 0.500
\end{tabular}
\end{ruledtabular}
\end{table}
Additionally the other ``dependent'' exponents, used throughout the paper and the implementation, can be derived from the exponents in Tab.~\ref{tab:dp-exponents} as follows:
\begin{equation}\label{eq:dp-exponents-relations}
\begin{aligned}
z &= \nu_\parallel / \nu_\perp \\
\alpha &= \beta/\nu_\parallel \\
\delta &= \beta'/\nu_\parallel \\
\gamma &= \nu_\parallel + d \nu_\perp - \beta - \beta' \\
\upsilon &= d \nu_\perp - \beta'
\end{aligned}
\end{equation}
Although the values used for the exponents not presented in Tab.~\ref{tab:dp-exponents} are calculated and used as double precision floating points in the implementation, their representative values up to three digits after decimal point is presented in Tab.~\ref{tab:dp-other-exponents} for the sake of comparison and quick reference.
\begin{table}[h!]
\caption{\label{tab:dp-other-exponents} Values used for the other critical exponents of the directed percolation universality class. The values presented here, calculated from Tab.~\ref{tab:dp-exponents} and Eq.\ref{eq:dp-exponents-relations} are truncated to three places after the decimal point for the purposes of this table.}
\begin{ruledtabular}
\begin{tabular}{crrrr}
\multicolumn{1}{c}{\textrm{exponent}} &
\multicolumn{1}{c}{\textrm{$d=1$}} &
\multicolumn{1}{c}{\textrm{$d=2$}} &
\multicolumn{1}{c}{\textrm{$d=3$}} &
\multicolumn{1}{c}{\textrm{Mean-field}}\\
\colrule
$z$ & 1.581 & 1.767 & 1.901 & 2.000 \\
$\alpha=\delta$ & 0.159 & 0.45 & 0.733 & 1.000 \\
$\gamma$ & 2.279 & 1.595 & 1.234 & 1.000 \\
$\upsilon$ & 0.821 & 0.883 & 0.938 & 1.000 \\
\end{tabular}
\end{ruledtabular}
\end{table}
\section{Implementation}
The experiments of this paper are implemented using mainly C++ and are made available online \cite{badie2021implementation}. The implementation and the paper itself makes use of the work of various other software projects \cite{tange2018gnu,yoo2003slurm,hagberg2008exploring,wolfram2021mathematica} as well as generous computational resources provided by the Aalto University Science-IT project and CSC -- IT Center for Science, Finland.
The different types of random networks are generated on the fly in the \verb|networks/| directory by the scripts provided in the \verb|scripts/| directory. Of the real-world datasets, analyzed in \cite{shortpaper}, the data for the US air transportation network \cite{bts2017air} and Helsinki public transportation network \cite{kujala2018collection} are provided along with the implementation.
The code can be compiled using the command \verb|make all|. Various C++17 features are used extensively throughout the code and it is only tested to compile with GCC 9.3.0, though it is expected to work with a more recent version of GCC as well.
You can generate a random temporal network using the \verb|random_network| executable:
\begin{verbatim}
$ ./random_network --seed 0 --nodes 512 --average-degree 9 --static-model regular \
--temporal-model poisson --mean-dt 1 --max-time 2048 > example/example-small.events
\end{verbatim}
or by passing a pre-generated static network to the \verb|--static-model|, instead of specifying one of the existing a static network model. To estimate quantities for source source limited-time spreading process on a synthetic temporal network with $\delta t = 0.8$, you can use a variation of the following command:
\begin{verbatim}
$ ./random_large_single --seed 0 --dt 0.8 --network example/example-small.events \
--summary /dev/stdout --window-min 0 --window-max 2048 \
--time-bins 16 --time-bins-min 0.25 --time-bins-max 2048
\end{verbatim}
which prints the results, including statistics on cluster mass, volume and lifetime and some basic information on the temporal network, in JSON format on standard output. The quantities are presented as average over all events, as well as binned logarithmically based on the time of the event and the time from the event to the end of the observation window. The lifetime and temporal distribution of events that produce ``censored'' percolation clusters, clusters that might span longer than the window of observation of the temporal network, are reported separately as well to facilitate estimating the probability of survival using Kaplan--Meier estimators \cite{kaplan1958nonparametric}. Similarly, if the same set of parameters is used with either of the executables \verb|random_large_homogeneous| or \verb|random_large_homogeneous_alternative|, it provides information about the distribution of occupied events in case of starting from homogeneous, fully occupied initial condition. The former accepts an additional parameter \verb|--field h| to simulate an external field of rate $h$ by spontaneously occupying the events, while the latter uses the much faster algorithm presented in this paper which calculates the average effect of occupying exactly one random event. The values reported for \verb|dt-band-homogeneous-occupied| and \verb|dt-band-external-field-occupied| corresponds to the number of normally occupied events in a $\delta t$ band before the end of the observation time window, and the number of occupied events in the presence of a minor external field that spontaneously occupies exactly one random event. The aggregation script \verb|aggregate_data.py| can provide more details on how the output of multiple runs of these executables results are aggregated and transformed into quantities that are reported in this manuscript.
Results for the synthetic networks were generated through the following iterative system: We started at an initial estimation for $\delta t_c$ and precision of $p=1$, where precision indicates the number of digits of certainty for $\delta t_c$ after the decimal point. At each step we calculated all the characteristic quantities for all possible values of $\delta t$ with one extra digit after decimal point ($\delta t_c - 9\times10^{-(p-1)} \geq \delta t \geq \delta t_c + 10\times10^{-(p-1)}$) and calculate total error of collapse for $P(t)$ $\rho(t)$ $V(t)$ and $M(t)$ trajectories assuming that value is the critical threshold $\delta t_c$. If one of the $\delta t_c$ candidates has a significantly higher quality of collapse than the current value, $\delta t_c$ is updated, precision is incremented and the process is repeated. After $\delta t_c$ is determined up to the desired level of precision or after determining that the precision cannot be increased further with the current method, system sizes or the number of realizations, we use all the data produced at every step to plot quantities as a function of $\delta t - \delta t_c$ and as a function of time at $\delta t = \delta t_c$. Please note that the main goal of the current manuscript is not to provide the best estimate of $\delta t_c$ for various synthetic temporal networks, but to validate the hypothesis that limited-time spreading on these networks belong to the same universality class as any directed percolation system. It might be possible to design more efficient numerical methods of estimating $\delta t_c$, but the current method and the estimates presented would suffice for the purposes of this manuscript.
\clearpage
\section{Full suit of empirical results for random \texorpdfstring{Erdős–Rényi}{Erdos–Renyi} based temporal networks}
\begin{table}[htbp!]
\caption{\label{tab:experiment-parameters-random} Experimental setup for Erdős–Rényi networks with $\langle k \rangle = 8$ and random 9-regular networks.}
\begin{ruledtabular}
\begin{tabular}{rrr}
\multicolumn{1}{c}{\textrm{Size $N$}} &
\multicolumn{1}{c}{\textrm{Time window $T$}} &
\multicolumn{1}{c}{\textrm{Realisations}}\\
\colrule
256 & 2048 & 4096 \\
512 & 2048 & 4096 \\
1024 & 2048 & 4096 \\
2048 & 1024 & 4096 \\
4096 & 1024 & 1024 \\
8192 & 1024 & 512 \\
16384 & 512 & 512 \\
32768 & 256 & 512 \\
65536 & 128 & 256 \\
131072 & 64 & 256 \\
\end{tabular}
\end{ruledtabular}
\end{table}
\begin{figure}[htbp!]
\centering
\includegraphics[width=0.65\linewidth]{ms-long-supplementary-figures/params/erdos-poisson-nu_para-nu_perp-all-beta-1.00-beta_prime-1.00-dtc-0.08421.pdf}
\caption{Total error of collapse of $M(t)$, $V(t)$, $\hat{P}(t)$ and $\rho(t)$ universal scaling functions for Erdős–Rényi networks $\langle k \rangle = 8$ and Poisson process activation $\lambda = 1$ with $\nu_\perp$ and $\nu_\parallel$ values assuming (a) $\beta \in \{0.84, 1, 1.16\}$, (b) $\beta' \in \{0.84, 1, 1.16\}$ and (c) $\delta t_c \in \{0.0840,0.08421,0.0844\}$ which shows a minimum around $\beta = \beta' = \nu_\parallel = 1$ and $\nu_\perp=0.5$}
\label{fig:erdos-grid-search-nu-para-nu-perp}
\end{figure}
\begin{figure}[htbp!]
\centering
\includegraphics[width=0.65\linewidth]{erdos-poisson-beta-beta_prime.pdf}
\caption{Total error of collapse of $M(t)$, $V(t)$, $\hat{P}(t)$ and $\rho(t)$ universal scaling functions for Erdős–Rényi networks $\langle k \rangle = 8$ and Poisson process activation $\lambda = 1$ with $\beta$ and $\beta'$ values assuming (a) $\nu_\perp \in \{0.34, 0.5, 0.66\}$, (b) $\nu_\parallel \in \{0.84, 1, 1.16\}$ and (c) $\delta t_c \in \{0.0840,0.08421,0.0844\}$ which shows a minimum around $\beta = \beta' = \nu_\parallel = 1$ and $\nu_\perp=0.5$}
\label{fig:erdos-grid-search-beta-beta-prime}
\end{figure}
\begin{figure}[htbp!]
\centering
\includegraphics[width=\linewidth]{ms-long-supplementary-figures/scaling/random-erdos-poisson-8.pdf}
\caption{Random Erdős–Rényi network $\langle k \rangle = 8$ with Poisson link activation with mean inter-event time of 1.}
\label{fig:erdos-poisson-scaling}
\end{figure}
\begin{figure}[htbp!]
\centering
\includegraphics[width=\linewidth]{ms-long-supplementary-figures/scaling/random-erdos-bursty_2.05-8.pdf}
\caption{Random Erdős–Rényi network $\langle k \rangle = 8$ with bursty link activation (i.e. power-law inter-event time distribution with minimum cutoff) with mean inter-event time of 1 and inter-event time exponents of 2.05.}
\label{fig:erdos-bursty-scaling-2.05}
\end{figure}
\begin{figure}[htbp!]
\centering
\includegraphics[width=\linewidth]{ms-long-supplementary-figures/scaling/random-erdos-bursty_2.2-8.pdf}
\caption{Random Erdős–Rényi network $\langle k \rangle = 8$ with bursty link activation (i.e. power-law inter-event time distribution with minimum cutoff) with mean inter-event time of 1 and inter-event time exponents of 2.20.}
\label{fig:erdos-bursty-scaling-2.2}
\end{figure}
\begin{figure}[htbp!]
\centering
\includegraphics[width=\linewidth]{ms-long-supplementary-figures/scaling/random-erdos-bursty_2.8-8.pdf}
\caption{Random Erdős–Rényi network $\langle k \rangle = 8$ with bursty link activation (i.e. power-law inter-event time distribution with minimum cutoff) with mean inter-event time of 1 and inter-event time exponents of 2.80.}
\label{fig:erdos-bursty-scaling-2.8}
\end{figure}
\begin{figure}[htbp!]
\centering
\includegraphics[width=\linewidth]{ms-long-supplementary-figures/scaling/random-erdos-bursty_5.2-8.pdf}
\caption{Random Erdős–Rényi network $\langle k \rangle = 8$ with bursty link activation (i.e. power-law inter-event time distribution with minimum cutoff) with mean inter-event time of 1 and inter-event time exponents of 5.2.}
\label{fig:erdos-bursty-scaling-5.2}
\end{figure}
\begin{figure}[htbp!]
\centering
\includegraphics[width=\linewidth]{ms-long-supplementary-figures/scaling/random-erdos-self-exciting_0.2_0.8_0.5-8.pdf}
\caption{Random Erdős–Rényi network with self-exciting Hawkes process with parameters $\mu=0.2$, $\alpha=0.8$ and $\theta=0.5$.}
\label{fig:erdos-self-exciting-scaling-0.2-0.8-0.5}
\end{figure}
\begin{figure}[htbp!]
\centering
\includegraphics[width=\linewidth]{ms-long-supplementary-figures/scaling/random-erdos-self-exciting_0.2_0.8_1.0-8.pdf}
\caption{Random Erdős–Rényi network with self-exciting Hawkes process with parameters $\mu=0.2$, $\alpha=0.8$ and $\theta=1.0$.}
\label{fig:erdos-self-exciting-scaling-0.2-0.8-1.0}
\end{figure}
\begin{figure}[htbp!]
\centering
\includegraphics[width=\linewidth]{ms-long-supplementary-figures/scaling/random-erdos-self-exciting_0.5_0.5_0.5-8.pdf}
\caption{Random Erdős–Rényi network with self-exciting Hawkes process with parameters $\mu=0.5$, $\alpha=0.5$ and $\theta=0.5$.}
\label{fig:erdos-self-exciting-scaling-0.5-0.5-0.5}
\end{figure}
\begin{figure}[htbp!]
\centering
\includegraphics[width=\linewidth]{ms-long-supplementary-figures/scaling/random-erdos-self-exciting_0.5_0.5_1.0-8.pdf}
\caption{Random Erdős–Rényi network with self-exciting Hawkes process with parameters $\mu=0.5$, $\alpha=0.5$ and $\theta=1.0$.}
\label{fig:erdos-self-exciting-scaling-0.5-0.5-1.0}
\end{figure}
\begin{figure}[htbp!]
\centering
\includegraphics[width=\linewidth]{ms-long-supplementary-figures/scaling/random-erdos-self-exciting_0.8_0.2_0.5-8.pdf}
\caption{Random Erdős–Rényi network with self-exciting Hawkes process with parameters $\mu=0.8$, $\alpha=0.2$ and $\theta=0.5$.}
\label{fig:erdos-self-exciting-scaling-0.8-0.2-0.5}
\end{figure}
\begin{figure}[htbp!]
\centering
\includegraphics[width=\linewidth]{ms-long-supplementary-figures/scaling/random-erdos-self-exciting_0.8_0.2_1.0-8.pdf}
\caption{Random Erdős–Rényi network with self-exciting Hawkes process with parameters $\mu=0.8$, $\alpha=0.2$ and $\theta=1.0$.}
\label{fig:erdos-self-exciting-scaling-0.8-0.2-1.0}
\end{figure}
\section{Full suit of empirical results for random regular based temporal networks}
\begin{figure}[htbp!]
\centering
\includegraphics[width=0.65\linewidth]{ms-long-supplementary-figures/params/regular-poisson-nu_para-nu_perp-all-beta-1.00-beta_prime-1.00-dtc-0.08808.pdf}
\caption{Total error of collapse of $M(t)$, $V(t)$, $\hat{P}(t)$ and $\rho(t)$ universal scaling functions for random 9-regular networks and Poisson process activation $\lambda = 1$ with $\nu_\perp$ and $\nu_\parallel$ values assuming (a) $\beta \in \{0.84, 1, 1.16\}$, (b) $\beta' \in \{0.84, 1, 1.16\}$ and (c) $\delta t_c \in \{0.0878,0.08808,0.0883\}$ which shows a minimum around $\beta = \beta' = \nu_\parallel = 1$ and $\nu_\perp=0.5$}
\label{fig:regular-grid-search-nu-para-nu-perp}
\end{figure}
\begin{figure}[htbp!]
\centering
\includegraphics[width=0.65\linewidth]{ms-long-supplementary-figures/params/regular-poisson-beta-beta_prime-all-nu_para-1.00-nu_perp-0.50-dtc-0.08808.pdf}
\caption{Total error of collapse of $M(t)$, $V(t)$, $\hat{P}(t)$ and $\rho(t)$ universal scaling functions for random 9-regular networks and Poisson process activation $\lambda = 1$ with $\beta$ and $\beta'$ values assuming (a) $\nu_\perp \in \{0.34, 0.5, 0.66\}$, (b) $\nu_\parallel \in \{0.84, 1, 1.16\}$ and (c) $\delta t_c \in \{0.0878,0.08808,0.0883\}$ which shows a minimum around $\beta = \beta' = \nu_\parallel = 1$ and $\nu_\perp=0.5$}
\label{fig:regular-grid-search-beta-beta-prime}
\end{figure}
\begin{figure}[htbp!]
\centering
\includegraphics[width=\linewidth]{ms-long-supplementary-figures/scaling/random-regular-poisson-9.pdf}
\caption{Random 9-regular network with Poisson link activation with mean inter-event time of 1.}
\label{fig:regular-poisson-scaling}
\end{figure}
\begin{figure*}[htbp!]
\centering
\includegraphics[width=1\linewidth]{ms-long-supplementary-figures/dists/regular/out-comps-0.07.pdf}
\includegraphics[width=1\linewidth]{ms-long-supplementary-figures/dists/regular/out-comps-0.08.pdf}
\includegraphics[width=1\linewidth]{ms-long-supplementary-figures/dists/regular/out-comps-0.08802.pdf}
\includegraphics[width=1\linewidth]{ms-long-supplementary-figures/dists/regular/out-comps-0.092.pdf}
\includegraphics[width=1\linewidth]{ms-long-supplementary-figures/dists/regular/out-comps-0.1.pdf}
\caption{Out-component size estimates of all events of a 9-regular network with Poisson process link activations $\lambda=1$ for (a) $\delta t=0.07$, (b) $\delta t=0.08$, (c) $\delta t=0.08802$, (d) $\delta t=0.092$ and (e) $\delta t=0.1$.}
\label{fig:regular-out-component-sizes}
\end{figure*}
\begin{figure}[htbp!]
\centering
\includegraphics[width=\linewidth]{ms-long-supplementary-figures/scaling/random-regular-bursty_2.05-9.pdf}
\caption{Random 9-regular network with bursty link activation (i.e. power-law inter-event time distribution with minimum cutoff) with mean inter-event time of 1 and inter-event time exponents of 2.05.}
\label{fig:regular-bursty-scaling-2.05}
\end{figure}
\begin{figure}[htbp!]
\centering
\includegraphics[width=\linewidth]{ms-long-supplementary-figures/scaling/random-regular-bursty_2.2-9.pdf}
\caption{Random 9-regular network with bursty link activation (i.e. power-law inter-event time distribution with minimum cutoff) with mean inter-event time of 1 and inter-event time exponents of 2.2.}
\label{fig:regular-bursty-scaling-2.2}
\end{figure}
\begin{figure}[htbp!]
\centering
\includegraphics[width=\linewidth]{ms-long-supplementary-figures/scaling/random-regular-bursty_2.8-9.pdf}
\caption{Random 9-regular network with bursty link activation (i.e. power-law inter-event time distribution with minimum cutoff) with mean inter-event time of 1 and inter-event time exponents of 2.8.}
\label{fig:regular-bursty-scaling-2.8}
\end{figure}
\begin{figure}[htbp!]
\centering
\includegraphics[width=\linewidth]{ms-long-supplementary-figures/scaling/random-regular-bursty_5.2-9.pdf}
\caption{Random 9-regular network with bursty link activation (i.e. power-law inter-event time distribution with minimum cutoff) with mean inter-event time of 1 and inter-event time exponents of 5.2.}
\label{fig:regular-bursty-scaling-5.2}
\end{figure}
\begin{figure}[htbp!]
\centering
\includegraphics[width=\linewidth]{ms-long-supplementary-figures/scaling/random-regular-self-exciting_0.2_0.8_0.5-9.pdf}
\caption{Random 9-regular network with self-exciting Hawkes process with parameters $\mu=0.2$, $\alpha=0.8$ and $\theta=0.50$.}
\label{fig:regular-self-exciting-scaling-0.2-0.8-0.5}
\end{figure}
\begin{figure}[htbp!]
\centering
\includegraphics[width=\linewidth]{ms-long-supplementary-figures/scaling/random-regular-self-exciting_0.2_0.8_1.0-9.pdf}
\caption{Random 9-regular network with self-exciting Hawkes process with parameters $\mu=0.2$, $\alpha=0.8$ and $\theta=1.0$.}
\label{fig:regular-self-exciting-scaling-0.2-0.8-1.0}
\end{figure}
\begin{figure}[htbp!]
\centering
\includegraphics[width=\linewidth]{ms-long-supplementary-figures/scaling/random-regular-self-exciting_0.5_0.5_0.5-9.pdf}
\caption{Random 9-regular network with self-exciting Hawkes process with parameters $\mu=0.5$, $\alpha=0.5$ and $\theta=0.5$.}
\label{fig:regular-self-exciting-scaling-0.5-0.5-0.5}
\end{figure}
\begin{figure}[htbp!]
\centering
\includegraphics[width=\linewidth]{ms-long-supplementary-figures/scaling/random-regular-self-exciting_0.5_0.5_1.0-9.pdf}
\caption{Random 9-regular network with self-exciting Hawkes process with parameters $\mu=0.5$, $\alpha=0.5$ and $\theta=1.0$.}
\label{fig:regular-self-exciting-scaling-0.5-0.5-1.0}
\end{figure}
\begin{figure}[htbp!]
\centering
\includegraphics[width=\linewidth]{ms-long-supplementary-figures/scaling/random-regular-self-exciting_0.8_0.2_0.5-9.pdf}
\caption{Random 9-regular network with self-exciting Hawkes process with parameters $\mu=0.8$, $\alpha=0.2$ and $\theta=0.5$.}
\label{fig:regular-self-exciting-scaling-0.8-0.2-0.5}
\end{figure}
\begin{figure}[htbp!]
\centering
\includegraphics[width=\linewidth]{ms-long-supplementary-figures/scaling/random-regular-self-exciting_0.8_0.2_1.0-9.pdf}
\caption{Random 9-regular network with self-exciting Hawkes process with parameters $\mu=0.8$, $\alpha=0.2$ and $\theta=1.0$.}
\label{fig:regular-self-exciting-scaling-0.8-0.2-1.0}
\end{figure}
\section{Full suit of empirical results for grid lattice based temporal networks}
Grid lattices is generated as a line, square-shaped, cubic and hyper-cubic with periodic boundary condition. Each system has same linear size across all dimensions, as presented in Tabs.~\ref{tab:experiment-parameters-grid-1}, \ref{tab:experiment-parameters-grid-2}, \ref{tab:experiment-parameters-grid-3} and \ref{tab:experiment-parameters-grid-4}.
\begin{table}[htbp!]
\caption{\label{tab:experiment-parameters-grid-1} Experimental setup for 1-dimensional grid.}
\begin{ruledtabular}
\begin{tabular}{rrr}
\multicolumn{1}{c}{\textrm{Size $N$}} &
\multicolumn{1}{c}{\textrm{Time window $T$}} &
\multicolumn{1}{c}{\textrm{Realisations}}\\
\colrule
256 & 8192 & 4096 \\
512 & 8192 & 4096 \\
1024 & 8192 & 4096 \\
2048 & 4096 & 4096 \\
4096 & 4096 & 1024 \\
8192 & 4096 & 512 \\
16384 & 2048 & 512 \\
32768 & 1024 & 512 \\
65536 & 512 & 256 \\
131072 & 256 & 256 \\
\end{tabular}
\end{ruledtabular}
\end{table}
\begin{figure}[htbp!]
\centering
\includegraphics[width=0.65\linewidth]{ms-long-supplementary-figures/params/grid_1-poisson-nu_para-nu_perp-all-beta-0.28-beta_prime-0.28-dtc-0.9919.pdf}
\caption{Total error of collapse of $M(t)$, $V(t)$, $\hat{P}(t)$ and $\rho(t)$ universal scaling functions for 1D grid lattice networks $\langle k \rangle = 8$ and Poisson process activation $\lambda = 1$ with $\nu_\perp$ and $\nu_\parallel$ values assuming (a) $\beta \in \{0.12, 0.28, 0.44\}$, (b) $\beta' \in \{0.12, 0.28, 0.44\}$ and (c) $\delta t_c \in \{0.98,0.9919,1.0\}$ which shows a minimum around $\beta = \beta' = 0.28$, $\nu_\parallel = 1.74$ and $\nu_\perp=1.10$}
\label{fig:grid-1-grid-search-nu-para-nu-perp}
\end{figure}
\begin{figure}[htbp!]
\centering
\includegraphics[width=0.65\linewidth]{ms-long-supplementary-figures/params/grid_1-poisson-beta-beta_prime-all-nu_para-1.74-nu_perp-1.10-dtc-0.9919.pdf}
\caption{Total error of collapse of $M(t)$, $V(t)$, $\hat{P}(t)$ and $\rho(t)$ universal scaling functions for 1D grid lattice networks $\langle k \rangle = 8$ and Poisson process activation $\lambda = 1$ with $\beta$ and $\beta'$ values assuming (a) $\nu_\perp \in \{0.94, 1.10, 1.26\}$, (b) $\nu_\parallel \in \{1.58, 1.74, 1.90\}$ and (c) $\delta t_c \in \{0.98,0.9919,1.0\}$ which shows a minimum around $\beta = \beta' = 0.28$, $\nu_\parallel = 1.74$ and $\nu_\perp=1.10$}
\label{fig:grid-1-grid-search-beta-beta-prime}
\end{figure}
\begin{figure}[htbp!]
\centering
\includegraphics[width=\linewidth]{ms-long-supplementary-figures/scaling/random-grid_1-poisson-2.pdf}
\caption{1-dimensional square grid lattice with periodic boundary condition network (i.e.~a ring of nodes) with Poisson link activation with mean inter-event time of 1.}
\label{fig:grid-1d-poisson-scaling}
\end{figure}
\begin{table}[htbp!]
\caption{\label{tab:experiment-parameters-grid-2} Experimental setup for 2-dimensional square grid.}
\begin{ruledtabular}
\begin{tabular}{rrr}
\multicolumn{1}{c}{\textrm{Size $N$}} &
\multicolumn{1}{c}{\textrm{Time window $T$}} &
\multicolumn{1}{c}{\textrm{Realisations}}\\
\colrule
256 & 4096 & 4096 \\
529 & 4096 & 4096 \\
1024 & 4096 & 4096 \\
2025 & 2048 & 4096 \\
4096 & 2048 & 1024 \\
8281 & 2048 & 512 \\
16384 & 1024 & 512 \\
32761 & 512 & 512 \\
65536 & 256 & 256 \\
131044 & 128 & 256 \\
\end{tabular}
\end{ruledtabular}
\end{table}
\begin{figure}[htbp!]
\centering
\includegraphics[width=0.65\linewidth]{ms-long-supplementary-figures/params/grid_2-poisson-nu_para-nu_perp-all-beta-0.58-beta_prime-0.58-dtc-0.28428.pdf}
\caption{Total error of collapse of $M(t)$, $V(t)$, $\hat{P}(t)$ and $\rho(t)$ universal scaling functions for 2D square grid lattice networks $\langle k \rangle = 8$ and Poisson process activation $\lambda = 1$ with $\nu_\perp$ and $\nu_\parallel$ values assuming (a) $\beta \in \{0.47, 0.58, 0.69\}$, (b) $\beta' \in \{0.47, 0.58, 0.69\}$ and (c) $\delta t_c \in \{0.2835,0.28428,0.2845\}$ which shows a minimum around $\beta = \beta' = 0.58$, $\nu_\parallel = 1.30$ and $\nu_\perp=0.74$}
\label{fig:grid-2-grid-search-nu-para-nu-perp}
\end{figure}
\begin{figure}[htbp!]
\centering
\includegraphics[width=0.65\linewidth]{ms-long-supplementary-figures/params/grid_2-poisson-beta-beta_prime-all-nu_para-1.30-nu_perp-0.74-dtc-0.28428.pdf}
\caption{Total error of collapse of $M(t)$, $V(t)$, $\hat{P}(t)$ and $\rho(t)$ universal scaling functions for 2D square grid lattice networks $\langle k \rangle = 8$ and Poisson process activation $\lambda = 1$ with $\beta$ and $\beta'$ values assuming (a) $\nu_\perp \in \{0.63, 0.74, 0.85\}$, (b) $\nu_\parallel \in \{1.19, 1.30, 1.41\}$ and (c) $\delta t_c \in \{0.2835,0.28428,0.2845\}$ which shows a minimum around $\beta = \beta' = 0.58$, $\nu_\parallel = 1.30$ and $\nu_\perp=0.74$}
\label{fig:grid-2-grid-search-beta-beta-prime}
\end{figure}
\begin{figure}[htbp!]
\centering
\includegraphics[width=\linewidth]{ms-long-supplementary-figures/scaling/random-grid_2-poisson-4.pdf}
\caption{2-dimensional square grid lattice with periodic boundary condition network with Poisson link activation with mean inter-event time of 1.}
\label{fig:grid-2d-poisson-scaling}
\end{figure}
\begin{table}[htbp!]
\caption{\label{tab:experiment-parameters-grid-3} Experimental setup for 3-dimensional nearest-neighbour square grid.}
\begin{ruledtabular}
\begin{tabular}{rrr}
\multicolumn{1}{c}{\textrm{Size $N$}} &
\multicolumn{1}{c}{\textrm{Time window $T$}} &
\multicolumn{1}{c}{\textrm{Realisations}}\\
\colrule
216 & 2048 & 4096 \\
512 & 2048 & 4096 \\
1000 & 2048 & 4096 \\
2197 & 1024 & 4096 \\
4096 & 1024 & 1024 \\
8000 & 1024 & 512 \\
15625 & 512 & 512 \\
32768 & 256 & 512 \\
64000 & 128 & 256 \\
132651 & 64 & 256 \\
\end{tabular}
\end{ruledtabular}
\end{table}
\begin{figure}[htbp!]
\centering
\includegraphics[width=0.65\linewidth]{ms-long-supplementary-figures/params/grid_3-poisson-nu_para-nu_perp-all-beta-0.82-beta_prime-0.82-dtc-0.15375.pdf}
\caption{Total error of collapse of $M(t)$, $V(t)$, $\hat{P}(t)$ and $\rho(t)$ universal scaling functions for 3D cubic grid lattice networks $\langle k \rangle = 8$ and Poisson process activation $\lambda = 1$ with $\nu_\perp$ and $\nu_\parallel$ values assuming (a) $\beta \in \{0.66, 0.82, 0.98\}$, (b) $\beta' \in \{0.66, 0.82, 0.98\}$ and (c) $\delta t_c \in \{0.1535,0.15375,0.1544\}$ which shows a minimum around $\beta = \beta' = 0.58$, $\nu_\parallel = 1.30$ and $\nu_\perp=0.74$}
\label{fig:grid-3-grid-search-nu-para-nu-perp}
\end{figure}
\begin{figure}[htbp!]
\centering
\includegraphics[width=0.65\linewidth]{ms-long-supplementary-figures/params/grid_3-poisson-beta-beta_prime-all-nu_para-1.10-nu_perp-0.58-dtc-0.15375.pdf}
\caption{Total error of collapse of $M(t)$, $V(t)$, $\hat{P}(t)$ and $\rho(t)$ universal scaling functions for 3D cubic grid lattice networks $\langle k \rangle = 8$ and Poisson process activation $\lambda = 1$ with $\beta$ and $\beta'$ values assuming (a) $\nu_\perp \in \{0.42, 0.58, 0.74\}$, (b) $\nu_\parallel \in \{0.94, 1.10, 1.26\}$ and (c) $\delta t_c \in \{0.1535,0.15375,0.1544\}$ which shows a minimum around $\beta = \beta' = 0.58$, $\nu_\parallel = 1.30$ and $\nu_\perp=0.74$}
\label{fig:grid-3-grid-search-beta-beta-prime}
\end{figure}
\begin{figure}[htbp!]
\centering
\includegraphics[width=\linewidth]{ms-long-supplementary-figures/scaling/random-grid_3-poisson-6.pdf}
\caption{3-dimensional square grid lattice with periodic boundary condition network with Poisson link activation with mean inter-event time of 1.}
\label{fig:grid-3d-poisson-scaling}
\end{figure}
\begin{table}[htbp!]
\caption{\label{tab:experiment-parameters-grid-4} Experimental setup for 4-dimensional nearest-neighbour square grid.}
\begin{ruledtabular}
\begin{tabular}{rrr}
\multicolumn{1}{c}{\textrm{Size $N$}} &
\multicolumn{1}{c}{\textrm{Time window $T$}} &
\multicolumn{1}{c}{\textrm{Realisations}}\\
\colrule
256 & 2048 & 4096 \\
625 & 2048 & 4096 \\
1296 & 2048 & 4096 \\
2401 & 1024 & 4096 \\
4096 & 1024 & 1024 \\
6561 & 1024 & 512 \\
14641 & 512 & 512 \\
28561 & 256 & 512 \\
65536 & 128 & 256 \\
130321 & 64 & 256 \\
\end{tabular}
\end{ruledtabular}
\end{table}
\begin{figure}[htbp!]
\centering
\includegraphics[width=0.65\linewidth]{ms-long-supplementary-figures/params/grid_4-poisson-nu_para-nu_perp-all-beta-1.00-beta_prime-1.00-dtc-0.1045.pdf}
\caption{Total error of collapse of $M(t)$, $V(t)$, $\hat{P}(t)$ and $\rho(t)$ universal scaling functions for 4D cubic grid lattice networks $\langle k \rangle = 8$ and Poisson process activation $\lambda = 1$ with $\nu_\perp$ and $\nu_\parallel$ values assuming (a) $\beta \in \{0.84, 1.00, 1.16\}$, (b) $\beta' \in \{0.84, 1.00, 1.16\}$ and (c) $\delta t_c \in \{0.103,0.1045,0.105\}$ which shows a minimum around $\beta = \beta' = \nu_\parallel = 1.00$ and $\nu_\perp=0.50$}
\label{fig:grid-4-grid-search-nu-para-nu-perp}
\end{figure}
\begin{figure}[htbp!]
\centering
\includegraphics[width=0.65\linewidth]{ms-long-supplementary-figures/params/grid_4-poisson-beta-beta_prime-all-nu_para-1.00-nu_perp-0.50-dtc-0.1045.pdf}
\caption{Total error of collapse of $M(t)$, $V(t)$, $\hat{P}(t)$ and $\rho(t)$ universal scaling functions for 4D cubic grid lattice networks $\langle k \rangle = 8$ and Poisson process activation $\lambda = 1$ with $\beta$ and $\beta'$ values assuming (a) $\nu_\perp \in \{0.84, 1.00, 1.16\}$, (b) $\nu_\parallel \in \{0.84, 1.00, 1.16\}$ and (c) $\delta t_c \in \{0.1535,0.15375,0.1544\}$ which shows a minimum around $\beta = \beta' = \nu_\parallel = 1.00$ and $\nu_\perp=0.50$}
\label{fig:grid-4-grid-search-beta-beta-prime}
\end{figure}
\begin{figure}[htbp!]
\centering
\includegraphics[width=\linewidth]{ms-long-supplementary-figures/scaling/random-grid_4-poisson-8.pdf}
\caption{4-dimensional square grid lattice with periodic boundary condition network with Poisson link activation with mean inter-event time of 1.}
\label{fig:grid-4d-poisson-scaling}
\end{figure}
\section{Introduction}\label{sec:introduction}
Connectivity is an essential characteristic of complex networks as it determines how far information or influence can spread in a network structure. Consequently, it governs the emergence and scale of any macroscopic phenomena often modelled on networks such as disease spreading, transportation, or information diffusion, to mention a few examples. Percolation theory provides a comprehensive understanding that characterizes network connectivity with various mathematical and algorithmic tools primarily developed for complex networks. For example, percolation can be mapped to late-stage results of specific epidemic processes \cite{newman2002spread, kenah2007second, kenah2011epidemic, pastor2015epidemic, rizi2021epidemic}, such that the size of percolating components determine the final size of the epidemic. Meanwhile, the percolation transition and its related critical behavior explain the disease outcome close to the epidemic threshold.
However, these theoretical descriptions commonly assume that the network is static, with links and nodes always present, ignoring the typical character of several complex structures where links may vary in time. Since information, disease or other effects can pass between two nodes in a network only at the time of their interactions, the temporal alternation of links may crucially influence the critical behavior and final outcome of any ongoing spreading processes \cite{liu2013contagion, karimi2013threshold, backlund2014effects, delvenne2015diffusion, bazzi2016community, aslak2018constrained, koher2019spreading}. To characterize these processes, one needs to measure connectivity in temporal networks across time, where components are defined in terms of network nodes and links and the temporal distribution of interactions. Consequently, beyond the well-studied structural heterogeneities of static networks, like in their node degrees, the effects of temporal correlations leading to temporal heterogeneities in the interaction dynamics, like burstiness, become important \cite{karsai2011small, rocha2011simulated, horvath2014spreading, ubaldi2017burstiness, hiraoka2018correlated, karsai2018bursty}. This is especially the case for so-called limited-waiting-time processes, where an effect or information, e.g.~a disease or a meme \cite{kuhn2014inheritance}, arriving at a node can pass over to another node only if an interaction appears within a time window $\delta t$. Otherwise, the pathogen times out, e.g., the patient recovers or the meme becomes irrelevant, making it impossible to reach other nodes.
Similar to static networks, the connectivity of temporal networks passes through a phase-transition. However, close to this critical threshold, temporal networks exhibit different critical behavior as compared to static structures~\cite{kivela2018mapping,parshani2010dynamic,shortpaper}.
For limited-waiting-time connectivity, where the control parameter is $\delta t$, this phase-transition can be theoretically understood under some simplifying assumptions about the homogeneous dynamics of connectivity \cite{shortpaper}. Since there is an embedded direction (or flow) of time, the microscopic dynamics can be fundamentally irreversible with a broken detailed balance and non-equilibrium steady-state. These results suggest that the dynamics of percolation on temporal networks are generically the same as any other system belonging to the Directed Percolation (DP) universality class, which is characterized by a one-component order parameter without additional symmetries and unconventional features such as quenched disorder \cite{henkel2008non}.
The homogeneity approximations used for the derivations presented in \cite{shortpaper}, however, become less grounded when the underlying structure deviates from a random graph or if the interaction dynamics become inhomogeneous. In this paper, our goal is to build on the theory laid down in \cite{shortpaper} to investigate further the relation between temporal networks and directed percolation. In other words, the primary objective of this manuscript is as follows: to show empirically that diverse classes of temporal networks, with various degrees of temporal and spatial heterogeneity, combined with the very general notion of limited-waiting-time reachability, will show an absorbing phase transition in connectivity that belongs to the directed percolation universality class.
In its epidemic interpretation, directed percolation can be one of the most basic non-equilibrium second-order phase transitions from fluctuating states into so-called absorbing states, which exhibit universal features, determined by symmetry properties and conservation laws. We demonstrate the precision of this mapping using extensive numerical simulations and provide further theoretical calculations to study synthetic temporal networks as directed percolation processes with a range of temporal and spatial inhomogeneities.
The remainder of Sec.~\ref{sec:introduction} will be dedicated to laying the groundwork and presenting the context in which this manuscript is set: in Sec.~\ref{sec:event-graph} we will discuss connectivity on temporal networks, the event graph representation and modeling spreading processes and Sec~\ref{sec:directed-percolation} will introduce directed percolation and its characteristics.
The next section, Sec.~\ref{sec:methods}, is dedicated to an overview of our contributions. Section \ref{sec:dp-analogous-in-temporal-networks} will describe our mapping of concepts of directed percolation and temporal networks. Section \ref{sec:mean-field-solution-dp-in-temporal-network} provides an overview of the theoretical results from Ref.~\cite{shortpaper}, which will be extended further in Sec.~\ref{sec:random-regular-solutions} by explicitly deriving some critical exponents and scaling relations. Section \ref{sec:estimation-tricks} will lay down the algorithmic techniques that make large-scale simulations of spreading processes on temporal networks possible.
Finally, in Sec.~\ref{sec:results} we will describe the experimental setup and provide numerical evidence for validity of our hypothesis by application of the methods described previously, while Sec.~\ref{sec:discussion} provides an overview of the implications of the results and the limitations of our study.
\subsection{Temporal networks and the event graph}\label{sec:event-graph}
A temporal network $G = (\mathcal{V}, \mathcal{E})$ provides representations of a dynamically changing complex system as a set of timed interactions known as events $\mathcal{E}$ between a set of entities $\mathcal{V} = \{v_1, v_2,\ldots, v_n\}$ known as nodes or vertices during an observation period $\mathcal{T}$. Each event indicates a time-dependent interaction between two nodes, e.g.~physical contact or communication between two people or trade between two commercial entities \cite{holme2019temporal}, i.e.~$e=(u, v, t_\text{start}, t_\text{end})$ such that $u, v \in \mathcal{V}$ between times of $t_\text{start}, t_\text{end} \in \mathcal{T}$ ($t_\text{start} < t_\text{end}$). Note that this definition can be easily extended to directed events and to directed or undirected temporal hypergraphs.
Two events $e, e' \in \mathcal{E}$ are \emph{adjacent} if they share at least one endpoint node in common, $\{u, v\} \cap \{u', v'\} \neq \emptyset$, and they follow each other in time such that $ \Delta t(e, e') = t'_\text{start} - t_\text{end} > 0$. Therefore, any temporal network can be represented as a higher-order static directed acyclic weighted graph known as the \textit{event graph} $D = (\mathcal{E}, E_D, \Delta t(e, e'))$~\cite{kivela2018mapping,Mellor}. Nodes of the event graph are the events of the original temporal network and the weight of a link between two connected nodes (adjacent events) is then defined as the time difference $\Delta t$ between the corresponding events.
Every path on the event graph constitutes a \emph{causal chain} as, by definition, a path constitutes a list of events where every two consecutive events are adjacent. Paths in event graphs are, therefore, equivalent to time-respecting paths in the corresponding temporal network representation \cite{saramaki2019weighted}. Therefore, calculating time-respecting reachability on a temporal network is equivalent to connectivity on its corresponding (static) event graph representation. The weakly connected components on an event graph determine \emph{causal domains}, disjoint sets of events where there can be no causal connections whatsoever between events if they belong to two different weakly connected components. In addition, as compared to reachability, the size and distribution of weakly connected components are quantities, which are much easier to measure for temporal networks and they characterize a percolation transition if we assume an undirected network. Moreover, the sizes of these components put an upper bound on how much an effect can spread starting from one of the events in that component \cite{kivela2018mapping}.
Temporal networks preserve the dynamic properties of the represented complex system, unlike aggregated static networks where this information is lost.
Through the studies of time-varying networks, several new phenomena in human dynamics have been explored over the last decades, such as node and link burstiness~\cite{karsai2018bursty,jo2020burst,urena2020estimating}, causal, temporal motifs \cite{kovanen2011temporal}, or the cyclic activation patterns of human interaction activities \cite{aledavood2018social}, to mention a few.
As opposed to systems governed by homogeneous and independent processes, these correlations and the induced temporal dynamics may have significant effects on various dynamical processes evolving on temporal networks such as spreading \cite{takaguchi2013bursty, sajjadi2020impact}, reachability \cite{holme2005network, lee2019concurrency}, diffusion \cite{moody2016interdependent, delvenne2015diffusion}, and opinion formation \cite{garimella2018polarization}.
The different dynamics of a temporal network are often straightforward to study through simulations. For example, in the case of spreading processes, transmission can be modeled by temporal network events \cite{karsai2011small, rocha2011simulated, horvath2014spreading, ubaldi2017burstiness, hiraoka2018correlated}. More concretely, in a physical interaction network, where nodes represent people and events represent two people coming to close proximity,
each of these contact events will have a probability of transmitting the disease. The disease then spreads to all the nodes that can be reached via such infecting events from the initially infected nodes. Similarly, in a network where events represent communication of information at a specific time, such as mobile phone calls or email exchanges, it is straightforward to model the spreading of information by keeping track of the information nodes have access to at each point in time.
Many dynamics evolving on top of networks, such as some spreading processes \cite{lambiotte2016guide, holme2012temporal, holme2015modern}, social contagion \cite{daley1964epidemics, castellano2009statistical,unicomb2021dynamics} ad-hoc message passing by mobile agents \cite{tripp2016special} or routing processes \cite{nassir2016utility}, can have a limited memory thus can only use paths constrained by limited waiting times. Limited waiting-time reachability can be modeled using the event graph, $D$, that contains a superposition of all temporal paths~\cite{kivela2018mapping, saramaki2019weighted, badiemodiri2020efficient}. In a limited waiting-time spreading process unfolding over a temporal network, either the spreading agent (e.g.~the pathogen in the disease spreading) must be transmitted onward from a node within some time $\delta t$ or the infection has to be renewed before that time. In other words, the node must participate in a possibly disease-carrying event in $\delta t$ time, or the process stops and the node reverts to susceptible. Therefore, all the spreading paths in the network are $\delta t$-constrained time-respecting paths. Let's call two adjacent events $e$ and $e'$ as $\delta t$-adjacent if $\Delta t(e, e') \leq \delta t$. A subset of the event graph $D$ with an upper threshold of weights no greater than $\delta t$, i.e., where directed links indicate $\delta t$-adjacency, enables us to calculate reachability for $\delta t$ limited-time spreading process for the corresponding temporal network. Therefore, the event graph encapsulates a complete set of $\delta t$-constrained time-respecting paths for all values of $\delta t$ simultaneously.
\subsection{Directed Percolation}\label{sec:directed-percolation}
The waiting-time limit $\delta t$ can be regarded as the control parameter of a continuous phase-transition, where connectivity in the event graph is determined by $\delta t$-connected paths of events. As the value of maximum waiting time decreases, more and more of the links of the event graph get removed, where each deleted link corresponding to an adjacency relationship between two events that are temporally more than $\delta t$ apart. This leads to a drop in connectivity in the event graph, which is exactly equivalent to the drop in connectivity on the temporal network. In order to characterize these phase transitions, unlike characterizing the superficially similar phase transitions that take place when removing links in static (undirected) networks, we need to consider a percolation framework that can explicitly model the one-way flow of time.
Directed percolation is a paradigmatic example of dynamical phase transitions into absorbing states with a well-defined set of universal critical exponents and is often used to model phenomena with inherent directionality, such as fluids passing through porous media \cite{hinrichsen2000non, odor2004universality, hinrichsen2006non, henkel2008non}. Originally introduced as a model for directed random connectivity \cite{broadbent1957percolation}, directed percolation attracted scrutiny in percolation theory in the late seventies \cite{blease1977directed}. Since then, a considerable body of work has been devoted to this approach of interpretation in the literature since the critical behavior of many stochastic many-particle non-equilibrium processes can be shown to belong to the directed percolation universality class. Directed percolation has applications in various domains at multiple scales ranging from galaxies to semiconductors \cite{schlogl1972chemical,gerola1980theory,van1981hopping,bak1993punctuated}.
As the simplest model exhibiting a transition between active and absorbing phases \cite{hinrichsen2000non}, it is straightforward to define and implement models governed by directed percolation, e.g. in the case of lattice models \cite{domany1984equivalence, kinzel1985phase, harris1974contact, jensen1993critical, mendes1994generalized, ziff1986kinetic, dhar1987collapse}. Directed percolation, however, does not appear to be an integrable model and its critical behavior is highly non-trivial. Moreover, it seems that the basic features of directed percolation, such as non-fluctuating states, are quite difficult to realize in nature \cite{hinrichsen2000possible}. Another fundamental problem is quenched disorder due to microscopic inhomogeneities of the system \cite{henkel2008non}. One of the earliest unambiguous and robust experimental realizations of a system exhibiting critical behavior in the directed percolation class was for the rather specific case of liquid crystal electrohydrodynamic convection \cite{takeuchi2007directed}. Another experimental evidence was reported in 2016 in the case of transition to turbulence \cite{lemoult2016directed}. Due to the simplicity and robustness of directed percolation, it seems to be a good model for explaining ubiquitous phase-transitions in many real-world phenomena, especially in the so-called contact processes \cite{barrat2008dynamical, pastor2015epidemic, kempe2002connectivity, moody2002importance, holme2005network, pan2011path, scholtes2014causality, stehle2011high, dai2020temporal, aleta2020data} in the realm of temporal networks \cite{holme2019temporal}.
Before presenting the mapping between reachability in temporal networks and the concepts in directed percolation, for the remainder of this section we will review these concepts for the case of the simple infinite lattice. Let us take the example of a spreading process across time in an infinitely large $d$-dimensional square lattice: assume that each infected (or occupied) node can infect any of its neighbors independently with probability $p$ at each tick of a discrete timer. Let us also assume that an infected node recovers (becomes unoccupied) in one tick of the clock after infection unless it is re-infected by a neighbor. This configuration is denoted in many sources as a $d+1$-dimensional lattice, substituting the temporal axis with another discrete spatial dimension with the only difference that, unlike the other $d$ dimensions, this one has an inherent directionality. Throughout the rest of this section, we will continue to use the space and time analogy to facilitate a better transition to modeling phenomena on temporal networks.
The dynamics of this spreading process is defined by the topology and dimensionality of the medium of percolation and competition between two processes: the probability that an infected node infects each of its neighbors in a single tick of the clock, or ``reproduction'' from the perspective of the spreading agent, and the time it takes for each infected node to recover, or ``self-annihilation'' or ``death'' of the spreading agent. In the many classic representations of directed percolation, the reproduction probability is often denoted by the parameter $p$ and the ``self-annihilation'' is set to happen in exactly one tick of the clock. For large enough values of $p$, the system will forever stay in an ``active state'' where there is a non-vanishing density of nodes infected (occupied) at all times. Conversely, if the annihilation process has the upper hand, the system eventually transitions irreversibly into an ``absorbing phase'' where no occupied nodes are left in the lattice and the spreading agent is extinct.
More generally, let us say the reproduction and self-annihilation process respectively happen at rates $\mu_p$ and $\mu_r$. Let us assume that at $t=0$, nodes are uniformly occupied with density $\rho_0$. To write a mean-field rate equation for occupation density $\rho(t)$, we need to take into account how often more than one spreading agent (pathogens) simultaneously occupies (infects) the same node, in which case only one new node is occupied. Let us only consider the rate $\mu_c$ at which two other nodes simultaneously infect a single node and assume the probabilities of three or more simultaneous infections are small. In this case, the rate equation is of the form
\begin{equation}\label{eq:general-mean-field-rate-equation}
\pderivative{t} \rho(t) = \tau \rho(t) - g \rho(t)^2\,,
\end{equation}
where the \emph{control parameter} $\tau = \mu_p - \mu_r$ is the manifestation of the competition between reproduction and death as described above and coupling constant $g = \mu_c$ describes the events of infecting a node already infected by another neighbour \cite{henkel2008non}. This equation has a steady-state at $\lim_{t\rightarrow\infty} \rho(t) = \rho_\text{stat}(\tau) = 0$ which corresponds to the aforementioned absorbing phase. Furthermore for $\tau > 0$ the value of $\rho(t)$ approaches a stationary occupation density of $\lim_{t\rightarrow\infty} \rho(t) = \rho_\text{stat}(\tau) = \tau/g$, which is identified as the order parameter of the directed percolation process. At exactly $\tau = 0$, occupation density decays algebraically with time $\rho(t) \sim (\rho_0^{-1} + gt)^{-1}$. Naturally, for values of $\tau < 0$ the system eventually arrives at the absorbing phase $\rho(t) \rightarrow 0$ in finite time.
More generally, starting from a homogeneously occupied initial condition, order parameter $\rho_\text{stat}(\tau)$ of a system in the directed percolation universality scales as $\rho_\text{stat}(\tau) \sim \tau^\beta$, when control parameter $\tau$ is close to $\tau_c = 0$. For $\tau > 0$, density decays algebraically as $\rho(t) \sim t^{-\alpha}$ where in the mean-field regime (i.e.~$d\geq4$), $\beta = \alpha = 1$. In the case of a spreading process controlled by a percolation probability $p$ introduced at the beginning of this section, it can be shown that $\tau \propto p - p_c$ where critical percolation probability $p_c$ is a function of topology and dimensionality of the percolation medium \cite{saberi2015recent, li2021percolation}.
Alternatively, we can focus on the ramifications of starting from a single seed of infection, as opposed to a homogeneous initial distribution of occupied nodes. A characteristic property of this scenario is survival probability $P(t)$: the probability that a spreading process starting from a single seed would still be in the active phase ($\rho(t) > 0$) at time $t$. Similar to occupation density $\rho(t)$, at criticality $\tau = \tau_c = 0$ survival probability also decays algebraically with time $P(t) \sim t^{-\delta}$. A second alternative for order parameter is the ultimate probability of survival $P_\text{surv}(\tau) = \lim_{t \rightarrow \infty} P(t)$. When the control parameter is close to the critical threshold $\tau \rightarrow 0^-$, the ultimate probability of survival scales algebraically as $P_\text{surv}(\tau) \sim \tau^{\beta'}$.
Continuous phase transitions in models with time-like dimensions generally have the same system of two separate order parameters, controlled by two different critical exponents $\beta$ and $\beta'$. For the case of directed percolation, however, ``rapidity-reversal symmetry'', an invariance property under time-reversal, ensures the two exponents have the same value $\beta = \beta'$ \cite{grassberger1979reggeon} which implies that $P(t)$ and $\rho(t)$ are at least asymptotically proportional as $t \rightarrow \infty$, and in some cases exactly equal $P(t) = \rho(t)$ \cite{henkel2008non}. Rapidity-reversal symmetry limits the number of independent critical exponents to three \cite{1982ZPhyB..47..365G, 1981ZPhyB..42..151J}.
\subsubsection{Characteristic quantities of the directed percolation}
The single-source initial condition also allows us to define additional interesting characteristic quantities in the absorbing phase, which might lend themselves to experimental observation. Let's define \emph{pair-connectedness function} $c(\Vec{r_1}, t_1, \Vec{r_2}, t_2)$ as the probability that a path exists from a node with spatial coordinates $\Vec{r_1}$ at time $t_1$ and another in $\Vec{r_2}$ at time $t_2$. Note that the definition of spatial coordinates for nodes as a $d$-dimensional vector $\Vec{r_i}$ implies that the percolation medium and node $i$ is embedded in a $d$-dimensional space, e.g.~a $d$-dimensional lattice. Assuming that the percolation medium is invariant with respect to translations across time and space, we can simplify the pair-connectedness function by fixing the origin on the source node and denote the pair-connectedness function as $c(\Vec{r}, t)$. \emph{Mean cluster mass} $M$ is defined as the integration of the pair-connectedness function across time and space:
\begin{equation}\label{eq:dp-mean-component-mass}
M = \int_0^\infty \dd{t} \int \dd{\Vec{r}} c(\Vec{r}, t) \,,
\end{equation}
which, with control parameter close to the critical threshold $\tau_c = 0$ scales like $M \sim (-\tau)^{-\gamma}$ where $\gamma = \nu_\parallel + d \nu_\perp - \beta - \beta'$. Similarly, \emph{mean spatial volume} $V$ can be defined as the number of unique nodes that will ever get infected in a single-source spreading scenario. As with the case of the cluster mass $M$, spatial volume scales through a power relationship $V \sim (-\tau)^{-\upsilon}$ close to the critical threshold where $\upsilon = d \nu_\perp - \beta'$. It is possible to think of spatial volume $V$ as the size of the projection of the percolation cluster over the $d$-dimensional spatial plane, i.e., over the original $d$-dimensional lattice. Projection of the same cluster on the temporal dimension will define the survival time of the cluster, which is distributed according to the probability of survival $P(t)$.
The homogeneous, fully-occupied initial condition, on the other hand, allows us to study the response of a system to an \emph{external field} $h$ on the order parameter static density $\rho_\text{stat}$. For the case of directed percolation, an external field can be implemented as the spontaneous occupation of nodes at a rate $h$. A positive external field deprives the system of the possibility of ever transitioning into an absorbing phase. \emph{Susceptibility} $\chi$ is defined as the magnitude of the response generated by a minuscule disturbance in the external field
\begin{equation}\label{eq:susceptibility}
\chi(\tau, h)=\frac{\partial}{\partial h} \rho_{stat}(\tau, h)\,,
\end{equation}
which diverges algebraically as the control parameter $\tau$ converges to the critical threshold $\tau_c = 0$, $\chi \sim \abs{\tau}^{-\gamma}$ where $\gamma$ is the same exponent as the mean cluster mass $M$. For the rest of this paper, when not specified, susceptibility $\chi$ is studied at minuscule values of external field ($h=0$) as $\tau$ converges to the critical threshold $\tau_c = 0$. In practical terms, susceptibility is a useful tool for finding the transition point, as unlike the order parameters, we do not need to define an arbitrary threshold for what constitutes a small or large value for a quantity such as $M(t)$ or $V(t)$ close to the transition point in a finite system. Instead, the susceptibility will typically show a peak even in finite systems, which are discussed in more detail in Sec.~\ref{sec:finite-size-scaling}.
\subsubsection{Finite size scaling properties of the system}\label{sec:finite-size-scaling}
While the dynamics described previously explain the behavior of an infinitely large system, measuring properties of infinitely large systems is a rather involved task. Verifying that the behavior of a system at criticality is explained by a specific set of critical exponents is often easier performed by studying the finite-size scaling properties of the system. This can be carried out by measuring a set of quantities for realizations at different scales and plotting the universal scaling function of each quantity as a function of scale-invariant ratios. If the exponents used are correct, all the scaling functions of different linear system sizes for the same quantity should collapse on top of each other.
The effect of the finite size of the system manifest themselves as deviations from the scaling laws as described before and their effects are measurable after some characteristic size-dependent amount of time has elapsed since the beginning of the simulation. For example, while in an infinitely large system in active phase $\tau > 0$ the system will forever stay in an active phase, a finite system will always have a non-vanishing probability of transitioning to the absorbing state due to fluctuation of the order parameter. These finite-size effects take place at a characteristic time $t_f$ that scales as $t_f \sim l^{z}$ where $z = \nu_\parallel / \nu_\perp$ is the so-called \emph{dynamical exponent} and $l$ is the lateral (or linear) size of the system as opposed to system size $N$ measured in number of nodes, where $N \propto l^d$.
In phenomenological scaling theory \emph{simple scaling} is assumed for absorbing phase transitions. This means that large-scale properties of the system are invariant under scale transformations with the control parameter close to the critical threshold. A multiplicative transformation, or ``concentration'', of the control parameter $\tau$ by a factor of $\lambda$, $\tau \mapsto \lambda \tau$ would result in re-scaling of other quantities as
\begin{equation}
\begin{aligned}
&t \mapsto \lambda^{-\nu_\parallel} t \; &&l \mapsto \lambda^{-\nu_\perp} l\\
&\rho \mapsto \lambda^\beta \rho \; &&P \mapsto \lambda^{\beta'} P \\
&h \mapsto \lambda^\sigma h \; &&\chi \mapsto \lambda^{-\gamma} \chi\,,
\end{aligned}
\end{equation}
where $t$ and $l$ stand for time-like and length-like quantities respectively.
More specifically, scale-invariance mandates very specifically how a quantity will change under multiplicative scale change. As an example, let's study changes of $\rho(t, l)$
\begin{equation}
\rho = f(t, l) \mapsto \lambda^\beta \rho = f(\lambda^{-\nu_\parallel} t, \lambda^{-\nu_\perp} l)\,,
\end{equation}
where $t$ is time from initial infection seed and $l$ is the linear system size.
Since this relationship is valid for all values of $\lambda$, we can remove one parameter of the function by selecting a special value $\lambda = l^{1/{\nu_\perp}}$
\begin{equation}\label{eq:scaling-for-rho}
l^{\beta/\nu_\perp} \rho = f(l^{-\nu_\parallel/\nu_\perp} t, 1) = F(l^{-\nu_\parallel/\nu_\perp} t)\,,
\end{equation}
where the function $F(x)$ is referred to as the ``(universal) scaling function'' of its corresponding quantity, in this case, density $\rho$. The parameter to this function $l^{-\nu_\parallel/\nu_\perp} t$ is in itself invariant to scale transformations. This parameter and those similarly derived for other quantities are oftentimes known as ``scale-invariant ratios''. The function $F(x)$ is universal, meaning that if measured to sufficient accuracy, we obtain exactly the same type of scaling function for systems with similar boundary conditions and shape for any phenomena in the directed percolation universality class \cite{henkel2008non}.
The value of each exponent is only a function of a few large-scale properties of the system, such as the number of spatial dimensions of the system. There exists an upper critical dimension $d_c$ where systems with spacial dimensionality $d \geq d_c$ all follow the same set of values for critical exponents, which are exactly equal to those derived through mean-field estimation. For the case of the directed percolation universality class the upper critical dimension has a value of $d_c = 4$ \cite{henkel2008non}.
\section{Methods}\label{sec:methods}
\subsection{Directed percolation in temporal networks}\label{sec:dp-analogous-in-temporal-networks}
Let us now take the case of $\delta t$ limited-time spreading from a single source on a temporal network. Similar to the classic directed percolation single-source spreading process, each temporal network node can participate in the spreading process by becoming infected, infecting others and recovering multiple times. Temporal networks are different from the archetypal directed percolation systems presented in Sec.~\ref{sec:directed-percolation} in that they do not present a regular lattice or metric space in the spatial dimension. Furthermore, there is typically no discrete structure in the temporal dimension, which is usually modeled as a continuous axis. Nevertheless, if the various concepts such as order parameter, control parameter and cluster sizes are defined carefully, temporal networks and limited waiting-time connectivity can be mapped to directed percolation \cite{shortpaper}.
To put it in the same reference frame as with other absorbing phase transitions, changing the parameter $\delta t$, in this scenario, controls the relative occurrence of ``annihilation'' and ``multiplication'' processes. A small enough value of $\delta t$, will lead to a situation where spreading scenarios will eventually die out, at which point the system enters an absorbing phase. Similarly, as $\delta t$ grows, a spreading agent will be able to avoid extinction for longer time, until after some threshold $\delta t > \delta t_c$ a random spreading scenario will not die out (in an infinitely large network). As discussed in Section~\ref{sec:event-graph}, such spreading scenarios are closely related to various properties of the $\delta t$ thresholded limited waiting-time event graph $D$.
\begin{figure*}[ht]
\centering
\includegraphics[width=\linewidth]{dag_schematic.pdf}
\caption{Two spreading scenarios starting from random events (marked with black circles on (b) and (c)) represented over (a) temporal network (b) $\delta t$-limited event graph and (c) reduced event graph of a temporal network built from a one-dimensional grid of 40 nodes (displayed on the left side) with Poisson activation of events with mean inter-event time 1 unit of time, simulated for 20 units of time. The adjacency relations have a maximum waiting time $\delta t = 0.8$ unit of time. Spatial volume $V$ can be visualized as the mean size of projection of a spreading cluster on the spatial plain, i.e.~the static base network on the vertical axis, whereas survival time $T$ is equivalent to the mean size of projection of the cluster on time (horizontal) axis. While measuring a direct analogue to component mass $M$, integrating the pair-connectedness function across time and space equivalent to the mean sum of lengths of colored horizontal lines in (a) is not straightforward with the event graph representation. It is possible to show that the mean number of uniquely counted events involved in the spreading process, corresponding to the cardinality of the out-component of the initial event here represented by the total number of colored nodes in the event graph, show the same scaling behavior. (d) Homogeneous, fully-occupied initial condition with the occupied events shown in a darker shade than unoccupied events shows the decline and eventual stabilization of the occupation density as time grows. In this scenario, all nodes are considered occupied for time $-\infty < t < 0$, which translates to the occupation of all events in period $0 \leq t < \delta t$ and all events in their out-components.}
\label{fig:dag-schematic}
\end{figure*}
As illustrated in Fig.\ref{fig:dag-schematic}, the projection of the spreading cluster over the spatial plane amounts to a subset of temporal network nodes $\mathcal{V}$ that has ever participated in the spreading process. This can be measured by calculating the mean number of unique temporal network nodes involved in the out-components of the event graph. The (ensemble) average number of unique nodes participating in single random source spreading processes is analogous to mean spatial volume $V$. The projection of the spreading cluster over the temporal axis is equal to the time window from the beginning of the spreading process to its end. The ensemble average of this time duration is analogous to mean survival time $T$. The sum of the duration of infectiousness for all the nodes, i.e.~the integration of the pair-connectedness function, would therefore be analogous to spatial and temporal integration of the pair-connectedness function or mean component mass $M$. Note that the duration of the infectiousness is equal in all of the events, therefore, we use the number of reachable (i.e.~possibly infection-carrying) events as a proxy for $M$, ignoring the overlaps. The above-defined quantities can be measured as features of the event graph. The average number of events in the out-component of a node in the event graph (equivalent to an event in the temporal network) measures the number of reachable events. The survival probability $P(t)$ can be similarly defined over an ensemble of single-source spreading instances based on the distribution of the lifetime of each spreading scenario, accounting for the finite temporal window of the simulation of the temporal network using a Kaplan--Meier estimator \cite{kaplan1958nonparametric}.
Another scenario is the simulation of the spreading process from homogeneous, fully occupied, initial conditions. Translating this from classic directed percolation poses a new problem; a homogeneous initial condition cannot translate to a ``full row'' of occupied nodes since we are dealing with continuous-time as opposed to the typical directed percolation case of discretized time presented in Sec.~\ref{sec:directed-percolation}. Rather, a better translation of the fully-occupied initial condition to continuous time is to assume all nodes to be occupied at the beginning of the observation period $t = min(\mathcal{T})$, or more accurately by assuming all nodes to be occupied for all values of $t$ where $t \leq min(\mathcal{T})$. Occupation density $\rho(t)$ is defined as the fraction of infected nodes at time $t$. Stationary density $\rho_\text{stat}(\tau)$ is therefore defined as occupation density after the system had enough time to reach a stationary state. We can also emulate the effects of an external field $h$ in this scenario: In continuous-time, this is equivalent to each node spontaneously becoming occupied through an independent Poisson point process with a rate of $h$. Susceptibility $\chi(\tau, h)$ can then be measured, from Eq.~\eqref{eq:susceptibility}, by the rate of change in stationary density as external field changes.
\subsection{Empirical methods for estimation of characteristic quantities}\label{sec:estimation-tricks}
In practice, we can estimate $M$, $V$, $T$ and $P(t)$ on the event graph by finding all the \emph{out-components}, i.e., every reachable event starting from every event (see, Figure \ref{fig:dag-schematic}(c)). Calculating the exact set of out-components for every event in the event graph is time and memory-intensive. However, if we are only interested number of events or number of unique nodes that participate in those events, as opposed to the full set of events in the out-components, we can use probabilistic cardinality estimation data-structures to estimate out-component sizes with arbitrary precision in $\mathcal{O}(\abs{E} \log \abs{E})$ time, as opposed to $\mathcal{O}(\abs{E}^2)$ time required for exact calculation \cite{badiemodiri2020efficient}. Minimum and maximum time of all events in the out-component can be exactly calculated in $\mathcal{O}(\abs{E} \log \abs{E})$ time. Calculating properties of the \emph{in-component} of an event is possible through a simple reversal of direction of all links in the event graph and applying the same algorithms.
Similarly, in the homogeneous fully-occupied initial condition scenario, we do not need to directly estimate occupation density $\rho(t)$, stationary occupation density $\rho_\text{stat}(\tau)$ and susceptibility $\chi(\tau, h)$ via naive algorithms, which would explicitly compute these measures by simulating propagation. The properties of homogeneous, fully occupied, $\delta t$-constrained reachability
can be estimated by marking as occupied any event that is in the out-component of at least one event with time $-\infty < t < t_0$. This can be accomplished by running the in-component size estimation algorithm \cite{badiemodiri2020efficient} once over the whole network, recording minimum observed time in in-component of each event and marking those with minimum in-component time smaller than $t_0$ as occupied. In practice, temporal networks are only recorded or generated for a finite window of time $t_\text{min} < t < t_\text{max}$. As there are no adjacency relationship between events more than $\delta t$ apart temporally, any event that has at least one event in its in-component with time $t_\text{min} < t < t_\text{min} + \delta t$ can be considered occupied. Figure \ref{fig:dag-schematic}(d) shows all occupied events (dark grey) with the initial condition that assumes all nodes are occupied from $-\infty < t < 0$. The density of occupied events, which corresponds to particle density $\rho(t)$, can be estimated from the event graph representation by the number of occupied nodes in a band of time divided by the area covered by the band, i.e.~number of nodes multiplied by the width of the band.
Normally, calculating the effects of an external field $h$ would require simulating a fully occupied initial condition, marking some nodes randomly selected with rate $h$ as occupied, computing their out-components, and measuring how many new events got occupied. As we are interested in the effects of a minuscule positive external field, indicated by susceptibility $\chi(\tau, 0)$, we can instead calculate the effects of spontaneously marking exactly one random event in the whole network as occupied using probabilistic counting and in-components of all events (i.e., looking back in time). If the number of events in the in-component of an event $e$ is denoted as $|E^\text{in}(e)|$ and the minimum time among all events in its in component as $t^\text{in}_\text{min}(e) = \min_{(u,v,t) \in E^\text{in}(e)} t$, the expected number of spontaneously occupied events when a minuscule external field $h$ is applied can be estimated as $\sum_{e \in \mathcal{E}} P_\text{occupied}(e)$ where
\begin{equation}
P_\text{occupied}(e) = \begin{cases}
1 & \text{if}\ t^\text{in}_\text{min}(e) < t_0\\
\frac{|E^\text{in}(e)|}{|\mathcal{E}|} & \text{otherwise}\,.
\end{cases}
\end{equation}
In this scenario, the respective value for the external field that would spontaneously occupy on average one event is proportional to $h \propto 1/|\mathcal{E}|$.
We approximate $\rho(t)$ by number of occupied events within a $\delta t$ time window divided by spatio-temporal hyper-volume of the time window $\delta t \times \left| \mathcal{V} \right|$. The estimate for $\rho(t)$ can in turn be used to approximate quantities like stationary density $\rho_{stat}(\tau)$ and susceptibility $\chi(\tau, h)$.
\section{Results}\label{sec:results}
\subsection{Experimental setup}\label{sec:experimental-setup}
In this section, we focus on validating and exploring the limits to our hypothesis that $\delta t$ limited-time spreading in many forms of temporal networks belongs to the directed percolation universality class. We do this by performing single-seed and homogeneous initial-condition spreading simulations following the method defined in Sec.~\ref{sec:dp-analogous-in-temporal-networks} and explained in detail in Sec.~\ref{sec:estimation-tricks}. By measuring various observables for networks of different sizes as described in Sec.~\ref{sec:finite-size-scaling}, we can verify whether for each quantity the corresponding universal scaling functions collapse for systems of different finite sizes when using the same values of critical exponents $\beta$, $\beta'$, $\nu_\parallel$ and $\nu_\perp$ as that of DP corresponding to the dimensions of the system as a previous mean-field approximation and experimental setups for the directed percolation.
The experiments are performed on a variety of synthetic temporal networks. The generation procedure consists of generating a static \emph{base network} corresponding to the aggregate network and generating events, i.e.~activations or timestamps, for each link based on some temporal dynamic. In total, we analyzed 26 combinations of base networks and link-activation processes. In order to perform the finite-size scaling analysis, we computed all the statistics for ten network sizes, starting from $N=2^8$ nodes and increasing the size by a factor of two until we reached $N=2^{17}$ nodes. For the case of $d$-dimensional square grids where $d \in \{ 2, 3, 4 \}$, however, closest powers of $d$ to the powers of two from $2^8$ to $2^{17}$ was used with a periodic boundary condition, to provide spatial translational invariance. Each statistic was calculated as the average of at least 256 (up to 4096) realizations and each realization of the largest configuration consists of around $3.7\times10^7$ events. No sampling of spreading scenarios was required for each network's realization, as the effect of starting a spreading process from any possible combinations of nodes and times could be gathered in one pass as described in Sec.~\ref{sec:estimation-tricks}. See supplementary material for a more detailed overview of the experimental setup.
Static base networks are either (a) one to four-dimensional square lattice grids with periodic boundary conditions (b) random regular graphs with specified average degree \cite{kim2003generating,steger1999generating} or (c) Erdős–Rényi $G(n, p)$ random networks with specified expected average degree \cite{batagelj2005efficient}. For the random networks, we chose the average degrees 8 for the Erdős–Rényi graphs and 9 for the random regular graphs (such that both networks have the same expected excess degree). The higher degrees of random networks ensure that the probability of generating networks with large isolated components remains negligible and that even locally, the network would be of high enough dimensionality to be in the mean-field regime above the upper critical dimension $d_c = 4$.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{timeseries-sample.pdf}
\caption{Sample timestamps from a single realization (activations of a single link) with different temporal dynamics. Each point represents a single activation at a specific time. The points are scattered over the vertical axis to avoid overlaps in the visualization. All timestamps were generated for 256 units of time with parameters or minimum cutoffs that would result in an expected inter-event time of 1. Equation \eqref{eq:hawkes-process-intensity} defines the parameters and the intensity function of the Hawkes univariate exponential self-exciting process.}
\label{fig:timestamps}
\end{figure}
Temporal dynamics of the links are either governed by (a) Poisson processes, i.e.~exponential inter-event times, (b) bursty processes, i.e.~renewal processes with power-law inter-event time distributed as $\propto \Delta t^-\gamma$ with exponents $\gamma \in \{ 2.05, 2.2, 2.8, 5.2 \}$ and minimum interval cutoff set so that the expected inter-event would be equal to one and (c) Hawkes independent self-exciting processes with different parameter sets. The Hawkes univariate exponential self-exciting process \cite{hawkes1971point}, is defined by the conditional intensity function
\begin{equation}\label{eq:hawkes-process-intensity}
\lambda^*(t) = \mu + \alpha \theta \sum_{t_i < t} \mathrm{e}^{-\theta (t - t_i)}\,.
\end{equation}
The parameters of this formulation of the Hawkes process are (1) background (or exogenous) intensity of events $\mu$ indicating the random probability of events happening without being caused through self-excitement, (2) the infectivity factor $\alpha$, which can be interpreted as the expected number of induced self-exciting events per each event, and (3) the rate parameter of the delay $\theta$. Based on the properties of exponential kernel used in defining Eq.~\ref{eq:hawkes-process-intensity}, $1/\theta$ is the expected inter-event time between an event (e.g.~a coincidental social interaction) and its corresponding induced self-exciting event (e.g.~the follow-up social interactions) \cite{laub2015hawkes}.
As the unit of time is arbitrary, temporal processes are scaled, without loss of generalization, so that they produce timestamps with a mean inter-event time equal to one. The processes are initialized in their stationary state, and in practice, the first timestamp for each event is generated through residual time distribution of each process, except for the case of Hawkes process where the process is allowed a burn-in time equal to the simulation time window before the first timestamp is recorded. The temporal processes of pairs of links are simulated independently of each other. Figure \ref{fig:timestamps} shows a visualization of the different methods of generating event activations. Temporal networks were simulated for a time window of at least $T = 64$ and up to $T = 8192$ units of time. See supplementary material for the exact experimental setup for each system size. The difference in system sizes and time windows for the simulations were necessitated by the limitations and optimal utilization of the computational facilities.
\subsection{Estimating the critical threshold \texorpdfstring{$\delta t_c$}{dt-c} and the critical exponents \texorpdfstring{$\beta$}{beta}, \texorpdfstring{$\beta'$}{beta'}, \texorpdfstring{$\nu_\parallel$}{nu-parallel} and \texorpdfstring{$\nu_\perp$}{nu-perpendicular}}\label{sec:estimating-dtc}
Best estimate of the critical exponents $\beta$, $\beta'$, $\nu_\parallel$, $\nu_\perp$ and critical threshold $\delta t_c$ can be determined by finding the values of these exponents that would produce the best data collapse for the universal scaling functions corresponding to $\rho(t)$, $\hat{P}(t)$, $M(t)$, and $V(t)$. The quality of collapse, in turn, can be assessed by comparing the deviation of the scaling function curves for different system sizes from the average trajectory. Here, for each of the quantities $\hat{P}(t)$, $\rho(t)$, $M(t)$ and $V(t)$, we calculated one trajectory for finite-size scaling function for each system size, as defined for example for the case of $\rho(t)$ by Eq.~\eqref{eq:scaling-for-rho}. As the tested value of critical exponents and $\delta t_c$ gets closer to the actual critical threshold, the curves for different sizes should more closely collapse on top of each other. Plotted with the correct values of critical exponents and critical threshold, we expect to see all trajectories collapse into one with the possible exception of very small values of $t$. To quantify the quality of a collapse, we measure the mean curve in the area where all system sizes have defined values for the scaling function and measure the root mean square difference of all points from all system sizes to the mean curve. The errors were measured after logarithmically scaling the values to account for the power-law nature of the scaling functions. Sum of errors for the collapse of $\hat{P}(t)$, $\rho(t)$, $M(t)$ and $V(t)$ was used in evaluating each set of parameters.
In order to assess collapse of the universal scaling functions, we first determine a value for $\delta t_c$ for each network configuration. That is, the best candidate for $\delta t_c$ is selected based on the least total error for collapse of $\hat{P}(t)$, $\rho(t)$, $M(t)$ and $V(t)$ assuming DP critical exponents. Figure \ref{fig:dtc-optimisation} shows this total error of collapse for two network configurations. This shows is a clear minimum for each configuration indicating the critical value $\delta t_c$, which is consistent across $\hat{P}(t)$, $\rho(t)$, $M(t)$, and $V(t)$ trajectories.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig-example-rmse.pdf}
\caption{Root mean square (logarithmic) deviation of scaling-corrected functions of probability of survival $\hat{P}(t)$, density $\rho(t)$, mass $M(t)$ and volume $V(t)$ for different system sizes from average trajectory shows a sharp drop at $\delta t_c$ due to data collapse. Each instance of the network is made through realizations of (a) Erdős–Rényi static network $\langle k \rangle = 8$ and Poisson process $\lambda=1$ activations, (b) random 9-regular networks with bursty (power-law with minimum cutoff) inter-event time distribution with mean 1 and exponent $\gamma = 2.8$, (c) Erdős–Rényi static network $\langle k \rangle = 8$ and Hawkes univariate exponential self-exciting process with parameters $\mu=0.2$, $\alpha=0.8$ and $\theta=1.0$ and (d) one-dimensional grid with periodic boundary conditions (a circle) and Poisson process $\lambda=1$ link activations. Refer to Sec.~\ref{sec:experimental-setup} for the definitions of the parameters.
}
\label{fig:dtc-optimisation}
\end{figure}
\begin{figure}[hb!]
\centering
\includegraphics[width=\linewidth]{erdos-poisson-beta-beta_prime.pdf}
\caption{Total error of collapse of universal scaling functions of $M(t)$, $V(t)$, $\hat{P}(t)$ and $\rho(t)$ for Erdős–Rényi networks $\langle k \rangle = 8$ and Poisson process activation $\lambda = 1$ as a function of $\beta$ and $\beta'$. In these visualizations we set $\nu_\perp=0.5$, $\nu_\parallel=1$ and $\delta t_c=0.08421$, and vary one of these parameters such that the three panels from left to right correspond to values (a) $\nu_\perp \in \{0.34, 0.5, 0.66\}$, (b) $\nu_\parallel \in \{0.84, 1, 1.16\}$ and (c) $\delta t_c \in \{0.0840,0.08421,0.0844\}$. Note that the center panel is repeated across the rows and always has parameter values $\nu_\perp=0.5$, $\nu_\parallel=1$ and $\delta t_c=0.08421$. We see that there is a minimum in the error close to $\beta = \beta' = \nu_\parallel = 1$ and $\nu_\perp=0.5$ within this five-dimensional space.}
\label{fig:grid-search-beta-beta-prime}
\end{figure}
\begin{figure*}[ht!]
\centering
\includegraphics[width=\linewidth]{scaling-paper.pdf}
\caption{Universal scaling functions for $\delta t$ limited-time reachability over (a,c,e,g,i,k,m,o) random 9-regular network with bursty (heavy-tail with minimum value cutoff) link activation with mean inter-event time of 1 and exponent $\gamma=2.8$ and (b,d,f,h,j,l,n,p) random Erdős–Rényi network $\langle k \rangle = 8$ with Hawkes univariate exponential self-exciting process link activation with parameters $\mu = 0.2$, $\alpha = 0.8$ and $\theta = 1.0$. The finite size scaling is performed for the following single-source scenarios: (a,b) the mean component mass $M$ as function of $\delta t$ close to critical point and (c,d) as function of time $t$ at the critical point, (e,f) the mean component volume $V$ as function of $\delta t$ close to critical point and (g,h) as function of time $t$ at the critical point and (o,p) Survival probability $\hat{P}$ as function of time $t$ at the critical point. For fully-occupied initial conditions the finite-size scaling is performed for (k,l) the occupation density $\rho$ as function of time $t$ at the criticality, and both (i,j) the static density $\rho_{\text{stat}}$ and (o,p) susceptibility $\chi$ as function of $\delta t$ close to the critical point. The collapse of the universal scaling functions validates the hypothesis that these systems are governed by the same critical exponents as in directed percolation in the mean-field regime.
See Sec.~\ref{sec:experimental-setup} for the full definitions of the parameters.}
\label{fig:regular-bursty-scaling}
\end{figure*}
The resulting estimates for $\delta t_c$ can be used to visually verify directed percolation critical exponents and our selected optimal value of $\delta t_c$ for each system by plotting the finite-size universal scaling functions of different system sizes. In total, we produce collapses for eight characteristic quantities measured in a single source or homogeneous initial conditions. Figure~\ref{fig:regular-bursty-scaling} shows these collapses measured for regular networks with bursty dynamics (renewal process with power-law inter-event times) and Erdős–Rényi networks with a Hawkes self-exciting process dynamics. The full set of plots for all of the 26 configurations are shown in Supplementary Materials. In all cases, a satisfactory collapse can be observed for at least probability of survival $\hat{P}(t)$ and density $\rho(t)$ and in most cases, other quantities show a good collapse as well. It is important to note that quantities that depend on measuring values as time approaches infinity, e.g., $\rho_\text{stat}(\delta t)$ and $\chi(\delta t)$ have generally lower quality of measurement and collapse since the time to reach a stable value for these increases substantially close to criticality \cite{hinrichsen2000non}.
Table \ref{tab:best-fits-and-errors} (column ``Est. $\delta t_c$'') shows our best estimate of the critical threshold $\delta t_c$ for each configuration using the method described above. As the systems become rapidly more and more connected after the critical threshold, a lower value for the critical threshold $\delta t_c$ indicates higher, or more robust, spatio-temporal connectivity, meaning that the same $\delta t$ limited-time spreading would result in a larger number of reachable nodes, $V(\delta t)$, or larger number of reachable events, $M(\delta t)$. When modeling infectious disease spreading as directed percolation on temporal networks, larger values for $V(\delta t)$ and $M(\delta t)$ may indicate larger epidemic sizes and the total number of human hours of infection in the population, respectively.
\begin{table*}[h!]
\caption{\label{tab:best-fits-and-errors}
Column ``Est. $\delta t_c$'' shows the best candidate for critical threshold $\delta t_c$ selected by minimizing the collapse error of the universal scaling functions for probability of survival $\hat{P}(t)$, density $\rho(t)$, mass $M(t)$ and volume $V(t)$ derived for different system sizes, assuming DP exponents. The collapse error is measured by the sum of root mean squared deviation of logarithmically scaled trajectories for all four scaling functions. For the best estimate for exponents $\alpha=\beta/\nu_\parallel$ and $\delta=\beta'/\nu_\parallel$, reported respectively in columns ``Est. $\alpha$'' and ``Est. $\delta$'', a power-law was fitted to the head of values of probability of survival $\hat{P}(t)$ and density $\rho(t)$ respectively for the largest system size simulated for the time period between $2 < t < 0.04 \times N^{\nu_\parallel/{d \nu_\perp}}$, where both functions are expected to be still mostly behaving, similar to an infinite system, according to power relations $t^{-\alpha}$ and $t^{-\delta}$ respectively. Directed percolation mean-field values for these critical exponents are $\delta = \alpha = 1$, which is close to the value estimated for random, high-dimensional networks. Furthermore the value of these critical exponents in a DP system are expected to be close to $\alpha=\delta=0.15946$ for 1+1 dimensional, $\alpha=\delta=0.450$ for 2+1 dimensional, $\alpha=\delta=0.732$ for 3+1 dimensional and equal to the mean-field estimates systems $\alpha=\delta=1$ for 4+1 dimensional \cite{hinrichsen2000non}, which is close to values estimated for 1 dimensional lattice and 2 to 4 dimensional square lattices.
}
\begin{ruledtabular}
\begin{tabular}{lddddddd}
\textrm{Configuration}&
\multicolumn{1}{c}{\textrm{Est. $\delta t_c$}}&
\multicolumn{1}{c}{\textrm{$\beta$ Error}}&
\multicolumn{1}{c}{\textrm{$\beta'$ Error}}&
\multicolumn{1}{c}{\textrm{$\nu_\parallel$ Error}}&
\multicolumn{1}{c}{\textrm{$\nu_\perp$ Error}}&
\multicolumn{1}{c}{\textrm{Est. $\alpha$}}&
\multicolumn{1}{c}{\textrm{Est. $\delta$}}\\
\colrule
\input{table}
\end{tabular}
\end{ruledtabular}
\end{table*}
These results indicate that within each spatial configuration, increased burstiness (as indicated by lower value for the power-law exponent $\gamma$) generally leads to a lower value for $\delta t_c$ threshold and higher connectivity. Furthermore, for the case of the self-exciting process, increasing the expected number of self-induced events, as indicated by $\alpha$, generally results in a lower value for $\delta t_c$ (higher connectivity). While it was previously understood that a wide range of temporal inhomogeneities slows down spreading processes over temporal networks \cite{karsai2011small}, these results demonstrate that certain temporal inhomogeneities, e.g.~a highly bursty or self-exciting temporal dynamic, can enable a \emph{more limited} spreading agent (expressed in terms of a maximum waiting time) to spread to a wider set of nodes. For example, spreading processes with maximum waiting time between $0.063 < \delta t < 0.084$ over an Erdős–Rényi networks $\langle k \rangle = 8$ will spread to a much larger set of nodes and span a longer span of time if the link activations are highly bursty ($\gamma = 2.05$) compared to a Poisson process with the same mean inter-event time, as the latter will be spreading in the sub-critical regime compared to the super-critical regime for the former.
It is also interesting to note that while the random spatial configurations, namely random 9-regular networks and the Erdős–Rényi networks $\langle k \rangle = 8$, both result in networks with the same expected excess degree value, the Erdős–Rényi networks with higher levels of spatial inhomogeneity, which manifests as a wider spread degree distribution, can be observed to have a lower $\delta t_c$ critical threshold. While testing on a wider range of spatial (structural) inhomogeneities would be required before a conclusion is reached, these results might hint at a similar behavior as with temporal inhomogeneities, namely that introducing certain spatial inhomogeneities might result in higher connectivity in the sense that the same limited-time spreading agent can eventually spread to a wider share of the network.
Additionally, we present a method to assess the quality of a collapse for a range of different values of critical exponents ($\beta$, $\beta'$, $\nu_\parallel$ and $\nu_\perp$) and $\delta t_c$. A five-dimensional grid search for optimal values for critical exponents and $\delta t_c$ based on the quality of collapse for $P(t)$, $\rho(t)$, $M(t)$ and $V(t)$ shows that the total error declines around critical exponent values close to that of directed percolation, i.e.~$\beta=\beta'=\nu_\parallel=1$ and $\nu_\perp = 0.5$ for mean-field regimes and their respective DP values for lower dimensionality square grid networks. Figure~\ref{fig:grid-search-beta-beta-prime} shows for Erdős–Rényi static networks with $\langle k \rangle = 8$ and Poisson process link activation, the $\beta\cross\beta'$ plane from the five-dimensional grid search with two sandwiching parallel planes along each of the $\nu_\parallel$, $\nu_\perp$ and $\delta t_c$ dimensions. This verifies that there is a minimum close to $\beta=\beta'=\nu_\parallel=1$, $\nu_\perp = 0.5$ and $\delta t_c = 0.08421$ for total error of collapse of $P(t)$, $\rho(t)$, $M(t)$ and $V(t)$. Similar plots for some other network configurations (along with a different two-dimensional slice, $\nu_\parallel \cross \nu_\perp$) can be viewed in the Supplementary Material. It is important to note that while other combinations of parameters in the grid might lead to other local optima, visual inspection of the resulting collapse show that to be mainly numerical artifacts where the total error changes rapidly close to extreme values of the parameters (i.e., critical exponents and $\delta t_c$) where only a very small fraction of the trajectories for different finite sizes actually overlap.
Furthermore, for each of the critical exponents, we can measure an estimation error based on this five-dimensional parameter grid. For each exponent, we find a range of values where, assuming that all other exponents are fixed at their DP values, would produce collapses of higher or equal quality compared to the DP value of that exponent. The sizes of these ranges, which by definition includes the DP value for all exponents, provides a confidence interval for the range of possible exponent values that are able to explain the behavior of the system with at least the same quality as that of directed percolation. As shown in Table~\ref{tab:best-fits-and-errors}, these errors are in most cases only a few percent, with a notable exception of the highly bursty renewal processes with $\gamma = 2.05$. Simulating power-law distributions becomes a much harder problem as the magnitude of the exponent approaches 2. Close to this exponent, it takes a larger and larger number of realizations for the properties of the population, e.g.~average inter-event time for bursty temporal dynamics, to converge. It is also possible that the large estimation error is an indicator that the system is approaching a breakdown of one of the key symmetries, with the most likely candidate being rapidity-reversal symmetry based on the fact that the estimation error for $\beta'$ is much larger than that of the other exponents.
\subsection{Estimating critical exponents by simulating very large systems}\label{sec:estimating-alpha-delta}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{fig-example-exponent.pdf}
\caption{An example of fitting power-law functions on empirical $\rho(t)$ and $\hat{P}(t)$ results on finite networks for deriving critical exponents $\alpha = \beta/\nu_\parallel$ and $\delta = \beta'/\nu_\parallel$. Power-law functions were fitted on experimental results of spreading over Erdős–Rényi networks with $\langle k \rangle = 8$ and $N=2^{17}$ nodes and bursty (power-law with minimum cutoff) inter-event time distribution with mean 1 and exponent $\gamma = 2.8$. The fitting was performed on values in range $1 < t < 0.04 \times N^{\nu_\parallel/{d \nu_\perp}}$ (i.e.~$1 < t < 14.48$) to limit the interference of finite-size effects with the scaling behavior.}
\label{fig:exponent-fit}
\end{figure}
As discussed before in Sec.~\ref{sec:directed-percolation}, the effects of the finite size of the system manifest at characteristic times $t_f \propto N^{\nu_\parallel/{d \nu_\perp}}$ in the form of fluctuations that causes the transition of the system to the absorbing phase. At times much smaller than $t_f$ the system shows approximately the scaling behavior of an infinitely large system where at criticality, $\rho(t) \sim t^{-\alpha}$ and $\hat{P}(t) \sim t^{-\delta}$ where $\alpha = \beta/\nu_\parallel$ and $\delta = \beta'/\nu_\parallel$. On the other hand, the power-law scaling behavior becomes visible at times comparable to the mean inter-event time of the dynamic process but not up to arbitrarily infinitesimal values of $t$. Given these properties, we fitted two power-law functions using the least-squares method to the results of experiment with the largest system size for the range of time $2 < t < 0.04 \times N^{\nu_\parallel/{d \nu_\perp}}$ on $\rho(t)$ and $\hat{P}(t)$ to derive exponent $\alpha$ and $\delta$. Figure \ref{fig:exponent-fit} shows one such fitting for a system made from Erdős–Rényi networks with $\langle k \rangle = 8$ and $N=2^{17}$ nodes and bursty (power-law with minimum cutoff) inter-event time distribution with mean 1 and exponent $\gamma = 2.8$. Table \ref{tab:best-fits-and-errors} (columns Est. $\alpha$ and Est. $\delta$) shows the best estimates of these exponents, which as expected are very close to respective directed percolation critical exponents of 1 (for the mean-field regime $d \geq 4$) for the case of random networks and 0.159, 0.450 and 0.732 for one-, two- and three dimensional lattices respectively \cite{hinrichsen2000non}.
\section{Discussions}\label{sec:discussion}
Through combining multiple methods of empirical and theoretical verification, we are able to confidently state that limited waiting-time connectivity percolation over a wide range of synthetic temporal networks incorporating a range of temporal and topological inhomogeneities show behavior compatible with the directed percolation universality class. It is of utmost importance to discuss the limitations of our method: chief among them, that our empirical finite-size simulation method, as described in Sec.~\ref{sec:estimation-tricks}, is not able to measure quantities which are defined at $t \rightarrow \infty$, such as the ultimate probability of survival $P_\text{surv}$ and static density $\rho_{stat}$ (and therefore susceptibility $\chi$) to the same standard of accuracy as the other quantities due to the finite size of the synthetic networks used for analysis. This is exacerbated close to the critical threshold where the equilibration time, the time required for the network to reach a stationary state, grows rapidly while the memory and computational cost of simulating a temporally larger temporal network grow linearly and log-linearly, respectively, with the increased simulated time \cite{hinrichsen2000non}. This is visible in Fig.~\ref{fig:regular-bursty-scaling}e,f,k,l as a worse collapse as compared to other quantities.
Also, while it is computationally much more feasible to measure susceptibility $\chi$ by inducing occupation of exactly one existing event in the temporal network (described in Sec.~\ref{sec:estimation-tricks}) as compared to inducing occupation of nodes at random times (as described in Sec.~\ref{sec:directed-percolation}), the latter method might be more robust, especially when dealing with a temporal network with a high degree of temporal inhomogeneity. Although our experiments with this alternative method were limited to smaller system sizes, we could not observe any significant difference between the two methods for the network configurations presented in this manuscript.
While a wide range of temporal dynamics and network structures with different levels of inhomogeneity are studied here, there is still a wide variety of systems that present computational and theoretical challenges. First, the effects of event-event correlations between links are not studied. It has been shown that event-event correlations, among other forms of inhomogeneity, can affect the rapidity of the spreading process on temporal networks \cite{karsai2011small}. Conceptually, local event-event correlations such as temporal motifs \cite{kovanen2011temporal}, are close to temporal event graphs, which are in practice computed using isomorphisms on slightly modified temporal event graphs. Thus, incorporating temporal motifs to the framework at the level of analytical computations is an interesting future direction, as that corresponds to modifying the frequency of appearance of structural motifs in the event graphs. Second, the effect of static base networks with heavy-tail degree distributions and other more complicated network topologies are absent from this study. Here, of especial interest are the networks with heavy-tailed degree distribution with static network reachability percolation threshold at zero occupied links, e.g.~$p(k) \propto k^{-2}$. While initial results did not support the conclusion that a $\delta t$ limited waiting time over this class of synthetic temporal networks would be in the directed percolation universality class, due to limitations on computational resources, we were not able to perform the analysis on the larger system sizes comparable to the other types of networks.
Depending on the physical mechanism involved in the modeled connectivity phenomenon or spreading process, alternative methods of defining the adjacency relationship might be more suitable than the one used here. For example, for the case of disease spreading over a physical contact network, the currently used definition of event graph causes a ``re-infection'' of the infected party, manifested as a restart of their $\delta t$ duration of disease. This can be resolved by substituting each undirected event in the temporal network with two simultaneous directed events. Similarly, for a disease spreading scenario over transportation networks, such as an airplane traffic network, the time between two events (the value that is compared to the maximum duration of disease $\delta t$ to determine whether two flights are adjacent) should be calculated from the departure of one flight to departure of the possibly adjacent flight and not, as it is currently presented, from the arrival of the latter to the departure of the former. This might be an important factor when dealing with scenarios in which the reasonable values for $\delta t$ are comparable to the delay or the duration of the events, e.g.~the time from the departure of a flight to the arrival in a spreading process over an air transport network.
For some spreading mechanisms, it might also be more suitable to replace the hard $\delta t$ limited-time cutoff of adjacency requirement used in this work with a probabilistic process by measuring quantities over an ensemble of event graphs. For example, using a Poisson process instead of a $\delta t$ limited-time cutoff would produce dynamics similar to simulations of SIS processes over networks while simulating results of the simulation starting at every possible starting point in one pass. Viewed this way, normal $\delta t$ limited-time cutoff can be seen as a probabilistic process where the probability of adjacency is a step function at $\Delta t = \delta t$. It is also possible to combine an occupation probability similar to classic directed percolation (see Sec.~\ref{sec:directed-percolation}) with a $\delta t$ limited-time cutoff (or a Poisson process cutoff or other forms of temporal locality constraint) to construct a two-dimensional phase diagram for each temporal network.
It would also be possible to define connectivity in the event graphs in a way that mimics the SIR process. In this case, one would need to prune some of the temporal paths in the event graph such that temporal network nodes are not repeated. This distinction is equivalent to paths and simple paths (or walks and paths, respectively) in static graphs. The algorithmic techniques employed in this work are not directly applicable to this case, and in fact, it has been recently shown that algorithmic problems in such settings can be computationally difficult. For example, in the SIR interpretation of the event graph, finding if it is possible for a node infected at a specific time to infect a given node is an NP-hard problem \cite{casteigts2019computational}. In any case, averaging over explicit simulations of spreading scenarios is always an alternative option to the algorithms that take advantage of the redundancies in computing reachability.
Connectivity, which encapsulates several important phenomena on complex systems such as spreading processes \cite{lambiotte2016guide, holme2012temporal, holme2015modern} and routing dynamics \cite{nassir2016utility}, has not yet undergone the same level of development on temporal networks as the static networks. It has been previously suggested that connectivity on temporal networks, or other adjacent representations such as dynamic networks, might show the same properties as any other directed percolation system \cite{parshani2010dynamic,kivela2018mapping}, a class of percolation models with built-in directionality which has enjoyed abundant attention in the past decades. In Ref.~\cite{shortpaper}, we laid formal foundations by providing one-to-one analogues between concepts from directed percolation and temporal network connectivity and provided theoretical evidence supporting this hypothesis. In this work, we presented multiple accounts of empirical evidence showing that connectivity on many model temporal networks belongs to the directed percolation universality class and that this hypothesis is robust for a range of temporal and spatial heterogeneities.
This work focused mainly on establishing the vocabulary and developing the required tools in the hopes of rendering studies of connectivity in temporal networks ripe for future analysis, especially from a critical phenomena perspective. It is important to note that this work has only scratched the surface of the analytical study of connectivity on temporal networks and still, a vast body of analytical and phenomenological topics, some of which were eluded to in the previous paragraphs, remains open for future study.
\section{Acknowledgement}
We would like to thank János Kertész and Géza Ódor for their helpful comments and suggestions. The authors wish to acknowledge CSC -- IT Center for Science, Finland, and Aalto University ``Science-IT'' project for generous computational resources. Márton Karsai acknowledges support from the DataRedux ANR project (ANR-19-CE46-0008) and the SoBigData++ H2020 project (H2020-871042).
|
\section{Conclusion}\label{conclusion}
The simultaneous inference of hidden dynamic states and model parameters of a stochastic reaction network is an important and difficult task.
One challenge lies in the high computational complexity of the exact simulation method for the stochastic reaction dynamics, which makes the particle filter, a simulation-based algorithm, computationally inefficient.
Moreover, the sample degeneracy occurring in conventional particle filters has a negative impact on the accuracy of the estimate.
To tackle these problems, this paper introduced regularized particle filters (RPFs) for stochastic reaction networks (SRNs) on different scales using computationally efficient simulation algorithms.
Specifically, the filter chooses the proper simulation algorithm for sampling based on the scale of the system considered so that the computational time is acceptable (\Cref{subsubsection tau-leaping and PDMP} and \cref{def RPFs for SCRNs}).
Also, an artificial evolution step is executed after resampling to mitigate sample degeneracy and cope with the fluctuation of model parameters in practical systems (\cref{remark artificial evolution}).
Furthermore, by parameter sensitivity analyses, we showed that the constructed RPF is convergent to the exact filter for a broad class of SRNs (\cref{thm main} and \cref{remark covered SCRNs}).
Finally, several numerical examples were shown to illustrate the superior performance of the constructed RPF compared with its sequential-importance-resampling (SIR) counterpart.
From the numerical examples, we can observe that the constructed RPF also has some drawbacks; for instance, not every parameter can be inferred accurately when the number of unknown parameters is large.
This suggests that the filter does not scale very well with the system dimension.
A possible solution to this problem is the Rao-Blackwellization method \cite{murphy2001rao}, in which some variables are marginalized out so that the filter only needs to simulate a model of reduced dimensionality.
For mass-action networks, such a scalable filter has been provided in \cite{zechner2014scalable}, where the model parameters are integrated out by assuming their priors to be Gamma distributions.
However, the theory for general kinetics is still undeveloped, and the combination of it with the RPF also remains to be investigated.
Furthermore, the numerical examples indicate that the posterior of some hidden dynamic variables can be either very noisy or extremely narrow disregarding the model parameters (see the discussion in \Cref{subsection gene transcription model} and \Cref{subsection transcription regulation network}).
This means that these variables are either impossible to infer accurately or can be learned in a more clever way with biological insights into the system.
Consequently, eliminating these variables can be another option to further reduce the system and improve the filter.
Lastly, it is known that unweighted particle filters can avoid the curse of dimensionality in some circumstances \cite{surace2019avoid}; therefore, it is worth applying this methodology to biological filtering problems.
However, conventional unweighted particle filters (e.g., the particle flow method \cite{daum2008particle} and feedback particle filters \cite{yang2013feedback}) are designed for systems of continuous state spaces.
For SRNs whose state space is discrete, one needs to significantly modify the theory and algorithms.
\section{Introduction}
The development of fluorescence technologies \cite{zhang2002creating} and modern microscopes \cite{stephens2003light,vonesch2006colored} in the past few decades has greatly improved scientists' ability to study various biological problems at the single-cell level \cite{locke2009using}.
For instance, researchers can now use fluorescent data to investigate the variability in single-cell gene expression \cite{rosenfeld2005gene}, and the role of transcription factors in shaping bursty dynamics \cite{rullan2018optogenetic}.
Despite these successes, a time-lapse microscope can only follow a few genes over time due to the limited availability of distinguishable reporters \cite{locke2009using};
consequently, many chemical species in the cell are not directly tracked.
This drawback greatly limits scientists' ability to further investigate and control the dynamical behaviors of single-cell systems.
Consequently, it is of immediate interest to establish efficient stochastic filters that can accurately infer the unobserved species in an intracellular reaction system by partial observations.
The filtering theory that aims to calculate the posterior of hidden dynamic states given the time-course observation has been extensively developed in the past few decades.
In the linear and Gaussian scenarios, the solution to the filtering problem is the well-known Kalman filter \cite{kalman1960new}, which can be calculated exactly in a computationally efficient way.
In contrast, the nonlinear and non-Gaussian systems usually lead to infinite-dimensional filters that cannot be calculated explicitly \cite{bain2008fundamentals}.
In this case, the particle filter (PF) originally introduced in \cite{gordon1993novel} is the most effective method to numerically solve the filtering problem, which approximates the posterior by a population of weighted samples generated by importance sampling (from the target dynamics) and resampling \cite{doucet2009tutorial}.
So far, PFs have been successfully applied to biochemical reaction systems under different problem settings, e.g., the heat shock response system \cite{liu2012state}, metabolic dynamics \cite{yang2007vivo}, transcriptional switch models \cite{hey2015stochastic}, multiscale reaction systems \cite{fang2021stochastic,fang2020stochastic}, biochemical processes of noise-free observations \cite{rathinam2020state}.
Moreover, the method has been shown to outperform other stochastic filters (e.g., extended Kalman filters and unscented Kalman filter) in general situations \cite{liu2012state}.
Usually, an intracellular reaction system is modeled by a continuous-time Markov Chain, also known as the stochastic reaction network (SRN), due to low molecular counts \cite{mcadams1997stochastic}.
Its dynamics can be exactly simulated by the Gillespie stochastic simulation algorithm \cite{gillespie1976general,gillespie1977exact} or the next reaction method \cite{gibson2000efficient}, whose computational complexity is proportional to the rate of the fastest reaction.
Commonly, an SRN has a multiscale nature, meaning that the concentration levels of the involved species differ by several orders of magnitude, and so do the reaction rates.
Consequently, the exact simulation of an intracellular reaction system can take an impractical amount of computational time, and the associated PF, as a simulation-based algorithm, can be computationally inefficient.
Accelerating algorithms for SRNs include the linear noise approximation \cite{van1976expansion}, chemical Langevin equation \cite{gillespie2000chemical} , tau-leaping algorithm \cite{gillespie2001approximate,cao2006efficient,rathinam2003stiffness}, piecewise deterministic Markov process (PDMP) \cite{kang2013separation,hepp2015adaptive,crudu2009hybrid}, etc., all of which
reduce the computational complexity via approximating the firing of fast reactions by more tractable processes.
(We refer interested readers to the literature \cite{schnoerr2017approximation} for a related discussion.)
Inspired by these methods, researchers also proposed computationally efficient particle filters using linear noise approximations \cite{sherlock2014bayesian}, chemical Langevin equations \cite{liu2012state,golightly2006bayesian}, and PDMPs \cite{fang2020stochastic,fang2021stochastic,hey2015stochastic}.
Another feature of intracellular reaction systems that needs to be considered is the large variability in model parameters from cell to cell, which suggests that model parameters should be inferred simultaneously with the state variables in the filtering problem.
A common way to combine these two tasks together is to view the model parameters as additional states of the system and, then, to use the PF to infer the augmented state vector.
In this situation, the conventional sequential importance resampling PF (SIRPF) suffers from sample degeneracy,
meaning that the particle diversity in parameters drops dramatically over time, and, therefore, the particles fail to represent the posterior accurately.
Good alternatives to SIRPFs include the resample-move method \cite{gilks2001following,berzuini2001resample} and the regularized particle filter (RPF) \cite{liu2001combined,oudjane2000progressive}, all of which mitigate sample degeneracy by perturbing particles with (additional) artificial noise and have been successful in various applications (see the aforementioned references).
This methodology is easy to implement but has the drawback that it treats model parameters (which are fixed over time) as time-varying variables and, consequently, throws aways some information about them \cite{liu2001combined}.
In the finite particle population case, researchers still cannot quantify the bias introduced by the artificial noise in an RPF \cite{kantas2015particle}.
However, in the large-particle limit, the RPF has already been shown to converge to the exact filter if the artificial noise is weak and the transition kernel satisfies some regularity conditions, e.g., the globally Lipschitz continuity \cite[Proposition 2.38]{del2000branching}, the mixing condition in the metric space \cite{le2004stability}, and the Lipschitz continuity of transition kernel's derivatives \cite{crisan2014particle}.
Motivated by the above facts, we present in this paper a computationally efficient RPF for multiscale SRNs based on reduced dynamical models and identify conditions under which the performance of the established RPF is guaranteed.
We note that the existing results about the convergence of RPFs cannot be directly applied to our problem.
First, the globally Lipschitz condition of the transition kernel required in \cite[Proposition 2.38]{del2000branching} is too restrictive for SRNs that only a limited class of networks, such as those having linear propensities or bounded trajectories, satisfy it.
In addition, the discreteness of the state space of an SRN suggests that the mixing condition in the whole metric space \cite{le2004stability} and the derivative related condition \cite{crisan2014particle} can never hold.
Instead, in this paper, we approach the problem by utilizing parameter sensitivity results in \cite{gupta2013unbiased,gupta2019sensitivity,gupta2018estimation}.
The rationale is that the parameter sensitivity can be used to quantify the effect of the artificial noise exerted to the model parameters; once we show that the error caused by artificial noise is small, the convergence of the RPF holds automatically.
Following this idea, we show that the established RPF converges to the exact filter in probability, if all the reactions that lead to an increase of the molecule population have linearly growing propensities and some other mild conditions are satisfied simultaneously.
The provided conditions are quite general, and they cover a large class of SRNs, including all mass-action networks in which only unimolecular reactions can lead to an increase of the total molecular population.
For the sake of simplicity, we only consider discrete-time observations in this paper.
However, we conjecture that the obtained result can be naturally extended to the continuous-time observation case with mild modifications, {as the parameter sensitivity analysis, the major technique used here, applies for continuous observations.}
Besides, compared with the SIRPF that we provided previously for multiscale SRNs in \cite{fang2021stochastic,fang2020stochastic}, the RPF established in this paper has a relatively similar computational cost but a superior performance in inferring model parameters when the parameter uncertainty is large.
The rest of this paper is organized as follows.
In \Cref{sec preliminary}, we briefly review the basics of SRNs and introduce the associated filtering problems.
Then, by using different dynamical models, we establish computationally efficient RPFs for SRNs on different scales and provide several mild conditions that guarantee the performance of the established algorithm in the limit of large particles (see \Cref{section RPF}).
When the particle population is finite, we also give a tuning rule for the hyperparameter (the intensity of artificial noise) so that a more precise estimate can be obtained.
In \Cref{Sec. numerical examples}, several numerical examples are presented to illustrate the effectiveness of the constructed RPF, particularly its superior performance to the SIRPF.
Finally, \Cref{conclusion} concludes this paper.
To improve the readability of the paper, all the proofs are presented in the appendix.
\section{Regularized particle filter (RPF) for SRNs} \label{section RPF}
\subsection{Constructing RPFs for SRNs on different scales}
Usually, the conditional distribution of an SRN cannot be computed explicitly due to the nonlinearity of the dynamics.
Therefore, we apply particle filters to the associated filtering problem.
The key idea of the particle filtering (also known as the sequential Monte Carlo method \cite{gordon1993novel}) is to approximate the posterior by a number of weighted trajectories sampled from the distribution of the target system.
A detailed particle filtering algorithm is provided in \Cref{alg discrete time regularized particle filters}.
Specifically, the sampling step in \Cref{alg discrete time regularized particle filters} mimics the recursive formulas \eqref{eq. recurssive expression 1} (the prediction step) and \eqref{eq. recurssive expression 2} (the adjustment step), in which weighted particles are generated to represent the conditional distribution.
The resampling step is executed to remove insignificant particles and, therefore, mitigate long-term weight degeneracy at the cost of adding additional noise at the current step \cite{doucet2009tutorial}.
In this paper, we use the residual resampling, as it introduces the minimum noise among all resampling methods \cite[Exerice 9.1]{bain2008fundamentals}.
Finally, particles are perturbed by artificial noise so that the particle diversity improves, and the sample degeneracy meaning that all particles become identical can be avoided.
Such a particle filtering algorithm with an artificial evolution step is called a regularized particle filter (RPF); if the artificial evolution is not executed, then this algorithm is called a sequential importance resampling particle filter (SIRPF).
\begin{algorithm}[htbp]
\caption{Regularized particle filter \cite{oudjane2000progressive}}
\label{alg discrete time regularized particle filters}
\begin{algorithmic}[1]
\STATE Input observations $\{Y(t_i)\}_{i\in\mathbb N}$, a dynamical model, and the initial distribution.
\STATE \textit{Initialization:} Sample $M$ particles $(\bar \kappa_1(0), \bar x_1(0)),\dots, (\bar \kappa_M(0), \bar x_M(0))$; set $i=1$ and $t_0=0$.
\WHILE{$t_i$ does not exceed the terminal time of the observations}
\STATE \textit{Sampling:} simulate $x_j(\cdot)$ ($j=1,\dots,M$) from time $t_{i-1}$ to $t_i$ according to the dynamical model with parameters $\bar \kappa_j(t_{i-1}) $ and initial conditions $\bar x_j (t_{i-1})$; Set $\kappa_j (t_{i})=\bar \kappa_j (t_{i-1})$; Calculate weights $w_j(t_i)\propto \ell_{Y^{N,\gamma}(t_i)}(\kappa_j(t_i), x_j(t_i))$.
\STATE \textit{Output the filter:} $\bar \pi_{M,t_i} (\phi) = {\sum_{j=1}^{M} w_j(t_i) \phi(\kappa_j, x_j(t_i))}$.
\STATE \textit{Resampling:} Resample
$\{ w_{j}(t_i), (\kappa_j(t_i), x_j(t_i))\}_j$ to obtain M equally weighted samples $\{{1}/{M}, (\hat \kappa_j(t_i), \bar x_j(t_i))\}_j$. (Residual resampling is applied)
\STATE \textit{Artificial evolution:} sample $\bar \kappa_j(t_i)$ from a kernel $\eta_{M}\left(\cdot | \hat \kappa_j(t_i) \right)$.
\ENDWHILE
\end{algorithmic}
\end{algorithm}
Conventionally, in the sampling step, particles are obtained by simulating the full dynamical model of the system.
However, it is not plausible for SRNs, especially the multiscale ones, where the exact simulation of \eqref{eq. scaling stochastic dynamics} can take an impractical amount of computational time.
Instead, for SRNs on different scales, we utilize different dynamical models in the sampling step so that the filter is computationally efficient.
The specific construction of such RPFs is as follows.
\begin{definition}[RPFs for SRNs on different scales] \label{def RPFs for SCRNs}
\begin{itemize}
\item If all reactions fire at the rate of $\mathcal O(1\sim 10)$ (i.e., all $N^{\gamma_1+ \rho_j}$ are $\mathcal O(1\sim 10)$), then we build the filter by utilizing exact simulation algorithms in the sampling step.
We denote such a RPF by $\bar \pi^{N,\gamma_1}_{M,t_i}$.
\item If the fastest reaction fires at the rate of $\mathcal O \left(10^{2}\sim 10^3\right)$, then we use the tau-leaping method in the sampling step.
We denote such a RPF by $\bar \pi^{N,\gamma_1,\tau}_{M,t_i}$.
\item If $N$ is extremely large, we use the PDMP in the sampling step. We denote such a RPF by $\bar \pi^{N,\gamma_1, H}_{M,t_i}$. (Here, $H$ stands for hybrid approximations. This filter has dependence on $N$ because the observation $Y^{N,\gamma_1}(\cdot)$ is input into the algorithm.)
\end{itemize}
\end{definition}
Another important factor for the RPF is the artificial noise used in the artificial evolution step.
Though this noise can mitigate sample degeneracy, it also introduces a bias to the estimate.
Therefore, if the kernel $\eta_{M}$ is not carefully chosen, the RPF can yield a very inaccurate estimate of the conditional distribution.
In this paper, we design the kernel $\eta_{M}$ according to the following assumptions so that the intensity of the artificial noise is contained.
\begin{assumption}{\label{cond. conditions on the artificial noise 1}}~
\begin{itemize}
\item $\mathcal K$'s prior distribution has a compact convex support. We denote it by $\Theta$.
\item We term $\mathcal V$ as the covariance matrix of $\mathcal K$ at the initial time. Then, for any $M>0$ and $\kappa\in\Theta$, the kernel $\eta_{M}\left(\cdot | \kappa\right)$ satisfies
\begin{equation*}
\eta_{M}\left( \tilde \kappa | \kappa\right) \propto \mathbbold 1_{\{\tilde \kappa\in \Theta \}} \mathcal N\left(\tilde \kappa \left| \kappa, C^2_{\eta} \mathcal V/M \right. \right)
\qquad \forall \tilde \kappa\in\mathbb R^{\tilde r},
\end{equation*}
where $C_{\eta}$ is a predetermined positive constant, and $\mathcal N\left(\cdot \left| \kappa, C^2_{\eta} \mathcal V/M \right. \right)$ is a Gaussian kernel with $\kappa$ mean and $C^2_{\eta} \mathcal V/M$ covariance matrix.
\end{itemize}
\end{assumption}
In general, the above assumption suggests that the artificial evolution keeps particles in the support of the prior distribution, and the artificial noise is in the order of $O\left( M^{-1/2}\right)$.
\begin{remark}{\label{remark artificial evolution}}
In references \cite{liu2001combined,liu2012state}, there is an alternative choice of the artificial noise where $\eta_{M}\left( \tilde \kappa | \kappa\right)=\mathcal N\left(\tilde \kappa \left| a\kappa+(1-a) \mu_t, (1-a) C^2_{\eta} \mathcal V_t/M \right. \right)$ with $\mu_t$ and $\mathcal V_t$ the mean and covariance matrix of particles $\{\hat k_j(t_i)\}_{1\leq j\leq M}$ and `$a$' a constant in $[0,1]$.
Compared with our setting, this one can adjust the artificial noise intensity dynamically according to the particle distribution, and, moreover, the associated filter can avoid ``variance inflation" \cite{liu2001combined}, meaning that the artificial evolution in its filter does not change the particle variance.
However, we think that variance inflation is beneficial, and with the method in \cite{liu2001combined,liu2012state}, the RPF can only delay sample degeneracy but never fully solve it.
The reasons are as follows.
With more observations collected, the exact posterior of the model parameter becomes narrower, and the variance of the particles corresponds to $\mathcal K$ inclines to decay dramatically over time.
Since the method in \cite{liu2001combined,liu2012state} avoids variance inflation, its particle variance decays strictly over time and finally degenerates to zero; in other words, the particles $\{ k_j(t_i)\}_{1\leq j\leq M}$ collapse into a single point in the end.
Ideally, the point should be close to the maximum of the posterior.
However, due to the randomness, this ideal case can rarely happen, and, consequently, sample degeneracy arises.
In contrast, the RPF satisfying \cref{cond. conditions on the artificial noise 1}
can greatly improve the particle variance (when it is too small) due to the constant covariance matrix of the artificial noise and, therefore, can fully solve the sample degeneracy problem.
Another fact to support our choice of artificial noise is that in reality the model parameters fluctuate dynamically in individual living cells due to environmental changes and the cell cycle \cite{rosenfeld2005gene}.
This requires the algorithm to generate diverse particles even if the exact posterior is very narrow.
Again, due to the constant covariance matrix of the artificial noise, our framework meets this requirement.
\end{remark}
\begin{remark}
Compared with our previous work \cite{fang2021stochastic,fang2020stochastic}, where an SIRPF was established for SRNs based on the PDMP, this paper constructs filters using one more reduced model, the tau-leaping method.
More importantly, the RPF constructed in this paper involves an artificial evolution step, which can mitigate sample degeneracy and make the filter perform better when the parameter uncertainty is large.
In the next section, we will illustrate this point using several numerical examples.
\end{remark}
\subsection{Convergence of RPFs}
In this subsection, we investigate conditions under which the established RPFs can provide reliable estimates of the hidden dynamic states and model parameters in the limit of large particles.
In the literature \cite{del2000branching,le2004stability,crisan2014particle}, a few conditions have been proposed to guarantee the convergence of RPFs; however, none of them can be directly applied to our setting for the following reasons.
\begin{remark}
It has been shown that an RPF converges to the exact filter if the artificial noise is weak and the transition kernel satisfies some regularity conditions, e.g., the globally Lipschitz continuity \cite[Proposition 2.38]{del2000branching}, the mixing condition in the metric space \cite{le2004stability}, and the Lipschitz continuity of transition kernel's derivatives \cite{crisan2014particle}.
For an SRN, the globally Lipschitz continuity of the transition kernel is very restrictive; it usually requires the propensities to be Lipschitz continuous (e.g., linear propensities), or the trajectory to evolve in a bounded region.
In addition, the discrete state space of an SRN invalidates the mixing condition in the whole metric space \cite{le2004stability} and the derivative related condition in \cite{crisan2014particle}.
\end{remark}
In this paper, we approach this problem in an alternative way.
Note that when the intensity of the artificial noise is zero, the constructed RPF degenerates to an SIRPF, and its convergence has been shown in part in our previous work \cite{fang2021stochastic}.
Therefore, the key to proving the convergence of the established RPFs lies in estimating the bias introduced by the artificial evolution and showing it to vanish as particle population tends to infinity.
This corresponds to calculating the parameter sensitivity for SRNs, as the artificial noise only perturbs the model parameter part of the particles.
Fortunately, neat expressions for this parameter sensitivity function have been provided in the literature \cite{gupta2013unbiased,gupta2019sensitivity,gupta2018estimation}, by which we can easily analyze the bias of interest.
Finally, through some technical analyses, we can show that both the parameter sensitivity and the state vector grow mildly under some weak conditions, and, consequently, the constructed RPF is convergent in probability to the exact filter.
More detailed discussions about it are presented in the Appendix.
We now specify a few assumptions needed in our main results.
In \Cref{remark covered SCRNs}, we illustrate that such conditions are so mild that they cover a large class of SRNs.
\begin{assumption}{\label{assumption for full model}
\begin{itemize}
\item All the propensities $\lambda^N_j (\kappa, x)$ are twice differentiable with respect to $\kappa$. Moreover, there exist positive constants $C_{\lambda,1}$ and $q$ such that
\begin{equation*}
\max\left\{
\left|\lambda^N_j (\kappa, x)\right|,
\left\|\frac{\partial \lambda^N_j (\kappa, x)}{\partial \kappa}\right\|_2,
\left\|\frac{\partial^2 \lambda^N_j (\kappa, x)}{\partial \kappa^2}\right\|_2
\right\} \leq C_{\lambda,1} \left(1+\left\|x\right\|_2^q\right)
\end{equation*}
for any $N>0$, $j\in\{1,\dots,r\}$, $\kappa\in \Theta$, and $x\in \mathbb{R}^{n}_{\geq 0}$.
\item For any $N>0$ and $j\in\{1,\dots, r\}$, if $\lambda^N_j (\kappa, x)>0$, then $x+\Lambda^{N} \zeta_j $ has no negative component.
\item For any $p>0$, there exists a constant $ C_{0,p}$ such that $\mathbb E \left[ \left\| X^{N}(0)\right\|_2^p \right]\leq C_{0,p}$
\item There exists a positive constant $C_{\lambda,2}$ such that for any $N$ and $j$ satisfying $ \bar 1^{\top} \Lambda^{N} \zeta_j >0$ ($\bar 1$ is the all-one vector) there holds
\begin{equation*}
\lambda^N_j (\kappa, x) \leq C_{\lambda,2}\left(1+\|x\|_2\right)\qquad \forall (\kappa,x)\in \Theta\times \mathbb R^{n}_{\geq 0}.
\end{equation*}
In other words, the propensity of the reaction that leads to an increase of the total population has at most a linear growth rate with the state argument.
\end{itemize}
\end{assumption}
\begin{assumption}{\label{assumption for PDMPs}}
For any reaction $j$ and parameter $\kappa\in\Theta$, the convergence $\lambda^N_j(\kappa, \cdot) \to \lambda'_j(\kappa, \cdot) $ (as $N\to\infty$) holds uniformly over compact sets. The same is true for derivatives $\frac{\partial \lambda^N_j (\kappa, \cdot)}{\partial \kappa} \to \frac{\partial \lambda'_j (\kappa, \cdot)}{\partial \kappa}$ and $\frac{\partial \lambda^N_j (\kappa, \cdot)}{\partial x} \to \frac{\partial \lambda'_j (\kappa, \cdot)}{\partial x}$.
\end{assumption}
Then, we present the main result of this paper, which states that the constructed RPFs can be very close to the exact filter when the particle population is large.
\begin{theorem}{\label{thm main}}
Assume that \Cref{cond. conditions on the artificial noise 1} and \Cref{assumption for full model} hold.
Also, we suppose the test function $\phi:\Theta\times \mathbb R^{n}_{\geq 0} \to \mathbb R$ to be bounded, continuous, and continuously differentiable in the first $m$-coordinates and satisfy
\begin{equation}\label{eq. sensitivity condition of phi}
\left\|\frac{\partial \phi (\kappa, x)}{\partial \kappa}\right\|_2 \leq C_{\phi} \left(1+\left\|x\right\|_2^q\right)
\qquad \forall (\kappa,x)\in \Theta \times \mathbb R^{n}_{\geq 0},
\end{equation}
where $C_{\phi}$ is a predetermined constant, and $q$ is the same as that in \Cref{assumption for full model}.
Then, the following results for the constructed RPFs hold for any $i\in\mathbb N_{>0}$.
\begin{itemize}
\item
$ \lim_{M\to\infty}\mathbb E \left[\left|\bar \pi^{N,\gamma_1}_{M,t_i}(\phi)- \pi^{N,\gamma_1}_{t_i}(\phi) \right| ^2 \right] =0$
for any scaling factor $N$, where $\bar \pi^{N,\gamma_1}_{M,t_i}(\phi)$ is the filter built on exact simulation algorithms, and $\pi^{N,\gamma_1}_{t_i}(\phi)$ is the exact filter.
\item If, moreover, \Cref{assumption non-negativity of tau-leaping algorithm} and \Cref{assumption convergence of tau-leaping algorithm} hold, then there is the relation
$ \lim_{M\to\infty} \lim_{|\tau|\to 0} \mathbb E \left[ \left|\bar \pi^{N,\gamma_1,\tau}_{M,t_i}(\phi)- \pi^{N,\gamma_1}_{t_i}(\phi) \right|^2 \right] =0$
for any $N$, where $\bar \pi^{N,\gamma_1,\tau}_{M,t_i}(\phi)$ is the filter built on the tau-leaping method.
\item If \Cref{assumption convergence of the initial condition} and \Cref{assumption for PDMPs} also hold, then there is the relation
$ \lim_{M\to\infty} \lim_{N\to +\infty} \mathbb E \left[\left|\bar \pi^{N,\gamma_1,H}_{M,t_i}(\phi)- \pi^{N,\gamma_1}_{t_i}(\phi) \right|^2\right] =0$,
where $\bar \pi^{N,\gamma_1,H}_{M,t_i}(\phi)$ is the filter built on the PDMP model.
\end{itemize}
\end{theorem}
\begin{proof}
The proof is shown in the {appendix}.
\end{proof}
\begin{remark}{\label{remark covered SCRNs}}
The conditions required in the above theorem are not restrictive.
\Cref{assumption non-negativity of tau-leaping algorithm}, \Cref{assumption convergence of tau-leaping algorithm}, and \Cref{cond. conditions on the artificial noise 1} can be easily achieved by properly designing the tau-leaping algorithm and the RPF.
\Cref{assumption convergence of the initial condition} and \Cref{assumption for PDMPs} are concerned with the well-definiteness of the PDMP, which are reasonable to assume for those systems where the PDMP has already been successfully applied for simulation.
For \Cref{assumption for full model}, the first claim (the polynomial growth rate of the propensity) is satisfied in most SRNs considered in the existing literature (following the mass-action kinetics or the Hill-type dynamics); the second claim is nothing but ensuring the state vector to be nonnegative;
the third requires all moments of initial conditions to be finite, which is natural in biology.
The last claim in \Cref{assumption for full model} indicates that all reactions that lead to an increase of the molecular population have linearly growing propensities.
Though the last one is not as weak as other assumptions, it still includes a lot of SRNs in systems biology, e.g., the simple birth-death model, the antithetic integral controller that interconnects to linear networks (see \cite{briat2016antithetic,aoki2019universal} for antithetic integral controllers), and all mass-action networks where only unimolecular reactions can lead to an increase of the total molecular population.
In summary, the conditions required in \Cref{thm main} are quite mild, and the proposed RPFs work for a large class of SRNs.
\end{remark}
\subsection{Hyperparameter tuning}
In the finite particle population case, the hyperparameter $C_{\eta}$, which depicts the intensity of the noise in artificial evolution (see \Cref{cond. conditions on the artificial noise 1}), has a significant influence on the performance of the established filter and, therefore, needs to be properly trained in advance.
Intuitively, this hyperparameter tuning task is an exploration-exploitation trade-off: by increasing the hyperparameter $C_\eta$, the RPF can explore a broad region of the parameter space and have a higher chance to discover all regions of high posterior, but at the cost of missing some knowledge conveyed in the current particles.
To optimize the hyperparameter, we generate a large number of simulated trajectories, $\{(\tilde \kappa_{\ell}, \tilde x_\ell(\cdot),\tilde y_\ell(\cdot))\}$, where $(\tilde \kappa_{\ell}, \tilde x_\ell(0))$ are sampled from the prior of $\left(\mathcal K, X^{N}(0)\right)$, $x_\ell(\cdot)$ is the simulated trajectory of the considered SRN, and $y_\ell(\cdot)$ is the simulated observation.
A good $C_{\eta}$ should make the associated particle filter close to the exact filter when applied to the simulated data.
However, this criterion cannot be directly used to optimize $C_{\eta}$ as the exact filter is unavailable.
Alternatively, we train the hyperparameter by the accuracy of the inferred parameters and the diversity of particles.
This strategy first requires the parameter identifiability assumption meaning that given long-time observations, the posterior of the model parameter is densely distributed around the true value.
This assumption is not very restrictive; usually, it can be checked by the frequency spectrum \cite{song2019frequency} and achieved by optimal experimental design \cite{faller2003simulation}.
According to the parameter identifiability, the accuracy of the RPF can be evaluated by the distance between the inferred model parameter (by the RPF) and the exact parameter.
As the true value of the model parameter is known in the simulated data, we can easily apply this rule to training the filter.
In addition, particle diversity in model parameters should also be considered in hyperparameter tuning, and we evaluate it by the entropy of the particles ($\{ \kappa_j\}_{j=1,\dots, M}$).
To describe particle diversity more accurately, we truncate the value of particles to four decimal places, so that two particles in close distance are regarded as identical ones.
In conclusion, the optimal hyperparameter, denoted by $C^*_{\eta}$, which balances accuracy and diversity, can be obtained by solving the following optimization problem:
\small{
\begin{align}\label{eq. performance of C eta}
&C^*_{\eta} = \\
& \notag \mathop{\arg\min}\limits_{C_{\eta}>0}
\sum_\ell \mathbb E \left[\left. {\bar \pi_{M,t_L,C_{\eta}}\left( \frac{\left\|\kappa-\tilde \kappa_\ell\right\|^2_2}{ \left\| \tilde \kappa_\ell \right\|^2_2} \right)}+\frac{1}{\exp\left( \mathcal{E}_{\pi_{M,t_L,C_{\eta}}}\right) }\right|~ Y^{N,\gamma_1}(t_i) = \tilde y_{\ell}(t_i), ~ \forall t_i \leq t_{L} \right],
\end{align}}where $t_L$ is a large time point, $\bar \pi_{M,t_L,C_{\eta}}$ represents the RPF established in \Cref{def RPFs for SCRNs} with the hyperparameter $C_{\eta}$, and $\mathcal{E}_{\pi_{M,t_L,C_{\eta}}}$ is the entropy of the particles $\{\bar \kappa_j\}_{j=1,\dots,M}$ at the time $t_L$.
In this formula, the term $\bar \pi_{M,t_L,C_{\eta}}\left( \left\|\kappa-\tilde \kappa_\ell\right\|^2_2/ \left\| \tilde \kappa_\ell \right\|^2_2 \right)$ evaluates the relative error between the inferred model parameter and the exact one in the simulated data, and $\exp( \mathcal{E}_{\pi_{M,t_L,C_{\eta}}})$ evaluates the particle diversity.
Note that this optimization problem is nonconvex; to ease the computational burden, we can find an approximation to $C^*_{\eta}$ by optimizing the above objective function on a finite-size grid.
\begin{remark}
Even though we only search for the optimal hyperparameter in a finite set, the above optimization problem still requires a lot of computational resources, because for each $(\tilde \kappa_{\ell}, \tilde x_\ell(\cdot),\tilde y_\ell(\cdot))$ it needs to generate many simulated trajectories to calculate the filter.
We argue that these computational resources spent on hyperparameter training are worthwhile.
First, the optimal hyperparameter can make the filter more accurate than that using a random $C_{\eta}$.
Second, the training is performed off-line, so it will not increase the computational burden of the RPF when solving a filtering problem online.
\end{remark}
\section{Numerical examples}{\label{Sec. numerical examples}}
In this section, we illustrate our approach using several numerical examples on different scales and show that the proposed regularized particle filter can significantly outperform the SIR counterpart.
The code applied to perform the analysis is available on GitHub: ``https://github.com/ZhouFang92/Regularized-particle-filter-for-stochastic-reaction-networks".
\subsection{Antithetic integral feedback controller}
The antithetic integral feedback (AIF) controller is a biological module that can achieve robust perfect adaptation for output species in arbitrary noisy bimolecular networks (see \cite{aoki2019universal,briat2016antithetic}).
In this example, we consider such a gene circuit (see \Cref{fig aic}) that has six reactions (see \Cref{fig aic_chemical_equations}) and four species: the control input species (denoted by $S_1$), the sensing species ($S_2$), the regulated species $(S_3)$, and the report species $(S_4)$.
Here, the dynamics are assumed to follow mass-action kinetics.
The first three reactions in \Cref{fig aic_chemical_equations} result in an ``integral controller" where the integrator is $\mathbb E\left[X_1(t)-X_2(t)\right]$ and ensures that the mean copy number of $S_3$ at the stationary distribution (if it exists) is ${k_1}/{k_2}$ regardless of any other parameters \cite{briat2016antithetic}.
In an \textit{E. coli} cell, such a network can be realized by choosing the sigma factor and anti-sigma factor as $S_1$ and $S_2$, respectively \cite{aoki2019universal}.
Here, we consider a filtering problem for this system in which $S_4$ is a fluorescent reporter and measured under a microscope, and our goal is to infer the dynamics of the regulated species ($S_3$) and its set point $k_1/k_2$.
\begin{figure}[htbp]
\centering
\subfigure[Gene circuit: $S_4$ is recorded.]{
\centering
\includegraphics[width= 0.45\textwidth]{figure/AIC_four_species.pdf}
\label{fig aic}
}
\subfigure[Chemical reactions]{
\vbox{\hsize=0.45\hsize
\begin{align*}
&\emptyset \stackrel{k_1}{\longrightarrow} S_1 && \text{reference} \\
&S_3 \stackrel{k_2}{\longrightarrow} S_2+ S_3 && \text{measurement} \\
&S_1+S_2 \stackrel{k_3}{\longrightarrow} \emptyset && \text{comparison} \\
&S_1 \stackrel{k_4}{\longrightarrow} S_1+S_3 +S_4 && \text{actuation} \\
&S_3 \stackrel{k_5}{\longrightarrow} \emptyset && \text{degradation}\\
&S_4 \stackrel{k_6}{\longrightarrow} \emptyset && \text{degradation}
\end{align*}}
\label{fig aic_chemical_equations}
}
\caption{{\bf Antithetic integral feedback controller.} (a) The diagram of an antithetic integral feedback controller where $S_1$ and $S_2$ form an integral controller with the integrator $\mathbb E\left[X_1(t)-X_2(t)\right]$ ensuring the mean copy number of $S_3$ to be constant at the stationary distribution.
$S_4$ is the report species that is visible and measured by a microscope.
(b) The chemical reactions contained in the antithetic integral feedback controller. Its dynamics is assumed to follow mass-action kinetics with reaction constants $k_i$ ($i=1,\dots,6$).
With these parameters, the set point of the regulated species $S_3$ is $k_1/k_2$.
}
\label{fig. AIC_four_species}
\end{figure}
\begin{table}[htbp]
\centering
\begin{tabular}{|l|l|l|l|}
\hline
\multicolumn{1}{|c|}{Model parameter} & \multicolumn{1}{|c|}{Exponent} & \multicolumn{1}{|c|}{Rescaled parameter} & \multicolumn{1}{|c|}{Initial condition}\\
\multicolumn{1}{|c|}{$\left(\text{min}^{-1}\right)$} & & \multicolumn{1}{|c|}{$\left(\text{min}^{-1}\right)$} &
\\
\hline
$k_1\sim \mathcal U (0.1, 1)$ & $\beta_1=0$ & $k'_1\sim\mathcal U (0.1, 1)$ &
$X^N_1(0) \sim \mathcal{UI}(0,10)$~~~~ \\
$k_2\sim \mathcal U (0.1, 1)$ & $\beta_2=0$ & $k'_2\sim\mathcal U (0.1, 1)$ &
$X^N_2(0)\sim \mathcal{UI}(0,10)$~~~~ \\
$k_3=0.2$~~ & $\beta_3=0$ & $k'_3=0.2$ &
$X^N_3(0)\sim \mathcal{UI}(0,20)$~~~~ \\
$k_4=0.5$ & $\beta_4=0$ & $k'_4=0.5$ &
$X^N_4(0)\sim \mathcal{UI}(0,20)$~~~~ \\
$k_5=0.7$ & $\beta_5=0$ & $k'_5=0.7$ & \\
$k_6=0.3$ & $\beta_6=0$ & $k'_6=0.3$ & \\
\hline
\end{tabular}
\caption{\noindent {\bf Model parameters of the AIF controller.} $\mathcal{U}$ is the uniform distribution, and $\mathcal{UI}$ is the integer uniform distribution.}
\label{table ex_AIC model_four_species parameters}
\end{table}
{
\begin{figure}[htbp]
\centering
\subfigure[Particle filters for the AIF controller.]{
\centering
\includegraphics[width= 0.9 \textwidth]{figure/aic_4_trajectory.pdf}
\label{fig aic_trajectory}
}
\subfigure[The distribution of particles at the final time.]{
\centering
\includegraphics[height= 0.2\textheight]{figure/aic_4_final_particles.pdf}
\label{fig aic_final_particles}
}
\hfill
\subfigure[The evolution of particles over time.]{
\centering
\includegraphics[height= 0.2 \textheight]{figure/aic_4_particle_evolution.pdf}
\label{fig aic_particle_evolution}
}
\caption{{\bf Numerical results of the AIF controller.} ``RPF" stands for the regularized particle filter, and ``SIRPF" stands for the sequential importance resampling particle filter.
Panel (a) shows the performance of both filters in estimating the regulated species $S_3$ and its set point $k_1/k_2$.
It tells that both filters can infer the dynamics of $S_3$ accurately, but the RPF has a much better performance in estimating the set point and model parameters.
Panel (b) presents the final time distribution of the particles associated with these filters in the parameter coordinate, and panel (c) shows the evolution of these particles over time.
These two panels tell that sample degeneracy happens to the SIRPF very early in time (around 48 minutes), and, therefore, the SIRPF cannot adjust its estimate of model parameters according to the new observations afterward.
This is the reason why the SIRPF fails to infer model parameters accurately.
In contrast, the particle diversity in the RPF is always large, and this enables the RPF to constantly refine its estimate of model parameters according to new observations.
Consequently, the RPF can infer model parameters and the set point accurately.
}
\label{fig aic_result}
\end{figure}
\begin{table}[htbp]
\centering
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline
$C_{\eta}$ & 0 & 0.2 & 0.4 & 0.6 & 0.8 & 1.0 & 1.2 & 1.4 & 1.6 & 1.8* & 2\\
\hline
Score & 1.00 & 0.14& 0.11& 0.11& 0.11 & 0.10& 0.09& 0.09& 0.10& 0.08*& 0.09\\
\hline
\end{tabular}
\caption{\noindent {\bf The performance of different $C_{\eta}$ (the AIF controller):} $t_L=240 ~(\rm{mins})$.}
\label{table exAIC_four_species hyperparameters}
\end{table}
}
In this example, we set the scaling factor ($N$) to be 100 and the abundance coefficients ($\alpha_1$, $\alpha_2$, $\alpha_3$, and $\alpha_4$) to be zero, meaning that the cell system has very few copies of the involved species.
Also, we assume the model parameters to satisfy \Cref{table ex_AIC model_four_species parameters}, where all the rate constants are known except for $k_1$ and $k_2$.
The observations are collected every 2 minutes and satisfy
\begin{equation*}
Y^{\gamma_1}(t_i) = 10 X_4(t_i) \wedge 1000 + W(t_i) \qquad i\in \mathbb N_{>0},
\end{equation*}
in which $t_i=2i$, and $\{W(t_i)\}_{i\in\mathbb N_{>0}}$ is a sequence of mutually independent standard Gaussian random variables.
In this setting, we can calculate that $\gamma_1=0$, and the most proper RPF to solve this filtering problem is the one utilizing the exact simulation algorithm in the sampling step (i.e., the first filter in \Cref{def RPFs for SCRNs}).
We set the particle population of the filter to be 2000 and trained the RPF following the rule \eqref{eq. performance of C eta}; see \Cref{table exAIC_four_species hyperparameters} for the training result.
Finally, we applied the RPF with the optimal hyperparameter to a simulated process and compared its result with the SIRPF; the numerical results are shown in \Cref{fig aic_result}.
\Cref{fig aic_trajectory} shows that both filters can infer the dynamics of the regulated species accurately; however, the RPF has a superior performance in estimating the model parameters and the set point
thanks to the larger particle diversity of the RPF (\Cref{fig aic_final_particles} and \Cref{fig aic_particle_evolution} ).
Moreover, the sample degeneracy of the SIRPF occurs very early in time (\Cref{fig aic_particle_evolution}), suggesting that the SIRPF fails to represent the posterior of the model parameters very quickly after the starting time.
All these facts suggest that the RPF outperforms the SIRPF in this numerical example.
\subsection{Gene transcription model}\label{subsection gene transcription model}
We then consider a gene transcription model (\Cref{fig. GT}) consisting of four reactions (see \Cref{fig GT_chemical_equations}) and three species: the inactivated gene (denoted by $S_1$), the activated gene ($S_2$), and the transcribed RNA ($S_3$).
Moreover, we assume that the RNA contains stem-loops recognized and bound by fluorescent reporters so that the produced RNAs can be visualized in a living cell.
This model resembles the gene circuit constructed in \cite{rullan2018optogenetic} for investigating and controlling transcriptional dynamics.
In this example, our goal is to infer both hidden dynamic states and model parameters from the history of the RNA signal.
\begin{figure}[htbp]
\centering
\subfigure[The circuit: the RNA is recorded.]{
\centering
\includegraphics[width= 0.38\textwidth]{figure/GT.pdf}
\label{fig GT circuit}
}
\hfill
\subfigure[Chemical reactions]{
\vbox{\hsize=0.5\hsize
\begin{align*}
&S_1 \stackrel{k_1}{\longrightarrow} S_2 && \text{gene activation} \\
&S_2 \stackrel{k_2}{\longrightarrow} S_1 && \text{gene inactivation} \\
&S_2 \stackrel{k_3}{\longrightarrow} S_2+S_3 && \text{transcription} \\
&S_3 \stackrel{k_4}{\longrightarrow} \emptyset && \text{degradation}
\end{align*}}
\label{fig GT_chemical_equations}
}
\caption{{\bf Gene transcription model.} (a) The diagram of a gene transcription circuit where the gene has ``on" and ``off" states and transcribes message RNAs only in the ``on" state.
The message RNAs are visible, and their light intensity signal is measured by a microscope.
(b) The chemical reactions contained in the gene transcription model, where $S_1$, $S_2$, and $S_3$ are the inactivated gene, the activated gene, and the message RNA, respectively. All the reactions follow mass-action kinetics with reaction constants $k_i$ ($i=1,\dots,4$).
}
\label{fig. GT}
\end{figure}
\begin{table}[htbp]
\centering
\begin{tabular}{|l|c|c|l|}
\hline
\multicolumn{1}{|c|}{Model parameter} & Exponent & Rescaled parameter & \multicolumn{1}{|c|}{Initial condition}\\
$\left(\text{min}^{-1}\right)$ & & $\left(\text{min}^{-1}\right)$ &
\\
\hline
$k_1\sim \mathcal U (0.03, 0.3)$ & $\beta_1=0$ & $k'_1\sim\mathcal U (0.03, 0.3)$ &
$X^N_1(0) \sim$ Bern($1/2$)~~~~ \\
$k_2\sim \mathcal U (0.03, 0.3)$ & $\beta_2=0$ & $k'_2\sim\mathcal U (0.03, 0.3)$ &
$X^N_2(0)=1-X^N_1(0)$~~~~ \\
$k_3\sim \mathcal U (10, 100)$~~ & $\beta_3=1$ & $k'_3\sim\mathcal U (0.2, 2)$~~~~ &
$X^N_3(0) \sim \frac{X^N_2(0)\text{Poiss(50)}}{50}$ \\
$k_4\sim \mathcal U (0.2, 2)$~~~~ & $\beta_4=0$ & $k'_4\sim\mathcal U (0.2, 2)$~~~~ & \\
\hline
\end{tabular}
\caption{\noindent {\bf Model parameters of the gene transcription model.} $\mathcal U$ is the uniform distribution, ``{\rm{Bern($\cdot$)}}" is the Bernoulli distribution, and ``{\rm{Poiss}}" is the Poisson distribution.}
\label{table exGT_tau-leaping model parameters}
\end{table}
\begin{figure}[htbp]
\centering
\subfigure[Particle filters for the gene transcription model.]{
\centering
\includegraphics[width= 0.9 \textwidth]{figure/GT_trajectory.pdf}
\label{fig GT_trajectory}
}
\subfigure[The distribution of particles at the final time.]{
\centering
\includegraphics[height= 0.2\textheight]{figure/GT_final_particles.pdf}
\label{fig GT_final_particles}
}
\hfill
\subfigure[The evolution of particles over time.]{
\centering
\includegraphics[height= 0.2 \textheight]{figure/GT_particle_evolution.pdf}
\label{fig GT_particle_evolution}
}
\caption{{\bf Numerical results of the gene transcription model.} ``RPF" stands for the regularized particle filter, and ``SIRPF" stands for the sequential importance resampling particle filter.
Panel (a) shows the performance of two filters in estimating the system states and model parameters.
Specifically, both filters infer the dynamics of DNAs and RNAs accurately, but the RPF has a better performance in estimating model parameters.
Panel (b) presents the final time distribution of the particles associated with these filters in the parameter coordinate, and panel (c) shows the time evolution of these particles.
Similar to the previous example, these two panels tell that sample degeneracy happens to the SIRPF very early in time (before 48 minutes), which disables the SIRPF to adjust its estimate of model parameters according to the new observations afterward.
In contrast, the RPF always has a large particle diversity and is able to refine its estimate of model parameters over time.
This explains why the RPF outperforms the SIRPF in parameter inference.
However, note that the RPF does not infer every parameter accurately; its estimate of $k'_2$ somehow deviates from the exact value (though the deviation is not too large).
}
\label{fig GT_result}
\end{figure}
\begin{table}[htbp]
\centering
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline
$C_{\eta}$ & 0 & 0.1 & 0.2 & 0.3 & 0.4 & 0.5 & 0.6 & 0.7 & 0.8 & 0.9* & 1\\
\hline
Score & 1.08 & 0.17& 0.13& 0.10& 0.10& 0.09 & 0.09& 0.09& 0.09& 0.08*& 0.09\\
\hline
\end{tabular}
\caption{\noindent {\bf The performance of different $C_{\eta}$ (the gene transcription model):} $t_L=240 ~(\rm{mins})$.}
\label{table exGT_tau-leaping hyperparameters}
\end{table}
For the considered gene transcription model, we set $N=50$, $\alpha_1=\alpha_2=0$, and $\alpha_3=1$, meaning that the system contains very few copies of DNAs but tens of mRNA copies.
Also, we assume the dynamics of the system to follow mass-action kinetics and the model parameters to satisfy \Cref{table exGT_tau-leaping model parameters} where all reaction constants are unknown.
The observations are measured every 0.5 minute and satisfy
\begin{equation*}
Y^{\gamma_1}(t_i) = 20 X^{N,\gamma_1}_3(t_i) \wedge 1000 + W(t_i) \qquad i\in \mathbb N_{>0},
\end{equation*}
where $\gamma_1=0$, $t_i=0.5i$, 1000 is the measurement range, and $\{W(t_i)\}_{i\in\mathbb N_{>0}}$ is a sequence of mutually independent standard Gaussian random variables.
Here, we utilized the tau-leaping based RPF to solve the associated filtering problem.
Again, we set the particle population to be 2000 and trained the filter following the rule \eqref{eq. performance of C eta}; see \Cref{table exGT_tau-leaping hyperparameters} for the training results.
Finally, we applied the filter to a simulated process, whose results are present in \Cref{fig GT_result}.
Similar to the conclusion obtained in the previous subsection, \Cref{fig GT_result} shows that the RPF has a superior performance in estimating model parameters due to its larger particle diversity, and the sample degeneracy of the SIRPF occurs very early in time.
These facts imply that RPF outperforms the SIRPF in this example.
However, we note that the RPF does not infer every parameter accurately (see $k'_2$ in \Cref{fig GT_final_particles}); this might be due to the fact that particle population is not large enough in this example.
Also, both filters have similar performances in estimating the gene dynamics (\Cref{fig GT_trajectory}) due to the simple structure of the system: once the observation increases/decreases sharply, the filter can learn that the gene is turned on/off whatever its estimates of model parameters.
\subsection{Transcription regulation network}\label{subsection transcription regulation network}
Finally, we consider a transcription regulation network (\Cref{fig. TR}) where the gene product inhibits the gene expression process by binding to the DNA molecule.
This system consists of six reactions (see \Cref{fig TR_chemical_equations}) and four chemical species: the inactivated gene (denoted by $S_1$), the activated gene ($S_2$), the mRNA ($S_3$), and the protein ($S_4$).
Also, we assume that the protein is fluorescent and, therefore, can be measured by a microscope.
In this example, our goal is to infer both the hidden dynamic states and model parameters from the time-course data.
\begin{figure}[htbp]
\centering
\subfigure[The circuit: the protein is recorded.]{
\centering
\includegraphics[width= 0.45\textwidth]{figure/TR.pdf}
\label{fig TR circuit}
}
\hfill
\subfigure[Chemical reactions]{
\vbox{\hsize=0.45\hsize
\begin{align*}
&S_2+S_4 \stackrel{k_1}{\longrightarrow} S_1 && \text{gene inactivation} \\
&S_1 \stackrel{k_2}{\longrightarrow} S_2+S_4 && \text{gene activation} \\
&S_2 \stackrel{k_3}{\longrightarrow} S_2+S_3 && \text{transcription} \\
&S_3 \stackrel{k_4}{\longrightarrow} \emptyset && \text{degradation} \\
&S_3 \stackrel{k_5}{\longrightarrow} S_3+S_4 && \text{translation} \\
&S_4 \stackrel{k_6}{\longrightarrow} \emptyset && \text{degradation}
\end{align*}}
\label{fig TR_chemical_equations}
}
\caption{{\bf Transcription regulation network.} (a) The diagram of a transcription regulation network, where the gene product (the protein) inhibits transcription by silencing the gene. The protein contains a fluorescent tag so that it is visible and measured by a microscope.
(b) The chemical reactions contained in the transcription regulation network, where $S_1$, $S_2$, $S_3$, and $S_4$ are the inactivated gene, the activated gene, the mRNA, and the protein, respectively.
All the reactions follow the mass-action kinetics with reaction constants $k_i$ ($i=1,\dots, 6$).
}
\label{fig. TR}
\end{figure}
\begin{table}[h]
\centering
\begin{tabular}{|l|c|c|l|}
\hline
\multicolumn{1}{|c|}{Model parameter} & Exponent & Rescaled parameter & \multicolumn{1}{|c|}{Initial condition}\\
$\left(\text{min}^{-1}\right)$ & & $\left(\text{min}^{-1}\right)$ &
\\
\hline
$k_1\sim \mathcal U (0.03, 0.3)$ & $\beta_1=0$ & $k'_1\sim\mathcal U (0.03, 0.3)$ &
$X^N_1(0) \sim$ Bern($1/2$)~~~~ \\
$k_2\sim \mathcal U (0.03, 0.3)$ & $\beta_2=0$ & $k'_2\sim\mathcal U (0.03, 0.3)$ &
$X^N_2(0)=1-X^N_1(0)$ \\
$k_3\sim \mathcal U (0.1, 1)$~~ & $\beta_3=0$ & $k'_3\sim\mathcal U (0.1, 1)$~~~~ &
$X^N_3(0) \sim \text{Poiss(10)}$ \\
$k_4\sim \mathcal U (0.1, 1)$~~~~ & $\beta_4=0$ & $k'_4\sim\mathcal U (0.1, 1)$~~~~ &
$X^N_4(0) \sim \frac{\text{Poiss}(10^4)}{10^4}$\\
$k_5\sim \mathcal U (10^3, 10^4)$~~~~ & $\beta_5=1$ & $k'_5\sim\mathcal U (0.1, 1)$~~~~ & \\
$k_6\sim \mathcal U (0.1, 1)$~~~~ & $\beta_6=0$ & $k'_6\sim\mathcal U (0.1, 1)$~~~~ & \\
\hline
\end{tabular}
\caption{\noindent {\bf Model parameters of the transcription regulation model}: $\mathcal U$ is the uniform distribution, ``Bern($\cdot$)" is the Bernoulli distribution, and ``Poiss" is the Poisson distribution.}
\label{table TR model parameters}
\end{table}
\begin{table}[htbp]
\centering
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline
$C_{\eta}$ & 0 & 0.2 & 0.4 & 0.6 & 0.8 & 1.0 & 1.2 & 1.4 & 1.6 & 1.8* & 2\\
\hline
Score & 1.29 & 0.41& 0.37& 0.34& 0.37& 0.36 & 0.36& 0.36& 0.35& 0.33*& 0.36\\
\hline
\end{tabular}
\caption{\noindent {\bf The performance of different $C_{\eta}$ (the transcription regulation network):} $t_L=240 ~(\rm{mins})$.}
\label{table exTR_PDMP hyperparameters}
\end{table}
\begin{figure}[htbp]
\centering
\subfigure[Particle filters for the gene transcription model.]{
\centering
\includegraphics[width= 0.9 \textwidth]{figure/TR_trajectory.pdf}
\label{fig TR_trajectory}
}
\subfigure[The distribution of particles at the final time.]{
\centering
\includegraphics[width= 0.45\textwidth]{figure/TR_final_particles.pdf}
\label{fig TR_final_particles}
}
\hfill
\subfigure[The evolution of particles over time.]{
\centering
\includegraphics[width= 0.4\textwidth]{figure/TR_particle_evolution.pdf}
\label{fig TR_particle_evolution}
}
\caption{{\bf Numerical results of the gene transcription model.} ``RPF" stands for the regularized particle filter, and ``SIRPF" stands for the sequential importance resampling particle filter.
Panel (a) shows the performance of both filters in estimating the hidden system states and model parameters.
Specifically, RPF has a superior performance in estimating the dynamics of mRNAs and model parameters. (Note that the SIRPF underestimates the mRNA dynamics many times around 200 minutes.)
However, both filters fail to infer the DNA states accurately.
Panel (b) shows the final time distribution of the particles associated with these filters in the parameter coordinate, and panel (c) shows the time evolution of these particles.
Similar to the previous examples, these two panels tell that sample degeneracy happens to the SIRPF very early in time (before 48 minutes), which disables the SIRPF to adjust its estimate of model parameters according to the new observations afterward.
In contrast, the RPF always has a large particle diversity and is enable to refine its estimate of model parameters to approach the exact value.
This is the reason why the RPF outperforms the SIRPF in estimating model parameters.
Also, note that the RPF does not infer every parameter very accurately; its estimate of $k'_2$ (the reaction constant for gene activation) somehow deviates from the exact value.
}
\label{fig TR_result}
\end{figure}
The modeling of the systems is as follows. We set $N=10^4$, $\alpha_1=\alpha_2=\alpha_3=0$, and $\alpha_4=1$; in other words, the system has very few copies of DNA molecules and mRNA molecules but a large number of proteins.
Also, we assume that the system dynamics to follow mass-action kinetics and the model parameters to satisfy \Cref{table TR model parameters}.
Based on it, we can easily calculate that the first timescale $\gamma_1$ equals to 0.
The observations are collected every 2 minutes and satisfy
\begin{equation*}
Y^{\gamma_1}(t_i) = 20 X^{N,\gamma_1}_4(t_i) \wedge 1000 + W(t_i) \qquad i\in \mathbb N_{>0},
\end{equation*}
where $t_i=2i$, 1000 is the measurement range, and $\{W(t_i)\}_{i\in\mathbb N_{>0}}$ is a sequence of mutually independent standard Gaussian random variables.
In this setting, the most proper filter for the considered system is the RPF using PDMPs in sampling (the third one in \Cref{def RPFs for SCRNs}).
Again, we set the particle population to be 2000 and trained the filter following the rule \eqref{eq. performance of C eta}; see \Cref{table exTR_PDMP hyperparameters} for the training results.
Finally, we applied the filter to a simulated process, whose results are present in \Cref{fig TR_result}.
Similar to the previous examples, the RPF outperforms the SIRPF in this example in estimating model parameters and the mRNA dynamics\footnote{The SIRPF underestimates the mRNA dynamics for many times; see the result around 200 minutes.} due to its large particle diversity (\Cref{fig TR_result}).
Moreover, even though the initial estimate is not very accurate, the particles in the RPF can gradually approach the exact value of the model parameters and, consequently, provide an accurate estimate of the model parameters at the final time (\Cref{fig TR_particle_evolution}).
Despite these successes, we can observe that both filters fail to precisely estimate the gene state (\Cref{fig TR_trajectory}).
This is attributed to the long time delay between the gene dynamics and the protein dynamics, which makes the exact posterior of the gene state very noisy.
Inspired by this observation, we argue that given an arbitrary gene circuit, it is not reasonable to expect the filter to infer every species accurately, especially those to whose changes the fluorescent reporters respond slowly.
\section{Preliminary}\label{sec preliminary}
\subsection{Notations}
In this paper, we denote the natural filtered probability space by $( \Omega, \mathcal F, \{\mathcal F_{t}\}_{t\geq 0}, \mathbb P)$, where $\Omega$ is the sample space, $\mathcal F$ is the $\sigma$-algebra, $\{\mathcal F_{t}\}_{t\geq 0}$ is the filtration, and $\mathbb P$ is the natural probability.
Also, we term $\mathbb N_{>0}$ as the set of positive integers, $\mathbb R^{n}$ with $n$ being a positive integer as the space of $n$-dimensional real vectors, $\|\cdot\|_2$ as the Euclidean norm, and $|\cdot|$ as the absolute value notation.
For any positive integer $n$ and any $t>0$, we term $\mathbb D_{ \mathbb R^{ n}}[0,t]$ as the Skorokhod space that consists of all $\mathbb R^n$ valued cadlag functions on $[0,t]$.
\subsection{Stochastic reaction networks}
In this paper, we consider an intracellular reaction system that consists of $n$ chemical species ($S_1, \dots ,S_n$)
and $r$ reactions:
\begin{equation*}
v_{1,j} S_1 + \dots + v_{n,j} S_n \ce{ ->} v'_{1,j} S_1 + \dots + v'_{n,j} S_n, \quad j=1,\dots, r,
\end{equation*}
where $v_{i,j}$ and $v_{i,j}$ are stoichiometric coefficients representing numbers of molecules consumed and produced in each reaction.
We term $X(t)=\left(X_1(t),X_2(t), \dots, X_n(t)\right)^{\top}$ as the vector of copy numbers of these species at time $t$.
Due to the small number of reactant molecules, an intracellular reaction system is usually model by a stochastic model, called the continuous-time Markov chain, whose dynamics are described by \cite{anderson2011continuous}
\begin{equation*}
X(t)= X(0)+\sum_{j=1}^{r} \zeta_j R_j\left( \int_0^t \lambda_j (k, X(s))\text{d} s \right)
\end{equation*}
where $\zeta_j\triangleq v'_{\cdot j}- v_{\cdot j}$ (for $j=1,\dots,r$), $R_{j}(t)$ are independent unit rate Poisson processes, $k= (k_1,\dots,k_{\tilde r})^{\top}$ is an $\tilde r$-dimensional vector of model parameters, and $\lambda_j (\kappa,x)$ are propensities indicating the speed of reaction firing.
The most common propensity in chemistry and biology is the mass-action type, where $\lambda_j (\kappa,x)= k_j \prod_{i=1}^n \frac{x_i!}{\left(x_i-v_{i,j}\right)!}\mathbbold 1_{\{x_i\geq v_{i,j}\}}$ with $\mathbbold{1}_{\{\cdot\}}$ the indicator function.
In the literature, the above model describing the dynamics of intracellular reaction systems is termed as a stochastic reaction network (SRN) due to the stochasticity of the dynamics and the network structure of the chemical reactions.
Usually, an SRN encountered in systems biology is multiscale in nature \cite{kang2013separation}, meaning that different species vary a lot in abundance, and so do reaction rates.
Following \cite{kang2013separation}, these quantities at different scales can be normalized as follows.
Let $N$ be a scaling factor, the parameter $\alpha_i$ the abundance factor of the $i$-th species such that $X^{N}_i(t)\triangleq N^{-\alpha_i}X_{i}(t) = O(1)$, and the parameter $\beta_j$ the magnitude of $k_{j}$ such that $k'_j\triangleq k_j N^{-\beta_j}=O(1)$.
Similarly, we term $\rho_j$ as the timescale of the $j$-th reaction such that $\lambda^{N}_j (\mathcal K, X^N(t))\triangleq N^{-\rho_j}\lambda_j (k, X(t))= O(1)$ where $\mathcal K\triangleq \left(k'_1,\dots,k'_{\tilde r}\right)^{\top}$.
Particularly, in the mass-action kinetics, these coefficients satisfy $\rho_j=\beta_j + \sum_{i=1}^{n}v_{ij} \alpha_i$.
Finally, by denoting $X^{N,\gamma}(t) \triangleq X^{N}(tN^{\gamma})$ with $\gamma$ the timescale of interest, we can re-express the dynamics as
\begin{align}{\label{eq. scaling stochastic dynamics}}
X^{N,\gamma}(t) =& X^{N,\gamma}(0) +\sum_{j=1}^{r} \Lambda^{N} \zeta_j R_j\left( \int_0^{t} N^{\gamma+ \rho_j} \lambda^N_{j}(\mathcal K, X^{N,\gamma}(s)) \text{d} s \right)
\end{align}
where $\Lambda^{N}\triangleq \text{diag}(N^{-\alpha_1},\dots,N^{-\alpha_s})$, and all terms but $\Lambda^{N}$ and $N^{\gamma+\rho_j}$ are in the order of $O(1)$.
In the sequel, we study the SRN in the normalized coordinate.
Also, for simplicity, we concern ourselves with the fastest timescale of the system, i.e., $\gamma_1\triangleq \min_{i}\left\{\alpha_i-\max_{j\in \{v_{i,j}\neq 0\}} \rho_j \right\}$, on which the system evolves at the rate of $\mathcal O (1)$.
\subsection{Filtering problems for SRNs}{\label{subsection filtering problems}}
Stochastic filtering for SRNs aims to infer hidden dynamic states of an SRN from partial observations of the system.
Usually, we need to simultaneously estimate the model parameter $\mathcal K$ in this filtering problem due to the uncertainty of $\mathcal K$.
In this paper, we suppose $\left(\mathcal K,X^{N}(0)\right)$ to have a known prior distribution.
For the observation, we assume that $m$ channels of light intensity signals, denoted by $Y^{N,\gamma_1}(\cdot)$, can be observed (with a microscope for instance) and satisfy
\begin{align*}
Y^{N,\gamma_1}(t_i)&= h\left(X^{N,\gamma_1}(t_i)\right) + W(t_i) \qquad && \forall i\in \mathbb N_{>0},
\end{align*}
where $\{t_i\}_{i\in\mathbb N_{>0}}$ is a strictly increasing sequence of time points at which the observation is collected, $h$ is an $m$-dimensional bounded Lipschitz continuous function indicating the relation between the observation and the SRN, and $\{W_{\ell}(t_i)\}_{i\in\mathbb N_{>0}}$ are independent $m$-variate standard Gaussian random variables.
Usually, the dimension of the observation vector is much less than that of the state vector, meaning that only a part of the system information is detected.
Besides, we assume that the observation noise is independent of the dynamical system \eqref{eq. scaling stochastic dynamics}, i.e., $\{W_{\ell}(t_i)\}_{i\in\mathbb N_{>0}}$ are independent of Poisson processes $R_{j} (\cdot)$ ($j=1,\dots,r$), system parameters $\mathcal K$, and the initial condition $X^{N}(0)$.
In the sequel, we term $\mathcal Y^{N,\gamma_1}_{t_i}$ as the $\sigma$-algebra generated by observations up to the time $t_i$.
In the filtering problem, the standard object to study is the conditional distribution process $\{\pi^{N,\gamma_1}_{t_i}\}_{i\in\mathbb N_{>0}}$ given by $\pi^{N,\gamma_1}_{t_i}(\phi)\triangleq \mathbb E_{\mathbb P}\left[\phi\left(\mathcal K, X^{N,\gamma_1}(t_i)\right) \left | \mathcal Y^{N,\gamma_1}_{t_i} \right.\right]$ for any measurable function $\phi:\mathbb R^{\tilde r} \times \mathbb R^{n} \to \mathbb R$.
This conditional distribution process satisfies the following relations \cite[Proposition 10.6]{bain2008fundamentals}
\begin{align}
p^{N,\gamma_1}_{t_i}(A) &= \int_{z\in \mathbb R^{\tilde r} \times \mathbb R^n } K^{N,\gamma_1}_{t_i-t_{i-1}} (z, A) \pi^{N,\gamma_1}_{t_{i-1}}(\text{d} z), && \forall A \subset \mathbb R^{\tilde r} \times \mathbb R^n \text{ and } \forall i \in\mathbb N_{>0}, \label{eq. recurssive expression 1}\\
\frac{\text{d} \pi^{N,\gamma_1}_{t_i}}{ \text{d} p^{N,\gamma_1}_{t_i} } &= \frac{ \ell_{Y^{N,\gamma_1}(t_i)} }{p^{N,\gamma_1}_{t_i} \left(\ell_{Y^{N,\gamma_1}(t_i)}\right) }, && \forall i \in\mathbb N_{>0}, \label{eq. recurssive expression 2}
\end{align}
where $t_0=0$, $\pi^{N,\gamma_1}_{t_0}$ is the initial distribution, $K^{N,\gamma_1}_{t}(z, A)$ is the transition kernel of the system \eqref{eq. scaling stochastic dynamics} being equal to $ \mathbb P\left( (\mathcal K, X^{N,\gamma_1}(t))\in A \left| (\mathcal K, X^{N,\gamma_1}(0))=z \right.\right)$, and $\ell_{y}(\kappa, x)\triangleq
\mathbb P \left( Y^{N,\gamma_1}(t_i)=y | X^{N,\gamma_1}(t_1)=x \right)$ is the likelihood function.
Here, equation \eqref{eq. recurssive expression 1} can be interpreted as a prediction step where one estimates hidden dynamic states at $t_i$ using observations up to the previous time points, and \eqref{eq. recurssive expression 2} is an adjustment step where the conditional distribution is obtained by reweighing the prediction according to the new observation.
Usually, the above equations cannot be solved explicitly; therefore, we intend to establish efficient particle filtering algorithms that can quickly and accurately solve these filtering problems.
\subsection{Model reduction for SRNs}
The dynamics of an SRN is a continuous-time Markov chain (CTMC) which can be exactly simulated using the Gillespie stochastic simulation algorithm \cite{gillespie1976general,gillespie1977exact} or the next reaction method \cite{gibson2000efficient}.
The computational complexity of these exact simulation algorithms is inversely proportional to the fastest timescale of reactions;
therefore, these algorithms can only be computationally efficient for those SRNs having no fast reaction (i.e., the maximum $N^{\gamma_1+\rho_j}$ is small), otherwise they take an impractical amount of computational time to simulate systems.
This fact can also have an adverse effect on particle filtering, as it is a simulation-based method (more details are illustrated in the next section).
To accelerate the simulation, a few algorithms have been invented for systems on different scales.
In the following, we review two well-known reduced models, the tau-leaping model and the piecewise deterministic Markov process (PDMP).
\subsubsection{Tau-leaping}
The tau-leaping method is a stochastic analog to the Euler method for ODEs, where the algorithm changes the value of the propensities periodically over time rather than instantaneously after each jump event.
Since the propensity is updated much less frequently in this method than in the exact simulation method, the tau-leaping algorithm consumes much less computational time to simulate the system.
At the timescale $\gamma_1$, the tau-leaping algorithm can be expressed as
\begin{align}
X^{N,\gamma_1}_{\tau}(t)= X^{N}(0)
+\sum_{j=1}^{r} \Lambda^{N} \zeta_j R_j\left( \sum_{i=0}^{\infty}
\left({\tau_{i+1}\wedge t}- {\tau_{i}\wedge t} \right) N^{\gamma_1+ \rho_j} \lambda^N_{j}\left(\mathcal K, X^{N,\gamma_1}_{\tau}\left( \tau_i \right)\right) \right)
\end{align}
where $X^{N,\gamma_1}_{\tau}(t)$ is the state vector, and $\{\tau_i\}_{i\in\mathbb N}$ is an increasing sequence of time points at which propensities are updated. Equivalently, it can also be written by $$X^{N,\gamma_1}_{\tau}\left(\tau_{i+1}\right)= X^{N,\gamma_1}_{\tau}\left(\tau_i\right)
+\sum_{j=1}^{r} \Lambda^{N} \zeta_j \mathcal P_{j,i}$$
where $\mathcal P_{j,i}$ are independent Poisson-distributed random variables with mean $(\tau_{i+1}-\tau_i) N^{\gamma_1+ \rho_j} \lambda^N_{j}\left(\mathcal K, X^{N,\gamma_1}_{\tau}(\tau_i)\right)$.
For simplicity, in this paper, we consider the case where the grid $\{\tau_i\}_{i\in\mathbb N}$ is deterministic and contains all the observation times.
Usually, the leaping time should be short enough so that the normalized propensity will not change much during $[\tau_i, \tau_{i+1})$.
Also, it should be greater than the timescale of the fastest reaction so that it leaps over several reaction firing events and, thus, is faster than the exact simulation method.
Often, the tau-leaping algorithm is applied to systems containing $ 10^{2}\sim 10^3$ molecules \cite{hu2011weak} (in other words, the maximum $N^{\gamma_1+ \rho_j}$ is $\mathcal O \left(10^{2}\sim 10^3\right)$).
In this paper, we only consider applying the tau-leaping algorithm to the system where the following assumptions hold.
\begin{assumption}[Nonnegativity of the tau-leaping algorithm]{\label{assumption non-negativity of tau-leaping algorithm}}
Denote $|\tau|\triangleq \max_{i} (\tau_{i+1}-\tau_i)$. Then, for any $t>0$ and any $|\tau|>0$, the state $X^{N,\gamma_1}_{\tau}(t)$ is always nonnegative.
\end{assumption}
\begin{assumption}[Consistency of the tau-leaping algorithm]{\label{assumption convergence of tau-leaping algorithm}}
For any fixed $\mathcal K$, there holds $X^{N,\gamma_1}_{\tau}(\cdot) \stackrel{|\tau| \to 0}{\Longrightarrow} X^{N,\gamma_1}(\cdot) $ on any interval $T>0$, in the sense of the Skorokhod topology.
\end{assumption}
In general, the above two properties assumed for the tau-leaping method are essential and non-restrictive.
The nonnegativity assumption ensures the algorithm to generate reasonable trajectories that evolve in the nonnegative orthant, and this property can be achieved by wisely designing the grid $\{\tau_i\}_{i\in\mathbb N}$ (see \cite{anderson2008incorporating}).
The consistency assumption states that the tau-leaping algorithm is accurate (in the weak sense) in simulating the target system when the coarseness of the time-discretization scheme is small enough.
A few sufficient conditions for the consistency assumption have been provided in the literature \cite{anderson2011error,li2007analysis,rathinam2005consistency} (for bounded processes) and \cite{rathinam2016convergence} (for unbounded processes).
\subsubsection{Piecewise deterministic Markov Processes (PDMP)}
For a system having an extremely large scaling factor $N$, the PDMP is a favorable approximate model, where the firing events of fast reactions are approximated by continuous processes, and those slow reactions are kept as what they are.
Specifically, for fast reactions ($j\in \{\gamma_1+ \rho_j>0\}$), this method utilizes Poisson's law of large numbers to approximate the reaction firing events by continuous processes, i.e.,
\begin{equation*}
N^{-\gamma_1- \rho_j} \zeta_j R_j\left( \int_{0}^{t} N^{\gamma_1+ \rho_j} \lambda^N_{j}\left(\mathcal K, X^{N,\gamma_1}\left( s \right)\right) \text{d} s \right)
\approx D^{\gamma_1+ \rho_j}\zeta_j \left(\int_0^t \lambda'_j( \mathcal K, X^{\gamma_1}(s)) \text{d} s\right)
\end{equation*}
where $\lambda'_j(\kappa, x)=\lim_{N\to+\infty} \lambda^N_j (\kappa, x)$, $X^{\gamma_1}(\cdot)$ is the limit process of $X^{N,\gamma_1}(\cdot)$, and $D^{\tilde \alpha} \triangleq \text{diag}( \mathbbold 1_{\{\alpha_1=\tilde \alpha\}}, \dots, \mathbbold 1_{\{\alpha_n=\tilde \alpha\}})$ indicates whether a species has the abundance factor $\tilde \alpha$.
In this regard, the reduced model can be written by \cite{kang2013separation}
\begin{align}{\label{eq. reduced model at the first time scale}}
X^{\gamma_{1}}(t)=& \lim_{N\to \infty} X^{N}(0) + \sum_{j : \gamma_1+ \rho_j>0} D^{\gamma_1+ \rho_j}\zeta_j \left(\int_0^t \lambda'_j( \mathcal K, X^{\gamma_1}(s)) \text{d} s \right) \\
& + \sum_{j : \gamma_1+ \rho_j=0 } D^0\zeta_j
R_{j}\left(\int_0^t \lambda'_j( \mathcal K, X^{\gamma_1}(s)) \text{d} s \right) \notag
\end{align}
where $\lim_{N\to \infty} X^{N}(0)$ is assumed to exist.
Since the above Markov process is deterministic between two neighboring jumping times, it is named the piecewise deterministic Markov process (PDMP).
For a slower timescale, such a reduced model can also be constructed in the same spirit if some fast species are first eliminated by quasi-stationary assumption.
A systematic procedure to construct a hierarchy of PDMPs representing dynamics in different timescales can be found in the literature \cite{kang2013separation}.
When $N$ is large, the simulation time of PDMPs is much less than that of the full model, because the PDMP avoids the exact simulation of fast reactions, and those continuous processes that approximate the firing events of fast reactions can be computed efficiently by the Euler method.
Detailed algorithms to simulate PDMPs have been provided in the literature \cite{crudu2009hybrid,duncan2016hybrid,hepp2015adaptive}.
The accuracy of the PDMP is ensured by the following proposition.
\begin{assumption}\label{assumption convergence of the initial condition}
$\lim_{N\to \infty} X^{N}(0)$ exists a.s., and $\lim_{N\to \infty} X^{N}_i(0)>0$ if $\alpha_i>0$.
\end{assumption}
\begin{proposition}[Adapted from \cite{kang2013separation}]{\label{prop Kang kurtz gamma1}}
For any fixed $\mathcal K$, if \cref{assumption convergence of the initial condition} holds, and $X^{N,\gamma_1}(t)$ (for all $N\in\mathbb N_{>0}$) and $X^{\gamma_1}(t)$ are almost surely non-explosive, then $X^{N,\gamma_1}(\cdot ) \stackrel{N\to\infty}{\Longrightarrow} X^{\gamma_{1}}(\cdot)$ in the sense of the Skorokhod topology on any finite time interval $[0,T]$.
\end{proposition}
\subsubsection{Working ranges of different simulation algorithms}{\label{subsubsection tau-leaping and PDMP}}
The working ranges of the aforementioned simulation algorithms are quite different.
When all reactions fire at the rate of $\mathcal O(1\sim 10)$ (i.e., all $N^{\gamma_1+ \rho_j}$ are $\mathcal O(1\sim 10)$), the exact simulation algorithms for the CTMC are recommended, as they are perfectly accurate and computationally efficient in this case.
The tau-leaping algorithm is usually applied to systems where the fastest reaction fires at the rate of $\mathcal O \left(10^{2}\sim 10^3\right)$;
in this scenario, the tau-leaping algorithm can be much faster than the exact simulation method while still capturing key behaviors of the system.
Finally, when the scaling factor $N$ is sufficiently large, the PDMP method is preferred to the others, because in this situation the randomness of fast reaction firing is averaged out due to the law of large numbers.
There is still one situation excluded here, that some reactions fire at extremely fast rates (e.g., $\mathcal O \left(10^{4}\right)$), but some others fire at the rate of $\mathcal O \left(10^{2}\sim 10^3\right)$.
In this case, the PDMP model fails to capture key behaviors of the system, as there is still randomness in those reactions on medium timescales.
On the other hand, the tau-leaping method is slow in this situation, because it needs to frequently generate Poisson random variables of large means.
Instead, the diffusion approximation is recommended for these systems, where the firing of a medium timescale reaction is approximated by a diffusion process.
Since this algorithm just needs to generate Gaussian random variables, the computational cost is low;
moreover, its accuracy is higher than that of the PDMP \cite{kang2014central}.
However, in this paper, we do not consider this type of approximation due to the difficulties in the analysis (particularly the parameter sensitivity analysis of this reduced model).
\section{The proof of \cref{thm main}}\label{section proof}
We first outline the strategy to prove our main result.
Note that the error of a particle filter comes from model reduction, finite sampling, resampling, and artificial evolution.
Here, we can easily analyze the error introduced by model reduction using the convergence of reduced models and investigate the error caused by finite sampling and resampling using the tools in \cite{del2000branching,crisan2001particle,bain2008fundamentals}.
Therefore, the key to proving our main result lies in the analysis of artificial evolution.
In this appendix, we show that the error caused by artificial evolution is of the order $\frac{1}{\sqrt{M}}$ (where $M$ is the particle population) by the the parameter sensitivity analysis, and, thus, this error vanishes as the particle population goes to infinity.
Finally, by combining all the error analysis results, we can prove the main theorem.
To make the idea more explicit, we formulate the proof outline as follows.
First, since the model reduction is not the key factor in this proof, we simplify the notation by removing the superscript that specifies the dynamical model applied.
Specifically, for a given RPF constructed in \Cref{def RPFs for SCRNs}, we term
\begin{equation}\label{eq. notations}
\begin{array}{lcl}
\bar \pi_{M,t_i}(\phi) & \text{as} &
\text{the particle filter }\bar \pi^{N,\gamma_1}_{M,t_i}(\phi),~ \bar \pi^{N,\gamma_1,\tau}_{M,t_i}(\phi), ~\text{or} ~\bar \pi^{N,\gamma_1,H}_{M,t_i}(\phi), \\
\hat \pi_{M,t_i}(\phi) &\text{as} & \text{the estimate of the function $\phi$ based on resampled particles} \\
&& (\hat \pi_{M,t_i}(\phi) \triangleq \frac{1}{M}{\sum_{j=1}^{M} \phi(\hat \kappa_j(t_i), \bar x_j(t_i))}),\\
\tilde \pi_{M,t_i}(\phi) &\text{as} &\text{the estimate of the function $\phi$ based on artificially perturbed} \\
&& \text{particles } (\tilde \pi_{M,t_i}(\phi) \triangleq \frac{1}{M}\sum_{j=1}^{M} \phi(\bar \kappa_j(t_i), \bar x_j(t_i))),\\
\pi_{t_{i+1}}(\phi) &\text{as} & \text{the exact filter $\pi^{N,\gamma_{1}}_{t_{i+1}}(\phi)$},\\
K_t & \text{as} & \text{the transition kernel of the applied dynamical system} \\ &&\text{(the CTMC, tau-leaping model, or PDMP)},\\
K^{N,\gamma_1}_t & \text{as} & \text{the transition kernel of the full model}\\
Y(t_i) &\text{as} & \text{the observation $Y^{N,\gamma_{1}}(t_{i})$},\\
\Delta t_i &\text{as} & t_{i+1}-t_{i}.
\end{array}
\end{equation}
With these notations, we can decompose the error between the RPF $\bar \pi_{M,t_{i+1}} (\phi)$ and the exact filter $\pi_{t_{i+1}}(\phi)$ as follows.
\begin{align}\label{eq. bias caused by artificial evolution}
\bar \pi_{M,t_{i+1}} (\phi) - \pi_{t_{i+1}}(\phi)
&= \bar \pi_{M,t_{i+1}} (\phi) - \frac{\pi_{t_{i}}( K^{N,\gamma_1}_{\Delta t_i} \ell _{Y(t_{i+1})}\phi)}{\pi_{t_{i}}( K^{N,\gamma_1}_{\Delta t_i} \ell _{Y(t_{i+1})})} \notag\\
&=\underbrace{\left(\bar \pi_{M,t_{i+1}} (\phi)- \frac{\tilde \pi_{M,t_{i}} ( K_{\Delta t_i} \ell _{Y(t_{i+1})}\phi)}{Z_{t}} \right)}_\text{the error caused by finite sampling} \\
&\quad+\frac{1}{Z_t}
\underbrace{\Big({\tilde \pi_{M,t_{i}} ( K_{\Delta t_i} \ell _{Y(t_{i+1})}\phi)} -
{\hat \pi_{M,t_{i}} ( K_{\Delta t_i} \ell _{Y(t_{i+1})}\phi)} \Big)}_\text{the error caused by the artificial evolution}
\notag\\
&\quad+\frac{1}{Z_t}
\underbrace{\Big(
{\hat \pi_{M,t_{i}} ( K_{\Delta t_i} \ell _{Y(t_{i+1})}\phi)}
- {\bar \pi_{M,t_{i}} ( K_{\Delta t_i} \ell _{Y(t_{i+1})}\phi)}
\Big)}_\text{the error caused by resampling} \notag\\
&\quad+\frac{1}{Z_t}
\underbrace{\Big(
{\bar \pi_{M,t_{i}} ( K_{\Delta t_i} \ell _{Y(t_{i+1})}\phi)}
- { \bar \pi_{M,t_{i}}( K^{N,\gamma_1}_{\Delta t_i} \ell _{Y(t_{i+1})}\phi)}
\Big)}_\text{the error caused by model reduction} \notag \\
&\quad+\frac{1}{Z_t}
\underbrace{\Big(
{ \bar \pi_{M,t_{i}} ( K^{N,\gamma_1}_{\Delta t_i} \ell _{Y(t_{i+1})}\phi)}
- {\pi_{t_{i}}( K^{N,\gamma_1}_{\Delta t_i} \ell _{Y(t_{i+1})}\phi)}
\Big)}_\text{the error of the particle filter at $t_i$} \notag
\end{align}
where the first equality follows from the recursive formulas \eqref{eq. recurssive expression 1} and \eqref{eq. recurssive expression 2}, the term $\ell_{Y(t_{i+1})} $ is the likelihood function (see \eqref{eq. recurssive expression 2} and the text below), and $Z_{t}=\pi_{t_{i}}( K^{N,\gamma_1}_{\Delta t_i} \ell _{Y(t_{i+1})})$.
We note that for given observations, both the likelihood function and $Z_t$ are lower bounded by a positive constant due to the boundedness of $h(\cdot)$ (see \cref{subsection filtering problems}), so $\frac{1}{Z_t}$ is uniformly upper bounded and, therefore, not a concern in the error analysis.
\eqref{eq. bias caused by artificial evolution} tells that the error analysis of a particle filter can be performed in a recursive fashion where one analyzes the error of the filter at time $t_{i+1}$ after estimating the error at time $t_i$.
Therefore, the key to proving our result lies in investigating the first four terms on the right hand side of this equality, which corresponds to finite sampling, artificial evolution, resampling, and model reduction, respectively.
We also note that the first and the third terms have already been analyzed in the classical literature \cite{del2000branching,crisan2001particle,bain2008fundamentals}, and the fourth can be analyzed by the convergence of the reduced model.
Consequently, the most challenging part is to analyze the second term, which corresponds to the error caused by the artificial evolution step.
Technically, we estimate the second moment of the error caused by artificial evolution.
Notice that for any function $\phi$ satisfying \eqref{eq. sensitivity condition of phi}, we can find a positive constant $\tilde C_{\phi}$ such that
\begin{align} \label{eq. tilde hat convergence}
&\mathbb E\left[ \left| \tilde \pi_{M,t_{i+1}}(\phi) - \hat \pi_{M,t_{i+1}}(\phi)\right|^2 \left| \mathcal Y_{t_{i+1}} \right. \right] \\
&= \mathbb E\left[ \left. \left| \frac{1}{M}\sum_{j=1}^M \phi(\hat \kappa_j(t_{i+1}), \bar x_j(t_{i+1})) - \phi(\bar \kappa_j(t_{i+1}), \bar x_j(t_{i+1}))\right|^2 \right| \mathcal Y_{t_{i+1}} \right] \notag\\
&\leq \frac{\tilde C^2_{\phi} }{M} \sum_{j=1}^{M} \mathbb E\left[ \|\bar \kappa_{j}(t_{i+1}) - \hat \kappa_{j}(t_{i+1}) \|^2_2 \left(1+\|\bar x_{j}(t_{i+1}) \|^q_2\right)^2 \big| \mathcal Y_{t_{i+1}}\right] \notag\\
&\leq \frac{\tilde C^2_{\phi}C^2_{\eta} \mathcal V}{M} \sum_{j=1}^{M}\frac{1}{M}\mathbb E\left[ \left(1+\|\bar x_{j}(t_{i+1})\|^q_2\right)^2\big| \mathcal Y_{t_{i+1}} \right] \notag
\end{align}
where the third line follows from the midpoint theorem and the Jensen inequality, and the last line follows from \Cref{cond. conditions on the artificial noise 1}.
Therefore, we can analyze the effect of artificial evolution by checking whether $K_{\Delta t_i} \ell _{Y(t_{i+1})}\phi$ satisfies \eqref{eq. sensitivity condition of phi} and estimating the conditional moments of particles $\sum_{j=1}^{M}\frac{1}{M}\mathbb E\left[ \left(1+\|\bar x_{j}(t_{i+1})\|^q_2\right)^2\big| \mathcal Y_{t_{i+1}} \right]$, of which the former corresponds to the parameter sensitivity analysis of the transition kernel.
Following this idea, in \Cref{subsection proof parameters}, we use the parameter sensitivity analysis to show that the second term on the right hand side of \eqref{eq. bias caused by artificial evolution} is of order $\frac{1}{\sqrt{M}}$.
Specifically, we first prove in \Cref{subsection proof parameters} that both the state vector and the parameter sensitivity for the transition kernel grow mildly under the assumed conditions (\cref{proposition finiteness of moments} and \cref{proposition Kphi}).
Then, based on these results and \eqref{eq. tilde hat convergence}, we can find a $\mathcal Y_{t_{i+1}}$ measurable random variable $\tilde C_{t_{i+1}}$ such that
\begin{align*}
\mathbb E\left[
\big| \text{the second term on the right hand side of }\eqref{eq. bias caused by artificial evolution} \big|^2 \big| \mathcal Y_{t_{i+1}} \right]
\leq \frac{\tilde C_{t_{i+1}}}{M},
\end{align*}
which indicates that the error caused by artificial evolution is of order $\frac{1}{\sqrt{M}}$ (see \Cref{remark bias}).
Finally, based on the obtained results, we prove the convergence of the established RPF by mathematical induction in \Cref{subsection proof main result}.
\subsection{Parameter sensitivities for transition kernels and the analysis of the error introduced by artificial evolution} \label{subsection proof parameters}
In this subsection, we use the parameter sensitivity analysis to estimate the error introduced by the artificial evolution and show that it vanishes as the particle population goes to infinity.
Recall that the key is to show $K_{\Delta t_i} \ell _{Y(t_{i+1})}\phi$ satisfying \eqref{eq. sensitivity condition of phi} and moments of particles being bounded from above (see \eqref{eq. tilde hat convergence}).
First, \Cref{assumption for full model} and \Cref{assumption non-negativity of tau-leaping algorithm} suggest that any moment of the state vector is bounded, and, therefore, the moments of particles are also bounded.
\begin{proposition}[Adapted from {\cite[Lemma 5.1]{gupta2013unbiased}} and {\cite[Theorem 4.5]{rathinam2016convergence}}]\label{proposition finiteness of moments}
If \Cref{assumption for full model} holds, then for all $p>0$, there exists a positive constant $C_p$ such that
\begin{align*}
\mathbb E\left[ \left. \left\|X^{N,\gamma_1}(t)\right\|^{p}_2 ~\right| \mathcal K= \kappa, X^{N,\gamma_1}(0)=x \right] &\leq
\left( \left\|x\right\|^{p}_2
+ C_p t \right) \text{e}^{C_p t}\\
\mathbb E\left[ \left. \left\|X^{\gamma_1}(t)\right\|^{p}_2 ~\right| \mathcal K = \kappa, X^{\gamma_1}(0)=x \right] &\leq
\left( \left\|x\right\|^{p}_2
+ C_p t \right) \text{e}^{C_p t}
\end{align*}
for any $t>0$ and $(\kappa,x) \in \Theta\times \mathbb R^{n}_{\geq 0}$.
If moreover \Cref{assumption non-negativity of tau-leaping algorithm} holds, then for $p>0$, there exists a constant $C_{p,\tau}$ such that
\begin{align*}
\sup_{t\in[0,T]}
\mathbb E \left[ \left. \left\| X^{N,\gamma_1}_{\tau}(t) \right\|_2^p \right| \mathcal K=\kappa, X^{N,\gamma_1}_{\tau}(t)=x \right]
\leq \left( \left\|x\right\|^{p}_2
+ 1\right) \text{e}^{C_{p,\tau} t }
\end{align*}
for all $(\kappa,x) \in \Theta\times \mathbb R^{n}_{\geq 0}$.
\end{proposition}
\begin{proof}
The results for the full model and the PDMP are adapted from {\cite[Lemma 5.1]{gupta2013unbiased}}; the result for the tau-leaping method is adapted from {\cite[Theorem 4.5]{rathinam2016convergence}}.
\end{proof}
Then, we look at the parameter sensitivity for the CTMC, tau-leaping model, and PDMP.
Given an initial condition $(\kappa,x)$ and a time point $t$, we term
\begin{align*}
\Psi^{N}_{t,g}(\kappa,x) &= \mathbb E \left[ g\left(X^{N,\gamma_1}(t)\right) \left|~\mathcal K\triangleq \kappa, X^{N,\gamma_1}(0)=x\right.\right], \\
\Psi^{N,\tau}_{t,g}(\kappa,x) &= \mathbb E \left[ g\left(X^{N,\gamma_1}_{\tau}(t)\right) \left|~\mathcal K\triangleq \kappa, X^{N}(0)=x\right.\right], \\
\Psi_{t,g}(\kappa,x) &= \mathbb E \left[ g\left(X^{\gamma_1}(t)\right) \left|~\mathcal K\triangleq \kappa, X^{N,\gamma_1}(0)=x\right.\right]
\end{align*}
as the conditional expectations of function $g$ for the CTMC, tau-leaping model, and PDMP, respectively.
Then, the parameter sensitivity for these models are defined, respectively, by $\frac{\partial \Psi^N_{t,g}(\kappa,x) }{\partial \kappa}$, $\frac{\partial \Psi^{N,\tau}_{t,g}(\kappa,x) }{\partial \kappa}$, and $\frac{\partial \Psi_{t,g}(\kappa,x) }{\partial \kappa}$.
The literature \cite{gupta2018estimation} shows that the parameter sensitivity for the full model can be computed by
\begin{small}
\begin{align}
&\frac{\partial \Psi^N_{t,g}(\kappa,x) }{\partial \kappa} \label{eq. parameter sensitivity for SCRNs} \\
&= \sum_{j=1}^{r} \mathbb{E}
\left[
\left.
\int_{0}^{t} \frac{\partial \lambda^{N}_{j}\left(\kappa, X^{N,\gamma_1}(s)\right)}{\partial \kappa}
\Delta^N_j \Psi^N_{t-s,g}\left(\kappa,X^{N,\gamma_1}(s)\right)
\text{d} s
\right| ~\mathcal K=\kappa, X^{N}(0)=x
\right] \notag
\end{align}
\end{small}where $\Delta^N_j \Psi^N_{t-s,g}\left(\kappa,x\right)\triangleq \Psi^N_{t-s,g}\left(\kappa,x+\Lambda^N \zeta_j\right)- \Psi^N_{t-s,g}\left(\kappa,x\right)$.
Similarly, by mathematical induction, we can show that the parameter sensitivity for the tau-leaping model (with a deterministic time-discretization scheme) can be computed by
\footnote{Note that $\frac{\partial \Psi^{N,\tau}_{t,g}(\kappa,x) }{\partial \kappa}$ computes the parameter sensitivity when the tau-leaping model is viewed as the ground truth. It is not necessarily equal or even close to $\frac{\partial \Psi^N_{t,g}(\kappa,x) }{\partial \kappa}$, the parameter sensitivity for the full model.}
\begin{small}
\begin{align}
&\frac{\partial \Psi^{N,\tau}_{t,g}(\kappa,x) }{\partial \kappa} \label{eq parameter sensitivity for the tau-leaping method} \\
&= \sum_{j=1}^{r} \mathbb{E}
\Bigg[
\sum_{\tau_i<t} (\tau_{i+1}\wedge t -\tau_i)\frac{\partial \lambda^{N}_{j}\left(\kappa, X^{N,\gamma_1}_\tau (\tau_i)\right)}{\partial \kappa}
\Delta^N_j \Psi^N_{t-s,g}\left(\kappa,X^{N,\gamma_1}_{\tau}(\tau_{i+1})\right) \notag \\
& \qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad
\bigg| ~\mathcal K=\kappa, X^{N,\gamma_1}_{\tau}(0)=x
\bigg]. \notag
\end{align}
\end{small}Finally, the literature \cite{gupta2019sensitivity} tells that under \Cref{assumption convergence of the initial condition}, \Cref{assumption for full model}, and \Cref{assumption for PDMPs}, the parameter sensitivity for the PDMP satisfies
\begin{equation}\label{eq. parameter sensitivity PDMP}
\frac{\partial \Psi_{t,g}(\kappa,x) }{\partial \kappa} = \lim_{N\to\infty} \frac{\partial \Psi^N_{t,g}(\kappa,x) }{\partial \kappa},
\end{equation}
if $g$ is bounded, continuously differentiable.
Using these results, we can further prove that the parameter sensitivity for the transition kernel is at most polynomially growing with respect to the state argument (see \Cref{proposition Kphi}).
\begin{proposition}(Polynomially growing rate of the parameter sensitivity for the transition kernel)\label{proposition Kphi}
For any measurable function $\phi$, we denote
\begin{align*}
\left(K^{N,\gamma_1}_t \phi \right)(\kappa, x) &\triangleq \mathbb E_{\mathbb P} \left[ \phi\left(\mathcal K, X^{N,\gamma_1}(t)\right) \left|~\mathcal K=\kappa, X^{N}(0)=x\right.\right], \\
\left(K^{N,\gamma_1,\tau}_t \phi \right)(\kappa, x) &\triangleq \mathbb E_{\mathbb P} \left[ \phi\left(\mathcal K, X^{N,\gamma_1}_{\tau}(t)\right) \left|~\mathcal K=\kappa, X^{N}(0)=x\right.\right], \\
\left(K^{\gamma_1}_t \phi \right)(\kappa, x) &\triangleq \mathbb E_{\mathbb P} \left[ \phi\left(\mathcal K, X^{\gamma_1}(t)\right) \left|~\mathcal K=\kappa, X^{\gamma_1}(0)=x\right.\right].
\end{align*}
If $\phi:\Theta\times \mathbb R^{n}_{\geq 0} \to \mathbb R$
satisfies the requirement in \cref{thm main}, then the following results hold.
\begin{itemize}
\item If \Cref{assumption for full model} holds, then for any $t>0$ there exists a constant $C_{t,\phi}$ such that $\left\| \frac{ \partial \left(K^{N,\gamma_1}_t \phi \right)(\kappa, x) }{\partial \kappa}\right\|_{2}
\leq C_{t,\phi} \left(1+ \|x\|^q_2\right)$ for all $ (\kappa,x) \in \Theta\times \mathbb R^{n}_{\geq 0}$.
\item If \Cref{assumption non-negativity of tau-leaping algorithm} and \Cref{assumption for full model} hold, then for any $t>0$ there exists a constant $C_{t,\phi,\tau}$ such that $\left\| \frac{ \partial \left(K^{N,\gamma_1,\tau}_t \phi \right)(\kappa, x) }{\partial \kappa}\right\|_{2}
\leq C_{t,\phi,\tau} \left(1+ \|x\|^q_2\right)$ for all $ (\kappa,x) \in \Theta\times \mathbb R^{n}_{\geq 0}$.
\item If \Cref{assumption convergence of the initial condition}, \Cref{assumption for full model}, and \Cref{assumption for PDMPs} hold, then for any $t>0$ and $ (\kappa,x) \in \Theta\times \mathbb R^{n}_{\geq 0}$ there holds $ \left\| \frac{ \partial \left(K^{\gamma_1}_t \phi \right)(\kappa, x) }{\partial \kappa}\right\|_{2} \leq C_{t,\phi} \left(1+ \|x\|^q_2\right)$,
where $C_{t,\phi}$ is the same as that in the first result.
\end{itemize}
\end{proposition}
\begin{proof}
We first prove the result for the transition kernel of the full model.
By definition, we can write the derivative of the transition kernel by
\begin{align}
\frac{ \partial K^{N,\gamma_1}_t \phi (\kappa, x) }{\partial \kappa} \notag
=& \mathbb E \left[ \left. \frac{\partial \phi\left(\kappa, X^{N,\gamma_1}(t)\right)}{\partial \kappa} \right|~\mathcal K=\kappa, X^{\gamma_1}(0)=x \right]
+
\left.\frac{\partial \Psi^{N}_{t,\phi(\kappa, \cdot)}(\theta,x) }{\partial \theta}\right|_{\theta=\kappa}.
\end{align}
By \cref{eq. sensitivity condition of phi} and \cref{proposition finiteness of moments}, we can find a positive constant $C^1_{t,\phi}$ such that
\begin{equation}\label{eq. sensitivity analysis help 1}
\mathbb E \left[ \left. \left\| \frac{\partial \phi\left(\kappa, X^{N,\gamma_1}(t)\right)}{\partial \kappa}\right\|_2 \right|~\mathcal K=\kappa, X^{\gamma_1}(0)=x \right]
\leq C^1_{t,\phi}\left( 1
+ \|x\|^q_2 \right)
\end{equation}
for all $(\kappa,x) \in \Theta\times \mathbb R^{n}_{\geq 0}$.
Also, by \eqref{eq. parameter sensitivity for SCRNs} and the boundedness of $\phi$, there holds
\begin{align}
\left\| \left.\frac{\partial \Psi^{N}_{t,\phi(\kappa, \cdot)}(\theta,x) }{\partial \theta}\right|_{\theta=\kappa} \right\|_2
&\leq 2\|\phi\|_{\infty}\sum_{j=1}^{r} \left\|\mathbb{E}
\left[
\left.
\int_{0}^{t} \frac{\partial \lambda^{N}_{j}\left(\kappa, X^{N,\gamma_1}(s)\right)}{\partial \kappa}
\text{d} s
\right| ~\mathcal K=\kappa, X^{N}(0)=x
\right]\right\|_2 \notag \\
& \leq 2\|\phi\|_{\infty}\sum_{j=1}^{r} \mathbb{E}
\left[
\left.
\int_{0}^{t} \left\| \frac{\partial \lambda^{N}_{j}\left(\kappa, X^{N,\gamma_1}(s)\right)}{\partial \kappa} \right\|_2
\text{d} s
\right| ~\mathcal K=\kappa, X^{N}(0)=x
\right] \notag
\end{align}
where the second inequality follows from Jensen's inequality and the triangle inequality.
Therefore by \Cref{proposition finiteness of moments} and \Cref{assumption for full model}, we can find a constant $C^2_{t,\phi}$ such that
\begin{equation}\label{eq. sensitivity analysis help 2}
\left\| \left.\frac{\partial \Psi^{N}_{t,\phi(\kappa, \cdot)}(\theta,x) }{\partial \theta}\right|_{\theta=\kappa}\right\|_2
\leq C^2_{t,\phi}\left( 1
+ \|x\|^q_2 \right)
\qquad \forall (\kappa,x) \in \Theta\times \mathbb R^{n}_{\geq 0}.
\end{equation}
Finally, by combing \eqref{eq. sensitivity analysis help 1} and \eqref{eq. sensitivity analysis help 2}, we prove the result for the transition kernel of the full model (the first result).
By the same argument, we can also prove the second result.
For the third result, we can learn from \eqref{eq. parameter sensitivity PDMP} that $\Psi_{t,\phi(\kappa, \cdot)}(\theta,x)$ satisfies \eqref{eq. sensitivity analysis help 2}.
Moreover, by \cref{eq. sensitivity condition of phi} and \cref{proposition finiteness of moments}, we can show that the quantity $\phi\left(\kappa, X^{\gamma_1}(t)\right)$ also satisfies \eqref{eq. sensitivity analysis help 1}.
Therefore, the third result is proven.
\end{proof}
\begin{remark}[Analysis of the error introduced by artificial evolution]{\label{remark bias}}
Now, we use the obtained propositions to show that the second term on the right hand side of \eqref{eq. bias caused by artificial evolution} is of the order $1/\sqrt{M}$.
First, we look at the conditional moments of the particles in the RPF.
We note that every weight $w_{j}(t_i)$ ($j=1,\dots,M$, $i=1,2,\dots,$) is upper bounded by the constant $\frac{1}{M}\exp\left( \left\|Y(t_i)\right\|^2_2+ m\|h\|^2_{\infty}\right)$, and the resampling does not change the expectation of the empirical mean of these weighted particles.
These facts together with the third claim of \cref{assumption for full model} and \cref{proposition finiteness of moments} suggests that given the observation process, the second conditional moment of the particles is bounded from above, i.e.,
\begin{align}\label{eq. boundedness of particles' moments}
\frac{1}{M} \sum_{j=1}^n \mathbb E\left[ \left(1+\|\bar x_{j}(t_{i+1})\|^q_2\right)^2\big| \mathcal Y_{t_i+1} \right] \leq \tilde C_{Y(t_{i+1})}
\end{align}
where $\tilde C_{Y(t_{i+1})}$ is a $\mathcal Y_{t_i+1} $ measurable random variable.
Second, we note that if $\phi$ satisfies the requirement in \Cref{thm main}, the function $\ell_{y} \phi$ (for any $y\in\mathbb R^{m}$) also satisfies this requirement, because the likelihood function $\ell_{y}$ is bounded and has no dependence on the argument $\kappa$.
Consequently, by \cref{proposition Kphi}, the function $K_{\Delta t_i} \ell _{Y(t_{i+1})}\phi$ also satisfies the requirement in \Cref{thm main} under the proposed conditions.
Finally, applying these results to \eqref{eq. tilde hat convergence}, we can find a $\mathcal Y_{t_i+1} $ measurable random variable $\tilde C_{t_{i+1}}$ such that
\begin{align}\label{eq error analysis the second term}
\mathbb E\left[
\big| \text{the 2nd term on the right hand side of }\eqref{eq. bias caused by artificial evolution} \big|^2 \big| \mathcal Y_{t_i+1} \right]
\leq \frac{\tilde C_{t_{i+1}}}{M},
\end{align}
which suggests that the bias contributed by artificial evolution is of order $\frac{1}{\sqrt{M}}$.
\end{remark}
\subsection{Proving the main result by mathematical induction} {\label{subsection proof main result}}
In this subsection, we prove \cref{thm main} using the results obtained in the previous section.
\begin{proof}[The Proof of \cref{thm main}]
For simplicity, we use the simplified notations in \eqref{eq. notations} to prove the result.
Moreover, we introduce another quantity $a$ to indicate the fidelity of the reduced model.
Specifically, if the RPF uses the tau-leaping model, then $a=|\tau|$;
if the RPF uses the PDMP model, then $a=1/N$;
if the RPF uses the full model, then $a\equiv 0$.
By the dominated convergence theorem, to prove the result, we only need to show
\begin{align}\label{eq. alternative quantity}
\lim_{M\to\infty} \lim_{a\to\infty} \mathbb E\left[ \left(\bar \pi_{M,t_{i}} (\phi) - \pi_{t_{i}}(\phi)\right)^2 \big| \mathcal Y_{t_{i}} \right] = 0
\end{align}
for any $i\in\mathbb N_{\geq 0}$ and any $\phi$ satisfying the requirement in \cref{thm main}.
Therefore, in the sequel, we focus on analyzing the conditional error of the filters.
We use mathematical induction to prove \eqref{eq. alternative quantity}.
For $i=1$, the relation \eqref{eq. alternative quantity} holds automatically by the law of large numbers and the convergence of the reduced models.
Then, we are going to show that if \eqref{eq. alternative quantity} holds for a positive integer $i$, then it also holds for the positive integer $i+1$.
Note that we can decompose the error of the filter as \eqref{eq. bias caused by artificial evolution}; therefore, the key is to separately analyze the error caused by finite sampling, artificial evolution, resampling, and model reduction.
\textit{Error caused by model reduction:}
Recall that for given observations, both the likelihood function and $Z_t$ in \eqref{eq. bias caused by artificial evolution} are lower bounded by a positive constant due to the boundedness of $h(\cdot)$ (see \cref{subsection filtering problems}), so $\frac{1}{Z_t}$ is uniformly upper bounded and, therefore, not a concern in the error analysis.
Then, by the convergence of the reduced models (\cref{assumption convergence of tau-leaping algorithm} and \cref{prop Kang kurtz gamma1}) and the conditional dominated convergence theorem, we have that for any bounded continuous function $\phi$, there holds
\begin{align}\label{eq. error 4th}
\lim_{a\to0} \mathbb E\left[ \left( \text{the 4th term on the RHS of \eqref{eq. bias caused by artificial evolution}}\right) ^2 \big | \mathcal Y_{t_{i+1}}\right] =0.
\end{align}
\textit{Error caused by resampling:}
If we use the multinomial resampling, then all the resampled particles are conditionally independently and identically distributed given the observation process, and, therefore, we can conclude \cite{crisan2001particle
\begin{equation*}
\mathbb E \left[ \left. \left| \hat\pi_{M,t_{i}}\left(K_{\Delta t_i} \ell _{Y(t_{i+1})}\phi\right) - \bar \pi_{M,t_{i+1}}\left(K_{\Delta t_i} \ell _{Y(t_{i+1})}\phi\right)\right|^2\right| \mathcal Y_{t_{i+1}} \right] \leq
\frac{ \|\ell _{Y(t_{i+1})}\phi\|^2_{\infty}}{M},
\end{equation*}
where $\|\ell _{Y(t_{i+1})}\phi\|_{\infty}$ is the uniform norm of the function $\ell _{Y(t_{i+1})}\phi$.
Moreover, by \cite[Exerice 9.1]{bain2008fundamentals} which states that residual resampling introduces the minimum noise among all resampling methods, we can also conclude the above relation for residual resampling, and, therefore,
\begin{align}\label{eq. error 3rd}
\lim_{M\to\infty} \lim_{a\to\infty} \mathbb E\left[ \left( \text{the 3rd term on the RHS of \eqref{eq. bias caused by artificial evolution}}\right) ^2 \big | \mathcal Y_{t_{i+1}}\right] =0,
\end{align}
for any bounded $\phi$.
\textit{Error caused by artificial evolution:} By \eqref{eq error analysis the second term} in \cref{remark artificial evolution}, we can conclude that for any $\phi$ satisfies the requirement in \Cref{thm main}, there holds
\begin{align}\label{eq. error 2nd}
\lim_{M\to\infty} \lim_{a\to\infty} \mathbb E\left[ \left( \text{the 2nd term on the RHS of \eqref{eq. bias caused by artificial evolution}}\right) ^2 \big | \mathcal Y_{t_{i+1}}\right] =0.
\end{align}
\textit{Error caused by finite sampling:}
For any $\phi$ satisfying the requirement in \cref{thm main}, the relations \eqref{eq. error 4th}, \eqref{eq. error 3rd}, \eqref{eq. error 2nd}, and \eqref{eq. alternative quantity} (for the integer $i$) imply
\begin{align}\label{eq. error 1st help}
\lim_{M\to\infty} \lim_{a\to\infty} \mathbb E \left[ \left. \left| \tilde \pi_{M,t_{i}}\left(K_{\Delta t_i} \ell _{Y(t_{i+1})}\phi\right) - \pi_{M,t_{i+1}}\left(K^{N,\gamma_1}_{\Delta t_i} \ell _{Y(t_{i+1})}\phi\right)\right|^2\right| \mathcal Y_{t_{i+1}} \right]=0.
\end{align}
Therefore, to analyze the error caused by finite sampling, we only need to compare $\bar \pi_{M,t_{i}}\left(\phi\right) $ and $\frac{\tilde \pi_{M,t_{i+1}}\left(K_{\Delta t_i} \ell _{Y(t_{i+1})}\phi\right)}{\tilde \pi_{M,t_{i+1}}\left(K_{\Delta t_i} \ell _{Y(t_{i+1})}\right)}$.
For any function $\phi$ satisfying the requirement in \Cref{thm main}, we define
\begin{align*}
\bar \rho_{M,t_{i+1}} (\phi) = {\sum_{j=1}^{M} \phi(\kappa_j(t_{i+1}), x_j(t_{i+1}))} \qquad \forall i\in\mathbb N,
\end{align*}
which is the estimate of the function $\phi$ right before the collection of $Y(t_{i+1})$.
Since the increments of the particles in the sampling step are independent, this quantity satisfies
\begin{align}\label{eq. estimate bar rho}
&\mathbb E \left[ \left. \left| \bar \rho_{M,t_i+1}(\phi) - \tilde \pi_{M,t_{i+1}}\left(K^{N,\gamma_1}_{\Delta t_i} \ell _{Y(t_{i+1})}\phi\right) \right|^2 \right| \mathcal Y_{t_{i+1}} \right] \leq \frac{\|\ell _{Y(t_{i+1})}\phi\|^2_{\infty}}{M} .
\end{align}
Moreover, by \cite[(2.10)]{crisan2001particle}, there is
\begin{align*}
&\left|\bar\pi_{M,t_{i+1}}(\phi) - \frac{\tilde \pi_{M,t_{i+1}}\left(K_{\Delta t_i} \ell _{Y(t_{i+1})}\phi\right)}{\tilde \pi_{M,t_{i+1}}\left(K_{\Delta t_i} \ell _{Y(t_{i+1})}\right)}\right| \\
&\leq
\frac{\left\|\phi\right\|_{\infty}}{\tilde \pi_{M,t_{i+1}}\left(K_{\Delta t_i} \ell _{Y(t_{i+1})}\right)}
\left|\bar \rho_{M,t_i+1}\left( \ell_{y_{i+1}}\right)
-
\tilde \pi_{M,t_{i+1}}\left(K_{\Delta t_i} \ell _{Y(t_{i+1})}\right)\right| \\
& \quad~ +
\frac{1}{\tilde \pi_{M,t_{i+1}}\left(K_{\Delta t_i} \ell _{Y(t_{i+1})}\right)}
\left|\bar \rho_{M,t_{i+1}}\left( \ell_{y_{i+1}}\phi\right)
-
\tilde \pi_{M,t_{i+1}}\left(K_{\Delta t_i} \ell _{Y(t_{i+1})}\phi\right)\right|.
\end{align*}
Recall that the likelihood function $\ell _{Y(t_{i+1})}$ has a $\mathcal Y_{t_{i+1}}$ measurable, positive lower bound due to the boundedness of $h(\cdot)$, so $\tilde \pi_{M,t_{i+1}}\left(K_{\Delta t_i} \ell _{Y(t_{i+1})}\right)$ is also greater than or equal to this bound.
Thus, by applying \eqref{eq. estimate bar rho} to this inequality, we have
\begin{equation*}
\lim_{M\to\infty} \lim_{a\to0}\mathbb E \left[
\left(\bar\pi_{M,t_{i+1}}(\phi) - \frac{\tilde \pi_{M,t_{i+1}}\left(K_{\Delta t_i} \ell _{Y(t_{i+1})}\phi\right)}{\tilde \pi_{M,t_{i+1}}\left(K_{\Delta t_i} \ell _{Y(t_{i+1})}\right)}\right)^2
\Bigg|
\mathcal Y_{t_{i+1}}
\right]
=0,
\end{equation*}
which together with \eqref{eq. error 1st help} suggest the error caused by finite sampling is controlled, i.e.,
\begin{align}\label{eq. error 1st}
\lim_{M\to\infty} \lim_{a\to\infty} \mathbb E\left[ \left( \text{the 1st term on the RHS of \eqref{eq. bias caused by artificial evolution}}\right) ^2 \big | \mathcal Y_{t_{i+1}}\right] =0.
\end{align}
Finally, by applying \eqref{eq. error 4th}, \eqref{eq. error 3rd}, \eqref{eq. error 2nd}, and \eqref{eq. error 1st} to \eqref{eq. bias caused by artificial evolution}, we can conclude that if \eqref{eq. alternative quantity} holds for a positive integer $i$, then it also holds for the positive integer $i+1$, which proves the result.
\end{proof}
\section{The proof of \cref{thm main}}\label{section proof}
\textcolor{blue}{
We first outline the strategy to prove our main result.
Note that the error of a particle filter comes from model reduction, finite sampling, resampling, and artificial evolution.
Here, we can easily analyze the error introduced by model reduction using techniques in \cite{calzolari2006approximation,fang2021stochastic} and the error caused by finite sampling and resampling using the tools in \cite{del2000branching,crisan2001particle,bain2008fundamentals}.
Therefore, the key to proving our main result lies in the analysis of artificial evolution.
In this appendix, we show that the error caused by artificial evolution is of the order $\frac{1}{\sqrt{M}}$ (where $M$ is the particle population) by the the parameter sensitivity analysis, and, thus, this error vanishes as the particle population goes to infinite.
Finally, by combing all the error analysis results, we can prove the main theorem.
}
\textcolor{blue}{
To make the idea more explicit, we formulate the proof outline as follows.
First, since the model reduction is not the key factor in this proof, we simplify the notation by removing the superscript that specifies the dynamical model applied.
Specifically, for a given RPF constructed in \Cref{def RPFs for SCRNs}, we term
\begin{equation}\label{eq. notations}
\begin{array}{lcl}
\bar \pi_{M,t_i}(\phi) & \text{as} &
\text{the particle filter }\bar \pi^{N,\gamma_1}_{M,t_i}(\phi),~ \bar \pi^{N,\gamma_1,\tau}_{M,t_i}(\phi), ~\text{or} ~\bar \pi^{N,\gamma_1,H}_{M,t_i}(\phi), \\
\hat \pi_{M,t_i}(\phi) &\text{as} & \text{the estimate of the function $\phi$ based on resampled particles} \\
&& (\hat \pi_{M,t_i}(\phi) \triangleq \frac{1}{M}{\sum_{j=1}^{M} \phi(\hat \kappa_j(t_i), \bar x_j(t_i))}),\\
\tilde \pi_{M,t_i}(\phi) &\text{as} &\text{the estimate of the function $\phi$ based on artificially perturbed} \\
&& \text{particles } (\tilde \pi_{M,t_i}(\phi) \triangleq \frac{1}{M}\sum_{j=1}^{M} \phi(\bar \kappa_j(t_i), \bar x_j(t_i))),\\
\pi_{t_{i+1}}(\phi) &\text{as} & \text{the exact filter $\pi^{N,\gamma_{1}}_{t_{i+1}}(\phi)$},\\
K_t & \text{as} & \text{the transition kernel of the applied dynamical system} \\ &&\text{(the CTMC, tau-leaping model, or PDMP)},\\
K^{N,\gamma_1}_t & \text{as} & \text{the transition kernel of the full model}\\
Y(t_i) &\text{as} & \text{the observation $Y^{N,\gamma_{1}}(t_{i})$},\\
\Delta t_i &\text{as} & t_{i+1}-t_{i}.
\end{array}
\end{equation}
With these notations, we can decompose the error between the RPF $\bar \pi_{M,t_{i+1}} (\phi)$ and the exact filter $\pi_{t_{i+1}}(\phi)$ as follows.
\begin{align}\label{eq. bias caused by artificial evolution}
\bar \pi_{M,t_{i+1}} (\phi) - \pi_{t_{i+1}}(\phi)
&= \bar \pi_{M,t_{i+1}} (\phi) - \frac{\pi_{t_{i}}( K^{N,\gamma_1}_{\Delta t_i} \ell _{Y(t_{i+1})}\phi)}{\pi_{t_{i}}( K^{N,\gamma_1}_{\Delta t_i} \ell _{Y(t_{i+1})})} \notag\\
&=\underbrace{\left(\bar \pi_{M,t_{i+1}} (\phi)- \frac{\tilde \pi_{M,t_{i}} ( K_{\Delta t_i} \ell _{Y(t_{i+1})}\phi)}{Z_{t}} \right)}_\text{the error caused by finite sampling} \\
&\quad+\frac{1}{Z_t}
\underbrace{\Big({\tilde \pi_{M,t_{i}} ( K_{\Delta t_i} \ell _{Y(t_{i+1})}\phi)} -
{\hat \pi_{M,t_{i}} ( K_{\Delta t_i} \ell _{Y(t_{i+1})}\phi)} \Big)}_\text{the error caused by the artificial evolution}
\notag\\
&\quad+\frac{1}{Z_t}
\underbrace{\Big(
{\hat \pi_{M,t_{i}} ( K_{\Delta t_i} \ell _{Y(t_{i+1})}\phi)}
- {\bar \pi_{M,t_{i}} ( K_{\Delta t_i} \ell _{Y(t_{i+1})}\phi)}
\Big)}_\text{the error caused by resampling} \notag\\
&\quad+\frac{1}{Z_t}
\underbrace{\Big(
{\bar \pi_{M,t_{i}} ( K_{\Delta t_i} \ell _{Y(t_{i+1})}\phi)}
- { \pi_{t_{i}}( K^{N,\gamma_1}_{\Delta t_i} \ell _{Y(t_{i+1})}\phi)}
\Big)}_\text{the error of the particle filter at $t_i$} \notag
\end{align}
where the first equality follows from the recursive formulas \eqref{eq. recurssive expression 1} and \eqref{eq. recurssive expression 2}, the term $\ell_{Y(t_{i+1})} $ is the likelihood function (see \eqref{eq. recurssive expression 2} and the text below), and $Z_{t}=\pi_{t_{i}}( K^{N,\gamma_1}_{\Delta t_i} \ell _{Y(t_{i+1})})$.
Here, the error caused by model reduction is hidden in the last term, because the transition kernels in the last line represent the reduced model and the full model, respectively, (see \eqref{eq. notations}).
Also, we note that both the likelihood function and $Z_t$ are lower bounded by a positive constant due to the boundedness of $h(\cdot)$ (see \cref{subsection filtering problems}), so $\frac{1}{Z_t}$ is uniformly upper bounded and, therefore, not a concern in the error analysis.
\eqref{eq. bias caused by artificial evolution} tells that the error analysis of a particle filter can be performed in a recursive fashion, where one analyzes the error of the filter at time $t_{i+1}$ after estimating the error at time $t_i$.
Therefore, the key to proving our result lies in investigating the first three terms on the right hand side of this equality.
We note that the first and the third terms can be easily analyzed using the techniques in the classical literature \cite{del2000branching,crisan2001particle,bain2008fundamentals}.
Consequently, the most challenging part is to analyze the second term, which corresponds to the error caused by the artificial evolution step.
}
\textcolor{blue}{Technically, we estimate the fourth order moment of the error caused by artificial evolution.
Notice that for any function $\phi$ satisfying \eqref{eq. sensitivity condition of phi}, we can find a positive constant $\tilde C_{\phi}$ such that
\begin{align} \label{eq. tilde hat convergence}
\mathbb E\left[ \left| \tilde \pi_{M,t_{i}}(\phi) - \hat \pi_{M,t_{i}}(\phi)\right|^4 \right]
&= \mathbb E\left[ \frac{1}{M^4}\left| \sum_{j=1}^M \phi(\hat \kappa_j(t_i), \bar x_j(t_i)) - \phi(\bar \kappa_j(t_i), \bar x_j(t_i))\right|^4 \right] \\
&\leq \frac{C^4_{\phi} }{M^4} \sum_{j=1}^{M} \mathbb E\left[\|\bar \kappa_{j}(t_{i}) - \hat \kappa_{j}(t_{i}) \|^4_2 \left(1+\|\bar x_{j}(t_{i})\|^q_2\right)^4\right] \notag\\
&\leq \frac{\tilde C_{\phi}}{M^2}\mathbb E\left[ \left(1+\|\bar x_{j}(t_{i})\|^q_2\right)^4\right] \notag
\end{align}
where the second line follows from the midpoint theorem, and the last line follows from \Cref{cond. conditions on the artificial noise 1}.
Therefore, we can analyze the effect of artificial evolution by checking whether $K_{\Delta t_i} \ell _{Y(t_{i+1})}\phi$ satisfies \eqref{eq. sensitivity condition of phi} and estimating $\mathbb E\left[ \left(1+\|\bar x_{j}(t_{i})\|^q_2\right)^4\right]$, the former of which corresponds to the parameter sensitivity analysis of the transition kernel.
}
\textcolor{blue}{Following this idea, in \Cref{subsection proof parameters}, we use the parameter sensitivity analysis to show that the second term on the right hand side of \eqref{eq. bias caused by artificial evolution} is of order $\frac{1}{\sqrt{M}}$.
Specifically, we first prove in \Cref{subsection proof parameters} that both the state vector and the parameter sensitivity for the transition kernel grow mildly under the assumed conditions (\cref{proposition finiteness of moments} and \cref{proposition Kphi}).
Then, based on these results and \eqref{eq. tilde hat convergence}, we can find a constant $\hat C_{y}$ for any $y\in\mathbb R^{m}$ such that
\begin{align*}
\mathbb E\left[
\big| \text{the second term on the RHS of }\eqref{eq. bias caused by artificial evolution} \big|^4 \Big| Y(t_{i+1})=y \right]
\leq \frac{\hat C_{y}}{M^2},
\end{align*}
which indicates that the error caused by artificial evolution is of order $\frac{1}{\sqrt{M}}$ (see \Cref{remark bias}).
Finally, based on the obtained results, we prove the convergence of the established RPF by mathematical induction in \Cref{subsection proof main result}.
}
\subsection{Parameter sensitivities for transition kernels and the analysis of the error introduced by artificial evolution} \label{subsection proof parameters}
\textcolor{blue}{In this subsection, we use the parameter sensitivity analysis to estimate the error introduced by the artificial evolution and show that it vanishes as the particle population goes to infinity.
Recall that the key is to show $K_{\Delta t_i} \ell _{Y(t_{i+1})}\phi$ satisfying \eqref{eq. sensitivity condition of phi} and moments of particles being bounded from above (see \eqref{eq. tilde hat convergence}).
}
First, \Cref{assumption for full model} and \Cref{assumption non-negativity of tau-leaping algorithm} suggest that any moment of the state vector is bounded, and, therefore, the moments of particles are also bounded.
\begin{proposition}[Adapted from {\cite[Lemma 5.1]{gupta2013unbiased}} and {\cite[Theorem 4.5]{rathinam2016convergence}}]\label{proposition finiteness of moments}
If \Cref{assumption for full model} holds, then for all $p>0$, there exists a positive constant $C_p$ such that
\begin{align*}
\mathbb E\left[ \left. \left\|X^{N,\gamma_1}(t)\right\|^{p}_2 ~\right| \mathcal K= \kappa, X^{N,\gamma_1}(0)=x \right] &\leq
\left( \left\|x\right\|^{p}_2
+ C_p t \right) \text{e}^{C_p t}\\
\mathbb E\left[ \left. \left\|X^{\gamma_1}(t)\right\|^{p}_2 ~\right| \mathcal K = \kappa, X^{\gamma_1}(0)=x \right] &\leq
\left( \left\|x\right\|^{p}_2
+ C_p t \right) \text{e}^{C_p t}
\end{align*}
for any $t>0$ and $(\kappa,x) \in \Theta\times \mathbb R^{n}_{\geq 0}$.
If moreover \Cref{assumption non-negativity of tau-leaping algorithm} holds, then for $p>0$, there exists a constant $C_{p,\tau}$ such that
\begin{align*}
\sup_{t\in[0,T]}
\mathbb E \left[ \left. \left\| X^{N,\gamma_1}_{\tau}(t) \right\|_2^p \right| \mathcal K=\kappa, X^{N,\gamma_1}_{\tau}(t)=x \right]
\leq \left( \left\|x\right\|^{p}_2
+ 1\right) \text{e}^{C_{p,\tau} t }
\end{align*}
for all $(\kappa,x) \in \Theta\times \mathbb R^{n}_{\geq 0}$.
\end{proposition}
\begin{proof}
The results for the full model and the PDMP are adapted from {\cite[Lemma 5.1]{gupta2013unbiased}}; the result for the tau-leaping method is adapted from {\cite[Theorem 4.5]{rathinam2016convergence}}.
\end{proof}
Then, we look at the parameter sensitivity for the CTMC, tau-leaping model, and PDMP.
Given an initial condition $(\kappa,x)$ and a time point $t$, we term
\begin{align*}
\Psi^{N}_{t,g}(\kappa,x) &= \mathbb E \left[ g\left(X^{N,\gamma_1}(t)\right) \left|~\mathcal K\triangleq \kappa, X^{N,\gamma_1}(0)=x\right.\right], \\
\Psi^{N,\tau}_{t,g}(\kappa,x) &= \mathbb E \left[ g\left(X^{N,\gamma_1}_{\tau}(t)\right) \left|~\mathcal K\triangleq \kappa, X^{N}(0)=x\right.\right], \\
\Psi_{t,g}(\kappa,x) &= \mathbb E \left[ g\left(X^{\gamma_1}(t)\right) \left|~\mathcal K\triangleq \kappa, X^{N,\gamma_1}(0)=x\right.\right]
\end{align*}
as the conditional expectations of function $g$ for the CTMC, tau-leaping model, and PDMP, respectively.
Then, the parameter sensitivity for these models are defined, respectively, by $\frac{\partial \Psi^N_{t,g}(\kappa,x) }{\partial \kappa}$, $\frac{\partial \Psi^{N,\tau}_{t,g}(\kappa,x) }{\partial \kappa}$, and $\frac{\partial \Psi_{t,g}(\kappa,x) }{\partial \kappa}$.
The literature \cite{gupta2018estimation} shows that the parameter sensitivity for the full model can be computed by
\begin{small}
\begin{align}
&\frac{\partial \Psi^N_{t,g}(\kappa,x) }{\partial \kappa} \label{eq. parameter sensitivity for SCRNs} \\
&= \sum_{j=1}^{r} \mathbb{E}
\left[
\left.
\int_{0}^{t} \frac{\partial \lambda^{N}_{j}\left(\kappa, X^{N,\gamma_1}(s)\right)}{\partial \kappa}
\Delta^N_j \Psi^N_{t-s,g}\left(\kappa,X^{N,\gamma_1}(s)\right)
\text{d} s
\right| ~\mathcal K=\kappa, X^{N}(0)=x
\right] \notag
\end{align}
\end{small}where $\Delta^N_j \Psi^N_{t-s,g}\left(\kappa,x\right)\triangleq \Psi^N_{t-s,g}\left(\kappa,x+\Lambda^N \zeta_j\right)- \Psi^N_{t-s,g}\left(\kappa,x\right)$.
Similarly, by mathematical induction, we can show that the parameter sensitivity for the tau-leaping model (with a deterministic time-discretization scheme) can be computed by
\begin{small}
\begin{align}
&\frac{\partial \Psi^{N,\tau}_{t,g}(\kappa,x) }{\partial \kappa} \label{eq parameter sensitivity for the tau-leaping method} \\
&= \sum_{j=1}^{r} \mathbb{E}
\Bigg[
\sum_{\tau_i<t} (\tau_{i+1}\wedge t -\tau_i)\frac{\partial \lambda^{N}_{j}\left(\kappa, X^{N,\gamma_1}_\tau (\tau_i)\right)}{\partial \kappa}
\Delta^N_j \Psi^N_{t-s,g}\left(\kappa,X^{N,\gamma_1}_{\tau}(\tau_{i+1})\right) \notag \\
& \qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad
\bigg| ~\mathcal K=\kappa, X^{N,\gamma_1}_{\tau}(0)=x
\bigg]. \notag
\end{align}
\end{small}\textcolor{blue}{(Note that $\frac{\partial \Psi^{N,\tau}_{t,g}(\kappa,x) }{\partial \kappa}$ computes the parameter sensitivity when the tau-leaping model is viewed as the ground truth. It is not necessarily equal or even close to $\frac{\partial \Psi^N_{t,g}(\kappa,x) }{\partial \kappa}$, the parameter sensitivity for the full model.)}
Finally, the literature \cite{gupta2019sensitivity} tells that under \Cref{assumption convergence of the initial condition}, \Cref{assumption for full model}, and \Cref{assumption for PDMPs}, the parameter sensitivity for the PDMP satisfies
\begin{equation}\label{eq. parameter sensitivity PDMP}
\frac{\partial \Psi_{t,g}(\kappa,x) }{\partial \kappa} = \lim_{N\to\infty} \frac{\partial \Psi^N_{t,g}(\kappa,x) }{\partial \kappa}
\end{equation}
if $g$ is bounded, continuously differentiable.
Using these results, we can further prove that the parameter sensitivity for the transition kernel is at most polynomially growing with respect to the state argument (see \Cref{proposition Kphi}).
\begin{proposition}(Polynomially growing rate of the parameter sensitivity for the transition kernel)\label{proposition Kphi}
For any measurable function $\phi$, we denote
\begin{align*}
\left(K^{N,\gamma_1}_t \phi \right)(\kappa, x) &\triangleq \mathbb E_{\mathbb P} \left[ \phi\left(\mathcal K, X^{N,\gamma_1}(t)\right) \left|~\mathcal K=\kappa, X^{N}(0)=x\right.\right], \\
\left(K^{N,\gamma_1,\tau}_t \phi \right)(\kappa, x) &\triangleq \mathbb E_{\mathbb P} \left[ \phi\left(\mathcal K, X^{N,\gamma_1}_{\tau}(t)\right) \left|~\mathcal K=\kappa, X^{N}(0)=x\right.\right], \\
\left(K^{\gamma_1}_t \phi \right)(\kappa, x) &\triangleq \mathbb E_{\mathbb P} \left[ \phi\left(\mathcal K, X^{\gamma_1}(t)\right) \left|~\mathcal K=\kappa, X^{\gamma_1}(0)=x\right.\right].
\end{align*}
If $\phi:\Theta\times \mathbb R^{n}_{\geq 0} \to \mathbb R$
satisfies the requirement in \cref{thm main}, then the following results hold.
\begin{itemize}
\item If \Cref{assumption for full model} holds, then for any $t>0$ there exists a constant $C_{t,\phi}$ such that $\left\| \frac{ \partial \left(K^{N,\gamma_1}_t \phi \right)(\kappa, x) }{\partial \kappa}\right\|_{2}
\leq C_{t,\phi} \left(1+ \|x\|^q_2\right)$ for all $ (\kappa,x) \in \Theta\times \mathbb R^{n}_{\geq 0}$.
\item If \Cref{assumption non-negativity of tau-leaping algorithm} and \Cref{assumption for full model} hold, then for any $t>0$ there exists a constant $C_{t,\phi,\tau}$ such that $\left\| \frac{ \partial \left(K^{N,\gamma_1,\tau}_t \phi \right)(\kappa, x) }{\partial \kappa}\right\|_{2}
\leq C_{t,\phi,\tau} \left(1+ \|x\|^q_2\right)$ for all $ (\kappa,x) \in \Theta\times \mathbb R^{n}_{\geq 0}$.
\item If \Cref{assumption convergence of the initial condition}, \Cref{assumption for full model}, and \Cref{assumption for PDMPs} hold, then for any $t>0$ and $ (\kappa,x) \in \Theta\times \mathbb R^{n}_{\geq 0}$ there holds $ \left\| \frac{ \partial \left(K^{\gamma_1}_t \phi \right)(\kappa, x) }{\partial \kappa}\right\|_{2} \leq C_{t,\phi} \left(1+ \|x\|^q_2\right)$,
where $C_{t,\phi}$ is the same as that in the first result.
\end{itemize}
\end{proposition}
\begin{proof}
We first prove the result for the transition kernel of the full model.
By definition, we can write the derivative of the transition kernel by
\begin{align}
\frac{ \partial K^{N,\gamma_1}_t \phi (\kappa, x) }{\partial \kappa} \notag
=& \mathbb E \left[ \left. \frac{\partial \phi\left(\kappa, X^{N,\gamma_1}(t)\right)}{\partial \kappa} \right|~\mathcal K=\kappa, X^{\gamma_1}(0)=x \right]
+
\left.\frac{\partial \Psi^{N}_{t,\phi(\kappa, \cdot)}(\theta,x) }{\partial \theta}\right|_{\theta=\kappa}.
\end{align}
By \cref{eq. sensitivity condition of phi} and \cref{proposition finiteness of moments}, we can find a positive constant $C^1_{t,\phi}$ such that
\begin{equation}\label{eq. sensitivity analysis help 1}
\mathbb E \left[ \left. \left\| \frac{\partial \phi\left(\kappa, X^{N,\gamma_1}(t)\right)}{\partial \kappa}\right\|_2 \right|~\mathcal K=\kappa, X^{\gamma_1}(0)=x \right]
\leq C^1_{t,\phi}\left( 1
+ \|x\|^q_2 \right)
\end{equation}
for all $(\kappa,x) \in \Theta\times \mathbb R^{n}_{\geq 0}$.
Also, by \eqref{eq. parameter sensitivity for SCRNs} and the boundedness of $\phi$, there holds
\begin{align}
\left\| \left.\frac{\partial \Psi^{N}_{t,\phi(\kappa, \cdot)}(\theta,x) }{\partial \theta}\right|_{\theta=\kappa} \right\|_2
&\leq 2\|\phi\|_{\infty}\sum_{j=1}^{r} \left\|\mathbb{E}
\left[
\left.
\int_{0}^{t} \frac{\partial \lambda^{N}_{j}\left(\kappa, X^{N,\gamma_1}(s)\right)}{\partial \kappa}
\text{d} s
\right| ~\mathcal K=\kappa, X^{N}(0)=x
\right]\right\|_2 \notag \\
& \leq 2\|\phi\|_{\infty}\sum_{j=1}^{r} \mathbb{E}
\left[
\left.
\int_{0}^{t} \left\| \frac{\partial \lambda^{N}_{j}\left(\kappa, X^{N,\gamma_1}(s)\right)}{\partial \kappa} \right\|_2
\text{d} s
\right| ~\mathcal K=\kappa, X^{N}(0)=x
\right] \notag
\end{align}
where the second inequality follows from Jensen's inequality and the triangle inequality.
Therefore by \Cref{proposition finiteness of moments} and \Cref{assumption for full model}, we can find a constant $C^2_{t,\phi}$ such that
\begin{equation}\label{eq. sensitivity analysis help 2}
\left\| \left.\frac{\partial \Psi^{N}_{t,\phi(\kappa, \cdot)}(\theta,x) }{\partial \theta}\right|_{\theta=\kappa}\right\|_2
\leq C^2_{t,\phi}\left( 1
+ \|x\|^q_2 \right)
\qquad \forall (\kappa,x) \in \Theta\times \mathbb R^{n}_{\geq 0}.
\end{equation}
Finally, by combing \eqref{eq. sensitivity analysis help 1} and \eqref{eq. sensitivity analysis help 2}, we prove the result for the transition kernel of the full model (the first result).
By the same argument, we can also prove the second result.
For the third result, we can learn from \eqref{eq. parameter sensitivity PDMP} that $\Psi_{t,\phi(\kappa, \cdot)}(\theta,x)$ satisfies \eqref{eq. sensitivity analysis help 2}.
Moreover, by \cref{eq. sensitivity condition of phi} and \cref{proposition finiteness of moments}, we can show that the quantity $\phi\left(\kappa, X^{\gamma_1}(t)\right)$ also satisfies \eqref{eq. sensitivity analysis help 1}.
Therefore, the third result is proven.
\end{proof}
\begin{remark}[Analysis of the error introduced by artificial evolution]{\label{remark bias}}
\textcolor{blue}{
Now, we use the obtained propositions to show that the second term on the right hand side of \eqref{eq. bias caused by artificial evolution} is of the order $1/\sqrt{M}$.}
\textcolor{blue}{
Note that if $\phi$ satisfies the requirement in \Cref{thm main}, the function $\ell_{y} \phi$ (for any $y\in\mathbb R^{m}$) also satisfies this requirement, because the likelihood function $\ell_{y}$ is bounded and has no dependence on the argument $\kappa$.
Consequently, by \cref{proposition Kphi}, the function $K_{\Delta t_i} \ell _{Y(t_{i+1})}\phi$ also satisfies the requirement in \Cref{thm main} under the proposed conditions.
Moreover, we can conclude by \cref{proposition finiteness of moments} that any moment of particles is bounded from above.
Finally, applying these results to \eqref{eq. tilde hat convergence}, we can find a positive constant $\hat C_{y}$ such that
\begin{align*}
\mathbb E\left[
\big| \text{the second term on the right hand side of }\eqref{eq. bias caused by artificial evolution} \big|^4 \left| Y^{N,\gamma_1}(t_{i+1})=y \right. \right]
\leq \frac{\hat C_{y}}{M^2},
\end{align*}
which suggests that the bias contributed by artificial evolution is of order $\frac{1}{\sqrt{M}}$.
}
\end{remark}
\subsection{Proving the main result by mathematical induction} {\label{subsection proof main result}}
\textcolor{blue}{In this subsection, we prove \cref{thm main} using the result obtained in the previous section.}
\textcolor{blue}{Before proving the main result, we first analyze the error caused by model reduction and show that it vanishes as the reduced model approaches the full model.
}
Here, we define observations for the tau-leaping algorithm and the PDMP by
\begin{align*}
Y^{N,\gamma_1,\tau}(t_i)&= h\left(X^{N,\gamma_1}_{\tau}(t_i)\right) + W(t_i) \qquad && \forall i\in \mathbb N_{>0}, \\
Y^{\gamma_1}(t_i)&= h\left(X^{\gamma_1}_{\tau}(t_i)\right) + W(t_i) \qquad && \forall i\in \mathbb N_{>0},
\end{align*}
and term $\mathcal Y^{N,\gamma_1,\tau}_{t_i}$ and $\mathcal Y^{\gamma_1}_{t_i}$ as the $\sigma$-algebras generated, respectively, by $Y^{N,\gamma_1,\tau}(\cdot)$ and $Y^{\gamma_1}(\cdot)$ up to the time $t_i$.
Note that for any bounded measurable function $\phi$, the conditional expectations $\mathbb E \left[ \phi \left( \mathcal K, X^{N,\gamma_1}(t_i) \right) \big| \mathcal Y^{N,\gamma_1}_{t_i} \right]$, $\mathbb E \left[ \phi \left( \mathcal K, X^{N,\gamma_1}_{\tau}(t_i) \right) \big| \mathcal Y^{N,\gamma_1,\tau}_{t_i} \right]$, and $\mathbb E \left[ \phi \left( \mathcal K, X^{\gamma_1}(t_i) \right) \big| \mathcal Y^{\gamma_1}_{t_i} \right]$, are respectively $\mathcal Y^{N,\gamma_1}_{t_i}$-measurable, $\mathcal Y^{N,\gamma_1,\tau}_{t_i}$-measurable, and $\mathcal Y^{\gamma_1}_{t_i}$-measurable.
So, there exist measurable functions $\hat f^{N,\gamma_1}_{\phi,t_i}$, $\hat f^{N,\gamma_1,\tau}_{\phi,t_i}$, and $\hat f^{\gamma_1}_{\phi,t_i}$, such that
\begin{align*}
\hat f^{N,\gamma_1}_{\phi,t_i}\left( Y^{N,\gamma_1}_{1:i}\right) &= \mathbb E_{\mathbb P} \left[ \phi \left( \mathcal K, X^{N,\gamma_1}(t_i) \right) \big| \mathcal Y^{N,\gamma_1}_{t_i} \right] \\
\hat f^{N,\gamma_1,\tau}_{\phi,t_i}\left( Y^{N,\gamma_1,\tau}_{1:i}\right) &= \mathbb E_{\mathbb P} \left[ \phi \left( \mathcal K, X^{N,\gamma_1}_{\tau}(t_i) \right) \big| \mathcal Y^{N,\gamma_1,\tau}_{t_i} \right] \\
\hat f^{\gamma_1}_{\phi,t_i}\left( Y^{\gamma_1}_{1:i}\right) &= \mathbb E_{\mathbb P} \left[ \phi \left( \mathcal K, X^{\gamma_1}(t_i) \right) \big| \mathcal Y^{\gamma_1}_{t_i} \right]
\end{align*}
where $Y^{N,\gamma_1}_{1:i}\triangleq \left(Y^{N,\gamma_1}(t_1),\dots, Y^{N,\gamma_1}(t_i)\right)$, $Y^{N,\gamma_1,\tau}_{1:i}\triangleq \left(Y^{N,\gamma_1,\tau}(t_1),\dots, Y^{N,\gamma_1,\tau}(t_i)\right)$, and $Y^{\gamma_1}_{1:i}\triangleq \left(Y^{\gamma_1}(t_1),\dots, Y^{\gamma_1}(t_i)\right)$.
Here, $ \hat f^{N,\gamma_1,\tau}_{\phi,t_i}\left( Y^{N,\gamma_1,\tau}_{1:i}\right)$ and $\hat f^{\gamma_1}_{\phi,t_i}\left( Y^{\gamma_1}_{1:i}\right)$ are exact filters for the tau-leaping model and PDMP, respectively.
The following proposition states that $\hat f^{N,\gamma_1,\tau}_{\phi,t_i}\left( Y^{N,\gamma_1}_{1:i}\right) $ and $\hat f^{\gamma_1}_{\phi,t_i}\left( Y^{N,\gamma_1}_{1:i}\right)$ converge in probability to the exact filter $\pi^{N,\gamma_1}_{t_i}(\phi)$ as $|\tau|\to 0$ and $N\to\infty$, respectively.
Therefore, to show the final result, we only need to check that $\bar \pi^{N,\gamma_1}_{M,t_i}(\phi)$, $\bar \pi^{N,\gamma_1,\tau}_{M,t_i}$, and $\bar \pi^{N,\gamma_1, H}_{M,t_i}$ converge respectively to $\hat f^{N,\gamma_1}_{\phi,t_i}\left( Y^{N,\gamma_1}_{1:i}\right)$, $\hat f^{N,\gamma_1,\tau}_{\phi,t_i}\left( Y^{N,\gamma_1}_{1:i}\right)$, and $\hat f^{\gamma_1}_{\phi,t_i}\left( Y^{N,\gamma_1}_{1:i}\right)$ as $M\to\infty$.
\begin{proposition}\label{prop fang}[Adapted from {\cite{fang2021stochastic}}]~
\begin{itemize}
\item If \Cref{assumption non-negativity of tau-leaping algorithm}, \Cref{assumption convergence of tau-leaping algorithm}, and \Cref{assumption for full model} hold, then for any $i\in\mathbb N_{>0}$, $N>0$, and bounded continuous function $\phi$, there is the convergence result $\hat f^{N, \gamma_1, \tau}_{\phi,t_i} \left(Y^{N,\gamma_1}_{1:i}\right) \stackrel{\mathbb P}{\to} \pi^{N,\gamma_1}_{t_i}(\phi)$ as $|\tau|\to0$.
\item If \Cref{assumption convergence of the initial condition} and \Cref{assumption for full model} hold, then there is the convergence result $\lim_{N\to\infty}\mathbb P \left( \left|\hat f^{\gamma_1}_{\phi,t_i} \left(Y^{N,\gamma_1}_{1:i}\right) - \pi^{N,\gamma_1}_{t_i}(\phi)\right| >\delta\right)=0$
for any $i\in\mathbb N_{>0}$, $\delta>0$ and bounded continuous function $\phi$.
\end{itemize}
\end{proposition}
\begin{proof}
The second result has been proven in \cite{fang2021stochastic}; following the technique in \cite{fang2021stochastic}, one can easily prove the first result.
Here, we give the outline of the proof.
Let's define Girsanov random variables $Z^{N,\gamma_1}(t_i)\triangleq \prod_{j=1}^i g\left(X^{N,\gamma_1}(t_j), Y^{N,\gamma_1}(t_j)\right)$ and $Z^{N,\gamma_1,\tau}(t_i)\triangleq \prod_{j=1}^i g\left(X^{N,\gamma_1}_{\tau}(t_j), Y^{N,\gamma_1,\tau}(t_j)\right)$ for the full dynamical model and the tau-leaping algorithm, respectively, where $g(x,y)=\exp\left(\sum_{\ell=1}^m h_{\ell}(x)y_{\ell}-h^2_{\ell}(x)/2\right)$.
Since the reciprocals of $Z^{N,\gamma_1}(t_i)$ and $Z^{N,\gamma_1,\tau}(t_i)$ are both martingales, we can then define reference probability $\mathbb P^{N,\gamma_1}$ and $\mathbb P^{N,\gamma_1,\tau}$ by
\begin{equation*}
\left.\frac{\text{d} \mathbb P^{N,\gamma_1}}{\text{d} \mathbb P}\right|_{\mathcal F_{t_i}} = \left(Z^{N,\gamma_1}(t_i)\right)^{-1}
\qquad
\text{and}
\qquad
\left.\frac{\text{d} \mathbb P^{N,\gamma_1,\tau}}{\text{d} \mathbb P}\right|_{\mathcal F_{t_i}} = \left(Z^{N,\gamma_1,\tau}(t_i)\right)^{-1}.
\end{equation*}
In the reference probability, the process and the observation are independent, and the observations become a sequence of independent $m$-variate standard Gaussian variables.
According to \cite[Theorem 2.6]{calzolari2006approximation}, the first result holds if for any $i>0$, we can construct random processes $\tilde X^{N,\gamma_1}(\cdot)$, $\tilde X^{N,\gamma_1,\tau}(\cdot)$, $\tilde Y (\cdot)$, $\tilde Z^{N,\gamma_1}(\cdot)$, and $Z^{N,\gamma_1,\tau}(\cdot)$ on a common probability space $\left(\tilde \Omega, \tilde {\mathcal F}, \mathbb Q \right)$ such that the following conditions hold.
\begin{enumerate} [label=a.\arabic*]
\item Random variables $\left( \tilde X^{N,\gamma_1}(\cdot), \tilde Y (\cdot), \tilde Z^{N,\gamma_1}(t_i) \right) \in \mathbb D_{\mathbb R^{n}}[0,t_i] \times \mathbb R^{m\times i} \times \mathbb R$ have the same law as $\left( X^{N,\gamma_1}(\cdot), Y^{N,\gamma_1}(\cdot), Z^{N,\gamma_1}(t_i) \right)$ on $\left(\Omega,\mathcal F, \mathbb P^{N,\gamma_1} \right)$. \label{a.1}
\item Random variables $\left( \tilde X^{N,\gamma_1,\tau}(\cdot), \tilde Y (\cdot), \tilde Z^{N,\gamma_1,\tau}(t_i) \right) \in \mathbb D_{\mathbb R^{n}}[0,t_i] \times \mathbb R^{m\times i} \times \mathbb R$ have the same law as $\left( X^{N,\gamma_1}_{\tau}(\cdot), Y^{N,\gamma_1,\tau}(\cdot), Z^{N,\gamma_1,\tau}(t_i) \right)$ on $\left(\Omega,\mathcal F, \mathbb P^{N,\gamma_1,\tau} \right)$. \label{a.2}
\item $\hat f^{N,\gamma_1,\tau}_{\phi,t_i} \left(\tilde Y_{1:i}\right) \stackrel{\mathbb P}{\to} \hat f^{N,\gamma_1}_{\phi,t_i} \left(\tilde Y_{1:i}\right) $ as $|\tau|\to0$, where $\tilde Y_{1:i}=\left(\tilde Y(t_1),\dots, \tilde Y(t_i)\right)$. \label{a.3}
\item $\lim_{|\tau|\to\infty} \mathbb E_{\mathbb Q} \left[ \left| \tilde Z^{N,\gamma_1}(t_i) - \tilde Z^{N,\gamma_1,\tau}(t_i) \right|\right]=0$. \label{a.4}
\end{enumerate}
These random variables and the common probability space can be constructed as follows.
By the Skorokhod representation theorem and \Cref{assumption convergence of tau-leaping algorithm}, we can find a common probability space $\left( \tilde \Omega_1, \tilde {\mathcal F}_1, \mathbb Q_1 \right)$, on which are defined $\mathbb D_{\mathbb R^{n}}[0,t_i]$-valued variables $\tilde X^{N,\gamma_1}(\cdot)$ and $\tilde X^{N,\gamma_1,\tau}(\cdot)$ that have the same law as $X^{N,\gamma_1}(\cdot)$ and $X^{N,\gamma_1}_{\tau}(\cdot)$ respectively and, moreover, satisfy $\lim_{|\tau|\to\infty}\tilde X^{N,\gamma_1}_\tau(\cdot) = X^{N,\gamma_1}(\cdot)$ $\mathbb Q$-almost surely.
Let us term $\left\{ \tilde Y (t_j) \right\}_{1\leq j \leq i}$ as a sequence of independent $m$-variate standard Gaussian random variables and $\left( \tilde \Omega_2, \tilde {\mathcal F}_2, \mathbb Q_2 \right)$ as the probability space on which they are defined.
Finally, the probability space $\left(\tilde \Omega, \tilde {\mathcal F}, \mathbb Q \right)$ is defined by the product of $\left( \tilde \Omega_1, \tilde {\mathcal F}_1, \mathbb Q_1 \right)$ and $\left( \tilde \Omega_2, \tilde {\mathcal F}_2, \mathbb Q_2 \right)$, and the Girsanov random variables are defined by $\tilde Z^{N,\gamma_1}(t_i)\triangleq \prod_{j=1}^i g\left(\tilde X^{N,\gamma_1}(t_j), \tilde Y(t_j)\right)$ and $\tilde Z^{N,\gamma_1,\tau}(t_i)\triangleq \prod_{j=1}^i g\left(\tilde X^{N,\gamma_1,\tau}(t_j), \tilde Y(t_j)\right)$.
Obviously, such random variables satisfy \ref{a.1} and \ref{a.2}. Moreover, \ref{a.4} follows immediately from the almost sure convergence of $\tilde X^{N,\gamma_1,\tau}(\cdot)$ to $\tilde X^{N,\gamma_1}(\cdot)$. Finally, \ref{a.3} can be proven by the Kallianpur-Striebel formula together with the aforementioned convergence result.
\end{proof}
Finally, we prove the main theorem using mathematical induction.
\begin{proof}[The Proof of \cref{thm main}]
Here, we prove the result using the simplified notations in \eqref{eq. notations}.
Moreover, for a given RPF, we briefly denote the associated measurable function ($\hat f^{N,\gamma_1}_{\phi,t_i}$, $\hat f^{N,\gamma_1,\tau}_{\phi,t_i}$, or $\hat f^{\gamma_1}_{\phi,t_i}$) by $\hat f_{\phi,t_i}$.
For $t_0$, we term $\hat f_{\phi, t_0} (Y^{N,\gamma_1}_{1:0})$ as the expectation of $\phi$ with respect to the initial distribution.
\Cref{prop fang} tells that to prove the result, we only need to show $\bar \pi_{M,t_i}(\phi) \to \hat f_{\phi, t_i} (Y^{N,\gamma_1}_{1:i})$ in probability as $M\to\infty$.
In the following, we prove that this convergence actually holds almost surely.
The proof is in the same spirit as that in the classical literature \cite{crisan2001particle,del2000branching,bain2008fundamentals}, with the difference that we need to additionally consider the moment of particles (see \ref{claim 1}).
In the sequel, we first make a few claims (see \ref{claim 1}, \ref{claim 2}, and \ref{claim 3} in the following paragraph) and then prove them by mathematical induction.
Here, proving \ref{claim 2} requires us to look at the finite sampling error (the first term on the right hand side of \eqref{eq. bias caused by artificial evolution}), and proving \ref{claim 3} requires us to analyze the error caused by artificial evolution and resampling (the second and third term on the right hand side of \eqref{eq. bias caused by artificial evolution}).
Finally, by applying the Borel-Cantelli lemma to \ref{claim 2}, we can show that $\bar \pi_{M,t_i}(\phi) \to \hat f_{\phi, t_i} (Y^{N,\gamma_1}_{1:i})$ almost surely as $M\to\infty$ for any fixed $N$ and $i$, which together with \cref{prop fang} proves the result.
Now, we only need to show that the following results hold.
\begin{enumerate}[label=R.\arabic*]
\item For $i=0$ and any $p>0$, there exists a positive constant $C_{0,p}$ such that the inequality $\mathbb E \left[ \sum_{j=1}^M \frac{\left\| \bar x_j(t_i)\right\|^p_2}{M}\right]\leq C_{0,p}$ holds; for any $i\in\mathbb N_{>0}$, $p>0$, and $y_{1:i}\in\mathbb R^{m\times i}$, there exits a positive constant $C_{i,p,y_{1:i}}$ such that $$\mathbb E \left[ \left. \sum_{j=1}^M \frac{\left\| \bar x_j(t_i)\right\|^p_2}{M} \right| Y^{N,\gamma_1}_{1:i}=y_{1:i} \right]\leq C_{i,p,y_{1:i}}.$$ \label{claim 1}
\item For any $i\in\mathbb N_{>0}$, $y_{1:i}\in\mathbb R^{m\times i}$ and function $\phi$ satisfying the requirement in \Cref{thm main}, there exists a positive constant $\bar C_{i,\phi,y_{1:i}}$ such that $$\mathbb E \left[ \left. \left| \bar\pi_{M,t_i}(\phi) - \hat f_{\phi, t_i} \left(Y^{N,\gamma_1}_{1:i}\right)\right|^4 \right| Y^{N,\gamma_1}_{1:i}=y_{1:i} \right] \leq\frac{\bar C_{i,\phi,y_{1:i}}}{M^2}.$$ \label{claim 2}
\item For $i=0$ and any function $\phi$ satisfying the requirement in \Cref{thm main}, there exists a constant $\tilde C_{0}$ such that $\mathbb E \left[ \left|\tilde \pi_{M,t_0}(\phi) - \hat f_{\phi, t_0} \left(Y^{N,\gamma_1}_{1:i}\right)\right|^4\right] \leq \frac{\tilde C_{0}}{M^2}$;
for any $i\in\mathbb N_{>0}$, $y_{1:i}\in\mathbb R^{m\times i}$ and function $\phi$ satisfying the requirement in \Cref{thm main}, there exists a positive constant $\tilde C_{i,\phi,y_{1:i}}$ such that $$\mathbb E \left[ \left. \left| \tilde\pi_{M,t_i}(\phi) - \hat f_{\phi, t_i} \left(Y^{N,\gamma_1}_{1:i}\right)\right|^4 \right| Y^{N,\gamma_1}_{1:i}=y_{1:i} \right] \leq \frac{\tilde C_{i,\phi,y_{1:i}}}{M^2}.$$\label{claim 3}
\end{enumerate}
For $i=0$, \ref{claim 1} and \ref{claim 3} follow immediately from the third claim of \Cref{assumption for full model} and the fact that initial particles are independently sampled from the initial distribution.
Then, we prove \ref{claim 1}, \ref{claim 2}, and \ref{claim 3} for a positive integer $i+1$, provide that \ref{claim 1} and \ref{claim 3} hold for the nonnegative integer $i$.
\textit{\ref{claim 1} (for the integer $i+1$):} By \cref{proposition finiteness of moments} and \ref{claim 1} (for the integer $i$), we can find a positive constant $\tilde C_{i+1,p,y_{1:i}}$ for each $p>0$ such that
$$\mathbb E \left[ \sum_{j=1}^M \frac{{\left\| x_j(t_{i+1})\right\|^p_2}}{M} \Bigg| Y^{N,\gamma_1}_{1:i+1}=y_{1:i+1}\right]\leq \tilde C_{i+1,p,y_{1:i}}.$$
Due to the boundedness of $h(\cdot)$, the weights $w_j(t_{i+1})$ are uniformly upper bounded by $\frac{1}{M}\exp\left(2\|h\|_{\infty} \|y_{i+1}\|_1+m\|h\|^2_{\infty}/2\right)$, where $y_{i+1}$ is the last column of $y_{1:i+1}$.
Therefore, by choosing $C_{i+1,p,y_{1:i}}=\tilde C_{i+1,p,y_{1:i}} \times \exp\left(2\|h\|_{\infty} \|y_{i+1}\|_1+m\|h\|^2_{\infty}/2\right) $, we have the relation $\mathbb E \left[ \sum_{j=1}^M w_j(t_{i+1}){\left\| x_j(t_{i+1})\right\|^p_2} \left| Y^{N,\gamma_1}_{1:i+1}=y_{1:i+1} \right. \right] \leq C_{i+1,p,y_{1:i}}.$
Note that the resampling does not change the expectation of the empirical mean, so \ref{claim 1} holds for the integer $i+1$.
\textit{\ref{claim 2} (for the integer $i+1$):}
According to \ref{claim 3} (for the integer $i$), the key is to analyze the finite sampling error (the first term on right hand side of \eqref{eq. bias caused by artificial evolution}), and we perform the analysis using the technique in \cite{crisan2001particle}.
Since we already know $\tilde \pi_{M,t_{i+1}}(\phi)$ converges to $\hat f_{\phi, t_{i+1}}\left(y_{1:i+1}\right)$, we directly compare $\bar\pi_{M,t_{i+1}}(\phi)$ and $\hat f_{\phi, t_{i+1}}\left(y_{1:i+1}\right)$ in this analysis.
For any function $\phi$ satisfying the requirement in \Cref{thm main}, we define
\begin{align*}
\bar \rho_{M,t_i} (\phi) = {\sum_{j=1}^{M} \phi(\kappa_j, x_j(t_i))} \qquad \forall i\in\mathbb N,
\end{align*}
which is the estimate of the function $\phi$ right before the time point $t_i$.
For any event $\left\{\mathcal Y^{N,\gamma_1}_{1:i+1}=y_{1:i+1}\right\}$, this quantity satisfies
\begin{align}\label{eq. estimate bar rho}
&\mathbb E \left[ \left. \left| \bar \rho_{M,t_i+1}(\phi) - \tilde f_{(K_{\Delta t_i}\phi), t_i} \left(Y^{N,\gamma_1}_{1:i}\right)\right|^4 \right| Y^{N,\gamma_1}_{1:i}+1=y_{1:i+1} \right] \\
&\leq 8\mathbb E \left[ \left. \left| \bar \rho_{M,t_i+1}(\phi) - \tilde \pi_{M,t_i}(K_{\Delta t_i} \phi)\right|^4 \right| Y^{N,\gamma_1}_{1:i+1}=y_{1:i+1} \right] \notag \\
&\quad+ 8 \mathbb E \left[ \left. \left| \tilde \pi_{M,t_i}(K_{\Delta t_i} \phi) - \tilde f_{(K_{\Delta t_i}\phi), t_i} \left(Y^{N,\gamma_1}_{1:i}\right)\right|^4 \right| Y^{N,\gamma_1}_{1:i}=y_{1:i} \right] \notag \\
& \leq \frac{3\|\phi\|^4_{\infty}}{M^2}+\frac{\tilde C_{i,(K_{\Delta t_i} \phi),y_{1:i}}}{M^2} \notag
\end{align}
where the first inequality follows from $(a+b)^4\leq \left( 2a^2+2b^2\right)^2\leq 8 \left(a^2 +b^2\right)$, and the second inequality follows from the independent increments of particles in the sampling step, \cref{proposition Kphi} (which implies $K_t \phi$ to satisfy \eqref{eq. sensitivity condition of phi}), and \ref{claim 3}.
Moreover, by \cite[(2.10)]{crisan2001particle}, there is
\begin{align*}
&\left|\bar\pi_{M,t_{i+1}}(\phi) - \hat f_{\phi, t_{i+1}}\left(y_{1:i+1}\right)\right| \\
&\leq
\frac{\left\|\phi\right\|_{\infty}}{\hat f_{ (K_{\Delta t_i}\ell_{y_{i+1}}), t_{i}}\left(y_{1:i}\right)}
\left|\bar \rho_{M,t_i+1}\left( \ell_{y_{i+1}}\right)
-
\hat f_{(K_{\Delta t_i}\ell_{y_{i+1}}), t_{i}}\left(y_{1:i}\right)\right| \\
& \quad~ +
\frac{1}{\hat f_{ (K_{\Delta t_i} \ell_{y_{i+1}}), t_{i}}\left(y_{1:i}\right)}
\left|\bar \rho_{M,t_{i+1}}\left( \ell_{y_{i+1}}\phi\right)
-
\hat f_{(K_{\Delta t_i} \ell_{y_{i+1}}\phi), t_{i}}\left(y_{1:i}\right)\right|.
\end{align*}
Recall that $(a+b)^4\leq 8 \left(a^4+b^4\right)$ for any real $a$ and $b$, and $\hat f_{(K_{\Delta t_i} \ell_{y_{i+1}}), t_i}\left(y_{1:i}\right)$ is lower bounded by a positive constant due to the boundedness of $h(\cdot)$.
Therefore, we can find a constant $\bar C_{\ell,\phi,y_{i+1}}$ such that
\begin{align*}
&\left|\bar\pi_{M,t_{i+1}}(\phi) - \hat f_{\phi, t_{i+1}}\left(y_{1:i+1}\right)\right| ^4 \\
&\leq
\bar C_{\ell,\phi,y_{i+1}}
\left|\bar \rho_{M,t_i+1}\left( \ell_{y_{i+1}}\right)
-
\hat f_{(K_{\Delta t_i} \ell_{y_{i+1}}), t_{i}}\left(y_{1:i}\right)\right|^4 \\
& \quad~ +\bar C_{\ell,\phi,y_{i+1}}
\left|\bar \rho_{M,t_{i+1}}\left( \ell_{y_{i+1}}\phi\right)
-
\hat f_{(K_{\Delta t_i}\ell_{y_{i+1}}\phi), t_{i}}\left(y_{1:i}\right)\right|^4.
\end{align*}
conditioned on $\left\{\mathcal Y^{N,\gamma_1}_{1:i+1}=y_{1:i+1}\right\}$.
Finally, applying \eqref{eq. estimate bar rho} to this formula, we show
\ref{claim 2} for the integer $i+1$.
\textit{\ref{claim 3} (for the integer $i+1$):}
According to the result \textit{\ref{claim 2} (for the integer $i+1$)}, in this part,
we need to analyze the error caused by artificial evolution and resampling, i.e., the second and third terms on the right hand side of \eqref{eq. bias caused by artificial evolution}.
If we use the multinomial resampling, then all the resampled particles are independently and identically distributed, and, therefore, we can conclude
\begin{equation}\label{eq convergence resampling 1}
\mathbb E \left[ \left. \left| \hat\pi^{N,\gamma_1}_{M,t_{i+1}}(\phi) - \bar \pi^{N,\gamma_1}_{M,t_{i+1}}(\phi)\right|^4 \right| Y^{N,\gamma_1}_{1:i+1}=y_{1:i+1} \right] \leq
\frac{\hat C \|\phi\|^4_{\infty}}{M^2}
\end{equation}
for a certain constant $\hat C$.
By \cite[Exerice 9.1]{bain2008fundamentals} which states that residual resampling introduces the minimum noise among all resampling methods, we can also conclude the above relation for residual resampling.
Moreover, by \ref{claim 1} (for the integer $i+1$) and \eqref{eq. tilde hat convergence}, we can also find a positive constant $\hat C_{i,\phi,y_{1:i}}$ such that
\begin{align}\label{eq convergence resampling 2}
\mathbb E \left[ \left. \left| \tilde \pi^{N,\gamma_1}_{M,t_{i+1}}(\phi) - \hat \pi^{N,\gamma_1}_{M,t_{i+1}}(\phi)\right|^4 \right| Y^{N,\gamma_1}_{1:i+1}=y_{1:i+1} \right]
\leq
\frac{\hat C_{i,\phi,y_{1:i}}}{M^2}
\end{align}
Finally, \ref{claim 3} (for the integer $i+1$) follows from \eqref{eq convergence resampling 1} and \eqref{eq convergence resampling 2}.
\end{proof}
\section{Proof of the main theorem}
In this section, we show the proof of \cref{thm main}.
First, we define an auxiliary process $\left(\tilde {\mathcal K}^{M,\gamma_1}(t), \tilde X^{M,\gamma_1}(t) \right)$ as follows
\begin{align*}
&\tilde X^{M,\gamma_1}(t)= \lim_{N\to\infty} X^{N}(0)+\sum_{j : \gamma_1+\tilde \rho_j>0} \int_0^t \lambda'_j \left(\tilde {\mathcal K}^{M,\gamma_1}(t),\tilde X^{M,\gamma_1}(s)\right) D^0\left(v'_{\cdot j}-v_{\cdot j}\right) \text{d} s \\
&\qquad\qquad\quad+\sum_{j : \gamma_1+\tilde \rho_j=0 } R_j\left( \int_0^t \lambda'_j \left(\tilde {\mathcal K}^{M,\gamma_1}(t), \tilde X^{M,\gamma_1}(s)\right) D^{\gamma_1+\tilde \rho_j}\left(v'_{\cdot j}-v_{\cdot j}\right) \text{d} s \right) \\
&\left\{\begin{array}{cclr}
\dot{\tilde {\mathcal K}}^{M,\gamma_1} (t) &=& 0 & t\notin \left\{ t_i\right\}_{i\in\mathbb N}\\
\tilde {\mathcal K}^{M,\gamma_1} (t) & \sim& \eta_{M} \left(\cdot\left| \tilde {\mathcal K}^{M,\gamma_1} (t^-)\right.\right) & t\in\left\{ t_i\right\}_{i\in\mathbb N}
\end{array}
\right.
\end{align*}
where $\tilde {\mathcal K}^{M,\gamma_1} (0)=\mathcal K$.
The above process can be seen as a variants of the hybrid model \eqref{eq. reduced model at the first time scale}, where the model parameter is perturbed by some artificial noise at time points $\left\{ t_i\right\}_{i\in\mathbb N}$.
Also, we define the observation of the process by
\begin{equation*}
\tilde Y^{M,\gamma_1}(t_i) = h\left(\tilde X^{M,\gamma_1}(t_i)\right) + W(t_i)
\end{equation*}
and $\tilde {\mathcal Y}^{M,\gamma_1}_{t_i}$ be the $\sigma$-algebra generated by the above readouts up to the time $t_i$.
For a bounded measurable function $\phi$, let $\tilde \pi ^{M,\gamma_1}_{t_i}(\phi)\triangleq \mathbb E_{\mathbb P} \left[ \left(\tilde {\mathcal K}^{M,\gamma_1}(t_i^-), \tilde X^{M,\gamma_1}(t_i) \right) \left| \tilde {\mathcal Y}^{M,\gamma_1}_{t_i} \right. \right]$ and $\hat f^{M}_{\phi,t}$ be a mapping such that
\begin{equation*}
\hat f^{M}_{\phi,t_i} \left( \tilde {\mathcal Y}^{M,\gamma_1}_{1:i} \right) = \tilde \pi ^{M,\gamma_1}_{t_i}(\phi)
\end{equation*}
where $\tilde {\mathcal Y}^{M,\gamma_1}_{1:i}=\left( \tilde {\mathcal Y}^{M,\gamma_1}_{t_1},\dots, \tilde {\mathcal Y}^{M,\gamma_1}_{t_i} \right)$.
Obviously, the constructed filter $\bar \pi^{N,\gamma_1}_{M,t_i}(\phi)$ is an SIR particle filter for the above system, and by \cite[Corollary 2.4.4]{crisan2001particle}, the filter $\bar \pi^{N,\gamma_1}_{M,t_i}(\phi)$ converges almost surely to $\hat f^{M}_{\phi,t_i} \left( {\mathcal Y}^{N,\gamma_1}_{1:i} \right)$ as $M\to\infty$ for any fixed $N$.
This fact together with \cref{pro. fang} suggests that to prove the main theorem, we only need to check $\hat f^{M}_{\phi,t_i} \left( {\mathcal Y}^{N,\gamma_1}_{1:i} \right) \to \hat f_{\phi, t_i} \left( {\mathcal Y}^{N,\gamma_1}_{1:i} \right) $ in probability as $M\to\infty$ for any fixed $N$.
In the following, we will show the result by borrowing \cite[Lemma 2.37]{del2000branching} which requires the transition kernel to be globally Lipschitz.
In our setting, the transition kernel of the hybrid model is not necessarily globally Lipschitz as the propensity is only locally Lipschitz (see \cref{cond. Lipschitz continuity of propensities}).
Therefore, we cannot borrow \cite[Lemma 2.37]{del2000branching} directly.
Instead, we look at the stopped process, whose transition kernel is then globally Lipschitz, and show the filters for such stopped processes converges by using \cite[Lemma 2.37]{del2000branching}.
Finally, by passing the stopping time to infinity, we can show the final result.
We first show that the transition kernel of the stopped hybrid process is globally Lipschitz.
Let us term $\tau_c$ as the first passage time of the process $X^{\gamma_1}(t)$ of existing a compact region $\{x| \left\|x\right\|_2\leq c\}$ and $\tau^M_c$ as the first passage time of $\tilde X^{M,\gamma_1}(t)$ of existing the same compact region.
According to \cite[Lemma 10.2]{bain2008fundamentals}, for any $i\in\mathbb N_{>0}$ and $y\in\mathbb R^{m\times i}$, there hold
\begin{align}
\hat f_{\phi, t_i}(y) &= \frac{\mathbb E_{\mathbb P}\left[ \prod_{j=1}^i \ell_y\left(\mathcal K, X^{\gamma_1}(t_j)\right) \phi \left(\mathcal K, X^{\gamma_1}(t_i)\right) \right]}{\mathbb E_{\mathbb P}\left[ \prod_{j=1}^i \ell_y\left(\mathcal K, X^{\gamma_1}(t_j)\right) \right]}, \label{eq. Kallianpur-Striebel formula hybrid model} \\
\hat f^M_{\phi, t_i}(y) &= \frac{\mathbb E_{\mathbb P}\left[ \prod_{j=1}^i \ell_y\left(\tilde {\mathcal K}^{M,\gamma_1}(t_j), \tilde X^{M,\gamma_1}(t_j) \right) \phi \left(\tilde {\mathcal K}^{M,\gamma_1}(t_i), \tilde X^{M, \gamma_1}(t_i)\right) \right]}{\mathbb E_{\mathbb P}\left[ \prod_{j=1}^i \ell_y\left(\tilde {\mathcal K}^{M,\gamma_1}(t_j), \tilde X^{M,\gamma_1}(t_j) \right) \right]} \label{eq. Kallianpur-Striebel formula auxiliary model}.
\end{align}
Then for any $i\in\mathbb N_{>0}$ and $y\in\mathbb R^{m\times i}$, we define filters for stopped processes as follows.
\begin{align}
\hat f^c_{\phi, t_i}(y) &\triangleq \frac{\mathbb E_{\mathbb P}\left[ \prod_{j=1}^i \ell_y\left(\mathcal K, X^{\gamma_1}(t_j\wedge \tau_c)\right) \phi \left(\mathcal K, X^{\gamma_1}(t_i\wedge \tau_c)\right) \right]}{\mathbb E_{\mathbb P}\left[ \prod_{j=1}^i \ell_y\left(\mathcal K, X^{\gamma_1}(t_j\wedge \tau_c)\right) \right]}, \label{eq. Kallianpur-Striebel formula hybrid model stopping time} \\
\hat f^{c,M}_{\phi, t_i}(y) &\triangleq \frac{\mathbb E_{\mathbb P}\left[ \prod_{j=1}^i \ell_y\left(\tilde {\mathcal K}^{M,\gamma_1}(t_j), \tilde X^{M,\gamma_1}(t_j\wedge \tau^{M}_c) \right) \phi \left(\tilde {\mathcal K}^{M,\gamma_1}(t_i), \tilde X^{M, \gamma_1}(t_i\wedge \tau_c^M)\right) \right]}{\mathbb E_{\mathbb P}\left[ \prod_{j=1}^i \ell_y\left(\tilde {\mathcal K}^{M,\gamma_1}(t_j), \tilde X^{M,\gamma_1}(t_j\wedge \tau^M_c) \right) \right]} \label{eq. Kallianpur-Striebel formula auxiliary model stopping time}.
\end{align}
\textcolor{red}{It is not easy to show the sensitivity analysis of a stopped process.}
\begin{lemma}{\label{lemma convergence of filters for stopped processes}}
If \cref{assumption convergence of the initial condition}, \cref{cond. conditions on the artificial noise 1} \cref{cond. Lipschitz continuity of propensities} hold, then for any $c>0$, $i\in\mathbb N_{>0}$, and $y\in\mathbb R^{m\times i}$, there holds $\lim_{M\to\infty}\hat f^{c,M}_{\phi, t_i}(y) \to \hat f^c_{\phi, t_i}(y) $ almost surely.
\end{lemma}
\begin{proof}
We first show that the transition kernel of the stopped hybrid process, i.e., $$\mathcal K^{c}_{t}(z,A)\triangleq \mathbb P \left( \left(\mathcal K, X^{\gamma_1}(t\wedge \tau_c) \right) \in A \left| \left(\mathcal K, X^{\gamma_1}(0) \right) =z \right.\right) \qquad \forall z \in \Theta \times \mathbb R^{n}_{\geq 0} \text{ and } \forall A\subset \mathbb R, $$
maps the set of globally Lipschitz function to its self. To do the sensitivity analysis, we follow the literature \cite{anderson2012efficient} to construct two highly coupled processes $(\hat{\mathcal K}_1, \hat X_1(t))$ and $(\hat{\mathcal K}_2, \hat X_2(t))$ as follows.
\begin{align}
&\hat X_1(t) \notag \\%\label{eq. K1 and X1} \\
&=\lim_{N\to \infty} X^{N}(0) + \sum_{j : \gamma_1+\tilde \rho_j>0} \int_0^t \lambda'_j( \hat{\mathcal K}_1, \hat {X}_1(s)) D^{\gamma_1+\tilde \rho_j}\left(v'_{\cdot j}-v_{\cdot j}\right)\text{d} s \notag \\
&\quad +\sum_{j : \gamma_1+\tilde \rho_j=0 } R^{(0)}_j\left( \int_0^t \lambda'_j (\hat{\mathcal K}_1, \hat X_1(s))\wedge \lambda_j (\hat{\mathcal K}_2, \hat X_2(s))\text{d} s \right) D^0\left(v'_{\cdot j}-v_{\cdot j}\right) \notag \\
&\quad + \sum_{j : \gamma_1+\tilde \rho_j=0 } R^{(1)}_j\left( \int_0^t \lambda'_j (\hat{\mathcal K}_1, \hat X_1(s))- \lambda'_j (\hat{\mathcal K}_1, \hat X_1(s))\wedge \lambda_j (\hat{\mathcal K}_2, \hat X_2(s)) \text{d} s \right) D^0\left(v'_{\cdot j}-v_{\cdot j}\right) \notag \\
&\hat X_2(t) \notag \\
&=\lim_{N\to \infty} X^{N}(0) + \sum_{j : \gamma_1+\tilde \rho_j>0} \int_0^t \lambda'_j( \hat{\mathcal K}_2, \hat {X}_2(s)) D^{\gamma_1+\tilde \rho_j}\left(v'_{\cdot j}-v_{\cdot j}\right)\text{d} s \notag \\
&\quad +\sum_{j : \gamma_1+\tilde \rho_j=0 } R^{(0)}_j\left( \int_0^t \lambda'_j (\hat{\mathcal K}_1, \hat X_1(s))\wedge \lambda_j (\hat{\mathcal K}_2, \hat X_2(s))\text{d} s \right) D^0\left(v'_{\cdot j}-v_{\cdot j}\right) \notag \\
&\quad + \sum_{j : \gamma_1+\tilde \rho_j=0 } R^{(2)}_j\left( \int_0^t \lambda'_j (\hat{\mathcal K}_2, \hat X_2(s))- \lambda'_j (\hat{\mathcal K}_1, \hat X_1(s))\wedge \lambda_j (\hat{\mathcal K}_2, \hat X_2(s)) \text{d} s \right) D^0\left(v'_{\cdot j}-v_{\cdot j}\right) \notag
\end{align}
where $\left(\hat{\mathcal K}_1, \hat X_1(0)\right)$ and $\left(\hat{\mathcal K}_2,\hat X_2(0)\right)$ have the same law as $(\mathcal K, X(0))$, and $\{R^{(0)}_{j}(\cdot)\}_{j}$, $\{R^{(1)}_{j}(\cdot)\}_{j}$, and $\{R^{(2)}_{j}(\cdot)\}_{j}$ are mutually independent, unit rate Poisson processes.
Obviously, the processes $\left(\hat{\mathcal K}_1, \hat X_1(\cdot)\right)$, $\left(\hat{\mathcal K}_2,\hat X_2(\cdot)\right)$ and $(\mathcal K, X(\cdot))$ have the same law under the natural probability measure.
According to the above equations we have that
\begin{align*}
&\left\| \hat X_1(t)-\hat X_2(t) \right\|_2\\
&\leq \left\| \hat X_1(0)-\hat X_2(0) \right\|_2+ \sum_{j=1}^{r} \left\|\int_0^t \left(\lambda_j (\hat{\mathcal K}_1, \hat X_1(s))- \lambda_j (\hat{\mathcal K}_2, \hat X_2(s))\right)\zeta_j \mathbbold{1}_{\{\delta_j=0\}}\text{d} s\right\|_2\\
& \quad+ \sum_{j=1}^{r} R^{(1)}_j\left( \int_0^t \lambda_j (\hat{\mathcal K}_1, \hat X_1(s))- \lambda_j (\hat{\mathcal K}_1, \hat X_1(s))\wedge \lambda_j (\hat{\mathcal K}_2, \hat X_2(s)) \text{d} s \right) \left\|\zeta_j \mathbbold{1}_{\{\delta_j=1\}} \right\|_2 \\
&\quad+ \sum_{j=1}^{r} R^{(2)}_j\left( \int_0^t \lambda_j (\hat{\mathcal K}_2, \hat X_2(s))- \lambda_j (\hat{\mathcal K}_1, \hat X_1(s))\wedge \lambda_j (\hat{\mathcal K}_2, \hat X_2(s)) \text{d} s \right) \left\| \zeta_j \mathbbold{1}_{\{\delta_j=1\}} \right\|_2
\end{align*}
~
~
\end{proof}
~
~
~
~
\newpage
One method to do the sensitivity analysis is to construct auxiliary random processes that have the same laws as the nominal and perturbed process and, meanwhile, are highly coupled \cite{anderson2012efficient} (also see \cite{gupta2013unbiased}).
Following the method in \cite{anderson2012efficient}, we term $(\hat{\mathcal K}_1, \hat X_1(t))$ and $(\hat{\mathcal K}_2, \hat X_2(t))$ as two stochastic processes satisfying
where $\left(\hat{\mathcal K}_1, \hat X_1(0)\right)$ and $\left(\hat{\mathcal K}_2,\hat X_2(0)\right)$ have the same law as $(\mathcal K, X(0))$, and $\{R^{(0)}_{j}(\cdot)\}_{j}$, $\{R^{(1)}_{j}(\cdot)\}_{j}$, and $\{R^{(2)}_{j}(\cdot)\}_{j}$ are mutually independent, unit rate Poisson processes.
Also, we set the above Poisson processes and system parameters to be independent of $(\mathcal K, X(\cdot))$ (see \eqref{eq. the hybrid crn}) and the observation noises, so that the above processes are independent of $Y_c(\cdot)$ and $Y_d(\cdot)$.
Obviously, the processes $\left(\hat{\mathcal K}_1, \hat X_1(\cdot)\right)$, $\left(\hat{\mathcal K}_2,\hat X_2(\cdot)\right)$ and $(\mathcal K, X(\cdot))$ have the same law under the natural probability measure.
The advantage of the above constructions is that the constructed processes are highly correlated and satisfy the following results.
\begin{proposition}{\label{prop. sensitivity}}
Let $\left(\hat{\mathcal K}_1, \hat X_1(t)\right)$ and $\left(\hat{\mathcal K}_2, \hat X_2(t)\right)$ satisfy \eqref{eq. K1 and X1} and \eqref{eq. K2 and X2}, respectively.
If \eqref{eq. assumption infinite explosion time} and \cref{cond. global Lipschitz continuity of propensities} hold, then for any $t>0$, $\kappa_1,\kappa_2\in\Theta$, and $x_1, x_2 \in\mathbb{X}$, there exist time dependent scalars $C_{1,t}$ and $C_{2,t}$, such that
\begin{align*}
&\mathbb E_{\mathbb P} \left[
\left\| \hat X_1(t)- \hat X_2(t)\right\|_{2}
\Big|
\hat{\mathcal K}_1=\kappa_1, ~\hat{\mathcal K}_2=\kappa_2,~ \hat X_1(0)=x_1,~ \hat X_2(0)=x_2
\right] \\
&\leq C_{1,t} \left(\left\|\kappa_1-\kappa_2\right\|_2+ \left\|x_1-x_2\right\|_2 \right).
\end{align*}
\end{proposition}
\begin{proof}
By subtracting \eqref{eq. K2 and X2} from \eqref{eq. K1 and X1} and taking the Euclidean norm on both sides, we have that
\begin{align*}
&\left\| \hat X_1(t)-\hat X_2(t) \right\|_2\\
&\leq \left\| \hat X_1(0)-\hat X_2(0) \right\|_2+ \sum_{j=1}^{r} \left\|\int_0^t \left(\lambda_j (\hat{\mathcal K}_1, \hat X_1(s))- \lambda_j (\hat{\mathcal K}_2, \hat X_2(s))\right)\zeta_j \mathbbold{1}_{\{\delta_j=0\}}\text{d} s\right\|_2\\
& \quad+ \sum_{j=1}^{r} R^{(1)}_j\left( \int_0^t \lambda_j (\hat{\mathcal K}_1, \hat X_1(s))- \lambda_j (\hat{\mathcal K}_1, \hat X_1(s))\wedge \lambda_j (\hat{\mathcal K}_2, \hat X_2(s)) \text{d} s \right) \left\|\zeta_j \mathbbold{1}_{\{\delta_j=1\}} \right\|_2 \\
&\quad+ \sum_{j=1}^{r} R^{(2)}_j\left( \int_0^t \lambda_j (\hat{\mathcal K}_2, \hat X_2(s))- \lambda_j (\hat{\mathcal K}_1, \hat X_1(s))\wedge \lambda_j (\hat{\mathcal K}_2, \hat X_2(s)) \text{d} s \right) \left\| \zeta_j \mathbbold{1}_{\{\delta_j=1\}} \right\|_2
\end{align*}
Let us denote an event $A\triangleq\left\{ \hat{\mathcal K}_1=\kappa_1, ~\hat{\mathcal K}_2=\kappa_2,~ \hat X_1(0)=x_1,~ \hat X_2(0)=x_2\right\}$, and a conditional expectation $\epsilon_1(t)\triangleq\mathbb E_{\mathbb P} \left[
\left\| \hat X_1(t)- \hat X_2(t)\right\|_{2}
\Big|
A
\right]$.
Then, by taking the conditional expectation on both sides of the above inequality and considering \cref{cond. global Lipschitz continuity of propensities}, we have that
\begin{equation*}
\epsilon_1(t) \leq \left\|x_1-x_2\right\|_2+ r C_{\lambda}\left(\max_{j\in\{1,\dots,r\}} \left\| \zeta_j\right\|_2\right)
\left( t \|\kappa_1-\kappa_2\|_2 +\int_0^{t} \epsilon_1(s) \text{d} s \right)
\end{equation*}
for any $ t\leq \sup\{s|\epsilon_1(s)<\infty \}$, which, together with Gronwall's inequality, proves the first inequality of this proposition.
\end{proof}
Notably, the mean squared error of $\hat X_1(t)$ and $\hat X_2(t)$ is of the order $O\left(\|k_1-k_2\|\right)$ rather than $O\left(\|k_1-k_2\|^2\right)$.
It is because we consider Poisson noise in this paper, and the mean and variance of a Poisson random variable are the same.
\section{The proof of \cref{pro. verify the second contion of del Moral's lemma}}
\begin{proof}[The proof of \cref{pro. verify the second contion of del Moral's lemma}]
Since $(\mathcal K, X(\cdot))$, $\left(\hat{\mathcal K}_1, \hat X_1(\cdot)\right)$, and $\left(\hat{\mathcal K}_2, \hat X_2(\cdot)\right)$ have the same law, we can write
\begin{align*}
&\Big| \mathbb E_{\mathbb P}\left[
f\left(\mathcal K, X(t_i)\right) \left| \mathcal K=\kappa_1, X(0)=x_1\right.
\right]
-
\mathbb E_{\mathbb P}\left[
f\left(\mathcal K, X(t_i)\right) \left| \mathcal K=\kappa_2, X(0)=x_2\right.
\right]
\Big| \\
&=\Big|\mathbb E_{\mathbb P}\left[
f\left( \hat {\mathcal K}_1, \hat X_1(t_i)\right) \left| \hat{\mathcal K}_1 =\kappa_1, \hat X_1(0)=x_1\right.
\right] \\
&\qquad
-
\mathbb E_{\mathbb P}\left[
f\left(\hat {\mathcal K}_2, \hat X_2(t_i)\right) \left| \hat{\mathcal K}_2=\kappa_2, \hat X_2 (0)=x_2\right.
\right]
\Big| \\
&=
\left| \mathbb E_{\mathbb P}\left[
f\left( \hat {\mathcal K}_1, \hat X_1(t_i)\right)-f\left(\hat {\mathcal K}_2, \hat X_2(t_i)\right) \left| \hat{\mathcal K}_1 =\kappa_1, \hat{\mathcal K}_2=\kappa_2, \hat X_1(0)=x_1, \hat X_2 (0)=x_2\right.
\right] \right| \\
&\leq
\mathbb E_{\mathbb P}\left[
\left| f\left( \hat {\mathcal K}_1, \hat X_1(t_i)\right)-f\left(\hat {\mathcal K}_2, \hat X_2(t_i)\right) \right| \left| \hat{\mathcal K}_1 =\kappa_1, \hat{\mathcal K}_2=\kappa_2, \hat X_1(0)=x_1, \hat X_2 (0)=x_2\right.
\right]
\end{align*}
where the forth line follows from the fact that the laws of $\left(\mathcal K_1, \hat X_1(\cdot)\right)$ and $\left(\mathcal K_2, \hat X_2(\cdot)\right)$ are irrelevant to each other's initial conditions, and the last line follows from Jensen's inequality.
Therefore, this proposition follows immediately from \cref{pro. the same law of X and hat X} and the Lipschitz continuity of the function $f$.
\end{proof}
\section{Proof for the continuous-time cases}
We define analogs to $Z_c(t)$ using $\hat{X}_1(\cdot)$ and $\hat X_2(\cdot)$.
For $\ell=1,2$, we denote
\begin{equation*}
\hat Z_{\ell, c} (t)\triangleq \exp\left( \int_{0}^{t} h^{\top}\left(\hat X_\ell(s)\right) \text{d} Y_c(s) - \frac{1}{2} \int_0^{t} \left\|h\left(\hat X_\ell (s)\right)\right\|^2_2 \text{d} s \right),
\qquad \forall t>0,
\end{equation*}
\begin{proposition}{\label{pro. the same law of X and hat X}}
Conditioned on $\mathcal Y_{c,t}$, random variables $\left(\hat X_{\ell}(\cdot), \hat Z_{\ell,c}(\cdot)\right)$ ($\ell=1,2$) under $\mathbb P$ have the same conditional law as $\left( X(\cdot), Z_{c}(\cdot)\right)$ under $\mathbb Q_c$.
\end{proposition}
\begin{proof}
It follows straightforwardly from the definition.
\end{proof}
Therefore, we can also conclude that
\begin{align}{\label{eq bar phi hat K hat X}}
\bar \phi_{\tau,t}(\kappa,x, y)
= & \mathbb E_{\mathbb P}\left[ \frac{\hat Z_{c,\ell}(t)}{\hat Z_{c,\ell}(\tau)} \phi(\hat{\mathcal K}_{\ell}, \hat X_{\ell}(t))\left| \hat{\mathcal K}_{\ell} =\kappa, \hat X_{\ell}(\tau)=x, Y_{c,\tau:t}=y \right. \right]
\end{align}
for $\ell=1,2$.
\begin{proof}[The proof of \cref{pro continuous-time inequality}]
To simplify the notations, let us first denote an event
$A\triangleq\left\{ \hat{\mathcal K}_1=\kappa_1, ~\hat{\mathcal K}_2=\kappa_2,~ \hat X_1(0)=x_1,~ \hat X_2(0)=x_2\right\}$.
By \eqref{eq bar phi hat K hat X}, we can show that
\begin{align*}
&\text{The forth square root of the left hand side of \eqref{eq. continuous-time inequality}}\\
&= \sqrt[4]{\mathbb E_{\mathbb P} \left[\left( \mathbb E_{\mathbb P}\left[ \left. \frac{\hat Z_{1,c}(t)}{\hat Z_{1,c}(\tau)} \phi\left(\hat{\mathcal K}_1, \hat X_1(t)\right) \right| \mathcal Y_{c,t} \vee A \right]
- \mathbb E_{\mathbb P}\left[ \left. \frac{\hat Z_{2, c}(t)}{\hat Z_{2,c}(\tau)} \phi\left(\hat{\mathcal K}_2, \hat X_2 (t)\right) \right| \mathcal Y_{c,t} \vee A \right]
\right)^4 \right] } \notag\\
&\leq \sqrt[4]{\mathbb E_{\mathbb P} \left[
\left.
\left(
\frac{\hat Z_{1,c}(t)}{\hat Z_{1,c}(\tau)} \phi\left(\hat{\mathcal K}_1, \hat X_1(t)\right)
-
\frac{\hat Z_{2, c}(t)}{\hat Z_{2,c}(\tau)} \phi\left(\hat{\mathcal K}_2, \hat X_2 (t)\right)
\right)^4
\right| A
\right]} \\
&\leq \sqrt[4]{\mathbb E_{\mathbb P} \left[
\left.\left(
\left(\frac{\hat Z_{1,c}(t)}{\hat Z_{1,c}(\tau)}
-
\frac{\hat Z_{2, c}(t)}{\hat Z_{2,c}(\tau)}\right)
\phi\left(\hat{\mathcal K}_1, \hat X_1(t)\right)
\right)^4
\right| A
\right]}
\\
&\qquad +
\sqrt[4]{\mathbb E_{\mathbb P} \left[
\left.
\left(\frac{\hat Z_{2,c}(t)}{\hat Z_{2,c}(\tau)} \right)^4
\left(\phi\left(\hat{\mathcal K}_1, \hat X_1(t)\right) - \phi\left(\hat{\mathcal K}_2, \hat X_2(t)\right) \right)^4
\right| A
\right]} \\
&\leq 8\| \phi\|_{\infty}^4e^{10\|h \|_{\infty}^2(t-\tau)}\sqrt[4]{\mathbb E_{\mathbb P} \left[
\left.
e^{4\|h\|_{\infty}\int_\tau^t \left|h\left(\hat X_1(s)\right)-h\left(\hat X_2(s)\right)\right|\text{d} s}
-1
\right| A
\right]} \\
& \qquad + 8\| \phi\|_{\infty}^3e^{12\|h \|_{\infty}^2(t-\tau)}
\sqrt[4]{\mathbb E_{\mathbb P} \left[
\left.
\left|\phi\left(\hat{\mathcal K}_1, \hat X_1(t)\right) - \phi\left(\hat{\mathcal K}_2, \hat X_2(t)\right) \right|
\right| A
\right]}
\end{align*}
\end{proof}
where the third line follows from Jensen's inequality and the law of total expectation, the forth line follows from the triangle inequality, and the last inequality follows from the independence between $\left(\mathcal K, \hat X_\ell (\cdot)\right)$ (for $\ell=1$ or $2$) and the observation noise $B(\cdot)$.
Finally, by applying the boundedness of $h$, the Lipschitz continuity of $\phi$ and $h$, and \cref{prop. sensitivity} to the above formulas, we can prove the result.
|
\section{Introduction}
Open superstring field theory (SFT) can be a powerful method to obtain non-perturbative information about superstring theory. The best example is tachyon condensation where tachyonic solutions to the SFT equations of motion have been shown to describe the decay of non-supersymmetric D-branes \cite{Sen:1999xm}. In addition to the non-supersymmetric tachyonic solutions, the equations of motion of SFT also contains spacetime supersymmetric solutions which can be studied. In particular, it would be interesting to use SFT to search for stringy generalizations of instanton solutions of super-Yang-Mills which give rise to non-perturbative effects in gauge theories.
After compactifying to four dimensions, the massless states of the open superstring include $D=4$ super-Yang-Mills fields and the instanton solution is obtained by requiring the four-dimensional Yang-Mills field-strength to be self-dual. However, it is difficult to generalize the concept of self-dual field strength to SFT since the only gauge-invariant quantity that can be constructed from the string field $\Phi$ is $Q \Phi$, which vanishes on-shell. Nevertheless, one can instead define the four-dimensional instanton solution as a localized half-BPS solution to the equations of motion, i.e. a solution which is annihilated by half of the $N=1$ $D=4$ spacetime supersymmetries.
In this paper, we will show that there is a unique generalization of this half-BPS definition of the instanton solution in SFT and will find the first stringy correction.
Since $D=4$ spacetime supersymmetry plays an important role in our construction, we will use the manifestly $N=1$ $D=4$ super-Poincaré invariant action for SFT of \cite{SuperPoincareSFT}, which
is based on the $D=4$ hybrid formalism of \cite{Berkovits:1994wr}. The massless contribution to this SFT action reproduces the usual $N=1$ $D=4$ superspace action for super-Yang-Mills theory, and the massive states are also described in terms of $N=1$ $D=4$ superfields. Although we will be unable to find an exact solution to the SFT equations of motion which generalizes the self-dual instanton, we will find the first stringy correction to the instanton solution which corresponds to turning on certain massive spin-2 and spin-0 fields.
In section \ref{sec_sym_instantons} of this paper, we review self-dual instanton solutions in super-Yang-Mills theory. In section \ref{sec_sft_instantons} we review the $D=4$ hybrid formalism for the superstring and introduce the central problem of this paper - solving the half-BPS condition in open superstring field theory. Our strategy for tackling this problem is to write a series expansion for the string field and solve order by order in the expansion parameter, as explained in section \ref{series_expansion}. In section \ref{sec_bpst_instanton} we review the BPST instanton in super-Yang-Mills and generalize it to SFT using the star product. In section \ref{sec_first_correction} we calculate the first stringy correction to the BPST instanton and show it corresponds to turning on certain off-shell massive fields by comparing with the vertex operators for the first massive level of the string.
\section{Super-Yang-Mills Instantons}
\label{sec_sym_instantons}
$N=1$ $D=4$ super-Yang-Mills can be described by a vector superfield $V^i(x,\theta,\bar{\theta})$ where $i$ is a gauge index and $(x^m, \theta^\alpha, \bar\theta^{\dot\alpha})$ are the usual $N=1$ $D=4$ superspace variables for $m=0$ to 3 and $\alpha, \dot\alpha$ = 1 to 2. We also define $V = V^iT^i$ where $T^i$ are the generators of the gauge group. Supersymmetry transformations are generated by acting on $V$ with
\begin{equation}
q_\alpha =
\frac{\partial}{\partial \theta^\alpha}
- \frac{i}{2} (\sigma^m \bar{\theta})_\alpha \partial_m
\end{equation}
and
\begin{equation}
\bar{q}^{\dot{\alpha}} =
\frac{\partial}{\partial \bar{\theta}_{\dot{\alpha}}}
+ \frac{i}{2} (\theta\bar{\sigma}^m)^{\dot{\alpha}} \partial_m.
\end{equation}
where $\sigma^m$ are the Pauli matrices and $\bar{\sigma}^{0}=\sigma^{0}$, $\bar{\sigma}^{i}=-\sigma^{i}$ for $i=1,2,3$.
We want to find field configurations that are preserved by half of this supersymmetry. Naively, this condition would be formulated as
\begin{equation}
(\epsilon q) V = 0,
\end{equation}
where $\epsilon^{\alpha}$ is a supersymmetry parameter and $\epsilon q = \epsilon^{\alpha} q_{\alpha}$.
However, that is not quite correct. We must keep in mind that the theory has a gauge symmetry, given by
\begin{equation}
\delta_\text{gauge} V =
\mathscr{L}_{V/2} [
\Lambda
- \bar{\Lambda}
-i \coth(\mathscr{L}_{V/2})(\Lambda + \bar{\Lambda})
],
\label{gauge_transformation}
\end{equation}
where $\Lambda$ and $\bar{\Lambda}$ are chiral and antichiral, respectively, and $\mathscr{L}$ is the Lie derivative:
\begin{equation}
\mathscr{L}_V (A) = [V, A]
\end{equation}
So the correct condition is not that $(\epsilon q) V$ vanishes, but that it is pure gauge. In other words, $V$ is preserved by half of the supersymmetry if there are $\Lambda$ and $\bar{\Lambda}$ such that
\begin{equation}
(\epsilon q) V = -\delta_\text{gauge} V.
\label{variations_equality}
\end{equation}
Of course, one could also consider states preserved by the antichiral supersymmetry, for which the analysis is similar.
To see what eq. \ref{variations_equality} implies, we first write $V$ in Wess-Zumino gauge and assume that the fermions vanish (since we are concerned with classical field configurations):
\begin{equation}
V = - \theta \sigma^m \bar{\theta} v_m + \frac{1}{2} \theta^2 \bar{\theta}^2 D.
\end{equation}
Since $V^3 = 0$ in Wess-Zumino gauge, the gauge transformation of eq. \ref{gauge_transformation} becomes
\begin{equation}
\delta_\text{gauge} V =
\Lambda + \bar{\Lambda}
+ \frac{1}{2}[V, \Lambda - \bar{\Lambda}]
+ \frac{1}{12} [V, [V, \Lambda + \bar{\Lambda}]].
\end{equation}
The most general $\Lambda$ can be written in components as
\begin{equation}
\Lambda =
A(x)
+ \sqrt{2} \theta\psi(x)
+ \theta^2 F(x)
+ i \theta\sigma^m\bar{\theta} \partial_m A(x)
+ \frac{i}{\sqrt{2}} \theta^2 \bar{\theta} \bar{\sigma}^m \partial_m \psi(x)
+ \frac{1}{2} \theta^2 \bar{\theta}^2\square A(x).
\end{equation}
Similarly, the most general $\bar{\Lambda}$ is
\begin{equation}
\bar{\Lambda} =
A^*
+ \sqrt{2} \bar{\theta}\bar{\psi}
+ \theta^2 F^*
- i \theta\sigma^m\bar{\theta} \partial_m A^*
+ \frac{i}{\sqrt{2}} \bar{\theta}^2 \theta \sigma^m \partial_m \bar{\psi}
+ \frac{1}{2} \theta^2 \bar{\theta}^2\square A^*.
\end{equation}
The gauge transformation is then
\begin{equation}
\begin{split}
\delta_\text{gauge}V =
A
+ A^*
+ \sqrt{2} \theta \psi
+ \sqrt{2} \bar{\theta} \bar{\psi}
+ \theta^2 F
+ \bar{\theta}^2 F^* + \\
\theta \sigma^m \bar{\theta} \left(
i \partial_m \left( A - A^* \right)
- \frac{1}{2} \left[ v_m, A + A^* \right]
\right)
+ \frac{i}{\sqrt{2}} \theta^2 \left(
\bar{\theta} \bar{\sigma}^m \partial_m \psi
+ \frac{1}{2} \left[ v_m, \psi \sigma^m \bar{\theta} \right]
\right) + \\
\frac{i}{\sqrt{2}} \bar{\theta}^2 \left(
\theta \sigma^m \partial_m \bar{\psi}
+ \frac{1}{2} \left[ v_m, \bar{\psi} \bar{\sigma}^m \theta \right]
\right) + \\
\frac{1}{2} \theta^2 \bar{\theta}^2 \left(
\frac{1}{2} \square \left( A + A^* \right)
- \frac{i}{2} \left[ v_m, \partial^m \left( A + A^* \right) \right]
+ \right. \\
\left.
v^2 \left( A + A^* \right)
- 2 v^m \left( A + A^* \right) v_m
+ \left( A + A^* \right) v^2
\right).
\label{gauge_var_compl}
\end{split}
\end{equation}
On the other hand, the supersymmetry transformation is
\begin{equation}
(\epsilon q) V =
- \epsilon \sigma^m\bar{\theta}v_m
+ \epsilon \theta \bar{\theta}^2 D
+ \frac{i}{2} \epsilon \sigma^m \bar{\sigma}^n \theta \bar{\theta}^2 \partial_m v_n
\label{susy_var_complete}
\end{equation}
Plugging eqs. \ref{gauge_var_compl} and \ref{susy_var_complete} into eq. \ref{variations_equality} and splitting it in components, we find the conditions:
\begin{equation}
A + A^* = \psi^\alpha = F = F^* =\partial_m (A - A^*) = 0,
\label{t00}
\end{equation}
\begin{equation}
\bar{\psi}_{\dot{\alpha}} =
\frac{1}{\sqrt{2}} \left( \epsilon \sigma^m \right)_{\dot{\alpha}} v_m,
\label{psi_cond}
\end{equation}
\begin{equation}
i \left( \sigma^{mn} \right)^\alpha_\beta F_{mn} =
\delta^\alpha_\beta D
\end{equation}
where $\sigma^{m n} = \frac{1}{4}\left(\sigma^{m} \bar{\sigma}^n - \sigma^{n} \bar{\sigma}^m\right)$. Since $\sigma^{mn}$ is traceless, we conclude that
\begin{equation}
D = 0
\label{sym_aux}
\end{equation}
and
\begin{equation}
\left( \sigma^{mn} \right)^\alpha_\beta F_{mn} = 0,
\label{sym_self_dual}
\end{equation}
which tells us that the gauge field must be anti-self-dual.
\section{String Field Theory Instantons}
\label{sec_sft_instantons}
To describe the SFT instantons, we'll use the hybrid formalism developed in \cite{SuperPoincareSFT} which describes a compactified superstring with manifest four dimensional super-Poicaré symmetry. We take the four-dimensional space to be Euclidean. Also, we will only consider states that are independent of the structure of the compactified manifold.
The hybrid formalism contains five free bosons ($x^m,\rho$) and eight free fermions $(\theta^{\alpha}, \bar{\theta}^{\dot{\alpha}},$
$ p^{\alpha}, \bar{p}^{\dot{\alpha}})$ with OPE's
\begin{gather}
x^{m}(z_1) x^{n}(z_2) \sim -\log |z_1-z_2| \delta^{m n}, \quad \rho(z_1) \rho(z_2) \sim \log (z_1-z_2) \\
p_{\alpha}(z_1) \theta^{\beta}(z_2) \sim \delta_{\alpha}^{\beta}(z_1-z_2)^{-1}, \quad \bar{p}_{\dot{\alpha}}(z_1) \theta^{\dot{\beta}}(z_2) \sim \delta_{\dot{\alpha}}^{\dot{\beta}}(z_1-z_2)^{-1} \nonumber
\end{gather}
and a field theory for the six-dimensional compactification manifold which is described by the $\hat{c}=3 \mathrm{~N}=2$ superconformal generators $\left[T_{C}, G_{C}^{+}, G_{C}^{-}, J_{C}\right]$.
The $\mathrm{N}=4$ superconformal generators are defined in terms of these free fields by
\begin{gather}
T=\frac{1}{2} \partial x^{m} \partial x_{m}+p_{\alpha} \partial \theta^{\alpha}+\bar{p}_{\dot{\alpha}} \partial \bar{\theta}^{\dot{\alpha}}+\frac{1}{2} \partial \rho \partial \rho-\frac{i}{2} \partial^{2} \rho+T_{C} \\
G^{+}=e^{i \rho} d^2+G_{C}^{+}, \quad G^{-}=e^{-i \rho} \bar{d}^2+G_{C}^{-} \\
\tilde{G}^{+}=e^{-2 i \rho+i H_{C}} \bar{d}^2+e^{-i \rho+i H_{c}}\left(G_{C}^{-}\right) \\
\tilde{G}^{-}=e^{2 i \rho-i H_{C}} d^2+e^{i \rho-i H_{c}}\left(G_{C}^{+}\right) \\
J=i \partial \rho+i J_{C}, \quad J^{++}=e^{-i \rho+i H_{C}}, \quad J^{--}=e^{+i \rho-i H_{C}}
\end{gather}
where
\begin{gather}
d_{\alpha}=p_{\alpha}+\frac{i}{2} \bar{\theta}^{\dot{\alpha}} \partial x_{\alpha \dot{\alpha}}-\frac{1}{4}(\bar{\theta})^{2} \partial \theta_{\alpha}+\frac{1}{8} \theta_{\alpha} \partial(\bar{\theta})^{2} \\
\bar{d}_{\dot{\alpha}}=\bar{p}_{\dot{\alpha}}+\frac{i}{2} \theta^{\alpha} \partial x_{\alpha \dot{\alpha}}-\frac{1}{4}(\theta)^{2} \partial \bar{\theta}_{\dot{\alpha}}+\frac{1}{8} \bar{\theta}_{\dot{\alpha}} \partial(\theta)^{2}
\end{gather}
$J_{C}=\partial H_{C}$, $x_{\alpha \dot{\alpha}} = \sigma_{\alpha \dot{\alpha}}^{m} x_{m}$, and the notation $e^{-i \rho+i H_{c}}\left(G_{C}^{-}\right)$ means the contour integral of $e^{-i \rho+i H_{c}}$ around $G_{C}^{-}$ .
In principle, instantons in SFT can be found using the methods of the previous section and are defined to be field configurations whose superymmetry variation is pure gauge:
\begin{equation}
(\epsilon q) V = -\delta_\text{gauge} V.
\label{variations_equality2}
\end{equation}
However, unlike the super-Yang-Mills case, there is no gauge choice where $V^3 = 0$. So the gauge transformation is now an infinite series in V:
\begin{equation}
\delta_\text{gauge} V =
\Lambda + \bar{\Lambda}
+ \frac{1}{2}[V, \Lambda - \bar{\Lambda}]
+ \frac{1}{12} [V, [V, \Lambda + \bar{\Lambda}]]
+ \mathcal{O}(V^3),
\label{gaugepar}
\end{equation}
where the products are to be interpreted as star ($*$) products, as defined in \cite{NonCommutative}. The gauge parameters $\Lambda$ and $\bar\Lambda$ in eq. \ref{gaugepar} have to satisfy
\begin{equation}
\tilde{G}^+(\Lambda) = 0
\end{equation}
and
\begin{equation}
G^+(\bar{\Lambda}) = 0,
\end{equation}
where the notation $G^+(\bar{\Lambda})$ denotes the contour integral of $G^+$ around $\bar{\Lambda}$.
Although we will be unable to analytically solve eq. \ref{variations_equality2} because of the infinite number of terms, one can solve eq. \ref{variations_equality2} perturbatively by expanding around the linearized massless solutions to $V$, $\Lambda$ and $\bar\Lambda$. In the rest of this paper, we will use this perturbative method to find the first stringy corrections to the super-Yang-Mills instanton solution.
\section{Series Expansion of the Half-BPS Condition}
\label{series_expansion}
To perform a perturbative expansion, first write the string field $V$ as a power series in some parameter $\lambda$:
\begin{equation}\label{V expansion}
V =
\lambda V_0
+ \lambda^2 V_1
+ \lambda^3 V_2
+ \ldots.
\end{equation}
Similarly, write
\begin{equation}
\Lambda = \lambda \Lambda_0 + \lambda^2 \Lambda_1 + \lambda^3 \Lambda_2 + \ldots
\end{equation}
and
\begin{equation}
\bar{\Lambda} = \lambda \bar{\Lambda}_0 + \lambda^2 \bar{\Lambda}_1 + \lambda^3 \bar{\Lambda}_2 + \ldots.
\end{equation}
We can then break up eq. \ref{variations_equality2} in powers of $\lambda$:
\begin{equation}
(\epsilon q) V_0 =
\Lambda_0 + \bar{\Lambda}_0,
\label{zeroth_order_g}
\end{equation}
\begin{equation}
(\epsilon q) V_1 =
\Lambda_1 + \bar{\Lambda}_1
+ \frac{1}{2} [V_0, \Lambda_0 - \bar{\Lambda}_0],
\label{first_order_g}
\end{equation}
and so on.
The goal is to first determine $V_0$ and then recursively determine $V_{i+1}$ from $V_i$.
Finding $V_0$ is relatively easy because eq. \ref{zeroth_order_g} is linear and thus the massless and massive states decouple. Actually, since we are working with off-shell string theory, the expressions ``massless" and ``massive" need some clarification. By ``massless" we mean any state that depends only on the worldsheet zero modes of $x^m$, $\theta^\alpha$ and $\bar{\theta}_{\dot{\alpha}}$ and not on their derivatives nor on $p^\alpha$. That is, they are the states that would be massless if they were on-shell. All remaining states will be called ``massive".
Let us choose a gauge such that
\begin{equation}
V_0 =
- \theta \sigma^m \bar{\theta} v_{0m}
+ \frac{1}{2} \theta^2 \bar{\theta}^2 D_0
+ V_{0, \text{massive}},
\end{equation}
where $V_{0, \text{massive}}$ contains the massive components.
It is trivial to adapt the computations of section \ref{sec_sym_instantons} to show that
\begin{equation}
D_0 = 0
\end{equation}
and
\begin{equation}
\sigma^{mn} (\partial_m v_{0n} - \partial_n v_{0m}) = 0.
\end{equation}
Later we will describe an explicit solution of this condition for the case of a $U(2)$ gauge group.
What about the massive part? We will argue here that massive on-shell states cannot contribute to the instanton solution since they are not localized in spacetime. Note that in Euclidean space, the solution to the linearized equation of motion $(\delta^{mn}\partial_m \partial_n - M^2)\Phi =0$
is
\begin{equation}
\Phi = \int d^4 k f(k) e^{k\cdot x} \delta (|k| - M).
\label{massphi}
\end{equation}
Suppose $f(\hat k)$ is non-zero for some $\hat k^m$ satisifying $|\hat k| = M$. Then it is easy to see that $\Phi(x)$ diverges exponentially as $e^{c M^2}$ in the direction $x^m = c \hat k^m$ for $c$ large. So there are no localized solutions to $( \delta^{mn}\partial_m \partial_n- M^2)\Phi =0$ in Euclidean space which can contribute to the instanton solution.
So the linearized condition of eq. \ref{zeroth_order_g} together with the requirement that the solution is localized in spacetime implies that
\begin{equation}
V_0 = -\theta \sigma^m \bar{\theta} v_{0m},
\end{equation}
where $v_{0m}$ satisfies
\begin{equation}
\sigma^{mn} (\partial_m v_{0n}
- \partial_n v_{0m})
= 0.
\label{linearized_self_duality}
\end{equation}
A specific choice for the compensating gauge transformation is
\begin{equation}
\Lambda_0 = 0,
\end{equation}
and
\begin{equation}
\bar{\Lambda}_0 =
-(\epsilon \sigma^m \bar{\theta}) v_{0m}
\end{equation}
Equipped with this, we can now try to tackle eq. \ref{first_order_g}. We will find that due to the complexity of the star product, $V_0 * \bar{\Lambda}$ contains massive components even though $V_0$ and $\bar{\Lambda}$ contain only massless components.
So unlike $V_0$, $V_1$ will contain massive components.
\section{BPST Instanton}
\label{sec_bpst_instanton}
To gain intuition, it will be useful to consider an explicit super-Yang-Mills instanton solution and see how it can be extended to string field theory. For an $SU(2)$ gauge group, the $k=1$ instanton can be written in a simple form which is the BPST instanton \cite{BPST}:
\begin{equation}
v_m^i
=2\frac{\bar\eta^i_{mn} (x - x_0)^n}{(x - x_0)^2 + \rho^2}.
\label{bpst_instanton}
\end{equation}
$x_0^m$ and $\rho$ are two fixed parameters representing the center and the size of the instanton, and $\bar\eta^i_{mn}$ are the 't Hooft symbols \cite{tHooft}:
\begin{equation}
\bar\eta_{imn}
= \epsilon_{imn4}
- \delta_{im} \delta_{n4}
+ \delta_{in} \delta_{m4}
\end{equation}
where $i = 1, 2, 3$ is an $SU(2)$ index and $m, n = 1, 2, 3, 4$ are Euclidean spacetime indices.
Note that the t'Hooft symbols are anti-self-dual in their Lorentz indices:
\begin{equation}
\bar\eta_{imn} = -\frac{1}{2} \epsilon_{mnpq} \bar\eta_{ipq}.
\end{equation}
To show that eq. \ref{bpst_instanton} does indeed represent an instanton, we have to show that the corresponding field strength,
\begin{equation}\label{Field-Strength}
F^i_{mn}
= \partial_m v^i_n
- \partial_n v^i_m
+ \epsilon^{ijk} v^j_m v^k_n,
\end{equation}
is anti-self-dual. The derivative of $v_m^i$ is
\begin{equation}
\partial_m v_n^i
= 2 \frac{\bar\eta_{inm}[(x - x_0)^2 + \rho^2]
- 2 \bar\eta_{inp} (x - x_0)^p (x - x_0)_m}{[(x - x_0)^2 + \rho^2]^2},
\end{equation}
whereas the quadratic part of the field strength is
\begin{equation}
\epsilon_{ijk} v_m^j v_n^i
=4 \frac{\epsilon_{ijk} \bar\eta_{jmp} \bar\eta_{knq} (x - x_0)^p (x - x_0)^q}{[(x - x_0)^2 + \rho^2]^2}.
\end{equation}
Using the identity
\begin{equation}
\epsilon_{ijk} \bar\eta_{jmp} \bar\eta_{knq}
= \delta_{mn} \bar\eta_{ipq} + \delta_{pq} \bar\eta_{imn}
- \delta_{mq} \bar\eta_{ipn}
- \delta_{pn} \bar\eta_{imq},
\label{tHooft_identity}
\end{equation}
we obtain
\begin{equation}
\epsilon_{ijk} v_m^j v_n^i
= 4 \frac{\bar\eta_{imn} (x - x_0)^2
- \bar\eta_{ipn} (x - x_0)^p (x - x_0)_m
- \bar\eta_{imp} (x - x_0)^p (x - x_0)_n}{[(x - x_0)^2 + \rho^2]^2}.
\end{equation}
Then the field strength is
\begin{equation}
F_{mn}^i = - 4 \frac{\bar\eta_{imn} \rho^2}{[(x - x_0)^2 + \rho^2]^2}
\end{equation}
and the anti-self-duality of the field strength follows from the anti-self-duality of $\bar\eta_{imn}$.
To connect this with what we were doing in the previous section, let us expand $v_m^i$ in powers of $1/ \rho^2$:
\begin{equation}
v_m^i = \sum_{r = 0}^\infty \frac{v_{rm}^i}{\rho^{2(r+1)}},
\label{expand}
\end{equation}
where
\begin{equation}
v_{rm}^i = 2 (-1)^r \bar\eta^i_{mn} (x - x_0)^n (x - x_0)^{2r}.
\label{rho_expansion}
\end{equation}
That is, we are working with a large instanton expansion which has the form of \ref{V expansion} where the expansion parameter is $\lambda = {1\over {\rho^2}}$.
Now we would like to find the corrections to this instanton solution in superstring field theory - that is, a solution to eq. \ref{variations_equality} where we take the massless part of the string field to have the form of the BPST instanton. Note that the gauge group should now be $U(2)$ rather than $SU(2)$. Since the star product in non-commutative, the massive levels of the string field will have singlet contributions, even if the massless level doesn't. We formally define the object $v^i_m$ as
\begin{equation}
v_m^i = 2\frac{\bar\eta^i_{mn} (x - x_0)^n}{(x - x_0)^2 + \rho^2},
\label{bpst_instanton_sft}
\end{equation}
where the products are to be interpreted as $\otimes$ products, and $\otimes$ denotes the truncation of the star product to the massless level (see appendix \ref{appendix}). To give meaning to division by $x^m$, we invoke the expansion in $1/\rho^2$ of eq. \ref{expand} where
\begin{equation}
v_{r m}^{i}=2(-1)^{r} \bar\eta_{m n}^{i}\left(x-x_{0}\right)^{n}\otimes\left(\left(x-x_{0}\right)^2\right)^{r_\otimes}
\end{equation}
and $r_\otimes$ in the exponent means we multiply $r$ times with the $\otimes$ product. Also, $x^2$ means the normal ordered product - that is, no contractions between the $x$'s.
Note that $V_{massless}^i = -\theta\sigma^m\bar\theta v_m^i$ is a solution to the massless part of eq. \ref{variations_equality}. Since, as shown in the last section, the massive part of $V_0$ is trivial,
\begin{equation}
V_0^0 = 0, \quad V_0^i = -2\theta\sigma^m\bar\theta\bar\eta^i_{mn}x^n
\end{equation}
where $V_0^0 $ denotes the $U(1)$ component of the $U(2)$ gauge group.
In the next section we'll compute the first massive level of $V_1$.
\section{The First Stringy Correction}\label{sec_first_correction}
We now want to compute the first stringy correction to the BPST instanton, which is the first massive level of $V_1$. To that end, we must solve the half-BPS equation at the quadratic level (i.e. $O(1/\rho^4)$ terms) which corresponds to eq. \ref{first_order_g}. Although $V_0$, $\Lambda_0$ and $\bar\Lambda_0$ are massless, the star products in the commutator will contain massive terms which will imply that $V_1$ has a nontrivial massive part.
The half-BPS solution is not unique - if $V_1$ is a solution to eq. \ref{first_order_g}, then $V_1'=V_1+q^2\Omega$ is also a solution for any field $\Omega$ since $ (\epsilon q) q^2 \Omega =0$. We can ask if there is a half-BPS solution which also satisfies the SFT equations of motion. The answer is yes, and we will show that this solution is unique if we also require that it vanishes at infinity, i.e. it is localized in spacetime like the Yang-Mills instanton. In other words, $q^2 \Omega$ will be uniquely determined if we require that $V_1'$ satisfies the SFT equations of motion and is localized in spacetime. This first stringy correction to the BPST instanton will be interpreted as turning on some off-shell massive fields. To see what kinds of fields these are, we will compare our result with the first massive states of the superstring which on-shell consists of a massive spin-two multiplet and two massive scalar multiplets \cite{FirstMassiveStateAction,FirstMassiveState}. We will find that our solution corresponds to turning on a spin-two and a scalar field.
In subsection \ref{Half-BPS solutions} we find the general solution to eq. \ref{first_order_g} when $V_0$ is given by the linearized BPST solution. In subsection \ref{Equations of motion} we find the unique half-BPS solution that also satisfies the equations of motion and vanishes at infinity. And in subsection \ref{Massive vertex operator} we compare our solution with the vertex operator for the first massive level.
\subsection{Half-BPS solutions}\label{Half-BPS solutions}
In this section we will solve the equation
\begin{equation}\label{BPS2}
(\epsilon q) V_{1} = \Lambda_{1} + \bar{\Lambda}_{1} + \frac{1}{2}\left[V_{0},\Lambda_{0}-\bar{\Lambda}_{0}\right]
\end{equation}
given the linearized BPST solution for $V_0$
\begin{align}
&V_0^i = -\theta\sigma^m\bar\theta v^i_{0m}(x) = -2\theta\sigma^m\bar\theta\bar\eta^i_{mn}x^n \\
&\bar\Lambda^i_0 = -2\epsilon\sigma^m\bar\theta\bar\eta^i_{mn}x^n \\
&\Lambda^i_0 = 0
\end{align}
The massless part of $V_1$ is simply
\begin{equation}
V_{1,massless}^i = 2\theta\sigma^m\bar\theta\bar\eta^i_{mn}x^n\otimes x^2.
\end{equation}
As noted before, we can add any $q^2$-exact term to a solution $V_1$ and still have a solution. For the massive part $V_{1,massive}$, we'll show that the solution for the first massive level is unique up to a $q^2$-exact term, and up to gauge transformations. If, in addition to eq. \ref{BPS2}, we require that the solution satisfies the equations of motion and vanishes at infinity, we will find that the $q^2$-exact term is fixed uniquely. From now on, we'll use $V_1$ instead of $V_{1,massive}$ to denote the first massive level since we'll only deal with this level in this section. Similarly, $\Lambda_1$ and $\bar\Lambda_1$ will denote the first massive level of the gauge parameters.
First we have to calculate the commutator on the right-hand side of eq. \ref{BPS2}, which involves evaluating a star product. Then we substitute the general form of the string field and gauge parameters at the first massive level. This is given by the fields that have conformal weight 1 at zero momentum, and which have mass squared $M^2=2$ when on-shell.
The commutator will contain terms that are not $\epsilon q$-exact, which must therefore be cancelled by the gauge terms. This will fix $\Lambda_1$ and $\bar \Lambda_1$, up to $\epsilon q$-exact gauge parameters (that is, up to $\epsilon q(\Lambda_1'+\bar\Lambda_1')$ where $\Lambda_1'$ and $\bar\Lambda_1'$ are gauge parameters). Once the gauge parameters are fixed, we have determined $\epsilon q V_1$, which means we have determined $V_1$ up to a $q^2$-exact term.
At the first massive level, the commutator $[V_0, \bar \Lambda_0]$ satisfies
\begin{equation}
\frac{1}{2}\left[V_{0},\bar{\Lambda}_{0}\right] = \frac{1}{2}V_{0}^i*\bar{\Lambda}_{0}^j\{\sigma^i,\sigma^j\} = V_{0}^i*\bar{\Lambda}_{0}^i \sigma^0
\end{equation}
where $\sigma^0$ is $i$ times the $2\times 2$ identity. This is because $(V_{0}^i*\bar{\Lambda}_{0}^j) = -(\bar{\Lambda}_{0}^j*V_{0}^i)$ at this level. Details on the calculation of the star product can be found in appendix \ref{appendix}, and one finds at the first massive level that
\begin{equation}
V_{0}^i*\bar{\Lambda}_{0}^i = \partial\theta^{\alpha}V_{\alpha} + \partial\bar\theta^{\dot\alpha}\bar V_{\dot\alpha} + \Pi^m U_m
\end{equation}
where
\begin{equation}
\Pi^m = \partial x^m - \frac{i}{2} \theta\sigma^m\partial\bar{\theta} - \frac{i}{2} \bar{\theta}\bar\sigma^m\partial\theta
\end{equation}
and
\begin{align}
&V_{\alpha} = -\frac{2}{3\sqrt{3}}(\sigma^m\bar\theta)_{\alpha}\epsilon\sigma^n\bar\theta v_{0m}^i\otimes v_{0n}^i \\
&\bar V_{\dot\alpha} = \frac{2}{3\sqrt{3}}\left((\theta\sigma^m)_{\dot\alpha}\epsilon\sigma^n\bar\theta - \theta\sigma^m\bar\theta(\epsilon\sigma^n)_{\dot\alpha}\right)v_{0m}^i\otimes v_{0n}^i(y) \\
&U_m = \frac{4}{3\sqrt{3}}\theta\sigma^p\bar\theta \epsilon\sigma^q\bar\theta \bar\eta^i_{pm}v^i_{0q} - p\leftrightarrow q
\end{align}
where $y^m=x^m+\frac{i}{2}\theta\sigma^m\bar\theta$.
The general form of the string field at the first massive level is \cite{FirstMassiveStateAction}
\begin{align}
V_1 = &d^{\alpha} W_{\alpha}(x, \theta, \bar{\theta})+\bar{d}^{\dot{\alpha}} \bar{W}_{\dot{\alpha}}(x, \theta, \bar{\theta})+\Pi^{m} V_{m}(x, \theta, \bar{\theta})+\partial \theta^{\alpha} V_{\alpha}(x, \theta, \bar{\theta}) \\
&+\partial \bar{\theta}^{\dot{\alpha}} \bar{V}_{\dot{\alpha}}(x, \theta, \bar{\theta})+i\left(\partial \rho-\partial H_{C}\right) B(x, \theta, \bar{\theta})+\left(\partial H_{C}-3 \partial \rho\right) C(x, \theta, \bar{\theta}) \nonumber
\end{align}
and the gauge transformations are
\begin{align}\label{gauge parameters}
\Lambda_1 =&G^{+}\left(e^{-i \rho}\left(\bar{d}^{\dot{\alpha}} \bar{E}_{\dot{\alpha}}(x, \theta, \bar{\theta})+\partial \theta^{\alpha} B_{\alpha}(x, \theta, \bar{\theta})\right)\right) \\
=& - 2 i \Pi_{\alpha \dot{\alpha}} D^{\alpha} \bar{E}^{\dot{\alpha}} + \bar{d}^{\dot{\alpha}} D^{2} \bar{E}_{\dot{\alpha}} + d^{\alpha} \left(2B_{\alpha} - i \partial_{\alpha \dot{\alpha}} \bar{E}^{\dot{\alpha}}\right) + \partial \theta^{\alpha} D^{2} B_{\alpha} \nonumber\\
& +\partial \bar{\theta}^{\dot{\alpha}} \left(-i\partial_{\alpha \dot{\alpha}} B^{\alpha} + 4 \bar{E}_{\dot{\alpha}} - \frac{1}{2}\partial^2\bar E_{\dot\alpha}\right) + i \partial \rho\left(2 D^{\alpha} B_{\alpha}-i \partial_{\alpha \dot{\alpha}} D^{\alpha} \bar{E}_{\dot{\alpha}}\right) \nonumber\\
\bar{\Lambda}_1 =&\tilde{G}^{+}\left(e^{2 i \rho-i H_{C}}\left(d^{\alpha} E_{\alpha}(x, \theta, \bar{\theta})+\partial \bar{\theta}^{\bar{\alpha}} \bar{B}_{\dot{\alpha}}(x, \theta, \bar{\theta})\right)\right)\nonumber \\
=&-2 i \Pi_{\alpha \dot{\alpha}} \bar{D}^{\dot{\alpha}} E^{\alpha} + d^{\alpha} \bar{D}^{2} E_{\alpha} + \bar{d}^{\dot{\alpha}} \left(2\bar{B}_{\dot{\alpha}} - i \partial_{\alpha \dot{\alpha}} E^{\alpha}\right) + \partial \bar{\theta}^{\bar{\alpha}} \bar{D}^{2} \bar{B}_{\dot{\alpha}} \nonumber\\
& +\partial \theta^{\alpha} \left(-i\partial_{\alpha \dot{\alpha}} \bar{B}^{\dot{\alpha}} + 4E_{\alpha}-\frac{1}{2}\partial^2 E_{\alpha}\right) + i \partial\left(H_{C}-2 \rho\right)\left(2 \bar{D}^{\dot{\alpha}} \bar{B}_{\dot{\alpha}}-i \partial_{\alpha \dot{\alpha}} \bar{D}^{\dot{\alpha}} E_{\alpha}\right) \nonumber
\end{align}
where $D_{\alpha}=\frac{\partial}{\partial \theta^{\alpha}}+\frac{i}{2} \bar{\theta}^{\dot{\alpha}}\partial_{\alpha \dot{\alpha}}$ and $\bar{D}_{\dot{\alpha}}=\frac{\partial}{\partial \bar{\theta}^{\dot{\alpha}}}+\frac{i}{2} \theta^{\alpha}\partial_{\alpha \dot{\alpha}}$. We can use the gauge freedom to fix $V_1$ in the form
\begin{align}
V_1 = \Pi^{m} V_{m}(x, \theta, \bar{\theta})+i\left(\partial \rho-\partial H_{C}\right) B(x, \theta, \bar{\theta})+\left(\partial H_{C}-3 \partial \rho\right) C(x, \theta, \bar{\theta})
\end{align}
Note that there are no terms proportional to $d$ or $\bar d$ in either $V_1$ or in the commutator. Fixing the $d$ and $\bar d$ terms to zero in \ref{gauge parameters}, the gauge parameters take the form
\begin{align}
\Lambda_1 =& -\frac{1}{2} i\Pi_{\alpha \dot{\alpha}} D^{\alpha} \bar{E}^{\dot\alpha} + \frac{1}{4}\partial\theta^{\alpha} i \partial_{\alpha \dot{\alpha}} D^{2} \bar{E}^{\dot{\alpha}} + \partial\bar\theta^{\dot\alpha}\left(\bar E_{\dot\alpha}-\frac{1}{8} \bar D^{2} {D}^{2} \bar E_{\dot\alpha}\right) - \frac{i}{4}\partial\rho D^{\alpha} \bar{D}^{2} E_{\alpha} \\
\bar\Lambda_1 =& -\frac{1}{2} i\Pi_{\alpha \dot{\alpha}} \bar D^{\dot\alpha} {E}^{\alpha} + \frac{1}{4}\partial\bar\theta^{\dot\alpha} i \partial_{\alpha \dot{\alpha}} \bar D^{2} E^{\alpha} + \partial\theta^{\alpha}\left(E_{\alpha}-\frac{1}{8} D^{2} \bar{D}^{2} E_{\alpha}\right) \\
&- \frac{i}{4}\partial(H_C-2\rho) \bar D^{\dot\alpha} {D}^{2} \bar E_{\dot\alpha} \nonumber
\end{align}
where we have rescaled $E_{\alpha}$ and $\bar E_{\dot\alpha}$ by a factor of four for later convenience.
Equation \ref{BPS2} can be decomposed into five equations by separating the terms proportional to $\partial\theta^{\alpha}$, $\partial\bar\theta^{\dot\alpha}$, $\Pi^m$, $\partial\rho$ and $\partial H_C$. The first two will fix $E_{\alpha}$ and $\bar E_{\dot\alpha}$, up to $\epsilon q$-exact terms. The other equations, in turn, will determine $V_1$.
The terms proportional to $\partial\theta^{\alpha}$ and $\partial\bar\theta^{\dot\alpha}$ give
\begin{align}
&V_{\alpha} = E_{\alpha}-\frac{1}{8} D^{2} \bar{D}^{2} E_{\alpha} + \frac{i}{4} \partial_{\alpha \dot{\alpha}} D^{2} \bar{E}^{\dot{\alpha}}\\
&\bar V_{\dot\alpha} = \bar E_{\dot\alpha} - \frac{1}{8} \bar D^{2} D^{2} \bar E_{\dot\alpha} + \frac{i}{4}\partial_{\alpha \dot{\alpha}} \bar D^{2} E^{\alpha}
\end{align}
which are solved by
\begin{align}
&E_{\alpha} = V_{\alpha}+\frac{1}{8} D^{2} \bar{D}^{2} V_{\alpha} - \frac{i}{4} \partial_{\alpha \dot{\alpha}} D^{2} \bar{V}^{\dot{\alpha}} + E'_{\alpha}\\
&\bar E_{\dot\alpha} = \bar V_{\dot\alpha} + \frac{1}{8} \bar D^{2} D^{2} \bar V_{\dot\alpha} - \frac{i}{4} \partial_{\alpha \dot{\alpha}} \bar D^{2} V^{\alpha} + \bar E'_{\dot\alpha}
\end{align}
where $E'_{\alpha}$ and $\bar{E'}_{\dot{\alpha}}$ are solutions to the homogeneous equations, i.e. they satisfy
\begin{align}
&E'_{\alpha} = -\frac{i}{2} \partial_{\alpha \dot{\alpha}} D^{2} \bar{E'}^{\dot{\alpha}},\quad \bar E'_{\dot\alpha} = -\frac{i}{2} \partial_{\alpha \dot{\alpha}} \bar D^{2} {E'}^{\alpha} \label{condition 1}\\
&\left(\partial^{n} \partial_{n}+1\right) E'_{\alpha} = \left(\partial^{n} \partial_{n}+1\right) \bar E'_{\dot\alpha} = 0. \label{condition 2}
\end{align}
As we'll see shortly, $E'_{\alpha}$ and $\bar{E'}_{\dot{\alpha}}$ are also $\epsilon q$-exact.
The terms proportional to $\Pi^m$ give the equation
\begin{equation}\label{Pi equation}
\epsilon q V_{m} + U_m = -\frac{i}{2}\left(\bar\sigma_{m}\right)^{\alpha \dot{\alpha}} D_{\alpha} \bar{E}_{\dot{\alpha}} - \frac{i}{2}\left(\bar\sigma_{m}\right)^{\alpha \dot{\alpha}} \bar{D}_{\dot{\alpha}} E_{\alpha}
\end{equation}
Note that
\begin{align}
\bar\sigma^{\dot\alpha\alpha}_mD_{\alpha}\bar V_{\dot\alpha} =& \frac{2}{3\sqrt{3}}\left[\left(-2\epsilon\sigma^q\bar\theta \delta^p_m + \epsilon\sigma^q\bar\sigma_m\sigma^p\bar\theta\right)v_{0p}^i\otimes v_{0q}^i(y) - 2i\theta\sigma^p\bar\sigma_m\sigma^n\bar\theta\epsilon\sigma^q\bar\theta\bar\eta^i_{pn}v_{0q}^i \right.\\
&\left. + 2i\theta\sigma^p\bar\theta\epsilon\sigma^q\bar\sigma_m\sigma^n v_{0p}^i\bar\eta^i_{qn} \right] \nonumber\\
=& \frac{2}{3\sqrt{3}}\left[(-4\epsilon\sigma^q\bar\theta\delta^{mp}+\epsilon\sigma_m\bar\theta\delta^{pq}) v_{0p}^i\otimes v_{0q}^{i}(y) + 4i\theta\sigma^p\bar\theta\epsilon\sigma^q\bar\theta \left(\bar\eta^i_{pm}v_{0q}^i - v_{0p}^i\bar\eta^i_{qm}\right) \right] \nonumber\\
=& \frac{2}{3\sqrt{3}}(-4\epsilon\sigma^q\bar\theta\delta^{mp}+\epsilon\sigma_m\bar\theta\delta^{pq}) v_{0p}^i\otimes v_{0q}^{i}(y) + 2iU_m \nonumber
\end{align}
In the second equality we used the fact that $\sigma^{mn}\bar\eta^i_{mn} = 0$ and $\left(\bar\sigma^{(m}\sigma^{n)}\right)^{\dot\alpha}_{\dot\beta} = -2\delta^{mn}\delta_{\dot\alpha}^{\dot\beta}$. So eq. \ref{Pi equation} becomes
\begin{align}\label{Pi equation 2}
\epsilon q V_{m} =& \frac{i}{3\sqrt{3}}(4\epsilon\sigma^q\bar\theta\delta^{mp}-\epsilon\sigma_m\bar\theta\delta^{pq}) v_{0p}^i\otimes v_{0q}^{i}(y) \\
&-\frac{i}{2}\bar\sigma_{m}^{\alpha \dot{\alpha}} D_{\alpha}\left(\frac{1}{8} \bar D^{2} D^{2} \bar V_{\dot\alpha} - \frac{i}{4} \partial_{\beta \dot{\alpha}} \bar D^{2} V^{\beta} + \bar E'_{\dot\alpha}\right)
- \frac{i}{2}\left(\bar\sigma_{m}\right)^{\alpha \dot{\alpha}} \bar{D}_{\dot{\alpha}} E_{\alpha} \nonumber
\end{align}
Finally, the equations for $C$ and $B$ are
\begin{align}\label{CB equation}
\epsilon q(C+iB) = \frac{i}{4}D^{\alpha} \bar{D}^{2} E_{\alpha} \\
\epsilon q(C-iB) = -\frac{i}{4}\bar D^{\dot\alpha}D^{2} \bar E_{\dot\alpha} \label{CB equation 2}
\end{align}
In order for equations \ref{Pi equation 2}-\ref{CB equation 2} to be consistent, we must verify that the right-hand sides are $\epsilon q$-exact. Note that any field of the form $\epsilon\psi(y,\bar\theta)$ is $\epsilon q$-exact: $\epsilon\psi(y,\bar\theta) = \epsilon q(\theta\psi(y,\bar\theta))$. It's then easy to verify that $V_{\alpha}$ and $D^2\bar V_{\dot\alpha}$ are $\epsilon q$-exact, as are any derivatives of these quantities, since $q$ anti-commutes with $D$ and $\bar D$. Thus, the equations are consistent as long as we require that $D_{\alpha}\bar E'_{\dot\alpha}+\bar D_{\dot\alpha}E'_{\alpha}$, $D^{\alpha} \bar{D}^{2} E'_{\alpha}$ and $\bar D^{\dot\alpha}D^{2} \bar E'_{\dot\alpha}$ are $\epsilon q$-exact.
One might suspect that the $E'_{\alpha}$ and $\bar E'_{\dot\alpha}$ terms simply give a gauge transformation of $V_1$. This is indeed the case, as we'll now show. First, note that requiring $D^{\alpha}\bar D^{2} E'_{\alpha} = \epsilon q(F)$ implies, by the first eq. in \ref{condition 1}, that $\bar E'_{\dot\alpha} = -\frac{1}{2}\epsilon q(\bar D_{\dot\alpha}F) = \epsilon q(\bar E''_{\dot\alpha})$. Similarly, $\bar D^{\dot\alpha}D^{2} \bar E'_{\dot\alpha} = \epsilon q(\bar F)$ implies $E'_{\alpha} = \epsilon q(E''_{\alpha})$. Note that $E''_{\alpha}$ and $\bar E''_{\dot\alpha}$ are only defined up to $q^2(something)$. Also, they satisfy eqs. \ref{condition 1} and \ref{condition 2} up to $q^2(something)$ - in other words, we can choose $E''_{\alpha}$ and $\bar E''_{\dot\alpha}$ such that they satisfy those conditions. That means they only change $\Lambda_1$ and $\bar\Lambda_1$ by $\epsilon q$-exact gauge parameters, which in turn means that they only contribute pure gauge terms to $V_1$.
Thus, from eqs. \ref{Pi equation 2}-\ref{CB equation 2}, we have uniquely determined $\epsilon q(V_1)$ up to a gauge transformation. This means we also have determined $V_1$, up to a gauge transformation and up to $q^2(something)$. The particular solution found by simply eliminating $\epsilon q$ in eqs. \ref{Pi equation 2}-\ref{CB equation 2} (after writing the right-hand sides as $\epsilon q(something)$ in the way described in the previous paragraphs) is
\begin{align}
\tilde V_1^0 =& \frac{2i}{3\sqrt{3}}\Pi^m\left[(2\theta\sigma^q\bar\theta\delta^p_m- \theta\sigma_m\bar\theta\delta^{pq})v_{0p}^i\otimes v_{0q}^{i}(y) - 72\theta\sigma_m\bar\theta\right] \\
& - \frac{8}{\sqrt{3}}\partial(5\rho-3H_C)\theta\sigma^m\bar\theta y_m \nonumber \\
V_1^i =& 0 \nonumber
\end{align}
where the $x$ products, again, are normal ordered, i.e. no contractions.
\subsection{Equations of motion}\label{Equations of motion}
We now want to verify that we can find a half-BPS solution which satisfies the equation of motion, given by \cite{SuperPoincareSFT}
\begin{equation}\label{E.o.m.}
\tilde{G}^{+}\left(G^{+}\left(V_{1}\right)\right)=\frac{1}{2}\left\{G^{+}\left(V_{0}\right), \tilde{G}^{+}\left(V_{0}\right)\right\}
\end{equation}
The solution $\tilde V_1$ written above does not satisfy this equation, but we can find a field $V_1$ that does, and which differs from $\tilde V_1$ by a $q^2$-exact term. This solution $V_1$ is unique if we also require that it vanishes at infinity.
Calculating the anticommutator between $G^{+}\left(V_{0}\right)$ and $\tilde{G}^{+}\left(V_{0}\right)$ again involves evaluating a star product:
\begin{equation}
\frac{1}{2}\left\{G^{+}(V_{0}), \tilde{G}^{+}(V_{0})\right\} = \frac{1}{2}G^{+}(V_{0}^i)*\tilde{G}^{+}(V_{0}^j)\{\sigma^i,\sigma^j\} = G^{+}(V_{0}^i)*\tilde{G}^{+}(V_{0}^i)\sigma^0
\end{equation}
where
\begin{gather}
G^+ (V_0) = e^{i\rho}(2d^{\alpha} D_{\alpha} - i\partial\bar\theta^{\dot\alpha}\partial_{\alpha\dot\alpha} D^{\alpha})V_0, \\
\tilde G^+ (V_0) = e^{-2i\rho+iH_C}2\bar d^{\dot\alpha}\bar D_{\dot\alpha}V_0, \nonumber
\end{gather}
and we have used that $\partial_{\alpha\dot\alpha}\bar D^{\dot\alpha}V_0 = 0$.
As before, let's separate the terms proportional to $\Pi^m$, $d_{\alpha}$, $\bar d_{\dot\alpha}$, $\partial\rho$ and $\partial H_C$. Note that acting with $G^+$ or $\tilde G^+$ on the equation gives zero, so not all terms are independent. For example, we don't need to check the terms proportional to $\partial\theta^{\alpha}$ and $\partial\bar\theta^{\dot\alpha}$. Details on the star product can again be found in appendix \ref{appendix}, and on the right-hand side of \ref{E.o.m.} we find:\\
Terms proportional to $\Pi^m$:
\begin{align}
&-\frac{16i}{3\sqrt{3}}\bar\sigma_m^{\dot\alpha\alpha} D_{\alpha}V_0^i\otimes \bar D_{\dot\alpha}V_0^i - \frac{8}{3\sqrt{3}}\left(\partial_m\bar D^{\dot\alpha}D_{\alpha}V_0^i\otimes D^{\alpha}\bar D_{\dot\alpha}V_0^i \right.\\
&\left. - \bar D^{\dot\alpha}D_{\alpha}V_0^i\otimes \partial_m D^{\alpha}\bar D_{\dot\alpha}V_0^i +
i\partial_{\alpha\dot\alpha}D^{\alpha}V_0^i\otimes \partial_m\bar D^{\dot\alpha}V_0^i \right) \nonumber \\
&= \frac{16i}{3\sqrt{3}}\left(2 \theta \sigma^{q} \bar{\theta} \delta_{m}^{p}-\theta \sigma_{m} \bar{\theta} \delta^{p q}\right) v_{0 p}^{i} \otimes v_{0 q}^{i}(y) + \frac{192i}{\sqrt{3}}\theta\sigma_m\bar\theta \nonumber
\end{align}
Terms proportional to $d^{\alpha}$:
\begin{align}
-\frac{16}{3\sqrt{3}} \bar D_{\dot\alpha}D_{\alpha}V_0^i\otimes \bar D^{\dot\alpha}V_0^i = -\frac{16}{3\sqrt{3}}\theta_{\alpha}v_{0 p}^{i} \otimes v_{0}^{ip}(y)
\end{align}
Terms proportional to $\bar d^{\dot\alpha}$:
\begin{align}
-\frac{16}{3\sqrt{3}} D_{\alpha}\bar D_{\dot\alpha}V_0^i\otimes D^{\alpha}V_0^i = -\frac{16}{3\sqrt{3}}\left(-\bar\theta_{\dot\alpha}v_{0 p}^{i} \otimes v_{0}^{ip}(y) + 2i\bar\theta^2(\theta\sigma^m)_{\dot\alpha}\bar\eta^i_{mn}v^{in}(y) \right)
\end{align}
Terms proportional to $\partial(H_C-3\rho)$:
\begin{align}
&\frac{8i}{3\sqrt{3}}\left(\bar D^{\dot\alpha}D_{\alpha}V_0^i\otimes D^{\alpha}\bar D_{\dot\alpha}V_0^i + i\partial_{\alpha\dot\alpha}D^{\alpha}V_0^i\otimes \bar D^{\dot\alpha}V_0^i \right) \\
&= -\frac{16i}{3\sqrt{3}}v_{0 p}^{i} \otimes v_{0}^{ip}(y) - \frac{64}{\sqrt{3}}\theta\sigma^m\bar\theta y_m \nonumber
\end{align}
And on the left-hand side of \ref{E.o.m.} we find:\\
Terms proportional to $\Pi^m$:
\begin{align}
&-\left\{\bar{D}^{2}, D^{2}\right\} V_{m} + 8 V_{m} - 4 \partial^{n} \partial_{n} V_{m} - 8 \partial_{m} B - 12\left(\bar{\sigma}_{m}\right)^{\dot{\alpha} \alpha}\left[\bar{D}_{\dot{\alpha}}, D_{\alpha}\right] C \\
&= \frac{16i}{3\sqrt{3}}\left(2 \theta \sigma^{q} \bar{\theta} \delta_{m}^{p}-\theta \sigma_{m} \bar{\theta} \delta^{p q}\right) v_{0 p}^{i} \otimes v_{0 q}^{i}(y) + \frac{192i}{\sqrt{3}}\theta\sigma_m\bar\theta + \frac{128}{\sqrt{3}}y_m \nonumber
\end{align}
Terms proportional to $d^{\alpha}$:
\begin{align}
&4 i\left(\sigma^{m}\right)_{\alpha \dot{\alpha}} \bar{D}^{\dot{\alpha}} V_{m} - 24 \partial_{\alpha \dot{\alpha}} \bar{D}^{\dot{\alpha}} C + D_{\alpha} \bar{D}^{2}(18iC + 2B) \\
&= -\frac{16}{3\sqrt{3}}\theta_{\alpha}v_{0 p}^{i} \otimes v_{0}^{ip}(y) \nonumber
\end{align}
Terms proportional to $\bar d^{\dot\alpha}$:
\begin{align}
&4 i\left(\sigma^{m}\right)_{\alpha \dot{\alpha}} D^{\alpha} V_{m} + 24 \partial_{\alpha \dot{\alpha}} D^{\alpha} C - \bar D_{\dot\alpha} D^{2}(18iC - 2B) \\
&=-\frac{16}{3\sqrt{3}}\left(-\bar\theta_{\dot\alpha}v_{0 p}^{i} \otimes v_{0}^{ip}(y) + 2\bar\theta^2(\theta\sigma^m)_{\dot\alpha}\bar\eta_{mn}^iv^{in}(y)\right) - \frac{256}{\sqrt{3}}\bar\theta_{\dot\alpha} \nonumber
\end{align}
Terms proportional to $i\partial(\rho-H_C)$:
\begin{align}
&-4 \partial^{m} V_{m}-8 B - \frac{1}{2}\left\{\bar{D}^{2}, D^{2}\right\} B+3 \partial_{\alpha \dot{\alpha}}\left[\bar{D}^{\dot{\alpha}}, D^{\alpha}\right] C \\
&= -\frac{128}{\sqrt{3}} \nonumber
\end{align}
Terms proportional to $\partial(H_C-3\rho)$:
\begin{align}
&-2 \bar{\sigma}_{m}^{\dot{\alpha} \alpha}\left[\bar{D}_{\dot{\alpha}}, D_{\alpha}\right] V^{m} - 8 C + \frac{11}{2}\left\{\bar{D}^{2}, D^{2}\right\} C + 16 \partial_{m} \partial^{m} C + \partial_{\alpha \dot{\alpha}}\left[\bar{D}^{\dot{\alpha}}, D^{\alpha}\right] B \\
&= -\frac{32i}{3\sqrt{3}}v_{0 p}^{i} \otimes v_{0}^{ip}(y) - \frac{64}{\sqrt{3}}\theta\sigma^m\bar\theta y_m + \frac{128i}{\sqrt{3}} \nonumber
\end{align}
We conclude that $\tilde V_1$ does not satisfy the equations of motion. However, the following $V_1$ can be defined which differs from $\tilde V_1$ by a $q^2$-exact term and which satisfies the equations of motion:
\begin{align}\label{V_1 solution}
V_1^0 =& \tilde V_1^0 -\frac{64}{\sqrt{3}}\Pi^m y_m - \frac{8i}{\sqrt{3}}(\partial H_C - 3\partial\rho)y^2 + \frac{64}{\sqrt{3}}i(\partial\rho - \partial H_C) \\
=& \frac{2i}{3\sqrt{3}}\Pi^m\left[(2\theta\sigma^q\bar\theta\delta^p_m- \theta\sigma_m\bar\theta\delta^{pq})v_{0p}^i\otimes v_{0q}^{i}(y) - 72\theta\sigma_m\bar\theta + 96iy_m \right] \nonumber\\
& +\frac{8}{\sqrt{3}}(\theta\sigma^m\bar\theta y_m - iy^2)(\partial H_C - 3\partial\rho) + \frac{16}{\sqrt{3}}(i\theta\sigma^m\bar\theta y_m + 4)i(\partial\rho - \partial H_C) \nonumber \\
V_1^i =& 0 \nonumber
\end{align}
We can still add to $V_1$ a term $q^2 \Phi$ where $\Phi$ satisfies $\tilde G^{+}\left(G^{+}(\Phi)\right)=0$, i.e. $\Phi$ describes an on-shell massive field. But as explained below eqn. \ref{massphi}, any non-zero on-shell massive field will diverge exponentially when $x\to \infty$. So in order to have an instanton solution that vanishes at infinity, we should set $\Phi$ to zero. Note that the stringy contribution $V_1$ of eqn.
\ref {V_1 solution} naively diverges quadratically in $x^m$ when $x\to\infty$, but this is expected since it was found using an expansion in ${1\over{\rho^2}}$. After summing over all orders in ${1\over{\rho^2}}$, one expects that the stringy contribution will vanish at infinity like the super-Yang-Mills instanton. However, adding an on-shell massive contribution to $V_1$ would diverge exponentially when $x\to\infty$ and could never cancel after summing over all orders in ${1\over{\rho^2}}$. We therefore have a unique solution.
\subsection{Meaning of the massive fields}\label{Massive vertex operator}
The first massive level of the on-shell string field describes two massive scalar multiplets and a massive spin-two multiplet. The on-shell field can be gauge-fixed to the form
\begin{equation}\label{on-shell first massive}
\Pi^m\left(\hat{V}_{m} + 2 i \partial_{m}(F-\bar{F})\right) - \frac{1}{8}\left(\partial \rho-\partial H_{C}\right) \left[D^{2}, \bar{D}^{2}\right] (F+\bar{F}) + \left(\partial H_{C}-3 \partial \rho\right) (F+\bar{F})
\end{equation}
where $F$ and $\bar F$ are the chiral and antichiral superfields for the scalar multiplets, and $\hat{V}_{m}$ is the superfield for the spin-two multiplet. In components,
\begin{equation}
F=X(y)+\theta^{\alpha} \xi_{\alpha}(y)+(\theta)^{2} Y(y), \quad \bar F=\bar X(\bar y)+\bar\theta^{\dot\alpha} \bar\xi_{\dot\alpha}(\bar y)+(\bar\theta)^{2} \bar Y(\bar y)
\label{Fonshell}
\end{equation}
and
\begin{align}
\hat V_{m}=&C_{m}+i \theta \chi_{m}-i \bar{\theta} \bar{\chi}_{m}
+\theta \sigma^{n} \bar{\theta} \left(v_{m n} + \frac{1}{4} \epsilon_{m n p q} \partial^{p} C^{q} \right)
\label{Vonshell} \\
&+i \theta \theta \bar{\theta} \bar{\lambda}_{m}-i \bar{\theta} \bar{\theta} \theta \lambda_{m}
-\theta \theta \bar{\theta} \bar{\theta}\frac{1}{16} \partial^{n} \partial_{n} C_{m} \nonumber
\end{align}
where $X$, $\bar X$, $Y$ and $\bar Y$ are massive real bosons, $\xi_{\alpha}$ and $\bar\xi_{\dot\alpha}$ are massive spinors, $v_{mn}$ is a massive symmetric traceless tensor, $C_m$ is a massive vector, and $\chi$ and $\lambda$ are massive spin-3/2 fermions \cite{FirstMassiveStateAction,FirstMassiveState}.
This is to be compared to our result of eq. \ref{V_1 solution}. Since $V_1$ is not on-shell, it doesn't take the form \ref{on-shell first massive}. Nevertheless, we can still identify the components of $V_1$ with certain components of $F$, $\bar F$ and $\hat V_m$. Comparing eqn. \ref{Fonshell} and \ref{Vonshell} with eqn. \ref{V_1 solution}, one finds that the non-zero component fields are
\begin{equation}
X^0(y) = -\frac{8i}{\sqrt{3}}y^2 + \frac{16i}{\sqrt{3}}, \quad v^0_{mn} = v_{0(m}^i\otimes v_{0n)}^i - 2v_{0p}^i\otimes v_0^{ip}\delta_{mn}
\end{equation}
where $v_{0(m}^i\otimes v_{0n)}^i = v_{0m}^i\otimes v_{0n}^i+v_{0n}^i\otimes v_{0m}^i$. We can therefore interpret the correction to the BPST instanton as turning on the $U(1)$ component of a massive scalar field and a massive spin-two field.
\section*{Acknowledgements}
NB acknowledges partial financial support from CNPq grant number 311434/2020-7 and FAPESP grant numbers 2016/01343-7, 2019/24277-8 and 2019/21281-4, VJ acknowledges support from CAPES grant number 88882.330780/2019-01, and UP acknowledges support from FAPESP grant number
2020/15902-3.
|
\subsection{\leftskip 0pt\@startsection{subsection}{2}{\z@}%
{-3.25ex\@plus -1ex \@minus -.2ex}%
{1.5ex \@plus .2ex}%
{\normalfont\normalsize\bfseries}}
\renewcommand\subsubsection{\@startsection{subsubsection}{3}{\z@}%
{-3.25ex\@plus -1ex \@minus -.2ex}%
{1.5ex \@plus .2ex}%
{\normalfont\normalsize\bfseries\leftskip 3ex}}
\makeatother
\def\underline{\mathbb{F}_p}{\underline{\mathbb{F}_p}}
\def\underline{\mathbb{Z}_p}{\underline{\mathbb{Z}_p}}
\setlist[enumerate]{label*=\arabic*.}
\title{On the parametrized Tate construction}
\author{J.D. Quigley}
\address{
Dept. of Mathematics \\
Cornell University \\
Ithaca, NY, U.S.A.
}
\email{<EMAIL>}
\author{Jay Shah}
\address{Fachbereich Mathematik und Informatik, WWU Münster, 48149 M\"{u}nster, Germany}
\email{<EMAIL>}
\begin{document}
\tikzcdset{arrow style=tikz, diagrams={>=stealth}}
\begin{abstract}
We introduce and study a genuine equivariant refinement of the Tate construction associated to an extension $\widehat{G}$ of a finite group $G$ by a compact Lie group $K$, which we call the parametrized Tate construction $(-)^{t_G K}$. Our main theorem establishes the coincidence of three conceptually distinct approaches to its construction when $K$ is also finite: one via recollement theory for the $K$-free $\widehat{G}$-family, another via parametrized ambidexterity for $G$-local systems, and the last via parametrized assembly maps. We also show that $(-)^{t_G K}$ uniquely admits the structure of a lax $G$-symmetric monoidal functor, thereby refining a theorem of Nikolaus and Scholze. Along the way, we apply a theorem of the second author to reprove a result of Ayala--Mazel-Gee--Rozenblyum on reconstructing a genuine $G$-spectrum from its geometric fixed points; our method of proof further yields a formula for the geometric fixed points of an $\cF$-complete $G$-spectrum for any $G$-family $\cF$.
\end{abstract}
\date{\today}
\maketitle
\tableofcontents
\section{Introduction}
Let $G$ be a finite group and $M$ a $G$-module. The \emph{Tate cohomology} $\widehat{H}^*(G;M)$ of $G$ with coefficients in $M$ was introduced by Tate in \cite{Tat52}. It combines information about three important invariants: group cohomology $H^*(G;M)$, group homology $H_*(G;M)$, and the additive norm map $\Nm: M_G \to M^G$ from coinvariants to invariants. Swan \cite{Swa60} used Tate cohomology to define an equivariant cohomology theory for $G$-spaces.
Greenlees \cite{Gre87} initiated the study of a vast generalization of Tate cohomology using equivariant homotopy theory. Let $X$ be a spectrum with $G$-action. The \emph{Tate construction} $X^{tG}$ is defined by
$$X^{tG} \coloneq (\widetilde{EG} \wedge F(EG_+, X) )^G,$$
where $EG$ is a contractible space with free $G$-action and $\widetilde{EG}$ is the cofiber of the map $EG_+ \to S^0$ which collapses $EG$ to the non-basepoint. This construction generalizes classical Tate cohomology by taking $X = HM$ to be the Eilenberg--MacLane spectrum for a $G$-module $M$, since
$$\pi_*(HM^{tG}) \cong \widehat{H}^{-*}(G;M).$$
The Tate construction was generalized to compact Lie groups and studied extensively by Greenlees and May in \cite{GM95}, where they constructed the primary computational tool for analyzing Tate constructions, the Tate spectral sequence.
The Tate construction has many applications in algebraic topology. In chromatic homotopy theory, the Tate spectrum often decreases chromatic complexity \cite{AMS98, BR19, DM84, DJKMW86, GM95, GS96, HS96} and its vanishing controls descent \cite{MM15, Rog08, mathew2017xamples}. In trace methods, the Tate construction arises when computing topological cyclic homology, an approximation to algebraic K-theory \cite{HM97, HM03, NS18}. In equivariant stable homotopy theory, the Tate construction naturally arises in the construction of equivariant Adams spectral sequences \cite{Gre87, HK01} and analysis of the Hill--Hopkins--Ravenel slice spectral sequence \cite{HHR, MSZ20, Ull13}.
The input of the Tate construction is a spectrum with $G$-action and the output is an ordinary (nonequivariant) spectrum. The purpose of this paper is to study the \emph{parametrized Tate construction}, a generalization of the Tate construction better suited for studying genuine\footnote{In this paper, we usually drop the qualifier ``genuine'' and simply refer to these as $G$-spectra.} $G$-spectra. The input of the parametrized Tate construction will be a $G$-spectrum (for a finite group $G$) with a twisted action by a compact Lie group $K$ and the output will be a $G$-spectrum.\footnote{Note the switch of notation: $K$ has now taken the role that $G$ used to play.} To make the idea of twisted action precise, we will use the formalism of \emph{parametrized $\infty$-categories} as developed by Barwick--Dotto--Glasman--Nardin and the second author \cite{Exp1,Exp2,Exp2b,nardin,paramalg}. Let $\sO_G$ denote the orbit category of $G$.
\begin{dfn}
A \emph{$G$-$\infty$-category} resp. \emph{$G$-space} is a cocartesian resp. left fibration $C \to \sO_G^{\op}$. A \emph{$G$-functor} $F: C \to D$ is then a morphism over $\sO_G^{\op}$ that preserves cocartesian edges.
\end{dfn}
\begin{rem}
Under the straightening correspondence, a $G$-$\infty$-category resp. $G$-space $C$ is equivalently specified by a presheaf on $\sO_G$ valued in $\infty$-categories resp. spaces, while a $G$-functor corresponds to a natural transformation of presheaves.
\end{rem}
\begin{exm}
We have the $G$-$\infty$-categories $\underline{\Spc}^G$ and $\underline{\Sp}^G$ of $G$-spaces and $G$-spectra. Their respective fibers over an orbit $U \cong G/H$ are given by the $\infty$-categories $\Spc^H \simeq \Fun(\sO_H^{\op}, \Spc)$ and $\Sp^H$ of $H$-spaces and $H$-spectra, and the cocartesian edges encode the functoriality of restriction and conjugation.
\end{exm}
Now let
$$\psi = [1 \to K \to \widehat{G} \to G \to 1]$$
be a group extension, regard $B K$ as a space with $G$-action via the Kan fibration $B \widehat{G} \to B G$, and let $B^{\psi}_G K$ be the $G$-space given by (the unstraightening of) the right Kan extension of $B K$ along the inclusion $B G \subset \sO_G^{\op}$.
\begin{dfn}[{\cref{dfn:BorelGSpectraRelativeToNormalSubgroup}}]
A \emph{$G$-spectrum with $\psi$-twisted $K$-action} is a $G$-functor $X: B_G^{\psi} K \to \underline{\Sp}^G$.
\end{dfn}
Suppose for now that $K$ is also finite. The most expeditious definition of the parametrized Tate construction proceeds through the following theorem, a parametrized analogue of the canonical embedding of spectra with $G$-action into $G$-spectra as the Borel complete (i.e., cofree) $G$-spectra. Recall that given a $G$-family $\cF$ with universal $G$-space $E \cF$, a $G$-spectrum $X$ is \emph{$\cF$-torsion} if $E \cF_+ \wedge X \xto{\simeq} X$ and \emph{$\cF$-complete} if $X \xto{\simeq} F(E \cF_+, X)$, cf. \cref{SS:families}.
\begin{ntn}[{\cref{ntn:NFreeFamily}}]
Let $\Gamma_K$ be the $\widehat{G}$-family of \emph{$K$-free} subgroups, i.e., those $H \leq \widehat{G}$ such that $H \cap K = 1$.
\end{ntn}
\begin{thmx}[{\cref{thm:BorelSpectraAsCompleteObjects}}] \label{thmA}
There exists a symmetric monoidal restriction functor\footnote{We endow $\Fun_{G}(B_{G}^\psi K, \underline{\Sp}^{G})$ with the pointwise symmetric monoidal structure of \cref{dfn:S-PointwiseMonoidal}.}
$$j^*: \Sp^{\widehat{G}} \to \Fun_{G}(B_{G}^\psi K, \underline{\Sp}^{G})$$
that participates in an adjoint triple
\[ \begin{tikzcd}[column sep=6ex]
\Fun_{G}(B_{G}^\psi K, \underline{\Sp}^{G}) \ar[shift left = 3, hookrightarrow]{r}{j_!} \ar[shift right = 3, hookrightarrow]{r}[swap]{j_*} & \Sp^{\widehat{G}} \ar{l}[description]{j^*}
\end{tikzcd} \]
in which $j_!$ and $j_*$ are fully faithful and embed as the $\Gamma_K$-torsion and $\Gamma_K$-complete $\widehat{G}$-spectra, respectively.
\end{thmx}
\begin{dfn}\label{dfn:FirstDefn}
The \emph{parametrized Tate construction}
$$(-)^{t_K G} : \Fun_{G}(B_{G}^\psi K, \underline{\Sp}^{G}) \to \Sp^{G}$$
is the composite lax symmetric monoidal functor
$$\Fun_{G}(B_{G}^\psi K, \underline{\Sp}^{G}) \overset{j_*}{\hookrightarrow} \Sp^{\widehat{G}} \xrightarrow{- \wedge \widetilde{E \Gamma_K}} \Sp^{\widehat{G}} \xrightarrow{\Psi^K} \Sp^{G},$$
where
\begin{enumerate}
\item $j_*$ is the embedding of \cref{thmA},
\item $\widetilde{E\Gamma_K}$ is the cofiber of the map ${E\Gamma_K}_+ \to S^0$, and
\item $\Psi^K$ is the categorical $K$-fixed points.\footnote{We write this as $\Psi^K$ to distinguish it from the spectrum-valued functor of categorical $K$-fixed points $(-)^K: \Sp^{\widehat{G}} \to \Sp$.}
\end{enumerate}
\end{dfn}
\begin{exm} \begin{enumerate}[leftmargin=*]
\item If we take
$$\psi = [1 \to K \to K \to 1 \to 1],$$
then the parametrized Tate construction recovers the ordinary $K$-Tate construction.
\item Let $\mu_{p^n}$ be the group of $p^n$th roots of unity with $C_2$-action given by inversion and let
$$\psi = [1 \to \mu_{p^n} \to D_{2p^n} = \mu_{p^n} \rtimes C^2 \to C_2 \to 1].$$
The associated parametrized Tate construction $(-)^{t_{C_2} \mu_{p^n}}$ features extensively in the authors' work on real cyclotomic spectra \cite{QS21b}. We discuss this further in \cref{Rmk:KR}.
\item The case
$$\psi = [1 \to \mu_2^{\times 2^{n-1}} \to \mu_2^{\times 2^{n-1}} \rtimes C_{2^n} \to C_{2^n} \to 1],$$
where $C_{2^n}$ acts by cyclically permuting the factors of $\mu_2^{\times 2^{n-1}}$, is analyzed in forthcoming work of the first author with Chatham--Li--Lorman. The $n=1$ case was used in \cite{LLQ19} to obtain a Tate splitting for real Johnson--Wilson theories and in \cite{Qui19b} to understand the $C_2$-equivariant stable stems. These applications are outlined further in \cref{Rmk:ERn}.
\end{enumerate}
\end{exm}
\begin{rem}[Extension to compact Lie groups]
With suitable modifications, our proof of \cref{thmA} (and consequently, \cref{dfn:FirstDefn}) also applies in the more general situation where $K$ is a compact Lie group. The reason that we state and prove \cref{thmA} only when $K$ is finite is due to our choice of foundations for equivariant stable homotopy theory. In this paper, we adopt the foundations laid down by Bachmann and Hoyois \cite[\S 9]{BachmannHoyoisNorms}, who attach to every profinite groupoid $X$ a presentable, stable, and symmetric monoidal $\infty$-category $\SH(X)$ such that for $X = BG$, $\SH(BG)$ is equivalent to the underlying $\infty$-category $\Sp^G$ of the category of orthogonal $G$-spectra. To avoid mixing different approaches to the foundations of $G$-spectra, we then wish to avoid any mention of $\Sp^G$ for an infinite compact Lie group $G$. Instead, we will define the parametrized Tate construction when $K$ is compact Lie via the machinery of parametrized assembly maps (cf. \cref{thmC}).
\end{rem}
To justify our claim that the parametrized Tate construction is a suitable genuine equivariant refinement of the Tate construction, we then have the following suite of basic results.
\begin{obs}[Norm cofiber sequence] \label{obs:NormCofiberSequence}
For the constant $G$-diagram functor
$$\delta: \Sp^G \to \Fun_G(B^{\psi}_G K, \underline{\Sp}^G),$$
define the \emph{parametrized homotopy orbits} $(-)_{h_G K}$ to be its left adjoint and the \emph{parametrized homotopy fixed points} to be its right adjoint.\footnote{These are the $G$-colimit and $G$-limit functors, respectively.} The parametrized Tate construction is then designed to ``measure the difference'' between $(-)_{h_G K}$ and $(-)^{h_G K}$.
More precisely, we have that the adjoint triple $j_! \dashv j^* \dashv j_*$ of \cref{thmA} is one half of the recollement on $\Sp^{\widehat{G}}$ determined by the idempotent $E_{\infty}$-algebra $\widetilde{E \Gamma_K}$. Recollement theory then yields the cofiber sequence
\begin{equation*} j_!(-) \to j_*(-) \to j_*(-) \wedge \widetilde{E \Gamma_K}. \end{equation*}
Since $\Psi^K$ is right adjoint to the inflation functor $\inf_G^{\widehat{G}}$ and $j^* \inf_G^{\widehat{G}} \simeq \delta$, we get that $\Psi^K j_* \simeq (-)^{h_G K}$. We also have the Adams-type isomorphism $\Psi^K j_! \simeq (-)_{h_G K}$ (\cref{AdamsIsomorphism}). Applying $\Psi^K$ then yields the \emph{norm cofiber sequence}
\[ (-)_{h_G K} \to (-)^{h_G K} \to (-)^{t_G K}. \]
\end{obs}
\begin{obs}[Geometric model] \label{obs:GeometricModel}
Let $Y$ be a $\widehat{G}$-spectrum. By monoidal recollement theory, we have that $j_! j^* Y \simeq E {\Gamma_K}_+ \wedge Y$ and $j_* j^* Y \simeq F(E {\Gamma_K}_+, Y)$. If we then consider $X = j^* Y$, we obtain a \emph{geometric model} for the parametrized Tate construction
\[ X^{t_G K} \simeq \Psi^K (F(E {\Gamma_K}_+, Y) \wedge \widetilde{E\Gamma_K}).\footnote{As we have already seen, classically it is customary to conflate $X$ and $Y$ and simply write $X^{t K} \simeq (F(E K_+, X) \wedge \widetilde{E K})^K$.} \]
This identifies $(-)^{t_G K}$ as a special instance of the Greenlees--May generalized Tate construction for a $\widehat{G}$-family \cite[Part IV]{GM95} and hence yields a spectral sequence with $E_2$-term given by certain Amitsur--Dress--Tate cohomology groups (\cref{rem:TateSS}).
\end{obs}
For the next observation, we write $\underline{\Fun}_G$ for the internal hom in $G$-$\infty$-categories and recall that for a $G$-cocomplete $G$-$\infty$-category such as $\underline{\Sp}^G$ and a (small) $G$-$\infty$-category $I$, the $G$-colimit functor
$$\colim^G: \Fun_G(I, \underline{\Sp}^G) \to \Sp^G$$
always refines to a $G$-functor $\underline{\Fun}_G(I, \underline{\Sp}^G) \to \underline{\Sp}^G$ such that the fiber over $G/H$ is given by the $H$-colimit. If $I = B_G^{\psi} K$, we then write $(-)_{\underline{h}_G K}$ for the $G$-colimit $G$-functor. Likewise, $\underline{\Sp}^G$ is also $G$-complete, and we write $(-)^{\underline{h}_G K}$ for the $G$-limit $G$-functor.
\begin{obs}[{Compatibility with restriction, \cref{rem:ParamTateCompatibleRestriction}}] \label{introobs1}
Let $X$ be a $G$-spectrum with $\psi$-twisted $K$-action, $H \leq G$ a subgroup, $\widehat{H} = \pi^{-1}(H)$ for the quotient map $\pi: \widehat{G} \to G$, and $\psi_H = [1 \to K \to \widehat{H} \to H \to 1].$ Regard $X$ as a $H$-spectrum with $\psi_H$-twisted $K$-action via the restriction functor
\[ \Fun_G(B^{\psi}_G K, \underline{\Sp}^G) \to \Fun_H(B^{\psi_H}_H K, \underline{\Sp}^H) \]
given by pulling back along $\sO_H^{\op} \simeq (\sO_G^{\op})^{(G/H)/} \to \sO_G^{\op}$. Then $\res^G_H X^{t_G K} \simeq X^{t_H K}$. In particular, the underlying spectrum of $X^{t_G K}$ is $X^{t K}$. More generally, $(-)^{t_G K}$ refines to a $G$-functor
\[ (-)^{\underline{t}_G K}: \underline{\Fun}_G(B^{\psi}_G K, \underline{\Sp}^G) \to \underline{\Sp}^G, \]
and the norm cofiber sequence refines to a sequence of natural transformations of $G$-functors
\[ (-)_{\underline{h}_G K} \to (-)^{\underline{h}_G K} \to (-)^{\underline{t}_G K} \]
whose fiber over $G/H$ is given by the norm cofiber sequence
\[ (-)_{h_H K} \to (-)^{h_H K} \to (-)^{t_H K}. \]
\end{obs}
\begin{obs}[{Residual action, \cref{prp:ResidualAction}}] \label{introobs2}
Let $X$ as above. Suppose that $\widehat{G} \cong K \rtimes G$ for a $G$-action on $K$, and let $L \trianglelefteq K$ be a normal subgroup such that the inclusion is $G$-equivariant. Consider the group extensions
$$\psi' = [L \to L \rtimes G \to G], \quad \psi''=[K/L \to K/L \rtimes G \to G],$$
and view $X$ as a $G$-spectrum with $\psi'$-twisted $L$-action via restriction along $B^{\psi'}_G L \to B^{\psi}_G K$. Then $X^{t_G L}$ canonically acquires a ``residual action'' in the sense of lifting to a $G$-spectrum with $\psi''$-twisted $K/L$-action, and we have a fiber sequence of $G$-spectra
\[ (X_{h_G L})^{t_{G} K/L} \to X^{t_G K} \to (X^{t_G L})^{h_G K/L}.\footnote{Note that $X_{h_G L}$ acquires a residual action by a more basic result of parametrized higher category theory. Namely, consider the $G$-left Kan extension $\rho_! X$ of $X$ along $\rho: B^{\psi}_G K \to B^{\psi}_G K/L$. By definition, $\rho_! X$ is a $G$-spectrum with $\psi''$-twisted $K/L$-action, and since $\rho$ has $G$-fiber $B^{\psi}_G L$ (with the basepoint of $B^{\psi}_G K/L$ given by the semidirect product splitting), the underlying $G$-spectrum of $\rho_! X$ is indeed $X_{h_G L}$.} \]
\end{obs}
Before stating our deeper results, we pause to explain some applications of the parametrized Tate construction.
\subsection{Motivation and applications}\label{SS:Applications}
Our study of $G$-spectra with $\psi$-twisted $K$-action and the parametrized Tate construction $(-)^{t_K G}$ is motivated by a number of topics in equivariant stable homotopy theory. In particular, these concepts play a central role in our work on real algebraic K-theory, hyperreal oriented cohomology theories, and generalized Mahowald invariants. Our guiding philosophy is that the parametrized Tate construction should replace the ordinary Tate construction whenever a statement in classical stable homotopy theory might admit a genuine equivariant refinement. We discuss the implementation of this philosophy in the following remarks.
\begin{rem}[Real algebraic K-theory]\label{Rmk:KR}
Hesselholt--Madsen \cite{HM13}, Schlichting \cite{Sch17}, Spitzweck \cite{spitzweck2016}, Heine--Spitzweck--Verdugo \cite{heine2019eal}, and Calm{\`e}s et. al. \cite{CDH+20a, CDH+20c, CDH+20b} have all defined \emph{real algebraic K-theory}.\footnote{Hesselholt--Madsen work with exact categories with duality and weak equivalences, while Schlichting works with dg categories with weak equivalences and duality. Spitzweck constructed a Grothendieck--Witt space and connective real K-theory $C_2$-spectrum for stable $\infty$-categories with duality, and subsequently Heine--Spitzweck--Verdugo defined the real algebraic K-theory of Waldhausen $\infty$-categories with genuine duality. In a separate line of development, Calm{\`e}s et. al. define the real algebraic K-theory of Poincar\'{e} $\infty$-categories, which were introduced by Lurie \cite{Lur13} in his course on algebraic $L$-theory and surgery.} All approaches ultimately yield a genuine $C_2$-equivariant refinement of algebraic K-theory whose categorical $C_2$-fixed points are Grothendieck--Witt theory (i.e., hermitian K-theory) and whose geometric $C_2$-fixed points are some flavor of algebraic L-theory.
By the Dundas--Goodwillie--McCarthy Theorem \cite{DGM12}, the algebraic K-theory of connective ring spectra is closely approximated by topological cyclic homology, an invariant obtained from the \emph{cyclotomic structure} on topological Hochschild homology. Cyclotomic structures were originally defined using genuine $S^1$-equivariant homotopy theory \cite{BHM93, HM97, BM16}. An incredible insight by Nikolaus--Scholze in \cite{NS18} is that for bounded-below spectra, cyclotomic structures can be described completely using Borel $S^1$-equivariant homotopy theory. The Tate construction features prominently in their perspective: given a bounded-below spectrum $X$ with $S^1$-action, a cyclotomic structure on $X$ amounts to a collection of $S^1$-equivariant \emph{Tate-valued Frobenius maps} $\varphi_p : X \to X^{tC_p}$, one for each prime $p$.
A genuine $C_2$-equivariant refinement of cyclotomic spectra, called \emph{real cyclotomic spectra}, was introduced by H{\o}genhaven in \cite{Hog16}. This notion mirrors the original definition of cyclotomic spectra and requires the use of genuine $O(2)$-equivariant homotopy theory. In \cite{QS21b}, we refine the approach of Nikolaus--Scholze to define real cyclotomic spectra using $C_2$-spectra with twisted $S^1$-action and the parametrized Tate construction as the receptacle of the real cyclotomic Frobenius, and prove that this recovers the genuine formulation provided that the underlying spectrum of the $C_2$-spectrum is bounded-below. This will be used to study real algebraic K-theory in future work; we refer the reader to the introduction of \cite{QS21b} for further discussion.
\end{rem}
The following result is a relatively straightforward application of the results mentioned above. It extends \cite[Thm. 16.1]{GM95} to the parametrized setting.
\begin{thmx}[{Inverse limit formula for the parametrized Tate construction, \cref{Thm:GM161}}]\label{MT:GM161}
Let $\psi = [K \to \widehat{G} \to G]$ be an extension with $K \subseteq H$ for each $H \notin \Gamma_K$. Suppose $V$ is as in \cref{Lem:VRestrictions} for $\cF = \Gamma_K$, i.e., $V$ is a $G$-representation with $V^H=0$ for $H \notin \Gamma_K$ and $V^H \neq 0$ for $H \in \Gamma_K$. For any $X \in \Sp^{\Phi \Gamma_K} \simeq \Sp^{G}$, we have
$$(j^* i_* X)^{t[\psi]} \simeq \lim_n ({B_{G}^{\psi} K}^{-nV} \otimes \Sigma X).$$
\end{thmx}
This formula for the parametrized Tate construction has several applications in equivariant stable homotopy theory.
\begin{rem}[Tate blueshift]\label{Rmk:ERn}
\emph{Hyperreal oriented cohomology theories} are genuine $C_{2^n}$-spectra which generalize complex oriented cohomology theories, or cohomology theories with Thom isomorphisms for complex vector bundles. When $n=1$, a hyperreal oriented cohomology theory is a real oriented cohomology theory as introduced by Hu--Kriz in \cite{HK01}. Hyperreal oriented cohomology theories for $n>1$ were first applied in the Hill--Hopkins--Ravenel solution to the Kervaire invariant one problem in geometric topology \cite{HHR}. They have since been applied to study the stable homotopy groups of spheres \cite{HS20, HSWX18, LSWX19}, where their fixed points model the fixed points of Lubin--Tate spectra.
The ordinary Tate construction was shown to decrease the height of certain complex oriented cohomology theories by Ando--Morava--Sadofsky \cite{AMS98}. The key computational input to their work is the isomorphism of graded rings
$$\pi_*(E^{tC_p}) \cong E^*((x))/[p](x),$$
where $[p](x)$ is the $p$-series of the formal group law associated to $E$. This isomorphism arises from a geometric description of the Tate construction proven by Greenlees--May in \cite[\S 16]{GM95}. The fact that it is an isomorphism of graded rings relies on the multiplicativity of the Tate construction (cf. \cite[Pg. 7]{GM95}).
The parametrized Tate construction was used to produce analogous results for real oriented cohomology theories in \cite{LLQ19} and will be applied in forthcoming work of the first author with Chatham--Li--Lorman to study hyperreal oriented cohomology theories. The starting point for these computations is the observation that the parametrized Tate construction, instead of the ordinary Tate construction, can be used to access formal group law techniques for hyperreal oriented cohomology. \cref{MT:GM161} is used to construct spectrum-level splittings of the parametrized Tate construction.
\end{rem}
\begin{rem}[Generalized Mahowald invariants]\label{Rmk:MI}
\cref{MT:GM161} and the generalized Segal conjecture (cf. \cite{AHJM88b} and \cref{Thm:Segal}) can be used to define a $G$-equivariant Mahowald invariant
$$M(-): \pi_\star^{G}(\s) \rightsquigarrow \pi_\star^{G}(\s)$$
whenever there exists a $\widehat{G}$-representation $V$ as in \cref{Lem:VRestrictions}. The classical Mahowald invariant may be used to produce Hopf invariant one elements in $\pi_*(\s)$: one has $\eta = M(2)$, $\nu = M(\eta)$, and $\sigma = M(\nu)$. Similarly, it was shown in \cite{Qui19b} that the $C_2$-equivariant Mahowald invariant gives rise to the $C_2$-equivariant Hopf invariant one elements in $\pi_\star^{C_2}(\s)$. It could be interesting to produce analogous elements for other groups.
The Mahowald invariant can often be approximated by studying $E^{t\mu_p}$, where $E$ is a ring spectrum equipped with a trivial $\mu_p$-action. The multiplicativity of the ordinary Tate construction is often useful in these analyses.\footnote{For instance, it implies that the inverse limit Adams spectral sequence for $E^{t\mu_p}$ is multiplicative.} The multiplicativity of the parametrized Tate construction should have similar consequences for computing approximations to generalized Mahowald invariants.
\end{rem}
\subsection{Deeper results and main theorems}
\cref{dfn:FirstDefn} of the parametrized Tate construction may be viewed as a generalization of Greenlees' geometric model for the ordinary Tate construction. Another important abstract perspective on Tate constructions is afforded by Lurie's theory of \emph{ambidexterity} \cite[\S 6.1.6]{HA}. Suppose $C$ is any semiadditive $\infty$-category that admits limits and colimits indexed by finite groupoids, suppose $K$ is a finite group, and let $X$ be an object in $C$ with $K$-action. Then one has an additive \emph{norm map}
$$\Nm: X_{h K} \to X^{h K}$$
whose cofiber defines the Tate construction $X^{t K}$. The vanishing of Tate (e.g., in the $T(n)$ or $K(n)$-local category \cite{Kuhn2004,Clausen2017}) then leads to a rich theory of \emph{higher semiadditivity} \cite{HL13,CSY18,CSY20,CSY21}.
Using the formalism of Beck--Chevalley fibrations, Hopkins and Lurie have set up a very general framework for producing norm maps in \cite[\S 4.1-2]{HL13}; to apply this theory in the above example, they then prove that the $\infty$-category $\LocSys(C)$ of local systems on $C$ constitutes a Beck--Chevalley fibration over $\Spc$ \cite[\S 4.3]{HL13}. We initiate the theory of \emph{parametrized ambidexterity} in this paper by proving that for a suitably $G$-cocomplete $G$-$\infty$-category $C$, the $\infty$-category $\LocSys^G(C)$ of $G$-local systems on $C$ is a Beck--Chevalley fibration over $\Spc^G$ (\cref{LocalSystemsIsBCFibration}). We then explicate the relevant $\pi$-finiteness and truncatedness conditions. These conditions turn out to be related to the notion of \emph{$G$-semiadditivity} studied in \cite{nardin}.
\begin{ntn}
Let $C$ be a $G$-$\infty$-category and let $U$ be a finite $G$-set with orbit decomposition $U \simeq \coprod_{i=1}^n U_i$. We then write $C_U \coloneq \prod_{i=1}^n C_{U_i}$. Moreover, for a map of finite $G$-sets $f: U \to V$, we write $f^*: C_V \to C_U$ for the evident restriction functor.
\end{ntn}
\begin{dfn} \label{dfn:admitCoproducts}
Let $C$ be a $G$-$\infty$-category. Then $C$ \emph{admits finite $G$-coproducts} if for every map of finite $G$-sets $f: U \to V$, $f^*: C_V \to C_U$ admits a left adjoint $f_!$, and the Beck--Chevalley condition is satisfied, i.e., for every pullback square of finite $G$-sets
\[ \begin{tikzcd}
U' \ar{r}{f'} \ar{d}{g'} & V' \ar{d}{g} \\
U \ar{r}{f} & V
\end{tikzcd} \]
the exchange transformation $f'_! g'^* \Rightarrow g^* f_!$ is an equivalence. Dually, we have the evident notion of when $C$ admits finite $G$-products, and we denote the resulting right adjoints as $f_*$.
Suppose now that $C$ admits finite $G$-products and $G$-coproducts. We say that $C$ is \emph{$G$-semiadditive} if $C$ is fiberwise pointed and for all maps of finite $G$-sets $f: U \to V$, the canonical natural transformation $\chi: f_! \Rightarrow f_*$ is an equivalence.\footnote{$\chi$ is constructed as in \cite[5.2]{nardin} or \cref{dfn:semiadditive}.}
\end{dfn}
We now have that for a suitably $G$-bicomplete $G$-semiadditive $G$-$\infty$-category $C$ and any $G$-functor $X: B^{\psi}_G K \to C$, one has a \emph{parametrized norm map}
\[ \Nm: X_{h_G K} \to X^{h_G K}. \]
\begin{thmx}[{\cref{thm:EquivalentTateConstructions}}] \label{thmB}
Suppose $C = \underline{\Sp}^G$. Then the two parametrized norm maps constructed via ambidexterity theory and recollement theory (\cref{obs:NormCofiberSequence}) coincide.
\end{thmx}
\begin{rem}
After \cref{thmB}, we establish \cref{introobs1} and \ref{introobs2} as a corollary of some general properties of norm maps established in \cite[\S 4.2]{HL13}.
\end{rem}
There is yet a third approach to the Tate construction via \emph{assembly maps} in the sense of Weiss--Williams \cite{WW95}, which was first studied by John Klein \cite{Klein2001} and later taken up in the $\infty$-categorical context by Nikolaus and Scholze \cite[\S I.3-4]{NS18}, who used it together with the multiplicative theory of the Verdier quotient to prove that the Tate construction uniquely admits a lax symmetric monoidal structure \cite[Thm.~I.4.1]{NS18}. We now let $K$ be a compact Lie group and write $\SS^{\mathfrak{a}}$ for $\Sigma^{\infty}$ of the one-point compactification of the adjoint representation $\mathfrak{a}$ of $K$. Suppose also that $\widehat{G} \cong K \rtimes G$ for some $G$-action on $K$,\footnote{We make this assumption to identify the dualizing $G$-spectrum as $\SS^{\mathfrak{a}}$; the parametrized assembly map itself exists for any $G$-space. We also don't need to suppose $\widehat{G} \cong K \rtimes G$ if $K$ is finite, since the dualizing $G$-spectrum is then necessarily the unit.} and note that $\SS^{\mathfrak{a}}$ canonically admits the structure of a $G$-spectrum with $\psi$-twisted $K$-action (\cref{obs:ActionsLieGroup}).
\begin{thmx} \label{thmC} There exists a parametrized assembly map
\[ (\SS^{\mathfrak{a}} \wedge -)_{\underline{h}_G K} \xto{\alpha} (-)^{\underline{h}_G K}: \underline{\Fun}_G(B^{\psi}_G K, \underline{\Sp}^G) \to \underline{\Sp}^G \]
such that:
\begin{enumerate}
\item If $K$ is finite, $\alpha$ coincides with the norm map $(-)_{\underline{h}_G K} \to (-)^{\underline{h}_G K}$ of \cref{introobs1}.
\item $\alpha$ is terminal among all natural transformations $\alpha': F \to (-)^{h_G K}$ from $G$-colimit preserving $G$-functors $F$.
\item $\alpha$ is an equivalence when restricted to the full $G$-subcategory $\underline{\Fun}_G(B^{\psi}_G K, \underline{\Sp}^G)^{\omega}$ of compact objects.
\end{enumerate}
Moreover, properties (2) and (3) uniquely specify $\alpha$.
\end{thmx}
\begin{proof}
Combine \cref{thm:paramTateGeneral}, \cref{rem:agreement}, and \cref{prp:DualizingSpectrum}.
\end{proof}
Given \cref{thmC}, we may unambiguously write $(-)^{\underline{t}_G K}$ for the cofiber of $\alpha$. We now wish to understand the multiplicative properties of $(-)^{\underline{t}_G K}$ as a $G$-functor. Note that, at least if $K$ is finite, \cref{dfn:FirstDefn} already endows $(-)^{t_G K}$ with the structure of a lax symmetric monoidal functor. However, in the context of $G$-$\infty$-categories, more elaborate multiplicative structure on $\underline{\Sp}^G$ in the form of the \emph{Hill-Hopkins-Ravenel norm functors} is present \cite{HHR,hillhopkins,BachmannHoyoisNorms}, and these endow $\underline{\Sp}^G$ with the structure of a \emph{$G$-symmetric monoidal $\infty$-category} (\cref{dfn:GSMC}). One also has a pointwise $G$-symmetric monoidal structure on $\underline{\Fun}_G(I, \underline{\Sp}^G)$ (\cref{exm:pointwiseGSMC}). We then don't expect to be able to uniquely characterize the lax symmetric monoidal structure on $(-)^{t_G K}$, but rather only a to-be-defined \emph{lax $G$-symmetric monoidal structure} on $(-)^{\underline{t}_G K}$. Indeed, by developing the theory of parametrized Verdier quotients, $G$-$\otimes$-ideals, and induced $G$-objects, we are able to prove:
\begin{thmx}[{\cref{cor:LaxGSymmetricMonoidalTate}}] \label{thmD}
The $G$-functor $(-)^{\underline{t}_G K}$ and natural transformation $\beta: (-)^{\underline{h}_G K} \to (-)^{\underline{t}_G K}$ uniquely admit the structure of a lax $G$-symmetric monoidal functor and morphism thereof.\footnote{As with \cite[Thm.~I.4.1(vi)]{NS18}, the uniqueness assertion is really about $\beta$ and not $(-)^{\underline{t}_G K}$ in isolation.}
\end{thmx}
\begin{wrn}
The universal property of the map $\beta$ in \cref{thmD} is with respect to mapping $(-)^{\underline{h}_G K}$ into lax $G$-symmetric monoidal functors that vanish on ``induced objects''. This turns out to be a subtle notion in the parametrized setting: see \cref{dfn:InducedObjects}.
\end{wrn}
\begin{rem}
In applications to real trace methods, \cref{thmD} is important since one wants the real topological cyclic homology of a $C_2$-$E_{\infty}$-algebra to again be a $C_2$-$E_{\infty}$-algebra.
\end{rem}
Lastly, it is imperative for applications to be able to understand the various geometric fixed points of the parametrized Tate construction in terms of computationally accessible spectra. To this end, we leverage recent advances in reconstructing $G$-spectra from their geometric fixed points \cite{Glasman17,AMGR-NaiveApproach} in order to give an explicit formula for the geometric fixed points of an $\cF$-complete $G$-spectrum, for $G$ a finite group and $\cF$ any family of subgroups of $G$. To explain our result, we need to introduce a few more definitions, cf. \cref{section:Reconstruction}.
\begin{dfn}
Given a preordered set $(S, <)$, we let $\sd(S)$ denote its \emph{barycentric subdivision}. This is the category whose objects are \emph{strings} $[x_0 < x_1 < \cdots < x_n]$ in $S$ and where a morphism
\[ \overline{\alpha} : [x_0 < x_1 \cdots < x_n] \to [y_0 < y_1 < \cdots < y_m] \]
is the data of an injective map $\alpha:\into{[n]}{[m]}$ of totally ordered sets and a commutative diagram in $S$ (regarded as a category)
\[ \begin{tikzcd}
x_0 \ar{r} \ar{d}{\cong} & x_1 \ar{r} \ar{d}{\cong} & \cdots \ar{r} & x_n \ar{d}{\cong} \\
y_{f(0)} \ar{r} & y_{f(1)} \ar{r} & \cdots \ar{r} & y_{f(n)}
\end{tikzcd} \]
where the vertical maps are isomorphisms.\footnote{Note that $\sd(S)$ is then a category associated to a preordered set.} By a slight abuse of terminology, we call such morphisms \emph{string inclusions}.
\end{dfn}
\begin{dfn}
Let $\mathfrak{S}$ be the preordered set whose objects are subgroups of $G$ and where the relation is that of subconjugacy.
\end{dfn}
For any subgroup $H \leq G$, let $W_G H = N_G H / H$ denote the Weyl group, and recall that the geometric $H$-fixed points functor
$$(-)^{\phi H}: \Sp^G \to \Sp^{h W_G H} = \Fun(B W_G H, \Sp)$$
admits a fully faithful right adjoint $\iota_H$.
\begin{dfn}
Let $H_0$ and $H_1$ be subgroups of $G$. The \emph{generalized Tate functor} $\tau^{H_1}_{H_0}$ is defined to be the composite
\[ \begin{tikzcd}
\tau^{H_1}_{H_0}: \Sp^{h W_G H_0} \ar[hookrightarrow]{r}{\iota_{H_0}} & \Sp^G \ar{r}{(-)^{\phi H_1}} & \Sp^{h W_G H_1}.
\end{tikzcd} \]
\end{dfn}
\begin{exm}
\begin{enumerate}[leftmargin=*]
\item If $H_0$ is not subconjugate to $H_1$, then $\tau^{H_1}_{H_0}$ is the zero functor. If $H_0$ is conjugate to $H_1$, then $\tau^{H_1}_{H_0}$ is an equivalence.
\item If $H_0 = 1$ and $H_1 = G$, then $\tau^G$ is the \emph{proper Tate construction}. This agrees with the ordinary Tate construction if $G$ is cyclic of prime order, but not in general. For instance, if $G = C_{p^2}$, then
$$(-)^{\tau C_{p^2}} \simeq ((-)^{h C_p})^{t C_{p^2}/ C_p}.$$
\end{enumerate}
\end{exm}
\begin{obs}\label{rem:canonical1}
Given a sequence of properly subconjugate subgroups $H_0 < H_1 < H_2$, generalized Tate functors only \emph{left laxly} (i.e., oplaxly) compose: that is, one has a canonical natural transformation
$$\can: \tau^{H_2}_{H_0} \Rightarrow \tau^{H_2}_{H_1} \circ \tau^{H_1}_{H_0}$$
which fails to be an equivalence in general. More generally, for any string inclusion
$$\overline{\alpha}: [H_0 < \cdots < H_n] \to [K_0 < \cdots < K_m]$$
such that $\alpha(0) = 0$ and $\alpha(n) = m$, one has a canonical natural transformation\footnote{Note that we implicitly make the identifications $\Sp^{h W_G H_0} \simeq \Sp^{h W_G K_0}$ and $\Sp^{h W_G H_n} \simeq \Sp^{h W_G K_m}$ using the conjugacy relations. All issues of conjugacy are dealt with rigorously in the main body of the paper.}
$$ \can_{\overline{\alpha}}: \tau^{H_n}_{H_{n-1}} \circ \cdots \circ \tau^{H_1}_{H_0} \Rightarrow \tau^{K_m}_{K_{m-1}} \circ \cdots \circ \tau^{K_1}_{K_0}.$$
\end{obs}
\begin{obs}\label{rem:canonical2}
For any string inclusion
$$\overline{\alpha}: [H_0 < \cdots < H_n] \to [K_0 < \cdots < K_m < H_0 < \cdots < H_n]$$
such that $\alpha$ is the inclusion of a convex subset, one has a natural transformation
\[ \can_{\overline{\alpha}}: \tau^{H_n}_{H_{n-1}} \circ \cdots \circ \tau^{H_{1}}_{H_0} \circ (-)^{\phi H_0} \Rightarrow \tau^{H_n}_{H_{n-1}} \circ \cdots \circ \tau^{K_1}_{K_0} \circ (-)^{\phi K_0} \]
given by applying $\tau^{H_n}_{H_{n-1}} \circ \cdots \circ \tau^{H_{1}}_{H_0} \circ (-)^{\phi H_0}$ to the composite of unit maps
\[ \id \Rightarrow \iota_{K_0} (-)^{\phi K_0} \Rightarrow \cdots \Rightarrow \iota_{K_n} (-)^{\phi K_n} \circ \cdots \circ \iota_{K_0} (-)^{\phi K_0}. \]
\end{obs}
\begin{ntn}
For a subgroup $H$ not in $\cF$, let $J_H \subset \sd(\mathfrak{S})$ be the full subcategory on strings $[K_0 < \cdots < K_n < H]$ such that $K_i \in \cF$ for all $1 \leq i \leq n$.
\end{ntn}
\begin{thmx}[{\cref{cor:FormulaForGeomFixedPointsOfCompleteSpectrum}}] \label{MT:GFP}
Suppose that $X \in \Sp^G$ is an $\cF$-complete $G$-spectrum. Then for each subgroup $H \notin \cF$, the $H$-geometric fixed points $ X^{\phi H} \in \Sp^{hW_G H}$ can be computed as a limit involving only the geometric fixed points $X^{\phi K}$ for subgroups $K \in \cF$ and generalized Tate functors thereon. More precisely, we may functorially associate to $X$ a functor $F: J_H \to \Sp^{h W_G H}$ such that $X^{\phi H} \simeq \lim_{J_H} F$ and
\begin{enumerate}
\item $F([K_0 < \cdots < K_n < H]) = \tau^{H}_{K_n} \tau^{K_n}_{K_{n-1}} \cdots \tau^{K_1}_{K_0} (X^{\phi K_0})$.
\item F sends a string inclusion $[K_0 < \cdots < K_n < H] \to [L_0 < \cdots < L_m < H]$ to the composite of the canonical maps of \cref{rem:canonical1} and \cref{rem:canonical2} associated to the factorization
\[ [K_0 < \cdots < K_n < H] \to [L_0 < \cdots < K_0 < \cdots < K_n < H] \to [L_0 < \cdots < L_m < H] \]
in which the first string inclusion is the inclusion of a convex subset.
\end{enumerate}
\end{thmx}
\begin{rem}
The formula of \cref{MT:GFP} in the case where $G = D_{2 p} = \mu_p \rtimes C_2$ and $\cF = \Gamma_{\mu_p}$ plays an important role in our proof of the \emph{dihedral Tate orbit lemma} in \cite{QS21b}, which is the key computational input needed to show that genuine and Borel real cyclotomic spectra agree in the ``underlying bounded-below'' case.
\end{rem}
\begin{rem}
Our proof of \cref{MT:GFP} is based off of an alternative proof of the Ayala--Mazel-Gee--Rozenblyum reconstruction theorem (\cref{thm:AMGRoriginal}) due to the second author. On the other hand, given their theorem and the appropriate cofinality arguments, it should not be difficult to derive \cref{MT:GFP}. We were motivated to take a slightly longer route in reproving their theorem in order to illustrate how a detailed understanding of the combinatorics of barycentric subdivision can substitute for any explicit usage of $(\infty,2)$-category theory in the proof.\footnote{Of course, using the theory of locally cocartesian fibrations may be thought of as implicitly using $(\infty,2)$-category theory.}
\end{rem}
\begin{rem}
Shortly after the original appearance of this article, Ayala--Mazel-Gee--Rozenblyum published a greatly expanded version of their work in \cite{AMGRb}. There, they derive a formula for the categorical fixed points of a $G$-spectrum in terms of its geometric fixed points \cite[Obs.~5.4.3]{AMGRb} that should be contrasted with \cref{MT:GFP}.
\end{rem}
\begin{rem}
In \cite[Thm.~A]{ayala2021derived}, Ayala--Mazel-Gee-Rozenblyum give a formula for the generalized Tate functors in terms of \emph{proper} Tate constructions.
\end{rem}
\subsection{Outline}
In \cref{section:FirstEquivariantSection}, we discuss background and the $\cF$-recollement on $\Sp^G$. In \cref{section:EquivariantConventions}, we recall the $\infty$-category of genuine $G$-spectra $\Sp^G$ from Bachmann--Hoyois \cite{BachmannHoyoisNorms}. In \cref{SS:families}, we recall the theory of families of subgroups of a finite group and describe the $\cF$-recollement on $\Sp^G$ for any family of subgroups $\cF$ of $G$. We describe the relationship between the $\cF$-recollement and the acyclization, completion, and localization functors of Mathew--Naumann--Noel \cite{MATHEW2017994}, as well as the canonical fracture of Ayala--Mazel-Gee--Rozenblyum \cite{AMGR-NaiveApproach}. In \cref{section:Reconstruction}, we reprove Thm. A from \emph{loc. cit.} which identifies genuine $G$-spectra in terms of their geometric fixed points. We then prove \cref{MT:GFP} and apply it in the cases relevant for \cite{QS21b}.
In \cref{Sec:RelThy}, we specialize to the $N$-free family $\Gamma_N = \{ H: H \cap N = 1 \}$ of subgroups for a normal subgroup $N \leq G$. In \cref{SS:NNaive}, we define the $\infty$-category of $N$-naive $G$-spectra $\Sp^G_{N\textnormal{-naive}}$ and discuss its fundamental properties. In \cref{SS:NBorel}, we define the $\infty$-category of $N$-Borel $G$-spectra $\Sp^G_{N\textnormal{-Borel}}$. We study the forgetful functor $\mathscr{U}_b[N]: \Sp^G \to \Sp^G_{N\textnormal{-Borel}}$, its adjoints, and the relationship to the $\Gamma_N$-recollement, proving \cref{thmA}. As an application, we obtain the stable symmetric monoidal recollement describing $\Sp^{D_{2p^n}}$, which we use in \cite{QS21b} to analyze real cyclotomic spectra.
In \cref{section:NormMaps}, we define the parametrized Tate construction in the finite group case. In \cref{SS:Ambi}, we extend the ambidexterity theory of Hopkins--Lurie \cite{HL13} to the parametrized setting. In \cref{SS:PTate}, we use parametrized ambidexterity to define a parametrized norm map between parametrized homotopy orbits and parametrized homotopy fixed points for any finite group extension of $G$. The parametrized Tate construction is then defined as the cofiber of this map. Reconciling this definition with \cref{dfn:FirstDefn}, we prove \cref{thmB} and deduce several useful properties mentioned above, such as Observations \ref{obs:NormCofiberSequence}--\ref{introobs2}. We also prove \cref{MT:GM161} regarding the inverse limit formula for the parametrized Tate construction.
In \cref{Sec:Assembly}, we define the parametrized Tate construction for an extension of a finite group by a compact Lie group by extending Klein's assembly map definition of the Tate construction (cf. \cite{Klein2001} and \cite[{\S I.4}]{NS18}) to the parametrized setting. In \cref{SS:GSM}, we recall several notions related to $G$-symmetric monoidal structures used throughout the sequel. We then discuss parametrized Verdier quotients and $G$-$\otimes$-ideals in \cref{SS:Verdier} and the parametrized notion of induced objects in \cref{SS:Induced}. Our key result is that the $G$-subcategory of induced objects is a $G$-$\otimes$-ideal (\cref{cor:InducedObjectsFormGTensorIdeal}). In \cref{SS:MainAssembly}, we apply all of these ideas to define the parametrized Tate construction for infinite groups (\cref{dfn:paramTateGeneral}) and prove Theorems~\ref{thmC} and \ref{thmD}. We also discuss an example (\cref{exm:CircleTate}) for $\widehat{G} = O(2)$ which plays an important role in our study of real cyclotomic spectra in \cite{QS21b}. Finally, we digress in the middle to upgrade the generalized Segal conjecture to a statement about Tambara functors (\cref{Thm:Segal}).
\subsection{Notation and terminology}
We assume knowledge of the theory of recollements in this paper and refer to \cite{ShahRecoll} as our primary reference. Given a stable $\infty$-category $\sX$ decomposed by a stable recollement $(\sU, \sZ)$, we will generically label the recollement adjunctions as
\[ \begin{tikzcd}[column sep=4em]
\sU \ar[hookrightarrow, shift left=2]{r}{j_!} \ar[hookrightarrow, shift right=4]{r}[swap]{j_*} & \sX \ar[shift left=1]{l}[description]{j^*} \ar[shift left=2]{r}{i^*} \ar[shift right=1, hookleftarrow]{r}[swap, description]{i_*} \ar[shift right=4]{r}[swap]{i^!} & \sZ.
\end{tikzcd} \]
Here $j^* i_* = 0$ determines the directionality of the recollement.
As is already apparent from the introduction, we will also use concepts from parametrized higher category theory in this paper, mostly in Sections \ref{section:NormMaps} and \ref{Sec:Assembly}. Given any $\infty$-category $S$, an \emph{$S$-$\infty$-category} is a cocartesian fibration over $S$, and we then have attendant notions of $S$-(co)limits and $S$-Kan extensions. Note then that the terminology $G$-$\infty$-category, $G$-functor, etc. is synonymous with $\sO_G^{\op}$-$\infty$-category, $\sO_G^{\op}$-functor, etc. Apart from the basic reference \cite{Exp2}, we refer the reader to \cite[Sec.~2]{Exp2b} for a quick overview of the theory of $S$-(co)limits and $S$-Kan extensions.\footnote{In that reference, we set $T = S^{\op}$ and instead speak of $T$-$\infty$-categories, $T$-functors, etc.} Let us also highlight the following example, which locates \cref{dfn:admitCoproducts} within the formalism of parametrized higher category theory:
\begin{exm}[Corepresentable $S$-diagrams] \label{exm:corepresentableDiagrams} Suppose that $T = S^{\op}$ admits \emph{multipullbacks}, i.e., the finite coproduct completion $\FF_T$ of $T$ admits pullbacks. For example, $\sO_G$ satisfies this condition. We call $\FF_T$ the $\infty$-category of \emph{finite $T$-sets} and $T \subset \FF_T$ the \emph{orbits}. For $U \in \FF_T$ with orbit decomposition $\coprod_{i \in I} U_i$, let
$$\ul{U} \coloneq \coprod_{i \in I} S^{U_i/} \to S$$
be the corresponding \emph{$S$-$\infty$-category of points}, and note that the assignment $\fromto{U}{\ul{U}}$ is covariant in morphisms in $\FF_T$. Let $\alpha: U \to V$ be a morphism in $\FF_T$ such that $V$ is an orbit. Let $x_i \in C_{U_i}$ be a set of objects for all $i \in I$ and write $(x_i): \ul{U} \to C_{\ul{V}}$ for the $S^{V/}$-functor determined by the $x_i$. Then the \emph{$S$-coproduct along $\alpha$}
$$\coprod_{\alpha} x_i \in C_V$$
is defined to be the $S^{V/}$-colimit of $(x_i)$. A \emph{finite $S$-coproduct} is any $S^{V/}$-colimit of this form. We have that $C$ admits all finite $S$-coproducts if and only if the following conditions obtain \cite[Prop.~5.12]{Exp2}:
\begin{enumerate}
\item For all $V \in S$, $C_V$ admits finite coproducts, and for all morphisms $\alpha: V \to W$ in $T$, the restriction functor $\alpha^{\ast}: C_W \to C_{V}$ preserves finite coproducts.
\item For all morphisms $\alpha: V \to W$ in $T$, $\alpha^{\ast}$ admits a left adjoint $\alpha_!$.
\item Given $U \in \FF_T$ with orbit decomposition $\coprod_{i \in I} U_i$, let $C_U \coloneq \prod_{i \in I} C_{U_i}$ and extend $\alpha^{\ast}$ and $\alpha_!$ to be defined for all morphisms $\alpha$ in $\FF_T$ in the obvious way.\footnote{e.g., if $\alpha: U \to V$ is a map with $V$ an orbit, then $\alpha_!(x_i) = \coprod_{i \in I}(\alpha_i)_!(x_i)$ for $\alpha_i: U_i \to V$ the restriction of $\alpha$ to $U_i$.} Then the \emph{Beck-Chevalley conditions} hold: for every pullback square
\[ \begin{tikzcd}[row sep=4ex, column sep=4ex, text height=1.5ex, text depth=0.25ex]
U' \ar{r}{\alpha'} \ar{d}{\beta'} & V' \ar{d}{\beta} \\
U \ar{r}{\alpha} & V
\end{tikzcd} \]
in $\FF_T$, the exchange transformation $(\alpha')_! (\beta')^{\ast} \Rightarrow \beta^{\ast} \alpha_!$ is an equivalence.
\end{enumerate}
In this case, the $S$-coproduct $\coprod_{\alpha} x_i$ above is computed by $\alpha_!(x_i)$.
Dually, $C$ admits all finite $S$-products if and only if the analogous conditions hold with respect to finite products in the fibers and right adjoints $\alpha_{\ast}$.
\end{exm}
We will also use the following terminology:
\begin{dfn}[{\cite[Def.~8.3]{Exp2}}]
Let $C,D$ be $S$-$\infty$-categories. Then an \emph{$S$-adjunction} is a relative adjunction
\[ \adjunct{F}{C}{D}{G} \]
in the sense of \cite[Def.~7.3.2.2]{HA} such that $F$ and $G$ are both $S$-functors.
\end{dfn}
\subsection{Acknowledgments}
This work is an expansion of \cite[Secs.~3-5]{QS19}. The main changes are as follows:
\begin{enumerate}
\item We extended the theory to compact Lie groups.
\item We added the material on parametrized assembly.
\item We proved that the parametrized Tate construction uniquely admits the structure of a lax $G$-symmetric monoidal functor.
\item We added an application to the generalized Segal conjecture.
\item We added an inverse limit formula for the parametrized Tate construction.
\end{enumerate}
We would like to thank Mark Behrens, Andrew Blumberg, Emanuele Dotto, Jeremy Hahn, Kristian Moi, Irakli Patchkoria, Dylan Wilson, Inna Zakharevich, and Mingcong Zeng for helpful discussions. The authors were partially supported by NSF grant DMS-1547292. J.S. was also funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) under Germany’s Excellence Strategy EXC 2044–390685587, Mathematics Münster: Dynamics–Geometry–Structure.
\section{The \texorpdfstring{$\cF$}{F}-recollement on \texorpdfstring{$\Sp^G$}{SpG}}
\label{section:FirstEquivariantSection}
Let $G$ be a finite group. In this section, we introduce and study recollements on the $\infty$-category $\Sp^G$ of $G$-spectra determined by a family $\cF$ of subgroups of $G$. We then apply \cite[Thm.~3.35]{ShahRecoll} to reprove a theorem of Ayala, Mazel-Gee, and Rozenblyum that reconstructs $\Sp^G$ from its geometric fixed points (\cref{thm:GeometricFixedPointsDescriptionOfGSpectra}). As a corollary, we deduce a limit formula (\cref{cor:FormulaForGeomFixedPointsOfCompleteSpectrum}) for the geometric fixed points of an $\cF$-complete spectrum by means of the pointwise formula of \cite[Thm.~3.29]{ShahRecoll}, which will play an important role in our proof of the dihedral Tate orbit lemma in \cite{QS21b} (cf. \cref{exm:DihedralEven} and \cref{exm:DihedralOdd}).
\subsection{Conventions on equivariant stable homotopy theory}
\label{section:EquivariantConventions}
At the outset, let us be clear about which foundations for equivariant stable homotopy theory are employed in this paper. In their monograph, Nikolaus and Scholze choose to work with the classical point-set model of orthogonal $G$-spectra \cite[Def.~II.2.3]{NS18}, then obtaining the $\infty$-category $\Sp^G$ of $G$-spectra via inverting equivalences \cite[Def.~II.2.5]{NS18}. In contrast, we will use the foundations laid out by Bachmann and Hoyois in \cite[\S 9]{BachmannHoyoisNorms}, which attaches to every profinite groupoid $X$ a presentable, stable, and symmetric monoidal $\infty$-category $\SH(X)$ such that for $X = B G$, $\SH(B G)$ is equivalent to $\Sp^G$ as defined in \cite{NS18} (cf. the remark prior to \cite[Lem.~9.5]{BachmannHoyoisNorms}). In fact, we will only need the Bachmann-Hoyois construction for finite groupoids.
\begin{dfn} \label{dfn:BachmannHoyoisFunctor} Let $\Gpd_{\fin}$ be the $(2,1)$-category of finite groupoids, and let $$\sH, \sH_{\sbullet}, \SH: \Gpd_{\fin}^{\op} \to \CAlg(\Pr^{\mr{L}})$$
denote the (restriction of the) functors constructed in \cite[\S 9.2]{BachmannHoyoisNorms}. For a map $f: X \to Y$ of finite groupoids, write $f^{\ast}$ for the associated functor and $f_{\ast}$ for its right adjoint.
\end{dfn}
\begin{rem} Let $X = BG$. Then $\sH(BG) \simeq \Spc^G \coloneq \Fun(\sO^{\op}_G, \Spc)$, the $\infty$-category of $G$-spaces defined as presheaves on the orbit category $\sO_G$, and likewise $\sH_{\sbullet}(BG)$ is the $\infty$-category $\Spc^G_{\ast}$ of pointed $G$-spaces. As we already mentioned, $\SH(BG) \simeq \Sp^G$ is the $\infty$-category of $G$-spectra, defined as the filtered colimit taken in $\Pr^L$
\[ \Spc^G_{\ast} \xto{\Sigma^{\rho}} \Spc^G_{\ast} \xto{\Sigma^{\rho}} \Spc^G_{\ast} \xto{\Sigma^{\rho}} \cdots, \]
where $\rho$ is the regular $G$-representation. In addition, by \cite[Exm.~9.11]{BachmannHoyoisNorms} $\Sp^G$ is equivalent to the $\infty$-category of \emph{spectral Mackey functors} on finite $G$-sets that was studied by Barwick \cite{M1} and Guillou-May \cite{guillou2}.
\end{rem}
Note that by definition, $f^{\ast}: \SH(Y) \to \SH(X)$ is the symmetric monoidal left Kan extension of $\sH_{\sbullet}(Y) \xto{f^{\ast}} \sH_{\sbullet}(X) \xto{\Sigma^{\infty}} \SH(X)$ along $\Sigma^{\infty}: \sH_{\sbullet}(Y) \to \SH(Y)$. Therefore:
\begin{enumerate} \item Suppose $f: BH \to BG$ is the map of groupoids induced by an injective group homomorphism $H \to G$. Then $f^{\ast}: \Sp^G \to \Sp^H$ is homotopic to the usual restriction functor, and $f_{\ast}: \Sp^H \to \Sp^G$ is homotopic to the usual induction functor. Instead of $f^{\ast} \dashv f_{\ast}$, we will typically write this adjunction as $\res^G_H \dashv \ind^G_H$. Note that this adjunction is ambidextrous and satisfies the projection formula (in fact, \cite[Lem.~9.4(3)]{BachmannHoyoisNorms} establishes the projection formula for any finite covering map).
\item Suppose $f: BG \to BG/N$ is the map of groupoids induced by a surjective group homomorphism $G \to G/N$. Then $f^{\ast}: \Sp^{G/N} \to \Sp^G$ is homotopic to the usual inflation functor, which we denote as $\inf^N$. The right adjoint to $\inf^N$ is the \emph{categorical fixed points} functor
$$ \Psi^N: \Sp^G \to \Sp^{G/N}. $$
Now suppose $H \leq G$ is any subgroup and let $W_G H = N_G H / H$ be the Weyl group of $H$. Then we will also write
$$ \Psi^H: \Sp^G \xtolong{\res^G_{N_G H}}{1.5} \Sp^{N_G H} \xto{\Psi^H} \Sp^{W_G H} $$
\end{enumerate}
Given a $G$-spectrum $X$, we introduce notation to distinguish the underlying spectrum of $\Psi^H X$.
\begin{ntn} For a $G$-spectrum $X$ and subgroup $H \leq G$, we let $X^H = \res^{W_G H} \Psi^H (X)$.\footnote{With respect to the description of $\Sp^G$ as spectral Mackey functors, $X^H$ is given by evaluation at $G/H$.}
\end{ntn}
Since the restriction functor $\Sp^{W_G H} \to \Sp$ lifts to $\Fun(B W_G H, \Sp)$, the spectrum $X^H$ also comes endowed with a $W_G H$-action.
\begin{rem} By stabilizing the adjointability relations in \cite[Lem.~9.4]{BachmannHoyoisNorms}, it follows that that for any pullback square of finite groupoids
\[ \begin{tikzcd}[row sep=4ex, column sep=4ex, text height=1.5ex, text depth=0.25ex]
W \ar{r}{f} \ar{d}{g} & Y \ar{d}{g} \\
X \ar{r}{f} & Z,
\end{tikzcd} \]
the canonical natural transformation $f^{\ast} g_{\ast} \to f_{\ast} g^{\ast}$ of functors $\SH(X) \to \SH(Y)$ is an equivalence. In particular, we have an equivalence $X^H \simeq \Psi^H \res^G_H (X)$.
\end{rem}
We now turn to the \emph{geometric fixed points} and \emph{Hill-Hopkins-Ravenel norm} functors.
\begin{dfn} \label{BachmannHoyoisFunctorNorms} Let $\sH^{\otimes}, \sH_{\sbullet}^{\otimes}, \SH^{\otimes}: \Span(\Gpd_{\fin}) \to \CAlg(\Cat_{\infty}^{\mr{sift}})$ be the (restrictions of the) functors defined as in \cite[\S 9.2]{BachmannHoyoisNorms}, which on the subcategory $\Gpd_{\fin}^{\op}$ restrict to the functors $\sH, \sH_{\sbullet}, \SH$ of \cref{dfn:BachmannHoyoisFunctor}. For a map of finite groupoids $f: X \to Y$, write $f_{\otimes}$ for the associated covariant functor.
\end{dfn}
Parallel to the discussion above, we note \cite[Rem.~9.9]{BachmannHoyoisNorms}:
\begin{enumerate} \item Suppose $f: BH \to BG$ for a subgroup $H \leq G$. Then $f_{\otimes}: \Sp^H \to \Sp^G$ is homotopic to the multiplicative norm functor $N^G_H$ introduced by Hill, Hopkins, and Ravenel \cite{HHR}.
\item Suppose $f: BG \to B(G/N)$. Then $f_{\otimes}: \Sp^G \to \Sp^{G/N}$ is homotopic to the usual geometric fixed points functor $\Phi^N$. For $H \leq G$ any subgroup, we also write
$$ \Phi^H: \Sp^G \xtolong{\res^G_{N_G H}}{1.5} \Sp^{N_G H} \xto{\Phi^H} \Sp^{W_G H}. $$
\end{enumerate}
\begin{ntn} For a $G$-spectrum $X$ and subgroup $H \leq G$, we let $X^{\phi H} = \res^{W_G H} \Phi^H (X)$. Also let
$$ \phi^H: \Sp^G \xto{\Phi^H} \Sp^{W_G H} \xto{\res} \Fun(B W_G H, \Sp). $$
\end{ntn}
\begin{rem} Because $\SH^{\otimes}$ is defined on $\Span(\Gpd_{\fin})$, we have that for any pullback square of finite groupoids
\[ \begin{tikzcd}[row sep=4ex, column sep=4ex, text height=1.5ex, text depth=0.25ex]
W \ar{r}{f} \ar{d}{g} & Y \ar{d}{g} \\
X \ar{r}{f} & Z,
\end{tikzcd} \]
there is a canonical equivalence $f^{\ast} g_{\otimes} \simeq f_{\otimes} g^{\ast}$ of functors $\SH(X) \to \SH(Y)$. In particular, we have an equivalence $X^{\phi H} \simeq \Phi^H \res^G_H X$.
\end{rem}
\begin{rem} We will use some additional features of these fixed points functors:
\begin{enumerate} \item For any subgroup $H \leq G$, the functor $\Psi^H$ is colimit-preserving, since the inflation functors preserve dualizable and hence compact objects; indeed, by equivariant Atiyah duality \cite[\S III.5.1]{MR866482} every compact object in $\Sp^G$ is dualizable, and conversely, since the unit in $\Sp^G$ is compact, all dualizable objects in $\Sp^G$ are compact.
\item The functors $\{ (-)^H : H \leq G \}$ are jointly conservative, since the orbits $\Sigma^{\infty}_+ G/H$ corepresent $(-)^H$ and form a set of compact generators for $\Sp^G$.
\item The functors $\{ \phi^H : H \leq G \}$ are jointly conservative, since the evaluation functors $\ev_{G/H}$ are jointly conservative for $\Spc^G$, $\phi^H \Sigma^{\infty}_+ \simeq \Sigma^{\infty}_+ \ev_{G/H}$, and suspension spectra generate $\Sp^G$ under desuspensions and sifted colimits.
\end{enumerate}
\end{rem}
We will also need to use some aspects of the theory of $G$-$\infty$-categories in this work.
\begin{dfn} Let $\omega_G: \FF_G \to \Gpd_{\fin}$ be the functor that sends a finite $G$-set $U$ to its action groupoid $U//G$.
\end{dfn}
\begin{dfn} \label{dfn:GCategoryGSpectra} Define the \emph{$G$-$\infty$-category of $G$-spectra} $\underline{\Sp}^G \to \sO^{\op}_G$ to be the cocartesian fibration classified by $\SH \circ (\omega^{\op}_G|_{\sO^{\op}_G})$. In addition, let $\underline{\Sp}^{G, \otimes} \to \sO^{\op}_G \times \Fin_{\ast}$ be the cocartesian $\sO_G^{\op}$-family of symmetric monoidal $\infty$-categories classified by $\SH \circ (\omega^{\op}_G|_{\sO^{\op}_G})$ (when viewed as valued in $\CMon(\Cat_{\infty})$).
\end{dfn}
\begin{ntn}
We will typically write $\underline{C}$ to distinguish a $G$-$\infty$-category from the $\infty$-category $C$ given by the fiber of $\underline{C}$ over $G/G$.
\end{ntn}
\begin{rem} \label{rem:sliceCategoryPassage} For a subgroup $H$ of $G$, let
\[ \adjunct{\ind^G_H}{\FF_H}{\FF_G}{\res^G_H} \]
denote the induction-restriction adjunction, where $\ind^G_H(U) = G \times_H U$. Then $\ind^G_H: \sO_H \to \sO_G$ factors as $\sO_H \simeq (\sO_G)_{/(G/H)} \to \sO_G$. Moreover, $\omega_G \circ \ind^G_H$ and $\omega_H$ are canonically equivalent, so we have an equivalence of $H$-$\infty$-categories
\[ \underline{\Sp}^H \simeq \sO_H^{\op} \times_{\sO_G^{\op}} \underline{\Sp}^G. \]
\end{rem}
\begin{rem} Given a $G$-$\infty$-category $K$, we may endow $\Fun_G(K, \ul{\Sp}^G)$ with the pointwise symmetric monoidal structure of \cref{dfn:S-PointwiseMonoidal} with respect to the construction $\underline{\Sp}^{G, \otimes}$ of \cref{dfn:GCategoryGSpectra}.
\end{rem}
Finally, we will later need the $G$-symmetric monoidal structure on $\underline{\Sp}^G$ furnished by the Hill--Hopkins--Ravenel norms; we record this as \cref{exm:GSMCSpectra}.
\subsection{Basic theory of families}\label{SS:families}
\begin{dfn} \label{dfn:subconjugacyPoset} Given a finite group $G$, its \emph{subconjugacy category} $\fS[G]$ is the category whose objects are subgroups $H$ of $G$, and whose morphism sets are defined by
\begin{align*} \Hom_{\fS[G]}(H,K) := \begin{cases} \ast \quad \text{ if } H \text{ is subconjugate to } K, \\
\emptyset \quad \text{ otherwise}.
\end{cases}
\end{align*}
We will also write $\fS = \fS[G]$ if the ambient group $G$ is clear from context.
\end{dfn}
\begin{dfn} \label{dfn:family} A \emph{$G$-family} $\cF$ is a sieve in $\fS$, i.e., a full subcategory of $\fS$ whose set of objects is a set of subgroups of $G$ closed under subconjugacy.
\end{dfn}
\begin{rem} Abusing notation, we will also denote the set of objects of $\fS$ or a family $\cF$ by the same symbol. If we view morphisms in $\fS$ as defining a binary relation $\leq$ on the set of subgroups of $G$, then $\fS$ is a preordered set, which is a poset if $G$ is abelian. Although we generally reserve the expression $H \leq K$ for $H$ a subgroup of $K$, when discussing strings in the preordered set $\fS$ we will also write $\leq$ for its binary relation -- we trust the meaning to be clear from context.
\end{rem}
\begin{cnstr} \label{cnstr:GspaceFromGfamily} Given a $G$-family $\cF$, define $G$-spaces $E \cF$ and $\widetilde{E \cF}$ by the formulas
\begin{align*} E \cF^K = \begin{cases} \emptyset \; \text{ if } K \notin \cF, \\
\ast \; \text{ if } K \in \cF,
\end{cases}, \quad \text{ and } \quad
\widetilde{E \cF}^K = \begin{cases} S^0 \; \text{ if } K \notin \cF, \\
\ast \; \text{ if } K \in \cF.
\end{cases}
\end{align*}
We have a cofiber sequence of pointed $G$-spaces
\[ E \cF_+ \to S^0 \to \widetilde{E \cF}. \]
The unit map $S^0 \to \widetilde{E \cF}$ exhibits $\widetilde{E \cF}$ as an idempotent object \cite[Def.~4.8.2.1]{HA} of $\Spc^G_{\ast}$ with respect to the smash product, hence $\widetilde{E \cF}$ is a idempotent $E_{\infty}$-algebra by \cite[Prop.~4.8.2.9]{HA}.\footnote{This is also obvious since we are considering presheaves of sets.} Let $E \cF_+$ and $\widetilde{E \cF}$ also denote $\Sigma^{\infty}$ of the same pointed $G$-spaces. Then $\widetilde{E \cF}$ is an idempotent $E_{\infty}$-algebra in $\Sp^G$, and hence by \cite[Obs.~2.36]{ShahRecoll} defines a stable symmetric monoidal recollement
\[ \begin{tikzcd}[row sep=4ex, column sep=4ex, text height=1.5ex, text depth=0.25ex]
\Sp^{h \cF} \ar[shift right=1,right hook->]{r}[swap]{j_{\ast}} & \Sp^G \ar[shift right=2]{l}[swap]{j^{\ast}} \ar[shift left=2]{r}{i^{\ast}} & \Sp^{\Phi \cF} \ar[shift left=1,left hook->]{l}{i_{\ast}}
\end{tikzcd} \]
such that $\Sp^{\Phi \cF} \simeq \Mod_{\Sp^G}(\widetilde{E \cF})$. By \cite[Cor.~2.35]{ShahRecoll}, for any $X \in \Sp^G$ we have the \emph{$\cF$-fracture square}
\[ \begin{tikzcd}[row sep=4ex, column sep=4ex, text height=1.5ex, text depth=0.25ex]
X \ar{r} \ar{d} & X \otimes \widetilde{E \cF} \ar{d} \\
F(E \cF_+, X) \ar{r} & F(E \cF_+, X) \otimes \widetilde{E \cF}.
\end{tikzcd} \]
Following standard terminology, we say that a $G$-spectrum $X$ is \emph{$\cF$-torsion}, \emph{$\cF$-complete}, or \emph{$\cF^{-1}$-local} if it is in the essential image of $j_!$, $j_{\ast}$, or $i_{\ast}$, respectively. Note that for a $G$-spectrum $X$,
\begin{itemize}
\item $X$ is $\cF$-torsion if and only if $X \otimes E \cF_+ \xto{\simeq} X$ or $X \otimes\widetilde{E \cF} \simeq 0$.
\item $X$ is $\cF$-complete if and only if $X \xto{\simeq} F(E \cF_+, X)$ or $F(\widetilde{E \cF},X) \simeq 0$.
\item $X$ is $\cF^{-1}$-local if and only if $X \xto{\simeq} X \otimes \widetilde{E \cF}$ or $X \otimes E \cF_+ \simeq 0$.
\end{itemize}
\end{cnstr}
\begin{ntn} For a $G$-family $\cF$, we have already set $\Sp^{h \cF} \subset \Sp^G$ to be the full subcategory of $\cF$-complete $G$-spectra and $\Sp^{\Phi \cF} \subset \Sp^G$ to be the full subcategory of $\cF^{-1}$-local $G$-spectra. We also let $\Sp^{\tau \cF} \subset \Sp^G$ denote the full subcategory of $\cF$-torsion $G$-spectra.
In addition, if $\cF$ is the trivial family $\{1\}$, we will also write $E \cF = E G$, $\Sp^{h \cF} = \Sp^{h G}$, and refer to $\cF$-torsion or complete objects as \emph{Borel} torsion or complete.\footnote{Other authors refer to Borel torsion spectra as \emph{free} and Borel complete spectra as \emph{cofree}.} It is well-known that $\Sp^{h G} \simeq \Fun(B G, \Sp)$ (\cite[Prop.~6.17]{MATHEW2017994}, \cite[Thm.~II.2.7]{NS18}) -- we will later give two different generalizations of this fact (\cref{lem:LocallyClosedFibersAreBorel} and \cref{thm:BorelSpectraAsCompleteObjects}).
\end{ntn}
\begin{rem} \label{rem:torsionCompleteEquivalence} The functor $j_! j^{\ast}: \Sp^{h \cF} \xto{\simeq} \Sp^{\tau \cF}$ implements an equivalence between $\cF$-complete and $\cF$-torsion objects \cite[Prop.~7]{BarwickGlasmanNoteRecoll}.
\end{rem}
\begin{rem} \label{rem:MathewComparison} The endofunctors $j_! j^{\ast}$, $j_{\ast} j^{\ast}$, and $i_{\ast} i^{\ast}$ of $\Sp^G$ attached to a family $\cF$ agree with the $A_\cF$-acyclization, $A_\cF$-completion, and $A_\cF^{-1}$-localization functors in \cite{MATHEW2017994} defined with respect to the $E_{\infty}$-algebra
$$A_{\cF} \coloneq \prod_{H \in \cF} F(G/H_+, 1)$$
by \cite[Prps.~6.5-6.6]{MATHEW2017994}. Moreover, the theory of $A$-torsion, $A$-complete, and $A^{-1}$-local objects for a dualizable $E_{\infty}$-algebra $A$ (\cite[Part 1]{MATHEW2017994} under the hypotheses \cite[2.26]{MATHEW2017994}) extends the more general monoidal recollement theory for the idempotent object $1 \to U_A$ of \cite[Constr.~3.12]{MATHEW2017994}. For example, the $\cF$-fracture square for $\Sp^G$ given by \cite[Cor.~2.12]{ShahRecoll} agrees with the $A_{\cF}$-fracture square given by \cite[Thm.~3.20]{MATHEW2017994} (although we additionally consider the monoidal refinement \cite[Thm.~2.30]{ShahRecoll}).
As a separate consequence, we also have that $\Sp^{\tau \cF} \subset \Sp^G$ is the localizing subcategory generated by the orbits $\{G/H_+ : H \in \cF \}$. Also, $G/H_+$ is both $\cF$-complete and $\cF$-torsion.
\end{rem}
In the remainder of this subsection, we collect some basic results concerning $\cF$-recollements that we will need in the sequel. Classical references for this material are \cite[\S II]{MR866482} and \cite[\S 17]{GM95}, and other references include \cite[\S 6]{MATHEW2017994} and \cite[\S 2]{AMGR-NaiveApproach}.
\begin{lem} \label{lem:GeometricFixedPointsDetectionCriterion} Let $\cF$ be a $G$-family and let $X \in \Sp^G$.
\begin{enumerate} \item $X$ is $\cF^{-1}$-local if and only if $X^{\phi K} \simeq 0$ for all $K \in \cF$.
\item $X$ is $\cF$-torsion if and only if $X^{\phi K} \simeq 0$ for all $K \notin \cF$.
\end{enumerate}
Therefore, for a map $f: X \to Y$ in $\Sp^G$, $f$ is a $j^{\ast}$-equivalence if and only if $f^{\phi K}$ is an equivalence for all $K \in \cF$, and $f$ is an $i^{\ast}$-equivalence if and only if $f^{\phi K}$ is an equivalence for all $K \notin \cF$.
\end{lem}
\begin{proof} First note that for any $X \in \Sp^G$ and subgroup $K$ of $G$,
\begin{align*} (X \otimes E \cF_+)^{\phi K} \simeq X^{\phi K} \otimes (E \cF_+)^{\phi K} & \simeq \begin{cases} 0 \; \text{ if } K \notin \cF, \\
X^{\phi K} \; \text{ if } K \in \cF,
\end{cases} \\
(X \otimes \widetilde{E \cF})^{\phi K} \simeq X^{\phi K} \otimes \widetilde{E \cF} {}^{\phi K} & \simeq \begin{cases} X^{\phi K} \; \text{ if } K \notin \cF, \\
0 \; \text{ if } K \in \cF.
\end{cases}
\end{align*}
Thus, if $X$ is $\cF^{-1}$-local so that $X \simeq X \otimes \widetilde{E \cF}$, then $X^{\phi K} \simeq 0$ for all $K \in \cF$. Conversely, if $X^{\phi K} \simeq 0$ for all $K \in \cF$, then $(X \otimes E \cF_+)^{\phi K} \simeq 0$ for all subgroups $K$, so by the joint conservativity of the functors $\phi^K$, $X \otimes E \cF_+ \simeq 0$ and $X$ is $\cF^{-1}$-local. This proves (1), and the proof of (2) is similar.
\end{proof}
\begin{rem}
The identification of $\cF^{-1}$-local objects in $\Sp^G$ above shows that if $N \leq G$ is a normal subgroup such that $N \subseteq H$ for each $H \notin \Gamma_N$, then there is an equivalence of $\infty$-categories
$$\Sp^{\Phi \Gamma_N} \simeq \Sp^{G/N}.$$
\end{rem}
\begin{rem}[Extension to $G$-recollement] \label{ParamRecollementFamily} Suppose $\cF$ is a $G$-family, and let $\cF^H \subset \fS[H]$ denote the $H$-family obtained by intersecting $\cF$ with $\fS[H] \subset \fS[G]$. For any map of $G$-orbits $f: G/H \to G/K$ with associated adjunction $\adjunct{f^{\ast}}{\Sp^K}{\Sp^H}{f_{\ast}}$, note that
\[ f^{\ast}(E \cF^K_+ \to S^0 \to \widetilde{E \cF^K}) \simeq E \cF^H_+ \to S^0 \to \widetilde{E \cF^H}. \]
By monoidality of $f^{\ast}$, it follows that $f^{\ast}$ preserves $\cF$-torsion and $\cF^{-1}$-local objects. Furthermore, the projection formula implies that
\[ f^{\ast} F(E \cF^K_+,X) \simeq F(E \cF^H_+, f^{\ast} X), \]
so $f^{\ast}$ preserves $\cF$-complete objects. Therefore, $\cF$ defines a lift of the functor $\SH: \sO_G^{\op} \to \Cat^{\st}_{\infty}$ to $\Recoll^{\stab}_{\str}$ (\cite[Def.~2.15]{ShahRecoll}), the $\infty$-category of stable recollements and strict morphisms thereof. Passing to Grothendieck constructions, let
\[ \begin{tikzcd}[row sep=4ex, column sep=4ex, text height=1.5ex, text depth=0.25ex]
\underline{\Sp}^{h \cF} \ar[shift right=1,right hook->]{r}[swap]{j_{\ast}} & \underline{\Sp}^G \ar[shift right=2]{l}[swap]{j^{\ast}} \ar[shift left=2]{r}{i^{\ast}} & \underline{\Sp}^{\Phi \cF} \ar[shift left=1,left hook->]{l}{i_{\ast}}
\end{tikzcd} \]
denote the resulting diagram of $G$-adjunctions. By \cite[Cor.~2.42]{ShahRecoll}, $\underline{\Sp}^{h \cF}$ and $\underline{\Sp}^{\Phi \cF}$ are $G$-stable $G$-$\infty$-categories and all $G$-functors in the diagram are $G$-exact. We thereby obtain a \emph{$G$-stable $G$-recollement} $(\ul{\Sp}^{h \cF}, \ul{\Sp}^{\Phi \cF})$ of $\ul{\Sp}^G$ in the sense of \cite[Def.~2.43]{ShahRecoll}. We also write $\underline{\Sp}^{\tau \cF} \subset \underline{\Sp}^G$ for the essential image of $j_!$.
\end{rem}
We may also consider $\cF$-recollements of the $\infty$-category of $G$-spaces (indeed, of any $\infty$-category of $\cE$-valued presheaves on $\sO_G$).
\begin{ntn} Given a $G$-family $\cF$, let $\sO_{G,\cF} \subset \sO_G$ be the full subcategory on those orbits with stabilizer in $\cF$, and let $\sO_{G,\cF}^c$ be its complement.
\end{ntn}
\begin{cnstr}[$\cF$-recollement of $G$-spaces] \label{cnstr:SpacesRecollement} Given a $G$-family $\cF$, we may define a functor $\pi: \sO_G^{\op} \to \Delta^1$ such that $(\sO_G^{\op})_1 = (\sO_{G,\cF})^{\op}$ and $(\sO_G^{\op})_0 = (\sO_{G,\cF}^c)^{\op}$. Let $\Spc^{h \cF} = \Fun((\sO_{G,\cF})^{\op}, \Spc)$ and $\Spc^{\Phi \cF} = \Fun((\sO^c_{G,\cF})^{\op}, \Spc)$. By \cite[Exm.~3.6]{ShahRecoll}, we obtain a symmetric monoidal recollement with respect to the cartesian product on $G$-spaces
\[ \begin{tikzcd}[row sep=4ex, column sep=4ex, text height=1.5ex, text depth=0.25ex]
\Spc^{h \cF} \ar[shift right=1,right hook->]{r}[swap]{j_{\ast}} & \Spc^G \ar[shift right=2]{l}[swap]{j^{\ast}} \ar[shift left=2]{r}{i^{\ast}} & \Spc^{\Phi \cF} \ar[shift left=1,left hook->]{l}{i_{\ast}}.
\end{tikzcd} \]
Moreover, if we instead take presheaves in $\Spc_{\ast}$, we obtain a symmetric monoidal recollement with respect to the smash product of pointed $G$-spaces
\[ \begin{tikzcd}[row sep=4ex, column sep=4ex, text height=1.5ex, text depth=0.25ex]
\Spc^{h \cF}_{\ast} \ar[shift right=1,right hook->]{r}[swap]{j_{\ast}} & \Spc^G_{\ast} \ar[shift right=2]{l}[swap]{j^{\ast}} \ar[shift left=2]{r}{i^{\ast}} & \Spc^{\Phi \cF}_{\ast} \ar[shift left=1,left hook->]{l}{i_{\ast}}.
\end{tikzcd} \]
where $\widetilde{E \cF} \simeq i_{\ast} i^{\ast}(S^0)$ and the unit map exhibits $\widetilde{E \cF}$ as the same idempotent object as above.
Given a map $f: X \to Y$ in $\Spc^G$, by definition $f$ is a $j^{\ast}$-equivalence if and only if $X^K \to Y^K$ is an equivalence for all $K \in \cF$, and $f$ is a $i^{\ast}$-equivalence if and only if $X^K \to Y^K$ is an equivalence for all $K \notin \cF$. Therefore, by \cref{lem:GeometricFixedPointsDetectionCriterion} and the compatibility of geometric fixed points with $\Sigma^{\infty}_+$, the functor $\Sigma^{\infty}_+$ is a morphism of recollements $(\Spc^{h \cF}, \Spc^{\Phi \cF}) \to (\Sp^{h \cF}, \Sp^{\Phi \cF})$, and likewise for $\Sigma^{\infty}$. In particular, we get induced functors
\[ \Sigma^{\infty}_+: \Spc^{h \cF} \to \Sp^{h \cF}, \quad \Sigma^{\infty}_+: \Spc^{\Phi \cF} \to \Sp^{\Phi \cF}. \]
On the other hand, $\Omega^{\infty}$ is not a morphism of recollements; indeed, if $X \in \Sp^G$ is $\cF$-torsion, then we may have that $i^{\ast} \Omega^{\infty} X$ is non-trivial, so $\Omega^{\infty}$ does not preserve $i^{\ast}$-equivalences. However, if $f: X \to Y$ is a $j^{\ast}$-equivalence in $\Sp^G$, so that $f^{\phi K}$ is an equivalence for all $K \in \cF$, then $\res^G_H(f)$ is an equivalence for all $H \in \cF$ because the functors $\phi^K$ for $K \leq H$ jointly detect equivalences in $\Sp^H$. Therefore, $\Omega^{\infty} (f)$ is a $j^{\ast}$-equivalence, and the $\Sigma^{\infty}_+ \dashv \Omega^{\infty}$ adjunction induces an adjunction
\[ \adjunct{\Sigma^{\infty}_+}{\Spc^{h \cF}}{\Sp^{h \cF}}{\Omega^{\infty}}. \]
Now suppose $X$ is $\cF^{-1}$-local, so that $X^{\phi K} \simeq 0$ for all $K \in \cF$. Then $\res^G_H X \simeq 0$ for all $H \in \cF$, so $(\Omega^{\infty} X)^H \simeq \ast$ for all $H \in \cF$ and thus $\Omega^{\infty} X$ lies in the essential image of $i_{\ast}$. We thereby obtain an adjunction
\[ \adjunct{\Sigma^{\infty}_+}{\Spc^{\Phi \cF}}{\Sp^{\Phi \cF}}{\Omega^{\infty}}. \]
To summarize the various compatibilities, we have that
\begin{enumerate} \item $j^{\ast} \Sigma^{\infty}_+ \simeq \Sigma^{\infty}_+ j^{\ast}: \Spc^{G} \to \Sp^{h \cF}$ and $j_{\ast} \Omega^{\infty} \simeq \Omega^{\infty} j_{\ast}: \Sp^{h \cF} \to \Spc^{G}$.
\item $j^{\ast} \Omega^{\infty} \simeq \Omega^{\infty} j^{\ast}: \Sp^G \to \Spc^{h \cF}$ and $j_! \Sigma^{\infty}_+ \simeq \Sigma^{\infty} j_!: \Spc^{h \cF} \to \Sp^G$.
\item $i^{\ast} \Sigma^{\infty}_+ \simeq \Sigma^{\infty}_+ i^{\ast}: \Spc^G \to \Sp^{\Phi \cF}$ and $i_{\ast} \Omega^{\infty} \simeq \Omega^{\infty} i_{\ast}: \Sp^{\Phi \cF} \to \Spc^{G}$.
\end{enumerate}
\end{cnstr}
Next, we study situations that arise in the presence of two $G$-families.
\begin{rem} \label{rem:FamilyIntersectionFormula} Let $\cF$ and $\cG$ be two $G$-families. Then their intersection $\cF \cap \cG$ is again a $G$-family. Note that $E(\cF \cap \cG) \simeq E \cF \times E \cG$ as $G$-spaces, so $E \cF_+ \otimes E \cG_+ \simeq E(\cF \cap \cG)_+$. Consequently, for any $X \in \Sp^G$, the $\cG$-fracture square for $F(E \cF_+, X)$ yields a commutative diagram
\[ \begin{tikzcd}[row sep=4ex, column sep=4ex, text height=1.5ex, text depth=0.25ex]
F(E \cF_+, X) \otimes {E \cG}_+ \ar{r} \ar{d}{\simeq} & F(E \cF_+, X) \ar{r} \ar{d} & F(E \cF_+, X) \otimes \widetilde{E \cG} \ar{d} \\
F(E (\cF \cap \cG)_+, X) \otimes {E \cG}_+ \ar{r} & F(E (\cF \cap \cG)_+, X) \ar{r} & F(E (\cF \cap \cG)_+, X) \otimes \widetilde{E \cG}
\end{tikzcd} \]
in which the righthand square is a pullback square.
\end{rem}
\begin{lem} \label{lem:KeyIntersectionPropertyFamilies} Let $\cF$ and $\cG$ be two $G$-families. Then $\Sp^{\Phi \cG} \cap \Sp^{h \cF} = \Sp^{\Phi(\cF \cap \cG)} \cap \Sp^{h \cF}$ and $\Sp^{\Phi \cG} \cap \Sp^{h \cF} = \Sp^{\Phi \cG} \cap \Sp^{h(\cF \cup \cG)}$.
\end{lem}
\begin{proof} We prove the first equality, the proof of the second being similar. If $X$ is $\cG^{-1}$-local, then $X$ is $(\cG \cap \cF)^{-1}$-local by \cref{lm:subfamilyProperties}(2), so we have the forward inclusion. On the other hand, by \cref{rem:FamilyIntersectionFormula}, for any $X \in \Sp^G$ we have that
\[ F(E \cF_+,X) \otimes E \cG_+ \simeq F(E(\cF \cap \cG)_+, X) \otimes E \cG_+. \]
But $F(E(\cF \cap \cG)_+, X) \simeq 0$ if $X$ is $(\cF \cap \cG)^{-1}$-local, and $X \simeq F(E \cF_+,X)$ if $X$ is $\cF$-complete. Thus, if $X$ is both $\cF$-complete and $(\cF \cap \cG)^{-1}$-local, then $X$ is $\cG^{-1}$-local. We thereby deduce the reverse inclusion.
\end{proof}
\begin{lem} \label{lm:subfamilyProperties} Suppose $\cG$ is a subfamily of $\cF$. Then
\begin{enumerate}
\item If $X$ is $\cG$-torsion, then $X$ is $\cF$-torsion.
\item If $X$ is $\cF^{-1}$-local, then $X$ is $\cG^{-1}$-local.
\item If $X$ is $\cG$-complete, then $X$ is $\cF$-complete.
\item If $X$ is $\cG^{-1}$-local, then its $\cF$-completion $F(E \cF_+, X)$ is again $\cG^{-1}$-local.
\item If $X$ is $\cG^{-1}$-local, then its $\cF$-acyclization $X \otimes E \cF_+$ is again $\cG^{-1}$-local.
\item If $X$ is $\cG$-complete and $\cF^{-1}$-local, then $X \simeq 0$.
\end{enumerate}
\end{lem}
\begin{proof} (1) and (2) follow immediately from \cref{lem:GeometricFixedPointsDetectionCriterion}. For (3), to show $X$ is $\cF$-complete, we need to show that for all $\cF^{-1}$-local $Y$, $\Map(Y,X) \simeq \ast$. But by (2), $Y$ is $\cG^{-1}$-local, so this mapping space is contractible since $X$ is $\cG$-complete by assumption. For (4), we need to show that for all $\cG$-torsion $Y$, $\Map(Y,F(E \cF_+,X)) \simeq \ast$. But
\[ \Map(Y,F(E \cF_+,X)) \simeq \Map(Y \otimes E \cF_+,X) \simeq \Map(Y,X) \simeq \ast \]
since $Y \otimes E \cF_+ \simeq Y$ by (1) and the assumption that $X$ is $\cG^{-1}$-local. The proof of (5) is similar: given $\cG^{-1}$-local $X$ and any $\cG$-complete $Y$, we have that $\Map(X \otimes E \cF_+, Y) \simeq \ast$ because $Y$ is also $\cF$-complete by (3), hence $X \otimes E \cF_+$ is $\cG^{-1}$-local. Finally, for (6) note that $X$ is then $\cG^{-1}$-local by (2), hence $X \simeq 0$.
\end{proof}
Supposing still that $\cG$ is a subfamily of $\cF$, by \cref{lm:subfamilyProperties}(1-3), the defining adjunctions of the $\cF$ and $\cG$-recollements on $\Sp^G$ restrict to adjunctions
\[ \adjunct{(i_{\cF})^{\ast}}{\Sp^{\Phi \cG}}{\Sp^{\Phi \cF}}{(i_{\cF})_{\ast}}, \adjunct{(j_{\cG})^{\ast}}{\Sp^{h \cF}}{\Sp^{h \cG}}{(j_{\cG})_{\ast}}, \adjunct{(j'_{\cG})_!}{\Sp^{\tau \cG}}{\Sp^{\tau \cF}}{(j'_{\cG})^{\ast}}. \]
By \cref{lm:subfamilyProperties}(4-5), the $\cF$-completion adjunction restricts to
\[ \adjunct{(j_{\cF})^{\ast}}{\Sp^{\Phi \cG}}{\Sp^{h \cF} \cap \Sp^{\Phi \cG}}{(j_{\cF})_{\ast}} \]
such that $(j_{\cF})^{\ast}$ admits a left adjoint $(j_{\cF})_!$ given by the inclusion of $\cF$-torsion and $\cG^{-1}$-local objects under the equivalence $\Sp^{h \cF} \cap \Sp^{\Phi \cG} \simeq \Sp^{\tau \cF} \cap \Sp^{\Phi \cG}$.
Next, let $(i_{\cG})^{\ast}: \Sp^{h \cF} \to \Sp^{h \cF} \cap \Sp^{\Phi \cG}$ be the composite $\Sp^{h \cF} \subset \Sp^G \xto{i^{\ast}} \Sp^{\Phi \cG} \xto{(j_{\cF})^{\ast}} \Sp^{h \cF} \cap \Sp^{\Phi \cG}$. Then $(i_{\cG})^{\ast}$ is left adjoint to the inclusion $(i_{\cG})_{\ast}$. Likewise, define the left adjoint $(i'_{\cG})^{\ast}$ to the inclusion $(i'_{\cG})_{\ast}: \Sp^{\tau \cF} \cap \Sp^{\Phi \cG} \to \Sp^{\tau \cF}$. Finally, note that $\Sp^{h \cF} \cap \Sp^{\Phi \cG}$ inherits a symmetric monoidal structure from the localization $(j_\cF)^{\ast} \dashv (j_{\cF})_{\ast}$, with respect to which $(i_{\cG})^{\ast}$ is symmetric monoidal. Under the equivalence of \cref{rem:torsionCompleteEquivalence}, this transports to a symmetric monoidal structure on $\Sp^{\tau \cF}$ and $\Sp^{\tau \cF} \cap \Sp^{\Phi \cG}$ for which the adjunction $(i'_{\cG})^{\ast} \dashv (i'_{\cG})_{\ast}$ is symmetric monoidal.
\begin{prp} \label{prp:RecollementsOfRecollements} Let $\cG$ be a subfamily of $\cF$. We have stable symmetric monoidal recollements
\[ \begin{tikzcd}[row sep=4ex, column sep=6ex, text height=1.5ex, text depth=0.5ex]
\Sp^{h \cG} \ar[shift right=1,right hook->]{r}[swap]{(j_{\cG})_{\ast}} & \Sp^{h \cF} \ar[shift right=2]{l}[swap]{(j_{\cG})^{\ast}} \ar[shift left=2]{r}{(i_{\cG})^{\ast}} & \Sp^{h \cF} \cap \Sp^{\Phi \cG} \ar[shift left=1,left hook->]{l}{(i_{\cG})_{\ast}},
\end{tikzcd}
\begin{tikzcd}[row sep=4ex, column sep=6ex, text height=1.5ex, text depth=0.5ex]
\Sp^{\tau \cG} \ar[shift right=1,right hook->]{r}[swap]{(j'_{\cG})_{\ast}} & \Sp^{\tau \cF} \ar[shift right=2]{l}[swap]{(j'_{\cG})^{\ast}} \ar[shift left=2]{r}{(i'_{\cG})^{\ast}} & \Sp^{\tau \cF} \cap \Sp^{\Phi \cG} \ar[shift left=1,left hook->]{l}{(i'_{\cG})_{\ast}},
\end{tikzcd} \],
\[ \begin{tikzcd}[row sep=4ex, column sep=6ex, text height=1.5ex, text depth=0.5ex]
\Sp^{h \cF} \cap \Sp^{\Phi \cG} \ar[shift right=1,right hook->]{r}[swap]{(j_{\cF})_{\ast}} & \Sp^{\Phi \cG} \ar[shift right=2]{l}[swap]{(j_{\cF})^{\ast}} \ar[shift left=2]{r}{(i_{\cF})^{\ast}} & \Sp^{\Phi \cF} \ar[shift left=1,left hook->]{l}{(i_{\cF})_{\ast}}.
\end{tikzcd} \]
Furthermore, the equivalence $\Sp^{h \cF} \xto{\simeq} \Sp^{\tau \cF}$ of \cref{rem:torsionCompleteEquivalence} is an equivalence of recollements under which $(j_{\cG})_!$ is the inclusion of $\cG$-torsion objects into $\cF$-torsion objects.
\end{prp}
\begin{proof} The defining properties of a stable symmetric monoidal recollement follow immediately from the same properties for the $\cF$ and $\cG$ recollements on $\Sp^G$. For the last assertion, the equivalence of $\cF$-complete and $\cF$-torsion objects is implemented by $j_! j^{\ast}$, and as such clearly restricts to equivalences $\Sp^{h \cG} \xto{\simeq} \Sp^{\tau \cG}$ and $\Sp^{h \cF} \cap \Sp^{\Phi \cG} \xto{\simeq} \Sp^{\tau \cF} \cap \Sp^{\Phi \cG}$ compatibly with the adjunctions in view of \cref{lm:subfamilyProperties}(4-5). Finally, the claim about $(j_{\cG})_!$ follows from a diagram chase of the right adjoints.
\end{proof}
\begin{rem}[Compact generation] \label{rem:CompactGenerationIntersection} Given a $G$-family $\cF$, the $\cF^{-1}$-local objects $\{ G/H_+ \otimes \widetilde{E \cF}: H \notin \cF \}$ form a set of compact generators for $\Sp^{\Phi \cF}$ because $\Sp^{\Phi \cF} = \Mod_{\Sp^G}(\widetilde{E \cF})$ and $G/H_+$ is $\cF$-torsion for all $H \in \cF$. Given two $G$-families $\cF$ and $\cG$, the essential image of $(j_\cF)_!$ is the localizing subcategory of $\Sp^{\Phi \cG}$ generated by $\{ G/H_+ \otimes \widetilde{E \cG}: H \notin \cG, H \in \cF \}$.
\end{rem}
\begin{rem} \label{NewRecollementSpaceCompatibility} The conclusions of \cref{prp:RecollementsOfRecollements} are also valid for the $\cF$ and $\cG$ recollements on the $\infty$-category of $G$-spaces. We likewise have the adjunction $\adjunct{\Sigma^{\infty}_+}{\Spc^{h \cF} \cap \Spc^{\Phi \cG} }{\Sp^{h \cF} \cap \Sp^{\Phi \cG}}{\Omega^{\infty}}$ and the same compatibility relations as in \cref{cnstr:SpacesRecollement}.
\end{rem}
\begin{rem} \label{rem:Fracture} Let us relate \cref{prp:RecollementsOfRecollements} to the `canonical fracture' of $G$-spectra studied in \cite[\S 2.4]{AMGR-NaiveApproach}. We say that a full subcategory $C_0 \subset C$ is \emph{convex} if given any $x,z \in C_0$ such that there exists a $2$-simplex $[x \to y \to z] \in C$, then $y \in C_0$. Let $\Conv(\fS)$ denote the poset of convex subcategories of $\fS$ and let $\Loc(\Sp^G)$ denote the poset of reflective subcategories of $\Sp^G$, with the order given by inclusion. Suppose $Q \in \Conv(\fS)$ and write $Q = \cF \setminus \cG$ for some $G$-family $\cF$ and subfamily $\cG$. Then the assignment $$\mathfrak{F}_G: \Conv(\fS) \to \Loc(\Sp^G)$$ of \cite[Prop.~2.69]{AMGR-NaiveApproach} sends $Q$ to $\Sp^{h \cF} \cap \Sp^{\Phi \cG}$. Indeed, if we let $\cK_{H}$ be the localizing subcategory of $\Sp^G$ generated by $G/H_+$ and examine \cite[Notn.~2.54]{AMGR-NaiveApproach}, we see that $\cK_{\leq Q} \simeq \Sp^{h \cF}$ and $\cK_{< Q} \simeq \Sp^{h \cG}$ under the equivalence between torsion and complete objects. Thus, $\Sp^G_Q$ defined as the presentable quotient of $\cK_{< Q} \to \cK_{\leq Q}$ is equivalent to $\Sp^{h \cF} \cap \Sp^{\Phi \cG}$ in view of \cref{prp:RecollementsOfRecollements}. Moreover, by inspection the functor $\rho: \Sp^G_Q \xto{\nu} \cK_{\leq Q} \xto{i_R} \Sp^G$ in \cite[Notn.~2.54]{AMGR-NaiveApproach} exhibiting $\Sp^G_Q$ as a reflective subcategory embeds $\Sp^G_Q$ as $\cF$-complete and $\cG^{-1}$-local objects.
By \cite[Prop.~2.69]{AMGR-NaiveApproach} the functor $\mathfrak{F}_G: \Conv(\fS) \to \Loc(\Sp^G)$ is a \emph{fracture} in the sense of \cite[Def.~2.32]{AMGR-NaiveApproach}. Thus, for any convex subcategory $Q = \cF \setminus \cG$ and sieve-cosieve decomposition of $Q$ into $Q_0 = \cF_0 \setminus \cG_0$ and $Q_1 = \cF_1 \setminus \cG_1$, we obtain a recollement $(\Sp^{h \cF_0} \cap \Sp^{\Phi \cG_0}, \Sp^{h \cF_1} \cap \Sp^{\Phi \cG_1})$ of $\Sp^{h \cF} \cap \Sp^{\Phi \cG}$. It is easily seen that these specialize to those considered in \cref{prp:RecollementsOfRecollements} in the case where $Q$ is itself a sieve or a cosieve.
\end{rem}
\begin{ntn} \label{ntn:locallyClosedFibers} Given a subgroup $H$ of $G$, let $\overline{H} = \fS_{\leq H}$ and $\partial \overline{H} = \fS_{<H}$ denote the $G$-family of subgroups that are subconjugate to $H$ and properly subconjugate to $H$, respectively.\footnote{This notation is consistent with viewing sieves as closed sets and cosieves as open sets for a topology on $\fS$.} Let $\fS^c_{\geq H}$ denote the $G$-family of subgroups $K$ such that $H$ is \emph{not} subconjugate to $K$.
\end{ntn}
\begin{lem} \label{lem:VanishingOutsideCone} Suppose $X \in \Sp^G$ is $\overline{H}$-complete and $(\partial \overline{H})^{-1}$-local. Then $X$ is in addition $(\fS^c_{\geq H})^{-1}$-local, i.e., for all subgroups $K$ such that $H$ is not subconjugate to $K$, $X^{\phi K} \simeq 0$.
\end{lem}
\begin{proof} Note that $\partial \overline{H} = \overline{H} \cap \fS^c_{\geq H}$ and use \cref{lem:KeyIntersectionPropertyFamilies}.
\end{proof}
The following two lemmas are explained in \cite[Obs.~2.11-14]{AMGR-NaiveApproach}) (and the first one also in \cite[Prop.~II.2.14]{NS18}), so we will omit their proofs.
\begin{lem} \label{lem:ClosedPartRecollementNormalSubgroup} Let $N$ be a normal subgroup of $G$. Then the geometric fixed points functor $\Phi^N: \Sp^G \to \Sp^{G/N}$ has fully faithful right adjoint with essential image $\Sp^{\Phi \fS^c_{\geq H}}$. Consequently, $\Sp^{G/N}$ is equivalent to the smashing localization $\Mod_{\Sp^G}(\widetilde{E \fS^c_{\geq N}})$.
\end{lem}
\begin{lem} \label{lem:LocallyClosedFibersAreBorel} The geometric fixed points functor $\phi^H: \Sp^G \to \Fun(B W_G H, \Sp)$ has fully faithful right adjoint with essential image $\Sp^{h \overline{H}} \cap \Sp^{\Phi (\partial \overline{H})} =\Sp^{h \overline{H}} \cap \Sp^{\Phi \fS^c_{\geq H}}$.
\end{lem}
\subsection{Reconstruction from geometric fixed points}
\label{section:Reconstruction}
We next aim to state the reconstruction theorem \cite[Thm.~A]{AMGR-NaiveApproach} of Ayala, Mazel-Gee, and Rozenblyum. For this, we need a few preliminary notions.
\begin{dfn} \label{dfn:GeometricLocus} The $G$-\emph{geometric locus} $$\Sp^G_{\locus{\phi}} \subset \Sp^G \times \fS[G]$$ is the full subcategory on objects $(X,H)$ such that $X \in \Sp^{h \overline{H}} \cap \Sp^{\Phi (\partial \overline{H})}$, i.e., $X$ is $\overline{H}$-complete and $(\partial \overline{H})^{-1}$-local (\cref{ntn:locallyClosedFibers}).
\end{dfn}
\begin{dfn}\label{dfn:GeneralizedTate} Given $H$ subconjugate to $K$, the \emph{generalized Tate construction}
\[ \tau^K_H: \Fun(B W_G H, \Sp) \to \Fun(B W_G K, \Sp) \]
is the functor given by the composition
\[ \Fun(B W_G H, \Sp) \to \Sp^G \xto{\phi^K} \Fun(B W_G K, \Sp) \]
where the first functor is the right adjoint to $\phi^H$. If $H=1$, then we will write $\tau^K \coloneq \tau^K_1$.
\end{dfn}
\begin{rem} \label{rem:GenTateResCompatibility} Evidently, the generalized Tate functors $\tau^K_H$ inherit some compatibility properties from the geometric fixed points functors. For example, for $H$ a subgroup of $K$ in $G$, the commutative diagrams
\[ \begin{tikzcd}[row sep=4ex, column sep=4ex, text height=1.5ex, text depth=0.25ex]
\Sp^K \ar{r}{\Phi^H} \ar{d}[swap]{\ind} & \Sp^{W_K H} \ar{d}[swap]{\ind} \ar{r} & \Fun(B W_K H, \Sp) \ar{d}[swap]{\ind} \\
\Sp^G \ar{r}{\Phi^H} & \Sp^{W_G H} \ar{r} & \Fun(B W_G H, \Sp)
\end{tikzcd},
\begin{tikzcd}[row sep=4ex, column sep=4ex, text height=1.5ex, text depth=0.25ex]
\Sp^G \ar{r}{\Phi^K} \ar{d}{\res} & \Sp^{W_K H} \ar{d}{\res} \\
\Sp^K \ar{r}{\Phi^K} & \Sp
\end{tikzcd}
\]
imply that the diagram of generalized Tate functors defined relative to $G$ and $K$
\[ \begin{tikzcd}[row sep=4ex, column sep=4ex, text height=1.5ex, text depth=0.25ex]
\Sp^{h W_G H} \ar{r}{\tau^K_H} \ar{d}{\res} & \Sp^{h W_G K} \ar{d}{\res} \\
\Sp^{h W_K H} \ar{r}{\tau^K_H} & \Sp
\end{tikzcd} \]
commutes. The notation is therefore unambiguous (or abusive) in the same sense as that for geometric fixed points.
Also, if $N_G K = N_G H$, then the composite (with the first functor right adjoint to $\Phi^H$)
\[ \Sp^{W_G H} \to \Sp^G \xto{\Phi^K} \Sp^{W_G K} \]
is homotopic to $\Phi^{K/H}$, and thus $\tau^K_H \simeq \tau^{K/H}$ for $K/H$ regarded as a normal subgroup of $W_G H$.
Finally, note that if $G = C_p$ is a cyclic group of prime order, then $\tau^{C_p} \simeq t^{C_p}$ is the ordinary Tate construction, but not generally otherwise.
\end{rem}
\begin{lem} The structure map $p: \Sp^G_{\locus{\phi}} \to \fS$ is a locally cocartesian fibration such that the functors $\tau^K_H$ are the pushforward functors encoded by $p$ under the equivalence of \cref{lem:LocallyClosedFibersAreBorel}.
\end{lem}
\begin{proof} This is \cite[Constr.~2.38]{AMGR-NaiveApproach} applied to the fracture $\mathfrak{F}_G$ of \cref{rem:Fracture}. To spell out a few more details, we need to show that for every edge $e: \Delta^1 \to \fS$ given by $H$ subconjugate to $K$, the pullback $p|_{e}$ of $p$ over $\Delta^1$ is a cocartesian fibration. Let $C' \subset \Sp^G \times \Delta^1$ be the full subcategory on objects $\{(X,i) \}$ where if $i=0$, then $X \in (\Sp^G_{\locus{\phi}})_H$. Then we have a factorization
\[ \Sp^G_{\locus{\phi}} \times_{\fS, e} \Delta^1 \xto{i''} C' \xto{i'} \Sp^G \times \Delta^1. \]
Note that $C' \to \Delta^1$ is a sub-cocartesian fibration of $\Sp^G \times \Delta^1$ via $i'$ (with cocartesian edges exactly those sent to equivalences via the projection to $\Sp^G$). As for the fiber over $1$, by definition we have that $(\Sp^G_{\locus{\phi}})_K$ is a localization of $\Sp^G$. By an elementary lifting argument, this extends to a localization functor $L: C' \to C'$ whose essential image is $\Sp^G_{\locus{\phi}} \times_{\fS, e} \Delta^1$. By \cite[Lem.~2.2.1.11]{HA}, we deduce that $p|_{e}$ is a cocartesian fibration.
\end{proof}
Recall the barycentric subdivision construction (\cite[Def.~3.19]{ShahRecoll} and \cite[Obs.~3.20]{ShahRecoll}). Unwinding that definition in our situation of interest for a preordered set, we see that $\sd(\fS)$ is the category whose objects are strings $\kappa = [H_0 < H_1 < \cdots < H_n]$ in $\fS$ with each $H_i$ properly subconjugate to $H_{i+1}$, and where a morphism
\[ \kappa = [H_0 < H_1 < \cdots < H_n] \to \lambda = [K_0 < K_1 < \cdots < K_m] \]
is the data of an injective map $\alpha: [n] \to [m]$ of totally ordered sets and a commutative diagram in $\fS$
\[ \begin{tikzcd}[row sep=4ex, column sep=4ex, text height=1.5ex, text depth=0.25ex]
H_{0} \ar{r} \ar{d} & H_{1} \ar{r} \ar{d} & \cdots \ar{r} & H_n \ar{d} \\
K_{\alpha(0)} \ar{r} & K_{\alpha(1)} \ar{r} & \cdots \ar{r} & K_{\alpha(n)}
\end{tikzcd} \]
whose vertical morphisms are equivalences. Note that if a morphism $\kappa \to \lambda$ exists, then $\alpha$ and the commutative ladder are uniquely determined. Thus, the morphism sets in $\sd(\fS)$ are either empty or singleton and $\sd(\fS)$ is also a preordered set. Regard $\sd(\fS)$ as a locally cocartesian fibration over $\fS$ via the functor which takes a string to its maximum element (\cite[Constr.~3.21]{ShahRecoll}).
\begin{rem} Given any locally cocartesian fibration $p: C \to \fS$ whose fibers $C_H$ are stable $\infty$-categories and whose pushforward functors are exact, the right-lax limit $\Fun^{\cocart}_{/\fS}(\sd(\fS),C)$ is a stable $\infty$-category by \cite[Lem.~3.34]{ShahRecoll}. Moreover, if the fibers are presentable and the pushforward functors are also accessible, then the right-lax limit is presentable by \cite[Prop.~3.38]{ShahRecoll}.
\end{rem}
We may now state \cite[Thm.~A]{AMGR-NaiveApproach}, rewritten in our notation.
\begin{thm} \label{thm:AMGRoriginal} There is a canonical equivalence $\Sp^G \simeq \Fun^{\cocart}_{/\fS}(\sd(\fS),\Sp^G_{\locus{\phi}})$.
\end{thm}
Examining the proof of \cite[Thm.~2.40]{AMGR-NaiveApproach}, we see that this equivalence is implemented by the \emph{right-lax} functor $\Sp^G \times \fS \dashrightarrow \Sp^G_{\locus{\phi}}$ that globalizes the left adjoints $\phi^H$. This is not expressible as a functor $\Sp^G \times \fS \to \Sp^G_{\locus{\phi}}$; rather, its construction derives from an existence and uniqueness theorem on adjunctions in $(\infty,2)$-categories (\cite[Lem.~1.34]{AMGR-NaiveApproach} and \cite[Cor.~3.1.7]{gaitsgory2017study}). However, by instead working with the defining inclusion $\Sp^G_{\locus{\phi}} \subset \Sp^G \times \fS$, we can avoid any explicit usage of $(\infty,2)$-category theory and still define a comparison functor, as in the following construction.
\begin{cnstr} \label{cnstr:ComparisonFunctorFromRightLaxLimitToGSpectra} Let $\cF$ be a $G$-family, $\cG$ a subfamily, and $\cH = \cF \setminus \cG$. Consider the composite functor
\[ \Theta'_{\cH}: \Fun^{\cocart}_{/\cH}(\sd(\cH), \cH \times_{\fS} \Sp^G_{\locus{\phi}}) \to \Fun(\sd(\cH), \Sp^G) \xto{\lim} \Sp^G \]
where the first functor is postcomposition by the projection to $\Sp^G$ and the second takes the limit. Note that by \cref{lm:subfamilyProperties}, if $X \in (\Sp^G_{\locus{\phi}})_H$ for any $H \in \cF \setminus \cG$, then $X \in \Sp^{h \cF} \cap \Sp^{\Phi \cG}$. Therefore, $\Theta'_{\cH}$ factors through the inclusion $\Sp^{h \cF} \cap \Sp^{\Phi \cG} \subset \Sp^G$. Denote that functor by $\Theta_{\cH}$.
In the case of $\cF = \fS$ and $\cG = \emptyset$, we also write $\Theta$ for the comparison functor.
\end{cnstr}
\begin{lem} \label{lm:GeometricFixedPointsOfComparisonFunctor} Let $\cF$ be a $G$-family, $\cG$ a subfamily, and $\cH = \cF \setminus \cG$. For every $H \in \cH$, the composition
\[ \Fun^{\cocart}_{/\cH}(\sd(\cH), \cH \times_{\fS} \Sp^G_{\locus{\phi}}) \xto{\Theta'_{\cH}} \Sp^G \xto{\phi^H} \Fun(B W_G H, \Sp) \]
is homotopic to evaluation at $H \in \sd(\cH)$ under the equivalence $(\Sp^G_{\locus{\phi}})_H \simeq \Fun(B W_G H, \Sp)$.
\end{lem}
\begin{proof} Let $f: \sd(\cH) \to \Sp^G_{\locus{\phi}}$ be an object in $\Fun^{\cocart}_{/\cH}(\sd(\cH), \cH \times_{\fS} \Sp^G_{\locus{\phi}})$, and let $f': \sd(\cH) \to \Sp^G$ denote the subsequent functor obtained by the projection to $\Sp^G$. We need to produce a natural equivalence
$$\phi^H \lim f' \simeq f'(H).$$
Since $\sd(\cH)$ is finite, it suffices instead to show $\lim \phi^H f' \simeq f'(H)$. Note that for any $X \in (\Sp^G_{\locus{\phi}})_K$, if $K$ is not in $\overline{H}$ then $X^{\phi H} \simeq 0$; indeed, $\Phi^L(X) \simeq 0$ for all $L \in \fS_{\geq K}^c$ by definition. Therefore, if we let $J \subset \sd(\cH)$ be the full subcategory on those strings $\sigma$ with $\max(\sigma) \leq H$, the functor $\phi^H f'$ is a right Kan extension of its restriction to $J$ (for this, also note that if $\tau = [K_0 < \cdots < K_n] \in \sd(\cH)$ with $K_n \notin \overline{H}$, then $\sd(\cH)_{\tau/} \times_{\sd(\cH)} J = \emptyset$).
Next, let $I \subset J$ be the full subcategory on those strings $\sigma$ with $\max(\sigma)$ conjugate to $H$. For a string $\tau = [K_0 < ... < K_n] \in J$ with $K_n$ properly subconjugate to $H$, the unique string inclusion
$$e: \into{[K_0 < ... < K_n]}{[K_0 < ... < K_n < H]}$$
is sent to an equivalence by $\phi^H f'$ by definition of the locally cocartesian edges in $\Sp^G_{\locus{\phi}}$; indeed, $f'(e)$ is a unit map of the localization for the reflective subcategory $(\Sp^G_{\locus{\phi}})_H \subset \Sp^G$. Observe also that $e$ is an initial object in $I \times_J J^{\tau/}$. We deduce that $\phi^H f'$ is a right Kan extension of its further restriction to $I$. Because $H$ is an initial object of $I$, we conclude that $\lim \phi^H f' \simeq f'(H)$, as desired.
\end{proof}
For the next proposition, recall from \cite[Thm.~3.35]{ShahRecoll} the recollement of a right-lax limit defined by a sieve-cosieve decomposition of the base.
\begin{prp} \label{prp:ComparisonFunctorIsStrictMorphismOfRecollements} Let $\cF$ be a $G$-family, $\cG$ a subfamily, and $\cH = \cF \setminus \cG$. The functor
\[ \Theta_{\cF}: \Fun^{\cocart}_{/\cF}(\sd(\cF), \cF \times_{\fS} \Sp^G_{\locus{\phi}}) \to \Sp^{h \cF} \]
is a strict morphism\footnote{A \emph{strict morphism of recollements} is one that commutes with the gluing functors; cf. \cite[Def.~2.6]{ShahRecoll}.} of stable recollements
\[ (\Fun^{\cocart}_{/\cG}(\sd(\cG), \cG \times_{\fS} \Sp^G_{\locus{\phi}}), \Fun^{\cocart}_{/\cH}(\sd(\cH), \cH \times_{\fS} \Sp^G_{\locus{\phi}})) \to (\Sp^{h \cG}, \Sp^{h \cF} \cap \Sp^{\Phi \cG}). \]
Moreover, the resulting functors between the open and closed parts are equivalent to $\Theta_{\cG}$ and $\Theta_{\cH}$.
\end{prp}
\begin{proof} We need to show that $\Theta_{\cF}$ sends the essential images of $j_!$, $j_{\ast}$, and $i_{\ast}$ to $\cG$-torsion\footnote{More precisely, $\cG$-torsion with respect to the embedding of $\Sp^{h \cF}$ in $\Sp^G$ as $\cF$-torsion objects.}, $\cG$-complete, and $\cG^{-1}$-local objects, respectively. Let
$$f: \sd(\cF) \to \cF \times_{\fS} \Sp^G_{\locus{\phi}}$$
be a functor that preserves locally cocartesian edges. By \cite[Prop.~3.32]{ShahRecoll}, if $f$ is in the essential image of $j_!$, then $f(H) = 0$ for all $H \in \cH$. By \cref{lm:GeometricFixedPointsOfComparisonFunctor}, we then have $\phi^H \Theta_{\cF}(f) \simeq 0$ for all $H \in \cH$, so $\Theta_{\cF}(f)$ is $\cG$-torsion. Similarly, using \cite[Prop.~3.33]{ShahRecoll} and \cref{lm:GeometricFixedPointsOfComparisonFunctor} again, the same proof shows that if $f$ is in the essential image of $i_{\ast}$, then $\phi^H \Theta_{\cF}(f) \simeq 0$ for all $H \in \cG$ and thus $\Theta_{\cF}(f)$ is $\cG^{-1}$-local. Finally, suppose that $f$ is in the essential image of $j_{\ast}$. By \cite[Thm.~3.29]{ShahRecoll}, $f$ is a relative right Kan extension of its restriction to the subcategory $\sd(\cF)_0$ of strings whose minimums lie in $\cG$. Because the inclusion $(\Sp^G_{\locus{\phi}})_H \subset \Sp^G$ of each fiber preserves limits, the further composition
$$f': \sd(\cF) \xto{f} \Sp^G_{\locus{\phi}} \to \Sp^G$$
is then a right Kan extension of its restriction to $\sd(\cF)_0$ (in the non-relative sense). Moreover, the inclusion $\sd(\cG) \subset \sd(\cF)_0$ is right cofinal. Indeed, for every string $\sigma = [K_0 < ... < K_n]$ in $\sd(\cF)_0$, if we let $\sigma'$ denote its maximal substring in $\sd(\cG)$, then $\sigma'$ is a terminal object in $(\sd(\cF)_0)^{/\sigma} \times_{\sd(\cF)_0} \sd(\cG)$, so these slice categories are weakly contractible and we may thus apply Joyal's version of Quillen's Theorem A \cite[Thm.~4.1.3.1]{HTT}. It follows that $\Theta_{\cF}(f)$ is computed as a limit of $\cG$-complete spectra and is hence itself $\cG$-complete.
The two functors on the open and closed parts induced by the morphism of stable recollements are then definitionally $(j_{\cG})^{\ast} \Theta_{\cF} j_{\ast}$ and $(i_{\cG})^{\ast} \Theta_{\cF} i_{\ast}$. These are equivalent to $\Theta_{\cG}$ and $\Theta_{\cH}$ by the same cofinality arguments as above.
\end{proof}
\begin{thm} \label{thm:GeometricFixedPointsDescriptionOfGSpectra} For every $G$-family $\cF$ and subfamily $\cG$, the functor $\Theta_{\cF \setminus \cG}$ is an equivalence of $\infty$-categories. In particular, we have an equivalence
\[ \Theta: \Fun^{\cocart}_{/\fS}(\sd(\fS), \Sp^G_{\locus{\phi}}) \xto{\simeq} \Sp^G. \]
\end{thm}
\begin{proof} Our strategy is to use \cref{prp:ComparisonFunctorIsStrictMorphismOfRecollements} in conjunction with the fact that given a strict morphism $F: \cX \to \cX'$ of stable recollements $(\cU, \cZ) \to (\cU', \cZ')$, if $F_{U}$ and $F_{Z}$ are equivalences then $F$ is an equivalence (\cite[Rem.~2.7]{ShahRecoll}).\footnote{This type of inductive argument is also used in the proof of \cite[Thm.~2.40]{AMGR-NaiveApproach}.} Let us first prove that $\Theta_{\cF}$ is an equivalence for all families $\cF$. We proceed by induction on the size of $\cF$. For the base case, if $\cF = \{ 1 \}$ is the trivial family, then $\sd(\cF) \cong \cF$ and $\Theta_{\cF}$ is definitionally an equivalence. Now suppose for the inductive hypothesis that $\Theta_{\cG}$ is an equivalence for all proper subfamilies $\cG$ of $\cF$. Let $H \in \cF$ be a maximal element and let $\cG \subset \cF$ be the largest subfamily excluding $H$. Then $\cF \setminus \cG = \overline{H} \setminus \partial \overline{H}$, so $\Theta_{\cF \setminus \cG}$ is definitionally an equivalence. By \cref{prp:ComparisonFunctorIsStrictMorphismOfRecollements}, we deduce that $\Theta_{\cF}$ is an equivalence.
Finally, to deal with the general case, we note that any strict morphism of stable recollements that is also an equivalence restricts to equivalences between the open and closed parts. Thus, having proven that $\Theta_{\cF}$ is an equivalence, we further deduce that $\Theta_{\cF \setminus \cG}$ is an equivalence for any subfamily $\cG$.
\end{proof}
\begin{rem} The generalized Tate functors $\tau^K_H$ are lax symmetric monoidal, and the various natural transformations among these functors encoded by the locally cocartesian fibration are also lax symmetric monoidal. In \cite[\S 5]{AMGRb}, Ayala--Mazel-Gee--Rozenblyum explain how this data assembles to a symmetric monoidal structure on the right-lax limit $\Fun^{\cocart}_{/\fS}(\sd(\fS),\Sp^G_{\locus{\phi}})$ such that the functor $\Theta$ of \cref{thm:GeometricFixedPointsDescriptionOfGSpectra} is an equivalence of symmetric monoidal $\infty$-categories.
\end{rem}
\cref{thm:GeometricFixedPointsDescriptionOfGSpectra} and \cref{prp:ComparisonFunctorIsStrictMorphismOfRecollements}, along with the explicit description of the functor $j_{\ast}$ given in \cite[Thm.~3.29]{ShahRecoll}, gives the following formula for the geometric fixed points of an $\cF$-complete spectrum in terms of a limit of generalized Tate constructions.
\begin{dfn}
For $H \notin \cF$, let $J_H \subset \sd(\fS)$ be the full subcategory on strings $[K_0 < \cdots < K_n < H]$ such that $K_i \in \cF$ for $1 \leq i \leq n$.
\end{dfn}
\begin{cor} \label{cor:FormulaForGeomFixedPointsOfCompleteSpectrum} Let $X$ be a $G$-spectrum and let $X^{\bullet}: \sd(\fS) \to \Sp^G_{\locus{\phi}}$ denote a lift of $X$ under the equivalence $\Theta$. Suppose that $X$ is $\cF$-complete. Then $$X^{\phi H} \simeq \lim_{J_H} X^{\bullet},$$
with the limit taken in the fiber $\Fun(B W_G H, \Sp) \simeq (\Sp^G_{\locus{\phi}})_H$.
\end{cor}
\begin{exm} Suppose that $G = C_{p^2}$ and let $\cP$ be the family of proper subgroups of $G$. Then $\sd(\cP) \cong \sd(\Delta^1)$, so the data of a $\cP$-complete spectrum $X$ amounts to
\begin{itemize}
\item A Borel $C_{p^2}$-spectrum $X^1$.
\item A Borel $C_{p^2}/ C_p$-spectrum $X^{\phi C_p}$.
\item A $C_{p^2}/ C_p \cong C_p$-equivariant map $\alpha: X^{\phi C_p} \to (X^1)^{t C_p}$.
\end{itemize}
The category $J_{C_{p^2}}$ as well as the functor $J_{C_{p^2}} \to \Sp$ is then identified as
\[ \left( \begin{tikzcd}[row sep=4ex, column sep=4ex, text height=1.5ex, text depth=0.25ex]
& \goesto{\left[ C_p < C_{p^2} \right] \ar{d} \\
\left[ 1 < C_{p^2} \right] \ar{r} & \left[ 1 < C_p < C_{p^2} \right]
\end{tikzcd} \right) }{ \left(
\begin{tikzcd}[row sep=4ex, column sep=4ex, text height=1.5ex, text depth=0.25ex]
& X^{\phi C_p t C_p} \ar{d}{\alpha^{t C_p}} \\
(X^1)^{\tau C_{p^2}} \ar{r}{\can} & (X^1)^{t C_p t C_p}
\end{tikzcd} \right) }, \]
where $\mit{can}$ is the canonical map encoded by the locally cocartesian fibration. Thus,
\[ X^{\phi C_{p^2}} \simeq (X^1)^{\tau C_{p^2}} \times_{(X^1)^{t C_p t C_p}} X^{\phi C_p t C_p}. \]
Moreover, it is not difficult to see that $(-)^{\tau C_{p^2}} \simeq (-)^{h C_p tC_p}$; in fact, we will explain the identification $(-)^{\tau C_{p^n}} \simeq (-)^{h C_{p^{n-1}} t C_p}$ in \cite{QS21b}.
\end{exm}
Let us now turn to the examples of interest for the dihedral Tate orbit lemma that we will prove in \cite{QS21b}.
\begin{exm} \label{exm:DihedralEven} Suppose that $G = D_4 = C_2 \times \mu_2$ is the Klein four-group and let $\Gamma = \{ 1, C_2, \Delta \}$ for $\Delta$ the diagonal subgroup. The data of a $\Gamma$-complete spectrum $X$ amounts to
\begin{itemize}
\item A Borel $D_4$-spectrum $X^1$, Borel $(D_4/C_2)$-spectrum $X^{\phi C_2}$, and Borel $(D_4/\Delta)$-spectrum $X^{\phi \Delta}$.
\item A $(D_4/C_2)$-equivariant map $\alpha: X^{\phi C_2} \to (X^1)^{t C_2}$ and $(D_4/ \Delta)$-equivariant map $\beta: X^{\phi \Delta} \to (X^1)^{t \Delta}$.
\end{itemize}
Since $J_{\mu_2} = \{ [1<\mu_2] \}$ and $J_{\mu_2} \to \Fun(B(D_4/\mu_2),\Sp)$ is the pushforward of $X^1$ by $(-)^{t \mu_2}$, we see that $X^{\phi \mu_2} \simeq (X^1)^{t \mu_2}$. On the other hand, $J_{D_4} \to \Sp$ is given by
\[ \left( \begin{tikzcd}[row sep=4ex, column sep=4ex, text height=1.5ex, text depth=0.25ex]
\goesto{\left[ \Delta < D_4 \right] \ar{r} & \left[ 1<\Delta<D_4 \right] \\
\left[ 1 < D_4 \right] \ar{ru} \ar{rd} & \\
\left[ C_2 < D_4 \right] \ar{r} & \left[ 1 < C_2 < D_4 \right]
\end{tikzcd} \right)}{\left(
\begin{tikzcd}[row sep=4ex, column sep=8ex, text height=1.5ex, text depth=0.25ex]
(X^{\phi \Delta})^{t(D_4/\Delta)} \ar{r}{\beta^{t(D_4/\Delta)}} & ((X^1)^{t \Delta})^{t (D_4/ \Delta)} \\
(X^1)^{\tau D_4} \ar{ru}[swap]{\can} \ar{rd}{\can} & \\
(X^{\phi C_2})^{t (D_4/C_2)} \ar{r}[swap]{\alpha^{t(D_4/C_2)}} & ((X^1)^{t C_2})^{t (D_4/C_2)}
\end{tikzcd} \right)}, \]
and $X^{\phi D_4}$ is the limit of this diagram.
\end{exm}
To handle the case of the dihedral group $D_{2p}$ of order $2p$ for $p$ an odd prime, we first record a vanishing property of the generalized Tate construction.
\begin{lem} \label{lem:GenTateVanishingMultiprime} Let $G$ be a finite group and suppose $K \leq G$ is a subgroup that is not a $p$-group. Then $\tau^K \simeq 0$.
\end{lem}
\begin{proof} By the compatibility of the generalized Tate functors with restriction (\cref{rem:GenTateResCompatibility}), we may suppose $K = G$ without loss of generality. Note that $\tau^G$ may be computed as the left Kan extension of $(-)^{h G}$ along the functor from $\Sp^{h G}$ to its Verdier quotient by orbits $\{ G/H_+ : H <G\}$ with $H$ proper \cite[Rem.~2.16]{AMGR-NaiveApproach}. If we let $\ul{All}$ be the family of subgroups $H$ such that $|H|=p^n$ for some prime $p$ and integer $n$ as in \cite[Fig.~1.7]{mathew2019}, then $\ul{All}$ is a subfamily of the proper subgroups under our assumption. However, by \cite[Thm.~4.25]{mathew2019}, the thick $\otimes$-ideal in $\Sp^G$ generated by $\{G/H_+: H \in \ul{All} \}$ includes the Borel completion of the unit. Therefore, the Verdier quotient in question is the trivial category, and we deduce that $\tau^G \simeq 0$.
\end{proof}
\begin{exm} \label{exm:DihedralOdd} Let $p$ be an odd prime, $G = D_{2p} = \mu_p \rtimes C_2$ the dihedral group of order $2p$, and $\Gamma$ the family of subgroups $H$ such that $H \cap \mu_p = 1$. Note that up to conjugacy, $\Gamma$ consists of the subgroups $1$ and $C_2$, and the Weyl group of $C_2$ is trivial. Thus, up to equivalence, the data of a $\Gamma$-complete spectrum $X$ amounts to
\begin{itemize}
\item A Borel $D_{2p}$-spectrum $X^1$ and a spectrum $X^{\phi C_2}$.
\item A map $\alpha: X^{\phi C_2} \to (X^1)^{t C_2}$.
\end{itemize}
Using that $J_{\mu_p} = [1 < \mu_p]$, we compute $X^{\phi \mu_p} \simeq (X^1)^{t \mu_p}$. As for $X^{\phi D_{2p}}$, by \cref{lem:GenTateVanishingMultiprime} we have that $(X^1)^{\tau D_{2p}} \simeq 0$. We further claim that the generalized Tate functor $\tau^{D_{2p}}_{C_2}$ vanishes:
\begin{itemize} \item[($\ast$)] Let $\cH = \{ 1, \mu_p \} = \fS^c_{\geq C_2}$. By \cref{ParamRecollementFamily} applied to $\ul{\Sp}^{\Phi \cH}$, the restriction and induction functors for $C_2 \subset D_{2p}$ descend to an adjunction
\[ \adjunct{\res'}{\Sp^{\Phi \cH}}{\Sp}{\ind'}, \]
where $(\ul{\Sp}^{\Phi \cH})_{D_{2p}/C_2} \simeq \Sp$ because the restriction of $\cH$ to a $C_2$-family yields the trivial family. Now consider the inclusion of the open fiber
\[ j_{\ast}: \Sp^{h W_G C_2} \to \Sp^{\Phi \cH}. \]
Because $W_{D_{2p}} C_2 \cong 1$, we have that $\res' \simeq j^{\ast}$, and we deduce that $j_! \simeq j_{\ast}$. Because $\cH^c = \fS_{\geq C_2}$ consists only of the two subgroups $C_2$ and $D_{2p}$ up to conjugacy, we may identify $\phi^{D_{2p}}: \Sp^{\Phi \cH} \to \Sp$ with the restriction $i^{\ast}$ to the closed complement of a recollement of $\Sp^{\Phi \cH}$ with $j_{\ast}$ as the inclusion of the open part. We then have $\tau^{D_{2p}}_{C_2} \simeq i^{\ast} j_{\ast}$. In view of the fiber sequence $$j_! \xto{\simeq} j_{\ast} \to i_{\ast} i^{\ast} j_{\ast} \simeq 0,$$ we deduce that $\tau^{D_{2p}}_{C_2} \simeq 0$.
\end{itemize}
Using \cref{cor:FormulaForGeomFixedPointsOfCompleteSpectrum}, we conclude that $X^{\phi D_{2p}} \simeq 0$.
\end{exm}
We conclude this section by indicating how the comparison functor $\Theta$ is functorial in the group $G$ with respect to restriction and geometric fixed points.
\begin{cnstr}[Restriction functoriality for geometric loci] \label{restrictionGeometricLoci} Let $H$ be a subgroup of $G$ and consider the map $i: \fS[H] \to \fS[G]$ that sends a subgroup $K$ of $H$ to the same $K$ viewed as a subgroup of $G$. Since $i$ preserves the subconjugacy relation, $i$ is a functor,\footnote{However, since there may be additional conjugacy relations in $G$, $i$ is not generally the inclusion of a subcategory.} and also let $i: \sd(\fS[H]) \to \sd(\fS[G])$ denote the induced functor on barycentric subdivisions. Next, consider the functor $\res^G_H \times \id: \Sp^G \times \fS[H] \to \Sp^H \times \fS[H]$. Since for any subgroup $K \leq H$, the restriction of the $G$-families $\fS[G]_{\leq K}$, $\fS[G]_{<K}$ to $H$ yields $H$-families $\fS[H]_{\leq K}$, $\fS[H]_{<K}$, by \cref{ParamRecollementFamily} we have an induced functor over $\fS[H]$
\[ \res^G_H: \Sp^G_{\locus{\phi}} \times_{\fS[G]} \fS[H] \to \Sp^H_{\locus{\phi}} \]
that preserves locally cocartesian edges. Precomposition by $i$ and postcomposition by $\res^G_H$ then defines a functor
\[ \res^G_H: \Fun^{\cocart}_{/\fS[G]}(\sd(\fS[G]), \Sp^G_{\locus{\phi}}) \to \Fun^{\cocart}_{/\fS[H]}(\sd(\fS[H]), \Sp^H_{\locus{\phi}}). \]
We have a lax commutative diagram
\[ \begin{tikzcd}[row sep=4ex, column sep=6ex, text height=1.5ex, text depth=0.5ex]
\Fun^{\cocart}_{/\fS[G]}(\sd(\fS[G]), \Sp^G_{\locus{\phi}}) \ar{d}[swap]{\res^G_H} \ar{r}{\Theta_G}[swap]{\simeq} \ar[phantom]{rd}{\NEarrow} & \Sp^G \ar{d}{\res^G_H} \\
\Fun^{\cocart}_{/\fS[H]}(\sd(\fS[H]), \Sp^H_{\locus{\phi}}) \ar{r}{\Theta_H}[swap]{\simeq} & \Sp^H
\end{tikzcd} \]
where the natural transformation $\eta: \res^G_H \circ \Theta_G \to \Theta_H \circ \res^G_H$ is defined using the contravariant functoriality of the limit for $i: \sd(\fS[H]) \to \sd(\fS[G])$.
We claim that $\eta$ is an equivalence, so that this diagram commutes. Indeed, suppose given
$$f: \sd(\fS[G]) \to \Sp^G_{\locus{\phi}}$$
and let $g = \res^G_H f: \sd(\fS[H]) \to \Sp^H_{\locus{\phi}}$. Let $f': \sd(\fS[G]) \to \Sp^G$ and $g': \sd(\fS[H]) \to \Sp^H$ be the functors obtained by postcomposition, so $g' = \res^G_H f' i$ by definition and $\eta_f$ is the comparison map
\[ \lim_{\sd(\fS[G])} \res^G_H f' \to \lim_{\sd(\fS[H])} \res^G_H f' i. \]
It suffices to check that for all subgroups $K \leq H$, $\phi^K(\eta_f)$ is an equivalence. But then by the commutativity of the diagram
\[ \begin{tikzcd}[row sep=4ex, column sep=4ex, text height=1.5ex, text depth=0.25ex]
\Sp^G \ar{d}[swap]{\res^G_H} \ar{r}{\phi^K} & \Fun(B W_G K, \Sp) \ar{d}{\res^{W_G K}_{W_H K}} \\
\Sp^H \ar{r}{\phi^K} & \Fun(B W_H K, \Sp),
\end{tikzcd} \]
and under the equivalences $\phi^K \Theta_G \simeq \ev_K$ and $\phi^K \Theta_H \simeq \ev_K$ of \cref{lm:GeometricFixedPointsOfComparisonFunctor}, we see that $\phi^K(\eta_f)$ is an equivalence.
\end{cnstr}
\begin{cnstr}[Geometric fixed points functoriality for geometric loci] \label{GeometricFixedPointsGeometricLoci} Let $N$ be a normal subgroup of $G$. Then we may embed $\fS[G/N]$ as a cosieve in $\fS[G]$ via the functor $i: \fS[G/N] \to \fS[G]$ that sends $M/N$ to $M$. We also let $i: \sd(\fS[G/N]) \to \sd(\fS[G])$ denote the induced functor on barycentric subdivisions, which is a cosieve inclusion. By \cref{lem:ClosedPartRecollementNormalSubgroup}, $\Phi^N: \Sp^G \to \Sp^{G/N}$ has fully faithful right adjoint with essential image given by the $(\fS[G] \setminus \fS[G/N])^{-1}$-local objects. Therefore, $\Phi^N$ implements an equivalence over $\fS[G/N]$
\[ \Sp^G_{\locus{\phi}} \times_{\fS[G]} \fS[G/N] \xto{\simeq} \Sp^{G/N}_{\locus{\phi}}. \]
Define $\Phi^N: \Fun^{\cocart}_{/\fS[G]}(\sd(\fS[G]), \Sp^G_{\locus{\phi}}) \to \Fun^{\cocart}_{/\fS[G/N]}(\sd(\fS[G/N]), \Sp^{G/N}_{\locus{\phi}})$ to be the functor obtained by $i^{\ast}$ under that equivalence. Then because $\Theta$ is a morphism of recollements, we have a commutative diagram
\[ \begin{tikzcd}[row sep=4ex, column sep=6ex, text height=1.5ex, text depth=0.5ex]
\Fun^{\cocart}_{/\fS[G]}(\sd(\fS[G]), \Sp^G_{\locus{\phi}}) \ar{d}[swap]{\Phi^N} \ar{r}{\Theta_G}[swap]{\simeq} & \Sp^G \ar{d}{\Phi^N} \\
\Fun^{\cocart}_{/\fS[G/N]}(\sd(\fS[G/N]), \Sp^{G/N}_{\locus{\phi}}) \ar{r}{\Theta_{G/N}}[swap]{\simeq} & \Sp^{G/N}.
\end{tikzcd} \]
\end{cnstr}
\section{Theories of \texorpdfstring{$G$}{G}-spectra relative to a normal subgroup \texorpdfstring{$N$}{N}}\label{Sec:RelThy}
In classical approaches to equivariant stable homotopy theory \cite{MR866482, AlaskaNotes}, one attaches to every $G$-universe $\cU$ a corresponding theory of $G$-spectra indexed with respect to $\cU$; upon inverting the weak equivalences, this yields a stable $\infty$-category $\Sp^G_{\cU}$. For the complete $G$-universe $\cU$, one obtains \emph{genuine} $G$-spectra $\Sp^G_{\cU} \simeq \Sp^G$, whereas for the trivial $G$-universe $\sU^G$, one obtains \emph{naive} $G$-spectra $\Sp^G_{\cU^G} \simeq \Fun(\sO^{\op}_G, \Sp)$.\footnote{We identify the $\infty$-category as the ordinary stabilization of $G$-spaces $\Spc^G = \Fun(\sO_G^{\op}, \Spc)$.} Interpolating between genuine and naive $G$-spectra, for every normal subgroup $N \trianglelefteq G$, one has the fixed points $G$-universe $\cU^N$ \cite[Ch.~XVI, \S 5]{AlaskaNotes} and the associated $\infty$-category $\Sp^G_{\cU^N}$. In this section, we will revisit these notions from a different and intrinsically $\infty$-categorical perspective that makes no reference to representation theory. Using the language of parametrized $\infty$-category theory, we define $\infty$-categories $\Sp^G_{\naive{N}}$ and $\Sp^G_{\Borel{N}}$ of \emph{$N$-naive} and \emph{$N$-Borel} $G$-spectra (\cref{dfn:NaiveGSpectraRelativeToNormalSubgroup} and \cref{dfn:BorelGSpectraRelativeToNormalSubgroup}). We then show $\Sp^G_{\Borel{N}}$ admits two canonical embeddings into $\Sp^G$ as the $\Gamma_N$-torsion and $\Gamma_N$-complete $G$-spectra for $\Gamma_N$ the $N$-free $G$-family (\cref{thm:BorelSpectraAsCompleteObjects}).
\begin{rem} Although we expect the $\infty$-category $\Sp^G_{\naive{N}}$ to be equivalent to $\Sp^G_{\cU^N}$, we will not give a precise comparison in this paper.
\end{rem}
\begin{wrn}
From this section until the end of \cref{section:NormMaps}, the role of $G$ is different from its role in the introduction. In particular, $G$ is always viewed as an extension of $G/N$ by a normal subgroup $N$ throughout this section, whereas $G$ was the quotient $\widehat{G}/K$ in the introduction. This change is made to emphasize the fact that $G$ is finite throughout the paper.
\end{wrn}
To begin with, we will need a lemma on cartesian fibrations arising from adjunctions.
\begin{lem} \label{lem:QuotientMapCartesianFibration} \begin{enumerate}[leftmargin=*] \item Let $\adjunct{L}{C}{D}{R}$ be an adjunction such that for all $c \in C$, $d \in D$, and $f: d \to L c$ the natural map
\[ L( Rd \times_{R L c} c ) \to d \]
adjoint to the projection $Rd \times_{R L c} c \to R d$ is an equivalence. Then $L$ is an essentially cartesian fibration\footnote{An \emph{essentially} cartesian fibration is the version of cartesian fibration that is stable under equivalence, defined to be the obvious generalization of a Street fibration to the $\infty$-categorical context.} with $L$-cartesian edges given by $Rd \times_{RLc} c \to c$, and hence a cartesian fibration if $L$ is assumed to be a categorical fibration.
\item Let
\[ \begin{tikzcd}[row sep=4ex, column sep=4ex, text height=1.5ex, text depth=0.25ex]
L'' = L' \circ L: C \ar[shift left=2]{r}{L} & C' \ar[shift left=1]{l}{R} \ar[shift left=2]{r}{L'} & D : R \circ R' = R'' \ar[shift left=1]{l}{R'}
\end{tikzcd} \]
\end{enumerate}
be a diagram of adjunctions such that $L \dashv R$, $L' \dashv R'$, and $L'' \dashv R''$ all satisfy the assumption in (1). Then $L$ sends $L''$-cartesian edges to $L'$-cartesian edges.
\end{lem}
\begin{proof} For (1), under our assumption, we need only show that $R d \times_{R L c} c \to c$ is a $L$-cartesian edge. But for this, for any $c' \in C$ we have the pullback square of spaces
\[ \begin{tikzcd}[row sep=4ex, column sep=4ex, text height=1.5ex, text depth=0.25ex]
\Map_{C}(c',Rd \times_{RL c} c) \ar{r} \ar{d} & \Map_C(c',Rd) \ar{d} \ar{r}{\simeq} & \Map_D(Lc',d) \ar{d} \\
\Map_C(c',c) \ar{r} & \Map_C(c',RLc) \ar{r}{\simeq} & \Map_D(Lc',Lc),
\end{tikzcd} \]
and the assertion follows from the definition of $L$-cartesian edge and a simple diagram chase.
For (2), let $c \in C$, $(f:d \to L'' c) \in D$, and consider the $L''$-cartesian edge $R'' d \times_{R'' L'' c} c \to c$ (this case suffices since all $L''$-cartesian edges are of this form up to equivalence). Note that the unit map for $L'' \dashv R''$ factors as the composition $R'' L'' c \simeq R R' L' L c \to R L c \to c$ of unit maps for $L' \dashv R'$ and $L \dashv R$. Thus, we have
\[ L(R'' d \times_{R'' L'' c} c) \simeq L(R \left(R' d \times_{R' L'' c} Lc \right) \times_{R L c} c) \xto{\simeq} R' d \times_{R' L'' c} L c \]
by our assumption on $L \dashv R$, and this equivalence respects the projection map to $L(c)$. But our assumption on $L' \dashv R'$ ensures that $R'(d) \times_{R' L''(c)} L(c) \to L(c)$ is a $L'$-cartesian edge.
\end{proof}
\subsection{\texorpdfstring{$N$}{N}-naive \texorpdfstring{$G$}{G}-spectra}\label{SS:NNaive}
\begin{obs} \label{inflationFunctors} Let $N$ be a normal subgroup of $G$ and let $\pi: G \to G/N$ denote the quotient map. We have the adjunction
\[ \adjunct{r_N}{\FF_G}{\FF_{G/N}}{\iota_N} \]
where $r_N(U) = U/N$ and $\iota_N(V) = V$ with $V$ regarded as a $G$-set via $\pi$. Note that $r_N \iota_N (V) = V$, so $\iota_N$ is fully faithful. For $U \in \FF_G$, $V \in \FF_{G/N}$, and a $G/N$-map $f: V \to U/N$, we also have
$$(V \times_{U/N} U)/N \cong V,$$
so $r_N$ is a cartesian fibration by \cref{lem:QuotientMapCartesianFibration}. Note also that the adjunction $r_N \dashv \iota_N$ restricts on orbits to
$$\adjunct{r_N}{\sO_G}{\sO_{G/N}}{\iota_N}$$
and $V \times_{U/N} U$ is transitive if $V$ and $U$ are, hence $r_N$ remains a cartesian fibration when restricted to $\sO_G$. Given a $G$-orbit $G/H$ and a $G/N$-map $f: \frac{G/N}{K/N} \cong G/K \to \frac{G/N}{H N/N} \cong G/ H N$, we may identify the pullback
$$G/H \times_{G/H N} G/K \cong G/(H \cap K),$$
and thus an $r_N$-cartesian edge lifting $f$ is given by $G/(H \cap K) \to G/H$.
\end{obs}
\begin{cvn} For $N$ a normal subgroup of $G$, we will regard $\sO_G^{\op}$ as a $G/N$-category via $r_N^{\op}$.
\end{cvn}
\begin{dfn} \label{dfn:NaiveGSpectraRelativeToNormalSubgroup} Let $\Sp^G_{\naive{N}} := \Fun_{G/N}(\sO_G^{\op}, \underline{\Sp}^{G/N})$ be the $\infty$-category of \emph{naive $G$-spectra relative to $N$}, or \emph{$N$-naive $G$-spectra}.
\end{dfn}
For example, if $N = G$ we have the usual $\infty$-category $\Fun(\sO_G^{\op}, \Sp)$ of naive $G$-spectra, and if $N = 1$ we instead have the $\infty$-category $\Sp^G$ itself.
\begin{cnstr} \label{cnstr:ForgetfulFunctorToNaiveGSpectra} We define a `forgetful' functor
$$\sU[N]: \Sp^G \to \Sp^G_{\naive{N}} = \Fun_{G/N}(\sO_G^{\op}, \underline{\Sp}^{G/N}).$$
First, let
$$q_N: \omega_G \circ \iota_N \Rightarrow \omega_{G/N}: \FF_{G/N} \to \Gpd_{\fin}$$
be the natural transformation defined on objects $U \in \FF_{G/N}$ by the functor $U//G \to U//(G/N)$ which sends objects $x \in U$ to the same $x \in U$ and morphisms $[g: x \rightarrow g \cdot x = \pi(g) \cdot x]$ to $[\pi(g): x \rightarrow \pi(g) \cdot x]$.
Recall the functor $\SH: \Gpd_{\fin}^{\op} \to \CAlg(\Pr^{\mr{L}})$ of \cref{dfn:BachmannHoyoisFunctor}. For any $\infty$-category $C$, the adjunction $r_N \dashv \iota_N$ induces an adjunction
\[ \adjunct{(r_N^{\op})^{\ast}}{\Fun(\FF_{G/N}^{\op},C)}{\Fun(\FF_G^{\op}, C)}{(\iota_N^{\op})^{\ast}} \]
where we may identify $(r_N^{\op})^{\ast}$ with the left Kan extension along $\iota_N^{\op}$. Let
$$\underline{\inf}[N]: \SH \omega_{G/N}^{\op} r_N^{\op} \to \SH \omega_G^{\op}$$
be the natural transformation adjoint to $\SH q_N^{\op}$ and let
\[ \underline{\inf}[N]: \sO_G^{\op} \times_{r_N^{\op},\sO_{G/N}^{\op}} \underline{\Sp}^{G/N} \to \underline{\Sp}^G \]
also denote the associated $G$-functor given by unstraightening. Note that for a $G$-orbit $G/H$, $\underline{\inf}[N]_{G/H}$ is given by the inflation functor $\inf^{H \cap N}: \Sp^{H/H \cap N} \to \Sp^H$. By the dual of \cite[Prop.~7.3.2.6]{HA}, $\underline{\inf}[N]$ admits a relative right adjoint
\[ \widehat{\Psi}[N]: \underline{\Sp}^G \to \sO_G^{\op} \times_{r_N^{\op},\sO_{G/N}^{\op}} \underline{\Sp}^{G/N} \]
over $\sO_G^{\op}$ that does \emph{not} preserve cocartesian edges; rather, for a map of $G$-orbits $f: G/K \to G/H$ we have a lax commutative square
\[
\begin{tikzcd}[row sep=6ex, column sep=6ex, text height=1.5ex, text depth=0.25ex]
\Sp^H \ar{r}{\gamma_{\ast}} \ar{d}{f^{\ast}} \ar[phantom]{rd}{\SWarrow} & \Sp^{H/H \cap N} \ar{d}{f^{\ast}} \\
\Sp^K \ar{r}{\gamma_{\ast}} & \Sp^{K/ K \cap N}
\end{tikzcd}
\]
for the square of $G$-orbits
\[
\begin{tikzcd}[row sep=6ex, column sep=4ex, text height=1.5ex, text depth=0.25ex]
G/K \ar{d}{f} \ar{r}{\gamma} & (G/N)/(K N/ N) \ar{d}{f} \\
G/H \ar{r}{\gamma} & (G/N)/(H N/N).
\end{tikzcd}
\]
However, for a map of $G/N$-orbits $f: \frac{G/N}{K/N} \to \frac{G/N}{H N/N}$ we have a homotopy commutative square
\[ \begin{tikzcd}[row sep=6ex, column sep=6ex, text height=1.5ex, text depth=0.25ex]
\Sp^H \ar{r}{\gamma_{\ast}} \ar{d}{f^{\ast}} & \Sp^{H/(H \cap N)} \ar{d}{f^{\ast}} \\
\Sp^{K \cap H} \ar{r}{\gamma_{\ast}} & \Sp^{(K \cap H)/(H \cap N)}
\end{tikzcd}
\]
for the pullback square
\[
\begin{tikzcd}[row sep=6ex, column sep=4ex, text height=1.5ex, text depth=0.25ex]
G/(K \cap H) \ar{d}{\gamma} \ar{r}{f} & G/H \ar{d}{\gamma} \\
G/K \ar{r}{f} & G/H N
\end{tikzcd}
\]
and hence the further composition $\underline{\Psi}[N] \coloneq \pr \circ \widehat{\Psi}[N]: \underline{\Sp}^G \to \underline{\Sp}^{G/N}$ \emph{does} preserve cocartesian edges over $\sO_{G/N}^{\op}$, where we regard $\underline{\Sp}^G$ as a $G/N$-$\infty$-category via $r_N^{\op}$. We also have the unit $\eta: \id \to \iota_N r_N$ which by precomposition yields the $G$-functor
\[ \underline{\res}[N]: \sO_G^{\op} \times_{(\iota_N r_N)^{\op}, \sO_G^{\op}} \underline{\Sp}^G \to \underline{\Sp}^G, \]
where for a $G$-orbit $G/H$, $\underline{\res}[N]_{G/H}$ is given by the restriction functor $\res^{H N}_H: \Sp^{H N} \to \Sp^H$. The composite
$$\underline{\Psi}[N] \circ \underline{\res}[N]: \sO_G^{\op} \times_{(\iota_N r_N)^{\op}, \sO_G^{\op}} \underline{\Sp}^G \cong \sO_G^{\op} \times_{r_N^{\op}, \sO_{G/N}^{\op}} (\sO_{G/N}^{\op} \times_{i_N^{\op},\sO_G^{\op}} \underline{\Sp}^G) \to \underline{\Sp}^{G/N} $$
is then a $G/N$-functor. This defines by adjunction the $G/N$-functor\footnote{The ad-hoc notation $\widetilde{\sU}[N]$ for this $G/N$-functor is employed so as not to conflict with the $G$-functor $\underline{\sU}[N]$ in \cref{ExtendingNaiveSpectraToGCategory} below.}
\[ \widetilde{\sU}[N]: \sO_{G/N}^{\op} \times_{i_N^{\op},\sO_G^{\op}} \Sp^G \to \underline{\Fun}_{G/N}(\sO_G^{\op}, \underline{\Sp}^{G/N}). \]
Define $\sU[N]$ to be the fiber of $\widetilde{\sU}[N]$ over $(G/N)/(G/N)$.
\end{cnstr}
\begin{obs}[Monoidality of forgetful functor] \label{LaxMonoidalForgetfulFunctorToNaiveSpectra} In \cref{cnstr:ForgetfulFunctorToNaiveGSpectra}, the monoidality of inflation and restriction implies that with respect to $\underline{\Sp}^{G, \otimes}$ and $\underline{\Sp}^{G/N, \otimes}$, the $G$-functors $\underline{\inf}[N]$ and $\underline{\res}[N]$ are symmetric monoidal and the $G/N$-functor $\underline{\Psi}[N]$ is lax symmetric monoidal. Therefore, $\sU[N]$ is lax symmetric monoidal with respect to the pointwise symmetric monoidal structure on $\Sp^G_{\naive{N}}$.
\end{obs}
\begin{obs}[Extension to $G$-$\infty$-category] \label{ExtendingNaiveSpectraToGCategory} For any subgroup $H$ of $G$, consider the commutative diagram of restriction functors
\[ \begin{tikzcd}[row sep=4ex, column sep=8ex, text height=1.5ex, text depth=0.25ex]
\FF_H & \FF_G \ar{l}[swap]{\res^G_H} \\
\FF_{H/(H \cap N)} \ar{u}[swap]{\iota_{H \cap N}} & \FF_{G/N} \ar{l}{\res^{G/N}_{H/(H \cap N)}} \ar{u}{\iota_N}
\end{tikzcd}
\]
that yields by adjunction
\[ \begin{tikzcd}[row sep=4ex, column sep=8ex, text height=1.5ex, text depth=0.25ex]
\FF_H \ar{r}{\ind^G_H} \ar{d}{r_{N \cap H}} & \FF_G \ar{d}{r_N} \\
\FF_{H/(H \cap N)} \ar{r}[swap]{\ind^{G/N}_{H/(H \cap N)}} & \FF_{G/N}.
\end{tikzcd} \]
Precomposition by $(\ind^G_H)^{\op}: \sO_H^{\op} \to \sO_G^{\op}$ yields functors
\[ \res^G_H: \Fun_{G/N}(\sO^{\op}_G, \underline{\Sp}^{G/N}) \to \Fun_{H/(H\cap N)}(\sO^{\op}_H, \underline{\Sp}^{H/(H \cap N)}) \]
that assemble to the data of a functor $\sO_G^{\op} \to \Cat_{\infty}$ and thereby define a $G$-$\infty$-category $\underline{\Sp}^G_{\naive{N}}$. Furthermore, $\sU[N]$ extends to a $G$-functor $\underline{\sU}[N]: \underline{\Sp}^G \to \underline{\Sp}^G_{\naive{N}}$, given on the fiber over $G/H$ by $\sU[N \cap H]$.
\end{obs}
\begin{obs}[Evaluation functors] \label{evaluationFactorizationNaiveSpectra} For any subgroup $H$ of $G$, evaluation on the orbit $G/H$ yields a functor
\[ s_H^{\ast}: \Fun_{G/N}(\sO^{\op}_G, \underline{\Sp}^{G/N}) \to \Sp^{H/(H \cap N)}. \]
By construction, this fits into a commutative diagram
\[ \begin{tikzcd}[row sep=4ex, column sep=4ex, text height=1.5ex, text depth=0.25ex]
\Sp^G \ar{r}{\sU[N]} \ar{d}[swap]{\res^G_H} & \Fun_{G/N}(\sO_G^{\op}, \underline{\Sp}^{G/N}) \ar{d}{s_H^{\ast}} \\
\Sp^H \ar{r}{\Psi^{H \cap N}} & \Sp^{H/(H \cap N)}.
\end{tikzcd} \]
Because $\underline{\Sp}^{G/N}$ is a $G/N$-presentable $G/N$-stable $G/N$-$\infty$-category, the same holds for $\underline{\Fun}_{G/N}(\sO^{\op}_G, \underline{\Sp}^{G/N})$ with $G/N$-limits and colimits computed as in \cite[Prop.~9.17]{Exp2}. Thus, $\Fun_{G/N}(\sO_G^{\op}, \underline{\Sp}^{G/N})$ is a presentable stable $\infty$-category such that the $s_H^{\ast}$ form a set of jointly conservative functors that preserve and detect limits and colimits. Since both the restriction and categorical fixed points functors preserve limits and colimits, it follows that $\sU[N]$ preserves limits and colimits and therefore admits both left and right adjoints.
We also have a partial compatibility relation as $H$ varies. Namely, given $H$, if $K$ is a subgroup such that $N \cap H \leq K \leq H$ (so $K \cap N = H \cap N$), then
\[ \res^{H/(N \cap H)}_{K/(N \cap H)} \circ s_H^{\ast} \simeq s_K^{\ast}: \Fun_{G/N}(\sO^{\op}_G, \underline{\Sp}^{G/N}) \to \Sp^{K/N \cap H}. \]
\end{obs}
Let us separate out the conclusion of \cref{evaluationFactorizationNaiveSpectra} into a definition.
\begin{dfn}
Let $\sF[N], \sF^{\vee}[N]: \Sp^G_{\naive{N}} \to \Sp^G$ denote the left resp. right adjoints to $\sU[N]$.
\end{dfn}
\begin{obs}[Interaction with $G$-spaces] By repeating the construction of $\sU[N]$ for $G$-spaces and using the compatibility of restriction and categorical fixed points with $\Omega^{\infty}$, we obtain a commutative diagram
\[ \begin{tikzcd}[row sep=4ex, column sep=4ex, text height=1.5ex, text depth=0.25ex]
\Sp^G \ar{r}{\sU[N]} \ar{d}[swap]{\Omega^{\infty}} & \Fun_{G/N}(\sO^{\op}_G, \underline{\Sp}^{G/N}) \ar{d}{\Omega^{\infty}} \\
\Spc^G \ar{r}{\sU'[N]} & \Fun_{G/N}(\sO^{\op}_G, \underline{\Spc}^{G/N})
\end{tikzcd} \]
where the righthand $\Omega^{\infty}$ functor denotes postcomposition by the $G/N$-functor $\Omega^{\infty}: \underline{\Sp}^{G/N} \to \underline{\Spc}^{G/N}$. Moreover, a diagram chase reveals that under the equivalence
\[ \Fun_{G/N}(\sO_G^{\op}, \underline{\Spc}^{G/N}) \simeq \Fun(\sO^{\op}_G, \Spc) = \Spc^G \]
of \cite[Prop.~3.10]{Exp2}, $\sU'[N]$ is an equivalence.
\end{obs}
To understand the compact generation of $N$-naive $G$-spectra, we need the following lemma.
\begin{lem} \label{lem:CompactGeneration} Let $C$ and $\{C_i: i \in I \}$ be presentable stable $\infty$-categories (with $I$ a small set) such that each $C_i$ has a (small) set $\{ x_{i \alpha}: \alpha \in \Lambda_i \}$ of compact generators. Suppose we have functors $U_i: C \to C_i$ that preserve limits and colimits and are jointly conservative. Let $F_i$ be left adjoint to $U_i$. Then $C$ has a (small) set of compact generators given by $\{ F_i x_{i \alpha}: i \in I, \alpha \in \Lambda_i \}$. In particular, $C$ is compactly generated.
\end{lem}
\begin{proof} We check directly that the indicated set generates $C$. Let $c \in C$ be any object and suppose that $\Hom_C(\Sigma^n F_i x_{i \alpha}, c) \cong 0$ for all choices of indices. Then by adjunction, $\Hom_{C_i}(\Sigma^n x_{i \alpha}, U_i c) \cong 0$, hence $U_i c \simeq 0$ for all $i \in I$. Invoking the joint conservativity of the $U_i$, we deduce that $c \simeq 0$. As for compactness, note that the assumption that each $U_i$ preserves colimits ensures that its left adjoint $F_i$ preserves compact objects.
\end{proof}
\begin{cor} \label{cor:CompactGenerationNaive} The $\infty$-category $\Fun_{G/N}(\sO^{\op}_G, \underline{\Sp}^{G/N})$ has a set of compact generators given by
$$\left\{ (s_H)_!(1) : H \leq G \right\}$$
where $(s_H)_!$ denotes the left adjoint to the functor $(s_H)^*$ of \cref{evaluationFactorizationNaiveSpectra}.
\end{cor}
\begin{proof} By applying \cref{lem:CompactGeneration} to the functors $s_H^{\ast}$, we deduce that
$$\{(s_H)_! \left( \frac{H/(H \cap N)}{K/(H \cap N)}_+ \right) : H \leq G \}$$ is a set of compact generators for $\Fun_{G/N}(\sO^{\op}_G, \underline{\Sp}^{G/N})$. Because $\res^{H/(N \cap H)}_{K/(N \cap H)} s_H^{\ast} \simeq s_K^{\ast}$, we may eliminate redundant expressions and reduce to the set $\{ (s_H)_!(1) : H \leq G \}$.
\end{proof}
\subsection{\texorpdfstring{$N$}{N}-Borel \texorpdfstring{$G$}{G}-spectra}\label{SS:NBorel}
We next consider Borel $G$-spectra relative to $N$. Let $\psi$ denote the extension $N \to G \to G/N$. Our main goal is to prove \cref{thm:BorelSpectraAsCompleteObjects}, which plays a key role in the sequel.
\begin{dfn} \label{dfn:TwistedClassifyingSpace} Let $B^{\psi}_{G/N} N \subset \sO^{\op}_G$ be the full subcategory on those $G$-orbits that are $N$-free.
\end{dfn}
Note that $B^{\psi}_{G/N} N$ is a cosieve in $\sO^{\op}_G$: this amounts to the observation that if $U$ is $N$-free and $f: V \to U$ is a $G$-equivariant map of $G$-sets, then $V$ is $N$-free.
\begin{lem} \label{lm:TwistedClassifyingSpaceIsSpace} The cocartesian fibration $r_N^{\op}: \sO^{\op}_G \to \sO^{\op}_{G/N}$ restricts to a left fibration
\[ \rho_N: B^{\psi}_{G/N} N \to \sO^{\op}_{G/N}. \]
\end{lem}
\begin{proof} Because $B^{\psi}_{G/N} N$ is a cosieve, the inclusion $B^{\psi}_{G/N} N \subset \sO^{\op}_G$ is stable under $r_N^{\op}$-cocartesian edges, so $\rho_N$ is a cocartesian fibration such that the inclusion preserves cocartesian edges. Furthermore, if $f: U \to V$ is a $G$-equivariant map of $N$-free $G$-sets such that $\overline{f}: U/N \to V/N$ is an isomorphism, then it is easy to check that $f$ is an isomorphism. Because $\rho_N$ is conservative, we deduce that $\rho_N$ is in addition a left fibration.
\end{proof}
\begin{rem} If $N$ yields a product decomposition $G \cong G/N \times N$, then $B^{\psi}_{G/N} N$ is spanned by those orbits of the form $G/\Gamma_{\phi}$ where $\Gamma_{\phi}$ is the graph of a homomorphism $\phi: M \to G/N$ for $M \leq N$. As a $G/N$-space, $B^{\psi}_{G/N} N$ then is the classifying $G/N$-space for $G/N$-equivariant principal $N$-bundles, which is usually denoted as $B_{G/N} N$. We thus think of $B^{\psi}_{G/N} N$ as a twisted variant of $B_{G/N} N$ for non-trivial extensions $\psi$. Many other authors have also studied equivariant classifying spaces in varying levels of generality -- for example, see \cite{Guillou2017, Luck05, Lck2014, AlaskaNotes}.
\end{rem}
\begin{dfn} \label{dfn:BorelGSpectraRelativeToNormalSubgroup} Let $\Sp^G_{\Borel{N}} := \Fun_{G/N}(B^{\psi}_{G/N} N, \underline{\Sp}^{G/N})$ be the $\infty$-category of \emph{Borel $G$-spectra relative to $N$}, or \emph{$N$-Borel $G$-spectra}. We will also refer to $G/N$-functors $$X: B^{\psi}_{G/N} N \to \ul{\Sp}^{G/N}$$
as \emph{$G/N$-spectra with $\psi$-twisted $N$-action}.
\end{dfn}
\begin{ntn} Given an abelian group $A$, we will use $B^t_{C_2} A$ as preferred alternative notation in lieu of $B^{\psi}_{C_2} A$ for the defining extension $\psi = [A \to A \rtimes C_2 \to C_2]$ of the semidirect product, where $C_2$ acts on $A$ by the inversion involution. We will also refer to $C_2$-functors $X:B^t_{C_2} A \to \ul{\Sp}^{C_2}$ as \emph{$C_2$-spectra with twisted $A$-action}, leaving $\psi$ implicit.
\end{ntn}
\begin{dfn}
Define the forgetful functor
\[ \sU_b[N]: \Sp^G \to \Sp^G_{\Borel{N}} = \Fun_{G/N}(B^{\psi}_{G/N} N, \underline{\Sp}^{G/N}) \]
to be the composition of $\sU[N]$ with restriction along the $G/N$-functor $i: B^{\psi}_{G/N} N \subset \sO^{\op}_G$. Also let
\[ \sF_b[N], \: \sF^{\vee}_b[N]: \Sp^G_{\Borel{N}} \to \Sp^G \]
denote the left resp. right adjoints to $\sU_b[N]$ (cf. \cref{VariousPropertiesForgetfulFunctorBorelSpectra}(1) below).
\end{dfn}
Parallel to the above discussion of $\sU[N]$, we now enumerate some of the properties of $\sU_b[N]$.
\begin{obs}[Properties of {$\sU_b[N]$}] \label{VariousPropertiesForgetfulFunctorBorelSpectra}
\leavevmode
\begin{enumerate}[leftmargin=6ex] \item Because both $\sU[N]$ and restriction along $i$ preserve limits and colimits, $\sU_b[N]$ preserves limits and colimits and thus admits left and right adjoints $\sF_b[N]$ and $\sF^{\vee}_b[N]$ that factor through $\sF[N]$ and $\sF^{\vee}[N]$.
\item For all $G/H \in B^{\psi}_{G/N} N$ we have $H \cap N = 1$. Therefore, the smaller collection of functors
$$\{ s_H^{\ast}: \Sp^G_{\Borel{N}} \to \Sp^H : H \cap N = 1 \}$$
is jointly conservative and preserves and detects limits and colimits. Moreover, from \cref{evaluationFactorizationNaiveSpectra} we get that $s_H^{\ast} \circ \sU_b[N] \simeq \res^G_H$.
\item Since $\sU[N]$ is lax symmetric monoidal by \cref{LaxMonoidalForgetfulFunctorToNaiveSpectra} and restriction is symmetric monoidal, we get that $\sU_b[N]$ is a lax symmetric monoidal functor. However, because each $s_H^{\ast} \circ \sU_b[N]$ for $H \cap N = 1$ is now symmetric monoidal, $\sU_b[N]$ is in fact symmetric monoidal.
\item As in \cref{cnstr:ForgetfulFunctorToNaiveGSpectra}, the functor $\sU_b[N]$ is the fiber over $(G/N)/(G/N)$ of a $G/N$-functor $$\widetilde{\sU}_b[N]: \sO_{G/N}^{\op} \times_{\sO_G^{\op}} \ul{\Sp}^G \to \ul{\Fun}_{G/N}(\sO_G^{\op}, \ul{\Sp}^{G/N}). $$
Also, as in \cref{ExtendingNaiveSpectraToGCategory}, $\Sp^G_{\Borel{N}}$ extends to a $G$-$\infty$-category $\underline{\Sp}^G_{\Borel{N}}$ and $\sU_b[N]$ extends to a $G$-functor $\underline{\sU}_b[N]: \underline{\Sp}^G \to \underline{\Sp}^G_{\Borel{N}}$, given on the fiber over $G/H$ by
\[ \sU_b[N \cap H]: \Sp^H \to \Fun_{H/(N \cap H)}(B^{\psi_H}_{H/(H \cap N)} (N \cap H), \underline{\Sp}^{H/(N \cap H)}) \]
for the restricted extension $\psi_H \coloneq [N \cap H \to H \to H/(N \cap H)]$.
\item As with $N$-naive $G$-spectra, we have a commutative diagram
\[ \begin{tikzcd}[row sep=4ex, column sep=4ex, text height=1.5ex, text depth=0.25ex]
\Sp^G \ar{r}{\sU_b[N]} \ar{d}[swap]{\Omega^{\infty}} & \Fun_{G/N}(B^{\psi}_{G/N} N, \underline{\Sp}^{G/N}) \ar{d}{\Omega^{\infty}} \\
\Spc^G \ar{r}{\sU'_b[N]} \ar[equal]{d} & \Fun_{G/N}(B^{\psi}_{G/N} N, \underline{\Spc}^{G/N}) \ar{d}{\simeq} \\
\Fun(\sO^{\op}_G,\Sp) \ar{r}{i^{\ast}} & \Fun(B^{\psi}_{G/N} N, \Spc).
\end{tikzcd} \]
where now we may identify $\sU'_b[N]$ with restriction along $i$. Consider the transposed lax commutative diagram
\[ \begin{tikzcd}[row sep=4ex, column sep=4ex, text height=1.5ex, text depth=0.25ex]
\Sp^G \ar{r}{\sU_b[N]} \ar[phantom]{rd}{\NWarrow} & \Fun_{G/N}(B^{\psi}_{G/N} N, \underline{\Sp}^{G/N}) \\
\Spc^G \ar{r}{\sU'_b[N]} \ar{u}{\Sigma^{\infty}_+} & \Fun_{G/N}(B^{\psi}_{G/N} N, \underline{\Spc}^{G/N}) \ar{u}{\Sigma^{\infty}_+}.
\end{tikzcd} \]
For any subgroup $H$ such that $H \cap N = 1$, we may extend this diagram to
\[ \begin{tikzcd}[row sep=4ex, column sep=4ex, text height=1.5ex, text depth=0.25ex]
\Sp^G \ar{r}{\sU_b[N]} \ar[phantom]{rd}{\NWarrow} & \Fun_{G/N}(B^{\psi}_{G/N} N, \underline{\Sp}^{G/N}) \ar{r}{\ev_H} & \Sp^H \\
\Spc^G \ar{r}{\sU'_b[N]} \ar{u}{\Sigma^{\infty}_+} & \Fun_{G/N}(B^{\psi}_{G/N} N, \underline{\Spc}^{G/N}) \ar{u}{\Sigma^{\infty}_+} \ar{r}{\ev_H} & \Spc^H \ar{u}{\Sigma^{\infty}_+}
\end{tikzcd} \]
where the horizontal composites are given by the restriction functor $\res^G_H$ since $H/(H \cap N) \cong H$ (cf. \cref{evaluationFactorizationNaiveSpectra} and the analogous setup for $G$-spaces). The righthand square commutes by definition, and the outer square commutes by the compatibility of restriction with $\Sigma^{\infty}_+$. Since the $\ev_H$ are jointly conservative, it follows that the the lefthand square commutes.
\end{enumerate}
\end{obs}
\begin{ntn} \label{ntn:NFreeFamily} Let $\Gamma_N$ be the \emph{$N$-free} $G$-family consisting of subgroups $H$ such that $H \cap N = 1$.
\end{ntn}
Recall the definitions of $\cF$-torsion and $\cF$-complete $G$-spectra from \cref{cnstr:GspaceFromGfamily}.
\begin{thm} \label{thm:BorelSpectraAsCompleteObjects} The functors $\sF_b[N]$ and $\sF_b^{\vee}[N]$ are fully faithful with essential image the $\Gamma_N$-torsion and $\Gamma_N$-complete $G$-spectra, respectively.
\end{thm}
\begin{proof} We first check that the unit $\eta: \id \to \sU_b[N] \sF_b[N]$ is an equivalence to show that the left adjoint $\sF_b[N]$ is fully faithful. Because $\Omega^{\infty} \sU_b[N] \simeq i^{\ast} \Omega^{\infty}$ and $\Sigma^{\infty}_+ i^{\ast} \simeq \sU_b[N] \Sigma^{\infty}_+$ by \cref{VariousPropertiesForgetfulFunctorBorelSpectra}(5), we have an equivalence of left adjoints $\Sigma^{\infty}_+ i_! \simeq \sF_b[N] \Sigma^{\infty}_+$, and for $X \in \Fun(B^{\psi}_{G/N} N, \Spc)$ we may identify $\eta_{\Sigma^{\infty}_+ X}$ with $\Sigma^{\infty}_+ \eta'_X$, where $\eta': \id \to i^{\ast} i_!$ is the unit of the adjunction $i_! \dashv i_{\ast}$. But $\eta'$ is an equivalence since $i_!$ is left Kan extension along the inclusion of a full subcategory. Thus, $\eta$ is an equivalence on all suspension spectra. In view of the commutative diagram for $H \in \Gamma_N$
\[ \begin{tikzcd}[row sep=4ex, column sep=4ex, text height=1.5ex, text depth=0.25ex]
\Fun_{G/N}(B^{\psi}_{G/N} N, \underline{\Sp}^{G/N}) \ar{r}{s_H^{\ast}} \ar{d}{\Omega^{\infty}} & \Sp^H \ar{d}{\Omega^{\infty}} \\
\Fun(B^{\psi}_{G/N} N, \Spc) \ar{r}{{s'_H}^{\ast}} & \Spc^H
\end{tikzcd} \]
where $s_H' = (\ind^G_H)^{\op}: \sO_H^{\op} \to B^{\psi}_{G/N} N \subset \sO_G^{\op}$, we have an equivalence of left adjoints ${s_H}_! \Sigma^{\infty}_+ \simeq \Sigma^{\infty}_+ {s'_H}_!$, so in particular ${s_H}_! (1)$ is a suspension spectrum. Elaborating upon \cref{cor:CompactGenerationNaive}, we observe that the set $\{ {s_H}_! (1) : H \in \Gamma_N \}$ constitutes a set of compact generators for $\Sp^G_{\Borel{N}}$. Because both the domain and codomain of $\eta$ commute with colimits, we conclude that $\eta$ is an equivalence. Moreover, because $(s_H)^{\ast} \circ \sU_b[N] \simeq \res^G_H$ as noted in \cref{evaluationFactorizationNaiveSpectra}, by adjunction $\sF_b[N] \circ (s_H)_! \simeq \ind^G_H$ and thus the essential image of $\sF_b[N]$ is the localizing subcategory generated by the set $\{ G/H_+ : H \in \Gamma_N \}$. This equals the full subcategory of $\Gamma_N$-torsion $G$-spectra. Because we already have the stable recollement
\[ \begin{tikzcd}[row sep=4ex, column sep=4ex, text height=1.5ex, text depth=0.25ex]
\Sp^{h \Gamma_N} \ar[shift right=1,right hook->]{r}[swap]{j_{\ast}} & \Sp^G \ar[shift right=2]{l}[swap]{j^{\ast}} \ar[shift left=2]{r}{i^{\ast}} & \Sp^{\Phi \Gamma_N} \ar[shift left=1,left hook->]{l}{i_{\ast}},
\end{tikzcd} \]
it follows that $\sF^{\vee}_b[N]$ is fully faithful with essential image the $\Gamma_N$-complete $G$-spectra. In more detail:
\begin{itemize} \item The composite $i^{\ast} \sF_b[N] \simeq 0$ because $\sF_b[N](X) \in j_!(\Sp^{h \Gamma_N})$ and $i^{\ast} j_! \simeq 0$. Passing to adjoints, we get $\sU_b[N] i_{\ast} \simeq 0$. Then for all $X \in \Sp^G_{\Borel{N}}$, $\sF^{\vee}_b[N](X)$ is $\Gamma_N$-complete by the equivalence
\[ \Map(i_{\ast} Z, \sF^{\vee}_b[N](X)) \simeq \Map(\sU_b[N] i_{\ast} Z, X) \simeq 0. \]
\item Using that $\sF_b[N]$ is an equivalence onto its essential image, we see that the composite $\sU_b[N] j_!$ is an equivalence from $\Sp^{h \Gamma_N}$ to $\Sp^G_{\Borel{N}}$. Its right adjoint $j^{\ast} \sF^{\vee}_b[N]$ is thus an equivalence.
\item Combining these two assertions, we have that the composite
\[ \begin{tikzcd}[row sep=4ex, column sep=4ex, text height=1.5ex, text depth=0.25ex]
\Sp^G_{\Borel{N}} \ar{r}{\sF^{\vee}_b[N]} \ar[bend right=15]{rr}[swap]{\simeq} & \Sp^G \ar{r}{j^{\ast}} & \Sp^{h \Gamma_N} \ar{r}{j_{\ast}} & \Sp^G
\end{tikzcd} \]
is equivalent to $\sF^{\vee}_b[N]$ via the unit $\sF^{\vee}_b[N] \xto{\simeq} j_{\ast} j^{\ast} \sF^{\vee}_b[N]$ and is fully faithful onto $\Gamma_N$-complete $G$-spectra.
\end{itemize}
\end{proof}
\begin{rem}[Monoidality] \label{rem:MonoidalIdentificationOfBorelSpectra} Because $\sU_b[N]$ is symmetric monoidal by \cref{VariousPropertiesForgetfulFunctorBorelSpectra}(3), its right adjoint $\sF^{\vee}_b[N]$ is lax symmetric monoidal. Therefore, the equivalence
$$\sF^{\vee}_b[N]: \Fun_{G/N}(B^{\psi}_{G/N} N, \underline{\Sp}^{G/N}) \xto{\simeq} \Sp^{h \Gamma_N}$$
of \cref{thm:BorelSpectraAsCompleteObjects} is one of symmetric monoidal $\infty$-categories with respect to the pointwise symmetric monoidal structure on the lefthand side and the symmetric monoidal structure induced by the $\Gamma_N$-symmetric monoidal recollement on the righthand side.
\end{rem}
\begin{cor}[Compatibility with restriction] \label{cor:BorelCompatibilityWithRestriction} The left and right adjoints $\sF_b[N]$ and $\sF_b^{\vee}[N]$ extend to $G$-left and right adjoints
\[ \underline{\sF}_b[N], \: \underline{\sF}_b^{\vee}[N]: \underline{\Sp}^G_{\Borel{N}} \to \underline{\Sp}^G. \]
\end{cor}
\begin{proof} Combine \cref{ParamRecollementFamily} and \cref{thm:BorelSpectraAsCompleteObjects}.
\end{proof}
We conclude this section by applying \cref{thm:BorelSpectraAsCompleteObjects} to decompose the $\infty$-category of $D_{2p^n}$-spectra; this example plays a crucial role in our study of real cyclotomic spectra in \cite{QS21b}.
\begin{exm} \label{exm:DihedralRecollement} Let $\Gamma = \Gamma_{\mu_{p^n}}$ be the $D_{2p^n}$-family that consists of those subgroups $H$ such that $H \cap \mu_{p^n} = 1$. Note that $H \notin \Gamma$ if and only if $\mu_p \leq H$. Therefore, $\Sp^{\Phi \Gamma} \simeq \Sp^{D_{2 p^{n-1}}}$ for $D_{2 p^{n-1}}$ viewed as the quotient $D_{2p^n}/\mu_p$. Together with \cref{thm:BorelSpectraAsCompleteObjects}, we obtain a stable symmetric monoidal recollement
\[ \begin{tikzcd}[row sep=4ex, column sep=10ex, text height=1.5ex, text depth=0.5ex]
\Fun_{C_2}(B^t_{C_2} \mu_{p^n}, \underline{\Sp}^{C_2}) \ar[shift right=1,right hook->]{r}[swap]{j_{\ast} = \sF^{\vee}_b[\mu_{p^n}]} & \Sp^{D_{2p^n}} \ar[shift right=2]{l}[swap]{j^{\ast} = \sU_b[\mu_{p^n}]} \ar[shift left=2]{r}{i^{\ast} = \Phi^{\mu_p}} & \Sp^{D_{2p^{n-1}}} \ar[shift left=1,left hook->]{l}{i_{\ast}}.
\end{tikzcd} \]
Furthermore, using \cref{ParamRecollementFamily}, this extends to a $C_2$-stable $C_2$-recollement
\[ \begin{tikzcd}[row sep=4ex, column sep=8ex, text height=1.5ex, text depth=0.5ex]
\ul{\Fun}_{C_2}(B^t_{C_2} \mu_{p^n}, \underline{\Sp}^{C_2}) \ar[shift right=1,right hook->]{r}[swap]{\widetilde{\sF}^{\vee}_b[\mu_{p^n}]} & \sO^{\op}_{C_2} \times_{\sO^{\op}_{D_{2p^n}}} \ul{\Sp}^{D_{2p^n}} \ar[shift right=2]{l}[swap]{ \widetilde{\sU}_b[\mu_{p^n}]} \ar[shift left=2]{r}{\Phi^{\mu_p}} & \sO^{\op}_{C_2} \times_{\sO^{\op}_{D_{2p^{n-1}}}} \ul{\Sp}^{D_{2p^{n-1}}} \ar[shift left=1,left hook->]{l}{i_{\ast}}
\end{tikzcd} \]
whose fiber over $C_2/1$ is the stable symmetric monoidal recollement
\[ \begin{tikzcd}[row sep=4ex, column sep=10ex, text height=1.5ex, text depth=0.5ex]
\Fun(B \mu_{p^n}, \Sp) \ar[shift right=1,right hook->]{r}[swap]{\sF^{\vee}_b[\mu_{p^n}]} & \Sp^{\mu_{p^n}} \ar[shift right=2]{l}[swap]{\sU_b[\mu_{p^n}]} \ar[shift left=2]{r}{\Phi^{\mu_p}} & \Sp^{\mu_{p^{n-1}}} \ar[shift left=1,left hook->]{l}{i_{\ast}},
\end{tikzcd} \]
with the $C_2$-action induced by the inversion action of $C_2$ on $\mu_{p^n}$.
\end{exm}
\section{Parametrized norm maps and ambidexterity}
\label{section:NormMaps}
In this section, we construct \emph{parametrized norm maps} that will permit us to define the parametrized Tate construction (\cref{dfn:ParamTateCnstr}). Our strategy is to mimic Lurie's construction of the norm maps \cite[\S 6.1.6]{HA} in a parametrized setting over a base $\infty$-category $S$, eventually specializing to $S = \sO_G^{\op}$. We first collect a few necessary aspects of the theory of $S$-colimits, limits, and Kan extensions from \cite{Exp2}.
\begin{obs} \label{GeneralTheoryParamColimits} Let $K$, $L$, and $C$ be $S$-$\infty$-categories (with $p: L \to S$ the structure map), and let $\phi: K \to L$ be an $S$-cocartesian fibration \cite[Def.~7.1]{Exp2}. We are interested in computing the left adjoint $\phi_!$ to the restriction functor
$$\phi^{\ast}: \Fun_S(L,C) \to \Fun_S(K,C)$$
as a (pointwise) $S$-left Kan extension \cite[Def.~10.1 or Def.~9.13]{Exp2}. In \cite[Thm.~9.15 and Prop.~10.8]{Exp2}, the second author gave a pointwise existence criterion and formula for $\phi_!$ of an $S$-functor $F: K \to C$. Namely, for all objects $x \in L$, let $\underline{x} \coloneq \{ x \} \times_{L, \ev_0} \Ar^{\cocart}(L)$ be as in \cite[Notn.~2.29]{Exp2} and let $i_x: \underline{x} \to L$ be the $S$-functor given by evaluation at the target (which serves as a canonical extension of the inclusion of the point $x \in L$ to an $S$-functor). Let $s = p(x)$ and note that $q = (\id, \Ar(p)): \trivfibto{\underline{x}}{S^{s/}}$ is a trivial fibration (cf. \cite[Lem.~12.10]{Exp2}), so we may think of $\underline{x}$ as an \emph{$S$-point} of $L$. Let
\[ F_x = (q \circ \pr_{\underline{x}}, F \circ \pr_K): K_{\underline{x}} \coloneq \underline{x} \times_L K \to C_{\underline{s}} \coloneq S^{s/} \times_S C \]
denote the resulting $S^{s/}$-functor. Then $\phi_! F$ exists if the $S^{s/}$-colimit of $F_x$ exists for all $x \in L$, after which $(\phi_! F)|_{\underline{x}}$ is computed as that $S^{s/}$-colimit. We will also say that $C$ \emph{admits the relevant $S$-colimits} with respect to $\phi: K \to L$ if for all $x \in L$ with $p(x)=s$, $C_{\underline{s}}$ admits all $S^{s/}$-colimits indexed by $K_{\underline{x}}$. In this case, the left adjoint $\phi_!$ to $\phi^{\ast}$ exists by \cite[Cor.~9.16]{Exp2}.
Now suppose instead that $\phi: K \to L$ is an $S$-cartesian fibration \cite[Def.~7.1]{Exp2}. In view of the discussion of vertical opposites in \cite[\S 5]{Exp2} and the observation that the formation of vertical opposites exchanges $S$-cocartesian and $S$-cartesian fibrations, we may dualize the above discussion to see that the $S$-right Kan extension $\phi_{\ast} F$ exists if the $S^{s/}$-limit of $F_x$ exists for all $x \in L$, after which $\phi_{\ast} F|_{\underline{x}}$ is computed as that $S^{s/}$-limit. Likewise, we have the dual notion of $C$ admitting the relevant $S$-limits with respect to $\phi$, in which case the right adjoint $\phi_{\ast}$ exists.
Finally, suppose that $K$ and $L$ are $S$-spaces. Using the cocartesian model structure on $s\Set^+_{/S}$ and the description of the fibrations between fibrant objects \cite[Prop.~B.2.7]{HA}, up to equivalence we may replace any $S$-functor $\phi: K \to L$ by a categorical fibration. But a categorical fibration between left fibrations over $S$ is necessarily both an $S$-cocartesian and $S$-cartesian fibration, hence both of the above formulas apply to compute $\phi_!$ and $\phi_{\ast}$.
\end{obs}
\begin{obs} We can also consider the $S$-functor $S$-$\infty$-category $\underline{\Fun}_S(K,C) \to S$ whose cocartesian sections are $\Fun_S(K,C)$. Let $\phi: K \to L$ be an $S$-cocartesian fibration and suppose that for every $s \in S$, the $S^{s/}$-$\infty$-category $C_{\underline{s}}$ admits the relevant $S^{s/}$-colimits with respect to $\phi_{\underline{s}}: K_{\underline{s}} \to L_{\underline{s}}$, so that the restriction functor
$$\phi_{\underline{s}}^{\ast}: \Fun_{S^{s/}}(L_{\underline{s}}, C_{\underline{s}}) \to \Fun_{S^{s/}}(K_{\underline{s}}, C_{\underline{s}})$$
admits a left adjoint $(\phi_{\underline{s}})_!$ computed as above. Then using the built-in compatibility of $S$-left Kan extension with restriction, by \cite[Prop.~7.3.2.11]{HA} these fiberwise left adjoints assemble to yield an $S$-adjunction \cite[Def.~8.3]{Exp2}
\[ \adjunct{\underline{\phi}_!}{\underline{\Fun}_S(K,C)}{\underline{\Fun}_S(L,C)}{\underline{\phi}^{\ast}} \]
(also see \cite[Cor.~9.16 and Thm.~10.5]{Exp2}). In particular, upon forgetting the structure maps\footnote{In other words, a relative adjunction yields an adjunction between the Grothendieck constructions.} we have an ordinary adjunction $\underline{\phi}_! \dashv \underline{\phi}^{\ast}$. Similarly, for $\phi$ an $S$-cartesian fibration we can consider the $S$-adjunction
\[ \adjunct{\underline{\phi}^{\ast}}{\underline{\Fun}_S(L,C)}{\underline{\Fun}_S(K,C)}{\underline{\phi}_{\ast}}. \]
For $\phi: X \to Y$ a map of $S$-spaces, we will consider $\underline{\phi}_! \dashv \underline{\phi}^{\ast} \dashv \underline{\phi}_{\ast}$.
\end{obs}
The key result that enables the construction of norm maps is the following lemma on adjointability. Note for the formulation of the statement that $S$-(co)cartesian fibrations are stable under pullback, and the property that $C$ admits the relevant $S$-(co)limits with respect to $\phi$ is stable under pullbacks in the $\phi$ variable.
\begin{lem} \label{lem:adjointability} Let $C$ be an $S$-$\infty$-category and let
\[ \begin{tikzcd}[row sep=4ex, column sep=4ex, text height=1.5ex, text depth=0.25ex]
K' \ar{r}{f'} \ar{d}[swap]{\phi'} & K \ar{d}{\phi} \\
L' \ar{r}{f} & L
\end{tikzcd} \]
be a pullback square of $S$-$\infty$-categories. Consider the resulting commutative square of $S$-functor categories and restriction functors
\[ \begin{tikzcd}[row sep=4ex, column sep=6ex, text height=1.5ex, text depth=0.25ex]
\Fun_S(K',C) & \Fun_S(K,C) \ar{l}[swap]{{f'}^{\ast}} \\
\Fun_S(L',C) \ar{u}{{\phi'}^{\ast}} & \Fun_S(L,C) \ar{u}[swap]{\phi^{\ast}} \ar{l}[swap]{f^{\ast}}.
\end{tikzcd} \]
\begin{enumerate} \item If $\phi$ is an $S$-cocartesian fibration and $C$ admits the relevant $S$-colimits, then the square is left adjointable, i.e., the natural map ${\phi'}_! {f'}^{\ast} \to f^{\ast} \phi_!$ is an equivalence.
\item If $\phi$ is an $S$-cartesian fibration and $C$ admits the relevant $S$-limits, then this square is right adjointable, i.e., the natural map $f^{\ast} \phi_{\ast} \to {\phi'}_{\ast} {f'}^{\ast}$ is an equivalence.
\item If $K, L, K', L'$ are $S$-spaces, then the square is both left and right adjointable provided that $C$ admits the relevant $S$-colimits and $S$-limits.
\end{enumerate}
Likewise, we have the same results for the commutative square of $S$-functor $S$-$\infty$-categories
\[ \begin{tikzcd}[row sep=4ex, column sep=6ex, text height=1.5ex, text depth=0.5ex]
\underline{\Fun}_S(K',C) & \underline{\Fun}_S(K,C) \ar{l}[swap]{\underline{f'}^{\ast}} \\
\underline{\Fun}_S(L',C) \ar{u}{\underline{\phi'}^{\ast}} & \underline{\Fun}_S(L,C) \ar{u}[swap]{\underline{\phi}^{\ast}} \ar{l}[swap]{\underline{f}^{\ast}}.
\end{tikzcd} \]
\end{lem}
\begin{proof} Let $F: K \to C$ be an $S$-functor. For (1), we need to check that ${\phi'}_! {f'}^{\ast} F \to f^{\ast} \phi_! F$ is an equivalence of $S$-functors. It suffices to evaluate on $S$-points $\underline{x}$ in $L'$, and we then have the map
\[ \colim^{S^{s/}} (F \circ f')_x \to \colim^{S^{s/}} F_{f(x)} \]
where $x$ lies over $s$. But since $\underline{x} \times_{L'} K' \simeq \underline{f(x)} \times_L K$ as $S^{s/}$-$\infty$-categories, these $S^{s/}$-colimits are equivalent under the comparison map. The proof of (2) is similar. For (3) we replace $\phi$ by a categorical fibration and then use (1) and (2). For the corresponding assertion about $\underline{\Fun}_S(-,C)$, it suffices to check that the natural transformations of interest are equivalences fiberwise, upon which we reduce to the prior assertion for $\Fun_{S^{s/}}(-,C_{\underline{s}})$ (ranging over all $s \in S$).
\end{proof}
\subsection{Ambidexterity of parametrized local systems}\label{SS:Ambi}
In this subsection, we extend Hopkins and Lurie's study of ambidexterity for local systems \cite[\S 4.3]{HL13} to the parametrized setting. The following definition generalizes \cite[Def.~4.3]{HL13}.
\begin{dfn} Let $C$ be an $S$-$\infty$-category. The $\infty$-category of \emph{$S$-local systems} on $C$ $$\LocSys^S(C) \to \Spc^S$$ is the cartesian fibration classified by the composite
\[ \Fun_S(-,C): (\Spc^S)^{\op} \subset \Cat_{\infty}^{S,\op} \to \Cat_{\infty}. \]
The \emph{$S$-$\infty$-category of $S$-local systems} on $C$ $$\ul{\LocSys}^S(C) \to \Spc^S$$ is the cartesian fibration classified by the composite
\[ \underline{\Fun}_S(-,C): (\Spc^S)^{\op} \subset \Cat_{\infty}^{S,\op} \to \Cat_{\infty}^S \xto{U} \Cat_{\infty}. \]
where $U$ forgets the structure map of a cocartesian fibration.
\end{dfn}
\begin{cor} \label{LocalSystemsIsBCFibration} Suppose that for all $s \in S$, $C_{\underline{s}}$ admits all $S^{s/}$-colimits indexed by $S^{s/}$-spaces. Then $\LocSys^S(C)$ and $\ul{\LocSys}^S(C)$ are Beck-Chevalley fibrations \cite[Def.~4.1.3]{HL13}.
\end{cor}
\begin{proof} Note that by the same argument as \cite[Rem.~5.14]{Exp2}, the hypothesis ensures that $C$ admits all $S$-colimits indexed by $S$-spaces. The corollary is then immediate from \cref{lem:adjointability}.
\end{proof}
\begin{rem}
The natural transformation
\[ \theta: \underline{\Fun}_S(-,C) \Rightarrow \mathrm{const}_S: (\Spc^S)^{\op} \to \Cat_{\infty}, \]
given objectwise by the structure map $\theta_X: \underline{\Fun}_S(X,C) \to S$, unstraightens to define a cocartesian fibration
$$\ul{\LocSys}^S(C) \to S,$$
so $\ul{\LocSys}^S(C)$ is indeed an $S$-$\infty$-category.
\end{rem}
We now have the general theory of ambidexterity \cite[\S 4.1-2]{HL13} for a Beck-Chevalley fibration, along with the attendant notions of ambidextrous and weakly ambidextrous morphisms \cite[Constr.~4.1.8 and Def.~4.1.11]{HL13} in $\Spc^S$. For the reader's convenience, let us recall the relevance of these notions for constructing norm maps, referring to \cite[\S 4.1]{HL13} for greater detail and precise definitions.
\begin{obs} Suppose $\sE \to \Spc^S$ is a Beck-Chevalley fibration, $f: X \to Y$ is a map of $S$-spaces, the left and right adjoints $f_!$ and $f_{\ast}$ to $f^{\ast}$ exist, and we wish to construct a norm map $\Nm_f: f_! \to f_{\ast}$. Consider the commutative diagram
\[ \begin{tikzcd}[row sep=4ex, column sep=6ex, text height=1.5ex, text depth=0.5ex]
X \ar{rd}{\delta} \ar[bend left=20]{rrd}{=} \ar[bend right=20]{rdd}[swap]{=} & \\
& X \times_Y X \ar{r}{\pr_2} \ar{d}{\pr_1} & X \ar{d}{f} \\
& X \ar{r}{f} & Y
\end{tikzcd} \]
and suppose we have already constructed a norm map $\Nm_{\delta}: \delta_! \to \delta_{\ast}$ and shown it to be an equivalence. If $\Nm_{\delta}^{-1}: \delta_{\ast} \xto{\simeq} \delta_!$ is a choice of inverse, then we have a natural transformation
\[ \pr_1^{\ast} \xto{\eta_{\delta}} \delta_{\ast} \delta^{\ast} \pr_1^{\ast} \simeq \delta_{\ast} \xto{\Nm_{\delta}^{-1}} \delta_! \simeq \delta_! \delta^{\ast} \pr_2^{\ast} \xto{\epsilon_{\delta}} \pr_2^{\ast}. \]
By adjunction and using the Beck-Chevalley property, we obtain a map
\[ f^{\ast} f_! \simeq (\pr_1)_! \pr_2^{\ast} \to \id. \]
Finally, we may adjoint this map in turn to define
\[ \Nm_f: f_! \to f_{\ast}. \]
Thus, for an inductive construction of norm maps, we may single out a class of `ambidextrous' morphisms for which a norm map has been constructed and shown to be an equivalence, and then define `weakly ambidextrous' morphisms to be those morphisms $f: X \to Y$ whose diagonal $\delta: X \to X \times_Y X$ is ambidextrous.
\end{obs}
Continuing our study, we henceforth suppose that $C_{\underline{s}}$ also admits all $S^{s/}$-limits indexed by $S^{s/}$-spaces, so that the right adjoints $f_{\ast}$, $\underline{f}_{\ast}$ exist for all maps $f$ of $S$-spaces. Then by \cite[Rem.~4.1.12]{HL13}, for $\LocSys^S(C)$ a map $f: X \to Y$ in $\Spc^S$ is ambidextrous if and only if the norm map $\Nm_{f'}: f'_! \to f'_{\ast}$ is an equivalence for all pullbacks $f': X' \to Y'$ of $f$, and similarly for $\ul{\LocSys}^S(C)$.
To simplify the following discussion, we will phrase all of our statements for $\LocSys^S(C)$. However, such statements have obvious implications for $\ul{\LocSys}^S(C)$ via checking fiberwise.
\begin{lem} \label{lem:AmbidexCheckedFiberwise} \begin{enumerate}[leftmargin=*] \item Let $f: X \to Y$ be a weakly ambidextrous morphism. Then $f$ is ambidextrous if and only if for all $y \in Y$, the norm map $\Nm_{f_y}$ for the pullback $f_y: X_{\underline{y}} \to \underline{y}$ is an equivalence.
\end{enumerate}
\begin{enumerate}
\setcounter{enumi}{1}
\item $f: X \to Y$ is weakly ambidextrous if and only if for all $y \in Y$, the pullback $f_y: X_{\underline{y}} \to \underline{y}$ is weakly ambidextrous.
\end{enumerate}
\end{lem}
\begin{proof} For (1), first note that the maps $f_y$ are weakly ambidextrous by \cite[Prop.~4.1.10(3)]{HL13}, so the statement is well-posed. The `only if' direction holds by definition. For the `if' direction, suppose given a pullback square of $S$-spaces
\[ \begin{tikzcd}[row sep=4ex, column sep=4ex, text height=1.5ex, text depth=0.25ex]
X' \ar{d}[swap]{f'} \ar{r}{\phi'} & X \ar{d}{f} \\
Y' \ar{r}{\phi} & Y.
\end{tikzcd} \]
For any point $y' \in Y'$, if we let $y=\phi(y')$ then we have an equivalence $X'_{\underline{y'}} \simeq X_{\underline{y}} \to \underline{y'} \simeq \underline{y}$. Therefore, without loss of generality it suffices to prove that $\Nm_f: f_! \to f_{\ast}$ is an equivalence. Let $y \in Y$ and denote the inclusion of the $S$-point as $i_y: \underline{y} \to Y$ and the $S$-fiber as $j_y: X_{\underline{y}} \to X$. By \cite[Rem.~4.2.3]{HL13} and \cref{lem:adjointability}, we have an equivalence
\[ i_y^{\ast} \Nm_f \simeq \Nm_{f_y} j_y^{\ast} : \Fun_S(X,C) \to \Fun_S(\underline{y},C) \simeq C_s \]
(where $y$ covers $s$), which by assumption is an equivalence. Because the evaluation functors $i_y^{\ast}$ are jointly conservative, it follows that $\Nm_f$ is an equivalence.
For (2), we only need to prove the `if' direction. We will show that the diagonal $\delta: X \to X \times_Y X$ is ambidextrous. Let $\delta_y$ denote the diagonal for $f_y$. For all $y$ we have a pullback square
\[ \begin{tikzcd}[row sep=4ex, column sep=6ex, text height=1.5ex, text depth=0.25ex]
X_{\underline{y}} \ar{r}{j_y} \ar{d}{\delta_y} & X \ar{d}{\delta} \\
X_{\underline{y}} \times_{\underline{y}} X_{\underline{y}} \ar{r}{(j_y,j_y)} & X \times_Y X.
\end{tikzcd} \]
Given any object $(x,x') \in X \times_Y X$ with $f(x)=f(x')=y$,\footnote{We write an equality here because we are implicitly modeling $f$ as a categorical fibration of left fibrations over $S$.} the inclusion of the $S$-fiber
$$i_{(x,x')}: \underline{(x,x')} \to X \times_Y X$$
factors through $X_{\underline{y}} \times_{\underline{y}} X_{\underline{y}}$. Therefore, if $\delta_y$ is ambidextrous, the norm map for $X_{\underline{(x,x')}} \to \underline{(x,x')}$ is an equivalence. By statement (1) of the lemma, we conclude that $\delta$ is ambidextrous.
\end{proof}
Recall from \cite[Prop.~4.1.10(6)]{HL13} that given a weakly $n$-ambidextrous morphism $f: X \to Y$ and $-2 \leq m \leq n$, $f$ is weakly $m$-ambidextrous if and only if $f$ is $m$-truncated \cite[Def.~5.5.6.1]{HTT}. To identify the $n$-truncated maps in $\Spc^S$, we have the following result.
\begin{lem} \label{lem:TruncationPresheaves} Let $X: S \to \Spc$ be an $S$-space. Then $X$ is $n$-truncated as an object of $\Spc^S$ if and only if for each $s \in S$, $X(s)$ is an $n$-truncated space. Similarly, for a map $f: X \to Y$ of $S$-spaces, $f$ is $n$-truncated if and only if $f(s)$ is an $n$-truncated map of spaces for all $s \in S$.
\end{lem}
\begin{proof} We repeat the argument of \cite[5.5.8.26]{HTT} for the reader's convenience. It suffices to prove the result for maps. Let $j: S^{\op} \to \Spc^S$ denote the Yoneda embedding. Then if $f$ is $n$-truncated, for any $s \in S$,
\[ f(s) \simeq \Map(j(s),f): \Map(j(s),X) \simeq X(s) \to \Map(j(s),Y) \simeq Y(s) \]
is $n$-truncated. Conversely, suppose each $f(s)$ is $n$-truncated. The collection of $S$-spaces $Z$ for which $\Map(Z,f)$ is $n$-truncated is stable under colimits, because limits of $n$-truncated spaces and maps are again $n$-truncated. Since the representable functors $j(s)$ generate $\Spc^S$ under colimits, it follows that $f$ itself is $n$-truncated.
\end{proof}
Let us now consider the $n=-1$ case.
\begin{dfn} Let $C$ be an $S$-$\infty$-category. $C$ is \emph{$S$-pointed} if for every $s \in S$, $C_s$ is pointed, and for every $\alpha: s \to t$, the pushforward functor $\alpha_{\sharp}: C_s \to C_{t}$ preserves the zero object. If $S = \sO_G^{\op}$, we also say that $C$ is \emph{$G$-pointed}.
\end{dfn}
\begin{lem} \label{lem:pointedAmbidex} $C$ is $S$-pointed if and only if for every $s \in S$, the weakly $(-1)$-ambidextrous morphism $0_s: \emptyset \to \underline{s}$ is ambidextrous.
\end{lem}
\begin{proof} For any $s \in S$, it is easy to see that the norm map $\Nm_{0_s}$ is the canonical map between the initial and final object in $C_s$. Moreover, for any $[\alpha: s \rightarrow t] \in \underline{s}$, the map $\underline{\alpha} \to \underline{s}$ is homotopic to $\alpha^{\ast}: \underline{t} \to \underline{s}$ and the pullback of $0_s$ along $\alpha^{\ast}$ is $0_t$. Thus $0_s$ is ambidextrous if and only if $C_s$ admits a zero object and for all $\alpha: s \to t$, the pushforward functor $\alpha_{\sharp}: C_s \to C_t$ preserves the zero object. The conclusion then follows.
\end{proof}
\begin{wrn} \label{wrn:ambidexCounterexample} In contrast to the non-parametrized case \cite[Prop.~6.1.6.7]{HA}, if $C$ is $S$-pointed then we may have weakly $(-1)$-ambidextrous morphisms that fail to be ambidextrous. For example, let $S = \sO_{C_2}^{\op}$ and let $p: J \to \sO_{C_2}^{\op}$ be the $C_2$-functor given by the inclusion of the full subcategory on the free transitive $C_2$-sets (so $J \simeq BC_2$). Then $J$ is a $C_2$-space via $p$, and for any $C_2$-$\infty$-category $C$, we have that $\Fun_{C_2}(J, C) \simeq (C_{C_2/1})^{h C_2}$ for the $C_2$-action on the fiber $C_{C_2/1}$ encoded by the cocartesian fibration. On the one hand, the $C_2$-diagonal functor $\delta: J \to J \times_{\sO_{C_2}^{\op}} J$ is an equivalence, so $J$ is a $(-1)$-truncated $C_2$-space. On the other hand, for $C = \underline{\Sp}^{C_2}$, the restriction $p^{\ast}$ may be identified with
\[ j^{\ast}: \Sp^{C_2} \simeq \Fun_{C_2}(\sO_{C_2}^{\op}, \underline{\Sp}^{C_2}) \to \Fun_{C_2}(J,\underline{\Sp}^{C_2}) \simeq \Fun(BC_2, \Sp) \]
which we saw has left and right adjoints $j_!$ and $j_{\ast}$ such that the norm map $j_! \to j_{\ast}$ is \emph{not} an equivalence.
\end{wrn}
We next consider the $n=0$ case. Let $T = S^{\op}$. Recall from \cite[Def.~4.1]{Exp4} that an $\infty$-category $T$ is said to be \emph{atomic orbital} if its finite coproduct completion $\FF_T$ admits pullbacks and $T$ has no non-trivial retracts (i.e., every retract is an equivalence). For example, $\sO_G$ is atomic orbital. We now assume that $T$ is atomic orbital, and we regard $V \in T$ as `orbits' and $U \in \FF_T$ as `finite $T$-sets'. Recall the notation $\underline{U}$ from \cref{exm:corepresentableDiagrams}.
\begin{lem} \label{lem:ComparisonCoproductProduct} Suppose $C$ is $S$-pointed. Then for any finite $T$-set $U$, orbit $V$ and morphism in $\Spc^S$
\[ f: \underline{U} \to \underline{V} \]
(necessarily specified by a morphism $f: U \to V$ in $\FF_T$), the diagonal $\delta: \underline{U} \to \underline{U} \times_{\underline{V}} \underline{U}$ is ambidextrous. Consequently, if $g: X \to Y$ is a morphism between finite coproducts of representables, then $g$ is weakly $0$-ambidextrous.
\end{lem}
\begin{proof} By our assumption on $T$, $\underline{U} \times_{\underline{V}} \underline{U}$ decomposes as a finite disjoint union of representables $\coprod_{i \in I} \underline{V_i}$. Moreover, because $T$ admits no non-trivial retracts, for some $J \subset I$ we have that $\underline{U} \simeq \coprod_{j \in J} \underline{V_j}$ with matching orbits, and $\delta$ is a summand inclusion $\underline{U} \to (\coprod_{j \in J} \underline{V_j}) \sqcup (\coprod_{i \in I-J } \underline{V_i})$. $\delta$ is then ambidextrous by \cref{lem:AmbidexCheckedFiberwise} and \cref{lem:pointedAmbidex}. The final consequence also follows by \cref{lem:AmbidexCheckedFiberwise}.
\end{proof}
By \cref{lem:ComparisonCoproductProduct}, the following definition is well-posed.
\begin{dfn} \label{dfn:semiadditive} Let $C$ be $S$-pointed. We say that $C$ is \emph{$S$-semiadditive} if for each morphism $f: U \to V$ in $\FF_T$, the norm map $\Nm_f$ for $f: \underline{U} \to \underline{V}$ is an equivalence. If $S = \sO_G^{\op}$, we will instead say that $C$ is \emph{$G$-semiadditive}.
\end{dfn}
Equivalently, in \cref{dfn:semiadditive} we could demand only that the norm maps for $f: U \to V$ with $V$ an orbit are equivalences.
\begin{rem} Unwinding the definition of the norm maps produced via our setup and in \cite[Constr.~5.2]{Exp4}, one sees that \cref{dfn:semiadditive} is the same as the notion of $T$-semiadditive given in \cite[Def.~5.3]{Exp4}. In particular, for $T = \sO_G^{\op}$, $\underline{\Sp}^G$ is an example of a $G$-semiadditive $G$-$\infty$-category. This amounts to the familiar fact that for each orbit $G/H$, $\Sp^H$ is semiadditive, and for each map of orbits $f: G/H \to G/K$, the left and right adjoints to the restriction functor $f^{\ast}: \Sp^K \to \Sp^H$ given by induction and coinduction are canonically equivalent.
\end{rem}
In the remainder of this subsection, we further specialize to the case $S = \sO_G^{\op}$ for $G$ a finite group. We have already encountered a potential problem in \cref{wrn:ambidexCounterexample} with developing a useful theory of $G$-ambidexterity. The issue is essentially due to the presence of fiberwise discrete $G$-spaces that do not arise from $G$-sets. To remedy this, we will restrict our attention to the Borel subclass of $G$-spaces.
\begin{dfn} \label{dfn:BorelGSpace} Suppose that $X \to \sO_G^{\op}$ is a $G$-space. Then $X$ is \emph{Borel} if the functor $\sO^{\op}_G \to \Spc$ classifying $X$ is a right Kan extension along the inclusion of the full subcategory $BG \subset \sO_G^{\op}$.
\end{dfn}
\begin{rem} \cref{dfn:BorelGSpace} is equivalent to the following condition on a $G$-space $X$: if we let $$X_H \coloneq \sO^{\op}_H \times_{(\ind^G_H)^{\op},\sO^{\op}_G} X$$ denote the restriction of $X$ to an $H$-space, then for every subgroup $H \leq G$, the natural map
\[ X_{G/H} \simeq \Map^{\cocart}_{/\sO^{\op}_H}(\sO^{\op}_H,X_H) \to \Map_{/BH}(BH,BH \times_{\sO^{\op}_H} X_H) \simeq (X_{G/1})^{h H} \]
is an equivalence.
\end{rem}
\begin{rem} Limits and coproducts of Borel $G$-spaces are Borel. Moreover, for every $G$-set $U$, the $G$-space $\underline{U}$ is Borel. Indeed, this amounts to the observation that $\Hom_G(G/H,U) \cong U^H$ for all subgroups $H \leq G$. In particular, since representables are Borel, the Borel property is stable under passage to $G$-fibers.
\end{rem}
\begin{dfn} Let $f: X \to Y$ be a map of Borel $G$-spaces and let $f_0: X_0 \to Y_0$ denote the underlying map of spaces. Then by \cref{lem:TruncationPresheaves}, $f$ is $n$-truncated if and only if $f_0$ is $n$-truncated. Furthermore, because every $G$-orbit is a finite set, the following two conditions are equivalent:
\begin{enumerate}
\item For every $y \in Y_0$, the homotopy fiber $(X_0)_y$ is a finite $n$-type \cite[Def.~4.4.1]{HL13}.
\item For every $y \in Y$, the underlying space of the homotopy $G$-fiber $X_{\underline{y}}$ is a finite $n$-type.
\end{enumerate}
In this case, we say that $f$ is \emph{$\pi$-finite $n$-truncated}. Note that if $f$ is $\pi$-finite $n$-truncated, then its diagonal is $\pi$-finite $(n-1)$-truncated; indeed, this property can be checked for the underlying spaces.
More generally, if $f:X \to Y$ is a map of $G$-spaces such that for every $y \in Y$, $X_{\ul{y}}$ is Borel, then we say that $f$ is \emph{($\pi$-finite) $n$-truncated} if the above conditions hold for all $y$ and $X_{\ul{y}}$.
\end{dfn}
\begin{lem} \label{lem:BorelAmbidex} Let $f: X \to Y$ be a map of $G$-spaces such that for all $y \in Y$, $X_{\ul{y}}$ is Borel.
\begin{enumerate}
\item Suppose that $C$ is $G$-pointed.
\begin{enumerate}
\item If $f$ is $(-1)$-truncated, then $f$ is $(-1)$-ambidextrous.
\item If $f$ is $0$-truncated, then $f$ is weakly $0$-ambidextrous.
\end{enumerate}
\item Suppose in addition that $C$ is $G$-semiadditive.
\begin{enumerate}
\item If $f$ is $\pi$-finite $0$-truncated, then $f$ is $0$-ambidextrous.
\item If $f$ is $\pi$-finite $1$-truncated, then $f$ is weakly $1$-ambidextrous.
\end{enumerate}
\end{enumerate}
\end{lem}
\begin{proof} By \cref{lem:AmbidexCheckedFiberwise} and under our hypothesis on the parametrized fibers, we may suppose in the proof that $Y = \underline{G/H}$ and $X$ is Borel. For (1), if $f$ is $(-1)$-truncated then the underlying space of $X$ is a discrete set that injects into $G/H$. But if $X$ is non-empty then $f$ must also be a surjective map of $G$-sets since the $G$-action on $G/H$ is transitive. Thus either $X = \emptyset$ or $f$ is an equivalence, so by \cref{lem:pointedAmbidex}, $f$ is $(-1)$-ambidextrous. If $f$ is $0$-truncated, then the $(-1)$-truncated diagonal $X \to X \times_Y X$ is $(-1)$-ambidextrous as just shown, so $f$ is weakly $0$-ambidextrous.
For (2), we employ the same strategy. If $f$ is $\pi$-finite $0$-truncated, then $X$ is necessarily a finite $G$-set, so $f$ is $0$-ambidextrous by hypothesis. If $f$ is $\pi$-finite $1$-truncated, then the diagonal $X \to X \times_Y X$ is $\pi$-finite $0$-truncated and hence $0$-ambidextrous, so $f$ is weakly $1$-ambidextrous.
\end{proof}
To apply the parametrized ambidexterity theory to our situation of interest, we need the following lemma.
\begin{lem} \label{lem:BorelClassifyingSpace} The $G/N$-space $B^{\psi}_{G/N} N$ of \cref{dfn:TwistedClassifyingSpace} is Borel.
\end{lem}
\begin{proof} For any subgroup $K/N$ of $G/N$, $(B^{\psi}_{G/N} N)_{K/N} \simeq (B^{\psi'}_{K/N} N)$ for $\psi' = [N \to K \to K/N]$. Therefore, without loss of generality it suffices to prove that the map of groupoids
\[ \chi: \Map^{\cocart}_{/\sO_{G/N}^{\op}}(\sO_{G/N}^{\op}, B^{\psi}_{G/N} N) \to \Map_{/B(G/N)}(B(G/N), (B^{\psi}_{G/N} N) \times_{\sO_{G/N}^{\op}} B(G/N) ) \]
is an equivalence. The fiber $E$ of $B^{\psi}_{G/N} N$ over the terminal $G/N$-set $\ast$ is spanned by those $N$-free $G$-orbits $U$ such that $U/N \cong \ast$, and an explicit inverse to the evaluation map
\[ \Map^{\cocart}_{/\sO_{G/N}^{\op}}(\sO_{G/N}^{\op}, B^{\psi}_{G/N} N) \xto{\simeq} E \]
is given by sending $U$ to the cocartesian section $s_U = (- \times U): \sO_{G/N}^{\op} \to B^{\psi}_{G/N} N$ that sends $V$ to $V \times U$: this follows from our identification of the cocartesian edges in \cref{lem:QuotientMapCartesianFibration}. Then
\[ \chi(s_U): B(G/N) \to (B^{\psi}_{G/N} N) \times_{\sO_{G/N}^{\op}} B(G/N) \] is the section which sends $G/N$ to the free transitive $G$-set $G/N \times U$. Let us now select a basepoint to identify $U \cong G/H$ for $H$ a subgroup such that $H \cap N = 1$ and $G = N H$. We have $W_G H \cong \Aut_G(G/H)$, where a coset $\overline{a} \in W_G H$ gives an automorphism $\theta_{\overline{a} }$ of $G/H$ that sends $1H$ to the well-defined coset $aH$, and under $\chi$ this is sent to the automorphism $\id \times \theta_{\overline{a} }$ of the section $\chi(s_{G/H})$.
By elementary group theory, each coset in $G/N$ has a unique representative $x N$ with $x \in H$, and each coset in $G/H$ has a unique representative $y H$ with $y \in N$. Moreover, the inclusion $N_G(H) \cap N \to N_G(H)$ yields an isomorphism $N_G(H) \cap N \cong W_G(H)$; the map is an injection because $N \cap H = 1$ and a surjection because $G = NH$. The two surjections $G/1 \to G/N$ and $G/1 \to G/H$ sending $1G$ to $1N$ and $1H$ define an isomorphism $G/1 \xto{\cong} G/N \times G/H$ for which an explicit inverse sends $(xN,yH)$ to $x \cdot y$. Under this isomorphism, $\id \times \theta_{\overline{a} }$ is sent to the unique element $a \in N_G(H) \cap N$ that is a representative for $\overline{a}$.
On the other hand, $(B^{\psi}_{G/N} N) \times_{\sO_{G/N}^{\op}} B(G/N) \simeq BG$ where we select $G/1$ to be the unique object of $BG$. We compute the groupoid of maps $\Map_{/B(G/N)}(B(G/N), BG)$ to have objects given by splittings $\tau: G/N \to G$ of the surjection $\pi: G \to G/N$ and morphisms $\tau \to \tau'$ given by $n \in N$ such that for every coset $bN$, $n \tau(b N) n^{-1} = \tau'(b N)$. In particular, $\Aut(\tau) = N \cap C_G(H)$ for $H = \tau(G/N)$. However, if $n h n^{-1} = h' \in H$, then $\pi(n h n^{-1} h^{-1}) = 1$ shows that $n h n^{-1} h^{-1} \in N \cap H =1$, so in fact $n h = h n$ and thus $\Aut(\tau) = N \cap N_G(H)$. Combining this with the explicit understanding of the comparison map given above, we deduce that $\chi$ is fully faithful. Essential surjectivity is also clear by the bijection between splittings of $\pi$ and subgroups $H$ with $N \cap H = 1$ and $G = N H$. We conclude that $\chi$ is an equivalence.
\end{proof}
\subsection{The parametrized Tate construction}\label{SS:PTate}
In view of \cref{lem:BorelClassifyingSpace}, we may define the parametrized Tate construction (\cref{dfn:ParamTateCnstr}) by applying the parametrized ambidexterity theory to the Beck-Chevalley fibration $$\LocSys^{G/N}(\underline{\Sp}^{G/N}) \to \Spc^{G/N}.$$ Let $\rho_N: B^{\psi}_{G/N} N \to \sO_{G/N}^{\op}$ be the structure map as in \cref{lm:TwistedClassifyingSpaceIsSpace}, and let
\[ {\rho_N}^{\ast}: \Sp^{G/N} \simeq \Fun_{G/N}(\sO_{G/N}^{\op},\underline{\Sp}^{G/N}) \to \Fun_{G/N}(B^{\psi}_{G/N} N, \underline{\Sp}^{G/N}) \]
be the functor given by restriction along $\rho_N$. We first introduce alternative notation for its left and right adjoints $(\rho_N)_!$ and $(\rho_N)_{\ast}$.
\begin{ntn} Given $X \in \Fun_{G/N}(B^{\psi}_{G/N} N, \underline{\Sp}^{G/N})$, we write
\begin{align*} X_{h[\psi]} \coloneq (\rho_N)_! (X) \text{ and } X^{h[\psi]} \coloneq (\rho_N)_{\ast} X
\end{align*}
for the parametrized homotopy orbits and fixed points functors, respectively.
\end{ntn}
\begin{dfn} \label{dfn:ParamTateCnstr} The $G/N$-functor $\rho_N$ has as its underlying map of spaces $B N \to \ast$, which is $\pi$-finite $1$-truncated. By \cref{lem:BorelAmbidex}, $\rho_N$ is weakly $1$-ambidextrous, so we can construct the norm map $\Nm_{\rho_N}: (\rho_N)_! \to (\rho_N)_{\ast}$. Let
\[ (-)^{t[\psi]}: \Fun_{G/N}(B^{\psi}_{G/N} N, \underline{\Sp}^{G/N}) \to \Sp^{G/N} \]
denote the cofiber of $\Nm_{\rho_N}$.
\end{dfn}
\begin{ntn} When the group extension $\psi$ is clear from context, we will also write
\[ X_{h_{G/N} N} = X_{h[\psi]} \:, \quad X^{h_{G/N} N} = X^{h[\psi]} \:, \quad X^{t_{G/N} N} = X^{t[\psi]}. \]
\end{ntn}
\begin{obs}[The norm vanishes on induced objects] \label{NormVanishesOnInduced} For $H \in \Gamma_N$ (so that $H \cap N = 1$), let $U = \rho_N(G/H) \cong \frac{G/N}{H N/N}$ be the $G/N$-orbit and $s_H: \underline{U} \to B^{\psi}_{G/N} N$ be the unique $G/N$-functor that selects $G/H$, so that the functor $s_H^{\ast}$ of \cref{evaluationFactorizationNaiveSpectra} is obtained by restriction along $s_H$. Note that the map of Borel $G/N$-spaces $s_H$ is $\pi$-finite $0$-truncated because its underlying map of spaces is $U \to B N$ with $U$ a finite discrete set. By \cref{lem:BorelAmbidex}, we see that $\Nm_{s_H}: {s_H}_! \xto{\simeq} (s_H)_{\ast}$.
Now consider the composite map $p_U = \rho_N \circ (s_H)$, which is also $\pi$-finite $0$-truncated. On the one hand, the associated norm map $\Nm_{p_U}$ is an equivalence (explicitly, between induction and coinduction from $\Sp^H$ to $\Sp^{G/N}$ for $H \cong H N/N$ viewed as a subgroup of $G/N$). On the other hand, by \cite[Rem.~4.2.4]{HL13}, we have that $\Nm_{p_U}$ is homotopic to the composite $((\rho_N)_{\ast} \Nm_{s_H}) \circ (\Nm_{\rho_N} (s_H)_!)$. We deduce that $\Nm_{\rho_N}$ is an equivalence on the image of $(s_H)_!$. This extends the observation that the ordinary norm map $X_{h G} \to X^{h G}$ is an equivalence on objects induced from $\Sp$ to $\Fun(B G, \Sp)$.
\end{obs}
By \cref{thm:BorelSpectraAsCompleteObjects}, we also have a norm map $\Nm': \sF_b[N] \to \sF^{\vee}_b[N]$ arising from the $\Gamma_N$-recollement of $\Sp^G$, with functors $\sF_b[N]$ and $\sF^{\vee}_b[N]$ as in \cref{VariousPropertiesForgetfulFunctorBorelSpectra}. We now proceed to show that the two norm maps $\Psi^N \Nm'$ and $\Nm_{\rho_N}$ are equivalent.
\begin{lem} \label{lm:IdentifyingDiagonalAsComposition} The functor $\rho_N^{\ast}$ is homotopic to the composite
\[ \sU_b[N] \circ {\inf}^N: \Sp^{G/N} \to \Sp^G \to \Fun_{G/N}(B^{\psi}_{G/N} N, \underline{\Sp}^{G/N}). \]
\end{lem}
\begin{proof} For the proof, we work in the setup of \cref{cnstr:ForgetfulFunctorToNaiveGSpectra}. Let
\[ \underline{\inf}'[N]: \underline{\Sp}^{G/N} \to \sO_{G/N}^{\op} \times_{i_N^{\op},\sO_G^{\op}} \underline{\Sp}^G \]
be the $G/N$-functor defined by unstraightening the natural transformation
$$\SH q_N^{\op}: \SH \omega_{G/N}^{\op} \to \SH \omega_G^{\op} \iota_N^{\op},$$
so for a $G/N$ orbit $V = \frac{G/N}{K/N}$, the fiber $\underline{\inf}'[N]_V: \Sp^{K/N} \to \Sp^K$ is given by the inflation functor $\inf^N$. By definition, the composite
\[ \begin{tikzcd}[row sep=4ex, column sep=6ex, text height=1.5ex, text depth=0.5ex]
\underline{\Sp}^{G/N} \ar{r}{\underline{\inf}'[N]} & \sO_{G/N}^{\op} \times_{i_N^{\op},\sO^{\op}_G} \Sp^G \ar{r}{\widetilde{\sU}[N]} & \underline{\Fun}_{G/N}(\sO_G^{\op}, \underline{\Sp}^{G/N})
\end{tikzcd} \]
is adjoint to the composite (abusively denoting $\underline{\inf}'[N]$ for its pullback along $r_N^{\op}$)
\[ \begin{tikzcd}[row sep=4ex, column sep=6ex, text height=1.5ex, text depth=0.5ex]
\sO^{\op}_G \times_{r_N^{\op},\sO_{G/N}^{\op}} \underline{\Sp}^{G/N} \ar{r}{\underline{\inf}'[N]} & \sO_G^{\op} \times_{(i_N r_N)^{\op}, \sO_G^{\op}} \underline{\Sp}^G \ar{r}{\underline{\res}[N]} & \underline{\Sp}^G \ar{r}{\widehat{\Psi}[N]} & \sO^{\op}_G \times_{r_N^{\op},\sO_{G/N}^{\op}} \underline{\Sp}^{G/N} \ar{r}{\pr} & \underline{\Sp}^{G/N}.
\end{tikzcd} \]
For a $G$-orbit $G/H$, the fiber of $\widehat{\Psi}[N] \circ \underline{\res}[N] \circ \underline{\inf}'[N]$ over $G/H$ is given by the composition
\[ \begin{tikzcd}[row sep=4ex, column sep=6ex, text height=1.5ex, text depth=0.5ex]
\Sp^{H N/N} \ar{r}{\inf^N} & \Sp^{H N} \ar{r}{\res^{H N}_H} & \Sp^H \ar{r}{\Psi^{H \cap N}} & \Sp^{H/(H \cap N)}.
\end{tikzcd} \]
Using that $H N/N \cong H/(H \cap N)$, the composition $\res^{H N}_H \circ \inf^N$ is homotopic to $\inf^{H \cap N}$. Therefore, if $H \in \Gamma_N$ so that $H \cap N = 1$, the entire composite is trivial. We deduce that the composite
\[ \begin{tikzcd}[row sep=4ex, column sep=6ex, text height=1.5ex, text depth=0.5ex]
B^{\psi}_{G/N} N \times_{\rho_N,\sO_{G/N}^{\op}} \underline{\Sp}^{G/N} \ar{r}{\underline{\inf}'[N]} & B^{\psi}_{G/N} N \times_{\rho_N, \sO_G^{\op}} \underline{\Sp}^G \ar{r}{\underline{\res}[N]} & \underline{\Sp}^G \ar{r}{\widehat{\Psi}[N]} & B^{\psi}_{G/N} N \times_{\rho_N, \sO_{G/N}^{\op}} \underline{\Sp}^{G/N}
\end{tikzcd} \]
is homotopic to the identity, which proves the claim.
\end{proof}
\begin{obs} Passing to right adjoints in \cref{lm:IdentifyingDiagonalAsComposition} we get that $(\rho_N)_{\ast} \simeq \Psi^N \sF^{\vee}_b[N]$. Let
$$(-)^{t'[\psi]} : \Fun_{G/N}(B^{\psi}_{G/N} N, \underline{\Sp}^{G/N}) \to \Sp^{G/N}$$
be the cofiber of $\Psi^N \Nm'$. Since the orbits $G/H_+ \in \Sp^G$ for $H \in \Gamma_N$ are both $\Gamma_N$-torsion and $\Gamma_N$-complete, $\Nm' \circ (s_H)_!(1)$ is an equivalence for all $H \in \Gamma_N$. Therefore, $(-)^{t'[\psi]}$ vanishes on each $(s_H)_!(1)$. Because $\{ (s_H)_!(1): H \in \Gamma_N \}$ is a set of compact generators for $\Fun_{G/N}(B^{\psi}_{G/N} N, \underline{\Sp}^{G/N})$ and $(\rho_N)_! $ is a colimit preserving functor, the composite
\[ (-)_{h[\psi]} = (\rho_N)_! \xtolong{\Nm_{\rho_N}}{1} (-)^{h[\psi]} = (\rho_N)_{\ast} \simeq \Psi^N \sF^{\vee}_b[N] \to (-)^{t'[\psi]} \]
is null-homotopic. We thereby obtain a natural transformation $\nu: (-)^{t[\psi]} \to (-)^{t'[\psi]}$. Taking fibers, we also have a natural transformation $\mu: (\rho_N)_! \to \Psi^N \sF_b[N]$. All together, for all $X \in \Sp^G_{\Borel{N}}$ we have a morphism of fiber sequences of $G/N$-spectra
\[ \begin{tikzcd}[row sep=4ex, column sep=8ex, text height=1.5ex, text depth=0.5ex]
X_{h[\psi]} \ar{r}{(\Nm_{\rho_N})_X} \ar{d}{\mu_X} & X^{h[\psi]} \ar{r} \ar{d}{\simeq} & X^{t[\psi]} \ar{d}{\nu_X} \\
\Psi^N \sF_b[N] (X) \ar{r}{\Psi^N \Nm'_X} & \Psi^N \sF_b^{\vee}[N](X) \ar{r} & X^{t'[\psi]}.
\end{tikzcd} \]
\end{obs}
\begin{thm} \label{thm:EquivalentTateConstructions} The natural transformations $\mu$ and $\nu$ are equivalences.
\end{thm}
\begin{proof} It suffices to show that $\mu$ is an equivalence. By \cref{NormVanishesOnInduced}, $\Nm_{\rho_N}$ is an equivalence on $(s_H)_!(1)$ for every $H \in \Gamma_N$, and we just saw the same property for $\Nm'$. Therefore, $\mu$ is an equivalence on each $(s_H)_!(1)$ by the two-out-of-three property of equivalences. Since both $(\rho_N)_!$ and $\Psi^N \sF_b[N]$ preserve colimits and the $(s_H)_!(1)$ form a set of compact generators, we conclude that $\mu$ is an equivalence.
\end{proof}
\begin{rem}[$\infty$-categorical Adams isomorphism] \label{AdamsIsomorphism} By \cref{thm:EquivalentTateConstructions}, for $X \in \Sp^G_{\Borel{N}}$, we have an equivalence of $G/N$-spectra $X_{h[\psi]} \simeq \Psi^N \sF_b[N] (X)$. Viewing $X$ as an `$N$-free' $G$-spectrum via the embedding $\sF_b[N]$, this amounts to the Adams isomorphism for a normal subgroup $N$ of a \emph{finite} group $G$ in our context (compare \cite[Ch.~XVI, Thm.~5.4]{AlaskaNotes}).
We also note that Sanders \cite{SandersCompactnessLocus} has recently introduced a different formal framework for producing the Adams isomorphism, in the more general situation of a closed normal subgroup of a compact Lie group. It would be interesting to understand the relationship between his results and ours.
\end{rem}
\begin{rem} In view of \cref{thm:EquivalentTateConstructions}, we could have defined the parametrized Tate construction as $(-)^{t'[\psi]}$ to begin with. However, we still need the Adams isomorphism to identify the fiber term $\Psi^N \sF_b[N]$ of $(\rho_N)_{\ast} \to (-)^{t'[\psi]}$ as the parametrized orbits functor $(\rho_N)_!$.
\end{rem}
\begin{rem}[Point-set models] \label{rem:PointSetModels} Let $X \in \Sp^G_{\Borel{N}}$ and consider the fiber sequence of $G/N$-spectra
\[ X_{h[\psi]} \to X^{h[\psi]} \to X^{t[\psi]}. \]
By \cref{thm:EquivalentTateConstructions} and the monoidal recollement theory for $\Gamma_N$, this fiber sequence is obtained as $\Psi^N$ of the fiber sequence of $G$-spectra
\[ \sF_b^{\vee}[N](X) \otimes {E \Gamma_N}_+ \to \sF_b^{\vee}[N](X) \to \sF_b^{\vee}[N](X) \otimes \widetilde{E \Gamma_N}. \]
If we let $X = \sU_b[N](Y)$ for $Y \in \Sp^G$, then we may also write this as
\[ Y \otimes {E \Gamma_N}_+ \simeq F( {E \Gamma_N}_+, Y) \otimes {E \Gamma_N}_+ \to F( {E \Gamma_N}_+, Y) \to F( {E \Gamma_N}_+, Y) \otimes \widetilde{E \Gamma_N}. \]
\end{rem}
\begin{rem}[Parametrized Tate spectral sequence]\label{rem:TateSS}
The point-set model for the parametrized Tate construction given in \cref{rem:PointSetModels} allows us to apply \cite[Thm.~22.6]{GM95} to define the \emph{parametrized Tate spectral sequence}
$$E^2_{p,q} = \widehat{H}^p_{\Gamma_N}(\pi_q(E)) \Rightarrow \pi_{q-p} E^{t_{G/N}N}$$
to study the $G/N$-parametrized $N$-Tate construction of $E \in \Sp^G_{N\text{-Borel}}$. The $E_2$-term is given by the \emph{$\Gamma_N$-Amitsur--Dress--Tate cohomology} of $N$ with coefficients in $\pi_*(E)$ as defined in \cite[Def.~21.1]{GM95}.\footnote{Here, we recall that coefficients for $\cF$-Amitsur--Dress--Tate cohomology with respect to a $G$-family $\cF$ are given by abelian group-valued presheaves on $\sO_{G,\cF}$, the full subcategory of $\sO_G$ on those orbits with stabilizer in $\cF$. Then for a $G/N$-functor $E: B^{\psi}_{G/N} N \to \underline{\Sp}^{G/N}$, under the equivalence $\Fun_{G/N}(B^{\psi}_{G/N} N, \underline{\Spc}^{G}) \simeq \Fun(B^{\psi}_{G/N} N, \Spc)$ of \cite[Prop.~3.10]{Exp2}, we have that $\pi_q(E) = \pi_0(\Omega^{\infty} \Sigma^{-q} E)$ renders as suitable input.} See also \cite[\S\S2.3, 3.2]{mathew2019}.
\end{rem}
\begin{obs}[Compatibility with restriction] \label{rem:ParamTateCompatibleRestriction} Note that the norm map $\Nm'$ also extends to a natural transformation of $G$-functors
\[ \left( \underline{\Nm'}: \underline{\sF}_b[N] \Rightarrow \underline{\sF}^{\vee}_b[N] \right): \underline{\Sp}^G_{\Borel{N}} \to \underline{\Sp}^G. \]
Postcomposing with the functor $\widehat{\Psi}[N]: \underline{\Sp}^G \to \sO_G^{\op} \times_{\sO_{G/N}^{\op}} \underline{\Sp}^{G/N}$ defined in \cref{cnstr:ForgetfulFunctorToNaiveGSpectra} and taking the cofiber, we may extend $(-)^{t[\psi]}$ to a functor over $\sO^{\op}_G$
\[ (-)^{\widehat{t}[\psi]}: \underline{\Sp}^G_{\Borel{N}} \to \sO_G^{\op} \times_{\sO_{G/N}^{\op}} \underline{\Sp}^{G/N} \]
that over an orbit $G/H$ is given by
\[ (-)^{t[\psi_H]}: \Fun_{H/(N \cap H)}(B^{\psi_H}_{H/(N \cap N)} (N \cap H), \underline{\Sp}^{H/(N \cap H)}) \to \Sp^{H/(N \cap H)}. \]
However, because $\underline{\Psi}[N]$ is not typically a $G$-functor, $(-)^{\widehat{t}[\psi]}$ may also fail to be a $G$-functor. If instead we precompose by the inclusion
\[ \underline{\Fun}_{G/N}(B^{\psi}_{G/N} N, \underline{\Sp}^{G/N}) \simeq \sO_{G/N}^{\op} \times_{i_N^{\op},\sO_G^{\op}} \underline{\Sp}^G_{\Borel{N}} \to \underline{\Sp}^G_{\Borel{N}} \]
and postcompose by the projection to $\underline{\Sp}^{G/N}$, then we obtain a $G/N$-functor
\[ (-)^{\underline{t}[\psi]}: \underline{\Fun}_{G/N}(B^{\psi}_{G/N} N, \underline{\Sp}^{G/N}) \to \underline{\Sp}^{G/N}. \]
By checking fiberwise, it is easy to verify that $(-)^{\underline{t}[\psi]}$ is equivalent to the cofiber of the norm map $\Nm_{\rho_N}: (\underline{\rho_N})_! \to (\underline{\rho_N})_{\ast}$ produced by the ambidexterity theory for the other Beck-Chevalley fibration
$$\underline{\LocSys}^{G/N}(\underline{\Sp}^{G/N}) \to \Spc^{G/N}.$$
In any case, we obtain a compatibility between the parametrized Tate construction and restriction. For example, given a $G/N$-functor $X: B^{\psi}_{G/N} N \to \underline{\Sp}^{G/N}$, we see that the underlying spectrum of $X^{t[\psi]}$ is $X^{tN} \coloneq (\res^{G/N} X)^{t N}$ for the underlying functor $\res^{G/N} X: B N \to \Sp$.
\end{obs}
A useful consequence of \cref{thm:EquivalentTateConstructions} is that it enables us to endow the functor $(-)^{t[\psi]}$ and the natural transformation $(-)^{h[\psi]} \to (-)^{t[\psi]}$ with lax symmetric monoidal structures, with respect to the pointwise symmetric monoidal structure on $\Fun_{G/N}(B^{\psi}_{G/N} N, \underline{\Sp}^{G/N})$ and the smash product on $\Sp^{G/N}$.
\begin{cor} \label{cor:TateLaxMonoidalStructure} The functor $(-)^{t[\psi]}$ and the natural transformation $(-)^{h[\psi]} \to (-)^{t[\psi]}$ are lax symmetric monoidal.
\end{cor}
\begin{proof} The cofiber of $\Nm'$ is the lax symmetric monoidal map $j_{\ast} \to i_{\ast} i^{\ast} j_{\ast}$ of the $\Gamma_N$-recollement of $\Sp^G$, where we use \cref{rem:MonoidalIdentificationOfBorelSpectra} to relate the pointwise symmetric monoidal structure on the domain $\Fun_{G/N}(B^{\psi}_{G/N} N, \underline{\Sp}^{G/N})$ to the symmetric monoidal recollement. Since the categorical fixed points functor $\Psi^N$ is also lax symmetric monoidal, we deduce that $\Psi^N \sF^{\vee}_b[N](-) \to (-)^{t'[\psi]}$ is lax symmetric monoidal. The conclusion now follows from \cref{thm:EquivalentTateConstructions}.
\end{proof}
On the other hand, one practical benefit of defining the parametrized Tate construction via the ambidexterity theory is that we may exploit the general naturality properties of norms as detailed in \cite[\S4.2]{HL13}. To state our next result, which involves two normal subgroups $M \trianglelefteq N$ of $G$, we first require a preparatory lemma.
\begin{lem} \label{lm:CategoricalFixedPointsProperties} Let $M \trianglelefteq N \trianglelefteq G$ be two normal subgroups of $G$ and let
\[ \psi = [N \to G \to G/N], \quad \psi' = [N/M \to G/M \to G/N] \]
denote the extensions.
\begin{enumerate}[leftmargin=6ex]
\item $\Psi^M: \Sp^G \to \Sp^{G/M}$ sends $\Gamma_N$-torsion spectra to $\Gamma_{N/M}$-torsion spectra.
\item Let $r_M: \FF_G \to \FF_{G/M}$, $r_M(U) = U/M$ be as in \cref{inflationFunctors}, and regard $\FF_G$, $\FF_{G/M}$ as cartesian fibrations over $\FF_{G/N}$ via $r_N$, $r_{N/M}$ respectively. Then $r_M$ preserves cartesian edges, so the restricted functor $r_M^{\op}: \sO_G^{\op} \to \sO_{G/M}^{\op}$ is a $G/N$-functor. Moreover, $r_M^{\op}$ further restricts to a $G/N$-functor $$\rho_M: B^{\psi}_{G/N} N \to B^{\psi'}_{G/N} N/M.$$
\item We have a commutative diagram
\[ \begin{tikzcd}[row sep=4ex, column sep=8ex, text height=1.5ex, text depth=0.5ex]
\Sp^G_{\Borel{N}} = \Fun_{G/N}(B^{\psi}_{G/N} N, \underline{\Sp}^{G/N}) & \Sp^G \ar{l}{\sU_b[N]} \\
\Sp^{G/M}_{\Borel{N/M}} = \Fun_{G/N}(B^{\psi'}_{G/N} (N/M), \underline{\Sp}^{G/N}) \ar{u}{(\rho_M)^{\ast}} & \Sp^{G/M} \ar{u}[swap]{\inf^G_{G/M}} \ar{l}{\sU_b[N/M]}
\end{tikzcd} \]
that yields a commutative diagram of right adjoints
\[ \begin{tikzcd}[row sep=4ex, column sep=8ex, text height=1.5ex, text depth=0.5ex]
\Sp^G_{\Borel{N}} = \Fun_{G/N}(B^{\psi}_{G/N} N, \underline{\Sp}^{G/N}) \ar{r}{\sF_b^{\vee}[N]} \ar{d}[swap]{(\rho_M)_{\ast}} & \Sp^G \ar{d}{\Psi^M} \\
\Sp^{G/M}_{\Borel{N/M}} = \Fun_{G/N}(B^{\psi'}_{G/N} (N/M), \underline{\Sp}^{G/N}) \ar{r}{\sF_b^{\vee}[N/M]} & \Sp^{G/M}
\end{tikzcd} \]
where the lefthand vertical functor is computed by the $G/N$-right Kan extension along $r^{\op}_M$.
\end{enumerate}
\end{lem}
\begin{proof} For (1), note that if $G/H$ is a $N$-free $G$-orbit, then $G/H$ is also $M$-free. Thus, we may compute $\Psi^M(G/H_+)$ as by taking the quotient by the $M$-action to obtain $\frac{G/M}{H M/M}_+$, which is $N/M$-free and thus $\Gamma_{N/M}$-torsion. Because the subcategory of $\Gamma_N$-torsion spectra is the localizing subcategory generated by such $G/H_+$ and $\Psi^M$ preserves colimits, the statement follows. (2) is a direct consequence of \cref{lem:QuotientMapCartesianFibration}(2). (3) is a relative version of \cref{lm:IdentifyingDiagonalAsComposition}, and also follows by an elementary diagram chase after unpacking the various definitions.
\end{proof}
Now suppose $G$ is a semidirect product of $N$ and $G/N$, so we have chosen a splitting $G/N \to G$ of the quotient map such that $G \cong N \rtimes G/N$, and with respect to the $G/N$-action on $N$, the inclusion $M \subset N$ is $G/N$-equivariant. Then $M \rtimes G/N$ is a subgroup of $G$, and we let $$\psi''=[M \to M \rtimes G/N \to G/N].$$ Also regard $B^{\psi'}_{G/N} N/M$ as a based $G/N$-space via the splitting. Then we have a homotopy pullback square of $G/N$-spaces
\[ \begin{tikzcd}[row sep=4ex, column sep=4ex, text height=1.5ex, text depth=0.25ex]
B^{\psi''}_{G/N} M \ar{r} \ar{d}{\rho_M} & B^{\psi}_{G/N} N \ar{d}{\rho_M} \\
\sO_{G/N}^{\op} \ar{r} & B^{\psi'}_{G/N} N/M
\end{tikzcd} \]
that arises from the fiber sequence $B M \to BN \to B N/M$ of spaces with $G/N$-action.
\begin{prp} \label{prp:ResidualAction} Suppose $X \in \Sp^{G}_{\Borel{N}}$ and also write $X$ for its restriction to $\Sp^G_{\Borel{M}}$. Then $X^{t[\psi'']}$ canonically acquires a `residual action' by lifting to an object in $\Sp^{G}_{\Borel{N/M}}$, and we have a fiber sequence of $G/N$-spectra
\[ (X_{h[\psi'']})^{t[\psi']} \to X^{t[\psi]} \to (X^{t[\psi'']})^{h[\psi']} \]
that restricts to a fiber sequence of Borel $G/N$-spectra
\[ (X_{h M})^{t(N/M)} \to X^{tN} \to (X^{t M})^{h(N/M)}. \]
\end{prp}
\begin{proof} We apply \cite[Rem.~4.2.3]{HL13} to the pullback square above to deduce the first assertion. For the second assertion, we apply \cite[Rem.~4.2.4]{HL13} to the factorization of $\rho_N$ as $\rho_{N/M} \circ \rho_M$ to obtain a commutative diagram of $G/N$-spectra
\[ \begin{tikzcd}[row sep=6ex, column sep=12ex, text height=1.5ex, text depth=1ex]
X_{h[\psi]} \ar{r}{\Nm_{\rho_{N/M}} \circ (\rho_M)_!} \ar{d} & (X_{h[\psi'']})^{h[\psi']} \ar{r}{(\rho_{N/M})_{\ast} \circ \Nm_{\rho_M}} \ar{d} & X^{h[\psi]} \ar{d} \\
0 \ar{r} & (X_{h[\psi'']})^{t[\psi']} \ar{r} \ar{d} & X^{t[\psi]} \ar{d} \\
& 0 \ar{r} & (X^{t[\psi'']})^{h[\psi']}
\end{tikzcd} \]
in which every rectangle is a homotopy pushout (using the two-out-of-three property of homotopy pushouts to show this for the upper righthand square and then the lower righthand square). The final assertion follows from \cref{rem:ParamTateCompatibleRestriction}.
\end{proof}
\subsection{Inverse limit formula for the parametrized Tate construction}
We prove that the parametrized Tate construction can be expressed as a limit involving certain Thom spectra, generalizing \cite[Thm. 16.1]{GM95}. This identification is used to prove parametrized Tate blueshift in \cite{LLQ19} and to define $C_2$-equivariant Mahowald invariants in \cite{Qui19b}.
\begin{lem}\label{Lem:VRestrictions}
Let $\cF$ be a family of subgroups of $G$. Suppose $V$ be a $G$-representation with $V^H = 0$ for $H \notin \cF$ and $V^H \neq 0$ for $H \in \cF$. Then $S(V^{\oplus \infty}) \simeq E\cF$ and $S^{\infty V} \simeq \widetilde{E\cF}$.
\end{lem}
The following lemma generalizes \cite[Lem.~2.6]{GM95}.
\begin{lem}\label{Lem:GM26}
There is an equivalence of $G$-spectra
$$\widetilde{E\cF} \otimes F(E\cF_+,X) \simeq F(\widetilde{E\cF}, E\cF_+ \otimes \Sigma X).$$
\end{lem}
\begin{proof}
We may identify
$$\widetilde{E\cF} \otimes F(E\cF_+,X) \simeq i_*i^*j_*j^*(X),$$
$$F(\widetilde{E\cF},E\cF_+ \otimes \Sigma X) \simeq i_*i^!j_!j^*(\Sigma X).$$
Writing $X = [U, Z \to \phi(U)]$ for the recollement decomposition of $X$, we find that $i_*i^*j_*j^*(X) = [0, \phi(U) \xrightarrow{=} \phi(U)]$. We also have $j_! j^* \Sigma X = [\Sigma U, 0 \to \Sigma \Phi(U)]$, so $i_*i^!j_!j^*(\Sigma X) = [0, \Phi(U) \xrightarrow{=} \Phi(U)]$ as desired.
\end{proof}
\begin{thm}\label{Thm:GM161}
Let $\psi = [N \to G \to G/N]$ be an extension with $N \subseteq H$ for each $H \notin \Gamma_N$. Suppose $V$ is as in \cref{Lem:VRestrictions} for $\cF = \Gamma_N$. For any $X \in \Sp^{\Phi \Gamma_N} \simeq \Sp^{G/N}$, we have
$$(j^* i_* X)^{t[\psi]} \simeq \lim_n ({B_{G/N}^{\psi} N}^{-nV} \otimes \Sigma X).$$
\end{thm}
\begin{proof}
We compute:
\begin{align*} F
(j^* i_* X)^{t[\psi]} & \simeq \Psi^N(\widetilde{E\Gamma_N} \otimes F({E\Gamma_N}_+, j^*i_* X)) & (\cref{rem:PointSetModels}) \\
& \simeq \Psi^N(F(\widetilde{E\Gamma_N}, {E\Gamma_N}_+ \otimes \Sigma j^*i_* X)) & (\cref{Lem:GM26}) \\
& \simeq \Psi^N(F(S^{\infty V}, {E\Gamma_N}_+ \otimes \Sigma j^*i_* X)) & (\cref{Lem:VRestrictions}) \\
& \simeq \Psi^N(F(\lim_n S^{nV}, {E\Gamma_N}_+ \otimes \Sigma j^*i_* X)) & \\
& \simeq \Psi^N( \lim_n F(S^{nV}, {E\Gamma_N}_+ \otimes \Sigma j^*i_* X)) & \\
& \simeq \Psi^N( \lim_n S^{-nV} \otimes {E\Gamma_N}_+ \otimes \Sigma j^*i_* X) & \\
& \simeq \lim_n (\Psi^N(S^{-nV} \otimes {E\Gamma_N}_+) \otimes \Psi^N(\Sigma j^*i_* X)) & \\
& \simeq \lim_n ((S^{-nV})_{h_{G/N}N} \otimes \Sigma X) & (\cref{thm:EquivalentTateConstructions}) \\
& \simeq \lim_n ({B_{G/N}^{\psi} N}^{-nV} \otimes \Sigma X). &
\end{align*}
\end{proof}
\begin{rem}
Although we have restricted to finite $G$ in this section, the results of the next section can be used to show that the evident analogue of \cref{Thm:GM161} holds for $G$ a compact Lie group. This agrees with the level of generality in the non-parametrized result of Greenlees--May \cite[Thm.~16.1]{GM95}.
\end{rem}
\section{Parametrized assembly}\label{Sec:Assembly}
Suppose $K$ is a compact Lie group. Then if $K$ is infinite, the Tate construction $(-)^{t K}$ cannot be defined via the Hopkins-Lurie ambidexterity theory since $B K$ is not $\pi$-finite. Instead, one defines $(-)^{t K}$ to be the cofiber of an \emph{assembly map}
\[ (\SS^{\mathfrak{a}} \otimes -)_{h K} \to (-)^{h K} \]
where $\SS^{\mathfrak{a}}$ is $\Sigma^{\infty}$ of the one-point compactification of the adjoint representation of $K$ (cf. \cite{Klein2001} and \cite[\S I.4]{NS18}). This construction is of particular interest when $K = S^1$. For example, given a cyclotomic spectrum $X$ (e.g., $\THH$ of a ring spectrum $A$), one defines the \emph{topological periodic homology} $\TP(X)$ to be $X^{t S^1}$, and this term (or rather, its profinite completion) participates in a fiber sequence computing the topological cyclic homology $\TC(X)$.
Since we are ultimately interested in applications to trace methods for real algebraic K-theory, we are thus motivated to develop a parametrized refinement of the above picture. More precisely, let $G$ be a finite group and $\psi: G \to \Aut(K)$ a group homomorphism (where $\Aut(K)$ denotes the group of continuous automorphisms of $K$).
\begin{dfn}
The $G$-space $B^{\psi}_G K$ is the Borel $G$-space obtained via right Kan extension along $BG \subset \sO_G^{\op}$ of $B K$ regarded as a space with $G$-action via $\psi$.
\end{dfn}
\begin{rem}
By \cref{lem:BorelClassifyingSpace}, this is consistent with our earlier definition of $B^{\psi}_G K$ when $K$ is finite.
\end{rem}
\begin{wrn}
The role of $G$ in this section is the same as its role in the introduction, but different from its role in the previous two sections. In particular, $G$ now plays the role of the quotient group $G = \widehat{G}/K$. Again, this is to emphasize that $G$ is finite; the notation for the normal subgroup has changed from $N$ to $K$ to emphasize that $K$ can be infinite, while $N$ was always finite.
\end{wrn}
In this section, we will construct a \emph{parametrized assembly map} (\cref{thm:paramTateGeneral} and \cref{prp:DualizingSpectrum})
\[ (\SS^{\mathfrak{a}} \otimes -)_{h_G K} \to (-)^{h_G K} \]
and then define the parametrized Tate construction $(-)^{t_G K}$ to be its cofiber (\cref{dfn:paramTateGeneral}). We unpack the example of $K = S^1$ with $G= C_2$ acting by complex conjugation as \cref{exm:CircleTate}. We will also equip $(-)^{t_G K}$ with a lax $G$-symmetric monoidal structure and thereby show that it preserves $G$-commutative algebras (\cref{cor:LaxGSymmetricMonoidalTate}); when $K$ is finite, this improves upon \cref{cor:TateLaxMonoidalStructure}.
We will use the following notation heavily in this section. Recall the notation $\underline{U}$ for a finite $G$-set $U$ from \cref{exm:corepresentableDiagrams} and let $\Cat_{\infty, \underline{U}}$ denote the $\infty$-category of $\underline{U}$-$\infty$-categories, i.e., cocartesian fibrations $C \to \underline{U}$.
\begin{ntn}
Let $f: U \to V$ be a map of finite $G$-sets and consider the pullback functor
\[ f^*: \Cat_{\infty,\underline{V}} \to \Cat_{\infty, \underline{U}} \:, \quad C \mapsto C_{\underline{U}}. \]
We write
\[ \begin{tikzcd}[column sep=4em]
\Cat_{\infty, \underline{U}} \ar[shift left = 3]{r}{f_!} \ar[shift right = 3]{r}[swap]{f_*} & \Cat_{\infty, \underline{V}} \ar{l}[description]{f^*}
\end{tikzcd} \]
for the adjoint triple.
\end{ntn}
\begin{rem} \label{rem:BaseChangeEquivalences}
Let
\[ \begin{tikzcd}
U' \ar{r}{f'} \ar{d}{g'} & V' \ar{d}{g} \\
U \ar{r}{f} & V
\end{tikzcd} \]
be a pullback square of finite $G$-sets. Then the exchange transformations
\[ g^* f_* \Rightarrow f'_* g'^*, \: f'_! g'^* \Rightarrow g^* f_!: \Cat_{\infty, \underline{U}} \to \Cat_{\infty, \underline{V'}} \]
are equivalences.
\end{rem}
\subsection{Recollections on \texorpdfstring{$G$}{G}-symmetric monoidal structures}\label{SS:GSM}
One of our main goals in this section is to explain how (for an extension $\psi$ of a finite group $G$ by a compact Lie group $K$) the parametrized Tate construction
\[ (-)^{t_G K}: \Fun_G(B^{\psi}_G K, \underline{\Sp}^G) \to \Sp^G \]
refines to a \emph{lax $G$-symmetric monoidal functor}. We begin by recalling the necessary terminology and concepts from the theory of $G$-symmetric monoidal $\infty$-categories, which will be explained in more detail in \cite{paramalg}.\footnote{Another basic reference is \cite{BachmannHoyoisNorms}, but note that there is some work involved in translating between their setup and ours.}
\begin{dfn} \label{dfn:GSMC}
A \emph{$G$-symmetric monoidal $\infty$-category} $C^\otimes$ is a cocartesian fibration over $\Span(\FF_G)$ whose straightening
\[ F_{C^\otimes}: \Span(\FF_G) \to \Cat_{\infty} \]
is a product-preserving functor. The \emph{underlying $G$-$\infty$-category} of $C^\otimes$ is the restriction $C = C^\otimes|_{\sO_G^{\op}}$ over the subcategory $\sO_G^{\op} \subset \Span(\FF_G)$. Conversely, given a $G$-$\infty$-category $C$, a \emph{$G$-symmetric monoidal structure} $C^\otimes$ on $C$ is such an extension of $C$ over $\Span(\FF_G)$. When the $G$-symmetric monoidal structure is understood, we refer to $C$ itself as a $G$-symmetric monoidal $\infty$-category.
For a map of finite $G$-sets $f: U \to V$, we define the (multiplicative) \emph{norm functor}
\[ f_{\otimes}: C^\otimes_U \to C^\otimes_V \]
as the pushforward functor associated to the span $[U \xot{=} U \xto{f} V]$.
A (strong) \emph{$G$-symmetric monoidal functor} $F: C^\otimes \to D^\otimes$ is a map of cocartesian fibrations over $\Span(\FF_G)$. If we only suppose that $F$ preserves cocartesian edges over $\FF_G^{\op}$, then $F$ is said to be \emph{lax $G$-symmetric monoidal}.
\end{dfn}
\begin{rem}
\cref{dfn:GSMC} is an $\infty$-categorical version of the $G$-symmetric monoidal categories introduced by Hill and Hopkins in \cite{hillhopkins}.
\end{rem}
Given a $G$-symmetric monoidal $\infty$-category $C^{\otimes}$, a \emph{$G$-commutative algebra} $A$ in $C^{\otimes}$ is a section $A^{\otimes}: \Span(\FF_G) \to C^{\otimes}$ whose restriction $A^{\otimes}|_{\FF^{\op}_G}$ preserves cocartesian edges. Though we won't discuss the notion of $G$-commutative algebra much in this paper, the reader should bear in mind that one of the main points of the theory of $G$-symmetric monoidal $\infty$-categories is to streamline arguments involving $G$-commutative algebras. For example, a lax $G$-symmetric monoidal functor necessarily preserves $G$-commutative algebras.
\begin{rem} \label{rem:GAdjSymmMon}
Let $C^{\otimes}, D^{\otimes}$ be $G$-symmetric monoidal $\infty$-categories and let $F: C^{\otimes} \to D^{\otimes}$ by a $G$-symmetric monoidal functor whose underlying $G$-functor is $G$-left adjoint. Then an easy argument with relative adjunctions shows that its $G$-right adjoint $R$ canonically refines to a lax $G$-symmetric monoidal functor.
\end{rem}
Note that if $V \cong \coprod_{j \in J} V_j$ for orbits $\{ V_j \}_{j \in J}$ and we write $f_j: U_{j} \coloneq U \times_V V_j \to V_j$ for the fiber, then we have
\[ f_{\otimes} \simeq \prod_{j \in J} (f_j)_{\otimes}: \prod_{j \in J} C^\otimes_{U_j} \to \prod_{j \in J} C_{V_j}. \]
Because of this, we may think of a $G$-symmetric monoidal structure as being defined by the collection of functors $\{ f_{\otimes} \}$ for those maps $[f: U \to V] \in \FF_G$ with $V$ an orbit, subject to relations encoded by $\Span(\FF_G)$.
\begin{rem} \label{rem:NormSymmetricMonoidal}
Given a $G$-symmetric monoidal $\infty$-category $C^\otimes$, for every finite $G$-set $V$ the fiber $C^\otimes_V$ is a symmetric monoidal $\infty$-category via restriction of $C^\otimes$ along the map
$$\Span(\FF) \to \Span(\FF_G), \quad n \mapsto V^{\sqcup n}.$$
In other words, the fold maps for $V$ define the symmetric monoidal structure on $C^\otimes_V$. Moreover, for any map $f: U \to V$ the norm functor $f_\otimes: C^\otimes_U \to C^\otimes_V$ is then symmetric monoidal.
\end{rem}
\begin{rem}
Suppose $C^\otimes$ is a $G$-symmetric monoidal $\infty$-category and $f: U \to V$ is a map of finite $G$-sets. Then in view of the compatibility between norms and restriction as encoded by $\Span(\FF_G)$, the norm functor $f_\otimes: C^\otimes_U \to C^\otimes_V$ canonically lifts to a \emph{norm $\underline{V}$-functor}
\[ f_\otimes: f_* f^* C^\otimes_{\underline{V}} = f_* C^\otimes_{\underline{U}} \to C^{\otimes}_{\underline{V}}. \]
\end{rem}
We next discuss the main examples of $G$-symmetric monoidal $\infty$-categories.
\begin{exm}
Suppose $C$ is a $G$-$\infty$-category that admits finite $G$-products (e.g., $\underline{\Spc}^G$). Then by applying Barwick's unfurling construction \cite[\S 11]{M1} to the Beck-Chevalley fibration $\widetilde{C} \to (\FF_G)^{\op}$ that encodes the functoriality of restriction and coinduction, we may define the \emph{$G$-cartesian} $G$-symmetric monoidal structure $C^{\times} \to \Span(\FF_G)$ such that the norm functors are given by coinduction.
\end{exm}
\begin{exm} \label{exm:GSMCSpectra}
The usual symmetric monoidal structures on $\{ \Sp^H \}_{H \leq G}$ together with the Hill--Hopkins--Ravenel norm functors furnish a $G$-symmetric monoidal structure on $\underline{\Sp}^G$. Indeed, we may define this via restricting $\SH^{\otimes}$ (\cref{BachmannHoyoisFunctorNorms}) along the map $\omega_G: \Span(\FF_G) \to \Span(\Gpd_{\fin}), U \mapsto U//G$ .
\end{exm}
\begin{exm}[Pointwise $G$-symmetric monoidal structure] \label{exm:pointwiseGSMC}
Let $C$ be a $G$-symmetric monoidal $\infty$-category and let $I$ be a $G$-$\infty$-category. Then we have a \emph{pointwise} $G$-symmetric monoidal structure on $\underline{\Fun}_G(I, C)$ such that for a map $f: U \to V$ of finite $G$-sets, the norm functor
\[ f_{\otimes}: \Fun_{\underline{U}}(I_{\underline{U}}, C_{\underline{U}}) \to \Fun_{\underline{V}}(I_{\underline{V}}, C_{\underline{V}}) \]
sends a $\underline{U}$-functor $[F: I_{\underline{U}} = f^* I_{\underline{V}} \to C_{\underline{U}} = f^* C_{\underline{V}}]$ to the $\underline{V}$-functor
\[ f_{\otimes} F: I_{\underline{V}} \xto{\eta} f_* f^* I_{\underline{V}} \xto{f_* F} f_* f^* C_{\underline{V}} \xto{f_{\otimes}} C_{\underline{V}}, \]
where $\eta$ is the unit map and $f_{\otimes}: f_* f^* C_{\underline{V}} \to C_{\underline{V}}$ is the norm $\underline{V}$-functor defined by $C^\otimes$. We will construct this in \cite{paramalg} as the cotensor in $G$-$\infty$-operads.\footnote{We note that this will supersede \cref{dfn:S-PointwiseMonoidal} in all examples of interest in this paper. However, since the theory of parametrized $\infty$-operads imposes strong restrictions upon the base $\infty$-category $S$ (namely, that $T = S^{\op}$ is atomic orbital), \cref{dfn:S-PointwiseMonoidal} itself is not superseded by the work done in \cite{paramalg}.} If we write $\underline{\CAlg}_G(C)$ for the $G$-$\infty$-category of $G$-commutative algebras, we then have $\underline{\CAlg}_G(\underline{\Fun}_G(I,C)) \simeq \underline{\Fun}_G(I, \underline{\CAlg}_G(C))$.
\end{exm}
We end this subsection by introducing the concept of \emph{$G$-distributivity}, which will play an important role in establishing the existence and uniqueness of a lax $G$-symmetric monoidal structure on $(-)^{t_G K}$ (\cref{cor:LaxGSymmetricMonoidalTate}).
\begin{dfn} \label{dfn:DistributiveFunctor}
Let $f: U \to V$ be a map of finite $G$-sets, let $C$ be a $\underline{U}$-$\infty$-category, and let $D$ be a $\underline{V}$-$\infty$-category. Let $F: f_* C \to D$ be a $\underline{V}$-functor. Then we say that $F$ is \emph{$\underline{V}$-distributive} if for every pullback square
\[ \begin{tikzcd}
U' \ar{r}{f'} \ar{d}{g'} & V' \ar{d}{g} \\
U \ar{r}{f} & V
\end{tikzcd} \]
of finite $G$-sets and $\underline{U'}$-colimit diagram $\overline{p}: K^{\underline{\rhd}} \to g'^* C$, the $\underline{V'}$-functor
\[ (f'_* K)^{\underline{\rhd}} \xto{\can} f'_* (K^{\underline{\rhd}}) \xto{f'_* \overline{p}} f'_* g'^* C \simeq g^* f_* C \xto{g^* F} g^* D \]
is a $\underline{V'}$-colimit diagram.\footnote{Given a $\underline{U'}$-$\infty$-category $K$, we write $K^{\underline{\rhd}}$ for the parametrized join $K \star_{\underline{U'}} \underline{U'}$ \cite[Def.~4.1]{Exp2}, and likewise $(f'_* K)^{\underline{\rhd}} = (f'_* K) \star_{\underline{V'}} \underline{V'}$ (so the notation $(-)^{\underline{\rhd}}$ implicitly involves the base). Using the compatibility of the parametrized join with restriction \cite[Lem.~4.4]{Exp2}, the canonical map $(f'_* K)^{\underline{\rhd}} \xto{\can} f'_* (K^{\underline{\rhd}})$ is then defined to be the adjoint to $\epsilon^{\underline{\rhd}}: (f'^* f'_* K)^{\underline{\rhd}} \to K^{\underline{\rhd}}$.}
\end{dfn}
\begin{rem}
In the situation of \Cref{dfn:DistributiveFunctor}, suppose that $C$ is $\underline{U}$-cocomplete and $D$ is $\underline{V}$-cocomplete. In view of the pointwise formula for parametrized Kan extensions [Shah, Thm.~10.3], we deduce the following apparently stronger conclusion:
\begin{itemize}
\item[($\ast$)] Suppose that
\[ \begin{tikzcd}
K \ar{r}{p} \ar{d}{\quad \Downarrow \eta}[swap]{\phi} & [3em] (g')^* C \\
L \ar{ru}[swap]{\phi_! p}
\end{tikzcd} \]
is a $\underline{U'}$-left Kan extension diagram. Then
\[ \begin{tikzcd}
f'_* K \ar{r}{f'_* p} \ar{d}{\quad \Downarrow \eta'}[swap]{f'_* \phi} & f'_* g'^* C \ar{r}{g^* F} & g^* D \\
f'_* L \ar{rru}[swap]{g^* F \circ f'_* (\phi_! p)}
\end{tikzcd} \]
is a $\underline{V'}$-left Kan extension diagram, where $\eta'$ is given by whiskering $f'_* \eta$ by $g^* F$.
\end{itemize}
\end{rem}
The following definition generalizes the notion of a cocomplete symmetric monoidal $\infty$-category in which the tensor product distributes over all colimits.
\begin{dfn} \label{dfn:DistributiveSMC}
Let $C^\otimes$ be a $G$-symmetric monoidal $\infty$-category. Then $C^\otimes$ is \emph{$G$-distributive} if:
\begin{enumerate}
\item $C$ is $G$-cocomplete.
\item For every map $f:U \to V$ of finite $G$-sets, the norm $\underline{V}$-functor $f_{\otimes}$ is $\underline{V}$-distributive (\Cref{dfn:DistributiveFunctor}).
\end{enumerate}
\end{dfn}
\begin{rem}
\Cref{dfn:DistributiveFunctor} is formulated so that the pullback of a $\underline{V}$-distributive functor along a map $g:V' \to V$ is again $\underline{V'}$-distributive. Since condition (2) of \Cref{dfn:DistributiveSMC} is checked against \emph{all} maps of finite $G$-sets, and the pullback along $g$ of a norm $\underline{V}$-functor is the corresponding $\underline{V'}$-norm functor, this base-change condition is superfluous in formulating \Cref{dfn:DistributiveSMC}.
\end{rem}
\begin{exm}
Nardin proved in his thesis \cite{nardin} that the $G$-symmetric monoidal $\infty$-category $\underline{\Sp}^G$ is $G$-distributive. In fact, he constructs the $G$-symmetric monoidal structure on $\underline{\Sp}^G$ as the initial $G$-distributive $G$-presentable $G$-stable $G$-symmetric monoidal $\infty$-category along the lines of Lurie's construction of the tensor product on spectra \cite[\S 4.8.2]{HA}; $G$-distributivity is thus essential to prove the \emph{uniqueness} of the $G$-symmetric monoidal structure on $\underline{\Sp}^G$. One may also establish $G$-distributivity by proving separately that the norm functors preserve sifted colimits and distribute over finite $G$-coproducts (cf. \cite[\S 5.2]{BachmannHoyoisNorms}).
\end{exm}
\begin{exm}
If $C$ is a $G$-distributive $G$-symmetric monoidal $\infty$-category, then the pointwise $G$-symmetric monoidal structure on $\underline{\Fun}_G(I,C)$ is also $G$-distributive.
\end{exm}
\begin{exm} \label{exm:CartesianDistributiveCommutationRelation}
Suppose $C$ is a presentable and cartesian closed $\infty$-category, so that the cartesian symmetric monoidal structure on $C$ is distributive. Then the $G$-cartesian symmetric monoidal structure on the $G$-$\infty$-category $\underline{C}_G$ of $G$-objects in $C$ is $G$-distributive.\footnote{Note that $\underline{C}_G$ is $G$-cocomplete and $G$-complete by \cite[Props.~5.5 and 5.6]{Exp2}.}
We may further unwind the meaning of $G$-distributivity in the case that $C = \Cat_{\infty}$. For a map of finite $G$-sets $f: U \to V$, write
$$f_{\times}: f^* f_* \underline{\Cat}_{\infty, \underline{V}} \to \underline{\Cat}_{\infty, \underline{V}}$$
for the norm $\underline{V}$-functor associated to the $G$-cartesian symmetric monoidal structure; over the fiber $W \xto{\pi} V$, we have that
$$f_{\times}: [C \to \underline{W \times_V U}] \mapsto [f'_* C \to \underline{W}]$$
where $f': W \times_V U \to W$ is the pullback of $f$ along $\pi$. Let $\alpha: A \to U$ be a map of finite $G$-sets, $C$ an $\underline{A}$-$\infty$-category, and $F_C: \underline{A} \to \underline{\Cat}_{\infty, \underline{U}}$ the $\underline{U}$-functor determined by $C$. Then
$$\colim^{\underline{U}} F_C = \alpha_! C \in \Cat_{\infty, \underline{U}}$$
and $G$-distributivity implies a formula for $f_* \alpha_! C \in \Cat_{\infty, \underline{V}}$. Namely, let $f^* \dashv f_*$ also denote the adjunction
\[ \adjunct{f^*}{(\FF_G)^{/V}}{(\FF_G)^{/U}}{f_*} \]
and consider the $\underline{V}$-functor
\[ \underline{f_* A} \simeq f_* \underline{A} \xtolong{f_*(F_C)}{1.5} f_* f^* \underline{\Cat}_{\infty,\underline{V}} \xto{f_{\times}} \underline{\Cat}_{\infty, \underline{V}}. \]
By $G$-distributivity, $f_* \alpha_! C \simeq \colim^{\underline{V}} (f_{\times} \circ f_* (F_C))$. Moreover, the $\underline{V}$-functor $f_* (F_C)$ is equivalent data to the $\underline{f^* f_* A}$-$\infty$-category $\ev^* C$, where $\ev: f^* f_* A \to A$ denotes the counit of the adjunction $f^* \dashv f_*$ (which may be concretely described as evaluation). Now let
$$\pr: f^* f_* A = (f_* A) \times_{V} U \to f_* A$$
denote the projection. Under the correspondence between $\underline{f_* A}$-points in $\underline{\Cat}_{\infty, \underline{V}}$ and $\underline{f_* A}$-$\infty$-categories, one then identifies $f_{\times} \circ f_* (F_C)$ with $\pr_* \ev^* C$. Finally, let $g: f_* A \to V$ denote the structure map. Summing up, given the commutative diagram
\[ \begin{tikzcd}
& f^* f_* A \ar{r}{\pr} \ar{d} \ar{ld}[swap]{\ev} & f_* A \ar{d}{g} \\
A \ar{r}{\alpha} & U \ar{r}{f} & V,
\end{tikzcd} \]
we obtain the equivalence
\[ f_* \alpha_! C \simeq g_! \pr_* \ev^* C. \]
\end{exm}
\subsection{Parametrized Verdier quotients}\label{SS:Verdier}
In this section, we set up a parametrized generalization of the theory of Verdier quotients. We first consider the $G$-Verdier quotient of a $G$-stable $G$-$\infty$-category by a full $G$-stable $G$-subcategory (\cref{prp:VerdierQuotient}), and then the $G$-Verdier quotient of a $G$-stable $G$-symmetric monoidal $\infty$-category by a $G$-$\otimes$-ideal (\cref{prp:VerdierQuotientInducedObjects}). We will use this theory to prove that the parametrized Tate construction uniquely admits a lax equivariant symmetric monoidal structure (cf. \cref{prp:MonoidalStructureTate}).
For the definition of $G$-Verdier quotient, recall from \cite{Hinich2016} that given a cocartesian fibration $\pi: C \to S$ and a class of fiberwise edges $\sW \subset C$ closed under cocartesian pushforward, one may form the Dwyer-Kan localization $\pi^W: C[\sW^{-1}] \to S$ as the fibrant replacement of the marked simplicial set $(C, \sW)$ in the cocartesian model structure on $s\Set^+_{/S}$. For every $s \in S$, one then has $(C[\sW^{-1}])_s \simeq C_s[\sW_s^{-1}]$; more generally, for every $\infty$-category $K \to S$, if we let $\sW_K \subset C_K = K \times_S C$ denote the restriction of $\sW$, then $C_K[\sW_K^{-1}] \simeq K \times_S C[\sW^{-1}]$.
\begin{dfn}
Let $C$ be a $G$-stable $G$-$\infty$-category and $D \subset C$ a full $G$-stable $G$-subcategory. We define the \emph{$G$-Verdier quotient} $C/D$ to be the Dwyer-Kan localization $C[\sW^{-1}]$, where $\sW \subset C$ is the class of fiberwise edges given over an orbit $V$ by those edges $x \to y$ with cofiber in $D_V$.
\end{dfn}
\begin{rem} \label{rem:IndCompletion}
Suppose $C$ is a small $G$-$\infty$-category that admits all finite $G$-colimits. Then the fiberwise Ind-completion $\underline{\Ind}_G(C)$ is fiberwise presentable, $G$-cocomplete and $G$-complete,\footnote{Given a $G$-$\infty$-category $E$ that is fiberwise presentable, to check that $E$ admits all small $G$-limits and $G$-colimits it suffices to verify that all restriction functors $f^*: E_V \to E_U$ admit left and right adjoints that satisfy the Beck-Chevalley conditions.} and the fiberwise Yoneda embedding $j^{\omega}_G: \into{C}{\underline{\Ind}_G(C)}$ preserves finite $G$-colimits. Moreover, by \cite[Thm.~9.11]{Exp2b} $\underline{\Ind}_G(C)$ identifies with the full $G$-subcategory
\[ i: \underline{\Fun}^{\lex}_G(C^{\vop}, \underline{\Spc}^G) \subset \underline{\PShv}_G(C) \]
whose fiber over an orbit $V \cong G/H$ is spanned by those $H$-presheaves that strongly preserve $H$-finite $H$-limits, and the $G$-Yoneda embedding $j_G: C \to \underline{\PShv}_G(C)$ factors as
\[ \begin{tikzcd}
C \ar[hookrightarrow]{r}{j^{\omega}_G} & \underline{\Ind}_G(C) \ar[hookrightarrow]{r}{i} & \underline{\PShv}_G(C).
\end{tikzcd} \]
If we further suppose that $C$ is $G$-stable, then since $j^{\omega}_G$ also preserves all finite $G$-limits (as $j_G$ does and $i$ is a right $G$-adjoint), it follows that $\underline{\Ind}_G(C)$ is $G$-stable and $j^{\omega}_G$ is $G$-exact.
\end{rem}
We have the following parametrized generalization of \cite[Thm.~I.3.3]{NS18}.
\begin{thm} \label{prp:VerdierQuotient} Let $C$ be a small $G$-stable $G$-$\infty$-category and $D \subset C$ a full $G$-stable $G$-subcategory.
\begin{enumerate}
\item The $G$-Verdier quotient $C/D$ is $G$-stable and the quotient $G$-functor $p: C \to C/D$ is $G$-exact. Moreover, for any $G$-stable $G$-$\infty$-category $E$, the restriction functor
\[ p^*: \into{\Fun_G^{\ex}(C/D, E)}{\Fun_G^{\ex}(C, E)} \]
is fully faithful with essential image spanned by those $G$-exact $G$-functors $F: C \to E$ such that $F|_{D} \simeq 0_G$. In particular, passing to mapping spaces we see that $C/D$ is the cofiber of the inclusion $D \subset C$ taken in the $\infty$-category of $G$-stable $G$-$\infty$-categories.
\item The restricted $G$-Yoneda $G$-functor $C/D \to \underline{\PShv}_G(C)$ factors through a $G$-exact $G$-functor
\[ r: C/D \to \underline{\Ind}_G(C) \]
given fiberwise by the formula of \cite[Thm.~I.3.3(ii)]{NS18}.
\item The quotient $G$-functor $p: C \to C/D$ prolongs to a $G$-functor
\[ p_!: \underline{\Ind}_G(C) \to \underline{\Ind}_G(C/D) \]
that preserves all $G$-colimits and admits a fully faithful right $G$-adjoint $p^*$ given by restriction along $p$, or equivalently as the prolongation of $r$. Moreover, $p^*$ also preserves all $G$-colimits and admits a right $G$-adjoint.
\item Let $E$ be a $G$-stable $G$-cocomplete $G$-$\infty$-category. The restriction functor
\[ p^*: \into{\Fun_G^{\ex}(C/D, E)}{\Fun_G^{\ex}(C, E)} \]
then admits a left adjoint $L$, which sends a $G$-exact $G$-functor $F: C \to E$ to the composite
\[ \begin{tikzcd}
C/D \ar{r}{r} & \underline{\Ind}_G(C) \ar{r}{\underline{\Ind}_G(F)} & [3em] \underline{\Ind}_G(E) \ar{r}{\colim^G} & E.
\end{tikzcd} \]
Moreover, this adjunction globalizes to a $G$-adjunction
\[ \adjunct{L}{\underline{\Fun}_G^{\ex}(C, E)}{\underline{\Fun}_G^{\ex}(C/D, E)}{p^*} \]
\end{enumerate}
\end{thm}
\begin{proof}
\begin{enumerate}[leftmargin=*]
\item Since $(C/D)_V \simeq C_V / D_V$, we already know that $C/D$ is fiberwise stable and $p$ is fiberwise exact. To see that $C/D$ is moreover $G$-stable and $p$ is $G$-exact, observe that since the inclusion $G$-functor $D \subset C$ is $G$-exact by assumption, it follows by the universal property of the Verdier quotient that the induction and coinduction functors $f_!$ and $f_*$ descend to $C/D$ (so then intertwine with $p$). Moreover, in view of the fiberwise surjectivity of $p$, the functors $\{ f_!, f_* \}_{f \in \Ar(\FF_G)}$ continue to satisfy the Beck-Chevalley condition for pullback squares in $\FF_G$, and the $G$-semiadditivity equivalence $f_! \simeq f_*$ in $C/D$ is implied by that in $C$.
For the second assertion, by the universal property of the Dwyer-Kan localization we already know that
\[ p^*: \into{\Fun^{\text{fib-ex}}_G(C/D, E)}{\Fun^{\text{fib-ex}}_G(C, E)} \]
is fully faithful with essential image spanned by those fiberwise exact $G$-functors $F: C \to E$ such that $F|_{D} \simeq 0_G$. Since we proved that $p$ is $G$-exact, it follows that after passage to subcategories of $G$-exact functors on both sides, $p^*$ remains fully faithful with essential image as described - note that if $F': C/D \to E$ is a $G$-functor such that $F' \circ p$ is $G$-exact, then it follows from the fiberwise surjectivity of $p$ that $F'$ is $G$-exact.
\item Note first that $\underline{\Ind}_G(C)$ is $G$-stable by \Cref{rem:IndCompletion}. In addition, under the identification
$$\underline{\Ind}_G(C) \simeq \underline{\Fun}^{\lex}_G(C^{\vop}, \underline{\Spc}^G)$$
of \Cref{rem:IndCompletion} (and ditto for $C/D$), the $G$-functor
\[ p^*: \PShv_G(C/D) \to \PShv_G(C) \quad \text{restricts to} \quad p^*: \underline{\Ind}_G(C/D) \to \underline{\Ind}_G(C). \]
The assertion then follows from \cite[Thm.~I.3.3(ii)]{NS18}.
\item In view of the universal property of $\underline{\Ind}_G$ (\cite[Thm.~9.11]{Exp2b}), we have a $G$-adjunction
\[ \adjunct{p_!}{\underline{\Ind}_G(C)}{\underline{\Ind}_G(C/D)}{p^*} \]
that fiberwise restricts to the adjunction of \cite[Prop.~I.3.5]{NS18}. It follows that $p^*$ is fully faithful from the fiberwise assertion. Moreover, since the inclusion $\underline{\Ind}_G(C/D) \subset \underline{\PShv}_G(C)$ preserves filtered $G$-colimits [ref] and restriction between functor $G$-categories preserves all $G$-colimits, $p^*$ preserves all filtered $G$-colimits. As $p^*$ is also $G$-exact, it thus preserves all $G$-colimits. Since $\underline{\Ind}_G(-)$ is in addition fiberwise presentable, we then see that $p^*$ admits a right $G$-adjoint.
\item By part (3), we have an adjunction
\[ \adjunct{(p^*)^*}{\Fun^L_G(\underline{\Ind}_G(C), E)}{\Fun^L_G(\underline{\Ind}_G(C/D),E)}{(p_!)^*} \]
in which $(p_!)^*$ is fully faithful. By the universal property of $\underline{\Ind}_G$, the restriction functor
\[ (j^{\omega}_G)^*: \Fun^L_G(\underline{\Ind}_G(C), E) \xto{\simeq} \Fun^{\ex}_G(C, E) \]
is an equivalence, and ditto for $C/D$. Under these equivalences, $p^*$ then corresponds to $(p_!)^*$ and $(p^*)^*$ yields the localization functor $L$ with the indicated formula. Finally, both assignments $L$ and $p^*$ are clearly compatible with base-change, hence the adjunction globalizes to a $G$-adjunction.
\end{enumerate}
\end{proof}
Next, given a $G$-symmetric monoidal structure on $C$, we would like to descend that structure to the $G$-Verdier quotient $C/D$. To do so, we need to suppose in addition that $D \subset C$ is a \emph{$G$-$\otimes$-ideal} in the sense of the following definition.
\begin{dfn} \label{dfn:Gtensorideal}
Let $C$ be a $G$-stable $G$-symmetric monoidal $\infty$-category and let $D \subset C$ be a $G$-stable $G$-subcategory. We say that $D$ is a \emph{$G$-$\otimes$-ideal} if
\begin{enumerate}
\item For every orbit $V$, $D_V \subset C_V$ is a $\otimes$-ideal.
\item For every map $f: U \to V$ between orbits,\footnote{Recall our convention that orbits are non-empty, so we aren't supposing that $1 \in D_V$.} if $x \in D_U$, then $f_{\otimes} x \in D_V$.
\end{enumerate}
\end{dfn}
\begin{rem}
In \Cref{dfn:Gtensorideal}, conditions (1) and (2) may be combined as follows:
\begin{itemize}
\item[($\ast$)] Suppose $V$ is an orbit, $U$ is a finite $G$-set with orbit decomposition $U \simeq \coprod_{i=1}^n U_i$, and $f: U \to V$ is a map. Then $D$ is a $G$-$\otimes$-ideal if and only if for all $n$-tuples $x=(x_1,...,x_n) \in C_U \simeq \prod_{i=1}^n C_{U_i}$ such that $x_i \in D_{U_i}$ for some $1 \leq i \leq n$, we have that $f_{\otimes} x \in D_V$.
\end{itemize}
\end{rem}
\begin{exm}
Let $\cF$ be a $G$-family. Then we claim that the $G$-stable $G$-subcategory $\underline{\Sp}^{\Phi \cF} \subset \underline{\Sp}^G$ of \cref{ParamRecollementFamily} is a $G$-$\otimes$-ideal. Indeed, since this is a fiberwise $\otimes$-ideal, it suffices to show stability under norms for any map of orbits $f:U \to V$. Without loss of generality, suppose $U= G/H$ and $V = G/G$. Given $X \in \Sp^H$, we need to show that if $\Phi^K(X) = 0$ for all $K \in \cF^H$, then $\Phi^L(N^G_H X) =0$ for all $L \in \cF$. But using that $N^G_H(E \cF_+) \simeq \Sigma^{\infty}_+ (\Coind^G_H E \cF)$ where $\Coind^G_H$ denotes right Kan extension along $\sO_H^{\op} \simeq (\sO_G^{\op})^{G/H} \to \sO_G^{\op}$, this follows from evaluating $(\Coind^G_H E \cF)^L = \ast$ using the pointwise formula for right Kan extension.
On the other hand, note that $\underline{\Sp}^{\tau \cF} \subset \underline{\Sp}^G$ is generally not a $G$-$\otimes$-ideal. Indeed, suppose $\cF$ is a proper nonempty $G$-family (and $G$ is nontrivial). Then $\Sp^{\tau \cF^e} = \Sp$, but $N^G: \Sp \to \Sp^G$ is split by $\Phi^G$ and hence doesn't send $\Sp^{\tau \cF^e}$ into $\Sp^{\tau \cF}$.
\end{exm}
To then understand the interaction of $G$-Verdier quotients with $G$-symmetric monoidal structures, we will need the following lemma, a $G$-version of \cite[Prop.~A.5]{NS18}.
\begin{lem} \label{lem:DKlocalization}
Let $C$ be a $G$-symmetric monoidal $\infty$-category. Suppose given classes of edges $\{ \sW_{V} \subset C_V \}_{V \in \sO_G}$ with each $\sW_V$ containing all the equivalences in $C_V$ and subject to the following condition:
\begin{itemize}
\item[($\ast$)] For every map of finite $G$-sets $f: U \to V$ with orbit decomposition $U \simeq \coprod_{i=1}^n U_i$ and edges $\{ \alpha_i \in (\sW_{U_i})_1 \}_{1 \leq i \leq n}$, we have that $f_{\otimes}(\alpha_1,...,\alpha_n) \in \sW_V$.\footnote{To obtain the Dwyer-Kan localization $C^\otimes[(\sW^\otimes)^{-1}]$, it would suffice to require the apparently weaker condition in which we suppose that all but one of the $\alpha_i$ are identity morphisms.}
\end{itemize}
Define $\sW^\otimes \subset C^\otimes$ to be the class of those fiberwise edges given over a finite $G$-set $U$ with orbit decomposition $U \simeq \coprod_{i=1}^n U_i$ by
$$\sW^\otimes_U \coloneq \sW_{U_1} \times ... \times \sW_{U_n}.$$
Then the Dwyer-Kan localization $C^\otimes[(\sW^\otimes)^{-1}]$ exists and comes equipped with a map to $\Span(\FF_G)$ that renders it a $G$-symmetric monoidal $\infty$-category. Moreover, $C^\otimes[(\sW^\otimes)^{-1}]$ has the following properties:
\begin{enumerate}
\item The underlying $G$-$\infty$-category of $C^\otimes[(\sW^\otimes)^{-1}]$ is equivalent to $C[\sW^{-1}]$,\footnote{Here, $\sW \subset C$ denotes the class of those fiberwise edges given over an orbit $V$ by $\sW_V$.} and its fiber over an orbit $V$ is given by $C_V[\sW_V^{-1}]$. More generally, the procedure of Dwyer-Kan localization is stable with respect to pullback in the base.
\item The localization functor $L^{\otimes}: C^\otimes \to C^\otimes[(\sW^\otimes)^{-1}]$ is a $G$-symmetric monoidal functor and restricts to the localization functor $L: C \to C[\sW^{-1}]$ over $\sO_G^{\op}$. More generally, $L^\otimes$ intertwines with pullback in the base.
\item For every $G$-symmetric monoidal $\infty$-category $D$, the pullback along the functor $L^\otimes$ induces fully faithful functors
\[ \into{\Fun^{\otimes}_G(C[\sW^{-1}], D)}{\Fun^{\otimes}_{G}(C,D)}, \quad \into{\Fun^{\lax}_G(C[\sW^{-1}], D)}{\Fun^{\lax}_{G}(C,D)} \]
with essential image spanned by those (lax) $G$-symmetric monoidal functors $F: C \to D$ that send $\sW$ to equivalences.
\end{enumerate}
\end{lem}
\begin{proof}
We may procede exactly as in \cite[Prop.~A.5]{NS18}, once we observe that Hinich's work on Dwyer-Kan localization \cite{Hinich2016} applies to the general context over working over a base $\infty$-category with a subclass of `inert' edges that distinguishes `strong' from `lax' morphisms: we apply his theory here to $(\Span(\FF_G), \FF_G^{\op})$.
\end{proof}
We now can prove the parametrized analogue of \cite[Thm.~I.3.6]{NS18}.
\begin{thm} \label{prp:VerdierQuotientInducedObjects}
Let $C$ be a small $G$-stable $G$-symmetric monoidal $\infty$-category and $D \subset C$ a $G$-$\otimes$-ideal.
\begin{enumerate}
\item The $G$-Verdier quotient $C/D$ uniquely inherits a $G$-symmetric monoidal structure from $C$ such that the projection $G$-functor $p: C \to C/D$ is $G$-symmetric monoidal. Moreover, for any $G$-symmetric monoidal $\infty$-category $E$, the restriction functor
\[ p^*: \into{\Fun_G^{\ex, (\otimes \: \mathrm{ or} \lax)}(C/D, E)}{\Fun_G^{\ex, (\otimes \: \mathrm{ or} \lax)}(C, E)} \]
is fully faithful with essential image spanned by those (lax) $G$-symmetric monoidal $G$-exact $G$-functors $F: C \to E$ such that $F|_{D} \simeq 0_G$.
\item Let $E$ be a $G$-stable $G$-distributive $G$-symmetric monoidal $\infty$-category. Then
$$p^*: \into{\Fun_G^{\ex, \lax}(C/D,E)}{\Fun_G^{\ex, \lax}(C,E)}$$
admits a left adjoint $L$ that sends a $G$-exact lax $G$-symmetric monoidal functor $F: C \to E$ to the composite
\[ \begin{tikzcd}
C/D \ar{r}{r} & \underline{\Ind}_G(C) \ar{r}{\underline{\Ind}_G(F)} & [3em] \underline{\Ind}_G(E) \ar{r}{\colim^G} & E
\end{tikzcd} \]
as in \Cref{prp:VerdierQuotient}(4). Moreover, the adjunction $L \dashv p^*$ globalizes to a $G$-adjunction.
\end{enumerate}
\end{thm}
\begin{proof}
\begin{enumerate}[leftmargin=*]
\item For an orbit $V$, let $\sW_V$ be the class of edges $\{ \alpha: x \to y, \: \mathrm{cof}(\alpha) \in D_V \}$ in $C_V$. Under our assumption that $D$ is a $G$-$\otimes$-ideal, \cref{lem:DKlocalization} then applies to prove all of the claims.
\item We first show that all the $G$-functors in the composite are lax $G$-symmetric monoidal. In \cite{paramalg}, the second author showed that for a lax $G$-symmetric monoidal functor $\phi: I \to J$, the restriction functor
\[ \phi^*: \Fun^{\lax}_G(J, E) \to \Fun^{\lax}_G(I,E) \]
admits a left adjoint $\phi_!$ given by $G$-operadic left Kan extension, such that $\phi_!$ is computed on underlying $G$-functors as left $G$-Kan extension. He also showed that $\underline{\PShv}_G(C)$ is a $G$-distributive $G$-symmetric monoidal $\infty$-category with respect to $G$-Day convolution and the $G$-Yoneda embedding $j: \into{C}{\underline{\PShv}_G(C)}$ is a $G$-symmetric monoidal functor. Using $G$-distributivity it then follows that $\underline{\Ind}_G(C) \subset \underline{\PShv}_G(C)$ is a $G$-symmetric monoidal subcategory and $j^{\omega}_G$ is also $G$-symmetric monoidal. We thus see that:
\begin{enumerate}
\item[(i)] $r = p^* \circ j^{\omega}_G$ is lax $G$-symmetric monoidal. Here we use that with respect to $G$-Day convolution, precomposition along a (strong or lax) $G$-symmetric monoidal functor is lax $G$-symmetric monoidal. Alternatively, as in the proof of \cite[Thm.~I.3.6]{NS18} we could observe that the prolongation of the $G$-symmetric monoidal functor $p$ to $p_!: \underline{\Ind}_G(C) \to \underline{\Ind}_G(C/D)$ is actually strong $G$-symmetric monoidal in view of $G$-distributivity, and hence its $G$-right adjoint $p^*$ is lax $G$-symmetric monoidal.
\item[(ii)] $\overline{F} = \colim^G \circ \underline{\Ind}_G(F)$, which is the left $G$-Kan extension of $F$ along $j^{\omega}_G$, canonically inherits the structure of a lax $G$-symmetric monoidal functor.
\end{enumerate}
The functor $L: \Fun^{\ex, \lax}_G(C,E) \to \Fun^{\ex, \lax}_G(C/D,E)$ is thus well-defined. Moreover, we may define the unit transformation $\eta: \id \to p^* L$ to be that given objectwise by the map
\[ F = \overline{F} \circ j^{\omega}_G \to \overline{F} \circ r \circ p = \overline{F} \circ p^* \circ j^{\omega}_G \circ p \simeq \overline{F} \circ p^* p_! \circ j^{\omega}_G \]
induced by the unit of the adjunction
\[ \adjunct{p_!}{\underline{\Ind}_G(C)}{\underline{\Ind}_G(C/D)}{p^*}, \]
where as noted above $p_!$ is strong $G$-symmetric monoidal and $p^*$ is lax $G$-symmetric monoidal. Since $\eta$ by definition covers the unit transformation of the adjunction in \Cref{prp:VerdierQuotient}(4), we see that $L \dashv p^*$ (for the other labeled $p^*$). Finally, the globalization assertion follows from similar ones established for all the functors used to construct $L$.
\end{enumerate}
\end{proof}
\begin{rem}
Roughly speaking, \cref{dfn:Gtensorideal} may be viewed as a categorification of the Tambara ideals introduced by Nakaoka in \cite[Def.~2.1]{Nak12}. \cref{prp:VerdierQuotientInducedObjects} can then be viewed as a categorification of the fact that quotients by Tambara ideals are Tambara functors \cite[Prop.~2.6]{Nak12}.
\end{rem}
\begin{exm} \label{exm:InducedGMSCBorel}
Applying \cref{prp:VerdierQuotientInducedObjects} to the $G$-$\otimes$-ideal $\underline{\Sp}^{\Phi \cF} \subset \underline{\Sp}^{G}$ we get an induced $G$-symmetric monoidal structure on $\underline{\Sp}^{h \cF}$.\footnote{We may get around the smallness assumption in \cref{prp:VerdierQuotientInducedObjects} by using that the inclusion is fiberwise accessible.} Consequently, in view of \cref{rem:GAdjSymmMon} given a $G$-commutative algebra $A$ in $\underline{\Sp}^G$, the unit map $A \to F(E \cF_+, A)$ canonically refines to a morphism of $G$-commutative algebras.
\end{exm}
\subsection{Digression on the Segal conjecture}
As an application of \cref{prp:VerdierQuotientInducedObjects}, we upgrade the generalized Segal conjecture to a statement about incomplete Tambara functors. Recall that the Segal conjecture (cf. \cite{Car84}) states that the map
$$(\pi_G^* \mathbb{S})_I^\wedge \to \pi^*(\Sigma^\infty_+ BG)$$
between the $G$-equivariant stable cohomotopy groups of a point completed at the augmentation ideal of the Burnside ring and the stable cohomotopy groups of $BG$ is an isomorphism. Adams, Haeberly, Jackowski, and May extended the Segal conjecture to families in \cite{AHJM88b}.
\begin{dfn}\cite[Intro.]{AHJM88b}
Let $G$ be a finite group and let $\cH$ be a class of subgroups of $G$. A functor $h$ defined on $G$-spaces and $G$-maps is \emph{$\cH$-invariant} if it carries $\cH$-equivalences to isomorphisms in the target category of $h$.
\end{dfn}
\begin{rem}\cite[Rem.~1.2]{AHJM88b}
If $\cH$ is a family of subgroups of $G$, a functor $h$ is $\cF$-invariant if and only if the map $h(X) \to h(E\cH \times X)$ induced by the projection $E\cH \times X \to X$ is an isomorphism for each $X$.
\end{rem}
\begin{thm}[{Generalized Segal conjecture, \cite[Thm.~1.5 and Cor.~1.6]{AHJM88b}}]
For any family $\cF$ the pro-group valued functor $\pi^*_G(-)_{I(\cF)}^\wedge$, given by equivariant cohomotopy completed at
$$I(\cF) := \bigcap_{H \in \cF} \ker(A(G) \to A(H)),$$
is $\cF$-invariant. In particular, there is a pro-isomorphism
\begin{equation}\label{Eqn:SegalMap}
\pi_G^*(X)_{I(\cF)}^\wedge \xrightarrow{\cong} \pi^*_G(E\cF \times X)
\end{equation}
natural in the $G$-space $X$.
\end{thm}
Taking $X = \ast$ to be a point and $\cF = \Gamma_N$ to be the $N$-free $G$-family, we have a pro-isomorphism
$$\pi_G^*(\s)_{I(\Gamma_N)}^\wedge \xrightarrow{\cong} \pi^*_G(E\Gamma_N).$$
The source can be identified with the $I(\Gamma_N)$-completion of $\pi_*(\s^G)$ by Spanier--Whitehead duality, while the point-set model for parametrized homotopy fixed points (\cref{rem:PointSetModels}) gives
$$\pi^*_G(E\Gamma_N) \cong \pi_*^G F({E\Gamma_N}_+,\s) \cong \pi_* \s^{h_{G/N}N}.$$
Therefore the map \eqref{Eqn:SegalMap} can be obtained by applying $\pi_*^G(-)$ and completing the source of the vertical map in the commutative diagram
\[
\begin{tikzcd}
{E\Gamma_N}_+ \arrow{r} \arrow{d}{\simeq} & \mathbb{S} \arrow{r} \arrow{d}{\alpha} & \widetilde{E\Gamma_N} \arrow{d}{\tilde{\alpha}} \\
F({E\Gamma_N}_+,\mathbb{S}) \otimes {E\Gamma_N}_+ \arrow{r} & F({E\Gamma_N}_+,\mathbb{S}) \arrow{r} & F({E\Gamma_N}_+,\mathbb{S}) \otimes \widetilde{E\Gamma_N}.
\end{tikzcd}
\]
The map $\alpha : \mathbb{S} \to F({E\Gamma_N}_+, \mathbb{S})$ is a map of $G$-commutative algebras (\cref{exm:InducedGMSCBorel}) and $\pi_0$ of a $G$-commutative algebra is a Tambara functor \cite{Brun2007}, so we obtain the following multiplicative refinement of the generalized Segal conjecture:
\begin{thm}[Multiplicative generalized Segal conjecture]\label{Thm:Segal}
The functor $\pi^*_{(-)}(-)_{\underline{I(\Gamma_N)}}$ valued in pro-Tambara functors, given by equivariant cohomotopy completed at the Tambara ideal $\uI$ defined by
$$\uI(G/K) := \bigcap_{H \in i^*_K \Gamma_N} \ker(A(K) \to A(H)),$$
is $\Gamma_N$-invariant. In particular, the pro-isomorphism of pro-groups \eqref{Eqn:SegalMap} refines to a pro-isomorphism of pro-Tambara functors
\[
\pi_{(-)}^*(X)_{\uI}^{\wedge} \xrightarrow{\cong} \pi^*_{(-)}(E \Gamma_N \times X)
\]
natural in the $G$-space $X$.
\end{thm}
\begin{exm}
Taking $G = D_4$ and $N = \mu_2$ recovers the $C_2$-equivariant analogue of Lin's Theorem \cite[Thm.~A.22]{Qui19b} proven using the $C_2$-equivariant Adams spectral sequence, i.e., the map
$$\s \to \s^{t_{C_2}\mu_2}$$
is a $2$-adic equivalence of $C_2$-spectra.
One might expect that taking $G = D_{2p}$ and $N = \mu_p$ for $p$ odd would yield a $C_2$-equivariant generalization of Gunawardena's Theorem \cite{AGM85}, i.e., the map
$$\s \to \s^{t_{C_2}\mu_p}$$
is a $p$-adic equivalence of genuine $C_2$-spectra. However, this is not the case: the geometric fixed points of $\s^{t_{C_2}\mu_p}$ are trivial by \cref{exm:DihedralOdd}, but the geometric fixed points of $\s$ are nontrivial. This discrepancy is explained in the next remark.
\end{exm}
\begin{rem}
The Tate diagram implies that the Segal conjecture for $N=G=C_p$ is equivalent to showing that the map
$$\pi_*(\s) \to \pi_*(\s^{tC_p})$$
is an isomorphism after $I$-completion of the source, where $I = \ker(A(C_p) \to \ZZ)$ is the augmentation ideal. Elementary commutative algebra shows that $I$-completion coincides with $p$-completion, so the Segal conjecture for $N=G=C_p$ is equivalent to showing that the map
$$\s \to \s^{tC_p}$$
is a $p$-adic equivalence.\footnote{The latter statement is occasionally referred to as the Segal conjecture for $C_p$.} More generally, two subtleties can occur in relating the Segal conjecture to a map to the Tate construction:
\begin{enumerate}
\item For a general finite group $G$, the $I$-adic topology and $p$-adic topology on $A(G)$ can be different. The relation between these topologies was studied in \cite[Sec.~1]{Lai79}, where it was shown the topologies coincide if $G$ is a finite $p$-group, but can differ in general (cf. \cite[Exm.~1.18]{Lai79}). This difference occurs in the generalized Segal conjecture for the $\mu_p$-free family of subgroups of $D_{2p}$, $p$ odd, as mentioned above.
\item Even when the $I$-adic and $p$-adic topologies coincide, the Segal conjecture is \emph{not} equivalent to showing that the map $\s \to \s^{tG}$ is a $p$-adic equivalence. For instance,
$$X^{tC_{p^2}} \simeq X^{tC_p hC_p} \simeq (\s_p^\wedge)^{hC_p} \not\simeq \s_p^\wedge.$$
This is because in the relevant Tate diagram, the right-hand vertical map $\tilde{\alpha}$ has the form
$$\tilde{\alpha} : \widetilde{EC_{p^2}} \to \s^{tC_{p^2}}.$$
\end{enumerate}
\end{rem}
\subsection{Induced objects}\label{SS:Induced}
Let $S_0$ be a space. One may define the full subcategory of \emph{induced objects} $\Fun(S_0, \Sp)_{\ind}$ in $\Fun(S_0, \Sp)$ to be the thick subcategory\footnote{One could also take the minimal stable subcategory as in \cite[Def.~I.3.7]{NS18}.} generated by the set $\{ s_! E : E \in \Sp \}_{s \in S_0}$. It then follows from the projection formula $s_!(E) \otimes F \simeq s_!(E \otimes s^* F)$ that $\Fun(S_0, \Sp)_{\ind}$ is a thick $\otimes$-ideal (cf. \cite[Lem.~I.3.8(ii)]{NS18}). Coupled with the multiplicative theory of the Verdier quotient \cite[Thm.~I.3.6]{NS18} and vanishing of the Tate construction on induced objects \cite[Lem.~I.3.8(i)]{NS18}, we then deduce that the Tate construction uniquely admits a lax symmetric monoidal structure \cite[Thm.~I.3.1]{NS18}.
More generally, let $S$ be a $G$-space. In order to ultimately apply \cref{prp:VerdierQuotientInducedObjects} in the context of the parametrized Tate construction, we now want a full $G$-stable $G$-subcategory
$$\underline{\Fun}_G(S, \underline{\Sp}^G)_{\ind} \subset \underline{\Fun}_G(S, \underline{\Sp}^G)$$
of induced objects which is a (fiberwise thick) $G$-$\otimes$-ideal in the sense of \cref{dfn:Gtensorideal}. Because such a full $G$-subcategory is necessarily closed under restriction, induction, and norms for the pointwise $G$-symmetric monoidal structure on $\underline{\Fun}_G(S, \underline{\Sp}^G)$, the correct formulation of `induced objects' becomes considerably more involved in the parametrized setting. We begin by specifying the set of induced generators for every fiber and then show that this yields first a $G$-stable $G$-subcategory (\cref{dfn:InducedObjectsGSubcat}) and subsequently a $G$-$\otimes$-ideal (\cref{cor:InducedObjectsFormGTensorIdeal}).
\begin{dfn} \label{dfn:InducedObjects}
Let $S$ be a $G$-space. Consider maps of finite $G$-sets
\[ U \xto{f} V \xto{p} W \]
where $W$ is an orbit. Let $s \in S_{U}$ be a point\footnote{If $U$ has orbit decomposition $\coprod_{i \in I} U_i$, then $S_{U} = \prod_{i \in I} S_{U_i}$ and $s$ is given as a tuple $(s_i)$.} and also write $s: \underline{U} \to S_{\underline{U}}$ for the unique $\underline{U}$-functor that selects $s \in S_U$. Let $s_f: \underline{V} \to f_* f^* S_{\underline{V}}$ be the $\underline{V}$-functor adjoint to $s$.\footnote{$s_f$ is also given by $f_*(s)$ since $f_*$ preserves terminal objects.} Let $\eta_{f}: S_{\underline{V}} \to f_* f^* S_{\underline{V}}$ denote the unit $\underline{V}$-functor and define the $\underline{V}$-$\infty$-category $\fib^f_{s}(\eta)$ to be the pullback
\[ \begin{tikzcd}[column sep=4em]
\fib^f_{s}(\eta) \ar{r}{\varphi^f_{s}} \ar{d}[swap]{\pi^f_{s}} & S_{\underline{V}} \ar{d}{\eta_f} \\
\underline{V} \ar{r}{s_f} & f_* f^* S_{\underline{V}}.
\end{tikzcd} \]
Now let $\epsilon_p: p_! p^* S_{\underline{W}} \to S_{\underline{W}}$ denote the counit and consider the composite of $\underline{W}$-functors
\[ p_! \fib^f_{s}(\eta) \xtolong{p_! \varphi^f_{s}}{3em} p_! p^* S_{\underline{W}} \xto{\epsilon_p} S_{\underline{W}}. \]
We let
\[ \cI_W(f,p,s) \coloneq \{ (\epsilon_p \circ p_! \varphi^f_s)_!(X): \: X \in \Fun_{\underline{W}}(\fib^f_{s}(\eta), (\underline{\Sp}^G)_{\underline{W}}) \} \subset \Fun_{\underline{W}}(S_{\underline{W}}, (\underline{\Sp}^G)_{\underline{W}}) \]
where $(\epsilon_p \circ p_! \varphi^f_s)_!$ denotes $\underline{W}$-left Kan extension. Now let $\cI_W \coloneq \bigcup_{f,p,s} \cI_W(f,p,s)$ and define the full subcategory of \emph{induced $\underline{W}$-objects}
\[ \Fun_{\underline{W}}(S_{\underline{W}}, (\underline{\Sp}^G)_{\underline{W}})_{\ind} \subset \Fun_{\underline{W}}(S_{\underline{W}}, (\underline{\Sp}^G)_{\underline{W}}) \]
to be the thick subcategory generated by the set $\cI_W$.
\end{dfn}
\begin{rem}
In terms of cocartesian fibrations, $p_!$ is implemented by postcomposition with the structure map $p: \underline{V} \to \underline{W}$, after which the counit $\epsilon_p: S_{\underline{V}} \simeq \underline{V} \times_{\underline{W}} S_{\underline{W}} \to S_{\underline{W}}$ identifies with the projection to $S_{\underline{W}}$.
\end{rem}
To show that the induced $\underline{W}$-objects assemble to define a full $G$-stable $G$-subcategory of induced objects, we first show that the induced generators are stable under restriction and induction.
\begin{lem} \label{lem:RestrictionInducedObjects}
The collection $\{ \cI_W \}_{W \in \sO_G}$ of induced generators is stable under restriction and induction (along maps of orbits) in $\underline{\Fun}_G(S, \underline{\Sp}^G)$.
\end{lem}
\begin{proof}
We first show closure under restriction. Suppose
\[ \begin{tikzcd}
U' \ar{r}{f'} \ar{d}{g''} & V' \ar{r}{p'} \ar{d}{g'} & W' \ar{d}{g} \\
U \ar{r}{f} & V \ar{r}{p} & W
\end{tikzcd} \]
is a commutative diagram of pullback squares of finite $G$-sets in which $W$ and $W'$ are orbits. In the setup of \cref{dfn:InducedObjects}, consider the commutative diagram of $\underline{W}$-$\infty$-categories
\[ \begin{tikzcd}
p_! \fib^f_s(\eta) \ar{r}{p_! \varphi^f_s} \ar{d}[swap]{p_! \pi^f_s} & p_! S_{\underline{V}} \ar{r}{\epsilon_p} \ar{d}{p_! \eta_f} & S_{\underline{W}} \\
\underline{V} \ar{r}{p_! s_f} & p_! f_* f^* S_{\underline{V}}.
\end{tikzcd} \]
Let $s' = g''^*(s)$. Applying $g^*$ then yields the commutative diagram
\[ \begin{tikzcd}
p'_! \fib^{f'}_{s'}(\eta) \ar{r}{p'_! \varphi^{f'}_{s'}} \ar{d}[swap]{p'_! \pi^{f'}_{s'}} & p'_! S_{\underline{V'}} \ar{r}{\epsilon_{p'}} \ar{d}{p'_! \eta_{f'}} & S_{\underline{W'}} \\
\underline{V'} \ar{r}{p'_! s'_{f'}} & p'_! f'_* f'^* S_{\underline{V'}}
\end{tikzcd} \]
in view of \cref{rem:BaseChangeEquivalences}. Using the built-in compatibility of parametrized left Kan extension with restriction, we thus see that $g^* (\cI_W(f,p,s)) \subset \cI_{W'}(f',p',s')$.
As for closure under induction, suppose that $h: W \to W''$ is a map of orbits. Then by \cref{rem:InductionOfParamColimits}, we have that $h_!(\cI_W(f,p,s)) \subset \cI_{W''}(f,h \circ p,s)$.
\end{proof}
\begin{rem} \label{rem:InductionOfParamColimits}
Suppose $C$ is a $G$-$\infty$-category that admits finite $G$-coproducts, $p: V \to W$ is a map of orbits, and $F: K \to C_{\underline{V}}$ is a $\underline{V}$-functor. Let $F' = \epsilon_p \circ p_! F: p_! K \to C_{\underline{W}}$ be the $\underline{W}$-functor adjoint to $F$. Then we have a canonical equivalence
\[ \colim^{\underline{W}} F' \simeq p_! (\colim^{\underline{V}} F) \in C_W \]
provided that the parametrized colimits exist. Indeed, this follows from the commutative diagram
\[ \begin{tikzcd}
C_W \ar{r}{p^*} \ar{d}{\delta} & C_V \ar{d}{\delta} \\
\Fun_{\underline{W}}(p_! K, C_{\underline{W}}) \ar{r}{\simeq} & \Fun_{\underline{V}}(K, C_{\underline{V}}).
\end{tikzcd} \]
More generally, let $L$ be a $\underline{W}$-$\infty$-category. Then $\Fun_{\underline{W}}(L, C_{\underline{W}})$ admits finite $\underline{W}$-coproducts; write
\[ \adjunct{p_!}{\Fun_{\underline{V}}(L_{\underline{V}},C_{\underline{V}})}{\Fun_{\underline{W}}(L, C_{\underline{W}})}{p^*} \]
for the resulting adjunction. Let $\varphi: K \to L_{\underline{V}}$ be a $\underline{V}$-functor and write $\varphi': p_! K \to L$ for the $\underline{W}$-functor adjoint to $\varphi$. Suppose that $\varphi_! F$ exists. Then we have a canonical equivalence
\[ \varphi'_! F' \simeq p_!(\varphi_! F): L \to C_{\underline{W}}. \]
\end{rem}
The next lemma highlights the relevance of \cref{lem:RestrictionInducedObjects}.
\begin{lem} \label{lem:GeneratorsDefineGStableSubcat}
Let $C$ be a $G$-stable $G$-$\infty$-category and suppose $\{ \cJ_W \subset C_W \}_{W \in \sO_G}$ is a collection of objects stable under restriction and induction. Let $D_W \subset C_W$ be the thick subcategory generated by the set $\cJ_W$. Then for every map of orbits $f: V \to W$, the adjunction
\[ \adjunct{f_!}{C_V}{C_W}{f^*} \]
restricts to an adjunction between $D_V$ and $D_W$. Consequently, the collection $\{ D_W \}_{W \in \sO_G}$ assembles to define a $G$-stable $G$-subcategory $D \subset C$.
\end{lem}
\begin{proof}
It suffices to check that the collection $\{ D_W \}_{W \in \sO_G}$ is stable under restriction and induction; closure under restriction ensures that $\{ D_W \}_{W \in \sO_G}$ assembles to a $G$-subcategory $D$, and closure under induction together with the assumption that $C$ is $G$-stable then ensures that $D$ is $G$-stable and the inclusion $D \subset C$ is $G$-exact. So let $f: V \to W$ be a map of orbits. Let $D^0_W \subset D_W$ be the full stable subcategory consisting of objects $x \in D_W$ such that $f^*(x) \in D_V$. By assumption, $f^*(\cJ_W) \subset \cJ_V$, so $\cJ_W \subset D^0_W$. Since $f^*$ is exact, it follows that $D^0_W$ is a thick subcategory and thus $D^0_W = D_W$. The same argument shows that $f_!(D_V) \subset D_W$.
\end{proof}
We are now ready to state the main definition of this subsection.
\begin{dfn} \label{dfn:InducedObjectsGSubcat}
By \cref{lem:RestrictionInducedObjects} and \cref{lem:GeneratorsDefineGStableSubcat}, the full subcategories of induced $\underline{W}$-objects ranging over all orbits $W$ assemble to define a $G$-stable full $G$-subcategory
\[ \underline{\Fun}_G(S, \underline{\Sp}^G)_{\ind} \subset \underline{\Fun}_G(S, \underline{\Sp}^G) \]
of \emph{$G$-induced objects}.
\end{dfn}
Our remaining goal in this subsection is to show that $\underline{\Fun}_G(S, \underline{\Sp}^G)_{\ind}$ is a $G$-$\otimes$-ideal. First, we show that the induced generators are closed under norms.
\begin{lem} \label{lem:ClosureUnderNorms}
The collection $\{ \cI_W \}_{W \in \sO_G}$ is stable under norms (taken along maps of orbits) in the pointwise $G$-symmetric monoidal structure on $\underline{\Fun}_G(S, \underline{\Sp}^G)$.
\end{lem}
\begin{proof}
Suppose
\[ U \xto{f} V \xto{p} W \xto{\gamma} W' \]
is a sequence of maps of finite $G$-sets in which $W$ and $W'$ are orbits, and
\[ \begin{tikzcd}
\fib^f_s(\eta) \ar{r}{\varphi^f_s} \ar{d}[swap]{\pi^f_s} & S_{\underline{V}} \ar{d}{\eta_{f}} \\
\underline{V} \ar{r}{s_f} & f_* f^* S_{\underline{V}}
\end{tikzcd} \]
is a pullback square of $\underline{V}$-$\infty$-categories as in \cref{dfn:InducedObjects}. Let $X: p_! \fib^f_s(\eta) \to (\underline{\Sp}^G)_{\underline{W}}$ be a $\underline{W}$-functor and consider the induced $\underline{W}$-object
$$Y = (\epsilon_p \circ p_!(\varphi^f_s))_! X: S_{\underline{W}} \to (\underline{\Sp}^G)_{\underline{W}}.$$
We want to show that $Y' = \gamma_{\otimes}((\epsilon_p \circ p_!(\varphi^f_s))_! X)$ is an induced $\underline{W'}$-object. To do so, we will need to consider the commutative diagram of finite $G$-sets
\[ \begin{tikzcd}
& U' \ar{r}{f'} \ar[bend left=25]{rr}{g} \ar{ld}[swap]{\ev'} & \gamma^* \gamma_* V \ar{d} \ar{ld}[swap]{\ev} \ar{r}{\gamma'} & V' = \gamma_* V \ar{d}{q} \\
U \ar{r}{f} & V \ar{r}{p} & W \ar{r}{\gamma} & W'
\end{tikzcd} \]
in which both parallelograms are pullback squares.\footnote{Here, notation is as in \cref{exm:CartesianDistributiveCommutationRelation}; we write $\gamma_*: (\FF_G)^{/W} \to (\FF_G)^{/W'}$ for the right adjoint to pullback along $\gamma$ and $\ev: \gamma^* \gamma_* \to \id$ for the counit of the adjunction.}
First note that by $G$-distributivity of $(\underline{\Sp}^G, \otimes)$, if we let $Y''$ be the $\underline{W'}$-left Kan extension of
\[ \gamma_* p_! \fib^f_s(\eta) \xtolong{\gamma_* X}{1.5} \gamma_* \gamma^* (\underline{\Sp}^G)_{\underline{W'}} \xtolong{\gamma_{\otimes}}{1.5} (\underline{\Sp}^G)_{\underline{W'}} \]
along
\[ \gamma_* p_! \fib^f_s(\eta) \xtolong{\gamma_* p_! \varphi^f_s}{1.5} \gamma_* p_! S_{\underline{V}} \xtolong{\gamma_* \epsilon_p}{1.5} \gamma_* S_{\underline{W}} \simeq \gamma_* \gamma^* S_{\underline{W'}}, \]
then we may identify $Y'$ as $Y'' \circ \eta_\gamma$. We then need to identify the pullback of $\gamma_* p_! \varphi^f_s$ and $\gamma_* \epsilon_p$ along $\eta_{\gamma}$. To deal with $\gamma_* \epsilon_p$, note that applying $\gamma_*$ to the pullback square of $\underline{W}$-$\infty$-categories
\[ \begin{tikzcd}
p_! p^* S_{\underline{W}} \ar{d} \ar{r}{\epsilon_p} & S_{\underline{W}} \ar{d} \\
\underline{V} \ar{r}{p} & \underline{W}
\end{tikzcd} \]
yields a pullback square of $\underline{W'}$-$\infty$-categories
\[ \begin{tikzcd}
q_! q^* (\gamma_* \gamma^* S_{\underline{W'}}) \ar{r}{\epsilon_q} \ar{d} & \gamma_* S_{\underline{W}} = \gamma_* \gamma^* S_{\underline{W'}} \ar{d} \\
\underline{V'} \ar{r}{q} & \underline{W'}
\end{tikzcd} \]
identifying $\gamma_* \epsilon_p$ with $\epsilon_q$. We then have the commutative diagram
\[ \begin{tikzcd}
q_! q^* S_{\underline{W'}} \ar{r}{q_! q^* \eta_q} \ar{d}{\epsilon_q} & q_! q^* \gamma_* \gamma^* S_{\underline{W'}} \ar{d}{\epsilon_q} \ar{r} & \underline{V'} \ar{d}{q} \\
S_{\underline{W'}} \ar{r}{\eta_q} & \gamma_* \gamma^* S_{\underline{W'}} \ar{r} & \underline{W'}
\end{tikzcd} \]
in which the outer rectangle and the righthand square are pullback squares, hence the lefthand square is a pullback square. Furthermore, under the equivalence $q^* \gamma_* \gamma^* S_{\underline{W'}} \simeq \gamma'_* \gamma'^* S_{\underline{V'}}$ we get that $q_! q^* \eta_q$ identifies with
\[ q_! \eta_{\gamma'}: q_! S_{\underline{V'}} \to q_! \gamma'_* \gamma'^* S_{\underline{V'}}. \]
We next have the sequence of equivalences
\begin{align*} \gamma_* p_! f_* f^* S_{\underline{V}} \simeq q_! \gamma'_* \ev^* f_* f^* S_{\underline{V}} \simeq q_! g_* g^* S_{\underline{V'}}
\end{align*}
where we use \cref{exm:CartesianDistributiveCommutationRelation} for the first equivalence and \cref{rem:BaseChangeEquivalences} for the second equivalence. Moreover, a diagram chase reveals that under this equivalence, $\gamma_* p_! s_f$ identifies with $q_! s'_g$, where $s' = \ev'^*(s) \in S_{U'}$. Putting these observations together, we obtain a commutative diagram
\[ \begin{tikzcd}
q_! \fib^g_{s'}(\eta) \ar{r}{\rho} \ar{d}[swap]{q_! \varphi^g_{s'}} & \gamma_* p_! \fib^f_s(\eta) \ar{r} \ar{d} & \underline{V'} \ar{d}{q_! s'_g} \\
q_! S_{\underline{V'}} \ar{r}{q_! \eta_{\gamma'}} & q_! \gamma'_* \gamma'^* S_{\underline{V'}} \ar{r} & q_! g_* g^* S_{\underline{V'}}
\end{tikzcd} \]
in which the righthand square is $\gamma_* p_!$ of the pullback square defining $\fib^f_s(\eta)$, hence a pullback square, and the outer rectangle is $q_!$ of the pullback square defining $\fib^g_{s'}(\eta)$, hence also a pullback square. It follows that the lefthand square is a pullback square. Finally, if we let $X' = \gamma_{\otimes} \circ \gamma_* X \circ \rho$, then by \cref{lem:adjointability}(3) we get that $Y' \simeq (\epsilon_q \circ q_! \varphi^g_{s'})_! X'$. We conclude that $\gamma_{\otimes} (\cI_W(f,p,s) ) \subset \cI_{W'}(g,q,s')$ and thus the collection $\{ \cI_W \}_{W \in \sO_G}$ is stable under norms.
\end{proof}
On the other hand, to show that $\underline{\Fun}_G(S, \underline{\Sp}^G)_{\ind}$ is fiberwise a $\otimes$-ideal, we will need the following projection formula.
\begin{lem} \label{lem:ProjectionFormula} Suppose $\varphi: I \to J$ is a map of $G$-spaces and let $C$ be a $G$-distributive $G$-symmetric monoidal $\infty$-category.\footnote{In fact, one only needs that for fold maps $p: V^{\sqcup n} \to V$, the norm $\underline{V}$-functor $p_* p^* C_{\underline{V}} \to C_{\underline{V}}$ is distributive. However, this is stronger than $C_V$ being distributive as a symmetric monoidal $\infty$-category.} Then
\[ \varphi_!: \Fun_G(I, C) \to \Fun_G(J, C) \]
satisfies the \emph{projection formula}: that is, for every pair of $G$-functors $X: I \to C$ and $Y: J \to C$, the canonical map
\[ \chi: \varphi_!(X \otimes \varphi^* Y) \to \varphi_! X \otimes Y \]
adjoint to $X \otimes \varphi^* Y \xto{\eta \otimes \id} \varphi^* \varphi_! X \otimes \varphi^* Y$ is an equivalence.
\end{lem}
\begin{proof}
It suffices to check that $\chi$ is an equivalence on all points of $J$. Let $j \in J_V$ and also write $j: \underline{V} \to J$ for the corresponding $G$-functor. We then want to show that $j^* \chi$ is an equivalence of $\underline{V}$-functors. Since $\varphi_!$, $\varphi^*$, and $\otimes$ are all compatible with restriction along $\underline{V} \to \sO_G^{\op}$, we may suppose $V = G/G$ without loss of generality. Now consider the pullback square of $G$-spaces
\[ \begin{tikzcd}
I_j \ar{r}{\iota} \ar{d}{\pi} & I \ar{d}{\varphi} \\
\ast_G \ar{r}{j} & J.
\end{tikzcd} \]
By \cref{lem:adjointability}(3), we have the base-change equivalence $j^* \pi_! \simeq \varphi_! \iota^*$, so we reduce to showing the map
\[ j^* \chi: \pi_! \iota^* (X \otimes \varphi^* Y) \simeq \pi_! (\iota^* X \otimes \pi^* j^* Y) \to \pi_! \iota^* X \otimes j^* Y \]
is an equivalence. For this, we can use the $G$-distributivity of $(C,\otimes)$. Let $p: U = (G/G)^{\sqcup 2} \to G/G$ be the fold map and consider the left $G$-Kan extension diagram
\[ \begin{tikzcd}
I_j \coprod \ast_G \ar{r}{\iota^* X \coprod j^* Y} \ar{d}{\qquad \Downarrow}[swap]{\pi \coprod \id} & [3em] C \coprod C \\
\ast_G \coprod \ast_G \ar[bend right=15]{ru}[swap]{\pi_! \iota^* X \coprod j^* Y}
\end{tikzcd} \]
as a diagram of $\underline{U}$-$\infty$-categories. Then we obtain a left $G$-Kan extension diagram
\[ \begin{tikzcd}
I_j \ar{d}{\qquad \Downarrow}[swap]{\pi} \ar{r}{(\iota^* X, \pi^* j^* Y)} & [3.5em] C \times C = p_* p^* C \ar{r}{\otimes} & C \\
\ast_G \ar[bend right=10]{rru}[swap]{\pi_! \iota^* X \otimes j^* Y},
\end{tikzcd} \]
and thus a canonical equivalence $\pi_! (\iota^* X \otimes \pi^* j^* Y) \xto{\simeq} \pi_! \iota^* X \otimes j^* Y$. A chase of the definitions shows this equivalence to be implemented by $j^* \chi$.
\end{proof}
The following lemma now suffices to show that $\underline{\Fun}_G(S, \underline{\Sp}^G)_{\ind}$ is a $G$-$\otimes$-ideal.
\begin{lem} \label{lem:GTensorIdealCondition}
Let $C$ be a $G$-stable $G$-distributive $G$-symmetric monoidal $\infty$-category and suppose $\{ \cJ_W \subset C_W \}_{W \in \sO_G}$ is a collection of objects stable under restriction, induction, and norms. Let $D_W \subset C_W$ be the thick subcategory generated by $\cJ_W$. Suppose in addition that each $D_W$ is a $\otimes$-ideal. Then for every map of orbits $f: V \to W$, $f_{\otimes}: C_V \to C_W$ restricts to a functor $f_{\otimes}: D_V \to D_W$.
\end{lem}
\begin{proof}
Let $D_V^0 \subset D_V$ be the full stable subcategory consisting of objects $x \in D_V$ such that $f_{\otimes}(x) \in D_W$. We claim that $D_V^0$ is a thick subcategory, for which it suffices to show that $D_V^0$ is closed under desuspensions, finite colimits, and retracts.\footnote{Of course, a simple argument as in \cref{lem:GeneratorsDefineGStableSubcat} no longer works since $f_{\otimes}$ is not an exact functor.} We deal with these cases in turn:
\begin{enumerate}[leftmargin=*]
\item Suppose $x \in D_V^0$. Then $\Sigma^{-1} x = S^{-1} \otimes x$ where $S^{-1} = \Sigma^{-1}(1)$ is the desuspension of the unit in $C_V$. Since $f_{\otimes}$ is symmetric monoidal (\cref{rem:NormSymmetricMonoidal}), we have $f_{\otimes}(\Sigma^{-1} x) \simeq f_{\otimes}(S^{-1}) \otimes f_{\otimes}(x)$. But by assumption $f_{\otimes}(x) \in D_W$ and $D_W \subset C_W$ is a $\otimes$-ideal. It follows that $f_{\otimes}(\Sigma^{-1} x ) \in D_W$ and thus $\Sigma^{-1} x \in D_V^0$.
\item To show that $D_V^0$ is stable under finite colimits, in order to use the $G$-distributivity of $(C, \otimes)$ it is convenient to establish a more general assertion regarding parametrized colimits. To this end, let $D \subset C$ be the $G$-stable full $G$-subcategory as in \cref{lem:GeneratorsDefineGStableSubcat}. For any map of orbits $W' \to W$, consider the pullback square
\[ \begin{tikzcd}
V' \ar{r}{f'} \ar{d} & W' \ar{d} \\
V \ar{r}{f} & W
\end{tikzcd} \]
and let $D^0_{V'} \subset D_{V'}$ be the full subcategory on those objects $x \in D_{V'}$ such that $f'_{\otimes}(x) \in D_{W'}$. Then using the compatibility of norms with restriction, we see that the collection
$$\{ D^0_{V'} \subset D_{V'} : V' \cong V \times_W W' \}_{W' \in (\FF_G)^{/W}}$$
assembles to define a full $\underline{W}$-subcategory $(f_* D_{\underline{V}})^0$ of $f_* D_{\underline{V}}$.
Now suppose that $K$ is a $\underline{V}$-finite $\underline{V}$-$\infty$-category and $\varphi: K \to D_{\underline{V}}$ is a $\underline{V}$-functor such that $f_* \varphi: f_* K \to f_* D_{\underline{V}}$ factors through $(f_* D_{\underline{V}})^0$; if $K \simeq L \times \underline{V}$ for a finite $\infty$-category $L$, then this hypothesis is equivalent to supposing that $\varphi_V: L \to D_V$ factors through $D^0_V$, so this is the case of interest for us. We aim to show that $\colim^{\underline{V}} \varphi \in D_V^0$. But using $G$-distributivity, we see that
\[ f_{\otimes}(\colim^{\underline{V}} \varphi) \simeq \colim^{\underline{W}}(f_{\otimes} \circ f_* \varphi: f_* L \to f_* D_{\underline{V}} \subset f_* f^* C_{\underline{W}} \to C_{\underline{W}}). \]
Since $f_* \varphi$ factors through $(f_* D_{\underline{V}})^0$, $f_{\otimes} \circ f_* \varphi$ factors through $D_{\underline{W}}$. Then since $f_* L$ is $\underline{W}$-finite and the inclusion $D_{\underline{W}} \subset C_{\underline{W}}$ is $\underline{W}$-exact, we get that $\colim^{\underline{W}}(f_{\otimes} \circ f_* \varphi) \in D_W$, as desired.
\item Since $f_{\otimes}$ preserves retract diagrams and $D_W$ is stable under retracts, it follows that $D^0_V$ is stable under retracts.
\end{enumerate}
We conclude that $D^0_V$ is a thick subcategory. Since $\cJ_W \subset D^0_V$ by assumption, it follows that $D^0_V = D_V$, as desired.
\end{proof}
\begin{cor} \label{cor:InducedObjectsFormGTensorIdeal}
Let $S$ be a $G$-space. Then the full $G$-subcategory $\underline{\Fun}_G(S, \underline{\Sp}^G)_{\ind}$ of induced objects in $\underline{\Fun}_G(S, \underline{\Sp}^G)$ is a $G$-$\otimes$-ideal.
\end{cor}
\begin{proof}
By a routine thick subcategory argument, \cref{lem:ProjectionFormula} implies that $\underline{\Fun}_G(S, \underline{\Sp}^G)_{\ind}$ is fiberwise a $\otimes$-ideal. Using \cref{lem:ClosureUnderNorms}, \cref{lem:GTensorIdealCondition} then applies to show that $\underline{\Fun}_G(S, \underline{\Sp}^G)$ is a $G$-$\otimes$-ideal.
\end{proof}
\subsection{Main results}\label{SS:MainAssembly}
We are now ready to define a genuine equivariant refinement of the Tate construction for infinite groups. Our discussion is parallel to that in \cite[\S I.4]{NS18}. First, we generalize work of Weiss--Williams \cite{WW95} by constructing parametrized assembly maps.
\begin{thm} \label{thm:paramTateGeneral}
Let $G$ be a finite group, $S$ a $G$-space, and let $p: S \to \ast_G$ be the projection to the terminal $G$-space.
\begin{enumerate}
\item The $G$-stable $G$-$\infty$-category $\underline{\Fun}_G(S, \underline{\Sp}^G)$ has compactly generated fibers, and taking compact objects fiberwise yields a $G$-stable full $G$-subcategory $\underline{\Fun}_G(S, \underline{\Sp}^G)^{\omega}$ such that
$$ \underline{\Ind}_G \left( \underline{\Fun}_G(S, \underline{\Sp}^G)^{\omega} \right) \xto{\simeq} \underline{\Fun}_G(S, \underline{\Sp}^G). $$
\item For every $G$-exact $G$-functor $F_1: \underline{\Fun}_G(S, \underline{\Sp}^G) \to \underline{\Sp}^G$, there exists a $G$-colimit preserving $G$-functor $F_0$ equipped with a natural transformation $\alpha: F_0 \to F_1$, which is an `assembly map' in the sense of being terminal among all such natural transformations. Moreover, $\alpha$ is an equivalence when restricted to the full $G$-subcategory $\underline{\Fun}_G(S, \underline{\Sp}^G)^{\omega}$ of compact objects, and this uniquely specifies $\alpha$.
\item Every $G$-colimit preserving functor $F: \underline{\Fun}_G(S, \underline{\Sp}^G) \to \underline{\Sp}^G$ is of the form $$F(-) \simeq p_!(D \otimes -)$$ for a uniquely determined $G$-functor $D: S \to \underline{\Sp}^G$.
\item For $F = p_*$, the assembly map takes the form $p_!(D_S \otimes -) \to p_*(-)$ for
\[ D_S: S \xto{j} \underline{\Fun}_G(S, \underline{\Spc}^G) \xto{\Sigma^{\infty}_+} \underline{\Fun}_G(S, \underline{\Sp}^G) \xto{p_*} \underline{\Sp}^G \]
where $j$ is the $G$-Yoneda embedding (under the equivalence $S \simeq S^{\vop}$).
\end{enumerate}
\end{thm}
\begin{proof} \begin{enumerate}[leftmargin=*]
\item Since $G$-colimits and $G$-limits are computed pointwise [Shah, Prop.~9.17], $\underline{\Fun}_G(S, \underline{\Sp}^G)$ is $G$-cocomplete and $G$-complete; it moreover has presentable fibers by \cite[Prop.~5.4.7.11]{HTT}. Now for every orbit $V \cong G/K$ and point $s \in S_{V}$, also write $s: \underline{V} \to S$ for the corresponding $G$-functor that selects $s$. We then have an adjunction
\[ \adjunct{s_!}{\Sp^K \simeq \Fun_G(\underline{V}, \underline{\Sp}^G)}{\Fun_G(S, \underline{\Sp}^G)}{s^*}, \]
and by the same argument as in the proof of \cite[Thm~I.4.1(i)]{NS18}, the set $\{ s_!(\mathbb{S}) : s \in S_{V} \}$ furnishes a set of compact generators for $\Fun_G(S, \underline{\Sp}^G)$. The argument for compact generation of the other fibers is identical.
Taking compact objects fiberwise then yields a $G$-stable full $G$-subcategory since restriction and induction are strongly cocontinuous, and the last claim is clear by definition of $\underline{\Ind}_G$ as fiberwise Ind-completion.
\item For any small $G$-stable $G$-$\infty$-category $C$ and $G$-cocomplete $G$-stable $G$-$\infty$-category $D$, since $G$-colimits decompose in terms of fiberwise filtered colimits and finite $G$-colimits, one has a colocalization adjunction \cite[Thm.~9.11]{Exp2b}
\[ \begin{tikzcd}
\Fun^{\ex}_G(C,D) \simeq \Fun^L_G(\underline{\Ind}_G C, D) \ar[hookrightarrow, shift left=.5ex]{r} & \Fun^{\ex}_G(\underline{\Ind}_G C, D) \ar[shift left=.5ex]{l}
\end{tikzcd} \]
with right adjoint given by restriction along the fiberwise Yoneda embedding and left adjoint given by fiberwise left Kan extension (which agrees in this case with $G$-left Kan extension). If we then let $C = \underline{\Fun}_G(S, \underline{\Sp}^G)^{\omega}$ and $D = \underline{\Sp}^G$, we may define the assembly map to be the counit of the adjunction, after which its claimed properties are immediate.
\item We have the equivalences
\begin{align*}
\Fun^L_{G}(\underline{\Fun}_G(S, \underline{\Sp}^G), \underline{\Sp}^G) & \simeq \Fun^L_{G}(\underline{\Fun}_G(S, \underline{\Spc}^G), \underline{\Sp}^G) \\
& \simeq \Fun_{G}(S, \underline{\Sp}^G)
\end{align*}
implemented by restriction along $\Sigma^\infty_+$ and the $G$-Yoneda embedding, respectively; the first equivalence follows from [Nardin, Theorem~7.4] and the second equivalence from the parametrized Yoneda lemma [Shah, Theorem~11.5]. The claim then follows by observing that $p_! (D \otimes -)$ restricts to $D$ under this equivalence.
\item The proof is identical to that of \cite[Thm.~I.4.1(iv)]{NS18}.
\end{enumerate}
\end{proof}
\begin{dfn} \label{dfn:paramTateGeneral}
In the situation of \Cref{thm:paramTateGeneral}, we define $p^T_*$ by extending the assembly map to a cofiber sequence
\[ p_!(D_S \otimes -) \xto{\alpha} p_*(-) \xto{\beta} p^T_*. \]
If $S = B^\psi_{G} K$ for a group extension $\psi$ of the finite group $G$ by a (possibly infinite) group $K$, then we will also write $(-)^{\underline{t}[\psi]}$ or $(-)^{\underline{t}_{G} K}$ for the $G$-functor $p^T_*$ and $(-)^{t[\psi]}$ or $(-)^{t_G K}$ for the fiber of $p^T_*$ over $G/G$.
\end{dfn}
\begin{rem}[Compatibility with restriction]
Suppose that $V \cong G/H$ is an orbit and let $\psi' = \psi|_{H}$. Then $(B^{\psi}_G K)_{\underline{V}} \simeq B^{\psi'}_{H} K$ and the fiber of $(-)^{\underline{t}_G K}$ over $V$ identifies with $(-)^{t_H K}$.
\end{rem}
\begin{rem}[Agreement of definitions] \label{rem:agreement}
Suppose that $\psi$ is an extension of $G$ by a \emph{finite} group $K$. Then the $G$-functor $(-)^{\underline{t}[\psi]}$ of \cref{dfn:paramTateGeneral} coincides with that in \cref{rem:ParamTateCompatibleRestriction} and hence the functor $(-)^{t[\psi]}$ coincides with that defined prior in \cref{dfn:ParamTateCnstr}. To see this, by \cref{thm:paramTateGeneral}(2) it suffices to show that the norm map $(-)_{\underline{h}[\psi]} \to (-)^{\underline{h}[\psi]}$ constructed via the parametrized ambidexterity theory (for the Beck-Chevalley fibration $\underline{\LocSys}^{G}(\underline{\Sp}^G) \to \Spc^{G}$) is an equivalence on compact objects. This follows from \cref{NormVanishesOnInduced} and the description of the compact generators given in \cref{thm:paramTateGeneral}(1).
\end{rem}
\begin{rem}[Point-set models]
As when all groups were finite, the parametrized Tate construction for a compact Lie group $K$ admits a point-set model using equivariant universal spaces, cf. \cref{rem:PointSetModels}.
\end{rem}
For the following pair of propositions, we refer to $p^T_*$ defined with respect to a fixed $G$-space $S$. First, in direct analogy to \cite[Thm.~I.4.1(iii)]{NS18}, we have the following universal property of $p_*^T$.
\begin{prp} \label{prp:UniversalPropertyTate}
The natural transformation $\beta: p_* \to p_*^T$ is initial among those natural transformations from $p_*$ to $G$-exact $G$-functors whose target vanishes on compact objects fiberwise.
\end{prp}
\begin{proof}
We first note that $p_*$ is fiberwise $\kappa$-accessible for some fixed regular cardinal $\kappa$ by the adjoint functor theorem [HTT, Cor.~5.5.2.9]; indeed, for every orbit $V$, $(p_*)_V$ is $\kappa_V$-accessible as a right adjoint between presentable $\infty$-categories, and we may then let $\kappa = \sup_{V \in \sO_G} \kappa_V$. Since the fiber of $\beta$ preserves all $G$-colimits by definition, it follows that $p_*^T$ is also fiberwise $\kappa$-accessible.
Now by \Cref{prp:VerdierQuotient}(4) applied to $\underline{\Fun}_G(S, \underline{\Sp}^G)^{\omega} \subset \underline{\Fun}_G(S, \underline{\Sp}^G)^{\kappa}$, there exists a natural transformation $\beta': p_* \to L(p_*)$ of (fiberwise $\kappa$-accessible) $G$-exact $G$-functors with the indicated universal property. We then have a map of fiber sequences
\[ \begin{tikzcd}
R(p_*) \ar{r}{\alpha'} \ar{d}{\alpha''} & p_* \ar{r}{\beta'} \ar{d}{=} & L(p_*) \ar{d}{\beta''} \\
p_!(D_S \otimes -) \ar{r}{\alpha} & p_* \ar{r}{\beta} & p_*^T.
\end{tikzcd} \]
We want to show that $\beta''$ is an equivalence, for which it suffices to show that $\alpha''$ is an equivalence. Since $\underline{\Fun}_G(S, \underline{\Sp}^G)$ is fiberwise compactly generated and $\alpha''$ is an equivalence when restricted to fiberwise compact objects, this will follow once we show that $R(p_*)$ preserves fiberwise filtered colimits. But this follows as in the proof of \cite[Thm.~I.4.1(iii)]{NS18}.
\end{proof}
We next discuss the interaction of parametrized assembly with $G$-symmetric monoidal structures. Note first that with respect to the pointwise $G$-symmetric monoidal structure of \Cref{exm:pointwiseGSMC} on $\underline{\Fun}_G(S, \underline{\Sp}^G)$, $p^*: \underline{\Sp}^G \to \underline{\Fun}_G(S, \underline{\Sp}^G)$ is a $G$-symmetric monoidal functor, hence its right $G$-adjoint $p_*$ is lax $G$-symmetric monoidal. For $p_*^T$, we then have the following proposition.
\begin{prp} \label{prp:MonoidalStructureTate}
Suppose that $p_*^T$ restricts to the zero functor on the full $G$-subcategory of induced objects. Then $p_*^T$ and the natural transformation $\beta: p_* \to p_*^T$ uniquely inherit the structure of a lax $G$-symmetric monoidal functor and morphism thereof.
\end{prp}
\begin{proof}
As in the proof of \Cref{prp:UniversalPropertyTate}, we may as well ignore size-theoretic issues in our proof. Now by \Cref{prp:VerdierQuotientInducedObjects} applied with respect to the $G$-$\otimes$-ideal of induced objects (cf. \cref{cor:InducedObjectsFormGTensorIdeal}), there exists a natural transformation $\beta': p_* \to L(p_*)$ of lax $G$-symmetric monoidal functors such that $L(p_*)$ vanishes on $G$-induced objects and $\beta'$ is the initial map with respect to this property. Moreover, since $\underline{\Fun}_G(S, \underline{\Sp}^G)^{\omega} \subset \underline{\Fun}_G(S, \underline{\Sp}^G)_{\ind}$ by \cref{lem:CompactObjectsAreInduced}, after forgetting $G$-symmetric monoidal structure we obtain a comparison map $p_*^T \to L(p_*)$ compatible with the maps from $p_*$. By our assumption on $p_*^T$, we likewise obtain a map $L(p_*) \to p_*^T$, and these are easily seen to be mutually inverse equivalences. We may thus equip $\beta$ (and hence $p_*^T$) with the indicated lax $G$-symmetric monoidal structure. Finally, the uniqueness assertion follows from the universal property of $\beta'$.
\end{proof}
\begin{lem} \label{lem:CompactObjectsAreInduced}
We have an inclusion
\[ \underline{\Fun}_G(S, \underline{\Sp}^G)^{\omega} \subset \underline{\Fun}_G(S, \underline{\Sp}^G)_{\ind}. \]
\end{lem}
\begin{proof}
We will show this inclusion on the fiber over $G/G$; the argument for the other fibers will then be identical. Since both sides are thick subcategories of $\Fun_G(S, \underline{\Sp}^G)$, it suffices to check that the compact generators described in the proof of \cref{thm:paramTateGeneral}(1) are induced. So let $V$ be an orbit and $s \in S_V$ a point. Then if we let $f = \id_V$ and $p: V \to G/G$ be the unique map, the $G$-functor
$$\epsilon_p \circ p_!(\varphi^f_s): p_! \fib^f_s(\eta) \to p_! p^* S \to S$$
in \cref{dfn:InducedObjects} coincides with $s: \underline{V} \to S$. Therefore, $s_! (\SS)$ is also an induced generator.
\end{proof}
We now specialize our discussion to the main case of interest. Let $K$ be a compact Lie group, $G$ a finite group, $\psi: G \to \Aut(K)$ a group homomorphism, and $S = B^{\psi}_G K$.
\begin{obs}[Parametrized actions on {$K$} and its Lie algebra] \label{obs:ActionsLieGroup}
Let $G$ act on $K \times K$ by $\psi \times \psi$ and consider the unique point-set action of $(K \times K) \rtimes G$ on $K$ that extends the $K \times K$-action given by $(k,k')\cdot x = k x k'^{-1}$ and the $G$-action given by $\psi$, so that $((k,k'),g) \cdot x = k \psi_g(x) k'^{-1}$.\footnote{This action is well-defined since $(1,g) \cdot ((k,k'),1) \cdot x = \psi_g(k x k'^{-1}) = \psi_g(k) \psi_g(x) \psi_g(k')_{-1} = ((\psi_g(k), \psi_g(k')), g) \cdot x$.} This defines $K$ first as an object in
$$\Spc^{(K \times K) \rtimes G} \simeq \Fun(\cO_{(K \times K) \rtimes G}^{\op}, \Spc)$$
and then as an object in
\[ \Fun_G(B^{\psi}_G (K \times K), \underline{\Spc}^G) \simeq \Fun(B^{\psi}_G (K \times K), \Spc) \]
via restriction along $B^{\psi}_G (K \times K) \subset \cO_{(K \times K) \rtimes G}^{\op}$. Let
$$\Sigma^{\infty}_+ K \in \Fun_G(B^{\psi}_G (K \times K), \underline{\Sp}^G)$$
denote the infinite $G$-suspension of $K$. Then the dualizing object $D_S$ of \cref{thm:paramTateGeneral}(4) identifies with $(\Sigma^{\infty}_+ K)^{h_G (K \times 1)}$.
Now let $\mathfrak{a}$ be the Lie algebra of $K$ and let $\mathbb{S}^{\mathfrak{a}}$ be $\Sigma^{\infty}$ of the one-point compactification $S^{\mathfrak{a}}$. We have that the $K \rtimes G$-action on $K$, uniquely specified as the action extending $k \cdot x = k x k^{-1}$ and $\psi$, defines a point-set $K \rtimes G$-action on $S^{\mathfrak{a}}$ and subsequently $\mathbb{S}^{\mathfrak{a}}$ as an object in $\Fun_G(B^{\psi}_G K, \underline{\Sp}^G)$, where (abusing notation) we also write $\mathbb{S}^{\mathfrak{a}}$ for the infinite $G$-suspension of $S^{\mathfrak{a}}$, which restricts to the prior spectrum $\mathbb{S}^{\mathfrak{a}}$.
\end{obs}
Let $S_0$ be the underlying space of $S$ and recall that John Klein identified the dualizing spectrum $D_{S_0}$ with $\mathbb{S}^{\mathfrak{a}}$ as a spectrum with $K$-action \cite[Thm.~10.1]{Klein2001}.
\begin{prp} \label{prp:DualizingSpectrum}
Klein's equivalence promotes to an equivalence
\[ D_S = (\Sigma^{\infty}_+ K)^{h_G (K \times 1)} \simeq \mathbb{S}^{\mathfrak{a}} \in \Fun_G(B^{\psi}_G K, \underline{\Sp}^G). \]
\end{prp}
\begin{proof}
This follows by inspecting the equivariance of the maps in Klein's proof along with invoking equivariant Atiyah duality in $\Sp^G$. In more detail, first note that Klein's point-set $K \times K$-equivariant maps\footnote{Note that Klein writes $G$ and not $K$ for the compact Lie group.} $\alpha: K_+ \to F(K_+, S^{\mathfrak{a}})$ and its adjoint $\widehat{\alpha}: (K \times K)_+ \to S^{\mathfrak{a}}$ are in fact $(K \times K) \rtimes G$-equivariant.\footnote{By choosing an invariant metric, we may assume that $G$ acts by isometries on $\mathfrak{a}$.} Suspending $\widehat{\alpha}$ and taking its adjoint, we then obtain a map
$$ \Sigma^{\infty}_+ K \to F(\Sigma^{\infty}_+ K, \mathbb{S}^{\mathfrak{a}}) \in \Fun_G(B^{\psi}_G (K \times K), \underline{\Sp}^G) $$
which is an equivalence by equivariant Atiyah duality. Taking $G$-parametrized homotopy fixed points with respect to the factor $K \times 1$ then shows the claim.
\end{proof}
\begin{exm}[Complex conjugation on the circle] \label{exm:CircleTate}
Suppose that $K=S^1$ with $G=C_2$ acting by complex conjugation, so that $K \rtimes G \cong O(2)$. Write $S = B^t_{C_2} S^1$ for the corresponding $C_2$-space. Since $B S^1 = \Omega^{\infty} H \ZZ[2]$, a group cohomology computation shows that $(B S^1)^{h C_2} \simeq B \mu_2$ and the map $(B S^1)^{h C_2} \to B S^1$ identifies with the delooping of the inclusion $\mu_2 = \{ \pm 1 \} \subset S^1$. Note also that we have a canonical isomorphism $\mu_2 \cong W_{O(2)} C_2$ of $\mu_2$ with the Weyl group of $C_2$.\footnote{We may choose any section $C_2 \to O(2)$ of the determinant map to view $C_2$ as a subgroup of $O(2)$, but also note that under the isomorphism $O(2) = S^1 \rtimes C_2$ one has a preferred section.} We then see that a $C_2$-functor $X: B^t_{C_2} S^1 \to \underline{\Sp}^{C_2}$ (with underlying $C_2$-spectrum $X$) is equivalent data to:
\begin{enumerate}
\item An $O(2)$-action on the underlying spectrum $X^{e}$.
\item A $\mu_2$-action on the geometric $C_2$-fixed points $X^{\phi C_2}$ such that the gluing map $X^{\phi C_2} \to (X^e)^{t C_2}$ is $\mu_2$-equivariant.
\end{enumerate}
In particular, if we let $D_S$ be the dualizing $C_2$-spectrum and $\sigma$ denote the sign $C_2$-representation, then under the equivalence of \cref{prp:DualizingSpectrum} we see that $D_S = \SS^{\sigma}$ with \emph{trivial} $C_2$-parametrized action, i.e., so that as a $C_2$-functor, $D_S$ factors as
\[ D_S: B^t_{C_2} S^1 \to \ast_{C_2} \xto{\SS^{\sigma}} \underline{\Sp}^{C_2}. \]
Therefore, we may write $D_S \otimes X$ as the $C_2$-tensor of $X$ by the $C_2$-spectrum $S^{\sigma}$. It follows that we may write the parametrized assembly map as
\[ (D_S \otimes X)_{h_{C_2} S^1} \simeq \SS^{\sigma} \otimes X_{h_{C_2} S^1} = \Sigma^{\sigma} X_{h_{C_2} S^1} \to X^{h_{C_2} S^1}. \]
The appearance of $\SS^{\sigma}$ here is closely related to the suspensions by the signed sphere appearing explicitly in \cite[Thm.~C]{Hog16} and implicitly in \cite[Thm.~B]{DMP21}.
\end{exm}
Finally, we establish lax equivariant symmetric monoidality of the parametrized Tate construction by showing that it vanishes on induced objects.
\begin{thm} \label{thm:TateVanishesOnInduced}
Let $K$ be a compact Lie group, $G$ a finite group, and $\psi: G \to \Aut(K)$ a group homomorphism. Then $(-)^{\underline{t}_G K}$ vanishes on the full $G$-subcategory of induced objects.
\end{thm}
\begin{proof}
We first recall the proof that $(-)^{t K}$ vanishes on induced objects (cf. \cite[Cor.~10.2]{Klein2001}). Since $(-)^{t K}$ is an exact functor, by a thick subcategory argument it suffices to show that $(-)^{t K}$ vanishes on induced generators. Given a point $s: \ast \to B K$ and writing $p: B K \to \ast$ for the unique functor to a point, we thus need to show that the assembly map
\[ \alpha_E: p_!(\SS^{\mathfrak{a}} \otimes s_! E) \to p_* s_! E \]
is an equivalence for all spectra $E$. By definition, $\alpha_E$ is an equivalence when $E$ is a compact spectrum, so it suffices to show that both sides commute with filtered colimits. For the lefthand side, this is obvious, while for the righthand side, we note that
\[ p_* s_! (-) = (\Sigma^{\infty}_+ K \otimes -)^{h K} \simeq F(D K_+, -)^{h K} \simeq F(K_+, \Sigma^{\mathfrak{a}} -)^{h K} \simeq \SS^{\mathfrak{a}} \otimes - \]
(invoking Atiyah duality for the middle equivalence), so $p_* s_!$ commutes with all colimits.
Now let $S = B^{\psi}_G K$ and write $\rho: S \to \ast_G$ for the unique $G$-functor (we don't use $p$ to avoid confusion with the notation of \cref{dfn:InducedObjects}). Since $(-)^{\underline{t}_G K}$ is fiberwise exact, it suffices to show that $(-)^{\underline{t}_G K}$ vanishes on induced generators. Moreover, since $(-)^{\underline{t}_G K}$ is $G$-exact, for any map of orbits $h: W \to W'$ and $X \in \Fun_{\underline{W}}(S_{\underline{W}}, (\underline{\Sp}^G)_{\underline{W}})$, if $\underline{(X)}^{t_G K} = 0$ then $\underline{(h_! X)}^{t_G K} = 0$. Therefore, considering induced generators as in the setup of \cref{dfn:InducedObjects}, we may suppose that $p = \id_V$ there.\footnote{The case where $V$ is not an orbit also follows since that only involves in addition taking fiberwise coproducts.} Furthermore, without loss of generality we may suppose $V = G/G$. So let $U$ be a finite $G$-set, let $s \in S_U$ be a point, and write $f: U \to G/G$ for the unique map. Consider the pullback square of $G$-$\infty$-categories
\[ \begin{tikzcd}
\fib^f_s(\eta) \ar{r}{\varphi} \ar{d}{\pi} & S \ar{d}{\eta} \\
\ast_G \ar{r}{s'} & f_* f^* S.
\end{tikzcd} \]
We need to show that for any $G$-functor $X: \fib_s^f(\eta) \to \underline{\Sp}^G$, the parametrized assembly map
\[ \alpha_{X}: \rho_!(\SS^{\mathfrak{a}} \otimes \varphi_! X) \to \rho_* \varphi_! X \]
is an equivalence in $\Sp^G$. We first note that by definition $\alpha_{X}$ is an equivalence if $X$ is compact. Since $\Fun_G(\fib^f_s(\eta), \underline{\Sp}^G)$ is compactly generated by \Cref{thm:paramTateGeneral}(1), it suffices to check that both sides commute with filtered colimits as functors $\Fun_G(\fib_s^f(\eta), \underline{\Sp}^G) \to \Sp^G$. For the lefthand side this is obvious, while for the righthand side we note that since $K$ is compact Lie it follows that $\fib_s^f(\eta)$ is a \emph{finite} $G$-space, hence
\[ \eta_* \varphi_! \simeq s'_! \pi_*: \Fun_G(\fib_s^f(\eta), \underline{\Sp}^G) \to \Fun_G(f_* f^* S, \underline{\Sp}^G), \]
and writing $\rho': f_* f^* S \to \ast_G$ for the unique $G$-functor, we get that
\[ \rho_* \varphi_! \simeq \rho'_* \eta_* \varphi_! \simeq \rho'_* s'_! \pi_* : \Fun_G(\fib_s^f(\eta), \underline{\Sp}^G) \to \Sp^G. \]
Now again using that $\fib_s^f(\eta)$ is finite, we get that $\pi_*$ commutes with filtered colimits, and it remains to check that $\rho'_* s'_!$ commutes with filtered colimits. For this, note first that if $U = G/G$, then by equivariant Atiyah duality applied as in \cref{prp:DualizingSpectrum} we would have
\[ \rho'_* s'_! (-) \simeq \SS^{\mathfrak{a}} \otimes (-). \]
In general, we may identify the indexed product $f_* f^* S = \prod_U B^{\psi}_G K$ as $B^{\psi_U}_G (\prod_{|U|} K)$ for the group homomorphism
\[ \psi_U: G \to \Aut\left(\textstyle\prod_{|U|} K\right), \quad (\psi_U)_g: (k_i)_{i \in U} \mapsto (\psi_g(k_{g^{-1} \cdot i}) )_{i \in U} \]
and thereby reduce to the logic of the prior case since $\prod_{|U|} K$ is still compact Lie.
\end{proof}
\begin{cor} \label{cor:LaxGSymmetricMonoidalTate}
Let $K$ be a compact Lie group, $G$ a finite group, and $\psi: G \to \Aut(K)$ a group homomorphism. Then the $G$-functor $(-)^{\underline{t}_G K}$ and natural transformation $(-)^{\underline{h}_G K} \to (-)^{\underline{t}_G K}$ uniquely inherit the structure of a lax $G$-symmetric monoidal functor and morphism thereof.
\end{cor}
\begin{proof}
Combine \cref{prp:MonoidalStructureTate} and \cref{thm:TateVanishesOnInduced}.
\end{proof}
|
\section{Introduction}
Wave turbulence theory (WTT) is a framework for describing the long-time statistical behavior of fields of many weakly-interacting waves. In such a system, over scales far from those of forcing and dissipation, self-similar wave-wave interactions drive an energy cascade between scales and lead to the development of a power-law spectrum. While wave turbulence shares much phenomenology with hydrodynamic turbulence, WTT enables an analytic treatment of the governing wave equation through which the evolution equation of the wave spectrum can be derived, yielding as a stationary solution the full functional form of the power-law spectrum. Due to the ubiquity of nonlinear waves in nature, WTT has found use in a diverse array of fields, including magnetohydrodynamics \citep{galtier_weak_2000,galtier_weak_2014}, physical oceanography \citep{zakharov_weak_1967,zakharov_stability_1968}, acoustics \citep{lvov_statistical_1997}, astrophysics \citep{galtier_direct_2021}, and others.
One of the primary results of WTT is the derivation of the Wave Kinetic Equation (WKE), which expresses the time evolution of wave action spectral density as an integral over wave-wave interactions. The WKE has an inertial-range stationary solution of wave action $n(k)\sim P^{\theta}k^{\gamma}$, where $k$ is the wave number, $P$ is the energy flux of the forward cascade, and $\theta$ and $\gamma$ are scaling exponents. Over the decades, many efforts have been made to numerically and experimentally study the scaling exponent $\gamma$ for a wide variety of physical systems \cite[e.g.][]{nazarenko_wave_2006,denissenko_gravity_2007,miquel_role_2014,during_wave_2017,hassaini_confinement_2018,monsalve_quantitative_2020}. The exponent $\theta$, and in general the properties of $P$, are however much less studied, despite the fact that they are more relevant to the formulation of the WKE. A small number of existing works on the scaling of $P$ \cite[e.g.][]{falcon_observation_2007,deike_energy_2014,pan_direct_2014} sometimes produce inconsistent conclusions, partly because of their indirect and inconsistent measurements of $P$ based on the energy input/dissipation rate that are inevitably complicated by the non-stationary spectrum and broadscale dissipation of the wave field \cite[e.g.][]{deike_energy_2014,pan_decaying_2015}.
A more direct approach for exact evaluation of $P$ can be formulated through the nonlinear terms in the governing equation which are responsible for the wave-wave interactions leading to the energy cascade \cite[e.g.][]{hrabski_effect_2020}. This approach allows the resolution of the probability distribution of $P$ at arbitrary scales that is not obtainable by previous input/dissipation-based methods. Moreover, this formulation enables a decomposition of the energy flux into contributions from exact and quasi resonances. In particular, we can perform a quartet-level decomposition of $P=\sum_{\Omega} P_{\Omega}$, with $P_{\Omega}$ representing contributions from a set of quartets with frequency mismatch $\Omega$. This decomposition technique will allow us to elucidate many mechanisms underlying the scaling and distribution of $P$, and provide a direct measure of nonlinear broadening by evaluating the contribution of quasi-resonances to the energy cascade. This new measure of nonlinear broadening can be more direct and physically intuitive than previous approaches based on the coherence function \cite[e.g.][]{aubourg_nonlocal_2015,pan_understanding_2017,zhang_numerical_2021} or the broadened dispersion relation \cite[e.g.][]{mordant_fourier_2010,deike_direct_2014}.
Another study enabled by this decomposition technique is one on the WTT closure model, which relates the high-order correlators (of frequency mismatch $\Omega$) to the product of pair correlators. This relation is usually argued in conjunction with a broadening function $f(\Omega)$ that approaches the delta function $\delta(\Omega)$ at the kinetic limit (of a large domain and small nonlinearity) \citep{nazarenko_wave_2011,zakharov_kolmogorov_2012,buckmaster_onset_2021,deng_full_2021,deng_derivation_2021}. We are particularly interested in the form of $f(\Omega)$ in a finite domain, which is more relevant to the situations in many numerical simulations and experiments. Interesting attempts in this direction include the development of a generalized kinetic equation (by implementing a numerical solution to the closure) \citep{annenkov_role_2006}, which however does not focus on the functional form $f(\Omega)$ in the stationary state of wave turbulence. A lack of understanding of the closure problem is a major obstacle for the advancement in the theory of wave turbulence, e.g., the obvious contradiction between the MMT closure and WTT closure (as well as the elusive numerical observations) raised more than 20 years ago is still not fully understood today \citep{majda_one-dimensional_1997,cai_spectral_1999,zakharov_wave_2001}.
The purpose of this paper, in general, is to establish a methodology such that all aforementioned analysis (including various properties of energy flux and the closely-related closure problem) can be directly studied using numerical data from simulations of the governing equations. We demonstrate our methodology in the context of the two-dimensional Majda-McLaughlin-Tabak (MMT) equation, but envision applications to much broader systems in wave turbulence. In a stationary state of MMT turbulence, our analysis shows that the energy flux $P$, as a time series, closely follows a Gaussian distribution, with a standard deviation of several times its mean value $\overline{P}$. The large standard deviation of $P$ is found to be dominated by fluctuations in time of the quasi-resonant components $P_{\Omega \neq 0}$. The scaling between the spectral level and $\overline{P}$ shows a $\overline{P}^{1/3}$ dependence at high nonlinearity levels and transits to $\overline{P}^{1/2}$ dependence at low nonlinearity levels. The $\overline{P}^{1/3}$ scaling is consistent with the kinetic scaling and is established when $\overline{P}$ is dominated by the quasi-resonant contributions. This is a remarkable result considering the fact that the WKE, based on which the kinetic scaling is developed, is formulated on the exact resonant manifold. The $\overline{P}^{1/2}$ scaling is consistent with the dynamic scaling derived from the MMT equation, established as a result of the dominance of exact resonances in $\overline{P}$. Finally, the WTT closure model on the fourth-order correlator, when evaluated in a time window of $O(500)$ fundamental periods, is found to be not valid on the individual quartet level. When we increase the number of quartets to the level that $P$ is formulated ($O(10^9)$ number of quartets for each $\Omega$), the average behavior of the fourth-order correlator lies much closer to WTT closure, but with $f(\Omega)\sim 1/\Omega^\beta$ (with $\beta$ between 1.3 and 1.6) observed instead of the previously argued functions (e.g., $sinc(\Omega)$) from WTT.
\section{Formulation of Energy Flux for the MMT Model}
We consider a two-dimensional (2D) MMT model \citep{majda_one-dimensional_1997} which has been widely used to study wave turbulence problems \citep{cai_spectral_1999,sheffield_ensemble_2017,chibbaro_weak_2017,hrabski_effect_2020}. The model describes the evolution of a complex scalar $\psi(\boldsymbol{x},t)$:
\begin{equation}
i\frac{\partial \psi}{\partial t} = |\partial_{\boldsymbol{x}}|^\alpha\psi + |\partial_{\boldsymbol{x}}|\left(\big| |\partial_{\boldsymbol{x}}|\psi\big|^2|\partial_{\boldsymbol{x}}|\psi\right),
\label{eqn:MMT}
\end{equation}
where $\boldsymbol{x}$ is the 2D spatial coordinates and $t$ is time, and the operator $|\partial_{\boldsymbol{x}}|^{\alpha}\psi$ corresponds to the multiplication of each Fourier component $\hat{\psi}_{\boldsymbol{k}}$ by $k^{\alpha}$ with $k$ being the magnitude of wave number vector $\boldsymbol{k}$. We choose $\alpha=2$, yielding a dispersion relation $\omega_{k}=k^2$ that is the same as the nonlinear Schrodinger equation.
\subsection{Exact Formulations of Instantaneous Energy Flux $P(t)$}
To formulate the energy flux $P$ across arbitrary wave number $k_b$, we start from a consideration of energy conservation in a control volume $k<k_b$ in spectral space:
\begin{equation}
P(t) = -\sum_{ \boldsymbol{k}\in \{ \boldsymbol{k}|k<k_b \} }\omega_{k}\frac{\partial(\hat{\psi}_{\boldsymbol{k}}\hat{\psi}_{\boldsymbol{k}}^*)}{\partial t}(t) = - \sum_{ \boldsymbol{k}\in \{ \boldsymbol{k}|k<k_b \} }\omega_{k}\left( \frac{\partial\hat{\psi}_{\boldsymbol{k}}}{\partial t}\hat{\psi}_{\boldsymbol{k}}^* + \frac{\partial \hat{\psi}_{\boldsymbol{k}}^*}{\partial t} \hat{\psi}_{\boldsymbol{k}}\right)(t).
\label{eqn:ptot}
\end{equation}
Substituting the Fourier-domain representation of (\ref{eqn:MMT}) into (\ref{eqn:ptot}) yields
\begin{equation}
P(t) = -\sum_{ \boldsymbol{k}\in \{ \boldsymbol{k}|k<k_b \} }\omega_{k} \sum_{(\boldsymbol{k}_{1}, \boldsymbol{k}_{2}, \boldsymbol{k}_{3})\in S_{\boldsymbol{k}}} 2 k_1 k_2 k_3 k \Imag\left(\hat{\psi}_{\boldsymbol{1}} \hat{\psi}_{\boldsymbol{2}} \hat{\psi}_{\boldsymbol{3}}^* \hat{\psi}_{\boldsymbol{k}}^*\right)(t),
\label{eqn:pquartet}
\end{equation}
where $S_{\boldsymbol{k}}$ is the set of all $(\boldsymbol{k}_{1}, \boldsymbol{k}_{2}, \boldsymbol{k}_{3})$ with $\boldsymbol{k}_{1} + \boldsymbol{k}_{2} - \boldsymbol{k}_{3} - \boldsymbol{k}=0$. We note that (\ref{eqn:pquartet}) is exact even if forcing and dissipation are added to (\ref{eqn:MMT}) (which may affect (\ref{eqn:ptot}) but not (\ref{eqn:pquartet})), because only the nonlinear term in (\ref{eqn:MMT}) is responsible for the inter-scale energy flux. Taking the wave action $n({\boldsymbol{k}})\sim \hat{\psi}_{\boldsymbol{k}}^2$, (\ref{eqn:pquartet}) implies a dynamic scaling of $n\sim P^{1/2}$ via a heuristic argument. This is a result of directly using the dynamic equation (\ref{eqn:MMT}) to formulate the energy flux.
With (\ref{eqn:pquartet}) available, we can further perform a decomposition $P(t)=\sum_{\Omega} P_{\Omega}(t)$ by partitioning the set $S_{\boldsymbol{k}}$ according to the frequency mismatch of each quartet interaction. Specifically, defining $S_{\Omega , \boldsymbol{k}} \equiv\ \{ (\boldsymbol{k}_1, \boldsymbol{k}_2, \boldsymbol{k}_3) \in S_{\boldsymbol{k}} \ \big| \ |\omega_{1}+\omega_{2}-\omega_{3}-\omega_{k}| = \Omega \}$, we have $\bigcup_{\Omega} S_{\Omega ,\boldsymbol{k}} = S_{\boldsymbol{k}}$ with all sets $S_{\Omega ,\boldsymbol{k}}$ as disjoint. Therefore, $P_{\Omega}(t)$ can be naturally formulated as
\begin{equation}
P_\Omega(t) = -\sum_{ \boldsymbol{k}\in \{ \boldsymbol{k}|k<k_b \} } \omega_{k} \sum_{(\boldsymbol{k}_{1}, \boldsymbol{k}_{2}, \boldsymbol{k}_{3})\in S_{\Omega , \boldsymbol{k}}} 2 k_1 k_2 k_3 k \Imag\left(\hat{\psi}_{\boldsymbol{1}} \hat{\psi}_{\boldsymbol{2}} \hat{\psi}_{\boldsymbol{3}}^* \hat{\psi}_{\boldsymbol{k}}^*\right)(t).
\label{eqn:Pex}
\end{equation}
The computation using (\ref{eqn:Pex}) allows us to measure the contribution to $P$ from resonances with different $\Omega$, as well as to separate the quasi-resonant and exact-resonant contributions by $P_{\Omega>0}$ and $P_{\Omega=0}$, respectively.
\subsection{Formulations of Energy Flux under the WTT Closure}
The WTT closure for (\ref{eqn:MMT}) relates the fourth-order correlator to pair correlators by
\begin{eqnarray}
\Imag \overline{\left(\hat{\psi}_{\boldsymbol{1}}\hat{\psi}_{\boldsymbol{2}}\hat{\psi}_{\boldsymbol{3}}^{*}\hat{\psi}_{\boldsymbol{k}}^{*}\right)} _{\Omega}
= 2 k k_1 k_2 k_3 (n_{\boldsymbol{1}} n_{\boldsymbol{2}} n_{\boldsymbol{3}} + n_{\boldsymbol{1}} n_{\boldsymbol{2}} n_{\boldsymbol{k}} - n_{\boldsymbol{1}} n_{\boldsymbol{k}} n_{\boldsymbol{3}} - n_{\boldsymbol{k}} n_{\boldsymbol{2}} n_{\boldsymbol{3}})f(\Omega),
\label{eqn:closure}
\end{eqnarray}
where an over bar denotes the ensemble average (or time average in numerical analysis) and $n_{\boldsymbol{k}} = \overline{\hat{\psi}_{\boldsymbol{k}} \hat{\psi}^*_{\boldsymbol{k}}}$ is wave action. We also use a subscript $\Omega$ for $\Imag \overline{\left(\hat{\psi}_{\boldsymbol{1}}\hat{\psi}_{\boldsymbol{2}}\hat{\psi}_{\boldsymbol{3}}^{*}\hat{\psi}_{\boldsymbol{k}}^{*}\right)} _{\Omega}$ to denote the frequency mismatch of the corresponding four wave modes. The closure (\ref{eqn:closure}) has been derived in various (heuristic) ways in the physics literature \citep{zakharov_kolmogorov_2012,janssen_nonlinear_2003,nazarenko_wave_2011,pan_understanding_2017-1}, all assuming quasi-Gaussian statistics and quasi-stationary spectra (to obtain an analytical solution of the differential equation for fourth-order correlators). Depending on different methods of derivation, $f(\Omega)$ takes the form of $\sin(\Omega t)/\Omega$ \citep{janssen_nonlinear_2003} or $\epsilon/(\Omega^2+\epsilon^2)$ \citep{zakharov_kolmogorov_2012}. The WKE is then derived in the limit of the parameters $t\rightarrow\infty$ and $\epsilon\rightarrow0$ for the first and second cases respectively, with both leading to $f(\Omega)\rightarrow \pi\delta(\Omega)$.
We are interested in the performance of the closure (\ref{eqn:closure}) in computing the energy flux $P_\Omega$. For this purpose, we substitute (\ref{eqn:closure}) into the average of (\ref{eqn:Pex}) to obtain
\begin{eqnarray}
\overline{P}_{\Omega} = -\sum_{ \boldsymbol{k}\in \{ \boldsymbol{k}|k<k_b \} } \omega_{k} & \nonumber \\
\times \sum_{\boldsymbol{k}_{1}, \boldsymbol{k}_{2}, \boldsymbol{k}_{3}\in S_{\Omega , \boldsymbol{k}}} 4 k_1^2 k_2^2 k_3^2 k^2 & (n_{\boldsymbol{1}} n_{\boldsymbol{2}} n_{\boldsymbol{3}} + n_{\boldsymbol{1}} n_{\boldsymbol{2}} n_{\boldsymbol{k}} - n_{\boldsymbol{1}} n_{\boldsymbol{k}} n_{\boldsymbol{3}} - n_{\boldsymbol{k}} n_{\boldsymbol{2}} n_{\boldsymbol{3}} )f(\Omega).
\label{eqn:Pke2}
\end{eqnarray}
In \S 4, we will evaluate the functional form $f(\Omega)$ in both (\ref{eqn:closure}) and (\ref{eqn:Pke2}) using our numerical data to compute all other terms in the two equations (e.g., $\overline{P}_{\Omega}$ in (\ref{eqn:Pke2}) can be computed through time average of (\ref{eqn:Pex})). The difference between the two evaluations of $f(\Omega)$ is that the latter involves the sum of an enormous number of interactions while the former is for an individual quartet. Through comparison of the numerically resolved $f(\Omega)$ with their WTT counterparts, the validity of the WTT closure can be assessed.
\section{Setup of Numerical Experiments} \label{sec:numer}
We compute the solution to (\ref{eqn:MMT}) via a pseudospectral method on a periodic domain of size $2\upi \times 2\upi$ containing $512\times512$ modes. The linear term is integrated analytically to reduce the system stiffness, while the nonlinear term is integrated via an explicit fourth order Runge-Kutta scheme. Our purpose is to generate a long stationary state so that the distributions of $P$ (and other quantities discussed in \S 2) are sufficiently resolved. Therefore, we force the system at large scales (in conjunction with small-scale dissipation) instead of considering free-decay turbulence. Specifically, we add a forcing term
\begin{equation}
F = \left\{ \begin{array}{l}
F_{r}+iF_{i} \ \ \ 7\le k\le9 \\
0 \ \ \ otherwise,
\end{array} \right.
\label{eqn:forcing}
\end{equation}
to the right hand side of (\ref{eqn:MMT}), with $F_{r}$ and $F_{i}$ taken from a Gaussian distribution with zero mean and a variance $\sigma^2_F$ (which determines the forcing magnitude). Dissipation is accounted for by adding two terms
\begin{eqnarray}
D_1 =& \left\{ \begin{array}{l}
-i\nu_1 \hat{\psi}_{\boldsymbol{k}} \ \ \ k\ge100 \\
0 \ \ \ otherwise,
\end{array} \right. \nonumber \\
D_2 =& \left\{ \begin{array}{l}
-i\nu_2 \hat{\psi}_{\boldsymbol{k}} \ \ \ k\le7 \\
0 \ \ \ otherwise,
\end{array} \right.
\label{eqn:dissipation}
\end{eqnarray}
at small and large scales respectively, with the latter included to prevent energy accumulation at large scales due to the inverse cascade. The parameters in (\ref{eqn:dissipation}) are chosen to be $\nu_1 = 6\times10^{-12} (k-100)^{8}$ and $\nu_2 = 30 k^{-4}$ throughout all the simulations. Forcing and dissipation of this type have been demonstrated to produce results compatible with the WTT predictions in the one-dimensional MMT model \citep{cai_spectral_1999}.
In addition, to accelerate convergence to the stationary state, we start the simulations from initial conditions described by Gaussian spectra $\hat{\psi}_{\boldsymbol{k}}=a_0e^{-0.1|k-10|+\mathrm{i}\phi_{\boldsymbol{k}}}$, with $\phi_{\boldsymbol{k}}$ the uniformly-distributed, decorrelated random phases, and $a_0$ a real constant chosen to provide an energy close to that of the expected stationary state. To obtain the scaling $\theta$ over a range of $\overline{P}$, we run a collection of 19 simulations, differing only in forcing strength $\sigma^2_F$ and initial spectral level $a_0$. These simulations cover a range of $\overline{P}$ spanning several orders of magnitude, with data collected in the stationary state for each case.
\section{Results} \label{sec:res}
Before presenting the results on energy flux, we first check the spectra at stationary states in simulations with different forcing magnitudes. Several typical spectra at different levels are shown in figure \ref{fig:niso}a, where we observe power-law ranges close to one decade for all spectra. Figure \ref{fig:niso}b shows the power-law exponent $\gamma$ evaluated in all 19 simulations, as a function of the spectral level computed by an integral measure of the (conservatively taken) power-law range
\begin{equation}
N = \sum_{ \boldsymbol{k}\in \{ \boldsymbol{k}|13<k<60 \} } n_{\boldsymbol{k}}.
\label{eqn:N}
\end{equation}
We see in figure \ref{fig:niso}b that $\gamma$ increases (i.e., the spectrum becomes shallower) with the decrease of $N$, reaching the WTT prediction $\gamma_0=-4.67$ for low spectral levels. This behavior of $\gamma$ is consistent with our previous study of the free-decay MMT turbulence (except that use of different dissipation schemes may have some slight effect). As analyzed by the authors \citet{hrabski_effect_2020}, the fact that $\gamma\rightarrow \gamma_0$ for small $N$ is a result of the dispersion relation $\omega_k=k^2$ which leads to a continuous resonant system at low nonlinearity \citep{faou_weakly_2016}. The deviation of $\gamma$ from $\gamma_0$ at high nonlinearity may result from coherent structures, as suggested by \citet{zakharov_wave_2001} and \cite{chibbaro_weak_2017} in the one-dimensional context, or some features of the 2D MMT model that are yet to be fully understood. For waves in different physical contexts, e.g., surface gravity waves \citep{zhang_numerical_2021} and capillary waves \citep{pan_direct_2014}, the behaviors of $\gamma$ are remarkably different.
\begin{figure}
\centerline{\includegraphics{niso}
\caption{(a) A representative collection of fully-developed, angle-averaged wave action spectra $n(k)$. (b) Spectral slope $\gamma$ as a function of $N$, with WTT value $\gamma_0=-4.67$ indicated (\dashed).}
\label{fig:niso}
\end{figure}
We next present our full study of energy flux, with results organized into three sections. \S 4.1 discusses the distributions of $P$ and its associated decomposition $P_\Omega$. \S 4.2 focuses on the scaling of spectral level with $P$, with the results explained by the contributions of quasi/exact resonances to $P$. The study related to the closure model is then presented in \S 4.3.
\subsection{Flux Distributions and Decomposition}
A typical distribution of energy flux $P(t)$, computed with $k_b=30$ from $2^{16}$ data points over a time window of $T_w=256T_0$ (with $T_0$ the fundamental period), is shown in figure \ref{fig:pk}a. We find that $P$ closely follows a Gaussian distribution, with a standard deviation $\sigma(P)=621.8$ several times larger than the mean value $\overline{P}=77.02$. The very large standard deviation is consistent with previous studies in wave turbulence \citep{falcon_fluctuations_2008} and hydrodynamic turbulence \citep{bandi_energy_2006}. However, the nearly perfect Gaussian form of the distribution has not been observed in these previous works, probably because of their approximation in evaluating $P$ (either from energy input rate or a filter-based method) and the different turbulent systems considered in their study.
\begin{figure}
\centerline{\includegraphics{pk}
\caption{(a) The histogram of stationary time series $P(t)$ evaluated over $256 T_0$, fitted with a Gaussian distribution of the same mean and standard deviation (\full) for reference. Figure inset: tail of the distribution in logarithmic scale. (b) the mean $\overline{P}$ and (c) standard deviation $\sigma(P)$ evaluated for different $k_b$. The dissipation-based estimate of $P_d$ is indicated in both (a) and (b) by (\dashed).}
\label{fig:pk}
\end{figure}
In addition, our method allows us to study the fully-resolved distribution of $P$ at any scale (i.e., with arbitrary $k_b$). In figures \ref{fig:pk}b and \ref{fig:pk}c, we plot the values of $\overline{P}$ and $\sigma(P)$ for $k_b$ varying in the inertial range from 20 to 90. The mean flux $\overline{P}$ remains almost constant for all $k_b$, which is consistent with the WTT constant flux argument in the inertial range (this is only possible by avoiding broad-scale dissipation in simulations). The standard deviation $\sigma(P)$ increases with $k_b$, because more quartet interactions are included (as $k$ becomes denser) resulting in more fluctuations in $P(t)$. We also include in figures \ref{fig:pk}a and \ref{fig:pk}b the energy flux $\overline{P}$ computed from the high-wave-number dissipation rate
\begin{equation}
\overline{P}_d = \sum_{ \boldsymbol{k}\in \{ \boldsymbol{k}|k>100 \} }\nu_1 \omega_{k} \overline{\hat{\psi}_{\boldsymbol{k}} \hat{\psi}_{\boldsymbol{k}}^*},
\label{eqn:pd}
\end{equation}
which agrees well with the majority values of $\overline{P}$, especially for larger $k_b$ (to a relative difference within $O(1\%)$). Because of this, we will use $\overline{P}_d$ to represent the values of $\overline{P}$ for all 19 simulations in the subsequent analysis, since $P_d$ yields a faster calculation due to an easier formulation and much smaller fluctuations (requiring less data points for averaging).
\begin{figure}
\centerline{\includegraphics{pscaling}
\caption{Dependence of $\sigma(P)$ (\fullcirc) and $\sigma(P_{\Omega=0})$ (\fullsquare) on $\overline{P}$, with the best fits indicated (\dashed).}
\label{fig:pscale}
\end{figure}
We next examine the relation between $\sigma(P)$ and nonlinearity level measured by $\overline{P}$, with $\sigma(P)$ as a function of $\overline{P}$ plotted in figure \ref{fig:pscale}. The result shows a power-law relation over two decades given by $\sigma(P) \sim \overline{P}^{0.8 \pm 0.05}$. Furthermore, we include in figure \ref{fig:pscale} the standard deviation of the exact-resonant contributions to energy flux, $\sigma(P_{\Omega=0})$, with $P_{\Omega=0}$ calculated by the decomposition method presented in \S 2. We observe a similar power-law relation between $\sigma(P_{\Omega=0})$ and $\sigma(P)$, but with the value of $\sigma(P_{\Omega=0})$ $O(10)$ times smaller than $\sigma(P)$ consistently for each nonlinearity level. This indicates that the large fluctuations in $P(t)$ are mainly generated due to quasi-resonant interactions.
A more detailed study about the contributions of exact and quasi resonances to $\overline{P}$ and $\sigma(P)$ can be conducted by looking into the components of $P_\Omega$ for varying values of $\Omega$. In figures \ref{fig:pdw1}a and \ref{fig:pdw1}b we plot $\overline{P}_\Omega$ and $\sigma(P_\Omega)$ for $\Omega\in [0,30]$ at four different levels of nonlinearity. We note that $\Omega$ can only take even integer values for the dispersion relation $\omega_k=k^2$ on a periodic domain of $2\upi\times2\upi$. The general trends in figure \ref{fig:pdw1}a and \ref{fig:pdw1}b show that $\overline{P}_\Omega$ decreases, but $\sigma(P_\Omega)$ increases with the increase of $\Omega$. This corresponds to a physical picture that as the interactions become more ``quasi'' (i.e., frequency mismatch $\Omega$ becomes larger), they contribute less to the mean flux but may contribute more to the fluctuations of the flux. We also emphasize here that while we always have $\sum_\Omega \overline{P}_\Omega = \overline{P}$, the quantity $\sum_\Omega \sigma^2(P_\Omega)$ is in general not equal to $\sigma^2(P)$ because $P_\Omega(t)$ with different $\Omega$ are not independent. Nevertheless, figure \ref{fig:pdw1}b in conjunction with figure \ref{fig:pscale} are sufficient to support the dominance of quasi-resonances in generating the large fluctuations in $P(t)$.
\begin{figure}
\centerline{\includegraphics{pdw}
\caption{(a) $\overline{P}_\Omega$ and (b) $\sigma(P_\Omega)$ as functions of $\Omega$, for four levels of nonlinearity with $\overline{P}=81.9$ (\fullcirc), $61.0$ (\fullsquare), $42.8$ (\fulltri), and $20.0$ (\fulltriangledown); (c) nonlinear broadening $\Gamma$ as a function of $\overline{P}$; (d) number of quartet interactions $\mathcal{N}_{\Omega}$ for different $\Omega$; (e) and (f) are similar to (a) and (b) but plotted for normalized flux $Q_n$. The computations to generate these results are conducted for $k_b=23$ to reduce the computational cost associated with the number of involved quartets.}
\label{fig:pdw1}
\end{figure}
We conclude this section by summarizing two additional important results regarding $P_\Omega$. First, the decomposition in terms of $\Omega$ enables a direct measure of nonlinear broadening by quantitatively considering the contribution of quasi-resonances to the total energy flux. For example, the nonlinear broadening $\Gamma$ based on the definition $\Gamma \equiv min \{\Omega|\overline{P}_{\Omega}< 0.1\times \overline{P}_{\Omega=0} \}$ is plotted in figure \ref{fig:pdw1}c, showing that $\Gamma$ increases with $\overline{P}$ (or the nonlinearity level). Second, the fluctuations seen in figure \ref{fig:pdw1}a and \ref{fig:pdw1}b can be removed by considering the normalized flux $Q_\Omega(t)=P_\Omega(t)/\mathcal{N}_\Omega$. With $\mathcal{N}_\Omega$ (shown in figure \ref{fig:pdw1}d) counting the number of elements in $\sum_{ \boldsymbol{k}\in \{ \boldsymbol{k}|k<k_b \} } S_{\Omega , \boldsymbol{k}}$ (see (\ref{eqn:Pex})), $Q_\Omega(t)$ calculates the quartet-averaged flux (over quartets with frequency mismatch $\Omega$). As plotted in figure \ref{fig:pdw1}e and \ref{fig:pdw1}f, both $\overline{Q}_{\Omega}$ and $\sigma(Q_\Omega)$ behave smoothly for the range of $\Omega$. We speculate that these smooth behaviors regarding $Q_\Omega$ may hold for different domain geometries (e.g., tori with different aspect ratios) with different functions $\mathcal{N}_\Omega$.
\subsection{Scaling of Spectral Level with Flux}
To understand the scaling of spectral level with energy flux, we plot in figure \ref{fig:ep} the spectral level $N$ (see (\ref{eqn:N})) as a function of both $\overline{P}$ and $\overline{P}_{\Omega=0}$ representing total and exact-resonant flux respectively. Two salient scalings are observed over the 3 decades of energy flux. At high nonlinearity level with $\overline{P} \in [30,100]$, we find a scaling approaching $N\sim \overline{P}^{1/3}$ (i.e., $\theta=1/3$ with $\theta$ the scaling exponent defined in \S 2) consistent with the kinetic scaling of WTT. At low nonlinearity level with $\overline{P} \in [0.3,3]$, the scaling behaves as $N\sim \overline{P}^{1/2}$ (i.e., $\theta=1/2$) consistent with the dynamic scaling from (\ref{eqn:pquartet}). We next discuss the mechanisms underlying these two scalings.
\begin{figure}
\centerline{\includegraphics{ep}
\caption{The scaling of inertial-range wave action $N$ with $\overline{P}$ (\fullcircline) and $\overline{P}_{\Omega =0}$ (\fulltriline). The dynamic scaling $\theta=1/2$ and kinetic scaling $\theta=1/3$ are indicated (\dashed).}
\label{fig:ep}
\end{figure}
For high nonlinearity, we see in figure \ref{fig:ep} that we have $\overline{P} \gg \overline{P}_{\Omega=0}$ consistent with observations in figure \ref{fig:pdw1}a. This suggests that the kinetic scaling, although developed from the WKE formulated on the exact resonant manifold, relies on the dominant contribution from quasi-resonances to be realized in a finite domain. This physical picture is consistent with the numerical study by \citet{annenkov_role_2006}, as well as the recent mathematical justification of the WKE \citep{deng_full_2021,deng_derivation_2021,buckmaster_onset_2021}. In particular, the rigorous mathematical derivation proves that the WKE describes the effective dynamics of both exact \emph{and} quasi resonances in the kinetic limit. The present results are supplemental to these previous works, which only deal with initial spectral evolution up to the kinetic time scale. Our study additionally suggests that this physical understanding holds for the long-time stationary state.
For low nonlinearity, $\overline{P}\approx \overline{P}_{\Omega=0}$ (as shown in figure \ref{fig:ep}) due to the elimination of quasi-resonant contributions to $\overline{P}$. This behavior has been rigorously proven by \citet{faou_weakly_2016} and numerically observed by \citet{hrabski_effect_2020}. Under this situation, the number of interactions contributing to $\overline{P}$ is substantially reduced, and the dynamic scaling from (\ref{eqn:pquartet}) becomes dominant. We further remark that the presence of an energy cascade, instead of frozen turbulence at low nonlinearity, is due to the dispersion relation $\omega_k=k^2$ leading to a continuous resonant system \citep{faou_weakly_2016} at $N\rightarrow 0$.
Finally, we note that the transition to the dynamic scaling from high to low nonlinearity seems to involve a small and non-smooth jump, as can be seen in figure \ref{fig:ep} (which is also observable in earlier figure \ref{fig:pscale}). This may suggest that the removal of quasi-resonant contributions is not smooth with the decrease of nonlinearity level, i.e., there exists a threshold below which the quasi-resonances are turned off non-incrementally. However, we will avoid overreaching with this point in our current work and consider it a hypothesis that requires further study.
\subsection{Investigation on the Closure Model}
In this section, we use our numerical data to study the WTT closure model, in particular the magnitude and functional form of $f(\Omega)$. We focus this study on the nonlinearity level associated with the kinetic scaling of $\overline{P}$, for which the WTT closure is developed. Theoretically one expects $f(\Omega)$ to take the form of either $\sin(\Omega t)/\Omega$ \citep{janssen_nonlinear_2003} or $\epsilon/(\Omega^2+\epsilon^2)$ \citep{zakharov_kolmogorov_2012}, with $\int_\Omega f(\Omega)d\Omega \sim O(1)$ since both forms are generalized delta functions. The numerical evaluation of $f(\Omega)$ will be performed at: (1) an individual quartet level using (\ref{eqn:closure}); (2) a family of quartets level using (\ref{eqn:closure}) in an average manner that will be introduced shortly; and (3) an inter-scale energy flux level using (\ref{eqn:Pke2}). The general procedure is to compute $f(\Omega)$ from (\ref{eqn:closure}) and (\ref{eqn:Pke2}) with all other terms determined from the numerical data. To distinguish these computations, we denote the numerically obtained $f(\Omega)$ from the three levels respectively as $f_Q(\Omega)$, $f_F(\Omega)$ and $f_P(\Omega)$.
Figure \ref{fig:q1} shows $f_Q(\Omega)$ evaluated for $O(50)$ quartets with $\Omega\in[0,30]$, and with average quantities in (\ref{eqn:closure}) evaluated over a time window $T_w=256 T_0$. It is clear that no obvious functional pattern can be found for $f_Q(\Omega)$ (i.e., with different values of $f_Q(\Omega)$ obtained for the same $\Omega$). Also, we see some values of $f_Q(\Omega)\sim O(10)$, likely in disagreement with the WTT result of $\int_\Omega f(\Omega) d\Omega \sim O(1)$. In general, figure \ref{fig:q1} indicates that the WTT closure for fourth-order correlators cannot be used to describe the behavior of a single quartet, regardless of its associated frequency mismatch $\Omega$.
To consider the average behavior of a family of quartets, we use a similar technique to that of \citet{annenkov_role_2006} to create a cluster of modes around each of the four modes of an exact resonant quartet. Specifically, for an exact quartet ($\boldsymbol{k}_0^e$, $\boldsymbol{k}_1^e$, $\boldsymbol{k}_2^e$, $\boldsymbol{k}_3^e$), we construct a family of quartets ($\boldsymbol{l}_0$, $\boldsymbol{l}_1$, $\boldsymbol{l}_2$, $\boldsymbol{l}_3$) at its vicinity by choosing all $\boldsymbol{l}_i$ with $|\boldsymbol{l}_i- \boldsymbol{k}_i^e|\le4$ for $i=0,1,2,3$. We then evaluate $f_F(\Omega)$ from (\ref{eqn:closure}) by summing over those quartets ($\boldsymbol{l}_0$, $\boldsymbol{l}_1$, $\boldsymbol{l}_2$, $\boldsymbol{l}_3$) with frequency mismatch $\Omega$ on both sides of the equation. Under this evaluation $f_F(\Omega)$ reflects the closure behavior averaged over $O(10^3)$ quartets. Figure \ref{fig:q2} shows $f_F(\Omega)$ computed from three representative families of quartets, defined via exact quartets ($\boldsymbol{k}_0^e$, $\boldsymbol{k}_1^e$, $\boldsymbol{k}_2^e$, $\boldsymbol{k}_3^e$). We see that $f_F(\Omega)$ is somewhat inversely proportional to $\Omega$ superposed with many fluctuations. While the general trend of $f_F(\Omega)$ seems consistent for different families, the details (e.g, the value of $f_F(0)$ as well as the fluctuation patterns) vary considerably across different families.
We finally examine the closure behavior considering the average over an enormous number of quartets, chosen as all quartets contributing to the energy flux across $k_b=23$. Under such a consideration, both sides of (\ref{eqn:Pke2}) are subject to summation over $O(10^9)$ elements in $\sum_{ \boldsymbol{k}\in \{ \boldsymbol{k}|k<k_b \} } S_{\Omega , \boldsymbol{k}}$ for each $\Omega$ (see figure \ref{fig:pdw1}d). The numerically resolved $f_P(\Omega)$ is plotted in figure \ref{fig:f} at several different nonlinearity levels. We remark that these results are convergent in the sense that they are not sensitive to the further increase of the time window $T_w$, i.e., a longer time average. It is clear that under this level of average, $\int_\Omega f(\Omega)d\Omega\sim O(1)$ consistent with WTT (we do not expect $\int_\Omega f(\Omega)d\Omega$ to be exactly one unless we can numerically reach the kinetic limit). To quantify the profile of $f_P(\Omega)$, we use a least-square method to fit the data to a general functional form of $f_P(\Omega)=C/(\rho + \Omega ^{\beta})$, where $\rho$ is needed as a desingularisation factor for $f_P(0)$. These fittings, as shown in figure \ref{fig:f}, agree with all data points remarkably well, with $C=[1.06,0.982,0.875]$, $\rho=[6.176,3.44,2.26]$, and $\beta=[1.35,1.47,1.52]$ from high to low nonlinearity level. Instead of a $sinc$ function (which involves negative values that are not observed), the functional form of $f_P(\Omega)$ is somewhat closer to the WTT form of $\epsilon/(\Omega^2+\epsilon^2)$, but with different exponents $\beta$. This is probably why \citet{pan_understanding_2017} find that the kinetic equation employing this WTT form of generalized delta function produces physically reasonable results in terms of the spectral slope and energy flux for capillary waves.
\begin{figure}
\centerline{\includegraphics{quartet1}
\caption{The function $f_Q(\Omega)$ evaluated for $O(50)$ selected quartets (with 3 quartets for each $\Omega$) with $T_w = 256 T_0$, for the highest nonlinearity level with $\overline{P}=81.9$.}
\label{fig:q1}
\end{figure}
\begin{figure}
\centerline{\includegraphics{quartet2}
\caption{The function $f_F(\Omega)$ for three representative families of quartets defined by $\boldsymbol{k}_0^e=(-2,8)$, $\boldsymbol{k}_1^e=(-10,0)$, $\boldsymbol{k}_2^e=(14+4j,-8-4j)$, and $\boldsymbol{k}_3^e=(6+4j,-16-4j)$ for $j=0$ (\fullcirc), $j=1$ (\fullsquare), and $j=2$ (\fulltri). The evaluation is for the highest nonlinearity case of $\overline{P}=81.9$ with $T_w = 256 T_0$.}
\label{fig:q2}
\end{figure}
\begin{figure}
\centerline{\includegraphics{pdw5}
\caption{The function $f_P(\Omega)$ evaluated by (\ref{eqn:Pke2}) with $k_b=23$ and $T_w = 256 T_0$, for different nonlinearity levels with $\overline{P}=81.9$ (\fullcirc), $42.8$ (\fullsquare), and $28.0$ (\fulltri). Fits to the data of the form $f(\Omega)=C/(\rho + \Omega^{\beta})$ are indicated (\dashed).}
\label{fig:f}
\end{figure}
\section{Conclusion and Discussions}
In this paper, we numerically study the properties of inter-scale energy flux $P$ for wave turbulence in the context of the 2D MMT equation. Unlike previous evaluations of $P$ based on energy input or dissipation rate, our formulation of $P$ computes the exact instantaneous energy flux across arbitrary scale $k_b$ directly from the nonlinear terms in the MMT equation, and allows a quartet-level decomposition of $P$ into $P_\Omega$ according to the frequency mismatch $\Omega$ of the quartets. Our results show that the energy flux $P$ across any scale in the inertial range closely follows a Gaussian distribution, with their mean value $\overline{P}$ almost a constant for any $k_b$, and standard deviation $\sigma(P)$ increasing with $k_b$. In addition, values of $\sigma(P)$ are generally several orders of magnitude larger than $\overline{P}$, mainly due to the contributions to $\sigma(P)$ from quasi-resonances, i.e., $P_{\Omega>0}$. The decomposition of $P$ into $P_\Omega$ also allows an alternative but more direct measure of nonlinear broadening by quantitatively considering the contribution of quasi-resonances to the mean energy flux. We further study the scaling of spectral level $N$ with the energy flux and find that $N\sim \overline{P}^{1/3}$ (consistent with the kinetic scaling) at high nonlinearity and $N\sim \overline{P}^{1/2}$ (dynamic scaling) at low nonlinearity. The former and latter are due to the dominance of the quasi-resonant and exact-resonant contributions to $\overline{P}$, respectively. Finally, our numerical study on the wave-turbulence closure model shows that the fourth-order correlator is in disagreement with the description by the theoretical closure on a single quartet level. When considering the average over all quartets contributing to the inter-scale energy flux (over $O(10^9)$ quartets for each $\Omega$), more consistent behavior to the theoretical closure is observed, but with the broadening function exhibiting $1/\Omega^\beta$ (with $\beta$ between $1.3$ and $1.6$) different from the forms derived in WTT.
While this work sheds new light on the physics of wave turbulence, more unanswered questions about wave turbulence, especially regarding the closure model, are raised. The closure model is a subject that has received insufficient attention from a numerical perspective, especially in terms of analysis utilizing data generated directly from the primitive dynamic equation. The few exceptions to this \cite[e.g.][]{majda_one-dimensional_1997,annenkov_role_2006,sheffield_ensemble_2017} have not considered the detailed functional form and magnitude of $f(\Omega)$, as studied in this work. Therefore, the authors consider the primary importance of this work as to provide a methodology such that many open questions in wave turbulence can be directly studied using vast numerical data. Within the presented results, for example, the function $f_Q(\Omega)$ in figure \ref{fig:q1} is still sensitive to the time window $T_w$ used for averaging quantities in (\ref{eqn:closure}) for an individual quartet. It is not clear whether a convergent behavior (close to the quartet-averaged result $f_P(\Omega)$) can be found if a time average over an extremely long time window is performed. In addition, will the smooth behavior associated with function $Q_\Omega$ in figure \ref{fig:pdw1}e and \ref{fig:pdw1}f be preserved in different domain geometries (e.g., an irrational torus as in \citet{hrabski_energy_2021}) where the normalization factor $\mathcal{N}_\Omega$ substantially varies? Last but not least, are we able to explore the kinetic limit in numerical simulations by varying both domain size and nonlinearity level? If this is possible, will we find more consistent behavior for $f_F(\Omega)$ for families centered on different quartets, and will all these functions eventually converge to a delta function?
These questions can be explored in greater detail with increasing computational resources, and it is certainly not unreasonable to think about studying wave turbulence with ``exascale computing'', an area under development for hydrodynamic turbulence \cite[e.g.][]{yeung_advancing_2020}. For wave turbulence, these resources may be better utilized in conjunction with an understanding of WTT, rather than simply boosting the resolution of simulations. It might also be beneficial to revisit the one-dimensional MMT equation, where some of these questions can be investigated with a reduced computational cost but with the new techniques developed here. In doing this we may also hope to settle the dispute between the WTT and MMT closure models that have been haunting the field for more than 20 years.
\section{Acknowledgements}
This material is based upon work supported by the National Science Foundation Graduate Research Fellowship under Grant No. DGE 1841052. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation. This work used the Extreme Science and Engineering Discovery Environment (XSEDE), which is supported by National Science Foundation grant number ACI-1548562. Computation was performed on XSEDE Bridges-2 at the Pittsburgh Supercomputing Center through allocation PHY200041. This research was supported in part through computational resources and services provided by Advanced Research Computing (ARC), a division of Information and Technology Services (ITS) at the University of Michigan, Ann Arbor.
\bibliographystyle{jfm}
|
\section{Introduction}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{teaser.pdf}
\caption{\textbf{Results of simultaneous novel view synthesis and appearance editing:} The top row shows our own scene captured with a handheld mobile phone while the bottom row is a scene from the DTU dataset. Our method can photo-realistically render novel view points while also modifying the scene's appearance (middle column, right column).}
\label{fig:teaser}
\end{figure}
The graphics pipeline renders high quality images of a scene given
well-defined scene geometry, lighting, and material properties. Creating
these for a new scene is a tedious process and realism is a distant
goal. The ability to render a scene from a few photographs remains a
long-sought goal of computer vision as a result.
Recently, the field of Neural Rendering \cite{Tewari2020StateOT} has achieved
success in representing and photo-realistically rendering a scene from
unstructured images. Neural rendering approaches replace components of
the graphics pipeline with learnable ones to achieve explicit or
implicit control over scene properties. Neural rendering makes it possible to generate
novel views and even relight the scene from either a single image or
multi-view images \cite{dnr, dnl, rnr, nerf, singlePortrait}.
The work of Mildenhall et al. \cite{nerf} proposed Neural Radiance Fields (NeRF) for volumetric scene representations. NeRF encodes a scene as a continuous
5D volume represented in terms of the learnt weights of a neural
network. The network was trained on unstructured photographs of a
scene and can be queried to generate novel views with volume rendering
techniques \cite{volrend}. Many follow-up
efforts extend NeRF to support unbounded scenes \cite{nerf++},
unstructured internet images with varying illumination \cite{nerf-w} and
deformable geometry \cite{dnerf, nerfies}. Recently, NeRV
\cite{nerv2020} and NeRD \cite{NeRD} proposed methods to explicitly
recover shape, illumination and materials in a NeRF-like model. While the representation of these methods is neural, image generation is still achieved through classical rendering techniques.
In contrast to the above, Deferred Neural Rendering (DNR) \cite{dnr} aims at
both representing and rendering a scene entirely using learnt
components. They propose \textit{neural textures}, a neural alternative to
classical texture maps and a corresponding neural renderer which
renders novel views from the neural texture. DNR recovers neural textures
of a scene by training on a set of unstructured photographs. More recent efforts allow view synthesis along with relighting of the scene, with either known or
unknown illumination \cite{dnl, rnr}. These methods use a network for rendering and therefore are more efficient. However, the material properties and the geometry of the scene remain fixed.
In this paper, we present a method that achieves simultaneous view
synthesis and appearance editing of a scene from unstructured captures
under known distant illumination. We tackle appearance editing of a scene instead of relighting, by extending the DNR framework \cite{dnr}.
We use a representation that disentangles the Bi-directional Reflectance Distribution Function (BRDF) and the local irradiance function (LIF) of the captured scene. The BRDF is estimated
using an optimization involving differentiable rendering and the
LIF is learnt using neural network from the input images. This
disentanglement allows the BRDF to be edited or modified independently. For a new viewpoint, the modified appearance is combined with the LIF produced by the network to render the final image.
The contributions of this paper are the following.
\begin{itemize}
\item A disentangled, appearance-independent, learnt representation of the scene.
\item A method to recover this representation from unstructured images
of a scene.
\item A novel pipeline to generate new views from this representation and edit material properties including changing its diffuse color texture, the bump map or the material model on synthetic and real scenes.
\end{itemize}
Figure \ref{fig:teaser} shows an example of simultaneous appearance
editing and view synthesis on our own scene and a scene from the DTU
\cite{dtu} dataset.
\section{Related Work}
We discuss work related to image-based rendering, neural rendering and pre-computed radiance transfer that has received significant interest in the recent past.
\paragraph{Image based rendering (IBR).}
Image based rendering utilizes 2D images to generate novel representation of the 3D world. Recently deep learning has proven to be a powerful tool for these class of methods for performing novel view synthesis. Strategies to perform view synthesis range from view interpolation\cite{Kalantari_2016} and extrapolation\cite{Srinivasan2019PushingTB} to multi plane images\cite{Flynn2019DeepViewVS,mildenhall2019llff,Zhou2018StereoM} to flow based warping methods\cite{Jin2018LearningTD,Liu2018GeometryAwareDN,Sun2018MultiviewTN} and to pure image based disentanglement\cite{Ji2017DeepVM,Olszewski2019TransformableBN}. Free View Synthesis \cite{Riegler2020FVS} performs view synthesis by mapping the encoded features from the source images into the target view and blending them via a neural network. Stable View Synthesis \cite{Riegler2020StableVS} does the same by mapping encoded image feature vectors onto the reconstructed geometry and aggregating these features for a novel view. Like IBR methods, we also work with imperfect reconstructed geometry.
\paragraph{Neural Radiance Fields.}
There has been a recent growing interest in representation of the scene
geometry as the parameters of a Neural Network. This is referred to as
Neural Implicit Representation. NeRF \cite{nerf} represents a static
scene as 5D continuous function by encoding it within a neural
network. To perform view synthesis, coordinates are queried along
camera rays and a classic volume rendering technique is used to
project color and density onto an image. NeRF-W \cite{nerf-w} extends
the method by accounting for unconstrained illumination and uses related
images from the internet. NeRF++ \cite{nerf++}
extends the original method for unbounded scenes. Nerfies \cite{nerfies} and
D-Nerf \cite{dnerf} augment the method by using a second neural network
which applies a deformation to every second of the video. NeRV
\cite{nerv2020} adds the relighting and material editing capability by adding a second ``visibility'' neural network to support
environment lighting and one-bounce indirect illumination. NeRD
\cite{NeRD} is an explicit decomposition model for shape, reflectance,
and illumination within a NeRF-like coordinate-based neural
representation framework. All these methods, though powerful, are still
different types of neural implicit representations of the scene, which
need to be rendered using classical techniques. Our method, on the other
hand, represents and renders the scene using learnable components,
thereby lifting the limitation of classic rendering.
\paragraph{Deferred Neural Rendering.}
Neural Rendering (NR) proposes to learn the end-to-end rendering process
from images captured under uncontrolled illumination. Deferred Neural
Rendering (DNR) \cite{dnr} demonstrates this by performing view
synthesis and face re-enactment with only images and an approximate 3D
reconstruction. DNR achieves novel view synthesis keeping the geometry,
appearance, and the lighting of the captured scene constant. Deferred
Neural Lighting (DNL) \cite{dnl} extended this to allow relighting by
projecting learnt neural textures onto a rough proxy geometry using a
scene-dependent neural rendering network for relighting. Relightable Neural Rendering (RNR) \cite {rnr} extends the DNR framework to allow lighting changes with arbitrary environment lighting
using a learnt light transport function. The network regresses on the
lighting, object intrinsics, and light transport function rather than
directly translating deep features to appearance. Our method extends the
DNR framework to allow material editing in novel view synthesis while the
geometry and lighting are fixed using a disentangled representation.
\paragraph{Pre-computed Radiance Transfer (PRT).}
Pre-computed radiance transfer frameworks \cite{prt} aim at
representing, storing, and rendering the light transport of a scene
under varying illumination. The work of Ben-Artzi et al. \cite{lif_separation}
re-formulates PRT to allow real-time appearance editing. Generally, PRT
methods decompose the rendering equation into parts and project each to
a compact basis representation. These approaches are however dependent on
well-defined geometry, lighting and materials. We instead aim to edit
the appearance of a scene from unstructured photographs. To this end, we
augment the approach of \cite{lif_separation} with learnable components
and use a similar decomposition of the rendering equation.
\section{Method}
Given the input data specified as $\{I_k, p_k, m_k, E, G\}_{k=1}^{N}$, consisting of a corpus of $N$ images $I_k$ of a scene with their corresponding pose $p_k$ and the object masks $m_k$ our aim is to render novel views of the scene and edit its appearance, assuming a fixed distant environment illumination $E$ and proxy geometry $G$. We achieve this by disentangling the BRDF and local irradiance. The latter is learnt using a neural network and the former is estimated from the training images and can be modified independently.
The incoming radiance $L$ at pixel $x$ as viewed from the direction $\omega_x$ is modeled by the rendering equation \cite{jim_kajiya} as:
\begin{equation}
\label{eq:rendering_eq}
\begin{split}
L(x) = \int_{\Omega} f_{r}(p_x, \omega_x, \omega_i) T(p_x, \omega_i) d\omega_i,\\
\end{split}
\end{equation}
as an integral over all directions on the hemisphere $\Omega$ of a product of two terms: the BRDF $f_r$, and the \textit{Local Irradiance Function} (LIF) \cite{lif_separation}. The LIF is defined as $T(p_x, \omega_i) = L_e(\omega_i)V(p_x, \omega_i)(\omega_i \cdot n)$ where $L_e$ is the illumination, $V$ is the visibility function, $\omega \cdot n$ is the cosine foreshortening factor, $p_x$ is the 3D point corresponding to the 2D pixel $x$, $\omega_i$ is the incoming light direction, $\omega_x$ is the direction from $p_x$ toward the camera.
A different method of separating the rendering equation into the light transport function and the illumination has also been used for relighting \cite{rnr}. The integral in Eq. \ref{eq:rendering_eq} can equivalently be evaluated as a dot product of the projections of $f_r$ and $T$ onto the Spherical Harmonic (SH) basis \cite{jarosz08thesis, prt_rr, rnr}:
\begin{equation}
\label{eq:dot_prod}
L(x) = \bar{f_r} \circ \bar{T},
\end{equation}
where $\bar{f_r}$ and $\bar{T}$ are the SH coefficient vectors of the BRDF and LIF respectively. Given a pixel $x$, the 4D BRDF $f_r$ is now a 2D function of $\omega_i$ over the unit sphere and can be readily projected to SH at run-time. Readers are referred to \cite{jarosz08thesis, prt_rr, prt} for a detailed explanation on integration of the product of two functions being evaluated as the dot product of their corresponding SH coefficients.
In our approach, we recover the BRDF parameters for every point of the scene with a differentiable renderer as a pre-processing step (Sect. \ref{sect:diff_rend}). We then learn the SH coefficients of the LIF using a network similar to DNR \cite{dnr}. The BRDF can then be independently modified to change the appearance of the scene. While generating new views, the SH projection $\bar{f_r}$ of the modified BRDF is computed and the SH projection $\bar{T}$ of the LIF is generated using the learnt network. A dot product of the two as given in Eq. \ref{eq:dot_prod} generates the final image (Sect. \ref{sect:learn_LIF}). More details on the exact SH computations are given in the supplementary document.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{material_opt_overview.pdf}
\caption{\textbf{Estimating the BRDF:} We initialize the scene with the captured environment map and perform differentiable rendering optimization to recover the BRDF parameters (diffuse colour).}
\label{fig:material_opt_overview}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{network_overview.pdf}
\caption{\textbf{Network architecture and training:} We use an architecture similar to DNR \cite{dnr} for our network, which outputs 25 SH coefficients per color channel corresponding to the LIF. We separately compute the BRDF's SH coefficients. The loss is backpropagated through the network only (dotted portion).}
\label{fig:network_overview}
\end{figure}
\subsection{Estimating the BRDF}
\label{sect:diff_rend}
Fig. \ref{sect:diff_rend} shows the overview of our BRDF estimation pipeline. To estimate the BRDF of the scene from a set of photographs, we use a differentiable rendering optimization on the proxy geometry $G$ initialized with the environment illumination $E$ \cite{invrend_li}. Since we use a known environment illumination, we do not need to optimize for light sources within the scene. Having a known environment illumination allows for better BRDF recovery by constraining the optimization.
We first assign a diffuse material model to the proxy geometry $G$ with a randomly initialized set of material parameters (here, the diffuse color texture $M_D$). Given a single training image $\{I_j, p_j, m_j, E, G\}$ as input, we render the proxy geometry $G$ with camera pose $p_j$ and under the environment illumination $E$ using path tracing. We then optimize for the parameter $M_D$ as:
\begin{equation}
\label{eq:diff_rend_loss}
\argmin_{M_D} \sum_{k=1}^{N} \left | PT(G, E, M_D, p_k) \odot m_k - I_k \odot m_k \right |_{1},
\end{equation}
where $PT$ is a differentiable path tracer. The minima is estimated by backpropagating the $\ell_1$ loss through the differentiable renderer to modify the diffuse colour texture $M_D$ for the scene. Note, however, that our method is independent of the algorithm used for BRDF estimation.
\subsection{Computing SH coefficients of the BRDF}
\label{sect:compute_sh}
Disentanglement of appearance and irradiance is at the core of our
method. From a given viewpoint, we find the first
intersection to proxy geometry $G$ for each pixel using ray tracing and
construct a BRDF using the estimated material parameters $M$ (Fig
\ref{fig:network_overview}, bottom section). The BRDF function $f_r$ is
exactly defined over the unit sphere given $M$. We can then obtain the
SH coefficients $\bar{f_r}$ by evaluating a projection integral over the
unit sphere using Monte Carlo (MC) techniques \cite{jarosz08thesis,
prt_rr}. We evaluate this integral for each pixel and thus obtain an SH
coefficient vector of the BRDF at that pixel.
During training, we use a diffuse material model for this step and use
it to disentangle LIF from material. At inference time, this model can
be changed to any well defined material model typically used in
traditional rendering applications.
\subsection{Learning the LIF}
\label{sect:learn_LIF}
The disentangled LIF is learned using a neural network, encoding the
geometry and environment lighting into it. The network directly learns
the SH coefficient volume of the LIF. We start with the input set as
$\{I_k, p_k, m_k, E, G, M\}_{k=1}^{N}$, where $M$ denotes the estimated
material parameters. We train our network to learn the LIF, in terms of
its SH coefficients.
Fig \ref{fig:network_overview} shows our network architecture, which
follows directly from DNR \cite{dnr}. For a given viewpoint, we project
and sample a high dimensional Neural Texture based on the UV mapping,
which is input to the network. The output of the network is an SH
representation of the LIF for each pixel. From the same camera view, we
obtain the SH coefficient vector of the BRDF, as described in
Sect. \ref{sect:compute_sh}. We then take its dot product with the
learnt LIF from the network to obtain the final image. The dot product
is performed in a per-pixel and per-colour channel fashion resulting in
a output RGB image, which is compared with the ground truth for
training.
\section{Experimental Evaluation}
In this section, we provide implementational details of our method (Sect. \ref{sect:impl}). We further analyze and evaluate our method on synthetic and real scenes (Sect. \ref{sect:syn}, Sect. \ref{sect:real} resp.).
\begin{figure}
\centering
\includegraphics[width=\linewidth]{view_synthesis_comparison_synthetic.pdf}
\caption{\textbf{Results of novel view synthesis on synthetic data:} Qualitative comparison of view synthesis results with previous methods and the ground truth. Our method performs at par with the others. Refer to Table. \ref{tbl:view_synthesis_comparison_metrics_synthetic} for quantitative metrics.}
\label{fig:view_synthesis_comparison_synthetic}
\end{figure}
\begin{table}
\centering
\includegraphics[width=\linewidth]{view_synthesis_comparison_metrics_synthetic.pdf}
\caption{Quantitative metrics (PSNR/SSIM) for two synthetic scenes compared to DNR. The values are averaged over the test set.}
\label{tbl:view_synthesis_comparison_metrics_synthetic}
\end{table}
\subsection{Implementation Details}
\label{sect:impl}
We use the differentiable renderer Mitsuba 2 \cite{mitsuba2} for optimizing BRDF parameters (Sect. \ref{sect:diff_rend}) using the training images set. In practice, we optimize for multiple iterations per image since the optimizer and the scene need to be reinitialized for each new image. We repeat this process for around five epochs. Before training our network, we precompute the UV maps and the SH projection of the optimized BRDF. We do a minimal manual cleanup of the proxy geometry and remove parts of it that are not the object of interest (eg. ground, background objects).
Our network is based on the Deep Residual UNet \cite{zhang2018road, resnet} architecture following Gao et al. \cite{dnl} and is implemented in PyTorch \cite{pytorch}. We follow a similar methodology as DNR \cite{dnr} and use a four-level mipmap Laplacian pyramid for the 16-channel neural texture and set the top level's resolution to 512x512. We also multiply channels 3-12 of the neural texture with nine SH coefficients evaluated at per-pixel view direction. The output of our network represents the per-pixel LIF with a 75 channel volume corresponding to order four ($l=4$) SH projection (25 coefficients for each colour channel). We similarly project the BRDF to order four SH. We also output a binary mask from our network and post multiply with the LIF and the BRDF SH representation as is done by Gao et al.\cite{dnl}. The final image is obtained by a dot product of the LIF SH representation and the BRDF SH representation, which is then compared to the ground truth for training. We train our network using the Adam \cite{Kingma2015AdamAM} optimizer with $lr=1 \cdot e^{-4}$, $\beta_1=0.9$, $\beta_2=0.999$ and $\epsilon=1 \cdot e^{-8}$ and train the network for around 100 epochs.
\textbf{Loss Functions.} We use the perceptual loss paired with the $\ell_1$ loss to preserve sharp details. Specifically, we use the feature reconstruction loss from a pre-trained VGG16 \cite{vgg16} network, which is given by :
\begin{equation}
\mathcal{L}_{feat}^{j}(y', y) = \frac{1}{C_{j}H_{j}W_{j}}\left \| \phi_{j}(y') - \phi_{j}(y) \right \|_{2}^{2},
\end{equation}
where $\phi_{j}$ is the activation of the \textit{jth} convolutional layer with dimensions $C_{j}\times H_{j}\times W_{j}$ representing number of channels, width and height of the feature map, respectively. Here, $y$ denotes the predicted output and $y'$ is the ground truth. We use the \textit{relu\_3\_3} (\textit{j=relu\_3\_3}) \cite{relu} feature representation in our experiments. Thus, our final composite loss is given by:
\begin{equation}
\mathcal{L}_{comp} = \lambda_1 \ell_1 + \lambda_2 \mathcal{L}_{feat}^{relu\_3\_3}.
\label{eq:network_loss}
\end{equation}
We use $\lambda_1=0.8$ and $\lambda_2=0.2$ for all our experiments. We also apply an additional binary cross-entropy loss on the mask prediction, following Gao et al.\cite{dnl}. Both our composite loss $\mathcal{L}_{comp}$ and the mask loss are weighed equally.
\subsection{Results on synthetic scenes}
\label{sect:syn}
We first evaluate our network's results and analyze the learnt representation of the LIF on synthetic datasets. We render three synthetic scenes using the physically based renderer Mitsuba 2 \cite{mitsuba2}: \textit{Stanford Buddha}, \textit{Stanford Bunny} and \textit{Monkey}. The scenes are rendered with the camera looking at the object and positions sampled on a trajectory on the upper hemisphere. The only light source in the scene is the environment map. We use different trajectories for the training and test sets, containing 400 and 360 frames respectively. We render all scenes at a resolution of 512x512.
\textbf{Novel View Synthesis.} We perform a qualitative and quantitative comparison of view synthesis with DNR \cite{dnr}. We train both networks on the same train trajectory for two scenes, Stanford Bunny and Stanford Buddha. Fig. \ref{fig:view_synthesis_comparison_synthetic} and Table. \ref{tbl:view_synthesis_comparison_metrics_synthetic} shows the visual comparison and quantitative metrics (PSNR, SSIM) for the two scenes, respectively. Our network performance in at par with DNR, both quantitatively and qualitatively. We would like to emphasize here that our method can not only perform view synthesis but also edit the scene's appearance.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{bump_map.pdf}
\caption{\textbf{Bump map modification:} We apply a bump map (highlighted in blue) and use the constructed BRDF with our network trained on \textit{Stanford Buddha} with a white material (no bump map, top left). The middle and the right columns show application of varying strength (amplitude of the shading normal) of the bump map. Our method is able to plausibly handle bump map changes and generates photo-realistic effects of shading normals.}
\label{fig:bump_map}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{synthetic_mat_change.pdf}
\caption{\textbf{Diffuse colour and Material model edits:} (a) We edit the diffuse colour texture to a different one as shown in green insets. (b) We edit the material model from Diffuse BRDF to Rough Conductor BRDF. The network is trained on the scene shown in the left column (Stanford Buddha Red, Monkey Yellow).}
\label{fig:synthetic_mat_change}
\end{figure}
\textbf{Appearance Editing.} We perform three types of appearance edits: (1) Editing the diffuse color texture, (2) Changing the material model and (3) Editing the bump map of the material. Fig. \ref{fig:synthetic_mat_change}(a) shows the effect of changing the diffuse color texture and using our network to render the modified scene. We compare this to the ground truth, which is rendered using Mitsuba. Our network produces highly plausible results as compared to the ground truth. Next, we change the entire material model of the object in Fig. \ref{fig:synthetic_mat_change}(b). The scene was trained with the \textit{Diffuse} material model, which we change to the \textit{Rough Conductor} model provided in Mitsuba 2 with $\alpha=1.0$ (roughness). Our method is successfully able to handle this change, and produces a plausible image w.r.t. the ground truth. Finally, we change the normal map of the material to include deformations (Fig. \ref{fig:bump_map}), and compare to the ground truth render from Mitsuba. Our method is directly able to handle such changes, thanks to the disentangled formulation and representation of the LIF.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{buddha_texture.pdf}
\caption{Applying two different textures (top row, bottom row) to the \textit{Stanford Buddha} scene trained on two different (white and red) materials (first column, second column resp.), shown in insets. The error maps are shown in the rightmost column. The material changes being almost identical indicate that the LIF representation is accurate to a high degree.}
\label{fig:buddha_texture_interchange}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{bunny_light_interchange.pdf}
\caption{\textbf{Interchanging the learnt LIF representation:} Here, we train the \textit{Stanford Bunny} with the same material (white cloth) but different envmap lighting (green insets), and interchange the learnt LIF. The envmaps for the two scenes are shown in insets. The LIF interchange results being almost identical indicate that the LIF representation is accurate to a high degree. The error maps are between the ground truth lit scene and the LIF interchange output for the same envmap.}
\label{fig:bunny_light_interchange}
\end{figure}
\textbf{Learnt LIF representation.} We now analyze the learnt representation of the LIF. We conduct two types of analysis: (1) Appearance editing of two networks trained on the same scene having two different materials and (2) Interchanging the learnt LIF trained on differently lit scenes of the same object having the same material.
We conduct the first analysis by rendering two variants of the \textit{Stanford Buddha} scene with same lighting but different material (Fig. \ref{fig:buddha_texture_interchange}, yellow insets). Denote the first variant as \textit{Buddha Red} and the other as \textit{Buddha White}. We train two separate networks on both variants and render a novel view with a new diffuse colour texture. Fig. \ref{fig:buddha_texture_interchange} shows the results of applying and rendering the a new but same diffuse colour texture to networks trained on \textit{Buddha Red} and \textit{Buddha White}. The error maps with the ground truth are also shown.
For the second analysis, we render two differently lit variants of the \textit{Stanford Bunny} scene with a white cloth material (Fig. \ref{fig:bunny_light_interchange}). Denote the first variant as \textit{Envmap 1 Bunny} and the second as \textit{Envmap 2 Bunny}. We train two separate networks on both variants. From a novel viewpoint, we then render the \textit{Envmap 1 Bunny} with the learnt LIF of \textit{Envmap 2 Bunny} and vice-versa. Results are shown in Fig. \ref{fig:bunny_light_interchange}. The error map for the result of LIF interchange for \textit{Envmap 1 Bunny} is computed using absolute difference between the \textit{Envmap 1 Bunny} regular and LIF changed output. Similarly, we compute the error map for \textit{Envmap 2 Bunny}.
As seen from Fig. \ref{fig:buddha_texture_interchange} and Fig. \ref{fig:bunny_light_interchange}, results of application of the same texture from the first analysis and the LIF interchanged results from the second analysis very closely match the ground truth, which suggests that the LIF is practically independent of the underlying appearance.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{view_synthesis_comparison_real.pdf}
\caption{\textbf{Results of novel view synthesis on real data:} Qualitative comparison of view synthesis results with previous methods and the ground truth. Our method performs at-par with the others. Refer to Table. \ref{tbl:view_synthesis_comparison_metrics_real} for quantitative metrics.}
\label{fig:view_synthesis_comparison_real}
\end{figure}
\begin{table}[t]
\centering
\includegraphics[width=\linewidth]{view_synthesis_comparison_metrics_real.pdf}
\caption{Quantitative metrics (PSNR/SSIM) for two real scenes, \textit{Fish} and \textit{Bird}, compared to DNR. The values are averaged over the test set.}
\label{tbl:view_synthesis_comparison_metrics_real}
\end{table}
\subsection{Results on Captured Scenes}
\label{sect:real}
We capture our own scene (\textit{Fish}) with a handheld mobile phone and use two publicly available scenes from the DTU \cite{dtu} dataset (\textit{Bird}, \textit{Hands}). We capture two trajectories for the \textit{Fish} scene, and sample 400 frames for train and 300 frames test sets. For the DTU scenes, we use the \textit{diffuse} lit images and manually remove frames with drastic illumination changes. We use 200 frames for training and 13 for testing. We use COLMAP \cite{colmap} to obtain camera parameters and dense point cloud and perform Poisson surface reconstruction \cite{poisson_recon} to obtain the 3D proxy geometry. We remove parts of the obtained mesh that are not the object of interest (eg. ground, background objects) and use U2Net \cite{u2net} to obtain masks from the ground truth images. We obtain the environment map using the Google Street View app on a consumer mobile phone. Since we do not have the envmaps for the \textit{Bird} and \textit{Hands} scenes, we approximate it with a diffuse white area light placed on top of the geometry, drawing on the observation that the captured images have diffuse uniform white light projecting from the top.
\textbf{Novel View Synthesis.} Fig. \ref{fig:view_synthesis_comparison_real} shows the qualitative results and Table. \ref{tbl:view_synthesis_comparison_metrics_real} show the quantitative values for our own scene and the \textit{Bird} scene in comparison with DNR. Even though our quantitative values are slightly lesser, our method produces plausible and photo-realistic view synthesis results.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{mat_edit_real.pdf}
\caption{\textbf{Appearance editing on real scenes:} Results of editing the appearance on \textit{Bird}, \textit{Hands} and \textit{Fish}. The modified diffuse textures are highlighted in green. We change the over all diffuse texture for the \textit{Bird} scene and paint over parts of the mesh for the \textit{Hands} scene and the \textit{Fish} scene. Naive colmap rendering is unable to render geometric detail and textures sharpness, which is accurately reproduced by our method. }
\label{fig:mat_edit_real}
\end{figure}
\textbf{Appearance Editing.} We perform appearance edits on all three real scenes. Results are shown in Fig. \ref{fig:mat_edit_real}, with the changed diffuse texture highlighted in green. Note that for real scenes, the UV map does not necessarily make semantic sense. We compare our method's results with the rendering of colmap geometry applied with the edited material. Our network preserves details that naive colmap geometry rendering is unable to reproduce, such as the texture detail and the underlying geometry variation.
\begin{figure}[t!]
\centering
\includegraphics[width=\linewidth]{loss_ablation.pdf}
\caption{\textbf{Loss function ablation:} Comparison between the $\ell_1$ photometric loss and our composite loss. The composite loss helps preserve the sharp details.}
\label{fig:loss_ablation}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\linewidth]{envmap_vis.pdf}
\caption{\textbf{Visualization of the learnt LIF:} We visualize the learnt LIF representation at the marked pixel location and compare with the ground truth environment map. Both scenes have the same material but are lit with different illumination. The learnt LIF representation follows the high level structures of the ground truth without dependence on the underlying material. Note that the \textit{normal} is pointing upward at the marked point.}
\label{fig:envmap_vis}
\end{figure}
More results on synthetic and real scenes with comparisons and analysis can be seen in the supplementary document and video.
\section{Discussions}
\textbf{Learnt LIF visualization.} In Fig. \ref{fig:envmap_vis}, we visualize two learnt LIF representations trained on the \textit{Stanford Bunny} scene with white cloth diffuse material with two different illuminations. We compare this to the ground truth environment map that was used to illuminate the scene. Fig. \ref{fig:envmap_vis} visualized the LIF representation for a point whose normal points upward. We deliberately choose such a point for easy alignment of the ground truth envmap and the LIF. The learnt LIF follows high level structures of the ground truth envmap. Even though the underlying material is a white cloth, there is minimal leaking of the appearance on the LIF trained on Bunny lit with the reddish envmap.
\textbf{Ablations.} We perform an ablative study on the choice of our loss function. Specifically, we train our method with only the $\ell_1$ loss and with our composite loss function described in Eq. \ref{eq:network_loss}. This comparison is shown in Fig. \ref{fig:loss_ablation}. Our composite loss is effective at retaining sharp and high frequency details.
\textbf{Limitations.} One of the major limitations of our method is the reliance on plausible BRDF estimation. Currently, we use a differentiable rendering optimization, but such an optimization is prone to errors if the geometry is highly inaccurate. This inaccuracy propagates to the learning stage, which results in over-compensation by the network. This directly affects its ability to learn a LIF representation that is truly appearance independent. Another limitation is that the material model cannot be highly glossy or specular. This is in part due to the inability of SH to represent high frequency details and in-part due to the inadequacy of the learnt LIF representation. Handling glossy edits to the appearance requires further analysis and possibly a different basis representation.
\section{Conclusions and Future work}
In this paper, we propose a neural rendering framework that achieves simultaneous appearance editing and view synthesis. This problem is orthogonal to current approaches that achieve relighting along with view synthesis. Our method proposes a BRDF independent formulation of the local irradiance function (LIF) and its subsequent projection to SH basis. We further present a method to to recover the BRDF and LIF from unstructured photographs, where only the latter component involves learning. This allows us to modify the underlying material of the scene to generate photo-realistic and plausible renderings, as demonstrated. Our method's performance is at par with the current state-of-the art approaches in view synthesis while adding the capability of appearance editing.
For future work, we would like to explore other basis representations for representing the LIF and the BRDF. Such bases could help extend our method to handle glossy or even nearly-specular material edits. Another interesting direction of research is to incorporate relighting into our method while retaining the ability to perform appearance editing and view synthesis.
{\small
\bibliographystyle{ieee_fullname}
|
\section{Auto-tuning for Production Simulations} \label{sec:online-preco} The
results of the preceding section provides a small window into the varieties of
performance behavior encountered in actual production cases, which span a large
range of problem sizes, domain topologies, mesh qualities. Moreover,
production simulations are solved across a range of architectures having
varying on-node and network performance, interconnect topologies, and processor
counts. One frequently encounters situations where certain communication
patterns might be slower under a particular MPI version on one platform versus
another. Unless a developer has access to that platform, it's difficult to
measure and quantify the communication overhead. Processor count alone can be
a major factor in preconditioner selection: large processor counts have
relatively high coarse-grid solve costs that can be mitigated by doing more
smoothing at the fine and intermediate levels. How much more is the open
question. The enormity of parameter space, particularly ``in the field''
(i.e., users working on unknown platforms) limits the effectiveness of standard
complexity analysis in selecting the optimal preconditioner for a given user's
application.
From the user's perspective, there is only one application (at a time,
typically), and one processor count of interest. Being able to provide
optimized performance---tuned to the application at hand, which includes the
processor count---is thus of paramount importance. Auto-tuning provides an
effective way to deliver this performance. Auto-tuning of preconditioners has
been considered in early work by Imakura {\em at al.}
\cite{imakura_auto-tuning_2012} and more recently by Yamada {\em at al.}
\cite{yamada2018preconditioner} and by Brown {\em et al.} \cite{brown_tuning_2021}.
The latter work couples their auto-tuning with local Fourier analysis to guide
the tuning process.
\input{tables/parameter-space.tex}
In large-scale fluid mechanics applications, auto-tuning overhead is typically
amortized over $10^4$--$10^5$ timesteps (i.e., pressure solves) {\em per run}
(and more, over an entire simulation campaign). Moreover, auto-tuning is of
particular importance for problems at large processor counts because these
cases often have long queue times, which preclude making multiple job
submissions in order to tweak parameter settings. Failure to optimize,
however, can result in significant opportunity costs. For example, in
\cite{fischer_highly_2021} the authors realized a factor of 2.8 speedup in time-per-step for a
352,000-pebble-bed simulation ($n$=51B gridpoints) through a sequence of tuning
steps. Even if there were 100 configurations in the preconditioner parameter
space, an auto-tuner could visit each of these in succession 5 times each within
the first 500 steps and have expended only a modest increment in overhead when
compared to the cost of submitting many jobs (for tuning) or to the cost of
10,000 steps for a production run.
As a preliminary step, the authors consider constructing
a small subset of the true search space (which could include, e.g., other cycles, schedules,
or smoothers) for use in a nascent auto-tuner.
Across all cases, with exception to the 67 pebble case, pMG preconditioning
with Chebyshev-accelerated ASM or RAS smoothing is the fastest solver or is comparable to SEMFEM.
The choice of Chebyshev order and multigrid schedule, moreover,
contributes only a modest $\approx$ 10-20\% improvement to the overall time-to-solution in most cases, all else being equal.
This makes the default Cheby-ASM(2),(7,3,1) or Cheby-ASM(2),(9,5,1) preconditioner
reasonably performant.
However, in order to avoid the situation encountered in the 67 pebble case, SEMFEM is added
to the considered search space.
The parameters in Table \ref{table:auto-tuner-parameter-space} are chosen as they include optimal or near-optimal preconditioner settings
for the results in Figs. \ref{fig:scaling-study-bsb}, \ref{fig:scaling-study-pb},
while still restricting the search space to something that is amenable to exhaustive search.
The authors note, however, that this parameter space may not reflect the various factors
affecting the performance of the preconditioners at especially large $P$ or on different machine architectures.
During the first timestep, our simple auto-tuner performs an exhaustive search over
the small parameter space identified in
Table \ref{table:auto-tuner-parameter-space}.
While this space is limited compared
to the true search space, the resultant preconditioners selected are effective.
In the 146 pebble case (Fig. \ref{fig:scaling-study-pb}a), Cheby-RAS(2),(7,3,1)
is identified as the preconditioner on each of the processor counts,
which was comparable in peformance to the best preconditioner.
The auto-tuner chose the optimal SEMFEM for the 1568 pebble case (Fig. \ref{fig:scaling-study-pb}b).
SEMFEM was identified at the preconditioner for the 67 pebble case on all processor counts,
which Fig. \ref{fig:scaling-study-pb}c confirms.
In the Boeing speed bump case (Fig. \ref{fig:scaling-study-bsb}), the auto-tuner
chose Cheby-RAS(2),(9,5,1) across all processor counts, which was either the optimal
or near-optimal preconditioner for the problem.
Note that the auto-tuner selects the fastest method at a {\em fixed-processor
count} (i.e., whatever the user has selected).
Consider, for example, the 67 pebble case on $P=18$ GPUs (dashed black line, Fig. \ref{fig:scaling-study-pb}c).
The goal of the auto-tuner is to select the preconditioner with lowest time-to-solution
along the user-specified iso-processor count line.
However, in the results discussed above, there was no change with respect to a change in the processor count.
We note that our principal
objective is not to squeeze out a few percent over a raft of good choices, but
rather to ensure that a case does not run with an unfortunate set of parameters
for which the performance is significantly substandard. This primitive
auto-tuning technique proves effective at preventing the selection of highly
suboptimal preconditioners. We have implemented this for production use and
will continue to update and refine the strategy moving forward.
\subsection{Poisson} \label{kershaw}
The Kershaw family of meshes \cite{kolev_ceed_2021,kershaw_differencing_1981}.
has been proposed as the basis for a high-order Poisson-solver benchmark by
Center for Efficient Exascale Discretization (CEED) within the DOE Exascale
Computing Project (ECP). This family is parametrized by an anisotropy measure,
$\varepsilon =\varepsilon_y=\varepsilon_z \in (0,1]$, that determines the
degree of deformation in the $y$ and $z$ directions. As $\varepsilon$
decreases, the mesh deformation and aspect ratio increase along with it.
The Kershaw mesh is shown in Fig. \ref{fig:kershaws} for several values
of $\varepsilon$. The domain $\Omega = [-1/2,1/2]^3$ with Dirichlet boundary
conditions on $\partial\Omega$. The right hand side for (\ref{eqn:poisson}) is
set to
\begin{equation}\label{eq:kershaw-rhs}
f(x,y,z) = 3\pi^2 \sin{(\pi x)}\sin{(\pi y)}\sin{(\pi z)} + g
\end{equation}
where $g(x,y,z)$ is a random, continous vector vanishing on $\partial\Omega$.
The linear solver terminates after reaching a relative residual reduction of
$10^{-8}$. Since this test case solves the Poisson equation, there is no
timestepper needed for the model problem.
\subsection{Navier-Stokes} \label{navier-stokes}
\begin{figure}
\begin{subfigure}{0.15\textwidth}
\includegraphics[width=\textwidth]{figs/kershaw-eps-1.0.png}
\end{subfigure}
\hfill
\begin{subfigure}{0.15\textwidth}
\includegraphics[width=\textwidth]{figs/kershaw-eps-0.3.png}
\end{subfigure}
\hfill
\begin{subfigure}{0.15\textwidth}
\includegraphics[width=\textwidth]{figs/kershaw-eps-0.05.png}
\end{subfigure}
\caption{
\label{fig:kershaws}
Kershaw, $E=12^3,p=1$. $\varepsilon=1.0,0.3,0.05$.
}
\end{figure}
\input{tables/problem-sizes}
\begin{figure}[h] \centering
{\setlength{\unitlength}{0.9in}
\begin{picture}(3.2,4.4)(0,0)
\put(0.00,2.27){\includegraphics[width=1.26in]{figs/pb146.png}}
\put(1.59,2.30){\includegraphics[width=1.44in]{figs/pb1568.png}}
\put(0.40,1.15){\includegraphics[width=2.70in]{figs/peb67-side-profile.png}}
\put(0.40,0.00){\includegraphics[width=2.70in]{figs/BoeingSpeedBump.png}}
\put(0.00,2.30){\large (a)}
\put(1.55,2.33){\large (b)}
\put(0.00,1.23){\large (c)}
\put(0.00,0.08){\large (d)}
\end{picture}}
\caption{Navier-Stokes cases: pebble-beds with
(a) 146, (b) 1568, and (c) 67 spheres;
(d) Boeing speed bump.
\label{fig:ns_cases}
}
\end{figure}
For the pressure-Poisson tests, four flow cases are considered, as depicted
in Fig. \ref{fig:ns_cases}.
The first three cases corresponds to turbulent flow through a cylindrical packed-bed
with 146, 1568, and 67 spherical pebbles.
The 146 and 1568 pebble cases are from Lan and coworkers \cite{lan_all-hex_2021}.
The 67 pebble case is constructed using a tet-to-hex meshing strategy by Yuan and coworkers \cite{yuan2020spectral}.
The first two bed flows are at Reynolds number $Re_D=5000$, based on sphere
diameter, $D$, while the 67 pebble case is at Reynolds number $Re_D=1460$. Time advancement is based on a two-stage 2nd-order
characteristics timestepper with CFL=4
($ \Delta t = 2\times10^{-3}$
$ \Delta t = 5\times 10^{-4}$, and
$ \Delta t = 5\times 10^{-5}$ for the 146, 1568, and 67 pebble cases).
An absolute pressure solver tolerance of $10^{-4}$ is used. A restart at
$t=10$, $t=20$, and $t=10.6$ convective time units is used for the 146, 1568, and 67 pebble
cases, respectively, to provide an initially turbulent flow.
The fourth case, shown in Fig. \ref{fig:ns_cases}d, is a direct numerical
simulation (DNS) of seperated turbulent flow over a speed bump at $Re = 10^6$.
This test case was designed by Boeing to provide a flow that exhibits
separation. A DNS of the full 3D geometry, however, remains difficult
\cite{shur_direct_2021}. Therefore, this smaller example proves a useful
application for benchmarking solver performance. This case uses a 2nd-order
timestepper with CFL=0.8 ($\Delta t = 4.5\times 10^{-6}$) and an absolute
pressure-solve tolerance of $10^{-5}$. A restart at $t=5.6$ convective time
units is used for the initial condition.
In all cases, solver results are collected over 2000 timesteps. At each step,
the solution is projected onto a space of up to 10 prior solution vectors to
generate a high-quality initial guess, ${\bar {\underline u}}$. Projection is standard
practice in nekRS as it can reduce the initial residual by orders of magnitude
at the cost of just one or two matrix-vector products in $A$ per step
\cite{fischer_projection_1998}.
The perturbation solution, $\delta {\underline u} := {\underline u} - {\bar {\underline u}}$, is
typically devoid of slowly evolving low wave-number content. Moreover, the
initial residual is oftentimes sufficiently small that the solution converges
in $k < 5$ iterations, such that the $O(k^2)$ overhead of GMRES is small.
Testing the preconditioners under these conditions ensures that the conclusions
drawn are relevant to the application space.
\subsection{SE Poisson Discretization}
Consider the Poisson equation in $\mathbb R^3$,
\begin{align}
-\nabla^2 u = f \text{ for } u,f \in \Omega \subset \mathbb R^3 \mapsto \mathbb R.
\label{eqn:poisson}
\end{align}
Boundary conditions for the pressure Poisson equation are either periodic,
Neumann, or Dirichlet, with the latter typically applicable only on a small
subset of the domain boundary, $\partial \Omega_D$, correspondong to an outflow condition.
Consequently, Neumann conditions apply over the majority (or all) of the domain
boundary, which makes the pressure problem more challenging than the standard
all-Dirichlet case.
The SE discretization of (\ref{eqn:poisson}) is based on the weak form:
{\em Find $u \in X_0^p$ such that,}
\begin{align}
(\nabla v, \nabla u)_p = (v,f)_p
\;\;
\forall v \in X_0^p,
\end{align}
where $X_0^p$ is a finite-dimensional approximation comprising the basis
functions used in the SE discretization, ${\phi_j(\mathbf x)}$, $j=1,\ldots,n$,
that vanish on $\partial \Omega_D$ and $(\cdot,\cdot)_p$ represents the discrete $L^2$
inner product based on Gauss-Lobatto-Legendre quadrature in the reference
element, ${\hat \Omega}:=[-1,1]^3$.
The basis functions allow us to represent the solution, $u$, as $u(\mathbf x) =
\sum u_j \phi_j(\mathbf x)$, leading to a linear system of unknown basis
coefficients,
\begin{align}
A \underline u = B \underline f,
\end{align}
with respective mass- and stiffness-matrix entries,
$B_{ij} := (\phi_i,\phi_j)_p$, and
$A_{ij} := (\nabla \phi_i, \nabla \phi_j)_p$.
$\Omega$ is tesellated into nonoverlapping hexahedral elements, $\Omega^e$,
for $e=1,\ldots,E$, with isoparametric mappings from ${\hat \Omega}$ to $\Omega^e$
provided by
$\mathbf x^e(r,s,t) = \sum_{i,j,k}\mathbf x_{ijk}^e h_i(r) h_j(s) h_k(t)$,
for $i,j,k\in[0,p]$.
Each $h_*(\xi)$ is a $p$th-order Lagrange cardinal polynomial on
the Gauss-Lobatto-Legendre (GLL) qudrature points, $\xi_j \in [-1,1]$.
Similarly, the test and trial functions $u,v$ are written in local form as
$u^e(r,s,t) = \sum_{i,j,k}u_{ijk}^e h_i(r) h_j(s) h_k(t)$.
Continuity is ensured across the interface between adjacent elements by
enforcing $u_{ijk}^e = u_{\hat i\hat j\hat k}^e$
when $\mathbf x_{ijk}^e = \mathbf x_{\hat i\hat j\hat k}^e$.
From this, a global-to-local degree-of-freedom mapping,
${\underline u} := \{u_l \} \longrightarrow {\underline u}_L := \{ u_{ijk}^e \}$,
can be represented by a Boolean matrix $Q$, such that ${\underline u}_L = Q{\underline u}$.
The assembled stiffness matrix is then $A=Q^TA_L Q$, where $A_L =
\operatorname{block-diag}(A^e)$ comprises the local stiffness matrices, $A^e$.
Similarly, $B=Q^TB_LQ$. The SE formulation uses coincident GLL quadrature and
nodal points, such that $B^e$ is diagonal. Moreover, $A^e$ is never formed, as
it would contain $O(p^6)$ nonzeros in the general case. Rather, the
tensor-product-sum factorization \cite{parter_spectral_1979} allows for
$A\underline{u}$ to be evaluated in $O(Ep^4)$ time with $O(Ep^3)$ storage,
as described in detail in \cite{deville_high-order_2002}.
\subsection{Preconditioners}
In the current study, all preconditioners are applied in the context of
restarted GMRES. Although $A$ is symmetric positive definite (SPD), many of the
preconditioners are asymmetric. Further, a recent study
\cite{fischer_highly_2021} has shown the benefits of projection-based GMRES
over flexible conjugate gradients (FCG) because the effectiveness of the overall
pressure solution strategy (which includes projection onto prior solutions
\cite{fischer_projection_1998}) generally ensures a low enough iteration
count, $k$, such that the $O(k^2)$ costs in GMRES are not overly onerous.
We note, however, that FCG may yield very low iteration counts
in certain cases (e.g., 1--2 iterations per step, as in \cite{fischer_nekrs_2021}),
in which case we use FCG rather than GMRES if it yields faster runtimes.
\subsubsection{SEMFEM} \label{sub:semfem}
In \cite{parter_spectral_1979}, Orszag suggested that constructing a sparse
preconditioner based on the low-order discretizations with nodes coinciding
with those of the high-order discretization would yield bounded condition
numbers and, under certain constraints, can yield $\kappa(M^{-1}A) \sim \pi^2 /
4$ for second-order Dirichlet problems. This observation has led to the
development of preconditioning techniques based on solving the resulting low-order system
\cite{olson_algebraic_2007,bello-maldonado_scalable_2019,canuto_finite-element_2010}.
In the current work, we employ the same low-order discretization
considered in \cite{bello-maldonado_scalable_2019}. Each of the vertices of
the hexahedral element is used to form one low-order, tetrahedral element,
resulting in a total of eight low-order elements for each GLL sub-volume
in each of the high-order hexahedral elements.
This low-order discretization is then used to form the sparse operator, $A_F$.
The so-called weak preconditioner, $A_F^{-1}$, is used to precondition the system.
Algebraric multigrid (AMG), implemented in CUDA in AmgX \cite{naumov_amgx_2015},
is used with the following setup to solve the low order system: \\[-2.5ex]
\begin{itemize}
\item PMIS coarsening \\[-3.90ex]
\item 0.25 strength threshold \\[-3.90ex]
\item Extended + i interpolation ($p_{max}=4$) \\[-3.90ex]
\item Damped Jacobi relaxation ($0.9$) \\[-3.90ex]
\item One V-cycle for preconditioning \\[-3.90ex]
\item Smoothing on the coarsest level \\[-3.90ex]
\end{itemize}
We denote this preconditioning strategy as SEMFEM.
\subsubsection{$p$-multigrid, Schwarz Smoothers}
Another preconditioning strategy for the SE-based Poisson problem is to use
geometric $p$-multigrid (pMG). The classical single pass V-cycle is
summarized in Algorithm \ref{alg:V-cycle}. In our application, we
limit pMG preconditioning to a single V-cycle pass. Alternative
strategies, such as F- or W-cycle multigrid are not considered.
\input{algs/v-cycle}
\begin{figure} \centering
{\setlength{\unitlength}{1.000in}
\begin{picture}(3.2,1.5)(0,1)
\put(0.00,1.3){\includegraphics[width=1.50in]{figs/schwarz-approximation.png}}
\put(1.59,1.30){\includegraphics[width=1.50in]{figs/overlapping-schwarz-diagram-p=5.png}}
\put(0.00,1.30){\large (a)}
\put(1.55,1.33){\large (b)}
\end{picture}}
\vspace{-0.5cm}
\caption{
\small
(a) Box-like approximation (b) overlapping domain
\label{fig:schwarz-overlapping}
}
\end{figure}
The SE-based additive Schwarz method (ASM) presented in
\cite{lottes_hybrid_2005,loisel_hybrid_2008} solves local Poisson
problems on subdomains that are extensions of the spectral elements.
The formal definition of the ASM preconditioner (or pMG smoother) is
\begin{eqnarray} \label{eq:asm}
{\underline s} &=& \sum_{e=1}^E W_e R_e^T {\bar A}_e^{-1} R_e {\underline r},
\end{eqnarray}
where $R_e$ is the restriction matrix that extracts nodal values
of the residual vector that correspond to each overlapping domain,
as indicated in Fig. \ref{fig:schwarz-overlapping}b.
To improve the smoothing properties of the ASM, we introduce
the diagonal weight matrix, $W_e$, which scales each nodal value by
the inverse of the number of subdomains that share that node.
Although it compromises symmetry, post-multiplication by $W_e$ was found to yield
superior results to pre- and post-multiplication by $W_e^{\frac{1}{2}}$
\cite{stiller_nonuniformly_2017,lottes_hybrid_2005}.
In a standard Galerkin ASM formulation, one would use ${\bar A}_{e} = R_e A
R_e^T$, but such an approach would compromise the $O(p^3)$ storage complexity
of the SE method. To construct fast inverses for ${\bar A}_e$, we approximate
each deformed element as a simple box-like geometry, as demonstrated in Fig.
\ref{fig:schwarz-overlapping}a. These boxes are then extended by a single
degree-of-freedom in each spatial dimension to form overlapping subdomains with
${\bar p}^3=(p+3)^3$ interior degrees-of-freedom in each domain.
(See Fig. \ref{fig:schwarz-overlapping}).
The approximate box domain enables the use of the fast diagonalization method
(FDM) to solve for each of the overlapping subdomains, which can be applied in
$O(Ep^4)$ time in $\mathbb R^3$\cite{lottes_hybrid_2005}. The
extended-box Poisson operator is separable, with the 3D form
\begin{equation}\label{eq:tensor-prod-poisson}
{\bar A} =
B_z \otimes B_y \otimes A_x+B_z \otimes A_y \otimes B_x+A_z \otimes B_y \otimes B_x,
\end{equation}
where each $B_*,A_*$ represents the extended 1D mass-stiffness matrix pairs
along the given dimension \cite{lottes_hybrid_2005}. The FDM begins with a
preprocessing step of solving a series of small,
${\bar p}\times {\bar p}$, generalized eigenvalue problems,
\begin{equation}\label{eq:generalized-eig}
A_* \underline{s}_i = \lambda_i B_* \underline{s}_i
\end{equation}
and defining
$S_*=(\underline{s}_1\ldots\underline{s}_{\bar p})$ and
$\Lambda_*=\operatorname{diag}(\lambda_i)$, to yield the
similarity transforms
\begin{equation}\label{eq:similarity}
S_*^T A_* S_* = \Lambda_*, \;\; S_*^T B_* S_* = I.
\end{equation}
From these, the inverse of the local Schwarz operator is \\[-2.5ex]
\begin{equation}\label{eq:inverse-tensor-prod-poisson}
{\bar A}^{-1} =
(S_z\otimes S_y\otimes S_x) D^{-1} (S_z^T \otimes S_y^T \otimes S_x^T),
\end{equation}
where
\begin{equation}
D=I\otimes I\otimes \Lambda_x+I\otimes \Lambda_y \otimes I+\Lambda_z \otimes I \otimes I
\end{equation}
is a diagonal matrix. This process is repeated for each element, at each
multigrid level save for the coarsest one. Note that the
per-element storage is only $3{\bar p}^2$ for the $S_*$ matrices and ${\bar
p}^3$ for $D$. At each multigrid level, the local subdomain solves are used as
a smoother. On the coarsest level ($p=1$), however, BoomerAMG
\cite{henson_boomeramg_2002} is used to solve the system with the same
parameters as the AmgX solver in section \ref{sub:semfem}, except using
Chebyshev smoothing. Unless otherwise noted, all pMG preconditioners use a
single BoomerAMG V-cycle iteration in the coarse-grid solve.
Presently, we also consider a restrictive additive Schwarz (RAS) version
of (\ref{eq:asm}), wherein overlapping values are not added after the action
of the local FDM solve, following \cite{ras99}. RAS has the added benefit of
reducing the amount of communication required in the smoother. (Formally,
RAS can be implemented by simply changing $W_e$.)
Note that, in the case of ASM and RAS smoothers without Chebyshev-acceleration,
an additive V-cycle with no post-smoothing is used to avoid residual
re-evaluation in Algorithm \ref{alg:V-cycle} \cite{fischer04}.
\subsubsection{Chebyshev Acceleration} \label{sub:cheby}
A notable improvement over standard Jacobi-smoothed multigrid is to use
Chebyshev-acceleration \cite{sundar_comparison_2015,adams_parallel_2003},
as described in Algorithm \ref{alg:cheby} for a given surrogate smoother, $S$.
While $S$ is typically based on Jacobi smoothing (e.g.,\cite{Kronbichler2019}),
it is also possible to consider the use of overlapping-Schwarz (\ref{eq:asm})
as the smoother, which is a new approach that we explore here.
\input{algs/cheby-acceleration}
Algorithm \ref{alg:cheby} requires approximate spectral bounds,
$(\lambda_{min}, \lambda_{max})$, of the smoothed operator, $SA$.
Using 10 rounds of Arnoldi iteration, we generate $\tilde\lambda$ as an initial
proxy for $\lambda_{max}$. The bounds employed in Algorithm \ref{alg:cheby}
are then determined through sensitivity analysis similar to that performed by
Adams and coworkers \cite{adams_parallel_2003}. The analysis is conducted
using a challenging model problem, namely the Kershaw case of subsection
\ref{kershaw}, with $\varepsilon=0.3$, $E=24^3,p=7$,
and relative ($2$-norm) residual tolerance of $10^{-8}$.
A second-order Chebyshev-accelerated ASM smoother is used in the $p=7,3,1$
pMG V-cycle preconditioner, which we denote as the current default
preconditioner in nekRS.
The results of the sensitivity study are shown in
Table \ref{table:kershaw-eig-multiplier-0.3}.
The required iteration counts are seen to be sensitive to underestimation of
$\lambda_{max}$, as also found by Adams and coworkers \cite{adams_parallel_2003}.
On the other hand, results are less sensitive to the minimum eigenvalue
estimate, provided $\lambda_{min} > 0$, as this delimits between the low
frequencies that are handled by the coarse grid correction and the high
frequencies that are eliminated by the action of the smoother.
The choice of the bounds (relative to ${\vspace*{.11in}{\tilde \lambda}}$) is not universal:
(1/30,1.1) \cite{adams_parallel_2003},
(0.3,1) \cite{baker_multigrid_2011},
(0.25,1) \cite{sundar_comparison_2015},
and
(1/6,1) \cite{zhukov_multigrid_2015} have all been considered.
Based on the results of Table \ref{table:kershaw-eig-multiplier-0.3}, we set
$(\lambda_{min},\lambda_{max})=(0.1,1.1){\vspace*{.11in}{\tilde \lambda}}$ as a conservative
choice.
\input{tables/kershaw-eig-multiplier-0.3}
\section{Introduction}
\label{introduction}
\input{introduction}
\section{Background and Implementation}
\label{implementation}
\input{implementation}
\section{Test Cases}
\label{cases}
\input{cases}
\section{Results}
\label{results}
\input{results}
\input{auto}
\section{Conclusions and Future Work}
\label{conclusions}
\input{conclusions}
\section{Acknowledgements}
\label{acknowledgements}
\input{acknowledgements}
\bibliographystyle{siamplain}
\subsection{Kershaw Mesh}
The Kershaw study comprises six tests. For each of two studies, we consider
the regular box case ($\varepsilon=1.0$), a moderately skewed case
($\varepsilon=0.3$), and a highly skewed case ($\varepsilon=0.05$).
The first study is a standard weak-scale test, where $P$ and
$E$ are increased, while the polynomial order is fixed at $p=7$ and the number
of gridpoints per GPU is set to $n/P = 2.67M$. The range of processors is
$P$=6 to 384. The second study is a test of the influence of polynomial order
on conditioning, with $P=24$ and $n/P=2.88M$ fixed, while $p$ ranges from 3 to 10.
Both cases use GMRES(20).
The results of the weak-scaling study are shown in Fig.
\ref{fig:kershaw-weak-scaling}.
For all values of $\varepsilon$, the iteration count exhibits a dependence
on problem size, as seen in Fig. \ref{fig:kershaw-weak-scaling}a,d,g,
especially in the highly skewed case ($\varepsilon = 0.05$).
The time-per-solve also increases with $n$, in part due to the increase
in iteration count, but also due to increased communication overhead as
$P$ increases.
This trend is not necessarily monotonic, as shown in Fig. \ref{fig:kershaw-weak-scaling}c,f.
In the case of $\varepsilon=0.3$ Cheby-RAS(2),(7,3,1), a minor fluctuation in the iteration count
from the $P=6$ to $P=12$ case causes the greater than unity parallel efficiency.
For $\varepsilon=1.0$ Cheby-Jac(2),(7,5,3,1), the effects of system noise
on the $P=6$ run causes the greater than unity parallel efficiency
for the $P=12$ run.
Table \ref{table:max-neighbors}
indicates the maximum number of of neighboring processors for the assembly
($QQ^T$) graph of $A$, which increases with $P$. In addition, the number of
grid points per GPU ($n/P=2.67M$) is relatively low. These two factors cause
an increased sensitivity of the problem to the additional communication
overhead as the number of GPUs is initially increased. The number of
neighbors, however, will saturate at larger (i.e., production-level) processor
counts.
Lastly, the relative preconditioner performance depends on $\varepsilon$.
Fig. \ref{fig:kershaw-weak-scaling}b, e, h show that, for the easy $\varepsilon=1.0$ case,
a pMG scheme with a smoother that is cheap to apply is best, such as
Cheby-RAS(1),(7,3,1), Cheby-Jac(2),(7,5,3,1), and ASM,(7,3,1).
However, as $\varepsilon$ decreases, more robust pMG smoothers such as
Cheby-RAS(2),(7,3,1) and SEMFEM result in lower time to solution.
Once $\varepsilon = 0.05$ (Fig. \ref{fig:kershaw-weak-scaling}h), the problem is sufficiently challenging that
SEMFEM overtakes the pMG based preconditioning schemes.
This indicates that, in the highly skewed case in which the maximum element aspect ratio
increases, the pMG preconditioner is not as effective as the SEMFEM preconditioner.
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{figs/kershaw-weak-scaling.png}
\vspace*{-1cm}
\caption{
\small
Kershaw weak scaling, GMRES(20).
$n/P = 2.67M$.
\label{fig:kershaw-weak-scaling}
}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{figs/kershaw-order-scaling.png}
\vspace*{-0.8cm}
\caption{
\small
Kershaw order dependence, GMRES(20).
$n/P = 2.88M$.
\label{fig:kershaw-order-dependence}
}
\end{figure*}
\input{tables/max-neighbors}
The influence of polynomial order is illustrated in Fig.
\ref{fig:kershaw-order-dependence}. For $\varepsilon=1.0$, iteration counts
are essentially $p$-independent, as seen in Fig.
\ref{fig:kershaw-order-dependence}a. For $\varepsilon=0.3$, however, a slight
upward trend in the iteration count is observed for SEMFEM and for the pMG
preconditioners with ASM, RAS, and Chebyshev-Jacobi smoothing
(Fig. \ref{fig:kershaw-order-dependence}b).
Similarly, $\varepsilon = 0.05$ exhibits a dependence between the iteration count and the polynomial order
for all preconditioners (Fig. \ref{fig:kershaw-order-dependence}c).
Figures \ref{fig:kershaw-order-dependence}d, e and f demonstrate that the {\em time} per
pressure solve is strongly dependent polynomial order.
For SEMFEM, there is an increase in time as a result of increased overhead for the underlying AMG solver.
For the pMG-based methods, higher orders are generally faster. This
performance gain can be attributed to surface-to-volume effects in the
evaluation of the operators and smoothers. Application of $A^e$ and (smoother)
$S^e$ is highly vectorizable, whereas application of $QQ^T$ (assembly) involves
a significant amount of indirect addressing. The discrete surface to volume
ratio for a spectral element of order $p=7$ is 296/512 $\approx$ 60\%. For
lower $p$ this value is larger. This situation is exacerbated in the case of
Schwarz-based smoothers, where the overlap contributes substantially to the
work and communication for the small-element (i.e., low-$p$) cases. In
addition, nekRS \cite{fischer_nekrs_2021} uses a single element per thread
block, which limits the amount of work available for a streaming multiprocessor
for relatively small polynomial orders.
\subsection{Navier-Stokes Results}
We consider scalability of nekRS for the cases of Fig. \ref{fig:ns_cases}.
All simulations except one use GMRES(15) with an initial guess generated by
$A$-conjugate projection onto 10 prior solutions \cite{fischer_projection_1998}.
Due to memory constraints, the 1568-pebble case with $P=24$ uses GMRES(10) with only
5 solution-projection vectors.
For each case, two pMG schedules are considered:
$(7,5,3,1)$ and $(7,3,1)$ for $p=7$;
and
$(9,7,5,1)$ and $(9,5,1)$ for $p=9$.
Other parameters, such as the Chebyshev order and the number of
coarse grid BoomerAMG V-cycles are also varied.
Results are shown in Figs. \ref{fig:scaling-study-bsb},\ref{fig:scaling-study-pb}.
The plots relate the effective work rate per node, measured as the
gridpoints $n$ (as shown in Table \ref{table:problem-sizes}) solved per second per node, to the time-to-solution.
The y-axis notes the drop in the relative work rate, which corresponds to a lower parallel efficiency, as the strong scale limit is reached,
while the x-axis denotes the time-to-solution.
Each node consists of 6 GPUs, hence $P=6 \times \text{nodes}$.
In all the performance tests conducted, the pMG preconditioner with
Chebyshev-Jacobi smoothing is outperformed by the other preconditioners,
whether using one or two V-cycle iterations in the AMG coarse-grid solve. For
each case, the fastest preconditioner scheme varies.
In the 146 pebble case (Fig. \ref{fig:scaling-study-pb}a),
using Cheby-RAS(2),(7,5,3,1) yields the smallest time per pressure solve.
However, in the 1568 pebble case (Fig. \ref{fig:scaling-study-pb}b), SEMFEM is a moderate improvement over the
second best preconditioner, Cheby-ASM(2),(7,5,3,1).
pMG with a
$(9,5,1)$ schedule and Chebyshev-RAS (of any order) yield the best scalability
and lowest time per pressure solve for the Boeing speed bump case (Fig. \ref{fig:scaling-study-bsb}).
The Chebyshev-accelerated Schwarz schemes are not always the fastest, however.
For the 67 pebble case (Fig. \ref{fig:scaling-study-pb}c), Cheby-Jac(2),(7,5,3,1) is comparable to Cheby-RAS(2),(7,5,3,1)
and are the two fastest pMG based preconditioners.
However, SEMFEM is significantly faster than the other preconditioners for this case.
Also considered is a hyrbid two-level pMG/SEMFEM approach wherein SEMFEM is used as the solver for the coarse level.
For $p=7$, a $(7,6)$ schedule with 2nd order Chebyshev-accelerated ASM smoothing on the $p=7$ level
and SEMFEM solver on the $p=6$ level, denoted as Cheby-ASM(2),(7,6) + SEMFEM, is used.
Similarly, Cheb-ASM(2),(7,5) + SEMFEM and Cheb-ASM(2),(7,3) + SEMFEM are considered.
For $p=9$, a $(9,8)$, $(9,7)$, $(9,5)$, and $(9,3)$ hybrid pMG/SEMFEM approach is considered,
denoted as Cheb-ASM(2),(9,8) + SEMFEM, Cheb-ASM(2),(9,7) + SEMFEM, Cheb-ASM(2),(9,5) + SEMFEM, and Cheb-ASM(2),(9,3) + SEMFEM,
respectively.
In the pebble cases shown in Fig. \ref{fig:scaling-study-pb}a,c, this hybrid approach
performs somewhere between the SEMFEM and Cheby-ASM(2),(7,3,1) preconditioners.
In the Boeing speed bump case, however, Fig. \ref{fig:scaling-study-bsb} demonstrates
that this approach is not as performant as either the SEMFEM or Cheby-ASM(2),(9,5,1) preconditioners.
Solver performance degrades whenever $n/P$ is sufficiently small,
regardless of the solver considered. For the various preconditioners
considered, at 2 nodes ($n/P=1.75M$), the strong-scale limit of 80\%
efficiency is far surpassed in the 146 pebble case. This leaves the effective
strong scale limit at 1-2 nodes ($n/P=3.5$ to $1.75M$).
For the 1568 pebble case, 12 nodes ($n/P=2.5M$) yields a parallel
efficiency around 60-70\%, depending on the specific solver.
The 67 pebble cases reaches 70\% effieciency on 3 nodes ($n/P = 2.3M$).
The
parallel efficiency for the Boeing speed bump case for the fastest
time-to-solution preconditioners drops below 70\% when using
more than 48 nodes ($n/P=2.24M$).
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{figs/bsb-scaling.png}
\vspace{-0.35cm}
\caption{
\small
Strong scaling results on Summit for the Navier-Stokes case of Fig. \ref{fig:ns_cases}d.
}
\label{fig:scaling-study-bsb}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=\textwidth]{figs/pb-scaling.png}
\vspace{-0.35cm}
\caption{
\small
Strong scaling results on Summit for the Navier-Stokes cases of Fig. \ref{fig:ns_cases}a,b,c.
Iso-processor count line illustrated in (c).
A user running on a specified number of processors should use the lowest time-to-solution preconditioner
along this line.
}
\label{fig:scaling-study-pb}
\end{figure*}
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{figs/pb67-zoom-scaling.png}
\vspace{-0.35cm}
\caption{
\small
Same as Fig. \ref{fig:scaling-study-pb}c, pMG preconditioners only.
}
\label{fig:scaling-study-pb-67-zoom}
\end{figure}
While the effect of mesh quality metrics, such as the max aspect ratio and scaled Jacobian, on solver convergence has been studied by Mittal {\em et al.}
\cite{mittal2019mesh}, predicting the optimal preconditioner settings from
mesh quality metrics is not obvious.
While the maximum aspect ratio is an important metric for mesh quality, it alone cannot
explain the apparent poor performance of the pMG preconditioners in the 67 pebble case, Fig. \ref{fig:scaling-study-pb}c.
Consider, for example, that the Boeing speed bump case has a larger maximum aspect ratio (255)
than the 67 pebble case (204), but does not exhibit this poor performance in the pMG preconditioners.
The minimum scaled Jacobian, however, is at least an order of magnitude smaller in the 67 pebble case ($5.97\times 10 ^{-3}$) as compared
to the other cases (e.g., .996 for the Boeing speed bump case, $4.31\times 10^{-2}$ and $2.59\times 10^{-2}$ for the 146, 1568 pebble cases, respectively).
This may, in turn, explain why the pMG preconditioners in the 67 pebble case
were signficiantly suboptimal compared to the SEMFEM preconditioner.
This demonstrates, however, that a user cannot simply rely on, e.g., the maximum aspect ratio
when deciding whether or not to use SEMFEM as a preconditioner.
This inability to correctly identify preconditioner settings based on mesh quality metrics
alone motivates the introduction of an auto-tuner to choose preconditioner parameters
during runtime.
\input{tables/mesh-metrics}
|
\section{Introduction}
Word forms are ambiguous, and derive meaning from the context in which they appear.
For example, the form ``bass'' can refer to a musical instrument, a low-frequency sound, a type of voice, or a kind of fish. The correct reference is determined by the surrounding linguistic context.
Traditionally, this kind of ambiguity was dealt via \emph{word sense disambiguation} (WSD),
a task that disambiguates word forms in context between symbolic sense-ids from a sense inventory such as WordNet \cite{miller1995wordnet} or, more recently, BabelNet \cite{navigli-ponzetto-2010-babelnet}. Such sense inventories rely heavily on manual curation, are labor intensive to produce, are not available in specialized domains and inherently unsuitable for words with emerging senses.\footnote{For example, in current WordNet~version, \textit{Corona} has 6 synsets, none of them relates to the novel \textit{Coronavirus}.}
This can be remedied by \emph{word sense induction} (WSI), a task where the input is a given word-type and a corpus, and the output is a derived sense inventory for that word. Then, sense disambiguation can be performed over the WSI-derived senses.
The introduction of large-scale pre-trained LMs and Masked LMs (MLM) seemingly made WSI/WSD tasks obsolete: instead of representing tokens with symbols that encode sense information, each token is associated with a contextualized vector embeddings that captures various aspects of its in-context semantics, including the word-sense. These contextualized vectors proved to be very effective as features for downstream NLP tasks. However, contextualized embeddings also have some major shortcomings: most notably for our case, they are expensive to store (\textit{e.g.} BERT embeddings are 768 or 1024 floating point numbers for each token), and are hard to index and query at scale. Even if we do manage to store and query them, they are not interpretable, making it impossible for a user to query for a particular sense of a word without providing a full disambiguating context for that word. For example, consider a user wishing to query a dataset for sentences discussing \emph{Oracle} in the mythology-prophet sense, rather than the tech company sense. It is not clear how to formulate such a query to an index of contextualized word vectors. However, it is trivial to do for an index that annotates each token with its derived sense-id (in terms of UI, after a user issues a query such as ``Oracle'', the system may show a prompt such as ``did you mean Oracle related to IBM; Sun; Microsoft, or to Prophet; Temple; Queen'', allowing to narrow the search in the right direction).
\citet{amrami2018word, amrami2019towards} show how contextualized embeddings can be used for achieving state-of-the-art WSI results. The core idea of their WSI algorithm is based on the intuition, first proposed by \citet{bacskaya2013ai}, that occurrences of a word that share a sense, also share in-context substitutes. An MLM is then used to derive top-$k$ word substitutes for each word, and these \emph{substitute-vectors} are clustered to derive word senses.
Our main contribution in this work is proposing a method that scales up \citet{amrami2018word}'s work to \emph{efficiently} annotate all tokens in a large corpus (e.g. Wikipedia) with automatically derived word-senses. This combines the high-accuracy of the MLM-based approach, with the symbolic representation provided by discrete sense annotations. The discrete annotations are interpretable (each sense is represented as a set of words), editable, indexable and searchable using standard IR techniques.
We show two applications of the discrete annotations, the first one is sense-aware information retrieval (\S\ref{section:senseful_ir}), and the second is high-quality senseful \emph{static} word embeddings we can derive by training a static embeddings model on the large sense annotated corpus (\S\ref{section:senseful_w2v}).
We first show how the method proposed by \citet{amrami2018word} can be adapted from deriving senses of individual lemmas to efficiently and cheaply annotating \emph{all the corpus occurrences} of \emph{all the words in a large vocabulary} (\S\ref{section:wsi_at_scale}). Deriving word-sense clusters for all of English Wikipedia words that appear as single-token words in \textsc{BERTlarge}'s \cite{devlin-etal-2019-bert} vocabulary, and assigning a sense to each occurrence in the corpus, required 100 hours of cheap P100 GPUs (5 hours of wall-clock time on 20 single GPU machines) followed by roughly 4 hours on a single 96-cores CPU machines. The whole process requires less than 50GB of disk space, and costs less than 150\$ on Google Cloud platform.
After describing the clustering algorithm (\S\ref{subsection:community_detection}), we evaluate the quality of our system and of the automatic sense tagging using SemEval datasets and a new manually annotated dataset we created (\S\ref{section:evaluation}). We show that with the produced annotated corpora it is easy to serve sense-aware information retrieval applications (\S\ref{section:senseful_ir}).
Another immediate application is feeding the sense-annotated corpora to a static embedding algorithm such as word2vec \cite{mikolov2013distributed}, for deriving \emph{sense-aware static embeddings} (\S\ref{section:senseful_w2v}). This results in state-of-the-art sense-aware embeddings, which we evaluate both on an existing WiC benchmark \cite{pilehvar-camacho-collados-2019-wic} and on a new challenging benchmark which we create (\S\ref{section:embs_eval}).
In contrast to WSD which relies on curated sense inventories, our method is data-driven, therefore resulting senses are corpus dependent. The method can be applied to any domain for which a BERT-like model is available, as we demonstrate by applying it to the PubMed Abstracts of scientific papers, using \textsc{SciBERT} \cite{Beltagy2019SciBERT}. The resulting senses cover scientific terms which are not typically found in standard sense inventories (\S\ref{section:biomedical_data}).
Figure \ref{fig:examples} shows examples of induced senses for selected words from the English Wikipedia corpus. For each sense we list 5 community-based representatives (\S\ref{section:wsi_at_scale}), as well as the 5 closest neighbours in the sense-aware embedding space (\S\ref{section:senseful_w2v}). Additional examples are available in Appendix \ref{appendix:more_examples}. Code and resources are available in \href{https://github.com/allenai/WSIatScale}{github.com/allenai/WSIatScale}.
\section{Related Work}
\label{section:related_work}
\paragraph{Word Sense Induction and Disambiguation}
Previous challenges like \citet{jurgens2013semeval} focused on word sense induction for small sized datasets. To the best of our knowledge we are the first to perform large-scale \textit{all-words} WSI. The closest work to our method is the substitution-based method proposed in \citet{amrami2018word, amrami2019towards} which is the starting point to our paper. In that paper, the authors suggested a WSI algorithm designed for a small dataset (SemEval 2010, 2013) with a predefined set of ambiguous target words (See (\S\ref{section:wsi_at_scale}) for more details on the algorithm). In our work, we change \citet{amrami2019towards} such that we can efficiently run sense induction on all the words in very large corpora.
An alternative approach for sense tagging is based on Word Sense Disambiguation (WSD). The two main WSD methods are Supervised WSD and Knowledge-based WSD. Supervised WSD suffers from the difficulty of
obtaining an adequate amount of annotated data. Indeed, even SemCor, the largest manually annotated tagged corpus, consists of only 226,036 annotated tokens.
Among different supervisied WSD methods, \citet{zhong-ng-2010-makes} suggested a SVM based approach and \citet{melamud-etal-2016-context2vec, yuan2016semi} suggested LSTMs paired with nearest neighbours classification.
Knowledge-base WSD \cite{moro2014entity, pasini2017train}, on the other hand, avoids the reliance on large annotated word-to-sense corpus and instead maps words to senses from a closed sense inventory (\textit{e.g.} WordNet~\cite{miller1995wordnet}, BabelNet~\cite{navigli-ponzetto-2010-babelnet}). As such, the quality of knowledge-based WSD heavily depends on the availability, quality and coverage of the associated annotated resources.
\paragraph{Sense Embeddings}
In \S\ref{section:senseful_w2v} we exploit the sense-induced corpus to train sense embeddings. \citet{reisinger2010multi} were the first to suggest creating multiple representations for ambiguous words. Numerous recent papers \cite{chen2014unified, rothe2015autoextend, iacobacci-etal-2015-sensembed, pilehvar2016conflated, mancini-etal-2017-embedding, iacobacci2019lstmembed} aim to produce similar embeddings, all of which use either WordNet~or BabelNet~as semantic network. Our method is similar to \citet{iacobacci-etal-2015-sensembed}, with the difference being that they rely on semantic networks (via Babelfy \cite{moro2014entity}). In contrast and similarly to us, \citet{pelevina2017making} does not rely on lexical resources such as WordNet.
The authors proposed splitting pretrained embeddings (such as word2vec) to a number of prototype sense-embeddings. Yet in our work, we directly learn the multi-prototype sense-embeddings which is only possible due to the large-scale corpus annotation. When comparing both methods in \S\ref{section:wic} we infer it is better to directly learn multi-prototype sense-embeddings.
\input{tables_and_figures/flow}
\section{Large Scale Sense Induction}
\label{section:wsi_at_scale}
\subsection{Definition}
We define large-scale sense induction as deriving sense clusters for all words in a large vocabulary and assigning a sense cluster to each corpus occurrence of these words.\footnote{
In \textit{BERT-large-cased-whole-word-masking} this corresponds to 16k vocabulary items, that match to 1.59B full words in English Wikipedia, or 92\% of all word occurrences. Analyzing the remaining words, only 0.01\% appear in Wikipedia more than 100 times.
We derive word senses to a substantial chunk of the vocabulary, which also corresponds to the most ambiguous words as less frequent words are substantially less polysemous \cite{hernandez2016testing, fenk2010frequency, zipf1945meaning}.
}
\subsection{Algorithm}
Contextualized BERT vectors contain sense information, and clustering the contextualized vectors results in sense clusters. However, storing a 1024 dimensional vector of 32bit floats for each relevant token in the English Wikipedia corpus requires over 8TB of disk-space, making the approach cumbersome and not-scalable. However, as shown by \citet{amrami2019towards}, MLM based word-substitutes also contain the relevant semantic information, and are much cheaper to store: each word-id in \textsc{BERTlarge}'s vocabulary can be represented by 2 bytes, and storing the top-5 substitutes for each corpus position requires less than 20GB of storage space.\footnote{The size can be reduced further using adaptive encoding techniques that assign fewer bits to frequent words. We did not implement this in this work.}
In order to perform WSI at scale, we keep the main intuition from \citet{amrami2019towards}, namely to cluster sparse vectors of lemmas of the top-k MLM-derived word substitutions. This results in vast storage saving, and also in a more interpretable representations. However, for scalability, we iterate over the corpus sentences and collect the top-k substitutes for all words in the sentence at once based on a single BERT call for that sentence. This precludes us from using the dynamic-patterns component of their method, which requires separately running BERT for each word in each sentence. However, as we show in Section \S\ref{subsection:semeval} we still obtain sufficiently high WSI results.
The steps for performing Scalable WSI are summarized in Fig. \ref{fig:flow}. We elaborate on each step below, using English Wikipedia as a running example.\footnote{The Wikipedia corpus is based on a dump from August 2020, with text extracted using WikiExtractor \cite{Wikiextractor2015}.}
\textbf{Annotation:} We run \textit{BERT-large-cased-whole-word-masking} on English Wikipedia, inferring substitutes for all corpus positions. For positions that correspond to single-token words,\footnote{We exclude single-character tokens, stopwords and punctuation.} we consider the predicted words, filter stop-words, lemmatize the remaining words \cite{spacy}, and store the top-5 most probable lemmas to disk. This step takes 5 hours on 20 cloud-based GPU machines (total of 100 GPU hours), resulting in 1.63B tokens with their corresponding top-5 lemmas.
\textbf{Inverted Word Index:} We create an inverted index mapping from each single-token word to its corpus occurrences (and their corresponding top-5 lemmas). This takes 5 minutes on a 96 cores CPU machine, and 10GB of disk.
\textbf{Sense Induction:}
For each of 16,081 lemmas corresponding to single-token words, we retrieve random 1000 instances,\footnote{The clustering algorithm scales super-linearly with the number of instances. To reduce computation cost for tokens that appear more than 1000 times in the dataset, we sample $\min($numOccur,$1000)$ instances for each token word, and cluster given the subset of instances. We then associate each of the remaining instances to one of the clusters as explained in the final step of the algorithm.} and induce senses using the community-based algorithm described in \S\ref{subsection:community_detection}. This process requires 30 minutes on the 96-core CPU machine, and uses 100MB of disk space. The average number of senses per lemma is 3.13. Each sense is associated with up to 100 representative words, which represent the highest-degree words in the sense's community.
Table \ref{table:bass_communities} shows the 5 senses found for the word \emph{bass} with their top-5 representative words. See additional examples in Fig. \ref{fig:examples} and Appendix \ref{appendix:more_examples}.
\input{tables_and_figures/communities_and_close_vecs/bass_communities}
\textbf{Tagging:}
Each of the remaining word-occurrences is associated with a sense cluster by computing the Jaccard similarity between the occurrences' top-5 lemmas and the cluster representatives, and choosing the cluster that maximizes this score. For example, an occurrence of the word \emph{bass} with lemmas \emph{tenor, baritone, lead, opera, soprano} will be associated with bass$_3$.
This takes 100 minutes on 96-core machine, and 25GB of storage.
\section{Sense Clustering Algorithm}
\label{subsection:community_detection}
We replace the hierarchical clustering algorithm used by \citet{amrami2018word,amrami2019towards} with a community-detection, graph-based clustering algorithm.
One major benefit of the community detection algorithms is that they naturally produces a dynamic number of clusters, and provide a list of interpretable discrete representative lemmas for each cluster.
We additionally found this method to be more stable.
Graph-based clustering for word-sense induction
typically constructs a graph from word occurrences or collocations, where the goal is to identify sense-specific sub-graphs within the graph that best induce different senses \cite{klapaftis2008word, klapaftis2010word}. We instead construct the graph based on word substitutes. Following \citet{jurgens2011word}, we pose identifying sense-specific clusters as a \emph{community detection problem}, where a community is defined as a group of connected nodes that are more connected to each other than to the rest of the graph.
\paragraph{Graph construction}
For each word $w$ in the vocabulary, we construct a graph $G_w=(V_w,E_w)$ where each vertex $v\in V_w$ is a substitute-word predicted by the MLM for $w$, and an edge $(u, v) \in E_w$ connects substitutes that are predicted for the same instance. The edge is weighted by the number of instances in which both $u$ and $v$ were predicted.
More formally, let $X = \{x_w^i\}_{i=1}^n$ bet the set of all top-$k$ substitutes for $n$ instances of word $w$, and $x_w^i = \{{w'}^j_{x_w^i}\}_{j=1}^k$ represents the $k$ top substitutes for the $i$th instance of word $w$.
The graph $G_w$ is defined as follows:
\begin{equation*}
\begin{split}
V_w &=\{u: \exists i~u \in x_w^i \} \\
E_w &=\{(u, v) : \exists i~u \in x_w^i \wedge v \in x_w^i\} \\
W(u, v) &= |\{ i : (u, v) \in x_w^i\}|
\end{split}
\end{equation*}
\noindent
\paragraph{Community detection} A community in a sub-graph corresponds to a set of tokens that tend to co-occur in top-$k$ substitutes of many instances, and not co-occur with top-$k$ substitutes of other instances. This corresponds well to senses and we take community's nodes as sense's representatives.
We identify communities using the fast \textit{``Louvain''} method \cite{blondel2008fast}. Briefly, Louvain searches for an assignment of nodes to clusters such that the \emph{modularity score} $Q$---which measures the density of edges inside communities compared to edges between communities---is maximized:
\begin{equation*}
Q = \frac{1}{2m}\sum_{u~v} \left[W(u, v)- \frac{k_u k_v}{2m} \right]\delta(c_u, c_v) \\
\end{equation*}
\noindent $m$ is the sum of all edge weights in the graph, $k_u=\sum_v W(u,v)$ is the sum of the weights of the edges attached to node $u$, $c_u$ is the community to which $u$ is assigned, and $\delta$ is Kronecker delta function. This objective is optimized using an iterative heuristic process.
For details, see \citet{blondel2008fast}.
\section{Intrinsic Evaluation of Clustering Algorithm}
\label{section:evaluation}
We start by \emph{intrinsically evaluating} the WSI clustering method on: (a) \semeval2010 and \semeval2013; and (b) a new test set we develop for large-scale WSI. In section \ref{section:embs_eval}, we additionally \textit{extrinsically evaluate} the accuracy of static embeddings derived from a sense-induced Wikipedia dataset.
When collecting word-substitutes, we lemmatize the top-k list, join equivalent lemmas, remove stopwords and the target word from the list, and keep the top-5 remaining lemmas.
\subsection{SemEval Evaluation}
\label{subsection:semeval}
We evaluate the community-based WSI algorithm on two WSI datasets:
SemEval 2010 Task 14 \cite{manandhar2010semeval} and SemEval 2013 Task 13 \cite{jurgens2013semeval}. Table \ref{table:semeval} compares our method to \citet{amrami2018word, amrami2019towards} and AutoSense \cite{amplayo2019autosense}, which is the second-best available WSI method.
Bert-noDP/DP are taken from \citet{amrami2019towards}. Bert-DP uses ``dynamic patterns'' which precludes wide-scale application. We follow previous work
\cite{manandhar2010semeval, komninos2016structured, amrami2019towards} and evaluate SemEval 2010 using F-Score and V-Measure and SemEval 2013 using Fuzzy Normalized Mutual Information (FNMI) and Fuzzy
B-Cubed (FBC) as well as their geometric mean (AVG).
Our method performs best on SemEval 2010 and comparable to state-of-the-art results on SemEval 2013. The algorithm performs on-par with the Bert-noDP method, and does not fall far behind the Bert-DP method.
We now turn to assess the end-to-end induction and tagging over Wikipedia.
\subsection{Large Scale Manual Evaluation}
\label{section:tagging_evaluation}
We evaluate our method on large corpora by randomly sampling 2000 instances from the sense-induced Wikipedia, focusing on frequent words with many senses. We manually annotate the samples' senses without access to the automatically induced senses, and then compare our annotations to the system's sense assignments. We publicly release our manual sense annotations.
\input{tables_and_figures/semeval_table}
\paragraph{Sampling and Manual Annotation}
We used a list of 20 ambiguous words from \textit{CoarseWSD-20} \cite{loureiro2021analysis}. The full list and per-word results can be found in Appendix \ref{appendix:tagging_evaluation}. For each word we sampled 100 passages from English Wikipedia with the target word, including inflected forms (case insensitive).
Unlike \textit{CoarseWSD-20}, we sampled examples without any respect to a predefined set of senses. For example, the only two senses that appear in \textit{CoarseWSD-20} for the target word \textit{arm} are \textit{arm (anatomy)}, and \textit{arm (computing)}, leaving out instances matching senses reflecting \textit{weapons, subdivisions, mechanical arms} etc.
With the notion that word sense induction systems should be robust to different annotations schemes, we gave two fluent English speakers 100 sentences for each of the 20 ambiguous words from \textit{CoarseWSD-20}. Annotators were not given a sense inventory. Each annotator was asked to label each instance with the matching sense \textit{according to their judgment}. For example, for the target word \textit{apple} in the sentence \textit{``The iPhone was announced by \textit{Apple} CEO."}, annotators can label the target sense with \textit{Apple Inc.}, \textit{Apple The Company} etc. Annotation Guidelines are available in Appendix \ref{appendix:annotation_guidelines}.
On average annotators labeled $6.65$ senses per word ($5.85$ and $7.45$ average clusters per word for the two annotators). This is more than the $2.65$ average senses according to \textit{CoarseWSD-20} and less than WordNet's $9.85$.
\paragraph{Results}
We report our system's performance alongside two additional methods: A strong baseline of the most frequent sense (MFS), and Babelfy \cite{moro2014entity}---the sense disambiguation system used in BabelNet~(Tested using Babelfy live version April 2021). Differently from the latter, our system does not disambiguates but induces senses, therefore, clusters are not labeled with a sense tag from a sense inventory. Instead, we represent senses to annotators using a list of common substitute words and a few examples. Thus, after annotating the Wikipedia passages, we additionally asked annotators to name the system's clusters with the same naming convention as in their annotations.
Given a similar naming convention between systems and annotators, we report F1 scores of systems' tagging accuracy with respect to the manual annotations. We report F1 averaged over words in Table \ref{table:tagging_evaluation}. Our system outperforms both baselines, despite Babelfy having access to a list of predefined word senses. A full by-word table and comprehensive results analysis are in Appendix \ref{appendix:tagging_evaluation}.
While a 1-to-1 mapping between system clusters and manual senses is optimal, our system sometimes splits senses into smaller clusters, thus annotators will name two system clusters with the same label. Therefore it is also important to report the number of clusters produced by the system comparing to the number of senses after the annotators merged similar clusters. Our system produced $7.25$ clusters with $2.25$ clusters on average merged by the annotators.\footnote{This is partially due to using clusters from two casing (\textit{e.g.} \textit{bank} and \textit{Bank}), some of the merges share sense meaning but of different casing.}
Additionally, in rare cases our system encapsulates a few senses in a single cluster: this happened 3 and 5 times for both annotators across all the dataset.
\input{tables_and_figures/tagging_evaluation}
\section{Application to Scientific Corpora}
\label{section:biomedical_data}
A benefit of a WSI approach compared to WSD methods is that it does not rely on a pre-specified sense inventory, and can be applied to any corpus for which a BERT-like model is available.
Thus, in addition to the Wikipedia dataset that has been presented throughout the paper, we also automatically induce senses over a corpus of 31 million PubMed Abstracts,\footnote{\href{https://www.nlm.nih.gov/databases/download/pubmed\_medline.html}{www.nlm.nih.gov/databases/download/pubmed\_medline}} using SciBERT \cite{Beltagy2019SciBERT}. As this dataset is larger than the Wikipedia dump, the process required roughly 145 GPU hours and resulting in $14,225$ sense-annotated lemmas, with an average number of $2.89$ senses per lemma.
This dataset highlights the data-driven advantages of sense-induction: the algorithm recovers many senses that are science specific and are not represented in the Wikipedia corpora. While performing a wide-scale evaluation of the scientific WSI is beyond our scope in this work, we do show a few examples to qualitatively demonstrate the kinds of induced senses we get for scientific texts.
For each of the words \emph{mosaic, race} and \emph{swine} we show the induced clusters and the top-5 cluster representatives for each cluster.
\input{tables_and_figures/communities_and_close_vecs/mosaic-pubmed}
While senses mosaic$_{0}$ (the common mosaic virus of plants) and mosaic$_{2}$ (``something resembling a mosaic", ``mosaic of..") are represented in Wikipedia, senses mosaic$_1$ (the mosaic genetic disorder) and mosaic$_3$ (mosaic is a quality, e.g., ``mosaic border'', ``mosaic pattern'') are specific to the scientific corpora (The Wikipedia corpora, on the other hand, includes a sense of mosaic as a decorative art-form, which is not represented in Pubmed).
\input{tables_and_figures/communities_and_close_vecs/race}
\noindent Senses race$_{0}$ (ethnic group), race$_{1}$ (competition) and race$_{2}$ (population/civilization) are shared with wikipedia, while the sense race$_3$ (``Rapid amplification of cDNA ends'', a technique for obtaining the sequence length of an RNA transcript using reverse transcription (RT) and PCR) is Pubmed-specific.
\input{tables_and_figures/communities_and_close_vecs/swine}
\noindent Here swine$_{1}$ captures the Swine Influenza pandemic, while swine$_{2}$ refers to swine as experimental Pigs.
\section{Sense-aware Information Retrieval}
\label{section:senseful_ir}
An immediate application of a high quality sense-tagged corpus is sense-aware retrieval.
We incorporate the sense information in the SPIKE extractive search system \cite{Shlain2020SyntacticSB}\footnote{\href{https://spike.apps.allenai.org}{spike.apps.allenai.org}} for Wikipedia and Pubmed datasets.
When entering a search term, suffixing it with @ triggers sense selection allowing to narrow the search for the specific sense.
Consider a scientist looking for PubMed occurrences of the word \textit{``swine"} in its influenza meaning. As shown in Figure \ref{fig:swine_in_spike}, this can be easily done by writing ``swine@'' and choosing the second item in the resulting popup window. The outputs are sentences with the word \textit{``swine"} in the matching sense. As far as we know, SPIKE is the first system with such WSI capabilities for IR. Similarly, \citet{blloshmi2021ir} suggested to enhance IR with sense information, but differently from us, this is done by automatically tagging words with senses from a predefined inventory.
\input{tables_and_figures/swine_in_spike}
\section{Sense-aware Static Embeddings}
\label{section:senseful_w2v}
Learning static word embeddings of sense-ambiguous words is a long standing research goal \cite{reisinger2010multi, huang2012improving}.
There are numerous real-world tasks where context is not available, precluding the use of contextualized-embeddings. These include Outlier Detection \cite{camacho2016find, blair2016automated}, Term Set Expansion \cite{roark2000noun} the Hypernymy task \cite{breit-etal-2021-wic}, etc.
Additionally, static embeddings are substantially more efficient to use, can accommodate larger vocabulary sizes, and can accommodate efficient indexing and retrieval.
Yet, despite their flexibility and success, common word embedding methods still represent ambiguous words as a single vector, and suffer from the inability to distinguish between different meanings of a word \cite{camacho2018word}.
Using our sense-tagged corpus we suggest a simple and effective method for deriving sense-aware static embeddings: We run an off-the-shelf embedding algorithm,\footnote{We use the CBOW variant of the word2vec algorithm \cite{mikolov2013distributed} as implemented in Gensim \cite{rehurek_lrec}. We derive 100-dimensional embeddings using the negative-sampling algorithm and a window size of 5.}
on the corpus where single-token words are replaced with a concatenation of the word and its induced sense (\textit{e.g.} \textit{``I caught a bass."} becomes \textit{``I caught@0 a bass@2."}). This makes the embedding algorithm learn embeddings for all senses of each word out-of-the-box.\footnote{A similar approach was used by \citet{iacobacci-etal-2015-sensembed} over a corpus which was labeled with BabelNet and WordNet senses.}
An integral property of the embedding algorithm is that it represents both the sense-annotated tokens and the other vocabulary items in the same embedding space --- this helps sense inferring about words that are represented in the MLM as multi-tokens words (Even though these correspond to less-frequent and often less ambiguous words \cite{hernandez2016testing, fenk2010frequency, zipf1945meaning}). For example, in the top-5 nearest neighbours for the different \emph{bass} senses as shown below, \textit{smallmouth} and \textit{pumpkinseed}, multi-token words in \textsc{BERTlarge}'s vocabulary, are close neighbours the \textit{bass} instances that correspond to the \textit{fish} sense.
\input{tables_and_figures/communities_and_close_vecs/bass_embs}
Note that some neighbours are sense annotated (single-token words that were tagged by our system), while others are not (multi-token words).
For English Wikipedia, we obtain a total vocabulary of 1.4M forms, $90,023$ of which are sense-annotated. Compared to the community-based representative words, the top neighbours in the embedding space tend to capture members of the same semantic class rather than direct potential replacements.
\section{Sense-aware Embeddings Evaluation}
\label{section:embs_eval}
\subsection{WiC Evaluation}
\label{section:wic}
\citet{pilehvar-camacho-collados-2019-wic} introduced the WiC~dataset for the task of classifying word meaning in context. Each instance in WiC~has a target word and two contexts in which it appears. The goal is to classify whether the word in the different contexts share the same meaning. \textit{e.g.} given two contexts: \textit{There's a lot of trash on the \underline{bed} of the river} and \textit{I keep a glass of water next to my \underline{bed} when I sleep}, our method should return \textit{False} as the sense of the target word \textit{bed} is different.
Our method is the following: Given the sense-aware embeddings, a target word $w$ and two contexts, we calculate the context vector as the average of the context words. The matching sense vector is the closest out of all $w$ embeddings. We then classify the contexts as corresponding to the same meaning if the cosine distance of the found sense embedding is more than threshold apart. We do not use the train set. The threshold is optimized over the development set and fixed to $0.68$.
\input{tables_and_figures/wic}
This task has a few tracks, we compare our embeddings systems to the best performing methods from the \textit{Sense Representations} track. Of these, JBT \cite{pelevina2017making}, a lexical embedding method, is the only one that does not use an external lexical resource (induction). The results in Table \ref{table:wic_results} show accuracy on this task. We outperform the induction method, and are on-par with the lexicon-based methods, despite not using any external lexical resource.
\input{tables_and_figures/outlier_results}
\subsection{Evaluation via Outlier Detection}
\label{subsection:outlier_detection}
Another setup for evaluating word embeddings is that of \emph{outlier detection}: given a set of words, identify which one does not belong to the set \cite{blair2016automated}. Outlier detection instances are composed of in-group elements and a set of outliers from a related semantic space. In each evaluation round, one outlier is added to the in-group items, and the algorithm is tasked with finding the outlier.
Existing outlier detection datasets either did not explicitly target sense-ambiguous words (\textit{8-8-8} \cite{camacho2016find}, \textit{WikiSem500} \cite{blair2016automated}) or explicitly \emph{removed} ambiguous words altogether (\textit{25-8-8-sem} \cite{andersen2020one}).
\paragraph{Ambiguity-driven Outlier Detection.} We construct a challenge set for outlier detection that specifically targets ambiguous cases.
In order to account for sense ambiguity, we add a \emph{distractor} to each of the in-group sets: the distractor is an item which has multiple senses, where the most salient sense does not belong to the group, while another sense does belong to the group. For example:
\textbf{In-group:} \textit{zeus, hades, poseidon, aphrodite, ares, athena, artemis}\\
\textbf{Outliers:} \textit{mercury, odysseus, jesus, sparta, delphi, rome, wrath, atlanta}\\
\textbf{Distractor:} \textit{nike}
Here, a model which does not explicitly represent the greek-god sense of \emph{nike} is likely to place it far away from the in-group instances, causing it to be mistakenly marked as the outlier.
The starting point for our dataset is \textit{25-8-8-Sem} \cite{andersen2020one}. This dataset contains 25 test groups, each with 8 in-group elements and 8 outliers, resulting in 200 unique test cases. The outliers are sorted in a decreasing degree of relatedness to the in-group elements. In our dataset we replace one of the in-group elements with an ambiguous distractor. For example, in the Greek-gods case above, we replaced the original 8\textsuperscript{th} item (\textit{``hera"}) with the ambiguous distractor \emph{nike}.
\footnote{We additionally changed terms that are debatably ambiguous and changed the \textit{``African animals"} group to the more general \textit{``animals"} as no distractors were found.}
The dataset consists of 25 groups of 7 non ambiguous group elements, 1 distractor and 8 outliers (\textit{25-7-1-8}), similarly resulting 200 unique test cases.
\noindent\textbf{Method}$\,\,\,$ Following \citet{camacho2016find}, we rank each word likelihood of being the outlier by the average of all pair-wise semantic similarities of the words in $W\backslash \{w\}$. Therefore if $w$ is an outlier, this score should be low. See Appendix \ref{appendix:outlier_detection_method} for additional details.
\noindent\textbf{Metrics}$\,\,\,$
\citet{camacho2016find} proposed evaluating outlier detection using the accuracy (The fraction of correctly classified outliers among the total cases) and Outlier Position Percentage (OPP) metric. OPP indicates how close outliers are to being classified correctly:
\begin{equation*}
OPP = \frac{\sum_{W \in D} \frac{OP(W)}{\vert W \vert -1}}{\vert D \vert} \times 100
\end{equation*}
\noindent where $OP(W)$ is the position of the outlier according to the algorithm.
\noindent\textbf{Results}$\,\,\,$
In Table \ref{table:outlier_results} we report performance of on the \textit{25-7-1-8}~set. Word2vec and GloVe~accuracy scores are low while having high OPP scores. This is the expected behaviour for embeddings without sense awareness. These will position the distractor and the outlier furthest away from the group items while not designed to make the hard decision required for high Accuracy. Our sense-aware embeddings strongly outperform GloVe and word2vec which do not include senses.
Our embeddings also outperform the word embeddings proposed in DeConf \cite{pilehvar2016conflated}, which are the best performing sense embeddings on WiC which are also publicly available.
\section{Conclusion}
\label{section:conclusion}
We show that substitution-based word-sense induction algorithms based on word-substitutions derived from MLMs are easily scalable to large corpora and vocabulary sizes, allowing to efficiently obtain high-quality sense annotated corpora.
We demonstrate the utility of such large-scale sense annotation, both in the context of a scientific search application, and for deriving high-quality sense-aware static word embeddings.
As a secondary contribution, we also develop a new variant of the Outlier Detection evaluation task, which explicitly targets ambiguous words.
\section{Acknowledgments}
This project has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme, grant agreement No. $802774$ (iEXTRACT).
\section{Additional Examples}
\label{appendix:more_examples}
\input{tables_and_figures/communities_and_close_vecs/appendix_examples}
\section{Annotation Guidelines for Manual Evaluation}
\label{appendix:annotation_guidelines}
The objective of this task is to annotate word-meanings of 20 ambiguous words in a total of 2000 different contexts.
What is word-meaning? Words have different meanings in different contexts, for example, in the sentence: \textit{``there is a \underline{light} that never goes out"}, the word \textit{``light"} refers to any device serving as a source of illumination. While \textit{``light"} in the sentence \textit{``\underline{light} as a feather"} refers to the comparatively little physical weight or density of an object.
\textbf{Step 1:}
In this dataset we examine 20 ambiguous words as targets. For each of these words we collected 100 sentences in which the target word appears. For every sentence in the 100 set per target word, you will be asked to write a short label expressing the meaning of the target word in that particular context.
For example, here are three sentences with the target word \textit{``light"}, each with its possible annotation.
1. \textit{``there is a \underline{light} that never goes out"} $\rightarrow$ visible light.
2. \textit{``\underline{light} as a feather"} $\rightarrow$ light as in weight.
3. \textit{``magnesium is a \underline{light} metal"} $\rightarrow$ light as in weight.
Note that in this example the annotator found the second and third meanings of the word \textit{``light"} to be the same and therefore labeled them with the same label.\footnote{For ease of use for future evaluators, at the end of this step, both annotators picked a single naming convention when two labels referred to the same sense. Names of labels that were used only by one annotator were not changed.}
While some annotations are indeed intuitive, labeling word-meanings when the target word is part of a name can be challenging. Here are a few guidelines for such use case:
Whenever a target word appeared as part of a name (Person, Organization etc.), one of three heuristics should be used\footnote{Some of the dissimilarities between the annotations are with respect the tension between the second and third guidelines.}:
1. If the target word is the surname of a person, the example should be tagged \textit{surname}.\footnote{As opposed to Babelfy, there was no attempt for entity linking, so all persons were tagged the same.}
2. If the entity (as a whole) refers to one of the word-meanings, it should be labeled as such. For example, \textit{Quitobaquito Springs} label should refer to a natural source of water.
3. If the target word is part of a name different from the original word-meaning, it should be tagged as \textit{Part of Name}. This includes song names, companies (\textit{Cold Spring Ice}), restaurants etc. Possible exceptions for this case are when a specific named entity is significantly frequent.
\textbf{Step 2:} \footnote{This step is presented to annotators once step 1 is done for all words}
For each of the target words you labeled, you will now receive a short list of indirect word-meaning definitions. Indirect word-meanings are composed of:
(a) A list of 10 words that may appear instead of the target word in specific contexts
(b) A list of 5 sentences in which the target word has this specific word-meaning.
For example, this is a possible indirect word-meaning for the target word \textit{``Apple"}, representing the fruit, as opposed to the tech company:
\textbf{Alternatives:}
orange, olive, cherry, lime, banana, emerald, lemon, tomato, oak, arrow,
\textbf{Sentences in which Apple appears in this word-meaning:}
\textit{``He and his new bride planted \underline{apple} trees to celebrate their marriage."}
\textit{``While visiting, Luther offers Alice an \underline{apple}."}
\textit{``When she picks the \underline{apple} up, it is revealed that Luther has stolen a swipe card and given it to Alice to help her escape."}
You will be asked to label the indirect word-meanings with one of the labels you used in step 1. If no label matches the indirect word-meaning you are allowed to propose a new label or define it to be “Unknown". Additionally, if you find several indirect word-meanings too close in meaning, label them the same.
\section{Analysis of Manual Evaluation}
\label{appendix:tagging_evaluation}
\input{tables_and_figures/tagging_evaluation_by_word}
In table \ref{table:tagging_evaluation_by_word} we report a by-word analysis of our manual evaluation results. For each word we detail F1 scores of the most frequent sense (MFS), Babelfy, and our proposed system. Similarly to \citet{loureiro2021analysis}, we report the ratio of the first sense with respect to the rest (F2R) and normalized entropy\footnote{
Computed as
$\frac{-\sum_{i=1}^{k} \frac{c_i}{n} \log \frac{c_i}{n}}{log(k)}$, where $k$ is the number of annotated senses, each of size $c_i$ and $n$ is the size of annotated examples per word, in our case $n=100$.
}
to reflect sense balance. All of which are reported per annotator.
\paragraph{Analysis}
Analysis of our system's error shows that for some words the system could not create a matching cluster for specific senses (to name a few examples, "yard" as a ship identifier and "impound/enclosure" sense for the word "pound"). It appears that a matching cluster was not created due to the low tally of these senses in the English Wikipedia, and indeed the two senses appeared only two and three times respectively in the 100 passages sample. Additionally, annotator 2 annotated in a more fine-grained manner that does not correspond to our system tendency to merge capitalized instances of the target word into a sense that corresponds to "part of named entity".
As described above, in rare cases our system merged two senses into a single cluster. For example, the same cluster of the word "trunk" contained occurrences which annotator 1 tagged either "human torso" or "tube-like organs" (like the pulmonary trunk). While such annotation was uncommon (3 out of 117 senses for annotator 1 and 5 out of 149 senses for annotator 2), it does affect our system's micro F1 score for the better. In case we do not allow such annotation our overall score drops from $87.52$ to $86.65$.
A comparison between Babelfy and our gold annotation shows a common mistake in its labeling where Babelfy attributes the vast majority of sentences to the same non-salient sense. For example, Babelfy attributes 77 out of 100 instances of \textit{hood} to "An aggressive and violent young criminal" - a sense that was not found even once in the manual annotation. While in a number of cases Babelfy used finer-grained sysnset groups than in our annotations we took into account any senses that are a subset of our annotated senses. For examples, Babelfy's \textit{"United States writer who lived in Europe; strongly influenced the development of modern literature (1885-1972)"} synset was attribute any instances from the senses \textit{surname} that refer to the writer Ezra Pound.
\section{Outlier Detection Method}
\label{appendix:outlier_detection_method}
When using a single-prototype vector-space models, \citet{camacho2016find} proposed a procedure for detecting outliers based on semantic similarity using \textit{compactness score}:
\begin{equation*}
c(w) = \frac{1}{n^2 - n} \sum_{w_i \in W \setminus \{w\}} \sum_{\substack{w_j \in W \setminus \{w\}\\w_i \ne w_j}} sim(w_i, w_j)
\end{equation*}
Where $D$ is the entire dataset and $W$ is defined as $\{w_1, w_2, \cdots, w_n, w_{n+1}\}$ where w.l.o.g. $\{w_1, w_2, \cdots, w_n\}$ are the group elements (including the distractor) and $w_{n+1}$ is the outlier.
We use the same procedure with an additional nuance, we expanded the procedure to receive more than a single vector representation per word such that it will fit multi-prototype embeddings (\textit{e.g.} our embeddings and DeConf) and case sensitive embeddings (\textit{e.g.} word2vec). When given as set of words (like $W \setminus \{w\}$ when calculating $c(w)$) we first find the relevant sense for each element before inferring the outlier.
\citet{camacho2016find} suggested calculating $c(w)$ using the \textit{pseudo inverted compactness score}.
|
\section{Definitions of Determinism \label{Determinism}}
The notion of determinism was most famously articulated by Laplace, who suggested that `\emph{we ought to regard the present state of the universe as the effect of its antecedent state and as the cause of the state that is to follow}.'(\cite{laplace1820theorie}) That is, a theory is deterministic if according to that theory, the present state of the universe, together with the laws of nature, is sufficient to fix everything that happens in the future. We will refer to this approach as `Laplacean determinism.'
Although the terminology has evolved somewhat, many modern definitions of determinism are still predicated on the basic idea of the past `determining' the future. For example, the model-theoretic definition of determinism, due originally to \cite{Montague1974-MONDT}, tells us that \emph{`a theory is deterministic if for any two of its models describing the possible temporal evolution of a system, coincidence of the state at one time brings with it coincidence at all times from then on, where “coincidence” is to be spelled out in terms of a suitable mapping.'}. Similarly, \cite{Butterfielddeterminism} suggests that `\emph{[A] theory is deterministic if, and only if: for any two of its models, if they have instantaneous slices that are isomorphic, then the corresponding final segments are also isomorphic.}' And \cite{doi:10.1093/bjps/axv049} provide a careful classification of three different approaches to defining determinism: we may require that the solution to the differential equations should always be unique, we may require that if two linear temporal realizations can be mapped at one time they can always be mapped at all future times, or we may require that the theory's models are not branching. As the authors note, all three of their approaches are based on the core idea \emph{`that given the way things are at present, there is only one possible way for the future to turn out'} and therefore all three approaches are still expressions of the temporally directed picture originally associated with Laplacean determinism.
There is an obvious metaphysics underlying this temporally directed picture, which enjoins us to envision the universe as something like a computer, taking some initial state and evolving it forward in time to generate the content of reality(\cite{Wharton}). We can imagine that in Laplace's time this description may have been understood more or less literally. Indeed, as argued by \cite{chen2021governing}, many physicists and philosophers still appear to take this kind of `dynamic production' picture seriously - for instance, \cite{Maudlin2002-MAUQNA} advocates a `\emph{metaphysical picture of the past generating the future}.' And of course, if one has a dynamic production picture in mind then questions about whether or not our best current theories obey Laplacean determinism have profound consequences for our understanding of the kind of world we live in. In particular, proponents of dynamic production will presumably consider that a failure of Laplacean determinism indicates that the universe probably contains some processes which are intrinsically chancy - for there must be some way of deciding what happens when the process of dynamic production arrives at a point where events are not fixed by the laws and the initial state.
\subsection{The end of dynamic production? \label{td}}
But there are well-known problems for the Laplacean way of thinking about determinism: in particular, it fails to properly disentangle the concepts of determinism and predictability. Laplace's original comments on the matter conflated the two, describing determinism metaphorically in terms of the predictive abilities of \emph{`an intellect which at a certain moment would know all forces that set nature in motion, and all positions of all items of which nature is composed,'}(\cite{laplace1820theorie}) and this way of thinking is still evident in many modern analyses of determinism (\cite{Popper1992-POPTOU-2}). Yet at least in theory, predictability and determinism are supposed to be quite distinct - predictability is an epistemic matter, whereas determinism is metaphysical, which is to say, it is supposed to capture something about the way the world really is, independent of our ability to find out things about it. Thus for example \cite{Clark1987-CLADAP-2}, considering the possibility that the functional dependence of the future on the present state might fail to be effectively computable, assert that \emph{`there is no reason at all to tie the claim of determinism to a thesis of global predictability.'} The Laplacean definition doesn't seem to do justice to this intuition.
Later work has sought to disentangle the notions of determinism and predictability - for example, \cite{Montague1974-MONDT} argues that a purely ontological characterisation of determinism would not necessarily imply deducibility, because deducing the state at a time $t$ from the state at some other time $t_0$ requires us to express those states in sentences, but there are only denumerably many sentences whereas presumably there are more than denumerably many instants, assuming that instants are in a one to one correspondence with real numbers. And \cite{Earmanbook} writes, \emph{`determinism and prediction need not work in tandem; for the evolution of the system may be such that some future states are not predictable [...] although any future complement than the one fixed from eternity is impossible.'} But as this quote makes clear, much of the relevant work still seeks to understand determinism in terms of forwards temporal evolution (albeit perhaps evolution which we ourselves can't predict), and thus it does not fully succeed in separating determinism and predictability. For the notion of time evolution is closely linked to prediction: we typically write down laws describing temporal evolution from a present state to future ones, and non-coincidentally, we ourselves have a strong practical interest in predicting the future from facts about the present. But if deteminism is genuinely to be regarded as a property of the world rather than as a function of our practical interests, there's no reason it should be defined in terms of (fowards) temporal evolution: from the point of view of the universe as a whole, it need not be the case that things are always determined by a kind of process closely modelled on the way in which in which human observers usually want to predict them.
Of course, defining determinism in terms of temporal evolution is harmless so long as it is taken for granted that all laws of nature take a dynamical, time-evolution form as suggested by the dynamic production picture. But as argued in refs \cite{adlam2021laws,chen2021governing}, the dynamic production picture is increasingly under threat in contemporary physics. As a case study, consider the procedure of Dirac quantisation (\cite{dirac2019principles}) employed in the canonical approach to quantum gravity (\cite{Kucha1993CanonicalQG}). In this procedure, we start with a Hilbert space consisting of all kinematically possible states, and then apply the Hamiltonian and diffeomorphism constraints to arrive at the physical Hilbert space, consisting of all those states which satisfy the constraints. For example, the Hamiltonian constraint is encoded in the Wheeler-deWitt equation, $H || \psi \rangle = 0$, where $H$ is the universal Hamiltonian and $|| \psi \rangle $ is the universal quantum state. This equation has the consequence that the state $|| \psi \rangle $ does not evolve - that is to say, it can't be regarded as a state at a time, but must be understood as a description of the whole of history at once (this is one manifestation of quantum gravity's famous `problem of time'). Suppose then that we take all of this at face value and we want to consider the constraints as potential laws of nature. But if we accept that there is no evolution in this picture, clearly they cannot be \emph{evolution} laws. In fact what the constraints do is narrow down the set of physical possibilities to a subset of the space we might originally have envisioned - that is, the constraints act directly on the space of possibilities. Moreover, clearly if we had enough constraints of this kind, we might eventually find the set of physical possibilities narrowed down to just a single physically possible universe. That, surely, ought to count as a form of determinism: indeed, it would be an example of what \cite{penrose2016emperor} refers to as `strong determinism,' where the laws of nature single out the course of history uniquely. But the way in which the constraints work is to rule out whole possibilities at once, so there is no guarantee that we will find that given the content of some region of spacetimes we can straightforwardly use the constraints to fill in the rest of the content of spacetime; nor is that a natural thing to try to do in this context, since the constraints do not treat the universe piecewise in this way. These laws, if they are indeed laws, seem to apply to the whole of history `all-at-once,' and so we need a different way of thinking about what it would mean for such laws to be deterministic.
Indeed, `all-at-once' laws occur even in classical physics. Although it is most common to conceptualise classical mechanics in terms of the Newtonian schema (\cite{Smolinref}) in which laws act on states to produce time evolution, there is also an alternative Lagrangian description of classical mechanics in which systems are required to take the path which optimizes a quantity known as the Lagrangian (\cite{brizard2008introduction}): the Lagrangian approach assigns probabilities to entire histories, rather than taking a state at a time and evolving it forward. And path integrals - the analogue of the Lagrangian method within quantum mechanics (\cite{feynman2010quantum}) - have become so important to quantum field theory that increasingly we are seeing calls to take the Lagrangian description more seriously (\cite{hartlespacetime, Sorkinpath}). As argued by \cite{Wharton}, taking Lagrangian methods seriously requires us to adopt an `all-at-once' approach to lawhood in which we think of these laws applying externally and atemporally to the whole of history at once (\cite{Adlamspooky}), so again, we are in need of a different way of thinking about determinism in the context of these sorts of laws.
A multitude of other examples are to be found in contemporary physics. Retrocausal approaches to the interpretation of quantum mechanics have been attracting significant attention in recent years; see for example the two-state vector interpretation (\cite{Aharonov}), the transactional interpretation (\cite{Cramer}), Kent's approach to Lorentzian beables (\cite{Kent}), Wharton's retrocausal path integral approach (\cite{Wharton_2018}), and Sutherland's causally symmetric Bohmian model (\cite{Sutherlandretro}). It is clear that these kinds of models do not fit comfortably into a metaphysical picture of time evolution occurring in only one direction. Likewise, large sectors of the growing field of quantum information science are concerned with discovering, not evolution laws, but rather general constraints on what information-processing tasks can be achieved using quantum systems. For example, there is the `no-signalling principle'(\cite{MRC}), `information causality'(\cite{Pawlowski}), and `no-cloning'(\cite{Scaranicloning}). And the constructor theory of \cite{Deutsch_2015} expresses physical law not in terms of time evolution, but by appeal to an atemporal characterisation of what tasks are possible and impossible (where \emph{impossible} is understood to mean, not that a process can never occur, but that it can't executed repeatably in a cycle with arbitrary accuracy). It is not our intention in this article to argue that all of these examples will definitely turn out to be amongst the fundamental laws of the universe, but we do contend that approaches outside the standard time evolution paradigm have become sufficiently prominent in contemporary physics that the philosophy of physics would do well to start addressing them. And since the Laplacean approach to determinism clearly isn't a useful way of thinking about constructor laws, retrocausal laws or constraints on information-processing tasks, there is an obvious need for a generalisation of determinism which makes sense once we start taking laws outside the time evolution paradigm seriously.
At this point we should reinforce that it is not our intention to have an argument over terminology - one could certainly make a case that determinism should be defined in terms of evolution forwards in time because that is the way the term has always been used, and proponents of this way of thinking might then be inclined to suggest that worlds governed by laws outside the time evolution paradigm should never qualify as deterministic. We have no particular quarrel with this position: our intention here is simply to argue that there is a meaningful distinction between worlds in which everything is determined by laws which may fail to take a time evolution form, and worlds in which some events are not determined by anything, so there is a need for terminology which recognises that distinction. For convenience and familiarity we will describe the former as `deterministic' and the latter as `indeterministic,' but readers who prefer to use the term determinism to refer only to determination by forwards-evolving states are welcome to adopt whatever alternative terminology they prefer.
\subsection{Generalisations of Laplacean Determinism}
One fairly obvious approach to generalising determinism involves allowing evolution backwards as well as forwards - for example, \cite{Maar2019-MAAKOD} distinguishes between systems that are `futuristically deterministic' (their evolution into the future is unique) and `historically deterministic' (their evolution into the past is unique). This is indeed an improvement, in that it no longer insists arbitrarily that determination must occur in one particular temporal direction. But clearly this approach is still predicated on an \emph{evolution} picture, and thus it won't work well for many of the types of laws that we have just discussed. For example, consider a retrocausal world in which the initial and final states of the universe together with the laws of nature suffice to fix everything which happens over the whole course of history: neither the initial or final state alone will fix the course of history, so if we write down a theory of this world in a time-evolution form it won't show unique evolution into either the future or the past. And yet there still seems to be a sense in which this world is not really indeterministic - after all nothing that happens in it is random and we have no need to invoke anything that looks like an objective chance \footnote{Note that here and throughout this article we will use `objective chance' to refer exclusively to chances which arise from the fundamental laws of nature, such as the probabilities arising from the Born rule within indeterministic interpretations of quantum mechanics - i.e. in this article `objective chance' does not include higher-level emergent chances, chances derived via the method of arbitrary functions, deterministic probabilities or anything else that might in another context be called an objective chance.}. This makes it clear that the generalisation we're seeking must involve a more radical conceptual shift than simply allowing for some backwards evolution.
Another possible generalisation has been proposed by \cite{articleDowe} to deal with the possibility of worlds in which the relationships between events are temporally non-local. For example, consider a non-Markovian world which is governed by laws such that the evolution at a given time depends not only on the present state, but also on some facts about the past which are not recorded in the present state. The non-Markovian world fails to satisfy any definition of Laplacean determinism based on states or instantaneous time-slices, since the state at a given time does not suffice to determine all future states. But it seems odd to refer to such a world as `indeterministic' - after all, the evolution at a given time is entirely determined by the past. Similarly, consider the example of a world whose ontology contains only pointlike events, so there are no `states' in this world. Clearly in such a world the present state of the universe can't determine the future, since there \emph{is} no present state, but nonetheless it might be the case that the distribution of future events is fully determined by the distribution of past events and under those circumstances one might be inclined to describe this world as `deterministic' even though it doesn't satisfy any definition of Laplacean determinism based on states or instantaneous time-slices,
In response to this problem, \cite{articleDowe} proposes several alternative definitions of determinism - for example, his modal-nomic definition proposes that a world $W$ is deterministic iff `\emph{for any time t and any other physically possible world W', if W and W' agree up until t then they agree for all times}.' Both the non-Markovian world and the stateless world can potentially be deterministic according to this definition. But Dowe's approach retains the temporally-directed features associated with dynamic production, and thus it still excludes many types of non-time-evolution laws from consideration. For example, the retrocausal world where initial and final states fix the course of history will be judged indeterministic by Dowe's criterion. Similarly, a world governed by `all-at-once' laws in the style of the Wheeler DeWitt equation will typically be judged as indeterministic by Dowe's criterion, because in the all-at-once picture, past and future events depend on one another mutually and reciprocally and therefore we will not typically be able to write these dependence relations wholly in terms of determination of the future by the past. Yet `all-at-once' laws could in principle fix the whole course of history uniquely, so it seems very reasonable to think that we ought to have a definition of determinism which allows retrocausal and all-at-once worlds to be deterministic.
\subsection{Region-based definitions}
The notion of determinism comes under particular pressure in the context of General Relativity and other spacetime theories, for within such theories it is often challenging to identify what counts as an `instantaneous slice' or even an `initial segment.' Even in special relativity there is no unique notion of a `state at a time,' and when we move to more general cases we have to deal with spacetimes which are not globally hyperbolic and thus have no Cauchy surfaces, i.e. they contain not a single spacetime region which could be regarded as an `initial slice' from which deterministic evolution could take place. Thus a generalisation of determinism which seems less closely aligned with the time-evolution picture has become common in the study of spacetime theories. Roughly speaking, this `region-based' definition says a theory is deterministic if `\emph{agreement on regions of a certain kind (typically sandwiches or slices) forces agreement elsewhere'} (\cite{10.2307/687461}). There has been some disagreement on the precise details of this definition, but here is one example, from \cite{10.2307/687461}:
\begin{definition}
A theory with models $\langle M,O_i \rangle$ is \textbf{S}-deterministic, where \textbf{S} is a kind of region that occurs in manifolds of the kind occurring in the models, iff: given any two models $\langle M,O \rangle$ and $\langle M' , O' \rangle $ containing regions S, S' of kind \textbf{S} respectively, and any diffeomorphism $\alpha$ from S onto S', if $\alpha^{*}(O_i) = O_i'$ on $\alpha (S) = S'$, then there is an isomorphism $\beta$ from M onto M' that sends S onto S', such that $\beta^{*}(O_i)= O_i'$ throughout M and $\beta(S) = S'$.
\end{definition}
Note in particular that this definition of determinism is relativized to a type of region $\textbf{S}$, so we are free to choose whatever kind of region seems most interesting to us - perhaps Cauchy surfaces, if such things occur in the relevant kinds of manifolds, or perhaps some other kind of structure altogether. Since there is no longer an insistence that the determining region should be in the past of the determined region, nor even that the determining region should be of some particular type, this approach does indeed seem to be moving away from the temporally directed dynamic production picture. But what alternative picture does it have to offer? Are we supposed to imagine the universe being provided with the content of the chosen region as an input and then filling in the content of the rest of spacetime? This is not a very compelling picture, not least because if the initial state is not privileged, there is no longer any natural way to identify which particular region we should start from: even in an extremely simple Newtonian spacetime we could start from any time-slice and generate the rest of history by deterministic evolution forwards and backwards, so if this is to be regarded as a literal description we would seem to have a serious case of underdetermination. But if we're not supposed to think of the universe as being `produced' from the chosen region, it's unclear what, if anything, we can conclude about reality from the failure of this sort of determinism. Certainly it does not seem to follow that if this sort of determinism fails for some choice of region then the universe must contain processes which are arbitrary or chancy, because one could always make the rejoinder that the content of reality is produced starting from some other kind of region, or indeed reject a `production' metaphysics altogether.
In fact, the literature seems to suggest that we are as far as possible to avoid associating any metaphysics at all with this notion of determinism. For example, \cite{Doboszewski2019-DOBRSA-3} contends that `\emph{the issue of determinism cannot be totally decoupled from metaphysical questions ... However one may nevertheless hope to avoid making substantial use of metaphysics in an analysis of determinism.}' And \cite{Muller+2013+47+62} observes that in the model-theoretic approach to determinism, `\emph{the discussion (of determinism) shifted from the metaphysical question to one about properties of theories.}' The region-based formulation of determinism, being one of these model-theoretic approaches, can only be understood as a property of a given theory; within the model-theoretic paradigm we are actively discouraged from thinking about determinism as a potential property of the world. But if this notion of determinism is not to be regarded as having metaphysical content, what exactly is its purpose? Clearly it is not to be regarded as a practical characterisation of the predictions we can actually make with the given theory: for as we have seen, determinism does not entail predictability, and moreover even in cases where there \emph{would} be predictability in principle, the definition offers no reason to think that the region $\textbf{S}$ will be one from which we could reasonably collect all the relevant data. Indeed, \cite{Doboszewski2019-DOBRSA-3} observes that if the region $\textbf{S}$ is supposed to be a Cauchy surface, we will not even be able to verify the existence of the right type of region in our actual spacetime. Thus it is somewhat unclear why this particular technical notion of determinism should be regarded as an interesting property of a theory; technical features of theories are typically interesting insofar as they are presumed to reflect some fact about the reality to which the theory corresponds, and/or some practical feature of how the theory works in application, neither of which obviously holds here.
One might hope to look to the actual practice of physics to understand the utility of the region-based definition. But the main application of the term `determinism' in the General Relativity community appears to be in the context of the initial-value problem, i.e. the problem of determining the course of history given the data on some time-slice $\Sigma$. If the data on $\Sigma$ satisfies certain constraints, then it can be embedded in a unique spacetime which is a solution of Einstein's equations; moreover the spacetime will be globally hyperbolic and $\Sigma$ will be a Cauchy surface of it (\cite{choquet1969global}). This is important to physicists because under these conditions it's possible to come up with a `time-evolution' version of General Relativity, and also to formulate canonical quantum gravity (\cite{Kucha1993CanonicalQG}), which is intended to furnish a description of the time evolution of the quantum state of the $3$-metric (although as we noted earlier, due to the famous `problem of time' it's not entirely clear that the evolution picture succeeds). That is to say, the notion of `determinism' employed by physicists working on General Relativity remains quite closely tied to Laplacean determinism and the time evolution paradigm - and thus, perhaps, involves an unarticulated commitment to a metaphysics of dynamic production. Therefore it's unclear that the justification for this region-based generalisation of determinism can come from within mainstream physics practice.
Moreover, even if there \emph{are} useful applications for a definition of determinism which is shorn of metaphysical content, we contend that there would also be many useful applications for a definition of determinism which \emph{does} aim to characterise a metaphysically interesting feature of reality. For as \cite{Muller+2013+47+62} puts, it, `\emph{The question of whether the world is deterministic or not, which is perceived by many to have enormous consequences for our understanding of ourselves as free agents, is also first and foremost a metaphysical one ... determinism and indeterminism are metaphysical notions.}' Determinism is frequently invoked in explicitly metaphysical discussions of topics like explanation, free will and probability (we will see some examples in section \ref{app}) and it seems unlikely that a definition of determinism which has been deliberately uncoupled from metaphysical commitments will be particularly useful in these applications. Furthermore, moving away from the time evolution paradigm has interesting consequences for a range of important debates within both physics and philosophy, and in order to fully appreciate these consequences it will be necessary to have some idea of what we might be moving \emph{towards} - that is, we must ask what alternative sort of metaphysical picture might replace the time evolution picture. Understanding the role of determinism in a post-time-evolution setting is surely an important part of that project, and for that we will need something more than the region-based formulation.
In addition to these concerns, the region-based definition of determinism, even considered as a purely technical criterion, becomes difficult to apply in many physically relevant contexts. For example, \cite{Doboszewski2019-DOBRSA-3} points out that when applying this definition to General Relativistic models, the verdict we reach depends sensitively on our choice of the type of region $\textbf{S}$: Doboszewski identifies ten types of cases which intuitively seem like they should perhaps be regarded as violations of determinism, and observes that there is no particular causal structure that they have in common, so there is no obvious way of identifying a specific type of region which will get the intuitively `right' verdict in all ten cases. Doboszewski concludes that we should simply adopt a pluralistic approach to determinism, but alternatively one might regard the fact that the region-based definition is so sensitive to a choice of region as evidence that it is perhaps not getting at any particularly interesting or relevant feature of these theories. Even more serious problems are likely to arise in the context of quantum gravity: for many approaches to quantum gravity treat spacetime itself as emergent from some underlying reality(\cite{cc,Blau2009}), and therefore we will not even be able to identify `regions' in the first place, so the definition of determinism will be applicable at best at an emergent level. This seems troubling - do we really believe that it is conceptually impossible for a theory of quantum gravity to be deterministic in a fundamental sense?
It's also clear that the region-based formulation will not work well in the context of many approaches outside the time evolution paradigm. For example, consider the retrocausal model where the content of reality is completely fixed by an initial and final boundary condition together, but not by either of them individually. The region-based definition of determinism will allow us to regard this theory as deterministic only if we are willing to accept a `region' made up of two disconnected parts, which seems to be stretching the definition of the term `region' quite far: and in any case, even if we are willing to accept the union of two disconnected regions as a region, one might well feel that a useful definition of determinism should not make questions about whether a possible world is deterministic depend on the essentially linguistic question of what we're willing to count as a `region.' One possible response to these troubles, as advocated by \cite{Earman2007-EARAOD}, is to adopt a form of quietism, i.e. say that the notion of determinism may simply be inapplicable in many cases that interest us in modern physics. But the alternative is to say that the `region-based' approach is perhaps not the right generalisation of the original idea of Laplacean determinism, or at least that the generalisation we propose in this article may be more appropriate in some cases - particularly cases where we are concerned not only with technical criteria but with metaphysical content.
\section{The Constraint Framework \label{LON}}
Because we are concerned here to analyse determinism as a metaphysical property of worlds, rather than a formal property of theories, we will follow \cite{Muller+2013+47+62} in eschewing model-theoretic language and instead characterising the notion in modal terms, since determinism and objective chance are patently modal notions: \emph{`They are not about what does or doesn’t happen, but about what can or what has to happen and are thus built upon the modal notions of possibility and necessity.'}( \cite{Muller+2013+47+62} ) Our approach to determinism will therefore be predicated on a fairly robust approach to modality - we will assume that for every possible world there exists some well-defined modal structure which we can invoke to determine whether or not that world is deterministic. We will not comment further on the nature of that modal structure - it might be defined by the axioms of the systematization of that world's Humean mosaic which is robustly better than all other systematizations (\cite{lewishumean}), or it could result from governing laws within any one of the realist approaches to lawhood (\cite{Armstrong1983-ARMWIA,Bird2005-BIRTDC, Ellis2001-ELLSE-2}), or it could be understood in terms of Lewis' modal realism (\cite{Lewis1986-LEWOTP-3}), or it could be something else altogether.
Objective modal structure is most commonly discussed in the context of ontic structural realism (\cite{Ladyman2007-LADETM}). In some cases this modal structure is characterised in terms of \emph{causal} structure (\cite{Berenstain2012-BEROSR,doi:10.1080/02698590903006917}), but as argued in \cite{adlam2021laws}, due to the asymmetrical nature of the standard notion of causation this approach doesn't work well when we're dealing with possibilities outside the time evolution paradigm, so in this paper we will instead adopt the more general approach to modal structure advocated by \cite{Ladyman2007-LADETM}. In order to formalise this notion of modal structure, we will employ the framework developed in \cite{adlam2021laws}, which characterises laws of nature in modal terms as \emph{constraints}. This approach is inspired by the increasing prominence of constraint-based laws in physics, as briefly surveyed in section \ref{td}; a very similar account of lawhood in terms of constraints was recently given by \cite{chen2021governing}.
To formalise the notion of a constraint we make use of the notion of a Humean mosaic, which is understood to consist of all and only the matters of local, non-modal particular fact in a given world. Obviously what counts as a matter of non-modal particular fact will vary according to one's other metaphysical commitments, and indeed this is our purpose in employing the general term `Humean mosaic' - our aim is to offer a definition of determinism which is compatible with a variety of accounts of the non-modal content of reality. We emphasize that the use of this terminology is not intended to reflect a commitment to the standard Humean ontology consisting only of the Humean mosaic - we use the phrase `Humean mosaic' to refer to all of the actual, non-modal content of reality, but we do not intend by that usage to rule out the possibility that reality also has genuine modal content: for example, a realist about modality might wish to maintain that constraints are ontologically prior to the actual Humean mosaic and thus play a role in determining its content.
Because we don't wish to make our approach to determinism contingent on any particular account of modality, we will henceforth define constraints \emph{extensionally}, appealing to techniques employed in modal logic: a constraint will simply be defined as a set of Humean mosaics, i.e. the set of all mosaics in which that constraint is satisfied. In some cases, a constraint will be expressible in simple English as a requirement like `no process can send information faster than light,' such that the constraint corresponds to exactly the set of mosaics in which the requirement is satisfied. But there are many sets of mosaics which will not have any straightforward English characterisation, so we will not be able to state the corresponding constraint in simple terms. Nonetheless, each set still defines a unique constraint.
Using these definitions, we postulate that every world has some set of laws of nature which are an objective fact about the modal structure of that world, and we characterise these laws in terms of \emph{probability distributions over constraints}. Within this picture, we can imagine that the laws of nature operate as follows: first, for each law a constraint is drawn according to the associated probability distribution, and then the constraints \emph{govern} by singling out a set of mosaics from which the Humean mosaic of the actual world must be drawn - i.e. the actual mosaic must be in the intersection of all the chosen constraints. That is to say, the laws of nature associated with a given world are understood to operate by narrowing down the set of physical possibilities for that world, thus dictating what properties the Humean mosaic for that world is required to have. We note that we could have given a similar definition in terms of probability distributions over mosaics rather than constraints, but we have chosen to use constraints here in order to acknowledge that a law may require (deterministically or probabilistically) that the actual Humean mosaic belongs to a given set of mosaics, but then say nothing further about which particular mosaic within the set will be selected - that is, the law prescribes no distribution within the set, not even the uniform distribution. This point will be a crucial feature of our approach to determinism.
As shown in \cite{adlam2021laws}, a very large class of possible laws can be written in this framework. The Wheeler deWitt equation can be analysed as a law assigning probability $1$ to the constraint consisting of all Humean mosaics instantiating a universal quantum state $|| \psi \rangle$ satisfying $H || \psi \rangle = 0$. A prohibition on superluminal signalling can be analysed as a law assigning probability $1$ to the constraint consisting of all the Humean mosaics in which no superluminal signalling occurs, and $0$ to all other constraints. A constructor law can be analysed as a law assigning probability $\epsilon$ to the set of all mosaics in which some `impossible' process occurs in a repeated cycle, where $\epsilon$ is very close to zero. And given a probabilistic evolution equation, we can use the equation to construct a probability distribution $p_s(h)$ over each possible history $h$ of a system $s$, and then analyse the equation as a law which assigns probability $p_s(h)$ to the constraint consisting of the set of all mosaics in which system $s$ has history $h$. Thus we can use these probability distributions over constraints to characterise the laws of nature without needing to be too specific about what sorts of laws we mean, and this makes it possible for us to give a general definition of determinism which is applicable across many different conceptions of the content of physical law.
We acknowledge that the idea of a constraint being selected according to some probability distribution may give rise to some scepticism, given that this process presumably only happens once and therefore no relative frequency interpretation is possible. Indeed, we share that scepticism, and would be inclined to adduce the difficulty of making sense of these sorts of probability distributions as a reason to believe that the world is likely to be deterministic in at least one of the senses described below. But we would argue that any conception of the content of physical reality which holds that there exist genuine objective chances is in effect committed to a probability distribution over Humean mosaics, and therefore if we are to articulate the notion of determinism it is necessary to make reference to the possibility of probability distributions of this kind in order to characterise the relevant contrast class - whether or not that possibility is actually a coherent one is a further question, and one which we will not have space to address here.
\section{Holistic Determinism \label{det}}
Now our task is to find a way of decoupling the notion of determinism from time evolution. One obvious way to achieve this would be to appeal to the intuitive idea that determinism and objective chance are mutually exclusive: that is, either it is the case that the fundamental laws of nature are deterministic, or it is the case that they involve some sort of `intrinsic randomness.' This line of thought leads to the idea that we should simply define determinism as the absence of objective chance. But the problem with this route is that we don't really have a strong grasp on what objective chances are either - the only generally agreed-upon characterisation is that they should obey the Principal Principle (\cite{Lewis1980-LEWASG}).
The constraint framework provides a precise way of formulating this idea. Indeed, it allows us to disambiguate several different ways in which a world may be deterministic. First, we will say a world (which, recall, can be associated with some set of laws of nature) obeys \emph{holistic determinism} iff the probability distributions induced by the laws are trivial:
\begin{definition}
A world satisfies \emph{holistic determinism} iff every one of its fundamental laws induces a probability distribution which assigns probability $1$ to a single constraint and zero to all disjoint constraints.
\end{definition}
When this condition is satisfied, we can get rid of the probability distributions altogether and simply assign to each law of nature a constraint, so the actual mosaic must lie in the intersection of all of these constraints. This is the analogue in the constraint framework of the idea that determinism is associated with the absence of objective chance: `no objective chance' is translated as `there are no objective probability distributions over constraints.' Of course the actual mosaic must still be selected from the intersection, but this selection does not need to involve any objective chances, since as noted in section \ref{LON} the laws do not define any particular probability distribution over the mosaics in the intersection, not even the uniform distribution.
However, it might be argued that defining determinism as the absence of objective chance is in some cases too broad. Consider for example a version of classical electromagnetism (\cite{Rousseaux:2005hn}) which includes the stipulation that the electromagnetic potentials are real fields which are a part of nature (and thus they feature in the Humean mosaic) - that is, different configurations of the potentials represent physically inequivalent situations. Intuitively, it seems clear that this world is not deterministic, since a large variety of physically inequivalent configurations of the electromagnetic potentials will be compatible with the laws in any given situation. But the theory defines no probability distribution over the different configurations of the electromagnetic potentials, so as far as we know there aren't really any `chances' here, not even the uniform distribution - some configuration must occur, and so some configuration does, but there is nothing that can be said about how likely different configurations are to occur. We will henceforth refer to cases of this kind, involving events which are not determined by anything but which also are not chancy events, as `arbitrary.' Thus it seems that a generalized definition of determinism should allow that worlds containing arbitrarinesss of this kind may be indeterministic even in the absence of objective chance.
In light of this possibility, we distinguish further between strong and weak holistic determinism:
\begin{definition}
A world satisfies \emph{strong holistic determinism} iff it satisfies holistic determinism and there is only one mosaic in the intersection of the set of constraints associated with its laws of nature
\end{definition}
\begin{definition}
A world satisfies \emph{weak holistic determinism} iff it satisfies holistic determinism and there is more than one mosaic in the intersection of the set of constraints associated with the laws of nature
\end{definition}
In the case of strong holistic determinism the actual mosaic is singled out uniquely by the laws and we have neither chance nor arbitrariness; whereas weak holistic determinism allows arbitrariness, but not chance. Thus classical electromagnetism with ontological electromagnetic potentials will satisfy weak holistic determinism but not strong holistic determinism. One might perhaps question whether `weak holistic determinism,' really ought to count as a form of determinism, given that it allows events to occur which are not determined by anything. However, even worlds satisfying Laplacean determinism exhibit some arbitrariness - typically in such a world the initial state of the universe is `arbitrary' in just the same way as the configuration of the electromagnetic potential, since we don't assign objective chance distributions over initial states but we don't usually take them to be determined by anything either. So when we set out to generalize determinism in a way that does not give any special status to the initial state of the universe, we have two options - either we allow arbitrariness elsewhere as well, or we allow no arbitrariness whatsoever. Weak holistic determinism takes the former route and strong holistic determinism the latter, so both have at least some claim to be the spiritual heir of Laplacean determinism.
That said, there does seem to be a sense in which a world satisfying Laplacean determinism is more strongly deterministic than electromagnetism with ontological electromagnetic potentials, since the arbitrariness involved in the former is limited to the start of time, whereas in the latter, even if we know the full content of spacetime except for some small region, we still won't be able to determine the electromagnetic potentials in that region (since it is always possible to perform a gauge transformation on the potentials which goes to zero outside the region). So ideally we would like to offer some further definition which distinguishes between worlds satisfying Laplacean determinism and worlds like the one governed by electromagnetism with ontological electromagnetic potentials. Of course, we could add a special exception for the initial state into our definition of determinism, but this would seem like a relic of the older temporally directed approach - if we're no longer insisting that laws must take the forward time-evolution form there's no good reason to treat the initial state as special. So instead we suggest the following definition:
\begin{definition}
A world associated with a set of laws of nature satisfies \emph{delocalised holistic determinism} iff it satisfies holistic determinism, and there is no pair of mosaics in the intersection of the set of constraints associated with its laws of nature which are identical everywhere except on some small subregion of spacetime
\end{definition}
Note that this definition assumes the existence of a natural map between the two spacetimes associated with two different mosaics. Of course in the case where two mosaics are associated with spacetimes having very different structure it may not be the case that there exists a unique natural map between their spacetimes (although perhaps such a thing could be arrived at using the best-matching techniques of \cite{articleBar}). However, in that case clearly we would not want to say that the two mosaics are identical everywhere except on some small subregion of spacetime, so there is no need to invoke a map; this definition becomes relevant only when we are dealing mosaics with spacetimes similar enough to support a natural map between the spacetimes, except possibly on some small subregion.
The point of this definition is that it ensures the arbitrariness involved in selecting the actual mosaic can't be localised in any particular region, which means we can't have any indeterministic `holes.' To see how the definition works, consider a world satisfying Laplacean determinism in which the laws of nature are time-reversal invariant. In the constraint framework, this world has a set of laws of nature which are each associated with a single constraint, and the intersection of these constraints is a set of mosaics such that each mosaic in the set has a different initial condition. Moreover, because the laws of nature of this world are deterministic and time-reversal invariant, it must be the case that no pair of mosaics in the set are the same on any time-slice after the beginning of time, since otherwise reversing the direction of time would take a single state into two different states, in violation of the assumption that the laws of this world satisfy Laplacean determinism. Thus when the actual mosaic is drawn from this set of mosaics, this has the effect of fixing the state of the world on the initial state, but since the mosaics differ at all subsequent times as well, one could equally well regard this process as fixing the state of the world on any other time slice. So there is no fact of the matter about which \emph{particular} time-slice is responsible for the rest of history: when we select one of the allowed mosaics we determine all the time-slices at once. Therefore the arbitrariness involved in a universe satisfying Laplacean determinism can't be located in any specific spacetime region, so the Laplacean universe does indeed satisfy weak delocalised holistic determinism.
On the other hand, a world governed by electromagnetism with ontological electromagnetic potentials will not satisfy delocalised holistic determinism, because given some mosaic in the intersection of the constraints, we can generate another mosaic which is identical except on some small region of spacetime by simply performing a gauge transformation which goes to zero everywhere outside the region in question, and since the laws don't assign any probabilities over gauge-related configurations, this other mosaic must also lie in the intersection of the constraints. Thus this definition captures what is distinctive about worlds satisfying Laplacean determinism without needing to give special significance to the initial condition - the important thing about such worlds is not that the arbitrariness occurs only at the start of time, but that the arbitrariness can never be localised anywhere, and thus in general such worlds have a high degree of predictability and consistency even though the laws of nature do not fix everything in the world uniquely.
Evidently all of these definitions are suitable for studying worlds with laws outside time evolution paradigm: worlds with non-Markovian laws, worlds with an ontology that does not include states, worlds with all-at-once laws and worlds with retrocausal laws are all capable of satisfying weak, delocalised and/or strong holistic determinism, depending on the specific details of the laws in question. The definitions encapsulate quite a different way of thinking about what it is for a world to be deterministic: rather than treating the initial conditions (or the conditions of some spacetime region) as an input and asking if the input is sufficient to determine the rest of the course of history, we have the laws induce constraints which select entire mosaics at once, and then we ask if these laws involve any objective chances and if they suffice to pick out the course of history uniquely. Thus these definitions also realise our desideratum of making a clean distinction between determinism and predictability, as they characterise determinism entirely in terms of objective modal structure, without explicitly or implicitly leaning on facts about the practical interests and epistemic limitations of human observers. A world which is deterministic in the holistic sense may indeed exhibit a high degree of predictability, but it also may not - the way in which the laws determine the course of events may be highly non-local in space and time, meaning that it may be impossible for limited local agents to gather enough data to see the whole picture. Thus this framework fulfils the mandate of providing a precise way of thinking about determinism as a metaphysical property of a world rather than a formal property of a theory, thus opening the door for that notion to be applied to a variety of ongoing metaphysical discussions.
\subsection{Other Comments}
\begin{enumerate}
\item The distinction we have made here is similar to one employed in \cite{penrose2016emperor}, which distinguishes between determinism, i.e. what we have referred to as Laplacean determinism (\emph{`if the state of the system is known at any one time, then it is completely fixed at all later (or indeed earlier) times by the equations of the theory'}) and strong determinism, i.e. what we have referred to as strong holistic determinism (\emph{` it is not just a matter of the future being determined by the past; the entire history of the universe is fixed, according to some precise mathematical scheme, for all time.'}). However, Penrose's scheme does not seem to accommodate the intermediate possibility of worlds which satisfy weak holistic determinism but not Laplacean determinism, such as worlds which have some degree of arbitrariness at some point other than the initial state of the universe but nonetheless have no objectively chancy events. Moreover, Penrose is rather vague about what it would take for the entire history of the universe to be fixed, but the constraint framework allows us to be precise: the history of the universe is fixed if and only the laws of nature induce trivial probability distributions over constraints, and the intersection of the constraints associated with these laws contains exactly one mosaic.
\item In one sense, weak holistic determinism, strong holistic determinism and delocalised holistic determinism are all weaker than Laplacean determinism, since worlds with laws outside the time evolution paradigm like those discussed in section \ref{td} could satisfy any of these forms of holistic determinism without satisfying Laplacean determinism. However, in another sense strong holistic determinism is much stronger than Laplacean determinism, because Laplacean determinism allows for the existence of arbitrariness (in the selection of initial conditions and the selection of parameter values) while strong holistic determinism insists that everything is fixed once and for all by the laws of nature.
\item We reinforce that all of these definitions for determinism depends crucially on the claim that reality has a unique, observer-independent modal structure. Given any Humean mosaic we could come up with an infinite number of sets of constraints whose intersection contains only this mosaic - trivially, for a mosaic $A$ we can pick two random mosaics $B$ and $C$ and choose the constraints $\{ A, B\}$ and $\{ A, C\}$. So simply finding such a set which is consistent with our observations doesn't tell us that the world must be deterministic: we must make the further claim that these constraints are \emph{in fact} the constraints which feature in the objective modal structure of reality, and therefore we must believe that there is some fact of the matter about what this objective modal structure really is. So for example, assessing a claim that the world satisfies strong holistic determinism has several steps: first we must verify that the proposed constraints do indeed single out a unique Humean mosaic, second, we must verify that the proposed constraints are consistent with our observations, and third, we must decide whether the proposed constraints seem like plausible candidates for laws. This last part of the process will naturally be shaped by whatever expectations we have about laws - for example, if we think the laws of nature are simple then we will tend to expect that the constraints induced by the laws can be expressed in a closed form in simple terms.
\item Our definitions also depend crucially on the stipulation that the laws of nature assign no probability distribution over the mosaics in the constraints that they induce, nor over the mosaics in the intersection of all of those constraints. However, this certainly does not entail that we ourselves may not assign any distribution over mosaics in the intersection. For we may have Bayesian priors which lead us to assign some distribution over mosaics before taking any of the laws of nature into account - for example we might simply assign the uniform distribution, or we might choose a distribution which favours simpler mosaics. Thus when we learn about a constraint induced by a law, we will update our beliefs by eliminating all mosaics not consistent with the corresponding constraint and renormalizing our credences, so after conditioning on all of the laws we will be left with a normalized probability distribution which is non-zero only on mosaics in the intersection of the constraints. In the case of strong holistic determinism, no matter what priors we start out with, after conditioning on the laws we will be left with a distribution that assigns probability $1$ to a single mosaic; but in the case of weak holistic determinism we will in general end up with a non-trivial distribution, which will be the uniform distribution if the original prior distribution was the uniform distribution, but which may be non-uniform if our original priors were non-uniform.
But this distribution is a \emph{subjective} probability distribution derived from our original subjective priors, not an objective chance distribution arising from the laws themselves: so in this case we still have determinism in the sense of `the absence of objective chance,' even though observers may still assign non-trivial probabilities over mosaics compatible with the laws. This is precisely analogous to the way in which we might assign subjective probability distributions over initial conditions in the context of Laplacean determinism, for example in analyses of statistical mechanical systems based on the principle of indifference (\cite{Meacham2010-MEACAT}).
\end{enumerate}
\subsection{Humean Mosaics}
None of our definitions are intended as a direct replacement of model-theoretic definitions of determinism, because our approach does not simply take a theory and decide algorithmically whether that theory is deterministic. For a start, the definition applies to worlds, not theories, so to apply it to a theory (expressed either as a set of models or a set of sentences) we would have to postulate a world governed by that theory, which would require us to translate the theory into putative `laws of nature' in the constraint formulation, and typically there will be more than one way of doing that. We would also have to decide whether we need to add additional laws of nature from outside the theory, since our definitions require us to consider the complete set of laws of nature associated with a given world. So in circumstances where one wants nothing more than a simple judgement as to whether some particular \emph{theory} is deterministic, the model-theoretic approach is obviously more suitable. This is not a weakness of our approach, because it was never our intention to define determinism as a formal property of theories. It would be interesting in future work to develop formal expressions of this approach to determinism which can be applied within a model-theoretic framework, but that is beyond the scope of the present project.
Applying our definitions to a specific theory would also require a judgement about what sorts of categorical properties feature in the set of Humean mosaics associated with the theory. For example, consider quantum mechanics: if we suppose that quantum mechanics is a complete theory of reality and we take it that wavefunction collapse is a part of the Humean mosaic, then the theory does not satisfy weak or strong holistic determinism, whereas if we say that only the unitarily evolving quantum state is part of the Humean mosaic, then we get the Everett interpretation (\cite{Wallace}), which satisfies weak holistic determinism. Note that the Everett interpretation in its usual form does not satisfy strong holistic determinism, since the initial state of the universe is arbitrary; however, the Everett interpretation coupled with something like Chen's `quantum Wentaculus,' which features a lawlike initial state (\cite{chen2018quantum}), could potentially satisfy strong holistic determinism.
One might worry that this ambiguity is a problem for our approach. However, it is in fact a deliberate choice, because our intention is to define in a general, metaphysical sense what it means for a world to be deterministic, and therefore it is important to avoid predicating the definition on specific ontological assumptions. Moreover, this definitional quirk is not unique to our approach - as emphasized by \cite{EarmanGR}, judgements about whether a theory satisfies Laplacean determinism likewise depend on our presuppositions about the possibility space that we're dealing with. Thus, for example, Laplacean determinism is similarly noncommittal about quantum mechanics. Moreover, this ambiguity has not prevented Laplacean determinism from yielding generally meaningful results in many cases, although of course there are a few interesting grey areas and ongoing debates (e.g. see \cite{EarmanGR, delsanto2021indeterminism}). We are happy for our definition to inherit this particular form of ambiguity - it would not make sense for our definition of determinism to be stricter than Laplacean determinism in this regard, since our intention here was to \emph{generalize} Laplacean determinism and therefore it is important for us to be able to recognise worlds satisfying Laplacean determinism as a subclass of worlds that are deterministic in our sense.
\subsection{Examples: Spacetime theories}
In order to show that the proposed approach to determinism captures some of the intuitive content that is not well addressed by the region-based formulation, we discuss how some of the problem cases cited by \cite{Doboszewski2019-DOBRSA-3} can be accommodated by the constraint approach.
\paragraph{Spacetimes violating weak causality conditions}
Spacetimes violating weak causality conditions do not allow for an initial value problem or anything remotely like one. This means we can't easily apply a region-based definition of determinism, or any other approach which asks us to take some physical data and construct the rest of the solution from it. However, since constraints pick out whole solutions at once, we can certainly have constraints which contain Humean mosaics corresponding to spacetimes which violate weak causality conditions, and therefore we can certainly postulate laws of nature which give rise to such spacetimes, and then check if those laws exhibit one or more of the forms of holistic determinisms we have postulated. It is clear that there do exist laws of nature which will produce such spacetimes, because the Einstein equations can produce them in special cases; and since the Einstein equations in their standard form are `all-at-once' laws whose solution is an entire course of history rather than a state at a time, they are exactly the kind of law for which the constraint formulation is particularly well-suited.
\paragraph{Spacetimes with closed timelike curves}
\cite{QMG, adlam2021tsirelsons} studied the compatibility of `closed loops' with a concept of `global determinism' which is similar to the notion of holistic determinism that we have used in this article. \cite{QMG, adlam2021tsirelsons} argues that the values of variables in these closed loops `come out of nowhere,' which is to say, they are not determined by anything, so closed causal loops are not consistent with a universe satisfying holistic determinism. Using the definitions set out in section \ref{det}, we can make this observation more precise. First, clearly variables which are not determined by anything are incompatible with a universe that obeys strong holistic determinism, so the argument of \cite{QMG, adlam2021tsirelsons} work perfectly well if `global determinism' is intended to mean `strong holistic determinism.' Moreover, indeterministic closed loops are also ruled out by delocalised weak holistic determinism, because the arbitrariness associated with the variables in a closed causal loop is localised to the events inside the loop. On the other hand indeterministic closed loops are not ruled out by weak holistic determinism, since the values of variables inside the loop could be arbitrary rather than chancy.
\paragraph{Extendible Spacetimes}
An extendible spacetime is a spacetime which is a subset of a larger spacetime(\cite{2012caus}). Suppose that a given world has a set of laws such that the intersection of the constraints associated with the laws contains a mosaic corresponding to an extendible spacetime. Then, assuming that there is no law specifically prohibiting extensions, it necessarily follows that the intersection of the constraints associated with the laws will contain more than one mosaic, since we can take the mosaic in this intersection which has an extendible spacetime and obtain another mosaic which will also be in the intersection by extending it. Therefore a world with laws allowing extendible mosaics will typically fail to satisfy strong holistic determinism. Moreover, the extended mosaic will differ from the original mosaic only on a small subregion of spacetime, in the sense that the new mosaic postulates a small region of spacetime that does not exist at all in the original mosaic, and therefore a world with laws allowing extendible mosaics will typically fail to satisfy delocalised holistic determinism. However a world with such laws may satisfy weak holistic determinism, since the laws need not define any probability distribution over the extent of spacetime.
\paragraph{Spacetimes with holes}
A spacetime is hole-free if `\emph{the Cauchy development of any spacelike surface is “as large as it can be” '}(\cite{pittphilsci4380})- i.e. it is not the case that the domain of dependence of some partial Cauchy hypersurface can be extended by imbedding it into a different spacetime (\cite{2012caus}). Suppose that a given world has a set of laws such that the intersection of the constraints associated with the law contains a mosaic corresponding to a spacetime which is not hole-free. Then, assuming that there is no law specifically prohibiting extending the domain of dependence of partial Cauchy hypersurfaces, it follows that the intersection of the constraints associated with the laws will contain more than one mosaic, since since we can take the mosaic in this intersection which has an extendible spacetime and obtain another mosaic which will also be in the intersection by extending the domain of dependence of some partial Cauchy hypersurface. So a world with laws allowing spacetimes that are not hole-free will typically fail to satisfy strong holistic determinism. Moreover, the extended mosaic will differ from the original mosaic only on a small subregion of spacetime, in the sense that the new mosaic postulates a small region of spacetime that does not exist at all in the original mosaic, and therefore such a world will typically not be compatible with delocalised holistic determinism either. However, a world with such laws may satisfy weak holistic determinism, since the laws need not define any probability distribution over the size of the domain of dependence.
\section{Applications \label{app}}
We have criticized Laplacean determinism for being inappropriately intertwined with our practical interests (in particular, our desire to predict the future based on facts about the present); but one might argue that being aligned with our practical interests is precisely what made Laplacian determinism a useful concept, and thus one might worry that generalizing determinism as we have done here serves only to make the concept less useful. Thus in this section we will explain how this generalization of determinism may have important consequences for scientific and philosophical thinking. The concept of Laplacean determinism has played a range of different intellectual roles in science and philosophy, so we will proceed by examining how a generalised notion of determinism may contribute to some of these domains of application.
\subsection{Scientific Explanation}
Determinism has played an important role in shaping our expectations for scientific explanation - for a long time it was standard in science to explain events by showing how they could be deterministically produced by past conditions. This is formalised in the `deductive-nomological' model for scientific explanation (\cite{popper2005logic,doi:10.1086/286983}), which posits that a valid explanation should be composed of a set of sentences from which the explanandum logically follows, where at least one of these sentences must express a law of nature and that law must be an essential element in the deduction of the explanandum. Most common examples of DN-explanations involve time-evolution laws or at least temporally directed laws, and thus DN explanations usually involve postulating some past conditions and showing that the deterministic forwards time-evolution of these conditions gives rise to the explanandum. This model is sometimes relaxed to allow for explanations involving statistical inference rather than exact logical deduction (\cite{hempel1965aspects}), but deductive-statistical and inductive-statistical explanations are simply a generalisation of forwards-evolving deterministic explanations and thus they typically preserve the expectation that the future will be explained by lawlike evolution from the past.
Holistic determinism opens up new possibilities for scientific explanation - for if the definition of Laplacean determinism is largely a function of our practical interests then there is no reason to expect that all valid scientific explanations will be predicated on forwards time evolution, since for scientific realists one of the main purpose of scientific explanation is to achieve \emph{understanding}, to which practical concerns are only tangentially relevant. So when we come upon phenomena for which there seems to be no satisfactory explanation within a forward time-evolution picture, we should be open to the possibility of explanations based on global laws using the constraint framework. A good example of this is the past hypothesis: by definition nothing evolves into the initial state of the universe, so the initial state can't be given a lawlike explanation if we assume that laws always take the forward time-evolution form. But the past hypothesis can straightforwardly be written as a constraint - it is simply the set of all Humean mosaics in which the arrangement of local matters of particular fact at one end of the mosaic has low entropy (or some other more sophisticated characterisation of the desired initial state - see \cite{pittphilsci8894}). The past hypothesis can therefore be explained in this framework by hypothesizing that there is some law of nature which induces this particular constraint - perhaps something like the `Mentaculus' of \cite{doi:10.1142/9789811211720_0001} or the quantum `Wentaculus' of \cite{chen2018quantum}) - thus ensuring that the initial state will necessarily be a low entropy state whilst the selection of one low entropy initial state in particular remains arbitrary.
Similar arguments can be made about parameter values: for example, cosmologists are currently much exercised over the problem of why the cosmological constant is so much smaller than their models suggest it ought to be (\cite{RevModPhys.61.1}). It's hard to see how this feature could be explained within the kinematical/dynamical picture, because the cosmological constant is usually considered to be a fundamental constant of nature which has always had the value that it has, and so there is no option to say that it has been produced by evolution from some earlier conditions. Thus most current approaches to explaining the value involve fairly exotic explanatory strategies, such as anthropic arguments in the context of a putative multiverse (\cite{Banks:2000pj}). However, within the constraint framework there is no reason why we can't simply explain the cosmological constant in a lawlike manner: all we need to do is suggest that there is some law of nature which induces the constraint consisting of all the Humean mosaics in which the cosmological constant is very small. In the constraint framework, this nomological explanation has exactly the same status as more familiar nomological explanations, such as explaining that apples fall due to the law of gravity: the key point is that nomological explanations do not necessarily have to have a temporal character.
Of course, some caution is required with this strategy. Seeking explanations for phenomena that we find surprising or conspiratorial is frequently a good way of making scientific progress, so we don't want to make explanation too cheap: answering every scientific question with `Because there's a law which makes it so,' isn't likely to lead to any new insights. But of course, the same is true with standard DN and inductive-statistical explanation - not every proposed DN or inductive-statistical explanation is interesting and informative, so we have a variety of criteria which can be applied to judge which explanations have merit. These criteria, including simplicity, unification, the absence of fine-tuning and so on, can equally well be applied to constraint-based explanations. Thus rather than simply imposing a constraint consisting of Humean mosaics in which the initial state is simple or a constraint consisting of Humean mosaics in which the cosmological constant is small, we might want to come up with some more general feature from which these constraints could be derived - for example, one could imagine deriving the past hypothesis from a more general constraint that singles out Humean mosaics which are sufficiently interesting or varied. This constraint rules out mosaics where everything is in thermal equilibrium throughout the whole of history, which does indeed entail that the initial state must have low entropy, but it may also have \emph{other} interesting consequences, and if the consequences turn out to be powerful enough we would have good reason to accept the proffered explanation. The overarching point is that holistic determinism offers us a new sort of explanatory framework: there can be good and bad explanations within that framework, and there's work to be done to establish the appropriate criteria for judging these sorts of explanations, but nonetheless this new approach is a promising route to answering questions which seem intractable within standard explanatory paradigms.
\subsection{Assessment of Theories}
Determinism has long functioned as a gold standard for an ideal scientific theory - for example, the idea that quantum mechanics might be indeterministic was accepted only begrudgingly by many physicists at the time of the theory's formulation, as witnessed by Einstein's complaint that `He (God) does not play dice.'(\cite{doi:10.1063/1.1995729}) Over the last century many attempts have been made to `complete' quantum mechanics so that it obeys Laplacean determinism after all (\cite{genovese2005research}), and indeed two of the most popular interpretations of quantum mechanics (the de Broglie-Bohm interpretation (\cite{holland1995quantum}) and the Everett interpretation (\cite{Wallace}) do satisfy Laplacean determinism.
However, if we accept that the standard definition of Laplacean determinism is to some degree a function of our practical interests, there is less justification for regarding Laplacean determinism as the ultimate goal of a scientific theory. Obviously, theories which satisfy Laplacean determinism are still desirable due to their practical utility for predicting the future, but from the point of view of the scientific realist with an interest in understanding how things \emph{really} are, theories which satisfy weak, strong or delocalised determinism may be just as well motivated. After all, all three of these notions of determinism do justice to Einstein's intuition that `He does not play dice,' in the sense that none of them allows the existence of genuine objective chances from the external, `god's-eye' point of view.
Thus this approach to determinism suggests new ways of thinking about what a good scientific theory should look like. For example, instead of coming up with theories which postulate a set of states and a set of differential equations, perhaps we should be looking more seriously at theories postulating laws which govern `all-at-once' from an external point of view. As noted in section \ref{td}, modern physics does already contain examples of such laws, but in general these examples have been arrived at by starting from a temporal evolution picture and subsequently generalizing it or reformulating it; things might look quite different if we started from the assumption that we are looking for a holistic, `all-at-once' theory and then proceeded without insisting on the existence of a time-evolution formulation of the theory. Adjusting our gold standard to more closely reflect the form of the true laws of nature is likely to be a good way to stimulate progress toward understanding those laws, so we have strong practical motivations for rethinking the gold standard.
\subsection{Free Will}
Laplace's ideas about determinism gave rise to a spirited debate over the possibility of free will in a deterministic universe which has continued to this day (\cite{Williams1980-WILFWA, Inwagen1975-INWTIO, Mickelson2019-MICTPO-41}). The possibility of holistic determinism certainly raises new questions in this debate. For example, even if you believe that we don't have free will in the context of Laplacean determinism, you might still be willing to say that we could have free will in the context of some sorts of \emph{holistic} determinism. After all, if the course of history is determined by laws which apply to the whole of history all at once, our actions are determined by other events, but also those events are partly determined by our actions, since each part of the history is optimized relative to all the other parts. So there is a degree of reciprocity which is lacking in the Laplacean context, where the present state determines our actions and our actions do not act back on the present state. This is an interesting direction of enquiry, but somewhat outside the scope of the present paper, so we will leave it as a topic for future research.
\subsection{Initial Conditions}
The dynamic production picture on which Laplacean determinism is predicated encourages us to take a particular attitude to initial conditions, regarding them as freely chosen inputs to an otherwise fully determined system. Thus in particular the Laplacean picture makes a sharp distinction between the freely chosen initial conditions and the values of any parameters which enter into the theory (e.g. the masses of fundamental particles, the gravitational constant and so on), which are normally taken to to be nomic (\cite{pittphilsci15757}). But in the constraint framework the case for this different status is significantly weaker. If the universe satisfies strong holistic determnism then both the initial conditions and the numerical values of the parameters must be determined by the laws of nature - that is, we must have state rigidity and parameter rigidity, in the terms of \cite{pittphilsci15757}. But if the universe is only weakly deterministic, it could well be the case that across the set of mosaics in the intersection of the constraints there is some variation in the initial conditions of the universe and also in the numerical value of the parameters, so drawing a mosaic from the set entails arbitrarily selecting both initial conditions and parameter values. In this context both initial conditions and parameters are simply different facets of the degrees of freedom left by the laws of nature, and thus they have precisely the same modal status. Of course, it \emph{could} also be the case that the parameters are in fact fixed by the laws of nature while the initial conditions are not, but likewise it \emph{could} be the case that the initial conditions are in fact fixed by the laws of nature while the parameters are not: until we have some specific evidence for either parameter rigidity or state rigidity, parameters and initial conditions should be afforded the same modal status.
\section{Objective Chance \label{objective}}
Given the close links between determinism and objective chance, our proposed alternative approach to determinism will also have consequences for the concept of objective chance. Famously, objective chance is a murky notion - though a number of analyses of the topic exist, including frequentist and propensity approaches, none has so far garnered universal approval and all seem to have serious problems to overcome (\cite{SEPprobability}). It is tempting to respond to this situation by simply insisting that there aren't any objective chances at the level of fundamental laws, but that strategy is hampered the fact that the current scientific evidence seems to be pointing away from Laplacean determinism. It may appear that we are faced with a dichotomic dilemma - either the world satisfies Laplacean determinism or there exist objective chances - and thus, since it doesn’t appear to be the case that quantum mechanics satisfies Laplacean determinism, it may seem that we have no choice but to accept the existence of objective chances.
But our discussion of holistic determinism demonstrates that this division of the possibilities is too rigid, for it turns out that we can postulate worlds which do not satisfy Laplacean determinism but which nonetheless satisfy holistic determinism, and therefore the probabilistic features of quantum mechanics do not force us to accept the existence of objective chance from the holistic, external point of view. There are a variety of ways in which such apparently probabilistic events might emerge within a world satisfying holistic determinism - they will appear wherever the world contains some events which depend on facts about reality that are not accessible to observers in the local region of those events - but for concreteness, in order to demonstrate how chances can emerge from a world satisfying holistic determinism we will henceforth focus on one particularly simple possibility: a law of the form `events of type A have outcome A with probability 0.8' may be translated to a law imposing a constraint of the form `the set of mosaics in which eighty percent of events of type E have the outcome A.' Evidently for the large majority of mosaics in this set, if we are in that mosaic and we observe a sufficiently large number of events of type E we will find the relative frequency of outcome A approaches 0.8, so this frequency-based law can be expected to give rise to the characteristic behaviour associated with the original probabilistic law.
The existence of these sorts of global constraints is entirely compatible with strong, weak or delocalised holistic determinism, so one could equally well say either that this approach reduces objective chance to frequency constraints, or that it simply eliminates objective chance altogether. We could even think of it as a way of reducing objective chances to subjective probabilities: the relevant events are sampled from a set of events with prespecified relative frequencies, so the objective chances attached to these events simply describe a process of sampling without replacement akin to the common textbook example of selecting a ball from a jar containing a specified mixture of black and white balls. And of course the probabilities involved in sampling without replacement can be described in terms of Bayesian credence functions with partially unknown initial conditions, since we know the initial proportions of black and white balls but not the order in which they will be drawn. Thus deriving chances from frequency constraints provides a satisfying explanation of the close conceptual and mathematical relationship between objective chance and subjective probability: objective chances behave like subjective probabilities because they really are just subjective probabilities over a large and temporally extended domain.
Frequency constraints are similar to a view that has been advanced by \cite{Roberts2009-ROBLAF} under the name of \emph{nomic frequentism}; much of Roberts' discussion is also relevant here, so we won't dwell on the points of agreement, but in appendix \ref{diff} we discuss a few points of difference. Frequency constraints are also clearly related to finite frequentism, which is which is the view that objective probabilities are by definition equal to the actual relative frequencies of the relevant sort of event across all of spacetime(\cite{SEPprobability}). But the frequency constraint approach avoids several of the main problems encountered by finite frequentism. For example, it has been objected that finite frequentism entails that probabilities can't be defined for processes which as a matter of fact occur only once or not at all (\cite{Hajek}). But this is not a problem for the frequency constraint view provided that we take a robust attitude to modality which maintains that constraints are ontologically prior to the Humean mosaic - in that case a constraint exists and is well-defined regardless of how many occurrences of the event type in question actually occur. Another common objection to finite frequentism is that it has the result that probability depends on a choice of reference class, and in the case of macroscopic everyday events this can be very nontrivial, because different descriptions of the event may suggest different natural reference classes which can lead to very different probabilities (\cite{Hajek}). However, this is not a problem for the frequency constraint view provided that we take constraints to be objective and mind-independent, because the reference class is defined within the constraint and therefore there is always a definite fact about which reference class is relevant. Of course, there is still an epistemological question about how we as observers can decide which reference class actually appears in the real underlying constraint, and it's always possible that we will get it wrong and will thus make the wrong predictions, but there is no ambiguity in the definition of the objective chance itself. \cite{Roberts2009-ROBLAF} gives more examples of the advantages of frequency constraints over finite frequentism.
In order to see if frequency constraints can give rise to chances which behave in the way we would expect, we will consider the following desiderata for an account of objective chance:
\begin{enumerate}
\item \emph{The Principal Principle:} Objective chances should satisfy the Principal Principle.
\item \emph{Probabilistic confirmation:} We should be able to confirm facts about objective chances by means of observing relative frequencies.
\item \emph{Exchangeability:} Given a sequence of events which are not causally related to one another, the chance for a given sequence should depend only on the relative frequencies in that sequence and not on the order in which the events occur.
\item \emph{Counterfactual independence:} Given a sequence of events which are not causally related to one another, the chance for the outcome of an individual event should not depend on the frequency of occurrence of other outcomes in the sequence.
\end{enumerate}
\subsection{The Principal Principle}
Do chances derived from frequency constraints satisfy the Principal Principle? Well, in accordance with the method of direct inference (\cite{McGrew}), if you know that you are in a Humean mosaic which belongs to the set of mosaics in which exactly eighty percent of As are Bs, and you have no other information, you should indeed set your credences for the next observed $A$ to be $B$ to eighty percent. Now in most real cases we will already have observed some As and therefore it is not quite the case that we are sampling at random from the entire class, so the correct probability to assign might actually be infinitesimally different from eighty percent. But provided that the class is large enough, this difference will be so small that it will make no practical difference to the way in which we use this probability in our reasoning processes and thus for all practical purposes the Principal Principle does indeed pick out the frequencies that appear in the constraints.
\subsection{Probabilistic Confirmation}
Next let us check for probabilistic confirmation: if we observe a set of events of a certain type and find that eighty percent of them have the outcome A, do we have grounds to confirm the hypothesis that we are in a Humean mosaic in which eighty percent of instances of this event type have outcome A? Of course, in general we will be able to observe only a very tiny proportion of the total number of such events occurring across the whole Humean mosaic, so we can't confirm this hypothesis by direct observation; rather we must assume that the As that we have observed are a \emph{representative sample} of the full set of As. Thus the
legitimacy of this inference depends crucially on the assumption that the As form a homogenous class which can be expected to exhibit stable relative frequencies across time - it is not the case that just any relative frequencies can be extrapolated in this way, as that would allow us to confirm virtually any hypothesis we like about future relative frequencies by simply labelling events in the right sort of way. Thus from observations of a relatively small subclass of instances of a given event type, we do not have grounds to confirm the hypothesis that eighty percent of instances of this event type have outcome A, but we \emph{do} have grounds to confirm the hypothesis that the laws of nature induce a constraint to the effect that eighty percent of As must be Bs. That is, the inference must include the hypothesis that we have identified a special class of events which appears directly in the laws of nature and that the relative frequencies we have seen are the consequence of a constraint which is induced by the laws of nature. Of course we could always be wrong about this hypothesis, but nonetheless, observing behaviour which is consistent with the hypothesis provides some degree of confirmation for it.
\subsection{Exchangeability and Counterfactual Independence }
Let us move to the final two desiderata, which are somewhat more mathematical in character. It is clear that chances obtained from frequency constraints will satisfy exchangeability, since these constraints govern only relative frequencies and not order of appearance. But what about counterfactual independence? Well, we have seen that in the frequency constraint picture, observing random events must be understood as a form of sampling without replacement, and so just as we would update our probabilities for the next event after each random draw when performing sampling without replacement, it seems we ought to do the same in the frequency constraint case: if I know that exactly 50 percent of events of a given type must have the outcome A, then if my first one hundred observations all have the outcome B I will conclude that the remaining set of events must contain one hundred more As than Bs, so I will perform the standard sort of inference we see in sampling without replacement, which involves updating my credences to reflect the fact that the next outcome is more likely to be A than B. Note that this remains true for any finite total number of events, so the conclusion does not change if I don't know the total number of events. It also doesn't matter if we have a constraint which prescribes the probability only approximately; there will always be some threshold number $n$ of Bs such that if I have observed $n$ more Bs than As, then the set of unobserved events must contain more As than Bs in order to get final frequencies which are correct to within the limits allowed by the relevant constraint.
This seems problematic, as the absence of counterfactual independence jars with many of our intuitions about probability - for example, it has the consequence that under certain circumstances the gambler's fallacy is not really a fallacy at all. Indeed Hajek argues against standard frequentism on similar grounds, pointing out that defining probabilities as relative frequencies prevents us from saying that the probability for an individual event is counterfactually independent of what happens in other similar events distant in space and time (\cite{Hajek}). So we must now ask ourselves whether we can accept an approach which fails to satisfy counterfactual independence as an acceptable analysis of objective chance. One possible response for the proponent of frequency constraints is to point out we will not get failures of counterfactual independence if the total number of instances is infinite, since sampling without replacement from an infinite set is mathematically equivalent to sampling \emph{with} replacement. And after all, we can never know for sure that the number of instances of an event type is finite, since that would require us to have illegitimate knowledge of the future, so it might be argued that we should never make updates to our probabilities which are premised on the number of instances of an event type being finite, meaning that the usual sort of inferences we see in sampling-without-replacement should not be allowed with regard to sequences of chancy events. However, although it is true that we can never know for sure that the number of instances is finite, nonetheless one can imagine cases where our best theories give us good reason to believe that this is so (for example, if we had a theory which implied that spacetime is discrete and the universe has a finite extent and time has both a start and an end), so at least in principle there are circumstances where counterfactual independence really would be violated in this picture.
Another possible response is to point out that we would only ever get violations of counterfactual independence with regard to sequences of chancy events if we had knowledge of frequency constraints that would trump the observed relative frequencies, and one might think that this could never come about. For example, in real life if we performed one hundred measurements and got the result B every time, we would probably form the working hypothesis that there is a constraint requiring that all events of this type must have outcome B, rather than supposing that exactly fifty percent of all events of this type must have outcome B and that therefore subsequent events must be more likely to have outcome A than outcome B in order to make up the total. However, probabilistic laws do not usually exist in isolation, and therefore one can imagine a case where we decide based on considerations of symmetry or coherence with the rest of our theory that the objective chance of B is 0.5 even though so far we have seen significantly more Bs than As. It would seem that under those circumstances, if we accepted the frequency constraint analysis and we had reason to believe that the total number of instances of the relevant event type across all of spacetime would be finite, we would have reason to assign a credence greater than 0.5 to the proposition that the next such event will have outcome A.
So in fact, probably the appropriate response here is to simply say that violations of counterfactual independence \emph{should} sometimes be allowed with respect to sequences of chancy events, because there is little reason to expect our intuitions to be a good guide on this point. For a start, most of our intuitions are based on probabilistic events that involve \emph{subjective} probabilities rather than objective chances of the kind that appear in the fundamental laws of nature - due to decoherence, macroscopic events do not typically depend sensitively on quantum mechanical measurement outcomes, and therefore paradigmatic probabilistic events like `rolling a die' or `flipping a coin' can be understood entirely in terms of our ignorance of the specific initial conditions. Provided that the relevant initial conditions are independent (or at least independent enough for practical purposes) these sorts of events do indeed obey counterfactual independence, which is likely the source of our intuitions on this point. Furthermore, even when we do observe events that may involve objective chances, such as the results of quantum measurements, we are presumably observing only a very small subset of the total number of events of this type, so we should not expect these observations to tell us very much about the class as a whole. If we know that exactly fifty percent of instances of a certain event type have the outcome $A$, and also that $10^{11}$ of these events occur across all of spacetime, then in principle the fact that we have observed one hundred more $B$ outcomes than $A$ outcomes does make it more likely that the next outcome will be an A, but in practice the change in probability is so small that it will not lead to any observable effects.
Therefore, the fact that counterfactual independence conflicts with our intuitions is no reason to reject chances obtained from frequency constraints, since our intuitions have been developed in entirely the wrong context of application. Indeed, counterfactual independence can be regarded as a way of expressing the expectation that objective chances should be intrinsic properties of individual entities which are spatially and temporally local; but in many ways chances actually look like properties not of individual entities but rather of large collections of entities. For example, we can't ascertain chances during a single observation, but rather we must observe a large number of events of the same type and then make an inference about what the chances are. A number of popular approaches to the analysis of objective chance - particularly the propensity interpretation, which holds that chances are intrinsic properties of individual entities akin to colours and masses (\cite{popper2005logic}) - struggle to make sense of this feature (\cite{Eagle}). But frequency constraints do justice to the apparently collective nature of objective chance by turning chances into a form of global coordination across space and time, where events are constrained to coordinate non-locally amongst themselves so as to ensure that outcomes occur with (roughly) the right frequency. \cite{Hajek} worries that `\emph{it's almost as if the frequentist believes in something like backward causation from future results to current chances,}' - and indeed, that is exactly what the proponent of frequency constraints does believe!
\section{Related Topics \label{rt}}
In this section, we discuss the relationship between holistic determinism and some other relevant research programmes.
\subsection{Superdeterminism}
Since the terms `holistic determinism' and `superdeterminism' are superficially similar, it's important to reinforce that these concepts are not identical. Broadly speaking, `superdeterminism' describes approaches to quantum mechanics which deny the existence of non-locality in quantum mechanics by rejecting the assumption of statistical independence which goes into Bell's theorem, i.e. the idea that our choice of measurement is independent of the state of the system which we are measuring (\cite{10.3389/fphy.2020.00139,hossenfelder2020superdeterminism,palmer2016invariant}). There are two typical ways to achieve this. First, given that our choice of measurement is always determined or at least influenced by facts about the past (e.g. the physical state of our brain in the time leading up to our choice) one could imagine the initial state of the universe being arranged such that our measurement choices are always correlated with the states of the systems we are measuring. Second, one could imagine there is some sort of retrocausality or `future outcome dependence' where the state of the system we are measuring is influenced by our future decisions about what measurement to perform.
It is immediately clear that neither of these approaches actually requires determinism: a \emph{probabilistic} dependency between choice of measurement and state is already enough to vitiate statistical independence and thus restore locality. Thus it is actually somewhat misleading to use the term `superdeterminism' to refer to the violation of statistical independence. The term `superdeterminism' contains the word `determinism' because Bell originally proved a theorem ruling out deterministic local hidden variable theories (\cite{Bell}), and so at that time superdeterminism was suggested as an approach to restore both locality and determinism; but a later version of Bell's theorem ruled out \emph{all} local hidden variable theories, either probabilistic or deterministic (\cite{articleEsfeld}), so in fact determinism is not really the key issue here, and modern proponents of superdeterminism are usually motivated more by the desire to preserve locality than any particular interest in preserving determinism.
That said, for the moment let us focus on superdeterministic approaches which \emph{are} also deterministic in some sense. What sort of determinism would that be? Well, if we take the first route where correlations between measurement choices and states are written into the initial state of the universe, then we have just standard Laplacean determinism (and thus also weak holistic determinism). If we take the second route involving retrocausality, the resulting theory would probably not be compatible with Laplacean determinism, because events which are determined by something in the future usually can't also be fully determined by the past; but such a theory certainly could exhibit either weak or strong holistic determinism. However, although superdeterminism can be a form of holistic determinism, the concept we have defined here is significantly more general than superdeterminism. In particular, a major motivation for our approach to was to understand what determinism could look like in a world which includes spatial and/or temporal non-locality, whereas superdeterminism is typically employed as a way of \emph{denying} the existence of either spatial or temporal non-locality, and therefore holistic determinism also accommodates realist approaches which take a very different approach from superdeterminism.
\subsection{Deterministic Probabilities}
There is a long-standing philosophical debate around the question of whether it is possible to have chances in a deterministic world (\cite{McCoyManuscript-MCCAOG-3,Glynn2010-GLYDC,GallowForthcoming-GALASG}), but this debate usually takes place against the backdrop of Laplacean determinism. Moving to a picture based on holistic determinism therefore opens up new ways in which we can have `chancy' events in a deterministic world, since we saw in section \ref{objective} that an event can be described by a non-trivial probability distribution when we condition on all of the information available to local observers, even if the event is not objectively chancy when we consider the way in which it is embedded in the complete objective modal structure of reality. This reinforces the point that `chanciness' is in many cases relative to a perspective: obviously if the outcome of a process is determined by something in the future it is for all practical purposes objectively chancy from the point of view of observers like ourselves who have directly epistemic access only to the past and present, but this does not necessarily mean that the process must be `chancy' from the point of view of the underlying modal structure.
\section{Conclusion}
We have argued that the metaphysical picture of determinism captured in Laplace's original vision is no longer fit for purpose, since it can't be usefully applied to the diverse range of laws outside the time evolution paradigm that appear in modern physics. However, we contend that even for a world governed by laws which are not time evolution laws, it is still meaningful to ask whether or not that world is deterministic. Thus in this article we have provided several generalized definitions of determinism which can be applied to a wide variety of nonstandard laws, thus rehabilitating determinism for the post-time-evolution era.
Using these definitions, we have shown that a world governed by laws outside the time evolution paradigm may fail to satisfy the standard conception of Laplacean determinism but may nonetheless exhibit a form of determinism on a global scale, where we say we have some form of `determinism' provided that we are never required to assign probability distributions over entire Humean mosaics or worlds. This has the consequence of dissolving the traditional dichotomic dilemma where we supposedly have to choose between Laplacean determinism and intrinsic randomness - a world governed by global aws may exhibit what appear to be objective chances from the point of view of local observers whilst requiring no irreducible objective chances from the external, holistic perspective. We have discussed one possible way of implementing this vision by means of constraints prescribing relative frequencies across spacetime, although frequency constraints are only one possible way in which chancy events could arise within a world satisfying holistic determinism.
Several interesting conceptual points have come up along this journey. In particular, the possibility of weak holistic determinism requires something of a conceptual shift, requiring us to distinguish between events which are chancy and events which are undetermined but nonetheless not chancy. This distinction is not entirely new - for example, it is common to hold that the initial conditions of the world are not determined by anything but also not governed by any objective probabilities - but it is not always well articulated, and often fails to be taken seriously as an option for any phenomena other than the initial conditions of the universe. Weak holistic determinism thus offers a welcome halfway house for those who find it implausible that everything in the universe is fully determined by the laws of nature, but who are also suspicious of the ill-defined concept of objective chance. We have briefly surveyed some of the consequences of this new perspective, nothing that it offers interesting insight into for long-standing debates in which determinism plays a role within both physics and philosophy.
|
\section{Introduction}
Let $K$ be a number field or a transcendence degree one function field of any characteristic, and let $T^0$ be a smooth, quasiprojective curve over $K$. Suppose we have an abelian scheme $A\to T^0$, so that every fiber $A_t$ is an abelian variety. By fixing a line bundle on $A$ which restricts to a symmetric and ample line bundle on each fiber, we may specify the N\'eron-Tate height $h_{A_t}$ on each fiber in a consistent way.
Fix a section $P:T^0\to A$ with specializations $P_t$. If we let $F=K(T^0)$ be the function field of $T$, this corresponds to an $F$ point on the generic fiber $A_{\eta}$.
Following a line of study begun by Silverman and Tate, we ask how the canonical height $h_{A_t}(P_t)$ varies as we vary the parameter $t$ along $T^0(\overline K)$. Let $T$ be the unique smooth projective curve containing $T^0$, and let $h_T(t)$ be any height on $T$ corresponding to a degree one line bundle on $T$. The general formulation of several theorems and conjectures is that
\[
h_{A_t}(P_t)=h_A(P)h_T(t)+(\text{error term}),
\]
Where $h_A$ is the N\'eron-Tate canonical height on $A_{\eta}$.
Silverman conjectured~\cite{SilvermanThesis}, and Tate proved~\cite[Main Theorem]{Tate83} that if $A$ is a family of elliptic curves, a divisor $D_P$ can be found on $T$ so that replacing $h_T$ with $h_{D_P}$ reduces the error term to $O(1)$. Call~\cite{Call87} and~\cite{Green89} prove similar results in higher dimension. A series of subsequent results of Silverman culminating in~\cite{Silverman94} further classify the nature of the bounded error function for families of elliptic curves, and work of Biesel, Holmes, and de Jong~\cite{deJong} makes similar refinements in higher dimension. More recently, DeMarco and Mavraki \cite[Theorem 1.1]{DeMM} showed how to eliminate the error term entirely for families of elliptic curves by replacing $h_T$ with a specific height constructed using adelic metrized line bundles. Here, we generalize their result to any dimension, resolving this question in full for abelian varieties.
\begin{theorem}\label{abelianvarieties}
Let $P\in A(F)$ such that $h_{A}(P)\ne 0$. The function
$$t\mapsto h_{A_t}(P_t),$$
defined for all but finitely many $t\in T$, can be extended to a height function $h_{\overline{M}}$ on all of $T$ which is given by a nef adelic line bundle $\overline{M}$ on $T$ whose underlying line bundle $M$ is ample.
\end{theorem}
This is proved by a completely different method from that of DeMarco and Mavraki. We use Yuan-Zhang's theory of adelic line bundles and vector heights~\cite{yz21} to construct a canonical adelic line bundle, essentially extending Tate's limiting method from $\mathbb{R}$-valued height functions to the underlying geometry. Once this theory is established, the existence of a N\'eron model for $A$ provides the needed global geometric setting on which to complete the argument.
\begin{remark}
By the Lang-N\'eron Theorem (see~\cite{conradtrace}), if $A_{\eta}$ has trivial $F/K$-trace, the condition that $h_A(P)\ne0$ is equivalent to $P$ being non-torsion. In general, it's equivalent to $P$ being outside a torsion coset of Tr$_{F/K}(A_{\eta})$.
\end{remark}
With a metrized line bundle giving $h_{A_t}(P_t)$ exactly, one could use equidistribution arguments to show, for example, that if $P,Q\in A(F)$ are two points whose heights are both non-zero and there exists a sequence $\{t_n\}\subset T(\overline \mathbb{Q})$ such that both $h_{A_{t_n}}(P_{t_n})$ and $h_{A_{t_n}}(Q_{t_n})$ tend to zero as $n\to\infty$, then $P$ and $Q$ specialize to torsion points at the exact same places. This is likely vacuous, however, due to the following conjecture of Zhang.
\begin{conj}(\cite{ZhangConj})
Suppose $A\to T^0$ is a non-isotrivial family of abelian varieties as above, with $K$ a number field, and that $A_{\eta}$ is simple of dimension at least two. For each non-torsion section $P:T^0\to A$ defined over $\overline\mathbb{Q}$, there exists an $\epsilon>0$ such that
\[
\{t\in T^0(\overline\mathbb{Q}): h_{A_t}(P_t)\le\epsilon\}
\]
is finite.
\end{conj}
Instead, using Theorem~\ref{abelianvarieties}, we can restate Zhang's conjecture as a positivity statement for self-intersections.
\begin{cor}\label{Bigness}
Let $\overline{M}$ be the metrized line bundle defined in Theorem~\ref{abelianvarieties} so that $h_{A_t}(P_t)=h_{\overline{M}}(t)$. Then the conclusion of Zhang's conjecture is equivalent to
\[
\overline{M}^2>0.
\]
\end{cor}
The author is extremely grateful to Xinyi Yuan, Laura DeMarco, Niki Myrto Mavraki, and Tom Tucker for several ideas and insightful discussion while preparing this paper.
\section{Vector-valued heights}
We begin with an overview of adelic line bundles and vector-valued height functions, in so much as is needed here. For a more complete treatment, see~\cite[Ch. 2]{yz21} and ~\cite[Sec. 2]{Carney2} for more development of the function field setting.
We will work in two parallel settings The first is for varieties over a field, in which all constructions are purely geometric, and which we will call the \emph{geometric case}. The second is for varieties over $\Spec \Z$, which we will call the \emph{arithmetic case}. Note that when $K$ is a number field, we will require both settings, as the height on $A_{\eta}$ is geometric, while that on each fiber is arithmetic.
\textbf{Geometric case:} Fix any field $k$, and let $\mathcal{U}$ be a quasi-projective $k$ variety. A \emph{projective model} for $\mathcal{U}$ is a projective variety $\mathcal{X}$ along with an open immersion $\mathcal{U}\hookrightarrow\mathcal{X}$ defined over $k$. We write $\Div(\mathcal{X})$ for the group of Cartier divisors on $\mathcal{X}$ and $\Prin(\mathcal{X})$ for the subgroup of principal Cartier divisors.
\textbf{Arithmetic case:} Let $\mathcal{U}\to\Spec \Z$ be a quasi-projective arithmetic variety. In this setting, a \emph{projective model} for $\mathcal{U}$ is a projective arithmetic variety $\mathcal{X}\to\Spec\Z$ with an open embedding $\mathcal{U}\to\mathcal{X}$ over $\Spec\Z$. Write $\widehat\Div(\mathcal{X})$ for the group of arithmetic divisors on $\mathcal{X}$, and $\widehat\Prin(\mathcal{X})$ for the subgroup of principal arithmetic divisors.
In both cases the projective models for $\mathcal{U}$ form an inverse system via dominating morphisms. Using pullbacks on that system, we define the following, first in the geometric case.
If $\mathcal{U}\to\Spec k$ is a quasi-projective variety,
$$\widehat\Div(\mathcal{U}/k)_{\Mod}:=\lim_{\mathcal{U}\hookrightarrow\mathcal{X}}\Div(\mathcal{X})_{\mathbb{Q}}\quad\text{and}\quad\widehat\Prin(\mathcal{U}/k)_{\Mod}:=\lim_{\mathcal{U}\hookrightarrow\mathcal{X}}\Prin(\mathcal{X})_{\mathbb{Q}}.$$
An element of $\widehat\Div(\mathcal{U}/k)_{\Mod}$ is called \emph{effective} if it comes from an effective divisor in some $\Div(\mathcal{X})_\mathbb{Q}$.
If $\mathcal{U}\to\Spec\Z$ is a quasi-projective arithmetic variety, we define
$$\widehat\Div(\mathcal{U}/\Z)_{\Mod}:=\lim_{\mathcal{U}\hookrightarrow\mathcal{X}}\widehat\Div(\mathcal{X})_{\mathbb{Q}}\quad\text{and}\quad
\widehat\Prin(\mathcal{U}/\Z)_{\Mod}:=\lim_{\mathcal{U}\hookrightarrow\mathcal{X}}\widehat\Prin(\mathcal{X})_{\mathbb{Q}}.$$
An element of $\widehat\Div(\mathcal{U}/\Z)_{\Mod}$ is called \emph{effective} if it comes from an effective arithmetic divisor in some $\widehat\Div(\mathcal{X})_\mathbb{Q}$, meaning that the finite part of the divisor is effective, and the Green's function is non-negative.
Next we define a topology on both of these groups, stemming from effectivity. Since the argument is identical in both the arithmetic and geometric cases we provide it only for the arithmetic case.
Let $\mathcal{U}$ be a quasi-projective arithmetic variety. Effectivity provides a partial ordering on $\widehat\Div(\mathcal{U}/\Z)_{\Mod}$. Fix some projective model $\mathcal{X}_0$ and a strictly effective arithmetic divisor $\overline\mathcal{D}_0$ on $\mathcal{X}_0$ such that the support of $\mathcal{D}_0$ is equal to $\mathcal{X}_0\backslash\mathcal{U}$. We call such $\mathcal{D}_0$ a \emph{boundary divisor}. For $\epsilon\in\mathbb{Q}_{>0}$, define a basis of epsilon balls around $0$ by
$$B(\epsilon,0):=\left\{\overline\mathcal{E}\in\widehat\Div(\mathcal{U}/\Z)_{\Mod}:\epsilon\overline\mathcal{D}_0\pm\overline\mathcal{E}\text{ are both effective}\right\}.$$
Via translation, this defines a topology on all of $\widehat\Div(\mathcal{U})_{\Mod}$, and it is easy to check that this topology does not depend on the choice of $\overline{\mathcal{D}}_0$.
We now define $\widehat\Div(\mathcal{U}/\Z)$ to be the completion of $\widehat\Div(\mathcal{U}/\Z)_{\Mod}$ with respect to this topology. $\widehat\Div(\mathcal{U}/k)$ is defined identically, except without a Green's function attached to each divisor. Finally, define the groups of \emph{adelic line bundles} in each setting:
$$\widehat\Pic(\mathcal{U}/\Z):=\widehat\Div(\mathcal{U}/\Z)/\widehat\Prin(\mathcal{U}/\Z)_{\Mod},$$
$$\widehat\Pic(\mathcal{U}/k):=\widehat\Div(\mathcal{U}/k)/\widehat\Prin(\mathcal{U}/k)_{\Mod}.$$
To further justify notating these as Picard groups, elements of $\widehat\Pic(\mathcal{U}/k)$ (resp. $\widehat\Pic(\mathcal{U}/\Z)$) can be represented by sequences $\{\mathcal{X}_i,\psi_i,\mathcal{L}_i,\ell_i\}_{i\ge1}$ (resp. $\{\mathcal{X}_i,\psi_i,\overline{\mathcal{L}}_i,\ell_i\}_{i\ge1}$) where $\mathcal{X}_i$ is a projective model for $\mathcal{U}$ with a morphism $\psi_i:\mathcal{X}_i\to\mathcal{X}_1$, where $\mathcal{L}_i$ (resp. $\overline{\mathcal{L}}_i$) is a $\mathbb{Q}$-line bundle (resp. Hermitian $\mathbb{Q}$-line bundle) on $\mathcal{X}$, and $\ell_i$ is a rational section of $\mathcal{L}_i\otimes\psi^*\mathcal{L}_1^{-1}$ with support contained in $\mathcal{X}_i\backslash\mathcal{U}$.
The equality of this representation with the definitions above is shown in~\cite[Lemma 2.5.1]{yz21}.
We will typically just write $\overline{\mathcal{L}}=\{\mathcal{X}_i,\mathcal{L}_i\}\in\widehat\Pic(\mathcal{U}/k)$ or $\overline{\mathcal{L}}=\{\mathcal{X}_i,\overline{\mathcal{L}}_i\}\in\widehat\Pic(\mathcal{U}/\Z)$. From the conditions on $\ell_i$, we get a well-defined restriction map $\overline{\mathcal{L}}\mapsto\mathcal{L}\in\Pic(\mathcal{U})_{\mathbb{Q}}$.
Such a sequence is \emph{Cauchy} provided that $\{\div(\ell_i)\}$ (resp. $\{\widehat\div(\ell_i)\}$) is Cauchy under the topology defined above. A sequence converges to zero if there exists a sequence of rational sections $s_i$ of $\mathcal{L}_i$ such that $\ell_i=s_i\otimes\psi_i^*s_1^{-1}$, and such that $\{\div(s_i)\}$ (resp. $\{\widehat\div(s_i)\}$) is itself Cauchy.
We call an adelic line bundle \emph{nef} if it is isomorphic to a sequence where every (Hermitian) line bundle is nef, and we call an adelic line bundle \emph{integrable} if it can be written as the difference of two nef ones. Denote the cones of nef elements
\[
\widehat\Pic(\mathcal{U}/k)_{\Nef} \quad\text{ and }\quad \widehat\Pic(\mathcal{U}/\Z)_{\Nef}
\]
and the subgroups of integrable elements
\[
\widehat\Pic(\mathcal{U}/k)_{\Int} \quad \text{ and }\quad \widehat\Pic(\mathcal{U}/\Z)_{\Int}.
\]
\subsection{The relative setting}
From here on out, we unify the notation for the base $k$ and base $\Z$ settings. Let $b$ be either $\Spec\Z$ or $\Spec k$. Let $K$ be a field which is finitely generated over $\mathbb{Q}$, if $b=\Spec\Z$, or finitely generated over $k$, when $b=\Spec k$.
An \emph{open model} for $K$ is a quasi-projective $b$-variety $\mathcal{V}$ with function field $K$.
Next, let $X$ be a quasi-projective variety over $K$. An open model for $X/K$ consists of an open model $\mathcal{V}$ for $K$, together with a quasi-projective and flat morphism $\mathcal{U}\to\mathcal{V}$ whose generic fiber is $X\to \Spec K$.
The open models for $X$ form an inverse system via inclusion. Taking the limit over this system, we define
\[
\widehat\Pic(X/b):=\lim_{\mathcal{U}\to\mathcal{V}}\widehat\Pic(\mathcal{U}/b).
\]
When $X=\Spec K$, we will simply write $\widehat\Pic(K/b)$. Restricting these limits to the nef and integrable elements, we define the nef cone and integrable subgroup
\[
\widehat\Pic(X/b)_{\Nef}\subset\widehat\Pic(X/b)_{\Int}\subset\widehat\Pic(X/b).
\]
We will write elements as $\overline L\in\widehat\Pic(X/b)$.
Given such $\overline L=\{\mathcal{X}_i,\overline{\mathcal{L}}_i\}_{i\ge0}\in\widehat\Pic(X/b)$, we can restrict via base change to
$$\overline L_K:=\{\mathcal{X}_{i,K},\mathcal{L}_{i,K}\}\in\widehat\Pic(X/K),$$
as each $\mathcal{X}_{i,K}$ is a projective model for $X$. Since, by construction, all $\mathcal{L}_{i,K}$ agree on $X\subset\mathcal{X}_{i,K}$, we have a well defined restriction to $X$ which we denote simply $L\in Pic(X)_{\mathbb{Q}}$.
If $X^0$ is a smooth, quasi-projective $K$ curve, since $X^0$ has a unique projective model $X$ over $K$,
\begin{equation}\label{curverestriction}
\widehat\Pic(X^0/K)=\Pic(X)_{\mathbb{R}}:=\Pic(X)\otimes_{\Z}\mathbb{R}.
\end{equation}
When $K$ is a number field or a function field of transcendence degree one over $k$ and $X$ is a projective $K$-variety, $\widehat\Pic(X/b)$ corresponds to the group of adelic metrized line bundles, as defined by Zhang~\cite{Z95}. There also exists an analytic theory of adelic line bundles over larger fields, for example~\cite[Sec. 3]{yz21},~\cite[Sec. 2]{Carney2}, but we will not require it here.
\subsection{Pullbacks}
Given a map $f:Y\to X$, since a priori the models defining $\widehat\Pic(Y/b)$ and $\widehat\Pic(X/b)$ may not be compatible, it is not immediately apparent that we can define pullbacks of adelic line bundles. We rectify that.
\begin{lemma}\label{pullbacklemma}
(C.f.~\cite[Sec. 2.5.5]{yz21})
Let $f:Y\to X$ a morphism of quasi-projective varieties which is flat over $b$.
Then there exists a well-defined pullback morphism
$$f^*:\widehat\Pic(X/b)\to\widehat\Pic(Y/b).$$
\end{lemma}
\begin{proof}
Assume for simplicity of notation that $b=\Spec\Z$; the only difference when $b=\Spec k$ is the lack of Hermitian metrics. Let $\overline L\in\widehat\Pic(X/b)$ be represented by a Cauchy sequence $\{\mathcal{X}_i,\overline{\mathcal{L}}_i\}$.
By moving to smaller open subvarieties as needed, we may extend $f$ to a morphism $f:\mathcal{W}\to\mathcal{U}$ of open models for $X$ and $Y$, and it suffices to prove that there exists a well defined pullback $\widehat\Pic(\mathcal{U}/b)\to\widehat\Pic(\mathcal{W}/b)$.
Let $\mathcal{Y}$ be a projective model for $\mathcal{W}$.
By Raynaud's flattening theorem~\cite{raynaudflattening}, by blowing up along $\mathcal{Y}\backslash\mathcal{W}$, we may assume that $f:\mathcal{W}\to\mathcal{X}_i$ extends to a flat morphism $f_i:\mathcal{Y}_i\to\mathcal{X}_i$. Replacing $\mathcal{Y}_i$ with a morphism which dominates it, we may further assume that there is a morphism $\psi_i:\mathcal{Y}_i\to\mathcal{Y}_0$ for every $i$, which is the identity on $\mathcal{W}$.
We now have a compatible system of models on which to pull back $\overline L$. Define
\[
f^*\overline L:=\left\{\mathcal{Y}_i,\psi_i,f_i^*\overline{\mathcal{L}}_i,f^*\ell_i\right\}_{i\ge0}.
\]
Since effective Cartier divisors pull back to effective Cartier divisors under both flat and dominant maps, and the topology on $\widehat\Pic(\mathcal{W}/b)$ does not depend on the choice of boundary divisor, the topology on $\widehat\Pic(\mathcal{U}/b)$ induces that on $\widehat\Pic(\mathcal{W}/b)$ via pullback of any boundary divisor on $\mathcal{X}_0$, and the sequence defining $f^*\overline L$ is Cauchy with respect to this topology.
\end{proof}
\section{Canonical heights}
\begin{definition}
Let $X$ be a quasi-projective $K$ variety, with $K$ and $b$ as in the previous section. Let $\overline L\in\widehat\Pic(X/b)_{\Int}$, and suppose that $L\in\Pic(X)$ is ample. Let $P\in X(K)$. We define the \emph{vector valued $\overline L$-height} of $P$ to be
$$\mathfrak h_{\overline L}(P):=P^*\overline L\in\widehat\Pic(K/b)_{\Int}.$$
If $P$ is defined over some finite extension $K'$, we compose with the map $\widehat\Pic(K'/b)_{\Nef}\to\widehat\Pic(K/b)_{\Nef}$ induced by the norm functor $N_{K'/K}$ and scale the result by $1/[K':K]$, so that this is a relative height which does not depend on the choice of $K'$. For simplicity, we will usually omit this formalism and assume without loss of generality that $P$ is defined over $K$.
\end{definition}
When $K$ is a number field or a transcendence degree one function field over $k$, taking the limit of the arithmetic degrees produces a Weil height~\cite{Z95}, which we notate
\[
h_{\overline L}(P):=\widehat\deg \left(\mathfrak h_{\overline L}(P)\right),
\]
and thus vector-valued heights extend the theory of $\mathbb{R}$-valued heights.
\begin{lemma}\label{tateslimiting}
(C.f.~\cite[Sec. 6.1.1]{yz21},~\cite[Sec. 4.1]{Carney2})
Let $f:X\to X$ be an endomorphism of projective varieties over $K$ with a \emph{polarization}, i.e. an ample $\mathbb{Q}$-line bundle $L\in\Pic(X)_{\mathbb{Q}}$ and a rational number $q>1$ such that $f^*L=qL$. Then there exists a nef adelic line bundle $\overline L_f\in\widehat\Pic(X/b)_{\Nef}$ extending $L$ such that
$$f^*\overline L_f=q\overline L_f.$$
In particular, there exists a canonical height $\mathfrak h_f:=\mathfrak h_{\overline L_f}$ such that for all $P\in X(\overline K)$,
$$\mathfrak h_f(f(P))=q\mathfrak h_f(P).$$
\end{lemma}
\begin{proof}
We again prove this only in the base $\Z$ case, as the base $k$ case is identical besides the lack of Hermitian metrics. To start, we can find an open model $\mathcal{U}\to\mathcal{V}$ over $b$ for $X\to\Spec K$ on which $f$ extends to a $\mathcal{V}$-morphism $f:\mathcal{U}\to\mathcal{U}$, and over which there exists an extension $\mathcal{L}\in\Pic(\mathcal{U})_{\mathbb{Q}}$ of $L$ such that $f^*\mathcal{L}=q\mathcal{L}$. Fix a projective model $\mathcal{X}_0\to \mathcal{B}$ for $\mathcal{U}\to\mathcal{V}$, and a nef extension $\overline{\mathcal{L}}_0\in\widehat\Pic(\mathcal{X}_0)_{\mathbb{Q}}$ of $\mathcal{L}$. By shrinking $\mathcal{V}$ if needed, we may assume $\mathcal{B}\backslash\mathcal{V}$ is an effective Cartier divisor.
Write $f^m$ to mean the $m$-th iterate of $f$. For each $m\ge1$, define $f_m:\mathcal{X}_m\to\mathcal{X}_0$ to be the normalization of the composition $X\xrightarrow{f^m}X\hookrightarrow\mathcal{X}_0$, and define
$$\overline{\mathcal{L}}_m:=\frac1{q^m}f_m^*\overline{\mathcal{L}}_0.$$
For each $m\ge0$, let $\mathcal{X}_m'$ be a projective model for $\mathcal{U}$ which dominates both $\mathcal{X}_m$ and $\mathcal{X}_{m+1}$ via $\phi_m:\mathcal{X}_m'\to\mathcal{X}_m$ and $\rho_m:\mathcal{X}_m'\to\mathcal{X}_{m+1}$, both of which restrict to the identity on $\mathcal{U}$. Further assume that for $m\ge1$, there exists a morphism $\tau_m:\mathcal{X}_m'\to\mathcal{X}_0'$ which commutes with $\phi_m,$ $\rho_m$ and $f_m$. By construction $\tau_m$ extends $f^m$ on $\mathcal{U}$.
Since $\phi_0^*\overline{\mathcal{L}}_0-\rho_0^*\overline{\mathcal{L}}_1$ is trivial on $\mathcal{U}$, we can find a boundary divisor $\overline{\mathcal{D}}_0$ supported on $\mathcal{B}\backslash\mathcal{V}$ such that for the topology defined by $\pi^*\overline{\mathcal{D}}_0$, we have $\phi_0^*\overline{\mathcal{L}}_0-\rho_0^*\overline{\mathcal{L}}_1\in B(1,0)$. Now
\begin{equation*}\label{tateslimitingequation}
\phi_m^*\overline{\mathcal{L}}_m-\rho_m^*\overline{\mathcal{L}}_{m+1}=q^{-m}\tau_m^*\left(\phi_0^*\overline{\mathcal{L}}_0-\rho_0^*\overline{\mathcal{L}}_1\right).
\end{equation*}
and thus
\[
\phi_m^*\overline{\mathcal{L}}_m-\rho_m^*\overline{\mathcal{L}}_{m+1}\in B\left(q^{-m},0\right).
\]
We get a Cauchy sequence
\[
\overline L_f:=\left\{\mathcal{X}_m,\overline{\mathcal{L}}_m\right\}_{m\ge0}\in\widehat\Pic(X/b)_{\Nef},
\]
which is easily seen to satisfy the desired conditions.
\end{proof}
\section{Proof of Theorem~\ref{abelianvarieties}}
Let $K$ be a number field and $b=\Spec\Z$, or let $K$ be a transcendence degree one function field with constant field $k$, and $b=\Spec k$. Let $T^0$ be a smooth quasiprojective curve over $K$, and let $F=K(T^0)$ be its function field. Let $A$ be a family of abelian varieties over $T^0$, and let $T$ be the unique smooth projective curve with function field $F$.
Write $\mathfrak h_{A}=\mathfrak h_{[2]}$ for the $\widehat\Pic(F/K)_{\Int}$-valued canonical height on the generic fiber $A_{\eta}$, and on each fiber $A_t$ over $t\in T^0(\overline K)$, write $\mathfrak h_{A_t}$ for the $\widehat\Pic(K/b)_{\Int}-$valued canonical height on $A_t$. As defined in the previous section, the usual $\mathbb{R}$-valued N\'eron-Tate canonical heights are
$$h_A(P)=\widehat\deg\left(\mathfrak h_A(P)\right),\qquad h_{A_t}(P_t)=\widehat\deg\left(\mathfrak h_{A_t}(P_t)\right)$$
We show that if $h_{A}(P)\ne0$, there exists a nef adelic line bundle $\overline{M}\in\widehat\Pic(K/b)_{\Nef}$ such that on the $\overline K$-points of $T$ over which $A$ has good reduction, the function
$$t\mapsto \mathfrak h_{A_t}(P_t)$$
is equal to a vector-valued height function $\mathfrak h_{\overline{M}}$, defined on all of $T$.
\begin{proof}
Let $\mathcal A\to T$ be the N\'eron model for $A_{\eta}$. This is a quasiprojective $K$-variety. Replacing $F$, and thus $T$, with a finite extension will not alter the result, as the heights involved are invariant under finite extensions. Thus, we may assume that $\mathcal A\to T$ has semiabelian reduction~\cite[Ch. 7.4, Thm 1]{RaynaudBook}. We also write $P:T\to\mathcal A$ for the unique section extending $P\in A(F)$, by the N\'eron mapping property.
Since $\mathcal A$ has abelian reduction at all but finitely many fibers, and the bad fibers have finitely many components, by replacing the point $P$ with $nP$ if needed, we may assume that the section $P$ is contained in the connected component of the identity, which we denote $\mathcal A^{\circ}$. This replacement scales the canonical height on both $A$ and on each abelian fiber $A_t=\mathcal A_t$ by $n^2$, and thus does not affect the result.
Let $\mathcal{X}\to b$ be a projective model for $\mathcal{A}^{\circ}$ over $b$. Fix a symmetric and ample line bundle $L$ on $A_{\eta}$, and extend this to a nef line bundle $\mathcal{L}$ on $\mathcal{X}$. By applying the normalization construction from the proof of Theorem~\ref{tateslimiting}, we can find a second projective model $\mathcal{X}'\to b$ with a morphism
\[
[-1]:\mathcal{X}'\to\mathcal{X}
\]
extending the automorphism $[-1]$ on $\mathcal{A}^{\circ}$. Fix a third model $\mathcal{X}_0$ dominating both $\mathcal{X}$ and $\mathcal{X}'$. We can then define a nef line bundle
\[
\mathcal{L}_0=\frac12\left(\mathcal{L}+[-1]^*\mathcal{L}\right)\in\Pic(\mathcal{X}_0)_{\mathbb{Q}},
\]
where we use the implied pullbacks from $\mathcal{X}$ and $\mathcal{X}'$ to $\mathcal{X}_0$.
We adapt an argument of Green~\cite{Green89}. Consider the difference $[n]^*\mathcal{L}_0|_{\mathcal{A}^{\circ}}-n^2\mathcal{L}_0|_{\mathcal{A}^{\circ}}$. By the theorem of the cube, this becomes trivial when restricted to $A_{\eta}=\mathcal A_{\eta}$. Then, since $\mathcal{A}^{\circ}$ is smooth and has integral fibers over $T$, this difference is equal to the pullback under $\mathcal{A}^{\circ}\to T$ of a line bundle on $T$, by~\cite[Cor. 21.4.13 of Ch. 4, Errata et Addenda]{EGA4}. Finally, pulling back by the identity section $E:T\to\mathcal{A}^{\circ}$, we see that $[n]^*\mathcal{L}_0|_{\mathcal{A}^{\circ}}-n^2\mathcal{L}_0|_{\mathcal{A}^{\circ}}$ must be trivial. Thus,
\[
[n]^*\mathcal{L}_0|_{\mathcal{A}^{\circ}}=n^2\mathcal{L}_0|_{\mathcal{A}^{\circ}}.
\]
\begin{remark}
For a different version of the above argument extending a polarization to $\mathcal{A}^{\circ}$, one can use Moret-Bailly's extension of cubical structure,~\cite[Thm. 3.5]{MoretBailly}, as in~\cite[Sec. 4]{Kunnemann}.
\end{remark}
Next, apply Lemma~\ref{tateslimiting} to extend $L$ to a nef adelic line bundle $\overline L=\{\mathcal{X}_m,\overline{\mathcal{L}}_m\}\in\widehat\Pic(A/b)_{\Nef}$ such that
$$[2]^*\overline L=4\overline L.$$
In general, the model adelic line bundles $\overline{\mathcal{L}}_m$ may all differ along a boundary divisor. Using the above construction extending the polarization to $\mathcal{A}^{\circ}$, however, the line bundles $\mathcal{L}_{m,K}$ can be made to all agree on $\mathcal{A}^{\circ}$. Since the section $P$ lands fully within $\mathcal{A}^{\circ}$, we have a well-defined pullback
\[
\overline{M}=P^*\overline L\in\widehat\Pic(T/b)_{\Int},
\]
since each $P^*(\mathcal{L}_{m,K})$ is the same line bundle on $T$.
By construction, $\mathfrak h_{\overline{M}}(t)=t^*\overline{M}=P_t^*\overline L_t=\mathfrak h_{A_t}(P_t)$ for all $t\in T^0(\overline K)$. Since $T$ is a curve over $K$, the $\mathbb{R}$-valued N\'eron-Tate height of $P$ is simply $\deg(M)=\deg((P^*\overline L)_K)\in\mathbb{Q}$. Since $P$ does not have height zero, we conclude that $M$ is ample, and that $\overline{M}$ is nef, as the pullback of a nef adelic line bundle.
\end{proof}
\begin{remark}
Without the N\'eron model, which exists for abelian varieties, but not in general for dynamical systems (see for example~\cite{Hsia96}), the above construction would instead produce only $\overline{M}\in\widehat\Pic(T^0/b)$. Since $T^0$ has a unique projective closure, this gives an $\mathbb{R}$-line bundle $M\in\Pic(T)_{\mathbb{R}}$, but it remains an open question whether the adelic structure extends over all of $T$, i.e. whether the metrics defined by each $\overline{\mathcal{L}}_m$ converge to an adelic metric on $T$.
\end{remark}
\subsection{Proof of Corollary~\ref{Bigness}}
Since $\overline{M}$ is nef, we know that $\overline{M}^2\ge0$. Define
\[
e_1:=\sup_{\text{open }U\subset T}\inf_{t\in U(\overline K)}h_{\overline{M}}(t).
\]
It can easily be seen that $e_1\ge0$, and is positive if and only if Zhang's conjecture holds.
By Zhang's essential inequalities~\cite[Thm 5.2]{Zhang95a},
\[
e_1\ge\frac{\overline{M}^2}{\deg M}\ge\frac{e_1}2.
\]
From this, the equivalence of the restatement of the conjecture in terms of the bigness of $\overline{M}$ is immediate.
|
\section{Introduction}
J1415+1320, a BL Lac has alluded to the astronomers for a very long time owing to its controversial properties. Some of these properties include (a) apparent yet rare association of the active galactic nuclei (AGN) with an optical spiral host~\cite{McHardy1994}, (b) detection of a counter jet in a blazar type AGN~\cite{Perlman1994}, (c) its association to the class of compact symmetric objects as well as blazar \cite{Perlman1996}. Some of these controversies were addressed in \cite{Readhead2021} and conclude that J1415+1320 is actually a background object in the redshift range 0.247 $<$ z $<$ 0.5 and is not associated with the previously known spiral host.\\
The radio VLBI structure of J1415+1320 shows a two-sided, bent core-jet structure of size $\sim$ 110 mas ($<$ 1 kpc). The core has an inverted spectrum ($\alpha$ $>$ 1) at frequencies between 1.4 $-$ 15 GHz \cite{Perlman1996}. At frequencies above 15 GHz where the core emission dominates \cite{Readhead2021}, the core has a flat spectrum ($\alpha$ $\sim$ 0.001) (estimated using 15 and 37 GHz flux densities). The spectrum of the jet and the counter-jet are generally steep ($\alpha$ $<$ $-$1) but varies from being flat at the knots to steep in the more diffused region \cite{Perlman1996}. The total flux density of the source at 1.4 GHz is $\sim$ 1 Jy, dominated by emission from the jets, and increases to 8 Jy at 80 MHz \cite{Lister2018}. \\
The source exhibits strong variability in its radio light curve \cite{Vedantham2017}. Variability in an AGN reveals crucial information on the size, structure, and dynamics of the radiating source down to scales that otherwise need extremely long baseline interferometers. While variability in a source is a complicated feature that is not yet fully understood, several possible mechanisms were discussed in the literature broadly categorized into intrinsic and extrinsic phenomena. Intrinsic variability occurs due to, but not limited to, $-$ (a) shocks waves forming and propagating relativistically along the jets \cite{Marscher1985}, (b) magnetohydrodynamic instabilities in the jet \cite{Marscher2014}, (c) variation in relativistic beaming as a result of viewing angle change in a twisted/bent jet \cite{Raiteri2017} (d) magnetic reconnection in turbulent jets \cite{Werner2016}. Variability is also observed owing to extrinsic phenomena such as refractive interstellar scintillation caused by large-scale irregularities in the interstellar medium (ISM) \cite{Quirrenbach1992}.\\
An interesting aspect of the radio light curve of J1415+1320 is that it displays a hitherto unrecognized form of variability, referred to as symmetric achromatic variability (SAV) \cite{Vedantham2017}. This variability is seen as a U-dip feature in the light curve, and is time-symmetric and achromatic over 15 to 234~GHz. This rare variability is possibly due to gravitational millilensing of the compact core emission in J1415+1320 by an intervening $10^2-10^5~M_{\odot}$ mass condensates \cite{Vedantham2017}. \\
So far most of the variability studies of J1415+1320 were made at frequencies $\ge$ 2.4~GHz. In this letter, we present long-term (1989.8 to 2017), low frequency (327~MHz) flux density observations of J1415+1320 taken using Ooty Radio Telescope (ORT), operated by the Radio Astronomy Centre, Tata Institute of Fundamental Research, India \cite{Swarup1971}. The details of observations and data reduction are given in Section~\ref{sec:data}. We report radio flux density variation of J1415+1320 at 327~MHz at epochs 2007.6 and 2008.6. Further, we use the data to investigate the achromatic nature of two previously reported SAV events down to 327~MHz, which is discussed in Section~\ref{sec:results}. The origin of the reported flux density variation is discussed in Section~\ref{sec:discussion} and our main conclusions are given in Section~\ref{sec:conclusions}.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{LC_allfreq.jpg}
\caption{Radio light curve of J1415+1320 at different frequencies taken from the literature together with the ORT data. The data was compiled from the following $-$ Owens Valley Radio Observatory (OVRO) \cite{Richards2014}, Green Bank interferometer (GBI) \cite{Lazio2001}, submillimetre common-user bolometer array (SCUBA) \cite{Jenness2010}, very long baseleine array (VLBA) \cite{Lister2009}, University of Michigan Radio Astronomy Observatory (UMRAO) \cite{Aller2003}.}
\label{fig:LC_allfreq}
\end{figure}
\section{Data}
\label{sec:data}
The data presented here comes from the extensive interplanetary scintillation (IPS) observations made with the ORT during the period 1989$-$2018. About 100 compact radio sources of angular size less than 250~mas were observed per day at 327~MHz as part of the IPS campaign. The main purpose of these IPS observations was to investigate the three-dimensional distributions of the solar wind density, turbulence, and speed at short time intervals of a few days as well as at different phases of solar activity \cite{Manoharan2012,Manoharan2017}. The IPS measurements on each source were made within the solar elongation range of about $\varepsilon \thickapprox \pm$ 60 degree with respect to the Sun, which spanned over an observing period of about 4 months. Typically a source was observed for about 2–3 minutes and the same source was likely observed more than once in a day at different hour-angles. Each observing session also included observations of several flux density calibrators distributed almost uniformly between the start and end of the session. \\
In this paper, the above discussed large IPS database was employed to study the variation of the total flux density of J1415+1320 over many years. The data for J1415+1320 and the control source B1345+125 are made available online in VizieR.
\section{Results}
\label{sec:results}
Figure \ref{fig:LC_allfreq} shows the radio light curves of J1415+1320 spanning nearly 40 years (1978$-$2018). The data at several frequencies ranging from 2.4 to 350~GHz was compiled from various monitoring programs found in the literature \cite{Lazio2001,Aller2003,Lister2009,Jenness2010,Richards2014}. The source exhibits variability at all frequencies. Strong variability is seen at higher frequencies compared to lower frequencies. The variability features seen at 15~GHz are followed at 8 and 5~GHz albeit at a lower level. Further, for the 15~GHz waveband, the strength of variability generally decreases over time.\\
We have included the 327~MHz radio light curve over the period 1989.8 to 2017 in Figure \ref{fig:LC_allfreq} for comparison. The data used for the 327~MHz light curve contains 1953 individual observations, which are obtained after excluding observations taken at small solar elongations (i.e., $\varepsilon \leq$ 5 degree) to avoid confusion caused by the telescope side lobe pointing at the Sun. Each data point represents the average of nearly 10 to 15 consecutive observations taken around an epoch (a few days). For years 1989, 1998, 1999, and 2002, the available number of observations is limited to only about 10 and their average is plotted. \\
In Figure \ref{fig:327MHz_LC}, we reproduce the 327~MHz light curve of J1415+1320 along with the flux density measurements of the control source B1345+125. The control source B1345+125 is a compact ($<$ 100 mas) steep spectrum source \cite{Lister2003}, monitored during a similar time period of observation of J1415+1320. The angular distance of the control source from J1415+1320 is about 10 deg. Both sources display strong IPS (scintillation index of the order of unity \cite{Manoharan1995, Manoharan2012}) and so the long-term epoch-to-epoch variations in both sources are dominated by scintillation. In other words, the normalized $\chi^2$ variability test as defined in \cite{Kesteven1976} gives a value close to unity when applied to the two datasets. The flux density of J1415+1320 interestingly shows a significant increase at epochs 2007.6 and 2008.6 (see Fig \ref{fig:327MHz_LC}); we refer to these flares as F1 and F2 respectively. The mean flux density of J1415+1320, estimated after excluding the data from the two epochs 2007.6 and 2008.6, is 2.70~Jy and the standard deviation $\sigma$, of flux density variation is 0.23~Jy. During flare F1, the flux density increased rapidly from 2.66~Jy to 3.99~Jy, and during F2, the flux density increased from 2.9~Jy to 4.19~Jy; an increase of $\sim$ 45\%. No variation in flux density is observed on the control source B1345+125; the mean and standard deviation for the control source are 7.6~Jy and 0.5~Jy respectively. The absence of flux density variation in the control source around epochs 2007.6 and 2008.6 rules out any instrumental effects or IPS causing the variation in the flux density of J1415+1320. \\
The maximum flux densities during flares F1 and F2 are 3.85~Jy (17$\sigma$) and 4.19~Jy (18$\sigma$) respectively. We perform $\chi^2$ test on the target and the control source to further quantify the significance of the flare. Figure \ref{fig:chisq} shows the chi-square test statistic per degree of freedom as a function of epoch. We define chi-square statistic $ \chi^2 = \sum_{i=1}^{N}\frac{\left(S_i-\bar{S}\right)^2}{\sigma_i^2}$ where S$_i$ and $\sigma_i$ are the flux density and corresponding measurement error at $i^{th}$ epoch and $\bar{S}$ is the mean flux density. The magnitude of $\sigma_i$ is determined by the IPS and is equal to the standard deviation (0.23~Jy) of the flux density variation away from the flares. $\bar{S} = 2.7$~Jy is the mean flux density estimated above. In Figure \ref{fig:chisq}, $ \chi^2$ is computed for the data points in a moving window with $N=4$ values and plotted against the epoch. The chi-square value is $>10$ during the flares and the p-value is less than 0.003. Thus we conclude that the flares detected at 2007.6 and 2008.6 in the J1415+1320 radio light curve at 327~MHz are statistically significant. \\
Figure \ref{fig:327MHz_flare_region} shows a close-up of the radio light curve from epoch 2004 to 2017. The flares F1 and F2 are marked in the figure. During the first flare F1, the flux density increases rapidly from 2.66~Jy to 3.99~ Jy (a variation of 1.33$\pm$0.22~Jy) in four days (all times given in this paper is relative to the observer). The flux density falls gradually in the next $\sim$ 84 days and reaches a low of 2.78~Jy on 2007.97. The second flare F2 is observed on 2008.67 where the flux density increases to 4.19~Jy. Unfortunately, no observations are present after the decay of the first flare and the onset of the second flare and so we are unable to estimate the rise time of this flare. The flux density of F2 falls in about 134 days and returns to 2.71~Jy on 2009.03. During F2, there is an abrupt fall in the flux density from 3.91~Jy to 2.92~Jy ($>$ 3$\sigma$ variation) on 2008.82. This lasted for about 11 days and then increased to 3.63~Jy on 2008.85.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{327MHz_target_control.jpg}
\caption{327~MHz light curve of J1415+1320 (blue) and control source B1345+125 (orange) obtained from ORT IPS measurements. The solid line and dashed line mark the mean and 3$\sigma$ standard deviation. F1 and F2 mark the flares at epoch 2007.6 and 2008.6 respectively.}
\label{fig:327MHz_LC}
\end{figure}
\begin{figure}[!h]
\centering
\includegraphics[width=\columnwidth]{chisqtest.jpg}
\caption{Chi-square test statistic normalized by the degrees of freedom ($\nu$) plotted for different epochs (see text).}
\label{fig:chisq}
\end{figure}
\section{Discussion}
\label{sec:discussion}
\subsection{Flares F1, F2}
We compare our data with available observations at higher frequencies. We find light curves at 15~GHz OVRO data and 37~GHz Mets\"{a}hovi radio observatory monitoring program overlapping with our low frequency observations. Figure \ref{fig:327MHz_flare_region} shows light curves measured at these high frequencies. An event with a rapid increase followed by a gradual decay nearer to the epochs of F1 and F2 is seen in the higher frequency light curves. The peak of the flare F1 at 327~MHz occurred $\sim 115$~days after the epoch at which the flux density peaked at 37~GHz. The time delay between 327~MHz and 15~GHz flares is $\sim 49$~days. For F2, the time delays are $\sim 15$ and $\sim 73$~days relative to 15 and 37~GHz flux density maxima. \\
The fractional increase in flux densities at higher frequencies is $\sim$ 80\% for both the flares; more than 1.6 compared to flux density increase obtained from the 327~MHz data. Also, the decay time scale at higher frequencies is larger by a factor of $\sim$ 2 compared to the decay time scale observed at 327~MHz. Further the spectrum of the increased flux density between 327 and 15~GHz is flatter ($\alpha=-0.2$), compared to the quiescent flux density spectrum ($\alpha=-0.3$; see Fig.~\ref{fig:spec}). For frequencies above 15~GHz, the increased flux density and quiescent values have similar spectral indices (see Fig.~\ref{fig:spec}). All these indicate that the flares might have originated outside the optically thick region of the source; the steeping of the flux density at 327~MHz during the quiescent state is due to increased contribution from the jet (and/or counter-jet) in J1415+1320 \cite{Perlman1996}. The observed flux density variation and the time evolution of the flare are generally consistent with a uniformly expanding cloud of relativistic electrons or shock-in-jet model \cite{van der Laan1966, Marscher1985}.
\subsection{Symmetric Achromatic Variability}
The flux density measurements at 327~MHz overlap with two symmetric achromatic variability (SAV) identified at higher frequencies ($>$ 15~GHz) by \cite{Vedantham2017}. These SAV events are marked in Figure \ref{fig:327MHz_flare_region} as SAV4 and SAV5 following the notation from \cite{Peirson2022}. The SAVs can be seen as a U$-$dip feature in the 15 and 37~GHz light curves. No variability of flux density within $\pm$0.7~Jy (3$\sigma$) is seen at 327~MHz during these two SAV events. \\
The SAV is possibly due to gravitational milli-lensing of the compact core emission by intervening mass condensates \cite{Vedantham2017, Peirson2022}. Therefore the variability is expected to be achromatic and would have been seen in the 327~MHz data. The spectral index of the core component is inverted with $\alpha=+1.7$ \cite{Perlman1996}. Thus the expected core flux density at 327~MHz is 1.2 mJy and the fractional change of $\sim$ 80\% inferred from higher frequency light curves during the SAV is much smaller than the measurement uncertainty of flux density at 327~MHz. Thus the non-detection of any variability at 327~MHz is consistent with the gravitational lensing model proposed to explain SAV.
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{flare_SAV.jpg}
\caption{327~MHz, 15 and 37~GHz light curves (from OVRO, UMRAO, and MRO) of J1415+1320 covering period between 2004$-$2017. The two flares are indicated by shaded regions F1 and F2 and the two SAVs found in 15~GHz light curve and reported in \cite{Peirson2022} are marked as shaded SAV4 and SAV5 following their notation.}
\label{fig:327MHz_flare_region}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\columnwidth]{spec.jpg}
\caption{Spectrum of the increased flux density during the flares F1 and F2 i.e. difference between maximum (representing the peak of the flare) and minimum (representing the quiescent state) flux density, and the average flux density.}
\label{fig:spec}
\end{figure}
\section{Conclusion}
\label{sec:conclusions}
We studied the 327~MHz multi-epoch data of J1415+1320, a blazar known for its high variability at higher frequencies. The total flux density time series was obtained from the IPS database from the ORT and covers the period between 1989 to 2018, although much of the data points exist after 2004. We report significant variability at two epochs 2007.6 (flare F1) and 2008.6 (flare F2) and establish the significance of these flares through statistical analysis. During these flares, the flux density rises rapidly, reaches a maximum, and then decreases gradually. Both F1 and F2 are seen in the overlapping datasets at 15 and 37~GHz, but with a time delay of a few months. The spectral index of the increased flux density during the flare is flatter than the quiescent spectral index, indicating the flare is likely to be associated with activity in a region with smaller optical depth. Our dataset also overlaps with two SAV events (SAV4 and SAV5) identified by \cite{Vedantham2017, Peirson2022}. No variability of flux density is seen within $\pm$ 0.7~Jy (3$\sigma$) level at 327~MHz during these events. Our non-detection of variability is consistent with a gravitational lensing origin for the SAVs if only the core flux density is affected by the lensing phenomenon.
\section{Acknowledgements}
Thanks to the anonymous referee for insightful suggestions which helped improve the manuscript. We acknowledge the help from the Radio Astronomy Centre, Ooty staff for their help during the observations. This research has made use of data from the University of Michigan Radio Astronomy Observatory which has been supported by the University of Michigan and by a series of grants from the National Science Foundation, most recently AST-0607523.
|
\section{Introduction}\label{sec:introduction}
}
\IEEEPARstart{D}{espite} the many recent improvements in automated software testing, testing is still a manual process in many industries. For example, in the gaming industry, game developers face several challenges and difficulties with writing automated tests~\cite{politowski2021survey,pascarella2018video,murphy2014cowboys}. As a result, test cases are often described in natural language and consist of a sequence of steps that must be manually performed to test the target game. Furthermore, those test cases are usually defined by employees from different departments, such as Quality Assurance (QA) engineers or developers, which may result in redundant test cases (i.e., test cases that are semantically similar or even duplicates) as the system evolves and the test suite grows~\cite{rothermel2002empirical}. Having redundant test cases is problematic in particular in a manual testing scenario, due to the tediousness and cost of executing such manual tests.
Manually identifying similar or duplicate test cases to reduce test redundancy is an expensive and time-consuming task. In addition, naive approaches (e.g., searching for exactly matching test cases) are not sufficient to capture all similarity, as different test case writers may use different terminology to specify a test case, even for similar test objectives. Therefore, an automated technique to identify similar test cases is necessary as it can prevent the QA and development teams from wastefully executing test cases that perform the same task. Throughout this paper, for brevity we adopt the term ``similar test cases'' to refer to semantically similar and duplicate test cases.
In this paper, we propose an approach to identify similar test cases that are specified in natural language. More specifically, (1)~we use text embedding, text similarity, and clustering techniques to cluster similar test steps that compose test cases and (2)~we compare test cases based on their similarity in terms of steps that belong to the same cluster.
In the first part of the study, we study how text embeddings obtained from different techniques, text similarity metrics, and different clustering algorithms can be leveraged to identify semantically similar test steps. We compare embeddings from five different techniques (Word2Vec, BERT, Sentence-BERT, Universal Sentence Encoder, and TF-IDF), two similarity metrics (Word Mover’s Distance and cosine similarity), and evaluate two different clustering techniques (Hierarchical Agglomerative Clustering and K-Means). In particular, we address the following research question for this part of the study:
\noindent\textbf{\rqone}\\
\textit{Understanding if we can effectively identify similar test steps automatically allows to know if we can rely on test step clusters to identify similarity between entire test cases. We found that we can achieve the highest performance (an F-score of 87.39\%) using an ensemble approach that consists of different embedding and clustering techniques. However, the best-performing single technique (Word2Vec) is very close in performance (with an F-score of 86.99\%).}
In the second part of the study, we leverage the previously detected clusters of test steps to identify similar test cases. We compared four different techniques to compute a similarity score (using the simple overlap, Jaccard, and cosine metrics) to measure the similarity of test cases based on the test step clusters that they have in common. In particular, we address the following research question for this part of the study:
\noindent\textbf{\rqtwo}\\
\textit{Given the difficulty of identifying similar test cases, which are usually composed of several steps, we use clusters of similar test steps to identify similar test cases. We found that test step clusters can be used to identify test case similarity with a high performance (an F-score of 83.47\%).}
Our work presents an approach to identify similar test cases based only on their natural language descriptions. We highlight that our approach is unsupervised as it does not require labelled data nor requires human supervision. In addition, no test source code or system model is necessary. QA engineers and developers can use our approach to obtain groups of similar test cases, which can be used to identify and remove redundant and legacy test cases from the test suite. Furthermore, existing groups of similar test cases can be leveraged to support the design of new test cases and help to maintain a more consistent and homogeneous terminology across the test suite. Finally, we provide access to the source code of our approach and the experiments that we performed.\footnote{\url{https://github.com/asgaardlab/test-case-similarity-technique}}
The remainder of the paper is organized as follows. In Section~\ref{sec:background}, we present the background information about text embedding and clustering techniques. We discuss related work in Section~\ref{sec:related_work} and our proposed approach in Section~\ref{sec:approach}. Section~\ref{sec:dataset} presents the dataset that we used to evaluate our approach. Sections~\ref{sec:approach_similar_steps} and~\ref{sec:approach_similar_cases} discuss the experiments that we performed to evaluate the two main stages of our approach. In Section~\ref{sec:discussions}, we discuss our results and the approach validation. Finally, Sections~\ref{sec:threats} and~\ref{sec:conclusion} present the threats to validity and conclude our work, respectively.
\section{Background}\label{sec:background}
In this section, we present an overview of the terminology and concepts that we use throughout the paper. In this work, we use ``test cases'' to refer to manual test cases that are described in natural language as a sequence of steps, i.e., test cases for which there is no source code associated.
\subsection{Text Representation}
In order to use text data as input for a machine learning algorithm, we first need to convert the text into a numeric vector through a process called \emph{text embedding}~\cite{weiss2010text,weiss2015fundamentals}. Different methods can be used to obtain a text embedding, and the embedding can be done at different granularity levels, such as at the word and sentence level. Below, we explain the different techniques that we use in this work to obtain the numeric representation of words and sentences.
\subsubsection{Word Embedding}
\label{sec:word_embedding}
A word embedding is the representation of a single word through a real-valued (and usually high-dimensional) numeric vector. In this study, we use two natural language processing techniques to obtain embeddings at the word level: Word2Vec~\cite{mikolov2013distributed} and BERT~\cite{DBLP:journals/corr/abs-1810-04805}. Figure~\ref{fig:word_embedding_example} presents two examples of pre-processed test steps along with part of their word embeddings obtained by the Word2Vec and BERT models. Next, we explain how each word embedding technique works and how the example embeddings presented in Figure~\ref{fig:word_embedding_example} are computed.
\begin{figure*}[!t]
\centering
\subfloat[Examples of word embeddings for test steps.]{
\label{fig:word_embedding_example}
\includegraphics[width=0.7\textwidth]{word_embedding_examples.pdf} }
\subfloat[Examples of sentence embeddings for test steps.]{
\label{fig:sentence_embedding_example}
\includegraphics[width=0.7\textwidth]{sentence_embedding_examples.pdf} }
\caption{Examples of test step embeddings. Note that we provide only the first two elements of the embedding vector due to space constraints as the actual vectors have a high dimension.}
\label{fig:text_embedding_examples}
\end{figure*}
\textbf{Word2Vec} transforms words into high-dimensional numeric vectors that are able to maintain the syntactic and semantic relationships between words in the vector space~\cite{mikolov2013distributed,mikolov2013efficient}. This means that embeddings of similar words will (most of the time) be close in the vector space (i.e., the distance between the embedding vectors is small). Furthermore, with Word2Vec, each word is assigned a single numeric vector regardless of the context in which it is used, as we can see for the words ``verify'' and ``item'' in the two steps in Figure~\ref{fig:word_embedding_example}. In this work, we used the continuous bag-of-words (CBOW) model architecture of Word2Vec, which is faster than the other possible architecture, called skip-gram~\cite{mikolov2013efficient}.
Differently from Word2Vec,~\textbf{BERT (Bidirectional Encoder Representations from Transformers)} is a transformer-based model that can be used to extract contextual word embeddings, i.e., embeddings that change depending on the context in which a word is present~\cite{DBLP:journals/corr/abs-1810-04805}. This means that the same word may have different embedding vectors, as we can see in Figure~\ref{fig:word_embedding_example}, where the BERT embeddings for the words ``verify'' and ``item'' are different in the two test steps because those words are in different contexts.
BERT is available as a model that was pre-trained on lower-cased English text (uncased BERT). This pre-trained model can further be trained with a domain-specific training set (known as domain-adaptive pre-training~\cite{gururangan2020don}). The BERT model uses WordPiece tokenization~\cite{wu2016google}, in which a word may be split into sub-words. For example, the word ``validate'' is composed of the sub-words ``valid'' and ``ate'', each one with its own embedding vector. Therefore, when extracting embeddings of words that are split into sub-words, we need to aggregate the embeddings of the sub-words (e.g., by averaging the embedding vectors).
\subsubsection{Sentence Embedding}
\label{sec:sentence_embedding}
Differently from word embedding, sentence embedding is the representation of a whole sentence with a real-valued (and usually high-dimensional) numeric vector. In this work, we use three different techniques to extract sentence embeddings (SBERT, USE, and TF-IDF). Figure~\ref{fig:sentence_embedding_example} presents two examples of pre-processed test steps along with part of their sentence embeddings obtained by the SBERT, USE, and TF-IDF techniques.
Next, we explain how each sentence embedding technique works.
\textbf{Sentence-BERT (SBERT)} is a BERT-based framework that allows us to directly extract numeric representations of full sentences~\cite{reimers2019sentence}. The embeddings of sentences that are semantically similar are close in the embedding space. We can use this information for different purposes, such as identifying paraphrases and clustering similar sentences. For instance, the SBERT embeddings of the two test steps presented in Figure~\ref{fig:sentence_embedding_example} are close in the embedding space (i.e., have a small distance between them). Among several generic and task-specific SBERT pre-trained models that are available\footnote{\href{https://www.sbert.net/docs/pretrained_models.html}{https://www.sbert.net/docs/pretrained$\_$models.html}}, three models are suitable for our task (identifying similar test steps): \textit{paraphrase-distilroberta-base-v1}, \textit{stsb-roberta-base}, and \textit{stsb-roberta-large}. While the first model is optimized to identify paraphrases and was trained on large scale paraphrase data, the second and third ones are the base and large versions of a model that was optimized for semantic textual similarity.
\textbf{Universal Sentence Encoder (USE)} is an embedding technique that can be used to directly extract embeddings from sentences, phrases, or short paragraphs to be used in another task, such as textual similarity and clustering tasks~\cite{cer2018universal}. With a similar behaviour to SBERT, the two examples presented in Figure~\ref{fig:sentence_embedding_example} have close embedding vectors.
Finally, we also used the \textbf{TF-IDF (Term Frequency–Inverse Document Frequency)} method to represent sentences. TF-IDF computes the importance of a word to a document by combining the word frequency in the document and the word frequency across all the other documents~\cite{joachims1996probabilistic,salton1991developments,jones1972statistical}. In our case, the test steps (i.e., sentences) are considered documents. We built a numeric vector for each test step using the word importance values. Words that are not present in the step are assigned a value of zero. We can observe a typical vector obtained with TF-IDF in the examples presented in Figure~\ref{fig:sentence_embedding_example}, in which the values different from zero correspond to the importance of the words presented in the ``verify item name'' and ``verify item description'' steps.
\subsection{Clustering techniques}
\textbf{Hierarchical Agglomerative Clustering (HAC)}~\cite{rokach2005clustering} is a clustering algorithm that works in a bottom-up manner. Initially, each data point corresponds to a single cluster itself, and as the algorithm iterates, different clusters are merged with the aim of minimizing a specific linkage criterion. The result of the iterative merging process is a tree structure that can represent the data points (and their clusters), known as a dendrogram. Although the dendrogram can be used to identify the number of clusters, in our work we determined that parameter empirically and used the number that maximizes our evaluation metric (as explained in Section~\ref{sec:clustering_test_steps}). Different linkage criteria can be used, such as single-linkage (the algorithm uses the minimum of the distances between all data points of two sets) and average-linkage (the algorithm uses the average of the distances between all data points of two sets).
The \textbf{K-means clustering}~\cite{duda1973pattern} algorithm splits the data points into $k$ different clusters. Different from HAC, no hierarchical cluster structure is generated with K-means. The goal of K-means is to group data points in order to minimize the distance between points belonging to the same cluster compared to the distance of points from different clusters. Using the Expectation-Maximization algorithm~\cite{moon1996expectation}, K-means starts with $k$ centroids. Then, the algorithm (1)~assigns each data point to the closest cluster (in terms of the distance between the point and the centroids) and (2)~computes the new centroids using the updated data point assignments. The execution finishes when there is no change to the allocation of data points.
\section{Related Work}\label{sec:related_work}
In this section, we discuss prior work that applied clustering techniques~\cite{li2020clustering,walter2018improving,chen2011using,zalmanovici2016cluster,b2016learning} and natural language processing (NLP)~\cite{masuda2016automatic,li2020clustering,wang2020automatic,wang2015automatic,mai2018natural,tan2020collaborative,masuda2015semantic} to software testing.
\subsection{Clustering techniques for software testing}
Our work is based on the study of~\citet{li2020clustering}, which proposed an approach to cluster test steps written in natural language based on the steps' similarities. The study used text embeddings (including embeddings obtained with the Word2Vec technique) together with the Relaxed Word Mover's Distance (RWMD) metric~\cite{kusner2015word} to measure similarity between embeddings. The test steps were then clustered with the hierarchical agglomerative and K-means clustering techniques. The approach was evaluated on a large-scale dataset of a mobile app and achieved an F-score of 81.55\% in the best case. The proposed approach also reduced the manual effort for implementing test-step API methods by 65.90\%. Differently from~\citeauthor{li2020clustering}'s work~\cite{li2020clustering}, we evaluated more recent NLP techniques to obtain word and sentence embeddings (BERT, SBERT, and Universal Sentence Encoder). Furthermore, we extended~\citeauthor{li2020clustering}'s work~\cite{li2020clustering} for the purpose of identifying similar test cases using the identified clusters of test steps.
\citet{walter2018improving} proposed an approach to improve the efficiency of test execution. The approach removes redundant test steps and uses clustering techniques to rearrange the remaining steps. To use the approach, the textual descriptions of test cases must be converted into a representation form of parameters concatenated by first order logic operators (AND, OR, NOT). The approach was evaluated in a case study with a system from an automotive industry company. The results indicated a test load reduction of 18\% due to the removal of redundant test steps and rearranging of the remaining steps.~\citet{chetouane2020using} proposed an approach to reduce a test suite by clustering similar test cases (based on their source code) with the K-means algorithm. 13 Java programs were used to evaluate if the approach could efficiently reduce the test suite and assess the impact on coverage metrics. The evaluation showed that the approach can reduce the test suite by 82.2\% while maintaining the same coverage metric as the original test suite.
\citet{pei2021dynamic} proposed distance-based Dynamic Random Testing (DRT) approaches with the goal of improving the fault detection effectiveness of DRT. The work clustered similar test cases based on their source code with three clustering methods: K-means, K-medoids, and hierarchical clustering. The information of distance between the test case groups was used to identify test cases that are closer to failure-causing groups. 12 versions of 4 open-source programs were used to evaluate the approaches. The evaluation showed that the proposed strategies achieve a larger fault detection effectiveness with a low computational cost compared to other DRT approaches.~\citet{arafeen2013test} investigated whether clustering of test cases based on similarities in their requirements could improve traditional test case prioritization techniques. The paper used TF-IDF and the K-means clustering algorithm to group test cases that have similar requirements. Two Java programs were used to evaluate the approach. The evaluation showed that the use of requirements similarity can improve the effectiveness of test case prioritization techniques but the improvements vary with the cluster size.
Differently from the works above, our study aims at finding redundant test cases by clustering similar test cases that are written in natural language. We experimented with different NLP and clustering techniques to find clusters of similar test steps, which are used with test case names to obtain similar test cases. Furthermore, differently from the work of~\citet{walter2018improving}, which converts natural language descriptions of test cases into a representation form of parameters concatenated by logic operators to be used with their approach, our proposed approach works in an unsupervised manner with the original test cases written in natural language.
\subsection{Natural Language Processing techniques for software testing}
\citet{wang2015automatic} proposed an approach to automate the generation of executable system test cases. The approach applies NLP techniques (such as tokenization and part-of-speech tagging) to textual data obtained from use case specifications. Furthermore, a domain model of the system under analysis is necessary to generate test data and oracles.~\citet{wang2015automatic} performed an industrial case study with automative software to demonstrate the feasibility of the proposed approach.~\citet{wang2020automatic} extended their previous work~\cite{wang2015automatic} by further providing empirical evidence about the scalability of the approach to generate executable, system-level test cases for acceptance testing from natural language requirements. In addition,~\citet{wang2020automatic} focused on embedded systems and demonstrated the effectiveness of the proposed approach using two industrial case studies, in which the approach correctly generated test cases that exercise different scenarios manually implemented by experts, including critical scenarios not previously considered.
\citet{yue2015rtcm} proposed a Test Case Specification (TCS) language, called Restricted Test Case Modeling (RTCM), and an automated test case generation tool, called \textit{aToucan4Test}, to transform textual test cases into executable test cases. RTCM provides a template that combines natural language with restriction rules and keywords for writing TCS. Two case studies were performed to assess the applicability of RTCM and a commercial video conferencing system was used to evaluate the \textit{aToucan4Test} tool. \textit{aToucan4Test} could correctly generate 246 executable test cases from 9 test case specifications of subsystems of the video conferencing system. The study also evaluated the effort to use RTCM and \textit{aToucan4Test} using the average time for deriving the executable test cases, which is 0.5 minutes.~\citet{mai2018natural} addressed the problem of automatically generating executable test cases from security requirements in natural language.~\citeauthor{mai2018natural} proposed an approach to generate security vulnerability test cases from use case specifications that capture malicious behavior of users. Similarly to previous work,~\citeauthor{mai2018natural} evaluated the approach with an industrial case study in the medical domain. The evaluation indicated that the proposed approach can automatically generate test cases detecting vulnerabilities.
The aforementioned works used different NLP techniques to automatically generate different types of test cases. In contrast, we propose an approach that leverages different NLP techniques to extract text embeddings and can automatically identify similar test cases. The approach can be used to identify redundant and legacy test cases written in natural language.
\begin{table*}[!t]
\centering
\caption{Running example.}
\label{tab:test_case_running_example}
\begin{tabularx}{\linewidth}{p{1cm}p{2cm}p{1.5cm}p{1cm}XX} \toprule
\textbf{\begin{tabular}[c]{@{}l@{}}Test case\\ identifier\end{tabular}} & \textbf{\begin{tabular}[c]{@{}l@{}}Test case\\ name\end{tabular}} & \textbf{\begin{tabular}[c]{@{}l@{}}Test case\\ type\end{tabular}} & \textbf{\begin{tabular}[c]{@{}l@{}}Test step\\ identifier\end{tabular}} & \textbf{\begin{tabular}[c]{@{}l@{}}Test step\\ (before pre-processing)\end{tabular}} & \textbf{\begin{tabular}[c]{@{}l@{}}Test step\\ (after pre-processing)\end{tabular}} \\ \midrule
\multirow{3}{*}{TC1} & \multirow{3}{*}{\begin{tabular}[c]{@{}l@{}}Log in to an\\ existing account\end{tabular}} & \multirow{3}{*}{Login} & TS1.1 & Login to the game using an existing account that has completed the tutorial & {[}login, game, using, existing, account, completed, tutorial{]} \\
& & & TS1.2 & Select the Playing from School portal & {[}select, playing, school, portal{]} \\ \midrule
\multirow{7}{*}{TC2} & \multirow{7}{*}{\begin{tabular}[c]{@{}l@{}}Assignment with\\ many students\end{tabular}} & \multirow{7}{*}{Education} & TS2.1 & Update the assignment adding students & {[}update, assignment, adding, student{]} \\
& & & TS2.2 & Request the next skill and question from the algorithm gateway for the 1st student on the assignment & {[}request, next, skill, question, algorithm, gateway, student, assignment{]} \\
& & & TS2.3 & Request the next skill and question from the algorithm gateway for the middle student on the assignment & {[}request, next, skill, question, algorithm, gateway, middle, student, assignment{]}
\\ \midrule
\multirow{10}{*}{TC3} & \multirow{10}{*}{\begin{tabular}[c]{@{}l@{}}Student has\\ multiple\\ assignments\end{tabular}} & \multirow{10}{*}{Education} & TS3.1 & Request the next skill and question from the algorithm gateway for one of the students that was on the assignment & {[}request, next, skill, question, algorithm, gateway, one, student, assignment{]} \\
& & & TS3.2 & Remove student from the first assignment & {[}remove, student, first, assignment{]} \\
& & & TS3.3 & Request the next skill and question from the algorithm gateway for one of the students that was on the assignment & {[}request, next, skill, question, algorithm, gateway, one, student, assignment{]} \\
& & & TS3.4 & Remove the student from the second assignment & {[}remove, student, second, assignment{]}
\\ \bottomrule
\end{tabularx}
\end{table*}
\section{Proposed approach}\label{sec:approach}
In this section, we demonstrate our proposed approach through a running example. Figure~\ref{fig:approach} presents an overview of our approach, which consists of three stages: (1) pre-processing of test cases, (2) clustering of similar test steps and (3) identification of similar test cases. Next, we explain the stages of our approach.
\begin{figure}[!t]
\centering
\includegraphics[width=0.5\textwidth]{methodology_running_example_fig.pdf}
\caption{Overview of our proposed approach.}
\label{fig:approach}
\end{figure}
\subsection{Stage 1: Test case pre-processing}
\label{sec:approach_preprocessing}
Our approach relies only on test cases that are written in natural language, which means that there is no source code available for our test cases. The input to our approach consists of unprocessed (raw) test cases. Table~\ref{tab:test_case_running_example} presents three test cases (TC1, TC2, and TC3) that we use as a running example to describe how our approach identifies similar test cases. As we can observe, each test case contains an identifier, a name and a type. In addition, a test case has one or more test steps, which are instructions that the tester must perform in order to achieve the overall objective of the test case. Note that this objective is generally not explicitly specified. The test steps that we collect to perform our experiments are explicitly identified (i.e., each test step has its own field within a test case). Therefore, we can directly collect the test steps and identify to which test case they belong.
Each test step is assigned a unique identifier and is pre-processed. Initially, we used tokenization to transform the step sentences into a list of words. To ensure that we have high-quality data, we obtained a list of the unique words in our data and manually inspected the list to identify misspelled words, which were used to build a list of [misspelled\_word, fixed\_word] tuples. The manually built tuple list was used to programmatically replace misspelled words with the corresponding fixed words across the entire dataset. We then removed stopwords (such as ``a'', ``of'', and ``the'') as they do not add meaning to the sentences. Also, we applied lemmatization to the words to have a consistent terminology across the data. Finally, similar to prior work~\cite{li2020clustering}, we removed words that occur only once in the whole dataset as they may introduce bias in the data (e.g., due to an imprecise numeric representation of those words). Overall, a test case instance can be represented by the triple: \[\textless test\_case\_name, test\_case\_type, test\_steps\textgreater\]
\vspace{-0.3cm}
\begin{figure*}[!t]
\centering
\includegraphics[width=0.80\textwidth]{approach_stage_2.pdf}
\caption{Overview of stage 2 of our approach with the running example.}
\label{fig:approach_stage_2}
\end{figure*}
\begin{figure*}[!t]
\centering
\includegraphics[width=0.80\textwidth]{approach_stage_3.pdf}
\caption{Overview of stage 3 of our approach with the running example.}
\label{fig:approach_stage_3}
\end{figure*}
\subsection{Stage 2: Test step clustering}
In the second stage, our approach clusters similar test steps. Figure~\ref{fig:approach_stage_2} shows how the steps of the three test cases are processed in this stage. Before applying a machine learning algorithm to text data, we need to transform the text into a numeric representation~\cite{weiss2010text,weiss2015fundamentals}. Our approach starts by transforming each test step into one or more numeric vectors (text embedding). The pairwise similarity between steps (in terms of embedding distance) is then computed. The computed distances between the text embeddings of the test steps can be used to capture their similarity. In particular, embeddings that are close in the embedding space should represent similar steps.
Finally, our approach leverages the computed distances to identify clusters of similar test steps. While steps that have a small distance between them should belong to the same cluster, steps with larger distances should be in different clusters.
\subsection{Stage 3: Test case similarity}
In the last stage, our approach leverages the clusters of test steps identified in stage 2 together with the test case name to find similar test cases. Figure~\ref{fig:approach_stage_3} shows how the TC1, TC2, and TC3 test cases are processed in this stage. The relationship between test cases and test step clusters is represented through a matrix in which each row is a test case (TC1, TC2, and TC3) and each column is a step cluster (C1, C2, C3, C4, and C5). Initially, for each test case (matrix row), the approach identifies the test step clusters (matrix column(s)) that contain one or more steps of the test case. Our approach supports the use of Boolean (which yield a matrix consisting of 0's and 1's) or numeric flags. Note that a numeric flag represents the number of test steps present in the identified cluster. After filling in the matrix, each test case is represented by the corresponding Boolean or numeric vector (a row in the matrix) with a length corresponding to the total number of test step clusters. Test cases are then compared to each other in terms of the similarity between their representation vectors. Finally, to incorporate knowledge from the test case name, the approach computes the pairwise similarity between test case name embeddings and combines this similarity metric with the one obtained from the test step clusters. The final test case similarity score is a weighted average between the test step cluster and the test case name metrics. For the running example, our approach identifies the TC2 and TC3 test cases as similar but both are different from the TC1 test case. A QA engineer can then investigate those test cases to decide, for example, whether they are redundant or should be improved.
\section{Dataset}\label{sec:dataset}
We collected 3,323 test case descriptions written in natural language. The test cases under study were manually designed to test the \textit{Prodigy Game}\footnote{https://www.prodigygame.com/main-en/}, a proprietary, educational math game with more than 100 million users around the world. Each test case is composed of one or more test steps and, in total, there are 15,644 steps. We also collected the predefined type of the test case regarding the part of the game that is being tested. The test case type is available for 2,053 test cases (62\% of the total number of test cases). All the test steps are pre-processed according to the pre-processing steps as explained in Section~\ref{sec:approach_preprocessing}.
To evaluate the performance of our approach (stage 2, for test step clustering, and stage 3, for test case similarity), we used our dataset to manually build a ground truth of similar test steps (stage 2) and similar test cases (stage 3), as we explain below.
\noindent\textbf{Ground truth of similar test steps (stage 2 of our approach).}
We selected a representative sample from all 15,644 test steps with a confidence level of 95\% and a confidence interval of 5\%, which corresponds to 394 steps. The test step samples were manually analyzed in an incremental manner: for each step under analysis, we looked at all the previously analyzed steps to identify the similar ones. When we found one step or a group of steps similar to the current step, we included the current one in the group of the previous steps. If there was no previous step similar to the current one, we assigned it to a new cluster. The ground truth of similar test steps ended up with a total of 213 clusters and an average of 1.9 (standard deviation of 2.0) test steps per cluster. We also found that the largest cluster has 15 test steps.
\noindent\textbf{Ground truth of similar test cases (stage 3 of our approach).}
We selected a representative sample of test cases with a confidence level of 95\% and a confidence interval of 5\%, which corresponds to 381 test cases. Similarly to the way that we built the ground truth of similar test steps, the test case samples were manually analyzed in an incremental manner: for each test case under analysis, we looked at all the previously analyzed cases to identify the similar ones in terms of what is being tested, the similarity between the test steps of the test cases, and similarity between the test case names. Note that, for each test case, we analyzed the test case name, test case type, and all the steps that compose the test case. The ground truth of similar test cases ended up with a total of 242 clusters and an average of 1.6 (standard deviation of 1.9) test cases per cluster. For this ground truth, we found that the largest cluster has 21 test cases.
\section{Evaluating our approach for clustering similar test steps}\label{sec:approach_similar_steps}
In this section, we discuss the experiments that we performed to evaluate our approach for clustering similar test steps in an industrial setting.
\subsection{Evaluated techniques}
Our approach consists of several steps that can be implemented through different techniques and models. To evaluate our approach, we performed experiments with combinations of five different text embedding techniques, two similarity metrics, and two clustering techniques. Figure~\ref{fig:test_step_similarity_experiments} presents an overview of the experiments that we performed to address RQ1. Different NLP techniques can be used for text embedding at different granularities, such as words, sentences, and short paragraphs~\cite{mikolov2013distributed,mikolov2013efficient,DBLP:journals/corr/abs-1810-04805,reimers2019sentence,cer2018universal,le2014distributed}. As our test steps usually consist of a single sentence and the test steps are transformed into a list of words after pre-processing, we adopt word-level and sentence-level text embedding. We used two techniques to obtain text embeddings at the word-level (Word2Vec~\cite{mikolov2013distributed,mikolov2013efficient} and BERT~\cite{DBLP:journals/corr/abs-1810-04805}) for the test steps and computed the embedding similarity using the Word Mover's Distance (WMD) metric~\cite{kusner2015word}. For text embeddings at the sentence-level, we used three techniques (SBERT~\cite{reimers2019sentence}, Universal Sentence Encoder~\cite{cer2018universal}, and TF-IDF~\cite{joachims1996probabilistic,salton1991developments}) and used the cosine similarity to compare the embeddings. For both types of embeddings, we applied the hierarchical agglomerative~\cite{sneath1973numerical} and K-means~\cite{duda1973pattern} clustering techniques to obtain clusters of similar steps.
\begin{figure}[!t]
\centering
\includegraphics[width=0.5\textwidth]{approach_rq1.pdf}
\caption{Overview of the experiments to identify clusters of similar test steps.}
\label{fig:test_step_similarity_experiments}
\end{figure}
\subsection{Configuration of the word embedding techniques}
\label{sec:word_embedding_evaluation}
\noindent\textit{Word2Vec.}
We trained a Word2Vec model using all 15,644 test steps that we collected. Furthermore, to provide more context to the embedding model during training, we concatenated the test case type (available for 2,053 test cases) and test case name to each step. We used an embedding vector of length 300 (as in the original study that proposed the Word2Vec model~\cite{mikolov2013distributed}). We used the continuous bag-of-words (CBOW) model architecture of Word2Vec with two context words as this configuration provides the highest test step clustering performance, which was determined through an experiment in which we varied the number of context words from one to ten. We initialized the word embeddings with the weights from the large-scale pre-trained model released by Google.\footnote{\href{https://code.google.com/archive/p/word2vec/}{https://code.google.com/archive/p/word2vec/}} This model contains 3 million word embeddings with dimension 300 and was trained on a Google News corpus with approximately 100 billion words. For words that are present in our dataset but not in the pre-trained model (and, therefore, cannot be initialized with pre-trained weights), we followed a process proposed by~\citet{li2020clustering} to initialize the word embeddings. We computed the mean and standard deviation of the initialized words and initialized the remaining words with samples of a normal distribution parameterized by the computed mean and standard deviation. Finally, the outcome of the training process is the word embeddings learned with our data.
\vspace{0.15cm}
\noindent\textit{BERT.}
In this work, we used the pre-trained model released by Google\footnote{\href{https://github.com/google-research/bert}{https://github.com/google-research/bert}} (\emph{pre-trained BERT}) to obtain contextual embeddings of the test steps. Furthermore, we used a model with additional pre-training using our own corpus of test steps (\emph{domain-adaptive pre-trained BERT}) to obtain the contextual embeddings. We explain the configurations of both models below.
\noindent\underline{Pre-trained BERT.}
For the pre-trained model, we used the uncased (case-insensitive) version of the base model~\cite{DBLP:journals/corr/abs-1810-04805,turc2019well}. We transformed the test step text into the BERT format by adding the \textit{[CLS]} and \textit{[SEP]} tokens respectively to the start and end of each test step text.
The test step was then tokenized with BERT's own tokenizer. Finally, we used the tokenized steps to extract the contextual embeddings. As explained in Section~\ref{sec:word_embedding}, we can adopt different pooling strategies to obtain the embedding vector for a word. We performed experiments with four different pooling strategies to combine the layers (as suggested by the original paper's authors~\cite{DBLP:journals/corr/abs-1810-04805}): using only the second-to-last layer, summing the last four layers, averaging the last four layers, and concatenating the last four layers. We found that summing the last four layers achieves the best performance with our data. Finally, we used the average of sub-word embeddings (see Section~\ref{sec:word_embedding}) to obtain the original out-of-vocabulary word embedding.
\noindent\underline{Domain-adaptive pre-trained BERT.}
We also performed additional pre-training of BERT with our corpus. For the additional pre-training, after experimenting with the base and large models, we decided to use the uncased version of the BERT large model as the initial checkpoint (i.e., we performed the additional pre-training on top of the pre-trained large model). We followed the same process to configure the test step text to a BERT-friendly format. However, differently from the pre-trained model, using the second-to-last layer (instead of summing the last four layers) achieves the best results for the domain-adaptive pre-trained BERT model.
\subsection{Configuration of the sentence embedding models}
\label{sec:sentence_embedding_evaluation}
\noindent\textit{Sentence-BERT (SBERT).}
We performed experiments with three available pre-trained SBERT models suitable for our task (see Section~\ref{sec:sentence_embedding}): \textit{paraphrase-distilroberta-base-v1}, \textit{stsb-roberta-base}, and \textit{stsb-roberta-large}. We decided to use the \textit{paraphrase-distilroberta-base-v1} model since it achieves the best results with our data. To obtain the embeddings for the test steps, we just provided the test steps directly as parameters to the SBERT model.
\vspace{0.15cm}
\noindent\textit{Universal Sentence Encoder (USE).}
To obtain the test step embeddings with the USE model, we provided the steps directly as parameters to the USE model.
\vspace{0.15cm}
\noindent\textit{TF-IDF.}
Finally, we also used TF-IDF to extract the numeric vector representations of the test steps. For each word, we computed its importance in a single test step relative to all the other test steps.
\subsection{Computing the test step similarity}
\label{sec:computing_step_similarity}
\noindent\textit{Word Mover's Distance (WMD).}
We used the Word Mover's Distance (WMD)~\cite{kusner2015word} metric to measure the similarity between test step embeddings at the word level. The WMD metric is suitable to be used together with the Word2Vec and BERT models because of the property that distances between embedded words in the embedding space are semantically meaningful, which is a property that WMD relies on~\cite{kusner2015word}. We computed the pairwise WMD metric between any two test steps and built a distance matrix of dimension \textit{[15,644 x 15,644]}. The more similar two steps are, the lower is the WMD metric, with the lowest bound being zero for exactly matching steps.
\vspace{0.15cm}
\noindent\textit{Cosine similarity.}
Since cosine is a widely used metric to measure similarity between text vectors~\cite{huang2008similarity,li2013distance,gomaa2013survey,salton1988term}, we used the cosine to measure the similarity between test step embeddings at the sentence level. Similarly to the way we computed the WMD metric, we computed the pairwise cosine similarity between any two test steps and built a distance matrix of dimension \textit{[15,644 x 15,644]}. As the cosine similarity value measure the cosine of the angle between two step numeric vectors, the smaller the angle, the larger its cosine and the more similar the two test steps are.
\subsection{Clustering test steps}
\label{sec:clustering_test_steps}
\noindent\textit{Hierarchical Agglomerative Clustering.}
We applied the hierarchical agglomerative clustering technique to the distance matrix that we built in the previous step (Section~\ref{sec:computing_step_similarity}). We used the average linkage criterion (with Euclidean distance), which means that the clustering algorithm merges pairs of test step clusters that minimize the average distance between each observation of the pairs.
\vspace{0.15cm}
\noindent\textit{K-means.}
To apply the K-means clustering technique, we used the test step embeddings obtained with the word/sentence embedding techniques (Sections~\ref{sec:word_embedding_evaluation} and~\ref{sec:sentence_embedding_evaluation}). Note that, for word-level embeddings, we transformed the embedding vectors of the words of a test step into a single vector to represent the whole test step by computing the word embeddings' average. Furthermore, to speed up the execution of K-means, we used the centroids of the clusters obtained by the hierarchical approach as the initialization centroids, similarly to prior work~\cite{li2020clustering,lu2008hierarchical}.
Regarding the number of clusters for both clustering techniques, we chose the number of clusters that maximized the F-score (which is our evaluation metric, as explained in Section~\ref{sec:evaluation_metric}). We performed a search by varying the number of clusters from 50 up to 15,000 with a step of 50, and for each value we executed both clustering approaches and computed the F-score. Finally, we selected the (optimal) number of clusters for which each clustering technique achieved the highest F-score. Note that the optimal number of clusters might be different for the hierarchical clustering and K-means.
\vspace{0.15cm}
\noindent\textbf{Ensemble approach.}
Each text embedding technique that we used has different characteristics and properties to extract word or sentence embeddings, which leads to different clusters of test steps. Therefore, attempting to mitigate each model's specific weaknesses, we built an ensemble approach that uses majority voting. That is, the ensemble approach assigns two test steps to the same cluster if at least three (out of the five) previously implemented approaches assign those two steps to the same cluster.
\vspace{0.15cm}
\noindent\textbf{Baseline.}
We used two baselines to evaluate the performance of our proposed approaches. The first baseline assigns test steps to the same cluster only if those steps are exactly the same. The second baseline uses the Word2Vec technique together with the WMD similarity metric and only assigns two test steps to the same cluster if the WMD similarity of those steps is zero (i.e., their embeddings are the same).
\subsection{Evaluation metric}
\label{sec:evaluation_metric}
We are interested in penalizing both the false positives (to avoid excessive suggestions of similar test steps when they are not similar) and false negatives (to avoid missing out many similar test steps). Therefore, we used the F-score metric (as shown in Equation~\ref{eq:f_score}) to evaluate the test step clustering approaches as this metric captures the trade-off between precision (related to false positives) and recall (related to false negatives). Using the test steps present in the manually built ground truth of similar test steps, we analyzed all the pairs of test steps, similarly to prior work~\cite{li2020clustering}:
\begin{itemize}
\item\textit{True positive (TP)}: when a pair of steps is included in the same cluster by our approach and the steps indeed belong to the same cluster in the ground truth.
\item \textit{False positive (FP)}: when a pair of steps is included in the same cluster by our approach but the steps do not belong to the same cluster in the ground truth.
\item \textit{True negative (TN)}: when a pair of steps is not included in the same cluster by our approach and the steps do not belong to the same cluster in the ground truth.
\item \textit{False negative (FN)}: when a pair of steps is not included in the same cluster by our approach but the steps belong to the same cluster in the ground truth.
\end{itemize}
We then computed the F-score metric as follows:
\begin{equation}
\label{eq:f_score}
\text{F-score} = 2 \times \frac{precision \times recall}{precision + recall }
\end{equation}
Where the precision corresponds to the proportion of true positives regarding all the pairs identified as positive ($\frac{TP}{TP + FP}$) and the recall corresponds to the proportion of true positives regarding all the existing positive instance ($\frac{TP}{TP + FN}$).
\subsection{Findings}
\textbf{Similar test steps that are written in natural language can be identified with a high performance by applying the ensemble approach that uses majority voting. Furthermore, we can achieve a similar high performance using a single technique (Word2Vec).} Table~\ref{tab:f_score_approaches} presents the F-score of all the approaches that we evaluated along with the clustering techniques and the optimal number of clusters.
All the proposed approaches achieve a similar and high performance, with an F-score between 83.96\% and 87.39\%, except for both baselines, which have the same F-score of 70.40\%. More specifically, the ensemble approach achieves the highest performance, with an F-score of 87.39\%. If we look at the performance of the single models, Word2Vec with K-means has the highest F-score (86.99\%), which is very close to the ensemble approach performance. TF-IDF achieves the second highest F-score (86.67\%) among the single models, followed closely by SBERT with K-means (86.10\%) and Domain-BERT with K-means (85.77\%)
Regarding the two versions of the BERT model, we observe that the domain-adaptive pre-trained BERT is a little better, with F-scores of 85.60\% (using HAC) and 85.77\% (using K-means), in comparison to the generic pre-trained BERT, with F-scores of 84.65\% (using HAC) and 85.15\% (K-means). One possible reason for the small gain is that we do not have large amounts of data for the domain-adaptive pre-training. However, our findings indicate that the additional pre-training is capable of improving the model performance and might be more helpful with larger datasets.
We can observe that for all the approaches except for TF-IDF, running K-means on top of HAC is beneficial as this increases the F-score. Note, however, that the gain in performance is minimal, such as 1.14\% and 0.50\% in absolute percentage point for Word2Vec and BERT, respectively. On average, applying K-means on top of hierarchical clustering increases the performance by 0.33\% in absolute percentage point.
Finally, even though the ensemble approach has the highest performance for clustering test steps, this approach is computationally expensive as it requires the implementation and execution of all the other approaches, which takes around 6 hours in total using a single core on an Intel i7-8700 CPU. However, we can achieve a very close performance with a single technique, such as Word2Vec (which takes around 2.5 hours to execute), TF-IDF (which takes around 25 minutes), or SBERT (which takes around 3 minutes). Our experiments showed that both Word2Vec and BERT present a large execution time due to the large computational cost of computing the Word Mover's Distance. The reported execution times are for the full test step clustering pipeline (test step pre-processing, word embedding training, test step similarity, and clustering).
\begin{table}[!t]
\centering
\caption{F-score of the test step clustering approaches along with the clustering techniques and the optimal number of clusters obtained (\# Clusters). Note that we abbreviate hierarchical agglomerative clustering as HAC. The best performing approaches are highlighted in bold.}
\label{tab:f_score_approaches}
\begin{tabular}{lp{2.099cm}ll} \toprule
Text embedding technique & Clustering & F-score & \# Clusters \\ \midrule
Baseline 1 & - & 70.40 & 4407 \\
Baseline 2 & - & 70.40 & 4393 \\
Word2Vec & HAC & 85.85 & 2650 \\
\textbf{Word2Vec} & \textbf{K-means} & \textbf{86.99} & \textbf{2650} \\
BERT & HAC & 84.65 & 3050 \\
BERT & K-means & 85.15 & 3050 \\
Domain-adaptive BERT & HAC & 85.60 & 3300 \\
Domain-adaptive BERT & K-means & 85.77 & 3300 \\
SBERT & HAC & 85.71 & 3350 \\
SBERT & K-means & 86.10 & 3350 \\
USE & HAC & 83.96 & 3050 \\
USE & K-means & 84.39 & 2900 \\
TF-IDF & HAC & 86.67 & 2500 \\
TF-IDF & K-means & 86.03 & 2500 \\
\textbf{Ensemble approach} & - & \textbf{87.39} & \textbf{3158} \\ \bottomrule
\end{tabular}
\end{table}
\section{Evaluating our approach for identifying similar test cases}\label{sec:approach_similar_cases}
In this section, we discuss the experiments that we performed to evaluate our approach for identifying similar test cases that are specified in natural language. Below, we discuss four different techniques to identify similar test cases using the previously identified clusters of test steps.
\subsection{Evaluated techniques}
We performed experiments with four different proposed techniques to identify similar test cases using the previously identified clusters of test steps. Figure~\ref{fig:test_case_similarity} gives an overview of the experiments. To explain how each technique works, we use the two example test cases presented in Table~\ref{tab:methods_case_similaririty}.
\begin{figure}[!h]
\centering
\includegraphics[width=0.5\textwidth]{approach_rq2.pdf}
\caption{Overview of the experiments to identify similar test cases.}
\label{fig:test_case_similarity}
\end{figure}
\begin{table*}[!t]
\centering
\caption{Examples of test case representations (through vectors) obtained with the experimented four techniques and the corresponding similarity scores.}
\label{tab:methods_case_similaririty}
\begin{tabular}{lllllll} \toprule
\textbf{Test case} & \textbf{Test step} & \textbf{Test step cluster} & \textbf{Technique 1} & \textbf{Technique 2} & \textbf{Technique 3} & \textbf{Technique 4} \\ \midrule
\multirow{2}{*}{TC1}
& \multirow{2}{*}{TS1, TS2, TS3, TS4}
& \multirow{2}{*}{C1, C2, C3, C1}
& \multirow{2}{*}{[C1, C2, C3]}
& \multirow{2}{*}{[1, 1, 1, 0, 0]}
& \multirow{2}{*}{[2, 1, 1, 0, 0]}
& \multirow{2}{*}{\shortstack[l]{[2, 1, 1, 0, 0] + \\
{[}TC1 Name embedding{]} }}
\\ \\
\multirow{2}{*}{TC2}
& \multirow{2}{*}{TS1, TS5, TS6, TS7, TS8}
& \multirow{2}{*}{C1, C4, C5, C2, C5}
& \multirow{2}{*}{[C1, C2, C4, C5]}
& \multirow{2}{*}{[1, 1, 0, 1, 1]}
& \multirow{2}{*}{[1, 1, 0, 1, 2]}
& \multirow{2}{*}{\shortstack[l]{[1, 1, 0, 1, 2] + \\
{[}TC2 Name embedding{]} }}
\\ \\ \bottomrule
\end{tabular}
\end{table*}
In the example, there are two test cases (TC1 and TC2). TC1 contains four steps (TS1, TS2, TS3, TS4) and TC2 contains five steps (TS1, TS5, TS6, TS7, TS8). As we can see, only the TS1 step is shared between the test cases. In the test step cluster column, we can see the cluster ID to which each step belongs (TS1 belongs to the C1 cluster, TS2 belongs to the C2 cluster, and so on), where \textit{Cn} is the ID of the cluster $n$. Note that different steps (such as TS2 and TS7) might belong to the same cluster (C2). Next, we explain each proposed technique using this example.
\noindent\textbf{Technique 1: Test step cluster overlap.}
For this technique, we used only the identifiers of the test step clusters to represent test cases. For each test case, we gathered the unique list of cluster IDs that contain the test steps. For our running example, the TC1 test case is represented through the \textit{[C1, C2, C3]} vector, while TC2 is represented through the \textit{[C1, C2, C4, C5]} vector. Finally, we computed the pairwise similarity of any two test cases using a simple overlap metric, which indicates the proportion of overlap that test cases have in terms of test step cluster IDs, as shown below:
\begin{equation}
\text{Overlap} = \frac{ length((TCn) \cap (TCm)) }{ max(length(TCn),length(TCm)) }
\end{equation}
Where $TCn$ and $TCm$ correspond to the representations of the test cases $n$ and $m$ through the unique cluster IDs, respectively. Intuitively, test cases that have a large overlap of test step clusters (even if the test steps themselves are different) should be similar since test steps in the same cluster are (most of the time) similar. For our example, the length of TC1 is three (C1, C2, C3), the length of TC2 is four (C1, C2, C4, C5), and the length of the intersection between TC1 and TC2 is two (C1, C2). Therefore, the overlap between the TC1 and TC2 test cases is: \( \frac{2}{max(3,4)} = \frac{2}{4} = 0.5\) (50\%).
We used the computed overlap as the similarity metric to compare the test cases. Furthermore, in order to determine the optimal similarity threshold (i.e., with the optimal tradeoff between false positives and false negatives) to be used to identify similar test cases, we performed a search by varying the threshold from 0.1 (10\% of overlap) up to 1.0 (100\% of overlap). Figure~\ref{fig:technique_1} shows how the F-score changes with the similarity threshold (the optimal threshold is indicated with the vertical red line). As we can see, our search showed that the threshold that provides the maximum F-score is 0.70, which means that two test cases should be considered similar if their overlap metric is at least 70\%.
\begin{figure*}[]
\centering
\begin{subfigure}{0.235\textwidth}
\centering
\includegraphics[width=0.95\linewidth]{appr_1.pdf}
\ExerciseCaption{F-score for different similarity thresholds for Technique 1.}
\label{fig:technique_1}
\end{subfigure} \hspace{0.01\textwidth} %
\begin{subfigure}{0.235\textwidth}
\centering
\includegraphics[width=0.95\linewidth]{appr_2.pdf}
\ExerciseCaption{F-score for different similarity thresholds for Technique 2.}
\label{fig:technique_2}
\end{subfigure}\hspace{0.01\textwidth} %
\begin{subfigure}{0.235\textwidth}
\centering
\includegraphics[width=0.95\linewidth]{appr_3.pdf}
\ExerciseCaption{F-score for different similarity thresholds for Technique 3.}
\label{fig:technique_3}
\end{subfigure}\hspace{0.01\textwidth} %
\begin{subfigure}{0.235\textwidth}
\centering
\includegraphics[width=0.95\linewidth]{appr_4.pdf}
\ExerciseCaption{F-score for different similarity thresholds for Technique 4.}
\label{fig:technique_4}
\end{subfigure}
\caption{F-score for different similarity thresholds for our four proposed techniques. The vertical red line indicates the threshold that maximizes the F-score.}
\label{fig:similarity_search_plots}
\end{figure*}
\noindent\textbf{Technique 2: Boolean representation of test cases.}
Similarly to Technique 1, for Technique 2 we used the test step clusters to represent test cases. However, instead of using the cluster IDs directly, we used a Boolean vector for each test case, in which we flagged the clusters that contain at least one test step of that case with a ``1''. Otherwise, we used ``0''. For our example, both test cases TC1 and TC2 are represented through a vector of length five because there are five different test step clusters in total (C1, C2, C3, C4, C5). TC1 is represented through the \textit{[1,1,1,0,0]} vector (because TC1 has steps that belong to the clusters C1, C2, and C3, but no step belongs to the clusters C4 and C5), while TC2 is represented through the \textit{[1,1,0,1,1]} vector. We built a matrix of dimension \textit{[\#test\_cases x \#test\_step\_clusters]}, where each row corresponds to a test case and each column corresponds to a test step cluster. Finally, we computed the pairwise similarity of any two test cases using the Jaccard index, as used in prior work~\cite{abreu2007accuracy, abreu2009practical} to calculate the similarity between Boolean vectors. Our search (see Figure~\ref{fig:technique_2}) shows that the optimal lower threshold for the Jaccard index is 0.60, which means that two test cases are similar if their Jaccard index is equal or larger than 0.60.
\noindent\textbf{Technique 3: Numeric representation of test cases.}
Using a Boolean vector to represent test cases might not be sufficient for situations where test cases have more than one step in a cluster. Therefore, we modified the previous technique so that, instead of representing test cases as a Boolean vector, we represent test cases as a numeric vector. This numeric vector corresponds to the number of test steps that the test case has in each cluster. For our example, TC1 is represented through the \textit{[2,1,1,0,0]} vector (because TC1 has two steps in the C1 cluster, one step in each of the C2 and C3 clusters, and no step in the C4 and C5 clusters). TC2 is represented through the \textit{[1,1,0,1,2]} vector.
We found that using a threshold of 0.85 (see Figure~\ref{fig:technique_3}) achieved the best performance in terms of F-score. This means that all the pairs of test cases that have a cosine similarity equal or larger than 0.85 are considered similar by Technique 3.
\begin{table*}[!t]
\centering
\caption{F-score of the test case similarity techniques along with the optimal similarity threshold. }
\label{tab:f_score_test_case_similarity}
\begin{tabularx}{1.5\columnwidth}{lp{5cm}rp{3.8cm}} \toprule
\textbf{Technique} & \textbf{Technique name} & \multicolumn{1}{r}{\textbf{F-score}} & \multicolumn{1}{r}{\textbf{Optimal similarity threshold}} \\ \midrule
- & Baseline 1 & 47.35 & \multicolumn{1}{r}{-} \\ [1pt]
- & Baseline 2 & 0.35 & \multicolumn{1}{r}{-} \\ [1pt]
Technique 1 & Test step cluster overlap & 80.54 & \multicolumn{1}{r}{0.70} \\ [1pt]
Technique 2 & Boolean representation of test cases & 76.90 & \multicolumn{1}{r}{0.60} \\ [1pt]
Technique 3 & Numeric representation of test cases & 77.42 & \multicolumn{1}{r}{0.85} \\ [1pt]
\multirow{2}{*}{\textbf{Technique 4}} & \textbf{Numeric representation of test cases with test case name embedding} & \multirow{2}{*}{\textbf{83.47}} &
\multicolumn{1}{r}{\multirow{2}{*}{\textbf{0.75}}} \\ \bottomrule
\end{tabularx}
\end{table*}
\noindent\textbf{Technique 4: Numeric representation of test cases with test case name embedding.}
Our last technique is similar to Technique 3, but here we included information about the test case name as well. For our example, both TC1 and TC2 are represented through the same vectors as for Technique 3. In this technique, we combined the test step clusters with the test case name embedding. To obtain the embeddings for the name, we used the best-performing text embedding technique from the experiments for test step clustering (which is Word2Vec). Following a similar process as we did for the test step clustering, we computed the pairwise distance for any two test case name embeddings. We computed the weighted average between the test case name distance and the similarity score between the clusters of steps to obtain the final similarity score for the test cases. In order to determine the best weights for the test case name and test step cluster distances, and to determine the optimal threshold for the cosine similarity of the final distance (as we did for the Techniques 1, 2, and 3), we performed a search similarly to the previous techniques. We found that the test case name and the test step clusters must contribute equally (i.e., a weight of 50\% for each) to achieve the highest performance. Furthermore, as Figure~\ref{fig:technique_4} shows, using a threshold of 0.75 for this similarity score achieved the best performance, which means that all the pairs of test case pairs that have a cosine similarity equal or larger than 0.75 are considered similar by Technique 4. Note that, due to space constraints and for a better visualization, Figure~\ref{fig:technique_4} only displays how the F-score changes with the threshold already using the optimal weight of 50\%.
\noindent\textbf{Baseline.}
We compared the performance of the four approaches with two baselines. The first baseline considers two test cases to be similar if they have the exact same steps (regarding the text of the step). The second baseline considers two test cases to be similar if they have the same name.
\subsection{Evaluation metric.}
To evaluate our approaches for finding similar test cases, we used the manually built ground truth of similar test cases to compute the F-score. We followed the exact same process as we did previously for the test step clustering (Section~\ref{sec:evaluation_metric}).
\subsection{Findings}
\textbf{Clusters of similar test steps and test case name embeddings together can be used to identify similar test cases with a high performance.} Table~\ref{tab:f_score_test_case_similarity} presents the F-score of all the techniques that we evaluated along with the optimal similarity threshold.
We observe that Technique 4 (\textbf{Numeric representation of test cases with test case name embedding}) achieves the highest performance in terms of F-score (83.47\%), followed by Technique 1 (\textbf{Test step cluster overlap}), which achieves an F-score of 80.54\%. We also observe that, even though Technique 3 (\textbf{Numeric representation of test cases}) achieves a higher performance than Technique 2 (\textbf{Boolean representation of test cases}), the improvement is very small (0.52 in absolute percentage point). This indicates that using the number of test steps in each cluster (instead of just flagging whether the cluster contains a test step) slightly improves the performance of the test case similarity technique. Further incorporating the test case name information significantly improves the performance of the technique. Regarding the baselines, we observe that Baseline 2 (test cases with the same name) achieves an extremely low F-score (0.35\%), while Baseline 1 (test cases with the same steps) achieves a better, but still low F-score (47.35\%). Furthermore, all the experimented techniques perform considerably better than both baseline methods.
For the similarity threshold, we observe that each technique has a different optimal threshold. For example, even though we may expect that a higher overlap would achieve a higher performance, this does not hold true, as we see that the optimal overlap threshold (Technique 1) is 0.70. We also observe that the threshold for the highest-performing technique (Technique 4) is 0.75. Using this similarity threshold, Technique 4 found that 64.8\% of the test cases (which corresponds to 2,153 test cases) have at least one similar test case and there are, in total, 429 groups of similar test cases. On average, each cluster has two similar test cases, with a standard deviation of four.
\begin{table*}[t]
\centering
\caption{Examples of the four types of test case similarity. Differences between test cases' steps are highlighted in bold.}
\label{tab:test_case_similarity_type}
\begin{tabular}{lll} \toprule
\textbf{Similarity type} & \textbf{Test case name}
& \textbf{Test steps} \\ \midrule
\multirow{4}{*}{\begin{tabular}[c]{@{}l@{}}(1) Same steps for different \\game assets\end{tabular}}
& \multirow{2}{*}{Check Hat - In Backpack}
& 1. Verify item name \\
& & 2. Verify item icon \\ \cmidrule{2-3}
& \multirow{2}{*}{Check Wand - In Backpack}
& 1. Verify item name \\
& & 2. Verify item icon \\ \midrule
\multirow{4}{*}{\begin{tabular}[c]{@{}l@{}}(2) Slightly different steps \\for different game assets\end{tabular}}
& \multirow{2}{*}{Equip Hat}
& 1. Trigger equip functionality via \textbf{backpack hat} item slot \\
& & 2. Trigger unequip functionality via \textbf{backpack hat} item slot
\\ \cmidrule{2-3}
& \multirow{2}{*}{Equip Wand}
& 1. Trigger equip functionality via \textbf{wand backpack} item slot \\
& & 2. Trigger unequip functionality via \textbf{wand backpack} item slot
\\ \midrule
\multirow{6}{*}{\begin{tabular}[c]{@{}l@{}}(3) Test cases with\\ additional/missing steps\end{tabular}}
& \multirow{4}{*}{Check Consumables (Water Resist)}
& 1. Use in battle \\
& & 2. Check battle bonus \\
& & 3. \textbf{Check item card name} \\
& & 4. \textbf{Check item card stats}
\\ \cmidrule{2-3}
& \multirow{2}{*}{Check Food (Popcorn)}
& 1. Use in battle
\\
& & 2. Check battle bonus \\ \midrule
\multirow{2}{*}{\begin{tabular}[c]{@{}l@{}}(4) Redundant test cases\end{tabular}}
& Catch Firefly in Forest
& 1. \textbf{Catch firefly in forest}
\\ \cmidrule{2-3}
& Firefly Forest - Catch Firefly
& 1. \textbf{Catch a firefly}
\\ \bottomrule
\end{tabular}
\end{table*}
Aiming at further understanding the output (groups of similar test cases) produced by our best technique (Technique 4), we manually inspected a representative sample of 100 of the obtained groups of similar test cases. We identified four main types of similar test cases, as shown in Table~\ref{tab:test_case_similarity_type}. While Type 1 corresponds to test cases with the same steps for different game assets, Type 2 regards test cases that have slightly different steps to indicate the asset being tested (e.g., \textbf{backpack hat} and \textbf{backpack wand}). Type 3 refers to test cases with a large overlap of steps but one of them has more/less steps, which might indicate extra (unnecessary) or missing steps. Finally, Type 4 regards redundant test cases, which are written differently and may have a different number of steps, but the testing objective is the same. The last type of similarity helps to identify test cases that might be completely removed from the test suite.
\section{Discussion}\label{sec:discussions}
In this section, we revisit the research questions and discuss the validation of our approach.
\noindent\textbf{\rqone}\\
Our experiments demonstrate that we can identify similar test steps with a high performance in terms of F-score. We showed that an ensemble approach using a combination (majority voting) of different techniques (five text embedding techniques with two similarity metrics and two clustering algorithms) achieves the highest performance. Such ensemble approach has a large computational cost as it requires the execution of several different techniques. However, we showed that using a single technique (such as Word2Vec or TF-IDF) can also provide a high performance while being less computationally expensive.
\noindent\textbf{\rqtwo}\\
Our experiments demonstrate that we can use the clusters of similar test steps identified in the first part of the study to represent test cases and identify the similar ones. More specifically, representing test cases through a vector that captures the number of test steps in each cluster boosts the similarity technique performance. Furthermore, we showed that combining the clusters of similar test steps with the embedding of the test case name achieves an even higher performance. Our experiments showed that the optimal weight (for our data) for the clusters and the test case names is 50\%. In addition, we can use a threshold of 0.75 for the similarity score to decide whether two test cases are similar.
\noindent\textbf{Validation with developers.}
To validate the results of our approach, we did an informal interview with a QA expert at Prodigy Education to discuss whether our results are valid and how they can be used in practice and improve the testing process. We selected a purposive sample~\cite{tongco2007purposive,guarte2006estimation,etikan2016comparison} to explicitly select test cases that cover the different types of similar test cases that we identified.
Overall, the expert validated the different types of test case similarity that we identified and mentioned that our approach can help the QA to improve the quality of the test cases. More specifically, the QA expert pointed out five practical usages of our approach, as we explain next. First, our approach can be used to identify redundant test cases. Such cases occur when test cases are described differently (e.g., because they were written by different professionals) but test exactly the same aspect/asset of the game. The second usage regards the reuse of existing test cases when creating new ones for new features of the game. In this case, existing test cases can either be fully or partially (e.g., a few test steps) reused. By reusing test cases, the overall quality of the test suite improves with more consistent and homogeneous descriptions in terms of terminology. Furthermore, reusing test cases reduces the manual effort and time required for designing and creating new test cases. The third application is to use our approach to identify legacy test cases, such as test cases that were created for a temporary feature and were kept in the test case base, even though the feature does not exist anymore. Such legacy test cases have a high impact in the (already time-consuming) manual testing effort. The fourth usage is to identify test cases with missing steps. A few test case samples that we discussed with the expert were indeed groups of similar test cases which perform the same task, but some of the cases had less steps than what is actually performed by a tester. We further investigated those cases with the QA expert and found out that the missing steps were scattered across the test suite (in different cases) and should be merged with the steps of the main test case. Finally, the fifth application is to identify test cases which are redundant but one of the cases has additional steps. This occurs when new test cases are created based on existing ones, but some steps are added for clarification purposes and the older test case is not removed from the test suite.
\section{Threats to Validity}\label{sec:threats}
\textbf{External validity} relates to the generalizability of our findings. One threat is regarding the methods that we used for text embedding, text similarity and clustering. Although we used five different text embedding techniques, three similarity metrics and two clustering algorithms, different results might be achieved with other methods. Future studies should further investigate additional methods for text embedding, text similarity and clustering. Another threat is that our findings rely on the test case descriptions of an educational game company. Test cases of organizations from different domains might be different (e.g., in terms of the used terminology and grammar complexity and structure) and might affect the results. Finally, our thresholds for optimal values (such as the number of clusters and the similarity metric) likely do not apply to other systems. However, our method for conducting the search for these values is generalizable
\textbf{Internal validity} concerns the bias and errors due to the experimental design. One threat is related to the manual analysis of the samples of test steps (to build the ground truth for the test step clustering) and test cases (to build the ground truth for the test case similarity). The manual analysis is subject to error and bias because of human factors. Furthermore, despite selecting statistically representative samples, the characteristics of the whole population may not be represented in those samples.
\section{Conclusion}\label{sec:conclusion}
Test cases written in natural language are often defined by different people who may use different terminology to refer to the same concept. As a result, many similar or redundant test cases may exist in the test suite, which increases the manual testing effort and the usage of development resources. Since manually identifying redundant test cases is a time-consuming task, an automated technique is necessary.
In this paper, we propose an approach to identify similar test cases specified in natural language. First, we evaluated different text embedding techniques, similarity metrics, and clustering algorithms to identify clusters of similar test steps (which compose test cases). We then leveraged the identified test step clusters together with the test case name to identify similar test cases. To evaluate the approach, we used test cases from an educational game company. We manually built a ground truth of similar test steps and test cases and computed the F-score metric. The approach evaluation shows that similar test steps can be identified with a high performance (an F-score of 87.39\%) using an ensemble approach which consists of different NLP techniques. We can also achieve a similar performance (an F-score of 86.99\%) using a single technique (Word2Vec). Furthermore, we identified similar test cases with a high performance (an F-score of 83.47\%) using clusters of similar test steps combined with the similarity between test case names.
In this work, we show how we can identify similar test cases based only on their description in natural language with an unsupervised approach, which requires no labelled data nor human supervision. As indicated in an informal interview with a QA engineer, our approach has several usages in practice, such as supporting QA and developers to identify and remove redundant and legacy test cases from the test suite. Furthermore, existing groups of similar test cases can be leveraged to create new test cases and help to maintain a more consistent and homogeneous terminology across the test suite.
\section*{Acknowledgments}
The research reported in this article has been supported by Prodigy Education and the Natural Sciences and Engineering Research Council of Canada under the Alliance Grant project ALLRP 550309.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\footnotesize
\bibliographystyle{IEEEtranSN}
|
\section{Introduction}
Most classical and quantum theories of modern physics are formulated
using the standard Lagrangians and Lagrangian formalism [1].
There are also nonstandard Lagrangians, whose applications to physics
are considered in [2], and null Lagrangians, whose relevance to
ordinary differential equations (ODEs) has already been studied [3].
The main objective of this paper is to derive a new set of nonstandard
null Lagrangians and their gauge functions, make them exact, and apply
them to the law of inertia.
In the calculus of variations, the action $\mathcal{A} [x (t)]$, where
$x (t)$ is a dynamical variable (or classical object's trajectory) that
depends on time $t$, is defined as an integral over a local real
function called a Lagrangian [4], which, for a second-order
ODE, is denoted as $L (\dot x, x, t)$, where the time derivative
$\dot x $ represents the particle's velocity in dynamics. The
Hamilton principle [1,4]
requires that $\mathcal{A} [x ]$ be stationary (to have either a
minimum or maximum or saddle point), which is mathematically
expressed as $\delta \mathcal{A} = 0$, where $\delta$ is the
functional (Fr\'echet) derivative of $\mathcal{A} [x (t)]$ with
respect to $x $. The necessary condition that $L (\dot x, x, t)$
satisfies the Hamilton principle is $\hat {EL} [ L (\dot x, x, t) ] = 0$,
where $\hat {EL}$ is the Euler--Lagrange operator [4].
A general second-order ODE with constant coefficients is of the
form $\hat D x (t) = 0$, where $\hat D = d^2 / dt^2 + b d / dt
+ c $ is a linear operator and $b$ and $c$ are constants. Let
$\hat D_o$ be the operator with $b = c = 0$, and $\hat D_c$
be the operator with $b = 0$. Then, for $\hat D_o x (t) = 0$,
its Lagrangian depends only on $\dot x^2$. However, for
$\hat D_c x (t) = 0$, its Lagrangian depends on both $\dot
x^2$, the kinetic energy-like term, and $x^2$, the potential
energy-like term, as originally shown by Lagrange [5].
Moreover, if $\hat D x (t) = 0$, its Lagrangian becomes the
Caldirola--Kanai Lagrangian [6,7]. Thus, Lagrangians
that depend either on $\dot x^2$ or on $\dot x^2$ and
$x^2$ are called {\it standard} Lagrangians (SLs).
Since the standard Lagrangians are not unique, it is also possible
to construct other Lagrangians that typically depend on $\dot x $
and $x $ but not on powers of these variables, and depend also
on arbitrary functions of the independent variable $t$; Arnold [8]
refers to such Lagrangians as non-natural Lagrangians.
Here, Lagrangians that depend on $\dot x $ and $x $, and on
functions of $t$ are called {\it nonstandard} Lagrangians (NSLs).
The existence of the standard and nonstandard Lagrangians is
guaranteed by the Helmholtz conditions [9]. The procedure
of finding these Lagrangians for given ODEs is called the inverse
(or Helmholtz) problem of the calculus of variations [10,11].
There are different methods to find the standard [12,13,14,15,16]
and nonstandard [15,16,17,18,19,20] Lagrangians. Generalized
nonstandard Lagrangians can also be obtained [21] and applied
to the ODEs, whose solutions are special functions of mathematical
physics [22]. Other generalizations of the nonstandard Lagrangians
have been applied to the Riccati equation [23] and to a Liénard-type
nonlinear oscillator [24].
In addition to the SLs and NSLs, there is also a family of null (or trivial)
Lagrangians (NLs) for which the Euler-Lagrange (E-L) equation vanishes
identically [3]. Another property of these NLs is that they can be
expressed as the total derivative of a scalar function [3], which is
called a gauge or gauge function [25,26]; in other words, all NLs
have their corresponding gauge functions (GFs). The NLs were constructed
and investigated in mathematics, specifically in Cartan and Lapage forms,
symmetries of Lagrangians, in Carath\'eodory's theory of fields, and
extremals and integral invariants [3,27,28,29,30,31,32]. The NLs
were also applied to elasticity, where they represent the energy density
function of materials [33,34].
The fact that the NLs and their GFs may also play an important role in
physics was shown recently by using them to restore Galilean invariance
of the standard Lagrangian for Newton's laws of dynamics [35,36]
and to add forces to an undriven harmonic oscillator [37]. Since
those previously constructed NLs resemble the standard Lagrangians, they
are called here {\it standard} NLs, and their corresponding GFs become
{\it standard} GFs. In this paper, the standard and nonstandard NLs and
their GFs are constructed and applied to the law of inertia in Galilean space
and time with the Galilean group of the metric [8,25].
The outline of the paper is as follows. In Section 2, previously
obtained standard null Lagrangians and their gauge functions are described.
In Section 3, the derived new nonstandard null Lagrangians
are presented and compared to the standard null Lagrangians. Invariance
of the action is used to define the exact nonstandard gauge functions in Section
4. Applications of the obtained results to the Newtonian law of
inertia are presented and discussed in Section 5. Section 6
summarizes the results of this paper.
\section{Standard Null Lagrangians} \label{section2}
For the considered ODEs of the form $\hat D x (t) = 0$, its standard Lagrangian
is given by
\begin{equation}
L_{\rm s} (\dot x , x, t) = {1 \over 2} ( \dot x^2 - c x^2 ) e^{bt}\ .
\label{S2eq1}
\end{equation}
This Lagrangian was first derived by Caldirola [6] and Kanai [7],
and it reduces to that given originally by Lagrange [5], if $b=0$.
One of the well-known null Lagrangians [4] is
\begin{equation}
L_{sn1} (\dot x , x ) = c_1 \dot x x \ ,
\label{S2eq2}
\end{equation}
where $c_1$ is an arbitrary constant. In this Lagrangian, the power of the
dependent variables is the same as in the standard Lagrangian given by
Equation (\ref{S2eq1}); however, the dependent variable and its derivative
are mixed.
Recently [35], $L_{sn1} (\dot x , x )$ was generalized to
\begin{equation}
L_{sn2} (\dot x , x , t) = c_1 \dot x x + c_2 ( \dot x t + x )
+ c_3 \dot x + c_4\ ,
\label{S2eq3}
\end{equation}
where $c_2$, $c_3$, and $c_4$ are arbitrary constants, and $L_{sn2}$
becomes $L_{sn1}$ if $c_2 = c_3 = c_4 = 0$. This Lagrangian was
constructed based on the principle that the power of the terms with the
dependent or independent variable, or their combination, does not
exceed the power of the terms in the original standard Lagrangian
given by Equation (\ref{S2eq1}).
Since
\begin{equation}
L_{sn2} (\dot x , x , t) = {{d \Phi_{sn2} ( x , t )} \over
{dt}}\ ,
\label{S2eq4}
\end{equation}
the gauge function $\Phi_{sn2} (x , t) $ is given \cite{35} by
\begin{equation}
\Phi_{sn2} (x , t) = {1 \over 2} c_1 x^2 + c_2 x t +
c_3 x + c_4 t\ .
\label{S2eq5}
\end{equation}
Following [36], the derived $\Phi_{sn2} (x , t)$ is generalized by
replacing its constant coefficients by arbitrary functions that depend
only on the independent variable. Then, the standard GF, $\Phi_{sn3}
(x , t)$, can be written as:
\begin{equation}
\Phi_{sn3} (x , t) = {1 \over 2} f_1 (t) x^2 + f_2 (t)
x t + f_3 (t) x + f_4 (t) t\ .
\label{S2eq6}
\end{equation}
Since $\Phi_{sn3} (x , t) $ is a function of the variables $x $
and $t$, and its total derivative results in the following general
standard null Lagrangian
\[
L_{sn3} (\dot x , x , t) = [ f_1 (t) \dot x + {1 \over 2}
\dot f_1 (t) x ] x + \left [ f_2 (t) \dot x + \dot f_2
(t) x \right ] t
\]
\begin{equation}
\hskip0.25in + f_2 (t) x + \left [ f_3 (t) \dot x +
\dot f_3 (t) x \right ] + \left [ f_4 (t) + \dot f_4 (t) t \right ]\ ,
\label{S2eq7}
\end{equation}
where $f_1 (t)$, $f_2 (t)$, $f_3 (t)$, and $f_4 (t)$ are arbitrary but
at least twice differentiable functions of the independent variable [37].
Additional constraints on these functions are presented in Section 4,
where invariance of the action is considered.
The generalization of the gauge function given by Equation (\ref{S2eq6}),
is one natural way to obtain a new NL, but there is also another way,
namely, by replacing the constant coefficients in $L_{sn2} (\dot x ,
x , t)$ (see Equation \eqref{S2eq3}) by the functions $f_1 (t)$, $f_2 (t)$,
$f_3 (t)$, and $f_4 (t)$. The result is:
\begin{equation}
L_{sn4} (\dot x , x , t) = f_1 (t) \dot x x + f_2 (t)
( \dot x t + x ) + f_3 (t) \dot x + f_4 (t)\ .
\label{S2eq8}
\end{equation}
Applying $\hat {EL} \{ L_{s,test} (\dot x , x , t) \} = 0$,
it is found that $L_{sn4} (\dot x , x , t) $ is a NL if, and only if,
the following condition
\begin{equation}
\dot f_1 (t) x + \dot f_2 (t) t + \dot f_3 (t) = 0\ .
\label{S2eq9}
\end{equation}
on the functions $f_1 (t)$, $f_2 (t)$, and $f_3 (t)$
is imposed.
There are several different solutions to Equation (\ref{S2eq9}); the
simplest one is $f_1 (t) = c_1$, $f_2 (t) = c_2$, and $f_3 (t) = c_3$,
which reduces $L_{sn4} (\dot x , x , t) $ to $L_{sn2} (\dot x , x ,
t)$ without any generalization, but with an additional requirement
that $f_4 (t) = c_4$. More interesting cases are: (i) $f_1 (t)
= c_1$, which gives $ \dot f_2 (t) t = - \dot f_3 (t)$; (ii) $f_2 (t)
= c_2$ and $ \dot f_1 (t) x = - \dot f_3 (t)$; and (iii) $f_3 (t)
= c_3$ with $ \dot f_1 (t) x = - \dot f_2 (t) t$. In all three
cases, three new standard NLs are obtained.
Since the functions $f_1 (t)$, $f_2 (t)$, and $f_3 (t)$ in $L_{sn4}
(\dot x , x , t) $ are limited by the auxiliary condition, given by
Equation (\ref{S2eq9}), and since $L_{sn3} (\dot x , x , t) $ does
not require any restrictions on these functions, the standard NL,
given by Equation (\ref{S2eq7}), is more general than
$L_{sn4} (\dot x , x , t) $; thus, the standard general NL
becomes $L_{sgn} (\dot x , x , t) = L_{sn3}
(\dot x , x , t) $.
The following Corollary summarizes (without a formal proof)
the results obtained in this Section.
{\bf Corollary:}
The Lagrangian $L_{sgn} (\dot x , x , t) $ is the general null Lagrangian
among all null Lagrangians that can be constructed
based on the principle that the power of the dependent variable
cannot exceed the power of this variable in the SL, given by Equation
(\ref{S2eq1}).
\section{Nonstandard Null Lagrangians} \label{section3}
Any Lagrangian different from $L_{\rm s} (\dot x , x)$ is a nonstandard
Lagrangian. Among different known nonstandard Lagrangians,
the most commonly used [15,16,17,18,19] is:
\begin{equation}
L_{ns} (\dot x , x , t) = {{1} \over {g_1 (t) \dot x +
g_2 (t) x + g_3 (t)}}\ ,
\label{S3eq1}
\end{equation}
where $g_1 (t)$, $g_2 (t)$, and $g_3 (t)$ are arbitrary and
differentiable functions to be determined.
Since there are no nonstandard NLs in the literature, the
objective of this paper is to find them. The procedure is
based on the two following conditions. First, for a null
Lagrangian to be called {\it nonstandard}, it must be of
different form than the standard NLs, given by
\mbox{Equations (\ref{S2eq3}) and (\ref{S2eq7}}), and
its form must be similar to that of Equation (\ref{S3eq1}).
The latter means that it must contain $\dot x $, $x $, and
arbitrary functions of $t$, or constants. The second condition
is similar to that used to
construct $L_{sn2} (\dot x , x , t) $, $L_{sn3} [\dot x , x ,
t]$ and $L_{sn4} (\dot x , x , t) $, namely, the power of the
dependent variable and its derivative must not exceed their
order in the nonstandard Lagrangian given by Equation
(\ref{S3eq1}). The obtained nonstandard null Lagrangians
are presented in Propositions 1 and 2, and in the Corollaries
that follow them.
{\bf Proposition:}
Let $a_1$, $a_2$, $a_3$, and $a_4$ be constants in the following
nonstandard test-Lagrangian
\begin{equation}
L_{ns,\rm test1} (\dot x , x , t) = {{a_1 \dot x } \over {a_2 x + a_3 t
+ a_4}}\ .
\label{S3eq2}
\end{equation}
Then, $L_{ns,\rm test1} (\dot x , x , t) $ is a null Lagrangian
if, and only if, $a_3 = 0$.
{\bf Proof:}
Since this Lagrangian must
satisfy the E-L equation, $\hat
{EL} \{ L_{ns,\rm test1} (\dot x , x , t) \} = 0$, the required condition
is $a_1 a_3 = 0$. With $a_1 \neq 0$, then $a_3 = 0$, and $L_{ns,\rm test1}
(\dot x , x , t) = L_{nsn1} (\dot x , x , t) $, where the latter is
the nonstandard NL. This concludes the proof.
{\bf Corollary:}
Let $L_{nsn1} [\dot x , x ]$ be the nonstandard null Lagrangian
given
by:
\begin{equation}
L_{nsn1} (\dot x , x , t) = {{a_1 \dot x } \over {a_2 x + a_4}}\ ,
\label{S3eq3}
\end{equation}
then its gauge function
$\Phi_{nsn1} ( x )$ is:
\begin{equation}
\Phi_{nsn1} ( x ) = {{a_1} \over {a_2}} \ln \vert a_2 x + a_4 \vert \ .
\label{S3eq4}
\end{equation}
{\bf Corollary:}
Another nonstandard null Lagrangian that can be constructed is $L_{nsn2}
(t) = b_1 / ( b_2 t + b_3 )$ and the corresponding gauge function is
$\Phi_{nsn2} (t) = (b_1 / b_2) \ln \vert b_2 t + b_3 \vert$; however,
the Lagrangian and gauge function do not obey the first condition; thus,
they will not be further considered.
Generalization of $L_{nsn1} (\dot x , x)$ is now presented in
Proposition 2.
{\bf Proposition:}
Let $h_1 (t)$, $h_2 (t)$, and $h_4 (t)$ be at least twice differentiable
functions and $L_{nsn1} (\dot x , x)$ be the nonstandard null Lagrangian
given by Equation (\ref{S3eq3}), with the corresponding nonstandard
gauge function given by Equation (\ref{S3eq4}). A more general
nonstandard null Lagrangian is obtained if, and only if, the constants
in $\Phi_{nsn1} ( x )$ are replaced by the functions $h_1 (t)$, $h_2 (t)$,
and $h_4 (t)$.
{\bf Proof:}
Replacing the constant coefficients $a_1$, $a_2$, and $a_4$ in
$L_{nsn1} (\dot x , x)$ by the functions $h_1 (t)$, $h_2 (t)$,
and $h_4 (t)$, respectively, the resulting Lagrangian is:
\begin{equation}
L_{ns,\rm test2} (\dot x , x , t) = {{h_1 (t) \dot x } \over {h_2 (t) x
+ h_4 (t)}}\ .
\label{S3eq5}
\end{equation}
Using $\hat {EL} \{ L_{ns,\rm test2} (\dot x , x , t) \} = 0$, it is found that
$L_{ns,\rm test2} (\dot x , x , t) $ is the nonstandard NL only when $h_1 (t)
= a_1$, $h_2 (t) = a_2$ and $h_4 (t) = a_4$, which reduces $L_{ns,\rm test2}
(\dot x , x , t) $ to $L_{nsn1} (\dot x , x , t) $ and shows that no
generalization of $L_{nsn1} (\dot x , x , t) $ can be accomplished this
way.
Now, replacing the constant coefficients in $\Phi_{nsn1} ( x )$ by the
functions $h_1 (t)$, $h_2 (t)$ and $h_4 (t)$ generalizes the gauge function
to
\begin{equation}
\Phi_{nsgn} ( x ) = {{h_1 (t)} \over {h_2 (t)}} \ln \vert h_2 (t)
x + h_4 (t) \vert \ .
\label{S3eq6}
\end{equation}
Since the total derivative of any differentiable scalar function that depends
on $x $ and $t$ is a null Lagrangian, the following nonstandard NL is
obtained:
\[
L_{nsgn} (\dot x , x , t) = {{h_1 (t) [ h_2 (t) \dot x +
\dot h_2 (t) x ] + \dot h_4 (t)} \over {h_2 (t) [h_2 (t) x +
h_4 (t)]}}
\]
\begin{equation}
\hskip0.5in + \left [ {{\dot h_1 (t)} \over {h_2 (t)}} - {{h_1
(t) \dot h_2 (t)} \over {h^2_2 (t)}} \right ]\ \ln \vert h_2 (t) x
+ h_4 (t) \vert\ .
\label{S3eq7}
\end{equation}
As expected, $\hat {EL} \{ L_{nsgn} (\dot x , x , t) \} = 0$;
thus, $L_{nsgn} (\dot x , x , t) $ is the general nonstandard
null Lagrangian when compared to Equation \eqref{S3eq3}.
This concludes the proof.
The derived $L_{nsgn} (\dot x , x , t) $ and $\Phi_{nsgn} (\dot
x , x , t)$ represent new families of nonstandard general NLs
and their GFs. These NLs and GFs were derived based on the
condition that the power of the dependent variable in the
nonstandard general NLs is either the same as, or lower than,
that displayed in the original NSL given by Equation (\ref{S3eq1}).
\section{Action Invariance and Conditions for Exactness} \label{section4}
Since the functions in the standard and nonstandard general null
Lagrangians are arbitrary, they require either mathematical or physical
constraints, or both. Among possible mathematical constraints
is invariance of the action, which is used to introduce exact gauge
functions [36], and symmetries of Lagrangians and the resulting
dynamical equations [38,39,40,41,42]. Moreover, by using
Galilean invariance [35], the additional physical constraints
are imposed on the GFs [36]. Here, only the invariance of
the action is applied to the obtained standard and nonstandard
general NLs and GFs; symmetries of Lagrangians are also briefly
discussed.
In the calculus of variations, the action is defined as:
\[
A [x ; t_e, t_o] = \int^{t_e}_{t_o} ( L + L_{\rm null} ) dt =
\int^{t_e}_{t_o} L dt + \int^{t_e}_{t_o} \left [ {{d \Phi_{\rm null}
(t)} \over {dt}} \right ] dt
\]
\begin{equation}
\hskip0.25in = \int^{t_e}_{t_o} L dt + [ \Phi_{\rm null} (t_e) -
\Phi_{\rm null} (t_o)]\ ,
\label{S4eq1}
\end{equation}
where $t_o$ and $t_e$ denote initial and final times, $L$ is a
Lagrangian that can be either any SL or any NSL, $L_{\rm null}$
is any null Lagrangian and $\Phi_{\rm null}$ is its gauge function.
Since both $\Phi_{\rm null} (t_e)$ and $\Phi_{\rm null} (t_o)$ are
constants, they do not affect the Hamilton principle that requires
$\delta A [x ] = 0$. However, the requirement that $\Delta
\Phi_{\rm null} = \Phi_{\rm null} (t_e) - \Phi_{\rm null} (t_o)$
= const adds this constant to the value of the action. In other
words, the value of the action is affected by the gauge function.
Using invariance of the action, the following definitions are introduced.
{\bf Definition:}
A null Lagrangian, whose $\Delta \Phi_{\rm null} = 0$, is called
the {\it exact} null Lagrangian (ENL).
{\bf Definition:}
A gauge function with
$\Delta \Phi_{\rm null} = 0$
is called the {\it exact} gauge function (EGF).
The condition $\Delta \Phi_{\rm null} = 0$ is satisfied when either
$\Phi_{\rm null} (t_e) - \Phi_{\rm null} (t_o) = 0$, or $\Phi_{\rm null}
(t_e) = 0$ and $\Phi_{\rm null} (t_o) = 0$; let the latter be valid. Then,
the exact null Lagrangians are those whose exact gauge functions make
the action invariant.
Invariance of the action may now be used to establish constraints on
the arbitrary functions in the standard NL, $L_{sgn} (\dot x , x , t) $
(see Equation \eqref{S2eq7}), and its gauge function, $\Phi_{sgn} (x , t) $
(see \eqref{S2eq6}), and make them exact. Taking $\Phi_{sgn} (t_e)
= 0$ and $\Phi_{sgn} (t_o) = 0$, the following conditions are obtained:
\begin{equation}
{1 \over 2} f_1 (t_e) x_e^2 + f_2 (t_e) x_e t_e + f_3 (t_e) x_e +
f_4 (t_e) t_e = 0\ ,
\label{S4eq2}
\end{equation}
and
\begin{equation}
{1 \over 2} f_1 (t_o) x_o^2 + f_2 (t_o) x_o t_o + f_3 (t_o) x_o +
f_4 (t_o) t_o = 0\ ,
\label{S4eq3}
\end{equation}
with $x_e = x (t_e)$ and $x_o = x (t_o)$ denoting the end points. If
the arbitrary functions satisfy these conditions, then $L_{sgn} (\dot x ,
x , t)$ and $\Phi_{sgn} (x , t) $ are exact. The first condition may
be solved by taking $f_3 (t_e) = - f_1 (t_e) x_e / 2$, and $f_4 (t_e) =
- f_2 (t_e) x_e$. Similar solutions are valid for $t_o$ showing
that the end values for the functions can be related to each other.
Applying the same procedure to $L_{nsgn} (\dot x , x , t)$, (see
Equation (\ref{S3eq7})), and to the resulting general gauge function
$\Phi_{nsgn} (\dot x , x , t)$, (see Equation (\ref{S3eq6})),
the conditions on the arbitrary functions are:
\begin{equation}
\left [ {{h_1 (t_e)} \over {h_2 (t_e)}} \right ] \ln \vert h_2 (t_e) x_e +
h_4 (t_e) \vert = 0\ ,
\label{S4eq4}
\end{equation}
and
\begin{equation}
\left [ {{h_1 (t_o)} \over {h_2 (t_o)}} \right ] \ln \vert h_2 (t_o) x_o +
h_4 (t_o) \vert = 0\ .
\label{S4eq5}
\end{equation}
Since $\ln [ h_2 (t_e) x + h_4 (t_e)] \neq 0$ and $\ln [ h_2 (t_e) x
+ h_4 (t_e)] \neq 0$, both conditions set up stringent limits on the function
$h_1 (t)$, whose end values must be: $h_1 (t_e) = 0$ and $h_1 (t_o) = 0$;
however, the procedure does not impose any constraint either on $h_2 (t)$
or on $h_4 (t)$.
Further constraints on all arbitrary functions that appear in the standard
and non-standard, general, exact null Lagrangians (ENLs) can be imposed
by considering symmetries of these Lagrangians and the resulting dynamical
equations. In general, Lagrangians posses less symmetry than the equations
they generate [38]. Among different symmetries, Noether and non-Noether
symmetries are identified [39,40,41,42]. The presence of NLs does not
affect the Noether symmetries [38,41]; however, it may effect the
non-Noether symmetries [42]. All these symmetries impose new
constraints on the functions.
\section{Applications to Newtonian Law of Inertia} \label{section5}
Let $(x,y,z)$ be a Cartesian coordinate system, and let $t$ be time
in all inertial frames; then the one-dimensional motion of a body in one
inertial frame is given by $\hat D_o x (t) = 0$, which represents the law
of inertia. Let $t_o = 0$ and $t_e = 1$ be the end conditions, and let $x (0)
= x_o = 1$, $x (1) = x_e = 2$ and $\dot x (0) = u_o$ be the initial conditions.
Then, the solution to $\hat D_o x (t) = 0$ is $x = u_o t + 1$.
The {\it standard} Lagrangian for this equation of motion is given by Equation
(\ref{S2eq1}), with the coefficients $b = c = 0$, and no arbitrary function to
be determined. However, the standard general NL and the corresponding GF
are given by Equations (\ref{S2eq7}) and (\ref{S2eq6}), respectively. To
make the NL and GF exact, the following conditions (see Equations
(\ref{S4eq2}) and (\ref{S4eq3})) must be imposed on the arbitrary
functions:
\begin{equation}
f_1 (1) + f_2 (1) + f_3 (1) + f_4 (1) = 0\ ,
\label{S5eq1}
\end{equation}
and
\begin{equation}
f_3 (0) = - {1 \over 2} f_1 (0)\ .
\label{S5eq2}
\end{equation}
These conditions guarantee that $L_{sgn} (\dot x , x , t) $
and $\Phi_{sgn} (x , t) $ are the standard general ENL and
the standard general
EGF, respectively.
The {\it nonstandard} Lagrangian
for the law of inertia is presented by
the \linebreak \mbox{following Proposition.}
{\bf Proposition:}
Let $g_1 (t)$, $g_2 (t)$, and $g_3 (t)$ be arbitrary but differentiable
functions, and let $\hat D_o x (t) = 0$ be the equation of motion for
the law of inertia. Then, the nonstandard Lagrangian for this equation
of motion is:
\begin{equation}
L_{ns} (\dot x , x , t) = {{1} \over {C_1 ( a_o t + v_o)^2}}\
{{1} \over {( a_o t + v_o) \dot x - a_o x + C_2}}\ ,
\label{S5eq3}
\end{equation}
where $a_o$, $v_o$, $C_1$, and $C_2$ are constants.
{\bf Proof:}
Following [18], the functions must satisfy
\begin{equation}
{{g_2 (t)} \over {g_1 (t)}} + {1 \over 3} {{\dot g_1 (t)} \over {g_1 (t)}}
= 0\ ,
\label{S5eq4}
\end{equation}
\begin{equation}
{{\dot g_2 (t)} \over {g_1 (t)}} - {1 \over 2} {{\dot g_1 (t)} \over {g_1 (t)}}
{{g_2 (t)} \over {g_1 (t)}} + {{g^2_2 (t)} \over {2 g^2_1 (t)}} = 0\ ,
\label{S5eq5}
\end{equation}
and
\begin{equation}
{{\dot g_3 (t)} \over {g_1 (t)}} - {1 \over 2} {{\dot g_1 (t)} \over {g_1 (t)}}
{{g_3 (t)} \over {g_1 (t)}} + {{g_3 (t)} \over {g_1 (t)}} {{g_2 (t)} \over
{2 g_1 (t)}}= 0\ .
\label{S5eq6}
\end{equation}
Eliminating $g_2 (t)$ from Equations (\ref{S5eq4}) and (\ref{S5eq5}), and
defining $u (t) = \dot g_1 (t) / g_1 (t)$, one obtains:
\begin{equation}
\dot u (t) + {1 \over 3} u^2 (t) = 0\ ,
\label{S5eq7}
\end{equation}
which is a special form of the Riccati equation. Following [22], the
solution to Equation (\ref{S5eq7}) is:
\begin{equation}
u (t) = 3 {{\dot v (t)} \over {v (t)}}\ ,
\label{S5eq8}
\end{equation}
with $v(t)$ representing a solution to $\ddot v (t) = 0$, which is the
auxiliary condition [21,22].
The initial conditions $v (t=0) = v_o$ and $\dot v (t=0) = a_o$ are
different from those used for $\hat D_o x (t) = 0$. Then, the solution
becomes $v (t) = a_o t + v_o$, and the functions $g_1 (t)$, $g_2 (t)$
and $g_3 (t)$ become
\begin{equation}
g_1 (t) = C_1 ( a_o t + v_o)^3\ ,
\label{S5eq9}
\end{equation}
where $C_1$ is an integration constant. Having obtained $g_1 (t)$,
$g_2 (t)$ becomes
\begin{equation}
g_2 (t) = - C_1 a_o ( a_o t + v_o)^2\ .
\label{S5eq10}
\end{equation}
Finally, $g_3 (t)$ can be found by eliminating $g_1 (t)$ and $g_2 (t)$
from Equation (\ref{S5eq5}). The solution is
\begin{equation}
g_3 (t) = C_1 C_2 ( a_o t + v_o)^2\ ,
\label{S5eq11}
\end{equation}
where $C_2$ is an integration constant.
Substituting $g_1 (t)$, $g_2 (t)$, and $g_3 (t)$ into Equation (\ref{S3eq1}),
for $\hat D_o x (t) = 0$, the following final form of the NSL is obtained
\begin{equation}
L_{ns} (\dot x , x , t) = {{1} \over {C_1 ( a_o t + v_o)^2}}\
{{1} \over {( a_o t + v_o) \dot x - a_o x + C_2}}\ ,
\label{S5eq12}
\end{equation}
which is the same as that given by Equation (\ref{S5eq3}). This concludes
the proof.
The derived NSL depends on two constants, $a_o$ and $v_o$,
which are given by the initial conditions for $\ddot v (t) = 0$,
and two arbitrary constants, $C_1$ and $C_2$, which may be
determined by the initial conditions for $\hat D_o x = 0$.
It is easy to verify that $L_{ns} (\dot x , x , t) $ gives
$\hat D_o x (t) = 0$ when substituted into the E-L equation.
This is the first example of the nonstandard Lagrangian for the
Newtonian law of inertia.
The {nonstandard general} null Lagrangian
$L_{nsgn} (\dot x ,
x , t) $, and its gauge function,
$\Phi_{nsgn} (x , t) $, are given by Equations
(\ref{S3eq7}) and (\ref{S3eq6}), respectively. To make this
NL and its GF exact, the following conditions must
be obeyed (see Equations \ref{S4eq4} and \ref{S4eq5})):
\begin{equation}
\left [ {{h_1 (1)} \over {h_2 (1)}} \right ] \ln \vert 2 h_2 (1) +
h_4 (1) \vert = 0\ ,
\label{S5eq13}
\end{equation}
and
\begin{equation}
\left [ {{h_1 (0)} \over {h_2 (0)}} \right ] \ln [ h_2 (0) + h_4 (0)] = 0\ .
\label{S5eq14}
\end{equation}
Since $\ln [ h_2 (1) + h_4 (1)] \neq 0$ and $\ln [ h_2 (0) + h_4 (0)] \neq 0$,
the end values of the function $h_1 (t) $ must be: $h_1 (1) = 0$ and
$h_1 (0) = 0$, but the end values of either $h_2 (t)$ or $h_4 (t)$ are
not limited by the conditions for exactness. Further constraints on the
functions $h_1 (t)$, $h_2 (t)$, and $h_4 (t)$ may be imposed by
considering symmetries and Lie groups [26] of the derived nonstandard
general ENL (see Section 4). With these constraints, the first
nonstandard EGF for the law of inertia is obtained.
The derived nonstandard null Lagrangians are of different forms when
compared to the standard null Lagrangians obtained in Section 3;
therefore, it is suggested that the NLs be divided into two separate sets. In
previous work [35,36], it was shown that standard null Lagrangians
and their gauge functions can be used to restore Galilean invariance of Lagrangians
in classical mechanics, and to introduce classical forces. The main physical
implication of the results obtained in this paper is that similar restoration
of invariance of Lagrangians and definition of forces can also be carried out by
using the derived general nonstandard NLs, and that the resulting forces will be
of different forms from those previously determined [36]. The fact
that not all general standard NLs contribute to the forces was shown by [37];
only NLs of special forms can be used to define forces [36,37]. In general,
most NLs have no influence on these forces. It remains to be determined
whether the derived general nonstandard null Lagrangian and its gauge function
define forces, and whether they can be used to convert the first law of dynamics
into the second law; however, such studies are beyond the scope of this paper.
The presented methods of finding general standard and nonstandard
ENLs and their EGFs can be extended to all second-order ODEs of the
form $\hat D x (t) = 0$, which includes the equations of motion of
undamped and damped oscillators, and other dynamical systems. In
previous work [43], the general standard ENLs and EGFs were
derived for the Bateman oscillators; however, the nonstandard ENLs
and EGFs are yet to be obtained. The presented methods may also
be generalized to partial differential equations of quantum mechanics,
such as the Schr\"odinger equation.
\section{Conclusions} \label{section6}
This paper presents methods to construct null Lagrangians. Using
these methods, two different sets of null Lagrangians were obtained
and classified as standard and nonstandard. The corresponding sets
of gauge functions were also derived. The presented general standard
null Lagrangians are known, but the general nonstandard null Lagrangians
obtained in this paper are new. Since there are differences in the forms
and properties of the two sets of null Lagrangians, it is suggested that
these Lagrangians be divided into two classes that correspond to
these sets.
The invariance of the action is used to introduce the exactness of both
general standard and nonstandard gauge functions. Having obtained the
exact gauge functions, they are used to derive the exact null Lagrangians.
All null Lagrangians and gauge functions, derived in this paper are exact,
which gives constraints on the end values of the arbitrary functions these
Lagrangians and gauge functions depend on. Further constraints can be
imposed by symmetries of the exact null Lagrangians and the corresponding
exact gauge functions, as well as their underlying Lie groups [26].
The obtained results are applied to the ordinary differential equation (ODE)
that represents the law of inertia for which the general exact nonstandard
null Lagrangian and the corresponding general exact gauge function are
derived. It is suggested that the derived Lagrangian and gauge function be
used to restore Galilean invariance of the standard Lagrangian for this law,
and to introduce forces as carried out in the previous work for the general
standard null Lagrangians [35,36,37]. Since there are significant
differences between the general standard and nonstandard null Lagrangians,
the resulting forces must also be different, which may allow establishing
a general procedure of defining forces in classical mechanics independently
from Newton's law of dynamics.
Finally, it must be pointed out that the same method can be used to obtain
the general exact nonstandard null Lagrangians and their gauge functions
for any ODE given by $\hat D x (t) = 0$, and that it can be extended to
homogeneous and inhomogeneous partial differential equations, and
applied to physical problems described by these equations.
\section{References}
|
\section{Introduction}
\label{sec:intro}
Diabetic retinopathy is an ocular disease that affects patients with undiagnosed, untreated, or undertreated diabetes mellitus. DR causes damage to the vessels inducing leakage of fluid within the retina, exudates and intraretinal hemorrhages. If the disease progresses it may lead to decreased vision and even blindness~\cite{Sengar2017}. Almost 40\% of diabetic patients have DR, from which about 5\% of patients face vision-threatening complications \cite{Antal2014}.
An ophthalmologist performs the DR diagnosis through a meticulously visual inspection of the retina according to the ICDR grading system \cite{Wilkinson2003}, which determines the DR level according to the presence of retinal lesions such as microaneurysms (MA), hemorrhages (H), exudates (EX), cotton wool spots (CWS), intraretinal microvascular abnormalities (IRMA), venous beading (VB), and neovascularization (NV) \cite{Chudzik2018, Sengar2017}. The initial clinical signs of DR are microaneurysms which allow leakage of blood from the affected capillaries, they appear as small reddish dots on the superficial retinal layers. Microaneurysms tend to have weak walls that can break, leading to hemorrhages of variable size and shape \cite{Paing2016}. Damaged retinal vessels allow lipoproteins precipitates to leak, which can be seen as yellowish exudates of irregular shape. These exudates exhibit a characteristic brightness in comparison to the contrast presented in microaneurysms and hemorrhages. CWS are lesions caused by retinal nerve fiber layer ischemia and emerge as yellowish-white spots with irregular edges. Signs of more advanced DR include irregular constriction and dilation of venous vessels in the retina, known as venous beading \cite{Chen2018a}, as well as neovascularization, which appears as red fronds of abnormal vascular networks. These vessels may show up arranged in a radiating pattern or without a distinct pattern \cite{talks2015}.
As much as 95\% of the cases of vision loss and blindness can be prevented with regular screening and appropriate management \cite{Bhaskaranand2019}. Despite the importance of a timely diagnosis fewer than a half of the patients with DR are aware of their condition \cite{Sengar2017}. Due to the asymptomatic nature of the disease in its early stages \cite{Mookiah2013}, patients should be screened with a fundus exam, to look for early signs of the disease and make a timely detection of the illness \cite{Beagley2014, Paing2016}. However, access to specialized care by an ophthalmologist is limited for some populations, given that most of the ophthalmologists in the world are concentrated mainly in urban areas and big cities. The development of an automated detection system for DR could improve access to specialized care by reducing the time, cost, and effort of screening \cite{Resnikoff2019, Orlando2018, Toledo2020}. In addition, the diabetic population is expected to increase 54\% by 2030, while the projected increase of ophthalmologists is only 2\%. Thus, the need to integrate methods to automate the screening process responds to these challenges in the present and future diabetic retinopathy panorama \cite{VanderHeijden2018}.
Dealing with the previous issues, we propose a model for the automatic detection of DR that initially relies on the identification of its ocular related lesions to later diagnose this illness. The proposed method is different from state-of-the-art models based on Convolutional Neural Networks (CNN), because it follows the clinician workflow, this approach has the advantage of providing additional information regarding the lesions found in the input image that give an improved interpretability. The experimental results also showed that this strategy also improves the overall classification performance of the system.
The main three contributions of this paper are as follows: a fine-grained lesions annotations for microaneurysms, hemorrhages, cotton wool spots, venous beading, neovascularization and exudates; global labels for ocular diseases and referable conditions for $3209$ images from the kaggle EyePACS dataset and the complete Messidor-2 dataset; and a strategy for the detection of DR that relies on the identification of its related retinal abnormalities.
The paper is organized as follows: Section~\ref{sec:previous_work} presents a review of the state-of-the-art including the most representative approaches that classify DR based on retinal lesions identification. Section~\ref{sec:databases} presents a description of the features of our new fine-grained labels for lesions of the retina related to DR and a comparison with the available databases. The proposed method is described in detail in Section~\ref{sec:method}, the results are summarized and discussed in Section~\ref{sec:results}, and conclusions and future work are presented in Section~\ref{sec:conclusion}.
\section{Previous work}
\label{sec:previous_work}
In this section, we present an overview of the most representative methods for the identification of DR based on the detection of individual and multiple ocular lesions.
The most common approach contains two stages: a first step extracts manually a set of features from DR-related lesions, and then a model uses them to classify or grade the DR.
Sharif et al. presented a model that combines independent component analysis with a curve fitting technique to remove retinal blood vessels and the optic disc from eye fundus images. A feature set of DR lesions is build, which includes the count number of EX, H, MA, mean, and standard deviations of candidate regions on the images. Then, a multi-class Gaussian Bayes classifier and a multi-SVM are trained with the feature set to grade DR~\cite{Sharif2019}. Similarly, Abdelmaksoud et al. reported a known segmentation method named as U-Net to discriminate among EX, MA, H, and blood vessels. Then a set of features as co-occurrence matrix, areas, and bifurcation points count are calculated to grade DR using a SVM~\cite{Abdelmaksoud2021}. Paing et al. proposed a method that uses histogram matching, morphological opening, and canny method to segment blood vessels, EX and MA. Some features as area and counts are estimated to classify the stages of DR using a customized artificial neural network~\cite{Paing2016}. Akram et al. presented a method that initially removes blood vessels and optic disc. Then, a set of morphological descriptors such as shape, color, and statistical features are used to train a weighted combination of multivariate m-Mediods and a Gaussian Mixture Model to identify potential candidates for MA, H, and EX. Finally, the fundus image is graded on a DR scale according to medical conditions that include the type and count of lesions~\cite{UsmanAkram2014}.
Alternatively, some researchers have proposed models to automatically perform feature extraction from DR-related lesions and use them to classify the illness. Yang et al. presented a two-stages CNN-based algorithm trained with image patches and a weighted lesion map on the input, which can detect DR severity and its related lesions~\cite{Ghafoorian2017b}. Wang et al. designed a hierarchical multi-task deep learning framework for the classification of DR severity and DR-related lesions such as MA, H, CWS, VB, NV, and others~\cite{Wang2020a}. Similarly, Zago et al. explored a lesion localization model using two CNN patch-based approaches. Additional information such as a lesion probability map and the maximum value of the probability map are used to classify DR~\cite{Zago2020}. Recently, Zhou et al. partially released fine-grained annotations of pixel-level and image-level lesions related to DR from the FGADR dataset. They reported individual modules for the segmentation of lesions to extract features and to automatically grade DR~\cite{Zhou2021}.
Our model does not require segmentation of retinal structures as a preprocessing step or any other special preprocessing of the images, unlike most of the methods in literature for the detection of DR based on lesions. Besides, we did not reannoted the DR labels, our DR detection model is evaluated with the already publicly available so it can be compared with state-of-the-art methods and all the additional labels of DR lesions used to train the model are going to be publicly available.
\section{Dataset description}
\label{sec:databases}
The kaggle EyePACS is a free real-world set of high-resolution photos of the posterior pole of the retina. The kaggle EyePACS dataset comprises $88702$ macula-centered eye fundus images acquired with different types of cameras and under a variety of imaging conditions. These images were labeled using a scale of 0, 1, 2, 3, 4, which stand for no DR, mild, moderate, severe and proliferative DR respectively~\cite{kaggle}.
The Messidor-2 dataset is a public dataset that contains 1748 macula-centered eye fundus images that were acquired with a 45-degree field of view and the sizes are ranged between $1440\times960$ and $2304\times1536$ pixels. Abramoff et al. released the images with a binary classification for presence or absence of DR~\cite{messidor_labels}.
These public datasets were used to create two new individual datasets with fine-grained labels of six ocular lesions used for ophthalmologists to diagnose DR. The construction details are explained as follows:
\emph{Dataset construction:} an ophthalmologist with supra-specialty in the retina from the Fundacion Oftalmologica Nacional of Colombia selected $3209$ retinal images from the kaggle-EyePACS dataset. The ophthalmologist performed the manual choice of these images based on the quality of fundus images and the presence of DR-related lesions. The Messidor-2 dataset was also analysed by the specialist who determined that $1689$ images are suitable for lesion level annotation. The expert labeled the images with six ocular findings: aneurysms, hemorrhages, cotton wool spots, venous beating, neovascularization, and exudates. In addition, these images were manually annotated with: binary labels for referable and non-referable patients and the grade of diabetic macular edema (DME) in a scale 0 to 3 for no DME, mild, moderate, and severe.
\emph{Annotation criteria:} the whole images were evaluated in their entire area as follows: first, a meticulous analysis looking for any lesions on the optic nerve and the macular region is performed. Then, the extra-macular retina and outside the vascular arches areas are examined in detail to find possible DR-related lesions. Finally, each image was classified in a fine-grained way as positive for that specific finding if at least one ocular lesion was identified. Regardless of whether it was a single injury or many, or whether it affected the center of the macula or its periphery. In addition, the images were labeled as DME in an image-level way according to the ETDRS scale. Also, images were classified as a referable image, if at least an ocular lesion was present in all image areas, regardless of which one.
\begin{table}[!hb]
\caption{Distribution of DR-related ocular lesions and image-level labels from the customized Kaggle EyePACS (training test) and Messidor-2 (test set) data sets.}
\begin{tabular}{c|c|c|c|c|c|c||c|c|c|}
\cline{2-10}
& \multicolumn{6}{c||}{Ocular lesions labels} & \multicolumn{3}{c|}{Image-level labels}\\
\cline{1-10}
\multicolumn{1}{|c|}{Set of images} & \multicolumn{1}{c|}{MA} & \multicolumn{1}{c|}{H} & \multicolumn{1}{c|}{CWS} & \multicolumn{1}{c|}{VB} & \multicolumn{1}{c|}{NV} & \multicolumn{1}{c||}{EX} & \multicolumn{1}{c|}{DR} & \multicolumn{1}{c|}{REF} & \multicolumn{1}{c|}{DME}\\ \hline
\multicolumn{1}{|l|}{\multirow{2}{*}{Training set}} & 1719 & 1418 & 653 & 720 & 140 & 1812 & 2016 & 2550 & 1812\\
\multicolumn{1}{|l|}{} & 53.55\% & 44.17\% & 20.34\% & 22.42\% & 4.36\% & 56.44\% & 62.80\% & 79.44\% & 56.44\%\\
\hline
\multicolumn{1}{|l|}{\multirow{2}{*}{Test set}} & 876 & 443 & 133 & 47 & 15 & 196 & 377 & 893 & 196\\
\multicolumn{1}{|l|}{} & 51.86\% & 26.22\% & 7.87\% & 2.78\% & 0.89\% & 11.60\% & 22.32\% & 52.87\% & 11.60\%\\
\hline
\end{tabular}
\label{tab:dataset_distribution}
\end{table}
A summary with the number of images and percentage per finding and image-level for the customized kaggle EyePACS dataset which is used as training set and for the Messidor-2 dataset, which is used as test set is presented in Table \ref{tab:dataset_distribution}.
The DR-related findings of MA, H, and CWS are the most common lesions as shown in Table~\ref{tab:dataset_distribution}. Moreover, these ocular findings are related to appear in the initial stages of subjects with DR. On the other hand, VB and NV findings presented few examples during the screening, but also these findings are scarce and rare to find because they are common in an advanced grade of DR. Finally, the Referable condition was the most common image-level labels among the datasets.
The comparison of our dataset with others publicly available with fined grained labels of DR related lesion is presented in Table \ref{tab:data_bases}. It can be noticed that the available datasets have fewer examples of DR related lesions, this is due to labeling images at a lesion level is costly, tedious and time consuming \cite{Orlando2018}. We hope that this new set of labels open opportunities to improve and develop new approaches for the detection of retina lesions caused by DR and for the detection of DR based on them, which leads to more relatable models for the clinicians.
\begin{table}[!ht]
\centering
\caption{Comparison of DR-related lesion labels from public databases. The DR-related lesion are: microaneurysm (MA), hemorrhages (H), cotton wool spots (CWS), venous beading (VB), neovascularization (NV), exudates (EX), and other lesions.}
\begin{tabular}{|p{4.7cm}|>{\centering\arraybackslash}p{1cm}|>{\centering\arraybackslash}p{1cm}|>{\centering\arraybackslash}p{1cm}|>{\centering\arraybackslash}p{1cm}|>{\centering\arraybackslash}p{1cm}|>{\centering\arraybackslash}p{1cm}|>{\centering\arraybackslash}p{1cm}|>{\centering\arraybackslash}p{1cm}|>{\centering\arraybackslash}p{1cm}|}
\hline
\diagbox{Database}{Ocular lesions} & MA & H & CWS & VB & NV & EX & Others\\
\hline
DRIVE & & X & & & & X & X \\
DIARETDB0 & X & X & X & & X & X & \\
DIARETDB1 & X & X & X & & & X & \\
STARE & X & & & & X & & X \\
IDRiD & X & X & X & & & X & \\
FGADR & X & X & X & & X & X & X \\
e-ophtha & X & & & & & X & \\
Our dataset & X & X & X & X & X & X & \\
\hline
\end{tabular}
\label{tab:data_bases}
\end{table}
\section{Method}\label{sec:method}
The pipeline of our proposed method is depicted in Figure~\ref{fig:model}. The method has two main stages, lesion detection and DR classification.
\begin{figure}[!hb]
\centering
\includegraphics[scale=0.42]{new_method.pdf}
\caption{An overview of the proposed method for ocular lesions detection and DR classification. The model is organized in three consecutive stages as follows: lesion detection (first block), lesion prediction (second block) and DR detection (third block). }
\label{fig:model}
\end{figure}
Initially, the eye fundus images are preprocessed to resize them to $512 \times 512$ pixels and remove the black edges. The lesion detection stage is based on a transfer learning strategy using the weights of pre-trained models for DR detection reported by He et al.~\cite{He2021}. The base models correspond to the embedding of two attention blocks into different backbones networks, namely, DenseNet121, Xception, ResNet50, and MobileNet for DR grading. Multiple classifiers are trained using the features extracted from the base networks and the fine-grained labels of the kaggle EyePACS dataset for the detection of DR-related lesion. The best overall performing method is selected as the classifier for the detection of lesions.
In the second stage the predictions of the lesions are concatenated to train multiple classifiers for the DR classification task. The best results are reported and compared with state-of-the-art methods using the Messidor-2 dataset as test set. The models and datasets used in this work will be released in a public repository to ensure reproducibility and encourage other researchers to create new methods.
The classifiers and the parameters explored in both stages are:
\begin{itemize}
\item Support vector machine (SVM): the parameters explored are the regularization parameter (C), different kernels and the gamma $(\gamma)$ kernel coefficient.
\item Gaussian Process (GP): the kernels `RBF' and `Mattern' are evaluated. The kernel parameters explored are lower length scale bound, upper length scale bound, lower noise level and upper noise level.
\item Multilayer Perceptron (MLP): the number of layers, number of neurons, activations and learning rate are explored.
\end{itemize}
\section{Results and Discussion}\label{sec:results}
\subsection{Lesions automatic detection}
A systematic exploration and evaluation of the detection of DR-related lesions using the different classifiers and the features obtained from the pre-trained models was performed. After an initial exploration, the best overall results are obtained using as backbone the {densenet121\_CAB\_EyePACS} model as feature extractor. The results of the detection of MA, H, CWS, VB, and NV using this model are reported in Table~\ref{tab:findings_results}, where the best performing classifiers are the GP regressor and the MLP. The MA and VB findings presented greater difficulty in comparison to H and NV findings to be detected. This may be because, hemorrhages are bigger and neovascularizations are usually accompanied by other lesions what can help the model to detect them since it appears in the latter stages of the disease.
\begin{table}[!ht]
\centering
\caption{Results of the detection of DR-related ocular lesions using the densenet121\_CAB\_EyePACS backbone model as a feature extractor.}
\begin{tabular}{|c|p{1.2cm}cc>{\centering\arraybackslash}p{1.5cm}>{\centering\arraybackslash}p{1.5cm}|}
\hline
Finding & Metric & GP classifier & GP regressor & SVM & MLP \\ \hline
\multirow{3}{*}{MA} & AUC & 0.7377 & 0.7865 & 0.7516 & 0.8055 \\
& Sp & 0.8794 & 0.7023 & 0.8733 & 0.7072 \\
& Se & 0.5959 & 0.7141 & 0.5833 & 0.7237 \\\hline
\multirow{3}{*}{H} & AUC & 0.8457 & 0.9332 & 0.8503 & 0.9317 \\
& Sp & 0.9622 & 0.8472 & 0.9775 & 0.8370 \\
& Se & 0.7291 & 0.8600 & 0.7088 & 0.8826 \\ \hline
\multirow{3}{*}{CWS} & AUC & 0.8791 & 0.8825 & 0.6884 & 0.9021 \\
& Sp & 0.7596 & 0.7622 & 0.9485 & 0.8303 \\
& Se & 0.8496 & 0.8571 & 0.4135 & 0.8496 \\ \hline
\multirow{3}{*}{VB} & AUC & 0.7525 & 0.7299 & 0.5348 & 0.7983 \\
& Sp & 0.6461 & 0.6565 & 0.9098 & 0.8026 \\
& Se & 0.6808 & 0.6808 & 0.1702 & 0.5957 \\
\hline
\multirow{3}{*}{NV} & AUC & 0.9401 & 0.9693 & 0.6205 & 0.9685 \\
& Sp & 0.8393 & 0.9074 & 0.9164 & 0.9127 \\
& Se & 0.8666 & 0.9333 & 0.3333 & 0.9333 \\ \hline
\end{tabular}
\label{tab:findings_results}
\end{table}
\subsection{DR detection using lesions prediction}
The proposed method trained with a reduced set of training samples shows competitive performance in AUC and consistent results in sensitivity and specificity evaluated in Messidor-2 when compared to the baseline models trained with larger datasets for the DR detection task as shown in Table~\ref{tab:dr_results_comparison}. Despite some state-of-the-art methods have better performances, they only classify DR, unlike our approach that also provides information about the DR-related lesions, which provides interpretability and makes the DR detection process more familiar for clinicians, since it follows a workflow similar to their own.
\begin{table}[!htbp]
\centering
\caption{Comparison of proposed method with state-of-the-art DR detection methods tested on Messidor-2 dataset. The top 3 results for each metric are marked in bold$^{(1)}$, italic$^{(2)}$, and underline$^{(3)}$ respectively.}
\begin{tabular}{|>{\centering\arraybackslash}p{2.5cm}|p{4.5cm}|>{\centering\arraybackslash}p{1.55cm}|>{\centering\arraybackslash}p{1.55cm}|>{\centering\arraybackslash}p{1.65cm}|}
\hline
\shortstack{Training set \\ (set of images) } & Method & AUC & Sensitivity & Specificity\\
\hline
EyePACS train (35126) & densenet121\_CAB\_DDR~\cite{He2021} & 0.765 & 0.359 & \textbf{0.990$^{(1)}$}\\
EyePACS train (35126)& densenet121\_CAB\_EyePACS~\cite{He2021} & 0.782 & 0.395 & \textit{0.954$^{(2)}$}\\
EyePACS train (35126)& resnet50\_CAB\_EyePACS~\cite{He2021} & 0.884 & 0.753 & 0.801\\
EyePACS train (35126)& xception\_CAB\_DDR~\cite{He2021} & 0.908 & 0.877 & 0.861\\
EyePACS train (35126)& xception\_CAB\_EyePACS~\cite{He2021} & 0.908 & \underline{0.880}$^{(3)}$ & 0.855\\
EyePACS custom (57146) & Voets et al.~\cite{Voets2018} & 0.800 & 0.737 & 0.697\\
EyePACS custom (28102) & Zhou et al.~\cite{Zhou2018} & \textbf{0.960$^{(1)}$} & - & - \\
EyePACS custom (75137) & Gargeya and Leng \cite{Gargeya2017} & \underline{0.940}$^{(3)}$ & \textbf{0.930$^{(1)}$} & 0.870\\
EyePACS custom (3209) & our model with MLP & \textit{0.948$^{(2)}$} & \textit{0.886$^{(2)}$} & \underline{0.875}$^{(3)}$\\
\hline
\end{tabular}
\label{tab:dr_results_comparison}
\end{table}
\section{Conclusion} \label{sec:conclusion}
We presented a strategy to improve the performance of a pre-trained model for the detection of DR by using it as feature extractor to classify DR related lesions, and use them to identify the illness. The best results are obtained using a MLP as classifier, which leads to an end-to-end model that is easier to handle and can be further improved.
The most remarkable aspect of our proposed method is the combination of domain knowledge from ophthalmologists and the deep learning versatility to support medical decision making. Moreover, the development of models that includes DR-related lesions and the clinical workflow to perform the ocular disease diagnosis may be a good solution to gain greater acceptance and possible use in real-world applications by clinical staff.
Finally, the validation of our method with a large number of images and other ocular diseases will be studied in a future work.
|
\section{Introduction}
For any finite group $G$, the category of rational $G$-spectra admits a splitting, which is first discussed by Greenlees and May:
\begin{theorem} \label{idemsplit}
\textbf{[GM95]} There is an orthogonal basis $\{e_H:H\subset G\}$ of the rational Burnside ring, which only contains idempotent elements. For any rational $G$-spectrum $X$, define $e_HX$ as
$$e_HX:=colim(X\xrightarrow{e_H}X\xrightarrow{e_H}X\xrightarrow{e_H}...)$$
Then we have
$$X\simeq\bigvee_H e_HX$$
and
$$[X,Y]^G\cong\prod_H[e_HX,e_HY]^G$$
with one $H$ chosen from each conjugacy class of subgroups.
\end{theorem}
This theorem is further studied by Barnes in \textbf{[Bar08]}, which reproves the splitting above by some topological constructions and sets up an algebraic model which completely describes the behavior of rational $G$-spectra.
One important idea in Barnes' proof is to use the universal $G$-space $E\lambda_H$, which is characterized by the fixed point subspaces:
$$(E\lambda_H)^K\simeq
\begin{cases}
S^0,\text{ if } K \text{ is conjugate to } H,\\
*, \text{ otherwise.}
\end{cases}$$
\begin{theorem} \label{computesplit}
\textbf{[Bar08]} For rational $G$-spectra $X,Y$, we have an isomorphism
$$[X,Y]^G\cong\prod_H [E\lambda_H\wedge X,E\lambda_H\wedge Y]^G$$
with one $H$ chosen from each conjugacy class. Moreover, the functor
$$X\mapsto\bigvee_H E\lambda_H\wedge X$$
is symmetric monoidal.
\end{theorem}
\medskip
However, the rational assumption is too strong. In this paper, we will reprove this theorem by further studying the construction and properties of universal spaces. We will show:
\begin{theorem} \label{orderinvert}
The splitting in \autoref{computesplit} happens when we only invert all prime factors of $|G|$, instead of applying a full rationalization.
\end{theorem}
\bigskip
Moreover, if we only invert some, but not all prime factors of $|G|$, we may still get a partial splitting
$$[X,Y]^G\cong[E\mathscr{F}_+\wedge X,E\mathscr{F}_+\wedge Y]^G\oplus [\widetilde{E\mathscr{F}}\wedge X,\widetilde{E\mathscr{F}}\wedge Y]^G$$
where $E\mathscr{F}$ is the universal space corresponding to the family $\mathscr{F}$.
The corresponding functor
$$X\mapsto (E\mathscr{F}_+\wedge X)\vee(\widetilde{E\mathscr{F}}\wedge X)$$
still turns out to be symmetric monoidal.
\medskip
The summands $E\mathscr{F}_+\wedge X$ and $\widetilde{E\mathscr{F}}\wedge X$ are usually more simple than $X$ since the product with either $E\mathscr{F}_+$ or $\widetilde{E\mathscr{F}}$ reduces the types of cells in $X$. When we want to compute the $RO(G)$-graded homotopy $\pi_\bigstar(X)$ of $X$, we can break it into more computable parts:
$$\pi_\bigstar(X)\cong\pi_\bigstar(E\mathscr{F}_+\wedge X)\oplus\pi_\bigstar(\widetilde{E\mathscr{F}}\wedge X)$$
after inverting proper prime factors of $|G|$. Inverting different prime factors in $|G|$ gives us different partial splittings. We can glue these data back to recover the unlocalized $\pi_\bigstar(X)$.
\medskip
One explicit example given in this paper is the computation of $\pi_\bigstar(H\underline{\mathbb{Z}})$ when $G=D_{2p}$ is a dihedral group. The additive structure is already computed in \textbf{[KL20]} and \textbf{[Zou18]}, while the latter also provides some partial multiplicative structure. With our new method, the complete multiplicative structure follows naturally since all functors in the splittings are symmetric monoidal.
The main idea is, we get two different partial splittings when $2$ and $p$ are inverted respectively. Both splittings break $\pi_\bigstar(H\underline{\mathbb{Z}})$ into two more computable parts, whose computations can be decomposed into some $RO(C_2)$ or $RO(C_p)$-graded cohomology theories. We can compute $\pi_\bigstar(H\underline{\mathbb{Z}})[1/2]$ and $\pi_\bigstar(H\underline{\mathbb{Z}})[1/p]$ from these two splittings, and hence recover $\pi_\bigstar(H\underline{\mathbb{Z}})$.
\bigskip
\paragraph{Structure of the paper:}
In section 2, we will introduce the universal space $E\mathscr{F}$. We are especially interested in the equivariant homology of $E\mathscr{F}$ with coefficients in the Burnside ring Mackey functor $A_G$. The most important homological information is given in \autoref{degree0} and \autoref{torsion}.
Section 3 contains a reproof of \autoref{computesplit} and a proof for \autoref{orderinvert}. We will show that the splitting in \autoref{computesplit} can be obtained by repeatedly applying the partial splitting
$$X\simeq (E\mathscr{F}_+\wedge X)\vee(\widetilde{E\mathscr{F}}\wedge X)$$
for all choices of $\mathscr{F}$. Since each partial splitting only requires some prime factors of $|G|$ to be inverted, \autoref{computesplit} works when we invert $|G|$. We will also give a criterion about which prime factors should be inverted in order to make the partial splitting happen. This criterion will be quite useful for our later computations.
In sections 4 and 5, we consider the dihedral group $G=D_{2p}$ and compute the $RO(G)$-graded homotopy of the Eilenberg-Maclane spectra $H\underline{\mathbb{Z}}$. We will discuss how this method can be applied to more general $D_{2p}$-spectra in section 6. In particular, we compute the $RO(G)$-graded homotopy of $HA_G$.
Finally, in section 7 and 8, we will explain how this idea can be applied for more general finite groups. We will focus on the case that $G=G_1\ltimes G_2$, and prove two partial splittings when either $|G_1|$ or $|G_2|$ is inverted.
\bigskip
Here is a list of main results of this paper:
The homotopy of $H\underline{\mathbb{Z}}$ as a $RO(D_{2p})$-graded ring: \autoref{hz}.
The homotopy computation of more general $D_{2p}$-spectra: \autoref{generald2p} and \autoref{hagrog}.
The splitting theorem for more general finite groups: \autoref{bigtwosplit}.
\bigskip
\paragraph{Notations:} In this paper, we use $*$ when the homotopy or homology is graded over $\mathbb{Z}$, and use $\bigstar$ when graded over $RO(G)$. We add an underline to express the Mackey functor valued homotopy or homology: $\underline{H}$, $\underline{\pi}$.
To be more precise, let $X$ be a $G$-space and $M$ be a Mackey functor. Define the Mackey functor valued equivariant homology of $X$ with coefficients in $M$ as
$$\underline{H}_*^G(X;M):=\underline{\pi}_*^G(X\wedge HM).$$
So we have
$$\underline{H}_*^G(X;M)(G/H)=[\Sigma^*G/H_+,HM\wedge X]^G,$$
$$H_*^G(X;M)=\underline{H}_*^G(X;M)(G/G).$$
Here $HM$ is the equivariant Eilenberg-Maclane spectrum corresponding to $M$.
\medskip
We discuss the properties of the universal spaces by computing their $\mathbb{Z}$-graded, Mackey functor valued homology $\underline{H}_*$ in section 2 and 3. But we compute the $RO(G)$-graded, abelian group (to be more precise, graded ring) valued homotopy $\pi_\bigstar$ in the remaining sections. Theoretically, our computational methods also work for $\underline{\pi}_\bigstar$, although the multiplicative structures expressed as box products may become quite complicated.
\section{Universal spaces}
We will introduce the universal space $E\mathscr{F}$ and study its equivariant homology in this section. The most important properties are given in \autoref{homsplit} and \autoref{torsion}.
\begin{definition} \label{clspace}
A \textbf{family} $\mathscr{F}$ is a collection of subgroups of $G$ which is closed under conjugation and taking subgroups. The corresponding \textbf{universal space} $E\mathscr{F}$ is an unbased $G$-space such that $(E\mathscr{F})^H$ is contractible for all $H\in\mathscr{F}$ and empty otherwise.
\end{definition}
\begin{remark}
When $\mathscr{F}$ only contains the trivial subgroup, $E\mathscr{F}$ becomes $EG$, which is a contractible space with free $G$-action.
\end{remark}
\smallskip
One construction of $E\mathscr{F}$ is given in \textbf{[Dieck72]} as a join of spaces. We will explain this idea later and use it to prove some important properties. An alternative construction is given by \textbf{[Elm83]} as a categorical bar construction.
\smallskip
We can fully characterize $E\mathscr{F}$ by fixed point subspaces:
\begin{lemma} \label{unique}
The universal space $E\mathscr{F}$ is unique up to weak equivalence.
\end{lemma}
\paragraph{Proof:} Assume that $X,Y$ are $G$-spaces with the same conditions on fixed point subspaces as $E\mathscr{F}$.
If we have a $G$-map $X\rightarrow Y$, its restriction on each fixed point subspace is a weak equivalence since $X,Y$ have the same types of fixed point subspaces as either $*$ or $\emptyset$. Thus this map is a $G$-weak equivalence.
If we do not have a map between $X,Y$, consider $X\times Y$, which is another $G$-space with the same conditions on fixed point subspaces. The above argument shows that the projections $X\times Y\rightarrow X$ and $X\times Y\rightarrow Y$ are $G$-weak equivalences. Therefore, $E\mathscr{F}$ is unique up to weak equivalence. $\Box$
\bigskip
We assume $E\mathscr{F}$ to be a $G$-CW complex by applying the $G$-CW approximation.
\begin{lemma} \label{uniqueef}
Any $G$-self map of $E\mathscr{F}$ is $G$-homotopic to the identity.
\end{lemma}
This lemma is implied by the following theorem:
\begin{theorem} \label{elmendorf}
Let $\mathcal{O}_G$ be the category of $G$-orbits. Define an $\mathcal{O}_G$-space to be a contravariant functor from $\mathcal{O}_G$ to topological spaces.
The following pair of functors
$$\Phi: G-spaces\rightleftarrows\mathcal{O}_G-spaces:\Psi$$
defined by $\Phi(X)(G/H):=X^H$ and $\Psi(T):=T(G/\{e\})$ form a Quillen equivalence.
\end{theorem}
We refer to [\textbf{May96}, VI.6] for more details.
\paragraph{Proof of \autoref{uniqueef}:} Construct an $\mathcal{O}_G$-space $T$ by setting
$$T(G/H)=
\begin{cases}
*, \text{ if }H\in\mathscr{F}\\
\emptyset, \text{ otherwise}
\end{cases}$$
Let $CT$ be a cofibrant approximation of $T$. Then
$$\Psi(CT)^H=\Phi\Psi(CT)(G/H)\cong CT(G/H)\cong
\begin{cases}
*, \text{ if }H\in\mathscr{F}\\
\emptyset, \text{ otherwise}
\end{cases}$$
Recall that
$$\Phi E\mathscr{F}(G/H)=(E\mathscr{F})^H\cong
\begin{cases}
*, \text{ if }H\in\mathscr{F}\\
\emptyset, \text{ otherwise}
\end{cases}$$
The unique map $\Phi E\mathscr{F}\rightarrow T$ induces $\Phi E\mathscr{F}\rightarrow CT$, whose adjuction $E\mathscr{F}\rightarrow\Psi(CT)$ becomes a $G$-equivalence.
So we have
$$[E\mathscr{F},E\mathscr{F}]^G\cong[E\mathscr{F},\Psi(CT)]^G\cong[\Phi E\mathscr{F},CT]^{\mathcal{O}_G}\cong[\Phi E\mathscr{F},T]^{\mathcal{O}_G},$$
which contains a single element.
Thus the identity map is the only self-map of $E\mathscr{F}$ up to $G
$-homotopy. $\Box$
\bigskip
Notice that $E\mathscr{F}\times E\mathscr{F}$ also appears as a universal space for $\mathscr{F}$. Thus the above two lemmas imply:
\begin{lemma} \label{monoid}
The universal space $E\mathscr{F}$ is a $G$-topological semigroup up to homotopy.
\end{lemma}
Usually $E\mathscr{F}$ is not a Hopf space since there is no unit. But with extra conditions, $\Sigma^\infty E\mathscr{F}_+$ will appear as a homotopy ring spectrum. We will explain this idea in section 3.
\medskip
Now we discuss the homological properties for $E\mathscr{F}$. First we recall the definition of the Burnside ring and the corresponding Mackey functor:
\begin{definition} \label{ag}
For any finite group $G$, the collection of isomorphism classes of finite $G$-sets forms a commutative monoid, with addition induced by disjoint unions. The \textbf{Burnside ring} $A(G)$ is defined to be the group completion of this monoid. We can express $A(G)$ as a free $\mathbb{Z}$-module, whose basis corresponds to $G$-orbits. We use $\{G/H\}$ to denote the basis element of the orbit $G/H$.
\medskip
The \textbf{Burnside ring Mackey functor} $A_G$ is defined by $A_G(G/H):=A(H)$. The Mackey functor structure is given by the following maps:
\smallskip
For any $L\subset H$, the transfer map
$$T_L^H: A(L)\rightarrow A(H)$$
sends $\{L/K\}$ to $\{H/K\}$ for any $K\subset L$.
The restriction map
$$R_L^H: A(H)\rightarrow A(L)$$
sends $\{H/N\}$ to itself, but viewed as an $L$-space.
When $L=g^{-1}Hg$, we also have an isomorphism
$$C_L^H:A(L)\rightarrow A(H)$$
sending $\{L/K\}$ to $\{H/gKg^{-1}\}$ for any $K\subset L$.
\medskip
The Cartesian product of finite $G$-sets makes $A(G)$ into a ring. This multiplicative structure also makes $A_G$ into a Green functor, which is a monoid under the box product.
\end{definition}
\begin{proposition} \label{burnsidehom}
For a $G$-CW complex $X$, the integer graded homology of $X$ with coefficients in $A_G$ can be computed as
$$H_*^G(X;A_G)\cong\bigoplus_H H_*(X^H/W_GH;\mathbb{Z})$$
with one $H$ chosen from each conjugacy class of subgroups of $G$. Here $W_GH$ is the Weyl group of $H$ in $G$.
As a Mackey functor,
$$\underline{H}_*^G(X;A_G)(G/L)\cong H_*^L(X;A_L)\cong\bigoplus_K H_*(X^K/W_LK;\mathbb{Z})$$
with one $K$ chosen from each conjugacy class of subgroups of $L$.
\end{proposition}
\paragraph{Proof:} It suffices to prove the first equation. The second one can be proved by the same argument.
\medskip
When we view $A_G$ as a covariant coefficient system, the restriction maps are removed. For any $G$-map $G/L_1\rightarrow G/L_2$ induced by the multiplication of $g\in G$ (where we require $g^{-1}L_1g\subset L_2$), we have an induced map
$$A(L_1)\rightarrow A(L_2)$$
sending each $\{L_1/H\}$ to $\{L_2/g^{-1}Hg\}$. Therefore, as a coefficient system, $A_G$ can be decomposed as
$$A_G=\bigoplus_{H}A_G^H$$
with one $H\subset G$ chosen from each conjugacy class. Here $A_G^H$ is the sub-coefficient system of $A_G$ such that each $A_G^H(G/L)\subset A(L)$ is generated by all $\{L/g^{-1}Hg\}$ with $g^{-1}Hg\subset L$.
\smallskip
Now we have
$$H_*^G(X;A_G)\cong\bigoplus_{H}H_*^G(X;A_G^H).$$
It suffices to prove
$$H_*^G(X;A_G^H)\cong H_*(X^H/W_GH;\mathbb{Z})$$
for any $H\subset G$.
\bigskip
As described in \textbf{[Wil75]}, the equivariant homology can be computed in a cellular way:
Let $\underline{C}_*(X)$ be the chain complex of contravariant coefficient systems such that
$$\underline{C}_*(X)(G/H):=C_*(X^H),$$
where $C_*(X^H)$ is the cellular chain complex of $X^H$ and the orbit maps are sent to conjugacies and inclusions of subcomplexes. For any covariant coefficient system $M$, $H_*^G(X;M)$ is the homology of the chain complex:
$$C_*^G(X;M):=\underline{C}_*(X)\otimes_{\mathcal{O}_G}M,$$
where the tensor product is taken over the category $\mathcal{O}_G$ of $G$-orbits. To be more precise, we have
$$C_*^G(X;M):=\left(\bigoplus_{L\subset G}C_*(X^L)\otimes M(G/L)\right)/\sim.$$
For any $G$-map $f:G/L_1\rightarrow G/L_2$, the equivalence relation identifies $f^*a\otimes b$ and $a\otimes f_*b$ for all $a\in C_*(X^{L_2})$ and $b\in M(G/L_1)$.
\medskip
Now we choose $M=A_G^H$. Since each $M(G/L)$ is freely generated by $\{L/g^{-1}Hg\}$ and the structure maps send these basis elements to each other, $C_*^G(X;M)$ is the free $\mathbb{Z}$-module generated by the equivalence classes of $e\otimes \{L/g^{-1}Hg\}$, for all cells $e$ in $X^L$ and $g^{-1}Hg\subset L$.
\medskip
We can eliminate the equivalence relation by the following four facts:
\textbf{(1)} The free $\mathbb{Z}$-module $A_G^H(G/H)$ is generated by a single element $\{H/H\}$.
\textbf{(2)} For any $e\otimes\{L/g^{-1}Hg\}$, it is identified with $ge\otimes\{H/H\}$ by the equivalence relation.
\textbf{(3)} If there exists another $g^\prime\in G$ such that
$$ge\otimes\{H/H\}\sim e\otimes\{L/g^{-1}Hg\}\sim g^\prime e\otimes\{H/H\},$$
then $\{L/g^{-1}Hg\}=\{L/(g^{\prime})^{-1}Hg^\prime\}$. Thus
$$g^{-1}Hg=l^{-1}((g^\prime)^{-1}Hg^\prime)l$$
for some $l\in L$. So $g^\prime lg^{-1}\in W_GH$. Since $e$ is $L$-fixed, we have $(g^\prime lg^{-1})(ge)=g^\prime e$. Thus $ge$ and $g^\prime e$ are in the same $W_GH$-orbit.
\textbf{(4)} The converse of \textbf{(3)} is also true: $e_1\otimes\{H/H\}$ and $e_2\otimes\{H/H\}$ are identified if $e_1,e_2$ are in the same $W_GH$-orbit.
\medskip
Now we have a 1-1 correspondence between the basis of $C_*^G(X;A_G^H)$ and the cells in $X^H/W_GH$. Therefore, we get
$$C_*^G(X;A_G^H)\cong C_*(X^H/W_GH).$$
Taking the homology on both sides gives us the required equation. $\Box$
\bigskip
\begin{remark}
\autoref{burnsidehom} follows from the fact that the underlying coefficient system of $A_G$ splits into $A_G^H$. However, such splitting cannot be lifted to the Mackey functor level. Thus the decomposition does not work for the cohomology with coefficients in $A_G$.
In fact, since we cannot define orbit spectra on the complete $G$-universe,
$$X\mapsto H_*(X^H/W_GH)$$
is a homology theory only for $G$-spaces, and hence cannot be represented by any $G$-spectrum.
\end{remark}
Since $E\mathscr{F}$ only has empty and contractible fixed point subspaces, \autoref{burnsidehom} helps us to compute the 0th degree equivariant homology of $E\mathscr{F}$ explicitly:
\begin{proposition} \label{degree0}
The trivial map $E\mathscr{F}\rightarrow\{*\}$ implies an inclusion
$$\underline{H}_0^G(E\mathscr{F},A_G)\hookrightarrow\underline{H}_0^G(*,A_G)\cong A_G.$$
The image of $\underline{H}_0^G(E\mathscr{F},A_G)(G/L)$ in $A(L)$ is generated by all $\{L/K\}\in A(L)$ with $K\in\mathscr{F}$.
\end{proposition}
We already have a small splitting here:
\begin{lemma} \label{homsplit}
When $|G|$ is inverted, $M_\mathscr{F}:=\underline{H}_0^G(E\mathscr{F};A_G)$ is a direct summand of $A_G$ as Mackey functors.
\end{lemma}
\paragraph{Proof:} Assume that $|G|$ is inverted everywhere. For any $H\subset G$, define a linear map
$$\varphi_H:A(G)\rightarrow\mathbb{Z}$$
which sends each $G$-set $S$ to $|S^H|$.
Let $s_{(K,H)}$ denote $\varphi_K(\{G/H\})=|G/H|^K$, which can be computed as the product between $|W_GH|$ and the number of subgroups of $G$ containing $K$ and in the conjugacy class of $H$. Thus $s_{(K,H)}\neq 0$ if and only if $K$ is sub-conjugate to $H$.
When $|G|$ is inverted, all $|W_GH|$ and non-zero $s_{(K,H)}$ become invertible. We choose elements $e_H\in A(G)$ inductively by defining
$$e_H=|W_GH|^{-1}\left(\{G/H\}-\sum_K s_{(K,H)}^{-1}e_K\right)$$
with one $K$ chosen from each conjugacy class that contains a proper subgroup of $H$.
By induction, we have
$$\varphi_K(e_H)=
\begin{cases}
1, \text{ if }K\text{ is conjugate to }H\\
0, \text{ otherwise}
\end{cases}$$
Since the dimension of $A(G)$ agrees with the number of conjugacy classes, the collection of $e_H$, with one $H$ chosen from each conjugacy class, forms a basis of $A(G)$.
\medskip
In general, for any $H\subset L\subset G$, define a linear map
$$\varphi_H^L:A(L)\rightarrow\mathbb{Z}$$
which sends each $L$-set to the size of its $H$-fixed subset. We can get a similar basis $\{e_H^L\}$, with one $H$ chosen from each conjugacy class of subgroups of $L$, such that
$$\varphi_K^L(e_H^L)=
\begin{cases}
1, \text{ if }K\text{ is conjugate to }H\text{ in }L\\
0, \text{ otherwise}
\end{cases}$$
\medskip
Now we discuss how these maps interact with transfer and restriction maps:
\smallskip
For any $K\subset L_1\subset L_2\subset G$, the restriction map $R_{L_1}^{L_2}$ keeps each $L_2$-set but views it as an $L_1$-set, hence does not change its $K$-fixed subset. So we have
$$\varphi_K^{L_2}=\varphi_K^{L_1}\circ R_{L_2}^{L_1}.$$
On the other hand, for any $K\subset L_1$, the elements in $\{L_1/K\}$ have isotropy groups conjugate to $K$ inside $L_1$. The elements in $T_{L_1}^{L_2}(\{L_1/K\})=\{L_2/K\}$ have isotropy groups conjugate to $K$ inside $L_2$. There may be more isotropy groups. But these additional isotropy groups are still chosen from the conjugacy class of $K$ inside $G$. According to the construction of the basis element $e_H^L$, we have
$$\varphi_H^{L_2}(T_{L_1}^{L_2}e_K^{L_1})\neq 0\text{ only if }H\text{ is conjugate to a subgroup of }K\text{ in }G$$
Let
$$N(L):=\bigcap_{H\in\mathscr{F},H\subset L}\ker\varphi_H^L$$
In other words, $N(L)$ is generated by $e_H^L$ for all $H\subset L$ and $H\notin\mathscr{F}$. The above discussion tells us that $N(L)$ is closed under transfer and restriction maps. Thus we get a sub-Mackey functor whose value at $G/L$ agrees with $N(L)$. Denote that as $N_\mathscr{F}$.
\smallskip
Notice that for each $L\subset G$, $e_H^L\in N(L)=N_\mathscr{F}(G/L)$ if $H\notin\mathscr{F}$, $e_H^L\in M_\mathscr{F}(G/L)$ if $H\in\mathscr{F}$, according to \autoref{degree0}. Moreover, consider any nontrivial element
$$a_1\{L/K_1\}+a_2\{L/K_2\}+...+a_n\{L/K_n\}\in M_\mathscr{F}(G/L)$$
with $a_1,a_2,...,a_n\neq 0$, $K_1,K_2,...,K_n$ in different conjugacy classes in $\mathscr{F}$, and $|K_1|\leq |K_2|\leq...\leq|K_n|$. The map $\varphi_{K_n}^L$ sends $\{L/K_1\},...,\{L/K_{n-1}\}$ to zero but $\{L/K_n\}$ to a positive value. Thus this element is not in $\ker\varphi_{K_n}^L$, and hence $M_\mathscr{F}(G/L)\cap N(L)=\emptyset$.
\medskip
In conclusion, we have $A_G=M_\mathscr{F}\oplus N_\mathscr{F}$ and $M_\mathscr{F}$ appears as a direct summand. $\Box$
\begin{remark} \label{injectivevar}
When the multiplicative structure is added into consideration, the maps $\varphi_H^L$ appear as the components of the ring isomorphism from $A(L)$ to several copies of $\mathbb{Z}[|G|^{-1}]$, with one $H$ chosen from each conjugacy class inside $L$. The generators $e_H^L$ become idempotent elements. So we can view $M_\mathscr{F}$ as a direct summand of $A_G$ as Green functors. This also explains why $e_HX$ in \autoref{idemsplit} can be expressed as a product with universal spaces. More details can be found in \textbf{[Bar08]} and \textbf{[LMS86}, Chapter V].
\end{remark}
For positive degrees, we have
\begin{theorem} \label{torsion}
For any family $\mathscr{F}$, the $\mathbb{Z}$-graded, Mackey functor valued homology
$$\underline{H}_*^G(E\mathscr{F};A_G)$$
contains only torsion when the degree is positive. Moreover, the torsion only has prime factors which divide $|G|$.
\end{theorem}
This is the most important property we want about $E\mathscr{F}$. We will use the rest of this section to prove it.
\bigskip
The main idea is an induction on the size of $\mathscr{F}$. For the base case, we have
\begin{lemma} \label{orduni}
Let $BG=EG/G$ be the classifying space of principal $G$-bundles. Then $H_*(BG;\mathbb{Z})$ contains only torsion when the degree is positive. Moreover, the torsion only has prime factors which divide $|G|$.
\end{lemma}
This is a standard result about $EG$. We give one possible proof below. The same idea will be generalized to other universal spaces $E\mathscr{F}$ in \autoref{eqorduni} and \autoref{genuniorb}.
\paragraph{Proof:} Give $EG$ the standard $G$-CW structure which only contains $G$-free cells. Consider the map between cellular chain complexes induced by the projection $EG\rightarrow EG/G=BG$:
$$C_*(EG)\rightarrow C_*(BG).$$
Define another map between chain complexes
$$C_*(BG)\rightarrow C_*(EG)$$
which sends each cell $e$ in $BG$ to the sum of all cells in $EG$ which are sent to $e$ under $EG\rightarrow BG$. Then the composition
$$C_*(BG)\rightarrow C_*(EG)\rightarrow C_*(BG)$$
is the multiplication by $|G|$ since all cells in $EG$ are $G$-free. The same $|G|$-multiplication passes to homology:
$$H_*(BG)\rightarrow H_*(EG)\rightarrow H_*(BG).$$
Since $EG$ is contractible, $H_*(EG)=\mathbb{Z}$, which is concentrated in degree $0$. Thus when the degree is positive, $H_*(BG)$ contains only torsion which divides $|G|$. $\Box$
\bigskip
In the case when $\mathscr{F}$ only contains the trivial subgroup, $E\mathscr{F}$ becomes $EG$ and only has orbit spaces homotopic to $BH$ for $H\subset G$. \autoref{orduni} and \autoref{burnsidehom} imply that it satisfies \autoref{torsion}.
\medskip
For general $\mathscr{F}$, we first give a construction of $E\mathscr{F}$:
\smallskip
\begin{definition} \label{joindef}
The \textbf{join} $X\ast Y$ of two spaces $X,Y$ is defined as
$$X\ast Y:= (X\times Y\times[0,1]\sqcup X\sqcup Y)/\sim,$$
with equivalence relation given by projections
$$X\times Y\times\{0\}\rightarrow X, \text{ }X\times Y\times\{1\}\rightarrow Y.$$
\end{definition}
It's not hard to check:
\begin{lemma} \label{join}
For $G$-CW complexes $X,Y$, $X\ast Y$ has a natural $G$-CW structure. For any subgroup $H$, $(X\ast Y)^H$ is
\textbf{(a)} contractible if and only if at least one of $X^H, Y^H$ is contractible.
\textbf{(b)} empty if and only if both $X^H, Y^H$ are empty.
\end{lemma}
Now in order to construct $E\mathscr{F}$, it suffices to find $G$-spaces $X_1,X_2,...,X_n$, such that:
\textbf{(a)} If $H\in\mathscr{F}$, $(X_i)^H$ is contractible for some $i$.
\textbf{(b)} If $H\notin\mathscr{F}$, $(X_i)^H$ is empty for all $i$.
Then $X_1\ast X_2\ast...\ast X_n$ will be a construction for $E\mathscr{F}$ by the above lemma.
\medskip
For any subgroup $H$, consider the $G$-space $G\times_{N_GH}EW_GH$. Here $N_GH$ is the normalizer of $H$ in $G$. We view $EW_GH$ as a $N_GH$-space. Define the product with $N_GH$ left-acting on $EW_GH$ and right-acting on $G$.
The fixed point subspaces of $G\times_{N_GH}EW_GH$ have the desired properties: For any $a,g\in G$ and $x\in EW_GH$, if the action of $g$ fixes the point $a\times x$:
$$g(a\times x)=ga\times x=a\times x,$$
then there is $h\in N_GH$, such that $hx=x$, $gah^{-1}=a$. Since all points in $EW_GH$ have isotropy group $H$, we have $h\in H$. Then $g=aha^{-1}\in aHa^{-1}$.
Therefore, the isotropy group of $a\times x$ is $aHa^{-1}$, and we have
\begin{lemma} \label{joinpiece}
The fixed point subspace $(G\times_{N_GH}EW_GH)^K$ is non-empty if and only if $K$ is sub-conjugate to $H$. Moreover, when $K=gHg^{-1}$, we have
$$(G\times_{N_GH}EW_GH)^K=\{g\}\times EW_GH,$$
which is contractible.
\end{lemma}
Therefore, $E\mathscr{F}$ can be constructed as the join of $G\times_{N_GH}EW_GH$, with one $H$ chosen from each conjugacy class in $\mathscr{F}$.
\begin{remark} \label{isotropytypes}
Notice that all points in $G\times_{N_GH}EW_GH$ have isotropy groups conjugate to $H$. The collection of all possible isotropy groups of points in $E\mathscr{F}$ agrees with $\mathscr{F}$.
\end{remark}
\medskip
The homology of a join can be computed by the Mayer-Vietoris sequence:
$$X\ast Y=(X\times Y\times[0,1]\sqcup X\sqcup Y)/\sim=(X\times Y\times\left[0,\frac{1}{2}\right]\sqcup X/\sim)\cup(X\times Y\times\left[\frac{1}{2},1\right]\sqcup Y/\sim).$$
We write $X\ast Y$ as the union of two mapping cylinders, which are homotopic to $X,Y$ respectively. The intersection of these two cylinders is $X\times Y\times\{1/2\}\simeq X\times Y$.
Thus we have a short exact sequence
$$0\rightarrow C_*(X\times Y)\rightarrow C_*(X)\oplus C_*(Y)\rightarrow C_*(X\ast Y)\rightarrow 0$$
which passes to fixed point subspaces. So we get a long exact sequence on equivariant homology with coefficients in $A_G$:
$$...\rightarrow \underline{H}_{n+1}^G(X\ast Y;A_G)\rightarrow \underline{H}_n^G(X\times Y;A_G)\rightarrow \underline{H}_n^G(X;A_G)\oplus \underline{H}_n^G(Y;A_G)\rightarrow \underline{H}_n^G(X\ast Y;A_G)\rightarrow...$$
On degree $0$, the map
$$\underline{H}^G_0(X\times Y;A_G)\rightarrow \underline{H}^G_0(X;A_G)\oplus \underline{H}^G_0(Y;A_G)$$
is always an inclusion since the 0th homology is determined by the number of connected components for each fixed point subspace. So we have
\begin{lemma} \label{joinhom}
If $X,Y,X\times Y$ satisfy \autoref{torsion}, then so does $X\ast Y$.
\end{lemma}
\medskip
\paragraph{Proof of \autoref{torsion}:} Assume that \autoref{torsion} is true for all families smaller than $\mathscr{F}$. Let $\mathscr{F}^\prime$ be a smaller family which is obtained by removing the conjugacy class of one largest subgroup $H\in\mathscr{F}$. Then $E\mathscr{F}$ can be constructed as
$$E\mathscr{F}=E\mathscr{F}^\prime\ast(G\times_{N_GH}EW_GH).$$
\medskip
For any $K\lhd L\subset G$, the discussion before \autoref{joinpiece} shows that the space $(G\times_{N_GH}EW_GH)^K$ is either empty or several copies of $EW_GH$. Since all points in each copy of $EW_GH$ have the same isotropy group, $(G\times_{N_GH}W_GH)^K/L$ is either empty or the disjoint union of $BN$ for some subgroups $N\subset N_GH\subset G$. \autoref{orduni} shows that its positive degree homology only contains torsion dividing $|G|$. Thus \autoref{burnsidehom} shows that $G\times_{N_GH}EW_GH$ satisfies \autoref{torsion}.
\medskip
Since $E\mathscr{F}^\prime$ satisfies \autoref{torsion}, according to \autoref{joinhom}, now it suffices to prove the theorem for $E\mathscr{F}^\prime\times(G\times_{N_GH}EW_GH)$. Using $X(H)$ to denote $G\times_{N_GH}EW_GH$, the homology of $E\mathscr{F}^\prime\times X(H)$ can be computed by the equivariant Künneth spectral sequence \textbf{[LM04]}:
$$E_{p,q}^2=\underline{Tor}^{A_G}_{p,q}(\underline{H}_*^G(E\mathscr{F}^\prime;A_G),\underline{H}_*^G(X(H);A_G))\Rightarrow\underline{H}_*^G(E\mathscr{F}^\prime\times X(H);A_G).$$
When we invert $|G|$, the homology of both $E\mathscr{F}^\prime$ and $X(H)$ is concentrated in degree 0. Moreover, $\underline{H}_0^G(E\mathscr{F}^\prime;A_G)=M_{\mathscr{F}^\prime}$ appears as a direct summand of $A_G$ by \autoref{homsplit}. Thus the $E_2$-page collapses into a single box product:
$$E^2=E^2_{0,0}=M_\mathscr{F}\text{ }\Box\text{ } \underline{H}_*^G(X(H);A_G).$$
Therefore, $E\mathscr{F}^\prime\times X(H)$ has trivial homology in positive degrees when $|G|$ is inverted. $\Box$
\bigskip
For special choices of $\mathscr{F}$, the torsion in $\underline{H}_*(E\mathscr{F};A_G)$ may not cover all prime factors of $|G|$. We emphasize the criterion which follows \autoref{burnsidehom}:
\begin{proposition} \label{torsiontype}
The torsion types in $\underline{H}_*^G(E\mathscr{F},A_G)$ come from $H_*(E\mathscr{F}^K/L)$ for all $K\lhd L\subset G$.
\end{proposition}
\section{Splittings of rational $G$-spectra}
In this section, we will reprove \autoref{computesplit} and prove \autoref{orderinvert}. The main idea is to consider the partial splitting induced by a cofiber sequence involving the universal space $E\mathscr{F}$:
$$E\mathscr{F}_+\rightarrow S^0\rightarrow\widetilde{E\mathscr{F}}$$
where the map $E\mathscr{F}_+\rightarrow S^0$ sends $E\mathscr{F}$ into the non-basepoint of $S^0$. The cofiber $\widetilde{E\mathscr{F}}$ can be viewed as the non-equivariant unbased suspension of $E\mathscr{F}$ but with a natural $G$-action.
\smallskip
\begin{theorem} \label{ringsplit}
When certain prime factors of $|G|$ are inverted, this cofiber sequence splits the category of $G$-spectra, in the sense that
$$X\simeq (E\mathscr{F}_+\wedge X)\vee(\widetilde{E\mathscr{F}}\wedge X),$$
$$[E\mathscr{F}_+\wedge X,\widetilde{E\mathscr{F}}\wedge Y]^G=[\widetilde{E\mathscr{F}}\wedge Y,E\mathscr{F}_+\wedge X]^G=0$$
for any $G$-spectra $X,Y$. Moreover, the functors
$$X\mapsto E\mathscr{F}_+\wedge X, \text{ }X\mapsto\widetilde{E\mathscr{F}}\wedge X$$
are symmetric monoidal.
\end{theorem}
We will give a criterion about which prime factors we need to invert in \autoref{invertprime}.
\medskip
The complete splitting in \autoref{computesplit} can be obtained by iterating \autoref{ringsplit} for all choices of $\mathscr{F}$. Instead of full rationalization, we will show that it suffices to invert all prime factors of $|G|$ in our proof.
\bigskip
As based $G$-spaces, both $E\mathscr{F}_+$ and $\widetilde{E\mathscr{F}}$ can be characterized by their fixed point subspaces:
$$E\mathscr{F}_+^H\simeq S^0 \text{ if } H\in\mathscr{F}, \text{ } E\mathscr{F}_+^H\simeq * \text{ if } H\notin\mathscr{F},$$
$$(\widetilde{E\mathscr{F}})^H\simeq S^0 \text{ if } H\notin\mathscr{F}, \text{ } (\widetilde{E\mathscr{F}})^H\simeq * \text{ if } H\in\mathscr{F}.$$
\autoref{monoid} also implies a multiplicative structure on $\widetilde{E\mathscr{F}}$. But unlike $E\mathscr{F}$, the map $S^0\rightarrow\widetilde{E\mathscr{F}}$ also provides a unit, which makes $\widetilde{E\mathscr{F}}$ into a based $G$-Hopf space.
\medskip
Consider the suspension of the cofiber sequence about $E\mathscr{F}$:
$$\Sigma^\infty E\mathscr{F}_+\rightarrow S\rightarrow\Sigma^\infty\widetilde{E\mathscr{F}}.$$
\begin{theorem} \label{cofibsplit}
When $|G|$ is inverted, there exists a left inverse of $\Sigma^\infty E\mathscr{F}_+\rightarrow S$.
\end{theorem}
\paragraph{Proof:} According to \autoref{degree0}, the map $\Sigma^\infty E\mathscr{F}_+\rightarrow S$ induces an inclusion in homology:
$$\underline{HA_G}_0(\Sigma^\infty E\mathscr{F}_+)=:M_\mathscr{F}\hookrightarrow\underline{HA_G}_0S=A_G.$$
When $|G|$ is inverted, \autoref{torsion} shows that the $HA_G$-homology is concentrated in degree 0. Thus it suffices to find a map $S\rightarrow\Sigma^\infty E\mathscr{F}_+$ which induces a projection from $A_G$ to $M_\mathscr{F}$ on their $HA_G$-homology. After composition with $\Sigma^\infty E\mathscr{F}_+\rightarrow S$, we get a self-map of $\Sigma^\infty E\mathscr{F}_+$ that induces an isomorphism on $HA_G$-homology. Since $\Sigma^\infty E\mathscr{F}_+$ is a connective spectrum, this self map must be a weak equivalence.
\smallskip
For each $H\in\mathscr{F}$, the transfer and restriction maps in Mackey functors between images of $G/G$ and $G/H$ are induced by stable maps
$$T: S=\Sigma^\infty S^0=\Sigma^\infty G/G_+\rightarrow\Sigma^\infty G/H_+$$
$$R: \Sigma^\infty G/H_+\rightarrow \Sigma^\infty G/G_+=S.$$
$R$ is induced by the space-level map $G/H_+\rightarrow S^0$, which sends $G/H$ into the non-basepoint of $S^0$. According to \autoref{isotropytypes}, $G/H$ can be mapped into $E\mathscr{F}$. Thus $R$ is factorized as
$$\Sigma^\infty G/H_+\rightarrow\Sigma^\infty E\mathscr{F}_+\rightarrow S.$$
Let $\iota_H$ be the composition of the first map above and $T$:
$$\iota_H:S\xrightarrow{T}\Sigma^\infty G/H_+\rightarrow\Sigma^\infty E\mathscr{F}_+.$$
The composition
$$S\xrightarrow{\iota_H}\Sigma^\infty E\mathscr{F}_+\rightarrow S$$
agrees with $R\circ T$, which is the equivariant Euler characteristic of $G/H$. We can fully describe the self-map on
$$\underline{HA_G}_0S\cong A_G\cong\underline{\pi}_0S$$
induced by the Euler characteristic as follows:
For each $L\subset G$, the induced self-map on $A_G(G/L)=A(L)$ is the multiplication by $\{G/H\}$ (which is viewed as an $L$-set). We refer to $\textbf{[LMS86}$, V.1 and V.2] for more details.
\medskip
Now we choose one $H$ from each conjugacy class in $\mathscr{F}$. Assign a number $c_H\in\mathbb{Z}[|G|^{-1}]$ for each such $H$. Consider the map
$$\sum_H c_H\iota_H:S\rightarrow\Sigma^\infty E\mathscr{F}_+.$$
The composition
$$S\xrightarrow{\sum_Hc_H\iota_H}\Sigma^\infty E\mathscr{F}_+\rightarrow S$$
induces a self-map on $A_G(G/L)=A(L)$ as the multiplication by $\sum_H c_H\{G/H\}$.
Recall that $\underline{HA_G}_0(\Sigma^\infty E\mathscr{F}_+)(G/L)=M_\mathscr{F}(G/L)\subset A(L)$ is generated by all $\{L/J\}$ with $J\in\mathscr{F}$. The statement that $\sum_Hc_H\iota_H$ induces the projection $A_G\rightarrow M_\mathscr{F}$ is equivalent to
$$\{L/J\}\cdot\sum_Hc_H\{G/H\}=\{L/J\}$$
for any $J\in\mathscr{F}$, $J\subset L$.
According to \autoref{injectivevar}, $\prod_{K\subset L}\varphi_K^L$ is an injective ring homomorphism from $A(L)$ to copies of $\mathbb{Z}$. Thus it suffices to check the equation above on the images under each $\varphi_K^L$. Recall that $\varphi_K^L$ is defined by sending each $L$-set to the number of $K$-fixed points. Thus the equation above under $\varphi_K^L$ becomes
$$|(L/J)^K|\cdot\sum_Hc_H|(G/H)^K|=|(L/J)^K|.$$
Notice that $|(L/J)^K|=0$ for any $K\notin\mathscr{F}$ (since $K$ is not sub-conjugate to $J\in\mathscr{F}$). Thus it suffices to choose $c_H$ such that
$$\sum_Hc_H|(G/H)^K|=\sum_H c_Hs_{(K,H)}=1, \text{ } \forall K\in\mathscr{F}$$
with one $H$ chosen from each conjugacy class in $\mathscr{F}$.
Recall that $s_{(K,H)}$ is non-zero if and only if $K$ is sub-congujate to $H$. Denote this relation by $[K]\leq[H]$. We have
$$\sum_{[K]\leq[H]\subset\mathscr{F}} c_Hs_{(K,H)}=1.$$
Thus
$$c_K=s_{(K,K)}^{-1}\left(1-\sum_{[K]\lneqq[H]\subset\mathscr{F}}c_Hs_{(K,H)}\right).$$
Since $s_{(K,K)}=|W_GK|$ is invertible when $|G|$ is inverted, $c_K$ can be chosen inductively from larger subgroups to smaller ones. Moreover, it's clear that the denominator of any $c_K$ only contains prime factors dividing $|G|$.
Now the map $\sum_Hc_H\iota_H$ gives us the required projection on the 0th homology, hence becomes the left inverse of $\Sigma^\infty E\mathscr{F}_+\rightarrow S$. $\Box$
\bigskip
\paragraph{Proof of \autoref{ringsplit}:} Assume that $|G|$ is inverted. The left inverse in \autoref{cofibsplit} and the multiplicative structure on space $E\mathscr{F}$ make $\Sigma^\infty E\mathscr{F}_+$ into a ring spectrum. So we have a splitting cofiber sequence of ring spectra:
$$\Sigma^\infty E\mathscr{F}_+\rightarrow S\rightarrow\Sigma^\infty\widetilde{E\mathscr{F}}.$$
Thus we have
$$X\simeq(E\mathscr{F}_+\wedge X)\vee(\widetilde{E\mathscr{F}}\wedge X).$$
Since
$$E\mathscr{F}_+\wedge E\mathscr{F}_+\simeq E\mathscr{F}_+\text{ and }\widetilde{E\mathscr{F}}\wedge\widetilde{E\mathscr{F}}\simeq\widetilde{E\mathscr{F}},$$
the functors
$$X\mapsto E\mathscr{F}_+\wedge X,\text{ }X\mapsto\widetilde{E\mathscr{F}}\wedge X$$
are symmetric monoidal.
Moreover, $E\mathscr{F}_+\wedge\widetilde{E\mathscr{F}}\simeq *$ since all its fixed point subspaces are contractible. Use $A,B$ to denote the suspensions of $E\mathscr{F}_+$ and $\widetilde{E\mathscr{F}}$ (in either order). The universal coefficient spectral sequence \textbf{[LM04]} tells us that for any $G$-spectra $X,Y$, we have
$$\underline{Ext}^{*,*}_{\underline{B}_*}(\underline{B}_*(A\wedge X),\underline{(B\wedge Y)}_*)\Rightarrow\underline{(B\wedge Y)}^*(A\wedge X)=[A\wedge X,\Sigma^* B\wedge Y]^G.$$
Since
$$\underline{B}_*(A\wedge X)=\underline{\pi}_*(B\wedge A\wedge X)=0,$$
the spectral sequence has trivial $E_2$-page. Thus we have
$$[E\mathscr{F}_+\wedge X,\widetilde{E\mathscr{F}}\wedge Y]^G=[\widetilde{E\mathscr{F}}\wedge Y,E\mathscr{F}_+\wedge X]^G=0.$$
$\Box$
\medskip
For special choices of $\mathscr{F}$, here is a criterion about which prime factors of $|G|$ we need to invert:
\begin{remark} \label{invertprime}
It suffices to invert the prime factors of the denominators of all $c_H$, and the torsion in $\underline{H}_*^G(E\mathscr{F};A_G)$.
According to \autoref{torsiontype}, the torsion in $\underline{H}_*^G(E\mathscr{F};A_G)$ comes from the torsion in $H_*(E\mathscr{F}^K/L)$ for all $K\lhd L\subset G$.
\end{remark}
\bigskip
Now we list all different families $\mathscr{F}_1,\mathscr{F}_2,...,\mathscr{F}_n$ of $G$. Use \autoref{ringsplit} repeatedly:
$$X\simeq((E\mathscr{F}_1)_+\wedge X)\vee(\widetilde{E\mathscr{F}_1}\wedge X)$$
$$\simeq((E\mathscr{F}_1)_+\wedge(E\mathscr{F}_2)_+\wedge X)\vee((E\mathscr{F}_1)_+\wedge\widetilde{E\mathscr{F}_2}\wedge X)$$
$$\vee(\widetilde{E\mathscr{F}_1}\wedge(E\mathscr{F}_2)_+\wedge X)\vee(\widetilde{E\mathscr{F}_1}\wedge\widetilde{E\mathscr{F}_2}\wedge X)\simeq...$$
Finally we have
$$X\simeq\bigvee_\lambda E\lambda\wedge X$$
where the wedge sum is computed over all functions
$$\lambda:\{1,2,...,n\}\rightarrow \text{\{based G-spaces\}}$$
sending each $i$ to either $(E\mathscr{F}_i)_+$ or $\widetilde{E\mathscr{F}_i}$. We define $E\lambda$ as the smash product of all elements in the image of $\lambda$.
\medskip
Since the fixed point subspaces of each $E\mathscr{F}_+$ and $\widetilde{E\mathscr{F}}$ are homotopic to either $S^0$ or $*$, the same thing is true for $E\lambda$.
\begin{lemma}
For any $\lambda$, there is at most one conjugacy class $[H]$ such that $(E\lambda)^H\simeq S^0$. Moreover, for each $[H]$, there is a unique $\lambda$ with $(E\lambda)^H\simeq S^0$. We call it $\lambda_H$.
\end{lemma}
\paragraph{Proof:} For any two subgroups $K,H$ which are not in the same conjugacy class, assume that $|K|\leq |H|$. Let $\mathscr{F}$ be the class of all subgroups which are sub-conjugate to $K$. Then
$$(E\mathscr{F}_+)^K\simeq(\widetilde{E\mathscr{F}})^H\simeq S^0,$$
$$(E\mathscr{F}_+)^H\simeq(\widetilde{E\mathscr{F}})^K\simeq *.$$
No matter which of $E\mathscr{F}_+$, $\widetilde{E\mathscr{F}}$ is chosen in $E\lambda$, we have to make one of $(E\lambda)^K,(E\lambda)^H$ contractible.
\smallskip
On the other hand, if $(E\lambda)^H\simeq S^0$, we have to choose $E\mathscr{F}_+$ if $H\in\mathscr{F}$, and $\widetilde{E\mathscr{F}}$ if $H\notin\mathscr{F}$. Thus $\lambda$ is unique. $\Box$
\bigskip
We can view $E\lambda_H$ as another kind of universal space and characterize $E\lambda_H$ by fixed point subspaces:
\begin{proposition} \label{moreclass}
The $K$-fixed subspace of $E\lambda_H$ is homotopic to $S^0$ if $K$ is conjugate to $H$, and contractible otherwise.
Any based $G$-CW complex with the same fixed point spaces is homotopic to $E\lambda_H$.
\end{proposition}
Now we get the splitting in \autoref{computesplit} with just inverting $|G|$. Thus \autoref{orderinvert} is proved.
\section{Splittings for dihedral groups}
The partial splitting in \autoref{ringsplit}, together with \autoref{invertprime}, is quite powerful in the study of $G$-spectra. When we just invert some of the prime factors of $|G|$, the splitting in \autoref{ringsplit} may still happen for special choices of $\mathscr{F}$. We can consider different splittings with different primes inverted in order to recover information about given spectra.
The following three sections are devoted to explicit computations for the dihedral group $G=D_{2p}$ with $p$ an odd prime. We will first compute the $RO(G)$-graded homotopy group $\pi_\bigstar(H\underline{\mathbb{Z}})$, which expresses the equivariant cohomology of a point with coefficients in $\underline{\mathbb{Z}}$. We will improve the computations in \textbf{[KL20]} and \textbf{[Zou18]} and compute $\pi_\bigstar(H\underline{\mathbb{Z}})$ as an $RO(G)$-graded ring with the idea explained above.
\bigskip
Use $\zeta, \tau$ to express the generators in $D_{2p}$ such that $\zeta^p=\tau^2=1$, $\zeta\tau=\tau\zeta^{-1}$.
The $RO(G)$ grading is generated by three different kinds of irreducible representations: The constant representation (denoted by $1$), the sign representation (denoted by $\sigma$), and $p-1$ different 2-dimensional dihedral representations (where $\tau$ acts as a reflection and $\zeta$ acts as a rotation).
The Periodicity Theorem in \textbf{[KL20]} shows that the cellular structures of the representation spheres for different dihedral representations are isomorphic. Let $\gamma$ be any single dihedral representation. It suffices to consider all degrees with form $k+m\sigma+n\gamma$, $\forall k,m,n\in\mathbb{Z}$. The additive and multiplicative structures of homotopy groups on these degrees already reveal all information about $\pi_\bigstar$.
\bigskip
\paragraph{Main Strategy:} Consider the families
$$\mathscr{F}_1=\{\{1\},\{1,\zeta^i\tau\}:i=1,2,...,p\},$$
$$\mathscr{F}_2=\{\{1\},\{1,\zeta,\zeta^2,...,\zeta^{p-1}\}\}.$$
We will prove:
\begin{proposition} \label{twosplit}
The splitting in \autoref{ringsplit} happens when
\textbf{(a)} $\mathscr{F}=\mathscr{F}_1$ and the prime $p$ is inverted, or
\textbf{(b)} $\mathscr{F}=\mathscr{F}_2$ and the prime $2$ is inverted.
\end{proposition}
Let $A$ be any $G$-ring spectrum. If we can compute the $RO(G)$-graded homotopy and corresponding ring structures, with either $2$ or $p$ inverted, for
$$(E\mathscr{F}_1)_+\wedge A,\text{ }\widetilde{E\mathscr{F}_1}\wedge A, \text{ }(E\mathscr{F}_2)_+\wedge A\text{ and }\widetilde{E\mathscr{F}_2}\wedge A,$$
then we can recover $\pi_\bigstar(A)\otimes\mathbb{Z}[1/2]$ and $\pi_\bigstar(A)\otimes\mathbb{Z}[1/p]$, hence also $\pi_\bigstar(A)$ itself.
Intuitively, we decompose the computation into $C_2$-spectra when $p$ is inverted, and into $C_p$-spectra when $2$ is inverted. Since $\pi_\bigstar(H\underline{Z})$ is already fully computed when $G=C_2$ or $C_p$, we can use these computations as building blocks to recover $\pi_\bigstar(H\underline{\mathbb{Z}})$ for $G=D_{2p}$.
\bigskip
In order to prove \autoref{twosplit}, we need an equivariant generalization of \autoref{orduni}:
\begin{lemma} \label{eqorduni}
Let $X$ be a $D_{2p}$-CW complex which only has $C_p$-free cells. Consider the projection $X\rightarrow X/C_p$, which is a $C_2$-map for any choice of $C_2\subset D_{2p}$ acting on $X$.
\textbf{(a)} The cokernel of
$$H_*(X;\mathbb{Z})\rightarrow H_*(X/C_p;\mathbb{Z})$$
only contains $p$-torsion;
\textbf{(b)} We have an isomorphism
$$H_*(X^{C_2};\mathbb{Z})\xrightarrow{\cong}H_*((X/C_p)^{C_2};\mathbb{Z}).$$
If $p$ is inverted and the map in part (a) is an injection, then $X\rightarrow X/C_p$ is a $C_2$-equivalence.
\end{lemma}
\paragraph{Proof:} Part (a) follows by the same argument as in \autoref{orduni}.
For part (b), any $D_{2p}$-cell in $X$ has the form of either $(D_{2p})_+\wedge e$ or $(D_{2p}/C_2)_+\wedge e$. In the first case, its image in $X/C_p$ is not $C_2$-fixed. In the second case, its image is $C_2$-fixed. But no matter which $C_2\subset D_{2p}$ is chosen, the action of $C_2$ fixes exactly one copy of $e$ and exchanges the rest in pairs. Thus we have an isomorphism $C_*(X^{C_2})\cong C_*((X/C_p)^{C_2})$, hence an isomorphism on homology.
If $p$ is inverted, the cokernel of the map in part (a) disappeared. If that map is also an injection, it becomes an isomorphism. Thus the map $X\rightarrow X/C_p$ induces equivalences on both the underlying spaces and $C_2$-fixed subspaces, hence becomes a $C_2$-equivalence. $\Box$
\medskip
\begin{remark} \label{eqordunimore}
\autoref{eqorduni} also works for $D_{2p}$-CW spectra when we replace the fixed point subspaces by geometric fixed point spectra.
Moreover, since the map $X\wedge Y\rightarrow (X/C_p)\wedge (Y/C_p)$ can be factorized into
$$X\wedge Y\rightarrow (X\wedge Y)/C_p\rightarrow (X/C_p)\wedge(Y/C_p)$$
as $C_2$-maps, the $C_2$-equivalence in \autoref{eqorduni} also preserves multiplicative structures.
\end{remark}
\medskip
\paragraph{Proof of \autoref{twosplit}:} According to \autoref{invertprime} and \autoref{torsiontype}, it suffices to show that the denominators of all $c_H$ and all torsion in $H_*(E\mathscr{F}^K/L)$ (with any $K\lhd L\subset G$) are inverted.
\medskip
As shown in the proof of \autoref{cofibsplit}, $c_H$ is computed by
$$\sum_H c_Hs_{(K,H)}=1,\text{ }\forall K\in\mathscr{F}$$
with one $H$ chosen from each conjugacy class in $\mathscr{F}$. Recall that $s_{(K,H)}=|(G/H)^K|$, which is non-zero if and only if $K$ is sub-conjugate to $H$.
When $\mathscr{F}=\mathscr{F}_1$, we have
$$2pc_{\{1\}}+pc_{\{1,\tau\}}=c_{\{1,\tau\}}=1.$$
Thus $c_{\{1\}}=(1-p)/2p$ and $c_{\{1,\tau\}}=1$. Since $1-p$ is even, the denominators only contain prime $p$.
\smallskip
When $\mathscr{F}=\mathscr{F}_2$, we have
$$2pc_{\{1\}}+2c_{\{1,\zeta,...\}}=2c_{\{1,\zeta,...\}}=1.$$
Thus $c_{\{1\}}=0$ and $c_{\{1,\zeta,...\}}=1/2$, which only have $2$ in denominators.
\bigskip
Up to homotopy, $E\mathscr{F}^K/L$ can be any of $E\mathscr{F}$, $E\mathscr{F}^{C_2}$, $E\mathscr{F}^{C_p}$, $E\mathscr{F}^{D_{2p}}$ $E\mathscr{F}^{C_p}/D_{2p}$, $E\mathscr{F}/C_2$, $E\mathscr{F}/C_p$, $E\mathscr{F}/D_{2p}$.
\bigskip
\textbf{When $\mathscr{F}=\mathscr{F}_1$:} As non-equivariant spaces, we have $E\mathscr{F}\simeq E\mathscr{F}^{C_2}\simeq *$, $E\mathscr{F}^{C_p}=E\mathscr{F}^{D_{2p}}=\emptyset$. Thus $E\mathscr{F}^{C_p}/D_{2p}=\emptyset$.
As a $C_2$-space, $E\mathscr{F}\simeq *$ since both its fixed point subspaces are contractible. Thus $E\mathscr{F}/C_2\simeq *$.
For $E\mathscr{F}/C_p$ and $E\mathscr{F}/D_{2p}$, we can apply \autoref{eqorduni} with $X=E\mathscr{F}$. Notice that $E\mathscr{F}\simeq *$ as a non-equivariant space. Thus the map in \autoref{eqorduni}(a) is an injection. When $p$ is inverted, we have a $C_2$-equivalence $E\mathscr{F}\simeq E\mathscr{F}/C_p$. So we have $E\mathscr{F}/C_p\simeq E\mathscr{F}\simeq *$ and $E\mathscr{F}/D_{2p}\simeq E\mathscr{F}/C_2\simeq *$.
Therefore, the homology groups of these spaces only have $p$-torsion when the degree is positive.
\bigskip
\textbf{When $\mathscr{F}=\mathscr{F}_2$:} We can give an explicit construction as $E\mathscr{F}\simeq EC_2$, with trivial $C_p$-action and the usual $C_2\simeq D_{2p}/C_p$ action.
Thus $E\mathscr{F}^{D_{2p}}=E\mathscr{F}^{C_2}=\emptyset$, $E\mathscr{F}=E\mathscr{F}^{C_p}=E\mathscr{F}/C_p\simeq *$, and
$$E\mathscr{F}^{C_p}/D_{2p}\simeq E\mathscr{F}/C_2\simeq E\mathscr{F}/D_{2p}\simeq BC_2.$$
According to \autoref{orduni}, the homology groups of all these spaces only have $2$-torsion when the degree is positive.
\bigskip
In conclusion, we have the splitting when $\mathscr{F}=\mathscr{F}_1$ and $p$ inverted, or when $\mathscr{F}=\mathscr{F}_2$ and $2$ inverted. $\Box$
\bigskip
Now it remains to compute the $RO(G)$-graded homotopy for
$$(E\mathscr{F}_1)_+\wedge A,\text{ }\widetilde{E\mathscr{F}_1}\wedge A, \text{ }(E\mathscr{F}_2)_+\wedge A\text{ and }\widetilde{E\mathscr{F}_2}\wedge A$$
with $A=H\underline{\mathbb{Z}}$.
\medskip
\section{Computation of $\pi_\bigstar(H\underline{\mathbb{Z}})$}
We will use the same way to define generators as in \textbf{[HHR17]} Definition 3.4:
\begin{definition} \label{generator}
For any actual $G$-representation $V$ with $V^G=0$, let $a_V\in\pi_{-V}(S^0)$ be the map $S^0\rightarrow S^V$ embedding $S^0$ to $0$ and $\infty$. We also use $a_V$ to denote its Hurewicz image in $\pi_{-V}(H\underline{\mathbb{Z}})$.
For any actual orientable representation $V$ of dimension $n$, let $u_V$ be the generator of $\pi_{n-V}(H\underline{\mathbb{Z}})=H_n^G(S^V;\underline{\mathbb{Z}})$ which restricts to the choice of orientation in
$$\underline{H}_n^G(S^V;\underline{\mathbb{Z}})(G/e)\cong H_n(S^n;\mathbb{Z}).$$
\end{definition}
Some important relations on these generators are given below:
\begin{proposition} \label{genrelation}
\textbf{(a)} For any $V_1,V_2$,
$$a_{V_1+V_2}=a_{V_1}a_{V_2},\text{ }u_{V_1+V_2}=u_{V_1}u_{V_2}.$$
\textbf{(b)} Let $G_V$ be the isotropy subgroup of $V$. Then $|G/G_V|a_V=0$.
\textbf{(c)} For $V,W$ both oriented with dimension $2$, with $G_V\subset G_W$, we have
$$a_Wu_V=|G_W/G_V|a_Vu_W.$$
\end{proposition}
\medskip
We consider the cases $G=C_2$ and $G=C_p$ first:
For $G=C_2$, we have the trivial representation and the sign representation $\sigma$.
For $G=C_p$, we have the trivial representation and $p-1$ different $2$-dimensional rotation representations. Again, the periodicity theorem in $[\textbf{KL20}]$ works since any two such rotation representation spheres have isomorphic cellular structures. It suffices to consider only one $2$-dimensional rotation, which we denote as $\lambda$.
\medskip
The computations of $\pi_\bigstar(H\underline{\mathbb{Z}})$ for these two cases can be found in \textbf{[Zen18]} and are listed below:
\begin{theorem} \label{primecase}
For $G=C_2$,
$$\pi_\bigstar(H\underline{\mathbb{Z}})=\mathbb{Z}[u_{2\sigma},a_\sigma]/(2a_\sigma)\oplus \left(\bigoplus_{i>0}2\mathbb{Z}\langle u_{2\sigma}^{-i}\rangle\right)\oplus\left(\bigoplus_{j,k>0}\mathbb{Z}/2\langle\Sigma^{-1}u_{2\sigma}^{-j}a_\sigma^{-k}\rangle\right).$$
For $G=C_p$,
$$\pi_\bigstar(H\underline{\mathbb{Z}})=\mathbb{Z}[u_\lambda,a_\lambda]/(pa_\lambda)\oplus\left(\bigoplus_{i>0}p\mathbb{Z}\langle u_\lambda^{-i}\rangle\right)\oplus\left(\bigoplus_{j,k>0}\mathbb{Z}/p\langle\Sigma^{-1}u_\lambda^{-j}a_\lambda^{-k}\rangle\right).$$
\end{theorem}
\medskip
Now let $G=D_{2p}$.
\subsection{Computation of $(E\mathscr{F}_1)_+\wedge H\underline{\mathbb{Z}}$}
We have $\mathscr{F}=\mathscr{F}_1$ and $p$ inverted in this case.
\medskip
For any virtual representation $V$, since $E\mathscr{F}_+\wedge S^V$ is $C_p$-free and $H\underline{\mathbb{Z}}$ is split, \autoref{ringsplit} and the Adams isomorphism tell us
$$[S^V,E\mathscr{F}_+\wedge H\underline{\mathbb{Z}}]^G\cong [E\mathscr{F}_+\wedge S^V, H\underline{\mathbb{Z}}]^G\cong [(E\mathscr{F}_+\wedge S^V)/C_p,H\underline{\mathbb{Z}}]^{C_2}.$$
We refer to \textbf{[May96}, XVI.2] for an explanation of split equivariant spectra.
Write $V=k\gamma+V_0$, where $V_0$ has no copies of $\gamma$. Consider the following composition:
$$E\mathscr{F}_+\wedge S^V=E\mathscr{F}_+\wedge(S^{sign(k)\gamma})^{\wedge|k|}\wedge S^{V_0}\rightarrow(E\mathscr{F}_+\wedge(S^{sign(k)\gamma})^{\wedge|k|}\wedge S^{V_0})/C_p=(E\mathscr{F}\wedge S^V)/C_p$$
$$\rightarrow (E\mathscr{F}_+/C_p)\wedge (S^{sign(k)\gamma}/C_p)^{\wedge|k|}\wedge(S^{V_0}/C_p).$$
Here $sign(k)=k/|k|$ if $k\neq 0$, and $sign(0)=0$.
Since $E\mathscr{F}_+\simeq (EC_p)_+$ as a $C_p$-space, when $p$ is inverted, \autoref{orduni} gives us a non-equivariant equivalence $E\mathscr{F}_+\simeq (E\mathscr{F})_+/C_p$.
The map $S^{sign(k)\gamma}\rightarrow S^{sign(k)\gamma}/C_p$ induces multiplication by $p^{sign(k)}$ on non-equivariant homology, hence is also a non-equivariant equivalence.
Moreover, since $C_p$ acts trivially on $S^{V_0}$, we have $S^{V_0}=S^{V_0}/C_p$. So the composition above is a non-equivariant equivalence. We know that
$$H\mathbb{Z}_*(E\mathscr{F}_+\wedge S^V)\rightarrow H\mathbb{Z}_*((E\mathscr{F}_+\wedge S^V)/C_p)$$
is an injection. According to \autoref{eqorduni} and \autoref{eqordunimore},
$$E\mathscr{F}_+\wedge S^V\rightarrow(E\mathscr{F}_+\wedge S^V)/C_p$$
is an equivalence of $C_2$-spectra. So we have
$$[S^V,E\mathscr{F}_+\wedge H\underline{\mathbb{Z}}]^G\cong [(E\mathscr{F}_+\wedge S^V)/C_p,H\underline{\mathbb{Z}}]^{C_2}\cong [E\mathscr{F}_+\wedge S^V, H\underline{\mathbb{Z}}]^{C_2}\cong [S^V,H\underline{\mathbb{Z}}]^{C_2}$$
where the last isomorphism is induced by the fact that $E\mathscr{F}\simeq *$ as a $C_2$-space.
To be more precise:
\begin{proposition} \label{z11}
For any integers $k,m,n$,
$$[S^{k+m\sigma+n\gamma},E\mathscr{F}_+\wedge H\underline{\mathbb{Z}}]^G\cong[S^{(k+n)+(m+n)\sigma},H\underline{\mathbb{Z}}]^{C_2}.$$
Moreover, this isomorphism preserves the multiplicative structures.
\end{proposition}
The isomorphism above induces a ring map from the $RO(D_{2p})$-graded ring $\pi_\bigstar^G(E\mathscr{F}_+\wedge H\underline{\mathbb{Z}})$ to the $RO(C_2)$-graded ring $\pi_\bigstar^{C_2}(H\underline{\mathbb{Z}})$. We can compute $\pi_\bigstar^G(E\mathscr{F}_+\wedge H\underline{\mathbb{Z}})$ by tracing the generators in the first equation of \autoref{primecase}. We have:
\medskip
\textbf{(1)}
$$\pi_0^{C_2}(H\underline{\mathbb{Z}})\cong\pi_{c(1+\sigma-\gamma)}^G(E\mathscr{F}_+\wedge H\underline{\mathbb{Z}}),\text{ }\forall c\in\mathbb{Z}.$$
We use $u_{\gamma-\sigma}$ to denote the generator of $\pi_{1+\sigma-\gamma}^G(E\mathscr{F}_+\wedge H\underline{\mathbb{Z}})$. Then $u_{\gamma-\sigma}$ is invertible.
\smallskip
\textbf{(2)}
$$\pi_{k+m\sigma}^{C_2}(H\underline{\mathbb{Z}})\cong\pi_{k+m\sigma+c(1+\sigma-\gamma)}^G(E\mathscr{F}_+\wedge H\underline{\mathbb{Z}}),\text{ }\forall c\in\mathbb{Z}.$$
Thus the pre-image of $u_{2\sigma},a_\sigma\in\pi_\bigstar^{C_2}(H\underline{\mathbb{Z}})$ are $u_{2\sigma}u_{\gamma-\sigma}^c,a_\sigma u_{\gamma-\sigma}^c\in \pi_\bigstar^G(E\mathscr{F}_+\wedge H\underline{\mathbb{Z}})$, for any $c\in\mathbb{Z}$.
\medskip
Therefore, $\pi_\bigstar^G(E\mathscr{F}_+\wedge H\underline{\mathbb{Z}})$ can be obtained by adding the invertible $u_{\gamma-\sigma}$ to $\pi_\bigstar^{C_2}(H\underline{\mathbb{Z}})$:
$$\pi_\bigstar(E\mathscr{F}_+\wedge H\underline{\mathbb{Z}})[\frac{1}{p}]=\mathbb{Z}[\frac{1}{p}][u_{2\sigma},a_\sigma,u_{\gamma-\sigma}^\pm]/(2a_\sigma)$$
$$\oplus\left(\bigoplus_{i>0}2\mathbb{Z}[\frac{1}{p}][u_{\gamma-\sigma}^\pm]\langle u_{2\sigma}^{-i}\rangle\right)\oplus\left(\bigoplus_{j,k>0}\mathbb{Z}/2[u_{\gamma-\sigma}^\pm]\langle\Sigma^{-1}u_{2\sigma}^{-j}a_\sigma^{-k}\rangle\right).$$
\subsection{Computation of $(E\mathscr{F}_2)_+\wedge H\underline{\mathbb{Z}}$}
We have $\mathscr{F}=\mathscr{F}_2$ and $2$ inverted in this case. We use the explicit construction of $E\mathscr{F}$ as $E(D_{2p}/C_p)=EC_2$.
The Adams isomorphism does not work since $C_2$ is not a normal subgroup of $D_{2p}$. Instead, we construct a collapsing spectral sequence.
\medskip
Give $(EC_2)_+$ the standard cellular structure with one $G$-cell $(G/C_p)_+\wedge e_n$ for each $n=0,1,2,...$, with boundary maps
$$e_0\xleftarrow{1-\tau}e_1\xleftarrow{1+\tau}e_2\xleftarrow{1-\tau}...$$
For any virtual representation $V$, give
$$[S^V,E\mathscr{F}_+\wedge H\underline{\mathbb{Z}}]^G_*\cong [E\mathscr{F}_+\wedge S^V,E\mathscr{F}_+\wedge H\underline{\mathbb{Z}}]^G_*\cong [E\mathscr{F}_+\wedge S^V, H\underline{\mathbb{Z}}]^G_*$$
a filtration induced by the skeleton of $E\mathscr{F}$. This filtration implies a cohomological spectral sequence, whose $E_1$-page, together with the boundary map, can be expressed as the following chain complex:
$$[(G/C_p)_+\wedge S^V,H\underline{\mathbb{Z}}]^G_*\xrightarrow{1-\tau}[(G/C_p)_+\wedge S^V,H\underline{\mathbb{Z}}]^G_*\xrightarrow{1+\tau}[(G/C_p)_+\wedge S^V,H\underline{\mathbb{Z}}]^G_*\xrightarrow{1-\tau}....$$
The action of $\tau$ on $[(G/C_p)_+\wedge S^V,H\underline{\mathbb{Z}}]^G_*$ sends each map $f:(G/C_p)_+\wedge S^V\rightarrow H\underline{\mathbb{Z}}$ to the composition
$$(G/C_p)_+\wedge S^V\xrightarrow{\tau\wedge id} (G/C_p)_+\wedge S^V\xrightarrow{f} H\underline{\mathbb{Z}}.$$
Since
$$[(G/C_p)_+\wedge S^V,H\underline{\mathbb{Z}}]^G_*\cong [S^V,H\underline{\mathbb{Z}}]^{C_p}_*,$$
the sequence above becomes
$$[S^V,H\underline{\mathbb{Z}}]^{C_p}_*\xrightarrow{1-\tau}[S^V,H\underline{\mathbb{Z}}]^{C_p}_*\xrightarrow{1+\tau}[S^V,H\underline{\mathbb{Z}}]^{C_p}_*\xrightarrow{1-\tau}....$$
The action of $\tau$ on $[S^V,H\underline{\mathbb{Z}}]^{C_p}_*$ is given by conjugation.
\medskip
We will show that the action of $\tau$ is multiplication by either $1$ or $-1$ on each $[S^V,\Sigma^tH\underline{\mathbb{Z}}]^{C_p}$. First we consider the case that $V$ has no copies of $\sigma$:
\medskip
\begin{proposition} \label{tauactionp}
Let $V=k+n\gamma$ for some integers $k,n$. Then the action of $\tau$ on $[S^V,\Sigma^t H\underline{\mathbb{Z}}]^{C_p}$ is multiplication by $-1$ if $|k-t|\equiv 2$ or $3$ (mod $4$). Otherwise, the action is multiplication by $1$.
\end{proposition}
\paragraph{Proof:} Notice that $[S^V,\Sigma^t H\underline{\mathbb{Z}}]^{C_p}$ can be viewed as the equivariant cohomology or homology of $S^{|n|\gamma}$ with coefficients in $\underline{\mathbb{Z}}$. The computation can be done in a cellular way. We first give $S^{|n|\gamma}$ a $C_p$-CW structure.
\smallskip
We view $S^\gamma$ as $\gamma$ compactified at $\infty$. The action of $\zeta$ is the counter-clockwise rotation by $2\pi/p$ and the action of $\tau$ is the reflection by the $x$-axis. The standard $CW$ structure of $S^\gamma$ can be described as follows:
\medskip
As a based non-equivariant space, $S^\gamma\cong S^2$ can be constructed by one single $2$-cell and the base point at the origin. Denote that $2$-cell by $a$.
\smallskip
As a based $C_p$-space, $S^\gamma\cong S^\lambda$ can be constructed by:
$2$-cells $b_1,b_2,...,b_p$, such that each $b_i$ has image
$$\left\{(r\cos\theta,r\sin\theta):\text{ }0\leq r\leq\infty, \frac{2\pi(i-1)}{p}\leq\theta\leq\frac{2\pi i}{p}\right\};$$
$1$-cells $c_1,c_2,...,c_p$, such that each $c_i$ as image
$$\left\{\left(r\cos\frac{2\pi i}{p},r\sin\frac{2\pi i}{p}\right):\text{ }0\leq r\leq\infty\right\};$$
One $0$-cell $d$ at $\infty$ and the base point at $0$.
Here the subscripts of $b_i$ and $c_i$ are defined modulo $p$.
\medskip
For both $CW$ structures, the action of $\tau$ on $S^\gamma$ can be made into a cellular map:
\smallskip
The action of $\tau$ on the whole $S^\gamma$ is a reflection. Thus $\tau a=-a$.
The action of $\tau$ gives a permutation among the images of $b_1,b_2,...,b_p$ and reverses the orientations of these $2$-cells. Thus $\tau b_i=-b_{p+1-i}$.
The action of $\tau$ gives a permutation among the images of $c_1,c_2,...,c_p$ and keeps the orientations of these $1$-cells. Thus $\tau c_i=c_{p+1-i}$.
Finally, $\tau d=d$ since the point $\infty$ is fixed.
\medskip
Since $S^{|n|\gamma}=S^\gamma\wedge S^\gamma\wedge...\wedge S^\gamma$, we can construct a CW structure as follows:
For $i\in\mathbb{Z}$ and $0\leq j\leq |n|-1$, let
$$e_{2j+2,i}:=(\zeta^ia,...,\zeta^ia,b_i,d,...,d)$$
$$e_{2j+1,i}:=(\zeta^ia,...,\zeta^ia,c_i,d,...,d)$$
with $j$ copies of $\zeta^ia$ and $|n|-1-j$ copies of $d$.
We also have one $0$-cell $e_0:=(d,...,d)$ and the base point.
Notice that $\zeta b_i=b_{i+1}$, $\zeta c_i=c_{i+1}$. This construction makes $S^{|n|\gamma}$ into a $C_p$-CW complex with one $C_p$-cell (except the base point) in each degree between $0$ and $2|n|$. The action of $\tau$ on $S^{|n|\gamma}$ is induced by the action on each copy of $S^\gamma$, hence is also made cellular:
$$\tau e_{2j+2,i}=(-1)^{j+1}e_{2j+2,p+1-i},\text{ }\tau e_{2j+1,i}=(-1)^je_{2j+1,p+1-i},\text{ }\tau e_0=e_0.$$
\medskip
Recall that $H_*^{C_p}(S^{|n|\gamma};\underline{\mathbb{Z}})$ and $H^*_{C_p}(S^{|n|\gamma};\underline{\mathbb{Z}})$ can be computed by the following chain and cochain complexes:
$$C_*^{C_p}(S^{|n|\gamma};\underline{\mathbb{Z}}):=\underline{C}_*(S^{|n|\gamma})\otimes_{\mathcal{O}_{C_p}}\underline{\mathbb{Z}};$$
$$C^*_{C_p}(S^{|n|\gamma};\underline{\mathbb{Z}}):=Hom_{\mathcal{O}_{C_p}}(\underline{C}_*(S^{|n|\gamma});\underline{\mathbb{Z}}),$$
where we view $\underline{\mathbb{Z}}$ as a covariant coefficient system in the first equation and a contravariant one in the second equation.
The actions of $\tau$ on the homology and cohomology are obtained by applying $\tau$ on both $\underline{C}_*(S^{|n|\gamma})$ and $\underline{\mathbb{Z}}$.
\smallskip
Since $S^{|n|\gamma}$ has only $C_p$-free cells in positive degrees and only $C_p$-fixed cells in degree $0$, and $\underline{\mathbb{Z}}$ is the constant coefficient system, we have
$$C_*^{C_p}(S^{|n|\gamma};\underline{\mathbb{Z}}):=\underline{C}_*(S^{|n|\gamma})\otimes_{\mathcal{O}_{C_p}}\underline{\mathbb{Z}}\cong C_*(S^{|n|\gamma})/C_p;$$
$$C^*_{C_p}(S^{|n|\gamma};\underline{\mathbb{Z}}):=Hom_{\mathcal{O}_{C_p}}(\underline{C}_*(S^{|n|\gamma});\underline{\mathbb{Z}})\cong Hom_\mathbb{Z}(C_*(S^{|n|\gamma})/C_p;\mathbb{Z}).$$
Let $e_{2j}$ and $e_{2j-1}$ be the orbits of cells $e_{2j,i}$ and $e_{2j-1,i}$. The induced $\tau$-action on $C_*(S^{|n|\gamma})/C_p$ is expressed as
$$\tau e_{2j}=(-1)^je_{2j},\text{ }\tau e_{2j-1}=(-1)^{j-1}e_{2j-1}.$$
Therefore, the $\tau$-action on the homology and cohomology is multiplication by $-1$ when the degree $\equiv$ $2$ or $3$ (mod $4$). Otherwise the $\tau$-action is multiplication by $1$. $\Box$
\bigskip
Now we consider the case that $V$ also contains copies of $\sigma$. For any $D_{2p}$-spectra $X,Y$, let $\tau$ act on $[X,Y]^{C_p}$ by conjugation. Since $S^\sigma\cong S^1$ as $C_p$-spaces, we have
$$[X,Y]^{C_p}\cong [X,S^{m(\sigma-1)}\wedge Y]^{C_p}.$$
Since $\tau$ acts on $S^\sigma$ as a reflection, the composition
$$[X,Y]^{C_p}\cong [X,S^{m(\sigma-1)}\wedge Y]^{C_p}\xrightarrow{\tau}[X,S^{m(\sigma-1)}\wedge Y]^{C_p}\cong [X,Y]^{C_p}$$
agrees with the $\tau$-action on $[X,Y]^{C_p}$ multiplied by $(-1)^m$.
\medskip
Together with \autoref{tauactionp}, we get
\begin{proposition} \label{tauaction}
Let $V=k+m\sigma+n\gamma$. The action of $\tau$ on $[S^V,\Sigma^tH\underline{\mathbb{Z}}]^{C_p}$ is multiplication by $-1$ if $[|k+m-t|/2]+m$ is odd. Otherwise, the action is multiplication by $1$.
\end{proposition}
\bigskip
Recall that in the spectral sequence computing $[S^V,E\mathscr{F}_+\wedge H\underline{\mathbb{Z}}]^G_*$, the $E_1$-page and the boundary maps are given by
$$[S^V,\Sigma^tH\underline{\mathbb{Z}}]^{C_p}\xrightarrow{1-\tau}[S^V,\Sigma^tH\underline{\mathbb{Z}}]^{C_p}\xrightarrow{1+\tau}[S^V,\Sigma^tH\underline{\mathbb{Z}}]^{C_p}\xrightarrow{1-\tau}...,$$
which can be written as either
$$[S^V,\Sigma^tH\underline{\mathbb{Z}}]^{C_p}\xrightarrow{0}[S^V,\Sigma^tH\underline{\mathbb{Z}}]^{C_p}\xrightarrow{2}[S^V,\Sigma^tH\underline{\mathbb{Z}}]^{C_p}\xrightarrow{0}....$$
or
$$[S^V,\Sigma^tH\underline{\mathbb{Z}}]^{C_p}\xrightarrow{2}[S^V,\Sigma^tH\underline{\mathbb{Z}}]^{C_p}\xrightarrow{0}[S^V,\Sigma^tH\underline{\mathbb{Z}}]^{C_p}\xrightarrow{2}....$$
depending on whether the $\tau$-action becomes the multiplication by $1$ or $-1$.
Since $2$ is inverted, the $E_2$-page vanishes except in the bottom line. Consider the case of $t=0$. We can compute the homotopy of $E\mathscr{F}_+\wedge H\underline{\mathbb{Z}}$ by this $E_2$-page:
\begin{proposition} \label{z12}
When $2$ is inverted, for any integers $k,m,n$,
$$[S^{k+m\sigma+n\gamma},E\mathscr{F}_+\wedge H\underline{\mathbb{Z}}]^G\cong
\begin{cases}
0,\text{ if }[|k+m|/2]+m\text{ is odd,}\\
[S^{(k+m)+n\lambda},H\underline{\mathbb{Z}}]^{C_p},\text{ if }[|k+m|/2]+m\text{ is even.}
\end{cases}$$
Moreover, this isomorphism preserves the multiplicative structures.
\end{proposition}
The isomorphism in the second case is obtained by sending each $G$-map
$$S^V\rightarrow E\mathscr{F}_+\wedge H\underline{\mathbb{Z}}$$
to the underlying $C_p$-map of
$$S^V\rightarrow E\mathscr{F}_+\wedge H\underline{\mathbb{Z}}\rightarrow S\wedge H\underline{\mathbb{Z}}\simeq H\underline{\mathbb{Z}}.$$
Thus the multiplicative structures are preserved.
\begin{remark}
According to the second equation in \autoref{primecase}, $\pi_{a+b\lambda}^{C_p}(H\underline{\mathbb{Z}})$ is non-zero only if $a$ is positive even, or negative odd. So we can rewrite \autoref{z12} as
$$[S^{k+m\sigma+n\gamma},E\mathscr{F}_+\wedge H\underline{\mathbb{Z}}]^G\cong
\begin{cases}
0,\text{ if }[(k+3m+1)/2]\text{ is odd,}\\
[S^{(k+m)+n\lambda},H\underline{\mathbb{Z}}]^{C_p},\text{ if }[(k+3m+1)/2]+m\text{ is even.}
\end{cases}$$
\end{remark}
The equation above induces a ring map from the $RO(D_{2p})$-graded ring $\pi_\bigstar^G(E\mathscr{F}_+\wedge H\underline{\mathbb{Z}})$ to the $RO(C_p)$-graded ring $\pi_\bigstar^{C_p}(H\underline{\mathbb{Z}})$. Similarly, in order to compute $\pi_\bigstar^G(E\mathscr{F}_+\wedge H\underline{\mathbb{Z}})$, we only need to trace the generators in the second equation in \autoref{primecase}:
\medskip
\textbf{(1)}
$$\pi_0^{C_p}(H\underline{\mathbb{Z}})\cong\pi_{2c(1-\sigma)}^G(E\mathscr{F}_+\wedge H\underline{\mathbb{Z}}),\text{ }\forall c\in\mathbb{Z}.$$
Thus $u_{2\sigma}$, as the generator of $\pi_{1-\sigma}^G(E\mathscr{F}_+\wedge H\underline{\mathbb{Z}})$, becomes invertible.
\smallskip
\textbf{(2)}
$$\pi_{2-\lambda}^{C_p}(H\underline{\mathbb{Z}})\cong\pi_{1+\sigma-\gamma+2c(1-\sigma)}^G(E\mathscr{F}_+\wedge H\underline{\mathbb{Z}}),\text{ }\forall c\in\mathbb{Z}.$$
Thus the pre-image of $u_\lambda\in\pi_\bigstar^{C_p}(H\underline{\mathbb{Z}})$ is $u_{\gamma-\sigma}u_{2\sigma}^c$, for any $c\in\mathbb{Z}$.
\smallskip
\textbf{(3)}
$$\pi_{-\lambda}^{C_p}(H\underline{\mathbb{Z}})\cong\pi_{-\gamma+2c(1-\sigma)}^G(E\mathscr{F}_+\wedge H\underline{\mathbb{Z}}),\text{ }\forall c\in\mathbb{Z}.$$
Thus the pre-image of $a_\lambda\in\pi_\bigstar^{C_p}(H\underline{\mathbb{Z}})$ is $a_\gamma u_{2\sigma}^c$, for any $c\in\mathbb{Z}$.
\medskip
Therefore, we have
$$\pi_\bigstar(E\mathscr{F}_+\wedge H\underline{\mathbb{Z}})[\frac{1}{2}]=\mathbb{Z}[\frac{1}{2}][u_{\gamma-\sigma},a_\gamma,u_{2\sigma}^\pm]/(pa_\gamma)$$
$$\oplus\left(\bigoplus_{i>0}p\mathbb{Z}[\frac{1}{2}][u_{2\sigma}^\pm]\langle u_{\gamma-\sigma}^{-i}\rangle\right)\oplus\left(\bigoplus_{j,k>0}\mathbb{Z}/p[u_{2\sigma}^\pm]\langle\Sigma^{-1}u_{\gamma-\sigma}^{-j}a_\gamma^{-k}\rangle\right).$$
\subsection{Computation of $\widetilde{E\mathscr{F}_1}\wedge H\underline{\mathbb{Z}}$ and $\widetilde{E\mathscr{F}_2}\wedge H\underline{\mathbb{Z}}$}
In these two cases, we can show that
$$\widetilde{E\mathscr{F}_1}\wedge H\underline{\mathbb{Z}}\simeq \widetilde{E\mathscr{F}_2}\wedge H\underline{\mathbb{Z}}\simeq *.$$
\medskip
We can compute $\underline{\pi}_*(\widetilde{E\mathscr{F}}\wedge H\underline{\mathbb{Z}})$ by the universal coefficient spectral sequence in \textbf{[LM04]}:
$$\underline{Tor}_{*,*}^{A_G}(\underline{HA_G}_*\widetilde{E\mathscr{F}},\underline{\mathbb{Z}})\Rightarrow\underline{H\underline{\mathbb{Z}}}_*\widetilde{E\mathscr{F}}=\underline{\pi}_*(\widetilde{E\mathscr{F}}\wedge H\underline{\mathbb{Z}}).$$
When either $2$ or $p$ is inverted and the splitting in \autoref{ringsplit} happens, $\underline{HA_G}_*\widetilde{E\mathscr{F}}$ is concentrated in degree $0$, and becomes the complement Green functor of $\underline{HA_G}_*(E\mathscr{F}_+)$ in $A_G$. We use the same notations $M_\mathscr{F}$ and $N_\mathscr{F}$ as in section 2 to denote $\underline{HA_G}_*(E\mathscr{F}_+)$ and $\underline{HA_G}_*\widetilde{E\mathscr{F}}$.
Recall that, as a sub-Green functor of $A_G$, the value of $N_\mathscr{F}$ on $G/L$ is generated by all $\{L/K\}$ with $K\notin\mathscr{F}$.
Since $N_\mathscr{F}$ is a direct summand of $A_G$,
$$\underline{Tor}_{s,t}^{A_G}(N_\mathscr{F},
\underline{\mathbb{Z}})=
\begin{cases}
\underline{\mathbb{Z}}\text{ }\Box\text{ }N_\mathscr{F},\text{ if }s=t=0,\\
0,\text{ otherwise.}
\end{cases}$$
Thus
$$\underline{\pi}_*(\widetilde{E\mathscr{F}}\wedge H\underline{\mathbb{Z}})=\underline{\mathbb{Z}}\text{ }\Box\text{ }N_\mathscr{F},$$
which is concentrated in degree $0$.
The box product can be computed as a coend, for which we refer to \textbf{[LM04]}. In the case when $G$ is finite, we have
$$(M\text{ }\Box\text{ }N)(G/H)=\left(\bigoplus_{K\subset H}M(G/K)\otimes N(G/K)\right)/\sim,$$
where the equivalence relation is generated by
$$a\otimes T_{K_1}^{K_2}(b)\sim R_{K_1}^{K_2}(a)\otimes b, \text{ } T_{K_1}^{K_2}(b)\otimes a\sim b\otimes R_{K_1}^{K_2}(a)$$
for any $K_1\subset K_2$, and
$$M(G/K_1)\otimes N(G/K_1)\cong M(G/K_2)\otimes N(G/K_2)$$
for all conjugacy pairs $K_1,K_2$.
\bigskip
When $\mathscr{F}=\mathscr{F}_1$ and $p$ is inverted, $N_\mathscr{F}(G/G)$ is generated by $\{G/G\}$ and $\{G/C_p\}$, $N_\mathscr{F}(G/C_p)$ is generated by $\{C_p/C_p\}$, and we have
$$N_\mathscr{F}(G/e)=N_\mathscr{F}(G/C_2)=0.$$
Thus
$$(\underline{\mathbb{Z}}\text{ }\Box\text{ }N_\mathscr{F})(G/e)=(\underline{\mathbb{Z}}\text{ }\Box\text{ }N_\mathscr{F})(G/C_2)=0.$$
For any $a\in\underline{\mathbb{Z}}(G/e)$ and $b\in N_\mathscr{F}(G/C_p)$, we have
$$Tr_e^{C_p}(a)\otimes b\sim a\otimes R_e^{C_p}(b)=a\otimes 0=0.$$
However, $Tr_e^{C_p}$ in $\underline{\mathbb{Z}}$ is the multiplication by $p$, which is invertible. Thus this equivalence relation identifies the whole $\underline{\mathbb{Z}}(G/C_p)\otimes N_\mathscr{F}(G/C_p)$ to zero.
By a similar argument, the equivalence relation between $G/G$ and $G/C_2$ identifies the whole $\underline{\mathbb{Z}}(G/G)\otimes N_\mathscr{F}(G/G)$ to zero. Thus we have $\underline{\mathbb{Z}}\text{ }\Box\text{ }N_\mathscr{F}=0$.
\medskip
When $\mathscr{F}=\mathscr{F}_2$ and $2$ is inverted, we just need to exchange $2$ and $p$ in the above argument. So we still have $\underline{\mathbb{Z}}\text{ }\Box\text{ }N_\mathscr{F}=0$.
\medskip
Therefore,
$$\widetilde{E\mathscr{F}_1}\wedge H\underline{\mathbb{Z}}\simeq \widetilde{E\mathscr{F}_2}\wedge H\underline{\mathbb{Z}}\simeq *.$$
\subsection{Final result}
We already computed $\pi_\bigstar(H\underline{\mathbb{Z}})$ with either $2$ or $p$ inverted. Now we just need to glue them together:
\begin{theorem} \label{hz}
For $G=D_{2p}$,
$$\pi_\bigstar(H\underline{\mathbb{Z}})=\mathbb{Z}[u_{\gamma-\sigma},u_{2\sigma},a_\sigma,a_\gamma]/(2a_\sigma,pa_\gamma)$$
$$\oplus\left(\bigoplus_{i>0}2\mathbb{Z}[u_{\gamma-\sigma}]\langle u_{2\sigma}^{-i}\rangle\right)\oplus\left(\bigoplus_{i>0}p\mathbb{Z}[u_{2\sigma}]\langle u_{\gamma-\sigma}^{-i}\rangle\right)\oplus\left(\bigoplus_{j,k>0}2p\mathbb{Z}\langle u_{2\sigma}^{-j}u_{\gamma-\sigma}^{-k}\rangle\right)$$
$$\oplus\left(\bigoplus_{j,k>0}\mathbb{Z}/p[u_{\gamma-\sigma}]\langle a_\gamma^j u_{2\sigma}^{-k}\rangle\right)\oplus\left(\bigoplus_{j,k>0}\mathbb{Z}/2[u_{2\sigma}]\langle a_\sigma^j u_{\gamma-\sigma}^{-k}\rangle\right)$$
$$\oplus\left(\bigoplus_{j,k>0}\mathbb{Z}/2[u_{\gamma-\sigma}^\pm]\langle\Sigma^{-1}u_{2\sigma}^{-j}a_\sigma^{-k}\rangle\right)\oplus\left(\bigoplus_{j,k>0}\mathbb{Z}/p[u_{2\sigma}^\pm]\langle\Sigma^{-1}u_{\gamma-\sigma}^{-j}a_\gamma^{-k}\rangle\right).$$
\end{theorem}
\section{Applications to general $D_{2p}$-spectra}
Let $G=D_{2p}$. In this section, we will generalize the computations in section 5 to other $G$-spectra. For any $G$-spectra $A$, \autoref{twosplit} shows that the information of $\pi_\bigstar(A)$ is covered by the homotopy of
$$(E\mathscr{F}_1)_+\wedge A,\text{ }\widetilde{E\mathscr{F}_1}\wedge A, \text{ }(E\mathscr{F}_2)_+\wedge A\text{ and }\widetilde{E\mathscr{F}_2}\wedge A$$
with either $2$ or $p$ inverted.
The main results of this section are given below:
\begin{theorem} \label{generald2p}
Let $k,m,n$ be arbitrary integers.
\textbf{(a)} Let $A$ be a split $G$-spectra. When $p$ is inverted, we have
$$[S^{k+m\sigma+n\gamma},(E\mathscr{F}_1)_+\wedge A]^G\cong[S^{(k+n)+(m+n)\sigma},A^{C_p}]^{C_2}.$$
\textbf{(b)} Let $A=HM$ for some Mackey functor $M$ such that $M(G/H)$ has a trivial $W_GH$-action for all $H\subset G$. When $2$ is inverted, we have
$$[S^{k+m\sigma+n\gamma},(E\mathscr{F}_2)_+\wedge A]^G\cong
\begin{cases}
0,\text{ if }[|k+m|/2]+m\text{ is odd,}\\
[S^{(k+m)+n\lambda},A]^{C_p},\text{ if }[|k+m|/2]+m\text{ is even.}
\end{cases}$$
\textbf{(c)} When $p$ is inverted, we have
$$[S^{k+m\sigma+n\gamma},\widetilde{E\mathscr{F}}_1\wedge A]^G\cong [S^{k+m\sigma},\widetilde{E\mathscr{F}}_1\wedge A]^G\cong [S^{k+m\sigma},\Phi^{C_p}A]^{C_2}.$$
\textbf{(d)} When $2$ is inverted, we have
$$[S^{k+m\sigma+n\gamma},\widetilde{E\mathscr{F}}_2\wedge A]^G\cong [S^{k+n},\widetilde{E\mathscr{F}}_2\wedge A]^G\cong
[S^{k+n},\Phi^{C_2}A]\oplus[S^{k+n},\Phi^GA].$$
Moreover, all isomorphisms above preserve the symmetric monoidal structure.
\end{theorem}
\begin{corollary} \label{hagrog}
In particular, when $A=HA_G$, we have
$$[S^{k+m\sigma+n\gamma},HA_G]^G\cong [S^{(k+n)+(m+n)\sigma},HA_{C_2}^2]^{C_2}\oplus [S^{k+m\sigma},HA_{C_2}]^{C_2}$$
when $p$ is inverted, and
$$[S^{k+m\sigma+n\gamma},HA_G]^G\cong [S^{k+n},H\mathbb{Z}^2]\oplus
\begin{cases}
0,\text{ if }[|k+m|/2]+m\text{ is odd,}\\
[S^{(k+m)+n\lambda},HA_{C_p}]^{C_p},\text{ if }[|k+m|/2]+m\text{ is even.}
\end{cases}$$
when $2$ is inverted.
\end{corollary}
Parts (a) and (b) in \autoref{generald2p} can be proved in the same way as \autoref{z11} and \autoref{z12}.
\medskip
In order to compute the homotopy of $\widetilde{E\mathscr{F}}\wedge A$, for either $\mathscr{F}=\mathscr{F}_1$ or $\mathscr{F}_2$, we need to consider the equivariant homology with coefficients in $N_\mathscr{F}$. We can use the same proof as in \autoref{burnsidehom} to show:
\begin{proposition} \label{nhom}
Let $X$ be a based $G$-CW complex or a $G$-CW spectrum. Then
$$(HN_\mathscr{F})_*X\cong\bigoplus_{H\notin\mathscr{F}}H\mathbb{Z}_*(\Phi^HX/W_GH)$$
where we only choose one $H$ from each conjugacy class of subgroups in $G$.
As a Mackey functor,
$$\underline{HN_\mathscr{F}}_*X(G/L)\cong\bigoplus_{K\notin\mathscr{F}}H\mathbb{Z}_*(\Phi^KX/W_LK)$$
where we only choose one $K$ from each conjugacy class of subgroups in $L$.
\end{proposition}
\bigskip
We will prove:
\begin{proposition} \label{efequiv}
\textbf{(a)} When $\mathscr{F}=\mathscr{F}_1$ and $p$ is inverted, $S^\gamma$ is $\widetilde{E\mathscr{F}}$-equivalent to $S^0$.
\textbf{(b)} When $\mathscr{F}=\mathscr{F}_2$ and $2$ is inverted, $S^\gamma, S^\sigma$ are $\widetilde{E\mathscr{F}}$-equivalent to $S^1,S^0$, respectively.
\end{proposition}
\paragraph{Proof:} Notice that
$$HN_\mathscr{F}=HA_G\wedge\widetilde{E\mathscr{F}}.$$
Since all bounded below $G$-spectra are $HA_G$-local, it suffices to prove \autoref{efequiv} with $\widetilde{E\mathscr{F}}$ replaced by $HN_\mathscr{F}$.
\medskip
\textbf{Part (a):} Assume that $\mathscr{F}=\mathscr{F}_1$ and $p$ is inverted.
\autoref{nhom} shows that:
$$\underline{HN_\mathscr{F}}_*X(G/e)=\underline{HN_\mathscr{F}}_*X(G/C_2)=0,$$
$$\underline{HN_\mathscr{F}}_*X(G/C_p)=H\mathbb{Z}_*(\Phi^{C_p}X),$$
$$\underline{HN_\mathscr{F}}_*X(G/G)=H\mathbb{Z}_*(\Phi^GX)\oplus H\mathbb{Z}_*(\Phi^{C_p}X/G).$$
Since these homology groups only depend on the $C_p$-fixed cells of $X$, $S^0\hookrightarrow S^\gamma$ is an $HN_\mathscr{F}$-equivalence.
\bigskip
\textbf{Part (b):} Assume that $\mathscr{F}=\mathscr{F}_2$ and $2$ is inverted.
\autoref{nhom} shows that:
$$\underline{HN_\mathscr{F}}_*X(G/e)=\underline{HN_\mathscr{F}}_*X(G/C_p)=0,$$
$$\underline{HN_\mathscr{F}}_*X(G/C_2)=H\mathbb{Z}_*(\Phi^{C_2}X),$$
$$\underline{HN_\mathscr{F}}_*X(G/G)=H\mathbb{Z}_*(\Phi^GX)\oplus H\mathbb{Z}_*(\Phi^{C_2}X/G).$$
for any choice of $C_2\subset G$.
These homology groups only depend on the $C_2$-fixed cells of $X$. So the following maps among $D_{2p}$-spaces become $HN_\mathscr{F}$-equivalences:
$$S^\gamma\rightarrow S^\gamma/C_p=S^{1+\sigma}\hookleftarrow S^1,$$
$$S^0\hookrightarrow S^\sigma.$$
$\Box$
\bigskip
\paragraph{Proof of \autoref{generald2p}(c)(d):} Part (c) follows after the fact that $S^{k+m\sigma+n\gamma}$ is $\widetilde{E\mathscr{F}_1}$-equivalent to $S^{k+m\sigma}$ according to \autoref{efequiv}.
\medskip
For part (d), \autoref{efequiv} shows that $S^{k+m\sigma+n\gamma}$ is $\widetilde{E\mathscr{F}_2}$-equivalent to $S^{k+n}$. Thus
$$[S^{k+m\sigma+n\gamma},\widetilde{E\mathscr{F}}_2\wedge A]^G\cong [S^{k+n},\widetilde{E\mathscr{F}}_2\wedge A]^G.$$
Let $\mathscr{F}^\prime$ be the family of all proper subgroups of $G$. Notice that $\widetilde{E\mathscr{F}_2}$ and
$$\widetilde{E\mathscr{F}^\prime}\vee(G/C_2)_+\wedge\widetilde{E(G/C_p)}$$
have the same homotopy types on their fixed point subspaces. \autoref{elmendorf} provides a weak equivalence between these two spaces. Together with \autoref{ringsplit}, we have
$$[S^{k+n},\widetilde{E\mathscr{F}}_2\wedge A]^G\cong[\widetilde{E\mathscr{F}}_2\wedge S^{k+n},A]^G$$
$$\cong[\widetilde{E\mathscr{F}^\prime}\wedge S^{k+n},A]^G\oplus[(G/C_2)_+\wedge\widetilde{E(G/C_p)}\wedge S^{k+n},A]^G$$
$$\cong[\widetilde{E\mathscr{F}^\prime}\wedge S^{k+n},A]^G\oplus[\widetilde{E(G/C_p)}\wedge S^{k+n},A]^{C_2}$$
$$\cong [S^{k+n},\widetilde{E\mathscr{F}^\prime}\wedge A]^G\oplus[S^{k+n},\widetilde{E(G/C_p)}\wedge A]^{C_2}\cong[S^{k+n},\Phi^GA]\oplus[S^{k+n},\Phi^{C_2}A].$$
$\Box$
\section{Splittings for more general finite groups}
In this section, we will generalize \autoref{twosplit} to more general finite groups $G$.
\medskip
Let $G=G_2\ltimes G_1$. In other words, $G_1\lhd G$, $G_2\subset G$, and $G_2\cong G/G_1$.
Let $\mathscr{F}_1$ be the family which contains all subgroups which are sub-conjugate to $G_2$, $\mathscr{F}_2$ be the family which contains all subgroups of $G_1$.
\begin{proposition} \label{gen11}
Assume that $|G_1|,|G_2|$ are relatively prime.
When $|G_1|$ is inverted, $\underline{H}_*^G(E\mathscr{F}_1;A_G)$ is concentrated in degree $0$.
\end{proposition}
\begin{proposition} \label{gen21}
When $|G_2|$ is inverted, $\underline{H}_*^G(E\mathscr{F}_2;A_G)$ is concentrated in degree $0$.
\end{proposition}
\begin{proposition} \label{gen12}
When $\mathscr{F}=\mathscr{F}_1$, the denominators of all $c_H$ only contain prime factors dividing $|G_1|$.
\end{proposition}
\begin{proposition} \label{gen22}
When $\mathscr{F}=\mathscr{F}_2$, the denominators of all $c_H$ only contain prime factors dividing $|G_2|$.
\end{proposition}
\begin{remark}
We do not require $|G_1|,|G_2|$ to be relatively prime for the last three propositions. We guess this requirement can also be removed for \autoref{gen11}, but do not have a proof at current time.
\end{remark}
According to \autoref{invertprime}, we can generalize \autoref{twosplit}:
\begin{theorem} \label{bigtwosplit}
Assume that $|G_1|,|G_2|$ are relatively prime. The splitting in \autoref{ringsplit} happens when
\textbf{(a)} $\mathscr{F}=\mathscr{F}_1$ and $|G_1|$ is inverted, or
\textbf{(b)} $\mathscr{F}=\mathscr{F}_2$ and $|G_2|$ is inverted.
\end{theorem}
\begin{remark}
When either $|G_1|$ or $|G_2|$ is inverted, it is possible to use the same techniques as in Sections 5 and 6 to decompose computations about $G$-spectra into computations about $G_1$ or $G_2$-spectra. Then we can glue back these two splittings and recover information of $G$-spectra.
\end{remark}
\medskip
We will prove Proposition \ref{gen11} to \ref{gen22} in the rest of the paper from the easiest to the hardest:
$$\ref{gen21}\rightarrow\ref{gen22}\rightarrow\ref{gen11}\rightarrow\ref{gen12}.$$
The proof of \autoref{gen12} will be given in a separate section according to its length. The cases of $\mathscr{F}_1$ and $\mathscr{F}_2$ are quite different since only $G_1$ is a normal subgroup of $G$.
\medskip
\paragraph{Proof of \autoref{gen21}:} Assume that $\mathscr{F}=\mathscr{F}_2$ and $|G_2|$ is inverted.
Recall that $E\mathscr{F}$ is characterized by its fixed point subspaces as $(E\mathscr{F})^H\simeq *$ if $H\subset G_1$, and $(E\mathscr{F})^H=\emptyset$ otherwise. Thus we can construct $E\mathscr{F}$ explicitly as $ E(G/G_1)$.
According to \autoref{burnsidehom}, $\underline{H}_*(E\mathscr{F};A_G)$ is determined by the homology of different orbit spaces of fixed point subspaces. However, the fixed point subspaces of $E(G/G_1)$ are either empty or the whole space. Thus the orbit spaces only have the form $E(G/G_1)/H\simeq BH$ for some $H\subset G/G_1$. According to \autoref{orduni}, we know that the positive-degree homology groups of these orbit spaces vanish when $|G_2|=|G/G_1|$ is inverted. $\Box$
\bigskip
\paragraph{Proof of \autoref{gen22}:} Let $\mathscr{F}=\mathscr{F}_2$. The coefficients $c_H$ are computed by
$$\sum_H c_Hs_{(K,H)}=1, \text{ }\forall K\in\mathscr{F}$$
with one $H$ chosen from each conjugacy class in $\mathscr{F}$. We can check that $c_{G_1}=|G_2|^{-1}$ and $c_H=0$ for all other $H$. Recall that $s_{(K,H)}$ is the product of $|W_GH|$ with the number of subgroups conjugate to $H$ and containing $K$. Thus for any $K\subset G_1$, $s_{(K,G_1)}=|G_2|$ since $G_1$ is normal and $|W_GG_1|=|G/G_1|=|G_2|$. We have
$$\sum_H c_Hs_{(K,H)}=c_{G_1}s_{(K,G_1)}+\sum_{G_1\neq H\in\mathscr{F}}c_Hs_{(K,H)}=|G_2|^{-1}\cdot|G_2|+0=1.$$
\autoref{gen22} is true since the only nontrivial denominator of any $c_H$ is $|G_2|$. $\Box$
\bigskip
For the other two propositions, we will need the following definition:
\begin{definition} \label{commsub}
For any $K\subset G_2$, define $P_K\subset G_1$ to be
$$P_K:=\{g\in G_1:\text{ }gh=hg\textbf{ },\forall h\in K\}.$$
\end{definition}
\begin{lemma} \label{completecomm}
For any $K\subset G_2$ and $g\in G_1$, $gKg^{-1}\subset G_2$ if and only if $g\in P_K$.
\end{lemma}
\paragraph{Proof:} For any $h\in K$, if $ghg^{-1}\in G_2$, then $ghg^{-1}h^{-1}\in G_2h^{-1}=G_2$. However, $g(hg^{-1}h^{-1})\in G_1$ since $G_1\lhd G$. Since $G_1\cap G_2=\{1\}$, we must have $ghg^{-1}h^{-1}=1$ and hence $gh=hg$, $g\in P_K$. $\Box$
\begin{lemma} \label{nnaction}
The subgroup $P_K$ is closed under the action of $N_{G_2}K$ by conjugation.
\end{lemma}
\paragraph{Proof:} For any $g\in P_K$, $h\in N_{G_2}K$, and $k\in K$, we have $h^{-1}kh\in K$. Thus
$$(h^{-1}gh)k=h^{-1}(g(hkh^{-1}))h=h^{-1}((hkh^{-1})g)h=k(h^{-1}gh)$$
and hence $h^{-1}gh\in P_K$. $\Box$
\bigskip
In order to prove \autoref{gen11}, we need a more general version of \autoref{eqorduni}:
\begin{lemma} \label{genuniorb}
Let $X$ be a $G$-CW complex whose cells only have isotropy groups sub-conjugate to $G_2$. Consider the orbit map $X\rightarrow X/G_1$, which is viewed as a map between $G_2$-spaces. Then for any $H\subset G_2$, the induced map
$$H_*(X^H;\mathbb{Z})\rightarrow H_*((X/G_1)^H;\mathbb{Z})$$
becomes a surjection when $|G_1|$ is inverted.
Moreover, if all such maps are also injective, then $X\rightarrow(X/G_1)$ is a $G_2$-equivalence.
\end{lemma}
\paragraph{Proof:} First, for any non-equivariant cell $e$ with isotropy group $aKa^{-1}$ for some $K\subset G_2$ and $a\in G$, write $a=gh$ with $g\in G_1$ and $h\in G_2$. Then $aKa^{-1}=g(hKh^{-1})g^{-1}$ and hence $hKh^{-1}$ is the isotropy group of $g^{-1}e$. Therefore, each $G_1$-cell in $X$ contains at least one non-equivariant cell with isotropy group inside $G_2$.
\medskip
Consider any cell $e$ in $X$ with isotropy group $K\subset G_2$. Fix $H\subset G_2$. If the image of $e$ in $X/G_1$ is $H$-fixed, then the action of $H$ is closed inside $\{ge:g\in G_1\}$.
This means that for any $h\in H$, $g_1\in G_1$, there exists $g_2\in G_1$ such that $hg_1e=g_2e$. So $g_2^{-1}hg_1\in K$, and hence $g_2^{-1}(hg_1h^{-1})\in Kh^{-1}\subset G_2$. Since $hg_1h^{-1}$ is also an element in $G_1$,
$$g_2^{-1}(hg_1h^{-1})\in G_1\cap G_2=\{1\}.$$
So $g_2=hg_1h^{-1}$. Then $g_2^{-1}hg_1\in K$ becomes $h\in K$. So the image of $e$ in $X/G_1$ is $H$-fixed if and only if $H\subset K$.
Now we assume that $H\subset K$. For any $g\in G_1$, $ge$ is $H$-fixed if and only if $hge=ge$ for all $h\in H$. Thus $g^{-1}hge=e$ and hence $g^{-1}hg\in K$, $g^{-1}Hg\subset K\subset G_2$. \autoref{completecomm} shows that $g\in P_H$. In this case, $g^{-1}Hg=H\subset K$ is always satisfied. So we have
$$\{ge:g\in G_1\}^H=\{ge:g\in P_H\}.$$
\bigskip
Now we can play the same trick as in \autoref{eqorduni}. Define a map between cellular chain complexes
$$C_*((X/G_1)^H;\mathbb{Z})\rightarrow C_*(X^H;\mathbb{Z})$$
as follows: For any $G_1$-cell $G/K_+\wedge e$ of $X$ with $K\supset H$, define the map above by $$(G/K_+\wedge e)/G_1\mapsto\sum_{g\in P_H}ge\in C_*(X^H;\mathbb{Z}).$$
Then the composition
$$C_*((X/G_1)^H;\mathbb{Z})\rightarrow C_*(X^H;\mathbb{Z})\rightarrow C_*((X/G_1)^H;\mathbb{Z})$$
becomes the multiplication by $|P_H|$, which can be passed to homology. Thus when $|G_1|$ is inverted, since $P_H\subset G_1$, $|P_H|$ is also inverted. The above composition becomes an isomorphism, and
$$H_*(X^H;\mathbb{Z})\rightarrow H_*((X/G_1)^H;\mathbb{Z})$$
becomes a surjection. $\Box$
\begin{remark}
The above argument still works when we replace $G_1, G_2$ by any subgroups $P\subset G_1$, $K\subset G_2$. Thus if $X$ is a $(K\ltimes P)$-CW complex with certain conditions, $X\rightarrow X/P$ will also be a $K$-equivalence when $|P|$ is inverted.
\end{remark}
We also need the powerful Schur–Zassenhaus theorem:
\begin{theorem} \label{schur}
(\textbf{[Iss08]} 3B) Let $K$ be a finite group and $K_1$ be a normal subgroup of $K$ such that $|K_1|$ and $|K/K_1|$ are relative prime. Then there exists $K_2\subset K$, such that $K=K_2\ltimes K_1$. Moreover, all such choices of $K_2$ are conjugate to each other.
\end{theorem}
\paragraph{Proof of \autoref{gen11}:} Assume that $|G_1|$ is inverted and $\mathscr{F}=\mathscr{F}_1$.
According to \autoref{burnsidehom}, it suffices to show that $H_*(E\mathscr{F}^H/L;\mathbb{Z})$ is concentrated in degree $0$, for any $H\lhd L\subset G$. Notice that $E\mathscr{F}^H=\emptyset$ if $H\notin\mathscr{F}$. Moreover, for any $g\in G$, $E\mathscr{F}^H/L$ is homotopy equivalent to $E\mathscr{F}^{g^{-1}Hg}/g^{-1}Lg$. Since $\mathscr{F}$ is the family of all subgroups sub-conjugate to $G_2$, we can assume $H\subset G_2$.
List all elements in $L$ as $L=\{a_1b_1,a_2b_2,...,a_nb_n\}$ with each $a_i\in G_1$ and $b_i\in G_2$. Let $\overline{L}$ be the subset of $G_2$ consisting of all $b_i$. Then $\overline{L}$ is closed under multiplication and inverses since
$$(a_ib_i)(a_jb_j)=(a_i(b_ia_jb_i^{-1}))(b_ib_j),$$
$$(a_ib_i)^{-1}=b_i^{-1}a_i^{-1}=(b_i^{-1}a_i^{-1}b_i)b_i^{-1}.$$
Thus $\overline{L}$ is a subgroup of $G_2$. The equations above also show that the map $L\rightarrow \overline{L}$ sending each $a_ib_i$ to $b_i$ is a surjective group homomorphism. Its kernel $L_0$ is the collection of all $a_ib_i$ with $b_i=1$. Thus $L_0$ is a subgroup of $G_1$ and a normal subgroup of $L$. Moreover, for any $i$, since $H\lhd L$, we have $b_i^{-1}a_i^{-1}Ha_ib_i=H$ and hence $a_i^{-1}Ha_i=b_iHb_i^{-1}\subset G_2$. \autoref{completecomm} tells us $a_i\in P_H$ and hence $b_i\in W_{G_2}H$.
Since $W_{G_2}H$ acts on $P_H$ by conjugation according to \autoref{nnaction}, the group generated by $P_H$ and $\overline{L}\subset W_{G_2}H$ is the semi-direct product $\overline{L}\ltimes P_H$.
\smallskip
Notice that $L_0\lhd L$ and $|L_0|$ are relatively prime to $|L/L_0|=|\overline{L}|$ (since $L_0,\overline{L}$ are subgroups of $G_1,G_2$). Apply \autoref{schur} on $L$ and $L_0$. We have $L=L^\prime\ltimes L_0$ for some $L^\prime$ isomorphic to $\overline{L}$. Notice that $L^\prime\subset L\subset\overline{L}\ltimes P_H$. Since $|L^\prime|=|\overline{L}|$ is relatively prime to $P_H$, the elements in $L^\prime$ are in different $P_H$-cosets. Thus $\overline{L}\ltimes P_H=L^\prime\ltimes P_H$. Apply \autoref{schur} again on $\overline{L}\ltimes P_H$, we know that $\overline{L}$ is conjugate to $L^\prime$. So there exists $g\in P_H$ such that $g^{-1}L^\prime g=\overline{L}$.
\smallskip
Now we have $g^{-1}Lg=\overline{L}\ltimes g^{-1}L_0g$ and
$$E\mathscr{F}^H/L\cong E\mathscr{F}^{g^{-1}Hg}/g^{-1}Lg=E\mathscr{F}^H/(\overline{L}\ltimes g^{-1}L_0g).$$
So we can assume $L=\overline{L}\ltimes L_0$ at the beginning. Otherwise we replace it by $\overline{L}\ltimes g^{-1}L_0g$.
\medskip
Recall that all points in $E\mathscr{F}$ have isotropy groups sub-conjugate to $G_2$. Thus when we view $X=E\mathscr{F}^H$ as an $L$-CW complex, all points have isotropy groups sub-conjugate to $\overline{L}$. So all cells are $L_0$-free. Apply \autoref{genuniorb} with $G=G_2\ltimes G_1$ replaced by $L=\overline{L}\ltimes L_0$. The maps between homology groups in \autoref{genuniorb} are always injective since all fixed point subspaces of $E\mathscr{F}^H$ are either contractible or empty. Thus the projection $E\mathscr{F}^H\rightarrow E\mathscr{F}^H/L_0$ is an $\overline{L}$-equivalence. So we have
$$E\mathscr{F}^H/L=(E\mathscr{F}^H/L_0)/\overline{L}\simeq E\mathscr{F}^H/\overline{L}$$
On the other hand, we have a $G_2$-equivalence $E\mathscr{F}\simeq *$ by checking all fixed point subspaces. Since both $H,\overline{L}$ are subgroups of $G_2$, $E\mathscr{F}^H/\overline{L}\simeq *$. Therefore, the homology of $E\mathscr{F}^H/L$ is concentrated in degree 0. $\Box$
\bigskip
\section{Proof of \autoref{gen12}}
Assume that $|G_1|$ is inverted and $\mathscr{F}=\mathscr{F}_1$ in this section.
Recall that $c_H$ is defined for each $H\in\mathscr{F}$ inductively: For any $K\in\mathscr{F}$, we have
$$\sum_H c_Hs_{(K,H)}=1$$
where we choose one $H$ from each conjugacy class inside $\mathscr{F}$, and $s_{(K,H)}=|(G/H)^K|$. We want to show that the denominators of $c_H$ for all $H\subset G$ are invertible.
Since $\mathscr{F}$ is the family of subgroups sub-conjugate to $G_2$, and $c_H=c_{H^\prime}$ when $H$ is conjugate to $H^\prime$, we can consider $K,H$ in the equation above as subgroups of $G_2$.
For any $abH\in G/H$ with $a\in G_1$, $b\in G_2$, if it is $K$-fixed, then $KabH=abH$ and hence
$$(a^{-1}Ka)(bHb^{-1})=(bHb^{-1})\Rightarrow a^{-1}Ka\subset bHb^{-1}\subset G_2.$$
\autoref{completecomm} shows that $a\in P_K$. So $KabH=abH$ becomes $KbH=bH$ and hence $bH\in(G_2/H)^K$. So we have
$$s_{(K,H)}=|P_K|\cdot|(G_2/H)^K|$$
where $|(G_2/H)^K|$ can be expressed as the product between $|W_{G_2}H|$ and the number of subgroups of $G_2$ which are conjugate to $H$ inside $G_2$ and contain $K$.
A direct consequence of \autoref{completecomm} shows that if $H_1,H_2\subset G_2$, then they are conjugate as subgroups of $G$ if and only if they are conjugate as subgroups of $G_2$. So for each $K\subset G_2$, we have
$$1=\sum_{\text{one }H\text{ from each conjugacy class in }G}c_Hs_{(K,H)}=\sum_{\text{one }H\text{ from each conjugacy class in }G_2}c_Hs_{(K,H)}$$
$$=\sum_{\text{one }H\text{ from each conjugacy class in }G_2}c_H\cdot|P_K|\cdot |W_{G_2}H|\cdot\#(\text{subgroups of }G_2\text{ conjugate to }H\text{ containing }K)$$
$$=\sum_{K\subset H\subset G_2}c_H\cdot|P_K|\cdot|W_{G_2}H|.$$
In the last sum, we choose all $H$ with $K\subset H\subset G_2$, not up to conjugacy. We highlight this new relation:
\begin{proposition} \label{chrelation}
For any $K\subset G_2$, we have
$$\sum_{K\subset H\subset G_2}c_H\cdot|P_K|\cdot|W_{G_2}H|=1.$$
Moreover, when $H,H^\prime$ are conjugate subgroups of $G_2$, we have $c_H=c_{H^\prime}$.
\end{proposition}
Before we start computing the value of each $c_H$, we need to study relations between subgroups of $G_1$ and subgroups of $G_2$.
\paragraph{Notation:} We will use letters $K,H,L$ to denote subgroups of $G_2$, and $P,Q$ to denote subgroups of $G_1$ in the rest of this section.
\medskip
First we extend \autoref{commsub}:
\begin{definition} \label{decent}
For any $K\subset G_2$, let
$$P_K:=\{g\in G_1:\text{ }gh=hg,\forall h\in K\}.$$
For any $P\subset G_1$, let
$$K_P:=\{h\in G_2:\text{ }gh=hg,\forall g\in P\}.$$
For any subgroup $K$ of $G_2$, it is called \textbf{G$_1$-good} if $K=K_P$ for some $P\subset G_1$.
For any subgroup $P$ of $G_1$, it is called \textbf{G$_2$-good} if $P=P_K$ for some $K\subset G_2$.
\end{definition}
Some of their properties are listed below:
\begin{proposition} \label{decentprop}
We have:
\textbf{(a)} $P\subset P_{K_P}$, with equality when $P$ is $G_2$-good.
\textbf{(b)} $K\subset K_{P_K}$, with equality when $K$ is $G_1$-good.
\textbf{(c)} For any $K_1,K_2\subset G_2$, $K_1\subset K_2$ implies $P_{K_2}\subset P_{K_1}$. The converse is true if $K_1,K_2$ are $G_1$-good.
\textbf{(d)} For any $P_1,P_2\subset G_1$, $P_1\subset P_2$ implies $K_{P_2}\subset K_{P_1}$. The converse is true if $P_1,P_2$ are $G_2$-good.
\textbf{(e)} For any $b\in G_2$, $K\subset G_2$, $P_{b^{-1}Kb}=b^{-1}P_Kb$. Thus if $K$ is $G_1$-good, so is $b^{-1}Kb$.
\textbf{(f)} If both $P_1,P_2\subset G_1$ are $G_2$-good, so is $P_1\cap P_2$.
\end{proposition}
\paragraph{Proof:} Parts (a) to (e) follow immediately after \autoref{decent}.
For part (f), let $K_1=K_{P_1}$ and $K_2=K_{P_2}$. Parts (a) and (b) show that $P_1=P_{K_1}$ and $P_2=P_{K_2}$.
Let $K$ be the smallest subgroup of $G_2$ which contains both $K_1$ and $K_2$. In other words, it is generated by elements in $K_1$ and $K_2$. We compare $P_K$ and $P_1\cap P_2$:
For any $g\in P_1\cap P_2$, we have $gh=hg$ for all $h\in K_1$ and all $h\in K_2$. Since $K$ is generated by $K_1,K_2$, $gh=hg$ for all $h\in K$. Thus $P_1\cap P_2\subset P_K$.
On the other hand, for any $g\in P_K$, since $K_1,K_2\subset K$, we have $gh=hg$ for all $h\in K_1$ and all $h\in K_2$. Thus $g\in P_1$ and $g\in P_2$. So we have $P_K\subset P_1\cap P_2$. Therefore, $P_1\cap P_2=P_K$ is also $G_2$-good. $\Box$
\bigskip
It suffices to consider $G_1$ and $G_2$-good subgroups since:
\begin{proposition} \label{onlydecent}
Let $\mathbb{D}$ be the collection of all $G_1$-good subgroups of $G_2$. Then $c_H=0$ for any $H\notin\mathbb{D}$.
\end{proposition}
\paragraph{Proof:} We use induction. Clearly $G_2$ itself is $G_1$-good since $G_2=K_{\{1\}}$.
Fix $K\subset G_2$ which is not $G_1$-good. Assume that the proposition is true for all $H\subset G_2$ and $H\supsetneqq K$. Let $K_0=K_{P_K}$. Then $K\subsetneqq K_0$ and $|P_K|=|P_{K_0}|$. \autoref{chrelation} gives us:
$$\sum_{K\subset H\subset G_2}c_H\cdot|W_{G_2}H|=|P_K|^{-1},$$
$$\sum_{K_0\subset H\subset G_2}c_H\cdot|W_{G_2}H|=|P_{K_0}|^{-1}=|P_K|^{-1}.$$
Taking the difference gives
$$\sum_{K\subset H,K_0\nsubseteq H, H\subset G_2}c_H\cdot|W_{G_2}H|=0.$$
For any $H$ which contains $K$ but not $K_0$, \autoref{decentprop}(c) shows that $P_H\subset P_K=P_{K_0}$. \autoref{decentprop}(d) then shows that $H$ cannot be $G_1$-good, since otherwise we must have $K_0\subset H$. Thus all summands in
$$\sum_{K\subset H,K_0\nsubseteq H, H\subset G_2}c_H\cdot|W_{G_2}H|=0$$
have $H\notin\mathbb{D}$. We already know that $c_H=0$ for all $K\subsetneqq H\notin\mathbb{D}$. Thus $c_K$ must also be zero. $\Box$
\bigskip
Now we can refine \autoref{chrelation}:
\begin{proposition} \label{finalrelation}
For any $K\in\mathbb{D}$, we have
$$\sum_{K\subset H\in\mathbb{D}}c_H\cdot|P_K|\cdot|W_{G_2}H|=1.$$
Moreover, $c_H=c_{H^\prime}$ if $H$ is conjugate to $H^\prime$. If $H\notin\mathbb{D}$, $c_H=0$.
\end{proposition}
\bigskip
We have finished all preparations and can start proving \autoref{gen12}.
\bigskip
We can still use induction. For $K=G_2$, \autoref{finalrelation} shows that
$$c_{G_2}\cdot|P_{G_2}|=1.$$
Since $P_{G_2}\subset G_1$, $|P_{G_2}|$ becomes invertible after we invert $|G_1|$. Thus the denominator of $c_{G_2}$ is invertible.
\medskip
Now fix $K\subset G_2$. Assume that the denominators of all $c_H$ with $K\subsetneqq H\subset G_2$ are invertible. We will prove the same thing for $c_K$. We can assume $K\in\mathbb{D}$ (otherwise $c_K=0$ by \autoref{finalrelation}).
\medskip
Let $P=P_K$. List all maximal proper $G_2$-good subgroups of $P$ as $P_1,P_2,...,P_n$. For any $S\subset\{1,2,...,n\}$. Define
$$P_S:=P\cap\left(\bigcap_{i\in S}P_i\right).$$
\autoref{decentprop}(f) shows that $P_S$ is $G_2$-good. Define $K_S:=K_{P_S}$. Then we have
$$P_{\{i\}}=P_i,\text{ }P_\emptyset=P,\text{ }K_\emptyset=K.$$
\medskip
We can also extend \autoref{nnaction}:
\begin{lemma} \label{naction}
Consider the action of $N_{G_2}K$ on $P_K$ by conjugation. The action of any $h\in N_{G_2}K$ permutes different $P_S$.
\end{lemma}
\paragraph{Proof:} For any $P_i$ and $h\in N_{G_2}K$, notice that $h^{-1}P_ih$ is also a maximal proper $G_2$-good subgroup of $P=P_K$. Thus $h^{-1}P_ih=P_j$ for some $j$. So the action of $h$ becomes a permutation among $P_1,P_2,...,P_n$, and also among general $P_S$, since each $P_S$ is defined as the intersection of some $P_i$. $\Box$
\bigskip
Write $K_i:=K_{\{i\}}=K_{P_i}$, $i=1,2,...,n$. A quick application of \autoref{decentprop} shows:
\begin{lemma} \label{maximalinclude}
For any $K\subsetneqq H\in\mathbb{D}$, $K_i\subset H$ for at least one $i$.
\end{lemma}
Apply \autoref{finalrelation} for each $S\subset\{1,2,...,n\}$:
$$\sum_{K_S\subset H\in\mathbb{D}}c_H\cdot|P_S|\cdot|W_{G_2}H|=1.$$
Add these equations together with additional signs:
$$\sum_{S}(-1)^{|S|}\left(\sum_{K_S\subset H\in\mathbb{D}}c_H\cdot|P_S|\cdot|W_{G_2}H|\right)=\sum_S(-1)^{|S|}=\binom{n}{0}-\binom{n}{1}+...+(-1)^n\binom{n}{n}=0.$$
Rewriting the sum in the left hand side:
$$\sum_{K\subset H\in\mathbb{D}}c_H\cdot|W_{G_2}H|\cdot\left(\sum_{K_S\subset H}(-1)^{|S|}|P_S|\right)=0.$$
Separate the summand of $K$:
$$-c_K\cdot|W_{G_2}K|\cdot|P_K|=\sum_{K\subsetneqq H\in\mathbb{D}}c_H\cdot|W_{G_2}H|\cdot\left(\sum_{K_S\subset H}(-1)^{|S|}|P_S|\right).$$
We divide all $H$ with $K\subsetneqq H\in\mathbb{D}$ into different classes: $H_1,H_2$ are in the same class if there exists $h\in N_{G_2}K$ such that $H_2=h^{-1}H_1h$.
For $K\subsetneqq H\in\mathbb{D}$ and $h\in N_{G_2}K$, $h^{-1}Hh$ is also a $G_1$-good subgroup which has $K$ as a proper subgroup. Moreover, for any $h_1,h_2\in N_{G_2}K$, we have $h_1^{-1}Hh_1=h_2^{-1}Hh_2$ if and only if $h_1,h_2$ are in the same $N_{G_2}H$-coset. Thus the class of $H$ contains $|N_{G_2}K|/|N_{G_2}K\cap N_{G_2}H|$ elements.
For any $H_1,H_2$ in the same class, since $H_1$ is conjugate to $H_2$, we have
$$c_{H_1}=c_{H_2},\text{ }|W_{G_2}H_1|=|W_{G_2}H_2|.$$
\autoref{naction} shows that
$$\sum_{K_S\subset H_1}(-1)^S|P_S|=\sum_{K_S\subset H_2}(-1)^S|P_S|.$$
Thus we have
$$-c_K\cdot|W_{G_2}K|\cdot|P_K|=\sum_{K\subsetneqq H\in\mathbb{D}}c_H\cdot|W_{G_2}H|\cdot\left(\sum_{K_S\subset H}(-1)^{|S|}|P_S|\right)$$
$$=\sum_{\text{one }H\text{ from each class}}c_H\cdot|W_{G_2}H|\cdot(\text{size of the class})\cdot\left(\sum_{K_S\subset H}(-1)^{|S|}|P_S|\right)$$
$$=\sum_{\text{one }H\text{ from each class}}c_H\cdot\frac{|N_{G_2}H|}{|H|}\cdot\frac{|N_{G_2}K|}{|N_{G_2}K\cap N_{G_2}H|}\cdot\left(\sum_{K_S\subset H}(-1)^{|S|}|P_S|\right).$$
When $|G_1|$ is inverted, $|P_K|$ becomes invertible. We also know that $c_H$ has invertible denominators for all $H\supsetneqq K$ by assumption. Now it suffices to prove:
\begin{proposition} \label{divide}
For any $K\subsetneqq H\in\mathbb{D}$, $|W_{G_2}K|$ divides
$$\frac{|N_{G_2}H|}{|H|}\cdot\frac{|N_{G_2}K|}{|N_{G_2}K\cap N_{G_2}H|}\cdot\left(\sum_{K_S\subset H}(-1)^{|S|}|P_S|\right).$$
\end{proposition}
\paragraph{Proof:} Notice that $H$ must contains at least one $K_i$ according to \autoref{maximalinclude}. Without loss of generality, assume that $H$ contains $K_1,K_2,...,K_m$, but does not contain $K_{m+1},K_{m+2},...,K_n$. Then \autoref{decentprop} shows that $P_H$ is contained in $P_1,P_2,...,P_m$, but not in $P_{m+1},...,P_n$.
We have
$$\sum_{K_S\subset H}(-1)^{|S|}|P_S|=\sum_{S\subset\{1,2,...,m\}}(-1)^{|S|}|P_S|=|P|+\sum_{\emptyset\neq S\subset\{1,2,...,m\}}(-1)^{|S|}\left|\bigcap_{i\in S}P_i\right|.$$
This is the number of elements in $P$ which are not contained in any of $P_1,P_2,...,P_m$.
\medskip
Now we consider the group action on $P_K$. But instead of the whole $N_{G_2}K$, we only use the action of $L:=N_{G_2}K\cap H$.
For any $g\in P-\bigcup_{i=1}^mP_i$ and $h\in L$, assume that $h^{-1}gh\in P_i$ for some $i\in\{1,2,...,m\}$. Then $g\in hP_ih^{-1}$. However, since $h\in N_{G_2}K$, $hP_ih^{-1}$ is another maximal proper $G_2$-good subgroup of $P$, denoted by $P_j$. Since $h\in H$ and $K_i\subset H$, $hK_ih^{-1}\subset H$ and hence $K_j\subset H$. This means that $j\in\{1,2,...,m\}$ and $g\in P_j$, which is a contradiction.
In other words, the subset $P-\bigcup_{i=1}^mP_i$ is closed under the action of $L$.
\medskip
We can also describe the orbit types of this $L$-action.
\smallskip
For any $g\in P-\bigcup_{i=1}^mP_i$ and $h\in L$, assume that $h^{-1}gh=g$ and $h\notin K$. Then $g\in P_{K^\prime}\subsetneqq P$, where $K^\prime$ is generated by $K$ and $h$. Since $h\in H$ and $K\subset H$, we have $K^\prime\subset H$ and hence $P_H\subset P_{K^\prime}\subsetneqq P_K$. Since $P_1,P_2,...,P_m$ are all the maximal $G_2$-good strict subgroups of $P=P_K$ which contain $P_H$, $P_{K^\prime}$ must be contained in one of them. Thus $g$ is contained in one of $P_1,...,P_m$, which is a contradiction.
Therefore, $h^{-1}gh=g$ if and only if $h\in K$. Thus $P-\bigcup_{i=1}^mP_i$ only have $L$-orbits as $L/K$. So we have
$$|L/K|=\frac{|N_{G_2}K\cap H|}{|K|}\text{ divides }\left|P-\bigcup_{i=1}^mP_i\right|=\sum_{K_S\subset H}(-1)^{|S|}|P_S|.$$
\smallskip
Thus it suffices to show that
$$|W_{G_2}K|=\frac{|N_{G_2}K|}{|K|}\text{ divides }\frac{|N_{G_2}H|}{|H|}\cdot\frac{|N_{G_2}K|}{|N_{G_2}K\cap N_{G_2}H|}\cdot\frac{|N_{G_2}K\cap H|}{|K|},$$
or equivalently,
$$\frac{|N_{G_2}K\cap N_{G_2}H|}{|N_{G_2}K\cap H|}\text{ divides }\frac{|N_{G_2}H|}{|H|}=|(N_{G_2}H)/H|.$$
This relation follows from the fact that $N_{G_2}K\cap H$ is the kernel of the group homomorphism
$$N_{G_2}K\cap N_{G_2}H\hookrightarrow N_{G_2}H\rightarrow (N_{G_2}H)/H.$$
$\Box$
\bibliographystyle{abbrv}
|
\section{Introduction}
\subsection{Environmental Microorganisms}
Today, Environmental Microorganisms (EMs) are inseparable from our lives~\cite{Li-2016-EMABC}.
Some EMs are conducive to the development of ecology and promote the progress of
human civilization. However, some EMs hinder ecological balance and even cause urban water pollution to affect human health~\cite{anand-2021-sc2a}. Many researchers use microscopes to observe and identify EMs, but sometimes they have deviations in analyzing EMs.
Computer vision analysis method is of great significance, which can help researchers to analyze the EMs with higher accuracy and efficiency.
For example, \textit{Oscillatoria} is a common EM, which can be observed in various freshwater environments and can thrive in various environments. When it reproduces vigorously, it will produce unpleasant odors, cause water pollution, consume oxygen in the water, and cause fish and shrimp to die of hypoxia~\cite{lu-202-1algicidal}.
In addition, \textit{Scenedesmus} is also a freshwater planktonic algae microorganism, which is usually composed of four to eight cells. \textit{Scenedesmus} has strong resistance to organic pollutants, and plays a vital role in water self-purification and sewage purification~\cite{kashyap-2021-scaao}.
The images of the EMs proposed above are shown in Fig.~\ref{Fig:EMs}.
\begin{figure*}
\centering
\includegraphics[scale=.80]{image/EM.pdf}
\caption{\centering{An example of EM images}}
\label{Fig:EMs}
\end{figure*}
\subsection{Application Scenarios of Environmental Microorganism Images}
Image noise appears in the process of acquiring and transmitting EMs images~\cite{Gonzalez-2002-DIP}. Image denoising can reduce the noise of the EM image while preserving the details of the image. Besides, in the process of deep learning based EMs image analyzation, we can extract the features of EM images, then send them to the deep learning network model for training, and match them with known data to classify, retrieve and detect EMs. In addition, EM images can also be applied in the field of image segmentation to separate microorganisms from the complex background of the image~\cite{pal-1993-AROIS}. Meanwhile, EM images can be used in the field of EM object detection. First, we can frame and mark the known EM objects in the original image, and then transfer the image to the object detection model for feature extraction and network training~\cite{zhang-2021-comprehensive}. Finally, the trained model can be applied for object detection of EMs.
\subsection{Contribution}
EMs are one of the important indicators for investigating the environment, so it is very essential to collect EM data and information~\cite{kosov-2018-EMCUC}. Environmental Microorganism Image Dataset SeventhVersion (EMDS-7) are all taken from urban areas, which can be used to monitor the pollution of the urban water environment.
Furthermore, due to the constant changes in conditions such as temperature and humidity, EMs are very sensitive to these conditions, so the biomass of EMs are easily affected~\cite{rodriguez2020temporal}. It is difficult to collect enough EM images. Currently, there are some EM data sets, but many of them are not open source. EMDS-7 is provided to researchers as an open source data set. In addition, we prepare high-quality corresponding object label files of EMDS-7 for algorithms and model evaluation. The label file of EMs can be directly used in multiple object detection and analysis. EMDS-7 has a variety of EM images, which provides sufficient data support for EMs object detection and achieves satisfactory detection results.
\section{Dataset Information of EMDS-7}
EMDS-7 consists of 2365 images of 42 EM categories and 13216 labeled objects. The EM sampling sources are images taken from different lakes and rivers in Shenyang (Northeast China) by two environmental biologists (Northeastern University, China) under a 400× optical microscope from 2018 to 2019. Then, four bioinformatics scientists (Northeastern University, China) manually prepared the object labeling files in ".XML" format corresponding to the original
2365 images from 2020 to 2021. In the EM object labeling files, 42 types of EMs are labeled by their categories. In addition, the unknown EMs and impurities are marked as Unknown, and a total of 13216 labeled objects are obtained. Table.~\ref{tbl1} shows the basic information of 41 types of EMs and unknown objects; Fig.~\ref{Fig:EMDS7} shows examples of 41 types of EMs and unknown objects in EMDS-7.
The labeled files of EMDS-7 images are manually labeled base on the following two rules:
Rule-A: All identifiable EMs that appear completely or more than 60$\%$ of their own in all images are marked with category labels corresponding to 41 categories.
Rule-B: Unknown EMs that are less than 40$\%$ of their own in all images and EMs other than 41 categories in this database. In addition, some obvious impurities in the background of the image are marked as unknown.
EMDS-7 is freely published for non-commercial purpose at: https://figshare.com/articles/dataset/EMDS-7\_DataSet/16869571
\begin{table*}[]
\caption{Basic information of 42 EM classes in EMDS-7. Number of original images (NoOI), Number of EM object NEMo), Visible characteristics (VC).}
\label{tbl1}
\resizebox{\textwidth}{40mm}{
\begin{tabular}{|l|l|l|l|l|l|l|l|}
\hline
Classes & NoOI & NEMo & VC & Classes & NoOI & NEMo & VC \\ \hline
\textit{Oscillatoria}& 41 & 178 &Cylindrical &\textit{Staurastrum}& 9 & 9 &Multi-radial symmetry \\ \hline
\textit{Ankistrodesmus}& 5 & 45 &Cell needle &\textit{Phormidium} & 276 & 1216 &Plant body gelatinous or leathery \\ \hline
\textit{Microcystis}& 307 & 826 &Spherical masses &\textit{Fragilaria} & 55 & 59 &Lanceolate \\ \hline
\textit{Gomphonema} & 87 & 108 &Linear-lanceolate &\textit{Anabaenopsis} & 22 & 37 &Filaments \\ \hline
\textit{Sphaerocystis}& 55 & 53 &Cell sphere &\textit{Coelosphaerium} & 77 & 165 &Group glue is generous and transparent \\ \hline
\textit{Cosmarium} & 17 & 28 &cell side flattened &\textit{Crucigenia} & 9 & 16 &Micelles with many cells \\ \hline
\textit{Cocconeis} & 14 & 15 &Flat, oval cells &\textit{Achnanthes} & 18 & 19 &Shell surface linear lanceolate \\ \hline
\textit{Tribonema} & 49 & 88 &Yellow-green cotton-like &\textit{Synedra}& 77 & 206 &Shell needle \\ \hline
\textit{Chlorella} & 80 & 155 &Small round or slightly oval &\textit{Ceratium} & 23 & 24 &Flat back and abdomen \\ \hline
\textit{Tetraedron} & 25 & 66 &Flat or pyramidal &\textit{Pompholyx} & 49 & 51 &Carapace Oval or Shield Shape \\ \hline
\textit{Ankistrodesmus} & 64 & 84 &Needle to spindle &\textit{Merismopedia} & 33 & 38 &Flat group \\ \hline
\textit{Brachionus} & 113 & 144 &quilt is wider and square&\textit{Spirogyra} & 89 & 134 &Strip, spiral \\ \hline
\textit{Chaenea} & 6 & 13 &Cylindrical or spindle &\textit{Coelastrum}& 29 & 30 &Spherical, oval or truncated pyramid \\ \hline
\textit{Pediastrum} & 95 & 105 &Disc or star &\textit{Raphidiopsis}& 9 & 19 &Curved \\ \hline
\textit{Spirulina} & 18 & 73 &Spiral &\textit{Gomphosphaeria}& 58 & 79 &Oval tiny groups \\ \hline
\textit{Actinastrum} & 23 & 181 &Wide round or pointed & \textit{Euglena} & 81 & 81 &Spindle to needle \\ \hline
\textit{Navicula} & 75 & 90 &Shell surface fusiform or oval &\textit{Euchlanis} & 14 & 13 &Oval or pear-shaped \\ \hline
\textit{Scenedesmus} & 86 & 139 &Oval or spindle &\textit{Keratella}& 65 & 69 &Irregular spines \\ \hline
\textit{Golenkinia} & 60 & 279 &Irregularly slender bristles &\textit{diversicornis} & 89 & 99 &Feet have toes and quilt \\ \hline
\textit{Pinnularia} & 36 & 38 &Oval to boat &\textit{Surirella}& 22 & 37 &False shell seam \\ \hline
\textit{unknown} & & 8088 &Unknown EM and impurities &\textit{Characium}& 5 & 19 &Spindle \\ \hline
\end{tabular}}
\end{table*}
\begin{figure*}
\centering
\includegraphics[scale=.70]{image/object.pdf}
\caption{\centering{An example of EM images in EMDS-7.}}\label{Fig:EMDS7}
\end{figure*}
\section{Object Detection Methods for EMDS-7 Evaluation in This Paper}
\subsection{Faster RCNN}
Faster RCNN generates candidate frames based on the Anchor mechanism by adding a region proposal networks (RPN), and finally integrates feature extraction, candidate frame selection, frame regression, and classification into one network, thereby the detection accuracy and efficiency can be effectively improved~\cite{ren-2015-FRTRO}. Faster RCNN achieves high-precision detection performance by adding RPN in two stages, which performs well on multiple data sets. Compared with other one-stage detection networks, the two-stage network is more accurate and can solve the problems of multi-scale and small object detection.
\subsection{YOLOv3}
The object detection model of the YOLO series is a one-stage detection network, which can locate and classify the objects at the same time. The advantage is that the training speed is fast with less time-consuming. One of the most types is YOLOv3. Joseph Redmon and others used the new basic network darknet-53 in the backbone of YOLOv3 for physical sign extraction~\cite{redmon-2016-yolo}. It contains 53 convolutional layers and introduces a residual structure so that the network can reach a deep level while avoiding the problem of gradient disappearance. In addition, darknet-53 removes the pooling layer and uses a convolutional layer with a step size of 2 to reduce the dimensionality of the feature map, which can maintain the information transmission better. And YOLOv3 also has excellent structures such as anchor and FPN~\cite{lin-2017-FPNFO}.
\subsection{YOLOv4}
YOLOv4 is an improved version based on YOLOv3, which adds CSP and PAN structures~\cite{bochkovskiy-2020-yolov4}. The backbone network of YOLOv3 is modified to CSPDarkne53, and add an spp (spatial pyramid pooling) idea behind the backbone network to expand the receptive field, using 1×1, 5×5, 9×9, 13×13 as the largest Pooling, multi-scale fusion, and improve the accuracy of the model. At the same time, in the neck network of Yolov4, there are Feature Pyramid Network (FPN)~\cite{lin-2017-FPNFO}, Path Aggregation Network (PAN), BiFPN, and NAS-FPN, which can collect different feature maps more effectively.
\subsection{SSD}
SSD is another striking object detection network after YOLO. SSD has two major advantages~\cite{liu-2016-ssd}. First, SSD extracts feature maps of different scales for detection. Large-scale feature maps can be used to detect small objects, while small-scale feature maps can be used to detect large objects. Second, SSD uses different Prior boxes (Prior boxes, Default boxes, Anchors) for scale and aspect ratio. It follows the method of direct regression box and classification probability in YOLO, and uses anchors to improve recognition accuracy referring to Faster R-CNN. By combining these two networks, SSD balances the advantages and disadvantages of Faster R-CNN and YOLO.
\subsection{RetinaNet}
The RetinaNet object detection model is also a one-stage object detection network. RetinaNet essentially consists of a backbone network (BackBone) and two subnets (SubNet)~\cite{lin-2017-focal}. The backbone network is responsible for calculating the convolution feature map on the entire input image, which is composed of the ResNet residual network and the FPN feature pyramid network. The two sub-networks use the features extracted from the backbone network to achieve their respective functions~\cite{lin-2017-FPNFO}. The first sub-network completes the classification task; the second sub-network completes the bounding box regression task.
\section{Evaluation of Deep Learning Object Detection Methods}
Object detection is an important part of image analysis. To prove the effectiveness of EMDS-7 in object detection and evaluation, we use five different deep learning object detection models to detect EMs in the EMDS-7 data set. The five models are Faster RCNN~\cite{ren-2015-FRTRO}, SSD~\cite{liu-2016-ssd}, RetinaNet~\cite{lin-2017-focal}, YOLOv3~\cite{redmon-2016-yolo} and YOLOv4~\cite{bochkovskiy-2020-yolov4}. Because the number of images of each category of EM in the EMDS-7 data set is different, we divide each category of Ems data set into the training, validation and test set according to 6:2:2 to ensure each sets has 42 types of EMs. We train EMDS-7 for five different kinds of deep learning object detection model, and then respectively predict the images of the test set. We calculate the number of EM objects in 456 EMS images. We set the threshold of the predictive frame confidence to be 0.5, when the model predicts the object's confidence is greater than 0.5, the prediction box is displayed, and Intersection over Union (IOU) is set to 0.3. In Fig.~\ref{Fig:tf} and Fig.~\ref{Fig:ap}, we summarize the number of predictive positive samples and negative samples, and the Average Precision (AP) value of each class of EMs. Analysis of predict results is shown in Tab.~\ref{tbl:3}. We also illustrate the location of the prediction box in the EMS image, and some samples are shown in Fig.\ref{Fig:model}.
In Tab.~\ref{tbl:3} and Fig.~\ref{Fig:ap} we can see that the EMDS-7 data set performs well in the task of object detection, and most of the EMs can be accurately identified. Meanwhile, different object detection models differ in the object detection effect of EMDS-7, which proves that EMDS-7 dataset provide performance analysis of different networks. We also calculate five models of the Mean Average Precision (MAP), which the Faster RCNN value of 76.05$\%$, the YOLOv3 value of 58.61$\%$, the YOLOv4 value of 40.30$\%$, the SSD value of 57.83$\%$, the RetinaNet value of 57.96$\%$. We can see that Faster RCNN has the highest detection performance of EMDS-7. YOLOv3, SSD, RetinaNet model predicts that the MAP value is similar. The lowest is YOLOv4. However, some EMS category predictive AP values can reach 100$\%$, while some kind of AP values are 0$\%$. We found that there are little AP value is 100$\%$ or 0$\%$, and the category of EMs in the image is basically consistent and the EMs characteristics are relatively large. In addition, different models are different from the method of extracting features, so differentiation occurs when model training. For example, in the object detection of Euglena category, the predictive sample has only two, and the AP value is 100$\%$ in the FASTER RCNN, but in the RetinaNET is 0$\%$. In synthesis, the EMDS-7 database can provide analytical performance for different object detection models.
\begin{table*}[]
\caption{A comparison of the object detection results on test set of EMs.
GTO (Ground Ture Object), tp (ture positive), fp (flase positive) and ap (Average Precision) (In [\%].)}
\label{tbl:3}
\resizebox{\textwidth}{100mm}{
\begin{tabular}{|l|l|l|l|l|l|l|l|l|l|l|l|l|l|l|l|l|}
\hline
& & \multicolumn{3}{l|}{Faster RCNN} & \multicolumn{3}{l|}{YOLOv3} & \multicolumn{3}{l|}{YOLOv4} & \multicolumn{3}{l|}{SSD} & \multicolumn{3}{l|}{RetinaNet} \\ \hline
EMS & GTO & tp & fp & ap & tp & fp & ap & tp & fp & ap & tp & fp & ap & tp & fp & ap \\ \hline
\textit{Oscillatoria} & 26 & 14 & 25 & 38.70 & 6 & 9 & 19.08 & 0 & 0 & 0 & 1 & 1 & 1.92 & 1 & 0 & 4.35 \\ \hline
\textit{Ankistrodesmus} & 9 & 1 & 2 & 5.56 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 \\ \hline
\textit{Microcystis } & 151 & 118 & 112 & 65.97 & 105 & 95 & 56.66 & 87 & 37 & 44.96 & 83 & 35 & 46.10 & 94 & 31 & 51.61 \\ \hline
\textit{Gomphonema} & 19 & 19 & 5 & 1.00 & 19 & 6 & 100 & 18 & 1 & 94.74 & 16 & 2 & 83.90 & 18 & 2 & 89.50 \\ \hline
\textit{Sphaerocystis} & 11 & 10 & 0 & 90.91 & 10 & 3 & 89.26 & 8 & 1 & 71.72 & 10 & 0 & 90.91 & 10 & 0 & 100 \\ \hline
\textit{Cosmarium} & 5 & 3 & 2 & 60.00 & 2 & 0 & 40.00 & 1 & 0 & 20.00 & 3 & 0 & 60.00 & 3 & 1 & 75.00 \\ \hline
\textit{Cocconeis } & 2 & 2 & 1 & 1.00 & 1 & 0 & 50.00 & 2 & 0 & 1.00 & 2 & 0 & 100 & 2 & 0 & 100 \\ \hline
\textit{Tribonema} & 17 & 8 & 5 & 43.03 & 8 & 7 & 45.45 & 5 & 3 & 28.43 & 5 & 0 & 29.41 & 8 & 7 & 37.11 \\ \hline
\textit{Chlorella} & 34 & 20 & 24 & 52.82 & 18 & 11 & 49.98 & 15 & 4 & 42.98 & 16 & 3 & 46.05 & 15 & 1 & 48.33 \\ \hline
\textit{Tetraedron} & 7 & 7 & 3 & 1.00 & 7 & 5 & 1.00 & 6 & 0 & 85.71 & 7 & 0 & 100 & 7 & 1 & 46.67 \\ \hline
\textit{Ankistrodesmus} & 13 & 13 & 6 & 98.97 & 13 & 5 & 97.20 & 12 & 4 & 82.75 & 11 & 2 & 79.94 & 12 & 1 & 98.72 \\ \hline
\textit{Brachionus} & 25 & 25 & 19 & 93.18 & 21 & 13 & 76.38 & 13 & 4 & 49.10 & 17 & 2 & 66.40 & 25 & 9 & 83.97 \\ \hline
\textit{Chaenea} & 2 & 2 & 1 & 1.00 & 1 & 1 & 25.00 & 2 & 0 & 100 & 0 & 0 & 0 & 1 & 0 & 50.00 \\ \hline
\textit{Pediastrum} & 24 & 23 & 1 & 95.83 & 23 & 3 & 94.30 & 20 & 1 & 82.34 & 23 & 0 & 95.83 & 20 & 2 & 95.24 \\ \hline
\textit{Spirulina} & 14 & 8 & 6 & 57.14 & 8 & 7 & 42.42 & 3 & 0 & 21.43 & 4 & 0 & 28.57 & 9 & 2 & 45.79 \\ \hline
\textit{Actinastrum} & 40 & 29 & 14 & 67.48 & 23 & 14 & 51.58 & 18 & 0 & 45.00 & 23 & 0 & 57.50 & 29 & 3 & 77.45 \\ \hline
\textit{Navicula} & 16 & 14 & 7 & 85.83 & 12 & 5 & 71.88 & 10 & 3 & 62.50 & 12 & 6 & 64.91 & 15 & 6 & 72.67 \\ \hline
\textit{Scenedesmus} & 25 & 19 & 7 & 76.00 & 19 & 10 & 72.88 & 19 & 3 & 74.40 & 20 & 1 & 80.00 & 18 & 0 & 75.00 \\ \hline
\textit{Golenkinia} & 41 & 31 & 25 & 71.15 & 32 & 35 & 71.30 & 20 & 8 & 37.52 & 29 & 8 & 68.88 & 29 & 9 & 60.57 \\ \hline
\textit{Pinnularia} & 7 & 7 & 3 & 1.00 & 7 & 4 & 83.47 & 6 & 0 & 85.71 & 5 & 0 & 71.43 & 6 & 1 & 75.00 \\ \hline
\textit{Staurastrum} & 3 & 0 & 1 & 0 & 0 & 0 & 0 & 1 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\ \hline
\textit{Phormidium} & 234 & 194 & 127 & 74.23 & 188 & 151 & 64.49 & 147 & 73 & 48.84 & 4 & 1 & 43.19 & 142 & 32 & 52.93 \\ \hline
\textit{Fragilaria} & 11 & 10 & 3 & 90.91 & 10 & 2 & 90.91 & 6 & 3 & 42.73 & 21 & 20 & 90.91 & 10 & 3 & 90.91 \\ \hline
\textit{Anabaenopsis} & 5 & 5 & 1 & 1.00 & 5 & 2 & 94.29 & 2 & 0 & 40.00 & 2 & 0 & 80.00 & 9 & 2 & 81.82 \\ \hline
\textit{Coelosphaerium} & 42 & 21 & 12 & 44.58 & 11 & 2 & 24.43 & 1 & 0 & 2.38 & 2 & 2 & 44.19 & 6 & 3 & 18.30 \\ \hline
\textit{Crucigenia} & 4 & 2 & 1 & 50.00 & 2 & 1 & 41.67 & 0 & 0 & 0 & 11 & 3 & 50.00 & 2 & 0 & 66.67 \\ \hline
\textit{Achnanthes} & 3 & 3 & 0 & 1.00 & 2 & 1 & 66.67 & 0 & 0 & 0 & 2 & 2 & 50.00 & 3 & 0 & 100 \\ \hline
\textit{Synedra} & 34 & 27 & 21 & 72.42 & 23 & 29 & 55.90 & 16 & 14 & 30.92 & 11 & 3 & 31.00 & 15 & 1 & 34.09 \\ \hline
\textit{Ceratium} & 4 & 2 & 0 & 50.00 & 2 & 2 & 50.00 & 0 & 0 & 0 & 1 & 1 & 12.50 & 1 & 1 & 33.33 \\ \hline
\textit{Pompholyx} & 9 & 8 & 5 & 87.65 & 7 & 5 & 70.83 & 2 & 0 & 22.22 & 8 & 4 & 86.67 & 8 & 3 & 85.28 \\ \hline
\textit{Merismopedia} & 6 & 6 & 1 & 1.00 & 4 & 1 & 63.33 & 5 & 0 & 83.33 & 6 & 1 & 97.62 & 5 & 0 & 71.43 \\ \hline
\textit{Spirogyra} & 25 & 21 & 9 & 81.31 & 19 & 16 & 70.23 & 12 & 16 & 28.14 & 20 & 4 & 78.18 & 21 & 2 & 73.50 \\ \hline
\textit{Coelastrum} & 6 & 4 & 0 & 66.67 & 4 & 1 & 66.67 & 4 & 0 & 66.67 & 4 & 1 & 66.67 & 3 & 0 & 42.86 \\ \hline
\textit{Raphidiopsis} & 2 & 2 & 1 & 83.33 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 1 & 50.00 & 0 & 0 & 0 \\ \hline
\textit{Gomphosphaeria} & 14 & 14 & 6 & 98.74 & 10 & 2 & 70.78 & 5 & 8 & 16.23 & 13 & 2 & 90.81 & 11 & 3 & 100 \\ \hline
\textit{Euglena} & 16 & 16 & 1 & 1.00 & 15 & 6 & 92.97 & 14 & 4 & 80.28 & 15 & 1 & 92.97 & 12 & 1 & 73.56 \\ \hline
\textit{Euchlanis} & 2 & 2 & 0 & 1.00 & 1 & 0 & 50.00 & 0 & 0 & 0 & 1 & 0 & 50.00 & 0 & 0 & 0 \\ \hline
\textit{Keratella} & 13 & 13 & 4 & 98.19 & 7 & 4 & 51.54 & 3 & 1 & 23.08 & 9 & 1 & 69.23 & 9 & 1 & 69.23 \\ \hline
\textit{diversicornis} & 18 & 17 & 4 & 94.44 & 17 & 17 & 81.71 & 13 & 14 & 40.30 & 14 & 3 & 72.47 & 19 & 11 & 92.78 \\ \hline
\textit{Surirella} & 6 & 4 & 3 & 66.67 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 16.67 & 1 & 1 & 16.67 \\ \hline
\textit{Characium} & 5 & 4 & 2 & 76.00 & 3 & 0 & 60.00 & 0 & 0 & 0 & 3 & 1 & 60.00 & 0 & 0 & 0 \\ \hline
\textit{unknown} & 1429 & 1049 & 1585 & 56.49 & 1068 & 1126 & 58.17 & 610 & 177 & 38.15 & 371 & 66 & 24.17 & 622 & 146 & 36.92 \\ \hline
\end{tabular}}
\end{table*}
\begin{figure*}
\centering
\includegraphics[scale=.50]{image/tf.pdf}
\caption{\centering{Five models predict the number of EMDS-7 positive and negative samples.}}\label{Fig:tf}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[scale=.50]{image/map.pdf}
\caption{\centering{Each category of EM object detection prediction AP value in EMDS--7}}\label{Fig:ap}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[scale=.50]{image/model.pdf}
\caption{\centering{Five object detection model prediction results in EMDS-7}}\label{Fig:model}
\end{figure*}
\section{Conclusion and Future Work}
EMDS-7 is an object detection data set containing 42 types of EMs, which contains the original image of EMs and object label data for corresponding EMs. EMDS-7 labeled 15342 EMs. At the same time, we further add some deep learning object detection experiments to the EMDS-7 database to prove the effectiveness.
During the object detection process, we divide the data set according to 6: 2: 2 for train, validation and test sets. We use five different deep learning object detection methods to test EMDS-7 and use multiple evaluation indices to evaluate the prediction results.
In the future, we will enlarge the category of EMs to increase the number of images of each EM. Make each class data balanced and sufficient. We hope to use the EMDS-7 database to achieve more features in the future. Meanwhile
\section*{Acknowledgements}
This work is supported by the ``National Natural Science Foundation of China'' (No.61806047)
and the ``Fundamental Research Funds for the Central Universities'' (No. N2019003).
We thank Miss Zixian Li and Mr. Guoxian Li for their important discussion.
\section*{Declaration of competing interest}
The authors declare that they have no conflict of interest.
\bibliographystyle{unsrt}
\section{Introduction}
\subsection{Environmental Microorganisms}
Today, Environmental Microorganisms (EMs) are inseparable from our lives~\cite{Li-2016-EMABC}.
Some EMs are conducive to the development of ecology and promote the progress of
human civilization. However, some EMs hinder ecological balance and even cause urban water pollution to affect human health~\cite{anand-2021-sc2a}. Many researchers use microscopes to observe and identify EMs, but sometimes they have deviations in analyzing EMs.
Computer vision analysis method is of great significance, which can help researchers to analyze the EMs with higher accuracy and efficiency.
For example, \textit{Oscillatoria} is a common EM, which can be observed in various freshwater environments and can thrive in various environments. When it reproduces vigorously, it will produce unpleasant odors, cause water pollution, consume oxygen in the water, and cause fish and shrimp to die of hypoxia~\cite{lu-202-1algicidal}.
In addition, \textit{Scenedesmus} is also a freshwater planktonic algae microorganism, which is usually composed of four to eight cells. \textit{Scenedesmus} has strong resistance to organic pollutants, and plays a vital role in water self-purification and sewage purification~\cite{kashyap-2021-scaao}.
The images of the EMs proposed above are shown in Fig.~\ref{Fig:EMs}.
\begin{figure*}
\centering
\includegraphics[scale=.80]{image/EM.pdf}
\caption{\centering{An example of EM images}}
\label{Fig:EMs}
\end{figure*}
\subsection{Application Scenarios of Environmental Microorganism Images}
Image noise appears in the process of acquiring and transmitting EMs images~\cite{Gonzalez-2002-DIP}. Image denoising can reduce the noise of the EM image while preserving the details of the image. Besides, in the process of deep learning based EMs image analyzation, we can extract the features of EM images, then send them to the deep learning network model for training, and match them with known data to classify, retrieve and detect EMs. In addition, EM images can also be applied in the field of image segmentation to separate microorganisms from the complex background of the image~\cite{pal-1993-AROIS}. Meanwhile, EM images can be used in the field of EM object detection. First, we can frame and mark the known EM objects in the original image, and then transfer the image to the object detection model for feature extraction and network training~\cite{zhang-2021-comprehensive}. Finally, the trained model can be applied for object detection of EMs.
\subsection{Contribution}
EMs are one of the important indicators for investigating the environment, so it is very essential to collect EM data and information~\cite{kosov-2018-EMCUC}. Environmental Microorganism Image Dataset SeventhVersion (EMDS-7) are all taken from urban areas, which can be used to monitor the pollution of the urban water environment.
Furthermore, due to the constant changes in conditions such as temperature and humidity, EMs are very sensitive to these conditions, so the biomass of EMs are easily affected~\cite{rodriguez2020temporal}. It is difficult to collect enough EM images. Currently, there are some EM data sets, but many of them are not open source. EMDS-7 is provided to researchers as an open source data set. In addition, we prepare high-quality corresponding object label files of EMDS-7 for algorithms and model evaluation. The label file of EMs can be directly used in multiple object detection and analysis. EMDS-7 has a variety of EM images, which provides sufficient data support for EMs object detection and achieves satisfactory detection results.
\section{Dataset Information of EMDS-7}
EMDS-7 consists of 2365 images of 42 EM categories and 13216 labeled objects. The EM sampling sources are images taken from different lakes and rivers in Shenyang (Northeast China) by two environmental biologists (Northeastern University, China) under a 400× optical microscope from 2018 to 2019. Then, four bioinformatics scientists (Northeastern University, China) manually prepared the object labeling files in ".XML" format corresponding to the original
2365 images from 2020 to 2021. In the EM object labeling files, 42 types of EMs are labeled by their categories. In addition, the unknown EMs and impurities are marked as Unknown, and a total of 13216 labeled objects are obtained. Table.~\ref{tbl1} shows the basic information of 41 types of EMs and unknown objects; Fig.~\ref{Fig:EMDS7} shows examples of 41 types of EMs and unknown objects in EMDS-7.
The labeled files of EMDS-7 images are manually labeled base on the following two rules:
Rule-A: All identifiable EMs that appear completely or more than 60$\%$ of their own in all images are marked with category labels corresponding to 41 categories.
Rule-B: Unknown EMs that are less than 40$\%$ of their own in all images and EMs other than 41 categories in this database. In addition, some obvious impurities in the background of the image are marked as unknown.
EMDS-7 is freely published for non-commercial purpose at: https://figshare.com/articles/dataset/EMDS-7\_DataSet/16869571
\begin{table*}[]
\caption{Basic information of 42 EM classes in EMDS-7. Number of original images (NoOI), Number of EM object NEMo), Visible characteristics (VC).}
\label{tbl1}
\resizebox{\textwidth}{40mm}{
\begin{tabular}{|l|l|l|l|l|l|l|l|}
\hline
Classes & NoOI & NEMo & VC & Classes & NoOI & NEMo & VC \\ \hline
\textit{Oscillatoria}& 41 & 178 &Cylindrical &\textit{Staurastrum}& 9 & 9 &Multi-radial symmetry \\ \hline
\textit{Ankistrodesmus}& 5 & 45 &Cell needle &\textit{Phormidium} & 276 & 1216 &Plant body gelatinous or leathery \\ \hline
\textit{Microcystis}& 307 & 826 &Spherical masses &\textit{Fragilaria} & 55 & 59 &Lanceolate \\ \hline
\textit{Gomphonema} & 87 & 108 &Linear-lanceolate &\textit{Anabaenopsis} & 22 & 37 &Filaments \\ \hline
\textit{Sphaerocystis}& 55 & 53 &Cell sphere &\textit{Coelosphaerium} & 77 & 165 &Group glue is generous and transparent \\ \hline
\textit{Cosmarium} & 17 & 28 &cell side flattened &\textit{Crucigenia} & 9 & 16 &Micelles with many cells \\ \hline
\textit{Cocconeis} & 14 & 15 &Flat, oval cells &\textit{Achnanthes} & 18 & 19 &Shell surface linear lanceolate \\ \hline
\textit{Tribonema} & 49 & 88 &Yellow-green cotton-like &\textit{Synedra}& 77 & 206 &Shell needle \\ \hline
\textit{Chlorella} & 80 & 155 &Small round or slightly oval &\textit{Ceratium} & 23 & 24 &Flat back and abdomen \\ \hline
\textit{Tetraedron} & 25 & 66 &Flat or pyramidal &\textit{Pompholyx} & 49 & 51 &Carapace Oval or Shield Shape \\ \hline
\textit{Ankistrodesmus} & 64 & 84 &Needle to spindle &\textit{Merismopedia} & 33 & 38 &Flat group \\ \hline
\textit{Brachionus} & 113 & 144 &quilt is wider and square&\textit{Spirogyra} & 89 & 134 &Strip, spiral \\ \hline
\textit{Chaenea} & 6 & 13 &Cylindrical or spindle &\textit{Coelastrum}& 29 & 30 &Spherical, oval or truncated pyramid \\ \hline
\textit{Pediastrum} & 95 & 105 &Disc or star &\textit{Raphidiopsis}& 9 & 19 &Curved \\ \hline
\textit{Spirulina} & 18 & 73 &Spiral &\textit{Gomphosphaeria}& 58 & 79 &Oval tiny groups \\ \hline
\textit{Actinastrum} & 23 & 181 &Wide round or pointed & \textit{Euglena} & 81 & 81 &Spindle to needle \\ \hline
\textit{Navicula} & 75 & 90 &Shell surface fusiform or oval &\textit{Euchlanis} & 14 & 13 &Oval or pear-shaped \\ \hline
\textit{Scenedesmus} & 86 & 139 &Oval or spindle &\textit{Keratella}& 65 & 69 &Irregular spines \\ \hline
\textit{Golenkinia} & 60 & 279 &Irregularly slender bristles &\textit{diversicornis} & 89 & 99 &Feet have toes and quilt \\ \hline
\textit{Pinnularia} & 36 & 38 &Oval to boat &\textit{Surirella}& 22 & 37 &False shell seam \\ \hline
\textit{unknown} & & 8088 &Unknown EM and impurities &\textit{Characium}& 5 & 19 &Spindle \\ \hline
\end{tabular}}
\end{table*}
\begin{figure*}
\centering
\includegraphics[scale=.70]{image/object.pdf}
\caption{\centering{An example of EM images in EMDS-7.}}\label{Fig:EMDS7}
\end{figure*}
\section{Object Detection Methods for EMDS-7 Evaluation in This Paper}
\subsection{Faster RCNN}
Faster RCNN generates candidate frames based on the Anchor mechanism by adding a region proposal networks (RPN), and finally integrates feature extraction, candidate frame selection, frame regression, and classification into one network, thereby the detection accuracy and efficiency can be effectively improved~\cite{ren-2015-FRTRO}. Faster RCNN achieves high-precision detection performance by adding RPN in two stages, which performs well on multiple data sets. Compared with other one-stage detection networks, the two-stage network is more accurate and can solve the problems of multi-scale and small object detection.
\subsection{YOLOv3}
The object detection model of the YOLO series is a one-stage detection network, which can locate and classify the objects at the same time. The advantage is that the training speed is fast with less time-consuming. One of the most types is YOLOv3. Joseph Redmon and others used the new basic network darknet-53 in the backbone of YOLOv3 for physical sign extraction~\cite{redmon-2016-yolo}. It contains 53 convolutional layers and introduces a residual structure so that the network can reach a deep level while avoiding the problem of gradient disappearance. In addition, darknet-53 removes the pooling layer and uses a convolutional layer with a step size of 2 to reduce the dimensionality of the feature map, which can maintain the information transmission better. And YOLOv3 also has excellent structures such as anchor and FPN~\cite{lin-2017-FPNFO}.
\subsection{YOLOv4}
YOLOv4 is an improved version based on YOLOv3, which adds CSP and PAN structures~\cite{bochkovskiy-2020-yolov4}. The backbone network of YOLOv3 is modified to CSPDarkne53, and add an spp (spatial pyramid pooling) idea behind the backbone network to expand the receptive field, using 1×1, 5×5, 9×9, 13×13 as the largest Pooling, multi-scale fusion, and improve the accuracy of the model. At the same time, in the neck network of Yolov4, there are Feature Pyramid Network (FPN)~\cite{lin-2017-FPNFO}, Path Aggregation Network (PAN), BiFPN, and NAS-FPN, which can collect different feature maps more effectively.
\subsection{SSD}
SSD is another striking object detection network after YOLO. SSD has two major advantages~\cite{liu-2016-ssd}. First, SSD extracts feature maps of different scales for detection. Large-scale feature maps can be used to detect small objects, while small-scale feature maps can be used to detect large objects. Second, SSD uses different Prior boxes (Prior boxes, Default boxes, Anchors) for scale and aspect ratio. It follows the method of direct regression box and classification probability in YOLO, and uses anchors to improve recognition accuracy referring to Faster R-CNN. By combining these two networks, SSD balances the advantages and disadvantages of Faster R-CNN and YOLO.
\subsection{RetinaNet}
The RetinaNet object detection model is also a one-stage object detection network. RetinaNet essentially consists of a backbone network (BackBone) and two subnets (SubNet)~\cite{lin-2017-focal}. The backbone network is responsible for calculating the convolution feature map on the entire input image, which is composed of the ResNet residual network and the FPN feature pyramid network. The two sub-networks use the features extracted from the backbone network to achieve their respective functions~\cite{lin-2017-FPNFO}. The first sub-network completes the classification task; the second sub-network completes the bounding box regression task.
\section{Evaluation of Deep Learning Object Detection Methods}
Object detection is an important part of image analysis. To prove the effectiveness of EMDS-7 in object detection and evaluation, we use five different deep learning object detection models to detect EMs in the EMDS-7 data set. The five models are Faster RCNN~\cite{ren-2015-FRTRO}, SSD~\cite{liu-2016-ssd}, RetinaNet~\cite{lin-2017-focal}, YOLOv3~\cite{redmon-2016-yolo} and YOLOv4~\cite{bochkovskiy-2020-yolov4}. Because the number of images of each category of EM in the EMDS-7 data set is different, we divide each category of Ems data set into the training, validation and test set according to 6:2:2 to ensure each sets has 42 types of EMs. We train EMDS-7 for five different kinds of deep learning object detection model, and then respectively predict the images of the test set. We calculate the number of EM objects in 456 EMS images. We set the threshold of the predictive frame confidence to be 0.5, when the model predicts the object's confidence is greater than 0.5, the prediction box is displayed, and Intersection over Union (IOU) is set to 0.3. In Fig.~\ref{Fig:tf} and Fig.~\ref{Fig:ap}, we summarize the number of predictive positive samples and negative samples, and the Average Precision (AP) value of each class of EMs. Analysis of predict results is shown in Tab.~\ref{tbl:3}. We also illustrate the location of the prediction box in the EMS image, and some samples are shown in Fig.\ref{Fig:model}.
In Tab.~\ref{tbl:3} and Fig.~\ref{Fig:ap} we can see that the EMDS-7 data set performs well in the task of object detection, and most of the EMs can be accurately identified. Meanwhile, different object detection models differ in the object detection effect of EMDS-7, which proves that EMDS-7 dataset provide performance analysis of different networks. We also calculate five models of the Mean Average Precision (MAP), which the Faster RCNN value of 76.05$\%$, the YOLOv3 value of 58.61$\%$, the YOLOv4 value of 40.30$\%$, the SSD value of 57.83$\%$, the RetinaNet value of 57.96$\%$. We can see that Faster RCNN has the highest detection performance of EMDS-7. YOLOv3, SSD, RetinaNet model predicts that the MAP value is similar. The lowest is YOLOv4. However, some EMS category predictive AP values can reach 100$\%$, while some kind of AP values are 0$\%$. We found that there are little AP value is 100$\%$ or 0$\%$, and the category of EMs in the image is basically consistent and the EMs characteristics are relatively large. In addition, different models are different from the method of extracting features, so differentiation occurs when model training. For example, in the object detection of Euglena category, the predictive sample has only two, and the AP value is 100$\%$ in the FASTER RCNN, but in the RetinaNET is 0$\%$. In synthesis, the EMDS-7 database can provide analytical performance for different object detection models.
\begin{table*}[]
\caption{A comparison of the object detection results on test set of EMs.
GTO (Ground Ture Object), tp (ture positive), fp (flase positive) and ap (Average Precision) (In [\%].)}
\label{tbl:3}
\resizebox{\textwidth}{100mm}{
\begin{tabular}{|l|l|l|l|l|l|l|l|l|l|l|l|l|l|l|l|l|}
\hline
& & \multicolumn{3}{l|}{Faster RCNN} & \multicolumn{3}{l|}{YOLOv3} & \multicolumn{3}{l|}{YOLOv4} & \multicolumn{3}{l|}{SSD} & \multicolumn{3}{l|}{RetinaNet} \\ \hline
EMS & GTO & tp & fp & ap & tp & fp & ap & tp & fp & ap & tp & fp & ap & tp & fp & ap \\ \hline
\textit{Oscillatoria} & 26 & 14 & 25 & 38.70 & 6 & 9 & 19.08 & 0 & 0 & 0 & 1 & 1 & 1.92 & 1 & 0 & 4.35 \\ \hline
\textit{Ankistrodesmus} & 9 & 1 & 2 & 5.56 & 0 & 1 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 0 & 0 & 0 \\ \hline
\textit{Microcystis } & 151 & 118 & 112 & 65.97 & 105 & 95 & 56.66 & 87 & 37 & 44.96 & 83 & 35 & 46.10 & 94 & 31 & 51.61 \\ \hline
\textit{Gomphonema} & 19 & 19 & 5 & 1.00 & 19 & 6 & 100 & 18 & 1 & 94.74 & 16 & 2 & 83.90 & 18 & 2 & 89.50 \\ \hline
\textit{Sphaerocystis} & 11 & 10 & 0 & 90.91 & 10 & 3 & 89.26 & 8 & 1 & 71.72 & 10 & 0 & 90.91 & 10 & 0 & 100 \\ \hline
\textit{Cosmarium} & 5 & 3 & 2 & 60.00 & 2 & 0 & 40.00 & 1 & 0 & 20.00 & 3 & 0 & 60.00 & 3 & 1 & 75.00 \\ \hline
\textit{Cocconeis } & 2 & 2 & 1 & 1.00 & 1 & 0 & 50.00 & 2 & 0 & 1.00 & 2 & 0 & 100 & 2 & 0 & 100 \\ \hline
\textit{Tribonema} & 17 & 8 & 5 & 43.03 & 8 & 7 & 45.45 & 5 & 3 & 28.43 & 5 & 0 & 29.41 & 8 & 7 & 37.11 \\ \hline
\textit{Chlorella} & 34 & 20 & 24 & 52.82 & 18 & 11 & 49.98 & 15 & 4 & 42.98 & 16 & 3 & 46.05 & 15 & 1 & 48.33 \\ \hline
\textit{Tetraedron} & 7 & 7 & 3 & 1.00 & 7 & 5 & 1.00 & 6 & 0 & 85.71 & 7 & 0 & 100 & 7 & 1 & 46.67 \\ \hline
\textit{Ankistrodesmus} & 13 & 13 & 6 & 98.97 & 13 & 5 & 97.20 & 12 & 4 & 82.75 & 11 & 2 & 79.94 & 12 & 1 & 98.72 \\ \hline
\textit{Brachionus} & 25 & 25 & 19 & 93.18 & 21 & 13 & 76.38 & 13 & 4 & 49.10 & 17 & 2 & 66.40 & 25 & 9 & 83.97 \\ \hline
\textit{Chaenea} & 2 & 2 & 1 & 1.00 & 1 & 1 & 25.00 & 2 & 0 & 100 & 0 & 0 & 0 & 1 & 0 & 50.00 \\ \hline
\textit{Pediastrum} & 24 & 23 & 1 & 95.83 & 23 & 3 & 94.30 & 20 & 1 & 82.34 & 23 & 0 & 95.83 & 20 & 2 & 95.24 \\ \hline
\textit{Spirulina} & 14 & 8 & 6 & 57.14 & 8 & 7 & 42.42 & 3 & 0 & 21.43 & 4 & 0 & 28.57 & 9 & 2 & 45.79 \\ \hline
\textit{Actinastrum} & 40 & 29 & 14 & 67.48 & 23 & 14 & 51.58 & 18 & 0 & 45.00 & 23 & 0 & 57.50 & 29 & 3 & 77.45 \\ \hline
\textit{Navicula} & 16 & 14 & 7 & 85.83 & 12 & 5 & 71.88 & 10 & 3 & 62.50 & 12 & 6 & 64.91 & 15 & 6 & 72.67 \\ \hline
\textit{Scenedesmus} & 25 & 19 & 7 & 76.00 & 19 & 10 & 72.88 & 19 & 3 & 74.40 & 20 & 1 & 80.00 & 18 & 0 & 75.00 \\ \hline
\textit{Golenkinia} & 41 & 31 & 25 & 71.15 & 32 & 35 & 71.30 & 20 & 8 & 37.52 & 29 & 8 & 68.88 & 29 & 9 & 60.57 \\ \hline
\textit{Pinnularia} & 7 & 7 & 3 & 1.00 & 7 & 4 & 83.47 & 6 & 0 & 85.71 & 5 & 0 & 71.43 & 6 & 1 & 75.00 \\ \hline
\textit{Staurastrum} & 3 & 0 & 1 & 0 & 0 & 0 & 0 & 1 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 \\ \hline
\textit{Phormidium} & 234 & 194 & 127 & 74.23 & 188 & 151 & 64.49 & 147 & 73 & 48.84 & 4 & 1 & 43.19 & 142 & 32 & 52.93 \\ \hline
\textit{Fragilaria} & 11 & 10 & 3 & 90.91 & 10 & 2 & 90.91 & 6 & 3 & 42.73 & 21 & 20 & 90.91 & 10 & 3 & 90.91 \\ \hline
\textit{Anabaenopsis} & 5 & 5 & 1 & 1.00 & 5 & 2 & 94.29 & 2 & 0 & 40.00 & 2 & 0 & 80.00 & 9 & 2 & 81.82 \\ \hline
\textit{Coelosphaerium} & 42 & 21 & 12 & 44.58 & 11 & 2 & 24.43 & 1 & 0 & 2.38 & 2 & 2 & 44.19 & 6 & 3 & 18.30 \\ \hline
\textit{Crucigenia} & 4 & 2 & 1 & 50.00 & 2 & 1 & 41.67 & 0 & 0 & 0 & 11 & 3 & 50.00 & 2 & 0 & 66.67 \\ \hline
\textit{Achnanthes} & 3 & 3 & 0 & 1.00 & 2 & 1 & 66.67 & 0 & 0 & 0 & 2 & 2 & 50.00 & 3 & 0 & 100 \\ \hline
\textit{Synedra} & 34 & 27 & 21 & 72.42 & 23 & 29 & 55.90 & 16 & 14 & 30.92 & 11 & 3 & 31.00 & 15 & 1 & 34.09 \\ \hline
\textit{Ceratium} & 4 & 2 & 0 & 50.00 & 2 & 2 & 50.00 & 0 & 0 & 0 & 1 & 1 & 12.50 & 1 & 1 & 33.33 \\ \hline
\textit{Pompholyx} & 9 & 8 & 5 & 87.65 & 7 & 5 & 70.83 & 2 & 0 & 22.22 & 8 & 4 & 86.67 & 8 & 3 & 85.28 \\ \hline
\textit{Merismopedia} & 6 & 6 & 1 & 1.00 & 4 & 1 & 63.33 & 5 & 0 & 83.33 & 6 & 1 & 97.62 & 5 & 0 & 71.43 \\ \hline
\textit{Spirogyra} & 25 & 21 & 9 & 81.31 & 19 & 16 & 70.23 & 12 & 16 & 28.14 & 20 & 4 & 78.18 & 21 & 2 & 73.50 \\ \hline
\textit{Coelastrum} & 6 & 4 & 0 & 66.67 & 4 & 1 & 66.67 & 4 & 0 & 66.67 & 4 & 1 & 66.67 & 3 & 0 & 42.86 \\ \hline
\textit{Raphidiopsis} & 2 & 2 & 1 & 83.33 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 1 & 50.00 & 0 & 0 & 0 \\ \hline
\textit{Gomphosphaeria} & 14 & 14 & 6 & 98.74 & 10 & 2 & 70.78 & 5 & 8 & 16.23 & 13 & 2 & 90.81 & 11 & 3 & 100 \\ \hline
\textit{Euglena} & 16 & 16 & 1 & 1.00 & 15 & 6 & 92.97 & 14 & 4 & 80.28 & 15 & 1 & 92.97 & 12 & 1 & 73.56 \\ \hline
\textit{Euchlanis} & 2 & 2 & 0 & 1.00 & 1 & 0 & 50.00 & 0 & 0 & 0 & 1 & 0 & 50.00 & 0 & 0 & 0 \\ \hline
\textit{Keratella} & 13 & 13 & 4 & 98.19 & 7 & 4 & 51.54 & 3 & 1 & 23.08 & 9 & 1 & 69.23 & 9 & 1 & 69.23 \\ \hline
\textit{diversicornis} & 18 & 17 & 4 & 94.44 & 17 & 17 & 81.71 & 13 & 14 & 40.30 & 14 & 3 & 72.47 & 19 & 11 & 92.78 \\ \hline
\textit{Surirella} & 6 & 4 & 3 & 66.67 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & 0 & 16.67 & 1 & 1 & 16.67 \\ \hline
\textit{Characium} & 5 & 4 & 2 & 76.00 & 3 & 0 & 60.00 & 0 & 0 & 0 & 3 & 1 & 60.00 & 0 & 0 & 0 \\ \hline
\textit{unknown} & 1429 & 1049 & 1585 & 56.49 & 1068 & 1126 & 58.17 & 610 & 177 & 38.15 & 371 & 66 & 24.17 & 622 & 146 & 36.92 \\ \hline
\end{tabular}}
\end{table*}
\begin{figure*}
\centering
\includegraphics[scale=.50]{image/tf.pdf}
\caption{\centering{Five models predict the number of EMDS-7 positive and negative samples.}}\label{Fig:tf}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[scale=.50]{image/map.pdf}
\caption{\centering{Each category of EM object detection prediction AP value in EMDS--7}}\label{Fig:ap}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[scale=.50]{image/model.pdf}
\caption{\centering{Five object detection model prediction results in EMDS-7}}\label{Fig:model}
\end{figure*}
\section{Conclusion and Future Work}
EMDS-7 is an object detection data set containing 42 types of EMs, which contains the original image of EMs and object label data for corresponding EMs. EMDS-7 labeled 15342 EMs. At the same time, we further add some deep learning object detection experiments to the EMDS-7 database to prove the effectiveness.
During the object detection process, we divide the data set according to 6: 2: 2 for train, validation and test sets. We use five different deep learning object detection methods to test EMDS-7 and use multiple evaluation indices to evaluate the prediction results.
In the future, we will enlarge the category of EMs to increase the number of images of each EM. Make each class data balanced and sufficient. We hope to use the EMDS-7 database to achieve more features in the future. Meanwhile
\section*{Acknowledgements}
This work is supported by the ``National Natural Science Foundation of China'' (No.61806047)
and the ``Fundamental Research Funds for the Central Universities'' (No. N2019003).
We thank Miss Zixian Li and Mr. Guoxian Li for their important discussion.
\section*{Declaration of competing interest}
The authors declare that they have no conflict of interest.
\bibliographystyle{unsrt}
|
\section{Conclusion}
The paper presented a semi-supervised learning approach for predicting abnormal events from multiple multivariate time-series data. The proposed approach learns separable embeddings by leveraging labeled samples to learn cluster centers in a contrastive way and enforces unlabeled samples to follow similar clustering distribution to the labeled samples in a semi-supervised manner. We plan to explore methods for further detecting the event root cause, for example, determining the devices or interfaces and their activities leading to an event. Another direction is to take advantage of other networks in a transfer learning way when the labeled data is scarce.
\section{Acknowledgement}
This material is based upon work supported in part by the NTT Global Networks and NVIDIA Corporation.
\setlength\bibitemsep{-0.1\itemsep}
\printbibliography
\end{document}
\section{Metrics}
We reported the precision, recall, and F1 score for label 1. The functions are described as follows:
\begin{displaymath}
\text{precision}(\text{label}=1) = \frac{TP(\text{label}=1)}{TP(\text{label}=1)+FP(\text{label}=1)}
\end{displaymath}
\vspace{-.2em}
\begin{displaymath}
\text{recall}(\text{label}=1) = \frac{TP(\text{label}=1)}{TP(\text{label}=1) + FN(\text{label}=1)}
\end{displaymath}
\vspace{-.2em}
\begin{displaymath}
\text{F1} = 2 \times \frac{\text{precision}(\text{label}=1) \times \text{recall}(\text{label}=1)}{\text{precision}(\text{label}=1) + \text{recall}(\text{label}=1)}
\end{displaymath}
where $TP$ is the number of true positives (i.e., the number of correctly predicted label 1), $FP$ is the number of false positives (i.e., the number of incorrectly predicted label 1), and $FN$ is the number of false negatives (i.e., the number of incorrectly predicted label 0). \\
\section{Model Performance}
Figure~\ref{figure: prec} and Figure~\ref{figure: rec} show the precision and recall comparison between baselines and our method. In general, our method achieved much better performance in precision and recall.
We also projected and plotted the embeddings in a 2D space using t-SNE with their labels to demonstrate our approach (see~Figure~\ref{figure: res}).
\begin{figure*}[htbp]
\centering
\includegraphics[width=\linewidth]{figures/prec.jpg}
\caption{The Precision for baselines and our method (red).}
\label{figure: prec}
\end{figure*}
\begin{figure*}[htbp]
\centering
\includegraphics[width=\linewidth]{figures/rec.jpg}
\caption{The Recall for baselines and our method (red).}
\label{figure: rec}
\end{figure*}
\begin{table}[htpb]
\centering
\renewcommand\arraystretch{1.1}
\caption{Precision and Recall between using different loss functions}
\label{table: prerec}
\begin{tabular}{ >{\centering\arraybackslash}p{2cm} || >{\raggedleft\arraybackslash}p{1.4cm} | >{\raggedleft\arraybackslash}p{1.4cm} | >{\raggedleft\arraybackslash}p{1.4cm} | >{\raggedleft\arraybackslash}p{1.4cm} | >{\raggedleft\arraybackslash}p{1.4cm} | >{\raggedleft\arraybackslash}p{1.4cm} }
\hline
& \multicolumn{3}{c|}{precision} & \multicolumn{3}{c}{recall} \\ \hline
Network ID & only-CC & only-KL & Ours & only-CC & only-KL & Ours \\ \hline \hline
1NTT18927 & 0.864 & 0.735 & 0.898 & 0.76 & 0.7 & 0.76 \\
1NTT19535 & 0.913 & 0.876 & 0.893 & 0.724 & 0.692 & 0.862 \\
1NTT67246 & 0.946 & 0.923 & 0.963 & 0.96 & 0.857& 0.929\\
1NTT19478 & 0.846 & 0.897 & 0.955 & 0.733 & 0.693& 0.7\\
143XB8CT & 0.667 & 0.621 & 0.674 & 0.353 & 0.373& 0.4\\
1NTT30990 & 0.267 & 0.254 & 0.294 & 0.267 & 0.333& 0.611\\
1NTT48337 & 0.682 & 0.629 & 0.691 & 0.833 & 0.722 & 0.829\\
1NTT19213 & 0.333 & 0.303 & 0.45 & 0.576 & 0.647 & 0.789\\
1NTT22885 & 0.875 & 0.81 & 0.85 & 0.73. & 0.895 & 0.949\\
1NTT18635 & 0.947 & 0.95 & 0.978 & 0.923 & 0.974 & 0.982\\ \hline
\end{tabular}
\end{table}
\begin{table}[htpb]
\centering
\renewcommand\arraystretch{1.1}
\caption{Precision, Recall, F1 score between using different graph construction methods}
\label{table: prerecf1}
\begin{tabular}{ >{\centering\arraybackslash}p{2cm} || >{\raggedleft\arraybackslash}p{1.9cm} | >{\raggedleft\arraybackslash}p{1.9cm} | >{\raggedleft\arraybackslash}p{1.9cm} | >{\raggedleft\arraybackslash}p{1.9cm} | >{\raggedleft\arraybackslash}p{1.9cm} | >{\raggedleft\arraybackslash}p{1.9cm} }
\hline
& \multicolumn{2}{c|}{precision} & \multicolumn{2}{c|}{recall} & \multicolumn{2}{c}{F1} \\ \hline
Network ID & Corr-Graph & Ours & Corr-Graph & Ours & Corr-Graph & Ours \\ \hline \hline
1NTT18927 & 0.714 & 0.898 & 0.746 & 0.76 & 0.73 & 0.823 \\
1NTT19535 & 0.775 & 0.893 & 0.701 & 0.862 & 0.736 & 0.877\\
1NTT67246 & 0.85 & 0.963 & 0.769& 0.929 & 0.807 & 0.946\\
1NTT19478 & 0.927 & 0.955 & 0.577& 0.7 & 0.711 & 0.808\\
143XB8CT & 0.495 & 0.674 & 0.366& 0.4 & 0.421 & 0.502\\
1NTT30990 & 0.184 & 0.294 & 0.477& 0.611 & 0.266 & 0.397\\
1NTT48337 & 0.47 & 0.691 & 0.614 & 0.829 & 0.532 & 0.754\\
1NTT19213 & 0.355 & 0.45 & 0.592 & 0.789 & 0.444 & 0.573\\
1NTT22885 & 0.553 & 0.85 & 0.791 & 0.949 & 0.651 & 0.897\\
1NTT18635 & 0.617 & 0.978 & 0.86 & 0.982 & 0.719 & 0.980\\ \hline
\end{tabular}
\end{table}
\begin{figure*}[htbp]
\centering
\includegraphics[width=0.95\linewidth]{figures/res1.jpg}
\includegraphics[width=0.95\linewidth]{figures/res2.jpg}
\caption{Examples of learned embedding space in 2D for different networks.}
\label{figure: res}
\end{figure*}
\end{document}
\section{Introduction}
Nowadays, many network management companies (e.g., Cisco, NTT Global Networks) collect large volumes of data logs to measure and monitor the status (e.g., healthy, overloaded, power outage) of devices (e.g., routers, switches) or interfaces (e.g., local area networks (LAN), wide area networks (WAN)) in a network. Figure~\ref{figure: network_deff} shows an example network. A network usually consists of multiple devices, each with some interfaces, e.g., an SDWAN device has LAN and WAN interfaces. Each interface is measured by several metrics or attributes, i.e., the number of in/out octets and in/out ucast packets, that describe network operations or activities as multivariate time series.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.95\linewidth]{figures/network_deff.jpg}
\caption{An example network contains two SDWAN devices, and the devices have some LAN and WAN interfaces, each measured by four attributes.}
\label{figure: network_deff}
\end{figure}
Efficiently and accurately predicting abnormal events from network data logs is important, which helps continuously monitor network states and raise alerts for potential incidents on time for network engineers to investigate and resolve. Traditional monitoring systems typically trigger a real-time notification when the network is experiencing some abnormal event (e.g., packet loss, intrusion) using pre-defined rules but cannot provide an early warning on a potential incident. Also, a network typically contain hundreds if not thousands of heterogeneous devices from different vendors, which can have diverse criteria (e.g., some rules and thresholds) indicating varying levels of abnormal status, making it challenging to build an early warning system and enable the system handling all kinds of networks.
Advanced data-driven methods on analyzing \textit{multiple multivariate time series} (e.g., machine learning approaches) can automatically learn and extract patterns from network operations that lead to a potential abnormal event, which has the potential to handle heterogeneous networks, devices, and interfaces. One challenge is how to effectively capture temporal dependencies across time and relationships between time series that can jointly represent network activities for predicting an abnormal event. Some existing approaches overlook the relationships between time series, e.g., univariate time-series anomaly detection algorithms~\cite{ren2019time}, fusing multiple time series directly as features~\cite{malhotra2016lstm, li2019mad}. In terms of modeling temporal dependencies, many approaches use recurrent neural networks (RNNs)~\cite{malhotra2016lstm, hundman2018detecting} yet they are often time-consuming and require large memory for computation. Transformers, oriented from natural language processing, become popular in time-series analysis problems~\cite{cai2020traffic}, but they have similar issues.
In addition, some labeled data might be available, but the labeling criteria could vary significantly among devices and interfaces, e.g., an abnormal event can last for several minutes, and sometimes only the beginning time point is marked as anomalous. Unsupervised approaches, including reconstruction-based methods~\cite{park2018multimodal, su2019robust}, forecasting-based methods~\cite{laptev2015generic, hundman2018detecting}, and hybrid methods~\cite{zhao2020multivariate}, identify infrequent behaviors in multivariate time series as an anomaly. However, this assumption is not always true, e.g., long-term peak usage of memory might be rare but normal. Supervised models~\cite{liu2015opprentice, shipmon2017time} can leverage labeled data to guide the process of learning the time-series patterns leading to an abnormal event, but they usually require sufficient labeled samples. Also, the number of labeled abnormal samples is usually much smaller compared to the normal samples. Therefore, limited and imbalanced labels make it challenging to capture the patterns leading to abnormal activities.
This paper proposes a semi-supervised learning approach that learns representative features for network activities and use the features to predict abnormal events (e.g., packet loss, intrusion).
Specifically, our approach first takes a given network and constructs a graph representation where each node contains some attributes of an interface (e.g., packet volume) at a time point. (There could be some dependencies (e.g., data transmission) between interfaces (nodes) via the linkage.) Considering a sequence of graphs (i.e., network data logs over a time period) leading to some abnormal event, our approach aims to generate an embedding to represent network behaviors and capture the unique pattern of this graph sequence. The embedding should contain effective temporal information by capturing temporal dependencies across time points (i.e., temporal patterns). The embedding also includes relevant information from neighboring nodes, capturing data communication patterns. Therefore, our approach jointly captures temporal dependencies across time points and spatial dependencies between nodes for generating an embedding of the input. Then these learned embeddings are refined to form two separable clusters representing normal and abnormal network behaviors. Our approach uses labeled samples to explicitly learn the cluster centers in a contrastive way. However, due to imbalanced and limited labeled samples, the learnt centers might not be applicable to all samples and hence poor embedding quality. Our approach updates the embeddings of all samples towards a high confidence clustering distribution to enforce cluster cleanness with a semi-supervised learning strategy. The goal is to improve separability between normal and abnormal network behaviors. Finally, our approach utilizes a support vector classifier on these embeddings to predict if the given graph sequence of network operations will lead to an abnormal event.
In sum, the main contribution of this paper is a semi-supervised learning approach that learns separable embeddings that jointly captures temporal dependencies between time points and relationships between time series for predicting abnormal events. The proposed approach can efficiently capture complex dependencies in multiple multivariate time-series data and learn effective features from imbalanced and limited labels, enabling accurate prediction of abnormal events. We demonstrate the proposed network architecture for abnormal event prediction on large real-world network operation time-series data.
\section{Methodology}
\subsection{Problem Definition}
Given a network containing multiple interfaces with unknown connectivity, our system first constructs a fully-connected graph, $\boldsymbol{G}=(\boldsymbol{V}, \boldsymbol{E})$, where each node $i \in \boldsymbol{V}$ represents an interface ($|\boldsymbol{V}|=n$) and the connection between any pair of nodes is initialized as one, meaning there can be data communication between interfaces. Let $\boldsymbol{G}^{(t)} \in \mathbb{R}^{n \times F}$ be the graph for network operation measurements at time~$t$, where~$n$ is the number of nodes and~$F$ is the number of attributes describing each node. Suppose the input is a sequence of graphs of length $T$, i.e., graphs from $t-T+1$ to $t$, our approach aims to learn a function $f$ that maps $T$ historical input signals to a binary signal $y \in \{0, 1\}$ indicating if there will be an abnormal event in the network at $t+1$.
\begin{displaymath}
[\boldsymbol{G}^{(t-T+1)}, \cdots, \boldsymbol{G}^{(t)}]\xrightarrow{\emph{f}} y
\end{displaymath}
Figure~\ref{figure: whole} shows the proposed network architecture for learning feature embeddings from the graph input. Specifically, our approach first builds an GatedConv-NGAT module to generate an initial feature embedding by jointly capturing dependencies among graphs (across time points in the sequence) and between nodes (between individual time series in a graph) (Section~\ref{sec: stconvngat}). Then our approach jointly refines the embeddings and cluster centers representing normal and abnormal activities by: (1) leveraging \textit{labeled} normal and abnormal samples to form clusters in a contrastive learning way; (2) improving the separability of clusters using both \textit{labeled} and \textit{unlabeled} samples in a semi-supervised manner (Section~\ref{sec: cc}). Finally, our approach trains a support vector classifier with these learned embeddings for the prediction task (Section~\ref{sec: svc}).
\begin{figure}[htbp]
\centering
\includegraphics[width=\linewidth]{figures/wholee.jpg}
\vspace{-.2in}
\caption{The training process of our system to generate feature representation for predicting abnormal events}
\vspace{-.2in}
\label{figure: whole}
\end{figure}
\subsection{GatedConv-NGAT Module.} \label{sec: stconvngat}
This section introduces the ST-Conv-NGAT module that generates embeddings from the input graph sequence (Figure~\ref{figure: stconvngat}). In general, the module contains temporal gated convolutional layers to capture temporal dependencies across time and a normalized graph attention layer to capture dependencies between nodes. The details of each sub-module are described as follows.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.95\linewidth]{figures/stconvngat.jpg}
\vspace{-.1in}
\caption{GatedConv-NGAT Module}
\vspace{-.2in}
\label{figure: stconvngat}
\end{figure}
\textbf{GatedConv layer for extracting temporal relationships.} Recurrent neural networks (RNNs), e.g., long short-term memory (LSTM), have been widely used for capturing temporal dependencies in time-series data, but these models typically suffer from complex gating mechanisms, time-consuming computation, and large memory requirement since every step relies on the previous steps. In comparison, convolutional neural networks (CNNs) have the advantages of fast training and parallel computation~\cite{gehring2017convolutional} that can process large network data logs more efficiently. Also, CNNs explicitly capture local features within a temporal neighborhood, and one or more local features might cause an abnormal event, instead of global features from the entire sequence.
Our approach applies gated-conv layers~\cite{gehring2017convolutional} with multiple kernel sizes consisting of 1D convolution operations on the time axis to extract local temporal features for each node. The input to the 1D convolution operation is a sequence of length $T$ with a channel size $F$ at each node $\in \mathbb{R}^{T \times F}$. With a kernel size~$K$ and a hidden size~$C$, the convolution kernel maps the input to the output $[\mathbf{m}\ \mathbf{n}] \in \mathbb{R}^{(T-K+1) \times (2C)}$, where $\mathbf{m}$, $\mathbf{n}$ split the output into halves with the same hidden size~$h$. The gated-conv layer leverages gated linear units (GLU)~\cite{dauphin2017language} as a non-linear gating mechanism over $\mathbf{m}$ and $\mathbf{n}$:
\begin{equation}
\text{GatedConv}(\mathbf{v}) = \mathbf{m} \odot \sigma(\mathbf{n})\ \in \mathbb{R}^{(T-K+1) \times (C)}
\end{equation}
where $\mathbf{m}$ and $\mathbf{n}$ are the input to the gates of GLU, $\odot$ denotes the element-wise product, $\sigma$ is the sigmoid function, and $\sigma(\mathbf{n})$ controls which information in $\mathbf{m}$ are relevant for generating temporal features.
\textbf{Normalized graph attention layer for extracting relationships between nodes.} After the first Gated-Conv layer, each node $i$ is represented by $(T-K+1)$ feature vectors of size $C$, where each vector contains the extracted temporal features for the corresponding sub-sequence. In addition to temporal dependencies, combining information from other nodes is also essential for determining an abnormal event. Thus, our approach introduces a normalized graph attention layer to capture relationships between nodes.
Since the initial graph topology is unknown, many graph convolution strategies, e.g., chebyshev spectral graph convolutional operator~\cite{defferrard2016convolutional}, might not be suitable because they require the graph connectivity as the input. Our approach leverages a graph attention (GAT) layer~\cite{velivckovic2017graph} to model the spatial relationships between nodes. A GAT layer computes the representation for each node:
\vspace{-.1in}
\begin{equation}
\text{GAT}(i) = \sigma \Big ( \sum_{k\in \mathcal{N}_i}\alpha_{ik}\mathbf{v_k} \Big )
\end{equation}
where node $i$ is the target node, $\mathbf{v_i} \in \mathbb{R}^{C}$ is the hidden feature vector of node $i$, $\mathcal{N}_i$ is the neighborhood set of node $i$, $\sigma$ is the sigmoid function, $\alpha_{ik}$ is the attention score for node $k$ to node $i$. The attention score $\alpha_{ik}$ is defined by:
\vspace{-.1in}
\begin{equation}
\alpha_{ik}=\frac{\text{exp}(\sigma(\mathbf{w}^T \cdot (\mathbf{v_i}\ \Vert\ \mathbf{v_k}))}{\sum_{j \in \mathcal{N}_i} \text{exp}(\sigma(\mathbf{w}^T \cdot (\mathbf{v_i}\ \Vert\ \mathbf{v_j}))))}
\end{equation}
where the operation $\Vert$ is to concatenate two vectors, $\mathbf{w} \in \mathbb{R}^{2C}$ is a learnable weight vector, $\mathcal{N}_i$ is the neighborhood set of node $i$, $\sigma$ is the LeakyReLU function.
Our approach then adapts graph normalization (GraphNorm)~\cite{cai2021graphnorm} to the output of the GAT layer, which makes the optimization more efficient. GraphNorm normalizes every feature dimension for individual graphs with a learnable shift, which significantly boosts the training upon InstanceNorm~\cite{ulyanov2016instance} (also does graph-targeted normalization):
\vspace{-.1in}
\begin{equation}
\text{GraphNorm}(h_{i, j})=\gamma_j \cdot \frac{h_{i, j} - \alpha_j \mu_j}{\sigma_j} + \beta_j
\end{equation}
\begin{displaymath}
\mu_j=\frac{\sum_{k=1}^n h_{k, j}}{n}, \ \sigma_j=\frac{\sum_{k=1}^n (h_{k, j} - \alpha_j \mu_j)}{n}
\end{displaymath}
where $h_{i, j}$ is the value of the $j$th feature dimension of $\text{GAT}(i)$ (the output of GAT layer for node $i$), $n$ is the number of nodes in the graph, $\alpha_j$ is the learnable parameter for the feature dimension $j$, which controls how much information to keep in the mean. $\gamma_j$ and $\beta_j$ are the affine parameters. \vspace{2pt}
\textbf{Putting everything together to learn embeddings.} Inspired by~\cite{yu2017spatio}, our approach adds another temporal gated-conv layer after the NGAT layer to fuse the features of node interactions on the time axis. This ``sandwich'' structure (i.e., a spatial layer in the middle of two temporal layers) can achieve fast spatial-state propagation through temporal convolutions. After the second temporal gated-conv layer, our approach concatenates the output to generate the node embeddings. Then our approach uses a global pooling layer to add node embeddings across the node dimension to generate the graph embedding. In this way, given [$\boldsymbol{G}^{(t-T+1)}, \cdots, \boldsymbol{G}^{(t)}] \in \mathbb{R}^{T \times n \times F}$, the GatedConv-NGAT module generates the embedding $\boldsymbol{E}^{(t)}$ of length ${(T-2(K-1)) \times C}$, that jointly captures the temporal dependencies and relationship between nodes.
\subsection{Learning clustering structure.} \label{sec: cc}
Our approach assumes two clusters in the embedding space: normal and abnormal events. The goal is to learn a representation space to form these two separable clusters.
\textbf{Learning clustering structure on labeled samples using contrastive center loss.} The labeled samples can explicitly guide the process of learning clustering centers by enforcing the embeddings of normal samples to be close to the normal center and far away from the abnormal center, and vice versa. Our approach applies the contrastive-center loss~\cite{qi2017contrastive} that considers intra-class compactness and inter-class separability simultaneously. This loss function aims to penalize the contrastive values between (1) the distances of samples to their corresponding class center and (2) the distances of samples to their non-corresponding class centers,
\vspace{-.1in}
\begin{equation}
\mathcal{L}_{cc} = \frac{1}{2} \sum^m_{i=1} \frac{\lVert \mathbf{E_i} - \mathbf{c_{y_i}} \rVert^2_2}{\lVert \mathbf{E_i} - \mathbf{c_{j}} \rVert^2_2 + \epsilon } \ (j\neq y_i)
\label{eq: cc}
\end{equation}
where $L_{cc}$ denotes the contrastive-center loss; $m$ is the number of labeled samples in a min-batch; $\mathbf{E_i}$ is the embedding of the $i$-th training sample; $\mathbf{c_{y_i}}$ is the corresponding center of $\mathbf{E_i}$. Since there are two clusters, when $y_i$ is 0, $j$ is 1, and vice versa ($j\neq y_i$). $\epsilon$ is a constant for preventing the denominator equal to 0 and set to ${1}\mathrm{e}{-6}$ empirically.
\textbf{Learning clustering structure on all samples using a semi-supervised learning strategy.} Using the contrastive learning strategy improves the clustering cleanness for labeled samples; however, due to limited and imbalanced labeled samples, the learned centers are easily guided by the given labeled samples, which might fail in inferencing. To overcome the scarcity of labeled training samples and ensure the clusters are applicable to all samples, our approach takes the idea from~\cite{xie2016unsupervised} to minimize the distance between the cluster assignment and some target cluster distribution. For labeled samples, the target distribution can be directly derived from the label, while for the unlabeled samples, the target distribution is derived from high confidence predictions of the current cluster distribution from all samples. Formally, our approach calculates the probability of assigning the embedding $\mathbf{E_i}$ to the cluster $j$ with a center $\mu_j$ as
\vspace{-.1in}
\begin{equation}
q_{ij} = \frac{(1+\rVert \mathbf{E_i} - \mu_j \rVert ^ 2) ^ {-1}}{\sum_{j'}(1+\rVert \mathbf{E_i} - \mu_{j'} \rVert ^ 2) ^ {-1})}
\label{eq: Q}
\end{equation}
Following the work in~\cite{xie2016unsupervised}, our approach derives an auxiliary distribution made up of the high confidence assignments of the current distribution (referred as $Q$). The main idea is to emphasize data points assigned with high confidence to ensure cluster purity. Equation~\ref{eq: P} defines the computation of auxiliary distribution (referred as $P$), where the second power of $q_{ij}$ places more weight on the instances near the center and the division of $\sum_{i} q_{ij}$ normalizes based on the cluster size, making the model robust to biased classes. Note that for labeled samples, our approach replaces the true distribution with the auxiliary distribution and hence enables the training process in a semi-supervised manner.
\vspace{-.1in}
\begin{equation}
p_{ij} = \frac{q_{ij}^2 / \sum_{i'} q_{i'j}}{\sum_{j'} (q_{ij'}^2 / \sum_{i'} q_{i'j'})}
\label{eq: P}
\end{equation}
To measure the distance between the distributions $P$ and $Q$, our approach computes the KL divergence to measure their distance. The clustering oriented loss is defined as,
\vspace{-.1in}
\begin{equation}
\mathcal{L_{KL}} = KL(P||Q) = \sum_i \sum_j p_{ij} \log \frac{p_{ij}}{q_{ij}}
\label{eq: Q}
\end{equation}
Therefore, the total loss is the sum of contrastive center loss and KL divergence loss with a hyper-parameter $\lambda$,
\vspace{-.1in}
\begin{equation}
\mathcal{L} = \mathcal{L}_{cc} + \lambda \times \mathcal{L}_{KL}
\label{eq: Q}
\end{equation}
\subsection{Predicting abnormal events from learned embeddings using SVC.}\label{sec: svc} Using the methods described so far in this section, our model generates separable embeddings (normal vs. abnormal). Then we take the embeddings and labels to train a support vector classifier to predict normal and abnormal events. Our approach does not train in an end-to-end manner, e.g., using a fully-connected layer for classification in the model because support vector classifier usually works better if the input are already separable in some space. Figure~\ref{figure: inference} shows the prediction process. Given a sequence of graphs representing network operations, the GatedConv-NGAT module generates the embedding and the support vector classifier takes the embedding to predict if the input would lead to an abnormal event.
\begin{figure}[htbp]
\centering
\vspace{-.1in}
\includegraphics[width=\linewidth]{figures/inference.jpg}
\caption{Event prediction}
\label{figure: inference}
\end{figure}
\section{Experiments and Results}
We implemented the proposed architecture with Python~3.8, PyTorch~1.8, and PyTorch Geometric framework~\cite{Fey/Lenssen/2019}. We conducted the experiments in a Docker container deployed on a GPU server and trained the model with one physical core and 16GB memory.
\subsection{Datasets.}
We used the network operation data from NTT-GN for the experiments. The dataset contains more than 150 million data logs from network interfaces (e.g., WAN routers and LAN switches) that NTT-GN collects, monitors, and manages. The data logs we used cover the time range from 2020-05-12 to 2020-06-12, with a time interval of 5 minutes. Figure~\ref{figure: network_deff} shows an example of the data structure. In this work, we assume a practical case that there could be dependencies between any interface pairs in a network (e.g., data transmissions) (Figure~\ref{figure: graph_construction}), and the network topology in not available. The graph representation can be modified if prior knowledge of the network topology is available.
In addition to the network operation data, NTT-GN also provides the packet loss event dataset that reports the ticket opening time at the device level. We assume that events are caused by some abnormal network activities in the network and can be captured by looking at the dynamics of the entire network. Figure~\ref{figure: label_construction} shows an example of assigning a label to a period of network activities (e.g., 6 hours, a sequence of 72 graphs). If there is an event reported at $t'$ and the closest graph before that is $\mathbf{G}^{(t)}$, we assigned label 1 (color blue) to the graph sequence $\mathbf{G}^{(t-71)} : \mathbf{G}^{(t)}$, denoting that these graphs contains some abnormal network activities directly leading to an event. Since we do not know how long the event lasts, the graph sequences overlapped with an event within 6 hours before and after (e.g., color yellow) were marked as ``unknown'' (i.e., unlabeled data). These graph sequences also help learn feature embeddings in our semi-supervised setting. The rest graph sequences that do not overlap with any event within 6 hours were labeled as 0, representing normal network activities.
\begin{figure}[htbp]
\centering
\subfigure[The graph representing the example network]{
\begin{minipage}[t]{\columnwidth}
\centering
\includegraphics[width=0.6\linewidth]{figures/graph_construction.jpg}
\label{figure: graph_construction}
\end{minipage}
}
\subfigure[An example of graph label construction]{
\begin{minipage}[t]{\columnwidth}
\centering
\includegraphics[width=\linewidth]{figures/label_construction.jpg}
\label{figure: label_construction}
\end{minipage}
}
\vspace{-0.2in}
\caption{Examples of sample construction}
\label{figure: data_construction}
\end{figure}
\subsection{Experiment Settings.} To evaluate the proposed approach, we selected the available networks with at least 50 events. We split the samples into training, validation, and testing sets, ensuring each set contains both label 0 and label 1. Table~\ref{table: dataset summary} shows the summary of the dataset. We observe that the number of samples labeled with 0 is much larger than label 1. Also, the ratio of labeled samples to unlabeled samples is around 0.2. We repeated the random splitting process three times for cross-validation and took the average as the results.
We set the length of the input graph sequence in a sample as 72. The batch size was 32, each containing both labeled and unlabeled samples. We set the temporal gated convolutional layers as 32 hidden states and the kernel size of 1D Conv as 12. The size of the output feature embeddings was 256. The initial learning rate was 0.001, with early stopping on the validation dataset. The hyper-parameters were chosen by grid search on the validation set, and we reported the results with the set of hyper-parameters yielding the best average performance, that is $\lambda=0.1$.
\begin{table}[h]
\centering
\caption{Summary of the datasets}
\label{table: dataset summary}
\begin{tabular}{ c | c | c | c }
\hline
Network ID & \#Nodes & \#Events & \# 0: \# 1 \\ \hline
1NTT18927 & 53 & 125 & 0.96:0.04 \\ \hline
1NTT19535 & 84 & 142 & 0.89:0.11 \\ \hline
1NTT67246 & 36 & 134 & 0.91:0.09 \\ \hline
1NTT19478 & 120 & 147 & 0.88:0.12 \\ \hline
143XB8CT & 97 & 78 & 0.98:0.02 \\ \hline
1NTT30990 & 20 & 71 & 0.98:0.02 \\ \hline
1NTT48337 & 89 & 83 & 0.95:0.05 \\ \hline
1NTT19213 & 119 & 78 & 0.96:0.04 \\ \hline
1NTT22885 & 91 & 91 & 0.95:0.05 \\ \hline
1NTT18635 & 143 & 189 & 0.73:0.27 \\ \hline
\end{tabular}
\end{table}
\noindent \textbf{Metrics.} We reported the precision, recall, and F1 score for label 1 since our goal is to predict events. The higher the values the better the performance. We put the detailed functions in the supplementary.
\noindent \textbf{Baselines.} We compared our proposed model to the following state-of-the-art baselines:
\begin{itemize}
\setlength\itemsep{-.1em}
\item Support Vector Classifier (SVC): We convert the problem to a binary classification task for SVC. We flatten an input graph sequence in time and node dimensions to be a large vector for predicting event or non-event.
\item Graph Attention Encoder (GAT): We concatenate features along the temporal dimension for each node and replace the encoder with only the graph attention module to capture interactions between time series. This method aims to demonstrate the necessity of capturing temporal dependencies.
\item Graph Attention LSTM (GAT-LSTM)~\cite{wu2018graph}: Instead of using the ST-Conv module, we apply LSTM in the encoder to capture interactions across time points. The rest architecture remains the same as our proposed approach.
\item MTAD-GAT~\cite{zhao2020multivariate}: This is an unsupervised anomaly detection approach that explicitly captures relationships between multiple features and across time points. We modify the approach so that the node value has more than one dimension.
\item DeepSAD~\cite{ruff2019deep}: This is a semi-supervised learning approach for anomaly detection that aims to learn a clustering center to represent normal cases. The original domain is computer vision. To extend the model on our dataset, we apply the ST-Conv-NGAT module as the encoder.
\item We select the following variations of the proposed model for ablation study: 1. ST-Conv-NGAT + only contrastive center loss (only-CC), which would be trained only with labeled samples; 2. ST-Conv-NGAT + only KL divergence loss (only-KL); 3. Constructing a correlation graph as the input~\cite{deng2021graph}, which computes pairwise Pearson correlation between nodes and assumes two nodes are connected only if the correlation value is above the mean plus one std.
\end{itemize}
\subsection{Model Performance.}
In general, our approach outperformed baseline methods in all networks (see F1 scores in Figure~\ref{figure: f1} and precision \& recall in the supplementary). The F1 scores in most networks are above 0.7. Even when the ratio between the number of label 0 and label 1 is high (e.g., ``143XB8CT'' and ``1NTT30990''), our model could still achieve far better results than the baseline models. Specifically, our approach achieved more than 50\% improvement in F1 scores compared to SVC. This is because SVC simply takes the input graph sequence as a flattened vector and does not explicitly capture relationships between time series or across time points.
Compared to using other encoding methods (GAT and GAT-LSTM) to generate representations for prediction, our approach with the ST-Conv-NGAT module as the encoder showed excellent advantage. First, GAT does not explicitly consider temporal dependencies but fuses features on the time dimension together and expects the model to extract useful information from these values for prediction directly. However, the temporal order of these values matters for event prediction, e.g., the position of a sudden increment might result in different network behaviors. Our approach leverages the 1D-CNN to summarize local temporal features, aiming to extract effective temporal patterns leading to an event. Second, although GAT-LSTM is commonly used for capturing spatiotemporal dependencies in multi-time-series data~\cite{wu2018graph}, its performance is poorer than our approach, which might be because GAT-LSTM is typically generating representations for the entire time series. However, in practice, only some subsequence(s) from the input graph sequence would cause an event. Some effective patterns (or features) at the early stage of the input sequence might be easily weakened or neglected during the sequential training of LSTM. Since the impacting subsequence(s) on the event are unknown (i.e., the leading time to an event could be varying), the sequence length (i.e., six hours in the experiment) was chosen to cover as many time points as possible. In comparison, the 1D-CNN in our approach focuses on manipulating temporal interactions on every subsequence, which empirically fulfills our assumption and achieves better performance on real datasets. Also, training GAT-LSTM for one epoch took more than 600 seconds while our approach required around 170 seconds under the same settings.
Furthermore, compared to existing work on anomaly detection, our approach achieved better performance than the unsupervised learning approach (MTAD-GAT), which demonstrates that assuming rare samples as anomalies might not be applicable in practice. Adding some labeled samples during training can guide the learning process to generate representative embeddings. Additionally, our approach outperformed the semi-supervised approach, DeepSAD. DeepSAD has been demonstrated to outperform some unsupervised approaches in the original paper. The loss function in DeepSAD enforces labeled normal samples to be close to the predefined center and abnormal samples to be as far as possible. For the unlabeled samples, DeepSAD has a strong assumption that the majority of data samples are normal and treats all unlabeled samples as normal in the loss. However, such an assumption might not hold in practice since one event can last for some period (i.e., the unlabeled data might contain abnormal network activities), which might confuse the model to learn effective representations for unlabeled samples and result in poor performance.
\begin{figure*}[htbp]
\centering
\includegraphics[width=0.95\linewidth]{figures/f1.jpg}
\caption{The F1 scores for baselines and our method (red). See precision and recall in the supplementary.}
\label{figure: f1}
\vspace{-.2in}
\end{figure*}
\vspace{-1em}
\subsection{Ablation Study.} \textbf{(1) Effect of Loss Functions.} Table~\ref{table: f1} shows that our approach achieved 12\% improvement in F1 scores on average compared to only using the contrastive center (CC) loss and 14\% improvement compared to only using the KL loss. The model with only the CC loss was trained with limited labeled samples, which worked well for the networks with large abnormal samples (e.g., more than 120 events in ``1NTT67246'' and ``1NTT18635''). However, the performance dropped in the cases with only a few abnormal samples. Our proposed approach trained on both labeled and unlabeled samples to force the learnt clustering structure was applicable to all data samples, resulting in better performance, e.g., 48\% improvement in F1 score for ``1NTT30990'' and 35\% for ``1NTT19213''. In the other case, the model with only KL loss was trained with all samples. Without the CC loss to explicitly optimize the centers, finding good centers was challenging even when the true clustering distribution could be estimated from labeled samples. \\
\vspace{-1em}
\noindent \textbf{(2) Effect of graph construction.} We compared the result of the model taking input graphs constructed based on the correlations between nodes. However, the performance is generally worse than taking the fully-connected graph as the input, which shows that the node correlations could not effectively describe the interactions between interfaces. Since the network connectivity is unknown, the proposed approach assumes that data communication (direct and indirect) could happen between any interface pairs and uses the graph attention module to automatically capture their relationships. The entire results are presented as Table 2 in the supplementary.
\begin{table}[htpb]
\centering
\caption{F1 scores between using different loss functions (see the full table in the supplementary)}
\label{table: f1}
\begin{tabular}{ >{\centering\arraybackslash}p{2cm} || >{\raggedleft\arraybackslash}p{1.4cm} | >{\raggedleft\arraybackslash}p{1.4cm} | >{\raggedleft\arraybackslash}p{1.4cm} }
Network ID & only-CC & only-KL & Ours \\ \hline \hline
1NTT18927 & 0.809 & 0.717 & \textbf{0.823}\\
1NTT19535 & 0.808 & 0.773 & \textbf{0.877}\\
1NTT67246 & \textbf{0.955} & 0.889 & 0.946\\
1NTT19478 & 0.786 & 0.782 & \textbf{0.808}\\
143XB8CT & 0.462 & 0.466 & \textbf{0.502}\\
1NTT30990 & 0.267 & 0.288 & \textbf{0.397}\\
1NTT48337 & 0.757 & 0.672 & 0.754\\
1NTT19213 & 0.422 & 0.413 & \textbf{0.573}\\
1NTT22885 & 0.800 & 0.850 & \textbf{0.897}\\
1NTT18635 & 0.935 & 0.962 & \textbf{0.980}\\ \hline
\end{tabular}
\end{table}
\section{Related Work}
There exists abundant literature on anomaly detection with time-series data. Deep learning approaches have shown promising results in capturing complex relationships between time series and across time points for extracting effective patterns that enable accurate anomaly prediction, including unsupervised~\cite{munir2018deepant,deng2021graph}, supervised~\cite{liu2015opprentice, shipmon2017time}, and semi-supervised approaches~\cite{vercruyssen2018semi}. For example,~\citet{hundman2018detecting} propose to use Long-Short-Term Memory (LSTM) to predict spacecraft telemetry and detect outliers in an unsupervised manner within each variable of a single multivariate time-series. However, the approach does not consider the dependencies between variables. Network operation data are typically multiple multivariate time-series data, which require careful considerations to to capture the interactions between variables and time series for event detection. RNNs and Transformers~\cite{cai2020traffic} have been widely used in modeling temporal dependencies in time-series data, but they typically suffer from time-consuming training process and large memory requirement~\cite{yu2017spatio}. \citet{zhao2020multivariate} propose an unsupervised anomaly detection approach on multivariate time-series data, with each variate/time series represented as one node in a graph. Their approach utilizes a feature attention layer and a temporal attention layer to explicitly capture the relationships between features and times respectively. Then the approach jointly optimizes a forecasting-based model and a reconstruction-based model to obtain better time-series representations. However, such unsupervised learning approaches usually make a strong assumption that infrequent behavior is anomalous and usually requires some scores for inference (e.g., see \cite{ruff2018deep}). Although there are some techniques, e.g., Peak Over Threshold (POT)~\cite{siffer2017anomaly}, that automatically determine the threshold during inference, the performance is still not competitive to supervised learning methods when labels are available (e.g., see~\cite{liu2015opprentice, shipmon2017time}). Supervised approaches can explicitly learn normal or abnormal patterns from labels yet requires a prohibitive large number of labeled examples (e.g., see~\cite{blazquez2021review}).
Semi-supervised learning approaches can overcome these limitations. For example, \citet{ruff2019deep} demonstrates that adding a few labeled samples in a semi-supervised manner would guide the model to achieve better performance than unsupervised learning methods for anomaly detection. \citet{xiao2020semi} propose a semi-supervised approach that uses an LSTM auto-encoder on both unlabeled data and labeled data with a reconstruction loss and adds prediction loss on the labeled data to guide the learning process. In comparison, our approach explicitly learns a separable embedding space for normal and abnormal samples instead of using a simple prediction loss (e.g., cross-entropy loss or focal loss for the imbalanced label situation). \citet{vercruyssen2018semi} propose a semi-supervised learning approach that first uses an unsupervised model, a clustering-based approach, to define typical normal behavior and identify anomalies. Then the approach actively adds labels from domain experts to guide the clustering approach in a semi-supervised manner. However, the performance would highly rely on the initialized clustering result. Our approach leverages the unlabeled samples to learn cluster centers in a self-learning way.
\section{Problem Specification.}In this paper, we consider the solution of the $N \times
N$ linear
system
\begin{equation} \label{e1.1}
A x = b
\end{equation}
where $A$ is large, sparse, symmetric, and positive definite. We consider
the direct solution of (\ref{e1.1}) by means of general sparse Gaussian
elimination. In such a procedure, we find a permutation matrix $P$, and
compute the decomposition
\[
P A P^{t} = L D L^{t}
\]
where $L$ is unit lower triangular and $D$ is diagonal.
\section{Design Considerations.}Several good ordering algorithms (nested dissection and
minimum degree)
are available for computing $P$ \cite{GEORGELIU}, \cite{ROSE72}.
Since our interest here does not
focus directly on the ordering, we assume for convenience that $P=I$,
or that $A$ has been preordered to reflect an appropriate choice of $P$.
Our purpose here is to examine the nonnumerical complexity of the
sparse elimination algorithm given in \cite{BANKSMITH}.
As was shown there, a general sparse elimination scheme based on the
bordering algorithm requires less storage for pointers and
row/column indices than more traditional implementations of general
sparse elimination. This is accomplished by exploiting the m-tree,
a particular spanning tree for the graph of the filled-in matrix.
\begin{theorem} The method was extended to three
dimensions. For the standard multigrid
coarsening
(in which, for a given grid, the next coarser grid has $1/8$
as many points), anisotropic problems require plane
relaxation to
obtain a good smoothing factor.\end{theorem}
Our purpose here is to examine the nonnumerical complexity of the
sparse elimination algorithm given in \cite{BANKSMITH}.
As was shown there, a general sparse elimination scheme based on the
bordering algorithm requires less storage for pointers and
row/column indices than more traditional implementations of general
sparse elimination. This is accomplished by exploiting the m-tree,
a particular spanning tree for the graph of the filled-in matrix.
Several good ordering algorithms (nested dissection and minimum degree)
are available for computing $P$ \cite{GEORGELIU}, \cite{ROSE72}.
Since our interest here does not
focus directly on the ordering, we assume for convenience that $P=I$,
or that $A$ has been preordered to reflect an appropriate choice of $P$.
\begin{proof} In this paper we consider two methods. The first method
is
basically the method considered with two differences:
first, we perform plane relaxation by a two-dimensional
multigrid method, and second, we use a slightly different
choice of
interpolation operator, which improves performance
for nearly singular problems. In the second method coarsening
is done by successively coarsening in each of the three
independent variables and then ignoring the intermediate
grids; this artifice simplifies coding considerably.
\end{proof}
Our purpose here is to examine the nonnumerical complexity of the
sparse elimination algorithm given in \cite{BANKSMITH}.
As was shown there, a general sparse elimination scheme based on the
bordering algorithm requires less storage for pointers and
row/column indices than more traditional implementations of general
sparse elimination. This is accomplished by exploiting the m-tree,
a particular spanning tree for the graph of the filled-in matrix.
\begin{Definition}{\rm We describe the two methods in \S 1.2. In \S\ 1.3. we
discuss
some remaining details.}
\end{Definition}
Our purpose here is to examine the nonnumerical complexity of the
sparse elimination algorithm given in \cite{BANKSMITH}.
As was shown there, a general sparse elimination scheme based on the
bordering algorithm requires less storage for pointers and
row/column indices than more traditional implementations of general
sparse elimination. This is accomplished by exploiting the m-tree,
a particular spanning tree for the graph of the filled-in matrix.
Several good ordering algorithms (nested dissection and minimum degree)
are available for computing $P$ \cite{GEORGELIU}, \cite{ROSE72}.
Since our interest here does not
focus directly on the ordering, we assume for convenience that $P=I$,
or that $A$ has been preordered to reflect an appropriate choice of $P$.
Our purpose here is to examine the nonnumerical complexity of the
sparse elimination algorithm given in \cite{BANKSMITH}.
As was shown there, a general sparse elimination scheme based on the
bordering algorithm requires less storage for pointers and
row/column indices than more traditional implementations of general
sparse elimination.
\begin{lemma} We discuss first the choice for $I_{k-1}^k$
which is a generalization. We assume that $G^{k-1}$ is
obtained
from $G^k$
by standard coarsening; that is, if $G^k$ is a tensor product
grid $G_{x}^k \times G_{y}^k \times G_{z}^k$,
$G^{k-1}=G_{x}^{k-1} \times G_{y}^{k-1} \times G_{z}^{k-1}$,
where $G_{x}^{k-1}$ is obtained by deleting every other grid
point of $G_x^k$ and similarly for $G_{y}^k$ and $G_{z}^k$.
\end{lemma}
To our knowledge, the m-tree previously has not been applied in this
fashion to the numerical factorization, but it has been used,
directly or indirectly, in several optimal order algorithms for
computing the fill-in during the symbolic factorization phase
[4] - [10], [5], [6]. In \S 1.3., we analyze the complexity of the old and new
approaches to the intersection problem for the special case of
an $n \times n$ grid ordered by nested dissection. The special
structure of this problem allows us to make exact estimates of
the complexity. To our knowledge, the m-tree previously has not been applied in this
fashion to the numerical factorization, but it has been used,
directly or indirectly, in several optimal order algorithms for
computing the fill-in during the symbolic factorization phase
[4] - [10], [5], [6].
In \S 1.2, we review the bordering algorithm, and introduce
the sorting and intersection problems that arise in the
sparse formulation of the algorithm.
In \S 1.3., we analyze the complexity of the old and new
approaches to the intersection problem for the special case of
an $n \times n$ grid ordered by nested dissection. The special
structure of this problem allows us to make exact estimates of
the complexity. To our knowledge, the m-tree previously has not been applied in this
fashion to the numerical factorization, but it has been used,
directly or indirectly, in several optimal order algorithms for
computing the fill-in during the symbolic factorization phase
[4] - [10], [5], [6].
For the old approach, we show that the
complexity of the intersection problem is $O(n^{3})$, the same
as the complexity of the numerical computations. For the
new approach, the complexity of the second part is reduced to
$O(n^{2} (\log n)^{2})$.
To our knowledge, the m-tree previously has not been applied in this
fashion to the numerical factorization, but it has been used,
directly or indirectly, in several optimal order algorithms for
computing the fill-in during the symbolic factorization phase
[4] - [10], [5], [6]. In \S 1.3., we analyze the complexity of the old and new
approaches to the intersection problem for the special case of
an $n \times n$ grid ordered by nested dissection. The special
structure of this problem allows us to make exact estimates of
the complexity. To our knowledge, the m-tree previously has not been applied in this
fashion to the numerical factorization, but it has been used,
directly or indirectly, in several optimal order algorithms for
computing the fill-in during the symbolic factorization phase
[4] - [10], [5], [6].
This is accomplished by exploiting the m-tree,
a particular spanning tree for the graph of the filled-in matrix.
To our knowledge, the m-tree previously has not been applied in this
fashion to the numerical factorization, but it has been used,
directly or indirectly, in several optimal order algorithms for
computing the fill-in during the symbolic factorization phase
\cite{EISENSTAT} - \cite{LIU2}, \cite{ROSE76}, \cite{SCHREIBER}.
\subsection{Robustness.}We do not
attempt to present an overview
here, but rather attempt to focus on those results that
are relevant to our particular algorithm.
This section assumes prior knowledge of the role of graph theory
in sparse Gaussian elimination; surveys of this role are
available in \cite{ROSE72} and \cite{GEORGELIU}. More general
discussions of elimination trees are given in
\cite{LAW} - \cite{LIU2}, \cite{SCHREIBER}.
Thus, at the $k$th stage, the bordering algorithm consists of
solving the lower triangular system
\begin{equation} \label{1.2}
L_{k-1}v = c
\end{equation}
and setting
\begin{eqnarray}
\ell &=& D^{-1}_{k-1}v , \\
\delta &=& \alpha - \ell^{t} v .
\end{eqnarray}
\begin{figure}
\vspace{14pc}
\caption{This is a figure 1.1.}
\end{figure}
\section{Robustness.} We do not
attempt to present an overview
here, but rather attempt to focus on those results that
are relevant to our particular algorithm.
\subsection{Versatility.}The special
structure of this problem allows us to make exact estimates of
the complexity. For the old approach, we show that the
complexity of the intersection problem is $O(n^{3})$, the same
as the complexity of the numerical computations
\cite{GEORGELIU}, \cite{ROSEWHITTEN}. For the
new approach, the complexity of the second part is reduced to
$O(n^{2} (\log n)^{2})$.
To our knowledge, the m-tree previously has not been applied in this
fashion to the numerical factorization, but it has been used,
directly or indirectly, in several optimal order algorithms for
computing the fill-in during the symbolic factorization phase
[4] - [10], [5], [6]. In \S 1.3., we analyze the complexity of the old and new
approaches to the intersection problem for the special case of
an $n \times n$ grid ordered by nested dissection. The special
structure of this problem allows us to make exact estimates of
the complexity. To our knowledge, the m-tree previously has not been applied in this
fashion to the numerical factorization, but it has been used,
directly or indirectly, in several optimal order algorithms for
computing the fill-in during the symbolic factorization phase
[4] - [10], [5], [6].
In \S 1.2, we review the bordering algorithm, and introduce
the sorting and intersection problems that arise in the
sparse formulation of the algorithm.
In \S 1.3., we analyze the complexity of the old and new
approaches to the intersection problem for the special case of
an $n \times n$ grid ordered by nested dissection. The special
structure of this problem allows us to make exact estimates of
the complexity. To our knowledge, the m-tree previously has not been applied in this
fashion to the numerical factorization, but it has been used,
directly or indirectly, in several optimal order algorithms for
computing the fill-in during the symbolic factorization phase
[4] - [10], [5], [6].
For the old approach, we show that the
complexity of the intersection problem is $O(n^{3})$, the same
as the complexity of the numerical computations. For the
new approach, the complexity of the second part is reduced to
$O(n^{2} (\log n)^{2})$.
To our knowledge, the m-tree previously has not been applied in this
fashion to the numerical factorization, but it has been used,
directly or indirectly, in several optimal order algorithms for
computing the fill-in during the symbolic factorization phase
[4] - [10], [5], [6]. In \S 1.3., we analyze the complexity of the old and new
approaches to the intersection problem for the special case of
an $n \times n$ grid ordered by nested dissection. The special
structure of this problem allows us to make exact estimates of
the complexity. To our knowledge, the m-tree previously has not been applied in this
fashion to the numerical factorization, but it has been used,
directly or indirectly, in several optimal order algorithms for
computing the fill-in during the symbolic factorization phase
[4] - [10], [5], [6].
This is accomplished by exploiting the m-tree,
a particular spanning tree for the graph of the filled-in matrix.
To our knowledge, the m-tree previously has not been applied in this
fashion to the numerical factorization, but it has been used,
directly or indirectly, in several optimal order algorithms for
computing the fill-in during the symbolic factorization phase
\cite{EISENSTAT} - \cite{LIU2}, \cite{ROSE76}, \cite{SCHREIBER}.
\section{Conclusion}
The paper presented a semi-supervised learning approach for predicting abnormal events from multiple multivariate time-series data. The proposed approach learns separable embeddings by leveraging labeled samples to learn cluster centers in a contrastive way and enforces unlabeled samples to follow similar clustering distribution to the labeled samples in a semi-supervised manner. We plan to explore methods for further detecting the event root cause, for example, determining the devices or interfaces and their activities leading to an event. Another direction is to take advantage of other networks in a transfer learning way when the labeled data is scarce.
\section{Acknowledgement}
This material is based upon work supported in part by the NTT Global Networks and NVIDIA Corporation.
\setlength\bibitemsep{-0.1\itemsep}
\printbibliography
\end{document}
\section{Problem Specification.}In this paper, we consider the solution of the $N \times
N$ linear
system
\begin{equation} \label{e1.1}
A x = b
\end{equation}
where $A$ is large, sparse, symmetric, and positive definite. We consider
the direct solution of (\ref{e1.1}) by means of general sparse Gaussian
elimination. In such a procedure, we find a permutation matrix $P$, and
compute the decomposition
\[
P A P^{t} = L D L^{t}
\]
where $L$ is unit lower triangular and $D$ is diagonal.
\section{Design Considerations.}Several good ordering algorithms (nested dissection and
minimum degree)
are available for computing $P$ \cite{GEORGELIU}, \cite{ROSE72}.
Since our interest here does not
focus directly on the ordering, we assume for convenience that $P=I$,
or that $A$ has been preordered to reflect an appropriate choice of $P$.
Our purpose here is to examine the nonnumerical complexity of the
sparse elimination algorithm given in \cite{BANKSMITH}.
As was shown there, a general sparse elimination scheme based on the
bordering algorithm requires less storage for pointers and
row/column indices than more traditional implementations of general
sparse elimination. This is accomplished by exploiting the m-tree,
a particular spanning tree for the graph of the filled-in matrix.
\begin{theorem} The method was extended to three
dimensions. For the standard multigrid
coarsening
(in which, for a given grid, the next coarser grid has $1/8$
as many points), anisotropic problems require plane
relaxation to
obtain a good smoothing factor.\end{theorem}
Our purpose here is to examine the nonnumerical complexity of the
sparse elimination algorithm given in \cite{BANKSMITH}.
As was shown there, a general sparse elimination scheme based on the
bordering algorithm requires less storage for pointers and
row/column indices than more traditional implementations of general
sparse elimination. This is accomplished by exploiting the m-tree,
a particular spanning tree for the graph of the filled-in matrix.
Several good ordering algorithms (nested dissection and minimum degree)
are available for computing $P$ \cite{GEORGELIU}, \cite{ROSE72}.
Since our interest here does not
focus directly on the ordering, we assume for convenience that $P=I$,
or that $A$ has been preordered to reflect an appropriate choice of $P$.
\begin{proof} In this paper we consider two methods. The first method
is
basically the method considered with two differences:
first, we perform plane relaxation by a two-dimensional
multigrid method, and second, we use a slightly different
choice of
interpolation operator, which improves performance
for nearly singular problems. In the second method coarsening
is done by successively coarsening in each of the three
independent variables and then ignoring the intermediate
grids; this artifice simplifies coding considerably.
\end{proof}
Our purpose here is to examine the nonnumerical complexity of the
sparse elimination algorithm given in \cite{BANKSMITH}.
As was shown there, a general sparse elimination scheme based on the
bordering algorithm requires less storage for pointers and
row/column indices than more traditional implementations of general
sparse elimination. This is accomplished by exploiting the m-tree,
a particular spanning tree for the graph of the filled-in matrix.
\begin{Definition}{\rm We describe the two methods in \S 1.2. In \S\ 1.3. we
discuss
some remaining details.}
\end{Definition}
Our purpose here is to examine the nonnumerical complexity of the
sparse elimination algorithm given in \cite{BANKSMITH}.
As was shown there, a general sparse elimination scheme based on the
bordering algorithm requires less storage for pointers and
row/column indices than more traditional implementations of general
sparse elimination. This is accomplished by exploiting the m-tree,
a particular spanning tree for the graph of the filled-in matrix.
Several good ordering algorithms (nested dissection and minimum degree)
are available for computing $P$ \cite{GEORGELIU}, \cite{ROSE72}.
Since our interest here does not
focus directly on the ordering, we assume for convenience that $P=I$,
or that $A$ has been preordered to reflect an appropriate choice of $P$.
Our purpose here is to examine the nonnumerical complexity of the
sparse elimination algorithm given in \cite{BANKSMITH}.
As was shown there, a general sparse elimination scheme based on the
bordering algorithm requires less storage for pointers and
row/column indices than more traditional implementations of general
sparse elimination.
\begin{lemma} We discuss first the choice for $I_{k-1}^k$
which is a generalization. We assume that $G^{k-1}$ is
obtained
from $G^k$
by standard coarsening; that is, if $G^k$ is a tensor product
grid $G_{x}^k \times G_{y}^k \times G_{z}^k$,
$G^{k-1}=G_{x}^{k-1} \times G_{y}^{k-1} \times G_{z}^{k-1}$,
where $G_{x}^{k-1}$ is obtained by deleting every other grid
point of $G_x^k$ and similarly for $G_{y}^k$ and $G_{z}^k$.
\end{lemma}
To our knowledge, the m-tree previously has not been applied in this
fashion to the numerical factorization, but it has been used,
directly or indirectly, in several optimal order algorithms for
computing the fill-in during the symbolic factorization phase
[4] - [10], [5], [6]. In \S 1.3., we analyze the complexity of the old and new
approaches to the intersection problem for the special case of
an $n \times n$ grid ordered by nested dissection. The special
structure of this problem allows us to make exact estimates of
the complexity. To our knowledge, the m-tree previously has not been applied in this
fashion to the numerical factorization, but it has been used,
directly or indirectly, in several optimal order algorithms for
computing the fill-in during the symbolic factorization phase
[4] - [10], [5], [6].
In \S 1.2, we review the bordering algorithm, and introduce
the sorting and intersection problems that arise in the
sparse formulation of the algorithm.
In \S 1.3., we analyze the complexity of the old and new
approaches to the intersection problem for the special case of
an $n \times n$ grid ordered by nested dissection. The special
structure of this problem allows us to make exact estimates of
the complexity. To our knowledge, the m-tree previously has not been applied in this
fashion to the numerical factorization, but it has been used,
directly or indirectly, in several optimal order algorithms for
computing the fill-in during the symbolic factorization phase
[4] - [10], [5], [6].
For the old approach, we show that the
complexity of the intersection problem is $O(n^{3})$, the same
as the complexity of the numerical computations. For the
new approach, the complexity of the second part is reduced to
$O(n^{2} (\log n)^{2})$.
To our knowledge, the m-tree previously has not been applied in this
fashion to the numerical factorization, but it has been used,
directly or indirectly, in several optimal order algorithms for
computing the fill-in during the symbolic factorization phase
[4] - [10], [5], [6]. In \S 1.3., we analyze the complexity of the old and new
approaches to the intersection problem for the special case of
an $n \times n$ grid ordered by nested dissection. The special
structure of this problem allows us to make exact estimates of
the complexity. To our knowledge, the m-tree previously has not been applied in this
fashion to the numerical factorization, but it has been used,
directly or indirectly, in several optimal order algorithms for
computing the fill-in during the symbolic factorization phase
[4] - [10], [5], [6].
This is accomplished by exploiting the m-tree,
a particular spanning tree for the graph of the filled-in matrix.
To our knowledge, the m-tree previously has not been applied in this
fashion to the numerical factorization, but it has been used,
directly or indirectly, in several optimal order algorithms for
computing the fill-in during the symbolic factorization phase
\cite{EISENSTAT} - \cite{LIU2}, \cite{ROSE76}, \cite{SCHREIBER}.
\subsection{Robustness.}We do not
attempt to present an overview
here, but rather attempt to focus on those results that
are relevant to our particular algorithm.
This section assumes prior knowledge of the role of graph theory
in sparse Gaussian elimination; surveys of this role are
available in \cite{ROSE72} and \cite{GEORGELIU}. More general
discussions of elimination trees are given in
\cite{LAW} - \cite{LIU2}, \cite{SCHREIBER}.
Thus, at the $k$th stage, the bordering algorithm consists of
solving the lower triangular system
\begin{equation} \label{1.2}
L_{k-1}v = c
\end{equation}
and setting
\begin{eqnarray}
\ell &=& D^{-1}_{k-1}v , \\
\delta &=& \alpha - \ell^{t} v .
\end{eqnarray}
\begin{figure}
\vspace{14pc}
\caption{This is a figure 1.1.}
\end{figure}
\section{Robustness.} We do not
attempt to present an overview
here, but rather attempt to focus on those results that
are relevant to our particular algorithm.
\subsection{Versatility.}The special
structure of this problem allows us to make exact estimates of
the complexity. For the old approach, we show that the
complexity of the intersection problem is $O(n^{3})$, the same
as the complexity of the numerical computations
\cite{GEORGELIU}, \cite{ROSEWHITTEN}. For the
new approach, the complexity of the second part is reduced to
$O(n^{2} (\log n)^{2})$.
To our knowledge, the m-tree previously has not been applied in this
fashion to the numerical factorization, but it has been used,
directly or indirectly, in several optimal order algorithms for
computing the fill-in during the symbolic factorization phase
[4] - [10], [5], [6]. In \S 1.3., we analyze the complexity of the old and new
approaches to the intersection problem for the special case of
an $n \times n$ grid ordered by nested dissection. The special
structure of this problem allows us to make exact estimates of
the complexity. To our knowledge, the m-tree previously has not been applied in this
fashion to the numerical factorization, but it has been used,
directly or indirectly, in several optimal order algorithms for
computing the fill-in during the symbolic factorization phase
[4] - [10], [5], [6].
In \S 1.2, we review the bordering algorithm, and introduce
the sorting and intersection problems that arise in the
sparse formulation of the algorithm.
In \S 1.3., we analyze the complexity of the old and new
approaches to the intersection problem for the special case of
an $n \times n$ grid ordered by nested dissection. The special
structure of this problem allows us to make exact estimates of
the complexity. To our knowledge, the m-tree previously has not been applied in this
fashion to the numerical factorization, but it has been used,
directly or indirectly, in several optimal order algorithms for
computing the fill-in during the symbolic factorization phase
[4] - [10], [5], [6].
For the old approach, we show that the
complexity of the intersection problem is $O(n^{3})$, the same
as the complexity of the numerical computations. For the
new approach, the complexity of the second part is reduced to
$O(n^{2} (\log n)^{2})$.
To our knowledge, the m-tree previously has not been applied in this
fashion to the numerical factorization, but it has been used,
directly or indirectly, in several optimal order algorithms for
computing the fill-in during the symbolic factorization phase
[4] - [10], [5], [6]. In \S 1.3., we analyze the complexity of the old and new
approaches to the intersection problem for the special case of
an $n \times n$ grid ordered by nested dissection. The special
structure of this problem allows us to make exact estimates of
the complexity. To our knowledge, the m-tree previously has not been applied in this
fashion to the numerical factorization, but it has been used,
directly or indirectly, in several optimal order algorithms for
computing the fill-in during the symbolic factorization phase
[4] - [10], [5], [6].
This is accomplished by exploiting the m-tree,
a particular spanning tree for the graph of the filled-in matrix.
To our knowledge, the m-tree previously has not been applied in this
fashion to the numerical factorization, but it has been used,
directly or indirectly, in several optimal order algorithms for
computing the fill-in during the symbolic factorization phase
\cite{EISENSTAT} - \cite{LIU2}, \cite{ROSE76}, \cite{SCHREIBER}.
\section{Introduction}
Nowadays, many network management companies (e.g., Cisco, NTT Global Networks) collect large volumes of data logs to measure and monitor the status (e.g., healthy, overloaded, power outage) of devices (e.g., routers, switches) or interfaces (e.g., local area networks (LAN), wide area networks (WAN)) in a network. Figure~\ref{figure: network_deff} shows an example network. A network usually consists of multiple devices, each with some interfaces, e.g., an SDWAN device has LAN and WAN interfaces. Each interface is measured by several metrics or attributes, i.e., the number of in/out octets and in/out ucast packets, that describe network operations or activities as multivariate time series.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.95\linewidth]{figures/network_deff.jpg}
\caption{An example network contains two SDWAN devices, and the devices have some LAN and WAN interfaces, each measured by four attributes.}
\label{figure: network_deff}
\end{figure}
Efficiently and accurately predicting abnormal events from network data logs is important, which helps continuously monitor network states and raise alerts for potential incidents on time for network engineers to investigate and resolve. Traditional monitoring systems typically trigger a real-time notification when the network is experiencing some abnormal event (e.g., packet loss, intrusion) using pre-defined rules but cannot provide an early warning on a potential incident. Also, a network typically contain hundreds if not thousands of heterogeneous devices from different vendors, which can have diverse criteria (e.g., some rules and thresholds) indicating varying levels of abnormal status, making it challenging to build an early warning system and enable the system handling all kinds of networks.
Advanced data-driven methods on analyzing \textit{multiple multivariate time series} (e.g., machine learning approaches) can automatically learn and extract patterns from network operations that lead to a potential abnormal event, which has the potential to handle heterogeneous networks, devices, and interfaces. One challenge is how to effectively capture temporal dependencies across time and relationships between time series that can jointly represent network activities for predicting an abnormal event. Some existing approaches overlook the relationships between time series, e.g., univariate time-series anomaly detection algorithms~\cite{ren2019time}, fusing multiple time series directly as features~\cite{malhotra2016lstm, li2019mad}. In terms of modeling temporal dependencies, many approaches use recurrent neural networks (RNNs)~\cite{malhotra2016lstm, hundman2018detecting} yet they are often time-consuming and require large memory for computation. Transformers, oriented from natural language processing, become popular in time-series analysis problems~\cite{cai2020traffic}, but they have similar issues.
In addition, some labeled data might be available, but the labeling criteria could vary significantly among devices and interfaces, e.g., an abnormal event can last for several minutes, and sometimes only the beginning time point is marked as anomalous. Unsupervised approaches, including reconstruction-based methods~\cite{park2018multimodal, su2019robust}, forecasting-based methods~\cite{laptev2015generic, hundman2018detecting}, and hybrid methods~\cite{zhao2020multivariate}, identify infrequent behaviors in multivariate time series as an anomaly. However, this assumption is not always true, e.g., long-term peak usage of memory might be rare but normal. Supervised models~\cite{liu2015opprentice, shipmon2017time} can leverage labeled data to guide the process of learning the time-series patterns leading to an abnormal event, but they usually require sufficient labeled samples. Also, the number of labeled abnormal samples is usually much smaller compared to the normal samples. Therefore, limited and imbalanced labels make it challenging to capture the patterns leading to abnormal activities.
This paper proposes a semi-supervised learning approach that learns representative features for network activities and use the features to predict abnormal events (e.g., packet loss, intrusion).
Specifically, our approach first takes a given network and constructs a graph representation where each node contains some attributes of an interface (e.g., packet volume) at a time point. (There could be some dependencies (e.g., data transmission) between interfaces (nodes) via the linkage.) Considering a sequence of graphs (i.e., network data logs over a time period) leading to some abnormal event, our approach aims to generate an embedding to represent network behaviors and capture the unique pattern of this graph sequence. The embedding should contain effective temporal information by capturing temporal dependencies across time points (i.e., temporal patterns). The embedding also includes relevant information from neighboring nodes, capturing data communication patterns. Therefore, our approach jointly captures temporal dependencies across time points and spatial dependencies between nodes for generating an embedding of the input. Then these learned embeddings are refined to form two separable clusters representing normal and abnormal network behaviors. Our approach uses labeled samples to explicitly learn the cluster centers in a contrastive way. However, due to imbalanced and limited labeled samples, the learnt centers might not be applicable to all samples and hence poor embedding quality. Our approach updates the embeddings of all samples towards a high confidence clustering distribution to enforce cluster cleanness with a semi-supervised learning strategy. The goal is to improve separability between normal and abnormal network behaviors. Finally, our approach utilizes a support vector classifier on these embeddings to predict if the given graph sequence of network operations will lead to an abnormal event.
In sum, the main contribution of this paper is a semi-supervised learning approach that learns separable embeddings that jointly captures temporal dependencies between time points and relationships between time series for predicting abnormal events. The proposed approach can efficiently capture complex dependencies in multiple multivariate time-series data and learn effective features from imbalanced and limited labels, enabling accurate prediction of abnormal events. We demonstrate the proposed network architecture for abnormal event prediction on large real-world network operation time-series data.
\section{Methodology}
\subsection{Problem Definition}
Given a network containing multiple interfaces with unknown connectivity, our system first constructs a fully-connected graph, $\boldsymbol{G}=(\boldsymbol{V}, \boldsymbol{E})$, where each node $i \in \boldsymbol{V}$ represents an interface ($|\boldsymbol{V}|=n$) and the connection between any pair of nodes is initialized as one, meaning there can be data communication between interfaces. Let $\boldsymbol{G}^{(t)} \in \mathbb{R}^{n \times F}$ be the graph for network operation measurements at time~$t$, where~$n$ is the number of nodes and~$F$ is the number of attributes describing each node. Suppose the input is a sequence of graphs of length $T$, i.e., graphs from $t-T+1$ to $t$, our approach aims to learn a function $f$ that maps $T$ historical input signals to a binary signal $y \in \{0, 1\}$ indicating if there will be an abnormal event in the network at $t+1$.
\begin{displaymath}
[\boldsymbol{G}^{(t-T+1)}, \cdots, \boldsymbol{G}^{(t)}]\xrightarrow{\emph{f}} y
\end{displaymath}
Figure~\ref{figure: whole} shows the proposed network architecture for learning feature embeddings from the graph input. Specifically, our approach first builds an GatedConv-NGAT module to generate an initial feature embedding by jointly capturing dependencies among graphs (across time points in the sequence) and between nodes (between individual time series in a graph) (Section~\ref{sec: stconvngat}). Then our approach jointly refines the embeddings and cluster centers representing normal and abnormal activities by: (1) leveraging \textit{labeled} normal and abnormal samples to form clusters in a contrastive learning way; (2) improving the separability of clusters using both \textit{labeled} and \textit{unlabeled} samples in a semi-supervised manner (Section~\ref{sec: cc}). Finally, our approach trains a support vector classifier with these learned embeddings for the prediction task (Section~\ref{sec: svc}).
\begin{figure}[htbp]
\centering
\includegraphics[width=\linewidth]{figures/wholee.jpg}
\vspace{-.2in}
\caption{The training process of our system to generate feature representation for predicting abnormal events}
\vspace{-.2in}
\label{figure: whole}
\end{figure}
\subsection{GatedConv-NGAT Module.} \label{sec: stconvngat}
This section introduces the ST-Conv-NGAT module that generates embeddings from the input graph sequence (Figure~\ref{figure: stconvngat}). In general, the module contains temporal gated convolutional layers to capture temporal dependencies across time and a normalized graph attention layer to capture dependencies between nodes. The details of each sub-module are described as follows.
\begin{figure}[htbp]
\centering
\includegraphics[width=0.95\linewidth]{figures/stconvngat.jpg}
\vspace{-.1in}
\caption{GatedConv-NGAT Module}
\vspace{-.2in}
\label{figure: stconvngat}
\end{figure}
\textbf{GatedConv layer for extracting temporal relationships.} Recurrent neural networks (RNNs), e.g., long short-term memory (LSTM), have been widely used for capturing temporal dependencies in time-series data, but these models typically suffer from complex gating mechanisms, time-consuming computation, and large memory requirement since every step relies on the previous steps. In comparison, convolutional neural networks (CNNs) have the advantages of fast training and parallel computation~\cite{gehring2017convolutional} that can process large network data logs more efficiently. Also, CNNs explicitly capture local features within a temporal neighborhood, and one or more local features might cause an abnormal event, instead of global features from the entire sequence.
Our approach applies gated-conv layers~\cite{gehring2017convolutional} with multiple kernel sizes consisting of 1D convolution operations on the time axis to extract local temporal features for each node. The input to the 1D convolution operation is a sequence of length $T$ with a channel size $F$ at each node $\in \mathbb{R}^{T \times F}$. With a kernel size~$K$ and a hidden size~$C$, the convolution kernel maps the input to the output $[\mathbf{m}\ \mathbf{n}] \in \mathbb{R}^{(T-K+1) \times (2C)}$, where $\mathbf{m}$, $\mathbf{n}$ split the output into halves with the same hidden size~$h$. The gated-conv layer leverages gated linear units (GLU)~\cite{dauphin2017language} as a non-linear gating mechanism over $\mathbf{m}$ and $\mathbf{n}$:
\begin{equation}
\text{GatedConv}(\mathbf{v}) = \mathbf{m} \odot \sigma(\mathbf{n})\ \in \mathbb{R}^{(T-K+1) \times (C)}
\end{equation}
where $\mathbf{m}$ and $\mathbf{n}$ are the input to the gates of GLU, $\odot$ denotes the element-wise product, $\sigma$ is the sigmoid function, and $\sigma(\mathbf{n})$ controls which information in $\mathbf{m}$ are relevant for generating temporal features.
\textbf{Normalized graph attention layer for extracting relationships between nodes.} After the first Gated-Conv layer, each node $i$ is represented by $(T-K+1)$ feature vectors of size $C$, where each vector contains the extracted temporal features for the corresponding sub-sequence. In addition to temporal dependencies, combining information from other nodes is also essential for determining an abnormal event. Thus, our approach introduces a normalized graph attention layer to capture relationships between nodes.
Since the initial graph topology is unknown, many graph convolution strategies, e.g., chebyshev spectral graph convolutional operator~\cite{defferrard2016convolutional}, might not be suitable because they require the graph connectivity as the input. Our approach leverages a graph attention (GAT) layer~\cite{velivckovic2017graph} to model the spatial relationships between nodes. A GAT layer computes the representation for each node:
\vspace{-.1in}
\begin{equation}
\text{GAT}(i) = \sigma \Big ( \sum_{k\in \mathcal{N}_i}\alpha_{ik}\mathbf{v_k} \Big )
\end{equation}
where node $i$ is the target node, $\mathbf{v_i} \in \mathbb{R}^{C}$ is the hidden feature vector of node $i$, $\mathcal{N}_i$ is the neighborhood set of node $i$, $\sigma$ is the sigmoid function, $\alpha_{ik}$ is the attention score for node $k$ to node $i$. The attention score $\alpha_{ik}$ is defined by:
\vspace{-.1in}
\begin{equation}
\alpha_{ik}=\frac{\text{exp}(\sigma(\mathbf{w}^T \cdot (\mathbf{v_i}\ \Vert\ \mathbf{v_k}))}{\sum_{j \in \mathcal{N}_i} \text{exp}(\sigma(\mathbf{w}^T \cdot (\mathbf{v_i}\ \Vert\ \mathbf{v_j}))))}
\end{equation}
where the operation $\Vert$ is to concatenate two vectors, $\mathbf{w} \in \mathbb{R}^{2C}$ is a learnable weight vector, $\mathcal{N}_i$ is the neighborhood set of node $i$, $\sigma$ is the LeakyReLU function.
Our approach then adapts graph normalization (GraphNorm)~\cite{cai2021graphnorm} to the output of the GAT layer, which makes the optimization more efficient. GraphNorm normalizes every feature dimension for individual graphs with a learnable shift, which significantly boosts the training upon InstanceNorm~\cite{ulyanov2016instance} (also does graph-targeted normalization):
\vspace{-.1in}
\begin{equation}
\text{GraphNorm}(h_{i, j})=\gamma_j \cdot \frac{h_{i, j} - \alpha_j \mu_j}{\sigma_j} + \beta_j
\end{equation}
\begin{displaymath}
\mu_j=\frac{\sum_{k=1}^n h_{k, j}}{n}, \ \sigma_j=\frac{\sum_{k=1}^n (h_{k, j} - \alpha_j \mu_j)}{n}
\end{displaymath}
where $h_{i, j}$ is the value of the $j$th feature dimension of $\text{GAT}(i)$ (the output of GAT layer for node $i$), $n$ is the number of nodes in the graph, $\alpha_j$ is the learnable parameter for the feature dimension $j$, which controls how much information to keep in the mean. $\gamma_j$ and $\beta_j$ are the affine parameters. \vspace{2pt}
\textbf{Putting everything together to learn embeddings.} Inspired by~\cite{yu2017spatio}, our approach adds another temporal gated-conv layer after the NGAT layer to fuse the features of node interactions on the time axis. This ``sandwich'' structure (i.e., a spatial layer in the middle of two temporal layers) can achieve fast spatial-state propagation through temporal convolutions. After the second temporal gated-conv layer, our approach concatenates the output to generate the node embeddings. Then our approach uses a global pooling layer to add node embeddings across the node dimension to generate the graph embedding. In this way, given [$\boldsymbol{G}^{(t-T+1)}, \cdots, \boldsymbol{G}^{(t)}] \in \mathbb{R}^{T \times n \times F}$, the GatedConv-NGAT module generates the embedding $\boldsymbol{E}^{(t)}$ of length ${(T-2(K-1)) \times C}$, that jointly captures the temporal dependencies and relationship between nodes.
\subsection{Learning clustering structure.} \label{sec: cc}
Our approach assumes two clusters in the embedding space: normal and abnormal events. The goal is to learn a representation space to form these two separable clusters.
\textbf{Learning clustering structure on labeled samples using contrastive center loss.} The labeled samples can explicitly guide the process of learning clustering centers by enforcing the embeddings of normal samples to be close to the normal center and far away from the abnormal center, and vice versa. Our approach applies the contrastive-center loss~\cite{qi2017contrastive} that considers intra-class compactness and inter-class separability simultaneously. This loss function aims to penalize the contrastive values between (1) the distances of samples to their corresponding class center and (2) the distances of samples to their non-corresponding class centers,
\vspace{-.1in}
\begin{equation}
\mathcal{L}_{cc} = \frac{1}{2} \sum^m_{i=1} \frac{\lVert \mathbf{E_i} - \mathbf{c_{y_i}} \rVert^2_2}{\lVert \mathbf{E_i} - \mathbf{c_{j}} \rVert^2_2 + \epsilon } \ (j\neq y_i)
\label{eq: cc}
\end{equation}
where $L_{cc}$ denotes the contrastive-center loss; $m$ is the number of labeled samples in a min-batch; $\mathbf{E_i}$ is the embedding of the $i$-th training sample; $\mathbf{c_{y_i}}$ is the corresponding center of $\mathbf{E_i}$. Since there are two clusters, when $y_i$ is 0, $j$ is 1, and vice versa ($j\neq y_i$). $\epsilon$ is a constant for preventing the denominator equal to 0 and set to ${1}\mathrm{e}{-6}$ empirically.
\textbf{Learning clustering structure on all samples using a semi-supervised learning strategy.} Using the contrastive learning strategy improves the clustering cleanness for labeled samples; however, due to limited and imbalanced labeled samples, the learned centers are easily guided by the given labeled samples, which might fail in inferencing. To overcome the scarcity of labeled training samples and ensure the clusters are applicable to all samples, our approach takes the idea from~\cite{xie2016unsupervised} to minimize the distance between the cluster assignment and some target cluster distribution. For labeled samples, the target distribution can be directly derived from the label, while for the unlabeled samples, the target distribution is derived from high confidence predictions of the current cluster distribution from all samples. Formally, our approach calculates the probability of assigning the embedding $\mathbf{E_i}$ to the cluster $j$ with a center $\mu_j$ as
\vspace{-.1in}
\begin{equation}
q_{ij} = \frac{(1+\rVert \mathbf{E_i} - \mu_j \rVert ^ 2) ^ {-1}}{\sum_{j'}(1+\rVert \mathbf{E_i} - \mu_{j'} \rVert ^ 2) ^ {-1})}
\label{eq: Q}
\end{equation}
Following the work in~\cite{xie2016unsupervised}, our approach derives an auxiliary distribution made up of the high confidence assignments of the current distribution (referred as $Q$). The main idea is to emphasize data points assigned with high confidence to ensure cluster purity. Equation~\ref{eq: P} defines the computation of auxiliary distribution (referred as $P$), where the second power of $q_{ij}$ places more weight on the instances near the center and the division of $\sum_{i} q_{ij}$ normalizes based on the cluster size, making the model robust to biased classes. Note that for labeled samples, our approach replaces the true distribution with the auxiliary distribution and hence enables the training process in a semi-supervised manner.
\vspace{-.1in}
\begin{equation}
p_{ij} = \frac{q_{ij}^2 / \sum_{i'} q_{i'j}}{\sum_{j'} (q_{ij'}^2 / \sum_{i'} q_{i'j'})}
\label{eq: P}
\end{equation}
To measure the distance between the distributions $P$ and $Q$, our approach computes the KL divergence to measure their distance. The clustering oriented loss is defined as,
\vspace{-.1in}
\begin{equation}
\mathcal{L_{KL}} = KL(P||Q) = \sum_i \sum_j p_{ij} \log \frac{p_{ij}}{q_{ij}}
\label{eq: Q}
\end{equation}
Therefore, the total loss is the sum of contrastive center loss and KL divergence loss with a hyper-parameter $\lambda$,
\vspace{-.1in}
\begin{equation}
\mathcal{L} = \mathcal{L}_{cc} + \lambda \times \mathcal{L}_{KL}
\label{eq: Q}
\end{equation}
\subsection{Predicting abnormal events from learned embeddings using SVC.}\label{sec: svc} Using the methods described so far in this section, our model generates separable embeddings (normal vs. abnormal). Then we take the embeddings and labels to train a support vector classifier to predict normal and abnormal events. Our approach does not train in an end-to-end manner, e.g., using a fully-connected layer for classification in the model because support vector classifier usually works better if the input are already separable in some space. Figure~\ref{figure: inference} shows the prediction process. Given a sequence of graphs representing network operations, the GatedConv-NGAT module generates the embedding and the support vector classifier takes the embedding to predict if the input would lead to an abnormal event.
\begin{figure}[htbp]
\centering
\vspace{-.1in}
\includegraphics[width=\linewidth]{figures/inference.jpg}
\caption{Event prediction}
\label{figure: inference}
\end{figure}
\section{Experiments and Results}
We implemented the proposed architecture with Python~3.8, PyTorch~1.8, and PyTorch Geometric framework~\cite{Fey/Lenssen/2019}. We conducted the experiments in a Docker container deployed on a GPU server and trained the model with one physical core and 16GB memory.
\subsection{Datasets.}
We used the network operation data from NTT-GN for the experiments. The dataset contains more than 150 million data logs from network interfaces (e.g., WAN routers and LAN switches) that NTT-GN collects, monitors, and manages. The data logs we used cover the time range from 2020-05-12 to 2020-06-12, with a time interval of 5 minutes. Figure~\ref{figure: network_deff} shows an example of the data structure. In this work, we assume a practical case that there could be dependencies between any interface pairs in a network (e.g., data transmissions) (Figure~\ref{figure: graph_construction}), and the network topology in not available. The graph representation can be modified if prior knowledge of the network topology is available.
In addition to the network operation data, NTT-GN also provides the packet loss event dataset that reports the ticket opening time at the device level. We assume that events are caused by some abnormal network activities in the network and can be captured by looking at the dynamics of the entire network. Figure~\ref{figure: label_construction} shows an example of assigning a label to a period of network activities (e.g., 6 hours, a sequence of 72 graphs). If there is an event reported at $t'$ and the closest graph before that is $\mathbf{G}^{(t)}$, we assigned label 1 (color blue) to the graph sequence $\mathbf{G}^{(t-71)} : \mathbf{G}^{(t)}$, denoting that these graphs contains some abnormal network activities directly leading to an event. Since we do not know how long the event lasts, the graph sequences overlapped with an event within 6 hours before and after (e.g., color yellow) were marked as ``unknown'' (i.e., unlabeled data). These graph sequences also help learn feature embeddings in our semi-supervised setting. The rest graph sequences that do not overlap with any event within 6 hours were labeled as 0, representing normal network activities.
\begin{figure}[htbp]
\centering
\subfigure[The graph representing the example network]{
\begin{minipage}[t]{\columnwidth}
\centering
\includegraphics[width=0.6\linewidth]{figures/graph_construction.jpg}
\label{figure: graph_construction}
\end{minipage}
}
\subfigure[An example of graph label construction]{
\begin{minipage}[t]{\columnwidth}
\centering
\includegraphics[width=\linewidth]{figures/label_construction.jpg}
\label{figure: label_construction}
\end{minipage}
}
\vspace{-0.2in}
\caption{Examples of sample construction}
\label{figure: data_construction}
\end{figure}
\subsection{Experiment Settings.} To evaluate the proposed approach, we selected the available networks with at least 50 events. We split the samples into training, validation, and testing sets, ensuring each set contains both label 0 and label 1. Table~\ref{table: dataset summary} shows the summary of the dataset. We observe that the number of samples labeled with 0 is much larger than label 1. Also, the ratio of labeled samples to unlabeled samples is around 0.2. We repeated the random splitting process three times for cross-validation and took the average as the results.
We set the length of the input graph sequence in a sample as 72. The batch size was 32, each containing both labeled and unlabeled samples. We set the temporal gated convolutional layers as 32 hidden states and the kernel size of 1D Conv as 12. The size of the output feature embeddings was 256. The initial learning rate was 0.001, with early stopping on the validation dataset. The hyper-parameters were chosen by grid search on the validation set, and we reported the results with the set of hyper-parameters yielding the best average performance, that is $\lambda=0.1$.
\begin{table}[h]
\centering
\caption{Summary of the datasets}
\label{table: dataset summary}
\begin{tabular}{ c | c | c | c }
\hline
Network ID & \#Nodes & \#Events & \# 0: \# 1 \\ \hline
1NTT18927 & 53 & 125 & 0.96:0.04 \\ \hline
1NTT19535 & 84 & 142 & 0.89:0.11 \\ \hline
1NTT67246 & 36 & 134 & 0.91:0.09 \\ \hline
1NTT19478 & 120 & 147 & 0.88:0.12 \\ \hline
143XB8CT & 97 & 78 & 0.98:0.02 \\ \hline
1NTT30990 & 20 & 71 & 0.98:0.02 \\ \hline
1NTT48337 & 89 & 83 & 0.95:0.05 \\ \hline
1NTT19213 & 119 & 78 & 0.96:0.04 \\ \hline
1NTT22885 & 91 & 91 & 0.95:0.05 \\ \hline
1NTT18635 & 143 & 189 & 0.73:0.27 \\ \hline
\end{tabular}
\end{table}
\noindent \textbf{Metrics.} We reported the precision, recall, and F1 score for label 1 since our goal is to predict events. The higher the values the better the performance. We put the detailed functions in the supplementary file.
\noindent \textbf{Baselines.} We compared our proposed model to the following state-of-the-art baselines:
\begin{itemize}
\setlength\itemsep{-.1em}
\item Support Vector Classifier (SVC): We convert the problem to a binary classification task for SVC. We flatten an input graph sequence in time and node dimensions to be a large vector for predicting event or non-event.
\item Graph Attention Encoder (GAT): We concatenate features along the temporal dimension for each node and replace the encoder with only the graph attention module to capture interactions between time series. This method aims to demonstrate the necessity of capturing temporal dependencies.
\item Graph Attention LSTM (GAT-LSTM)~\cite{wu2018graph}: Instead of using the ST-Conv module, we apply LSTM in the encoder to capture interactions across time points. The rest architecture remains the same as our proposed approach.
\item MTAD-GAT~\cite{zhao2020multivariate}: This is an unsupervised anomaly detection approach that explicitly captures relationships between multiple features and across time points. We modify the approach so that the node value has more than one dimension.
\item DeepSAD~\cite{ruff2019deep}: This is a semi-supervised learning approach for anomaly detection that aims to learn a clustering center to represent normal cases. The original domain is computer vision. To extend the model on our dataset, we apply the ST-Conv-NGAT module as the encoder.
\item We select the following variations of the proposed model for ablation study: 1. ST-Conv-NGAT + only contrastive center loss (only-CC), which would be trained only with labeled samples; 2. ST-Conv-NGAT + only KL divergence loss (only-KL); 3. Constructing a correlation graph as the input~\cite{deng2021graph}, which computes pairwise Pearson correlation between nodes and assumes two nodes are connected only if the correlation value is above the mean plus one std.
\end{itemize}
\subsection{Model Performance.}
In general, our approach outperformed baseline methods in all networks (see F1 scores in Figure~\ref{figure: ff1} and precision \& recall in the supplementary file). The F1 scores in most networks are above 0.7. Even when the ratio between the number of label 0 and label 1 is high (e.g., ``143XB8CT'' and ``1NTT30990''), our model could still achieve far better results than the baseline models. Specifically, our approach achieved more than 50\% improvement in F1 scores compared to SVC. This is because SVC simply takes the input graph sequence as a flattened vector and does not explicitly capture relationships between time series or across time points.
Compared to using other encoding methods (GAT and GAT-LSTM) to generate representations for prediction, our approach with the ST-Conv-NGAT module as the encoder showed excellent advantage. First, GAT does not explicitly consider temporal dependencies but fuses features on the time dimension together and expects the model to extract useful information from these values for prediction directly. However, the temporal order of these values matters for event prediction, e.g., the position of a sudden increment might result in different network behaviors. Our approach leverages the 1D-CNN to summarize local temporal features, aiming to extract effective temporal patterns leading to an event. Second, although GAT-LSTM is commonly used for capturing spatiotemporal dependencies in multi-time-series data~\cite{wu2018graph}, its performance is poorer than our approach, which might be because GAT-LSTM is typically generating representations for the entire time series. However, in practice, only some subsequence(s) from the input graph sequence would cause an event. Some effective patterns (or features) at the early stage of the input sequence might be easily weakened or neglected during the sequential training of LSTM. Since the impacting subsequence(s) on the event are unknown (i.e., the leading time to an event could be varying), the sequence length (i.e., six hours in the experiment) was chosen to cover as many time points as possible. In comparison, the 1D-CNN in our approach focuses on manipulating temporal interactions on every subsequence, which empirically fulfills our assumption and achieves better performance on real datasets. Also, training GAT-LSTM for one epoch took more than 600 seconds while our approach required around 170 seconds under the same settings.
Furthermore, compared to existing work on anomaly detection, our approach achieved better performance than the unsupervised learning approach (MTAD-GAT), which demonstrates that assuming rare samples as anomalies might not be applicable in practice. Adding some labeled samples during training can guide the learning process to generate representative embeddings. Additionally, our approach outperformed the semi-supervised approach, DeepSAD. DeepSAD has been demonstrated to outperform some unsupervised approaches in the original paper. The loss function in DeepSAD enforces labeled normal samples to be close to the predefined center and abnormal samples to be as far as possible. For the unlabeled samples, DeepSAD has a strong assumption that the majority of data samples are normal and treats all unlabeled samples as normal in the loss. However, such an assumption might not hold in practice since one event can last for some period (i.e., the unlabeled data might contain abnormal network activities), which might confuse the model to learn effective representations for unlabeled samples and result in poor performance.
\begin{figure*}[htbp]
\centering
\includegraphics[width=0.95\linewidth]{figures/f1.jpg}
\caption{The F1 scores for baselines and our method (red). See precision and recall in the supplementary file.}
\label{figure: ff1}
\vspace{-.2in}
\end{figure*}
\vspace{-1em}
\subsection{Ablation Study.} \textbf{(1) Effect of Loss Functions.} Table~\ref{table: f1} shows that our approach achieved 12\% improvement in F1 scores on average compared to only using the contrastive center (CC) loss and 14\% improvement compared to only using the KL loss. The model with only the CC loss was trained with limited labeled samples, which worked well for the networks with large abnormal samples (e.g., more than 120 events in ``1NTT67246'' and ``1NTT18635''). However, the performance dropped in the cases with only a few abnormal samples. Our proposed approach trained on both labeled and unlabeled samples to force the learnt clustering structure was applicable to all data samples, resulting in better performance, e.g., 48\% improvement in F1 score for ``1NTT30990'' and 35\% for ``1NTT19213''. In the other case, the model with only KL loss was trained with all samples. Without the CC loss to explicitly optimize the centers, finding good centers was challenging even when the true clustering distribution could be estimated from labeled samples. \\
\vspace{-1em}
\noindent \textbf{(2) Effect of graph construction.} We compared the result of the model taking input graphs constructed based on the correlations between nodes. However, the performance is generally worse than taking the fully-connected graph as the input, which shows that the node correlations could not effectively describe the interactions between interfaces. Since the network connectivity is unknown, the proposed approach assumes that data communication (direct and indirect) could happen between any interface pairs and uses the graph attention module to automatically capture their relationships. The entire results are presented as Table 2 in the supplementary file.
\begin{table}[htpb]
\centering
\caption{F1 scores between using different loss functions (see the full table in the supplementary file)}
\label{table: f1}
\begin{tabular}{ >{\centering\arraybackslash}p{2cm} || >{\raggedleft\arraybackslash}p{1.4cm} | >{\raggedleft\arraybackslash}p{1.4cm} | >{\raggedleft\arraybackslash}p{1.4cm} }
Network ID & only-CC & only-KL & Ours \\ \hline \hline
1NTT18927 & 0.809 & 0.717 & \textbf{0.823}\\
1NTT19535 & 0.808 & 0.773 & \textbf{0.877}\\
1NTT67246 & \textbf{0.955} & 0.889 & 0.946\\
1NTT19478 & 0.786 & 0.782 & \textbf{0.808}\\
143XB8CT & 0.462 & 0.466 & \textbf{0.502}\\
1NTT30990 & 0.267 & 0.288 & \textbf{0.397}\\
1NTT48337 & 0.757 & 0.672 & 0.754\\
1NTT19213 & 0.422 & 0.413 & \textbf{0.573}\\
1NTT22885 & 0.800 & 0.850 & \textbf{0.897}\\
1NTT18635 & 0.935 & 0.962 & \textbf{0.980}\\ \hline
\end{tabular}
\end{table}
\section{Related Work}
There exists abundant literature on anomaly detection with time-series data. Deep learning approaches have shown promising results in capturing complex relationships between time series and across time points for extracting effective patterns that enable accurate anomaly prediction, including unsupervised~\cite{munir2018deepant,deng2021graph}, supervised~\cite{liu2015opprentice, shipmon2017time}, and semi-supervised approaches~\cite{vercruyssen2018semi}. For example,~\citet{hundman2018detecting} propose to use Long-Short-Term Memory (LSTM) to predict spacecraft telemetry and detect outliers in an unsupervised manner within each variable of a single multivariate time-series. However, the approach does not consider the dependencies between variables. Network operation data are typically multiple multivariate time-series data, which require careful considerations to to capture the interactions between variables and time series for event detection. RNNs and Transformers~\cite{cai2020traffic} have been widely used in modeling temporal dependencies in time-series data, but they typically suffer from time-consuming training process and large memory requirement~\cite{yu2017spatio}. \citet{zhao2020multivariate} propose an unsupervised anomaly detection approach on multivariate time-series data, with each variate/time series represented as one node in a graph. Their approach utilizes a feature attention layer and a temporal attention layer to explicitly capture the relationships between features and times respectively. Then the approach jointly optimizes a forecasting-based model and a reconstruction-based model to obtain better time-series representations. However, such unsupervised learning approaches usually make a strong assumption that infrequent behavior is anomalous and usually requires some scores for inference (e.g., see \cite{ruff2018deep}). Although there are some techniques, e.g., Peak Over Threshold (POT)~\cite{siffer2017anomaly}, that automatically determine the threshold during inference, the performance is still not competitive to supervised learning methods when labels are available (e.g., see~\cite{liu2015opprentice, shipmon2017time}). Supervised approaches can explicitly learn normal or abnormal patterns from labels yet requires a prohibitive large number of labeled examples (e.g., see~\cite{blazquez2021review}).
Semi-supervised learning approaches can overcome these limitations. For example, \citet{ruff2019deep} demonstrates that adding a few labeled samples in a semi-supervised manner would guide the model to achieve better performance than unsupervised learning methods for anomaly detection. \citet{xiao2020semi} propose a semi-supervised approach that uses an LSTM auto-encoder on both unlabeled data and labeled data with a reconstruction loss and adds prediction loss on the labeled data to guide the learning process. In comparison, our approach explicitly learns a separable embedding space for normal and abnormal samples instead of using a simple prediction loss (e.g., cross-entropy loss or focal loss for the imbalanced label situation). \citet{vercruyssen2018semi} propose a semi-supervised learning approach that first uses an unsupervised model, a clustering-based approach, to define typical normal behavior and identify anomalies. Then the approach actively adds labels from domain experts to guide the clustering approach in a semi-supervised manner. However, the performance would highly rely on the initialized clustering result. Our approach leverages the unlabeled samples to learn cluster centers in a self-learning way.
\section{Metrics}
We reported the precision, recall, and F1 score for label 1. The functions are described as follows:
\begin{displaymath}
\text{precision}(\text{label}=1) = \frac{TP(\text{label}=1)}{TP(\text{label}=1)+FP(\text{label}=1)}
\end{displaymath}
\vspace{-.2em}
\begin{displaymath}
\text{recall}(\text{label}=1) = \frac{TP(\text{label}=1)}{TP(\text{label}=1) + FN(\text{label}=1)}
\end{displaymath}
\vspace{-.2em}
\begin{displaymath}
\text{F1} = 2 \times \frac{\text{precision}(\text{label}=1) \times \text{recall}(\text{label}=1)}{\text{precision}(\text{label}=1) + \text{recall}(\text{label}=1)}
\end{displaymath}
where $TP$ is the number of true positives (i.e., the number of correctly predicted label 1), $FP$ is the number of false positives (i.e., the number of incorrectly predicted label 1), and $FN$ is the number of false negatives (i.e., the number of incorrectly predicted label 0). \\
\section{Model Performance}
Figure~\ref{figure: prec} and Figure~\ref{figure: rec} show the precision and recall comparison between baselines and our method. In general, our method achieved much better performance in precision and recall.
We also projected and plotted the embeddings in a 2D space using t-SNE with their labels to demonstrate our approach (see~Figure~\ref{figure: res}).
\begin{figure*}[htbp]
\centering
\includegraphics[width=\linewidth]{figures/prec.jpg}
\caption{The Precision for baselines and our method (red).}
\label{figure: prec}
\end{figure*}
\begin{figure*}[htbp]
\centering
\includegraphics[width=\linewidth]{figures/rec.jpg}
\caption{The Recall for baselines and our method (red).}
\label{figure: rec}
\end{figure*}
\begin{table}[htpb]
\centering
\renewcommand\arraystretch{1.1}
\caption{Precision and Recall between using different loss functions}
\label{table: prerec}
\begin{tabular}{ >{\centering\arraybackslash}p{2cm} || >{\raggedleft\arraybackslash}p{1.4cm} | >{\raggedleft\arraybackslash}p{1.4cm} | >{\raggedleft\arraybackslash}p{1.4cm} | >{\raggedleft\arraybackslash}p{1.4cm} | >{\raggedleft\arraybackslash}p{1.4cm} | >{\raggedleft\arraybackslash}p{1.4cm} }
\hline
& \multicolumn{3}{c|}{precision} & \multicolumn{3}{c}{recall} \\ \hline
Network ID & only-CC & only-KL & Ours & only-CC & only-KL & Ours \\ \hline \hline
1NTT18927 & 0.864 & 0.735 & 0.898 & 0.76 & 0.7 & 0.76 \\
1NTT19535 & 0.913 & 0.876 & 0.893 & 0.724 & 0.692 & 0.862 \\
1NTT67246 & 0.946 & 0.923 & 0.963 & 0.96 & 0.857& 0.929\\
1NTT19478 & 0.846 & 0.897 & 0.955 & 0.733 & 0.693& 0.7\\
143XB8CT & 0.667 & 0.621 & 0.674 & 0.353 & 0.373& 0.4\\
1NTT30990 & 0.267 & 0.254 & 0.294 & 0.267 & 0.333& 0.611\\
1NTT48337 & 0.682 & 0.629 & 0.691 & 0.833 & 0.722 & 0.829\\
1NTT19213 & 0.333 & 0.303 & 0.45 & 0.576 & 0.647 & 0.789\\
1NTT22885 & 0.875 & 0.81 & 0.85 & 0.73. & 0.895 & 0.949\\
1NTT18635 & 0.947 & 0.95 & 0.978 & 0.923 & 0.974 & 0.982\\ \hline
\end{tabular}
\end{table}
\begin{table}[htpb]
\centering
\renewcommand\arraystretch{1.1}
\caption{Precision, Recall, F1 score between using different graph construction methods}
\label{table: prerecf1}
\begin{tabular}{ >{\centering\arraybackslash}p{2cm} || >{\raggedleft\arraybackslash}p{1.9cm} | >{\raggedleft\arraybackslash}p{1.9cm} | >{\raggedleft\arraybackslash}p{1.9cm} | >{\raggedleft\arraybackslash}p{1.9cm} | >{\raggedleft\arraybackslash}p{1.9cm} | >{\raggedleft\arraybackslash}p{1.9cm} }
\hline
& \multicolumn{2}{c|}{precision} & \multicolumn{2}{c|}{recall} & \multicolumn{2}{c}{F1} \\ \hline
Network ID & Corr-Graph & Ours & Corr-Graph & Ours & Corr-Graph & Ours \\ \hline \hline
1NTT18927 & 0.714 & 0.898 & 0.746 & 0.76 & 0.73 & 0.823 \\
1NTT19535 & 0.775 & 0.893 & 0.701 & 0.862 & 0.736 & 0.877\\
1NTT67246 & 0.85 & 0.963 & 0.769& 0.929 & 0.807 & 0.946\\
1NTT19478 & 0.927 & 0.955 & 0.577& 0.7 & 0.711 & 0.808\\
143XB8CT & 0.495 & 0.674 & 0.366& 0.4 & 0.421 & 0.502\\
1NTT30990 & 0.184 & 0.294 & 0.477& 0.611 & 0.266 & 0.397\\
1NTT48337 & 0.47 & 0.691 & 0.614 & 0.829 & 0.532 & 0.754\\
1NTT19213 & 0.355 & 0.45 & 0.592 & 0.789 & 0.444 & 0.573\\
1NTT22885 & 0.553 & 0.85 & 0.791 & 0.949 & 0.651 & 0.897\\
1NTT18635 & 0.617 & 0.978 & 0.86 & 0.982 & 0.719 & 0.980\\ \hline
\end{tabular}
\end{table}
\begin{figure*}[htbp]
\centering
\includegraphics[width=0.95\linewidth]{figures/res1.jpg}
\includegraphics[width=0.95\linewidth]{figures/res2.jpg}
\caption{Examples of learned embedding space in 2D for different networks.}
\label{figure: res}
\end{figure*}
\end{document}
|
\subsection{Model}
We consider a tight-binding lattice model for the interlayer excitons on the TMD \moire{} superlattice, as proposed in previous works \cite{Yu2017, WuLovornMacDonald2018, Jin2019, Seyler2019, Lagoin2021}. The superlattice inherits the triangular symmetry of the underlying monolayers, and has three high-symmetry locations labeled A, B, and C, as seen in Fig.~\ref{fig:geometry}a \cite{supplementary}. Sites A and B are local energy minima hosting bound states \revised{while C is a higher local energy maximum. A and B are generally not degenerate, and their energetic ordering may depend on the choice of monolayer compounds and whether they are stacked near 0 or 180 degrees} \cite{Yu2017, WuLovornMacDonald2018}. However, for simplicity we consider only the lowest-energy locale and our results hold whether it is A or B.
\revised{Additionally, the triangular symmetry of each monolayer lends it a hexagonal Brillouin zone, with gapped valleys at the two inequivalent $\pm$K-point corners.
e--h pairs are pumped optically via vertical interband transitions, and so excitons can be photogenerated in either valley, labelled by $\tau = \pm 1$ \cite{Jin2019b}.
}
We thus consider the Bose--Hubbard \cite{Fisher1989} Hamiltonian
\begin{align} \label{eq:two-valley-Hamiltonian}
\hat{H} &
\!\!= \ns \ns \sum_{\k,\tau} \! (\!E_{0} \ns+\ns \epsilon_{\k}\!) \xdop{\k \tau\!} \xop{\k \tau\!}
\!+\!\!\ns\ns \sum_{\R, \tau,\sigma} \!\!\! \frac{U_{\!\tau \hspace{-0.075em} \sigma\!}}{2} \! \xdop{\R \tau\!} \xdop{\R \sigma\!} \xop{\R \sigma\!} \xop{\R \tau\!}
\!+\!
\hat{V}_\mathrm{LMI} \!\!\!\!
\end{align}
where $\xdop{\R \tau}$ is the bosonic creation operator of the lowest Wannier state in supercell $\R$ (at either locale A or B) \revised{and valley $\tau$}.
$\xdop{\k \tau}$, defined shortly, creates plane-wave states with dispersion $\epsilon_\k$. We use $\epsilon_\k~=~-~t[4 \cos ( k_x a /2) \cos (\sqrt{3} k_y a /2) + 2 \cos(k_x a) - 6]$, corresponding to nearest-neighbor hopping with amplitude $t>0$ and \moire{} period $a$ \cite{supplementary}, and plotted in Fig.~\ref{fig:geometry}b.
$E_0$ is the exciton formation energy (bandgap minus binding energy). $U_{\tau\sigma}>0$ are valley-dependent \cite{Rivera2016} on-site repulsion strengths; due to \moire{} localization a strongly-interacting system with $t/U_\mathrm{\tau\sigma} < 0.1$ is predicted \cite{Yu2017, Lagoin2021}, which we will treat accordingly.
$\hat{V}_\mathrm{LMI}$ is the light--matter interaction, addressed below.
A specific gauge is fixed in Eq.~\eqref{eq:two-valley-Hamiltonian}:
Exciton momentum eigenstates superpose e--h pairs with fixed momentum transfer $\p_e-\p_h=~\p$. Due to the indirect gap, the lowest-energy transition occurs at $\p=-\tau{\vec{K}}\neq\vec{0}$, where ${\vec{K}} = \vec K_h - \vec K_e$ is the momentum mismatch between the $\tau=+1$ valley extrema of the two layers \cite{Yu2015, Yu2017, IntervalleyExcitons}, coinciding with the \moire{} Brillouin zone (MBZ) corner.
It is convenient to define the shifted wave vector $\k=\p+\tau{\vec{K}}$, so real-~and $\k$-space states are related by
\begin{equation} \label{eq:fourier_transform}
\xop{{\k} \tau} = \frac{1}{\sqrt{N}} \sum_{\revised{\R}} e^{{-}i (\k - \tau{\vec{K}})\cdot \R} \xop {{\R} \tau}
\end{equation}
with $N$ the number of supercells. Indeed, the hopping amplitudes $\braket{\R \tau | \hat{H} | \R' \tau}$ are complex. Their phases are fixed by the momentum mismatch \cite{Yu2017},
which guarantees \citeapp{[cf. Appendix \ref{app:symmetries}]} that in this unique gauge $\epsilon_\k$ is valley-independent and minimal at $\k=\vec{0}$.
\revised{This momentum mismatch implies that the excitonic ground states cannot recombine radiatively: The nearly-vertical photon dispersion gives rise to the well-known optical light cone (LC) of states with $\left|\p\right| \apprle E_0 / \hbar c$ that can recombine, with $c$ the speed of light in the surrounding medium. The light cone of each valley is thereby centered at $\k = \tau {\vec{K}}$, i.e. the MBZ corners, see Fig.~\ref{fig:geometry}b. Crucially, the $\k=\vec{0}$ states lie outside both cones, and thus they are momentum-dark. }
\revised{Under a broad set of conditions, the Hamiltonian \eqref{eq:two-valley-Hamiltonian} has a many-body ground state that is Bose-condensed, with a large phase-coherent occupation of the single-particle ground states at $\k=\vec{0}$. In direct-gap systems, where these states are bright, this would lead to a pronounced phase-coherent emission similar to superradiance \cite{MoskalenkoSnoke2000}. Yet the twist-induced momentum mismatch renders the BEC a ``dark condensate'' from which direct optical emission is forbidden by translation symmetry.}
\revised{Breaking translation symmetry can enable emission from momentum-dark excitons.
This has been realized externally in TMDs, with electronic charge-order \cite{Shimazaki2021, Smolenski2021} and incommensurate substrates \cite{Joshi2021}.
Indeed, translation symmetry breaking is a hallmark of indirect-gap exciton coherence \cite{Jerome1967,Halperin1968,Kogar2017}. The interplay of two interacting valleys in Hamiltonian \eqref{eq:two-valley-Hamiltonian} can cause the excitons to break translation symmetry and form exciton density waves, which we explore in the Supplementary Material \citeapp{}. We find that while these density waves have precisely the required geometry, three-fold rotation symmetry prevents direct emission and the condensate remains dark.}
\subsection{Leaky condensates}
While its coherent component cannot emit directly, the many-body condensed ground state is not completely dark. Rather, exciton--exciton interactions induce an incoherent component that can radiate.
Emission is driven by excitonic collective modes \cite{Ozeri2005}, which have attracted recent attention as insightful probes of excitonic many-body states \cite{Utsunomiya2008,Kogar2017,Werdehausen2018ScienceAdvances,Remez2020, GoleZ2020, Estrecho2021,Bretscher2021}. Here, the collective modes supply the \revised{momentum} necessary for the excitons to recombine.
\revised{Collective modes exist in normal phases, and so Bose--coherence is not prerequisite for this mechanism. However, below we demonstrate that in twisted bilayers this is the dominant emission channel at low temperatures. Thus, many-body interaction effects will determine the optical character of the low-temperature condensed phases, which we call ``leaky condensates''.}
The essential physics of our mechanism is captured by a single-valley model.
This may be realized experimentally by pumping a valley-contrasting \revised{circularly-polarized} intralayer exciton resonance, followed by rapid interlayer charge transfer \cite{Zhang2019}.
\revised{Excitons exhibit long valley depolarization times \cite{Jiang2018,Seyler2019, Zhang2019, Scuri2020} thanks to the large e--h vertical separation \cite{Rivera2018}, suggesting the exciton population can be treated as valley-polarized over radiative timescales.}
Projecting onto valley $\tau=+1$ and suppressing $\tau$ henceforth, $\hat{H}$ reduces t
\begin{equation} \label{eq:single_valley_hamiltonian}
\hat{H} = \!\sum_{\k} \! \left(E_0 + \epsilon_\k \right)\! \xdop{\k} \xop{\k} + \frac{U}{2} \! \sum_{\R} \! \xdop{\R} \xdop{\R} \xop{\R} \xop{\R}+\hat{V}_\mathrm{LMI}. \!\!
\end{equation}
Eq.~\eqref{eq:fourier_transform} notwithstanding, all excitons now carry the same momentum mismatch so it is gauge-eliminable, and Eq.~\eqref{eq:single_valley_hamiltonian} realizes the usual Bose--Hubbard Hamiltonian.
We model the light--matter interaction by
\begin{align} \label{eq:V_LMI}
\hat{V}_\mathrm{LMI} = \!\!\!\sum_{\p_\parallel, p_\perp, \sigma}\!\!\! \hbar \omega_\p \adop{\p \sigma} \aop{\p \sigma} + (g_{\p\sigma} \adop{\p \sigma}\xop{{\vec{K}} + \p_\parallel} + \hc),
\end{align}
where $\adop{\p\revised{\sigma}}$ creates a photon with momentum $\p$ with the indicated in-~and out-of-plane components and polarization $\sigma$.
We have incorporated the exciton momentum mismatch, and made the rotating-wave approximation \cite{ScullyZubairy1997}. The coupling constants $g_{\p\sigma}$ are obtained from electronic interband transition dipole matrix elements, and depend on the exciton pairing wavefunction \cite{Choi2021} etc. The recombination rate $\Gamma_\k$ of each $\k$ mode inside the light cone can be computed \cite{Choi2021} using Fermi's Golden Rule \cite{Sakurai2017}. This defines a natural timescale, the lifetime of a localized exciton in a single \moire{} site,
$\tau^{-1}_\mathrm{loc} = \tfrac{1}{N} \sum_{\k \in \mathrm{LC}} \Gamma_\k $. \revised{The relative size of the light cone compared to the total MBZ $=\tfrac{1}{N} \sum_{\k \in \mathrm{LC}} \sim (E_0 a / \hbar c)^2$ gives the fraction of the localized wavefunction that is contained in the light cone, and provides the estimate $\tau_\mathrm{loc} \sim (\hbar c / E_0 a )^2 / \Gamma_{\vec{K}} \sim 10~\mathrm{ns}$ \cite{Yu2017}}.
Note $\tau_\mathrm{loc}$ is not the exciton mean radiative lifetime, which is associated with a thermal averaging over $\Gamma_\k$ \cite{Choi2021}.
Let us assume that exciton recombination is sufficiently slow to maintain quasi-equilibrium, with an associated chemical potential $\mu$ and grand-canonical potential $\hat{\Xi}~=~\hat{H} - \mu \hat{\mathcal{N}}$.
We will study the ground state of the excitonic sector of $\hat{\Xi}$ and treat $\hat{V}_\mathrm{LMI}$ as a weak perturbation that generates photons which probe it. This picture is made consistent by shifting the photon energies to $(\hbar \omega_\p - \mu)$ \citeapp{}.
\subsection{Weak interactions: Bogoliubov theory}
Though the excitons we consider are strongly-interacting \cite{Yu2017, Lagoin2021}, we first present our emission mechanism in the familiar weakly-interacting Bogoliubov theory to build our intuition.
A BEC with all excitons at $\k=\vec{0}$ will be depleted by interactions that eject pairs of counter-propagating excitons from the condensate.
If one lands within the light cone, it may recombine.
Bogoliubov theory lets us neatly resum these virtual processes and find the depleted ground state.
Consider a state with total filling $\nu$ and condensate filling $\nu_c(\nu)$.
Eq.~\eqref{eq:single_valley_hamiltonian} leads to the standard \cite{BECPitaevskiiStringari} mean-field (MF) Bogoliubov--de-Gennes (BdG) Hamiltonian
$\hat{\Xi}_\mathrm{MF} = \sum_\k \Omega_\k \bdop{\k}\bop{\k}$
with the familiar dispersion $ \Omega_\k^2 = {\epsilon_\k (\epsilon_\k + 2 \nu_c U)}$, $\mu=E_0 + \nu_c U$, and the Bogoliubov modes
\begin{align} \label{eq:bogoliubov_modes}
\bop{\k} & = \cosh(\theta_\k) \xop{\k} \!+\! \sinh (\theta _\k) \xdop{-\k}, \,\,\,
\sinh \theta_\k
= \frac{\Omega_\k \!-\! \epsilon_\k}{2 \sqrt{ \Omega_\k \epsilon_\k}}.\!\!
\end{align}
%
The ground state of the theory is the BdG vacuum. Yet the mixing of particle creation and annihilation in Eq.~\eqref{eq:bogoliubov_modes} implies that it nevertheless contains some excitons, most notably inside the light cone.
Consider recombination in terms of the collective modes. Transcribed into BdG modes, Eq.~\eqref{eq:V_LMI} contains terms such as
$- g_{\p\sigma} \sinh (\theta_{{\vec{K}} + \p_\parallel}\!) \adop{\p \sigma} \bdop{-{\vec{K}} - \p_\parallel}$, which
represent a spontaneous emission of a photon and a BdG mode, the latter assuring momentum conservation in analogy to phonon-assisted exciton recombination \cite{MoskalenkoSnoke2000}. Therefore, interactions enable an otherwise-dark exciton condensate to ``leak'' photons with
\begin{equation} \label{eq:leaky_emission_spectrum}
\hbar \omega_\p = \mu - \Omega_{-{\vec{K}} - \p_\parallel} = E_0 + \nu_c U - \Omega_{-{\vec{K}} - \p_\parallel} < E_0\,.
\end{equation}
As expected, some energy is lost to the Bogoliubov mode.
We compute the total emission rate $\Gamma$ with Fermi's Golden Rule via transitions between the BdG vacuum and single quasiparticle states. The redshift in Eq.~\eqref{eq:leaky_emission_spectrum} is negligible compared to $E_0$,
so this is equivalent to counting the number of excitons present within the light cone,
\begin{align} \label{eq:leaky_emission_rate}
\Gamma &
\approx \sum_{\k \in \mathrm{LC}} \Gamma_\k n_\k
\approx N \tau^{-1}_\mathrm{loc} \, n_{\vec{K}}
\approx
N \tau^{-1}_\mathrm{loc} \Big(\dfrac{1}{18} \dfrac{\nu_c}{t/U}\Big)^2.
\end{align}
Here we approximated $n_{\k} = \braket{\xdop{\k}\xop{\k}}\approx n_{\vec{K}} =
\sinh^2 \theta_{{\vec{K}}}$ inside the light cone, assuming it is much smaller than the MBZ, and expanded around small densities. The total filling is found by integrating $n_\k$ \cite{BECPitaevskiiStringari}, and in two dimensions $\nu_c = \nu [1 - \mathcal{O}(U/t)]$. Thus, unlike spontaneous decay, $\Gamma$ is quadratic in density.
The generality of this construction suggests that exciton condensates in any indirect-gap system are leaky. Moreover, emission is accomplished without additional degrees of freedom. This contrasts with external optical probing \cite{Combescot2014a} and other mechanisms that involve phonons \cite{Danovich2016} or carrier exchange in larger exciton complexes \cite{Danovich2017}.
Interactions deplete excitons into excited bands as well, and so leaky condensates can also occur in systems that are dark due to a spin-forbidden transition, etc. Previous studies have shown that, for sufficiently strong interactions (or, equivalently, above a threshold density), a dark condensate can transition into a so-called ``gray condensate'' \cite{Combescot2012, Alloing2014, Combescot2017, MazuzHarpaz2019}. Our finding of leaky emissions is distinct from these previous works:
(i)~Emission from the leaky condensate grows continuously with increasing density, with no threshold value.
\revised{(ii)~In indirect-gap materials, the bright and (momentum-)dark states are smoothly connected along the same Bloch band, unlike the usual scenario where they are spin-split and form separate bands. This prevents fragmentation into a gray condensate.}
(iii)~The gray condensate emission is coherent whereas the leaky condensate emission is not, due to its entanglement with the generated collective modes.
Condensate depletion by interactions has recently gained significant experimental attention \cite{Xu2006, Chang2016,Lopes2017, Pieczarka2020, Steger2021}.
Relatedly, BdG modes are used to renormalize phonon-assisted photoluminescence line shapes, e.g. in Cu$_2$O \cite{MoskalenkoSnoke2000, Haug1983, Shi1994}. However, the role of collective modes in {\emph{enabling}} recombination of indirect-gap excitons, to the best of our knowledge, has not been pointed out so far.
Additionally, these descriptions focus on weakly-interacting excitons.
\subsection{Strong interactions: hard-core bosons}
We now explore how leaky condensates manifest under strong interactions. Consider the $U \to \infty$ limit corresponding to hard-core bosons, which with the transformation ($S=\frac{1}{2}$ henceforth implied)
\begin{equation} \label{eq:spin_operators}
\sop{\R}^- = e^{-i{\vec{K}}\cdot\R} \xop{\R},\,\,
\sop{\R}^+ = e^{i{\vec{K}}\cdot\R} \xdop{\R},\,\,
\sop{\R}^z= \xdop{\R} \xop{\R} \!-\! S,
\end{equation}
map to a spin-$\frac{1}{2}$ XX ferromagnet in a transverse field. The grand-canonical potential becomes
\begin{align} \label{eq:XXZ-Hamiltonian}
\hat{\Xi} &
\!=\! -t \!\!\!\! \sum_{\langle \R,\R'\rangle} \!\!\!\! ( \sop{\R'}^+ \sop{\R}^- \!+\!
\sop{\R}^+ \sop{\R'}^- )
+ (E_0 \!-\!\mu ) \! \sum_{\R} \! (\sop{\R}^z \!+\! S ).
\end{align}
The recombination rate remains $\Gamma\tau_\mathrm{loc} /N~\approx n_{{\vec{K}}}~=~\braket{\sop{{\vec{K}}}^+ \sop{{\vec{K}}}^-}$ where $\sop{{\vec{K}}}^-~=~\frac{1}{\sqrt{N}} \sum_\R e^{-i{\vec{K}}\cdot\R} \sop{\R}^-$.
This limit readily manifests ground state emission: With $t = 0$, Eq.~\eqref{eq:XXZ-Hamiltonian} factorizes into independent sites each with MF solution $\sqrt{\nu}\ket{\Uparrow} + \sqrt{1-\nu} \ket{\Downarrow}$, yielding \citeapp{} $n_{{\vec{K}}}(\nu)=\nu^2$. As nonzero hopping allows repelling excitons to separate, leading to \emph{anti}correlations, this is an upper bound. Additionally, an emergent particle--vacancy duality \citeapp{} $\sop{\R}^\pm \to \sop{\R}^\mp$ connects the ground states of Eq.~\eqref{eq:XXZ-Hamiltonian} with fillings $\nu$ and $1-\nu$, providing the identity $n_{\vec{K}} (\nu) - n_{\vec{K}} (1-\nu) = 2\nu - 1$. Thus, $n_{\vec{K}}(\nu)>\mathrm{max}(2\nu -1,0)$. These bounds already demonstrate the interaction-driven nonlinearity of $\Gamma (\nu)$.
A quantitative treatment of emission is again found in terms of spontaneously excited collective modes, now taking the form of spin waves. We perform a Holstein-Primakoff (HP) $1/S$ expansion \cite{HolsteinPrimakoff1940} similar to Bernardet et~al.~\cite{Bernardet2002} with details provided in the Supplemental Material \citeapp{}.
The qualitative features of Bogoliubov theory are reproduced: we obtain a quadratic $\hat{\Xi}=\sum\Omega_\k \bdop{\k}\bop{\k}$, now with dispersion $\Omega_\k^2 = \epsilon_\k [(2\nu_\mathrm{MF}-1)^2\epsilon_\k + 24\nu_\mathrm{MF}(1-\nu_\mathrm{MF}) t] $ and $\mu = E_0 + 6(2\nu_\mathrm{MF}-1)$, where the filling $\nu_\mathrm{MF}$ is the mean-field order parameter [cf. $\nu_c$ in the previous section].
The exciton occupation \revised{in the spin-wave vacuum, or equivalently at temperature $T=0$, is now}
\begin{equation}
n_\k
= \left[\nu_\mathrm{MF} \cosh \theta_\k + (1-\nu_\mathrm{MF}) \sinh \theta_\k\right]^2.
\end{equation}
$\theta_\k$ and the emission spectra are still given by Eqs.~\eqref{eq:bogoliubov_modes} and \eqref{eq:leaky_emission_spectrum}.
\begin{figure}[tb]
\includegraphics[width=\apsfigurewidth]{fig2}
\caption{Emission rate $\Gamma(\nu)$ \revised{at $T=0$} versus filling \cite{densities} within the Holstein--Primakoff expansion \citeapp{} (solid black), compared with exact diagonalization on a 6$\times$3 lattice (filled circles) and a least-squares fit derived from two-body states (dashed blue).
The solid and dashed gray lines are upper and lower bounds explained in the text.
Inset: The crossover temperature $T^\star$ between interaction- and thermally-dominated emission.
The solid and dashed thin lines mark the Berezinskii–Kosterlitz–Thouless superfluid transition and quantum degeneracy crossover, estimated at $kT/t=\sqrt{3}\pi\nu$ and $4\sqrt{3}\pi\nu$, respectively \citeapp{}. Interactions dominate the cold regimes.}
\label{fig:hard-core_boson}
\end{figure}
We plot the \revised{$T=0$} emission rate $\Gamma\propto n_{\vec{K}}$ in Fig.~\ref{fig:hard-core_boson} \cite{densities}. Comparing it against small-scale exact diagonalization, we find very good agreement across a wide range of fillings.
In the dilute limit
\begin{equation} \label{eq:hard-core_bosons_emission_rate}
\Gamma(\nu \ll 1) \approx
\frac{4}{9} N \tau_\mathrm{loc}^{-1} \nu_\mathrm{MF}^2 .
\end{equation}
The apparent quadratic dependence in Eq.~\eqref{eq:hard-core_bosons_emission_rate} does not imply an absence of correlations, which cause
$\nu_\mathrm{MF}(\nu)$ and $\nu$ to differ. Exciton correlations can be inferred from two-body states \cite{MazuzHarpaz2019}, from which we deduce \citeapp{} the asymptotic form $\Gamma \sim (\nu/\log \nu)^2$, revealing the expected suppression.
Numerics confirm \citeapp{} that $\nu_\mathrm{MF}\sim(\nu/\log \nu)$ for $\nu\ll1$, indicating correlations are successfully captured.
Our HP theory also allows us to treat nonzero temperatures. Thermal excitations enhance emission and unlock a second channel whereby a collective mode is absorbed instead of emitted, leading to two emission lines.
In the inset of Fig.~\ref{fig:hard-core_boson} we plot the crossover temperature $T^\star$ at which depletion (due to interactions) and thermal excitations contribute equally \citeapp{} to the total emission rate, and below which interactions dominate.
This crossover occurs above the Berezinskii--Kosterlitz--Thouless (BKT) transition \cite{Berezinskii1972, KosterlitzThouless1973}, indicating that \revised{leaky emission will be the dominant emission channel characterizing the superfluid phase and much of the quantum-degenerate regime}.
\revised{Remarkably, in this strongly-interacting system, at small filling ``leaks'' dominate emission even in the hot gas phase that can be treated semi-classically.}
Unlike the Stokes and anti-Stokes lines in phonon-assisted emission \cite{MoskalenkoSnoke2000}, the two BdG processes have unequal matrix elements with different density dependences. The anti-Stokes-like line dominates above $T^\star$, and the net annihilation of BdG modes may evaporatively cool the BEC, similarly to a mechanism recently suggested \cite{Misra2021}.
Finally, the strength of interactions can be inferred from the ratio of emission line intensities \citeapp{}.
The leaky condensate picture thus predicts a distinctive property of \moire{} exciton emission: A dominant redshifted emission line with quadratic density dependence below $T^\star$, compared to a dominant blueshifted line with linear density dependence above $T^\star$.
We do not expect a qualitative change at the BKT transition.
\subsection{Experimental Consequences}
We assess the parameters under which a leaky condensate may be observed.
Comparing Eq.~\eqref{eq:leaky_emission_rate} with \eqref{eq:hard-core_bosons_emission_rate} suggests that this mechanism saturates once $U \apprge 12 t$, which should hold across a wide range of twist angles \cite{Yu2017}.
For $a \approx 10~\mathrm{nm}$ (twist $\approx 2^\circ$) and corresponding $t \approx 0.2~\mathrm{meV}$ \cite{Yu2017}, and at a demonstrated \cite{Wang2018a} photoexcited exciton densities of $n \approx 10^{11}~\mathrm{cm}^{-2}$ ($\nu \approx 0.1$), $T^\star \approx 5~\mathrm{K}$.
Larger twist angles or intercalated hBN spacers would increase $t$ and thus $T^\star$ \cite{Yu2017, Lagoin2021}, and mitigate inhomogeneity effects. Furthermore, electrostatic gating might modify the \moire{} symmetry \cite{Yu2017} or its elastic reconstruction \cite{Andersen2021, BennettRemez2021}, allowing in-situ tunability.
Above $T^\star$ the two emission lines are split by $\sim 20t\sim 4~\mathrm{meV}$ and should be resolvable.
Leaky emission will also manifest in a quadratic loss $\partial_t n = - \gamma n^2 $, with $n$ the exciton number density. With Eq.~\eqref{eq:hard-core_bosons_emission_rate} and $\tau_\mathrm{loc}\sim10~\mathrm{ns}$ at $a\sim10~\mathrm{nm}$ \cite{Yu2017} we estimate a rate constant $\gamma\sim a^2/\tau_\mathrm{loc} \approx 10^{-4}~ \mathrm{cm^2/s}$.
Exciton density is controlled with pumping fluence, and tracked with transient absorption \cite{Sun2014, Poellmann2015}, emission blueshift [cf.~Eq.~\eqref{eq:leaky_emission_spectrum}] \cite{Alloing2014,Combescot2014a,Wang2018a}, or time-resolved photoluminescence (PL) \cite{Wang2018a}. These reveal such quadratic dependence, which is attributed to Auger recombination. In TMDs this process is associated with \revised{a high-energy conduction band that supports bound excitons at energies close to $2 E_0$, making Auger recombination nearly resonant} \cite{Steinhoff2021, Lin2021}. Yet crucially, for nonzero detuning Auger recombination freezes out at $T\to0$ \cite{Erkensten2021} and leaky emission will dominate. Comparisons at nonzero temperature are not straightforward; estimates are available mostly for room-temperature monolayers, suggesting an intrinsic Auger constant $\gamma \sim 10^{-3}~\mathrm{cm^2/s}$ \cite{Hoshi2017, Zipfel2020, Steinhoff2021}. However, moving to cryogenic temperatures or to bilayers could each reduce $\gamma$ by orders of magnitude \cite{Erkensten2021,Yuan2015}. Thus, the dominance of leaky recombination over Auger could feasibly extend to $T^\star$ and above.
This unique regime of emission linear in continuous fluence yet nonlinear in \emph{instantaneous} density might be important in interpreting PL experiments. In settings demonstrating population lifetimes approaching microseconds \cite{Miller2017, Nagler2017, Jiang2018, Montblanch2021}, the short "leaky lifetime" $\tau_\mathrm{leaks}\sim (\gamma n)^{-1} \sim 10$ ns induced at high fluences could also play a significant role in the complex population dynamics.
\subsection{Conclusions}
We have shown that strong interactions challenge the picture of dark condensates in \moire{} bilayers, where they dominate optical processes. This prompts further material-specific modelling to compare these effects to other recombination and loss mechanisms.
\revised{Additionally, while here we mostly considered a single-valley model, the physics of two-valley \moire{} condensates is very rich. As we remark above, excitonic density waves put such condensates on the verge of optical activation which is prevented only by vestigial rotational symmetry \citeapp{}. Therefore, direct emission with long-range phase coherence may be achieved in these systems by external fields, strain, layer separation, and pressure, with possible sensing applications. We leave this and other novel intervalley phenomena to future work.}
\subsection{Acknowledgments}
We thank M.~Atat\"ure, D.~Kara, C.~M.~Pursar, A.~R.-P.~Montblanch, T.~F.~Heinz, O.~Karni, E.~Barr\'e, A.~Camacho-Guardian and D.~Bennett for fruitful discussions. The support of the Cambridge International Trust, of EPSRC Grant Nos. EP/P009565/1, EP/P034616/1 and of a Simons Investigator Award are gratefully acknowledged.
\bibliographystyle{apsrev4-1}
|
\section{Introduction}
\label{sec:intro}
\textbf{\textit{Ex vivo}} MRI of the brain provides remarkable advantages over \textit{in vivo} MRI for visualizing detailed neuroanatomy and linking macroscopic morphometric measures such as cortical thickness to underlying cytoarchitecture and pathology \cite{mancini2020multimodal}. It helps in characterizing the underlying anatomy at the scale of subcortical layers \cite{augustinack2013medial}, such as hippocampal subfields in the medial temporal lobe (MTL) \cite{yushkevich2021three, ravikumar2020building}. Compared to \textit{in vivo} MRI, \textit{ex vivo} MRI is not affected by head or respiratory motion artifacts and has much less stringent time and specific absorption rate constraints. Compared to histology, it does not suffer from distortion or tearing of brain tissue, thereby giving flexibility in acquiring ultra-high resolution images. Indeed, \textit{ex vivo} MRI is often used to provide a 3D reference space onto which to map 2D histological images. Combined analysis of \textit{ex vivo} MRI and histology makes it possible to link morphological changes in the brain to underlying pathology as well as to generate anatomically correct parcellations of the brain based on cytoarchitecture \cite{schiffer2021convolutional}, and pathoarchitecture \cite{augustinack2013medial}.
There has been substantial work in brain MRI parcellation such as \emph{FreeSurfer} \cite{fischl2012freesurfer} and recent efforts based on deep learning \cite{henschel2020fastsurfer, chen2018voxresnet}. However, these approaches focus on \textit{in vivo} MRI, and limited work has focused on developing automated segmentation methods for \textit{ex vivo} MRI segmentation. \textit{Ex vivo} segmentation methods have been region specific. Recent developments include automated deep learning methods for high resolution cytoarchitectonic mapping of the occipital lobe in 2D histological sections \cite{schiffer2021convolutional}. The work by \cite{iglesias2015computational} has developed an atlas to segment the MTL using manual segmentations in \textit{ex vivo} images. Yet, an \textit{ex vivo} segmentation method applicable to a variety of brain regions has yet to be described. Limited availability of \textit{ex vivo} 3D MRI segmentation algorithms may be explained by only a few groups focusing on whole brain \textit{ex vivo} image analysis and hence limited availability of specimens, scans, and labeled ground truth segmentations; greater heterogeneity in scanning protocols when compared to \textit{in vivo} structural MRI; larger image dimensions, greater textural complexity, and more profound imaging artifacts than in \textit{in vivo} MRI.
\begin{figure*}[t!]
\includegraphics[width=\textwidth,height=8cm]{cortex.png}
\caption{Thickness measurement pipeline and patch-level gray matter ground truth data. (A) Cortical thickness is measured at the 16 out of the 18 landmarks. \textit{Note that we do not measure thickness at Cornu Ammonis 1 (CA1) and subiculum as their segmentation requires Stratum Radiatum, Lacunosum, and Moleculare (SRLM) segmentation, which is currently not performed.} (B) A dot (shown: the motor cortex dot) is first placed to define an anatomical landmark, around which a semi-automatic level set segmentation of the surrounding cortical ribbon is provided. A maximally inscribed sphere is then computed using Voronoi skelentonization, and the diameter of the sphere gives thickness at that landmark. (C) Ground truth training data acquired by manually segmenting gray matter in 64 x 64 x 64 patches shown for two subjects (1 and 2) in three viewing planes with 3D renderings displayed in 1a and 2a respectively.}
\label{fig:dots}
\end{figure*}
\textbf{In this work, we present a novel dataset of 32 high resolution (0.3 x 0.3 x 0.3 mm$^{3}$) 7~Tesla \textit{ex vivo} MRI scans of whole brain hemispheres of older adult patients with Alzheimer's Disease or Related Dementias (ADRD) or cognitively normal adults. We then benchmark nine deep learning neural architectures to segment cortical and sub-cortical gray matter in whole brain hemispheres, with limited patch-based training data.} We measure cortical thickness at several key locations in the cortex and correlate these automated measures with thickness measurements obtained using a user-guided semi-automated protocol. High consistency between these two sets of measures supports the use of deep learning based automated thickness measures for \textit{ex vivo} brain morphometry. Additionally, we show that networks trained on T2w images acquired at 7~Tesla are able to generalize to \textit{ex vivo} images obtained with T2*-weighted (T2*w) gradient echo images acquired at 7~Tesla, and \textit{ex vivo} images acquired at a lower field strength of 3~Tesla.
\begin{figure*}[!t]
\includegraphics[width=\textwidth,height=6.5cm]{figure_2_with_Arrows.png}
\caption{Cortical mantle segmentation across whole brain hemispheres of the different neural network architectures, on the 5th cross validation fold, for a given subject on the left panel. (A) an axial slice. (B) Attention Unet. (C) VNet. (D) VoxResNet. (E) 3D Unet. (F) AnatomyNet (CE + SE). (G) AnatomyNet (Vanilla). (H) AnatomyNet (CE). (I) AnatomyNet (SE). (J) nnU-Net. Right panel: (K-L) nnU-Net based segmentation shown on a sagittal plane for two subjects. (M) 3D rendering of the cortex from lateral and medial views for the subject shown in K.}
\label{fig:all_results_compare}
\end{figure*}
\section{Materials}
\label{sec:materials}
\textbf{Image Acquisition.} We analyze a dataset of 32 \textit{ex vivo} whole-hemisphere MRI scans. Patients were selected for the study from our ongoing research autopsy program. Data was drawn from 11 females (Age: 64-94) and 21 males (Age: 54-97) with Alzheimer's Disease or Related Dementias (ADRD) or cognitively normal adults. Human brain specimens were obtained in accordance with the University of Pennsylvania Institutional Review Board guidelines. Specimens were scanned after atleast a 4 week fixation period. T2w images were acquired using a 3D-encoded T2 SPACE sequence with 0.28 mm isotropic resolution, 3 s repetition time (TR), echo time (TE) 383 ms, turbo factor 188, echo train duration 951 ms, bandwidth 348 Hz/px. All data was acquired on a Siemens MAGNETOM Terra 7~Tesla scanner using a custom birdcage transmit/receive coil. Sample slices are shown in Fig. \ref{fig:all_results_compare}.
\textbf{Patch-level Gray Matter Segmentation.} To train the neural networks, we sampled five 3D image patches of size 64 x 64 x 64 around the orbitofrontal, anterior temporal, inferior prefrontal, primary motor, and primary somatosensory cortices from 6 brain hemispheres, resulting in a total of 30 patches. Fig. \ref{fig:dots} C shows sample patch images and the corresponding ground truth labels with 3D renderings. Five manual raters, divided into groups of two and three, labeled gray matter as the foreground, and rest of the image as the background using a combination of manual tracing and the semi-automated segmentation tool in ITK-SNAP software \cite{yushkevich2019user}. Inter-rater reliability scores were computed for these manual segmentations in terms of Dice Coefficient (DSC): Raters 1\&2: 95.26 $\pm$ 1.37 \%, Raters 1\&3: 94.64 $\pm$ 1.64 \%, Raters 2\&3: 94.54 $\pm$ 1.20 \%, Raters 4\&5: 92.04 $\pm$ 4.26 \%.
\begin{figure*}[!t]
\centering
\includegraphics[scale=1]{plots.png}
\caption{Cortical thickness measurements at the 16 cortical regions. Shown are correlation plots between cortical thickness measured by automated nnU-Net (y-axis) and ground truth manual segmentations (x-axis). Within each plot, we tabulate Pearson's correlation coefficient (r), p-value, and the Average fixed raters Intra-class correlation coefficient (ICC) scores.}
\label{fig:graphs}
\end{figure*}
\textbf{Thickness measurements at key cortical locations.}
To obtain localized quantitative signatures of cortical morphometry at our Center, in each of the 32 hemispheres, we identified 18 cortical landmarks including 5 in the medial temporal lobe (Fig. \ref{fig:dots} A): visual, midfrontal, orbitofrontal, motor, anterior and posterior cingulate, superior and ventrolateral temporal, anterior temporal pole, anterior insula, inferior frontal, angular gyrus, superior parietal, Entorhinal Cortex (ERC), Brodman Area 35 (BA35), Parahippocampal Cortex (PHC), Cornu Ammonis 1 (CA1) and subiculum. These locations were chosen to gather neuropathology data which is part of a separate ongoing research project. To measure cortical thickness at these locations, we use the pipeline developed in \cite{wisse2021downstream}, shown in Fig. \ref{fig:dots} B.
\section{Methods}
\label{sec:methods}
We benchmark variants of popular biomedical image segmentation deep learning models: (1.) nnU-Net \cite{isensee2021nnu}; four variants of AnatomyNet \cite{zhu2019anatomynet} based on squeeze-and-excitation blocks \cite{rickmann2019project}: (2.) Spatial excitation AnatomyNet (SE), (3.) Channel excitation AnatomyNet (CE), (4.) AnatomyNet (Vanilla) \cite{zhu2019anatomynet}, (5.) Channel-spatial excitation AnatomyNet (CE + SE); (6.) 3D Unet-like network \cite{khandelwal2020domain}; (7.) VoxResNet \cite{chen2018voxresnet}; (8.) VNet \cite{milletari2016v} ; and (9.) Attention Unet \cite{oktay2018attention}. We use PyTorch 1.5.1 and Nvidia Quadro RTX 5000 GPUs to train the models using user-annotated patches described above. Patches were standardized, and then normalized between 0 and 1. We implemented all the network architectures within the nnU-Net framework \cite{isensee2021nnu} and thus systematically evaluated the nine deep learning architectures in a five-fold cross-validation experiment under matched conditions (i.e., same split of data into training/validation/testing subsets; same data augmentation strategy, same hyper-parameter tuning strategy).
\begin{comment}
For all the models, except the off-the-shelf nnU-Net, which has its own internal optimization procedure, we use a batch size of 2, Adam optimizer with learning rate 0.001, weight decay of 0.00005 for 30 epochs until convergence with Generalized Dice Coefficient as the loss function. We performed the following random data augmentation to each image patch: flipping, rotation, and elastic deformation. Inference was performed on whole hemisphere images, using the model which gave best validation accuracy across the epochs, via the patch-wise sliding window.
\end{comment}
\textbf{Evaluation.} First, we compare the performance of the deep learning architectures at patch-level by reporting Dice Coefficient (DSC) and Hausdorff Distance 95th percentile (HD95) in a five-fold cross-validation setting for the 30 patches. We then employ the best performing model, based on qualitative results as shown in Fig. \ref{fig:all_results_compare}, to segment the cortical mantle in whole hemispheres. We compute the thickness of the cortical mantle around 13 landmarks as described in Section \ref{sec:materials}. We then correlate the cortical thickness of manual, and automated segmentations via Pearson's correlation coefficient with t-distribution as the test statistic reporting the p-value (with 0.05 significance level), and the Average fixed raters Intra-class Correlation Coefficient (ICC) for the 13 cortical locations.
\section{Results and Discussion}
\label{sec:results}
\subsection{Deep learning segmentations}
\label{sec:results_DL}
Table 1 tabulates the performance of different networks across 6-fold cross validation. In terms of DSC, AnatomyNet and its variants performs at par with VoxResNet, and outperforms the rest of the networks. But, superior DSC performance does not translate equally well to whole hemisphere segmentations as evident by qualitative results in Fig. \ref{fig:all_results_compare}. We observe that inferior performing models (\ref{fig:all_results_compare} B-E) mislabel white matter as gray matter. The AnatomyNet and its variants (\ref{fig:all_results_compare} F-I) are able to distinguish gray matter from white matter, but fail to segment the low intensity anterior and posterior regions, virtue of coil limitations, shown in white arrows. There are also some under-segmentations of the cortex (white arrows). Fig. \ref{fig:all_results_compare} J-L depicts that the best performing model is nnU-Net, which clearly demarcates GM/WM boundary, segments regions with low signal, which were \textit{not} included in the training patches, making the performance of nnU-Net even more remarkable. Fig. \ref{fig:all_results_compare} M shows 3D renderings of whole cortical segmentation. Hence, we use nnU-Net to segment the cortical mantle in whole hemisphere across the 32 subjects.
\begin{comment}
Clearly, nnU-Net outperforms the rest by a large margin in terms of DSC of 98.52 $\pm$ 5.84 $\%$ and HD95 of 0.28 $\pm$ 0.00 mm.
\end{comment}
\begin{table}[t]
\label{table:results}
\caption{Five-fold cross validation Dice Coefficient (DSC) and Hausdorff Distance 95th percentile (HD95) scores between ground truth and automated patch-level segmentations.}
\centering
{\begin{tabular}{c|c|c}
\makecell{\textbf{Deep learning method}} & \makecell{\textbf{DSC} \textbf{(\%)}} & \makecell{\textbf{HD95} \textbf{(mm)}} \\
\hline
nnU-Net &
93.98 $\pm$ 5.25 &
0.49 $\pm$ 0.45 \\
\hline
AnatomyNet (SE) &
94.84 $\pm$ 3.84 &
0.45 $\pm$ 0.42 \\
\hline
AnatomyNet (CE) &
94.91 $\pm$ 3.27 &
0.45 $\pm$ 0.42 \\
\hline
AnatomyNet (Vanilla) &
94.86 $\pm$ 3.83 &
0.46 $\pm$ 0.44 \\
\hline
AnatomyNet (CE + SE) &
94.66 $\pm$ 3.79 &
0.47 $\pm$ 0.44 \\
\hline
3D Unet &
93.57 $\pm$ 5.22 &
0.58 $\pm$ 0.51 \\
\hline
VoxResNet &
94.84 $\pm$ 4.00 &
0.45 $\pm$ 0.42 \\
\hline
VNet &
90.84 $\pm$ 5.93 &
0.99 $\pm$ 0.56 \\
\hline
Attention Unet &
93.65 $\pm$ 4.91 &
0.62 $\pm$ 0.66
\end{tabular}}
\end{table}
\subsection{Cortical Thickness Measurements}
\label{sec:results_cortical}
We correlate thickness (mm) between ground truth and automated nnU-Net based segmentations at 16 out of the 18 landmarks. Note that we do not measure thickness at Cornu Ammonis 1 (CA1) and subiculum as their segmentation requires Stratum Radiatum, Lacunosum, and Moleculare (SRLM) segmentation, which is currently not performed. Fig. \ref{fig:graphs} shows good agreement between ground truth and automated thickness with 8 regions having correlation coefficient (r) greater than 0.6 with 15 regions reaching statistical significance at p $<$ 0.05, except in inferior frontal region. We also observe high ICC scores with 9 regions having ICC greater than 0.7, which confirms that automated segmentations are accurate to give desirable cortical thickness measurements.
\begin{comment}
\begin{figure}[t!]
\begin{minipage}[b]{1.0\linewidth}
\centering
\centerline{\includegraphics{figure_3.png}}
\end{minipage}
\caption{Generalization to other unseen imaging sequences. (A-B) image acquired using T2*w gradient echo, (C-D) 3~Tesla image at lower resolution of 0.5 x 0.5 x 0.5 mm$^{3}$.}
\label{fig:results_other_data}
\end{figure}
\end{comment}
\begin{figure*}[!t]
\centering
\includegraphics[scale=0.6]{figure_3.png}
\caption{Generalization to other unseen imaging sequences. (A-B) image acquired using T2*w gradient echo, (C-D) 3~Tesla image at lower resolution of 0.5 x 0.5 x 0.5 mm$^{3}$.}
\label{fig:results_other_data}
\end{figure*}
\subsection{Generalization to other imaging sequences}
\label{sec:results_generalize}
In Fig. \ref{fig:results_other_data}, we qualitatively show that our model trained on 7~Tesla 0.3 x 0.3 x 0.3 mm${^3}$ T2w images, is able to generalize well to MRI sequences and resolutions unseen during training. Segmentation results on T2*w gradient echo \textit{ex vivo} images, acquired \cite{tisdall2021joint} at 0.28 mm and 0.16 mm isotropic resolution, are shown in Fig. \ref{fig:results_other_data} A and B respectively. Fig. \ref{fig:results_other_data} C and D show that our model is able to segment gray matter in the publicly available \textit{ex vivo} T2w image acquired at 3~Tesla at a lower resolution of 0.5 x 0.5 x 0.5 mm$^{3}$ \cite{mancini2020multimodal}.
\section{Conclusion and Future Work}
\label{sec:future_work}
Our results show that even using limited patch-level training data from six subjects, nnU-Net (and to a lesser extent AnatomyNet) is able to generate high-quality segmentations of cortical and subcortical gray matter in \textit{ex vivo} MRI of brain hemispheres, generalizing well to areas of low contrast unseen during training, as well as to other MRI protocols, field strengths, and resolutions. Moreover, thickness measures derived from nnU-Net segmentations concur with user-supervised thickness measurements, suggesting the feasibility of fully automated cortical thickness analysis in \textit{ex vivo} MRI analogous to the way \emph{FreeSurfer} is used for \textit{in vivo} MRI morphometry. A limitation of our approach is that it does not separate subcortical gray matter from the cortex. In future work, we intend to address this limitation using anatomical priors, and develop techniques for groupwise normalization of \textit{ex vivo} MRI and correlate cortical thickness with neuropathology.
\vfill
\pagebreak
\bibliographystyle{IEEEbib}
|
\section{Introduction}
\Ila{Copiato dall'art vecchio}
One of the main research topics in Business Process Management (BPM) concerns regulatory or Compliance Management (CM), i.e. the analysis of compliance to norms [Dumas et al., 2018; Van der Aalst, 2013]. The necessity of satisfying regulations or laws forces organizations in redesign their internal processes, in the context of change management [Hayes, 2014]. The increasing pressure from regulatory authorities to organizations led to the development and application of Compliance Management Frameworks (CMFs). In this context, CM can be addressed at the operational level by focusing on business processes, intended as the set of activities accomplishing a specific organizational goal.
Business process analysis usually introduces performance objectives to be considered in addition to constraints imposed by external pressures (e.g., regulatory issues). The investigation of undesirable events and norm violations adopted traditional techniques, e.g. root cause analysis (commonly used in manufacturing processes to improve performance). More recently, CMFs explore the relationship between the formal representation of a process model and the relevant regulations. There are many different adoptable CM strategies consisting in approaches to automate the checking whether a business process complies with the actual regulation [Governatori and Sadiq, 2009]. The goal is to ensure that such approaches properly model business processes as well as norms. Moreover, in the past decades many CM approaches in the context of digitization to automatize business processes have been proposed [Sackmann et al., 2018]. We describe here a CM approach to support regulatory compliance for healthcare business processes based on a compliance-by-design methodology [Governatori and Sadiq, 2009] and using a business process compliance checker called Regorous [Governatori, 2015b]. In particular, this paper explores the adoption of a two-step pipeline introducing a CMF applied to an innovative hospital service. In a first step, business process analysis can be performed by adopting standard modeling language to investigate healthcare processes at operational level. In a second step, a regulatory CM is proposed on the top of the model by applying a logic-based approach to automate checking whether the process complies with the new General Data Protection Regulation (GDPR).
The rest of this paper is structured as follows: background and related work, detail of the analysis of the case study on Business Process Management prospective and Regulatory Compliance prospective and finally, results and discussions on future work.
\section{RELATED WORK}
\Ila{Copiato da articolo vecchio}
\subsection{Risk management and regulatory compliance}
Risk is part of every business activity and therefore part of every business process [Van der Aalst, 2013; Hashmi et al., 2016]. The occurrence of a risk may lead to loss of quality, increased costs, time delays, complaints, and legal problems [Betz et al., 2011] as well as, in healthcare, serious and permanent damages up to death. There are several type of risks, as legal, procedural, economical, financial, etc. The Risk Management is the discipline that allows the management of these different kinds of risks by the applications of some principles [Sadgrove, 2016; McNeil et al., 2015; Haimes, 2015].
Regarding legal risks, it should be considered that the process has to be compliant to law, whereas norms and regulations are constantly evolving and new reorganizations must be implemented with the introduction of new procedures [Hayes, 2014], i.e., for privacy control, AI technologies.
Compliance in healthcare considers the conformity of care processes with laws, regulations and standards related to patient safety, privacy of patient information and administrative practices.
Ultimately, health compliance is about providing safe and high quality patient care. Healthcare organizations are also required to comply with strict standards, regulations and laws at regional and state level. Violations of these laws may result in legal action, heavy fines or loss of licenses.
It is possible to find several studies on compliance with laws, rules or regulations in the case of processes related to patient health [Buddle et al., 2005; Racz et al., 2010; Amantea et al., 2018].
The intensive use of ICT solutions to collect, share and digitize data of a health process, makes it necessary to prepare tools able to identify any possible risk scenario related to the use of computer systems and lack of awareness on the agents, as well as to facilitate the adoption of appropriate counter-measures. Previous research on IT in healthcare explored digitalization challenges for organization [Amantea et al., 2018]. These innovations may require the application of new regulations, such as the GDPR, without forgetting that the health sector is full of strictly health regulations in constant evolution.
\subsection{Business process compliance and logic}
Regulatory compliance is the set of activities an enterprise does to ensure that its core business does not violate relevant regulations, in the jurisdictions in which the business is situated, governing the (industry) sectors where the enterprise operates. The activities an organization does to achieve its business objectives can be represented by the business processes of the company. On the other hand, a normative document (e.g., a code, a guide line, an act) can be understood as a set of clauses, and these clauses can be represented in an appropriate formal language.
\subsubsection{Business process modeling}
In order to analyse the business processes of this type of home hospitalization, we exploit a Business Process Management (BPM) methodology. One of the central issues in BPM is change management [Van der Aalst et al., 2010; Abo-Hamad and Arisha, 2013; Di Leva et al., 2020]. Using a process-centric approach, due to describe the diagram of the process, we will adopt the Business Process Model and Notations (BPMN) standard language [Allweyer, 2016]. Primarily, in the context of healthcare studies, BPMN standard language acquires a peculiar consideration [Muller and Rogge-Solti, 2011; Amantea et al., 2020; Sulis and Di Leva, 2018].
The business process analysis aims to define and engineer a model to be verified and validated with system experts. One of the main output is the creation of visual models of processes (i.e., process map or flowchart). These diagrams depict the sequence of activities and various crossroads (gateways), which lead to different routes depending on choices made.
A business process model is a self-contained, temporal and logical order in which a set of activities are expected to be executed to achieve a business goal. Typically, a process model describes what needs to be done and when (control flow), who is going to do what (resources), and on what it is working on (data). In this context, a possible execution, called process trace or simply trace, is a sequence of tasks and events respecting the order given by the connectors.
\subsubsection{The automation of compliance}
Business process compliance is a relationship between the formal representation of a process model and the formal representation of the relevant regulations [Governatori et al., 2006a]. Any approach to automate the checking whether a business process complies with the regulation governing has to ensure that it is able to properly model business processes as well as norms. In the past decades many approaches to automatize business process compliance have been proposed [Ly et al., 2015; Becker et al., 2012] and legal informatics is experiencing growth in activity, also at the industrial level [Boella et al., 2016; Ajani et al., 2016].
However, a challenging research topic is the possibility of modeling standards in a conceptually valid, detailed and exhaustive way that can be used in practice for companies and, at the same time, have the ability to be used generically for any type of standard also taking into account the regulatory environment as a whole [Governatori, 2015b].
This shifts the focus to the adoption formalisms. Temporal logic and Event Calculus have been used in several frameworks. However, it has been shown that when norms are formalized in Linear Temporal Logic the evaluation whether a process is compliant produces results that are not compatible with the intuitive and most natural legal interpretation [Hashmi et al., 2014; Governatori, 2015a]. Furthermore, it was argued that, while such logics can properly model norms such formalizations would be completely useless from a process compliance point of view insofar they would require an external oracle to identify the compliant executions of the process, and build the formalization from the traces corresponding to the traces deemed legal by the oracle. This means that, there is no need of the formalization to determine if the process is compliant or not, since this is done by the oracle [Governatori, 2015a; Governatori and Hashmi, 2015].
Some studies had focused on the application of Natural Language Processing (NLP) methods to design legal document management system to assist legal professionals in navigate legislation and retrieving the information they are interested in [Boella et al., 2013a; Boella et al., 2013b]. An example is Eunomos [Boella et al., 2012; Boella et al., 2016]. These type of systems classify, index, and discover inter-links between legal documents, retrieved through Web-crawling tools, by exploiting NLP tools, such as parsers and statistical algorithms, and semantic knowledge bases, such as legal ontologies in Web Ontology Language (OWL) . This is often done by transforming the source legal documents into XML standards and tagging the relevant information to then allow later archiving and querying the XML files.
However, the overall usefulness of these systems are limited due to their focus on terminological issues and information retrieval while disregarding the specific semantic aspects, which allow for legal reasoning. Just as standard deontic logic focused mostly on the notion of obligation, subsequent developments in deontic logic also adopt an abstract view of law, with a very loose connection with the texts of regulations. For lawyers, the meaning of laws can only be really understood in the rich expressiveness of natural language since “like language generally, legal discourse can never escape its own textuality” [Peller, 1985].
There is thus a gap between a powerful reasoning mechanism on the formalization of law and the textuality of law, which can be addressed with solutions coming from the literature on Natural Language Semantics.
A new standardization initiative called LegalRuleML [Athan et al.,2013; Athan et al., 2015] tries to address these issues. LegalRuleML is an XML format that extends the RuleML standard to define a rule interchange language for the legal domain. While legal XML standards are used to tag the original textual content of the legal documents, LegalRuleML separately represents and stores the logical content of the provisions. Specifically, LegalRuleML allows to specify semantic/logical representations and associate them with both the structural elements of the documents or with tasks in a business process.
LegalRuleML allows to encode RuleML representations of formulas in Defeasible Deontic Logic (DDL) [Governatori, 2018]. This is an extension of standard Defeasible Logic with deontic operators, and the operators for compensatory obligation [Governatori and Rotolo, 2006]. Defeasible Logic is an efficient and simple rule based computationally oriented non-monotonic formalism, designed for handling exception in a natural way. According to the formalization proposed in [Antoniou et al., 2001], Defeasible Logic is a constructive logic with its proof theory and inference condition as its core. The logic exploits both positive proofs, a conclusion has been constructively prove using the given rules and inference conditions (also called proof conditions), and negative proofs: showing a constructive and systematic failure of reaching particular conclusions, or in other terms, constructive refutations. The logic uses a simple language, that proved successful in many application area, due to the scalability and constructiveness of the logic. These elements are extremely important for normative reasoning, where an answer to a verdict is often nor enough, and full traceability is needed.
\subsection{Legal reasoning and Defeasible Deontic Logic}
Norms describe general cases and what behavior should be taken, or the consequences, if the real facts are similar to the general case described in the norm. Therefore, norms describe the conditions under which they are applicable and the normative effects they produce when applied. Simply put, the scope of norms is to regulate the behavior of their subjects and to define what is legal and what is illegal.
In a compliance perspective, the normative effects of importance are the deontic effects (also called normative positions). The basic and more important deontic effects are: obligation, prohibition and permission.
\begin{itemize}
\item Obligation: when there is a situation, an act, or a course of action to which a bearer is legally bound, and if it is not achieved or performed results in a violation.
\item Prohibition: when there is a situation, an act, or a course of action which a bearer should avoid, and if it is achieved results in a violation.
\item Permission: when something is permitted if the prohibition of it or the obligation to the contrary do not hold.
\end{itemize}
This gives rise to some considerations:
\begin{itemize}
\item Obligations and prohibitions are constraints that limit the space of action of processes.
\item They can be violated, and a violation does not imply an inconsistency within a process with the consequent termination of or impossibility to continue the business process.
\item Violations can be generally compensated for, and processes with compensated violations are still compliant [Governatori and Sadiq, 2009; Governatori and Milosevic, 2005] (e.g. contracts typically contain compensatory clauses specifying penalties and other sanctions triggered by breaches of other contract clauses [Governatori, 2005]).
\item Not all violations are compensable, and uncompensated violations means that a process is not compliant.
\item Permissions cannot be violated. They can be used (indirectly) to determine that there are no obligations or prohibitions to the contrary, or to derive other deontic effects.
\item Legal reasoning and legal theory typically assume a strong relationship between obligations and prohibitions: the prohibition of A is the obligation of ¬A (the opposite of A), and then if A is obligatory, then ¬A is forbidden [Pattaro, 2005].
\end{itemize}
Taking in consideration the notion of obligation, compliance means to identify whether a process violates or not a set of obligations. Thus, the first step is to determine whether and when an obligation is in force. Hence, an important aspect of the study of obligations is to understand the lifespan of an obligation and its implications on the activities carried out in a process. A norm can specify if there is:
\begin{itemize}
\item Punctual obligations: an obligation is in force for a particular time point.
\item Persistent obligations: a norm indicates when an obligation enters in force. An obligation remains in force until terminated or removed.
\begin{itemize}
\item For persistent obligations we can ask if to fulfil an obligation we have to obey to it for all instants in the interval in which it is in force, maintenance obligations, or
\item Whether doing or achieving the content of the obligation at least once is enough to fulfil it, achievement obligations.
\end{itemize}
\begin{itemize}
\item For achievement obligations another aspect to consider is whether the obligation could be fulfilled even before the obligation is actually in force. If this is admitted, then there is a preemptive obligation, otherwise the obligation is non-preemptive.
\end{itemize}
\item Termination of obligations: norms can specify the interval in which an obligation is in force.
\end{itemize}
As said, what differentiates obligations and other constraints is that obligations can be violated.
\begin{itemize}
\item If we still have to comply with a violated obligation (the obligation persists after being violated) we speak of a perdurant obligation.
\item Otherwise, we speak of a non-perdurant obligation [Governatori, 2015b].
\end{itemize}
\section{THE CASE STUDY}
Before you begin to format your paper, first write and save the content as a
separate text file. Complete all content and organizational editing before
formatting. Please note sections \ref{AA}--\ref{SCM} below for more information on
proofreading, spelling and grammar.
Keep your text and graphic files separate until after the text has been
formatted and styled. Do not number text heads---{\LaTeX} will do that
for you.
\subsection{Abbreviations and Acronyms}\label{AA}
Define abbreviations and acronyms the first time they are used in the text,
even after they have been defined in the abstract. Abbreviations such as
IEEE, SI, MKS, CGS, ac, dc, and rms do not have to be defined. Do not use
abbreviations in the title or heads unless they are unavoidable.
\subsection{Units}
\begin{itemize}
\item Use either SI (MKS) or CGS as primary units. (SI units are encouraged.) English units may be used as secondary units (in parentheses). An exception would be the use of English units as identifiers in trade, such as ``3.5-inch disk drive''.
\item Avoid combining SI and CGS units, such as current in amperes and magnetic field in oersteds. This often leads to confusion because equations do not balance dimensionally. If you must use mixed units, clearly state the units for each quantity that you use in an equation.
\item Do not mix complete spellings and abbreviations of units: ``Wb/m\textsuperscript{2}'' or ``webers per square meter'', not ``webers/m\textsuperscript{2}''. Spell out units when they appear in text: ``. . . a few henries'', not ``. . . a few H''.
\item Use a zero before decimal points: ``0.25'', not ``.25''. Use ``cm\textsuperscript{3}'', not ``cc''.)
\end{itemize}
\subsection{Equations}
Number equations consecutively. To make your
equations more compact, you may use the solidus (~/~), the exp function, or
appropriate exponents. Italicize Roman symbols for quantities and variables,
but not Greek symbols. Use a long dash rather than a hyphen for a minus
sign. Punctuate equations with commas or periods when they are part of a
sentence, as in:
\begin{equation}
a+b=\gamma\label{eq}
\end{equation}
Be sure that the
symbols in your equation have been defined before or immediately following
the equation. Use ``\eqref{eq}'', not ``Eq.~\eqref{eq}'' or ``equation \eqref{eq}'', except at
the beginning of a sentence: ``Equation \eqref{eq} is . . .''
\subsection{\LaTeX-Specific Advice}
Please use ``soft'' (e.g., \verb|\eqref{Eq}|) cross references instead
of ``hard'' references (e.g., \verb|(1)|). That will make it possible
to combine sections, add equations, or change the order of figures or
citations without having to go through the file line by line.
Please don't use the \verb|{eqnarray}| equation environment. Use
\verb|{align}| or \verb|{IEEEeqnarray}| instead. The \verb|{eqnarray}|
environment leaves unsightly spaces around relation symbols.
Please note that the \verb|{subequations}| environment in {\LaTeX}
will increment the main equation counter even when there are no
equation numbers displayed. If you forget that, you might write an
article in which the equation numbers skip from (17) to (20), causing
the copy editors to wonder if you've discovered a new method of
counting.
{\BibTeX} does not work by magic. It doesn't get the bibliographic
data from thin air but from .bib files. If you use {\BibTeX} to produce a
bibliography you must send the .bib files.
{\LaTeX} can't read your mind. If you assign the same label to a
subsubsection and a table, you might find that Table I has been cross
referenced as Table IV-B3.
{\LaTeX} does not have precognitive abilities. If you put a
\verb|\label| command before the command that updates the counter it's
supposed to be using, the label will pick up the last counter to be
cross referenced instead. In particular, a \verb|\label| command
should not go before the caption of a figure or a table.
Do not use \verb|\nonumber| inside the \verb|{array}| environment. It
will not stop equation numbers inside \verb|{array}| (there won't be
any anyway) and it might stop a wanted equation number in the
surrounding equation.
\subsection{Some Common Mistakes}\label{SCM}
\begin{itemize}
\item The word ``data'' is plural, not singular.
\item The subscript for the permeability of vacuum $\mu_{0}$, and other common scientific constants, is zero with subscript formatting, not a lowercase letter ``o''.
\item In American English, commas, semicolons, periods, question and exclamation marks are located within quotation marks only when a complete thought or name is cited, such as a title or full quotation. When quotation marks are used, instead of a bold or italic typeface, to highlight a word or phrase, punctuation should appear outside of the quotation marks. A parenthetical phrase or statement at the end of a sentence is punctuated outside of the closing parenthesis (like this). (A parenthetical sentence is punctuated within the parentheses.)
\item A graph within a graph is an ``inset'', not an ``insert''. The word alternatively is preferred to the word ``alternately'' (unless you really mean something that alternates).
\item Do not use the word ``essentially'' to mean ``approximately'' or ``effectively''.
\item In your paper title, if the words ``that uses'' can accurately replace the word ``using'', capitalize the ``u''; if not, keep using lower-cased.
\item Be aware of the different meanings of the homophones ``affect'' and ``effect'', ``complement'' and ``compliment'', ``discreet'' and ``discrete'', ``principal'' and ``principle''.
\item Do not confuse ``imply'' and ``infer''.
\item The prefix ``non'' is not a word; it should be joined to the word it modifies, usually without a hyphen.
\item There is no period after the ``et'' in the Latin abbreviation ``et al.''.
\item The abbreviation ``i.e.'' means ``that is'', and the abbreviation ``e.g.'' means ``for example''.
\end{itemize}
An excellent style manual for science writers is \cite{b7}.
\subsection{Authors and Affiliations}
\textbf{The class file is designed for, but not limited to, six authors.} A
minimum of one author is required for all conference articles. Author names
should be listed starting from left to right and then moving down to the
next line. This is the author sequence that will be used in future citations
and by indexing services. Names should not be listed in columns nor group by
affiliation. Please keep your affiliations as succinct as possible (for
example, do not differentiate among departments of the same organization).
\subsection{Identify the Headings}
Headings, or heads, are organizational devices that guide the reader through
your paper. There are two types: component heads and text heads.
Component heads identify the different components of your paper and are not
topically subordinate to each other. Examples include Acknowledgments and
References and, for these, the correct style to use is ``Heading 5''. Use
``figure caption'' for your Figure captions, and ``table head'' for your
table title. Run-in heads, such as ``Abstract'', will require you to apply a
style (in this case, italic) in addition to the style provided by the drop
down menu to differentiate the head from the text.
Text heads organize the topics on a relational, hierarchical basis. For
example, the paper title is the primary text head because all subsequent
material relates and elaborates on this one topic. If there are two or more
sub-topics, the next level head (uppercase Roman numerals) should be used
and, conversely, if there are not at least two sub-topics, then no subheads
should be introduced.
\subsection{Figures and Tables}
\paragraph{Positioning Figures and Tables} Place figures and tables at the top and
bottom of columns. Avoid placing them in the middle of columns. Large
figures and tables may span across both columns. Figure captions should be
below the figures; table heads should appear above the tables. Insert
figures and tables after they are cited in the text. Use the abbreviation
``Fig.~\ref{fig}'', even at the beginning of a sentence.
\begin{table}[htbp]
\caption{Table Type Styles}
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline
\textbf{Table}&\multicolumn{3}{|c|}{\textbf{Table Column Head}} \\
\cline{2-4}
\textbf{Head} & \textbf{\textit{Table column subhead}}& \textbf{\textit{Subhead}}& \textbf{\textit{Subhead}} \\
\hline
copy& More table copy$^{\mathrm{a}}$& & \\
\hline
\multicolumn{4}{l}{$^{\mathrm{a}}$Sample of a Table footnote.}
\end{tabular}
\label{tab1}
\end{center}
\end{table}
\begin{figure}[htbp]
\centerline{\includegraphics{fig1.png}}
\caption{Example of a figure caption.}
\label{fig}
\end{figure}
Figure Labels: Use 8 point Times New Roman for Figure labels. Use words
rather than symbols or abbreviations when writing Figure axis labels to
avoid confusing the reader. As an example, write the quantity
``Magnetization'', or ``Magnetization, M'', not just ``M''. If including
units in the label, present them within parentheses. Do not label axes only
with units. In the example, write ``Magnetization (A/m)'' or ``Magnetization
\{A[m(1)]\}'', not just ``A/m''. Do not label axes with a ratio of
quantities and units. For example, write ``Temperature (K)'', not
``Temperature/K''.
\section*{Acknowledgment}
The current research was conducted in the project ``CANP - CAsa Nel Parco'' of Regione Piemonte funded by POR FESR PIEMONTE 2014-2020. We are grateful for the collaboration of the ``City of Health and Science'' of Torino (Italy). We are also grateful to the partners involved in the project
\section*{References}
Please number citations consecutively within brackets \cite{Dumas-etal:2018}. The
sentence punctuation follows the bracket \cite{b2}. Refer simply to the reference
number, as in \cite{b3}---do not use ``Ref. \cite{b3}'' or ``reference \cite{b3}'' except at
the beginning of a sentence: ``Reference \cite{b3} was the first $\ldots$''
Number footnotes separately in superscripts. Place the actual footnote at
the bottom of the column in which it was cited. Do not put footnotes in the
abstract or reference list. Use letters for table footnotes.
Unless there are six authors or more give all authors' names; do not use
``et al.''. Papers that have not been published, even if they have been
submitted for publication, should be cited as ``unpublished'' \cite{b4}. Papers
that have been accepted for publication should be cited as ``in press'' \cite{b5}.
Capitalize only the first word in a paper title, except for proper nouns and
element symbols.
For papers published in translation journals, please give the English
citation first, followed by the original foreign-language citation \cite{b6}.
\section{Introduction}
\noindent The necessity of satisfying regulations or laws forces business organizations to redesign their internal processes, in the context of change management \cite{Hayes:2018}. In the field of Business Process Management (BPM) research, one of the main topic concerns regulatory or Compliance Management (CM), i.e. the analysis of compliance to norms \cite{Dumas-etal:2018}\cite{vanderAalst:2013}.
The increasing pressure from regulatory authorities to organizations led to the development and application of Compliance Management Frameworks (CMFs). In this context, CM can be addressed at the operational level by focusing on business processes, intended as the set of activities accomplishing a specific organizational goal.
Business process analysis usually introduces performance objectives to be considered in addition to constraints imposed by external pressures (e.g., regulatory issues). The investigation of undesirable events and norm violations adopted traditional techniques, e.g. root cause analysis (commonly used in manufacturing processes to improve performance). More recently, CMFs explore the relationship between the formal representation of a process model and the relevant regulations. There are many different adoptable CM strategies consisting in approaches to automate the checking whether a business process complies with the actual regulation \cite{Governatori-Sadiq:2009}. The goal is to ensure that such approaches properly model business processes as well as norms. Moreover, in the past decades many CM approaches in the context of digitization to automatize business processes have been proposed \cite{Sackmann-etal:2018}. We describe here a CM approach to support regulatory compliance for healthcare business processes based on a compliance-by-design methodology \cite{Governatori-Sadiq:2009} and using a business process compliance checker called Regorous \cite{Governatori:2015}. Regorous implements a propositional logic-based approach of Defeasible Deontic Logic (DDL) \cite{Nute:1997} to automate CM.
This work is based on a Regional project with European funds called ``CANP"\footnote{CANP ``La CAsa Nel Parco": http://casanelparco-project.it/}. The goal of the project is using Artificial Intelligence and telemedicine solutions to enhance e-Health procedures within the ``Città della Salute e della Scienza", one of the biggest hospitals in Europe \cite{Sulis-etal:2019}. The application of innovative telemedicine technologies supports the care of elderly patients in the context of a Hospital at Home (HaH). The use of communication systems in the remote management of the patient could improve treatment outcomes, increase access to care, and reduce health costs \cite{Caplan-etal:2012}.
In the paper, we will exemplify our approach by showing how some General Data Protection Regulation (GDPR) provisions to safeguard the personal data of the patients are enforced on the BPM representing the HaH process. However, it will be evident that the approach is general enough to handle any kind of legal constraint involved in BPM procedures.
\section{Background and Related work}
\noindent Legal risks play a crucial roles in business activities, as non-compliant processes may incur heavy sanctions, in particular in the healthcare domain. On the other hand, compliance checking is known to be a difficult task due to the fact that norms are constantly evolving and new reorganizations must be often implemented with the introduction of new procedures
\cite{Hayes:2018}. For this reason, legal technology
is experiencing growth in activity, also at the industrial level \cite{Ajani-etal:2016}, \cite{Boella-etal:2016}.
Researching the interplay between BPM and CM approaches can lead to the quick identification of processes that need restructuring because they are no longer compliant when regulations are introduced or updated \cite{Racz-etal:2010}.
It is possible to find several studies on compliance with laws, rules or regulations in the case of processes related to patient health \cite{Buddle-etal:2005}, \cite{Amantea-etal:2018}. Compliance in healthcare considers the conformity of care processes with laws, regulations and standards related to patient safety, privacy of patient information and administrative practices.
Ultimately, health compliance is about providing safe and high quality patient care. Healthcare organizations are also required to comply with strict standards, regulations and laws at regional and state level.
These innovations may require the application of new regulations, such as the GDPR, without forgetting that the health sector is full of strictly health regulations in constant evolution.
\subsection{Business process compliance and logic}
\noindent A business process model is a self-contained, temporal and logical order in which a set of activities are expected to be executed. Typically, a process model describes what needs to be done and when (control flow), who is going to do what (resources), and on what it is working on (data).
Business Process Management (BPM) is a discipline standardly used to model and analyse business processes, being Business Process Model and Notations (BPMN) the formal language used therein \cite{Allweyer:2016}.
Change management is one of the central issues in BPM research \cite{Abo-Hamad-Arisha:2013}, \cite{Sulis-DiLeva:2018}, \cite{DiLeva-etal:2020}. In the context of healthcare studies, change management is usually handled by creating visual models of processes (i.e., process map or flowchart) and implement them in BPMN \cite{Amantea-etal:2020}, \cite{Sulis2019risk},\cite{DiLeva:2017}, \cite{Muller-Rogge-Solti:2011}. These diagrams depict the sequence of healthcare activities and various crossroads (gateways), which lead to different routes depending on choices made \cite{DiLeva:2017}.
In order to check compliance on business processes, the formal representation of a process and the formal representation of the relevant regulations need to be related \cite{Governatori-etal:2006}.
In the past decades many approaches to automatize business process compliance have been proposed \cite{Becker-etal:2012}, \cite{Ly-etal:2015}.
However, a challenging research topic is the possibility of modeling standards in a conceptually valid, detailed and exhaustive way that can be used in practice for companies and, at the same time, have the ability to be used generically for any type of standard also taking into account the regulatory environment as a whole \cite{Governatori:2015}.
This shifts the focus to the formalisms used to model regulations.
Linear Temporal Logic and Event Calculus have been used in several frameworks \cite{Awad-etal:2008}. However, it has been shown that when norms are formalized in Linear Temporal Logic the evaluation whether a process is compliant produces results that are not compatible with the intuitive and most natural legal interpretation \cite{Hashmi-etal:2014}, \cite{Governatori:2015a}. Furthermore, it was argued that, while such logics can properly model norms such formalizations would be completely useless from a process compliance point of view insofar they would require an external oracle to identify the compliant executions of the process.
Defeasible Deontic Logic (DDL) \cite{Governatori:2018} has been proposed to overcome these limitations. This is an extension of standard Defeasible Logic with deontic operators, and the operators for compensatory obligation \cite{Governatori-Rotolo:2006}. According to the formalization proposed in \cite{Antoniou-etal:2001}, Defeasible Logic is a constructive logic with its proof theory and inference condition as its core. The logic exploits both positive proofs, a conclusion has been constructively prove using the given rules and inference conditions (also called proof conditions), and negative proofs: showing a constructive and systematic failure of reaching particular conclusions, or in other terms, constructive refutations. The logic uses a simple language, that proved successful in many application area, due to the scalability and constructiveness of the logic. These elements are extremely important for normative reasoning, where an answer to a verdict is often nor enough, and full traceability is needed.
\subsection{Modelling norms in DDL}
\noindent Norms are usually modelled as if-then rules representing the conditions under which norms are applicable and the normative effects these produce when applied. In a compliance perspective, the normative effects of importance are the deontic effects (also called normative positions). The basic deontic effects are: permission, obligation, and
prohibitions. Thus, an if-then rule in the form ``$a$$\rightarrow$$b$'' that represents a norm reads
``if $a$ holds, then $b$ is permitted/obligatory/prohibited''.
However, it is standardly assumed that prohibitions are indeed obligations; something is prohibited if and only if there is obligation to its contrary \cite{Pattaro-etal:2005}. In this paper, for the sake of simplicity, we assume the term ``obligation'' also encompasses prohibitions.
Obligations are constraints that limit the space of action of processes. If the antecedent of an if-then rule representing a obligation holds in the context and the bearer of an obligation does not comply with the consequent, then s/he violates the
obligation. However, a violation does not imply an inconsistency within a process with the consequent termination of or impossibility to continue the business process.
Violations can be compensated, and processes with compensated violations are still compliant \cite{Governatori-Sadiq:2009} \cite{Governatori-Milosevic:2005}. E.g., contracts typically contain compensatory clauses specifying penalties and other sanctions triggered by breaches of other contract clauses \cite{Governatori:2005}. Not all violations are compensable; if a process contain uncompensated violations, then it is not compliant.
On the other hand, permissions cannot be violated. They can be used (indirectly) to determine that there are no obligations or prohibitions to the contrary, or to derive other deontic effects. Permissions can be
explicitly encoded in the legislation (positive permissions) or they can be derived, e.g., by observing that there is nothing in the code that forbids them (negative permissions) (cf. \cite{Robaldo-Sun:2017}).
Instead, compliance usually means to identify whether a process violates or not a set of obligations. Thus, the first step is to determine whether and when an obligation is in force. In other words, an important aspect of the study of obligations is to understand the lifespan of an obligation and its implications on the activities carried out in a process.
A first broad classification is between punctual obligations and persistent obligations, the former holding on a particular time point, the latter holding over a time interval. In our research, we were only concerned with persistent obligations.
The latter are further sub-classified into maintenance obligations and achievement obligations, the former being in force for all instants in the interval, the latter only until they are achieved/satisfied. Finally, achievement obligations are further sub-classified into preemptive and non-preemptive obligations and between perdurant and non-perdurant obligations. Preemptive obligations are those that could be fulfilled even before the obligation is actually in force. On the other hand, perdurant obligations are those that are still in force even if/after they have been violated.
Regorous, the propositional DDL system we used in our research, formalize the different subclasses of persistent obligations as follows, where ``p'' is a propositional symbol:
\vspace{3pt}
\begin{itemize}
\item ``[OM]p'': there is a maintenance obligation for p.
\item ``[OAPP]p'': there is an achievement preemptive and perdurant obligation for p.
\item ``[OAPNP]p'': there is an achievement preemptive and non-perdurant obligation for p.
\item ``[OANPP]p'': there is an achievement non preemptive and perdurant obligation for p.
\item ``[OANPNP]p'': there is an achievement non preemptive and non-perdurant obligation for p.
\end{itemize}
\vspace{3pt}
On the other hand, positive permissions are represented as ``[P]p'', meaning that ``p'' is permitted. As mentioned above, also positive permissions are involved in the compliance checking process because positive permissions entail that there are no obligations or prohibitions to the contrary or other deontic effects. Furthermore, we may further restrict the process by imposing that it is permitted only what it is explicitly asserted by positive permissions. However, these
entailments are not relevant here as the case study used in this paper do not encompass deontic reasoning on permissions.
Regorous is implemented as a plug-in of Eclipse. The BPMN is externally built and then uploaded in the platform together with a set of rules in Regorous XML format. Subsequently, in each task of the process it is possible to specify which terms of the vocabulary are true or false via special Eclipse windows provided by the plug-in. Both the BPMN file and the XML file for Regorous are available on GitHub\footnote{\url{https://github.com/liviorobaldo/TrECECpaperAttachments}}.
\vspace{3pt}
\section{Case study: Hospital at Home (HaH) process}
\noindent For more than 30 years, the Hospital ``Città della Salute e della Scienza'' of Turin has operated the Hospital at Home (HaH) service. In this healthcase service, the patient in acute disease (but who do not require equipment with high technological complexity and intensive or invasive monitoring), is taking in charge from the hospital. Specifically, the hospital provides for the organization of care but in the patient's house.
This service is made by two main processes, i.e. the acceptance and the real process of care. We modelled, implemented, and tested both processes, but, for space constraints, this paper will only focus on the acceptance process \cite{amantea2020pm}. The BPMN describing the acceptance process is shown in Figure \ref{RegorousForComplianceChecking} and it is available on GitHub.
The requests are made by the general medical doctor of emergency or regular departments (generator ``Request HaH''). Each case refers to some guideline to understand if the patient has some characteristics to be taken in charge to this type of hospitalization (tasks ``Make preliminary analysis'', ``Talk to doctor'', ``Talk to patient'', ``Talk to caregiver''). During this meetings, the case manager explains to the patient and to the family the characteristics, organization, and requirements of the service. On the other hand, she collects information out of clinical and functional evaluations as well as cognitive aspects.
At the end of this evaluation process, the CM decides if either the patient is suitable for an HaH service or if another healthcare process would be more appropriate for his case (``Evaluate alternative route''). In the latter case, the patient will start a new separate healthcare process.
On the other hand, in case the HaH service is selected, the patient or his caregiver need to express his will to access this service or not. From now onward, the process is specifically dealing with the administrative/legal constraints of the HaH service, including those connected with the processing of the patient's personal data. The patient or his caregiver will sign some documents (tasks ``Sign policy of admission'' and ``Compile emergency report'') including, the most important in our case, the informed consent (task ``Fill out the nurse form + Pick up informed consent''), in which the patient or his caregiver will have to provide consent for the processing of the patient's personal data, as required by the GDPR.
After the informed consent is collected, the patient will be registered in the system and taken in charge from the HaH service (tasks ``Make HaH prescription'', ``Make taking in load'', ``Transfer of power''). Other separate healthcare processes will be subsequently carried out in the future in order to implement the HaH service at the patient's home.
\subsection{Encoding GDPR norms in Regorous}
As point out above, Regorous allow to specify permissions and (different) types of obligations holding on propositional symbols ``p''. ``p'' is called a “term” in Regorous terminology.
Regorous lists all terms used in a set of formalization, together with their description, in a special XML tag
{\tt <Vocabulary>}. Below, we will show how we modelled Article 6, paragraph 1, point 1 of the GDPR\footnote{\url{https://eur-lex.europa.eu/eli/reg/2016/679/oj}}, which specifies one condition for lawfulness of processing (and indeed the main one): processing of personal data is lawful if ``the data subject has given consent to the processing of his or her personal data for one or more specific purposes;''.
In order to model this norm, we need the concept of personal data processing and the one of consent. These corresponds to two propositional symbols, i.e., in Regorous, to the following two terms that we add to the vocabulary:
\vspace{3pt}
\begin{alltt}
<Vocabulary>
<Term atom="Proc"
description="Processing: means
any operation or set of operations
which is performed on personal data
..."/>
<Term atom="GiveConsent"
description="Consent given by
the data subject means any freely
given, specific, informed and
unambiguous indication ..."/>
</Vocabulary>
\end{alltt}
\vspace{3pt}
Given the two terms in the vocabulary, we decided to implement the GDPR norms connected with the lawfulness of processing by forbidding processing of personal via a general maintenance obligation, unless a more specific (and stronger) if-then rule, based on given consent, permits it.
The maintenance obligation is formalized as follows:
\vspace{3pt}
\begin{alltt}
<Rule xmlns:xsi="..." ruleLabel="Art.6.0"
xsi:type="DflRuleType">
<ControlObjective>Personal data
processing is prohibited.
</ControlObjective>
<FormalRepresentation>
=>[OM]-Proc
</FormalRepresentation>
</Rule>
\end{alltt}
\vspace{3pt}
In Regorous, negation and logical entailment are represented via the symbols ``{\tt -}'' and ``{\tt =>}''. Therefore, the antecedent of the if-then rule ``{\tt =>[OM]-Proc}'' is empty, meaning that the consequent always holds, i.e., that the negation of the term ``{\tt Proc}'' is obligatory for all over the process.
On the other hand, the permission based on given consent is formalized as follows:
\vspace{3pt}
\begin{alltt}
<Rule xmlns:xsi="..." ruleLabel="Art.6.1a"
xsi:type="DflRuleType">
<ControlObjective>Processing shall be
lawful if the data subject has given
consent to the processing of his or
her personal data for one or more
specific purposes.</ControlObjective>
<FormalRepresentation>
GiveConsent=>[P]Proc
</FormalRepresentation>
</Rule>
\end{alltt}
\vspace{3pt}
In this case, the antecedent of the if-then rule is not empty: the rule will triggers only when the term
{\tt GiveConsent} is asserted. In our BPMN, this will be done in the task ``Fill out the nurse form + Pick up informed consent''.
Of course, the two formulas cannot hold together as the first entails that the processing is prohibited while the latter entails that it is permitted. In order to solve these conflicts, DDL implements overriding relations between norms. In our example, the permission {\tt [P]Proc} must override the obligation {\tt [OM]-Proc} for the business process to be compliant with the GDPR.
In Regorous, DDL overriding relations are implemented as ``superiority relations'', encoded via the homonym tag, in which the ``superiorRuleLabel'' overrides the ``inferiorRuleLabel''. Therefore, in the example under consideration we have:
\begin{alltt}
<SuperiorityRelation
superiorRuleLabel="Art.6.1a"
inferiorRuleLabel="Art.6.0"/>
\end{alltt}
By asserting these superiority relations, Regorous is able to infer that if both the terms {\tt Proc} and {\tt GiveConsent} holds, the process is compliant.
The assertion of {\tt GiveConsent} does not only entail that processing of personal data is permitted. According to Article 7(1) of the GDPR, ``Where processing is based on consent, the controller shall be able to demonstrate that the data subject has consented to processing of his or her personal data.''. Therefore, the assertion of
{\tt GiveConsent} also entails another maintenance obligation for the hospital, formalized in Regorous as follows:
\vspace{3pt}
\begin{alltt}
<Rule xmlns:xsi="..." ruleLabel="Art.7.1"
xsi:type="DflRuleType">
<ControlObjective>If processing is
based on consent, the controller shall
be able to demonstrate the consent.
</ControlObjective>
<FormalRepresentation>
GiveConsent=>[OM]DemonstrateConsent
</FormalRepresentation>
</Rule>
\end{alltt}
\vspace{3pt}
The obligation is again a maintenance obligation, rather than an achievement one, because the GDPR requires the controller (the hospital, in our case) to store and provide evidence the given consent {\it at any time}, upon request of the data protection authority or any other appointed auditor. In the HaH process, the personal data record for the patient, which includes the evidence for his given consent, is prepared and stored in the hospital digital archive in the task ``Make taking in load'', together with all other administrative details referring to the HaH service that will be put in place. At the same time, the term {\tt DemonstrateConsent} will be asserted in Regorous. The term will remain true until the record is kept in the hospital archive, so that Regorous will be always able to assess compliance with {\tt [OM]DemonstrateConsent} that, being a maintainance obligation, must be satisfied in every instant of the execution.
Note that in our use case, we are focusing only on Article 6.1(a) of the GDPR. However, given consent is only one of the legal bases that allows for processing of personal data, the one that is used by the HaH acceptance process. On the other hand, our formalization includes other legal bases, specified in other norms of the GDPR, and used in other healthcare processes within the hospital.
For instance, Article 6.1(b) of the GDPR permits processing if it is ``necessary for the performance of a contract to which the data subject is party or in order to take steps at the request of the data subject prior to entering into a contract'' while Article 6.1(d) of the GDPR permits processing if it is ``necessary in order to protect the vital interests of the data subject or of another natural person.''.
Article 6.1(b) and Article 6.1(d) provides then other two sufficient conditions to permit processing of personal data, so that they are respectively formalized in the if-then rules ``{\tt Contract=>[P]Proc}'' and ``{\tt VI=>[P]Proc}''. Two corresponding superiority relations are also added to the Regorous knowledge base in order to allow the overriding of {\tt [OM]-Proc} also in these two cases.
The fact that these two sufficient conditions are not used in the BPMN under consideration highlights the fact that the Regorous file is not 1:1 associated with the HaH acceptance process. In other words, the formalization of the legislation is kept independent from the modeling of the healthcare processes and shared among multiple ones.
In case new legislation will enter into force, or, alternatively, the existing one is updated, only the {\tt <Rule>}(s) will need to be modified at first. Regorous will be then able to identify which business processes are affected by the changes: those that are no longer compliant with the new legal setting.
\section{Executing Regorous on the business process}
Given a set of well-formed rules and superiority relations encoded in the XML format briefly seen above, Regorous allows to check whether a Business Process in the BPMN standard is compliant with them.
Regorous is implemented as a plug-in of Eclipse.
After the creation or the import of BPMN process it is possible to upload of the file or the files with the set of rules in Regorous XML format. Afterwards, in each activity of the process it is possible to specify the resources used and for each of them the detailed tasks that are included in the activity. Both the resourses and the actions have to be express using the name of the vocabulary terms.
In this way, via a special Eclipse windows provided by the plug-in, it is possible to see if the process is compliant, not compliant, or also maybe compliant (with some warning).
Just to show a practical example, Figure \ref{RegorousForComplianceChecking} shows how Regorous performs compliance checking on the HaH process. The BPMN file is uploaded in Eclipse together with the ruleset formalizing the GDPR norms in Regorous XML format.
Precisely, as shown on the bottom of Figure \ref{RegorousForComplianceChecking}, the plug-in includes special tabs that allow to specify, for each task, the values of the terms. For instance, by specifying “GiveConsent” and “Proc” in the task “Fill out the nurse form + Pick up informed consent”, Regorous infers that the process is compliant with the ruleset, as the superiority relation seen above will make the processing of personal data permitted. Conversely, by specifying the single action “Proc”, Regorous infers that the process is not compliant with the ruleset because the rule with ruleLabel=“Art.6.0” asserts the processing of personal data as prohibited and, contrary to the previous case, that prohibition is not overridden by a stronger permission.
After setting the activities, the compilation is launched. At this point it is important to underline some things. The more important is that the check follows the process flow. Therefore, it starts from the generator (the rod on the left) and runs the various flows for each branch (gateway) of the process and till all the various possible ends of the process. In this way, the first import thing is that all possible individual paths are checked.
Moreover, each path is not run once, but several times. Here the rules of superiority come into play and therefore it is possible to verify even if some activities are generically controlled sooner or later within the process, if they are for all possible branches and if they are (for each branch) in the sequence established by law.
In fact, a rule can, for example:
\begin{itemize}
\item generically establish that an action must be carried out. In this case, by running the run more than once it is possible to check whether the action is carried out sooner or later in an activity of the process, but also if this activity is not only present, but is present for all possible branches of the process. For example, it is possible that one requirement may be overridden if another occurs. For example, in our case the rule is that personal data cannot be processed, this is the primary rule. But then the rule allows exceptions, for example if consent is given. Of course, we don't know from the beginning, by law, if and when (in what activity) consent will be given.Therefore, the primary rule will apply until the requirement contained in the superior rule of consent arrives. If there were no rules of superiority, the primary requirement that personal data cannot be processed would continue to apply. Furthermore, if for example the consent were revoked, this would be a rule superior to the "give consent" rule which would therefore annul the consent previously given. In this case, however, it should be noted that it is as if the consent had not been obtained from the beginning, therefore, if, for example, other branches of the process are in progress that are in the meantime being carried out in parallel, they may no longer be compliant. If the Regorous control stops at the first run, without possibly having some subsequent steps in memory, it could provide incorrect output. It would be enough for the consent to be given before by a gateway which provides for two parallel activities / paths and then to be revoked in only one of two branches (for example, the patient would probably communicate the revocation to the doctor or nurse, but not to any laboratory).
\item provide that one activity must be done following or before another. For example, in our case the consent must be signed after having received the information and before having the treatment. Therefore, it is not only important that the activities of speaking, signing, being taken in charge exist, but it is important that, in any branch of the process, if there is the signing activity, it is preceded by having had the information and if a patient is registered is preceded by having signed consent.
\end{itemize}
In this way, at the end of the execution, an output will appear in which it will be explained if:
\begin{itemize}
\item The process is compliant (in green). It means that the whole process, therefore all the paths that form the process are completely compliant. In this case there will be no other information.
\item The process is not compliant (in red). It means that there is one or more lack of mandatory controls or activities. In this case they will also be explained what they are. They will be indicated with the names used in the Vocabulary and the missing control indicated in the formalized law.
\item The process may not be compliant (in orange). This is a warning. It means that, for example, there could be the mandatory activities but in some branch they may not be in the right sequence, therefore it is indicated where to carry out the check.
\end{itemize}
This is obviously an example case to show the methodology. however, it is important to underline two other aspects:
\begin{itemize}
\item Once a law has been formalized, the file remains saved, so in the event of partial changes by the legislator it will be sufficient to modify only the new part.
\item More than one file can be uploaded, one per regulation. Probably any company must be subject to more than one law (for example that of detail, the national one, the European one. Or in different fields, such as that of the sector, that on employment contracts, that on security, and that on privacy). Being able to enter and therefore check compliance on more than one regulation at the same time, it is also possible that Regorous (in orange) signals steps that could be compliant for a regulation but not fully correct for a second.
\end{itemize}
\begin{figure*}
\includegraphics[width=1\textwidth]{Reg.png}
\caption{Regorous screenshot example for compliance checking}
\label{RegorousForComplianceChecking}
\end{figure*}
\section{CONCLUSION AND FUTURE WORKS}
The proposed pipeline addresses a specific risk management application in a selected healthcare process. However, of course much further work is needed to formalize all the laws, regulations, and guidelines involved in all healthcare processes, in order to have a full exhaustive analysis on how legal compliance is handled in medical procedures.
The research presented here requires manual formalization of the regulations, a task that we aim at semi-automatizing in our future work by following the methodology presented in \cite{Bartolini-etal:2016}. On the other hand, following \cite{Robaldo:2021}, we aim at integrating automatic reasoning on the formalized norms fit to further assist and guide the creation of large repositories or rules, to be possibly shared among multiple business processes.
In addition, the aim is to combine this compliance checking methodology with a framework that also includes a change in the flow of activity in a context of re-organization and optimization of processes. Maintaining the already formalized norms as a background, the purpose is to obtain a methodology able to balance the managerial aspect with that of regulatory compliance.
In summary, this methodology, which is based on processes, uses the modeling of them in the standard BPMN language and which allows you to export the process or import one already created previously via a .bpm file, has a particular advantage. The .bpm extension is quite common for business programs, therefore, the file containing the process could be used by other tools, such as business simulators or value management platforms, to consider other aspects of the company.
In this way, it would be easier to balance the purely organizational aspects of the business (resources, production times and costs or qualitative values of the service) with the aspect of regulatory compliance.
This aspect is becoming more and more important as the legislative evolution is in continuous and increasingly rapid evolution and companies must adapt and change more and more rapidly to remain competitive on the market.
On the other side, it is then important to note that the BPMN and the XML configuration file of Regorous will always need human intervention to be modified. In other words, Artificial Intelligence is only conceived as a tool to only assist and monitor humans in their work activities, but not to take decisions in their place.
The different modules required by the Artificial Intelligence can be prepared and updated by hospital departments with different expertise and authority, thus enhancing efficiency and communication. For instance, the Regorous configuration files can be prepared and maintained by the legal department of the hospital, while the BPMNs by the department concerned with logistics. Regorous will then provides an implicit and efficient communication channel between the two departments, so that each of them will be able to focus on its own competence field while minimizing interactions with the other departments (arrange specific inter-department meetings, etc.)
In the context of the project CANP, the authors have also started working on two directions of improvement for the framework presented here, which will become the main object of future works:
\begin{itemize}
\item The first is to try other deontic logics, for instance reified I/O logic \cite{Robaldo-etal:2020} that, being first-order, provides an higher degree of expressivity as well as a direct interface with OWL \cite{OWL:2012}, the W3C standard for the semantic web.
\item Secondly, as mentioned above we are carrying out researches to verify whether it is possible to semi-automate the manual encoding step of the law via NLP \cite{Sulis-etal:2021,amantea-model:2019}. This is the most critical point, as it is the most crucial for companies but, at the same time, the most complex to automate as it is necessary to take into account the interpretation of the laws.
\end{itemize}
\section*{Acknowledgment}
The current research was conducted in the project ``CANP - CAsa Nel Parco'' of Regione Piemonte funded by POR FESR PIEMONTE 2014-2020. We are grateful for the collaboration of the ``City of Health and Science'' of Torino (Italy). We are also grateful to the partners involved in the project. Livio Robaldo has been supported by the Legal Innovation Lab Wales
operation within Swansea University’s HRC School of Law; the operation has been part-funded by the European Regional Development Fund through the Welsh Government.
|
\section{Introduction}
{
Kohn-Sham density functional theory (KS-DFT) and plane wave (PW) based {\it ab initio} molecular dynamics (AIMD) techniques
are widely used in investigating structural and dynamical properties of condensed matter systems.\cite{marx-hutter-book,Tuckerman-book,leach,frenkel_smit,md_review_tuckerman}
The accuracy of the KS-DFT calculations crucially depends on the choice of exchange-correlation (XC) functionals.
Owing to improved accuracy, hybrid functionals are preferred over the commonly used XC functionals using the Generalized Gradient Approximation (GGA).\cite{Science_DFT_limitations,PRL_DFT_errors,Chemist's_Guide}
Hybrid functionals incorporate a certain fraction of the Hartree-Fock (HF) exchange to the GGA exchange.\cite{Chemist's_Guide,Martin-book,JCP_B3LYP,JCP_PBE0,JCP_HSE}
They give better prediction of energies, structures, electronic properties, reaction barriers, band gap of solids and dynamical properties of liquids.\cite{JCP_B3LYP,JCP_PBE0,JCP_PBE0_model,JCP_HSE,PCCP_HFX,PCCP_HSE,Galli_RSB_JPCL,Chem_Rev_Cohen,jpcl_2011_hfx,jpcl_2019_hfx,jpcl_2020_hfx,Scuseria:JCP:2008,Adamo:2012,JPCB_AIMD_HFX,JCTC_AIMD_HFX,JCP_AIMD_HFX,Mol_Phy_Car_MLWF,Mol_Phy_Car_MLWF_1,JPCB_water_hfx,sagar_JCC,sagar_JCTC,JCP_sagar}
However, the prohibitively high computational cost associated with HF exchange energy evaluation makes the hybrid functionals and PW based AIMD simulations extremely time consuming.\cite{JCP_HFX_Voth}
This limits the routine use of hybrid functionals and PW based AIMD simulations for large condensed matter systems.
A number of promising strategies have been proposed so far to speed-up such calculations, which can be broadly divided into two categories.
The first set of techniques
introduce some approximations in the evaluation of the HF exchange energy, thereby reducing the computational cost.
Several works have been reported in this direction
using localized orbitals,\cite{PRB_Car_Wannier,JCP_AIMD_HFX,Mol_Phy_Car_MLWF,Mol_Phy_Car_MLWF_1,Nature_Car_MLWF,PRL_RSB,JCTC_RSB,JCTC_RSB_1,Galli_RSB_CPL,Galli_RSB_JPCL,Galli_RSB_JPCL1,JCP_sagar,Car_hfx_2019,enabling_part2}
multiple time step (MTS) algorithms,\cite{HFX_Hutter_JCP,MTS_AIMD_Ursula,MTS_AIMD_Steele_3,JCP_2019_sagar,sagar_JCC,sagar_JCTC}
{coordinate scaling,}\cite{JPCL_2018_Bircher,CPC_BIRCHER_2020}
and other {strategies}.\cite{jctc_Bolnykh_2019,single_precision_hfx,HFX_JB,HFX_Goedecker}
The other group of methods improve the performance
by employing massively parallel algorithms.\cite{HFX_Curioni,DUCHEMIN_2010,VARINI_2013,BARNES_2017}
A combination of both the strategies has been also used to achieve remarkable speed-up.\cite{Car_hfx_2019,enabling_part2}
Recently, we proposed an efficient and robust method\cite{JCP_2019_sagar,sagar_JCC} for performing hybrid functionals and PW based AIMD.
We employed a MTS integrator\cite{Tuckerman-book}
scheme based on the adaptively compressed exchange (ACE)~\cite{ACE_Lin,ACE_Lin_1} operator formalism.
To take advantage of the ACE operator formalism, we partitioned
the ionic forces
into computationally cheap fast forces using an approximated ACE operator and computationally costly slow forces due to corrections to the approximated ACE operator.
We denote this method as MTACE hereafter.
This approach
provided a significant speed-up in AIMD simulations by decreasing the number of exact exchange evaluations.
Subsequently, we improved the efficiency of this method\cite{sagar_JCTC}
by employing localized orbitals.
In particular, we used the selected column of the density matrix (SCDM)\cite{SCDM_main} method to obtain localized orbitals, and we used these localized orbitals
to build the ACE operator.\cite{Carnimeo_2019}
We will be denoting this method as s-MTACE.
These methods could speed-up the calculations up to an order of magnitude without
compromising on accuracy.\cite{sagar_JCTC,sagar_JCC,JCP_2019_sagar}
Both MTACE and s-MTACE methods are found to reproduce the structure and dynamics of bulk water and free energetics of chemical reactions in solutions correctly.
In the PW based KS-DFT codes, wavefunctions and KS potentials can be in real and/or reciprocal space, and these representations are inter-converted with the help of three dimensional (3D) FFTs.
For the optimal performance of the 3D parallel FFTs, PW implementations use a slab decomposition of the real 3D FFT grids to distribute the data.
For instance, the 3D FFT grids are distributed along the $X$-direction and the $YZ$ planes are distributed among the MPI tasks (or compute-cores).
Weak scaling performance of such implementations is limited by the number of grid points along the $X$ direction.
For typical DFT calculations today,
the number of grid points along any direction is few hundreds,
while the number of available compute cores on any modern day supercomputing resource is of the order of few thousands to millions.
Thus, slab decomposition based FFTs cannot take the full advantage of the large computational
resources available today.
To overcome this, a task group based parallelization strategy was proposed.\cite{HFX_Curioni,CPG_curioni}
This strategy implemented in the CPMD\cite{cpmd} program is called {{CP~Group}}.
In the {{CP~Group}} implementation of HF exchange computation, the available processors are divided into several task groups and the array that holds the wavefunctions is replicated among these groups.
The total workload of the HF exchange energy computation is divided into several parts and they are distributed evenly among these task groups.
Finally, a global summation across these groups provides the total contribution to the HF exchange energy.
In the present work, we report the implementations of the MTACE and s-MTACE methods together with {{CP~Group}}
within the CPMD program and we present their performance.
In particular, we present the scaling behaviour of the MTACE and s-MTACE methods on a large number of CPU compute cores. %
We will be demonstrating here that a significant improvement in the performance of these methods can be achieved through such an approach.
}
\section{Methods}
\subsection{HF Exchange Operator}
\label{s:methods:reweight}
\begin{figure*}
\centering
\includegraphics[scale=0.5]{flowchart.pdf}
\caption{Flowcharts of the MTACE and s-MTACE algorithms:
%
(a)
SCF procedure in the s-MTACE method, construction of ${\mathbf V}_{\rm X}^{\rm ACE}$ in the first SCF step is done using SCDM-localized orbitals.
%
%
(b) MTS integration used to perform AIMD simulations.
%
}
\label{FIG:1}
\end{figure*}
In conventional KS-DFT calculations using hybrid density functionals and plane waves,
evaluation of the HF exchange contributes the most to the total computational time.
The HF exchange operator ${\mathbf V}_{\rm X}$ is defined as
\begin{equation}
{\mathbf V}_{\rm X}= -\sum_{j}^{N_{\rm orb}} \frac{| \psi _{j} \rangle \langle \psi _{j} |}{r_{12}} \enspace ,
\end{equation}
where, $\{|\psi_{j} \rangle\}$ is the set of occupied KS orbitals.
$N_{\rm orb}$ is the total number of occupied orbitals and $r_{12}=\left | \mathbf r_1 - \mathbf r_2 \right | $.
The ${\mathbf V}_{\rm X}$ operator is applied on a KS orbital $|\psi _{i} \rangle$ as
\begin{equation}
\label{vxonpsi}
\begin{split}
{\mathbf V}_{\rm X}|\psi _{i}\rangle & =- \sum_{j}^{N_{\rm orb}} |\psi _{j} \rangle \left \langle\psi _{j} \left | \left ( r_{12}\right )^{-1} \right | \psi _{i}\right \rangle
\\ & =- \sum_{j}^{N_{\rm orb}} v_{ij}(\mathbf{r}_{1}) |\psi _{j}\rangle \enspace, ~~i=1,....,N_{\rm orb}
\end{split}
\end{equation}
with
\begin{equation}
\label{e1:vij}
v_{ij}(\mathbf {r}_{1})=\left \langle\psi _{j} \left | \left ( r_{12}\right )^{-1} \right | \psi _{i}\right \rangle \enspace.
\end{equation}
The HF exchange energy is calculated as
\begin{equation}
\label{e:xc2}
E^{\textrm{HF}}_{\textrm{X}} = - \sum_{i,j}^{N_{\rm orb}}
\left \langle \psi _{i} \left | v_{ij}({\mathbf {r}}_{1}) \right | \psi _{j}
\right \rangle \enspace .
\end{equation}
For an optimal performance, $v_{ij}(\mathbf{r})$ is usually evaluated in reciprocal space.\cite{JCP_HFX_Voth,PRB_Car_Wannier}
The computational cost for doing Fourier transform scales as $N_{\rm PW}\log N_{\rm PW}$ using the Fast Fourier Transform (FFT) technique, where $N_{\rm PW}$ is total number of PWs.
Therefore, the total computational cost for the evaluation of exchange energy scales as $N_{\rm orb}^2 N_{\rm PW}\log N_{\rm PW}$,\cite{JCP_HFX_Voth} as $v_{ij}(\mathbf{r})$ has to be evaluated $N_{\rm orb}^{2}$ times.
This scaling is the reason behind the computational time required for hybrid functional calculations.
\subsection{ACE Operator
}
\label{s:methods:reweight:ts}
Lin~Lin~\cite{ACE_Lin,ACE_Lin_1} introduced
a low rank decomposition of the
${\mathbf V}_{\rm X}$ operator
in the form
\[ {\mathbf V}_{\rm X}^{\rm ACE}= -\sum_{k}^{N_{\rm orb}} | P_{k} \rangle \langle P_{k} | \enspace . \]
Here, the set of ACE projection vectors $\{|P_{k} \rangle\}$ can be computed by a decomposition of the ${\mathbf V}_{\rm X}$ operator, see Ref.\cite{ACE_Lin,JCP_2019_sagar} for details.
The construction of $\{|P_{k} \rangle\}$ requires evaluation of $\{{\mathbf V}_{\rm X}|\psi _{i}\rangle \}$, which follows the same computationally demanding procedure
requiring $N_{\rm orb}^{2}$ evaluations of $v_{ij}(\mathbf{r})$ as discussed
in the previous section.\cite{ACE_Lin}
However, once the ${\mathbf V}_{\rm X}^{\rm ACE}$ operator is constructed, it can be easily applied on KS orbitals through the evaluation of $N_{\rm orb}^{2}$
{inner} products as
\begin{equation}
{\mathbf V}_{\rm X}^{\rm ACE}|\psi _{i}\rangle=- \sum_{k}^{N_{\rm orb}} |P_{k} \rangle \left \langle P_k | \psi_{i} \right \rangle , \enspace ~~i=1,....,N_{\rm orb} \enspace .
\end{equation}
The advantage of the ACE approach lies in the fact that the application of the ${\mathbf V}_{\rm X}^{\rm ACE}$ operator on each KS orbitals consumes much less time as compared to the ${\mathbf V}_{\rm X}$ operator.
Such a low rank decomposition can be used in multiple ways to speed-up
the HF exchange energy calculations.
\subsection{MTACE Method}
The MTACE method introduced by some of the authors of this paper~\cite{JCP_2019_sagar,sagar_JCC} uses the ACE formalism in the framework of the MTS scheme for speeding-up hybrid functional based AIMD.
The self consistent field (SCF) procedure is
modified to take the benefits of the ACE operator.
In the first SCF step, the ACE operator is constructed after the decomposition of $\mathbf V_{\rm X}$,
which is a computationally demanding step as it involves the computation of $\mathbf V_{\rm X}$.
However, for the remaining SCF steps, we use the same ACE operator constructed
at the first SCF step without recalculating it.
After reaching complete SCF convergence, we compute the
ionic forces $\mathbf F^{\rm ACE}$.
It is to be noted that the optimized wavefunction is most certainly different
to the wavefunction which one would obtain if the ACE operator is updated every SCF step.
As a result, $\mathbf F^{\rm ACE} \neq \mathbf F^{\rm exact}$, where $\mathbf F^{\rm exact}$ is the
ionic force computed using the HF $\mathbf V_{\rm X}$ operator.
We take care of this difference in forces within the MTS algorithm as explained later.
The flowchart for this modified SCF procedure is shown in Figure~\ref{FIG:1}(a).
In the MTACE scheme,
ionic force component can be written
as
\begin{equation}
F^{\rm exact}_K= F^{\textrm{ACE}}_K+ \Delta F_K \enspace , \enspace K=1,\cdots,3N_{\rm atom} \enspace
\end{equation}
with $\Delta F_K = F^{\rm exact}_K - F^{\textrm{ACE}}_K $.
In our earlier works,\cite{JCP_2019_sagar,sagar_JCC}
we have shown that differences in the ionic force components of $\mathbf F^{\rm exact}$ and $\mathbf F^{\textrm{ACE}}$ are very small and
it is justified to consider
$\mathbf F^{\textrm{ACE}}$ as fast and $\Delta \mathbf F$ as slow:
\begin{eqnarray}
F^{\rm slow}_K &\equiv& \Delta F_K = F^{\rm exact}_K - F^{\textrm{ACE}}_K, ~~\mathrm{and} \nonumber \\[1ex]
F^{\rm fast}_K &\equiv& F^{\rm ACE}_K \enspace .
\end{eqnarray}
Finally, we employ the reversible reference system propagator algorithm (r-RESPA) scheme\cite{r-RESPA} which allows us to compute the computationally costly $\Delta \mathbf F$ less frequently, in fact every $n$ MD steps, as compared to computationally cheaper $\mathbf F^{\textrm{ACE}}$, resulting in speeding-up the calculations as shown in Figure~\ref{FIG:1}(b).
For more details, see Ref.~\cite{JCP_2019_sagar}.
We could achieve a speed-up of about 7 times using this method for a periodic system containing $\sim$100 atoms employing only 120 compute cores.
\subsection{s-MTACE Method}
A modification of the MTACE method, namely s-MTACE, was
subsequently proposed,\cite{sagar_JCTC} wherein localized SCDM orbitals\cite{SCDM_main} are used for the construction of the $\mathbf V_{\rm X}^{\rm ACE}$ operator, see Figure~\ref{FIG:1}a.
Based on a rank-revealing QR factorization of ${\mathbf \Psi}^{*}$, where ${\mathbf \Psi}$ is
the matrix with all the occupied KS orbitals, the SCDM method constructs linearly
independent columns of the density matrix ${\mathbf \Gamma}={\mathbf \Psi} {\mathbf \Psi}^{*}$
without computing the full ${\mathbf \Gamma}$ matrix.
The selected columns of ${\mathbf \Gamma}$ are then used to construct the set of orthonormal localized SCDMs
$\{\phi_i\}$.
By screening $\{\phi_i\}$ based on their spatial overlap, it is possible to achieve a substantial reduction in the number of orbitals
involved in the evaluation of $\{{\mathbf V}_{\rm X}|\psi _{i}\rangle \}$ in Equation~(\ref{vxonpsi}).
We screen the orbitals using the criteria
\[ \int d\textbf{r} \left | \phi_{i}(\textbf{r}) \phi_{j}^{*}(\textbf{r}) \right | \geqslant \rho_{\textrm {cut}} \enspace .\]
We have reported that this procedure substantially reduces the cost of $\mathbf V_{\rm X}^{\rm ACE}$ operator construction.\cite{sagar_JCTC}
It has been shown that s-MTACE can achieve one order of magnitude speed-up for a system
containing $\sim$100 atoms using 120 compute cores.
\subsection{s$^\prime$-MTACE Method}
The rank-revealing QR factorization required for the construction of SCDMs is the most time consuming step of the SCDM procedure.
To speed-up these calculations, we employ parallel ScaLAPACK routines.
However, these procedures scale poorly when using a large number of processors.
Giannozi and co-workers\cite{Carnimeo_2019} proposed a way to
improve the computational efficiency of
finding localized SCDMs.
In this method, one pre-selects a column of the density matrix based on the electron density and the gradient of the electron density.
A column with index $i$ is selected only if
\begin{equation}
\rho(\mathbf r_i) > \left \langle \rho \right \rangle \nonumber
~~\mathrm{and}
~~ \nabla\rho (\mathbf r_i) < \left \langle \nabla\rho \right \rangle \enspace .
\end{equation}
Here, $ \left \langle \rho \right \rangle$ and $\left \langle \nabla\rho \right \rangle$ are the average electron density and average gradient of the electron density over the grid points.
This pre-screening scheme substantially reduces the number of grid points which are to be considered for QR factorization.
Now, the call to the ScaLAPACK routine involves a smaller size matrix $\tilde{\mathbf \Psi}^{*}$, thereby improving the
performance.
A successful implementation of this procedure is already available in the Quantum ESPRESSO code.\cite{Espresso_hfx,QE_new_JCP}
We have implemented the same approach in the CPMD program to improve the performance of the
s-MTACE method.
{This
will be referred as s$^\prime$-MTACE method hereafter.
}
\subsection{CP Groups Approach}
\begin{figure*}
\centering
\includegraphics[scale=.25]{cpgrp.jpeg}
\caption{(a) Conventional distribution of the ${\mathbf \Psi}$ matrix with $N_{\rm core}$ compute-cores. The total number of rows ($N_{\rm PW}$) is distributed among these available compute-cores. The part of the matrix residing in different compute-core is shown with different color.
%
(b) The {{CP~Group}} data distribution is shown for $N_{\rm core}$ compute-cores. Total number of compute-cores are divided into $N_{\rm CPG}$ groups. Each group contains $n_{\rm core}$ compute-cores to distribute the ${\mathbf \Psi}$ matrix. Each compute-core contains $N_{\rm PW}/n_{\rm core}$ rows of the ${\mathbf \Psi}$ matrix.}
\label{FIG:2}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[scale=1.2]{scaling.pdf}
\caption{Scaling of the average computational time per BOMD step and average computational time per SCF step for periodic systems containing 32, 64 and 128 water molecules.
%
%
$t_{\rm PBE}$, $t_{\rm PBE0}$, $t_{\rm MTACE}$, $t_{\rm s-MTACE}$ and $t_{\rm s'-MTACE}$ are the average computing time per MD step in {\bf PBE}, {\bf PBE0}, {\bf MTACE}, {\bf s-MTACE} and {\bf s'-MTACE} runs.
$t_{\rm exact}^{\rm SCF}$ is the average computing time per SCF step during the computation of $\mathbf F^{\rm exact}$.
$t_{\rm s-ACE}^{\rm SCF}$ and $t_{\rm s'-ACE}^{\rm SCF}$ are the average computing time for the first SCF step during the computation of $\mathbf F^{\rm s-ACE}$ and $\mathbf F^{\rm s'-ACE}$.
The number of compute cores per {{CP~Group}} ($n_{\rm core}$) are 120, 144 and 192 for systems containing 32, 64 and 128 water molecules, respectively.
}
\label{FIG:3}
\end{figure*}
\begin{figure}
\centering
\includegraphics[scale=0.22]{with-without-cpgrp.jpg}
\caption{CPU time per MD step for {\bf MTACE} with increasing number of compute cores ($N_{\rm core}$) for a periodic 32-water model.
%
Black line shows the scalability of the method without {CP~Group}, and the blue bars display the scaling when {{CP~Group}} was used. One MD step is 0.48~fs.
}
\label{no-cpg-scaling}
\end{figure}
Let us consider that each of the $N_{\rm orb}$ KS orbitals possesses $N_{\rm PW}$ PW coefficients and $N_{\rm core}$
compute-cores are available, then
each compute core stores $N_{\rm PW}/N_{\rm core}$ rows of
the wavefunction matrix ${\mathbf \Psi}$ in a typical implementation of the slab decomposition, see Figure~\ref{FIG:2}(a).
In the {{CP~Group}} approach, as shown in Figure~\ref{FIG:2}(b), the total number of available compute cores $N_{\rm core}$ divided into $N_{\rm CPG}$ groups.
Each of such a {{CP~Group}} possesses $n_{\rm core} = N_{\rm core} / N_{\rm CPG}$ compute cores.
A copy of the whole ${\mathbf \Psi}$ matrix is kept with every group, distributed among the $n_{\rm core}$ compute cores within that group.
As a result, each compute core of a group keeps $N_{\rm PW}/n_{\rm core}$ rows of the ${\mathbf \Psi}$ matrix.
The workload across the task groups
is parallelized over the orbital pairs entering
the exchange integral in such a manner that computations within each {{CP~Group}} is restricted to a subset of orbital pairs.
Through this, computation of $v_{ij}(\mathbf{r})$ (in Equation~\ref{e1:vij}), which has to be performed for all the orbital pairs, can now be done in chunks across the
{{CP~Group}}.
Finally, a global sum across the groups is performed to
evaluate the full exchange operator.
The replication of the whole ${\mathbf \Psi}$ matrix among the {{CP~Groups}} minimizes the inter group communication.
It has been shown that the {{CP~Group}} approach can
be made use to achieve excellent scaling performance in
hybrid functional based calculations on several thousands of compute cores.\cite{HFX_Curioni,CPG_curioni}
\section{Results and Discussion}
We are presenting here the results of scaling
tests of the MTACE, s-MTACE and s$^\prime-$MTACE methods using the {{CP~Groups}} implementation.
Benchmark calculations were carried out for periodic supercells with 32, 64, and 128 water molecules (Table~\ref{tbl1})
\begin{table
\caption{Details of the liquid water systems used in our benchmarking studies. $N_{\rm water}$ is the number of water molecules, $N_{\rm atom}$ is the number of atoms, $N_{\rm orb}$ is the number of orbitals, $N_{\rm grid}$ is the number of grid points, $N_{\rm PW}^{\rm wave}$ is the number of PWs for wavefunction cutoff, and $N_{\rm PW}^{\rm density}$ is the number of PWs for density cutoff.
}\label{tbl1}
\begin{tabular*}{\tblwidth}{@{} CCCCCCC@{} }
\toprule
$N_{\rm water}$ & $N_{\rm atom}$ & $N_{\rm orb}$ & $N_{\rm grid}$ & Cell size ({\AA}) & $N_{\rm PW}^{\rm wave}$ & $N_{\rm PW}^{\rm density}$ \\
\midrule
32 & 96 & 128 & {120} & 9.85 & 39103 & 311563 \\
64 & 192 & 256 & 144 & 12.41 & 77978 & 623469 \\
128 & 384 & 512 & 192 & 15.64 & 156181 & 1247311 \\
\bottomrule
\end{tabular*}
\end{table}
\begin{table
\caption{Different simulation runs;
Note: for the case of MTS based BOMD, we have $n>1$, where $n=\Delta t/ \delta t$.
}\label{tbl:labels}
\begin{tabular*}{\tblwidth}{@{} CCCC @{} }
\toprule
Simulation Label & Functional & $n$ & BOMD Scheme \\
\midrule
{\bf PBE} & GGA/PBE & 1 & Conventional \\
{\bf PBE0} & Hybrid/PBE0 & 1 & Conventional \\
{\bf MTACE} & Hybrid/PBE0 & 15 & MTACE \\
{\bf s-MTACE} & Hybrid/PBE0 & 15 & s-MTACE \\
{\bf s'-MTACE} & Hybrid/PBE0 & 15 & s$^\prime$-MTACE \\
\bottomrule
\end{tabular*}
\end{table}
\begin{table
\caption{
The number of BOMD steps over which the average compute-times were calculated.
$N_{\rm MD}^{\rm X}$ is the total number of MD steps with method {X}.
$\rho_{\rm cut}$ is the cutoff used for the screening of the SCDMs in {\bf s-MTACE} and {\bf s$^\prime$-MTACE} runs. $n_{\rm core}$ is the number of compute cores per task group.
}\label{tbl_nmd}
\begin{tabular*}{\tblwidth}{@{} CCCCCCCC@{} }
\toprule
$N_{\rm water}$ & $N_{\rm MD}^{\rm PBE}$ & $N_{\rm MD}^{\rm PBE0}$ & $N_{\rm MD}^{\rm MTACE}$ & $N_{\rm MD}^{\rm s-MTACE}$ & $N_{\rm MD}^{\rm s\prime -MTACE}$ & $\rho_{\rm cut}$ & $n_{\rm core}$ \\
\midrule
32 & 500 & 300 & 300 & 300 & 300 & $2.5\times10^{-2}$ & 120 \\
64 & 300 & 50 & 150 & 150 & 150 & $1.0\times10^{-2}$ & 144 \\
128 & 300 & 20 & 75 & 75 & 75 & $2.0\times10^{-3}$ & 192 \\
\bottomrule
\end{tabular*}
\end{table}
\begin{table*
\caption{Average computational time per BOMD step and average computational time per SCF step for periodic systems containing 32, 64 and 128 water molecules.
$N_{\rm water}$ is the number of water molecules, $N_{\rm core}$ is the total number of CPU compute cores, $N_{\rm node}$ is the total number of compute nodes, $N_{\rm CPG}$ is the number of TASK groups.
$t_{\rm PBE}$ is the average computing time per MD step using GGA (PBE) functional.
$t_{\rm PBE0}$, $t_{\rm MTACE}$, $t_{\rm s-MTACE}$ and $t_{\rm s'-MTACE}$ are the average computing time per MD step using {\bf PBE0}, {\bf MTACE}, {\bf s-MTACE} and {\bf s$^\prime-$MTACE} methods.
$t_{\rm exact}^{\rm SCF}$ is the average computing time per SCF step during the computation of $\mathbf F^{\rm exact}$.
$t_{\rm s-ACE}^{\rm SCF}$ and $t_{\rm s'-ACE}^{\rm SCF}$ are the average computing time for the first SCF step during the computation of $\mathbf F^{\rm s-ACE}$ and $\mathbf F^{\rm s'-ACE}$.
All the times reported are in seconds.
Calculations were done with 48 compute cores per node, except for the case with $N_{\rm water}=32$ and $N_{\rm CPG}=1$ case where 24 compute cores per node were used. }
\label{tbl3}
\begin{tabular*}{\tblwidth}{@{} LCCCCCCCCCCC@{} }
\toprule
$N_{\rm water}$ & $N_{\rm core}$ & $N_{\rm node}$ & $N_{\rm CPG}$ & $t_{\rm PBE}$ & $t_{\rm PBE0}$ & $t_{\rm MTACE}$ & $t_{\rm s-MTACE}$ & $t_{\rm s'-MTACE}$ & $t_{\rm exact}^{\rm SCF}$ & $t_{\rm s-ACE}^{\rm SCF}$ & $t_{\rm s'-ACE}^{\rm SCF}$ \\
\midrule
32 & 120 & 5 & 1 & 0.68 & 110.19 & 16.20 & 11.05 & 11.12 & 9.92 & 2.14 & 2.27 \\
& 240 & 5 & 2 & 0.58 & 77.88 & 11.62 & 8.61 & 8.06 & 7.00 & 2.17 & 1.67 \\
& 480 & 10 & 4 & 0.43 & 37.38 & 5.71 & 4.88 & 4.15 & 3.35 & 1.60 & 0.87 \\
& 960 & 20 & 8 & 0.36 & 27.64 & 4.27 & 3.90 & 3.12 & 2.47 & 1.39 & 0.62 \\
& 1920 & 40 & 16 & 0.33 & 14.19 & 2.34 & 2.74 & 1.89 & 1.25 & 1.24 & 0.39 \\
& 3840 & 80 & 32 & 0.32 & 8.04 & 1.44 & 2.19 & 1.31 & 0.70 & 1.15 & 0.28 \\
& 7680 & 160 & 64 & 0.34 & 3.60 & 0.86 & 1.87 & 0.98 & 0.30 & 1.12 & 0.22 \\
\midrule
64 & 144 & 3 & 1 & 2.79 & 974.08 & 144.42 & 90.77 & 92.14 & 91.98 & 19.61 & 20.27 \\
& 288 & 6 & 2 & 1.66 & 523.59 & 76.90 & 50.97 & 48.85 & 49.42 & 12.52 & 10.46 \\
& 576 & 12 & 4 & 1.12 & 239.36 & 36.25 & 26.86 & 23.38 & 22.52 & 8.66 & 5.11 \\
& 1152 & 24 & 8 & 0.83 & 113.13 & 17.28 & 15.93 & 11.66 & 10.55 & 6.90 & 2.61 \\
& 2304 & 48 & 16 & 0.76 & 51.79 & 8.40 & 10.92 & 6.31 & 4.81 & 6.16 & 1.54 \\
& 4608 & 96 & 32 & 0.76 & 44.14 & 7.40 & 10.20 & 5.55 & 4.11 & 5.95 & 1.28 \\
& 9216 & 192 & 64 & 0.79 & 20.91 & 4.04 & 8.46 & 3.67 & 1.90 & 5.72 & 0.93 \\
& 18432 & 384 & 128 & 0.81 & 11.19 & 2.69 & 7.73 & 2.90 & 0.98 & 5.62 & 0.76 \\
\midrule
128 & 192 & 4 & 1 & 11.12 & 8776.64 & 1285.86 & 699.56 & 708.00 & 837.51 & 183.01 & 212.59 \\
& 384 & 8 & 2 & 5.83 & 4039.04 & 600.21 & 355.45 & 348.78 & 391.59 & 104.11 & 102.83 \\
& 768 & 16 & 4 & 3.66 & 1829.83 & 268.16 & 176.99 & 157.06 & 175.60 & 65.23 & 47.26 \\
& 1536 & 32 & 8 & 2.57 & 823.68 & 124.82 & 102.89 & 75.66 & 78.96 & 49.24 & 23.61 \\
& 3072 & 64 & 16 & 2.22 & 383.01 & 58.84 & 67.84 & 38.82 & 36.80 & 41.70 & 12.94 \\
& 6144 & 128 & 32 & 2.27 & 294.42 & 45.25 & 60.08 & 30.44 & 27.96 & 39.44 & 9.95 \\
& 12288 & 256 & 64 & 2.31 & 145.67 & 24.25 & 49.56 & 19.02 & 13.73 & 37.41 & 6.81 \\
& 24576 & 512 & 128 & 2.35 & 75.79 & 14.11 & 44.41 & 13.43 & 7.03 & 36.32 & 5.20 \\
& 49152 & 1024 & 256 & 2.54 & 44.80 & 9.55 & 42.37 & 11.22 & 4.08 & 35.92 & 4.61 \\
\bottomrule
\end{tabular*}
\end{table*}
\begin{figure}
\centering
\includegraphics[scale=0.18]{best-performance.jpg}
\caption{ Best performance of the methods discussed in this article for the 32 water system.
%
Performance has been measured as the length of the trajectory (in ps) generated in one day.
%
A time step of 0.48~fs was only used as H atom mass was taken as 1~amu.
%
%
Grey area highlights the schemes where {{CP~Group}} was not used.
120 compute cores were used for all the calculations.
%
Dotted black line indicates the performance of the GGA calculations with 120 compute cores.
The best performance with {{CP~Group}} was obtained for the {\bf MTACE} on 7680 compute cores.
%
Effective speed-up compared to PBE0 is also indicated here (blue).
%
}
\label{best-perf}
\end{figure}
\begin{table*}[]
\caption{Decomposition of total computational time per SCF step for periodic systems containing 32, 64 and 128 water molecules.
Various contributions to $t_{\rm s-ACE}^{\rm SCF}$ and $t_{\rm s'-ACE}^{\rm SCF}$ are reported.
$t_{\rm SCDM}$ or $t_{\rm SCDM}^{'}$ is the time for the localization procedure.
$t_{\rm QR}$ is the time for the QR factorization and $t_{\rm other}$ is the compute time for the rest of the part.
$t_{\rm comput}$ is the compute time for the actual computation of the HF exchange energy.
All compute times reported here are in seconds.
}\label{tbl4}
\begin{tabular}{cc|cccc|cccc}
\toprule
\multicolumn{1}{c}{\multirow{2}{*}{$N_{\rm water}$}} & \multicolumn{1}{c|}{\multirow{2}{*}{$N_{\rm CPG}$}} & \multicolumn{1}{c}{\multirow{2}{*}{$t_{\rm s-ACE}^{\rm SCF}$}} & \multicolumn{2}{c}{$t_{\rm SCDM}$} & \multicolumn{1}{c|}{\multirow{2}{*}{$t_{\rm comput}$}} & \multicolumn{1}{c}{\multirow{2}{*}{$t_{\rm s'-ACE}^{\rm SCF}$}} & \multicolumn{2}{c}{$t_{\rm SCDM}^{'}$} & \multicolumn{1}{c}{\multirow{2}{*}{$t_{\rm comput}$}} \\ \cline{4-5} \cline{8-9}
\multicolumn{1}{c}{} & \multicolumn{1}{c|}{} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{$t_{\rm QR}$} & \multicolumn{1}{c}{$t_{\rm other}$} & \multicolumn{1}{c|}{} & \multicolumn{1}{c}{} & \multicolumn{1}{c}{$t_{\rm QR}$} & \multicolumn{1}{c}{$t_{\rm other}$} & \multicolumn{1}{c}{} \\ \midrule
\multirow{7}{*}{32} & 1 & 2.14 & 0.46 & 0.04 & 1.65 & 2.27 & 0.00 & 0.03 & 2.23 \\
& 2 & 2.17 & 0.93 & 0.06 & 1.18 & 1.67 & 0.01 & 0.05 & 1.62 \\
& 4 & 1.60 & 0.94 & 0.06 & 0.60 & 0.87 & 0.01 & 0.05 & 0.82 \\
& 8 & 1.39 & 0.93 & 0.06 & 0.40 & 0.62 & 0.01 & 0.05 & 0.56 \\
& 16 & 1.24 & 0.94 & 0.05 & 0.25 & 0.39 & 0.01 & 0.05 & 0.33 \\
& 32 & 1.15 & 0.93 & 0.05 & 0.17 & 0.28 & 0.01 & 0.05 & 0.22 \\
& 64 & 1.12 & 0.94 & 0.05 & 0.13 & 0.22 & 0.01 & 0.05 & 0.16 \\ \midrule
\multirow{8}{*}{64} & 1 & 19.61 & 5.10 & 0.24 & 14.27 & 20.27 & 0.01 & 0.20 & 20.06 \\
& 2 & 12.52 & 5.00 & 0.23 & 7.29 & 10.46 & 0.01 & 0.19 & 10.25 \\
& 4 & 8.66 & 4.92 & 0.21 & 3.52 & 5.11 & 0.02 & 0.19 & 4.90 \\
& 8 & 6.90 & 4.90 & 0.22 & 1.78 & 2.61 & 0.02 & 0.19 & 2.40 \\
& 16 & 6.16 & 4.89 & 0.22 & 1.05 & 1.54 & 0.02 & 0.19 & 1.33 \\
& 32 & 5.95 & 4.89 & 0.21 & 0.85 & 1.28 & 0.03 & 0.18 & 1.07 \\
& 64 & 5.72 & 4.90 & 0.21 & 0.61 & 0.93 & 0.04 & 0.18 & 0.70 \\
& 128 & 5.62 & 4.90 & 0.21 & 0.51 & 0.76 & 0.04 & 0.18 & 0.53 \\ \midrule
\multirow{9}{*}{128} & 1 & 183.01 & 32.57 & 1.26 & 149.18 & 212.59 & 0.13 & 1.03 & 211.43 \\
& 2 & 104.11 & 32.01 & 1.24 & 70.87 & 102.83 & 0.13 & 1.02 & 101.68 \\
& 4 & 65.23 & 31.72 & 1.19 & 32.32 & 47.26 & 0.13 & 1.01 & 46.11 \\
& 8 & 49.24 & 31.67 & 1.19 & 16.38 & 23.61 & 0.13 & 1.01 & 22.46 \\
& 16 & 41.70 & 31.64 & 1.18 & 8.89 & 12.94 & 0.14 & 1.01 & 11.79 \\
& 32 & 39.44 & 31.63 & 1.17 & 6.65 & 9.95 & 0.15 & 0.99 & 8.81 \\
& 64 & 37.41 & 31.64 & 1.17 & 4.60 & 6.81 & 0.16 & 0.99 & 5.66 \\
& 128 & 36.32 & 31.65 & 1.17 & 3.50 & 5.20 & 0.18 & 0.98 & 4.03 \\
& 256 & 35.92 & 31.69 & 1.17 & 3.06 & 4.61 & 0.26 & 0.99 & 3.36 \\ \bottomrule
\end{tabular}
\end{table*}
\begin{figure*}
\centering
\includegraphics[scale=0.25]{SCF-time.jpeg}
\caption{Scaling performance of computational time per SCF step for {\bf s-MTACE} and {\bf s$^\prime$-MTACE} with number of {{CP~Group}} ($N_{\rm CPG}$) for the periodic 32 water model.
%
Pie charts show the percentage of time spent for different contributions in one SCF step.
%
Here, $t_{\rm QR}$ is the computational time for QR factorization, $t_{\rm other}$ is the computational time for the rest,
%
and $t_{\rm comput}$ is the time for the actual computation of the HF exchange energy.
}
\label{SCF-decompose}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[scale=0.25]{ps-per-day.jpg}
\caption{Performance measured in units of ps per day for periodic systems containing (a) 32, (b) 64 and (c) 128 water molecules.
%
$X$ axis denotes number of compute cores ($N_{\rm core}$) and
%
%
$Y$ axis is indicating different methods described in this article.
%
$Z$ axis is ps of trajectory that can be generated per day using all of these schemes.
%
}
\label{perf}
\end{figure*}
All the methods and algorithms presented earlier were implemented in { a modified version of}
the {\tt CPMD 4.3} program,\cite{cpmd,KLOFFEL2021}
and adapted for the existing {{CP~Group}} implementation within the program.
We used PBE0~\cite{JCP_PBE0_model} (hybrid) and PBE~\cite{PRL_GGA_PBE} (GGA) functionals
for all the computations.
Core electrons were accounted by using
norm-conserving Troullier Martin pseudopotentials.~\cite{PRB_TM}
A cutoff energy of 80~Ry was used to expand the wavefunctions in the PW basis set.
We carried out Born-Oppenheimer molecular dynamics (BOMD) simulations and the wavefunction convergence criteria in SCF calculations was set to $10^{-6}$ au for the wavefunction gradients.
At every MD step, initial guess of the wavefunctions was constructed based on the Always Stable Predictor Corrector extrapolation scheme\cite{JCC_ASPC} with order 5.
For standard PBE and PBE0 calculations, the standard Velocity Verlet scheme was employed with a timestep of $\Delta t = 0.48$~fs.
However, for the PBE0 runs with the MTS scheme, $\delta t = 0.48$~fs and $\Delta t = 7.2$~fs (i.e., $n=15$) were taken as the smaller and larger time steps, respectively.
We use the labels {\bf PBE}, {\bf PBE0}, {\bf MTACE}, {\bf s-MTACE}, and {\bf s$^\prime$-MTACE}
to indicate different methods used in this work, see Table~\ref{tbl:labels}.
All the benchmark calculations presented here were performed on SuperMUC-NG located at Leibniz Supercomputing Center (LRZ).
The compute nodes are equipped with two Intel\textsuperscript{\textregistered} Skylake Xeon Platinum 8174 processors (24 compute cores per processor).
Each compute node has 48 compute cores and 96~GB memory.
The nodes are interconnected through a fast Intel\textsuperscript{\textregistered} OmniPath network with 100 {\tt Gbit/s} speed.
In Table~\ref{tbl_nmd}, we give the number of steps for which average compute timings were calculated
and the values of $\rho_{\rm cut}$ for the s-MTACE runs.
For an ideal load balancing, the number of compute cores per {{CP~Group}} ($n_{\rm core}$) is chosen to be equal to the number of grid points in the $X$-direction.
We chose $n_{\rm core}$ as 120, 144 and 192 for systems containing 32, 64 and 128 water molecules, respectively, see Table~\ref{tbl_nmd}.
The average computational time per BOMD step is reported in Table~\ref{tbl3} for the {\bf PBE}, {\bf PBE0}, {\bf MTACE} and {\bf s-MTACE} methods.
Also, the scaling is shown in Figure~\ref{FIG:3}.
First, we observe that the {\bf PBE} calculations have poor scaling with {{CP~Groups}} for all the systems.
This is expected as these calculations lack enough computationally scalable work that can be distributed over the {{CP~Groups}}.
In fact due to the extra overhead of communication and synchronisation, we notice small increase in
computational time per MD step with large number of {{CP~Groups}}.
However, for all the other three methods which use hybrid functionals, we notice considerable improvements in the performance with increase in {{CP~Groups}}.
Noticeably, {\bf PBE0} scales almost perfectly with the number of {{CP~Groups}}.
The scaling behavior of {\bf MTACE} is also
as good as {\bf PBE0}.
It is clear that the scaling of {\bf MTACE} without the {{CP~Group}} approach is poor when more than 120 compute cores
are used for the 32 water system (Figure~\ref{no-cpg-scaling}).
From Table~\ref{tbl3}, we observe that the ratio of the average computing time per MD step for {\bf MTACE} and {\bf PBE} ($t_{\rm MTACE}/t_{\rm GGA}$) decreases with increasing number of {CP~Group}.
This is a consequence of the fact that the {\bf MTACE} calculations are scaling well with {{CP~Groups}} as compared to {\bf PBE}.
Depending on the system size, {\bf MTACE} is only 2-4 times slower than {\bf PBE} runs by employing the combination of MTACE and {{CP~Groups}} when a sufficiently large number of processors
is used.
The best case scenarios for the systems with 32, 64 and 128 water molecules are having
{\bf MTACE} runs clocking only 2.7, 3.5 and 4.3 times slower than {\bf PBE}, respectively.
Further, {\bf MTACE} is giving a 4-5 fold speed-up compared to {\bf PBE0} when using the highest number of {{CP~Groups}} we employed.
For a 32 water system, the computing time for one MD step is now only 0.86~s for hybrid
functional based BOMD with the {MTACE} method when 7680 compute-cores were taken.
Figure~\ref{best-perf} shows the improved performance of the method with and without the {{CP~Group}} approach.
Similar enhancement in performance is also seen for
systems with 64 and 128 water molecules.
These results are encouraging as we can generate long trajectories at the level of hybrid functionals within a shorter time by making use of large computing resources.
The best computing performances we obtained are 48, 15 and 4 ps of trajectory per day for systems with 32, 64 and 128 water molecules, respectively (see also Figure~\ref{perf}).
We would like to emphasis that $\delta t=0.48$~fs was taken considering that H atoms were assigned 1~amu mass.
By using a deuterium mass for H atoms, a $\delta t=1$~fs could be used, resulting in doubling the simulation performance (ps/day).
Our calculations show that the scaling behavior of {\bf s-MTACE} deteriorates beyond a certain number of compute cores.
In order to understand this poor scaling behavior, we have looked at the average computing time per SCF step during different modes of force calculations.
We label the average computing times per SCF step during the computation of $\mathbf F^{\rm exact}$
and $\mathbf F^{\rm s-ACE}$ as $t_{\rm exact}^{\rm SCF}$, and
$t_{\rm s-ACE}^{\rm SCF}$, respectively.
We observe that $t_{\rm s-ACE}^{\rm SCF}$ scales poorly when {{CP~Groups}} is large (Table ~\ref{tbl3}).
To scrutinize the poor scaling of $t_{\rm s-ACE}^{\rm SCF}$, we decomposed the time for various stages of computation (Table ~\ref{tbl4} and Figure~\ref{SCF-decompose}).
The parallel QR factorization with the ScaLAPACK routines ($t_{\rm QR}$) contributes mostly to the
computational overhead for the SCDM localization procedure ($t_{\rm SCDM}$).
It is clear that $t_{\rm QR}$ scales poorly with $N_{\rm CPG}$, resulting in an overall poor scaling of $t_{\rm SCDM}$.
On the other hand, computation of the HF exchange integrals ($t_{\rm comput}$) scales
well with $N_{\rm CPG}$.
To overcome this and to improve the scalability
we adopted the {s$^\prime$-MTACE} method.
The reported results in Table~\ref{tbl4} and Figure~\ref{SCF-decompose} suggest that the QR factorization with the ScaLAPACK routine now consumes negligible amount of computing time.
As a result,the poor scaling of $t_{\rm QR}$ has no significant effect on
the overall scaling of the method.
It has to be noted that the pre-screening procedure used in the {s$^\prime$-MTACE}
method slightly deteriorates the localization properties of the computed SCDMs, resulting
in a large number of overlapping pairs during the construction of the ACE operator.
Consequently, $t_{\rm comput}$ turns out to be higher than that of the s-MTACE scheme.
The overall scaling behavior of MD timings ($t_{\rm s^\prime-MTACE}$) is satisfactory as
can be seen in Table~\ref{tbl3} and Figure~\ref{FIG:3}.
\section{Summary}
We have presented a detailed benchmarking study on the computational performance of the
{\bf MTACE} and {\bf s-MTACE} algorithms with task-groups ({CP~Group}) for performing AIMD simulations
with hybrid density functionals and plane waves.
In our implementations of the {\bf MTACE} and {\bf s-MTACE} methods using the {{CP~Group}} environment in the CPMD program, orbital pairs are distributed across the processor groups to achieve a better scaling performance.
Through this implementation, we are able to accomplish excellent scaling behavior beyond $\sim$100 compute cores,
even for typical system sizes with $\sim$100 atoms.
Further, excellent speed-up has been also seen while using this implementation.
In the best performance achieved for a model system containing 32 water molecules,
computational overhead for doing hybrid density functional based AIMD is only 3 times more expensive than with
GGA.
Our implementation has resulted in boosting
the performance of hybrid functional based AIMD of
this system by a factor of 121 (see Figure~\ref{best-perf}).
The performance of the {s-MTACE} method was better than the {MTACE} for a small number of {{CP~Groups}}, however, it deteriorated with increasing number of the {{CP~Groups}}.
This problem was overcome by the implementation of the { s$^\prime$-MTACE} method.
Our results suggest that either the {MTACE} or {s$^\prime$-MTACE} method in combination with {{CP~Groups}}
is ideal for running hybrid density functional based AIMD simulations on high-performance computers.
For system with finite band gap, {s$^\prime$-MTACE} should perform better than {MTACE}.
\section*{Acknowledgement}
{Financial support from the National Supercomputing Mission (Subgroup Materials and Computational Chemistry) and Science and Engineering Research Board (India)
under the MATRICS scheme (Ref. No. MTR/2019/000359) and from the German
Research Foundation (DFG) through Research Unit FOR 1878 (funCOS) and
Collaborative Research Center SFB 953 (project number 182849149) are
gratefully acknowledged.
RK thanks the Council of Scientific \& Industrial Research (CSIR), India for
her Junior Research Fellowship (JRF).
Computational resources were provided by
SuperMUC-NG (project pn98fa) at Leibniz Supercomputing Centre (LRZ).
}
\bibliographystyle{apsrev}
|
\section{Introduction}
In the early 1980s, Nieh and Yan derived a four-dimensional topological invariant involving torsion~\cite{Nieh-Yan1982,NiehIJMPA}, which has been since then of great importance in different branches of physics including general relativity~\cite{Date2009,Perez2009,Kaul2012,Sengupta_2013,Corichi2}, $f(\mathfrak{R})$ gravity~\cite{Bombacigno_2021_fR}, and condensed matter physics~\cite{Nissinen_2020_PRL}. Furthermore, there has also been a remarkable interest in exploring the generalizations of the Nieh-Yan topological invariant to higher dimensions. In last direction, we can mention the work of Mardones and Zanelli~\cite{Mardones_1991} that introduces torsional invariants in higher dimensions and the higher dimensional generalizations of the Nieh-Yan invariant studied by Chand\'{\i}a and Zanelli~\cite{Zanelli1997}, where explicit expressions for four Nieh-Yan-like invariants in eight dimensions were reported. Subsequently, in Refs.~\cite{Zanelli1998,HanYing1999} it was shown how to systematically construct Nieh-Yan-like invariants in $4k$ ($k\in Z$) dimensions. Another natural, and perhaps more interesting, direction to generalize the Nieh-Yan topological invariant is to modify its structure but remaining in four-dimensional manifolds. Generalizations of this kind, in particular, would be expected to play a relevant role in the realm of general relativity and related theories of gravity. Recently, it has been reported new torsional topological invariants in four dimensions~\cite{Nieh2018,Montesinos2021}. However, to our knowledge, there are no such four-dimensional generalizations of the Nieh-Yan topological invariant up to now. In this paper we fill out this gap.
More precisely, we present a systematic construction of generalizations of the Nieh-Yan topological invariant in four dimensions. Our approach follows the same spirit of Refs.~\cite{Nieh-Yan1982,Zanelli1997,Nieh2018,Montesinos2021} that leads to the Nieh-Yan invariant and the new torsional invariants in four dimensions recently introduced, but involving an $SO(4,1)$~[$SO(5)$] connection constructed out from an $SO(3,1)$~[$SO(4)$] connection and three $SO(3,1)$~[$SO(4)$] tensor $1$-forms. The new topological invariants, like the Nieh-Yan topological invariant, are integrals of 4-forms that arise from the difference of two Pontrjagin forms. We report explicit expressions for the generalizations of the Nieh-Yan topological invariant and additionally show how the Nieh-Yan invariant emerges as a particular case.
\section{Preliminaries}
We begin by fixing the notation. In the first-order formalism of general relativity, the fundamental variables to describe the gravitational field are an orthonormal frame of $1$-forms $e^I$ and a connection 1-form $\omega^I{}_J$ that is compatible with the metric $(\eta_{IJ}) =\mbox{diag} (\sigma, 1, 1, 1)$, i.e., $d \eta_{IJ} - \omega^K{}_I \eta_{KJ} - \omega^K{}_J \eta_{IK}=0$. This means that $\omega_{IJ}=-\omega_{JI}$ since the frame indices $I,J,K, \ldots$ (taking the values $0, 1, 2, 3$) are raised and lowered with $\eta_{IJ}$. The frame rotation group corresponds to $SO(3,1)$ for $\sigma=-1$ and to $SO(4)$ for $\sigma=1$. The $SO(3,1)$-invariant [$SO(4)$-invariant] tensor $\epsilon_{IJKL}$ is totally antisymmetric and such that $\epsilon_{0123}=1$. The dual basis of $e^I$ is $\partial_I$ and satisfies $\partial_J \intprod \, e^I = \delta^I_J$ where $\intprod$ stands for the contraction of a differential form with a vector field. Furthermore, $R^I{}_J =d \omega^I{}_J + \omega^I{}_K \wedge \omega^K{}_J$ is curvature of $\omega^I{}_J$ and the covariant derivative of an internal tensor that is also a $p$-form $T^{I_1 \cdots I_k}{}_{J_1 \cdots J_l}$ with respect to $\omega^I{}_J$ is given by $D T^{I_1 \cdots I_r}{}_{J_1 \cdots J_l} = d T^{I_1 \cdots I_r}{}_{J_1 \cdots J_l}+ \omega^{I_1}{}_{K_1} \wedge T^{K_1 \cdots I_r}{}_{J_1 \cdots J_l} + \cdots + \omega^{I_r}{}_{K_r} \wedge T^{I_1 \cdots K_r}{}_{J_1 \cdots J_l} - \omega^{K_1}{}_{J_1} \wedge T^{I_1 \cdots I_r}{}_{K_1 \cdots J_l} - \cdots - \omega^{K_l}{}_{J_l} \wedge T^{I_1 \cdots I_r}{}_{J_1 \cdots K_l} $. Likewise, the spin connection $\Gamma^I{}_J$ is defined by $d e^I + \Gamma^I{}_J \wedge e^J=0$ and $d \eta_{IJ} - \Gamma^K{}_I \eta_{KJ} - \Gamma^K{}_J \eta_{IK}=0$, and then $\Gamma_{IJ}=-\Gamma_{JI}$. In addition, $\star$ denotes the Hodge dual operator $\star(e_{I_1} \wedge \dots\wedge e_{I_r}) =(1/(4-r)!) \, \epsilon_{I_1\dots I_rI_{r+1} \dots I_4} e^{I_{r+1}} \wedge \dots \wedge e^{I_4}$ and we define the internal dual of an antisymmetric object $U_{IJ}=-U_{JI}$ as $\ast U_{IJ}:=(1/2) \epsilon_{IJKL} U^{KL}$.
Consider the $SO(4,1)$ [$SO(5)$] connection 1-form $\mathcal{A}^a{}_b$ ($\mathcal{A}^{ab}=-\mathcal{A}^{ba}$) constructed out from the connection $\omega^I{}_J$ and the frame $e^I$~\cite{Nieh-Yan1982,Zanelli1997},
\begin{eqnarray}
\left ( \mathcal{A}^a{}_b \right ) =
\left( {\begin{array}{rc}
\omega^I{}_J & \frac{1}{l}e^I\\ -\frac{1}{l}e_J & 0 \end{array} } \right ), \label{NYconn}
\end{eqnarray}
where $l$ is an arbitrary constant and the indices $a,b,\dots$ run from 0 to 4 and are raised and lowered with the metric $(\eta_{ab}) =\mbox{diag} (\sigma, 1, 1, 1, 1)$. The curvature $\mathcal{F}^a{}_b=d\mathcal{A}^a{}_b + \mathcal{A}^a{}_c \wedge \mathcal{A}^c{}_b$ of $\mathcal{A}^a{}_b$ takes the form
\begin{eqnarray}
\mathcal{F}^I{}_J=R^I{}_J-\frac{1}{l^2}e^I\wedge e_J,\hspace{5mm} \mathcal{F}^I{}_4=\frac{1}{l} De^I. \label{NYcurv}
\end{eqnarray}
It is straightforward to check that the connection $\mathcal{A}^a{}_b$ fulfills the identity
\begin{equation}
\mathcal{F}^{a}{}_{b}\wedge \mathcal{F}^{b}{}_{a}=d\left[ \mathcal{A}^{a}{}_{b} \wedge \left( \mathcal{F}^{b}{}_{a} - \frac{1}{3} \mathcal{A}^{b}{}_{c} \wedge \mathcal{A}^{c}{}_{a} \right) \right], \label{PontrF}
\end{equation}
which involves the Pontrjagin form $\mathcal{F}^{a}{}_{b}\wedge \mathcal{F}^{b}{}_{a}$. On account of~\eqref{NYconn} and~\eqref{NYcurv}, the identity~\eqref{PontrF} leads to
\begin{align}
&R^{I}{}_{J}\wedge R^{J}{}_{I}-\frac{2}{l^2} \left( De^I \wedge De_I - e^I \wedge e^J \wedge R_{IJ} \right) \nonumber \\
&=d\left[ \omega^{I}{}_{J} \wedge \left( R^{J}{}_{I} - \frac{1}{3} \omega^{J}{}_{K} \wedge \omega^{K}{}_{I} \right) - \frac{2}{l^2} e^I \wedge De_I \right]. \label{PontrF2}
\end{align}
Since the parameter $l$ is arbitrary, we can equate the terms with identical powers of $l$ on both sides of~\eqref{PontrF2}. From the zero power of $l$ we get the identity
\begin{equation}
R^{I}{}_{J}\wedge R^{J}{}_{I}=d\left[ \omega^{I}{}_{J} \wedge \left( R^{J}{}_{I} - \frac{1}{3} \omega^{J}{}_{K} \wedge \omega^{K}{}_{I} \right) \right], \label{PontrR}
\end{equation}
satisfied by the Pontrjagin form $R^{I}{}_{J}\wedge R^{J}{}_{I}$, and from the second power of $1/l$ we obtain the identity
\begin{equation}
De^I \wedge De_I - e^I \wedge e^J \wedge R_{IJ} =d\left(e^I \wedge De_I \right), \label{NYiden}
\end{equation}
satisfied by the Nieh-Yan form~\cite{Nieh-Yan1982,NiehIJMPA}
\begin{equation}
N\equiv De^I \wedge De_I - e^I \wedge e^J \wedge R_{IJ}.\label{NYinv}
\end{equation}
In the next section, we apply the methodology outlined in the previous paragraphs to derive 4-forms that are generalizations of the Nieh-Yan form, whose integrals in four dimensions are topological invariants.
\section{Generalizations of the Nieh-Yan topological invariant}\label{sec-general}
We begin by defining the $SO(4,1)$ [$SO(5)$] connection 1-form $\Omega^a{}_b$ ($\Omega^{ab}=-\Omega^{ba}$) by
\begin{eqnarray}
\left ( \Omega^a{}_b \right ) =
\left( {\begin{array}{rc}
\omega^I{}_J & \frac{1}{l}Q^I\\ -\frac{1}{l}Q_J & 0 \end{array} } \right ), \label{conn1}
\end{eqnarray}
where $l$ is an arbitrary constant and the $1$-form $Q^I$ is an $SO(3,1)$~[$SO(4)$] tensor whose specific expression is not needed to what follows. The curvature $F^a{}_b=d\Omega^a{}_b + \Omega^a{}_c \wedge \Omega^c{}_b$ of $\Omega^a{}_b$ takes the form
\begin{eqnarray}
F^I{}_J=R^I{}_J-\frac{1}{l^2}Q^I\wedge Q_J,\hspace{5mm} F^I{}_4=\frac{1}{l} DQ^I. \label{curv1}
\end{eqnarray}
Following the method employed in~\cite{Nieh2018,Montesinos2021}, we further introduce the $SO(4,1)$ [$SO(5)$] connection 1-form $\Omega'^a{}_b$ defined by
\begin{equation}
\Omega'^a{}_b\equiv\Omega^a{}_b + \xi K^a{}_b,\label{conn2}
\end{equation}
where $\xi$ is an arbitrary parameter and $K^a{}_b$ ($K^{ab}=-K^{ba}$) is an $SO(4,1)$ [$SO(5)$] 1-form. Also, as in the case of $Q^I$, the particular expression for $K^a{}_b$ does not need to be specified beforehand. The corresponding curvature of $\Omega'^a{}_b$ is $F'^a{}_b =d \Omega'^a{}_b + \Omega'^a{}_c \wedge \Omega'^c{}_b $ and can be written as
\begin{equation}
F'^a{}_b =F^a{}_b + \xi \mathbb{D} K^a{}_b + \xi^2 K^a{}_c \wedge K^c{}_b, \label{curv2}
\end{equation}
where $\mathbb{D}$ is the covariant derivative computed with respect to $\Omega^a{}_b$ given by~\eqref{conn1}, and explicitly reads $\mathbb{D} K^a{}_b=d K^a{}_b + \Omega^a{}_c \wedge K^c{}_b - \Omega^c{}_b \wedge K^a{}_c$.
The identity~\eqref{PontrF} for the connection $\Omega'^a{}_b$ reads
\begin{equation}
F'^{a}{}_{b}\wedge F'^{b}{}_{a}=d\left[ \Omega'^{a}{}_{b} \wedge \left( F'^{b}{}_{a} - \frac{1}{3} \Omega'^{b}{}_{c} \wedge \Omega'^{c}{}_{a} \right) \right], \label{PontrFprime}
\end{equation}
which involves the Pontrjagin form $F'^{a}{}_{b}\wedge F'^{b}{}_{a}$. Substituting~\eqref{conn2} and~\eqref{curv2} into~\eqref{PontrFprime}, we arrive at
\begin{eqnarray}
&&F^{a}{}_{b}\wedge F^{b}{}_{a} + 2 \xi F^{a}{}_{b} \wedge \mathbb{D} K^b{}_a + \xi^2 \big( \mathbb{D} K^a{}_b \wedge \mathbb{D} K^b{}_a \nonumber \\
&& \hspace{3mm} + 2 K^a{}_b \wedge K^b{}_c \wedge F^c{}_a \big) + 2 \xi^3 K^a{}_b \wedge K^b{}_c \wedge \mathbb{D} K^c{}_a \nonumber\\
&& = d\bigg[ \Omega^{a}{}_{b} \wedge \bigg( F^{b}{}_{a} - \frac{1}{3} \Omega^{b}{}_{c} \wedge \Omega^{c}{}_{a} \bigg)+ \xi \big( \Omega^{a}{}_{b} \wedge \mathbb{D} K^b{}_a \nonumber\\
&&\hspace{3mm} -\Omega^a{}_b \wedge \Omega^b{}_c \wedge K^c{}_a + K^{a}{}_{b} \wedge F^b{}_a \big) + \xi^2 K^a{}_b \wedge \mathbb{D} K^b{}_a \nonumber\\
&&\hspace{3mm} + \frac{2}{3} \xi^3 K^a{}_b \wedge K^b{}_c \wedge K^c{}_a \bigg].\label{PontrFprime2}
\end{eqnarray}
Bearing in mind that $\xi$ is an arbitrary parameter, we can equate the terms with the same powers of $\xi$ on both sides of the identity~\eqref{PontrFprime2}. By doing this, from the zero, first, second, and third powers of $\xi$ we obtain, respectively, the following identities:
\begin{equation}
F^{a}{}_{b}\wedge F^{b}{}_{a}=d\left[ \Omega^{a}{}_{b} \wedge \left( F^{b}{}_{a} - \frac{1}{3} \Omega^{b}{}_{c} \wedge \Omega^{c}{}_{a} \right) \right], \label{Iden-I}
\end{equation}
\begin{eqnarray}
2 F^{a}{}_{b} \wedge \mathbb{D} K^b{}_a &=& d \big(\Omega^{a}{}_{b} \wedge \mathbb{D} K^b{}_a -\Omega^a{}_b \wedge \Omega^b{}_c \wedge K^c{}_a \nonumber \\
&&+ K^{a}{}_{b} \wedge F^b{}_a \big), \label{Iden-II}
\end{eqnarray}
\begin{equation}
\mathbb{D} K^a{}_b \wedge \mathbb{D} K^b{}_a + 2 K^a{}_b \wedge K^b{}_c \wedge F^c{}_a = d \big(K^a{}_b \wedge \mathbb{D} K^b{}_a\big), \label{Iden-III}
\end{equation}
\begin{equation}
K^a{}_b \wedge K^b{}_c \wedge \mathbb{D} K^c{}_a= d \left(\frac{1}{3} K^a{}_b \wedge K^b{}_c \wedge K^c{}_a \right). \label{Iden-IV}
\end{equation}
The identity~\eqref{Iden-I} involves the Pontrjagin form $F^{a}{}_{b}\wedge F^{b}{}_{a}$, whereas the identities~\eqref{Iden-II},~\eqref{Iden-III}, and~\eqref{Iden-IV} involve, respectively, the 4-forms $F^{a}{}_{b} \wedge \mathbb{D} K^b{}_a$, $\mathbb{D} K^a{}_b \wedge \mathbb{D} K^b{}_a + 2 K^a{}_b \wedge K^b{}_c \wedge F^c{}_a$, and $K^a{}_b \wedge K^b{}_c \wedge \mathbb{D} K^c{}_a$. The integrals of these 4-forms are topological invariants in four dimensions.
In the following subsections, we use the connection~\eqref{conn1} and its curvature~\eqref{curv1} to write down~\eqref{Iden-I}-\eqref{Iden-IV} and obtain new identities from the different powers of the parameter $l$.
\subsection{Identities coming from~\eqref{Iden-I}}
Substituting~\eqref{conn1} and~\eqref{curv1} into~\eqref{Iden-I}, we obtain
\begin{align}
&R^{I}{}_{J}\wedge R^{J}{}_{I}-\frac{2}{l^2} \left( DQ^I \wedge DQ_I - Q^I \wedge Q^J \wedge R_{IJ} \right) \nonumber \\
&=d\left[ \omega^{I}{}_{J} \wedge \left( R^{J}{}_{I} - \frac{1}{3} \omega^{J}{}_{K} \wedge \omega^{K}{}_{I} \right) - \frac{2}{l^2} Q^I \wedge DQ_I \right].\label{Iden2-I}
\end{align}
As before we proceed to match the terms of~\eqref{Iden2-I} with identical powers of $1/l$. From the zero power of $1/l$ we recover the original identity~\eqref{PontrR} and from the second power of $1/l$ we obtain the new identity
\begin{equation}
DQ^I \wedge DQ_I - Q^I \wedge Q^J \wedge R_{IJ} =d\left(Q^I \wedge DQ_I \right). \label{Iden-I-2}
\end{equation}
This identity involves the 4-form
\begin{equation}
I_1\equiv DQ^I \wedge DQ_I - Q^I \wedge Q^J \wedge R_{IJ}, \label{NewInv1}
\end{equation}
whose integral in four dimensions is a topological invariant. Notice that $I_1$ constitutes a generalization of the Nieh-Yan form~\eqref{NYinv}. In fact, it is direct to see that in the particular case $Q^I\equiv e^I$, the 4-form~\eqref{NewInv1} reduces to the Nieh-Yan form~\eqref{NYinv}, $I_1=N$. Furthermore, notice from~\eqref{Iden-I} and~\eqref{Iden2-I} that $I_1$ is proportional to the difference of the Pontrjagin forms $F^{a}{}_{b}\wedge F^{b}{}_{a}$ and $R^{I}{}_{J}\wedge R^{J}{}_{I}$.
As an illustration, let us consider some other particular cases of the 4-form~\eqref{NewInv1}:
\begin{enumerate}[(i)]
\item Setting $Q^I\equiv f e^I$ where $f$ is a real function, the 4-form $I_1$ reads
\begin{eqnarray}
I_1&=& D(f e^I) \wedge D(f e_I) - f^2 e^I \wedge e^J \wedge R_{IJ} \nonumber\\
&=&f^2\left( De^I \wedge De_I - e^I \wedge e^J \wedge R_{IJ} \right) \nonumber\\
&&+2f Df \wedge e^I \wedge De_I. \,\,\,\, \label{exampleA1}
\end{eqnarray}
Note that this 4-form vanishes if there is no torsion, $De^I=0$. See the conclusions for a comment about the 4-form~\eqref{exampleA1}.
\item Setting $Q^I\equiv \star(R^I{}_J \wedge e^J)$, the 4-form $I_1$ takes the form
\begin{eqnarray}
I_1&=&D[\star(R^I{}_J \wedge e^J)] \wedge D[\star(R_{IK} \wedge e^K)] \nonumber \\
&&- \star(R^I{}_K \wedge e^K) \wedge \star(R^J{}_L \wedge e^L) \wedge R_{IJ}. \label{exampleA2}
\end{eqnarray}
Notice that, using $D^2e^I=R^I{}_J\wedge e^J$, the 4-form~\eqref{exampleA2} vanishes when there is no torsion, $De^I=0$.
\item Setting $Q^I\equiv \star[(\star R^I{}_J) \wedge e^J]=-\sigma \mathfrak{R}^I{}_J e^J$ where $\mathfrak{R}_{IJ}=R^K{}_{IKJ}$ with $R^I{}_J=(1/2)R^I{}_{JKL} e^K \wedge e^L$, the 4-form $I_1$ is
\begin{eqnarray}
I_1&=& D(\mathfrak{R}^I{}_J e^J) \wedge D(\mathfrak{R}_{IK} e^K) \nonumber\\
&&- \mathfrak{R}^I{}_K \mathfrak{R}^J{}_L e^K \wedge e^L \wedge R_{IJ} \nonumber\\
&=&\mathfrak{R}^I{}_K \mathfrak{R}^J{}_L \left(\eta_{IJ} De^K \wedge De^L -e^K \wedge e^L \wedge R_{IJ}\right) \nonumber\\
&&-D\mathfrak{R}^I{}_J \wedge D\mathfrak{R}_{IK} \wedge e^J \wedge e^K \nonumber\\
&& + 2 \mathfrak{R}^I{}_J D\mathfrak{R}_{IK} \wedge De^J \wedge e^K.
\end{eqnarray}
It is worth noting that $\mathfrak{R}_{IJ}\neq \mathfrak{R}_{JI}$ generically because nothing has been stated about torsion.
\item Setting $Q^I\equiv \star[(\ast R^I{}_J) \wedge e^J]=\sigma G_J{}^I e^J$ where $G_{IJ}=\mathfrak{R}_{IJ}-(1/2) \mathfrak{R} \eta_{IJ}$ with $\mathfrak{R}=\mathfrak{R}^I{}_I$, the 4-form $I_1$ takes the form
\begin{eqnarray}
I_1&=& D(G_J{}^I e^J) \wedge D(G_{KI} e^K) \nonumber\\
&& - G_K{}^I e^K \wedge G_L{}^J e^L \wedge R_{IJ} \nonumber\\
&=&G_K{}^I G_L{}^J \left(\eta_{IJ} De^K \wedge De^L -e^K \wedge e^L \wedge R_{IJ}\right) \nonumber\\
&&-DG_J{}^I \wedge DG_{KI} \wedge e^J \wedge e^K \nonumber\\
&& + 2 G_J{}^I DG_{KI} \wedge De^J \wedge e^K.
\end{eqnarray}
We also remark here that $G_{IJ}\neq G_{JI}$ generically because nothing has been stated about torsion.
\item Setting $Q^I\equiv \epsilon^{IJKL} \partial_J \intprod \, R_{KL}$, the 4-form $I_1$ reads
\begin{eqnarray}
I_1&=& \epsilon^{IJKL} \epsilon_{IMNP} D(\partial_J \intprod \, R_{KL}) \wedge D(\partial^M \intprod \, R^{NP})\nonumber\\
&& - \epsilon^{IKLM} \epsilon^{JPQS}(\partial_K \intprod \, R_{LM}) \wedge (\partial_P \intprod \, R_{QS}) \wedge R_{IJ} \nonumber\\
&=& 2 \sigma D(\partial_I \intprod \, R_{JK}) \wedge \big[ D(\partial^I \intprod \, R^{JK}) - 2 D(\partial^J \intprod \, R^{IK}) \big] \nonumber\\
&& + 2 \sigma (\partial_I \intprod \, R_{JK}) \wedge \big[ 2 (\partial^J \intprod \, R^{KL}) \wedge R^{I}{}_{L} \nonumber\\
&& + (\partial^L \intprod \, R^{JK}) \wedge R^{I}{}_{L}-2 (\partial^I \intprod \, R^{KL}) \wedge R^{J}{}_{L} \nonumber\\
&&+ 2 (\partial^K \intprod \, R^{IL}) \wedge R^{J}{}_{L}- 2 (\partial^L \intprod \, R^{IK}) \wedge R^{J}{}_{L}\big].
\end{eqnarray}
\end{enumerate}
It is not difficult to realize that two other possible choices for $Q^I$ are $Q^I\equiv \epsilon^{IJKL} \partial_J \intprod \, (\ast R_{KL})$ and $Q^I\equiv \epsilon^{IJKL} \partial_J \intprod \, (\star R_{KL})$. Nevertheless, since $ \epsilon^{IJKL} \partial_J \intprod \, (\ast R_{KL})=-2\sigma \mathfrak{R}^I{}_J e^J$ and $\epsilon^{IJKL} \partial_J \intprod \, (\star R_{KL})=2\sigma G_J{}^I e^J$, they correspond to the cases (iii) and (iv) and thus there is no need to consider them. It is worth mentioning that the choices for $Q^I$ presented are not meant to be exhaustive, but rather illustrative of the variety of 4-forms that can arise from~\eqref{NewInv1}.
\subsection{Identities coming from~\eqref{Iden-II}}
Using~\eqref{conn1} and~\eqref{curv1}, the identity~\eqref{Iden-II} leads to
\begin{align}
&2 R^{I}{}_{J}\wedge DK^{J}{}_{I} -\frac{2}{l} \big( 2DK^{I}{}_{4} \wedge DQ_I - K^I{}_4 \wedge Q^J \wedge R_{IJ} \nonumber\\
&\hspace{3mm}- Q^I \wedge K^J{}_4 \wedge R_{IJ} \big) \nonumber\\
&\hspace{3mm} - \frac{2}{l^2} \big( Q^I \wedge Q_J \wedge DK^J{}_I + 2 DQ^I \wedge Q_J \wedge K^J{}_I \big) \nonumber\\
&= d\bigg[ \omega^{I}{}_{J} \wedge DK^{J}{}_{I} - \omega^{I}{}_{J}\wedge \omega^{J}{}_{K}\wedge K^{K}{}_{I} + K^{I}{}_{J}\wedge R^{J}{}_{I} \nonumber\\
&\hspace{3mm} -\! \frac{2}{l} \big( DK^{I}{}_{4} \wedge Q_I + K^{I}{}_{4} \wedge DQ_I \big) \!-\! \frac{2}{l^2} Q^I \wedge Q_J \wedge K^J{}_{I} \bigg]\!. \label{Iden2-II}
\end{align}
Next, equating the terms of~\eqref{Iden2-II} with the same powers of $1/l$, from the zero, first, and second powers of $1/l$ we obtain, respectively, the following identities:
\begin{eqnarray}
2 R^{I}{}_{J}\wedge DK^{J}{}_{I} &=& d \big( \omega^{I}{}_{J} \wedge DK^{J}{}_{I} - \omega^{I}{}_{J}\wedge \omega^{J}{}_{K}\wedge K^{K}{}_{I} \nonumber\\
&&+ K^{I}{}_{J}\wedge R^{J}{}_{I} \big), \label{Iden-II-1}
\end{eqnarray}
\begin{eqnarray}
&&2DK^{I}{}_{4} \wedge DQ_I - K^I{}_4 \wedge Q^J \wedge R_{IJ} - Q^I \wedge K^J{}_4 \wedge R_{IJ} \nonumber\\
&&=d \left( DK^{I}{}_{4} \wedge Q_I + K^{I}{}_{4} \wedge DQ_I \right) ,\label{Iden-II-2}
\end{eqnarray}
\begin{eqnarray}
&&Q^I \wedge Q_J \wedge DK^J{}_I + 2 DQ^I \wedge Q_J \wedge K^J{}_I \nonumber\\
&&= d \left( Q^I \wedge Q_J \wedge K^J{}_{I} \right). \label{Iden-II-3}
\end{eqnarray}
The identity~\eqref{Iden-II-1} involves the 4-form
\begin{equation}
R^{I}{}_{J}\wedge DK^{J}{}_{I},
\end{equation}
which was first reported in~\cite{Montesinos2021}, whereas the identities~\eqref{Iden-II-2} and~\eqref{Iden-II-3} show, respectively, that the integrals of the 4-forms
\begin{eqnarray}
I_2 &\equiv& 2DK^{I}{}_{4} \wedge DQ_I - K^I{}_4 \wedge Q^J \wedge R_{IJ} \nonumber \\
&& - Q^I \wedge K^J{}_4 \wedge R_{IJ}, \label{NewInv2}
\end{eqnarray}
\begin{equation}
I_3\equiv Q^I \wedge Q_J \wedge DK^J{}_I + 2 DQ^I \wedge Q_J \wedge K^J{}_I, \label{NewInv3}
\end{equation}
are topological invariants. Notice that the 4-form $I_2$ is a generalization of the 4-form~\eqref{NewInv1}, and then, in turn, is a generalization of the Nieh-Yan topological form~\eqref{NYinv}. In fact, if we set $K^{I}{}_{4}\equiv Q^I$, it is straightforward to see that $I_2=2 I_1$. Clearly, there are several plausible choices for $K^{I}{}_{4}$ and $Q^I$. However here, for illustrative purposes, we limit ourselves to two particular cases of~\eqref{NewInv2}:
\begin{enumerate}[(i)]
\item Setting $K^{I}{}_{4}\equiv e^I$ and $Q^I\equiv \star(R^I{}_J \wedge e^J)$, the 4-form $I_2$ takes the form
\begin{eqnarray}
I_2 &=& 2De^I \wedge D[\star(R_{IJ} \wedge e^J)] - e^I \wedge \star(R^J{}_K \wedge e^K) \wedge R_{IJ} \nonumber \\
&&- \star(R^I{}_K \wedge e^K) \wedge e^J \wedge R_{IJ}. \label{ExI2}
\end{eqnarray}
Notice that this 4-form vanishes if there is no torsion, $De^I=0$.
\item Setting $K^{I}{}_{4}\equiv e^I$ and $Q^I\equiv \star[(\star R^I{}_J) \wedge e^J]=-\sigma \mathfrak{R}^I{}_J e^J$, the 4-form $I_2$ can be written as
\begin{align}
I_2 =& -\sigma \big( 2 \mathfrak{R}_{IJ} De^I \wedge De^J + 2 D\mathfrak{R}_{IJ} \wedge De^I \wedge e^J \nonumber \\
& - \mathfrak{R}^J{}_K e^I \wedge e^K \wedge R_{IJ} - \mathfrak{R}^I{}_K e^K \wedge e^J \wedge R_{IJ} \big).
\end{align}
\end{enumerate}
To illustrate~\eqref{NewInv3} we also have a wide variety of choices for $Q^I$ and $K^I{}_J$. In particular, it is interesting to note that taking $Q^I\equiv e^I$ and $K^I{}_J\equiv\omega^I{}_J - \Gamma^I{}_J$ (contorsion), the 4-form~\eqref{NewInv3} reduces to the Nieh-Yan form~\eqref{NYinv}, $I_3=N$. This can be checked directly by using $De^I=K^I{}_J \wedge e^J$ and $D^2e^I=R^I{}_J\wedge e^J$. Among the other possible alternatives for $Q^I$ and $K^I{}_J$, we consider here the following two cases:
\begin{enumerate}[(i)]
\item Setting $Q^I\equiv \star[(\star R^I{}_J) \wedge e^J]$ and $K^I{}_J\equiv\omega^I{}_J - \Gamma^I{}_J$, the 4-form $I_3$ reads
\begin{eqnarray}
I_3 &=& - \mathfrak{R}^I{}_K \mathfrak{R}^J{}_L e^K \wedge e^L \wedge DK_{IJ} \nonumber\\
&& - D(\mathfrak{R}^I{}_K e^K) \wedge \mathfrak{R}^J{}_L e^L \wedge K_{IJ} \nonumber\\
&=&-\mathfrak{R}^I{}_K \mathfrak{R}^J{}_L \big(e^K \wedge e^L \wedge DK_{IJ}+2De^K \wedge e^L \wedge K_{IJ}\big) \nonumber\\
&&-2 \mathfrak{R}^J{}_L D\mathfrak{R}^I{}_K e^K \wedge e^L \wedge K_{IJ}.
\end{eqnarray}
Since $De^I=K^I{}_J \wedge e^J$, this 4-form vanishes in the absence of torsion, $De^I=0$.
\item Setting $Q^I\equiv \star(R^I{}_J \wedge e^J)$ and $K^I{}_J\equiv T_{K}{}^I{}_J e^K$ where $De^I\equiv (1/2) T^{I}{}_{JK} e^J\wedge e^K$ with $T^I{}_{JK}=-T^I{}_{KJ}$, the 4-form $I_3$ takes the form
\begin{align}
I_3 =& \star(R^I{}_K \wedge e^K) \wedge \star(R_{JL} \wedge e^L) \wedge D\big(T_{K}{}^J{}_I e^K\big) \nonumber\\
&+ 2 D\big[\star(R^I{}_K \wedge e^K)\big] \wedge \star(R_{JL} \wedge e^L) \wedge \big(T_{K}{}^J{}_I e^K\big).
\end{align}
Clearly, this 4-form vanishes when torsion vanishes, $De^I=0$.
\end{enumerate}
\subsection{Identities coming from~\eqref{Iden-III}}
Taking into account~\eqref{conn1} and~\eqref{curv1}, the identity~\eqref{Iden-III} leads to
\begin{align}
&DK^{I}{}_{J}\wedge DK^{J}{}_{I} + 2 K^{I}{}_{J}\wedge K^{J}{}_{K}\wedge R^{K}{}_{I} + 2 DK^I{}_4 \wedge DK^4{}_I \nonumber\\
&\hspace{3mm}+ 2 K^I{}_4 \wedge K^4{}_J \wedge R^{J}{}_{I} - \frac{4}{l} \big(K^I{}_4 \wedge Q_J \wedge DK^{J}{}_{I} \nonumber\\
&\hspace{3mm}+ DK^I{}_4 \wedge Q_J \wedge K^{J}{}_{I}- K^I{}_4 \wedge DQ_J \wedge K^{J}{}_{I} \big) \nonumber\\
&= d\bigg( K^{I}{}_{J}\wedge DK^{J}{}_{I} + 2 K^I{}_4 \wedge DK^4{}_I \nonumber\\
&\hspace{3mm}-\frac{4}{l} K^I{}_4 \wedge Q_J \wedge K^{J}{}_{I} \bigg). \label{Iden2-III}
\end{align}
Matching the terms of~\eqref{Iden2-III} with identical powers of $1/l$, from the zero and first powers of $1/l$ we get, respectively, the following identities:
\begin{align}
&DK^{I}{}_{J}\wedge DK^{J}{}_{I} \! + \! 2 K^{I}{}_{J}\wedge K^{J}{}_{K}\wedge R^{K}{}_{I} \! + \! 2 DK^I{}_4 \wedge DK^4{}_I \nonumber\\
&\hspace{3mm}+ 2 K^I{}_4 \wedge K^4{}_J \wedge R^{J}{}_{I} \nonumber\\
&= d\big( K^{I}{}_{J}\wedge DK^{J}{}_{I} + 2 K^I{}_4 \wedge DK^4{}_I \big), \label{Iden-III-1}
\end{align}
\begin{eqnarray}
&&K^I{}_4 \wedge Q_J \wedge DK^{J}{}_{I} + DK^I{}_4 \wedge Q_J \wedge K^{J}{}_{I} \nonumber\\
&&- K^I{}_4 \wedge DQ_J \wedge K^{J}{}_{I} = d\left( K^I{}_4 \wedge Q_J \wedge K^{J}{}_{I} \right). \label{Iden-III-2}
\end{eqnarray}
The identities~\eqref{Iden-III-1} and~\eqref{Iden-III-2} show, respectively, that the integrals of the 4-forms
\begin{eqnarray}
I_4&\equiv& DK^{I}{}_{J}\wedge DK^{J}{}_{I} + 2 K^{I}{}_{J}\wedge K^{J}{}_{K}\wedge R^{K}{}_{I} \nonumber\\
&&+ 2 DK^I{}_4 \wedge DK^4{}_I + 2 K^I{}_4 \wedge K^4{}_J \wedge R^{J}{}_{I}, \label{NewInv4}\\
I_5&\equiv& K^I{}_4 \wedge Q_J \wedge DK^{J}{}_{I} + DK^I{}_4 \wedge Q_J \wedge K^{J}{}_{I} \nonumber\\
&&- K^I{}_4 \wedge DQ_J \wedge K^{J}{}_{I} , \label{NewInv5}
\end{eqnarray}
are topological invariants. Notice that $I_4$ is a linear combination of the 4-form
\begin{equation}
DK^{I}{}_{J}\wedge DK^{J}{}_{I} + 2 K^{I}{}_{J}\wedge K^{J}{}_{K}\wedge R^{K}{}_{I},
\end{equation}
which was reported in~\cite{Montesinos2021} and the 4-form~\eqref{NewInv1} with $K^I{}_4=Q^I$. On the other hand, the 4-form $I_5$ is readily recognized as a generalization of the 4-form~\eqref{NewInv3}. Indeed, by taking $K^I{}_4=Q^I$, it is direct to see that~\eqref{NewInv5} reduces to~\eqref{NewInv3}, $I_5=I_3$. Two particular cases of~\eqref{NewInv5} are
\begin{enumerate}[(i)]
\item Setting $K^{I}{}_{4}\equiv e^I$, $Q^I\equiv \star(R^I{}_J \wedge e^J)$ and $K^I{}_J\equiv\omega^I{}_J - \Gamma^I{}_J$, we find that $I_5=(1/2)I_2$ with $I_2$ given by~\eqref{ExI2}. This can be verified by using $De^I=K^I{}_J \wedge e^J$ and $D^2e^I=R^I{}_J\wedge e^J$.
\item Setting $K^{I}{}_{4}\equiv \star[(\star R^I{}_J) \wedge e^J]$, $Q^I\equiv \star(R^I{}_J \wedge e^J)$ and $K^I{}_J\equiv T_{K}{}^I{}_J e^K$, the 4-form $I_5$ reads
\begin{eqnarray}
I_5&=& -\sigma \big\{ \mathfrak{R}^I{}_{K} e^K \wedge \star(R_{JL} \wedge e^L) \wedge D(T_{M}{}^J{}_I e^M) \nonumber\\
&&+ D\big(\mathfrak{R}^I{}_{K} e^K\big) \wedge \star(R_{JL} \wedge e^L) \wedge T_{M}{}^J{}_I e^M \nonumber\\
&&- \mathfrak{R}^I{}_{K} e^K \wedge D[\star(R_{JL} \wedge e^L)] \wedge T_{M}{}^J{}_I e^M \big\}.
\end{eqnarray}
It is direct to see that this 4-form vanish if there is no torsion, $De^I=0$.
\end{enumerate}
\subsection{Identity coming from~\eqref{Iden-IV}}
Using~\eqref{conn1} and~\eqref{curv1}, the identity~\eqref{Iden-IV} can be written as
\begin{align}
&K^{I}{}_{J}\wedge K^{J}{}_{K}\wedge DK^{K}{}_{I} + DK_{IJ} \wedge K^I{}_4 \wedge K^{J}{}_4 \nonumber\\
& \hspace{3mm} + 2 K_{IJ} \wedge K^I{}_4 \wedge DK^J{}_4 \nonumber\\
&= d\left( \frac{1}{3} K^{I}{}_{J}\wedge K^{J}{}_{K}\wedge K^{K}{}_{I} + K_{IJ} \wedge K^I{}_4 \wedge K^J{}_4 \right). \label{Iden2-IV}
\end{align}
Notice that~\eqref{Iden2-IV} does not depend on the parameter $l$, and hence no additional identities emerge. Then, directly from~\eqref{Iden2-IV} it follows that the integral of the 4-form
\begin{eqnarray}
I_6 &\equiv& K^{I}{}_{J}\wedge K^{J}{}_{K}\wedge DK^{K}{}_{I} + DK_{IJ} \wedge K^I{}_4 \wedge K^{J}{}_4 \nonumber\\
&& + 2 K_{IJ} \wedge K^I{}_4 \wedge DK^J{}_4
\end{eqnarray}
is a topological invariant. The 4-form $I_6$ turns out to be a linear combination of the 4-form
\begin{equation}
K^{I}{}_{J}\wedge K^{J}{}_{K}\wedge DK^{K}{}_{I},
\end{equation}
which was reported in~\cite{Montesinos2021} and the 4-form~\eqref{NewInv3} with $Q^I=K^I{}_4$; therefore, no new 4-form whose integral leads to new topological invariant arises in this case.
\vspace{3mm}
\section{Conclusions}
In this paper we have reported generalizations of the Nieh-Yan topological invariant in four dimensions. We have followed a systematic method along the lines of Refs.~\cite{Nieh-Yan1982,Zanelli1997,Nieh2018,Montesinos2021}, obtaining a family of 4-forms resulting from cases A, B, and C of Sec. \ref{sec-general}. In each case, explicit expressions for some of the resulting invariants are given and we showed how the Nieh-Yan form is obtained from particular choices of the 1-forms $Q^I$, $K^{I}{}_{4}$, and $K^I{}_J$. We have considered all the possible cases that emerge from the method, namely cases A, B, C, and D of Sec. \ref{sec-general}; this despite the fact that case~D turned out to be very simple and does not give rise to new topological invariants.
Regarding the invariant~\eqref{exampleA1}, it is worth noting that $f$ can be a functional depending on $e^I$ and $\omega^I{}_J$ through $e^I$ itself, $R^I{}_J$, and $De^I$. In this way, for instance, $f$ could be a function of the scalar curvature $ \mathfrak{R}= \mathfrak{R}^I{}_I $, i.e., $f=f(\mathfrak{R})$, or it could be a function of $\tau_I \tau^I$ where $\tau_I=T^J{}_{IJ}$, i.e., $f=f(\tau_I \tau^I)$. Additionally, in those cases where a scalar field $\phi$ is coupled to gravity, $f$ could be a function of the scalar field, i.e., $f=f(\phi)$. Furthermore, it would be also interesting to compute the topological invariants presented here in theories in which torsion emerges, as is the case of general relativity coupled to spin matter fields~\cite{Freidel2005,Perez-Rovelli-2006,Mercuri2006,Alexandrov_2008,Romero_2021} and $f(\mathfrak{R})$ gravity with torsion~\cite{Rubilar_1998,Capozziello2010,MontesinosfR_2020}.
Lastly, another interesting aspect to investigate is the addition of these topological invariants to actions for general relativity such as Palatini or Holst~\cite{Holst9605} actions and also to actions of alternative theories of gravity, and to study the implications of this fact both at Lagrangian and Hamiltonian levels. In fact, at the Hamiltonian level, it would be expected that these topological invariants would induce canonical transformations on the phase space~\cite{Perez2009}, which would introduce new phase space variables that could have geometric or physical relevance.
\begin{acknowledgments}
We thank Mariano Celada and Jorge Romero for carefully reading the manuscript and for their fruitful comments. This work was partially supported by Fondo SEP-Cinvestav and by Consejo Nacional de Ciencia y Tecnolog\'{i}a (CONACyT), M\'{e}xico, Grant No.~A1-S-7701.
\end{acknowledgments}
|
\section{Introduction}
Given a vertex operator algebra $V$, Zhu introduced in \cite{Z}, an associative algebra which was used to prove modular invariance of graded characters for $V$-modules in the case when $V$ is $C_2$-cofinite and rational. In \cite{DLM}, Dong, Li, and Mason generalized these algebras, defining for every vertex operator algebra $V$, a family of associative algebras denoted $A_n(V)$, for $n\in \mathbb{N}$. The $n=0$ algebra in this family is the original (level zero) associative algebra that Zhu introduced, and the algebras for $n>0$ are now known as ``higher level Zhu algebras". These Zhu algebras are powerful tools in the study of both the representation theory of vertex operator algebras and the number-theoretic properties of the graded traces and ``pseudo-traces" of these modules. However a major obstacle in realizing the full potential of these Zhu algebras is that they are extremely difficult to calculate. It is thus an important problem to develop methods for and insight into the construction of these algebras.
In this paper, we prove general results regarding the generators and relations for the level $n$ Zhu algebra, $n \in \mathbb{N}$, for a vertex operator algebra $V$, and show how these results apply to the level $n$ Zhu algebras for the Heisenberg and Virasoro vertex operator algebras, as well as to the conformal vector or Heisenberg-like vectors in any vertex operator algebra. We also give some clarifications on the necessity of a certain condition in the definition of the higher level Zhu algebras, i.e. for $n >0$.
As was mentioned above, Zhu algebras are important tools in studying the representation theory of a vertex operator algebra. In \cite{FZ}, Frenkel and Zhu proved that there is a one-to-one correspondence between isomorphism classes of irreducible representations of a vertex operator algebra $V$ and isomorphism classes of irreducible representations of its (level zero) Zhu algebra $A_0(V)$. However if $V$ admits indecomposable non irreducible modules (i.e., if $V$ is irrational), then $A_0(V)$ will not in general be sufficient for the correspondence between isomorphism classes of indecomposable modules for $V$ and for $A_0(V)$ to hold. In \cite{DLM}, Dong, Li, and Mason define a functor, denoted $\Omega_n$, from the category of admissable $V$-modules, to the category of $A_n(V)$ modules and another functor, denoted $L_n$ from the category of $A_n(V)$-modules to the category of admissable $V$-modules and (under an added necessary condition provided in \cite{BVY}) prove that these functors realize a bijection between the isomorphism classes of irreducibles for $V$ and for $A_n(V)$. In \cite{BVY}, the second author along with Vander Werf and Yang prove several results in the theory of higher level Zhu algebras which help clarify some statements of \cite{DLM} and further illuminate the relationship between $V$-modules and $A_n(V)$-modules in the non semi-simple setting of indecomposable non irreducible modules for irrational vertex operator algebras. In particular, in \cite{BVY} the nature of the level $n$ Zhu algebra in relation to the level $n-1$ Zhu algebra was shown to determine whether new information about certain types of modules can be gleaned from the level $n$ Zhu algebra not already detected by the level $n-1$ Zhu algebra, and the nature of that information.
The second major application of higher level Zhu algebras, is to construct the graded pseudo-traces and analyze their number-theoretic properties. An analogue of Zhu's modular invariance for graded characters (i.e., graded traces) for modules for a rational $C_2$-cofinite vertex operator algebra was provided by Miyamoto in \cite{Miyamoto2004}, where the notion of a graded ``pseudo-trace" was defined for irrational $C_2$-cofinite vertex operator algebras. While the graded characters of modules for rational vertex operator algebras encode information about the $L(0)$ eigenspaces of the modules, these graded pseudo-traces give information about the generalized eigenspaces. Where the original (level $0$) Zhu algebra plays an important role in Zhu's proof of modular invariance in the rational setting, the higher level Zhu algebras are central to Miyamoto's construction.
In \cite{AN}, Arike and Nagatomo studied graded pseudo-traces for the $\mathbb{Z}_2$-orbifold of symplectic fermions. While their construction notably does not use Zhu algebras, it is not known if their graded pseudo-traces coincide with those of Miyamoto. To the authors' knowledge, no examples of graded pseudo-traces have been explicitly determined using Miyamoto's construction, where the higher level Zhu algebras are necessary. Examples of these pseudo-traces using Miyamoto's construction will be provided by the authors and their collaborators in subsequent publications.
A major obstacle in utilizing higher level Zhu algebras in the study of vertex operator algebras and related questions of modular invariance, is that determining their structure is incredibly difficult. In \cite{BVY-Heisenberg} and \cite{BVY-Virasoro}, the second author along with Vander Werf and Yang gave the first explicit constructions of higher level Zhu algebras, determining the structure of $A_1(V)$ for $V$ the rank one Heisenberg vertex operator algebra and the Virasoro vertex operator algebras, respectively. They also proved that the level one Zhu algebra for the Virasoro vertex operator algebra can be used to induce new modules not induced by the level zero Zhu algebra. Later, in \cite{AB-details-of-Heisenberg}, the present authors determined the structure of the level two Zhu algebra for the Heisenberg vertex operator algebra; see Theorem \ref{A_2-theorem} below.
Notably in the constructions provided in \cite{BVY-Heisenberg}, \cite{BVY-Virasoro}, and \cite{AB-details-of-Heisenberg}, only very minimal facts about modules for the relevant vertex operator algebras are used. The purpose of this paper is to expand on the insight gained from these papers to provide tools for determining the structure of higher level Zhu algebras. In fact, the results of this paper require no knowledge of the modules of any of the vertex operator algebras to which our results apply. This is particularly important, as one would like to use higher level Zhu algebras to study the modules of vertex operator algebras for which very little information is known, e.g. irrational $C_2$-cofinite vertex operator algebras which are central to logarithmic conformal field theory but for which very few examples are currently known.
In this paper, we present results on generators and relations for higher level Zhu algebras that depend only on the definition of higher level Zhu algebras and, for some results, some relatively mild assumptions about the vertex operator algebra at hand. The main results of this paper are presented in Sections \ref{general-n-section} and \ref{definition-clarification-section}. In the first subsection of Section \ref{general-n-section}, we prove a recursion formula for elements in $A_n(V)$ which applies to any vertex operator algebra $V$, and provides relations which are key to calculating $A_n(V)$ explicitly, in particular in determining minimal generating sets. In the second subsection of Section \ref{general-n-section}, we prove further results on generators and relations for $A_n(V)$, which apply to a vertex operator algebra $V$ which has generators with certain nice properties. We give a general formula for multiplication in higher level Zhu algebras for certain elements of the vertex operator algebra. In particular, we prove that terms in multiplication formulas in the level $n$ Zhu algebras for these vertex operator algebra elements are indexed by partitions of $2n$.
In Section \ref{definition-clarification-section}, we provide an important clarification about the construction of higher level Zhu algebras and discuss the implications of this clarification for future constructions of these algebras. We note that in the definition of the level $n$ Zhu algebra $A_n(V)= V/O_n(V)$ for $n \in\mathbb{Z}_+$ in some instances, cf. \cite{Miyamoto2004}, the ideal $O_n(V)$ is defined to be span$\{ u \circ_n v \; | \; u,v \in V\}$ and it is implicitly assumed that this ideal contains the elements of the form $(L(-1) + L(0))v$ for $v\in V$. This assumption is likely due to the fact that in the original Zhu algebra, i.e., the $n=0$ Zhu algebra, the vectors $(L(-1)+L(0))v\in \mathrm{span} \{ u \circ_n v \; | \; u,v \in V\}$ for all $v\in V$, see Remark \ref{first-remark}. In this paper, we prove that for Zhu algebras of level $n>0$, it is a necessary extra condition to include the vectors $(L(-1) + L(0))v$ for $v\in V$ in the definition of $O_n(V)$, as they are not in general in span$\{ u \circ_n v \; | \; u,v \in V\}$. In addition, we give certain general cases for when these vectors will not be in this spanning set and thus give new relations in $A_n(V)$.
In Section 5, we highlight some aspects of Zhu algebras that provide techniques for calculating higher level Zhu algebras. These techniques include utilizing the structure of the level $n-1$ Zhu algebra and the $V$-modules induced at level $n-1$ to help determine the relations for the level $n$ Zhu algebra.
In Section 6, we give applications of our results for the vertex operator algebras associated to the Heisenberg and Virasoro algebras, respectively, recall the cases for which the higher level Zhu algebras have been explicitly calculated for these vertex operator algebras, and present a conjecture from \cite{AB-details-of-Heisenberg} about the structure for all higher level Zhu algebras for the Heisenberg vertex operator algebra.
{\bf Acknowledgments: } The authors thank the Simons Foundation for their support. The first author was a recipient of an AMS-Simons Foundation Travel Grant, and the second author was the recipient of a Simons Foundation Collaboration Grant 282095.
{\bf Notation:} Throughout this paper, $\mathbb{N}$ denotes the nonnegative integers and $\mathbb{Z}_+$ denotes the positive integers.
\section{The algebras $A_n(V)$, and the functors $\Omega_n$ and $L_n$}\label{zhu-algebra-definition-section}
In this section, letting $V$ be a vertex operator algebra, we recall the definition and some properties of the algebras $A_n(V)$ for $n \in \mathbb{N}$, first introduced in \cite{Z} for $n = 0$, and then generalized to $n >0$ in \cite{DLM}. We then recall the functors $\Omega_n$ and $L_n$ defined in \cite{DLM} and some results from \cite{BVY}.
For $n \in \mathbb{N}$, let $O_n(V)$ be the subspace of $V$ spanned by elements of the form
\begin{equation}\label{define-circ}
u \circ_n v =
\mbox{\rm Res}_x \frac{(1 + x)^{\mathrm{wt}\, u + n}Y(u, x)v}{x^{2n+2}}
\end{equation}
for all homogeneous $u \in V$ and for all $v \in V$, and by elements of the form $(L(-1) + L(0))v$ for all $v \in V$. The vector space $A_n(V)$ is defined to be the quotient space $V/O_n(V)$.
That is, using the notation introduced in \cite{BVY}, we define
\begin{equation}\label{define-O-subsets}
O^L (V) = \{ (L(-1) + L(0))v \, | \, v \in V \} \quad
\mathrm{and}
\quad O_n^\circ(V) = \mathrm{span} \{ u \circ_n v \; | \; u,v \in V\} ,
\end{equation}
then
\begin{equation}\label{define-O}
O_n(V) = O^L (V) + O_n^\circ(V) .
\end{equation}
We will sometimes refer to $\circ_n$ as the ``circle product".
\begin{rem}\label{first-remark}{\em
For $n=0$, since $v \circ_0 \mathbf{1} = v_{-2} \mathbf{1} + (\mathrm{wt} \, v) v = L(-1) v + L(0) v$, it
follows that $(L(-1) + L(0)) v \in O^\circ_0(V)$ for all $v \in V$ and thus $O_0(V) = O_0^\circ(V)$.
However, we show in Section \ref{definition-clarification-section} that for $n > 0$, in
general, $O^L (V) \not\subset O_n^\circ(V)$. Thus for $A_n(V)$ to have a well-defined action on
modules via zero modes, it is necessary to not just define $O_n(V)$ to consist of $O_n^\circ(V)$ as is
sometimes done, but to also include $O^L(V)$ as part of the definition; cf. \cite{DLM},
\cite{Miyamoto2004}. This is simply due to the fact that from the $L(-1)$-derivative property, the zero
mode of $L(0)v$ which is $(\mathrm{wt} \, v)v_{\mathrm{wt} \, v - 1}$ is equal to the negative of the zero mode of $L(-1)v$. We show in Section 4 that in general $O^L (V) \not\subset O_n^\circ(V)$, by showing that, for instance, any time $V$ is strongly generated (as defined in Section \ref{generators-section}) by an element $u$ such that $1 \leq \mathrm{wt} \, u \leq n$, then $(L(-1) + L(0) )u \notin O_n^\circ(V)$. More generally, for any $V$ with $u \in V$ of positive weight or of weight zero such that $u_{-2} {\bf 1} \neq 0$, and such that the vertex subalgebra generated by $u$ is strongly generated, then one can have $(L(-1) + L(0) )u \notin O_n^\circ(V)$ if $n >0$. For instance, this holds for the Heisenberg vertex operator algebra generated by a weight one element $u = \alpha(-1) \mathbf{1}$, or for such weight one elements $u \in V$ where there is not linear dependence between $(L(-1) + L(0))u$ and circle products of vectors in $V$ not generated by $u$. This also holds for the vertex operator subalgebra of $V$ generated by the conformal element if $n>1$. And finally we note that our result also pertains to the case when $V$ has a nontrivial vacuum (weight zero) space containing any vectors $u$ such that $u_{-2} {\bf 1} \neq 0$. Thus there are common settings in which a vertex operator algebra $V$ has the property that $O^L (V) \not\subset O_n^\circ(V)$. More importantly, as we point out in Remark \ref{OL-remark} below, the presence of these $(L(-1) + L(0)) v \in
O_n(V)$ is used heavily in the constructions of the higher level Zhu algebras in \cite{BVY-Heisenberg}, \cite{BVY-Virasoro}, and \cite{AB-details-of-Heisenberg}. }
\end{rem}
We define the following multiplication on $V$
\[
u *_n v = \sum_{m=0}^n(-1)^m\binom{m+n}{n}\mbox{\rm Res}_x \frac{(1 + x)^{\mathrm{wt}\, u + n}Y(u, x)v}{x^{n+m+1}},
\]
for $v \in V$ and homogeneous $u \in V$, and for general $u \in V$, $*_n$ is
defined by linearity. It is shown in \cite{DLM} that with this multiplication, the subspace $O_n(V)$ of $V$ is a two-sided ideal of $V$, and $A_n(V)$ is an associative algebra, called the {\it level $n$ Zhu algebra}.
\begin{rem} {\em Due to the truncation condition on $V$, the expressions $u \circ_n v$ and $u *_n v$ in the definition of $A_n(V)$ are indeed well-defined elements of $V$, even if $\mathrm{wt} \, u + n <0$, i.e. they are finite sums of elements in $V$. Thus, it is unnecessary (as is often done in the literature) to assume that $V$ is positively graded. In fact, notably many of our results in this paper apply to elements in $V$ of negative weight.}
\end{rem}
Next we prove a lemma stating that certain elements in $V$, namely those of the form (\ref{more-elements}) given below, lie in $O_n(V)$, and in fact, more precisely in $O_n^\circ(V)$. In \cite{Z}, Zhu proved that such elements lie in $O_n(V)$ in the case when $n = 0$, and his proof extends easily to the $n>0$ case. In \cite{DLM} this result without proof was used to show that there is a surjective homomorphism from $A_n(V)$ onto $A_{n-1}(V)$ for any $n \in \mathbb{Z}_+$. Here we provide the proof so as to point out a refinement of the result given in \cite{DLM}, namely that these elements lie not just in $O_n(V)$, but in $O_n^\circ(V)$. We will use this more precise statement later in Section \ref{general-n-section}, in particular to prove the useful recursion formula in Proposition \ref{recursion-n}.
\begin{lem}\label{Zhu-lemma}
For every homogeneous element $u \in V$ and for $m, k \in \mathbb{N}$ with $m \geq k \geq 0$, elements of the form
\begin{equation}\label{more-elements}
\mbox{\rm Res}_x \frac{(1 + x)^{\mathrm{wt}\, u + n+ k}Y(u, x)v}{x^{m+2n+2}}
\end{equation}
lie not only in $O_n(V)$, but in $O_n^\circ(V)$.
\end{lem}
\begin{proof}
The proof follows from the definition of $O_n^\circ(V)$ and the $L(-1)$-derivative property for $V$. First we note that $m\geq k$ implies that $m-j \geq 0$ for $j =0,\dots,k$, and thus since
\begin{equation}
\mbox{\rm Res}_x \frac{(1 + x)^{\mathrm{wt}\, u + n+ k}Y(u, x)v}{x^{m+2n+2}} = \sum_{j=0}^k
\mbox{\rm Res}_x \binom{k}{j} \frac{(1 + x)^{\mathrm{wt}\, u + n}Y(u, x)v}{x^{m+2n+2-j}}
\end{equation}
we need only prove the statement for the case when $k = 0$ and $m\geq 0$.
We then proceed by induction on $m$. For $k= 0$ and $m = 0$, the statement follows immediately from the definition of $\circ_n$. Assume the statement holds for $k = 0$ and $m= j$. Then for $m = j+1$, by the inductive assumption, we have that
\begin{equation}\label{inductive-assumption}
\mbox{\rm Res}_x \frac{(1 + x)^{ \mathrm{wt}\, w + n}Y(w, x)v}{x^{j+2n+2}} \in O_n^\circ(V) \quad \mbox{ for both $w = u$ and $w = L(-1)u$,}
\end{equation}
where we note that $\mathrm{wt} \, L(-1) u = \mathrm{wt} \, u + 1$. Then by the definition of $\mbox{\rm Res}_x$, and the $L(-1)$-derivative property in $V$ , we have
\begin{eqnarray*}
0 &=& \mbox{\rm Res}_x \frac{d}{dx} \left( \frac{(1 + x)^{\mathrm{wt}\, u + n + 1}Y(u, x)v}{x^{j+2n+2}}\right) \\
&=& \mbox{\rm Res}_x (\mathrm{wt} \, u + n + 1) \frac{(1+x)^{ \mathrm{wt} \, u + n } Y(u, x)v}{x^{j+2n+2}} - \mbox{\rm Res}_x
(j + 2n + 2) \frac{(1 + x)^{\mathrm{wt}\, u + n + 1} Y(u, x)v}{x^{j+2n+3}} \\
& & \quad + \, \mbox{\rm Res}_x \frac{(1 + x)^{\mathrm{wt}\, u + n + 1}\frac{d}{dx}Y(u, x)v}{x^{j+2n+2}} \\
&=& \mbox{\rm Res}_x (\mathrm{wt} \, u + n + 1 ) \frac{(1+x)^{ \mathrm{wt} \, u + n } Y(u, x)v}{x^{j+2n+2}} - (j + 2n + 2) \left( \mbox{\rm Res}_x
\frac{(1 + x)^{\mathrm{wt}\, u + n} Y(u, x)v}{x^{(j+1)+2n+2}} \right. \\
& & \quad \left. + \, \mbox{\rm Res}_x
\frac{(1 + x)^{\mathrm{wt}\, u + n} Y(u, x)v}{x^{j+2n+2}} \right)
+ \mbox{\rm Res}_x \frac{(1 + x)^{\mathrm{wt}\, u + n + 1 }Y(L(-1) u, x)v}{x^{j+2n+2}} \\
&=& (\mathrm{wt} \, u -j-n-1 ) \mbox{\rm Res}_x \frac{(1+x)^{ \mathrm{wt} \, u + n } Y(u, x)v}{x^{j+2n+2}} - (j + 2n + 2) \mbox{\rm Res}_x
\frac{(1 + x)^{\mathrm{wt}\, u + n} Y(u, x)v}{x^{(j+1)+2n+2}} \\
& & \quad + \, \mbox{\rm Res}_x \frac{(1 + x)^{\mathrm{wt}\, u + n + 1}Y(L(-1) u, x)v}{x^{j+2n+2}} .
\end{eqnarray*}
By Eqn. \eqref{inductive-assumption}, the first and last terms of the right hand side are in $O^\circ_n(V)$ , and thus
\[ (j + 2n + 2) \mbox{\rm Res}_x \frac{(1 + x)^{\mathrm{wt}\, u + n} Y(u, x)v}{x^{(j+1)+2n+2}} \in O_n^\circ(V).\]
Since $j + 2n + 2 >0$, the result holds.
\end{proof}
Lemma \ref{Zhu-lemma} implies that $O^\circ_n(V) \subset O^\circ_{n-1}(V)$, and thus $O_n(V) \subset O_{n-1}(V)$. In fact, from Proposition 2.4 in \cite{DLM}, we have that the map
\begin{eqnarray}\label{surjection}
A_n(V) & \longrightarrow & A_{n-1}(V) \\
v + O_n(V) & \mapsto & v + O_{n-1}(V) \nonumber
\end{eqnarray}
is a surjective algebra homomorphism.
Next we recall the definitions of various $V$-module structures and then briefly introduce the functors that associate modules for $A_n(V)$ with certain types of modules for $V$. We assume the reader is familiar with the notion of weak $V$-module for a vertex operator algebra $V$ (cf. \cite{LL}).
\begin{defn}\label{N-gradable-definition}
{\em An {\it $\mathbb{N}$-gradable weak $V$-module} (also often called an {\it admissible $V$-module} as in \cite{DLM}) $W$ for a vertex operator algebra $V$ is a weak $V$-module that is $\mathbb{N}$-gradable, $W = \coprod_{k \in \mathbb{N}} W(k)$, with $v_m W(k) \subset W(k + \mathrm{wt} v - m -1)$ for homogeneous $v \in V$, $m \in \mathbb{Z}$ and $k \in \mathbb{N}$, and without loss of generality, we can and do assume $W(0) \neq 0$, unless otherwise specified. We say elements of $W(k)$ have {\it degree} $k \in \mathbb{N}$.
An {\it $\mathbb{N}$-gradable generalized weak $V$-module} $W$ is an $\mathbb{N}$-gradable weak $V$-module that admits a decomposition into generalized eigenspaces via the spectrum of $L(0) = \omega_1$ as follows: $W=\coprod_{\lambda \in{\mathbb{C}}}W_\lambda$ where $W_{\lambda}=\{w\in W \, | \, (L(0) - \lambda \, id_W)^j w= 0 \ \mbox{for some $j \in \mathbb{Z}_+$}\}$, and in addition, $W_{n +\lambda}=0$ for fixed $\lambda$ and for all sufficiently small integers $n$. We say elements of $W_\lambda$ have {\it weight} $\lambda \in \mathbb{C}$.
A {\it generalized $V$-module} $W$ is an $\mathbb{N}$-gradable generalized weak $V$-module where $\dim W_{\lambda}$ is finite for each $\lambda \in \mathbb{C}$.
An {\it (ordinary) $V$-module} is a generalized $V$-module such that the generalized eigenspaces $W_{\lambda}$ are in fact eigenspaces, i.e., $W_{\lambda}=\{w\in W \, | \, L(0) w=\lambda w\}$.}
\end{defn}
We will often omit the term ``weak" when referring to $\mathbb{N}$-gradable weak and $\mathbb{N}$-gradable generalized weak $V$-modules.
The term {\it logarithmic} is also often used in the literature to refer to $\mathbb{N}$-gradable weak generalized modules or generalized modules.
Let $W$ be an $\mathbb{N}$-gradable $V$-module, and let
\begin{equation}\label{Omega}
\Omega_n(W) = \{w \in W \; | \; v_iw = 0\;\mbox{if}\; \mbox{\rm wt}\ v_i < -n \;
\mbox{for $v\in V$ of homogeneous weight}\}.
\end{equation}
It was shown in \cite{DLM} that $\Omega_n(W)$ is an $A_n(V)$-module
via the action $o(v+O_n(V)) = v_{\mathrm{wt} \, v -1}$ for $v \in V$. In particular, this action satisfies $o(u *_n v) = o(u)o(v)$ for $u,v \in A_n(V)$.
Furthermore, it was shown in \cite{DLM} and \cite{BVY} that there is a bijection between the isomorphism classes of irreducible $A_n(V)$-modules which cannot factor through $A_{n-1}(V)$ and the isomorphism classes of irreducible $\mathbb{N}$-gradable $V$-modules with nonzero degree $n$ component.
The functor $L_n$ from the category of $A_n(V)$-modules to the category of $\mathbb{N}$-gradable $V$-modules is more complicated to define and thus we refer the reader to \cite{DLM} and \cite{BVY} for the definition.
We have the following theorem from \cite{BVY}:
\begin{thm}\label{mainthm}\cite{BVY}
For $n \in \mathbb{N}$, let $U$ be a nonzero $A_n(V)$-module such that if $n>0$, then $U$ does not factor through $A_{n-1}(V)$. Then $L_n(U)$ is an $\mathbb{N}$-gradable $V$-module with $L_n(U)(0) \neq 0$. If we assume further that there is no nonzero submodule of $U$ that factors through $A_{n-1}(V)$, then $\Omega_n/\Omega_{n-1}(L_n(U)) \cong U$.
\end{thm}
One of the main reasons we are interested in Theorem \ref{mainthm} is what it implies about when modules for the higher level Zhu algebras give rise to indecomposable nonsimple modules for $V$ not detected by the lower level Zhu algebras, {cf. \cite{BVY}, \cite{BVY-Heisenberg}, \cite{BVY-Virasoro}. A major motivation for the present paper is to provide methods for construction of those higher level Zhu algebras which can be used to induce new indecomposable nonsimple modules.
\section{Some general results for generators and relations of the algebra $A_n(V)$}\label{general-n-section}
In this section we prove the main results of the paper regarding generators and relations
that occur in $A_n(V)$ in certain settings. In Section \ref{recursion-section}, we prove a general relation that holds for all elements in a vertex operator algebra, as well as a recursion formula that holds for ``most" elements in a vertex operator algebra. In Section \ref{generators-section} we use this recursion formula to identify sets of generators for $A_n(V)$ if $V$ is a strongly generated cyclic vertex operator algebra, such as the Heisenberg or Virasoro vertex operator algebras, but this also applies for broad classes of elements in a more general vertex operator algebra.
Recall the notation (\ref{define-O-subsets}) and (\ref{define-O}). To distinguish between relations in $A_n(V)$ arising from $O^L(V)$ versus $O^\circ_n(V)$ in our results below, we use the notation $u \approx v$ if $u \equiv v \, \mathrm{mod} \,O^L(V)$. And we will write $u \sim_n v $ if $u \equiv v \, \mathrm{mod} \, O_n^\circ(V)$. More broadly if $u$ is equivalent to $w$ modulo $O_n(V)$, we will write $u \equiv_n w$.
\subsection{General reduction relations in $A_n(V)$ arising from $O^L(V)$ and $O_n^\circ(V)$}\label{recursion-section}
We have the following reduction relation in $A_n(V)$ arising from $O^L(V)$.
\begin{lem}\label{L-reduction-lemma}
Let $V$ be a vertex operator algebra, $r \in \mathbb{N}$, $u^{(1)}, \dots, u^{(r)} \in V$, and $ k_1, \dots, k_r \in \mathbb{Z}$. For $w = u_{k_1}^{(1)} u_{k_2}^{(2)} \cdots u_{k_r}^{(r)} \mathbf{1} \in V$, we have
\begin{equation}
(\mathrm{wt} \, w )u_{k_1}^{(1)} u_{k_2}^{(2)} \cdots u_{k_r}^{(r)} \mathbf{1} \approx
\sum_{j = 1}^r k_j u_{k_1}^{(1)} u_{k_2}^{(2)} \cdots u_{k_{j-1}}^{(j-1)} u_{k_j - 1}^{(j)} u_{k_{j+1}}^{(j+1)} \cdots u_{k_r}^{(r)} \mathbf{1} .
\end{equation}
In particular, for $u,v \in V$ and $j,k \in \mathbb{Z}$, $j \neq 0$, we have
\begin{equation}\label{L-reduction-special-case}
u_{-j-1} v_{-k} {\bf 1} \approx - \left( \frac{\mathrm{wt} \, u + \mathrm{wt} \, v +j +k -2}{j} \right) u_{-j} v_{-k} {\bf 1} - \frac{k}{j} u_{-j} v_{-k-1} {\bf 1} .
\end{equation}
\end{lem}
\begin{proof} From the $L(-1)$-derivative property and the $L(-1)$-bracket derivative property for $V$, we have that for $k \in \mathbb{Z}$ and $v \in V$.
\begin{equation}\label{L(-1)-properties}
\left[ L(-1), v_{k} \right] = (L(-1) v)_{k} = -k v_{k-1}.
\end{equation}
Thus we have for $w = u_{k_1}^{(1)} u_{k_2}^{(2)} \cdots u_{k_r}^{(r)} \mathbf{1}$
\begin{eqnarray*}
0 &\approx& (L(-1) + L(0)) u_{k_1}^{(1)} u_{k_2}^{(2)} \cdots u_{k_r}^{(r)} \mathbf{1}\\
&=& (\mathrm{wt} \, w )u_{k_1}^{(1)} u_{k_2}^{(2)} \cdots u_{k_r}^{(r)} \mathbf{1}
- k_1 u_{k_1 -1}^{(1)} u_{k_2}^{(2)} \cdots u_{k_r}^{(r)} \mathbf{1} \\
& & \quad - \, k_2 u_{k_1}^{(1)} u_{k_2-1}^{(2)} \cdots u_{k_r}^{(r)} \mathbf{1} - \cdots -k_r u_{k_1}^{(1)} u_{k_2}^{(2)} \cdots u_{k_r-1}^{(r)} \mathbf{1} \\
&=& (\mathrm{wt} \, w )u_{k_1}^{(1)} u_{k_2}^{(2)} \cdots u_{k_r}^{(r)} \mathbf{1}
- \sum_{j = 1}^r k_j u_{k_1}^{(1)} u_{k_2}^{(2)} \cdots u_{k_{j-1}}^{(j-1)} u_{k_j - 1}^{(j)} u_{k_{j+1}}^{(j+1)} \cdots u_{k_r}^{(r)} \mathbf{1}
\end{eqnarray*}
proving the first equation.
The second equation follows from the first by letting $r = 2$, $u^{(1)} = u$, $u^{(2)} = v$, $k_1 = -j$, and $k_2 = -k$.
\end{proof}
For certain elements of $V$, we have the following relations in $A_n(V)$ arising from a recursion in $O_n^\circ(V)$.
\begin{prop}\label{recursion-n}
Let $V$ be a vertex operator algebra and fix $n \in \mathbb{N}$.
If $u \in V$ satisfies $\mathrm{wt} \, u > -n$, then for $v \in V$ and $m \geq n+ 1$
\begin{equation}\label{recursion-n-equation}
u_{\mathrm{wt} \, u - m - 1}v \ \sim_n \ (-1)^{m+ \mathrm{wt} \, u} \sum_{j = 1}^{n+ \mathrm{wt} \, u} \binom{m-n-1}{j-1} \binom{m-n-j-1}{n+\mathrm{wt} \, u -j } u_{\mathrm{wt} \, u - n - j - 1} v
\end{equation}
where $\sim_n$ denotes equivalence modulo $O^\circ_n(V)$. In particular, this gives $u_{\mathrm{wt} \, u - m - 1} v \ \sim_n \ u_{\mathrm{wt} \, u - m - 1} v$, and thus no new relations in $A_n(V)$ for $n+1 \leq m \leq 2n + \mathrm{wt} \, u$, but does give additional relations in $A_n(V)$ arising from $O_n^\circ(V)$ for $m > 2n + \mathrm{wt} \, u$.
If $u \in V$ satisfies $\mathrm{wt} \, u = -n$, then for $v \in V$, and $m \geq n+ 1$,
\begin{equation}\label{recursion-n-equation-lowest}
u_{\mathrm{wt} \, u - m -1}v \sim_n 0.
\end{equation}
\end{prop}
\begin{proof}
For $\mathrm{wt} \, u > -n$, and $n+1 \leq m \leq 2n + \mathrm{wt} \, u$, if $j \neq m - n$ on the righthand side of Eqn.\ (\ref{recursion-n-equation}) then one of the two binomial coefficients for that term in the sum is zero. Whereas if $j = m-n$, the coefficient of that term is one. Thus for $n+1 \leq m \leq 2n + \mathrm{wt} \, u$, the righthand side of this equation reduces to $u_{\mathrm{wt}\, u - m - 1}v$, giving a trivial relation in $A_n (V)$ in this case, as stated.
By (\ref{more-elements}) with $k = 0$, and letting $m = l \geq 0$ in Eqn.\ (\ref{more-elements}), we have
\begin{eqnarray*}
0 &\sim_n& \mathrm{Res}_x x^{-l - 2n-2} (1+x)^{\mathrm{wt} \, u + n} \sum_{ j \in \mathbb{Z}} u_j x^{-j-1} v \\
&=& \mathrm{Res}_x \sum_{ j \in \mathbb{Z}} \sum_{ i \in \mathbb{N}} \binom{\mathrm{wt} \, u + n}{i} x^{i-l-2n-j-3} u_j v \\
&=& \sum_{i \in \mathbb{N}} \binom{\mathrm{wt} \, u + n}{i} u_{i- l - 2n - 2}v.
\end{eqnarray*}
Thus for $\mathrm{wt} \, u + n > 0$ and $l \geq 0$, we have
\begin{equation}\label{for-recursion-proof}
u_{-l-2n-2}v \ \sim_n \ - \sum_{i = 1}^{n + \mathrm{wt} \, u} \binom{n + \mathrm{wt} \, u}{i} u_{i - l - 2n - 2}v,
\end{equation}
and for $\mathrm{wt} \, u = -n$ and $l\geq 0$, we have $u_{-l-2n-2} v \sim_n 0$. This finishes the proof of the last statement of the proposition, i.e. the case when $\mathrm{wt} \, u = -n$.
We are left with proving Eqn.\ (\ref{recursion-n-equation}) holds when $\mathrm{wt} \, u > -n$ and $m > 2n + \mathrm{wt} u$. Since the equation holds for $m = 2n + \mathrm{wt} \, u$, we make the inductive assumption that Eqn.\ (\ref{recursion-n-equation}) holds for $u_{\mathrm{wt} \, u -m' - 1}$ for $2n+ \mathrm{wt} \, u \leq m' < m$. Then by Eqn.\ (\ref{for-recursion-proof}) with $l = m - \mathrm{wt} \, u - 2n - 1$ (which is greater than or equal to zero when $m>2n+\mathrm{wt}\, u$), and the inductive assumption, as well as the fact that we already have shown that Eqn.\ (\ref{recursion-n-equation}) holds for $u_{\mathrm{wt} \, u -m' - 1}$ when $n+ 1 \leq m' \leq 2n+ \mathrm{wt} \, u$, we have
\begin{eqnarray*}
\lefteqn{u_{\mathrm{wt} \, u - m - 1}v \ = u_{-l - 2n - 2}v }\\
& \sim_n & - \sum_{i= 1}^{n+ \mathrm{wt} \, u} \binom{n + \mathrm{wt} \, u}{i} u_{i- l - 2n - 2}v\\
&=& - \sum_{i = 1}^{n + \mathrm{wt} \, u} \binom{n + \mathrm{wt} \, u}{i} u_{\mathrm{wt} \, u - (m-i) - 1}v \\
&\sim_n& - \sum_{i = 1}^{n + \mathrm{wt} \, u} \binom{n + \mathrm{wt} \, u}{i}
(-1)^{m-i+\mathrm{wt} \, u} \sum_{j= 1}^{n+ \mathrm{wt} \, u} \binom{m-i-n-1}{j-1} \binom{m-i-n-j-1}{n+\mathrm{wt} \, u - j} \\
& & \qquad u_{\mathrm{wt} \, u - n - j - 1} v\\
&=& (-1)^{m+\mathrm{wt} \, u} \sum_{j= 1}^{n+ \mathrm{wt} \, u} \ \sum_{i = 1}^{ n + \mathrm{wt} \, u } (-1)^{i+1} \binom{n + \mathrm{wt} \, u }{i}
\binom{m-i-n-1}{j-1} \binom{m-i-n-j-1}{n+\mathrm{wt} \, u - j} \\
& & \qquad u_{\mathrm{wt} \, u - n - j - 1} v.
\end{eqnarray*}
Thus the result holds if we can show that for $j = 1, 2, \dots, n + \mathrm{wt} \, u$, and $m > 2 n + \mathrm{wt} \, u$
\begin{multline*}
\sum_{i = 1}^{ n + \mathrm{wt} \, u } (-1)^{i+1} \binom{n + \mathrm{wt} \, u}{i}
\binom{m-i-n-1}{j-1} \binom{m-i-n-j-1}{n+\mathrm{wt} \, u - j } \\
= \binom{m-n-1}{j-1} \binom{m-n-j-1}{n+\mathrm{wt} \, u - j }
\end{multline*}
or equivalently that for $j = 1, \dots, n + k$, with $k = \mathrm{wt} \, u$, and $m > 2n + k$ that
\begin{equation}\label{combinatorial-formula}
\sum_{i = 0}^{ n + k} (-1)^{i+1} \binom{ n + k }{i}
\binom{m-i-n-1}{j-1} \binom{m-i-n-j-1}{n+ k-j} =0.
\end{equation}
Since the proof of this last equality is quite long, and only a combinatorial identity, we leave it for the appendix; see Proposition \ref{recursion-n-identity}.
\end{proof}
The recursion (\ref{recursion-n-equation}) can be used to determine certain elements in $O_n(V)$ that are useful in determining the generators and relations for $A_n(V)$. For instance,
letting $Y^+(v,x) \in (\mathrm{End}(V))[[x]]$ and $Y^-(v,x)\in x^{-1}(\mathrm{End}(V))[[x^{-1}]]$ denote the regular and singular parts, respectively, of $Y(v,x)$ for $v \in V$, one can use the recursion to investigate which elements are of the form $u \circ_n v$ for $u,v \in V$, and thus lie in $O_n(V)$. To that end, we note the following lemma.
\begin{lem}\label{Y+-lemma} Let $V$ be a vertex operator algebra and fix $n \in \mathbb{N}$. If $u \in V$ satisfies wt $u > -n$, then for $v \in V$, we have
\begin{multline}\label{Y+}
Y^+(u,x)v \sim_n
\sum_{k=1}^{n - \mathrm{wt} \, u + 1} u_{-k} v x^{k-1} \\
+ \sum_{m\ge n+1} \sum_{k = n - \mathrm{wt} \, u + 2}^{2n + 1} (-1)^{m- \mathrm{wt} \, u} \binom{m-n-1}{k-n + \mathrm{wt} \, u -2} \binom{m-k-\mathrm{wt} \, u}{2n-k + 1 } u_{-k} v x^{-\mathrm{wt} \, u + m} ,
\end{multline}
and if $\mathrm{wt} \, u = -n$, then
\begin{equation}\label{Y+-neg}
Y^+(u,x)v \sim_n
\sum_{k=1}^{2n + 1} u_{-k} v x^{k-1} .
\end{equation}
\end{lem}
\begin{proof}
Applying the recursion given in Eqn.\ (\ref{recursion-n-equation}), we have that if $\mathrm{wt} \, u > -n$, then
\begin{eqnarray*}
\lefteqn{Y^+(u, x) v }\\
&=& \sum_{k\ge1}u_{-k}vx^{k-1} = \sum_{m\ge \mathrm{wt} \, u}u_{\mathrm{wt} \, u - m - 1}vx^{-\mathrm{wt} \, u + m} \\
&=& \sum_{m=\mathrm{wt} \, u}^n u_{\mathrm{wt} \, u - m - 1}vx^{-\mathrm{wt} \, u + m} + \sum_{m\ge n+1} u_{\mathrm{wt} \, u - m - 1}vx^{-\mathrm{wt} \, u + m } \\
&\sim_n&
\sum_{m=\mathrm{wt} \, u}^n u_{\mathrm{wt} \, u - m - 1} v x^{-\mathrm{wt} \, u + m} \\
& & \quad + \sum_{m\ge n+1} (-1)^{m+ \mathrm{wt} \, u} \sum_{j = 1}^{n+ \mathrm{wt} \, u} \binom{m-n-1}{j-1} \binom{m-n-j-1}{n+\mathrm{wt} \, u -j } u_{\mathrm{wt} \, u - n - j - 1} v x^{-\mathrm{wt} \, u + m} \\
&\sim_n&
\sum_{k=1}^{n - \mathrm{wt} \, u + 1} u_{-k} v x^{k-1} \\
& & \quad + \sum_{m\ge n+1} \sum_{k = n - \mathrm{wt} \, u + 2}^{2n + 1} (-1)^{m- \mathrm{wt} \, u} \binom{m-n-1}{k-n + \mathrm{wt} \, u -2} \binom{m-k-\mathrm{wt} \, u}{2n-k + 1 } u_{-k} v x^{-\mathrm{wt} \, u + m} ,
\end{eqnarray*}
giving Eqn.\ (\ref{Y+}). Eqn.\ (\ref{Y+-neg}) follows easily from the first two equalities in the calculation above and Eqn.\ (\ref{recursion-n-equation-lowest}).
\end{proof}
\begin{rem}{\em In practice, as in \cite{AB-details-of-Heisenberg}, one can often find a closed form for (\ref{Y+}) in terms of rational functions in $x$ which allows for ease of calculations. Then using the $L(-1)$-derivative property, one can obtain similar expressions of equivalence modulo $O_n^\circ(V)$ for $Y^+ (u_{-k} \mathbf{1}, x)$, in particular for $1 \leq k \leq 2n+1$. As we show below in Proposition \ref{general-generators-prop},
in many circumstances these $u_{-k}$, for $1 \leq k \leq 2n+1$, are enough to form a generating set for elements in $A_n(V)$ generated by $u$.
}
\end{rem}
\subsection{Generators and relations for subalgebras of $A_n(V)$ arising from certain elements of $V$ using the reduction and recursion}\label{generators-section}
In this section, we use this recursion relation in $A_n(V)$ given in Proposition \ref{recursion-n} to obtain results regarding the generators for certain subalgebras of $A_n(V)$ arising from elements in $V$ with certain nice properties. Calculating a minimal set of generators for a given higher level Zhu algebra is in general quite difficult. The results that follow give some guidance in how to do this.
Recall (e.g. \cite{LL}) that for $V$ a vertex algebra, a subset $S \subset V$ is said to generate $V$, if $V$ is spanned by elements of the form
\[ V = \mathrm{span} \{u_{k_1}^{(1)} u_{k_2}^{(2)} \cdots u_{k_r}^{(r)} \mathbf{1} \; | \; r \in \mathbb{N}, \ u^{(1)}, \dots, u^{(r)} \in S, \ k_1, \dots, k_r \in \mathbb{Z} \},\]
and this is denoted by $V = \langle S \rangle$. For $V$ a nontrivial vertex algebra, we say that $V$ is {\it positively generated at $K$ by $S$} for $K \in \mathbb{Z}$, if
\[ V = \mathrm{span} \{u_{k_1}^{(1)} u_{k_2}^{(2)} \cdots u_{k_r}^{(r)} \mathbf{1} \; | \; r \in \mathbb{N}, \ u^{(1)}, \dots, u^{(r)} \in S, \ k_1, \dots, k_r \leq -K \},\]
and denote this by $V = \langle S \rangle^K$. If $S$ is finite, and $V$ is positively generated at $K = 1$ by $S$, then we say that $V$ is {\it strongly generated} by $S$.
Suppose $V = \langle S \rangle^K$. For $v \in V$ and $r \in \mathbb{N}$, we define $F_r(v)$ to be the subspace of $V$ linearly spanned by all elements of the form
\begin{equation}\label{general-span}
u_{-k_1}^{(1)} u_{-k_2}^{(2)} \cdots u_{-k_m}^{(m)} v ,
\end{equation}
for $k_1, \dots, k_m \geq K$ and $m\in\mathbb{N}$ with $m \leq r$. Thus $F_r(v) \subseteq F_{r+1}(v)$, and $F_0(v) = \mathbb{C}v$.
We say that $V$ satisfies the {\it permutation property} if for all $m \in \mathbb{Z}_+$, for $\sigma$ any permutation on $m$ letters, for any $u^{(1)}, u^{(2)}, \dots, u^{(m)}, v \in V$, and for any $k_1, \dots, k_m \in \mathbb{Z}$, then
\begin{equation}\label{general-permutation}
u_{-k_1}^{(1)}\cdots u_{-k_m}^{(m)}v - u_{-k_{\sigma(1)}}^{(\sigma(1))} \cdots u_{-k_{\sigma(m)}}^{(\sigma(m))}v \in F_{m-1}(v).
\end{equation}
If Eqn.\ \eqref{general-permutation} holds only for some particular subspaces of $V$, say for $u^{(1)}, \dots, u^{(m)} \in U \subset V$ and $v \in W \subset V$, then we say that the {\it the permutation property holds for $U$ acting on $W$}.
If $V$ is a vertex algebra, $U = \langle S \rangle^1 \subset V$, the permutation property holds for $U$ acting on $W \subset V$, and for all $u \in U$, $w \in W$, we have $u_jw = 0$ if $j\geq0$, then we say that {\it $U$ acts strongly on $W$.}
For the following results, we let $V$ be a vertex operator algebra and for fixed $n \in \mathbb{N}$, we suppose there exists $u \in V$, such that $V^u = \langle u \rangle^1$ is a subvertex algebra of $V$ generated by $u$ and $V^u$ satisfies the permutation property on $W \subset V$. Note that we do not assume that $V^u$ is necessarily a subvertex operator algebra. We introduce the notation $A_n^u$ to denote the subalgebra of $A_n(V)$ generated by elements in $V^u$.
And we use the notation $A_n^u . W$ to denote the elements in $A_n(V)$ of the form $u_{-k_1} \cdots u_{-k_m} v + O_n(V)$ for $v \in W$, and $k_1, \dots, k_m \in \mathbb{Z}_+$.
\begin{rem} {\em Note that examples of vertex algebras which are strongly generated as $\langle u \rangle^1$ include the rank one Heisenberg vertex operator algebra and the vertex operator algebras arising from the Virasoro algebra. Also note that in these cases, the vertex algebra is indeed a vertex operator algebra. In particular, for any vertex operator algebra $V$ with conformal element $\omega$, we have the subvertex operator algebra of $V$ given by $\langle \omega \rangle^1$. }
\end{rem}
\begin{prop}\label{general-generators-prop}
Let $V$ be a vertex operator algebra. Fix $n \in \mathbb{N}$, and suppose there is an element $u \in V$ with $\mathrm{wt} \, u \geq -n$, and such that the subvertex algebra $V^u = \langle u \rangle^1$ satisfies the permutation property. Let $A_n^u$ be the subalgebra of $A_n(V)$ generated by elements in $V^u$. Then elements in $A_n^u$ are generated by elements of the form
\[ u_{-2n - 1}^{i_{2n+1}} u_{- 2n}^{i_{2 n}} \cdots u_{ - 1}^{i_1 }\mathbf{1} + O_n(V) ,\]
for $i_1, i_2, \dots, i_{2 n+1} \in \mathbb{N}$. Moreover, if $V^u$ acts strongly on $W \subset V$, then any element in $A_n^u . W$ can be written, modulo $O_n^\circ(V)$, as a linear combination of such elements of the form $u_{-2n - 1}^{i_{2n+1}} u_{- 2n}^{i_{2 n}} \cdots u_{ - 1}^{i_1 }v + O_n(V)$.
\end{prop}
\begin{proof}
By Proposition \ref{recursion-n}, the fact that $V^u$ is spanned by elements of the form (\ref{general-span}) with $K = 1$, and $V^u$ acts strongly on $W$, we have that modulo $O^\circ_n(V)$, elements in $A_n^u . W$ can be written as a linear combination of elements of the form $u_{-k_1}\cdots u_{-k_r} v$ for $r \in \mathbb{N}$ and $k_j \in \{ 1, \dots, 2 n+1 \}$. By induction on $r$, and the assumption that $V^u$ satisfies the permutation property (\ref{general-permutation}) on $W$, the result follows.
\end{proof}
In order to reduce the set of generators further, we observe the following:
\begin{lem}\label{strongly-acting-lemma}
Let $V$ be a vertex operator algebra, and $U = \langle S \rangle^1 \subset V$, such that $U$ acts strongly on $W \subset V$. If $v \in W$, then modulo $O^L(V)$, $U$ acts strongly on $v_{-k} {\bf 1}$ for all $k \in \mathbb{Z}_+$.
\end{lem}
\begin{proof} We prove the result by induction on $k$. The result is true for $k = 1$ since $v_{-1} {\bf 1} = v$. Assume the result holds for $1\leq k' <k$. For $i_1, \dots, i_m \in \mathbb{Z}_+$, $u^{(1)}, \dots, u^{(m)} \in U$, and $w = u^{(1)}_{-i_1} \cdots u^{(m)}_{-i_m} v_{-(k-1)} {\bf 1}$, Lemma \ref{L-reduction-lemma} implies that
\begin{eqnarray*}
\lefteqn{ (k-1) u^{(1)}_{-i_1} u^{(2)}_{-i_2} \cdots u^{(m)}_{-i_m} v_{-k} {\bf 1} } \\
&\approx& - (\mathrm{wt} \, w ) u^{(1)}_{-i_1} \cdots u^{(m)}_{-i_m} v_{-(k-1)} {\bf 1} - \sum_{j = 1}^m i_j u^{(1)}_{-i_1} \cdots u_{-i_{j-1}}^{(j-1)}u_{-i_{j-1}} u^{(j)}_{-i_j - 1} u^{(j+1)} _{-i_{j+1} } \cdots u^{(m)}_{-i_m} v_{-(k-1)} {\bf 1} ,
\end{eqnarray*}
so that modulo $O^L(V)$, $u^{(1)}_{-i_1} u^{(2)}_{-i_2} \cdots u^{(m)}_{-i_m} v_{-k} {\bf 1}$ is a linear combination of elements comprised of $U$ acting strongly on $v_{-(k-1)}{\bf 1}$. Therefore, modulo $O^L(V)$, $U$ acts strongly on $v_{-k} {\bf 1}$.
\end{proof}
Now combining the restriction of the generating set given by Proposition \ref{general-generators-prop} and the additional relations given by $O^L(V)$, we can further restrict the generators as follows:
\begin{thm}\label{reducing-general-generators-thm}
Let $V$ be a vertex operator algebra. Fix $n \in \mathbb{Z}_+$, and suppose there is a $u \in V$ with $\mathrm{wt} \, u \geq -n$, and such that $V^u = \langle u \rangle^1$ satisfies the permutation property and acts strongly on $W \subset V$. Let $A_n^u$ be the subalgebra of $A_n(V)$ generated by elements in $V^u$. Then elements in $A_n^u$ are generated by elements of the form
\[ u_{-2n}^{i_{2n}} u_{-2 n + 1}^{i_{2n -1}} \cdots u_{ - 1}^{i_1 }\mathbf{1} + O_n(V),\]
for $i_1, i_2, \dots, i_{2n} \in \mathbb{N}$, and elements in $A_n^u . W$ can be expressed as linear combinations of elements of the form $u_{-2n}^{i_{2n}} u_{-2 n + 1}^{i_{2n -1}} \cdots u_{ - 1}^{i_1 }v_{-k} {\bf 1} + O_n(V)$ for $v \in W$ and $k \in \mathbb{Z}_+$.
Furthermore,
\begin{equation}\label{-2n-1-reduction}
u_{-2n-1} v_{-k} {\bf 1} \approx -\left( \frac{\mathrm{wt} \, u + \mathrm{wt} \, v + 2n + k -2 }{2n} \right)u_{-2n} v_{-k} {\bf 1} - \frac{k}{2n} u_{-2n} v_{-k-1} {\bf 1}.
\end{equation}
And for $i_1, \dots, i_{2n} \in \mathbb{N}$ and $i_{2n+1}, k \in \mathbb{Z}_+$, we have
\begin{eqnarray}\label{reduce-equation}
\lefteqn{\ \ \ \ \ \ \ \ u_{-2n-1}^{i_{2n + 1}} u_{-2n}^{i_{2n}} u_{-2n + 1}^{i_{2n-1}} \cdots u_{-1}^{i_1} v_{-k} {\bf 1} }\\
&\equiv_n & \! \! -\frac{1}{2n( i_{2n} + 1) } \Biggl( \biggl( \mathrm{wt} \, v_{-k} {\bf 1} - 1 + \sum_{j = 1}^{2n+1} i_j (\mathrm{wt} \, u + j -1) \biggr) u_{-2n - 1}^{i_{2n+1}-1} u_{-2 n}^{i_{2n}+1} u_{-2n+1}^{i_{2n-1}} \cdots u_{ - 1}^{i_1 } v_{-k} {\bf 1} \nonumber \\
& & - (2n+1) (i_{2n+1} -1)\biggl( \sum_{j = 1}^{n+ \mathrm{wt} \, u} \! \binom{n + \mathrm{wt} \, u}{j-1} u_{\mathrm{wt} \, u - n - j - 1} \biggr) u_{-2n-1}^{i_{2n+1} -2} u_{-2n}^{i_{2n}+1}u_{-2n+1}^{i_{2n-1}} \cdots u_{-1}^{i_1} v_{-k} {\bf 1} \nonumber \\
& & + \, (2n-1) i_{2n-1} u_{-2n-1}^{i_{2n + 1}-1} u_{-2n}^{i_{2n} + 2}u_{-2n + 1}^{i_{2n-1} - 1} u_{-2n + 2}^{i_{2n-2}} \cdots u_{-1}^{i_1} v_{-k} {\bf 1} + \cdots \nonumber \\
& & \cdots + 2 i_2 u_{-2n - 1}^{i_{2n+1}-1} u_{-2 n}^{i_{2n}+1}u_{-2n+1}^{i_{2n-1}} \cdots u_{-4}^{i_4} u_{-3}^{i_3 + 1} u_{-2}^{i_2 - 1} u_{ - 1}^{i_1 } v_{-k} {\bf 1} \nonumber \\
& & + \, i_1 u_{-2n - 1}^{i_{2n+1}-1} u_{-2 n}^{i_{2n}+1} u_{-2n+1}^{i_{2n-1}} \cdots u_{-3}^{i_3} u_{-2}^{i_2 + 1} u_{ - 1}^{i_1-1 } v_{-k} {\bf 1} + k u_{-2 n-1}^{i_{2n+1}- 1} u_{-2n}^{i_{2n}+1} u_{-2n +1}^{i_{2n-1}} \cdots u_{ - 1}^{i_1 } v_{-k-1} {\bf 1} \nonumber \\
& & + f_{i_{2n+1} + i _{2n} + \cdots i_1 -1}( v_{-k} {\bf 1} ) \Biggr).\nonumber
\end{eqnarray}
for some $f_{i_{2n+1} + i _{2n} + \cdots i_1 -1}( v_{-k} {\bf 1} ) \in F_{i_{2n+1} + i _{2n} + \cdots i_1 -1}(v_{-k} {\bf 1} )$, where the term with the second summation in Eqn.\ \eqref{reduce-equation} is zero if $\mathrm{wt} \, u = -n$, where if $n = 1$, this is to be interpreted as
\begin{eqnarray}\label{reduce-equation-n=1}
\lefteqn{u_{-3}^{i_{3}} u_{-2}^{i_{2}} u_{-1}^{i_1} v_{-k} {\bf 1} }\\
&\equiv_1 & \! \! -\frac{1}{2( i_{2} + 1) } \Biggl( \biggl( \mathrm{wt} \, v_{-k} {\bf 1} - 1 + \sum_{j = 1}^{3} i_j (\mathrm{wt} \, u + j -1) \biggr) u_{-3}^{i_{3}-1} u_{-2}^{i_{2}+1} u_{ - 1}^{i_1 } v_{-k} {\bf 1} \nonumber \\
& & - 3 (i_{3} -1)\biggl( \sum_{j = 1}^{1+ \mathrm{wt} \, u} \! \binom{1 + \mathrm{wt} \, u}{j-1} u_{\mathrm{wt} \, u - j - 2} \biggr) u_{-3}^{i_{3} -2} u_{-2}^{i_{2}+1} u_{-1}^{i_1} v_{-k} {\bf 1} \nonumber \\
& & + \, i_1 u_{-3}^{i_{3}-1} u_{-2}^{i_{2}+2} u_{ - 1}^{i_1-1 } v_{-k} {\bf 1} + k u_{-3}^{i_{3}- 1} u_{-2}^{i_{2}+1} u_{ - 1}^{i_1 } v_{-k-1} {\bf 1} \nonumber \\
& & + f_{i_{3} + i _{2} + i_1 -1}( v_{-k} {\bf 1} ) \Biggr),\nonumber
\end{eqnarray}
and where $f_{i_{2n+1} + i_{2n} + \cdots i_1 -1} (v_{-k} \mathbf{1}) = 0$ in both (\ref{reduce-equation}) and (\ref{reduce-equation-n=1}) if the $u_{-j}$'s for $j \in \mathbb{Z}_+$ commute.
\end{thm}
\begin{proof} We first prove Eqns.\ (\ref{-2n-1-reduction}) and \eqref{reduce-equation}. Eqn.\ (\ref{-2n-1-reduction}) follows from Eqn.\ (\ref{L-reduction-special-case}) applied to the case when $j = 2n$.
Now note that for $u_{-2n -2} = u_{\mathrm{wt} \, u - m - 1}$ and $m = \mathrm{wt} \, u + 2n + 1 \geq n + 1$, we have by Proposition \ref{recursion-n}, that for any $v \in V$,
\begin{eqnarray}\label{using-n-recursion}
u_{-2n-2} v &\sim_n & - \sum_{j = 1}^{n+ \mathrm{wt} \, u} \binom{n + \mathrm{wt} \, u}{j-1} \binom{n + \mathrm{wt} \, u -j}{n+\mathrm{wt} \, u -j } u_{\mathrm{wt} \, u - n - j - 1} v\\
&=& - \sum_{j = 1}^{n+ \mathrm{wt} \, u} \binom{n + \mathrm{wt} \, u}{j-1} u_{\mathrm{wt} \, u - n - j - 1} v, \nonumber
\end{eqnarray}
if $\mathrm{wt} \, u > -n$ and $u_{-2n-2}v \sim_n 0$ if $\mathrm{wt} \, u = -n$.
Thus using (\ref{L(-1)-properties}), the permutation property of $V^u$ (\ref{general-permutation}), the fact that $V^u$ acts strongly on $W$, and (\ref{using-n-recursion}), we have that for $i_1, i_2, \dots, i_{2n+1} \in \mathbb{N}$, $k \in \mathbb{Z}_+$, and $\mathrm{wt} \, u > -n$
\begin{eqnarray*}
0 &\approx& (L(-1) + L(0)) u_{-2n - 1}^{i_{2n+1}} u_{-2 n}^{i_{2n}} \cdots u_{ - 1}^{i_1 } v_{-k} {\bf 1} \\
&=& \left( \mathrm{wt} \, v_{-k} {\bf 1} + \sum_{j = 1}^{2n+1} i_j (\mathrm{wt} \, u + j -1) \right) u_{-2n - 1}^{i_{2n+1}} u_{-2 n}^{i_{2n}} \cdots u_{ - 1}^{i_1 } v_{-k} {\bf 1} \\
& & \quad + \, (2n+1) i_{2n+1} u_{-2n -2} u_{-2n-1}^{i_{2n+1} -1} u_{-2n}^{i_{2n}} \cdots u_{-1}^{i_1} v_{-k} {\bf 1} \\
& & \quad + \, 2n i_{2n} u_{-2n-1}^{i_{2n + 1} + 1} u_{-2n}^{i_{2n} -1} u_{-2n + 1}^{i_{2n-1}} \cdots u_{-1}^{i_1} v_{-k} {\bf 1} + \cdots \\
& & \quad \cdots + 2 i_2 u_{-2n - 1}^{i_{2n+1}} u_{-2 n}^{i_{2n}} \cdots u_{-3}^{i_3 + 1} u_{-2}^{i_2 - 1} u_{ - 1}^{i_1 } v_{-k} {\bf 1} \\
& & \quad + \, i_1 u_{-2n - 1}^{i_{2n+1}} u_{-2 n}^{i_{2n}} \cdots u_{-2}^{i_2 + 1} u_{ - 1}^{i_1-1 }v_{-k} \mathbf{1} +k u_{-2 n-1}^{i_{2n+1}} u_{-2n}^{i_{2n}} \cdots u_{ - 1}^{i_1 } v_{-k-1} {\bf 1} \\
& & \quad + f_{i_{2n+1} + i _{2n} + \cdots i_1 -1}( v_{-k} {\bf 1} )\\
&\sim_n& \left( \mathrm{wt} \, v_{-k} {\bf 1} + \sum_{j = 1}^{2n+1} i_j (\mathrm{wt} \, u + j -1) \right) u_{-2n - 1}^{i_{2n+1}} u_{-2 n}^{i_{2n}} \cdots u_{ - 1}^{i_1 } v_{-k} {\bf 1} \\
& & \quad - \, (2n+1) i_{2n+1} \left( \sum_{j = 1}^{n+ \mathrm{wt} \, u} \binom{n + \mathrm{wt} \, u}{j-1} u_{\mathrm{wt} \, u - n - j - 1} \right) u_{-2n-1}^{i_{2n+1} -1} u_{-2n}^{i_{2n}} \cdots u_{-1}^{i_1} v_{-k} {\bf 1} \\
& & \quad + \, 2n i_{2n} u_{-2n-1}^{i_{2n + 1} + 1} u_{-2n}^{i_{2n} -1} u_{-2n + 1}^{i_{2n-1}} \cdots u_{-1}^{i_1} v_{-k} {\bf 1} \\
& & \quad + \, (2n-1) i_{2n-1} u_{-2n-1}^{i_{2n + 1}} u_{-2n}^{i_{2n} + 1}u_{-2n + 1}^{i_{2n-1} - 1} u_{-2n + 2}^{i_{2n-2}} \cdots u_{-1}^{i_1} v_{-k} \mathbf{1} + \cdots \\
& & \quad \cdots + 2 i_2 u_{-2n - 1}^{i_{2n+1}} u_{-2 n}^{i_{2n}} \cdots u_{-3}^{i_3 + 1} u_{-2}^{i_2 - 1} u_{ - 1}^{i_1 } v_{-k} {\bf 1} \\
& & \quad + \, i_1 u_{-2n - 1}^{i_{2n+1}} u_{-2 n}^{i_{2n}} \cdots u_{-2}^{i_2 + 1} u_{ - 1}^{i_1-1 } v_{-k} \mathbf{1} + k u_{-2 n-1}^{i_{2n+1}} u_{-2n}^{i_{2n}} \cdots u_{ - 1}^{i_1 } v_{-k-1} {\bf 1} \\
& & \quad + f_{i_{2n+1} + i _{2n} + \cdots i_1 -1}( v_{-k} {\bf 1} ),\\
\end{eqnarray*}
where $ f_{i_{2n+1} + i _{2n} + \cdots i_1 -1}( v_{-k} {\bf 1} ) \in F_{i_{2n+1} + i _{2n} + \cdots i_1 -1}(v_{-k} {\bf 1} )$, and this is zero if the $u_{-2n-1}, \dots, u_{-1}$ commute since these terms are only arising from reordering the $u_{-m}$'s after acting by $L(-1)$.
Therefore if $i_{2n} \neq 0$, we have that
\begin{eqnarray*}
\lefteqn{ u_{-2n-1}^{i_{2n + 1} + 1} u_{-2n}^{i_{2n} -1} u_{-2n + 1}^{i_{2n-1}} \cdots u_{-1}^{i_1} v_{-k} {\bf 1} }\\
&\equiv_n& -\frac{1}{2n i_{2n} } \Biggl( \left( \mathrm{wt} \, v_{-k} {\bf 1} + \sum_{j = 1}^{2n+1} i_j (\mathrm{wt} \, u + j -1) \right) u_{-2n - 1}^{i_{2n+1}} u_{-2 n}^{i_{2n}} \cdots u_{ - 1}^{i_1 } v_{-k} {\bf 1} \\
& & \quad - \, (2n+1) i_{2n+1} \left( \sum_{j = 1}^{n+ \mathrm{wt} \, u} \binom{n + \mathrm{wt} \, u}{j-1} u_{\mathrm{wt} \, u - n - j - 1} \right) u_{-2n-1}^{i_{2n+1} -1} u_{-2n}^{i_{2n}} \cdots u_{-1}^{i_1} v_{-k} {\bf 1} \\
& & \quad + \, (2n-1) i_{2n-1} u_{-2n-1}^{i_{2n + 1}} u_{-2n}^{i_{2n} + 1}u_{-2n + 1}^{i_{2n-1} - 1} u_{-2n + 2}^{i_{2n-2}} \cdots u_{-1}^{i_1} v_{-k} {\bf 1} + \cdots \\
& & \quad \cdots + 2 i_2 u_{-2n - 1}^{i_{2n+1}} u_{-2 n}^{i_{2n}} \cdots u_{-4}^{i_4} u_{-3}^{i_3 + 1} u_{-2}^{i_2 - 1} u_{ - 1}^{i_1 } v_{-k} {\bf 1} \\
& & \quad + \, i_1 u_{-2n - 1}^{i_{2n+1}} u_{-2 n}^{i_{2n}} \cdots u_{-3}^{i_3} u_{-2}^{i_2 + 1} u_{ - 1}^{i_1-1 } v_{-k} {\bf 1} + k u_{-2 n-1}^{i_{2n+1}} u_{-2n}^{i_{2n}} \cdots u_{ - 1}^{i_1 } v_{-k-1} {\bf 1} \\
& & \quad + f_{i_{2n+1} + i _{2n} + \cdots i_1 -1}( v_{-k} {\bf 1} ) \Biggr).
\end{eqnarray*}
Equivalently, shifting $i_{2n+1} \mapsto i_{2n + 1} -1$ and $i_{2n} - 1 \mapsto i_{2n}$, we have that for $i_{2n+1} \in \mathbb{Z}_+$ and $i_1, \dots, i_{2n} \in \mathbb{N}$, gives Eqn.\ \eqref{reduce-equation}.
We are left to prove the statements in the first paragraph of the theorem.
By Proposition \ref{general-generators-prop}, we have that $A_n^u$ is generated by elements of the form
\[ u_{-2n - 1}^{i_{2n+1}} u_{- 2n}^{i_{2 n}} \cdots u_{ - 1}^{i_1 }\mathbf{1} + O_n(V) ,\]
for $i_1, i_2, \dots, i_{2 n+1} \in \mathbb{N}$, modulo $O_n^\circ(V) \subset O_n(V)$, and $A_n^u . W$ consists of linear combinations of elements of the form $u_{-2n - 1}^{i_{2n+1}} u_{- 2n}^{i_{2 n}} \cdots u_{ - 1}^{i_1 } v + O_n(V) $ for $v \in W$, also modulo $O_n^\circ(V) \subset O_n(V)$.
For fixed $v \in W$, and $i_1, i_2, \dots, i_{2n+1} \in \mathbb{N}$, we need to show that
$u_{-2 n - 1}^{i_{2n +1}} u_{-2n}^{i_{2n}} \cdots u_{ - 1}^{i_1 }v$ can be expressed as a linear combination of elements of the form $ u_{-2n}^{j_{2n}} \cdots u_{ - 1}^{j_1 }v_{-k} {\bf 1}$, for $k \in \mathbb{Z}_+$ and $j_1, j_2, \dots, j_{2n} \in \mathbb{N}$. We prove this by induction on the total degree of elements of the form $u_{-2 n - 1}^{i_{2n +1}} u_{-2n}^{i_{2n}} \cdots u_{ - 1}^{i_1 }v_{-k} {\bf 1}$ where the total degree is defined as follows:
For fixed $v \in W$, $k \in \mathbb{Z}_+$ and $i_1, i_2, \dots, i_{2n+1} \in \mathbb{N}$, we define the total degree of $u_{-2 n - 1}^{i_{2n +1}} u_{-2n}^{i_{2n}} \cdots u_{ - 1}^{i_1 }v_{-k} {\bf 1}$ with respect to $v_{-k} {\bf 1}$ to be
\[ \mathrm{deg} \, u_{-2 n - 1}^{i_{2n +1}} u_{-2n}^{i_{2n}} \cdots u_{ - 1}^{i_1 } v_{-k} {\bf 1} = \sum_{j=1}^{2n+1} i_j .\]
Note that the degree is independent of $k$.
For degree one elements acting on $v_{-k} {\bf 1}$, including $v_{-1} {\bf 1} = v$, if $i_{2n+1} = 0$, the result follows trivially, and if $i_{2n+1} = 1$, the result follows from Eqn.\ \eqref{-2n-1-reduction}.
Assume the result holds for elements of total degree less than or equal to $m$ acting on $v_{-k} {\bf 1}$ for $k \in \mathbb{Z}_+$, and let $i_{1} + \cdots + i_{2n+1} = m+1$. If $i_{2n+1} = 0$ then the result holds trivially. If $i_{2n+1} >0$, the result follows by Eqn.\ (\ref{reduce-equation}), by induction on $i_{2n+1}$ at degree $m+1$, by the inductive assumption for degree less than or equal to $m$, and by Lemma \ref{strongly-acting-lemma}. This finishes the result in the case $\mathrm{wt} \, u > -n$. In the case $\mathrm{wt} \, u = -n$, we simply note that since $u_{-2n-2}v \sim_n 0$, in this case the second term on the right hand side of Eqn.\ \eqref{reduce-equation} vanishes, and the results hold.
\end{proof}
\subsection{Reducing generators of $A_n(V)$ further by establishing multiplication formulas}\label{mult-section}
We proceed to reduce the number of generators further by considering the types of elements produced from multiplying certain generators in $A_n(V)$. We first prove a general multiplication formula in the lemma below, and then a proposition employing one case of this formula to rewrite certain linear combinations of certain generators in $A_n(V)$ in terms of products of other generators. This then ultimately allows us to further reduce the set of generators in Theorem \ref{last-generator-theorem}.
\begin{lem}\label{multiplication-lemma}
Let $V$ be a vertex operator algebra. Fix $n\in \mathbb{N}$. Let $t\in \mathbb{Z}_+$, and $i_1,\dots, i_t \in \mathbb{N}$. For $1\leq s \leq t$, set $p_s=\sum_{j=1}^{s}i_j$, $p_0 = 0$, and $r=\sum_{j=1}^{t}(\text{wt }u+j-1)i_j$. We have the following formula for $u_{-t}^{i_{t}}\dots u_{-1}^{i_{1}}{\bf{1}}*_nv$, where $u, v\in V$
\begin{multline}\label{multiplication-formula-infinity}
u_{-t}^{i_{t}}\cdots u_{-1}^{i_{1}}{\bf{1}}*_nv\\
= \sum_{m=0}^n \sum_{j=-m}^n (-1)^m\binom{m+n}{n}\binom{n+r}{j+m}\displaystyle\sum_{\substack{k_1,\dots ,k_{p_t}\in \mathbb{N}\\k_1+\cdots +k_{p_t}=n-j}} \! \! {}_{\circ}^{\circ}P_{i_1,\dots, i_{t}}(k_1,\dots, k_{p_t}) {}_{\circ}^{\circ} v +g_{i_1,\dots, i_{t}}(v),
\end{multline}
where
\begin{equation}\label{define-P}
P_{i_1,\dots, i_{t}}(k_1,\dots, k_{p_t})=\prod_{s=1}^{t}\prod_{l=1}^{i_{s}}\binom{k_{l+p_{s-1}}+s-1}{s-1}u_{-k_{l+p_{s-1}}-s},
\end{equation}
and $g_{i_1,\dots, i_{t}}(v)$ is a linear combination of elements of the form ${}_{\circ}^{\circ}u_{q_1}\cdots u_{q_{p_t}}{}_{\circ}^{\circ}v$, where $q_1, \dots,q_{p_t} \in \mathbb{Z}$ with at least one of the $q$'s nonnegative.
Furthermore if $n + r \geq 0$ and $r <0$, then the range for the sum in $j$ in Eqn.\ \eqref{multiplication-formula-infinity} can be considered to be $-n\leq j \leq n + r$.
\end{lem}
\begin{proof}
From the definition of multiplication in the level $n$ Zhu algebra, we have
\begin{eqnarray*}
\lefteqn{ u_{-t}^{i_{t}}\cdots u_{-1}^{i_{1}}{\bf{1}}*_n v}\\
&=& \mbox{\rm Res}_x\sum_{m=0}^n(-1)^m\binom{m+n}{n}x^{-m-n-1}\sum_{j=0}^{\infty}\binom{n+r}{j}x^j{}_{\circ}^{\circ}\prod_{s=1}^t\Big(\sum_{l \in \mathbb{Z}}\binom{l +s-1}{s-1}u_{-l-s}x^l\Big)^{i_s}{}_{\circ}^{\circ}v\\
&=& \sum_{m=0}^n\sum_{j=0}^{\infty}(-1)^m\binom{m+n}{n}\binom{n+r}{j}\displaystyle\sum_{\substack{k_1,\dots ,k_{p_t}\in \mathbb{Z}\\k_1+\cdots +k_{p_t}=n+m-j}}{}_{\circ}^{\circ}P_{i_1,\dots, i_{t}}(k_1,\dots, k_{p_t}){}_{\circ}^{\circ}v\\
&=&\sum_{m=0}^n\sum_{j=-m}^{\infty}(-1)^m\binom{m+n}{n}\binom{n+r}{j+m}\displaystyle\sum_{\substack{k_1,\dots ,k_{p_t}\in \mathbb{Z}\\k_1+\cdots +k_{p_t}=n-j}}{}_{\circ}^{\circ}P_{i_1,\dots, i_{t}}(k_1,\dots, k_{p_t}){}_{\circ}^{\circ}v,\
\end{eqnarray*}
where $P$ is given by Eqn.\ \eqref{define-P}.
The binomial coefficient $\binom{k_{l+p_{s-1}}+s-1}{s-1}$ appearing in the formula for $P_{i_1,\dots, i_{t}}(k_1,\dots, k_{p_t})$ is zero if $k_{l+p_{s-1}}$ is such that $-s< k_{j+p_{s-1}}<0$. Thus restricting the $k_{j+p_{s-1}}$'s to be nonnegative includes all the terms corresponding to products of $u_{i}$'s where all $i$'s are negative. In other words, we have
\begin{multline*}
u_{-t}^{i_{t}}\cdots u_{-1}^{i_{1}}{\bf{1}}*_nv \\
=\sum_{m=0}^n\sum_{j=-m}^{\infty}(-1)^m\binom{m+n}{n}\binom{n+r}{j+m}\displaystyle\sum_{\substack{k_1,\dots,k_{p_t}\in \mathbb{N}\\k_1+\cdots +k_{p_t}=n-j}} \! \! {}_{\circ}^{\circ} P_{i_1,\dots, i_{t}}(k_1,\dots, k_{p_t}) {}_{\circ}^{\circ} v+g_{i_1,\dots, i_{t}}(v),
\end{multline*}
where $g_{i_1,\dots, i_{t}}(v)$ is a linear combination of elements of the form ${}_{\circ}^{\circ}u_{q_1}\cdots u_{q_{p_t}}{}_{\circ}^{\circ}v$, for $q_1, \dots, q_{p_t} \in \mathbb{Z}$ with at least one of the $q$'s nonnegative. Observing then that we must have $n-j= k_1 + \cdots + k_{p_t} \geq 0$ if $k_1, \dots, k_{p_t} \in \mathbb{N}$, we see that $j$ ranges only from $-m$ to $n$, proving Eqn.\ \eqref{multiplication-formula-infinity}.
If $n + r \geq 0$, then $\binom{n+r}{j+m}$ is zero when $j>n+r-m$, and thus if $r<0$, the expression becomes
\begin{eqnarray*}
\lefteqn{ u_{-t}^{i_{t}}\cdots u_{-1}^{i_{1}}{\bf{1}}*_n v}\\
&=& \sum_{m=0}^n\sum_{j=-m}^{n+r}(-1)^m\binom{m+n}{n}\binom{n+r}{j+m}\displaystyle\sum_{\substack{k_1,\dots ,k_{p_t}\in \mathbb{N}\\k_1+\cdots +k_{p_t}=n-j}}{}_{\circ}^{\circ}P_{i_1,\dots, i_{t}}(k_1,\dots, k_{p_t}){}_{\circ}^{\circ}v +g_{i_1,\dots, i_{t}}(v),
\end{eqnarray*}
proving the last statement.
\end{proof}
\begin{rem} {\em The formula above shows that the terms in the multiplication formula that are not part of $g_{i_1,\dots, i_{t}}(v)$ are indexed by ordered partitions of $2n$ and less. }
\end{rem}
\begin{rem}{\em
$g_{i_1,\dots, i_{t}}(v)$ can be relatively manageable in the case that $V$ is, for example, the rank one Heisenberg vertex operator algebra, $M(1)$. For instance, in \cite{BVY-Heisenberg} and \cite{AB-details-of-Heisenberg}, respectively, where $A_1(M(1))$ and $A_2(M(1))$, respectively, are explicitly calculated, an ordering is defined on elements of $M(1)$ and Lemma \ref{multiplication-lemma} is used to induct on this ordering and reduce the number of required generators. A key part of this argument uses the fact that terms in $g_{i_1,\dots, i_{t}}(v)$ are of lower order than certain other terms in the corresponding multiplication formula. In Proposition \ref{reduce-prop} and Theorem \ref{last-generator-theorem} below, we give an example of how this can be done.}
\end{rem}
\begin{rem}{\em In Proposition \ref{reduce-prop} and Theorem \ref{last-generator-theorem} below, we give just one example of how Eqn.\ \eqref{multiplication-formula-infinity} can be used to find further relations in $A_n(V)$ and to reduce the set of generators of $A_n(V)$. However, one difficulty in using Eqn.\ \eqref{multiplication-formula-infinity} in this way is that it is sometimes necessary to argue that certain coefficients are nonzero and this can be difficult, especially in cases where Proposition \ref{recursion-n} or Lemma \ref{multiplication-lemma} are being applied. In particular, one often obtains terms whose coefficients are rather complicated combinations of binomial coefficients that must be simplified or at least shown to be nonzero. }
\end{rem}
\begin{cor}\label{multiplication-cor-special-case-more-general}
Let $V$ be a vertex operator algebra. Fix $n \in \mathbb{N}$, and suppose there is a $u \in V$, such that $\mathrm{wt}\, u \geq -n$, and $V^u = \langle u \rangle^1 \subset V$ acts strongly on $W \subset V$. For $i,t \in \mathbb{Z}_+$ and $v \in W$, we have the following multiplication formula
\begin{multline}\label{multiplication-cor-formula}
u_{-t}^i \mathbf{1} *_n v = \sum_{m=0}^n \sum_{j =0}^{m+n} \sum_{\stackrel{p_1, p_2, \dots, p_i \in \mathbb{N}}{p_1 + p_2 + \cdots + p_i = j} } (-1)^m \binom{m+n}{n} \binom{n + i (\mathrm{wt}\, u + t - 1) }{m + n - j} \\
\left( \prod_{l = 1}^i \binom{p_l + t - 1}{t-1} u_{-p_l -t} \right) \, v + g_{i}(v) ,
\end{multline}
where $g_{i}(v)$ is a linear combination of elements of the form ${}_{\circ}^{\circ}u_{q_1}\cdots u_{q_{i}} {}_{\circ}^{\circ}v$ where at least one of the $q$'s is nonnegative.
That is, modulo $g_{i}(v)$, the product $u_{-t}^i \mathbf{1} *_n v$ involves a linear combination of terms of the form $u_{-p_1-t} u_{-p_2-t} \cdots u_{-p_i-t} \, v$ where $\mathbf{p} = (p_1, p_2, \dots, p_i)$ is an ordered partition of $j$ into $i$ nonnegative integer parts for $0 \leq j \leq m+ n \leq 2n$.
Furthermore, when $i = 1$ and $u_{k-t}v = 0$ for $k\in \mathbb{Z}_+$, or if $n + \mbox{\rm wt}\ u + t - 1 \geq 0$ and $\mbox{\rm wt}\ u + t <2$, then $g_1(v) = 0$. In addition, if $n \geq \mbox{\rm wt}\ u + t \geq 2$, then letting $k = \mathrm{wt} \, u + t$, we have
\[g_1(v) = \sum_{m=0}^{k - 2} \sum_{j=-1}^{m -k + 1}(-1)^m\binom{m+n}{n}\binom{n+ k -1 }{m + n - j }\binom{j + t-1}{t-1} u_{-j - t} v , \]
and more generally
\[ g_1(v) = \sum_{m=0}^n\sum_{j=-1}^{-\infty}(-1)^m\binom{m+n}{n}\binom{n+ \text{wt }u + t -1 }{m + n - j }\binom{j + t-1}{t-1} u_{-j - t} v .\]
\end{cor}
\begin{proof} From Eqn.\ \eqref{multiplication-formula-infinity} applied in this case, so that $i_j = 0$ for $1\leq j <t$, $i_t = i$, $p_s=\sum_{j=1}^{s}i_j = 0$ if $0\leq s <t$ and $p_t = i$, and $r=\sum_{j=1}^{t}(\text{wt }u+j-1)i_j =( \text{wt }u + t -1)i$, we have that for any $u,v \in V$, $t \in \mathbb{Z}_+$, and $i \in \mathbb{N}$
\begin{multline*}
u_{-t}^{i}{\bf{1}}*_nv = \sum_{m=0}^n \sum_{j=-m}^n (-1)^m\binom{m+n}{n}\binom{n + i (\mathrm{wt}\, u + t - 1) }{j+m} \\
\displaystyle\sum_{\substack{k_1,\dots ,k_{i}\in \mathbb{N}\\k_1+\cdots +k_{i}=n-j}} \! \! {}_{\circ}^{\circ}
\prod_{l=1}^{i} \binom{ k_l + t - 1}{t-1} u_{-k_{l}-t} {}_{\circ}^{\circ} v +g_{i}(v),
\end{multline*}
where $g_{i}(v)$ is a linear combination of elements of the form ${}_{\circ}^{\circ}u_{q_1}\cdots u_{q_{i}} {}_{\circ}^{\circ}v$ where at least one of the $q$'s is nonnegative, and the formula follows by shifting $j$ to $n - j$.
To prove the last statement of the Corollary, we note that if $i = 1$, then Eqn.\ (\ref{multiplication-cor-formula}) becomes
\begin{eqnarray*}
u_{-t} \mathbf{1} *_n v &= & \sum_{m=0}^n \sum_{j =0}^{m+n} (-1)^m \binom{m+n}{n} \binom{n + \mathrm{wt}\, u + t - 1 }{m + n - j} \binom{j + t - 1}{t-1} u_{-j -t} \, v
+ g_1(v) ,
\end{eqnarray*}
whereas from the first equation in the proof of Lemma \ref{multiplication-lemma}, or computing directly from the definition of $*_n$, we have
\begin{eqnarray*}
\lefteqn{ u_{-t} {\bf{1}}*_n v}\\
&=& \sum_{m=0}^n\sum_{j=-m}^{\infty}(-1)^m\binom{m+n}{n}\binom{n+ \text{wt }u + t -1 }{j+m}\binom{n-j + t-1}{t-1} u_{j - n - t} v\\
&=& \sum_{m=0}^n\sum_{j=m+n}^{-\infty}(-1)^m\binom{m+n}{n}\binom{n+ \text{wt }u + t -1 }{m + n - j }\binom{j + t-1}{t-1} u_{-j - t} v \\
&=& \sum_{m=0}^n \sum_{j =0}^{m+n} (-1)^m \binom{m+n}{n} \binom{n + \mathrm{wt}\, u + t - 1 }{m + n - j} \binom{j + t - 1}{t-1} u_{-j -t} \, v \\
& & \quad + \sum_{m=0}^n\sum_{j=-1}^{-\infty}(-1)^m\binom{m+n}{n}\binom{n+ \text{wt }u + t -1 }{m + n - j }\binom{j + t-1}{t-1} u_{-j - t} v ,
\end{eqnarray*}
where the last line is $g_1(v)$. Thus if $u_{k-t}v = 0$ for $k\in \mathbb{Z}_+$, then $g_1(v) = 0$. But in addition, the middle binomial term in $g_1(v)$ is zero if $n + \mbox{\rm wt}\ u + t - 1 \geq 0$ and $m+n - j > n + \mbox{\rm wt}\ u + t -1$. Thus if $n + \mbox{\rm wt}\ u + t - 1 \geq 0$ and $\mbox{\rm wt}\ u + t <2$, it also follows that $g_1(v) = 0$. If $n \geq \mbox{\rm wt}\ u + t \geq 2$, then letting $k = \mathrm{wt} \, u + t$, we have
\[g_1(v) = \sum_{m=0}^{k - 2} \sum_{j=-1}^{m -k + 1}(-1)^m\binom{m+n}{n}\binom{n+ k -1 }{m + n - j }\binom{j + t-1}{t-1} u_{-j - t} v . \]
This proves the last statement of the Corollary.
\end{proof}
Using this Corollary applied to the case that $v$ is a particular type of element in $A_n^u$, multiplied on the left by $u_{-1}^{i_1} {\bf 1}$, we have the following proposition:
\begin{prop}\label{reduce-prop}
Let $V$ be a vertex operator algebra. Fix $n \in \mathbb{Z}_+$, and suppose there is a $u \in V$ with $\mathrm{wt} \, u \geq -n$, and such that $V^u = \langle u \rangle^1$ satisfies the permutation property. Then for $i_1, i_2, \dots, i_{2n} \in \mathbb{N}$,
\begin{multline}\label{reduce1-prop}
\binom{i_1 \mathrm{wt} \, u +n}{n} \sum_{m=0}^n (-1)^m \binom{i_1 \mathrm{wt} \, u}{m} u_{-2n}^{i_{2n}} u_{-2 n + 1}^{i_{2n -1}} \cdots u_{ - 2}^{i_2 } u_{-1}^{i_1} \mathbf{1} \\
\equiv_n u_{-1}^{i_1} \mathbf{1} *_n u_{-2n}^{i_{2n}} u_{-2 n + 1}^{i_{2n -1}} \cdots u_{ - 2}^{i_2 }\mathbf{1} + w',
\end{multline}
where $w'$ is a linear combination of terms of the form $u_{-2n}^{i_{2n}'} \cdots u_{-1}^{i_1'} \mathbf{1}$ where either $ i_1' + \cdots + i_{2n}' < i_1 + \cdots + i_{2n}$, or $ i_1' + \cdots + i_{2n}' = i_1 + \cdots + i_{2n}$ and $i_1'< i_1$.
Furthermore, the coefficient of the term on the lefthand side of Eqn.\ (\ref{reduce1-prop}) is zero if and only if $0< | i_1 \mathrm{wt} \, u | \leq n$, and is equal to
\begin{equation}\label{reduce2-prop}
(-1)^n \binom{i_1 \mathrm{wt} \, u +n}{n} \binom{i_1 \mathrm{wt} \, u -1}{n}
\end{equation}
if $i_1 \mathrm{wt} \, u >n$.
\end{prop}
\begin{proof}
We apply Corollary \ref{multiplication-cor-special-case-more-general} to the case when
$v = u_{-2n}^{i_{2n}} u_{-2 n + 1}^{i_{2n -1}} \cdots u_{ - 2}^{i_2 }\mathbf{1} \in F_{i_2 + \cdots i_{2n}} ({\bf 1}) \subset V^u$, $t = 1$, and $i = i_1$. Thus Eqn.\ \eqref{multiplication-cor-formula} holds with $g_{i_1} (v) \in F_{i_1 + i_2 + \cdots i_{2n} -1} ( {\bf 1})$, and the only term in $u_{-1}^{i_1} \mathbf{1} *_n v$ that is not in $F_{i_1 + \cdots + i_{2n} -1}({\bf 1})$ and involves $u_{-2n}^{j_{2n}} u_{-2 n + 1}^{j_{2n -1}} \cdots u_{ - 1}^{j_1}\mathbf{1}$ with $j_1 \geq i_1$ is the one corresponding to the partition of $j = 0$ given by $\mathbf{p} = (0,0,\dots, 0)$. That is, we have
\begin{eqnarray*}
\lefteqn{u_{-1}^{i_1} \mathbf{1} *_n u_{-2n}^{i_{2n}} u_{-2 n + 1}^{i_{2n -1}} \cdots u_{ - 2}^{i_2 }\mathbf{1}}\\
&=& \sum_{m=0}^n (-1)^m \binom{m+n}{n} \binom{n + i_1 \mathrm{wt}\, u }{m + n } u_{-1}^{i_1} \, ( u_{-2n}^{i_{2n}} u_{-2 n + 1}^{i_{2n -1}} \cdots u_{ - 2}^{i_2 }\mathbf{1}) \\
& & \quad + \sum_{m=0}^n \sum_{j =1}^{m+n} \ \ \sum_{\stackrel{p_1, p_2, \dots, p_{i_1} \in \mathbb{N}}{p_1 + p_2 + \cdots + p_{i_1} = j} } (-1)^m \binom{m+n}{n} \binom{n + i_1 \mathrm{wt}\, u}{m + n - j} u_{-p_1-1} u_{-p_2-1} \cdots \\
& & \quad \cdots u_{-p_{i_1}-1} \, ( u_{-2n}^{i_{2n}} u_{-2 n + 1}^{i_{2n -1}} \cdots u_{ - 2}^{i_2 }\mathbf{1}) + f_{i_1 + \cdots i_{2n} -1}({\bf 1}) \\
&=& \sum_{m=0}^n (-1)^m \binom{m+n}{n} \binom{n + i_1 \mathrm{wt}\, u}{m + n } u_{-2n}^{i_{2n}} u_{-2 n + 1}^{i_{2n -1}} \cdots u_{ - 2}^{i_2 } u_{-1}^{i_1} \mathbf{1} + f'_{i_1 + i_2 + \cdots + i_{2n} - 1}( \bf 1) \\
& & \quad + \sum_{m=0}^n \sum_{j =1}^{m+n} \ \ \sum_{\stackrel{p_1, p_2, \dots, p_{i_1} \in \mathbb{N}}{p_1 + p_2 + \cdots + p_{i_1} = j} } (-1)^m \binom{m+n}{n} \binom{n + i_1 \mathrm{wt}\, u }{m + n - j} u_{-p_1-1} u_{-p_2-1} \cdots \\
& & \quad \cdots u_{-p_{i_1}-1} \, ( u_{-2n}^{i_{2n}} u_{-2 n + 1}^{i_{2n -1}} \cdots u_{ - 2}^{i_2 }\mathbf{1}) + f_{i_1 + i_2 + \cdots + i_{2n} - 1}( \bf 1)
\end{eqnarray*}
for some $f_{i_1+ \cdots + i_{2n} -1}({\bf 1}), f'_{i_1 + i_2 + \cdots + i_{2n} - 1}( \bf 1) \in F_{i_1 + \cdots + i_{2n} -1}({\bf 1})$. Therefore
\begin{eqnarray*}
\lefteqn{\sum_{m=0}^n (-1)^m \binom{m+n}{n} \binom{n+ i_1 \mathrm{wt}\, u}{m + n } u_{-2n}^{i_{2n}} u_{-2 n + 1}^{i_{2n -1}} \cdots u_{ - 2}^{i_2 } u_{-1}^{i_1} \mathbf{1} }\\
&=& u_{-1}^{i_1} \mathbf{1} *_n u_{-2n}^{i_{2n}} u_{-2 n + 1}^{i_{2n -1}} \cdots u_{ - 2}^{i_2 }\mathbf{1} \\
& & \quad - \sum_{m=0}^n \sum_{j =1}^{m+n} \ \ \sum_{\stackrel{p_1, p_2, \dots, p_{i_1} \in \mathbb{N}}{p_1 + p_2 + \cdots + p_{i_1} = j} } (-1)^m \binom{m+n}{n} \binom{n+ i_1 \mathrm{wt}\, u}{m + n - j} u_{-p_1-1} u_{-p_2-1} \cdots \\
& & \quad \cdots u_{-p_{i_1}-1} \, ( u_{-2n}^{i_{2n}} u_{-2 n + 1}^{i_{2n -1}} \cdots u_{ - 2}^{i_2 }\mathbf{1}) + h_{i_1+ \cdots + i_{2n}-1}({\bf 1}),\\
\end{eqnarray*}
where $h_{i_1+ \cdots + i_{2n}-1}({\bf 1}) \in F_{i_1+ \cdots + i_{2n}-1}({\bf 1})$.
Using the permutation property on the terms in the sum involving ordered partitions, and breaking out the partitions in the sum over ordered partitions that involve a $u_{-2n-1}$ term (namely those with $m = n$, $j = 2n$, and $\mathbf{p} = (p_1, p_2, \dots, p_{i_1})$ consisting of $i_1-1$ zeros, and one $2n$), we have
\begin{multline*}
\sum_{m=0}^n (-1)^m \binom{m+n}{n} \binom{n+ i_1 \mathrm{wt}\, u}{m + n } u_{-2n}^{i_{2n}} u_{-2 n + 1}^{i_{2n -1}} \cdots u_{ - 2}^{i_2 } u_{-1}^{i_1} \mathbf{1} \\
= u_{-1}^{i_1} \mathbf{1} *_n u_{-2n}^{i_{2n}} u_{-2 n + 1}^{i_{2n -1}} \cdots u_{ - 2}^{i_2 }\mathbf{1} - (-1)^n i_1 \frac{(2n)!}{n!n!} u_{-2n-1} u_{-2n}^{i_{2n}} u_{-2 n + 1}^{i_{2n -1}} \cdots u_{ - 2}^{i_2 } u_{-1}^{i_1 - 1} \mathbf{1} \\
+ w + h'_{i_1 + i_2 + \cdots + i_{2n} - 1}(\mathbf{1}) ,
\end{multline*}
for some $h'_{i_1 + i_2 + \cdots + i_{2n} - 1}(\mathbf{1}) \in F_{i_1 + i_2 + \cdots + i_{2n} - 1}(\mathbf{1})$, and
where $w$ is a linear combination of terms of the form $u_{-2n}^{i_{2n}'} \cdots u_{-1}^{i_1'} \mathbf{1}$ where $ i_1' + \cdots + i_{2n}' \leq i_1 + \cdots + i_{2n}$ but $i_1'< i_1$.
Since $\mathrm{wt} \, u \geq -n$, we can use Proposition \ref{recursion-n} to replace any terms which occur in $h'_{i_1+ \cdots + i_{2n}-1}({\bf 1})$ involving $u_{-k}$ for $k \geq 2n+2$ with terms involving sums of $u_{-j}$ for $1 \leq j \leq 2n+1$, and thus replace $h'_{i_1+ \cdots + i_{2n}-1}({\bf 1})$ with $h''_{i_1+ \cdots + i_{2n}-1}({\bf 1}) \in F_{i_1+ \cdots + i_{2n}-1}({\bf 1})$ which only involves $u_{-k}$ for $1 \leq k \leq 2n + 1$.
Using Theorem \ref{reducing-general-generators-thm}, and in particular repeatedly using Eqn.\ (\ref{reduce-equation}), we can rewrite the terms involving $u_{-2n-1}$ as linear combinations of terms involving only $u_{-k}$ for $1\leq k \leq 2n$ while not introducing any higher powers of $u_{-1}$ except for possibly in lower degree terms, meaning terms in $F_r (\mathbf{1})$ for $r< i_1 + \cdots + i_{2n}$. Therefore we have
\begin{multline}\label{reduce1}
\sum_{m=0}^n (-1)^m \binom{m+n}{n} \binom{n + i_1 \mathrm{wt}\, u}{m + n } u_{-2n}^{i_{2n}} u_{-2 n + 1}^{i_{2n -1}} \cdots u_{ - 2}^{i_2 } u_{-1}^{i_1} \mathbf{1} \\
\equiv_n u_{-1}^{i_1} \mathbf{1} *_n u_{-2n}^{i_{2n}} u_{-2 n + 1}^{i_{2n -1}} \cdots u_{ - 2}^{i_2 }\mathbf{1} + w',
\end{multline}
where $w'$ is a linear combination of terms of the form $u_{-2n}^{i_{2n}'} \cdots u_{-1}^{i_1'} \mathbf{1}$ where either $ i_1' + \cdots + i_{2n}' < i_1 + \cdots + i_{2n}$, or $ i_1' + \cdots + i_{2n}' = i_1 + \cdots + i_{2n}$ and $i_1'< i_1$.
To prove that the coefficient on the lefthand side of Eqn.\ (\ref{reduce1}) is that given as in Eqn.\ (\ref{reduce1-prop}), we let $j = i_1 \mathrm{wt} \, u$ and observe that
\begin{eqnarray}\label{coefficient-eqn}
C_j &=& \sum_{m=0}^n (-1)^m \binom{m+n}{n} \binom{j + n}{m + n } = \sum_{m=0}^n (-1)^m \frac{(m+n)!}{n!m!} \frac{(j + n)!}{(m+n)!(j-m)!} \\
&=& \sum_{m=0}^n (-1)^m \frac{(j+n)!}{n!m!(j-m)!} = \frac{(j+n)!}{n!j!} \sum_{m=0}^n (-1)^m \frac{j!}{m!(j-m)!} \nonumber \\
&=& \binom{j+n}{n} \sum_{m=0}^n (-1)^m \binom{j}{m} \nonumber .
\end{eqnarray}
Thus, if $0< j \leq n$
\begin{eqnarray*}
C_j = \binom{j+n}{n} \sum_{m=0}^n (-1)^m \binom{j}{m} x^m \Big|_{x = 1} = \binom{j+n}{n} (1-x)^j |_{x=1} = 0.
\end{eqnarray*}
If $-n\leq j <0$, $\binom{j+n}{n} = 0$, and thus $C_j = 0$.
If $j >n$, then noting that it follows from
\[ \binom{j}{m} = \binom{j-1}{m} + \binom{j-1}{m-1}\]
and induction on $j$ that
\[ \sum_{m=0}^n (-1)^m \binom{j}{m} = (-1)^n \binom{j-1}{n} \]
which gives
\[
C_j = (-1)^n \binom{j+n}{n} \binom{j-1}{n} \neq 0 \]
in this case, proving Eqn.\ (\ref{reduce2-prop}).
Finally if $j<-n$, then using the fact that $\binom{j}{m} = (-1)^m \binom{-j+m-1}{m}$, we have
\begin{eqnarray*}
C_j &=& \binom{j+n}{n} \sum_{m=0}^n (-1)^m \binom{j}{m} \ = \ \binom{j+n}{n} \sum_{m=0}^n \binom{-j+m -1}{m}
\end{eqnarray*}
which is always nonzero since $\binom{j+ n}{n} \neq 0$, and the summation consists only of positive terms when $j<-n$, proving the last statement of the Proposition.
\end{proof}
Proposition \ref{reduce-prop} allows us to reduce the generators further. In particular, we have the following refinement of Theorem \ref{reducing-general-generators-thm}:
\begin{thm}\label{last-generator-theorem}
Let $V$ be a vertex operator algebra. Fix $n \in \mathbb{Z}_+$, and suppose there is an element $u \in V$ with $\mathrm{wt} \, u \geq -n$, and such that the vertex subalgebra $V^u = \langle u \rangle^1 \subset V$ satisfies the permutation property. Let $A_n^u$ be the subalgebra of $A_n(V)$ generated by elements in $V^u$. Then $A_n^u$ is generated by elements of the form
\[ u_{-2n}^{i_{2n}} u_{-2 n + 1}^{i_{2n -1}} \cdots u_{ - 1}^{i_1}\mathbf{1}+ O_n(V) , \mbox{and} \ u_{-1}^{j} \mathbf{1} + O_n(V),\]
for $i_1, i_2, \dots, i_{2n} \in \mathbb{N}$, with $-n \leq i_1 \mathrm{wt} \, u \leq n$, and $j \in \mathbb{Z}_+$.
\end{thm}
\begin{proof} By Theorem \ref{reducing-general-generators-thm}, $A_n^u$ is generated by $u^{i_{2n}}_{-2n} u^{i_{2n-1}}_{-2n+1} \cdots u_{-1}^{i_1} \mathbf{1} + O_n(V)$ for $i_1, i_2, \dots, i_{2n} \in \mathbb{N}$. We proceed by induction on the following lexicographical ordering of such generators to show that each of these elements $u^{i_{2n}}_{-2n} u^{i_{2n-1}}_{-2n+1} \cdots u_{-1}^{i_1} \mathbf{1} + O_n(V)$ in $A_n^u$ can be generated by the elements prescribed in the statement of the present theorem. The lexicographical ordering is as follows:
\begin{multline}\label{ordering}
u^{i_{2n}'}_{-2n} u^{i_{2n-1}'}_{-2n+1} \cdots u_{-1}^{i_1'} \mathbf{1} + O_n(V) < u^{i_{2n}}_{-2n} u^{i_{2n-1}}_{-2n+1} \cdots u_{-1}^{i_1} \mathbf{1} + O_n(V) \\
\iff \ \left\{ \begin{array}{l}
i_{2n}' + i_{2n-1}' + \cdots + i_1' < i_{2n} + i_{2n-1} + \cdots + i_1 \\
\mbox{or}\\
i_{2n}' + i_{2n-1}' + \cdots + i_1' = i_{2n} + i_{2n-1} + \cdots + i_1
\ \mbox{and $i_1'<i_1$.}
\end{array} \right.
\end{multline}
First note that the theorem immediately holds for elements $ u^{i_{2n}}_{-2n} u^{i_{2n-1}}_{-2n+1} \cdots u_{-1}^{i_1} \mathbf{1} + O_n(V) \in A_n^u \subset A_n(V)$ of total degree $i_{2n} + i_{2n-1} + \cdots + i_1 = 0$ or $1$, as well as for any total degree as long as $i_1 = 0$, or $i_{2n} = i_{2n-1} = \cdots = i_2 = 0$.
We make the inductive assumption that the result holds for $u_{-2n}^{i_{2n}'} \cdots u_{-1}^{i_1'} \mathbf{1} + O_n(V)$ of total degree $i_{2n}' + \cdots + i_1'$, and we proceed to prove the statement holds for $u_{-2n}^{i_{2n}} \cdots u_{-1}^{i_1} \mathbf{1} + O_n(V)$ of total degree $i_{2n} + \cdots + i_1 = i_{2n}' + \cdots + i_1' + 1$. By Proposition \ref{reduce-prop}, the coefficient on the lefthand side of Eqn.\ (\ref{reduce1-prop}) is zero if and only if $0 < | i_1 \mathrm{wt} \, u | \leq n$. Thus as long as either $i_1 \mathrm{wt} \, u > n$ or
$i_1 \mathrm{wt} \, u <-n$, we have
\begin{multline}\label{reduce2}
u_{-2n}^{i_{2n}} u_{-2 n + 1}^{i_{2n -1}} \cdots u_{ - 2}^{i_2 } u_{-1}^{i_1} \mathbf{1} \\
\equiv_n \left( \binom{i_1 \mathrm{wt} \, u +n}{n} \sum_{m=0}^n (-1)^m \binom{i_1 \mathrm{wt} \, u}{m} \right)^{-1} u_{-1}^{i_1} \mathbf{1} *_n u_{-2n}^{i_{2n}} u_{-2 n + 1}^{i_{2n -1}} \cdots u_{ - 2}^{i_2 }\mathbf{1} + w',
\end{multline}
where $w'$ is a linear combination of terms involving only elements of the form $u_{-2n}^{i_{2n}'} \cdots u_{-1}^{i_1'} \mathbf{1}$ where either $ i_1' + \cdots + i_{2n}' < i_1 + \cdots + i_{2n}$, or $ i_1' + \cdots + i_{2n}' = i_1 + \cdots + i_{2n}$ and $i_1'< i_1$.
By the inductive assumption, the terms in (\ref{reduce2}) of total degree less than $i_{2n} + \cdots + i_1$ can be expressed in terms of the generators specified in the statement of the theorem. For the terms of equal total degree but with $i_1'<i_1$, we induct on $i_1'$. That is, we must show that the result holds for terms of the form $u_{-2n}^{i_{2n}'} \cdots u_{-1}^{i_1'} \mathbf{1}$ where $ i_1' + \cdots + i_{2n}' = i_1 + \cdots + i_{2n}$ and $i_1' <i_1$.
We consider the two cases: Case I. $-n \leq i_1' \mathrm{wt} \, u \leq n$; Case II. $i_1' > | \frac{n}{\mathrm{wt} \, u}|$.
The result holds if $i_1' \leq | \frac{n}{\mathrm{wt} \, u}|$, since then $-n\leq i_1' \mathrm{wt} \, u \leq n$, and so $u_{-2n}^{i_{2n}'} \cdots u_{-1}^{i_1'} \mathbf{1}$ is in the list of generators specified by the theorem. This includes the base case for our induction on $i_1'$. We make the inductive assumption that the result holds for $i_1'$ and prove the result for $i_1'+ 1 > | \frac{n}{\mathrm{wt} \, u}|$. In this case, by Proposition \ref{reduce-prop}, we have
\begin{multline}\label{reduce3}
u_{-2n}^{i_{2n}'} u_{-2 n + 1}^{i_{2n -1}'} \cdots u_{ - 2}^{i_2' } u_{-1}^{i_1' + 1} \mathbf{1} \\
\equiv_n \left( \binom{(i_1'+1) \mathrm{wt} \, u +n}{n} \sum_{m=0}^n (-1)^m \binom{(i_1'+1) \mathrm{wt} \, u}{m} \right)^{-1} u_{-1}^{i_1' + 1} \mathbf{1} *_n u_{-2n}^{i_{2n}'} u_{-2 n + 1}^{i_{2n -1}'} \cdots u_{ - 2}^{i_2' }\mathbf{1} + w'',
\end{multline}
where $w''$ is a linear combination of terms of the form $u_{-2n}^{i_{2n}''} \cdots u_{-1}^{i_1''} \mathbf{1}$ where either $ i_1'' + \cdots + i_{2n}'' < ( i_1' + 1) + i_2' + \cdots + i_{2n}'$, or $ i_1'' + \cdots + i_{2n}'' = (i_1' + 1) + i_2' + \cdots + i_{2n}'$ and $i_1''< i_1'$. By the inductive assumptions on the total degree, and on $i_1'$, the result follows.
\end{proof}
\section{On the definition of the higher level Zhu algebras and applications}\label{definition-clarification-section}
In this section, we prove a result that helps motivate the definition of $O_n(V)$ when $n \in \mathbb{Z}_+$, namely that one must define $O_n(V)$ to include all elements of the form $(L(-1) + L(0))v$ for $v \in V$. In addition, as we note in Remark \ref{OL-remark} below, this theorem also singles out some of the relations that must hold in $A_n(V)$ that are contained in $O^L(V)$, but not necessarily in $O_n^\circ(V)$. Thus the theorem is useful in determining the structure of $A_n(V)$.
\begin{thm}\label{clarify-definition-thm}
Suppose that $V$ is a vertex operator algebra that is strongly generated by $u \in V$. Then
\[O_n^L(V)\not\subset O_n^\circ(V)\]
as long as $n \in \mathbb{Z}_+$ and one of the following hold:
(i) $1\leq \mathrm{wt} \, u \leq n$;
or
(ii) $\mathrm{wt} \, u = 0$, and $u_{-2} \mathbf{1} \neq 0$.
\end{thm}
\begin{proof} Note that if $V$ is a vertex operator algebra, then for any $u \in V$, we have that $(L(0)+L(-1)) u = (L(0)+L(-1)) u_{-1}{\bf{1}}= (\mathrm{wt} \, u) u_{-1}{\bf{1}}+ u_{-2}{\bf{1}}$. Thus if $\mathrm{wt} \, u \neq 0$ then $(L(0)+L(-1)) u$ is a linear combination of vectors in $V$ that contains a scalar multiple of $u=u_{-1}{\bf{1}}$. We will show that for all $n>0$, if one of the two cases, (i) or (ii) hold, then the vector $(\mathrm{wt} \, u) u_{-1}{\bf{1}}+ u_{-2}{\bf{1}} \notin O_n^\circ(V)$, by showing that for any $v,w \in V$ the expression $v \circ_n w$ written as a linear combination of vectors in $V$ either cannot contain a nonzero scalar multiple of $u = u_{-1}{\bf{1}}$ or cannot contain a nonzero scalar multiple of $u_{-2} \mathbf{1}$.
Since $V$ is strongly generated by $u$, any $v \in V$ can be written as a linear combination of elements of the form $u_{-k_1} u_{-k_2 } \cdots u_{ - k_m} \mathbf{1}$, for $m \in \mathbb{N}$, and $k_1, \dots, k_m \in \mathbb{Z}_+$. Thus it is sufficient to consider $v \circ_n w$ of the form $u_{-k_1} u_{-k_2 } \cdots u_{ - k_m} \mathbf{1} \circ_n w$.
If $m = 0$, we have $\mathbf{1} \circ_n w = \mbox{\rm Res}_x\frac{(1+x)^{n}}{x^{2n+2}} w = 0$. If $m >0$, we
let
\begin{equation}\label{define-m}
t = \sum_{r=1}^{m}(\mathrm{wt} \, u + k_r -1), \quad \mbox{and} \quad s = \sum_{r = 1}^{m} k_r , \quad \mbox{and note that $t = (\mathrm{wt} \, u -1)m + s$.}
\end{equation}
If $\mathrm{wt} \, u \geq 0$, then $\mathrm{wt} \, (u_{-k_1} u_{-k_2 } \cdots u_{ - k_m} \mathbf{1} )= t \geq 0$, and we have
\begin{eqnarray*}
\lefteqn{ u_{-k_1} u_{-k_2 } \cdots u_{ - k_m} \mathbf{1} \circ_n w}\\
&= & \mbox{\rm Res}_x\frac{(1+x)^{t+n}}{x^{2n+2}} \, {}_{\circ}^{\circ} \, (Y(u_{-k_1}{\bf{1}}, x)) \cdots (Y(u_{-k_m}{\bf{1}}, x)) \, {}_{\circ}^{\circ} \, w\\
&=& \mbox{\rm Res}_x \sum_{l = 0}^{t+n} \binom{t +n }{l}x^{l-2n-2} \, {}_{\circ}^{\circ} \, \left( \frac{1}{(k_1-1)!} \left(\frac{d}{dx}\right)^{k_1 -1}\sum_{j \in \mathbb{Z}} u_{-j}x^{j-1} \right)\cdot\\
& & \quad \left( \frac{1}{(k_2-1)!} \left(\frac{d}{dx}\right)^{k_2-1}\sum_{j \in \mathbb{Z}} u_{-j}x^{j-1} \right) \cdots
\left( \frac{1}{(k_m - 1)!} \left(\frac{d}{dx}\right)^{k_m - 1} \sum_{j \in \mathbb{Z}} u_{-j}x^{j-1}\right) \, {}_{\circ}^{\circ} \, w\\
&=& \mbox{\rm Res}_x \sum_{l =0 }^{t+n} \binom{t+n }{l}x^{l-2n-2} \, {}_{\circ}^{\circ} \, \left(\sum_{j\in \mathbb{Z}}\binom{j-1}{k_1-1} u_{-j}x^{j-k_1}\right) \left( \sum_{j \in \mathbb{Z}} \binom{j-1}{k_2 - 1} u_{-j}x^{j-k_2} \right) \cdot \\
& &
\cdots \left( \sum_{j \in \mathbb{Z}} \binom{j-1}{k_m - 1} u_{-j}x^{j-k_m} \right) \, {}_{\circ}^{\circ} \, w\\
&=& \mbox{\rm Res}_x \sum_{l = 0}^{t+n} \binom{t+n }{l}x^{l-2n-2-s} \, {}_{\circ}^{\circ} \, \left(\sum_{j\in \mathbb{Z}}\binom{j-1}{k_1 - 1} u_{-j}x^{j}\right) \left( \sum_{j \in \mathbb{Z}} \binom{j-1}{k_2 - 1} u_{-j}x^{j} \right) \cdot \\
& & \cdots \left( \sum_{j \in \mathbb{Z}} \binom{j-1}{k_m -1 } u_{-j}x^{j} \right) \, {}_{\circ}^{\circ} \, w.
\end{eqnarray*}
Each term in this expression is a scalar multiple of something of the form
\begin{equation}\label{terms}
{}_{\circ}^{\circ} \, u_{j_1}\cdots u_{j_m}\, {}_{\circ}^{\circ} \, w,
\end{equation}
where the normal ordering is given by
\[ {}_{\circ}^{\circ} \, u_j u_ k \, {}_{\circ}^{\circ} \, = \left\{ \begin{array}{ll}
u_j u_k & \mbox{if $j<0$}\\
u_k u_j & \mbox{if $j \geq 0$}
\end{array}
\right. .\]
After taking the residue above, the only terms appearing in $u_{-k_1} u_{-k_2} \cdots u_{ - k_m} \mathbf{1} \circ_n w$ are ones of this form (\ref{terms}) where
\begin{equation*}
-j_1-\cdots-j_m +l-2n-2- s =-1
\end{equation*}
or equivalently,
\begin{equation}\label{equation-for-sum}
j_1+\cdots+j_m =l-2n-1-s
\end{equation}
with $0\leq l \leq n+t$, and $t$ and $s$ as in (\ref{define-m}). That is, we have
\begin{equation}\label{equation-for-sum2}
-2n-1-s \leq j_1+\cdots+j_m \leq t-n-1-s = (\mathrm{wt} \, u - 1)m -n - 1 .
\end{equation}
Since $w \in V = \langle u \rangle^1$, we have that $w$ is a linear combination of elements of the form $u_{-p_1} \cdots u_{-p_q} \mathbf{1}$,
for some $p_1, \dots, p_q \in \mathbb{Z}_+$ and $q \in \mathbb{N}$, and
\[ \mathrm{wt} \, u_{-p_1} \cdots u_{-p_q} \mathbf{1} = q \, \mathrm{wt} \, u + p_1 + \cdots + p_q - q.\]
For one of the terms (\ref{terms}) to be a nonzero scalar multiple of $u_{-1} \mathbf{1} = u$, which is homogeneous of weight $\mathrm{wt}\, u$, we would have to have some linear component satisfying
\begin{eqnarray*}
\mathrm{wt} \, u &=& \mathrm{wt} \, {}_{\circ}^{\circ} \, u_{j_1}\cdots u_{j_m}\, {}_{\circ}^{\circ} \, u_{-p_1} \cdots u_{-p_q} \mathbf{1} \\
&=& (m + q) \mathrm{wt} \, u - j_1 - \cdots - j_m + p_1 + \cdots + p_q - m - q,
\end{eqnarray*}
that is
\begin{equation}\label{weights}
j_1+\cdots+j_m = (m + q -1) \mathrm{wt} \, u + p_1 + \cdots + p_q - m - q.
\end{equation}
This with Eqn. (\ref{equation-for-sum2}) would imply
\begin{equation}\label{weights2}
-2n-1-s \leq (m + q -1) \mathrm{wt} \, u + p_1 + \cdots + p_q - m - q \leq (\mathrm{wt} \, u - 1)m -n - 1
\end{equation}
or equivalently
\begin{equation}\label{weights3}
-2n-s - (\mathrm{wt} \, u - 1) m \leq (q -1) \, \mathrm{wt} \, u + p_1 + \cdots + p_q - q + 1 \leq -n <0 .
\end{equation}
If $q>0$, then since $p_1 + \cdots + p_q - q \geq 0$, and $(q-1) \, \mathrm{wt} \, u \geq 0$ when $q>0$, we have that
\[ (q -1) \mathrm{wt} \, u + p_1 + \cdots + p_q - q + 1 \geq 1\]
which can not be less than or equal to $-n$ in Eqn.\ (\ref{weights3}). Thus there exist no such combination of $p_1, \dots, p_q$ with $q>0$ such that $u_{-1} \mathbf{1}$ appears as a linear component of $v \circ_n w$.
If $q = 0$, i.e. if $w = \mathbf{1}$, then Eqn.\ (\ref{weights3}) becomes
\begin{equation}\label{weights4}
-2n-s - (\mathrm{wt} \, u - 1) m \leq - \mathrm{wt} \, u + 1 \leq -n <0,
\end{equation}
which is possible if and only if $\mathrm{wt} \, u \geq n + 1$, proving that $(\mathrm{wt} \, u) u_{-1}{\bf{1}}+ u_{-2} {\bf{1}}\notin O_n^\circ(V)$ if $1 \leq \mathrm{wt} \, u \leq n$.
If $\mathrm{wt} \, u = 0$, then as previously noted $(L(0)+L(-1)) u_{-1}{\bf{1}}= (\mathrm{wt} \, u) u_{-1}{\bf{1}}+ u_{-2}{\bf{1}} = u_{-2} \mathbf{1}$, and thus in this case we have to show that $u_{-2} \mathbf{1} \notin O_n^\circ(V)$, which in particular, imposes the condition that $u_{-2} \mathbf{1} \neq 0$. In this case, for one of the terms (\ref{terms}) to be a nonzero scalar multiple of $u_{-2}{\bf{1}}$, we would have to have some $u_{-p_1} \cdots u_{-p_q} \mathbf{1}$ such that
\begin{eqnarray*}
\mathrm{wt} \, u_{-2} \mathbf{1} &=& \mathrm{wt} \, {}_{\circ}^{\circ} \, u_{j_1}\cdots u_{j_m}\, {}_{\circ}^{\circ} \, w \\
&=& (m + q) \mathrm{wt} \, u - j_1 - \cdots - j_m + p_1 + \cdots + p_q - m - q,
\end{eqnarray*}
that is
\begin{equation}\label{weights-0}
j_1+\cdots+j_m = p_1 + \cdots + p_q - m - q -1 .
\end{equation}
This with Eqn. (\ref{equation-for-sum2}) would imply
\begin{equation}\label{weights2-0}
-2n-1-s \leq p_1 + \cdots + p_q - m - q - 1 \leq -m -n - 1
\end{equation}
or equivalently
\begin{equation}\label{weights3-0}
-2n-s + m \leq p_1 + \cdots + p_q - q \leq -n <0
\end{equation}
However, since $p_1 + \cdots + p_q - q \geq 0$, this is impossible, proving that if $\mathrm{wt} \, u = 0$, no $v,w \in V$ exist such that $v\circ_n w$ has a term that is a nonzero scalar multiple of $u = u_{-2}{\bf 1}$. This proves that $(\mathrm{wt} \, u) u_{-1}{\bf{1}}+ u_{-2} {\bf{1}}\notin O_n^\circ(V)$ if $\mathrm{wt} \, u = 0$, and $u_{-2} \mathbf{1} \neq 0$.
\end{proof}
\begin{rem}\label{OL-remark}
{\em As stated, Theorem \ref{clarify-definition-thm} applies directly to vertex operator algebras strongly generated by a vector of weight greater than or equal to one, e.g. the rank one Heisenberg vertex operator algebra, or the Virasoro vertex operator algebras, or indirectly for such positive weight elements $u \in V$ where there is not linear dependence between $(L(-1) + L(0))u$ and circle products of vectors in $V$ not generated by $u$, in particular when the level $n$ of the Zhu algebra in consideration becomes greater than or equal to the weight of the vector being considered. The Theorem also pertains to the case when $V$ has a nontrivial vacuum (weight zero) space containing any vectors $u$ such that $u_{-2} {\bf 1} \neq 0$. Thus there are common settings in which a vertex operator algebra $V$ has the property that $O^L (V) \not\subset O_n^\circ(V)$. In practice, the theorem implies that when determining $A_n(V)$ for $n \in \mathbb{Z}_+$ and for $V$ any vertex operator algebra, one should at least single out vectors of positive weight less than the level $n$ being considered, any nontrivial weight zero vectors, as well as potentially other vectors, and add in these $(L(-1) + L(0))u \approx 0$ relations that may not arise from $O_n^\circ(V)$ but nevertheless arise from $O^L(V) \subset O_n(V)$. That is, one should expect to use the fact that these can give nontrivial extra relations in $A_n(V)$ when determining a minimal set of generators and relations of $A_n(V)$. For instance, the presence of these $(L(-1) + L(0)) v \in
O_n(V)$ is used heavily in the constructions of the higher level Zhu algebras in \cite{BVY-Heisenberg}, \cite{BVY-Virasoro}, and \cite{AB-details-of-Heisenberg}. }
\end{rem}
\section{Using $A_{n-1}(V)$ to calculate $A_n(V)$}\label{technique-section}
In this Section, we provide some further techniques which have proved helpful in calculating higher level Zhu algebras once a reasonable set of generators has been determined by using the techniques of the previous sections. The main two techniques in calculating $A_n(V)$, assuming $A_{n-1}(V)$ is already known, are: I. Make use of the relationship between these two algebras which imposes restrictions on the possible relations between the generators of $A_n(V)$ in terms of the relations of these generators in $A_{n-1}(V)$; II. Make use of the action of the zero modes of the generators of $A_n(V)$ on the $V$-modules already determined by $A_{n-1}(V)$ to help further determine relations amongst the generators in $A_n(V)$.
We present two ``Theorems" below, which gather together the pertinent facts or general principles that allow Techniques I and II to be of use. For Technique I, we note the following:
\begin{thm}\label{use-lower-A-thm}
Suppose that $S_n = \{v_j + O_n(V)\}_{j \in J}$ for $v_j \in V$ is a set of generators for $A_{n}(V)$ and let $R= \mathbb{C}\langle x_j \; | \; j \in J \rangle$ denote the free algebra generated by the formal variables $x_j$ for $j \in J$. Then
(i) $S_{n-1}$ is a generating set for $A_{n-1}(V)$.
(ii) There exists a two-sided ideal $I_{n-1}$ of $R$ such that $A_{n-1}(V) \cong R/I_{n-1}$.
(iii) There exists a two-sided ideal $I_n$ of $R$ such that $A_n(V) \cong R/I_n$.
(iv) $I_n \subset I_{n-1}$.
\end{thm}
\begin{proof} From Proposition 2.4 in \cite{DLM}, we have the surjective homomorphism given by Eqn.\ (\ref{surjection}), which follows from the fact that $O_n(V) \subset O_{n-1}(V)$. If $S_n$ generates $A_n(V)$, then there is a surjective homomorphism from $R$ onto $A_n(V)$ with $x_j \mapsto v_j + O_n(V)$ whose kernel is a two-sided ideal $I_n$, giving $A_n(V) \cong R/I_n$. Then the composition of the surjection of $R$ onto $A_n(V)$ and the surjection of $A_n(V)$ onto $A_{n-1}(V)$ gives a surjective homomorphism from $R$ onto $A_{n-1}(V)$ with $x_j \mapsto v_j + O_{n-1}(V)$ whose kernel is $I_{n-1}$, so that $A_{n-1}(V) \cong R/I_{n-1}$, and such that $I_n \subset I_{n-1}$.
\end{proof}
For Technique II, we note the following Theorem
\begin{thm}\label{use-zero-mode-thm}
Suppose that $S_n = \{v_j +O_n(V)\}_{j \in J} \subset A_n(V)$ is a set of generators for $A_{n}(V)$ and let $R= \mathbb{C}\langle x_j \; | \; j \in J \rangle$ denote the free algebra generated by the formal variables $x_j$ for $j \in J$. If $W = L_{n-1}(U)$ is a $V$-module induced at level $n-1$ from the $A_{n-1}(V)$-module $U$, $P(x_j \; | \; j \in J)$ is a word in $R$, and $P(v_j + O_n(V) \; | \; j \in J) \in I_n$ where $A_n \cong R/I_n$, then $o(P).W(k) = 0$ for $k = 0, \dots, n$.
\end{thm}
\begin{proof} Recall from Section 2, or \cite{DLM}, that for any $\mathbb{N}$-gradable $V$-module $W$, the subspace $\Omega_n(W)$ which contains $\coprod_{k = 0}^n W(k)$ (see \cite{BVY}), is an $A_n(V)$-module with action given via zero modes, i.e. $(v + O_n(V)).w = o(v)w = v_{\mathrm{wt} \, v - 1}w$, where $v_{\mathrm{wt}, v - 1}$ acting on $W$ is defined via the $V$-module structure of $W$.
Thus if $P \in I_n$, then $o(P)$ acts trivially on any $A_n(V)$-module, in particular on any $\coprod_{k = 0}^n W(k)$ for $W$ a $V$-module induced from $A_{n-1}(V)$.
\end{proof}
In particular, if a relation $P$ can be found or guessed and proved to hold up to lower order terms in the total degree of the words in the generators $S$, then often acting via the zero modes of the generators, one can establish what the lower degree terms of $P$ are exactly. In addition, using the fact that $P \in I_n \subset I_{n-1}$ as established in Theorem \ref{use-lower-A-thm}, allows one to further limit the possible form of these lower degree terms, thus helping establish the relation.
We illustrate these techniques in detail in \cite{AB-details-of-Heisenberg} and in the following section by giving some of the results which were used in \cite{AB-details-of-Heisenberg} to calculate the level two Zhu algebra for the Heisenberg vertex operator algebra using the relations of the level two generators inside the level one Zhu algebra for the Heisenberg vertex operator algebra which was calculated in \cite{BVY-Heisenberg}.
\section{Applications}
The applications of the results of the previous sections are wide ranging. In particular they apply to the subalgebra of $A_n(V)$ generated by the conformal element $\omega$ of $V$, for any vertex operator algebra $V$, since $\langle \omega \rangle$ is a strongly generated vertex operator algebra. They also apply to any elements $u \in V$ that generate a Heisenberg vertex subalgebra. We discuss these two settings below.
\subsection{The Heisenberg vertex operator algebra}\label{Heisenberg-section}
We denote by $\mathfrak{h}$ a one-dimensional abelian Lie algebra spanned by $\alpha$ with a bilinear form $\langle \cdot, \cdot \rangle$ such that $\langle \alpha, \alpha \rangle = 1$, and by
\[
\hat{\mathfrak{h}} = \mathfrak{h}\otimes \mathbb{C}[t, t^{-1}] \oplus \mathbb{C} \mathbf{k}
\]
the affinization of $\mathfrak{h}$ with bracket relations
\[
[a(m), b(n)] = m\langle a, b\rangle\delta_{m+n, 0}\mathbf{k}, \;\;\; a, b \in \mathfrak{h},
\]
\[
[\mathbf{k}, a(m)] = 0,
\]
where we define $a(m) = a \otimes t^m$ for $m \in \mathbb{Z}$ and $a \in \mathfrak{h}$.
Set
\[
\hat{\mathfrak{h}}^{+} =\mathfrak{h} \otimes t\mathbb{C}[t] \qquad \mbox{and} \qquad \hat{\mathfrak{h}}^{-} = \mathfrak{h} \otimes t^{-1}\mathbb{C}[t^{-1}].
\]
Then $\hat{\mathfrak{h}}^{+}$ and $\hat{\mathfrak{h}}^{-}$ are abelian subalgebras of $\hat{\mathfrak{h}}$. Consider the induced $\hat{\mathfrak{h}}$-module given by
\[
M(1) = U(\hat{\mathfrak{h}})\otimes_{U(\mathbb{C}[t]\otimes \mathfrak{h} \oplus \mathbb{C} c)} \mathbb{C}{\bf 1} \simeq S(\hat{\mathfrak{h}}^{-}) \qquad \mbox{(linearly)},
\]
where $U(\cdot)$ and $S(\cdot)$ denote the universal enveloping algebra and symmetric algebra, respectively, $\mathfrak{h} \otimes \mathbb{C}[t]$ acts trivially on $\mathbb{C}\mathbf{1}$ and $\mathbf{k}$ acts as multiplication by $1$. Then $M(1)$ is a vertex operator algebra, often called the {\it vertex operator algebra associated to the rank one Heisenberg}, or the {\it rank one Heisenberg vertex operator algebra}, or the {\it one free boson vertex operator algebra}. Here, the Heisenberg Lie algebra in question being precisely $\hat{\mathfrak{h}} \diagdown \mathbb{C}\alpha(0)$.
Any element of $M(1)$ can be expressed as a linear combination of elements of the form
\begin{equation}\label{generators-for-V}
\alpha(-k_1)\cdots \alpha(-k_j){\bf 1}, \quad \mbox{with} \quad k_1 \geq \cdots \geq k_j \geq 1.
\end{equation}
In other words, $M(1)$ is strongly generated by $\alpha(-1) \mathbf{1}$ implying that most of the results of this paper pertain to $M(1)$.
The conformal element for $M(1)$ is given by $\omega = \frac{1}{2} \alpha(-1)^2{\bf 1}$
However the element $\omega_a = \frac{1}{2} \alpha(-1)^2{\bf 1} + a \alpha(-2) \mathbf{1}$
for any $a \in \mathbb{C}$, also gives $M(1)$ with the conformal element $\omega_a$ the structure of a vertex operator algebra with central charge $c = 1 - 12 a^2$. We denote this vertex operator algebra with shifted conformal elements by $M_a(1)$. So, for instance $M(1) = M_0(1)$.
Next, we recall the construction of $A_0(M_a(1))$ from \cite{FZ} and the construction of $A_1(M_a(1))$ from \cite{BVY}:
\begin{thm}\label{A_0-theorem} \cite{FZ} As algebras
\[A_0(M_a(1)) \cong \mathbb{C}[x,y]/(x^2-y) \cong \mathbb{C}[x]\]
under the indentifications
\[\alpha(-1)\mathbf{1} + O_0(M_a(1)) \longleftrightarrow x, \qquad \mbox{and} \qquad \alpha(-1)^2\mathbf{1} + O_0(M_a(1)) \longleftrightarrow y.\]
In addition, there is a bijection between isomorphism classes of irreducible admissible $M_a(1)$-modules and irreducible $\mathbb{C}[x]$-modules given by $M_a(1,\lambda) \longleftrightarrow \mathbb{C}[x]/(x - \lambda)$.
\end{thm}
\begin{thm}\label{A_1-theorem}\cite{BVY-Heisenberg}
Let $I$ be the ideal generated by the polynomial $(x^2-y)(x^2-y+2)$ in $\mathbb{C}[x, y]$.
Then we have the following isomorphisms of algebras
\begin{eqnarray}\label{first-characterization-level-one}
A_1(M_a(1)) &\cong& \mathbb{C}[x, y]/I \ \cong \ \mathbb{C}[x,y]/(x^2 - y) \oplus \mathbb{C}[x,y]/(x^2 - y + 2) \\
& \cong & A_0(M_a(1)) \oplus \mathbb{C}[x,y]/(x^2 - y+ 2) \ \cong \ \mathbb{C}[x] \oplus \mathbb{C} [x] \label{second-characterization}
\end{eqnarray}
under the identifications
\[
\alpha(-1){\bf 1} + O_1(M_a(1)) \longleftrightarrow x + I, \quad \qquad \alpha(-1)^2{\bf 1} + O_1(M_a(1)) \longleftrightarrow y + I.
\]
\end{thm}
More recently, using results from this paper, we prove in \cite{AB-details-of-Heisenberg} the following:
\begin{thm}\label{main-generators-thm}\cite{AB-details-of-Heisenberg}
The algebra $A_2(M_a(1))$ is generated by $v + O_2(M_a(1))$ for $v$ in
\[S = \{ \alpha(-1) {\bf 1}, \, \alpha(-1)^2 {\bf 1}, \, \alpha(-1) \alpha(-4) {\bf 1}, \, \alpha(-1)^2\alpha(-4){\bf 1}, \, \alpha(-1)\alpha(-4)^2{\bf 1} \}.\]
\end{thm}
We introduce the notation:
\begin{eqnarray}
x_n &=& \alpha(-1) \mathbf{1} + O_n(M_a(1)) \\
y_n &=& \alpha(-1)^2 \mathbf{1} + O_n(M_a(1)) \\
\tilde y_n &=& \alpha(-1)\alpha(-4) \mathbf{1} + O_n(M_a(1)) \\
z_n &=& \alpha(-1)^2 \alpha(-4) \mathbf{1} + O_n(M_a(1)) \\
\tilde z_n &=& \alpha(-1)\alpha(-4)^2 \mathbf{1} + O_n(M_a(1)).
\end{eqnarray}
Using Technique I of Section \ref{technique-section}, Theorem \ref{use-lower-A-thm}, and results of the proof of Theorem \ref{A_1-theorem} in \cite{BVY-Heisenberg}, in \cite{AB-details-of-Heisenberg} we prove
\begin{prop}\label{use-level-one-prop}\cite{AB-details-of-Heisenberg}
For $V = M_a(1)$, and $R_1 = \mathbb{C}[x_1,y_1]\langle \tilde y_1, z_1, \tilde z_1\rangle$,
we have
\[A_1(V) \cong R_1/I_1 \]
where $I_1$ is the two-sided ideal
\begin{equation}\label{I_1}
I_1 = \big( (x_1^2 -y_1)(x_1^2 - y_1 + 2) , (x_1^2 -2y_1 - \tilde{y}_1), (4x_1^3 -5x_1y_1 - z_1), \\
(3x_1^3 - 4x_1y_1 + \tilde{z}_1) \big).
\end{equation}
\end{prop}
Motivated by this Proposition \ref{use-level-one-prop}, and the action of the zero modes for the generating set $S$ for $A_2(M_a(1))$ on the irreducible modules already determined by $A_1(M_a(1))$ as proscribed by Technique II and Theorem \ref{use-zero-mode-thm}, we make the change of variables:
\begin{eqnarray}\label{change-var1-n}
Y_n &=& \frac{1}{12} (x_n^2 - 2y_n - \tilde y_n)\\
Z_n &=& \frac{1}{32} ( x_n^3 + 2x_n\tilde y_n + \tilde z_n) \\
&=& \frac{1}{32} \left( (3x_n^3 - 4x_ny_n + \tilde z_n) - 2x_n(x_n^2 - 2y_n - \tilde y_n) \right), \\
\ \ \ W_n &=& -\frac{1}{40} \left(2z_n + \tilde z_n + 2x_ny_n - 2x_n \tilde y_n - 3x_n^3 \right) \label{change-var3-n} \\
&=& - \frac{1}{40} \left( 2x_n(x_n^2 - 2y_n - \tilde y_n)- 2(4x_n^3 - 5x_ny_n - z_n) + (3x_n^3 - 4x_n y_n + \tilde z_n)\right) \nonumber,
\end{eqnarray}
so that $R_n = \mathbb{C}[x_n,y_n]\langle \tilde y_n, z_n, \tilde z_n \rangle = \mathbb{C}[x_n,y_n]\langle Y_n, Z_n, W_n \rangle$
With these change of variables , in \cite{AB-details-of-Heisenberg} we note the following Corollary to Proposition \ref{use-level-one-prop}
\begin{cor}\label{level-one-cor}\cite{AB-details-of-Heisenberg}
For $V = M_a(1)$ and $R = \mathbb{C}[x,y]\langle Y, Z, W \rangle$ we have $A_1(V) \cong R/I_1$ where $I_1$ is the two-sided ideal
\begin{equation}
I_1 = \big( (x_1^2 - y_1)(x_1^2 - y_1 + 2), Y_1, Z_1, W_1 \big).
\end{equation}
\end{cor}
Then in \cite{AB-details-of-Heisenberg}, motivated by the zero-mode action of these generators on the already determined irreducible $M_a(1)$-modules, we prove relations for the generators of $A_2(M_a(1))$ up to lower order terms in $F_r(\mathbf{1})$ for $r$ less than the top degree of the conjectured relations. In particular we prove
\begin{lem}\label{lower-order-lemY} \cite{AB-details-of-Heisenberg}
\begin{eqnarray}
(x_2^2 - y_2) Y_2, \ Y_2^2 & \in & O_2(V) + F_3(\mathbf{1}) , \\
(x_2^2 - y_2)Z_2, \ (x_2^2 - y_2)W_2, \ Z_2Y_2, \ Y_2W_2 & \in & O_2(V) + F_4(\mathbf{1}), \\
(x_2^2 - y_2)^3, \ Z_2^2, \ W_2^2 &\in & O_2(V) + F_5(\mathbf{1}).
\end{eqnarray}
\end{lem}
Then using the zero-mode action of these generators on the irreducible $M_a(1)$-modules, we prove
\begin{thm}\label{A_2-theorem}\cite{AB-details-of-Heisenberg}
Let $\mathbb{C}[x, y]\langle Y, Z, W \rangle$ denote the algebra generated over $\mathbb{C}$ by the two commuting variables $x$ and $y$, and three non-commuting variables $Y, Z$, and $W$. Let $I_2$ be the ideal generated by the polynomials
\begin{eqnarray}
\ \ \ \ & (x^2-y)(x^2-y+2)(x^2-y+4), \ \ (x^2 - y + 4)Y, \ \ (x^2 - y + 4)Z, \ \ (x^2 - y + 4)W, &\\
&Y^2 - Y, \ \ Z^2, \ \ W^2, \ \ ZY, \ \ YW, \ \ ZW - Y, & \\
& YZ - Z, \ \ WY - W, \ \ Y + WZ - \frac{1}{8}(x^2 - y)(x^2 - y + 2) .&
\end{eqnarray}
Then we have the following isomorphism of algebras
\begin{equation}
A_2(M_a(1)) \cong \mathbb{C}[x, y]\langle Y, Z, W \rangle/I_2 \\
\end{equation}
under the identification
\begin{equation}\label{identification1}
\alpha(-1){\bf 1} + O_2(M_a(1)) \longleftrightarrow x + I_2, \quad \alpha(-1)^2{\bf 1} + O_2(M_a(1)) \longleftrightarrow y + I_2,
\end{equation}
\begin{equation}
\alpha(-1)\alpha(-4){\bf 1} + O_2(M_a(1)) \longleftrightarrow \tilde y + I_2, \quad \alpha(-1)^2\alpha(-4){\bf 1} + O_2(M_a(1)) \longleftrightarrow z + I_2,
\end{equation}
\begin{equation}
\alpha(-1)\alpha(-4)^2{\bf 1} + O_2(M_a(1)) \longleftrightarrow \tilde z + I_2. \label{identification3}
\end{equation}
and the change of variables
\begin{equation}
\ Y = \frac{1}{12} (x^2 - 2y - \tilde y), \ \ Z =\frac{1}{32} ( x^3 + 2x\tilde y + \tilde z), \ \ W = -\frac{1}{40} \left(2z + \tilde z + 2xy -2x \tilde y - 3x^3 \right).
\end{equation}
Furthermore
\begin{eqnarray}
A_2(M_a(1)) & \cong& \mathbb{C}[x]\oplus \mathbb{C}[x]\oplus (\mathbb{C}[x]\otimes M_2(\mathbb{C})) \label{second-characterization-level-two} \\
& \cong & A_1(M_a(1))\oplus (\mathbb{C}[x]\otimes M_2(\mathbb{C})),
\label{third-characterization-level-two}
\end{eqnarray}
where $M_2(\mathbb{C})$ denotes the algebra of $2\times 2$ complex matrices.
\end{thm}
\subsection{Generators and relations for the algebra $A_n(M_a(1))$}\label{A_n-construction-section}
As an illustration of applications of this paper, we note that $M_a(1)=\langle \alpha(-1){\bf 1}\rangle$ and satisfies the permutation property. Therefore, the results of Sections \ref{general-n-section} and 4 can be used in the calculation of the higher level Zhu algebras $A_n(M_a(1))$, for $n>0$, and as mentioned above, have been used by the authors to explicitly determine the structure of $A_2(M_a(1))$ \cite{AB-details-of-Heisenberg}, using only the internal structure of $M_a(1)$ to obtain a minimal set of generators, and then some basic facts about irreducible modules for $M_a(1)$ to prove the relations.
In particular, the recursion given in Proposition \ref{recursion-n} can be used to reduce the collection of potential generators in order to determine a finite generating set for $A_n(M_a(1))$. Lemma \ref{Y+-lemma} allows one to obtain expressions that are equivalent modulo $O_n^\circ(V)$ to $Y^+(\alpha(-m){\bf 1}, x)$ and are rational functions of $x$, which is helpful in simplifying the calculations of circle products, $u\circ_nv$, that can be used in determining the structure of $A_n(M_a(1))$. Proposition \ref{general-generators-prop} and Theorem \ref{reducing-general-generators-thm} further reduce the set of generators. Eqns.\ \eqref{reduce-equation} and \eqref{reduce-equation-n=1} are important in simplifying many of the expressions needed to determine relations satisfied by elements of $A_2(M_a(1))$, and one expects to similarly use such expressions to calculate $A_n(M_a(1))$ for $n>2$.
The multiplication formulas from Section \ref{mult-section} and the subsequent results about generators are also heavily used in obtaining the structure of $A_2(M_a(1))$ and can be used to calculate $A_n(M_a(1))$ for $n>2$. For instance, Theorem \ref{last-generator-theorem} applied to the case that $V=M_a(1)$ gives the following result:
\begin{thm}\label{last-generator-theorem-heisenberg} For $n \in \mathbb{Z}_+$,
$A_n(M_a(1))$ is generated by elements of the form
\[ \alpha({-2n})^{i_{2n}}\alpha({-2 n + 1})^{i_{2n -1}} \cdots \alpha({ - 1})^{i_1}\mathbf{1}+ O_n(V) , \mbox{and} \ \alpha({-1})^{j} \mathbf{1} + O_n(V),\]
for $i_1, i_2, \dots, i_{2n} \in \mathbb{N}$, with $0 \le i_1 \leq n$, and $j \in \mathbb{Z}_+$.
\end{thm}}
This immediately reduces the collection of elements one must consider in determining finite generating sets for the higher level Zhu algebras for $M_a(1)$.
Theorem \ref{clarify-definition-thm} shows that vectors of the form $(L(-1)+L(0))v$, $v\in M_a(1),$ must be included in the definition of $O_n(M_a(1))$ when one defines $A_n(M_a(1))$ for any $n>0$ and, as is illustrated in \cite{AB-details-of-Heisenberg}, provides important insight into determining the structure of these higher level Zhu algebras.
In \cite{AB-details-of-Heisenberg}, we make the following conjecture for the general structure of the higher level Zhu algebras $A_n(M_a((1))$, which has been proven for $n = 1$ in \cite{BVY-Heisenberg} and for $n= 2$ in \cite{AB-details-of-Heisenberg}
\begin{conj} For $n \in \mathbb{Z}_+$
\begin{eqnarray*}
A_n(M_a(1)) &\cong& A_{n-1}(M_a(1)) \oplus \left( \mathbb{C}[x] \otimes M_{p(n)}(\mathbb{C}) \right)\\
&\cong& \mathbb{C}[x] \oplus \mathbb{C}[x] \oplus \left( \mathbb{C}[x] \otimes M_2(\mathbb{C}) \right) \oplus \left(\mathbb{C}[x] \otimes M_3(\mathbb{C}) \right) \oplus \left(\mathbb{C}[x] \otimes M_5(\mathbb{C}) \right) \\
& & \quad \oplus \cdots \oplus \left(\mathbb{C}[x] \otimes M_{p(n)} (\mathbb{C}) \right)
\end{eqnarray*}
where $p(n)$ denotes the unordered partitions of $n$ into nonnegative integers, and $M_{p(n)}(\mathbb{C})$ denotes the algebra of $p(n) \times p(n)$ matrices.
\end{conj}
\subsection{The Virasoro vertex operator algebra}
The Virasoro vertex operator algebra is another example of a vertex operator algebra for which the results of this paper can be readily and easily applied to aid in determining the structure of higher level Zhu algebras.
Let $\mathcal{L}$ be the Virasoro algebra with central charge $\mathbf{c}$, that is, $\mathcal{L}$ is the vector space with basis $\{\bar{L}_n \,|\, n\in\mathbb{Z}\}\cup \{\mathbf{c}\}$ with bracket relations
\begin{align*}
[\bar{L}_m,\bar{L}_n]=(m-n)\bar{L}_{m+n}+\frac{m^3-m}{12} \delta_{m+n,0} \, \textbf{c},\quad\quad [\textbf{c},\bar{L}_m]=0
\end{align*}
for $m,n\in \mathbb{Z}$. Here we use a bar over the Virasoro generators to distinguish between these Virasoro elements and the functor $L_n$ defined earlier.
Let $\mathcal{L}^{\geq 0}$ be the Lie subalgebra with basis $\{ \bar{L}_n \,|\, n\geq 0 \} \cup \{\mathbf{c}\}$, and let $\mathbb{C}_{c,h}$ be the $1$-dimensional $\mathcal{L}^{\geq 0}$-module where $\mathbf{c}$ acts as $c$ for some $c\in \mathbb{C}$, $\bar{L}_0$ acts as $h$ for some $h\in \mathbb{C}$, and $\bar{L}_n$ acts trivially for $n\geq 1$. Form the induced $\mathcal{L}$-module
\begin{equation}\label{V-module-M}
M(c,h)= \mathcal{U}(\mathcal{L})\otimes_{\mathcal{L}^{\geq 0}} \mathbb{C}_{c,h} .
\end{equation}
We shall write $L(n)$ for the operator on a Virasoro module corresponding to $\bar{L}_n$, and $\mathbf{1}_{c,h} = 1 \in \mathbb{C}_{c,h}$. Then
\[ V_{Vir}(c,0)= M(c,0)/\langle L(-1)\mathbf{1}_{c,0}\rangle
\]
has a natural vertex operator algebra structure with vacuum vector $1=\mathbf{1}_{c,0}$, and conformal element $\omega=L(-2)\mathbf{1}_{c,0}$, satisfying $Y(\omega,x)= \sum_{n\in\mathbb{Z} } L(n)x^{-n-2}$. In addition, for each $h\in\mathbb{C}$, we have that $M(c,h)$ is an ordinary $V_{Vir}(c,0)$-module with $\mathbb{N}$-gradation
\[M(c,h)=\coprod_{k\in \mathbb{N}} M(c,h)_k\]
where $M(c,h)_k$ is the $L(0)$-eigenspace with eigenvalue $h + k$. We say that $M(c,h)_k$ has degree $k$ and weight $h+k$.
We now fix $c \in \mathbb{C}$, and denote by $V$, the vertex operator algebra $V_{Vir}(c,0)$.
It was shown in \cite{W} that
\begin{equation}
A_0(V) \cong \mathbb{C}[x,y]/(y-x^2 - 2x) \cong \mathbb{C}[x]
\end{equation}
under the identification
\begin{eqnarray}
L(-2)\mathbf{1} + O_0(V) &\longleftrightarrow& x + (q_0(x,y)), \\
L(-2)^2\mathbf{1} + O_0(V) &\longleftrightarrow& y + (q_0(x,y)),
\end{eqnarray}
where $q_0(x,y) = y - x^2 - 2x$.
It was shown in \cite{BVY-Virasoro} that
\begin{equation}
A_1(V) \cong \mathbb{C}[x,y]/(y-x^2 - 2x)(y - x^2 -6x + 4)\cong \mathbb{C}[\tilde{x},\tilde{y}]/(\tilde{x}\tilde{y})
\end{equation}
under the identification
\begin{eqnarray}
L(-2)\mathbf{1} + O_1(V) &\longleftrightarrow& x + (q_0(x,y)q_1(x,y) ), \\
L(-2)^2\mathbf{1} + O_1(V) &\longleftrightarrow& y + (q_0(x,y)q_1(x,y)),
\end{eqnarray}
where $\tilde{x} = q_0(x,y) = y - x^2 - 2x$, and $\tilde{y} = q_1(x,y) = y - x^2 - 6x + 4$.
Fix $V = V_{Vir}(c,0)$, and recall (e.g., \cite{LL}, \cite{KarL}) that $V$ is spanned by vectors of the form
\begin{equation}\label{span}
L(-k_1) \cdots L(-k_m) \mathbf{1} \qquad \mbox{for $k_1 \geq \cdots \geq k_m \geq 2$ and $m \in \mathbb{N}$} .
\end{equation}
For $v \in V$ and $r \in \mathbb{N}$, we define $F_r(v)$ to be the subspace of $V$ linearly spanned by all elements of the form
\[
L(-k_1)\cdots L(-k_m)v,
\]
where $k_1, \dots, k_m \geq 2$ and $m \leq r$. Thus $F_r(v) \subseteq F_{r+1}(v)$. Then from \cite{BVY-Virasoro}, we have the following:
\begin{lem}\label{permute}
For $m \in \mathbb{Z}_+$, let $\sigma$ be a permutation on $m$ letters, and let $v \in V$. Then for $k_1, \dots, k_m \geq 2$,
\[
L(-k_1)\cdots L(-k_m)v - L(-k_{\sigma(1)})\cdots L(-k_{\sigma(m)})v \in F_{m-1}(v).
\]
In particular,
\[ L(-k_1) \cdots L(-k_m) v \in F_{m-1}(v) \]
if $k_j \leq 1$ for any $j = 1, \dots, m$.
\end{lem}
In particular, $V_{Vir}(c,0)$ satisfies the permutation property and is strongly generated by $\omega$. Therefore, the results listed in Section \ref{A_n-construction-section}, which we noted were important in determining the structure of $A_2(M(1))$ can also be applied to $A_n(V_{Vir}(c,0))$ for $n>1$. For instance, Theorem \ref{last-generator-theorem} applied to the case $V_{Vir}(c,0)$ gives the following result:
\begin{thm}\label{last-generator-theorem-virasoro}
Let $V=V_{Vir}(c,0)$ and fix $n \in \mathbb{Z}_+$. Then $A_n(V)$ is generated by elements of the form
\[ L({-2n-1})^{i_{2n+1}} L({-2 n})^{i_{2n}} \cdots L({ - 2})^{i_2}\mathbf{1}+ O_n(V) , \mbox{and} \ L({-2})^{j} \mathbf{1} + O_n(V),\]
for $i_2, i_3, \dots, i_{2n+1} \in \mathbb{N}$, with $0 \leq 2i_2 \leq n$, and $j \in \mathbb{Z}_+$.
\end{thm}
As in the case of $M_a(1)$, one can further reduce generators and calculate relations satisfied by these generators by using the multiplication formulas found in Section \ref{mult-section} as well as by applying Lemma \ref{Y+-lemma} to calculate formulas equivalent modulo $O_n^\circ(V)$ to various circle products, $u\circ_nv$. Finally, we remark that Theorem \ref{clarify-definition-thm} shows the importance of including the vectors of the form $(L(-1)+L(0))v$, $v\in V_{Vir}(c,0)$ in $O_n(V)$ for $n>1$.
|
\section{Introduction}
New developments in large-scale language models have produced models that are capable of generating coherent, convincing text in a wide variety of domains \cite{vaswani_attention_nodate, brown_language_2020, adiwardana_towards_2020}. Their success has spurred improvements on many tasks, from classification to question answering to summarization \cite{brown_language_2020}, as well as creative writing support \cite{coenen_wordcraft_2021}.
Language models have the potential to be powerful writing tools that can support writers in real-world, high-impact domains. These new models are task agnostic, making them applicable to many tasks without requiring more training, and we believe such models are the future of AI technologies.
Despite their successes, language models continue to exhibit known problems, such as generic outputs \cite{holtzman_curious_2020}, lack of diversity in their outputs \cite{ippolito_comparison_2019}, and factually false or contradictory information \cite{lin_truthfulqa_nodate}. There remain many unknowns about how this technology will interface with people in real-world tasks, such as what interactions best serve writers, how language models can best contribute to different writing forms \cite{calderwood_how_2018}, and how to mitigate the bias that language models encode \cite{bender_dangers_2021}.
To that end, we study how language models can be applied to a real-world, high-impact writing task. In particular, we use a science writing form called “tweetorials” which explain technical concepts on Twitter for a general audience \cite{breu_tweetstorm_2020}. Tweetorials are short explanations of around 500 words which have a low-barrier to entry and are gaining popularity as a science writing medium \cite{soragni_scientists_2019}.
Working on science writing requires a system to demonstrate proficiency within an area of expertise. This is much more difficult than traditional creative writing tasks, such as stories and poetry, which tend to deal with common objects and relations. Thus, we present the following research question:
\begin{quote}
\textit{RQ: How can outputs from a language model support writers in a creative but constrained writing task?}
\end{quote}
In this paper, we present a system that aims to inspire domain experts when writing tweetorials on a topic of their expertise. This system provides what we call “sparks” -- sentences intended to spark ideas in the writer. Our system generates sparks using a mid-sized language model (GPT-2 \cite{radford_language_nodate}) and a custom decoding method to encourage specific and diverse outputs. Additionally, we develop a set of 10 prompts based on narratology and expository theory that participants can use when interacting with the model.
We run two evaluations. In the first, we compare the outputs from a custom decoding method to a competitive baseline as well as to a human-created gold standard, reporting on the diversity and coherence of all outputs. In our second study, we have 13 PhD students from five STEM disciplines write tweetorials with our system and report on how they thought about and made use of the sparks.
We make the following contributions:
\begin{itemize}
\item a system that uses a language model to generate “sparks” related to a scientific concept;
\item a custom decoding method for generating sparks from a pre-trained language model;
\item an evaluation
demonstrating that the sparks are more coherent and diverse than an off-the-shelf system, and approach a human gold standard; and
\item an exploratory study with 13 PhD students showing three main use cases of sparks.
\end{itemize}
We end by discussing how to best use language models in constrained writing tasks.
\section{Related Work}
\subsection{Natural Language Generation}
A language model is any model that predicts the likelihood of a sequence of words. This can be used to generate text by giving the model a prefix -- either a sequence of words or a special `start of sequence' word -- and having it calculate the likelihood of all words in its vocabulary as the next word. This can be used to select the next word, and thus generate text. Language models are trained on a dataset of text that does not need to be annotated in any way, as the model can simply be trained on what the next word in a sentence will be \cite{stanford-nlp-textbook}.
Language models are getting larger: they are being trained on more text and the models have more parameters \cite{radford_language_nodate,brown_language_2020, adiwardana_towards_2020}. Much recent work has been on how to make the best use of these large language models, which have shown to be much more general purpose than previous ones \cite{keskar_ctrl_2019}, even showing promise in generating code \cite{austin_program_2021}. Additionally, it is useful to be able to take one language model and use it for many tasks, rather than having to train a new model for each task.
It has been shown that a well-selected prefix, or `prompt', can dramatically increase the performance of a language model on a specific task \cite{reynolds_prompt_2021}. A resulting line of research has been automatically creating either natural language prompts or continuous vector prompts, to perform well on tasks \cite{gao_making_2021, li_prefix-tuning_2021}. However, automatically learned prompts have yet to consistently outperform manually crafted prompts \cite{gao_making_2021}. Additionally there have been very few studies on prompt selection for open ended, generative tasks \cite{coenen_wordcraft_2021}.
Despite the successes of language models, problems remain. When considering the generation of text, language models tend to output repetitive and vague responses \cite{holtzman_curious_2020, ippolito_comparison_2019}. Language models also have no model of the truth; they are learning correlations from large amounts of text. Thus they are able to produce text that includes falsehoods and offensive language \cite{bender_dangers_2021}.
\subsection{Generative Writing Support}
Technological writing support has a long history, but has seen an increase in attention as language models have improved. Early work on language models for creative writing focused on activities such as storytelling \cite{roemmele_writing_nodate} and metaphor writing \cite{chakrabarty_mermaid_2021}. While these tools proved helpful for writers, they were narrow in what they could provide. An exploratory study found that generic auto completion from a language model did not provide enough control for novelists \cite{calderwood_how_2018}. More recent work on writing support for creative tasks has varied the ways in which technology can support the writer, for instance by providing description, plot points, or even asking questions, depending on the desires of the writer \cite{coenen_wordcraft_2021, arnold_generative_nodate}.
Writing support for nonfiction writing tasks tends to be much more constrained, for instance as sentence completion \cite{chen_gmail_2019}. A good example is Gmail's smart reply function, which aims to suggest only text that the writer would have written in anyway \cite{kannan_smart_2016}, although it has been shown that even these suggestions can change what people write \cite{arnold_sentiment_nodate}. Work on helping people craft responses to those in mental health crisis focuses on providing writers feedback, and suggested words, rather than complete phrases or sentences \cite{peng_exploring_2020}.
While natural language generation is a large and growing field, few of its technologies are studied in the context of how they will be used by writers. For instance, although there is much work on automatic summarization \cite{zhang_pegasus_2020, he_ctrlsum_2020}, there's less work on how the summaries might be used by people. Our work aims to study how text generated by language models might be used by writers in a science writing task. There's some relation to a natural language generation task like summarization, because we are concerned with real facts, but we take a human centered approach where the language model provides suggestions, rather than a completed output.
\subsection{Science Communication on Social Media}
Science communication helps the public understand scientific contributions -- consider how it has been applied to tackle vaccine misinformation \cite{shelby_story_2013}, the COVID-19 pandemic \cite{yang_covid-19_2020}, and climate change \cite{hart_boomerang_2012}. Traditionally, science communication took place through journals, conferences, articles, and books -- places where peer review was an implicit part of the publication process. However, the rise of digital networks has made science establish a virtual presence through electronic journals and digital records. The ubiquity of social media further presented opportunities for scientists to have direct channels to the public. Now any scientist can conduct science communication online by posting about their work online \cite{soragni_scientists_2019}, engaging in the `Ask' communities on Reddit \cite{gilbert_i_2020} or explaining something on Youtube \cite{welbourne_science_nodate}. Even PhD students or undergraduate researchers have the ability to disseminate their scientific knowledge at any time without depending on a venue or a publication.
This emerging trend, where the scientist can now partake in conversations outside of an implicitly gated, peer-review process, reflects one of the many broad shifts away from traditional science communication. Scholars of science communication have reified this emerging form of communication as ``post-normal science communication'' \cite{bruggemann_post-normal_2020}. Defining characteristics of post-normal science communication include a tolerance for subjectivity, an insertion of the self, the integration of advocacy, and call to actions. Despite these dramatic shifts, the original tenets of science communication such as storytelling, analogies, figures, and citations remain valuable, and storytelling in particular is a driving principle within our system.
\subsection{Expository and Narrative Theory}
In studying how narratives are embedded in text, we turn to a rich body of literature about narratives and knowledge structure in semiotics and discourse theory. These domains inform our search for structures we could use to prompt language models.
We looked at frameworks for both expository and narrative writing, because tweetorials are a hybrid of both. Specifically, we draw from the constructionist theory for narrative text, discourse theory for narrative text, and discourse theory for expository text.
The constructionist framework of narratology states that all reading comprehension is “a search for meaning'' \cite{Graesser1994}. Readers infer as they build a mental model of why certain actions, events, and states are involved in a situation. The constructionist framework has a classification of inferences that we borrow from for many of our prompt templates. Our prompts exemplify a subset of these inference classes such as case structure role assignment, causal antecedent, the presence of superordinate goals, and instantiation of a noun category.
Concurrently, we examined expository text discourse theory for knowledge structures that would lend well to prompt templates. One framework for expository text introduced a taxonomy of didactic methods (evaluation, explanation, occasion, and expansion) to enumerate the different conversational moves a writer can make to "influence the inference process of the reader" \cite{Tucker1986}. An alternative and popular framework was put forth by Meyer et. al., who enumerated signal phrases that distinguish expository texts, such as ‘specifically’ or ‘attributes of’. We chose to incorporate multiple signal phrases from Meyer’s framework into our prompt templates \cite{meyer_structure_2017}.
\section{Formative Study}
In order to understand how a language model might best support the task of writing a tweetorial, we ran a formative study where participants were first given a technique for coming up with a compelling introduction, before being asked to write the first tweet of a tweetorial on a technical topic they were familiar with. Since the first tweet tends to set up the context and intention of the tweetorial \cite{breu_tweetstorm_2020} we found this to be an effective and efficient way to understand what participants found difficult in the writing process, even when provided with writing strategies.
\subsection{Methodology}
We recruited 10 students in Computer Science.\footnote{6 women / 4 men; 7 undergraduates--no first years / 3 PhD students.}
Participants were required to go through tutorials on how to write an engaging science writing introduction on two example topics -- recursion and virtual private networks -- which included several examples and a step-by-step process for coming up with ideas. These tutorials focused on coming up with an intriguing question for the first tweet, and were developed in consultation with a science journalist.
The process was: 1) brainstorm three concrete situations related to the topic, 2) turn each situation into a question for the reader, 3) select the most engaging question.\footnote{Links to tutorials will be released after anonymous review.} These tutorials were intended to provide the participants with as much “unintelligent” support as possible, such that we could identify where language models may be able to add benefit.
After the tutorials, participants were asked to select a topic from one of six Computer Science topics and write the first tweet for a tweetorial that would explain that topic.\footnote{The topics were: hashing, sorting algorithms, Bayes theorem, HTTP, transistors, and Turning Machines. We selected these topics as ones that a) most computer science students should have learned in a formal setting, and b) could make for an interesting tweetorial.} Participants were asked to think aloud during the writing process. They were not allowed to browse the web. Afterwards, they were asked a series of questions about their writing process in a semi-structured interview.
After all participants had completed the study, the research team reviewed their writing with a science journalist. No formal coding was done, but general areas of success and areas for development were discussed.
\subsection{Results}
\subsubsection{Participants reported that the task required creativity, and that it was difficult to come up with ideas.}
Although we didn’t frame the task as a creative writing task, many participants described the task as difficult because it required creativity to come up with something that would engage the reader. Most participants said they don’t typically do creative writing, so they found the task difficult and outside of their area of comfort. This supported our selection of tweetorials as a writing task, as we want to study a task that is both constrained and creative.
Participants found the tutorials helpful, though for a variety of reasons. Some liked seeing the examples, others appreciated a process to follow, and still others found it comforting to see writing get better with brainstorming and revision. Several commented that the tutorials made the task look easy, but when they began to write about their own topic it was surprisingly difficult to come up with ideas.
Most participants (9 out of 10) said that making the topic interesting to a general audience was the most difficult part of the writing task. When pressed to be more specific, participants mentioned coming up with concrete examples/situations and creating an engaging question as hard tasks. Though this was surely influenced by the process the tutorials introduced, this confirmed that tutorials are not enough to fully support writers in this task.
\subsubsection{Participants struggled to come up with ideas that created suspense.}
When reviewing what the participants had written, all the tweets mimicked the tone of the examples. However, the science journalist had critiques for all of them, and most of the critiques at the core were the same: the tweet lacked suspense. By this he meant, the tweet did not introduce a compelling problem or gap in the reader’s understanding that would make the reader want to read more. Often this was because the example used wasn’t particularly compelling or didn’t reflect a real use case of the topic.
Additionally, participants tended to repeat similar ideas to others who had selected the same topic. For instance, all the people writing about HTTP used either Google or Twitter as their example, suggesting that participants may converge on similar, easy to reach ideas.
Given that participants reported coming up with ideas difficult, it's likely that participants could have done better if given help coming up with more ideas.
Members of the research team also noted that many of the tweets written might be difficult to turn into full length tweetorials. For instance, if the question couldn’t really be answered with an explanation about their chosen topic. For this reason, in future studies we had participants write more than just the first tweet.
\subsection{Design Goals}
Based on our formative study, we developed two design goals for our system:
\subsubsection{Support writers with idea generation.}
Given that language models have no model of truth, we want our system to come up with “sparks”, intended to spark ideas in the writer, rather than having the system provide the ideas themselves. This aligns with prior work on creativity support tools, where users make use of system outputs as initial directions that are then interpreted and diverged from in the users' actual creation \cite{gero_metaphoria_2019}.
Additionally, this also encourages the writer to feel more ownership over their final product, which has shown to be a concern in past work \cite{peng_exploring_2020}.
\subsubsection{Generate outputs that are coherent and diverse.}
In order for writers to make use of outputs, even if they are not always perfectly accurate, they should be coherent -- well-formed and generally reflecting accurate knowledge. Additionally, to support idea generation, outputs should also be diverse, such that writers have a variety of outputs to make use of.
\section{System Design}
\subsection{Generating Sparks}
\subsubsection{Language model selection.}
To generate sparks we use GPT-2, an open source, mid-sized (1.5 billion parameters), transformer language model trained on 40GB of text from the web \cite{radford_language_nodate}. We use the huggingface implementation \cite{wolf_transformers_2020}. While larger open source models are available, e.g. GPT-3 \cite{brown_language_2020} or Megatron-LM \cite{shoeybi_megatron-lm_2020}, we wanted to limit the size of the model we used as larger models are more expensive to run and take more time to generate text. Additionally, there have been many critiques of the super-large language models \cite{bender_dangers_2021}, and thus we wanted to use the smallest language model able to perform well for our use case. Anecdotally, we found that DistilGPT2, a `distilled', smaller version of GPT-2 \cite{sanh_distilbert_2020}, was not able to produce coherent responses to our prompts.
We experimented with fine-tuning GPT-2 on a data set of science writing, but found that this made little difference, especially compared to modifying the decoding method or the prompts. For this reason most of our design effort focused on decoding and prompt engineering.
\subsubsection{Decoding method.}
In addition to selecting a model, we had to design a decoding method -- how to select the next token given the probability distribution the model outputs. There are several common ways of decoding from language models: greedy search, beam search, top-k sampling \cite{holtzman_curious_2020}, and top-n sampling \cite{fan_hierarchical_2018}, to name a few. Different methods have different strengths and weaknesses. Beam search tends to produce high quality results \cite{meister_if_2021} but also tends to produce very similar results for the same prompt. Sampling methods can produce much more varied results, but at the cost of being less coherent. We designed a method that attempts to further increase the coherence of beam search while also increasing its diversity.
First, we modify the probability distribution using a normalized inverse word frequency, in order to increase the likelihood of infrequent words. Normalized inverse word frequency is often used in natural language generation to improve the specificity of outputs \cite{ko_domain_nodate, zhang_learning_nodate}, which is one method for increasing the overall quality of results. To our knowledge this is the first work to use normalized inverse word frequency purely as a decoding method as opposed to during training.
To calculate the word frequencies, we wanted a corpus that doesn't over-represent uncommon science words, like a science writing dataset might, but also reflects modern word usage. For these reasons,
we use a corpus of Vox news articles that includes all articles published before March, 2017.\footnote{\url{https://data.world/elenadata/vox-articles}}
\autoref{fig:logits} shows an example of the probability distribution being modified. In this figure you can see that words like ``governments'', ``Bitcoin'', ``software'', and ``developers'' have an increased weight, while words like ``many'', ``both'', and ``all'', are not modified.
\begin{figure}
\includegraphics[width=.6\textwidth]{fig/logits.png}
\caption{This graph shows how the likelihood of the 20 next most likely words given the prompt "cryptography is used by". The orange line shows the distribution after it has been rewieghted with normalized inverse word frequencies (NIWF). Words like ``governments'', ``Bitcoin'', ``software'', and ``developers'' have an increased probability, while words like ``many'', ``both'', and ``all'' are not modified.}
\Description{adfsdfasdfsdfsad.}
\label{fig:logits}
\end{figure}
Second, we use only the top 50 highest ranking tokens. This is sometimes called top-k sampling, as only the top $k$ tokens are used \cite{fan_hierarchical_2018}. However, since we're not using a sampling method, the effect of this is to ensure that the modified probability distribution doesn't introduce any incoherencies, for example by dramatically increasing the rank of a token very far down in the original probability distributions.
Third, we increase the diversity of outputs by forcing the first token of each output to be unique, but attempt to retain coherence generating the rest of the tokens with beam search. While several more sophisticated methods have been proposed to increase diversity while retaining the coherence of beam search (e.g. \cite{vijayakumar_diverse_2018}), in testing we found none were as effective as simply enforcing the first token to be unique.
Finally, in order to keep the sparks succinct and generating quickly, we only generate 10 tokens after the prompt, and cut off the generation as soon as a sentence has been completed.
We implement our decoding method using the huggingface transformers \cite{wolf_transformers_2020}.\footnote{Link to code to be added after anonymous review.}
\subsubsection{Prompt design.}
Designing prompts for language models has become an active area of research, with many automatic methods being proposed \cite{gao_making_2021, li_prefix-tuning_2021}. However, any automatic method requires at least some training data, and it's yet to be seen that automatically developed prompts can outperform hand-crafted prompts \cite{gao_making_2021}. For these reasons, we hand-craft our prompts.
First we craft a `prefix' prompt to pre-pend to any prompt used by a writer. Prefix prompts have been shown to greatly improve performance by providing the language model with appropriate context \cite{reynolds_prompt_2021}. We found early on in development that simply providing the model with a technical topic was not enough -- also providing a context area was necessary for it to appropriately interpret technical terms. For instance, if you use a prompt like "Natural language generation is used for", the model is likely to talk about linguistic research on languages, rather than computational methods. If instead you use the prompt, "Natural language generation, a topic in computer science, is used by" the results are much more likely to refer to computational language generation. Given this, we pre-pend all prompts with the following:
``\{topic\} is an important topic in \{context area\}''
where \{topic\} and \{context area\} are provided by the writer.
In hand-crafting our prompts, we wanted to make sure our prompts captured a range of relevant angles, so our system could flexibly work with any technical discipline. To do so, we synthesized work from expository and narrative theory into prompts capturing five categories: expository, instantiation, goal, causal, and role. Each category represented an angle that a writer might want to explore. All prompts can be seen in \autoref{tab:prompts}.
\begin{table}
\caption{Prompt templates designed for science writing task.}
\label{tab:prompts}
\begin{tabular}{ll}
\toprule
category & prompt\\
\midrule
expository
&One attribute of \{topic\} is\\
&Specifically, \{topic\} has qualities such as\\
instantiation
&One application of \{topic\} in the real world is \\
&\{topic\} occurs in the real world when\\
goal
&For instance, people use \{topic\} to \\
&\{topic\} is used for\\
causal
&\{topic\} happen because \\
&For example, \{topic\} causes\\
role
&\{topic\} is used by \\
&\{topic\} is studied by\\
\bottomrule
\end{tabular}
\end{table}
We manually developed these prompts according to established frameworks within narrative and expository theory that we referenced in our related work. Our prompts within the categories of instantiation, goal, antecedent, and role were crafted based upon the constructionist framework of inferences, specifically the following categories: case structure role assignment, causal antecedent, the presence of superordinate goals, and the instantiation of a noun category (respectively). Less formally, \textit{Instantiation} prompt templates suggest completions that instantiate where and in what ways topic X may occur in the real world. \textit{Goals} prompt templates suggest completions that represent how topic X is used in the real world. \textit{Causes} prompt templates suggest completions for how topic X might interact in cause and effect chains. \textit{Roles} prompt templates cover entities involved with topic X.
As tweetorials exhibit both elements of narrative and expository writing, we also borrowed signal phrases from Meyer’s framework for expository text \cite{meyer_structure_2017} -- e.g. “specifically", ``such as", ``attribute” -- and folded them within our prompt templates.
In testing we found that participants often wanted to `follow up' on an output by entering in their own prompt. For this reason, we added the ability for writers to add their own prompts, though this prompt would also be pre-pended with our prefix.\footnote{One intriguing area of research is `meta-prompting', where the language model is used to generate the prefix for the next generation \cite{reynolds_prompt_2021}. While we found that this produced intriguing results for our use case, for example by having the model first produce a list of types of people who interact with a topic, and then putting those phrases into a downstream template, we thought it added too much complexity.}
\subsection{Interface}
We design a website that takes in a writer's topic and context area. \autoref{fig:system} shows a screenshot of the system with its important features marked. The website consists of a single textbox for writing, and a `prompt box' above it that allows writers to interact with the sparks. Writers can select a templated prompt from a dropdown menu, or type in their own prompt and add it to the dropdown list. When a prompt is selected, if they press `GENERATE' the language model will generate a single spark. Writers can `star' a spark by clicking on the lightbulb icon -- this fills in the lightbulb and also pastes the spark into the textbox. If a writer selects a different prompt, the sparks already generated are preserved such that if they return to a previous prompt their generated sparks will be shown again.
\begin{figure}
\includegraphics[width=\textwidth]{fig/system-diagram.png}
\caption{Example screenshot of our system generates sparks. A: writers can select from 10 template of prompts in a drop-down menu. B: writers can add their own prompt to the drop-down menu. C: sparks are generated with a lightbulb icon to the left, if writers click the lightbulb it will highlight and the spark is copied into the text area. D: writers can hit the generate button in order to generate a new spark.}
\Description{adfsdfasdfsdfsad.}
\label{fig:system}
\end{figure}
The textbox contains some features useful for the tweetorial writing task. The textbox is split into two sections with a line of dashes. Above the line is reserved for brainstorming and notes, a feature writers requested and found useful during pilot studies. Below the line is the text area for the tweetorial writing. A word count for the writer's tweetorial draft is displayed at the top of the textbox, and a character count for each tweet (separated by line breaks and two forward slashes) is displayed to the left. \autoref{fig:textbox} shows these features with an example from our user study.
The website is implemented using Python 3.7 and the Flask web framework.\footnote{Link to demo to be released after anonymous review.}
\begin{figure}
\includegraphics[width=.6\textwidth]{fig/textbox.png}
\caption{Screenshot of the text area from our user study. At the top is a word count, which counts only the words below the dashed line. Text above the dashed line is interpreted as brainstorming or notes. Participants can separate tweets with a double `//', and the character count for each tweet is shown to its left.}
\Description{adfsdfasdfsdfsad.}
\label{fig:textbox}
\end{figure}
\section{Study 1: Spark Quality}
We wanted to evaluate the quality of the sparks generated by our system. In particular, we wanted to evaluate how well the sparks, in isolation (i.e. not in a writing task), met our design goals of coherent and diverse. We also wanted to test how well the sparks could support a wide range of topics, and if certain prompts supported some topics better than others. To do so, we compared the sparks generated by the custom decoding method to a baseline system, as well as a human-created gold standard.
We have three hypotheses:
\begin{itemize}
\item H1: The custom decoding produces more coherent and diverse outputs than a baseline system, but less coherent and diverse outputs than a human-created gold standard.
\item H2: The custom decoding performs consistently across many different topics.
\item H3: Some prompts work better for some topics.
\end{itemize}
\subsection{Methodology}
We wanted to evaluate the quality of ideas for a variety of topics. We selected three disciplines that have a glossary of terms page on Wikipedia, and that have been demonstrated to be a rich discipline for science writing on social media.\footnote{e.g. \url{https://twitter.com/dannydiekroeger/status/1281100866871648256}, \url{https://twitter.com/GeneticJen/status/897153589193441281}, and \url{https://twitter.com/meehancrist/status/1197527975379505152}} These disciplines were computer science, environmental science, and biology. For each discipline we randomly sampled 10 topics from their glossary of terms page. See the appendix for the full list of topics studied.
\subsubsection{Collecting a human-created gold standard.}
We wanted to collect human responses to our prompts to represent a gold standard or upper limit on the quality of ideas these prompts can generate. To do this, we recruited 2-3 PhD or senior undergraduate students in each discipline and had them complete the same prompts the language model did. Each student was paid \$20/hour for as long as it took them to finish the task.
We explained to them that the purpose of the prompts was to generate ideas to support an expert writing about the topic for a general audience. Each student had to complete 5 prompts per topic in 3 different ways, and was told to make the completions for a given prompt+topic combination maximally different. They were also instructed to ensure their completions were accurate, given their understanding of the topic, and that they could reference the web if they needed to check anything, as well as use web search results for inspiration. Finally, we explained that their ideas should be as concrete and specific as possible. Each student completed 5 prompts for the 10 topics in their discipline, for a total of 5 x 10 x 3 = 150 completions per person. It took them on average 3.5 hours to come up with completions for all 10 topics in their discipline, and in the end we had 6 high quality completions per prompt+topic combination.
\subsubsection{Baseline language model condition.}
We compare the custom decoding to a language model baseline: group beam search with hamming diversity penalty. This is a strong baseline that encourages diversity in the way \cite{vijayakumar_diverse_2018} recommends, and can be implemented using arguments in the ‘generate’ function in the huggingface transformer library. Both the custom decoding and baseline model use the same underlying language model.
\subsubsection{Measuring coherence and diversity.}
Coherence is notoriously difficult to measure automatically, especially without training data -- measures like perplexity merely measure an output's likelihood under the model itself. For this reason we recruited 10 domain experts to annotate outputs for coherence on a 0 - 4 scale, in line with knowledge graph evaluations \cite{li_commonsense_2016}. For biology we had 3 senior undergraduate students majoring in biology; for environmental science we had 2 senior undergraduate students majoring in environmental science; for computer science we had 2 PhD students from the computer science department.\footnote{The students could not have also participated in the generation portion.}
Each discipline had 900 sentences to annotate (300 human generated, 300 from the baseline model, and 300 from the custom decoding).
250 randomly selected outputs from each discipline were annotated by two different domain experts, and the Cohen's weighted kappa was calculated as: $\kappa=.54$ for biology, $\kappa=.51$ for environmental science, and $\kappa=3.4$ for computer science. Given that the agreement was moderate, we had a single annotation for the remaining sentences.
We measure diversity with sentence embeddings \cite{reimers_sentence-bert_2019}, in particular we report the average distance between outputs within a given prompt. A higher average distance means that outputs are more dissimilar, and therefore more diverse.
\subsection{Results}
Overall, the baseline had low diversity and coherence across all disciplines, while the human-created outputs perform much better. \autoref{fig:diversity} and \autoref{fig:coherence} show that the custom decoding method outperforms the baseline, but does not reach the performance of the human-created outputs. For diversity, two-tailed t-tests show this to be a significant difference for all disciplines (computer science: $p<.001$, climate science $p<.001$, biology: $p < .001$); for coherence, mann-whitney U tests show this to be a significant difference for all disciplines (computer science: $p<.001$, climate science $p<.001$, biology: $p < .001$).
\begin{figure}
\centering
\hfill
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{fig/diversity1.png}
\caption{Distribution of diversity, split by discipline. Diversity is measured as the average sentence embedding distance per prompt+topic combination.}
\label{fig:diversity}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{fig/coherence.png}
\caption{Mean coherence per prompt+topic combination, split by discipline. Each prompt completion was scored by a domain expert on a scale of 0 to 4.}
\label{fig:coherence}
\end{subfigure}
\caption{Diversity and coherence measures across three test disciplines for three conditions: a baseline language model, a language model with the custom decoding, and a human-created gold standard. The custom decoding improves upon the baseline and approaches the human gold standard.}
\label{fig:three graphs}
\end{figure}
\begin{table}
\begin{tabular}{llll}
\toprule
condition & coh & div & One attribute of source code is...\\
\midrule
&&&it is typically written in a human-readable format.\\
human created&4&.38& editability, so that programmers can easily change it to suit their needs.\\
&&&it is a description a computer program.\\
\hline
&&&that it contains code written by humans.\\
custom decoding & 4 &.37&its modularity - code modules contain reusable code components.\\
&&& complexity.\\
\hline
&&&that it can be used as a source of information.\\
baseline model & 2.6&.08&that it can be used as a source of inspiration.\\
&&&its modularity.\\
\bottomrule
\end{tabular}
\caption{Example outputs from our three conditions for a single prompt+topic combination, and the average coherence (coh) and diversity (div) scores for each set of three outputs.}
\label{tab:outputs}
\end{table}
\autoref{tab:outputs} shows some example outputs from each conditions for a single prompt+topic. These examples demonstrate the quality of the human generated outputs: they are long, detailed, and diverse. Comparatively both language model methods are shorter, less specific, and more repetitive. However, the custom method seems to improve the overall quality of the outputs.
\begin{figure}
\centering
\includegraphics[width=.8\textwidth]{fig/prompt-coherence.png}
\caption{This graph shows the coherence per topic for the custom decoding and the human-created gold standard, where 0 is nonsensical or untrue and 4 is generally true. The black dot shows the average coherence of all responses for a given topic, while the colored dots show the average coherence for a given topic per prompt. Topics are ordered by average coherence in the custom decoding. This graph shows that some topics perform much better than others with custom decoding, while the human outputs are generally high quality regardless of topic. It also shows that within a topic there can be a large variation between prompt templates.}
\label{fig:prompt-coherence}
\end{figure}
It is important to acknowledge that the variation in both the diversity and coherence measures are quite large. This means that while on average the custom decoding is an improvement over the baseline, and on average the human-created outputs are better than the language model outputs, for any given prompt+topic combination the output could be very high quality or of a much lower quality. People using the system will not necessarily see this huge variation; they will only see the 10 or so model outputs that they generate.
For this reason, we dig into the variation by topic and prompt. \autoref{fig:prompt-coherence} shows the average coherence per topic for the custom decoding method and the human-created outputs. It plots the average coherence for each topic with the black dots, and the coherence for each prompt+topic combination in the colored dots. From this we can see that the variation in quality over the topics for the custom decoding method. For instance, the "computer security" outputs score an average of 3.7 in coherence, while "automata theory" outputs score 2.1. When looking at the human-created outputs, the quality is far more consistent, with no topics dropping below an average of 3 in coherence.
This demonstrates that our system works well for some topics and less well for others. While we expected that our system would not perform as well as a human would, we did expect that the system would perform more consistently across topics. It is unclear why the language model performs significantly better on some topics, and given the way that these language models are trained it is difficult to inspect or even predict how well the model will perform on a given topic.
\autoref{fig:prompt-coherence} also shows that some prompt templates work better for some topics than others. In our system, the quality of outputs vary significantly with the prompt template. In the human generated outputs, the variation is smaller, but still we see some range. For instance, let's look at the topic "dynein", the worst performing topic. The prompt "Dynein happens because" scores an almost 0 on the 0 to 4 coherence scale, while the prompt "One attribute of dynein is" scores a 3. Dynein is a family of proteins important in cell behavior. Given this, it makes sense that the system is more likely to produce coherent outputs on attributes, rather than why this family of proteins "happens".
However, it's notable that the human outputs scored 3 or above for all prompts for "dynein". Here is a human output about why dynein happens: "Dynein happens because organelles, such as the Golgi complex, need to be positioned in cells." Though this sentence structure is a little convoluted, it's clear that the human was able to compensate for the prompt and still write something coherent and meaningful.
Seeing the difference a prompt template can make highlights the importance of using a prompt that works well for the topic. Since we wanted to test our system with unseen topics, we ensure that participants can add their own prompts in case the templated prompts don't work well for their topic.
\section{Study 2: User Evaluation}
We evaluated how our system supported PhD students in writing tweetorials. Tweetorials are short explanations of around 500 words which have a low-barrier to entry and are gaining popularity as a science writing medium \cite{soragni_scientists_2019}. We use PhD students as they are eager to participate in science writing \cite{howell_engagement_2019} and many tweetorials are already written by PhD students, demonstrating that this a writing task our participants may conceivably want to engage in on their own. This study was approved by the relevant IRB.
\subsection{Methodology}
We recruited 13 participants, all students from five different STEM disciplines, to write tweetorials on a topic of their own choice, related to their area of study. By letting the participants pick their own topic, we ensured that they were writing within their area of expertise, and we were able to test our system on unseen topics.
Participants were first asked to read an introduction to tweetorials, which explained what tweetorials are and walked through an example tweetorial. They were then introduced to the system and watched a short video that demonstrated the system's features, and showed an example use case of the system when writing about 'machine learning algorithms'. Participants could ask clarifying questions to the facilitator. If participants asked to learn more about how the system worked, the facilitator said that it was an algorithm that could generate text in response to a prompt, and that they could discuss the system further after they completed the writing task.
At this point the participant was asked to pick a topic to write about, as well as provide a `context area' to aid the system in correctly interpreting their topic. Then they were given 15 - 20 minutes to interact with the system and write approximately the first 100 words of their tweetorial. Mouse clicks and key presses while the participant interacted with the system were collected, as well as all sparks generated.
After this, the participant filled out a short survey and partook in a semi-structured interview with the facilitator. The survey questions and the questions that structured the interview can be found in the appendix. The study took about an hour and participants were compensated \$40 USD for their time.
Participant interviews were transcribed and the authors performed a thematic analysis \cite{cooper_thematic_2012} on the interview transcripts. The analysis centered on three areas: how sparks were helpful, how sparks were unhelpful, and ownership concerns in response to writing with a machine. Relevant quotes were selected from the transcripts and collated in a shared document, where is the author discussed and collected the quotes into emergent themes.
\subsection{Results}
We report on participant demographics and topic selection in \autoref{tab:participants}, as well as the prominent themes that emerged through our analysis in \autoref{tab:themes}. Our thematic analysis covered three main areas: ways in which sparks were helpful, ways in which sparks were unhelpful, and how participants felt about incorporating sparks into their writing. For each area, we report themes with prevalence greater than 20\%. Prevalence is measured by the number of participants who brought up that theme in their interview.
In the case of ownership, there was a very high variability in responses, such that no one theme had over 20\%. For this reason we report the reasons that people brought up in the text, but do not list them in the table. Behavioral data, such as time spent writing and how many sparks were starred, are also reported and timelines for each participants' activity can be found in \autoref{fig:timeline}.
\begin{table}
\caption{Participant Demographics. Low = once a year or so. Med = Once a month or so. High = once a week or so.}
\label{tab:participants}
\begin{tabular}{ccC{2.3cm}cc}
\toprule
ID & Discipline & Science Writing (general / twitter)&Topic&Context Area\\
\midrule
P1 & Climate Science & Low / Low & rainfall variability & climate science \\
P2 & Climate Science & Low / Never & predicting climate change & climate science\\
P3 & Climate Science & Never / High & sea level change & geophysics\\
P4 & Climate Science & Low / Low & glacier retreat over the holocene & paleoclimate\\
P5 & Computer Science & Low / Never & computationally hard problems & computer science\\
P6 & Computer Science & Never / Never & pseudorandomness & theoretical computer science\\
P7 & Political Science & Med / Med & document embeddings & natural language processing\\
P8 & Psychology & Never / Low & regulatory fit & psychology\\
P9 & Psychology & Low / Low & motivated impression updating & social psychology\\
P10 & Public Health & Low / Low & measurement of sexism & sociology\\
P11 & Public Health & Never / Never & logistic regression & epidemiology\\
P12 & Public Health & Low / Never & deprivation indices & public health\\
P13 & Public Health & Med / Med & threat multiplier & environmental health\\
\bottomrule
\end{tabular}
\end{table}
\renewcommand{\arraystretch}{1.25}
\begin{table}
\caption{Thematic analysis}
\label{tab:themes}
\begin{tabular}{|p{4.5cm}cp{8cm}|}
\hline
\textbf{Code} & \textbf{Prevalence} & \textbf{Example Quote}\\
\hline
\multicolumn{3}{c}{\textit{reasons sparks were helpful}}\\
\hline
Crafted concise, detailed sentences. &54\%&
Most of the time it [the system] was articulating the ideas that were already in my head in a way that's short and concise.\\
Came up with ideas or angles. &46\%&
It [the system] reminded me, Oh, it's not just my application, there's these other people using the same technology, but working on other problems. \\
Showed reader perspectives. &31\%&
It [the system] reminded me that there might be a more common understanding of this thing that I'm writing about, that's different from the highly specific one I've been living in.\\
\hline
\multicolumn{3}{c}{\textit{reasons sparks were unhelpful}}\\
\hline
Incorrectly interpreted the topic. &38\%&
It [the spark] just wasn't helpful, but only because it was using the different sense of `embedded'.\\
Inaccuracies.&23\%&
Some of the sparks said, like, logistic regressions are used to estimate relative risks, which is completely not true.\\
Not desired angle.&23\%&
Someone probably does really care about measuring sexist attitudes ... but it just isn't my focus.\\
Vagueness.&23\%&
I would say about 20\% of them were just not specific enough to warrant talking about.\\
\hline
\end{tabular}
\end{table}
\subsubsection{Participant demographics and topic selection.}
The 13 participants came from five STEM disciplines, with the most common disciplines being Climate Science and Public Health. All but one were doing PhD (the remaining doing a research Master's) and varied from their 2nd year to their 7th year in their program. Participants were asked how often they wrote about technical topics for a public audience, and how often they did so on Twitter. Most participants rarely or never did so, though a few did so on a monthly or even weekly basis.
Participants were asked to select a topic they understood well that was related to their research. The facilitator attempted to aid participants in selecting a topic that wasn't too broad, but also not too specific, but as the facilitator did not necessarily have the same expertise as the participant this was at times difficult. Participants selected a wide range of topics, with no overlap. The full demographics and topic selection can be found in \autoref{tab:participants}.
Given the diversity in the participants' topics, how well the system generated sparks on their topics, and how they articulated or responded to questions in the interview, there was a high variability in how participants felt about the system. For this reason, a prevalence of 50\% or above is considered very high. This would mean that over 50\% of participants independently responded in the same way to an open ended question, despite writing about a unique topic and seeing a unique response from the system.
\begin{figure}
\includegraphics[width=\textwidth]{fig/timeline.png}
\caption{Timelines of all participants from the study, with time writing versus time generating sparks marked in different colors. Participants are grouped by their engagement pattern.}
\Description{adfsdfasdfsdfsad.}
\label{fig:timeline}
\end{figure}
\subsubsection{Participants had different engagement patterns when interacting with the system.}
Overall we saw many different ways in which participants interacted with the system. Some participants generated and starred many sparks, while others generated only a handful and starred two or three.
The average number of sparks generated was 17.2 (std=10) and the average number `starred' was 5.1 (std=4.3). Additionally, it did not seem that generating or starring more sparks necessarily meant a participant found the system more helpful. Some participants who starred just two or three sparks talked at length in their interview about how useful the system was. Others who starred more than 10 complained that the system produced only vague or inaccurate outputs.
However, we did see themes of how people tended to interact with the system.
\autoref{fig:timeline} shows all participants' task timeline, split into three archetypal patterns: `prompt and flow', `flow with breaks', and 'shifting'. We talk about each in detail:
\textbf{`Prompt and flow'} participants look at the sparks at the beginning, before they start writing, and then spend the rest of their time writing, never going back to generate more sparks. This kind of participant used ideas to jump start the writing, but once they were writing had no need to return to the sparks. Several participants described getting some kind of inspiration from the sparks and then "flowing through an argument". Once these participants got into the `flow' of writing, they didn't need any further support.
Although these participants don't generate more sparks, they may be using several sparks that they had `starred' earlier. For instance, here are the first three tweets written by P8, where sentences or phrases inspired or highlighted:
\begin{quote}
\small
Have you ever wondered why certain activities just feel right to you? Like you are able to do well and feel good in the process? \hl{Research on regulatory fit helps explain why certain activities feel right versus wrong...} (1/n)
//
\hl{Regulatory fit happens because people can be driven by a range of different fundamental motives, and distinct activities "fit" each of these motives.} For instance, people who are most motivated by their duties and obligations to others prefer activities that (2/n)
//
Involve carefully defending and protecting what they find important and methodically scrutinizing information. In contrast, people who are most motivated by hopes and aspirations prefer activities that involve eagerly making progress and engaging in curious exploration. (3/n)
\end{quote}
Although they wrote all of this without generating more sparks, they did make use of two sparks -- one for the first tweet (``Research on regulatory fit reveals that regulatory fit varies considerably across individuals depending upon factors.") and another for the second (``Regulatory fit happens because individuals behave differently under varying conditions."). So while they did not return to generating more sparks, they did return to the sparks they had previously starred.
\textbf{`Flow with breaks'} style participants are similar, but they pause in the middle of the writing and go back to sparks. This type of participant talked about returning to the sparks when they "would get stuck" or at some kind of inflection point in the writing process. P2 described this as, "there are a bunch of ideas that I need to kind of weave together before using the tool again, when I get more stuck." In both of these types of interactions, we see that the writer is using the sparks to jump-start their writing whenever something makes them pause, but when they are writing they are writing continuously, even if they are drawing on sparks they had earlier selected as potentially useful.
\textbf{`Shifting'} style participants are often moving between writing and generating sparks. They are constantly returning to the sparks for different details, or for help with crafting a sentence. They tended to star a lot of sparks and use a lot of custom prompts, as what they were looking for from the sparks changed as they proceeded through their writing, and a lot of language in their writing draws directly from a spark. For example, consider the start of P1's tweetorial on rainfall variability, where sentences or phrases inspired by sparks are highlighted:
\begin{quote}
\small
Do you like coffee??
//
Want to understand how coffee and climate change are related?
//
\hl{Rainfall variability happens because rain falls differently depending on temperature and humidity.}
//
\hl{Since temperature and humidity vary around the globe, we observe different rainfall patterns across the world.}
//
\hl{For example, rainfall variability may cause extreme precipitation events leading to flooding or a lack of precipitation will lead to drought.}
\end{quote}
The last three tweets shown here are all taken directly from sparks, with minimal changes. It almost seems like collage, where the writer is mostly arranging language that came from the system.
This suggests there are different engagement patterns that people use when interacting with machine-generated suggestions. Although we expected participants would mostly use sparks for the beginning of their tweetorial, we found that participants used sparks at all points of their writing process. In the next sections, we report on the various ways in which participants made use of the sparks.
\subsubsection{Sparks helped participants craft concise and detailed sentences quickly.}
Although we intended the system to inspire participants with new ideas, the most prevalent reason the participants cited for the spark being helpful was for crafting sentences. Many participants remarked that although the sparks were showing them information that they already knew well, it was much faster and easier to draw on language from the sparks than to write a sentence from scratch.
For instance, P12 said:
\begin{quote}
Most of the time [the system] was articulating the ideas that were already in my head in a way that's short and concise, which is useful. Like `deprivation index measures the relative deprivation experienced by an individual relative to others,' that would have probably taken me like three sentences to write, then I'd have to spend time editing it down. And then yeah... this is a lot quicker.
\end{quote}
\autoref{fig:ex1} shows another example in which P12 drew on a spark in order to write a clear and concise definition.
Several participants noted that they often go to Google or Wikipedia simply to get a well-written definition of a topic they understand well. This is something that the system was able to do for them without requiring a click away from the writing interface and incurring a change in context. P7 noted that it did a good job compressing what he would have looked for or found a Google search. P8 noted that all the sparks were similar to what she would have found on Wikipedia or via a Google search, but that they were ``bite-sized" or ``sound bite ready".
\begin{figure}
\centering
\begin{subfigure}[b]{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{fig/ex1.png}
\caption{An example of how a participant in our study used a spark for crafting a detailed sentence. Highlighted text was inspired by the spark.}
\label{fig:ex1}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{fig/ex2.png}
\caption{An example of how a participant in our study used a spark to make their tweetorial more engaging to a general audience. Highlighted text was inspired by the spark.}
\label{fig:ex2}
\end{subfigure}
\hfill
\end{figure}
\subsubsection{Sparks reminded participants of other ideas or angles about their topic.}
Several participants noted that the sparks provided good ideas or angles for discussing or introducing their topic. P2 noted that `weather prediction models' were useful entry point to their research. They said, ``that's something within my field that the general public might be more familiar with than what I actually do.''
\autoref{fig:ex2} shows how P4 drew on a spark about the `sea level rise' in order to make their topic of `glacier retreat over the holocene' interesting to the reader. P4 said in their interview, ``It's often hard to figure out how to spin things in ways that feel relevant to people who don't study this," and that the sparks helped her find ways to make her research relevant. P7 said, ``[the system] definitely generated multiple [ideas] that I could have written different tweetorials about."
When asked if the sparks were giving them new ideas, many participants said that the system was helping them get to ideas they likely would have come up with themselves, but faster. For instance P4 said, ``It was definitely faster. I think I would have gotten there, but it would have taken me longer."
\subsubsection{Sparks encouraged participants to think about common reader perspectives.}
Several participants noted that the sparks reminded them of how people reading their tweetorial might be interpreting their topic. For instance P10, who was writing about measuring sexism, noted that many of the sparks talked about sexist attitudes. She said the while that certainly is an aspect of measuring sexism, it isn't the aspect that she actually studies. And so the sparks reminded her that people's main assumptions when thinking about sexism is probably about attitudes and therefore that might be an assumption that she will have to address in her tweetorial.
P5, writing about computationally hard problems, noted that one of the sparks talked about NP-completeness. He said that while at first he thought this might be too technical, it then made him wonder if someone who was reading a tweetorial about computational hard problems might already know at least some of the keywords about this topic. In this way the sparks made him reflect on what knowledge his readers might already have.
\subsubsection{Sparks failed in different ways for different participants.}
The ways in which participants found sparks to be unhelpful varied highly. The most common reason participants said sparks were unhelpful was that they incorrectly interpreted their topic. In this case, the sparks were not necessarily incorrect, but rather they reflected some alternate interpretation of their topic. For instance, P12, who was writing about deprivation indices, said that some of the sparks were about obesity. Obesity has little to do with deprivation indices, but they thought the algorithm may have been associating deprivation with nutrition. Similarly P8, who was writing about regulatory fit, commented on several sparks about government regulation, which is unrelated to her psychology topic, but she assumed the algorithm was simply free associating with the word `regulatory'.
Other reasons participants found the sparks to be unhelpful were factual inaccuracies, dealing with aspects of their topic that they were not trying to explain or that they did not study, and vague outputs. Participants also mentioned that some sparks were nonsensical, tautological, had too much jargon, or were simply "bizarre".
Overall participants varied highly in how useful they found the sparks. Some participants found that the sparks were so low quality that they found the system completely unhelpful. Others said that even though some of the sparks were not helpful the ones that were helpful were so helpful they were unconcerned with a few that didn't make sense or were off-topic.
\subsubsection{While most participants had no ownership concerns for science writing, a few felt apprehensive about accepting machine-generated ideas.}
Most participants had no concerns about incorporating sparks into their writing. Several participants cited being very familiar with the material as a reason for not being concerned. Others said that since they are writing about public knowledge, it was unimportant where their ideas came from. One participant articulated that coming up with ideas is not the hardest part of science writing, but rather putting in the time and energy into building an audience and writing something engaging, so incorporating sparks would simply be one small part of a much larger endeavor that she took on. One participant compared the sparks to searching on Google; another compared it to Grammarly (a grammar-checking service). One participant said that the sparks were simply elaborating on his own idea, and thus he still felt ownership of the material.
A few participants did have ownership concerns. One participant talked about how he considered outreach and science writing to be part of his job as an academic, and thus any system that automated some aspect of this felt like it was taking over something that he found fundamental to his work. Several participants noted that machine writing was simply the future, and while they may have some apprehensions, they knew they would have to get over it.
Across all participants, whether or not they had ownership concerns, they wanted to ensure that anything they wrote was in their own voice and that they did not plagiarize. Several participants brought up that they were unsure exactly where the sparks were coming from, and they wanted to make sure that anything they took from the sparks was adequately changed, to alleviate any concerns about plagiarism.
P2 described this as, "I think if I was using something like this, I would probably never use an entire sentence verbatim. Just because, if you don't know where it's pulling it from... I wouldn't want to run the risk of plagiarizing something accidentally even."
When asked if these edits would make the sparks less helpful, for instance if it meant it would take more time in order to edit what they chose to incorporate, most participants said that sparks would still overall speed up the writing process. But not everyone agreed; some participants thought the need to edit the sparks made them less helpful.
\subsubsection{Custom prompts allowed participants to iterate on how they interacted with the system.}
Many participants noted that they were unsure how to get the best sparks out of the system. Several compared the system to Google but said it was easier to navigate Google because they already knew what kind of queries would get the best results. In the study though, they noted that they were unsure which prompts would produce the best sparks. Several participants even said that some of the sparks that were unhelpful may have been because their custom prompts didn't properly prime the system.
One way participants used the custom prompts was to get the system to better understand their topic.
P13's discussed how the model didn't seem to correctly interpret his topic of "threat multiplier". You can see him iterating on custom prompts in an attempt to produce more accurate results:
\begin{itemize}
\item ``threat multipliers are''
\item ``climate change is considered a threat multiplier because''
\item ``climate change is considered a threat multiplier for health disparities because''
\end{itemize}
Participants also used the prompts to generate more specific sparks.
In the case of P7, the system seemed to understand his topic "document embeddings" well enough, but he was curious about specific aspects of this topic. You can see him looking for these in the following custom prompts:
\begin{itemize}
\item ``Document embeddings help researchers''
\item ``Document embeddings help search engines''
\item ``Linguists can use document embeddings to''
\end{itemize}
Finally, others participants entered custom prompts that seemed to be used less for idea generation, and more as an auto complete. For instance, P12 entered several custom props designed to generate a definition:
\begin{itemize}
\item ``The Townsend Index is''
\item ``In sociology, a deprivation index measures''
\item ``the variables in the Townsend Index are''
\end{itemize}
Overall, we saw many different uses of the custom prompts, and found them to be an important aspect of the system.
\section{Discussion}
\subsection{How to best use large language models in constrained writing tasks.}
Creativity requires both convergent and divergent thinking \cite{taura_creativity_2011}. We expected that the sparks would support participants with divergent thinking, by showing them how their topic might connect to things of interest to readers. And indeed many participants cited this as a main use case of the sparks. However, we found that many participants also found that sparks helped with convergent thinking. Sparks helped participants put the implicit, detailed, and often messy knowledge they had in their mind into concise sentences.
Existing work on large language models for storytelling tend to focus on divergent thinking. Language models provide writers with new ideas for plot points, character attributes, scene descriptions, and the like \cite{sudowrite, aidungeon}. However, what has typically been ignored is the convergent aspect of storytelling, for instance the tying together of plot points into a satisfactory ending. This is likely because of the difficulty language models have had in the past with staying on topic, though this may be changing as language models improve.
One of the reasons that the sparks may have been able to help participants with convergent thinking is because participants already knew what they were trying to explain. If we think of storytelling in response to a writing prompt as starting somewhere and perhaps not knowing the destination, we can think of writing tweetorials as knowing the destination but not where to start. In this case, participants were able to look at the sparks and recognize some of them as a convergence of the ideas already in their mind.
We sometimes saw this when participants would enter custom prompts into the system. They often knew the correct way to complete the prompt, but didn't necessarily know the best phrasing. For instance, P12 entered in the prompts "The Townsend Index is" and "the variables in the Townsend Index are" not because they didn't know what the Townsend Index is or what its variables are, but because they wanted the system to provide a well-crafted sentence for them.\footnote{The authors of this paper did not know what the Townsend Index is. For the curious, it is a measure of material deprivation levels in a population, and it's variables are: unemployment, non–car ownership, non-home ownership, and housing overcrowding.} In this instance we might think of the language model as performing convergent thinking for the writer and letting the writer recognize what is correct and well-phrased.
An area of future development we see is more explicitly using the language model to help writers with different aspects of the writing process, for instance by developing prompts intended to directly spark new ideas and other prompts to return concise definitions or summaries. We are seeing this in the storytelling space, where systems are being developed to prompt plot ideas or specific scene descriptions or focus on rewriting in a different style.
\subsection{Three areas for technical development: better controls, increased breath, and graceful failing.}
However, to support science writing we will need to improve the underlying technology. We see three main areas of technical development that may be fruitful.
The first is improving the controllability of language models. While it is clear that pre-trained language models can generate sentences that span the spectrums of specificity and complexity and take on a number of styles and opinions, it is less clear how to ensure that a language model generates the kind of outputs a writer wants. Prompt engineering seems like a promising avenue for this, especially as prompt engineering can be done in natural language. If prompt engineering is done with vectors, it will be completely opaque to users, but if we are able to control language models through natural language prompts alone, users will have some intuitive understanding as to how to control the language model and are more likely to be able to use this control themselves.
The second is increasing the breadth of a language model's knowledge. Again, it is clear that pre-trained language models contain much world knowledge, but it is also clear that this knowledge is not evenly distributed. There have been several calls to be more careful and explicit with the data that these pre-trained models are trained on \cite{bender_dangers_2021, bandy_addressing_2021, caswell_quality_2021}. If we better understood the training data, then we might better understand how much a model can be reasonably expected to know and can better augment the training data to increase a model's breadth.
Third, language models do not always fail gracefully or in a way that a writer might understand. In our study, when reflecting on sparks from a custom prompt that were tautological or redundant, P1 said, "maybe it's that either way you slice it `atmospheric circulation patterns' is not the best way to start when you're talking to a general audience." In this case the participant was trying to make sense of why their custom prompt was producing low quality outputs -- sentences that were syntactically correct but contained no real knowledge. They assumed that if the model was unable to generate a reasonable sentence then perhaps the topic was too specialized for her audience. But we are not sure why the system produced redundant results; perhaps the prompt wasn't engineered properly or something else went wrong. Studies could be done on a model's knowledge of the training data to see how a model responds differently when prompted with topics it has varying levels of knowledge of. Then we might be able to methodologically design methods that detect why model produce certain low quality outputs.
\subsection{More experience with language models may improve results for users.}
When asked how the system compared to using a web search engine, many participants noted that they had so much experience with web search that they were more comfortable navigating it to find what they needed. In contrast, when using the system, they were unsure what kinds of prompts would produce the results they wanted. How users craft web search queries, and how they overall use web search engines for information retrieval, is a long and active area of study \cite{yamamoto_exploring_2018, smith_domain-independent_2017}. For instance, advanced users query less frequently in a session, compose longer queries, and click on results further down on the list \cite{white_investigating_2007}.
It's possible that some of the issues participants encountered, for instance outputs that incorrectly interpreted the topic or did not reflect participants desired angle, could be improved if participants had more experience and a better understanding of how to prompt language models. For example, communities interested in text-to-image generative models have been crowd-sourcing "tricks" (such as the appending of a phrase) that help steer generations to their liking. People are discovering workarounds by probing the system, in a way that is separate from the research on better prompt engineering. They are learning how to use the tool and its depth of understanding to their full advantage. No amount of prompt engineering will remove the need for user input, in whatever form that may be, just like no matter how much work is put into delivering better search results, users still have to learn how to craft good queries and build expertise and understanding of the system.
\subsection{Limitations: the effects of personality, writing style, topic, and skill.}
Our user study was intended to be exploratory. The 13 participants came to the writing task with very different levels of experience in science writing, and very different topics for the system to respond to. Future work, and more developed systems, may be able to study these differences more carefully, perhaps with a larger set of participants, such that the impact of things like writing style or topic granularity may be able to be taken into consideration.
Additionally, it was clear that some participants were far more forgiving of system than others. For instance, some participants saw one or two inaccurate or redundant sparks and then considered that prompt or even the whole system to be a "dead end". Others were not phased by seeing sparks that didn't meet their needs. They even found ways to re-purpose unhelpful sparks to be useful, for example by interpreting inaccurate sparks as an indication that their topic may have been too niche or as a signal for how readers might also misinterpret their topic. Work on how users develop mental models of AI systems has shown that some people are more likely to blame the system when something goes wrong, and others are more likely to blame themselves \cite{gero_mental_2020}. Future work may want to study if personality type is a predictor of how useful participants find these kinds of systems, such that it does not remain a confounding factor.
Finally, our system uses just one particular language model with one particular decoding method. Improving the generative abilities of language models is a large and growing area work. We expect that their abilities will continue to improve, and this will obviously change the ways in which people react to them.
\section{Conclusion}
In this work we investigated how to use a large language model to support writers in the creative but constrained task of science writing. We developed a system that generates ``sparks'', sentences about a scientific concept intended to inspire writers. We found that our sparks were higher quality than a baseline system, and approached a human-created gold standard. We also found in an exploratory study with 13 PhD students that participants used the sparks in many different ways: to craft detailed sentences, to get ideas for how to engage the reader, and to better understand common reader perspectives. Finally, we discussed how language models might be used as support tools for writing in the future, what areas of technical development we believe will be fruitful, and how users might learn to interact with language models in a writing context.
\bibliographystyle{ACM-Reference-Format}
\section{Introduction}
New developments in large-scale language models have produced models that are capable of generating coherent, convincing text in a wide variety of domains \cite{vaswani_attention_nodate, brown_language_2020, adiwardana_towards_2020}. Their success has spurred improvements on many tasks, from classification to question answering to summarization \cite{brown_language_2020}, as well as creative writing support \cite{coenen_wordcraft_2021}.
Language models have the potential to be powerful writing tools that can support writers in real-world, high-impact domains. These new models are task agnostic, making them applicable to many tasks without requiring more training, and we believe such models are the future of AI technologies.
Despite their successes, language models continue to exhibit known problems, such as generic outputs \cite{holtzman_curious_2020}, lack of diversity in their outputs \cite{ippolito_comparison_2019}, and factually false or contradictory information \cite{lin_truthfulqa_nodate}. There remain many unknowns about how this technology will interface with people in real-world tasks, such as what interactions best serve writers, how language models can best contribute to different writing forms \cite{calderwood_how_2018}, and how to mitigate the bias that language models encode \cite{bender_dangers_2021}.
To that end, we study how language models can be applied to a real-world, high-impact writing task. In particular, we use a science writing form called “tweetorials” which explain technical concepts on Twitter for a general audience \cite{breu_tweetstorm_2020}. Tweetorials are short explanations of around 500 words which have a low-barrier to entry and are gaining popularity as a science writing medium \cite{soragni_scientists_2019}.
Working on science writing requires a system to demonstrate proficiency within an area of expertise. This is much more difficult than traditional creative writing tasks, such as stories and poetry, which tend to deal with common objects and relations. Thus, we present the following research question:
\begin{quote}
\textit{RQ: How can outputs from a language model support writers in a creative but constrained writing task?}
\end{quote}
In this paper, we present a system that aims to inspire domain experts when writing tweetorials on a topic of their expertise. This system provides what we call “sparks” -- sentences intended to spark ideas in the writer. Our system generates sparks using a mid-sized language model (GPT-2 \cite{radford_language_nodate}) and a custom decoding method to encourage specific and diverse outputs. Additionally, we develop a set of 10 prompts based on narratology and expository theory that participants can use when interacting with the model.
We run two evaluations. In the first, we compare the outputs from a custom decoding method to a competitive baseline as well as to a human-created gold standard, reporting on the diversity and coherence of all outputs. In our second study, we have 13 PhD students from five STEM disciplines write tweetorials with our system and report on how they thought about and made use of the sparks.
We make the following contributions:
\begin{itemize}
\item a system that uses a language model to generate “sparks” related to a scientific concept;
\item a custom decoding method for generating sparks from a pre-trained language model;
\item an evaluation
demonstrating that the sparks are more coherent and diverse than an off-the-shelf system, and approach a human gold standard; and
\item an exploratory study with 13 PhD students showing three main use cases of sparks.
\end{itemize}
We end by discussing how to best use language models in constrained writing tasks.
\section{Related Work}
\subsection{Natural Language Generation}
A language model is any model that predicts the likelihood of a sequence of words. This can be used to generate text by giving the model a prefix -- either a sequence of words or a special `start of sequence' word -- and having it calculate the likelihood of all words in its vocabulary as the next word. This can be used to select the next word, and thus generate text. Language models are trained on a dataset of text that does not need to be annotated in any way, as the model can simply be trained on what the next word in a sentence will be \cite{stanford-nlp-textbook}.
Language models are getting larger: they are being trained on more text and the models have more parameters \cite{radford_language_nodate,brown_language_2020, adiwardana_towards_2020}. Much recent work has been on how to make the best use of these large language models, which have shown to be much more general purpose than previous ones \cite{keskar_ctrl_2019}, even showing promise in generating code \cite{austin_program_2021}. Additionally, it is useful to be able to take one language model and use it for many tasks, rather than having to train a new model for each task.
It has been shown that a well-selected prefix, or `prompt', can dramatically increase the performance of a language model on a specific task \cite{reynolds_prompt_2021}. A resulting line of research has been automatically creating either natural language prompts or continuous vector prompts, to perform well on tasks \cite{gao_making_2021, li_prefix-tuning_2021}. However, automatically learned prompts have yet to consistently outperform manually crafted prompts \cite{gao_making_2021}. Additionally there have been very few studies on prompt selection for open ended, generative tasks \cite{coenen_wordcraft_2021}.
Despite the successes of language models, problems remain. When considering the generation of text, language models tend to output repetitive and vague responses \cite{holtzman_curious_2020, ippolito_comparison_2019}. Language models also have no model of the truth; they are learning correlations from large amounts of text. Thus they are able to produce text that includes falsehoods and offensive language \cite{bender_dangers_2021}.
\subsection{Generative Writing Support}
Technological writing support has a long history, but has seen an increase in attention as language models have improved. Early work on language models for creative writing focused on activities such as storytelling \cite{roemmele_writing_nodate} and metaphor writing \cite{chakrabarty_mermaid_2021}. While these tools proved helpful for writers, they were narrow in what they could provide. An exploratory study found that generic auto completion from a language model did not provide enough control for novelists \cite{calderwood_how_2018}. More recent work on writing support for creative tasks has varied the ways in which technology can support the writer, for instance by providing description, plot points, or even asking questions, depending on the desires of the writer \cite{coenen_wordcraft_2021, arnold_generative_nodate}.
Writing support for nonfiction writing tasks tends to be much more constrained, for instance as sentence completion \cite{chen_gmail_2019}. A good example is Gmail's smart reply function, which aims to suggest only text that the writer would have written in anyway \cite{kannan_smart_2016}, although it has been shown that even these suggestions can change what people write \cite{arnold_sentiment_nodate}. Work on helping people craft responses to those in mental health crisis focuses on providing writers feedback, and suggested words, rather than complete phrases or sentences \cite{peng_exploring_2020}.
While natural language generation is a large and growing field, few of its technologies are studied in the context of how they will be used by writers. For instance, although there is much work on automatic summarization \cite{zhang_pegasus_2020, he_ctrlsum_2020}, there's less work on how the summaries might be used by people. Our work aims to study how text generated by language models might be used by writers in a science writing task. There's some relation to a natural language generation task like summarization, because we are concerned with real facts, but we take a human centered approach where the language model provides suggestions, rather than a completed output.
\subsection{Science Communication on Social Media}
Science communication helps the public understand scientific contributions -- consider how it has been applied to tackle vaccine misinformation \cite{shelby_story_2013}, the COVID-19 pandemic \cite{yang_covid-19_2020}, and climate change \cite{hart_boomerang_2012}. Traditionally, science communication took place through journals, conferences, articles, and books -- places where peer review was an implicit part of the publication process. However, the rise of digital networks has made science establish a virtual presence through electronic journals and digital records. The ubiquity of social media further presented opportunities for scientists to have direct channels to the public. Now any scientist can conduct science communication online by posting about their work online \cite{soragni_scientists_2019}, engaging in the `Ask' communities on Reddit \cite{gilbert_i_2020} or explaining something on Youtube \cite{welbourne_science_nodate}. Even PhD students or undergraduate researchers have the ability to disseminate their scientific knowledge at any time without depending on a venue or a publication.
This emerging trend, where the scientist can now partake in conversations outside of an implicitly gated, peer-review process, reflects one of the many broad shifts away from traditional science communication. Scholars of science communication have reified this emerging form of communication as ``post-normal science communication'' \cite{bruggemann_post-normal_2020}. Defining characteristics of post-normal science communication include a tolerance for subjectivity, an insertion of the self, the integration of advocacy, and call to actions. Despite these dramatic shifts, the original tenets of science communication such as storytelling, analogies, figures, and citations remain valuable, and storytelling in particular is a driving principle within our system.
\subsection{Expository and Narrative Theory}
In studying how narratives are embedded in text, we turn to a rich body of literature about narratives and knowledge structure in semiotics and discourse theory. These domains inform our search for structures we could use to prompt language models.
We looked at frameworks for both expository and narrative writing, because tweetorials are a hybrid of both. Specifically, we draw from the constructionist theory for narrative text, discourse theory for narrative text, and discourse theory for expository text.
The constructionist framework of narratology states that all reading comprehension is “a search for meaning'' \cite{Graesser1994}. Readers infer as they build a mental model of why certain actions, events, and states are involved in a situation. The constructionist framework has a classification of inferences that we borrow from for many of our prompt templates. Our prompts exemplify a subset of these inference classes such as case structure role assignment, causal antecedent, the presence of superordinate goals, and instantiation of a noun category.
Concurrently, we examined expository text discourse theory for knowledge structures that would lend well to prompt templates. One framework for expository text introduced a taxonomy of didactic methods (evaluation, explanation, occasion, and expansion) to enumerate the different conversational moves a writer can make to "influence the inference process of the reader" \cite{Tucker1986}. An alternative and popular framework was put forth by Meyer et. al., who enumerated signal phrases that distinguish expository texts, such as ‘specifically’ or ‘attributes of’. We chose to incorporate multiple signal phrases from Meyer’s framework into our prompt templates \cite{meyer_structure_2017}.
\section{Formative Study}
In order to understand how a language model might best support the task of writing a tweetorial, we ran a formative study where participants were first given a technique for coming up with a compelling introduction, before being asked to write the first tweet of a tweetorial on a technical topic they were familiar with. Since the first tweet tends to set up the context and intention of the tweetorial \cite{breu_tweetstorm_2020} we found this to be an effective and efficient way to understand what participants found difficult in the writing process, even when provided with writing strategies.
\subsection{Methodology}
We recruited 10 students in Computer Science.\footnote{6 women / 4 men; 7 undergraduates--no first years / 3 PhD students.}
Participants were required to go through tutorials on how to write an engaging science writing introduction on two example topics -- recursion and virtual private networks -- which included several examples and a step-by-step process for coming up with ideas. These tutorials focused on coming up with an intriguing question for the first tweet, and were developed in consultation with a science journalist.
The process was: 1) brainstorm three concrete situations related to the topic, 2) turn each situation into a question for the reader, 3) select the most engaging question.\footnote{Links to tutorials will be released after anonymous review.} These tutorials were intended to provide the participants with as much “unintelligent” support as possible, such that we could identify where language models may be able to add benefit.
After the tutorials, participants were asked to select a topic from one of six Computer Science topics and write the first tweet for a tweetorial that would explain that topic.\footnote{The topics were: hashing, sorting algorithms, Bayes theorem, HTTP, transistors, and Turning Machines. We selected these topics as ones that a) most computer science students should have learned in a formal setting, and b) could make for an interesting tweetorial.} Participants were asked to think aloud during the writing process. They were not allowed to browse the web. Afterwards, they were asked a series of questions about their writing process in a semi-structured interview.
After all participants had completed the study, the research team reviewed their writing with a science journalist. No formal coding was done, but general areas of success and areas for development were discussed.
\subsection{Results}
\subsubsection{Participants reported that the task required creativity, and that it was difficult to come up with ideas.}
Although we didn’t frame the task as a creative writing task, many participants described the task as difficult because it required creativity to come up with something that would engage the reader. Most participants said they don’t typically do creative writing, so they found the task difficult and outside of their area of comfort. This supported our selection of tweetorials as a writing task, as we want to study a task that is both constrained and creative.
Participants found the tutorials helpful, though for a variety of reasons. Some liked seeing the examples, others appreciated a process to follow, and still others found it comforting to see writing get better with brainstorming and revision. Several commented that the tutorials made the task look easy, but when they began to write about their own topic it was surprisingly difficult to come up with ideas.
Most participants (9 out of 10) said that making the topic interesting to a general audience was the most difficult part of the writing task. When pressed to be more specific, participants mentioned coming up with concrete examples/situations and creating an engaging question as hard tasks. Though this was surely influenced by the process the tutorials introduced, this confirmed that tutorials are not enough to fully support writers in this task.
\subsubsection{Participants struggled to come up with ideas that created suspense.}
When reviewing what the participants had written, all the tweets mimicked the tone of the examples. However, the science journalist had critiques for all of them, and most of the critiques at the core were the same: the tweet lacked suspense. By this he meant, the tweet did not introduce a compelling problem or gap in the reader’s understanding that would make the reader want to read more. Often this was because the example used wasn’t particularly compelling or didn’t reflect a real use case of the topic.
Additionally, participants tended to repeat similar ideas to others who had selected the same topic. For instance, all the people writing about HTTP used either Google or Twitter as their example, suggesting that participants may converge on similar, easy to reach ideas.
Given that participants reported coming up with ideas difficult, it's likely that participants could have done better if given help coming up with more ideas.
Members of the research team also noted that many of the tweets written might be difficult to turn into full length tweetorials. For instance, if the question couldn’t really be answered with an explanation about their chosen topic. For this reason, in future studies we had participants write more than just the first tweet.
\subsection{Design Goals}
Based on our formative study, we developed two design goals for our system:
\subsubsection{Support writers with idea generation.}
Given that language models have no model of truth, we want our system to come up with “sparks”, intended to spark ideas in the writer, rather than having the system provide the ideas themselves. This aligns with prior work on creativity support tools, where users make use of system outputs as initial directions that are then interpreted and diverged from in the users' actual creation \cite{gero_metaphoria_2019}.
Additionally, this also encourages the writer to feel more ownership over their final product, which has shown to be a concern in past work \cite{peng_exploring_2020}.
\subsubsection{Generate outputs that are coherent and diverse.}
In order for writers to make use of outputs, even if they are not always perfectly accurate, they should be coherent -- well-formed and generally reflecting accurate knowledge. Additionally, to support idea generation, outputs should also be diverse, such that writers have a variety of outputs to make use of.
\section{System Design}
\subsection{Generating Sparks}
\subsubsection{Language model selection.}
To generate sparks we use GPT-2, an open source, mid-sized (1.5 billion parameters), transformer language model trained on 40GB of text from the web \cite{radford_language_nodate}. We use the huggingface implementation \cite{wolf_transformers_2020}. While larger open source models are available, e.g. GPT-3 \cite{brown_language_2020} or Megatron-LM \cite{shoeybi_megatron-lm_2020}, we wanted to limit the size of the model we used as larger models are more expensive to run and take more time to generate text. Additionally, there have been many critiques of the super-large language models \cite{bender_dangers_2021}, and thus we wanted to use the smallest language model able to perform well for our use case. Anecdotally, we found that DistilGPT2, a `distilled', smaller version of GPT-2 \cite{sanh_distilbert_2020}, was not able to produce coherent responses to our prompts.
We experimented with fine-tuning GPT-2 on a data set of science writing, but found that this made little difference, especially compared to modifying the decoding method or the prompts. For this reason most of our design effort focused on decoding and prompt engineering.
\subsubsection{Decoding method.}
In addition to selecting a model, we had to design a decoding method -- how to select the next token given the probability distribution the model outputs. There are several common ways of decoding from language models: greedy search, beam search, top-k sampling \cite{holtzman_curious_2020}, and top-n sampling \cite{fan_hierarchical_2018}, to name a few. Different methods have different strengths and weaknesses. Beam search tends to produce high quality results \cite{meister_if_2021} but also tends to produce very similar results for the same prompt. Sampling methods can produce much more varied results, but at the cost of being less coherent. We designed a method that attempts to further increase the coherence of beam search while also increasing its diversity.
First, we modify the probability distribution using a normalized inverse word frequency, in order to increase the likelihood of infrequent words. Normalized inverse word frequency is often used in natural language generation to improve the specificity of outputs \cite{ko_domain_nodate, zhang_learning_nodate}, which is one method for increasing the overall quality of results. To our knowledge this is the first work to use normalized inverse word frequency purely as a decoding method as opposed to during training.
To calculate the word frequencies, we wanted a corpus that doesn't over-represent uncommon science words, like a science writing dataset might, but also reflects modern word usage. For these reasons,
we use a corpus of Vox news articles that includes all articles published before March, 2017.\footnote{\url{https://data.world/elenadata/vox-articles}}
\autoref{fig:logits} shows an example of the probability distribution being modified. In this figure you can see that words like ``governments'', ``Bitcoin'', ``software'', and ``developers'' have an increased weight, while words like ``many'', ``both'', and ``all'', are not modified.
\begin{figure}
\includegraphics[width=.6\textwidth]{fig/logits.png}
\caption{This graph shows how the likelihood of the 20 next most likely words given the prompt "cryptography is used by". The orange line shows the distribution after it has been rewieghted with normalized inverse word frequencies (NIWF). Words like ``governments'', ``Bitcoin'', ``software'', and ``developers'' have an increased probability, while words like ``many'', ``both'', and ``all'' are not modified.}
\Description{adfsdfasdfsdfsad.}
\label{fig:logits}
\end{figure}
Second, we use only the top 50 highest ranking tokens. This is sometimes called top-k sampling, as only the top $k$ tokens are used \cite{fan_hierarchical_2018}. However, since we're not using a sampling method, the effect of this is to ensure that the modified probability distribution doesn't introduce any incoherencies, for example by dramatically increasing the rank of a token very far down in the original probability distributions.
Third, we increase the diversity of outputs by forcing the first token of each output to be unique, but attempt to retain coherence generating the rest of the tokens with beam search. While several more sophisticated methods have been proposed to increase diversity while retaining the coherence of beam search (e.g. \cite{vijayakumar_diverse_2018}), in testing we found none were as effective as simply enforcing the first token to be unique.
Finally, in order to keep the sparks succinct and generating quickly, we only generate 10 tokens after the prompt, and cut off the generation as soon as a sentence has been completed.
We implement our decoding method using the huggingface transformers \cite{wolf_transformers_2020}.\footnote{Link to code to be added after anonymous review.}
\subsubsection{Prompt design.}
Designing prompts for language models has become an active area of research, with many automatic methods being proposed \cite{gao_making_2021, li_prefix-tuning_2021}. However, any automatic method requires at least some training data, and it's yet to be seen that automatically developed prompts can outperform hand-crafted prompts \cite{gao_making_2021}. For these reasons, we hand-craft our prompts.
First we craft a `prefix' prompt to pre-pend to any prompt used by a writer. Prefix prompts have been shown to greatly improve performance by providing the language model with appropriate context \cite{reynolds_prompt_2021}. We found early on in development that simply providing the model with a technical topic was not enough -- also providing a context area was necessary for it to appropriately interpret technical terms. For instance, if you use a prompt like "Natural language generation is used for", the model is likely to talk about linguistic research on languages, rather than computational methods. If instead you use the prompt, "Natural language generation, a topic in computer science, is used by" the results are much more likely to refer to computational language generation. Given this, we pre-pend all prompts with the following:
``\{topic\} is an important topic in \{context area\}''
where \{topic\} and \{context area\} are provided by the writer.
In hand-crafting our prompts, we wanted to make sure our prompts captured a range of relevant angles, so our system could flexibly work with any technical discipline. To do so, we synthesized work from expository and narrative theory into prompts capturing five categories: expository, instantiation, goal, causal, and role. Each category represented an angle that a writer might want to explore. All prompts can be seen in \autoref{tab:prompts}.
\begin{table}
\caption{Prompt templates designed for science writing task.}
\label{tab:prompts}
\begin{tabular}{ll}
\toprule
category & prompt\\
\midrule
expository
&One attribute of \{topic\} is\\
&Specifically, \{topic\} has qualities such as\\
instantiation
&One application of \{topic\} in the real world is \\
&\{topic\} occurs in the real world when\\
goal
&For instance, people use \{topic\} to \\
&\{topic\} is used for\\
causal
&\{topic\} happen because \\
&For example, \{topic\} causes\\
role
&\{topic\} is used by \\
&\{topic\} is studied by\\
\bottomrule
\end{tabular}
\end{table}
We manually developed these prompts according to established frameworks within narrative and expository theory that we referenced in our related work. Our prompts within the categories of instantiation, goal, antecedent, and role were crafted based upon the constructionist framework of inferences, specifically the following categories: case structure role assignment, causal antecedent, the presence of superordinate goals, and the instantiation of a noun category (respectively). Less formally, \textit{Instantiation} prompt templates suggest completions that instantiate where and in what ways topic X may occur in the real world. \textit{Goals} prompt templates suggest completions that represent how topic X is used in the real world. \textit{Causes} prompt templates suggest completions for how topic X might interact in cause and effect chains. \textit{Roles} prompt templates cover entities involved with topic X.
As tweetorials exhibit both elements of narrative and expository writing, we also borrowed signal phrases from Meyer’s framework for expository text \cite{meyer_structure_2017} -- e.g. “specifically", ``such as", ``attribute” -- and folded them within our prompt templates.
In testing we found that participants often wanted to `follow up' on an output by entering in their own prompt. For this reason, we added the ability for writers to add their own prompts, though this prompt would also be pre-pended with our prefix.\footnote{One intriguing area of research is `meta-prompting', where the language model is used to generate the prefix for the next generation \cite{reynolds_prompt_2021}. While we found that this produced intriguing results for our use case, for example by having the model first produce a list of types of people who interact with a topic, and then putting those phrases into a downstream template, we thought it added too much complexity.}
\subsection{Interface}
We design a website that takes in a writer's topic and context area. \autoref{fig:system} shows a screenshot of the system with its important features marked. The website consists of a single textbox for writing, and a `prompt box' above it that allows writers to interact with the sparks. Writers can select a templated prompt from a dropdown menu, or type in their own prompt and add it to the dropdown list. When a prompt is selected, if they press `GENERATE' the language model will generate a single spark. Writers can `star' a spark by clicking on the lightbulb icon -- this fills in the lightbulb and also pastes the spark into the textbox. If a writer selects a different prompt, the sparks already generated are preserved such that if they return to a previous prompt their generated sparks will be shown again.
\begin{figure}
\includegraphics[width=\textwidth]{fig/system-diagram.png}
\caption{Example screenshot of our system generates sparks. A: writers can select from 10 template of prompts in a drop-down menu. B: writers can add their own prompt to the drop-down menu. C: sparks are generated with a lightbulb icon to the left, if writers click the lightbulb it will highlight and the spark is copied into the text area. D: writers can hit the generate button in order to generate a new spark.}
\Description{adfsdfasdfsdfsad.}
\label{fig:system}
\end{figure}
The textbox contains some features useful for the tweetorial writing task. The textbox is split into two sections with a line of dashes. Above the line is reserved for brainstorming and notes, a feature writers requested and found useful during pilot studies. Below the line is the text area for the tweetorial writing. A word count for the writer's tweetorial draft is displayed at the top of the textbox, and a character count for each tweet (separated by line breaks and two forward slashes) is displayed to the left. \autoref{fig:textbox} shows these features with an example from our user study.
The website is implemented using Python 3.7 and the Flask web framework.\footnote{Link to demo to be released after anonymous review.}
\begin{figure}
\includegraphics[width=.6\textwidth]{fig/textbox.png}
\caption{Screenshot of the text area from our user study. At the top is a word count, which counts only the words below the dashed line. Text above the dashed line is interpreted as brainstorming or notes. Participants can separate tweets with a double `//', and the character count for each tweet is shown to its left.}
\Description{adfsdfasdfsdfsad.}
\label{fig:textbox}
\end{figure}
\section{Study 1: Spark Quality}
We wanted to evaluate the quality of the sparks generated by our system. In particular, we wanted to evaluate how well the sparks, in isolation (i.e. not in a writing task), met our design goals of coherent and diverse. We also wanted to test how well the sparks could support a wide range of topics, and if certain prompts supported some topics better than others. To do so, we compared the sparks generated by the custom decoding method to a baseline system, as well as a human-created gold standard.
We have three hypotheses:
\begin{itemize}
\item H1: The custom decoding produces more coherent and diverse outputs than a baseline system, but less coherent and diverse outputs than a human-created gold standard.
\item H2: The custom decoding performs consistently across many different topics.
\item H3: Some prompts work better for some topics.
\end{itemize}
\subsection{Methodology}
We wanted to evaluate the quality of ideas for a variety of topics. We selected three disciplines that have a glossary of terms page on Wikipedia, and that have been demonstrated to be a rich discipline for science writing on social media.\footnote{e.g. \url{https://twitter.com/dannydiekroeger/status/1281100866871648256}, \url{https://twitter.com/GeneticJen/status/897153589193441281}, and \url{https://twitter.com/meehancrist/status/1197527975379505152}} These disciplines were computer science, environmental science, and biology. For each discipline we randomly sampled 10 topics from their glossary of terms page. See the appendix for the full list of topics studied.
\subsubsection{Collecting a human-created gold standard.}
We wanted to collect human responses to our prompts to represent a gold standard or upper limit on the quality of ideas these prompts can generate. To do this, we recruited 2-3 PhD or senior undergraduate students in each discipline and had them complete the same prompts the language model did. Each student was paid \$20/hour for as long as it took them to finish the task.
We explained to them that the purpose of the prompts was to generate ideas to support an expert writing about the topic for a general audience. Each student had to complete 5 prompts per topic in 3 different ways, and was told to make the completions for a given prompt+topic combination maximally different. They were also instructed to ensure their completions were accurate, given their understanding of the topic, and that they could reference the web if they needed to check anything, as well as use web search results for inspiration. Finally, we explained that their ideas should be as concrete and specific as possible. Each student completed 5 prompts for the 10 topics in their discipline, for a total of 5 x 10 x 3 = 150 completions per person. It took them on average 3.5 hours to come up with completions for all 10 topics in their discipline, and in the end we had 6 high quality completions per prompt+topic combination.
\subsubsection{Baseline language model condition.}
We compare the custom decoding to a language model baseline: group beam search with hamming diversity penalty. This is a strong baseline that encourages diversity in the way \cite{vijayakumar_diverse_2018} recommends, and can be implemented using arguments in the ‘generate’ function in the huggingface transformer library. Both the custom decoding and baseline model use the same underlying language model.
\subsubsection{Measuring coherence and diversity.}
Coherence is notoriously difficult to measure automatically, especially without training data -- measures like perplexity merely measure an output's likelihood under the model itself. For this reason we recruited 10 domain experts to annotate outputs for coherence on a 0 - 4 scale, in line with knowledge graph evaluations \cite{li_commonsense_2016}. For biology we had 3 senior undergraduate students majoring in biology; for environmental science we had 2 senior undergraduate students majoring in environmental science; for computer science we had 2 PhD students from the computer science department.\footnote{The students could not have also participated in the generation portion.}
Each discipline had 900 sentences to annotate (300 human generated, 300 from the baseline model, and 300 from the custom decoding).
250 randomly selected outputs from each discipline were annotated by two different domain experts, and the Cohen's weighted kappa was calculated as: $\kappa=.54$ for biology, $\kappa=.51$ for environmental science, and $\kappa=3.4$ for computer science. Given that the agreement was moderate, we had a single annotation for the remaining sentences.
We measure diversity with sentence embeddings \cite{reimers_sentence-bert_2019}, in particular we report the average distance between outputs within a given prompt. A higher average distance means that outputs are more dissimilar, and therefore more diverse.
\subsection{Results}
Overall, the baseline had low diversity and coherence across all disciplines, while the human-created outputs perform much better. \autoref{fig:diversity} and \autoref{fig:coherence} show that the custom decoding method outperforms the baseline, but does not reach the performance of the human-created outputs. For diversity, two-tailed t-tests show this to be a significant difference for all disciplines (computer science: $p<.001$, climate science $p<.001$, biology: $p < .001$); for coherence, mann-whitney U tests show this to be a significant difference for all disciplines (computer science: $p<.001$, climate science $p<.001$, biology: $p < .001$).
\begin{figure}
\centering
\hfill
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{fig/diversity1.png}
\caption{Distribution of diversity, split by discipline. Diversity is measured as the average sentence embedding distance per prompt+topic combination.}
\label{fig:diversity}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.48\textwidth}
\centering
\includegraphics[width=\textwidth]{fig/coherence.png}
\caption{Mean coherence per prompt+topic combination, split by discipline. Each prompt completion was scored by a domain expert on a scale of 0 to 4.}
\label{fig:coherence}
\end{subfigure}
\caption{Diversity and coherence measures across three test disciplines for three conditions: a baseline language model, a language model with the custom decoding, and a human-created gold standard. The custom decoding improves upon the baseline and approaches the human gold standard.}
\label{fig:three graphs}
\end{figure}
\begin{table}
\begin{tabular}{llll}
\toprule
condition & coh & div & One attribute of source code is...\\
\midrule
&&&it is typically written in a human-readable format.\\
human created&4&.38& editability, so that programmers can easily change it to suit their needs.\\
&&&it is a description a computer program.\\
\hline
&&&that it contains code written by humans.\\
custom decoding & 4 &.37&its modularity - code modules contain reusable code components.\\
&&& complexity.\\
\hline
&&&that it can be used as a source of information.\\
baseline model & 2.6&.08&that it can be used as a source of inspiration.\\
&&&its modularity.\\
\bottomrule
\end{tabular}
\caption{Example outputs from our three conditions for a single prompt+topic combination, and the average coherence (coh) and diversity (div) scores for each set of three outputs.}
\label{tab:outputs}
\end{table}
\autoref{tab:outputs} shows some example outputs from each conditions for a single prompt+topic. These examples demonstrate the quality of the human generated outputs: they are long, detailed, and diverse. Comparatively both language model methods are shorter, less specific, and more repetitive. However, the custom method seems to improve the overall quality of the outputs.
\begin{figure}
\centering
\includegraphics[width=.8\textwidth]{fig/prompt-coherence.png}
\caption{This graph shows the coherence per topic for the custom decoding and the human-created gold standard, where 0 is nonsensical or untrue and 4 is generally true. The black dot shows the average coherence of all responses for a given topic, while the colored dots show the average coherence for a given topic per prompt. Topics are ordered by average coherence in the custom decoding. This graph shows that some topics perform much better than others with custom decoding, while the human outputs are generally high quality regardless of topic. It also shows that within a topic there can be a large variation between prompt templates.}
\label{fig:prompt-coherence}
\end{figure}
It is important to acknowledge that the variation in both the diversity and coherence measures are quite large. This means that while on average the custom decoding is an improvement over the baseline, and on average the human-created outputs are better than the language model outputs, for any given prompt+topic combination the output could be very high quality or of a much lower quality. People using the system will not necessarily see this huge variation; they will only see the 10 or so model outputs that they generate.
For this reason, we dig into the variation by topic and prompt. \autoref{fig:prompt-coherence} shows the average coherence per topic for the custom decoding method and the human-created outputs. It plots the average coherence for each topic with the black dots, and the coherence for each prompt+topic combination in the colored dots. From this we can see that the variation in quality over the topics for the custom decoding method. For instance, the "computer security" outputs score an average of 3.7 in coherence, while "automata theory" outputs score 2.1. When looking at the human-created outputs, the quality is far more consistent, with no topics dropping below an average of 3 in coherence.
This demonstrates that our system works well for some topics and less well for others. While we expected that our system would not perform as well as a human would, we did expect that the system would perform more consistently across topics. It is unclear why the language model performs significantly better on some topics, and given the way that these language models are trained it is difficult to inspect or even predict how well the model will perform on a given topic.
\autoref{fig:prompt-coherence} also shows that some prompt templates work better for some topics than others. In our system, the quality of outputs vary significantly with the prompt template. In the human generated outputs, the variation is smaller, but still we see some range. For instance, let's look at the topic "dynein", the worst performing topic. The prompt "Dynein happens because" scores an almost 0 on the 0 to 4 coherence scale, while the prompt "One attribute of dynein is" scores a 3. Dynein is a family of proteins important in cell behavior. Given this, it makes sense that the system is more likely to produce coherent outputs on attributes, rather than why this family of proteins "happens".
However, it's notable that the human outputs scored 3 or above for all prompts for "dynein". Here is a human output about why dynein happens: "Dynein happens because organelles, such as the Golgi complex, need to be positioned in cells." Though this sentence structure is a little convoluted, it's clear that the human was able to compensate for the prompt and still write something coherent and meaningful.
Seeing the difference a prompt template can make highlights the importance of using a prompt that works well for the topic. Since we wanted to test our system with unseen topics, we ensure that participants can add their own prompts in case the templated prompts don't work well for their topic.
\section{Study 2: User Evaluation}
We evaluated how our system supported PhD students in writing tweetorials. Tweetorials are short explanations of around 500 words which have a low-barrier to entry and are gaining popularity as a science writing medium \cite{soragni_scientists_2019}. We use PhD students as they are eager to participate in science writing \cite{howell_engagement_2019} and many tweetorials are already written by PhD students, demonstrating that this a writing task our participants may conceivably want to engage in on their own. This study was approved by the relevant IRB.
\subsection{Methodology}
We recruited 13 participants, all students from five different STEM disciplines, to write tweetorials on a topic of their own choice, related to their area of study. By letting the participants pick their own topic, we ensured that they were writing within their area of expertise, and we were able to test our system on unseen topics.
Participants were first asked to read an introduction to tweetorials, which explained what tweetorials are and walked through an example tweetorial. They were then introduced to the system and watched a short video that demonstrated the system's features, and showed an example use case of the system when writing about 'machine learning algorithms'. Participants could ask clarifying questions to the facilitator. If participants asked to learn more about how the system worked, the facilitator said that it was an algorithm that could generate text in response to a prompt, and that they could discuss the system further after they completed the writing task.
At this point the participant was asked to pick a topic to write about, as well as provide a `context area' to aid the system in correctly interpreting their topic. Then they were given 15 - 20 minutes to interact with the system and write approximately the first 100 words of their tweetorial. Mouse clicks and key presses while the participant interacted with the system were collected, as well as all sparks generated.
After this, the participant filled out a short survey and partook in a semi-structured interview with the facilitator. The survey questions and the questions that structured the interview can be found in the appendix. The study took about an hour and participants were compensated \$40 USD for their time.
Participant interviews were transcribed and the authors performed a thematic analysis \cite{cooper_thematic_2012} on the interview transcripts. The analysis centered on three areas: how sparks were helpful, how sparks were unhelpful, and ownership concerns in response to writing with a machine. Relevant quotes were selected from the transcripts and collated in a shared document, where is the author discussed and collected the quotes into emergent themes.
\subsection{Results}
We report on participant demographics and topic selection in \autoref{tab:participants}, as well as the prominent themes that emerged through our analysis in \autoref{tab:themes}. Our thematic analysis covered three main areas: ways in which sparks were helpful, ways in which sparks were unhelpful, and how participants felt about incorporating sparks into their writing. For each area, we report themes with prevalence greater than 20\%. Prevalence is measured by the number of participants who brought up that theme in their interview.
In the case of ownership, there was a very high variability in responses, such that no one theme had over 20\%. For this reason we report the reasons that people brought up in the text, but do not list them in the table. Behavioral data, such as time spent writing and how many sparks were starred, are also reported and timelines for each participants' activity can be found in \autoref{fig:timeline}.
\begin{table}
\caption{Participant Demographics. Low = once a year or so. Med = Once a month or so. High = once a week or so.}
\label{tab:participants}
\begin{tabular}{ccC{2.3cm}cc}
\toprule
ID & Discipline & Science Writing (general / twitter)&Topic&Context Area\\
\midrule
P1 & Climate Science & Low / Low & rainfall variability & climate science \\
P2 & Climate Science & Low / Never & predicting climate change & climate science\\
P3 & Climate Science & Never / High & sea level change & geophysics\\
P4 & Climate Science & Low / Low & glacier retreat over the holocene & paleoclimate\\
P5 & Computer Science & Low / Never & computationally hard problems & computer science\\
P6 & Computer Science & Never / Never & pseudorandomness & theoretical computer science\\
P7 & Political Science & Med / Med & document embeddings & natural language processing\\
P8 & Psychology & Never / Low & regulatory fit & psychology\\
P9 & Psychology & Low / Low & motivated impression updating & social psychology\\
P10 & Public Health & Low / Low & measurement of sexism & sociology\\
P11 & Public Health & Never / Never & logistic regression & epidemiology\\
P12 & Public Health & Low / Never & deprivation indices & public health\\
P13 & Public Health & Med / Med & threat multiplier & environmental health\\
\bottomrule
\end{tabular}
\end{table}
\renewcommand{\arraystretch}{1.25}
\begin{table}
\caption{Thematic analysis}
\label{tab:themes}
\begin{tabular}{|p{4.5cm}cp{8cm}|}
\hline
\textbf{Code} & \textbf{Prevalence} & \textbf{Example Quote}\\
\hline
\multicolumn{3}{c}{\textit{reasons sparks were helpful}}\\
\hline
Crafted concise, detailed sentences. &54\%&
Most of the time it [the system] was articulating the ideas that were already in my head in a way that's short and concise.\\
Came up with ideas or angles. &46\%&
It [the system] reminded me, Oh, it's not just my application, there's these other people using the same technology, but working on other problems. \\
Showed reader perspectives. &31\%&
It [the system] reminded me that there might be a more common understanding of this thing that I'm writing about, that's different from the highly specific one I've been living in.\\
\hline
\multicolumn{3}{c}{\textit{reasons sparks were unhelpful}}\\
\hline
Incorrectly interpreted the topic. &38\%&
It [the spark] just wasn't helpful, but only because it was using the different sense of `embedded'.\\
Inaccuracies.&23\%&
Some of the sparks said, like, logistic regressions are used to estimate relative risks, which is completely not true.\\
Not desired angle.&23\%&
Someone probably does really care about measuring sexist attitudes ... but it just isn't my focus.\\
Vagueness.&23\%&
I would say about 20\% of them were just not specific enough to warrant talking about.\\
\hline
\end{tabular}
\end{table}
\subsubsection{Participant demographics and topic selection.}
The 13 participants came from five STEM disciplines, with the most common disciplines being Climate Science and Public Health. All but one were doing PhD (the remaining doing a research Master's) and varied from their 2nd year to their 7th year in their program. Participants were asked how often they wrote about technical topics for a public audience, and how often they did so on Twitter. Most participants rarely or never did so, though a few did so on a monthly or even weekly basis.
Participants were asked to select a topic they understood well that was related to their research. The facilitator attempted to aid participants in selecting a topic that wasn't too broad, but also not too specific, but as the facilitator did not necessarily have the same expertise as the participant this was at times difficult. Participants selected a wide range of topics, with no overlap. The full demographics and topic selection can be found in \autoref{tab:participants}.
Given the diversity in the participants' topics, how well the system generated sparks on their topics, and how they articulated or responded to questions in the interview, there was a high variability in how participants felt about the system. For this reason, a prevalence of 50\% or above is considered very high. This would mean that over 50\% of participants independently responded in the same way to an open ended question, despite writing about a unique topic and seeing a unique response from the system.
\begin{figure}
\includegraphics[width=\textwidth]{fig/timeline.png}
\caption{Timelines of all participants from the study, with time writing versus time generating sparks marked in different colors. Participants are grouped by their engagement pattern.}
\Description{adfsdfasdfsdfsad.}
\label{fig:timeline}
\end{figure}
\subsubsection{Participants had different engagement patterns when interacting with the system.}
Overall we saw many different ways in which participants interacted with the system. Some participants generated and starred many sparks, while others generated only a handful and starred two or three.
The average number of sparks generated was 17.2 (std=10) and the average number `starred' was 5.1 (std=4.3). Additionally, it did not seem that generating or starring more sparks necessarily meant a participant found the system more helpful. Some participants who starred just two or three sparks talked at length in their interview about how useful the system was. Others who starred more than 10 complained that the system produced only vague or inaccurate outputs.
However, we did see themes of how people tended to interact with the system.
\autoref{fig:timeline} shows all participants' task timeline, split into three archetypal patterns: `prompt and flow', `flow with breaks', and 'shifting'. We talk about each in detail:
\textbf{`Prompt and flow'} participants look at the sparks at the beginning, before they start writing, and then spend the rest of their time writing, never going back to generate more sparks. This kind of participant used ideas to jump start the writing, but once they were writing had no need to return to the sparks. Several participants described getting some kind of inspiration from the sparks and then "flowing through an argument". Once these participants got into the `flow' of writing, they didn't need any further support.
Although these participants don't generate more sparks, they may be using several sparks that they had `starred' earlier. For instance, here are the first three tweets written by P8, where sentences or phrases inspired or highlighted:
\begin{quote}
\small
Have you ever wondered why certain activities just feel right to you? Like you are able to do well and feel good in the process? \hl{Research on regulatory fit helps explain why certain activities feel right versus wrong...} (1/n)
//
\hl{Regulatory fit happens because people can be driven by a range of different fundamental motives, and distinct activities "fit" each of these motives.} For instance, people who are most motivated by their duties and obligations to others prefer activities that (2/n)
//
Involve carefully defending and protecting what they find important and methodically scrutinizing information. In contrast, people who are most motivated by hopes and aspirations prefer activities that involve eagerly making progress and engaging in curious exploration. (3/n)
\end{quote}
Although they wrote all of this without generating more sparks, they did make use of two sparks -- one for the first tweet (``Research on regulatory fit reveals that regulatory fit varies considerably across individuals depending upon factors.") and another for the second (``Regulatory fit happens because individuals behave differently under varying conditions."). So while they did not return to generating more sparks, they did return to the sparks they had previously starred.
\textbf{`Flow with breaks'} style participants are similar, but they pause in the middle of the writing and go back to sparks. This type of participant talked about returning to the sparks when they "would get stuck" or at some kind of inflection point in the writing process. P2 described this as, "there are a bunch of ideas that I need to kind of weave together before using the tool again, when I get more stuck." In both of these types of interactions, we see that the writer is using the sparks to jump-start their writing whenever something makes them pause, but when they are writing they are writing continuously, even if they are drawing on sparks they had earlier selected as potentially useful.
\textbf{`Shifting'} style participants are often moving between writing and generating sparks. They are constantly returning to the sparks for different details, or for help with crafting a sentence. They tended to star a lot of sparks and use a lot of custom prompts, as what they were looking for from the sparks changed as they proceeded through their writing, and a lot of language in their writing draws directly from a spark. For example, consider the start of P1's tweetorial on rainfall variability, where sentences or phrases inspired by sparks are highlighted:
\begin{quote}
\small
Do you like coffee??
//
Want to understand how coffee and climate change are related?
//
\hl{Rainfall variability happens because rain falls differently depending on temperature and humidity.}
//
\hl{Since temperature and humidity vary around the globe, we observe different rainfall patterns across the world.}
//
\hl{For example, rainfall variability may cause extreme precipitation events leading to flooding or a lack of precipitation will lead to drought.}
\end{quote}
The last three tweets shown here are all taken directly from sparks, with minimal changes. It almost seems like collage, where the writer is mostly arranging language that came from the system.
This suggests there are different engagement patterns that people use when interacting with machine-generated suggestions. Although we expected participants would mostly use sparks for the beginning of their tweetorial, we found that participants used sparks at all points of their writing process. In the next sections, we report on the various ways in which participants made use of the sparks.
\subsubsection{Sparks helped participants craft concise and detailed sentences quickly.}
Although we intended the system to inspire participants with new ideas, the most prevalent reason the participants cited for the spark being helpful was for crafting sentences. Many participants remarked that although the sparks were showing them information that they already knew well, it was much faster and easier to draw on language from the sparks than to write a sentence from scratch.
For instance, P12 said:
\begin{quote}
Most of the time [the system] was articulating the ideas that were already in my head in a way that's short and concise, which is useful. Like `deprivation index measures the relative deprivation experienced by an individual relative to others,' that would have probably taken me like three sentences to write, then I'd have to spend time editing it down. And then yeah... this is a lot quicker.
\end{quote}
\autoref{fig:ex1} shows another example in which P12 drew on a spark in order to write a clear and concise definition.
Several participants noted that they often go to Google or Wikipedia simply to get a well-written definition of a topic they understand well. This is something that the system was able to do for them without requiring a click away from the writing interface and incurring a change in context. P7 noted that it did a good job compressing what he would have looked for or found a Google search. P8 noted that all the sparks were similar to what she would have found on Wikipedia or via a Google search, but that they were ``bite-sized" or ``sound bite ready".
\begin{figure}
\centering
\begin{subfigure}[b]{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{fig/ex1.png}
\caption{An example of how a participant in our study used a spark for crafting a detailed sentence. Highlighted text was inspired by the spark.}
\label{fig:ex1}
\end{subfigure}
\hfill
\begin{subfigure}[b]{0.49\textwidth}
\centering
\includegraphics[width=\textwidth]{fig/ex2.png}
\caption{An example of how a participant in our study used a spark to make their tweetorial more engaging to a general audience. Highlighted text was inspired by the spark.}
\label{fig:ex2}
\end{subfigure}
\hfill
\end{figure}
\subsubsection{Sparks reminded participants of other ideas or angles about their topic.}
Several participants noted that the sparks provided good ideas or angles for discussing or introducing their topic. P2 noted that `weather prediction models' were useful entry point to their research. They said, ``that's something within my field that the general public might be more familiar with than what I actually do.''
\autoref{fig:ex2} shows how P4 drew on a spark about the `sea level rise' in order to make their topic of `glacier retreat over the holocene' interesting to the reader. P4 said in their interview, ``It's often hard to figure out how to spin things in ways that feel relevant to people who don't study this," and that the sparks helped her find ways to make her research relevant. P7 said, ``[the system] definitely generated multiple [ideas] that I could have written different tweetorials about."
When asked if the sparks were giving them new ideas, many participants said that the system was helping them get to ideas they likely would have come up with themselves, but faster. For instance P4 said, ``It was definitely faster. I think I would have gotten there, but it would have taken me longer."
\subsubsection{Sparks encouraged participants to think about common reader perspectives.}
Several participants noted that the sparks reminded them of how people reading their tweetorial might be interpreting their topic. For instance P10, who was writing about measuring sexism, noted that many of the sparks talked about sexist attitudes. She said the while that certainly is an aspect of measuring sexism, it isn't the aspect that she actually studies. And so the sparks reminded her that people's main assumptions when thinking about sexism is probably about attitudes and therefore that might be an assumption that she will have to address in her tweetorial.
P5, writing about computationally hard problems, noted that one of the sparks talked about NP-completeness. He said that while at first he thought this might be too technical, it then made him wonder if someone who was reading a tweetorial about computational hard problems might already know at least some of the keywords about this topic. In this way the sparks made him reflect on what knowledge his readers might already have.
\subsubsection{Sparks failed in different ways for different participants.}
The ways in which participants found sparks to be unhelpful varied highly. The most common reason participants said sparks were unhelpful was that they incorrectly interpreted their topic. In this case, the sparks were not necessarily incorrect, but rather they reflected some alternate interpretation of their topic. For instance, P12, who was writing about deprivation indices, said that some of the sparks were about obesity. Obesity has little to do with deprivation indices, but they thought the algorithm may have been associating deprivation with nutrition. Similarly P8, who was writing about regulatory fit, commented on several sparks about government regulation, which is unrelated to her psychology topic, but she assumed the algorithm was simply free associating with the word `regulatory'.
Other reasons participants found the sparks to be unhelpful were factual inaccuracies, dealing with aspects of their topic that they were not trying to explain or that they did not study, and vague outputs. Participants also mentioned that some sparks were nonsensical, tautological, had too much jargon, or were simply "bizarre".
Overall participants varied highly in how useful they found the sparks. Some participants found that the sparks were so low quality that they found the system completely unhelpful. Others said that even though some of the sparks were not helpful the ones that were helpful were so helpful they were unconcerned with a few that didn't make sense or were off-topic.
\subsubsection{While most participants had no ownership concerns for science writing, a few felt apprehensive about accepting machine-generated ideas.}
Most participants had no concerns about incorporating sparks into their writing. Several participants cited being very familiar with the material as a reason for not being concerned. Others said that since they are writing about public knowledge, it was unimportant where their ideas came from. One participant articulated that coming up with ideas is not the hardest part of science writing, but rather putting in the time and energy into building an audience and writing something engaging, so incorporating sparks would simply be one small part of a much larger endeavor that she took on. One participant compared the sparks to searching on Google; another compared it to Grammarly (a grammar-checking service). One participant said that the sparks were simply elaborating on his own idea, and thus he still felt ownership of the material.
A few participants did have ownership concerns. One participant talked about how he considered outreach and science writing to be part of his job as an academic, and thus any system that automated some aspect of this felt like it was taking over something that he found fundamental to his work. Several participants noted that machine writing was simply the future, and while they may have some apprehensions, they knew they would have to get over it.
Across all participants, whether or not they had ownership concerns, they wanted to ensure that anything they wrote was in their own voice and that they did not plagiarize. Several participants brought up that they were unsure exactly where the sparks were coming from, and they wanted to make sure that anything they took from the sparks was adequately changed, to alleviate any concerns about plagiarism.
P2 described this as, "I think if I was using something like this, I would probably never use an entire sentence verbatim. Just because, if you don't know where it's pulling it from... I wouldn't want to run the risk of plagiarizing something accidentally even."
When asked if these edits would make the sparks less helpful, for instance if it meant it would take more time in order to edit what they chose to incorporate, most participants said that sparks would still overall speed up the writing process. But not everyone agreed; some participants thought the need to edit the sparks made them less helpful.
\subsubsection{Custom prompts allowed participants to iterate on how they interacted with the system.}
Many participants noted that they were unsure how to get the best sparks out of the system. Several compared the system to Google but said it was easier to navigate Google because they already knew what kind of queries would get the best results. In the study though, they noted that they were unsure which prompts would produce the best sparks. Several participants even said that some of the sparks that were unhelpful may have been because their custom prompts didn't properly prime the system.
One way participants used the custom prompts was to get the system to better understand their topic.
P13's discussed how the model didn't seem to correctly interpret his topic of "threat multiplier". You can see him iterating on custom prompts in an attempt to produce more accurate results:
\begin{itemize}
\item ``threat multipliers are''
\item ``climate change is considered a threat multiplier because''
\item ``climate change is considered a threat multiplier for health disparities because''
\end{itemize}
Participants also used the prompts to generate more specific sparks.
In the case of P7, the system seemed to understand his topic "document embeddings" well enough, but he was curious about specific aspects of this topic. You can see him looking for these in the following custom prompts:
\begin{itemize}
\item ``Document embeddings help researchers''
\item ``Document embeddings help search engines''
\item ``Linguists can use document embeddings to''
\end{itemize}
Finally, others participants entered custom prompts that seemed to be used less for idea generation, and more as an auto complete. For instance, P12 entered several custom props designed to generate a definition:
\begin{itemize}
\item ``The Townsend Index is''
\item ``In sociology, a deprivation index measures''
\item ``the variables in the Townsend Index are''
\end{itemize}
Overall, we saw many different uses of the custom prompts, and found them to be an important aspect of the system.
\section{Discussion}
\subsection{How to best use large language models in constrained writing tasks.}
Creativity requires both convergent and divergent thinking \cite{taura_creativity_2011}. We expected that the sparks would support participants with divergent thinking, by showing them how their topic might connect to things of interest to readers. And indeed many participants cited this as a main use case of the sparks. However, we found that many participants also found that sparks helped with convergent thinking. Sparks helped participants put the implicit, detailed, and often messy knowledge they had in their mind into concise sentences.
Existing work on large language models for storytelling tend to focus on divergent thinking. Language models provide writers with new ideas for plot points, character attributes, scene descriptions, and the like \cite{sudowrite, aidungeon}. However, what has typically been ignored is the convergent aspect of storytelling, for instance the tying together of plot points into a satisfactory ending. This is likely because of the difficulty language models have had in the past with staying on topic, though this may be changing as language models improve.
One of the reasons that the sparks may have been able to help participants with convergent thinking is because participants already knew what they were trying to explain. If we think of storytelling in response to a writing prompt as starting somewhere and perhaps not knowing the destination, we can think of writing tweetorials as knowing the destination but not where to start. In this case, participants were able to look at the sparks and recognize some of them as a convergence of the ideas already in their mind.
We sometimes saw this when participants would enter custom prompts into the system. They often knew the correct way to complete the prompt, but didn't necessarily know the best phrasing. For instance, P12 entered in the prompts "The Townsend Index is" and "the variables in the Townsend Index are" not because they didn't know what the Townsend Index is or what its variables are, but because they wanted the system to provide a well-crafted sentence for them.\footnote{The authors of this paper did not know what the Townsend Index is. For the curious, it is a measure of material deprivation levels in a population, and it's variables are: unemployment, non–car ownership, non-home ownership, and housing overcrowding.} In this instance we might think of the language model as performing convergent thinking for the writer and letting the writer recognize what is correct and well-phrased.
An area of future development we see is more explicitly using the language model to help writers with different aspects of the writing process, for instance by developing prompts intended to directly spark new ideas and other prompts to return concise definitions or summaries. We are seeing this in the storytelling space, where systems are being developed to prompt plot ideas or specific scene descriptions or focus on rewriting in a different style.
\subsection{Three areas for technical development: better controls, increased breath, and graceful failing.}
However, to support science writing we will need to improve the underlying technology. We see three main areas of technical development that may be fruitful.
The first is improving the controllability of language models. While it is clear that pre-trained language models can generate sentences that span the spectrums of specificity and complexity and take on a number of styles and opinions, it is less clear how to ensure that a language model generates the kind of outputs a writer wants. Prompt engineering seems like a promising avenue for this, especially as prompt engineering can be done in natural language. If prompt engineering is done with vectors, it will be completely opaque to users, but if we are able to control language models through natural language prompts alone, users will have some intuitive understanding as to how to control the language model and are more likely to be able to use this control themselves.
The second is increasing the breadth of a language model's knowledge. Again, it is clear that pre-trained language models contain much world knowledge, but it is also clear that this knowledge is not evenly distributed. There have been several calls to be more careful and explicit with the data that these pre-trained models are trained on \cite{bender_dangers_2021, bandy_addressing_2021, caswell_quality_2021}. If we better understood the training data, then we might better understand how much a model can be reasonably expected to know and can better augment the training data to increase a model's breadth.
Third, language models do not always fail gracefully or in a way that a writer might understand. In our study, when reflecting on sparks from a custom prompt that were tautological or redundant, P1 said, "maybe it's that either way you slice it `atmospheric circulation patterns' is not the best way to start when you're talking to a general audience." In this case the participant was trying to make sense of why their custom prompt was producing low quality outputs -- sentences that were syntactically correct but contained no real knowledge. They assumed that if the model was unable to generate a reasonable sentence then perhaps the topic was too specialized for her audience. But we are not sure why the system produced redundant results; perhaps the prompt wasn't engineered properly or something else went wrong. Studies could be done on a model's knowledge of the training data to see how a model responds differently when prompted with topics it has varying levels of knowledge of. Then we might be able to methodologically design methods that detect why model produce certain low quality outputs.
\subsection{More experience with language models may improve results for users.}
When asked how the system compared to using a web search engine, many participants noted that they had so much experience with web search that they were more comfortable navigating it to find what they needed. In contrast, when using the system, they were unsure what kinds of prompts would produce the results they wanted. How users craft web search queries, and how they overall use web search engines for information retrieval, is a long and active area of study \cite{yamamoto_exploring_2018, smith_domain-independent_2017}. For instance, advanced users query less frequently in a session, compose longer queries, and click on results further down on the list \cite{white_investigating_2007}.
It's possible that some of the issues participants encountered, for instance outputs that incorrectly interpreted the topic or did not reflect participants desired angle, could be improved if participants had more experience and a better understanding of how to prompt language models. For example, communities interested in text-to-image generative models have been crowd-sourcing "tricks" (such as the appending of a phrase) that help steer generations to their liking. People are discovering workarounds by probing the system, in a way that is separate from the research on better prompt engineering. They are learning how to use the tool and its depth of understanding to their full advantage. No amount of prompt engineering will remove the need for user input, in whatever form that may be, just like no matter how much work is put into delivering better search results, users still have to learn how to craft good queries and build expertise and understanding of the system.
\subsection{Limitations: the effects of personality, writing style, topic, and skill.}
Our user study was intended to be exploratory. The 13 participants came to the writing task with very different levels of experience in science writing, and very different topics for the system to respond to. Future work, and more developed systems, may be able to study these differences more carefully, perhaps with a larger set of participants, such that the impact of things like writing style or topic granularity may be able to be taken into consideration.
Additionally, it was clear that some participants were far more forgiving of system than others. For instance, some participants saw one or two inaccurate or redundant sparks and then considered that prompt or even the whole system to be a "dead end". Others were not phased by seeing sparks that didn't meet their needs. They even found ways to re-purpose unhelpful sparks to be useful, for example by interpreting inaccurate sparks as an indication that their topic may have been too niche or as a signal for how readers might also misinterpret their topic. Work on how users develop mental models of AI systems has shown that some people are more likely to blame the system when something goes wrong, and others are more likely to blame themselves \cite{gero_mental_2020}. Future work may want to study if personality type is a predictor of how useful participants find these kinds of systems, such that it does not remain a confounding factor.
Finally, our system uses just one particular language model with one particular decoding method. Improving the generative abilities of language models is a large and growing area work. We expect that their abilities will continue to improve, and this will obviously change the ways in which people react to them.
\section{Conclusion}
In this work we investigated how to use a large language model to support writers in the creative but constrained task of science writing. We developed a system that generates ``sparks'', sentences about a scientific concept intended to inspire writers. We found that our sparks were higher quality than a baseline system, and approached a human-created gold standard. We also found in an exploratory study with 13 PhD students that participants used the sparks in many different ways: to craft detailed sentences, to get ideas for how to engage the reader, and to better understand common reader perspectives. Finally, we discussed how language models might be used as support tools for writing in the future, what areas of technical development we believe will be fruitful, and how users might learn to interact with language models in a writing context.
\bibliographystyle{ACM-Reference-Format}
|
\section{Introduction}
Quantum mechanics does not treat space and time on the same footing, even in
its relativistic version. For instance, a meaningful probabilistic
interpretation of quantum states requires that time evolution is unitary such
that an evolving state is always normalized. Unitarity implies that time will
keep going on forever. Positions in space, by contrast, can easily be limited
to finite regions or periodic boundary conditions, as in the basic examples of
an infinite square well and of a particle required to move on a circle.
Such a distinction between space and time appears to violate relativity, in
particular the general covariance of arbitrary transformations of time and
space coordinates realized in general relativity. In this context, attempts to
combine quantum properties with general relativity in a theory of quantum
gravity have indeed encountered several obstacles collectively referred to as
the problem of time \cite{KucharTime,Isham:Time,AndersonTime}. For instance,
the perpetual nature of time, encoded mathematically in the condition of
unitarity in quantum mechanics, is incompatible with the possibility of the
universe (and therefore time) having a beginning at the big bang, or an end if
the universe happens to collapse in some distant future. Moreover, time in
general relativity is a local coordinate that, in general, need not be defined
in the same way everywhere in space-time even if the universe does not
encounter a physical boundary.
The picture of unitary evolution in which monotonic time seems to be required
to never cease increasing is also add odds with our physical measurements of
time, which are based on periodic phenomena such as planetary orbits, the
moving hands on a clock, or the vibrations of a quartz. (For a detailed
discussion of physical clocks, see for instance \cite{Clocks}.) Time is
measured by periodic processes, but represented mathematically in a linear,
monotonic fashion. There is an interesting dichotomy between periodic and
monotonic phenomena in the context of time, which is hard to resolve because
we have a very intuitive understanding of how we experience time but do not
know well what time is on a fundamental level. We experience time as pointing
from the past to the future without being able to move back, as perhaps
indicated by the second law of thermodynamics. The temporal labels we attach
to events are accordingly based on conventions that imply monotonic behavior:
While the numbers we conventionally assigned to what we call time in a strict
sense (second, minute, hour) reflect the periodic nature of how we measure
brief intervals, the complete date (day, month, year) of an event renders the
assingment monotonic.
The appearance of the traditional periodic processes in measurements of time
is not fundamental but determined by their utility. It is easier to recognize
change by observing a periodic system returning to a fixed state multiple
times, compared with the gradual motion of a monotonic process. In addition,
the compactness of a periodic scale makes it easier to construct portable or
wearable clocks. The prevalence of periodic processes in time measurements
therefore does not imply that time must fundamentally be based on a periodic
process. Then again, the possibility of constructing a monotonic label of time
(and date) by counting the cycles that a hierarchy of periodic processes goes
through, as in our actual time-and-date measurements, shows that the
monotonicity of our experienced time does not imply either that time
fundamentally must be monotonic.
In this situation, any statement about fundamental clocks requires fundamental
physics. Here, we use two main ingredients to argue that time should
fundamentally be based on a periodic process. At the same time, we define what
we mean by a ``periodic process'' in a formulation that does not use a
monotonic background time. First, in order to bring space and time on a more
equal footing in quantum mechanics, a decades-old treatment postulates that
time $t$ should be represented by an operator (or a phase-space coordinate in
a classical theory) just like the spatial position $x$. Such a formulation of
evolution is called relational because it describes how one physical degree of
freedom, $x$, evolves with respect to another physical degree of freedom, $t$,
rather than how a single physical degree of freedom evolves with respect to
some external time parameter. (Of course, in a relational formulation we could
equally well describe how $t$ evolves with respect to $x$, but in keeping with
conventions, we choose to call the reference degree of freedom $t$.) This idea
goes back to Dirac \cite{GenHamDyn1} and Bergmann \cite{BergmannTime} and has
seen much recent interest in the context of combining quantum physics with
relativity and gravity
\cite{GeomObs2,PartialCompleteObs,PartialCompleteObsII,EffTime,EffTimeLong,EffTimeCosmo,MultChoice,ClocksDyn,TwoTimes,BianchiInternal,AlgebraicTime,QuantumRef1,QuantumRef2,QuantumRef3,QuantumRef4,QuantumRef5}.
Secondly, if time as a physical degree of freedom is on an equal footing with
space, or with matter degrees of freedom, it may in general be expected to
have interactions with itself or with other degrees of freedom. In a
fundamental theory, the reference degree of freedom we use to describe time
could be selected from one of the fundamental fields of the standard model of
particle physics, or from a physical degree of freedom that determines the
structure of space-time. If we require that the degree of freedom we call time
is such that it does not have any self-interactions, as in traditional
formulations of relational evolution, this degree of freedom takes on a
special form devised just for the purpose of being able to play the role of
time as we think we know it. Such an assumption would forgo any possibility of
determining fundamental properties of time. More generally, if time is based
on a fundamental clock, it should generically be expected to have
self-interactions, perhaps described by a potential. The reference clock
degree of freedom could then have ``turning points'' which confine its values
to a certain finite range. At this point we also have to address a common
language problem: Our standard experience of a monotonic background time is so
common that it is built into several physics concepts, such as ``turning
points,'' that we still have to refer to even when we try to formulate an
oscillating fundamental clock. The conventional term ``turning points''
assumes motion with respect to some background time, but here we are not
interested in this motion. We merely refer to the confinement to a finite
range that may be implied by a potential and, for lack of an alternative term,
use the established dynamical term. Any process with turning points in this
sense will be referred to as ``periodic.''
In a fundamental description, we should therefore be able to make sense of
relational evolution with respect to a confined or periodic reference degree
of freedom. For the sake of clarity, we will reserve the word ``time'' for a
monotonic label in accordance with our common experience of time. A
non-monotonic reference system on which the measure of time is based will be
called a ``clock.'' The distinction between ``time'' and ``clock'' does not
appear in traditional relational evolution, but it is relevant for a
discussion on a fundamental level as presented in this paper.
We will describe and evaluate relational evolution with respect to a periodic
clock degree of freedom in what follows, demonstrating that it is not only
consistent with standard requirements on quantum mechanics but also implies
new and potentially observable effects, as announced in \cite{LocalTime}.
\section{Relational evolution}
Relational evolution as a formal device, used often in quantum gravity and
quantum cosmology, helps to bring time conceptually closer to space by
accompanying the canonical pair of position and momentum, $x$ and $p$, with a
second canonical pair of energy and time, $E$ and $t$. For consistent signs,
not that time $t$ is the momentum of $E$, or $-E$ is the momentum of $t$, as a
consequence of the usual negative sign in the time components of the Minkowski
metric $\eta_{\mu\nu}$. In 4-dimensional notation, one can write these
canonical relationships through a Poisson bracket,
$\{x_{\mu},p_{\nu}\}= \eta_{\mu\nu}$.
The extended description by canonical variables implies that the corresponding
quantum theory should include a time operator $\hat{t}$, in addition to an
energy operator $\hat{E}$, such that $[\hat{t},\hat{E}]=-i\hbar$. The
classical energy equation $E=H(x,p;t)$ with the Hamiltonian $H(x,p;t)$ of the
system, possibly having an explicit time dependence, is then quantized to the
Schr\"odinger equation by representing the energy as a derivative operator
$\hat{E}\psi=i\hbar \partial\psi/\partial t$, acting on wave functions
depending on $x$ and $t$.
In order to make the relationship between time and space more apparent,
we can formulate the energy equation as a constraint,
\begin{equation}
C_1=-E+H(x,p;t)=0\,,
\end{equation}
moving time and space to the same side of the equation. We initially
introduced a new degree of freedom into the usual formulation, $t$ with
momentum $-E$, and now impose a constraint to make sure that the correct
number of independent parameters is maintained.
In spite of a certain formal semblance between $x$ and $t$ in this
formulation, it does not manage to put space and time on an equal footing.
One remaining difference between these two variables is that the latter's canonical
momentum, $E$, appears linearly in the constraint, while the former's
canonical momentum, $p$, usually appears in a quadratic form. A relativistic
energy equation, such as
\begin{equation} \label{C2free}
C_2=-E^2+p^2+m^2=0
\end{equation}
for a free particle with mass $m$,
helps to reduce this difference.
\subsection{Classical formulation}
The reformulation of standard Hamiltonian evolution as a constraint linear in $E$
does not change the assumption that time is monotonic. A constraint generates
equations of motion with respect to an auxiliary parameter $\epsilon$ in the
same form as a Hamilton function generates Hamilton's equations in time. For
instance, the equations generated by $C_1$ are
\begin{eqnarray}
\frac{{\rm d}x}{{\rm d}\epsilon}&=& \frac{\partial C_1}{\partial
p}=\frac{\partial H}{\partial p}\,,\\
\frac{{\rm d}p}{{\rm d}\epsilon}&=& -\frac{\partial C_1}{\partial x}=
-\frac{\partial H}{\partial x}\,,\\
\frac{{\rm d}E}{{\rm d}\epsilon}&=& \frac{\partial C_1}{\partial t}=
\frac{\partial H}{\partial t}\,,\\
\frac{{\rm d}t}{{\rm d}\epsilon}&=& -\frac{\partial C_1}{\partial
E}=1\,. \label{dtdl}
\end{eqnarray}
The last equation implies that $t$ is monotonic with respect to $\epsilon$; in
fact, it can be identified with $\epsilon$ up to a constant shift. The
remaining equations then obtain their usual Hamiltonian form, demonstrating
the equivalence of the Hamiltonian and constrained formulations.
Equation~(\ref{dtdl}), derived from the non-relativistic constraint, means
that time in non-relativistic mechanics is, unlike space, inevitably given by
a function that is monotonic in $\epsilon$, and therefore perpetual.
The relativistic constraint introduced so far, $C_2$ in equation
(\ref{C2free}), implies a similar monotonic behavior of time. Initially,
\begin{equation} \label{tepsilon}
\frac{{\rm d}t}{{\rm d}\epsilon}= -\frac{\partial C_2}{\partial E}=2E
\end{equation}
is not just a numerical constant. However, the constraint $C_2$ for a free
particle implies, via ${\rm d}E/{\rm d}\epsilon=\partial C_2/\partial t=0$,
that $E$ is constant. The time variable $t$ is therefore still monotonic in
$\epsilon$, although its rate of change, given by $2E$, is no longer universal
but depends on the energy.
This observation shows how we can move closer to a local notion of time based
on a periodic clock: If we find a relativistic model in which there is a
time-dependent potential added to $C_2$, the energy will no longer be
constant. If its value can move through zero and change sign,
${\rm d}t/{\rm d}\epsilon=2E$ would change sign, and $t$ might oscillate for a
suitable potential. These oscillations would be with respect to an external
parameter $\epsilon$, but this parameter is only auxiliary because it can
locally be eliminated from solutions. Yet, in spite of this auxiliary nature
of the parameter in which oscillations may unfold, dynamics with respect to an
oscillating clock degree of freedom representing time would be markedly
different from a monotonic time $t$ because of the presence of a potential,
which we take as the defining feature of an oscillating clock. Recall that
this definition only refers to the form of the constraint and does not require
a background time.
Systems with time-dependent potentials in a relational interpretation
indeed exist in fundamental physics. For instance, the cosmological dynamics
of an expanding universe on large scales is determined by the Friedmann
equation
\begin{equation}
\left(\frac{1}{a}\frac{{\rm d}a}{{\rm d}t}\right)^2= \frac{8\pi G}{3c^2}\rho\,,
\end{equation}
for the scale factor $a$, whose time derivative is related to the energy
density $\rho$ of matter, Newton's constant $G$, and the speed of light
$c$. It can be formulated as a constraint
\begin{equation} \label{FriedmannConstraint}
C_3= - H(\phi,p_{\phi};V)+\frac{6\pi G}{c^2} Vp_V^2=0
\end{equation}
in canonical variables \cite{Foundations} given by the expanding volume, $V$,
and its momentum, $p_V=-c^2{\cal H}/(4\pi G)$ related to the Hubble parameter
${\cal H}=a^{-1}{\rm d}a/{\rm d}t$. The matter variables are often described
by another canonical pair, $\phi$ and $p_{\phi}$, which appear in the matter
Hamiltonian $H=V\rho$. A common example of an isotropic matter degree of
freedom $\phi$ is a scalar field with mass $m$, in which case the energy
density equals
\begin{equation} \label{rho}
\rho=\frac{c^2}{2} \frac{p_{\phi}^2}{V^2}+ \frac{1}{2}m^2\phi^2\,.
\end{equation}
For $m=0$, equations of motion imply that $p_{\phi}$ is conserved and $\phi$
is monotonic, much like $E$ and $t$ as determined by the previous constraint,
$C_2$. But the more generic case of $m\not=0$ leads to a representation of
time through a clock degree of freedom $\phi$ that evolves in a non-monotonic,
periodic fashion.
Simplifying some coefficients in the cosmological example, we will now work
with a constraint of the form
\begin{equation} \label{C4}
C_4= -p_{\phi}^2-\lambda^2 \phi^2+ H(x,p)^2
\end{equation}
with some system Hamiltonian $H(x,p)$, keeping the quadratic dependence on
$\phi$ and $p_{\phi}$ but applying it to non-cosmological models. For
$\lambda=0$, in which case $\phi$ is not confined or periodic, we can
factorize the quadratic constraint into two factors linear in $p_{\phi}$ such
that solutions to the constraint equation $C_4=0$ are equivalent to solutions
of the non-relativistic constraint $C_1=0$ with $p_{\phi}=\pm E$. For
$\lambda\not=0$, $\phi$ is confined for a given system energy, allowing us to
generalize monotonic time behavior to an oscillating clock, $\phi$. Imposing
the constraint $C_4=0$ then couples the system degrees of freedom, $x$ and
$p$, to the clock degrees of freedom, $\phi$ and $p_{\phi}$. Even though there
is no force between system and clock for a constraint of the form (\ref{C4}),
their dynamics are related by the energy-balance constraint $C_4=0$. The
absence of a coupling force relieves us from the obligation to justify any
specific form from fundamental physics. If there were such a coupling force in
addition to the energy-balance constraint, it would only strengthen the
effects of clock-system interactions that we will observe in what follows.
\subsection{Quantum formulation}
Quantum cosmology aims to quantize constraints such as
(\ref{FriedmannConstraint}) or the simpler (\ref{C4}) by solving and
interpreting the quantum constraint equation
$\hat{C}_4\psi(x,\phi)=0$. However, solutions of this equation for the wave
function $\psi$ do not evolve in an obvious way because the imposed quantum
constraint implies that the evolution operator associated with $\hat{C}_4$,
given by $\exp(-i\hat{C}_4\epsilon/\hbar)$ where $\epsilon$ is analogous to
the auxiliary parameter of the same letter used in the classical formulation,
acts trivially on solutions $\psi(x,\phi)$ of $\hat{C}_4\psi(x,\phi)=0$.
\subsubsection{Ordering questions}
A common method to address this problem goes back to Dirac \cite{GenHamDyn1},
called deparameterization and implemented in detail in \cite{Blyth} for
quantum cosmology. This method, which assumes $\lambda=0$ in (\ref{C4}) or
$m=0$ in (\ref{rho}), amounts to an inversion of the process that led us from
Hamiltonians to constraints: We factorize the quantum constraint equation as
\begin{equation} \label{Factor}
\hat{C}_4^{(\lambda=0)}\psi=
(-\hat{p}_{\phi}^2+ H(\hat{x},\hat{p})^2)\psi=
(-\hat{p}_{\phi}+H(\hat{x},\hat{p}))(\hat{p}_{\phi}+H(\hat{x},\hat{p}))\psi=0\,,
\end{equation}
such that it can be solved by either of the parentheses being zero when acting
on $\psi$: $\hat{p}_{\phi}\psi=\mp H(\hat{x},\hat{p})\psi$, or
\begin{equation}\label{pmSchroed}
i\hbar\frac{\partial\psi}{\partial\phi}= \pm H(\hat{x},\hat{p})\psi\,.
\end{equation}
We have obtained Schr\"odinger evolution (for both choices of the orientation
of time) from the quantum constraint.
For $\lambda\not=0$, however, the procedure suggested by Dirac does not go
through precisely because time is no longer monotonic. Deparameterized
evolution with respect to $\phi$ then cannot be unitary because the classical
$\phi$ oscillates. A further difficulty appears at a formal level, noting
that the factorization (\ref{Factor}) is not correct if a Hamiltonian
$\hat{\bar{H}}$, such as a quantization of $\bar{H}=
\sqrt{H(x,p)^2-\lambda^2\phi^2}$ which classically solves $C_4=0$ for
$p_{\phi}$, depends on $\phi$. Because $[\hat{p},\hat{\bar{H}}]\not=0$,
\begin{equation} \label{Factor2}
\hat{C}_4':=(-\hat{p}_{\phi}+\hat{\bar{H}})(\hat{p}_{\phi}+\hat{\bar{H}})=
-\hat{p}_{\phi}^2-[\hat{p}_{\phi},\hat{\bar{H}}]+ \hat{\bar{H}}{}^2=
-\hat{p}_{\phi}^2+ \hat{\bar{H}}{}^2+ i\hbar
\widehat{\frac{\partial\bar{H}}{\partial\phi}}
\end{equation}
does not agree with the constraint $-\hat{p}_{\phi}^2+\hat{\bar{H}}{}^2$; see
also \cite{EffTime,EffTimeLong}.
The commutator term, being proportional to $\hbar$, could be interpreted as a
quantum correction (although a complex-valued one), modifying the classical
constraint $C_4$. The modified constraint equation $\hat{C}_4'\psi=0$ can then
be solved by $-\hat{p}_{\phi}\psi=\hat{\bar{H}}\psi$, using the rightmost
factor in (\ref{Factor2}) next to the wave function $\psi$ in
$\hat{C}_4'\psi=0$. However, if $\phi$ is a local oscillating clock, both
factors in a version of (\ref{Factor}) are required for forward and backward
evolution with respect to $\phi$. But exchanging the factors in
(\ref{Factor2}), such that $-\hat{p}_{\phi}+\hat{\bar{H}}$ now acts directly
on a wave function $\psi$, modifies the constraint:
\begin{equation} \label{Factor3}
\hat{C}_4'':=(\hat{p}_{\phi}+\hat{\bar{H}})(-\hat{p}_{\phi}+\hat{\bar{H}})=
-\hat{p}_{\phi}^2+[\hat{p}_{\phi},\hat{\bar{H}}]+ \hat{\bar{H}}{}^2=
-\hat{p}_{\phi}^2+ \hat{\bar{H}}{}^2- i\hbar
\widehat{\frac{\partial\bar{H}}{\partial\phi}} \not=\hat{C}_4'\,.
\end{equation}
It therefore seems impossible to include both signs in a quantized
$p_{\phi}=\pm \bar{H}$ for a unique quantum model, based on a single quantum
constraint.
\subsubsection{Gribov horizons}
In a classical treatment, as $\phi$ evolves through its turning points in the
quadratic potential $\lambda^2\phi^2$, the sign of $p_{\phi}$
alternates. Different half-cycles of this periodic evolution are therefore
governed by not just one but both factors in the classical version of
(\ref{Factor}). The factorization of quadratic quantum constraints appears to
be in conflict with this elementary behavior. However, quantum mechanics is
more subtle. As shown in \cite{Gribov}, the problem of time is a special case
of the Gribov problem of gauge theories \cite{GribovQuant,GribovRev}, where
the $\epsilon$-flow generated by the constraint plays the role of the gauge
flow, and selecting a variable such as $\phi$ as time fixes the gauge as long
as $p_{\phi}\not=0$: Setting $\phi=\tau$ to a constant value $\tau$ of a
global time parameter then gives a cross-section of the flow. When
$p_{\phi}=0$ at a turning point of $\phi$, $\phi$ is at an extremum and the
condition $\phi=\tau$ is not transversal to the flow. Moreover, because $\phi$
is not monotonic, the condition $\phi=\tau$ evaluated on the full evolution
does not have a unique solution.
These issues are common to all gauge theories with Gribov problems, in which
transition surfaces in phase space, such as $p_{\phi}=0$, are called Gribov
horizons. The usual solution to this problem in quantized gauge theories is to
ensure that Gribov horizons are never crossed on a single gauge orbit in order
to avoid double-counting gauge-fixed solutions in a path integral. In the
present case, this means that quantum mechanics cannot allow $p_{\phi}$ to
change sign on a single gauge orbit. However, since our gauge fixing,
$\phi=\tau$, is time-dependent, we may choose a different Gribov region at
different times, such that we choose the region with $p_{\phi}<0$ when $\phi$
moves forward and the region with $p_{\phi}>0$ for backward motion. (According
to (\ref{tepsilon}), forward motion with respect to $\epsilon$ implies $E>0$,
which corresponds to $p_{\phi}<0$.)
In a canonical treatment and with positive $H$, these two regions correspond
to solutions of the constraint annihilated by the left and right factors,
respectively, in (\ref{Factor2}). Acting on wave functions, we therefore seem
subject to the ordering problem, $\hat{C}_4'\not=\hat{C}_4''$, when $\hat{H}$
is $\phi$-dependent. However, when $\phi$ runs backwards ($p_{\phi}>0$),
evolution with respect to $\phi$ is reversed compared with forward motion
($p_{\phi}<0$). Since time reversal in quantum mechanics is associated with
complex or Hermitian conjugation, we may impose the constraint on wave
functions by acting to the left in this case, $\psi\hat{C}_4'=0$, while using
the standard action to the right for forward motion. A single constraint in a
fixed ordering, $\hat{C}_4'$, can then be used to describe both forward and
backward motion of $\phi$. Implementing this concept formally on a Hilbert
space is subtle because, as we will discuss in more detail, turning points
where $p_{\phi}=0$ are energy dependent according to the constraint
equation. (Algebraic formulations of quantum mechanics \cite{LocalQuant}
that generalize Hilbert-space treatments may be useful in this context, as
they turned out to be in other questions about time as well \cite{AlgebraicTime}.)
Time reversals therefore happen at different times for different energy
eigenstates that are superimposed in an evolving wave function. Our specific
constructions will demonstrate that such a formulation is meaningful and
feasible, but we postpone a detailed general discussion of such time-reversal
states to later work.
\subsubsection{Clock and time}
The specific implementation of cycles in which a local clock such as $\phi$
may move forward or backward, solving the problem of oscillating clocks, is
perhaps obvious, with hindsight, but it has been noticed only recently
\cite{Gribov}: We should distinguish carefully not only between background
time $\epsilon$ and a clock variable $\phi$, as formalized by
deparameterization which would also identify $\phi$ with time, but rather
between {\em three} conceptually different notions: background time
$\epsilon$, a clock variable $\phi$, and (as a new ingredient) global
monotonic time $\tau$. The roles of $\phi$ and $\tau$ are indistinguishable in
the usual treatment of deparameterization in which $\phi$ is monotonic and
can be assumed to be identical with $\tau$. If $\phi$ has turning points,
however, it can be identified with a linear function of $\tau$ only locally,
for periods of evolution that do not contain a turning point of $\phi$. After
a turning point, we should realign the relationship between $\phi$ and $\tau$
such that $\tau$ keeps on going forward while $\phi$ moves back. For instance,
if $\phi(\tau)=\tau+A$ before a turning point at $\tau_{\rm t}$ (with some
constant $A$), $\phi(\tau)=-\tau+A+2\tau_{\rm t}$ rewinds $\phi$ in a way that
is connected continuously to $\phi(\tau)$ before the turning point.
The new distinction between three types of variables related to time has
conceptual implications that will not be the focus of this paper, and which
we mention only briefly in this paragraph. For instance, there may be no time
operator because time $\tau$, as outlined below, is a constructed, effective
parameter and not fundamental. There would only be a clock operator
$\hat{\phi}$ which can measure the direction in which $\phi$ points in its
cycle, but not which cycle it is in. The number of cycles, and therefore time,
would have to be determined by keeping track of a suitable succession of clock
measurements. It would not be possible to determine time by a single
measurement because, unlike the clock, it is not represented by a fundamental
degree of freedom. As we will also see, a quantum clock is generically in a
superposition of different cycles, such that the cycle is not a sharply
defined observable.
Since classical physics, not restricted by the condition of unitarity, can
easily be formulated with local times, we usually do not have to introduce a
time parameter such as $\tau$. In this context, the transition from a
background parameter $\epsilon$ to a clock variable $\phi$ is often motivated
as moving a step closer toward a fundamental description of time, no longer
given by a mathematical coordinate but rather by a physical measurement by
means of a clock. The value of $\phi$ is then the position of a periodic
phenomenon used as a clock, modeled in our example with regular periods by the
harmonic-oscillator Hamiltonian $p_{\phi}^2+\lambda^2\phi^2$ added to our
constraints. However, observing $\phi$ no longer corresponds to a physical
measurement of time. It is merely the reading of a clock instant without putting
it into the context of a constructed time-and-date label of events. The
variable $\phi$, like a Cartesian coordinate used to determine the position of
the hands on a clock, oscillates back and forth, but the time we infer from
this motion always increases. This perpetually increasing time is the global
time $\tau$ introduced here.
Because ${\rm d}\phi/{\rm d}\tau$ changes sign at the turning points of
$\phi$, a parameterization with respect to global time is consistent with
having alternating signs of $p_{\phi}$ (related to our specification of Gribov
regions) even while the energy (of a stable system) should always be positive:
Building on (\ref{pmSchroed}), the Schr\"odinger equation
\begin{equation} \label{Schroedingertau}
i\hbar\frac{\partial\psi}{\partial\tau}= i\hbar\frac{{\rm d}\phi}{{\rm
d}\tau} \frac{\partial\psi}{\partial\phi}= -\frac{{\rm
d}\phi}{{\rm d}\tau} {\rm sgn}(\hat{p}_{\phi})
H(\hat{x},\hat{p},\phi(\tau))\psi
=H(\hat{x},\hat{p},\phi(\tau))\psi
\end{equation}
with respect to global time $\tau$ contains both branches of wave-function
evolution with $\hat{p}_{\phi}=\mp H(\hat{x},\hat{p})$, even if the
Hamiltonian $\hat{H}$ is always positive. Since we have forward motion (now
with respect to our new parameter $\tau$) for $p_{\phi}<0$ and backward motion
for $p_{\phi}>0$, we always obey the condition
$-({\rm d}\phi/{\rm d}\tau){\rm
sgn}(p_{\phi})=1$. Equation~(\ref{Schroedingertau}) is therefore the
standard Schr\"odinger equation with a positive Hamiltonian. Global,
unitary time evolution with respect to $\tau$ is then defined by taking
into account the sign changes of $p_{\phi}$ in strict correspondence with the
sign changes of ${\rm d}\phi/{\rm d}\tau$ for forward and backward motion of
$\phi(\tau)$.
Still, even though (\ref{Schroedingertau}) looks like a standard Schr\"odinger
equation, it depends on a piecewise linear but not strictly linear
parameterization $\phi(\tau)$ and therefore implies new features. At turning
points of $\phi$, the dependence of $\phi$ on $\tau$ changes abruptly, which
can be implemented in solutions by concatenating evolution operators
$\exp(-i\int \hat{H}{\rm d}\tau/\hbar)$ derived for strictly linear branches
of $\phi(\tau)$, or their transition amplitudes. The sudden changes in a wave
function constructed from such concatenated evolution operators imply that the
time-dependent phase lacks smoothness, but continuity and unitarity are never
compromised.
While this procedure introduced in \cite{Gribov} presents a solution to some
aspects of the problem of time, the first to make sense of a local notion of
an oscillating clock variable $\phi$, the question of its physical viability
remained open. For instance, one might worry that the rather sudden changes of
$\tau$-evolution operators at turning points, and correspondingly of the phase
of the wave function, could destroy coherence faster than in standard quantum
mechanics with a background time. They might then be in conflict with
sensitive experimental observations, for instance in atomic clocks. In
particular, the turning-point condition $p_{\phi}=0$ is met at different clock
values $\phi$ for different system energies because of the constraint, and
therefore at different global times $\tau$. The independent energy
contributions in a coherent state would therefore be affected differently by
turning points, endangering their delicate balance required for long-term
coherence.
In the next section we will show that this concern is unwarranted: Coherence
remains intact over long time scales, provided the fundamental clock is
sufficiently fast, with durations of cycles much shorter than the typical rate
of change of the non-time observable, $x$. The procedure of oscillating clocks
is therefore physically viable, and it is testable by measurements of quantum
coherence. Detailed calculations presented in what follows impose a tight
upper bound on the possible fundamental period of time.
\section{Global evolution}
Our formal results are valid for a constraint of the form (\ref{C4}) with an
arbitrary Hamiltonian $H(x,p)$ of a bound-state system. Since any initial
state can be written as a superposition of eigenstates of
$\hat{H}=H(\hat{x},\hat{p})$, it is sufficient to compute evolving wave
functions or transition amplitudes by solving the ordinary differential
equations
\begin{equation} \label{SchroedingerE}
i\hbar \frac{{\rm d}\psi_k(\phi)}{{\rm d}\phi}=\pm
\sqrt{E_k^2 -\lambda^2\phi^2} \ \psi_k(\phi)\,,
\end{equation}
in the energy representation, where $E_k$ is one of the energy eigenvalues of
$\hat{H}$ and $\psi_k$ the corresponding eigenfunction. This differential
equation is straightforward to solve, giving
\begin{equation} \label{psik}
\psi_k(\phi) =\psi_k (0)
\exp \left( \mp\frac{i }{2 \hbar} \left(
\phi
\sqrt{E_k^2-\lambda^2\phi^2}+\frac{E_k^2}{\lambda}
\arcsin\left(\frac{\lambda\phi}{E_k}\right) \right)
\right)\,.
\end{equation}
(Without loss of generality, we assume that the initial phase of
$\psi_k(\phi)$ with respect to the energy eigenstate $\psi_k$ vanishes.)
As a function of the energy eigenvalues, the phase function
\begin{equation} \label{Theta}
\Theta_k(\phi)= -\frac{1}{2 \hbar} \left(
\phi
\sqrt{E_k^2-\lambda^2\phi^2}+\frac{E_k^2}{\lambda}
\arcsin\left(\frac{\lambda\phi}{E_k}\right)\right)
\end{equation}
can be used for any bound-state system, provided the clock Hamiltonian is
given by $p_{\phi}^2+\lambda^2\phi^2$.
\begin{figure}
\centering
\includegraphics[width=0.6\textwidth]{Figures/Phipphi.eps}
\caption{Harmonic clock degree of freedom in phase
space. \label{Fig:Phipphi}}
\end{figure}
When $E_k^2 < \lambda^2\phi^2$, Eq.~(\ref{psik}), taken at face value,
produces a wave function that is not normalized, highlighting the unitarity
problem of deparameterization with an oscillating clock. Unitarity starts
being violated precisely when $\phi=\pm\phi_{\rm t}$, where
\begin{equation} \label{phit}
\phi_{\rm t}=\frac{E_k}{\lambda}\,,
\end{equation}
reaches a turning point corresponding to the energy $E_k$. For later
reference, we illustrate the phase-space trajectory of the clock degree of
freedom in phase space in Fig.~\ref{Fig:Phipphi}. For a complete cycle, we
clearly need both positive and negative $p_{\phi}$.
\subsection{Unwinding time}
In order to solve the unitarity and sign problems, following \cite{Gribov}, we
introduce a monotonic global time $\tau$ related to the clock variable $\phi$
in a continuous and piecewise linear fashion:
\begin{equation} \label{phitau}
\phi(\tau)=\left\{\begin{array}{cl}
\tau -4n\phi_{\rm t} & \mbox{if }\quad 4n-1 \leq \tau/\phi_{\rm t} \leq
4n+1\\
(4n+2)\phi_{\rm t}-\tau & \mbox{if }\quad 4n+1 \leq \tau/\phi_{\rm t} \leq
4n+3
\end{array}\right.\,.
\end{equation}
Here, the integer
\begin{equation} \label{n}
n =\left \lfloor \frac{1+\tau/\phi_{\rm t}}{4} \right \rfloor
\end{equation}
equals the number of clock cycles, starting with $n=0$ at $\tau=0$. This
parameterization, illustrated in Fig.~\ref{Fig:phitau}, is constructed such
that $\phi$ (i) is related to $\tau$ in a piecewise linear fashion with equal
rates for $\phi$ and $\tau$ (${\rm d}\phi/{\rm d}\tau=\pm 1$), and (ii) never
takes values outside of the range delimited by its turning points,
$\pm\phi_{\rm t}$. Time $\tau$ therefore progresses at the same rate as the
clock, and it keeps track of the number of clock cycles that have passed while
it unwinds the periodic behavior of the clock.
Implicitly, each energy eigenstate contained in a system state dictates its
own clock period $4\phi_{\rm t}$ through the $\phi_{\rm t}$-dependence in
$\phi(\tau)$, where $\phi_{\rm t}$ depends on $E_k$ according to
(\ref{phit}). For simplicity, we dropped the subscript $k$ in the more
complete notation $\phi_{k}(\tau)$ because we will for some time be working
with individual energy eigenstates. However, when we bring different energy
eigenstates back in superposition, the $E_k$-dependence of $\phi_{\rm t}$
implies that a unique global time $\tau$ for the entire state requires
different energy eigenstates to be at different clock values $\phi$ and in
different cycles. The combined clock-system state therefore evolves into a
superposition of different clock cycles whenever the system is in a
superposition of different energy eigenstates.
\begin{figure}
\centering
\includegraphics[width=0.8\textwidth]{Figures/phitau.eps}
\caption{A periodic clock degree of freedom $\phi$ as a function of global
monotonic time, as constructed in Eq.~(\ref{phitau}). The parameter $n$
counts the clock cycle according to Eq.~(\ref{n}), starting at $\tau=0$,
while $\pm$ indicate the sign of ${\rm d}\phi/{\rm d}\tau$, determining
which one of the two options (\ref{Theta1}) and (\ref{Theta2}) should be
used at a given $\tau$. \label{Fig:phitau}}
\end{figure}
Inserting $\phi(\tau)$ in
$\psi_k(\phi)=\psi_k(\phi_0)\exp(\pm i(\Theta_k(\phi)-\Theta_k(\phi_0)))$,
with some initial $\phi_0$ in a given half-cycle of the clock, results in the
local solutions
\begin{equation}\label{Theta1}
\psi_k(\tau)= \psi_k(\phi_0) \exp(\pm i(\Theta_k(\phi(\tau))-\Theta_k(\phi_0)))
\end{equation}
of (\ref{Schroedingertau}). In any given half-cycle of the clock, the sign in
the exponent has to be chosen such that it cancels the sign
${\rm d}\phi/{\rm d}\tau$ produced by acting on $\Theta_k(\phi(\tau))$ with a
$\tau$-derivative in the Schr\"odinger equation, using the chain rule. This
condition, as introduced before, ensures that $\tau$-evolution is generated by
a positive Hamiltonian for a stable system.
Therefore,
\begin{equation}
\psi_k(\tau)= \psi_k(\tau_0) \exp( i{\rm sgn}({\rm d}\phi/{\rm d}\tau)
(\Theta_k(\phi(\tau))-\Theta_k(\phi(\tau_0))))
\end{equation}
or, in a piecewise description,
\begin{equation}\label{Theta1}
\psi_k(\tau)= \psi_k(\tau_1) \exp\left( i( \Theta_k(\phi(\tau))-
\Theta_k(\phi(\tau_1)))\right)
\end{equation}
during any half-cycle with ${\rm d}\phi/{\rm d}\tau>0$ (starting at some
$\tau_1$), while
\begin{equation}\label{Theta2}
\psi_k(\tau)= \psi_k(\tau_2) \exp\left( i( \Theta_k(\phi(\tau))-
\Theta_k(\phi(\tau_2)))\right)
\end{equation}
when ${\rm d}\phi/{\rm d}\tau<0$ ({\em ending} at some $\tau_2$). The latter
equation implies that
\begin{equation} \label{Theta2p}
\psi_k(\tau_2)= \psi_k(\tau) \exp\left(- i( \Theta_k(\phi(\tau))-
\Theta_k(\phi(\tau_2)))\right)
\end{equation}
with the opposite sign in the phase for forward evolution in a backward
half-cycle, as required. This equation may be used whenever $\tau_2>\tau$ in
the same backward half-cycle.
The concatenated solution then has a continuous phase because the phases of
the two wave functions (\ref{Theta1}) and (\ref{Theta2}) indeed meet in the
middle: At a turning point, if $\phi_k(\tau_1)$ marks the beginning of a
forward half-cycle and $\phi_k(\tau_2)$ marks the end of the next backward
half-cycle, we may choose $\tau$ to be in both half-cycles, interpreted either
as the end of the first one or the beginning of the second
one. Correspondingly, the phase $\Theta_k(\phi(\tau))- \Theta_k(\phi(\tau_1))$
added to $\psi_k(\tau_1)$ and the phase
$\Theta_k(\phi(\tau))- \Theta_k(\phi(\tau_2))$ added to $\psi_k(\tau_2)$
produce the same state. This solution is globally valid because $|\phi(\tau)|$
never surpasses $\phi_{\rm t}$, demonstrating unitarity. While phase
continuity is guaranteed by construction, smoothness or even differentiability
is not. The system maintains unitary evolution as the wavefunction itself is
smooth during any half-cycle, whereas different evolution operators are
concatenated (rather than extended by solving a single differential equation)
precisely where the phase is not differentiable.
In our specific example, according to the phase $\Theta_k(\phi(\tau))$ in
(\ref{psik}), each half-clock cycle of $\phi$, changing monotonically from
$-\phi_{\rm t}$ to $\phi_{\rm t}$ or back, adds an amount of
\begin{equation} \label{DeltaTheta}
\frac{1}{2}\Delta\Theta_k=\Theta_k(\phi_{\rm t})-\Theta_k(-\phi_{\rm t})=
-\frac{E_k^2}{2\lambda\hbar}
\left(\arcsin(1)-\arcsin(-1)\right)= -\frac{\pi E_k^2}{2\lambda\hbar}
\end{equation}
to the phase. A forward half-cycle starts at $-\phi_{\rm t}$ and ends at
$\phi_t$ and has a phase changing according to $\Theta_k(\phi)$, while a
backward half-cycle starts at $\phi_{\rm t}$ and ends at $-\phi_{\rm t}$ but
has a phase changing according to $-\Theta_k(\phi)$. Therefore, in both cases
the phase added per half-cycle equals
$\Theta_k(\phi_{\rm t})-\Theta_k(-\phi_{\rm t})$. A full clock cycle, going
from $-\phi_{\rm t}$ to $\phi_{\rm t}$ and back, adds twice this phase. Going
back in $\phi$ does not cancel out the phase of the previous half cycle
because of our specific construction in which we flip the sign of $\Theta_k$
according to the sign of ${\rm d}\phi/{\rm d}\tau$, dictated by positivity of
the Hamiltonian for a stable system.
As an example, starting at the beginning of the zeroth cycle according to
Fig.~\ref{Fig:phitau}, such that $\tau_0=-\phi_{\rm t}$, $\tau$ increases from
$-\phi_{\rm t}$ to $\phi_{\rm t}$ during the first monotonic phase of $\phi$,
where the latter also increases from $-\phi_{\rm t}$ to $\phi_{\rm t}$. The
next monotonic phase then completes the zeroth cycle and has $\tau$ increasing
from $\phi_{\rm t}$ to $3\phi_{\rm t}$ while $\phi$ decreases back to its
initial value, $-\phi_{\rm t}$. The next clock cycles ($n=1,2,\ldots$) repeat
this process. Since a full clock cycle adds $-\pi E_k^2/(\lambda\hbar)$ to
the phase, according to (\ref{DeltaTheta}), at the end of the $n$th cycle the
wave function equals
\begin{equation}
\psi_k(\tau_1)=\psi_k(\tau_0)
\exp \left( \frac{-i n\pi E_k^2}{\lambda \hbar} \right)
\end{equation}
where $\phi(\tau_1)=-\phi_{\rm t}=\tau_0$.
It then proceeds by
\begin{eqnarray} \label{psik1}
\psi_k(\tau) &=& \psi_k(\tau_0)
\exp \left( \frac{-i n\pi E_k^2}{\lambda \hbar} \right)
\exp\left(i(\Theta_k(\phi(\tau))-\Theta_k(-\phi_{\rm
t}))\right)\nonumber\\
&=& \psi_k(\tau_0)
\exp \left( \frac{-i (n+1/4)\pi E_k^2}{\lambda \hbar} \right)
\exp(i\Theta_k(\phi(\tau)))
\end{eqnarray}
as long as $4n-1 \leq \tau/\phi_{\rm t} \leq 4n+1$ ($\phi(\tau)$
increasing). This half-cycle, ending at a time $\tau_2'$ when
\begin{equation}
\psi_k(\tau_2')= \psi_k(\tau_0) \exp \left( \frac{-i (n+1/4)\pi E_k^2}{\lambda
\hbar} \right)
\exp(i\Theta_k(\phi_{\rm t})) = \psi_k(\tau_0) \exp \left( \frac{-i
(n+1/2) \pi E_k^2}{\lambda
\hbar} \right)
\end{equation}
adds an additional $-\pi E_k^2/(2\lambda\hbar)$ to the phase. According to
(\ref{Theta2p}), the wave function in the next half-cycle is then given by
\begin{eqnarray}\label{psik2}
\psi_k(\tau) &=& \psi_k(\tau_0)
\exp \left( \frac{-i (n+1/2)\pi E_k^2}{\lambda \hbar} \right)
\exp\left(-i(\Theta_k(\phi(\tau))-\Theta_k(\phi_{\rm
t}))\right)\nonumber\\
&=& \psi_k(\tau_0)
\exp \left( \frac{-i (n+3/4)\pi E_k^2}{\lambda \hbar} \right)
\exp(-i\Theta_k(\phi(\tau)))
\end{eqnarray}
($4n+1 \leq \tau/\phi_{\rm t} \leq 4n+3$, $\phi(\tau)$ decreasing). Because
this half-cycle ends when $\phi(\tau)=-\phi_{\rm t}$, the final phase increase
during the combination of two half-cycles equals $-\pi E_k^2/(\lambda\hbar)$,
in agreement with (\ref{DeltaTheta}).
\subsection{Harmonic oscillators}
\begin{figure}
\centering
\includegraphics[width=0.7\textwidth]{Figures/psimedlam.eps}\\
\includegraphics[width=0.7\textwidth]{Figures/psihilam.eps}
\caption{Density plot of the wave function for intermediate (top) and large
$\lambda$ (bottom), using a harmonic-oscillator Hamiltonian
$\hat{H}=\frac{1}{2}(\hat{p}^2+\hat{x}^2)$ and a standard coherent initial
state. Coherence is quickly lost in $\tau$-evolution for intermediate
$\lambda$ as soon as the first turning points are reached (around
$\tau\sim 10$ for $\lambda=10^{-1}$, top). By contrast, coherence is
maintained for long times for large $\lambda$ even though billions of
turning points are crossed during a single system period for
$\lambda=10^{10}$ (bottom). Strong coherence is maintained in this case even
though the actual $\phi$-Hamiltonian $(\hat{H}^2-\lambda^2\phi^2)^{1/2}$ is
not harmonic if $\lambda\not=0$. \label{Fig:psilam}}
\end{figure}
In order to highlight features of coherence, we now consider the example of a
harmonic system Hamiltonian. Since there is perfect coherence in suitable
states of the standard harmonic oscillator, any loss of coherence implied by
our treatment of local clocks can easily be discerned.
Section~\ref{s:NonHarm} will show analogous properties in non-harmonic
examples in order to confirm the general qualitative features.
The parameter $\lambda$ in the harmonic clock Hamiltonian $p_{\phi}^2+\lambda^2\phi^2$
acts as a scaling factor for the frequency of the clock variable. It does not
have the units of frequency but rather of frequency times energy. In fact, the
number of clock cycles in any range of global time depends, through the value
$\phi_{\rm t}=E_k/\lambda$, on the energy of the system, or rather on the
contributions of energy eigenstates to an evolving system state. Since
$\lambda$ determines how often the clock turns around, it must therefore refer
to the energy.
The parameter $\lambda$ directly affects how often an energy eigenstate (of
the system Hamiltonian) moves through turning points of $\phi$ in a given
$\tau$-interval. As usual, any initial state can be expanded as a
superposition of energy eigenstates, but since different eigenstates imply
different times for the turning points of a clock, an oscillating clock makes
the superposition evolve in a modified way compared with absolute time in
standard quantum mechanics. Figures~\ref{Fig:psilam} and \ref{Fig:PSLambda01}
show how complicated this new dynamics can be, even for a coherent initial
state of a standard harmonic system Hamiltonian. However, the behavior
simplifies not only when $\lambda \rightarrow 0$, in which case we have
standard quantum mechanics with a monotonic time, but also, surprisingly, when
$\lambda \rightarrow \infty$ as evidenced by Fig.~\ref{Fig:psilam}. The
different behaviors of quantum fluctuations and deviations of the expectation
values from classical sinusoidal behavior are also illustrated in
Figs.~\ref{Fig:PSLambda01} and \ref{Fig:PSLambda1}.
\begin{figure}
\centering
\includegraphics[width=0.7\textwidth]{Figures/PSLambda01.eps}
\caption{Phase-space trajectory and quantum uncertainty (bottom bar) for a
harmonic system Hamiltonian with intermediate
$\lambda=0.1$. \label{Fig:PSLambda01}}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.7\textwidth]{Figures/PSLambda1.eps}\\[-10mm]
\includegraphics[width=0.7\textwidth]{Figures/PSLambda5.eps}
\caption{Phase-space trajectory and quantum uncertainty (bottom bars) for a
harmonic system Hamiltonian. Compared with Fig.~\ref{Fig:PSLambda01},
coherence and semiclassical behavior are regained as
$\lambda$ is increased, here showing the examples of $\lambda=1$ and
$\lambda=5$. \label{Fig:PSLambda1}}
\end{figure}
\subsubsection{Small-$\lambda$ approximation}
As $\lambda \rightarrow 0$, the state never approaches a turning point in any
finite range of $\tau$. Far away from these turning points, the phase of the
state resembles that of an unconstrained system, solving
(\ref{SchroedingerE}). Indeed, rewriting the constraint equation (\ref{C4}) by
substituting $\lambda \approx 0$ yields the relation $\hat{p}_\phi \approx
\pm\hat{H}$, belonging to a system with global
time $\phi$.
Utilizing this approximation, the equation of motion is equivalent to the
familiar Schr\"odinger equation
\[
i\hbar \frac{\partial\psi_k(\phi)}{\partial\phi}=
\hat{H} \psi_k(\phi) = E_k\psi_k(\phi)
\]
(choosing a sign suitable for stability from a positive Hamiltonian). This
equation generates the expected global time-dependent phase for a stationary
state,
\begin{equation}
\psi_k(\tau) = \psi_k(0)
\exp \left( \frac{-i \tau E_k }{ \hbar} \right)
\end{equation}
if we identify clock $\phi$ and time $\tau$ in this case.
\subsubsection{Large-$\lambda$ approximation}
\label{s:Largel}
As $\lambda \rightarrow \infty$, the stationary state passes through turning
points at a very high rate. The evolving state is obtained by concatenating
many branches of wave functions (\ref{psik1}) and (\ref{psik2}). The fact that
concatenations happen at different times for different energy eigenstates in a
superposition makes it hard to understand the long-term behavior of an
evolving state by analytic means. Nevertheless, numerical features, seen in
several model systems, show surprising simplifications, as demonstrated by the
example of a harmonic-oscillator system Hamiltonian in Figs.~\ref{Fig:psilam}
and \ref{Fig:PSLambda1}. According to numerical results, each stationary
state as a function of $\tau$ is again governed by a sinusoidally varying
phase, except for small intervals around turning points. The main visible
difference with standard evolution is that the sinusoidal frequency is a
multiple of the expected frequency at $\lambda =0$ by a factor of
$\frac{1}{4}\pi$. An example is shown in Fig.~\ref{Fig:Phases}, where
movements of the clock through turning points (about two per system cycle) are
clearly visible for intermediate $\lambda$, while the many turning points the
clock goes through per system cycle for large $\lambda$ merely rescale the
system period but do not lead to noticeable deviations from sinusoidal
behavior.
\begin{figure}
\centering
\includegraphics[width=0.8\textwidth]{Figures/Phases.eps}
\caption{Phases of the ground state of the harmonic potential
for different values of $\lambda$. Small plateaus caused by individual
turning points of $\phi$ are clearly visible for intermediate
$\lambda=10^{-1}$. These plateaus stretch out the curve in the time
direction. As a consequence, the system period for a large
$\lambda=10^3$ is greater than the period for a small $\lambda=10^{-5}$.
Plateaus are no longer visible for large $\lambda$ because the clock
period is much smaller in this case compared with intermediate
$\lambda$. Accordingly, the phases behave sinusoidally for large
$\lambda$, just as for the standard harmonic oscillator approached at
small $\lambda$.}
\label{Fig:Phases}
\end{figure}
The same features are shown by numerical evolution of expectation values of
$\hat{x}$ and $\hat{p}$ and their second-order moments (fluctuations and the
covariance), seen in Fig.~\ref{Fig:SmallFirst} for small $\lambda$ (close to
the standard harmonic oscillator), Fig.~\ref{Fig:MedFirst} for intermediate
$\lambda$, and Fig.~\ref{Fig:LargeFirst} for large $\lambda$. A single
stationary state does not show time-dependent expectation values and
moments. For this analysis, we have therefore chosen an initial state of
minimum uncertainty which would be a dynamical coherent state of the harmonic
oscillator ($\lambda=0$ or small).
For intermediate $\lambda$, the system rapidly loses coherence, as expected
because the $\phi$-Hamiltonian $\sqrt{H^2-\lambda^2\phi^2}$ is no longer
harmonic. As noted before, the phase of a wave function, obtained from
concatenated evolutions for each half-cycle of the clock, is not smooth. For
large $\lambda$, the high frequency at which the system crosses turning points
smoothes out the phase on scales larger than the clock period, which may have
been expected. More surprisingly, the same non-harmonic Hamiltonian
$\sqrt{H^2-\lambda^2\phi^2}$ that implies rapid loss of coherence at
intermediate $\lambda$ leads to strongly coherent behavior at large
$\lambda$. This feature, as well as its explanation and applications below,
are the main results of \cite{LocalTime}, for which the present paper provides
a detailed discussion.
\begin{figure}
\centering
\includegraphics[width=0.8\textwidth]{Figures/Small_First.eps}\\
\includegraphics[width=0.8\textwidth]{Figures/Small_Second.eps}
\caption{Basic expectation values (top) and second-order moments (bottom) for
small $\lambda$, using a harmonic system Hamiltonian. \label{Fig:SmallFirst}
}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.8\textwidth]{Figures/Med_First.eps}\\
\includegraphics[width=0.8\textwidth]{Figures/Med_Second.eps}
\caption{Basic expectation values (top) and second-order moments (bottom)
for
intermediate $\lambda$, using a harmonic system
Hamiltonian. \label{Fig:MedFirst} }
\end{figure}
\begin{figure}
\centering
\includegraphics[width=0.8\textwidth]{Figures/Large_First.eps}\\
\includegraphics[width=0.8\textwidth]{Figures/Large_Second.eps}
\caption{Basic expectation values (top) and second-order moments (bottom)
for large $\lambda$, using a harmonic system
Hamiltonian. \label{Fig:LargeFirst} }
\end{figure}
To explain this behavior, the large-$\lambda$ limit of Eqs.~(\ref{psik1}) and
(\ref{psik2}) can be simplified by ignoring the last exponential factors as
they approach zero in this limit: The first term in each of the last
exponentials is reduced to zero as $\lambda$ increases since the amplitude of
$\phi(\tau)$ approaches zero as $\lambda \rightarrow \infty$. The second term
in each of the last exponentials is an arcsine function divided by $\lambda$,
which also approaches zero as $\lambda \rightarrow \infty$. The reduced
equations therefore become
\begin{equation}\label{psik1large}
\psi_k(\tau) \approx \psi_k(0)
\exp \left( -\frac{i (n+1/4)\pi E_k^2}{\lambda \hbar} \right)
\end{equation}
if $4n-1 \leq \tau/\phi_{\rm t} \leq 4n+1$, and
\begin{equation}\label{psik2large}
\psi_k(\tau) \approx \psi_k(0)
\exp \left( -\frac{i (n+3/4)\pi E_k^2}{\lambda \hbar} \right)
\end{equation}
otherwise. Since $\phi_{\rm t}\to 0$ in this limit for fixed $E_k$, even small
changes in $\tau$ imply transitions between different clock cycles. Any
extended range of $\tau$ therefore leads to large numbers of clock cycles,
$n$, such that the remaining exponentials in (\ref{psik1large}) and
(\ref{psik2large}) are non-trivial even for large $\lambda$. However, there is
a negligible difference beteen $n+3/4$ and $n+1/4$ in the exponents, which
therefore are nearly identical in this limit.
As $\lambda \rightarrow \infty$ it is possible to approximate the floor
function in (\ref{n}), defining $n$, in a continuous form:
\begin{equation} \label{napprox}
\frac{n}{\lambda} = \frac{\left \lfloor 1/4+\lambda \tau/(4E_k)
\right \rfloor}{\lambda}
\approx \frac{\tau}{4E_k}\,.
\end{equation}
Using this result in Eq.~(\ref{psik1large}) yields
\begin{equation}
\psi_k(\tau) \approx \psi_k(0)
\exp \left(- \frac{i \pi E_k}{4 \hbar} \tau \right) \,.
\end{equation}
Comparing this time-dependent phase with the usual solution
\begin{equation}
\psi_k(t)= \psi_k(0) \exp(-iE_kt/\hbar)
\end{equation}
of the time-independent Schr\"odinger equation shows that the frequency of the
phase in the large-lambda limit is $\frac{1}{4}\pi$ times the frequency of the
system for $\lambda\to0$, in agreement with our numerical plots for the
harmonic oscillator with Hamiltonian
$\hat{H}=\frac{1}{2}(\hat{p}^2+\hat{x}^2)$: When $\lambda = 10^{-5} $
(approximating the limit $\lambda \rightarrow 0$), a sinusoidal function with
a period of $2\pi$ is returned, as expected for the ground state of the
harmonic oscillator of frequency parameter $\omega=1$. For large values of
lambda, $\lambda = 10^3$ (approximating $\lambda \rightarrow \infty $), the
period of the oscillations is multiplied by a factor of $4/\pi$ resulting in a
value of $8$. With intermediate values, turning points are spaced at easily
visible intervals, for instance located at $\tau = 5 + 10j$ with integer $j$
in the case of $\lambda = 10^{-1}$ as shown in Fig.~\ref{Fig:Phases}.
\subsection{Non-harmonic examples}
\label{s:NonHarm}
Our specific equations for the $\tau$-dependent phase can be used for any
system Hamiltonian. The harmonic example enjoys the most coherent dynamics and
therefore highlights any loss of coherence implied by an oscillating
fundamental clock. Non-harmonic systems cannot regain strong coherence for
large $\lambda$ simply because their dynamics is not coherent, but it is
nevertheless possible to see an approach to standard quantum mechanics for a
periodic clock with large $\lambda$.
\begin{figure}
\centering
\includegraphics[width=0.8\textwidth]{Figures/Lambda05Low.eps}\\
\includegraphics[width=0.8\textwidth]{Figures/Lambda10Low.eps}
\caption{Expectation value and quantum uncertainty of the orbital radius
with a standard hydrogen Hamiltonian. The quantum uncertainty is similar
for all $\lambda$, while the time-dependent expectation value has an
additional high-frequency contribution for intermediate $\lambda$ (top)
compared with large $\lambda$ (bottom) or standard quantum
mechanics. \label{Fig:Hydro} }
\end{figure}
Figure~\ref{Fig:Hydro} shows the example of a basic hydrogen Hamiltonian,
based on the Coulomb potential, with a non-coherent initial state, chosen as a
certain superposition of finitely many energy eigenstates. Quantum
fluctuations vary rather strongly for any $\lambda$, as they would also do in
standard quantum mechanics in this case. There is a notable difference in the
expectation values for different $\lambda$, shown by a high-frequency signal
visibly superimposed for intermediate $\lambda$ that disappears for large
$\lambda$ where the expectation values approach the standard behavior.
Similar features are obtained for a non-harmonic clock Hamiltonian. The main
difference of a non-harmonic system compared with a harmonic clock is that its
period depends on the energy. However, in our coupled system the clock period
$T_{\rm C}=4\phi_{\rm t}=4 E_k/\lambda$ already depends on the system energy,
which equals the clock energy by the energy-balance constraint. Non-harmonic
behavior of the clock therefore does not introduce crucial new features,
although it makes explicit calculations more complicated.
\section{Observational bounds}
So far, we have found one quantitative difference between the evolution of a
harmonic system with respect to an oscillating clock one one hand, and with
respect to an absolute time on the other: A rescaling of the system period by
a factor of $4/\pi$ for the clock Hamiltonian used here, as described in
Section~\ref{s:Largel}. However, this difference is not observable because it
depends only on the clock Hamiltonian and therefore rescales all system
frequencies or their characteristic time scales in the same way. The rescaling
factor can therefore be removed by absorbing it in a bare frequency
$\omega'= 4 \omega/\pi$ that appears in the mathematical expression of the
system Hamiltonian, and therefore gives rise to the observed frequency
$\omega$ after our rescaling. The scaling factor depends neither on the
energy or initial state for a given system, nor on the system itself. It
depends only on the clock dynamics, which is the same for all systems if the
clock is fundamental.
The second implication of an oscillating clock is its effect on the
decoherence time, depending on the clock parameter $\lambda$. Since a coherent
state of the harmonic oscillator has an infinite decoherence time in standard
quantum mechanics, a finite decoherence time as shown by our numerical results
cannot be eliminated by a simple rescaling. Moreover, the time of how long a
system can maintain coherence is under good observational control. For
instance, the current relative precision of $10^{-19}$ of atomic clocks
\cite{LatticeClock} could not be obtained if nature had provided us with an
intermediate $\lambda$ in our fundamental clock that would destroy coherence
or a stable system period over many system cycles.
\begin{figure}
\centering
\includegraphics[width=0.8\textwidth]{Figures/GreyAnalyticalvsNumerical.eps}
\caption{Relative standard deviation of the system period over many system
cycles as a function of $\lambda$. The analytical approximation
(\ref{sigma}) agrees well with a numerical computation of many system
periods, both confirming a $1/\lambda$-behavior. The analytical result
is rescaled by a factor of $2/\pi$ to take into account the average of a
sine fuction over a quarter-cycle. \label{Fig:AnalyticalvsNumerical}}
\end{figure}
Taking (\ref{napprox}) to the next order in $1/\lambda$ indicates that
relative deviations from strict sinusoidal behavior should be of this
order. More precisely, for given finite $\lambda$, we can compute the variance
of the phase $\Theta_k(\phi(\tau))$, given in (\ref{Theta}) around
\begin{equation}
\Theta_k^{\infty}(\tau):=\lim_{\lambda\to\infty}\Theta_k(\phi(\tau))=
\frac{\pi}{4}E_k\tau
\end{equation}
averaged over a quarter-cycle of $\phi$:
\begin{eqnarray}
\sigma^2 &=& \frac{1}{\phi_{\rm t}} \int_0^{\phi_{\rm t}}
\left(\Theta_k(\phi(\tau))- \Theta_k^{\infty}(\tau)\right)^2{\rm
d}\tau\nonumber\\
&=& \frac{E_k^4 (21\pi^2-1024/5)}{24^2\lambda^2\hbar^2} \,. \label{sigma}
\end{eqnarray}
Introducing the clock period $T_{\rm C}=4\phi_{\rm t}=4E_k/\lambda$ and the system
period $T_{\rm S}=2\pi\hbar/E_k$ allows us to eliminate the parameters
$\lambda$ and $E_k$ in favor of more general clock
characteristics. Equation~(\ref{sigma}) then takes the form
\begin{equation} \label{TT}
T_{\rm C}= \frac{48 \sigma T_{\rm S}}{\pi\sqrt{21\pi^2-1024/5}}\approx 9.7
\sigma T_{\rm S}\,.
\end{equation}
The dependence on $\sigma$ implies a strong magnification factor that can make
precision measurements sensitive to a small period of the fundamental clock
even if they operate on a larger system period. Indirect bounds on the clock
period therefore come within reach.
Even though (\ref{TT}) no longer refers directly to an energy eigenvalue, its
derivation remains strictly valid only for a system period of an eigenstate. It
is more difficult to obtain analytical control over the variance of the system
period of a superposition of several energy eigenstates, such as a coherent
state. A natural expectation is that $E_k$ in (\ref{sigma}) should then be
replaced with a certain expectation value determined by the Hamilton operator
$\hat{H}$, for which there are different options. As shown by
Fig.~\ref{Fig:AnalyticalvsNumerical}, replacing $E_k^2$ in (\ref{sigma}) with
the expectation value $\langle\hat{H}^2\rangle$ is in good agreement with
numerical computations of the relative standard deviation, $\sigma$, of the
system period over many system cycles, shown as a function of $\lambda$ in the
figure. While the close agreement is encouraging, it is also surprising and
remains incompletely understood: The quarter-cycle calculation in
(\ref{sigma}) does not directly refer to turning points of the clock and is
therefore insensitive to flipping the signs of the phase according to
(\ref{Theta1}) and (\ref{Theta2}), while the latter is important for global
evolution over many cycles during which we notice the restoration of
coherence. The quarter-cycle calculation does, however, depend on the
non-linearity of the phase implied by an oscillating clock, which is most
prominent in the approach to a turning point.
For very large $\lambda$, one should evolve through many system cycles in
order to determine the standard deviation accurately, considering the
collection of cycles as a statistical ensemble for the observable period. Such
long-term evolution was beyond the numerical capacity available to
us. However, since the upper bound on the standard deviation shown in the
figure follows a simple $1/\lambda$-behavior, we are justified in
extrapolating it to values of $\lambda$ even larger than those shown in the
plot. According to this extrapolation to large $\lambda$, a relative accuracy
of $10^{-19}$, as reported for recent atomic clocks in \cite{LatticeClock},
requires ten orders of magnitude less than the smallest relative standard
deviations shown in Fig.~\ref{Fig:AnalyticalvsNumerical}. Since one order of
magnitude in $\lambda$ corresponds to about one order of magnitude in
$\sigma$, taking into account the numerical factor of about ten in (\ref{TT}),
we need $\lambda$ at least as large as $10^{18}$ times the system frequency,
or a fundamental clock period of at most $10^{-18}$ times the system period.
The system period (or atomic clock period) used in \cite{LatticeClock} is
based on the transition from the $^3P_0$ state to the $^1S_0$ ground state of
Strontium, with a wave length of $698\,{\rm nm}$, amounting to a system period
of about $2\,{\rm fs}$. The upper bound on the fundamental clock frequency is
therefore
\begin{equation}\label{Upper}
T_{\rm C}<10^{-18}\cdot 2\,{\rm fs}= 2\times 10^{-33}\, {\rm s}\,.
\end{equation}
Although this upper bound is still several orders of magnitude larger than the
Planck time $t_{\rm P}= 5 \times 10^{-44}\,{\rm s}$, which is often suggested
as a fundamental period of time based on dimensional arguments, it is much
smaller than any value that could at present be obtained from a direct time
measurement. For instance, the current value of the shortest time interval
measured directly, given by the photon travel time $247\cdot 10^{-21}{\rm s}$
across a hydrogen molecule \cite{Zepto}, is more than ten orders of magnitude
larger than our upper bound.
Our new upper bound is also stronger than previous indirect measurements of
short time scales. In particular, the shortest length measurements currently
possible are of the order $10^{-19}\,{\rm m}$, achieved at high-energy
particle accelerators. Using the speed of light, this value translates into an
upper bound of $10^{-19}\,{\rm m}/c\approx 3\cdot 10^{-28}\,{\rm
s}$ on the time scale, about five orders of magnitude above our new upper
bound. This value, like ours, is based on an indirect measurement because it
translates a direct measurement of a scattering cross section into a length,
and then into a time parameter. By exploiting the dephasing time, our indirect
measurement is much more sensitive even than indirect measurements at high
energy.
\section{Conclusions}
We have analyzed a combination of an oscillating clock variable $\phi$ and an
evolving system degree of freedom $x$, coupled minimally through an
energy-balance constraint (\ref{C4}). Expressed as relational evolution of $x$
with respect to $\phi$, the dynamics is governed by a standard Schr\"odinger
equation (\ref{SchroedingerE}) with time-dependent Hamiltonian
$\sqrt{\hat{H}^2 -\lambda^2\phi^2}$ if $\hat{H}$ is the Hamiltonian that
determines the energy of the system. Therefore, an oscillating clock implies
that the usual equality between the energy operator
$i\hbar\partial/\partial t$ and the system Hamiltonian no longer holds. The
operator $i\hbar\partial/\partial \phi$ instead determines the momentum of the
clock variable $\phi$ and therefore its kinetic energy, but not the full clock
energy because an oscillating clock also has potential energy. The
energy-balance constraint makes sure that the combined energy of the system
and the clock is conserved. But the clock momentum, measured in quantum
mechanics by $i\hbar\partial/\partial \phi$, no longer equals the system
Hamiltonian.
Experience with standard quantum mechanics would suggest that evolution with a
Hamiltonian $\sqrt{\hat{H}^2 -\lambda^2\phi^2}$ is rather complicated for any
$\lambda\not=0$, even if $\hat{H}$ belongs to the harmonic oscillator as in
our main example. (Fractional powers of Hamiltonians can imply additional
subtleties; see for instance \cite{FractionalHarmonic}.) This expectation is
confirmed by Fig.~\ref{Fig:psilam} for small and intermediate values of
$\lambda$, defined such that $\lambda T_{\rm S}/\langle\hat{H}\rangle$ is not
very large, where $T_{\rm S}$ is the system period and the expectation value
$\langle\hat{H}\rangle$ of the system energy is taken in an initial system
state. For this range of $\lambda$, the coherence of an initial standard
coherent state is quickly lost as soon as the term $\lambda^2\phi^2$ becomes
relevant in the action of $\sqrt{\hat{H}^2 -\lambda^2\phi^2}$ on an evolving
state. For large $\lambda$, one would then expect that coherence is lost even
faster, well before the system can complete a single period or just move in a
noticeable way. Surprisingly, however, coherence is restored for very large
$\lambda$, as also shown in Fig.~\ref{Fig:psilam}.
The unexpected restoration of coherence for small periods of the fundamental
clock demonstrates that an oscillating fundamental clock is consistent not
only conceptually, as already shown in \cite{Gribov}, but also physically:
Even though the potential required for a periodic clock affects the coupling
between system and clock through the energy-balance constraint, leading to a
non-harmonic system Hamiltonian of the form
$\sqrt{\hat{H}^2-\lambda^2\phi^2}$, coherence can be maintained for
surprisingly long times for large $\lambda$. Here, it is important to note
that the dynamics is governed not only by the linear Schr\"odinger equation
(\ref{SchroedingerE}), but also by a discrete process given by flipping the
sign of the phase according to (\ref{Schroedingertau}). The precise
mathematical origin of restored coherence remains to be understood. This
restoration of coherence is not simply a perturbation of the standard harmonic
oscillator because it occurs for large $\lambda$, where perturbation theory
cannot be used in $\sqrt{\hat{H}^2-\lambda^2\phi^2}$. While intermediate
$\lambda$ are ruled out by observations of long coherence in isolated quantum
systems, large $\lambda$ and therefore sufficiently small fundamental periods
are consistent with current observations. A slight dephasing persists,
however, even at large $\lambda$, giving rise to our upper bound
(\ref{Upper}).
The origin of the coherence effect lies in quantum mechanics with an
oscillating fundamental clock. It does not have a complete classical analog. A
fundamental period of time would imply that a system period that is not an
integer multiple of the fundamental period cannot be sampled
precisely. Successive system periods may therefore appear slightly longer or
shorter depending on which system cycle an incomplete clock period is
attributed to in a measurement. This classical model would also lead to a
certain variance in system periods, but any such effect would quickly average
out over a few system cycles. Moreover, two systems starting at the same time
would be affected by the over/undercounting of complete clock cycles in the
same way. If they are synchronized, like an atom and a photon of the right
energy to generate a transition of energy levels or like two atomic clocks,
they would therefore not get out of tune by classical variations of the system
period. Observable implications of such a classical model would be
insignificant.
Our new quantum effect, which also implies variations of the system frequency,
is of a different nature. It acts on subtle coherence properties in a
superposition of energy eigenstates of the system. If, again, we have two
synchronized systems starting at the same time, they do not have identical
quantum states, and therefore are affected in different ways by the new
coherence and dephasing effects. Their system periods still vary, and
subsequent cycles of one system present a statistical ensemble independent of
the cycles of the other system because the initial states are largely
independent except in certain macroscopic properties that have been arranged
to agree in the synchronization procedure. Therefore, detuning can in
principle be observed by comparing the two systems. Similarly, if the two
systems are an atom and a photon which are ``synchronized'' in the sense that
the energy of the photon matches a transition energy of the atom, the atom and
photon states are necessarily different and therefore react differently to the
new coherence effect.
Unfortunately, it seems difficult to derive the new coherence effect at large
$\lambda$ in a controlled analytical approximation, even though it is clearly
presented by numerical simulations. We have been able to compute the variance
of the system period (\ref{sigma}) in good agreement with a numerical analysis
of the statistical ensemble given by the periods of a simulated wave function,
as demonstrated by Fig.~\ref{Fig:AnalyticalvsNumerical}. However, the
agreement remains somewhat mysterious because the calculation in (\ref{sigma})
is based on a quarter cycle of the fundamental clock, which is a much shorter
time than spanned by the large number of system periods that are used in the
numerical analysis. The agreement is encouraging and supports the relevance of
the new effect, but for a detailed analysis and further predictions, as well
as stricter upper bounds, it would be desirable to have an analytical
approximation that could accurately describe the statistical and coherence
properties of the system over many periods. Developing such an approximation
is challenging because the dynamics over many clock cycles is governed not
only by the linear partial differential equation (\ref{SchroedingerE}), but
also by the phase ``reflections'' in Eqs.~(\ref{Theta1}) and (\ref{Theta2}).
Our results have several conceptual implications for the quantum nature of
clocks and time. The fact that evolution for large $\lambda$ agrees with
$\lambda=0$ to a good degree demonstrates that deparameterization, a procedure
going back to Dirac and now widely used to evade the problem of time in
quantum gravity and quantum cosmology, can be considered a controlled
approximation of quantum dynamics at least as long as system periods are not
Planckian. The deparameterization procedure may therefore be applied at low curvature, but it
remains questionable at Planckian curvature where the system period (or any
rate of change if the system is not periodic, like the expanding universe)
itself is Planckian and compariable with the period of a fundamental
clock. New, unexpected effects may therefore be implied by a fundamental clock
at the big bang, which remain to be evaluated.
Finally, our analysis has shown that a periodic fundamental clock is, in
general, in a superposition of different clock cycles. We reach this
conclusion because the turning points of the clock variable, minimally coupled
to the system through energy balance, depend on the system energy as shown by
(\ref{phit}). Since a system is generically in a superposition of different
energy eigenstates, a fundamental period will quickly evolve into a
superposition of different clock cycles even if it is assumed to start in a
state that is sharply peaked around a given clock value. In particular, an
application of equation (\ref{phitau}) shows that the requirement of having a
unique value of global time $\tau$ for all energy eigenstates in superposition
implies different clock values $\phi_k(\tau)$ for different energy eigenvalues
$E_k$. The more time $\tau$ progresses, the more the various $\phi_k$ for a
given system state differ, potentially stretching over many clock cycles. Our
procedure of concatenating half-cycle evolutions for each eigenstate and then
bringing them back into superposition allows us to avoid dealing directly
with a complicated clock state, but such a state is indirectly realized.
This effect does not appear in non-periodic clocks such as those used in
deparameterization. To the best of our knowledge, our model is the first in
which the clock is truly quantum, understood in the sense that it is by
necessity in a superposition of different cycles.
\section*{Acknowledgements}
We thank Ding Ding, Bianca Dittrich, Kurt Gibble and Philipp Hoehn for
discussions. This work was supported in part by NSF grant PHY-1912168. LM was
supported by a Gates Scholarship.
\begin{appendix}
\section{Ingredients of the code}
We present crucial parts of the MATLAB program used for our results as
pseudocode:
\begin{itemize}
\item Calculating $\phi(\tau)$:
\begin{verbatim}
if ((tau/Phi_t) <= 4*n+1)
Phi_tau = t-4*n*Phi_t;
elseif ((tau/Phi_t) > (4*n+1))
Phi_tau = (4.*n+2)*Phi_t-t;
\end{verbatim}
\item Calculating $n$:
\begin{verbatim}
n = floor((1+(tau/Phi_t))/4);
\end{verbatim}
\item Calculating $\Theta(\phi(\tau))$:
\begin{verbatim}
if ((tau/Phi_t)>(4*n+1))
Theta = exp(pi*i*(n+1/2)*((k+1/2)^2)/lambda/hbar)*
exp(-i/2/lambda/hbar*(lambda*Phi_tau*((k+1/2)^2-lambda^2*Phi_tau^2)^.5))*
exp(-i/2/lambda/hbar*(k+1/2)^2*asin(lambda*Phi_tau/(k+1/2)));
elseif ((tau/Phi_t(k,lambda))<=(4*n(k,lambda,tau)+1))
Theta = exp(pi*1i*n*(k+1/2)^2/lambda/hbar)*
exp(i/2/lambda/hbar*(lambda*Phi_tau*((k+1/2)^2-lambda^2*Phi_tau^2)^.5))*
exp(i/2/lambda/hbar*(k+1/2)^2*asin(lambda*Phi_tau/(k+1/2)));
end
\end{verbatim}
\item Calculating zero crossing statistics:
\begin{verbatim}
for (a given range of lambda)
Qbar(t)=Integral(|Psi(q,t)|^2*q)
for (a given number of q intervals)
distance_from_exact= (lambda is infty zero) - findzero(Qbar,q interval)
std_dev(distance_from_exact)
average(distance_from_exact)
\end{verbatim}
\end{itemize}
\end{appendix}
|
\section{Introduction}\label{sec:intro}
\citet{townsend1976structure} proposed a conceptual model for wall-bounded turbulence, the attached eddy hypothesis (AEH), which idealizes structures as a collection of inertia-driven self-similar eddies that are randomly distributed in the plane of the wall. Details of key assumptions and limitations associated with the AEH are covered in a recent review by \citet{marusic2019attached}. Based on the AEH, \citet{perry1982mechanism} proposed that coherent wall-attached eddies scale with the distance from the wall $z$, and their heights comprise a geometrical progression. Evidence in support of a self-similarity and wall-scaling of wall-attached vortices has been reported in recent turbulent boundary layer (TBL) studies \citep[\emph{e.g.},][]{jimenez2012cascades, hwang2015statistical, baars2017self}. Figure~\ref{intro_sketch} shows an idealization of a self-similar hierarchy of wall-attached structures within the logarithmic region of a TBL \citep[][]{baidya2019simultaneous,marusic2019attached,deshpande2019streamwise}. Here, we consider three hierarchy levels of randomly positioned regions of coherent velocity fluctuations with each hierarchy shown in a different color. For simplicity, we consider the volume of influence of eddies, in each level, to be characterized by $\mathcal{L}_{i}$, $\mathcal{W}_{i}$ and $\mathcal{H}_{i}$ in the $x$, $y$ and $z$ directions, respectively, with $i=1,2,3$ denoting the $i^{\rm th}$ hierarchy level. Figures~\ref{intro_sketch}($b$) and~\ref{intro_sketch}($c$) denote the aspect ratios in the streamwise/wall-normal plane \AR$^{z}\equiv \mathcal{L}_i/\mathcal{H}_i\propto \lambda_{x}/\Delta{z}$ and in the streamwise-spanwise plane \AR$^{y}\equiv \mathcal{L}_i/\mathcal{W}_i\propto \lambda_{x}/\Delta{y}$, respectively. \citet{baars2017self} reported that in the neutral laboratory zero-pressure gradient TBL the self-similarity is described by a streamwise/wall-normal aspect ratio of $\lambda_{x}/\Delta{z} \approx 14$. A recent study by \citet{baidya2019simultaneous}, in high Reynolds number pipe and boundary layer flows, indicated that the self-similar wall-attached structures follow a three-dimensional aspect ratio of 14:1:1 in the streamwise, spanwise and wall-normal directions, respectively. More recently, \citet{krug2019vertical} explored the coherence for both velocity and temperature signals in the ASL. They found that the streamwise/wall-normal aspect ratio (\AR$^{z} \equiv \lambda_x/\Delta z$) decays with a logarithmic trend with increasing unstable thermal stratification; spanwise information was not explored in their study.
\begin{figure}
\vspace{0pt}
\centering
\includegraphics[width = 0.999\textwidth]{fig1.pdf}
\caption{($a$) Isometric view of three hierarchies of self-similar wall-attached eddies (simplified as slanted cuboids) in the logarithmic region of an ASL. ($b$) $x$,$z$--plane and ($c$) $x$,$y$--plane of one structure. Here, $\mathcal{L}_i$, $\mathcal{W}_i$ and $\mathcal{H}_i$ denote the streamwise, spanwise and wall-normal extent of the $i^{\rm th}$ hierarchy structure.}
\label{intro_sketch}
\end{figure}
Coherent structures reported in TBLs have been associated with characteristic inclination angles because of the mean shear. In the idealized (statistical) view of figure~\ref{intro_sketch}($b$), inclination angle $\alpha$ in the $x$,$z$--plane reflects a phase shift $\Delta \tau$ in time series of velocity fluctuations at different $z$. \citet{perry1982mechanism} used a vortex skeleton approach and the Biot-Savart law to determine the inviscid velocity field of a representative eddy, termed the $\Lambda$-vortex. In neutral TBLs, hairpin vortices have commonly been invoked as the representative eddy, and \citet{adrian2000vortex} suggested that these vortices are arranged together in groups called vortex packets \citep[for a comprehensive review of hairpin structures and their generating mechanisms, see][]{adrian2007hairpin}. \citet{christensen2001statistical} found that the sequence of individual vortex heads forms an interface or shear layer that is, statistically, inclined away from the wall at angles between $12^\circ$ and $20^\circ$. Laboratory results indicate the most probable inclination angle to be around $10^\circ$ to $15^\circ$ \citep[][]{christensen2001statistical,adrian2000vortex,baars2016spectral}. In the neutral surface layer, inclination angles ranging from 10$^\circ$ to 20$^\circ$ have been reported \citep[][]{boppe1999large, carper2004role,chauhan2013structure,liu2017variation}. Also, \citet{marusic2007reynolds} demonstrated the invariance of the inclination angle in wall-bounded flows with zero buoyancy (neutral conditions) over a wide range of Reynolds number through laboratory and field experiments.
The aforementioned studies all refer to the near-neutral case. However, in studies of the atmospheric surface layer (ASL) it has been observed that the inclination angle changes drastically under different stability conditions \citep[with steeper angles in increasingly buoyant cases, see][]{chauhan2013structure, liu2017variation, lotfy2018effect}. The thermal stability of the ASL is generally characterized by the Monin-Obukhov stability parameter $z_s/L$ \citep[][]{obukhov1946turbulence,monin1954basic}, where $L=-u_{\tau}^3{\overline{\theta}}/\kappa{\overline{w\theta}}{g}$ is the Obukhov length, $\kappa=0.41$ the von K\'{a}rm\'{a}n constant, $g$ the gravitational acceleration, $\overline{w\theta}$ the surface heat flux with $w$ and $\theta$ the fluctuating wall-normal velocity and temperature components, $\overline{\theta}$ the mean temperature, $u_{\tau}$ the friction velocity, and $z_s$ the reference height for evaluating this parameter. \citet{chauhan2013structure} found that under stable conditions, the inclination angle of structures reduced below the near-neutral angle to approximately $10^\circ$. Recently, \citet{salesky2020coherent} introduced an additional parameter to account for the loading and unloading of surface layer flux-gradient relations imposed by the passage of large-scale motions (LSMs). Meanwhile, \citet{salesky2020revisiting} developed a prognostic model for large-scale structures, where the inclination angle is the sum of the inclination angle observed in a neutrally stratified wall-bounded turbulent flow and the stability-dependent inclination angle of the wedge. \citet{baars2016spectral} indicates that in the neutral case, and for all scales $\lambda_x/\delta > 0.5$, the coherent scales obey a virtually constant inclination angle. In unstable conditions in the atmosphere, positive buoyancy lifts the structure away from the surface leading to an increase in the statistical inclination angle \citep[as averaged across all scales, see][]{chauhan2013structure,liu2017variation}. Now, in the unstable case, the dominance of buoyancy over shear is a function of wall-normal height, and hence one expects the inclination angle to be scale-dependent.
Since the coherent structure in the ASL has a strong relationship with the stability parameter, this paper will specifically address the influence of stability on: (1) the streamwise/wall-normal aspect ratio \AR$^{z}\equiv\lambda_{x}/\Delta z$ and the streamwise/spanwise aspect ratio \AR$^{y}\equiv\lambda_{x}/\Delta y$ in \S\,\ref{sec:ratio}, and (2) the scale-dependent angle $\alpha$ in \S\,\ref{sec:angle}, particularly under unstable conditions. Statistical relations for the aspect ratio and inclination angle for coherent turbulence fluctuations in the ASL are particularly relevant when analysing wind loading in the field of wind engineering \citep[see][]{davenport:1961phd,davenport:2002a}.
\section{Turbulence dataset of the atmospheric surface layer}\label{sec:data}
\subsection{QLOA facility and available data}\label{sec:qloa}
The measurement data used throughout this article were acquired at the QLOA facility in western China, Gansu province during three-month long measurement campaigns over two years (March to May in 2014 and 2015). The QLOA consists of wall-normal and spanwise arrays of sonic anemometers, performing synchronous measurements of the three-dimensional turbulent flow field. Sonic anemometers (Gill Instruments R3-50 installed from $s_2$ to $s_7$ and Campbell CSAT3B installed from $h_1$ to $h_{11}$, figure~\ref{experiment_site}) were employed to acquire the three components of velocity, as well as the static temperature, at a sampling frequency of 50\,Hz. Continuous observations were conducted at the QLOA site for a duration of more than 3000 hours, from which 89 hours of data were selected to analyze the characteristics of the large-scale coherent structures under different stratification stability conditions. The wall-normal array consists of 11 sonic anemometers that were placed with a logarithmic spacing on a vertical radio-type tower. The spanwise array covered an overall distance of 30\,m with 7 anemometers that were placed at a constant height of $z = 5$\,m, with an equi-distant spanwise spacing of 5\,m. The spanwise and wall-normal coordinates for each of the 17 anemometers are provided in figure~\ref{experiment_site}($b$). It should be noted that the first sonic anemometer in the spanwise array ($s_1$) also functions as the fifth on the main tower ($h_5$), which means we have 7 available anemometers in the spanwise array. The friction velocity $u_{\tau}$ is inferred from $u_{\tau}=(-\overline{uw})^{1/2}$ at $z = 5$\,m (calculated by the mean value from 7 sonic anemometers in the spanwise array). We assume an estimate for the surface-layer thickness of $\delta = 60$\,m, following \citet{hutchins2012towards}. The 89 hours of data remaining after preselection include 69 hours of unstable data ($z_s/L<-0.01$), 10 hours of near neutral data ($-0.01\leq{z}_s/L<0.01$) and 10 hours of stable data ($z_s/L>0.01$). Recall that $z_s$ is the reference height used to define the stability parameter $z_{s}/L$. For the benefit of comparison with previous works \citep{chauhan2013structure,liu2017variation,krug2019vertical}, the majority of our work uses $z_s = 2.5$\,m, unless otherwise specified. The demarcation of $z_s/L = 0.01$ to distinguish between neutral and unstable thermal stratification is commonly found in the literature, but for all analysis in this paper we present results as a function of $z_s/L$. The preselection criteria included: wind direction (the wind direction had to be aligned with the $x$ axis of the anemometer to within $\pm 30^\circ$) and steadiness \citep[statistically steady conditions based on the high-quality requirement by][]{fiken2004post}. A de-trending operation is also added (to remove the large-scale synoptic trend). See \citet{hutchins2012towards} and \citet{wang2016very} for full details of the preselection criteria.
\begin{figure}
\vspace{0pt}
\centering
\includegraphics[width = 0.999\textwidth]{fig2.pdf}
\caption{($a$) Three-dimensional view of the measurement setup at the QLOA site. ($b$) Northwest view of the sonic anemometer array; Campbell CSAT3B and Gill Instruments R3-50 sonics were installed at positions $h_1$ to $h_{11}$ and $s_2$ to $s_7$, respectively.}
\label{experiment_site}
\end{figure}
\subsection{Processing method}\label{sec:processing}
\begin{figure}
\vspace{0pt}
\centering
\includegraphics[width = 0.999\textwidth]{fig3.pdf}
\caption{($a$) An example of time series data for $z_s/L = -0.52$ at heights $h_5$ and $h_6$. The blue lines are low-pass filtered at $f = U/\delta$. ($b$) Portion of the filtered signals for inspecting the coherence and time shift. ($c$) Coherence spectrum for height $h_6$, relative to $h_5$, as function of temporal frequency. ($d$) Coherence spectrum for all heights $h_6$\,...\,$h_{11}$, again relative to $h_5$. Here the abscissa is converted to a spatial wavelength with the mean velocity at height $z$. ($e$) Scale-dependent phase of the cross-spectrum between $h_5$ and $h_6$, which is converted to a physical inclination angle $\alpha$ as a function of a spatial wavelength in sub-figure ($f$). The blue dots and blue solid lines in ($c,f$) indicate the frequency or wavelength corresponding to $\gamma _L^2=0.1$. Filtered spectra (black line), overlaid on the grey raw spectra, utilize a bandwidth moving filter of 25\%.}
\label{fig:processing}
\end{figure}
Two-point correlations can be computed on a per-scale basis in the Fourier domain using the linear coherence spectrum (LCS). For the fluctuating streamwise velocity signals $u$ the LCS is defined as:
\begin{equation}
\gamma_{L}^{2}(z,z_{\rm lcs};\lambda_{x})\equiv\frac{|\langle{\widetilde{U}(z;\lambda_{x})\widetilde{U}^{*}(z_{\rm lcs};\lambda_{x})}\rangle|^{2}}{\langle{|\widetilde{U}(z;\lambda_{x})|^{2}}\rangle\langle{|\widetilde{U}(z_{\rm lcs};\lambda_{x})|^{2}}\rangle}=\frac{|\phi_{uu}^{'}(z,z_{\rm lcs};\lambda_{x})|^{2}}{\phi_{uu}(z;\lambda_{x})\phi_{uu}(z_{\rm lcs};\lambda_{x})}.
\label{equ:spectrum}
\end{equation}
Here, $\widetilde{U}(z;\lambda_{x})=\mathcal{F}[u(z)]$ is the Fourier transform of $u(z)$, in either $x$ or time. The spatio-temporal transformation uses Taylor's hypothesis \citep{taylor1938spectrum}, where the local mean velocity is taken as the convection velocity. The asterisk $*$ indicates the complex conjugate, $\langle\rangle$ denotes ensemble averaging and $||$ designates the modulus. Scale-dependent phase information is explicitly embedded in the phase of the cross-spectrum $\phi_{uu}^{'}$. For (\ref{equ:spectrum}), the LCS is defined based on $u$ at two positions ($z_{\rm lcs}$ and $z$), separated in the wall normal direction by $\Delta z \equiv z - z_{\rm lcs}$. This coherence can also be computed across all other measured signals ($v$ and $\theta$) and also across spanwise separations $\Delta y$. The reference signal's height is denoted with the subscript `lcs' and is thus stated as $z_{\rm lcs}$. Since the LCS considers the magnitude of the complex-valued cross-spectrum, only the magnitude of coherence is considered (phase is covered later). Based on assumptions from the AEH \citep[][]{baars2017self,krug2019vertical}, the coherence magnitude within a self-similar region in $\lambda_x,z$--space is expected to adhere to
\begin{equation}
\gamma_{L}^{2}=C_{1}{\ln}\left( \frac{\lambda_{x}}{\Delta z} \right) +C_{2},
\label{EQgamma}
\end{equation}
\noindent from which the statistical aspect ratio (in this case streamwise/wall-normal) then follows
\begin{equation}
\AR^z = \left. \frac{\lambda_x}{\Delta z}\right\vert_{\gamma_{L}^2=0}=\exp\left(\frac{-C_{2}}{C_{1}} \right).
\label{EQgamma2}
\end{equation}
Here, $C_1$ and $C_2$ are fitted parameters. Figure \ref{fig:processing} shows an example of data obtained from the ASL to illustrate the process of the coherence spectrum. Figure \ref{fig:processing}($a$) indicates the raw data for the streamwise velocity collected at $h_5 = 5$\,m and $h_6 = 7.15$\,m, under unstable conditions with $z_{s}/L=-0.52$. A shorter time-history of corresponding filtered signals are shown in figure~\ref{fig:processing}($b$), evidencing that signal $h_6$ leads $h_5$. Thus, a coherent velocity fluctuation is first sensed at the higher wall-normal location as a result of the structure inclination angle. The LCS for $h_5$ and $h_6$ is presented in figure~\ref{fig:processing}($c$) as a function of temporal frequency, as computed from the time series data. Using Taylor’s frozen turbulence hypothesis, the frequency axis can be converted to a streamwise wavelength: $\lambda_x \equiv U_c/f$. Here, $U_{c}$ is a convective speed, taken as the mean velocity at local height $z$. The coherence spectrum of figure~\ref{fig:processing}($c$) can now be presented as a function of the streamwise wavelength, relative to the wall-normal separation distance $\Delta z$, as shown in figure~\ref{fig:processing}($d$). In addition to the coherence spectrum for $h_5$ and $h_6$, the coherence spectra for all heights above $h_6$ (relative to $h_5$ again) are also shown to illustrate the wall-similarity that is to be investigated. \citet{krug2019vertical} noticed that only data at neutral and unstable thermal stratification conditions complied with (\ref{EQgamma}), and thus an aspect ratio was only found for those conditions (a similar conclusion was reached for our QLOA site data and therefore we do not consider stable stratification with $z_s/L > 0$). In addition, \citet{krug2019vertical} found that the self-similar scaling applies also to fluctuations of the spanwise velocity $v$ and the static temperature $\theta$. \citet{baidya2019simultaneous} demonstrated that a scaling similar to (\ref{EQgamma}) and (\ref{EQgamma2}) occurs in the spanwise direction, resulting in a streamwise/spanwise aspect ratio \AR$^y$ for the self-similar structure.
Scale-dependent phase information is explicitly embedded in the phase of the cross-spectrum $\phi_{uu}^{'}$, given by
\begin{equation}
\Phi\left(z,z_{\rm lcs};\lambda_x\right)= \tan^{-1}\left(\frac{\Imag [\phi_{uu}^{'}(z,z_{\rm lcs};\lambda_x) ]}{\Real\left[\phi_{uu}^{'}(z,z_{\rm lcs};\lambda_x)\right]}\right).
\label{equ:phase}
\end{equation}
The phase spectrum aids in assessing the temporal shift between signals. Phase $\Phi(f)$ in (\ref{equ:phase}) is shown in figure~\ref{fig:processing}($e$) and can be used to extract a scale-by-scale inclination angle $\alpha$ (as shown in figure~\ref{fig:processing}f). That is, the temporal shift $\tau=\Phi(f)/(2\pi{f})$ where $f$ is the mode frequency and aids in computing the physical inclination angle through $\alpha=\tan[\Delta{z}/(\tau U_{c})]$. For the spectral analysis, the highest frequency resolved is set by the Nyquist frequency $f_s/2 = 25$\,Hz, where $f_s = 50$\,Hz is the sampling frequency. The lowest frequency is dictated by the interval length $I$ used in the spectral analysis and the longest internal used was $I = 2^N$ samples with $N = 15$ (an interval length of $\approx 650$\,s). A composite approach with varying interval length ($N = 8\,...\,15$) was used to generate the full spectra with as many ensembles as possible for the higher frequency portions of the spectrum (for $N = 8$ a total of around 1000 ensembles were used).
\section{Results}\label{sec:results}
\subsection{Stability dependence of aspect ratio}\label{sec:ratio}
\begin{figure}
\vspace{0pt}
\centering
\includegraphics[width = 0.999\textwidth]{fig4.pdf}
\caption{$\gamma^2_L$ in the range $2.15\,{\rm m} \leq \Delta{z} \leq 25$\,m and $5\,{\rm m} \leq \Delta{y} \leq 30$\,m (with increasing $\Delta{z}$ and $\Delta{y}$ indicated by lighter shades of grey) for the ($a$) wall-normal coordinate $z$ and ($c$) spanwise coordinate $y$, respectively, for the unstable case with $z_s/L = -0.52$. Here, the range between $\Delta z_{5,6} = h_6 - h_5 = 2.15$\,m and $\Delta z_{5,11} = h_{11} - h_5 = 25$\,m; $\Delta y_{1,2} = s_2 - s_1 = 5$\,m and $\Delta y_{1,7} = s_7-s_1 = 30$\,m. The blue line is a fit to obtain the aspect ratio according to (\ref{EQgamma}) with $C_{1}=0.302$ fixed; the fitting region used is bounded by $\gamma_{L}^2>0.1$ and $\lambda_x < 100$\,m and is indicated in red lines. Subscript $i=u,v,\theta$ in \AR$_{i}^k$ signifies the aspect ratio for streamwise, spanwise velocity components as well as temperature component; superscript $k=z,y$ in \AR$_{i}^k$ indicates either the streamwise/wall-normal or streamwise/spanwise aspect ratio. ($b,d$) Streamwise/wall-normal and streamwise/spanwise aspect ratios, respectively, as a function of $z_s/L$. The blue dots are our ASL results and the black solid lines denote the semi-log fitting. The dashed lines in ($b$) come from \citet{krug2019vertical} with $z_{\rm s}=2.14$\,m and $z_{\rm lcs}=1.41$\,m. Note that in this work $z_{\rm lcs}=z_s=5$\,m and so the dashed lines of \citet{krug2019vertical} were shifted along the $z_s$ axis to compare the trends at matched $-z_s/L$.}
\label{result_gamma}
\end{figure}
The linear coherence spectrum for $u$, $v$ and $\theta$ as a function of $\lambda_x/\Delta z$ and $\lambda_x/\Delta y$ for the unstable case with $z_s/L = -0.52$ is given in figures~\ref{result_gamma}($a$) and \ref{result_gamma}($c$), respectively. As reported by \citet{krug2019vertical}, the LCS collapse on one common curve over a range of $\lambda_x/\Delta z$ and $\lambda_x/\Delta y$. By fitting (\ref{EQgamma}) to these regions to obtain $C_1$ and $C_2$, the aspect ratio $\AR$ can be assessed. For this particular unstable case, we might expect the positive buoyancy to cause the self-similar structures in the hierarchy to lift more aggressively from the wall, extending the wall-normal coherence for a given $\lambda_x$ scale, and hence reducing \AR$^z_u$. Indeed, for the unstable case considered in figure~\ref{result_gamma}($a$) this yields an $\AR_u^z$ that is significantly lower ($ \AR^z_{u}= 3.6$) than the value of \AR$^z_{u}\approx14$ reported for laboratory neutral conditions \citep{baars2017self} and in close agreement with the aspect ratio reported in \citet{krug2019vertical} for similar values of the stability parameter. The resulting aspect ratios for $u$, $v$ and $\theta$ from all 79 hour datasets (covering a range of stabilities from $0.007 \leq -z_{\rm lcs}/L \leq 1.04$ are plotted as a function of the stability parameter in figure~\ref{result_gamma}($b$) for streamwise/wall-normal aspect ratios $\AR^z$ and in figure~\ref{result_gamma}($d$) for streamwise/spanwise aspect ratios $\AR^y$. In all cases, a clear trend emerges between aspect ratio and stability parameter, and a log-linear trend is fitted to the extracted data (black solid curves). These fitted trends are consistent with those of \citet{krug2019vertical} (black dashed curves in figure~\ref{result_gamma}$b$) indicating that the self-similar scaling under near-neutral and unstable conditions is a universal phenomenon. As an extension to the results of \citet{krug2019vertical}, the streamwise/spanwise aspect ratios of figure~\ref{result_gamma}($d$) seem to also exhibit log-linear trends, although in these cases the scatter in results is greater.
\citet{baidya2019simultaneous} indicates that the aspect ratio \AR$^z_{u}:$\AR$^y_{u}=1:1$ in the laboratory neutral condition. The results shown in figure~\ref{result_gamma}($b,d$) are also supportive of this. The curve fits to the data suggest that $\AR_u^z:\AR_u^y = 0.91:1$ for the near-neutral case ($\vert z_s/L \vert < 0.03$) changing slightly to 0.96:1 for the strong unstable ($z_s/L < -0.40$) stability conditions. Moreover, the aspect ratio for $u$ shows that $\lambda_x:\Delta y:\Delta z\approx 4.25:0.96:1 $ under the strong unstable condition, demonstrating that positive buoyancy has a lifting effect, increasing the size of coherent structures in the wall-normal and spanwise directions relative to its streamwise extent. The data here indicate that these self-similar eddies for $u$ follow an aspect ratio of $\lambda_x:\Delta y:\Delta z\approx 6.4:0.91:1 $ in the near-neutral condition ($\vert z_s/L \vert < 0.03$). It is worth highlighting again that the aspect ratio is sensitive to even very weakly unstable conditions. Therefore, the value for $u$ measured in the ASL is less than the result $\lambda_x/\Delta{z} = 14$ from \citet{baars2017self} and \citet{baidya2019simultaneous} in neutral laboratory conditions, as was also noticed by \citet{krug2019vertical}, whose prediction implies that $\lambda_x/\Delta z = 14$ will only be attained for $\vert z_s/L \vert \approx 0.0003$ (but for the QLOA dataset we only have data for $\vert z_s/L \vert \geq 0.007$).
\subsection{Stability dependence of structure inclination angle}\label{sec:angle}
\begin{figure}
\vspace{0pt}
\centering
\includegraphics[width = 0.999\textwidth]{fig5.pdf}
\caption{($a$) The phase expressed as a physical inclination angle $\alpha$ as a function of wavelength $\lambda_{x}/\delta$ and stability parameter $-z_s/L$. The LCS reference height, $z_{\rm lcs}=0.90$\,m, and inclination angles are presented for distances $\Delta z = h_5 - h_1 = 4.1$\,m. The solid increasing shades of blue dots indicate the scales at $\lambda_{x}/\delta=1$; $\lambda_{x}/\delta=2$ and $\lambda_{x}/\delta=6$, which is addressed in detail in figure~\ref{fig:angle1}. The solid blue lines show the trend fitted to these data given by the ($b$) surface fit to the data of the form (\ref{equation_zL}). }
\label{fig:phase1surf}
\end{figure}
Per the phase spectrum in figure~\ref{fig:processing}($f$), the scale-dependent phase is now analyzed for scales within the range $1 < \lambda_x/\delta < 10$ and for a similar $\Delta z$ as before: $\Delta z = h_5 - h_1 = 4.1$\,m. Larger wavelength information in the phase spectra is prone to noise issues due to the limited ensembles available for constructing the spectra, while smaller wavelengths generally have a lower coherence. Figure~\ref{fig:phase1surf}($a$) indicates the phase spectra in terms of inclination angle as a function of wavelength $\lambda_{x}/\delta$, for all different stability parameters $-z_s/L$. Note that all measured angles in figure~\ref{fig:phase1surf}($a$) are positive, corresponding to forward leaning structures. Though there are some clear outliers in these plots, certain trends are visible as evidenced by the coloured symbols which show data at constant scales of $\lambda_x/\delta$ = 1, 2 and 6. Clearly these different wavelengths exhibit different dependencies of $\alpha$ with stability parameter, with $\lambda_x/\delta = 1$ exhibiting much steeper angles $\alpha$ in the most unstable cases and markedly shallower angles as near-neutrality is approached. In general, it is also noted that longer structures (larger wavelength) will exhibit smaller inclination angles, especially noticeable at stronger convective conditions. Similar to \citet{chauhan2013structure} and \citet{baars2016spectral}, an extended parametric equation is fitted to the log-linear trend of convective data to model the variation of $\alpha$ with stability $z_s/L$, scale $\lambda_{x}$, following
\begin{center}
\begin{equation}
\alpha \left(\frac{z_s}{L},\frac{\lambda_x}{\delta}\right) =\alpha_{0}+C_{0}\left(\frac{\lambda_x}{\delta}\right)\ln \left(1+70\left| \frac{z_s}{L} \right| \right),
\label{equation_zL}
\end{equation}
\end{center}
where $z_s = 2.5$\,m, and $\alpha_{0}$ is the constant inclination angle selected as $\alpha_{0}=12^\circ$ in the neutral surface layer, $C_0$ is a function of the outer-scaled wavelength $\lambda_x/\delta$, which will be illustrated later. A curve-fitted plane to those spectra is shown in figure~\ref{fig:phase1surf}($b$), in which this fit has the form in (\ref{equation_zL}). It should be noted from figure~\ref{fig:phase1surf}($a$) that for small near-neutral values of the stability parameter it is not always possible to compute an inclination angle $\alpha$ for the smaller wavelength ($\lambda_x/\delta = 1$) from the phase spectra since the coherence across $\Delta z$ drops below $\gamma_L^2 = 0.1$.
\begin{figure}
\vspace{0pt}
\centering
\includegraphics[width = 0.999\textwidth]{fig6.pdf}
\caption{Variation of $\alpha$ with the stability parameter $z_s/L$ ($z_s = 2.5$\,m, $z_{\rm lcs} = 0.90$\,m and $z = 5$\,m, so that $\Delta z = 4.1$\,m). The solid color lines are the fitting lines based on (\ref{equation_zL}) for the corresponding scales of $\lambda_x/\delta = 1$, 2 and 6. Current data are shown by black circles (based on a two-point correlation, including all scales). The open black square correspond to near-neutral and unstable ASL data of \citet{chauhan2013structure}; asterisks from $R_{u_{\tau}u}$ of \citet{marusic2007reynolds}; diamonds of \citet{carper2004role}; open pentagram from $R_{uu}$ of \citet{marusic2007reynolds}.}
\label{fig:angle1}
\end{figure}
Figure~\ref{fig:angle1} shows the variation of $\alpha$ as a function of the stability parameter $-z_s/L$ for three different length scales $\lambda_x/\delta = 1,2,6$. Inclination angles computed from the phase spectra are shown by the colored circles and the blue lines show the surface fit to the data given by (\ref{equation_zL}). In addition, the black open circles shown in figure~\ref{fig:angle1}, which show the inclination computed from the two-point correlations for all 11 measurement heights indicate that the current data are consistent with the similarly computed results from \citet{chauhan2013structure}, shown by the black open squares. The black solid line in figure~\ref{fig:angle1} shows the fit proposed by \citet{chauhan2013structure} based on the two-point correlation contours. As such this original fit includes \emph{all scales}, and is disproportionately skewed towards larger scale features at higher $z$. It is clear from figure~\ref{fig:angle1} that the increasing buoyancy lifts all scales to larger inclination angles, although the smaller scale structures considered ($\lambda_x/\delta = 1$) exhibit steeper angles at all values of the stability parameter as compared to the larger features ($\lambda_x / \delta = 6$). This means that the fitting parameter $C_{0}$ in (\ref{equation_zL}) increases systematically as we focus on smaller scales. The coloured curves in figure~\ref{fig:angle1} show the curve fits to the data based on the fit proposed in (\ref{equation_zL}), with the constants $\alpha_0$ and $C_0$ given in the figure legend. It should be noted here that for the limited scale range discussed here $1<\lambda_{x}/\delta<10$, all of these scales are large, and associated with the upper end of attached motions and superstructures \citep{hutchins2007evidence}. Although with increasing stability all scales are lifted compared with the neutral condition, the low end of this range ($\lambda_x/\delta = 1$) exhibits the steepest angles, reaching $\alpha \approx 70^\circ$ at $z_s/L= -1.0$. \citet{baars2016spectral} indicated that the inclination angle of the large-scale structures in the neutral laboratory boundary layer is scale independent and equal to $\alpha = 14.7^\circ$. Though the data in figure~\ref{fig:angle1} do suggest that $\alpha$ becomes scale-independent in the limit of small $|z_s/L|$, the angle seems to be closer to $\alpha \approx 12^\circ$ for the current data.
\begin{figure}
\vspace{0pt}
\centering
\includegraphics[width = 0.999\textwidth]{fig7.pdf}
\caption{Summary of the different heights involved in the analysis. Heights for which the analyses were performed up to this point are listed in the box.}
\label{heights}
\end{figure}
The results summarised in figure~\ref{fig:angle1} are computed for a linear coherence spectrum between the reference location $z_{\rm lcs} = 0.9$\,m and the location $z = 5$\,m ($\Delta z = 4.1$\,m). In addition, the stability parameter presented on the abscissa of figure~\ref{fig:angle1} is computed based on conditions at $z_s = 2.5$\,m. Before considering further the form of the fit described by (\ref{equation_zL}), we must consider the sensitivity to $z_s$, and $\Delta z$. Generally, the stability parameter $\vert z/L \vert$ is linearly increasing with height, which would suggest that the scale-dependent structure inclination angles $\alpha$ will also increase with wall-normal height $z$ (since as we move away from the wall, buoyancy effects will increase in dominance relative to shear). To account for this, and to form a stability parameter that better reflects the altitude at which the linear coherence is evaluated, we propose a fractional stability parameter, dubbed $z_F/L$, where $z_F$ is fixed at a constant fraction of the wall-normal offset $\Delta z$. Hence stability is always assessed at a fixed fractional height between the lower and upper probes used to compute the linear coherence spectrum. This should minimise the sensitivity of inclination angle-stability parameter relationship to changes in $\Delta z$. Since the data in figure~\ref{fig:angle1} are given based on $z_{\rm lcs} = 0.90$\,m, $z_{s} = 2.5$\,m and $z=5$\,m (figure~\ref{heights}), the fractional stability parameter $z_F$ can be assessed as
\begin{eqnarray}
z_F &=& \overbrace{\left(\frac{z_s-z_{\rm lcs}}{z-z_{\rm lcs}}\right)}^{\rm \scalebox{0.80}{values fig.\,\ref{heights}}}\Delta z+z_{\rm lcs} = 0.39\Delta z + z_{\rm lcs}.
\label{equation_zF}
\end{eqnarray}
Maintaining $z_F$ at the value given by (\ref{equation_zF}), when changing $\Delta z$ or $z_{\rm lcs}$, ensures that we assess stability at the same fractional location 0.39$\Delta z$ when varying the height of inspection $z$. We can now substitute $z_F$ for $z_s$ in (\ref{equation_zL}),
\begin{equation}
\alpha\left(\frac{z_F}{L},\frac{\lambda_{x}}{\delta}\right) =\alpha_{0}+C_{0}\left(\frac{\lambda_{x}}{\delta}\right) \ln\left(1+70\left|\frac{z_F}{L}\right| \right).
\label{equation_last1}
\end{equation}
\begin{figure}
\vspace{0pt}
\centering
\includegraphics[width = 0.999\textwidth]{fig8.pdf}
\caption{($a$) Parameter $C_{0}$ as a function of wavelength $\lambda_x$ with increasing $\Delta z$ indicated by thinner lines and darker shades of grey. Here $z_{\rm lcs}$ remains fixed at $h_1 = 0.91$\,m and $z$ increases from 1.71\,m to 30\,m ($h_2$ to $h_{11}$), the blue solid line indicates a log-linear fit to all grey curves. ($b$) The scale-dependent $\alpha$ for $u$ as a function of stability parameter, for $\lambda_x/\delta=1,2,6$ at $h_5$, $h_7$ and $h_9$ ($z_s = 2.5$\,m and $z_{\rm lcs} = 0.90$\,m); lines come from (\ref{equation_final}) and each set of profiles is offset by $15^\circ$.}
\label{fig:c0}
\end{figure}
The grey curves in figure~\ref{fig:c0}($a$) show the fitting parameter $C_0$ from (\ref{equation_last1}) extracted from fits to the data with various $\Delta z$. Here $z_{\rm lcs}$ remains fixed at 0.90\,m and $z$ covers all heights above ranging from 1.71\,m to 30\,m, with thinner grey lines indicating lower values of $z$. We can only extract $\alpha$ from the phase spectra in situations where the coherence $\gamma_L^2$ is greater than 0.1, this condition gives rise to the sharp drops/rises in figure~\ref{fig:c0}($a$) as at different $\lambda_x/\delta$, varying subsets of data are available for the calculation of $\alpha$. In general it is noted that the fractional location for the stability height described above, does a reasonable job of collapsing the $C_0$ curves for all $\Delta z$; $C_0$ decreases with $\lambda_x/\delta$ indicating that the smaller scale structures tend to have larger inclination angles in the convective conditions. Based on the approximate collapse observed in figure~\ref{fig:c0}($a$), we can crudely approximate the $\lambda_x/\delta$ dependence of $C_0$ with,
\begin{equation}
C_{0}=13.1-3.8\ln\left(\frac{\lambda_x}{\delta}\right),
\label{equation_last2}
\end{equation}
which is shown by the blue solid line in figure~\ref{fig:c0}($a$). Finally, by combining (\ref{equation_last2}) with (\ref{equation_last1}), we obtain the scale-dependent structural inclination angle as,
\begin{equation}
\alpha\left(\frac{z_F}{L},\frac{\lambda_{x}}{\delta}\right) =\alpha_{0}+\left(13.1-3.8\ln\left(\frac{\lambda_{x}}{\delta}\right)\right)\ln\left(1+70\left|\frac{z_F}{L}\right|\right).
\label{equation_final}
\end{equation}
Figure~\ref{fig:c0}($b$) shows the influence of wall-normal offset $\Delta z$ (with $z_{\rm lcs}$ fixed at 0.90\,m) on the computed inclination angle $\alpha$ as a function of stability for the wavelengths $\lambda_{x}/\delta=1,2,6$ for $u$. A larger $\Delta z$ leads to higher $\alpha$, but this can be accounted for by considering the fractional stability parameter. The curves, showing (\ref{equation_final}), describe the variation of $\alpha$ with $z_s/L$, $\lambda_x/\delta$ and $\Delta z$ reasonably well.
By way of a summary, figure~\ref{figure_sketch} shows an illustration of the scale-dependent structure inclination angle and aspect ratio for both neutral (subscript $n$) and unstable (subscript $u$) thermal stratification conditions. The illustrations show the streamwise extent of the structure (in physical space following the simplification that its length $\mathcal{L}$ scales with half the wavelength, thus $\mathcal{L} \sim \lambda_x/2$) and its wall-normal extent with height $\mathcal{H}$; the streamwise/wall-normal aspect ratio of the structure in the $x$,$z$--plane adheres to \AR$_u^z$ of figure~\ref{result_gamma}($b$). When concentrating on the neutral stability condition first (bottom figure), both scales drawn exhibit the same inclination angle $\alpha_n$. That is, the statistical inclination angle of a large- and small-scale comprise the same forward leaning behavior.
\begin{figure}
\vspace{0pt}
\centering
\includegraphics[width = 0.999\textwidth]{fig9.pdf}
\caption{Summary of aspect ratio $\AR^z$ and scale-dependent angle $\alpha$ in a neutral and unstable stratified ASL. Two hierarchy levels are considered, each of a different scale: $L$ and $H$ show the structure's streamwise length and wall-normal height. The subscripts' first letter $n$ or $u$ signifies a `neutral' or `unstable' condition, and the second designation $s$ or $l$ denotes `small-scale' ($\lambda_x/\delta = 2$) or `large-scale' ($\lambda_x/\delta = 6$). The dashed lines in the top figure indicate a reference location $z_{\rm lcs} = 0.90$\,m and heights $z_1 = z_{\rm lcs} + \Delta z = 15$\,m and $z_2 = z_{\rm lcs} + \Delta z = 60$\,m, for which the corresponding $\alpha$ are indicated.}
\label{figure_sketch}
\end{figure}
When considering unstable thermal stratification---in the illustration solely one value is considered of $z_s/L = -0.50$ evaluated at $z_s = 2.5$\,m---the inclination angle is both scale-dependent and varies with $z$ (the procedure for plotting this structure is outlined in Appendix~A). First of all, note that the coherent wall-attached structures are still self-similar per the definition used in this paper based on the structure's streamwise wavelength relative to its wall-normal extent. But, the aspect ratio reduces compared to the neutral case and that this depends on the degree of thermal stratification per the relations shown in figure~\ref{result_gamma}($b$), thus: $(\mathcal{L}_{ns}/\mathcal{H}_{ns} = \mathcal{L}_{nl}/\mathcal{H}_{nl}) > (\mathcal{L}_{us}/\mathcal{H}_{us} = \mathcal{L}_{ul}/\mathcal{H}_{ul})$. Here subscript $s$ and $l$ refer to the small- and large-scale structure visualized, respectively. Concentrating on the inclination angle (\emph{e.g.}, the phase shift between a reference height $z_{\rm lcs}$ and a height of inspection $z$), it was found that for unstable stratification it increases with increasing $z$. Hence, $\alpha_{ul2} > \alpha_{ul1}$ and the coherent shape of the structure is characterized by a leading/trailing edge front that is curved and becomes steeper with height. Moreover, smaller scale structures exhibit steeper angles for the case of equal $\Delta z$ (\emph{e.g.}, $\alpha_{us} > \alpha_{ul1}$). The trends for this one stability condition are visualized with the `$\alpha$-surface' in the centre of figure~\ref{figure_sketch}. There the fractional stability $z_F$ is a surrogate for the $\Delta z$ trend.
\section{Conclusion}\label{sec:concl}
Wall-normal and spanwise arrays of sonic anemometers deployed in the atmospheric surface layer enable examination of the linear coherence spectrum, $\gamma_{L}^2$, as a function of the streamwise wavelength ($\lambda_{x}$), spanwise offset ($\Delta{y}$) and wall-normal offset ($\Delta{z}$). This in turn offers the opportunity to explore the three-dimensional form of the wall-attached self-similar structure for the streamwise velocity $u$, which illustrates that the self-similar wall-attached structures follow an aspect ratio of $\lambda_{x}/\Delta{z}:\lambda_{x}/\Delta{y}\approx1$ under near-neutral and unstable conditions. It is found that the aspect ratio $\lambda_x/\Delta z$ is greater for the near-neutral case, and becomes progressively smaller as instability increases. Hence similar length ($\lambda_x$) structures in unstable conditions will be taller and wider than their near-neutral counterparts. The phase of the cross-spectrum provides a scale-by-scale structure inclination angle. We find that this inclination angle is invariant with scale for the near-neutral case, but with increasing positive buoyancy becomes increasingly scale dependent. For unstable conditions (\emph{e.g.}, $z_s/L = -0.5$), all scales are inclined at steeper angles compared to the near-neutral case, with the smaller scales with $\lambda_x/\delta = 1$ exhibiting inclination angles that are approximately twice that of larger scales of $\lambda_x/\delta = 6$.
\section*{Acknowledgements}
We acknowledge support from the National Natural Science Foundation of China (No.92052202) and the financial support from the Australian Research Council. Also, XBL was supported by a CSC scholarship (File No.\,201706180037).\\[-8pt]
\noindent \textbf{Declaration of Interests.} The authors report no conflict of interest.
\section*{Appendix A: Outline of a coherent structure}\label{sec:appA}
Visualizing the outline of a coherent, statistical structure of a streamwise velocity fluctuation in the streamwise-wall-normal plane relies on a simple model based on (\ref{equation_final}) and (\ref{equation_zF}). The procedure for plotting a structure such as the example one in figure~\ref{fig:plottingcode} starts with the following steps:\\[-8pt]
\begin{enumerate}[labelwidth=0.65cm,labelindent=0pt,leftmargin=0.65cm,label=(\roman*),align=left]
\item \noindent A wavelength of the structure should be chosen, \emph{i.e.}, $\lambda_x = 2\delta$. Note that the structure is visualized in physical space, through the assumption that its streamwise extent spans half the wavelength, $\mathcal{L} \sim \lambda_x/2$.\\[-10pt]
\item \noindent The degree of unstable stratification should be chosen, \emph{i.e.}, $z_s/L = -0.40$ (here $z_s = 2.5$\,m).\\[-10pt]
\item \noindent A reference height, above which the structure is visualized, should be chosen.\\[-8pt]
\end{enumerate}
A structure outline is generated through considering a sequence of local heights $z$. For every height, the angle relative to the fixed reference height is determined from (\ref{equation_final}) and (\ref{equation_zF}). Note that the structure is only defined up to a height that is dictated by the aspect ratio-condition, following the trend line in figure~\ref{result_gamma}($b$). Thus, the wall-normal extent up to which the structure is defined follows from inferring the value of $\Delta z$ from \AR$_u^z \equiv \lambda_x/\Delta z = -1.9\log\left(-z_s/L\right) + 3.5$.
\begin{figure}
\vspace{0pt}
\centering
\includegraphics[width = 0.999\textwidth]{fig10.pdf}
\caption{Example of an outline of a coherent, streamwise/wall-normal structure of a particular scale and set of ASL parameters (indicated in the figure).}
\label{fig:plottingcode}
\end{figure}
\bibliographystyle{jfm}
\section{Package usage}
The symbol \AR\ is accessed in \LaTeX\ by means of the package \pack{ar} and is produced by the command \cs{AR}, in both math and text modes.
The package is loaded by the usual command:
\begin{flushleft}
\cs{usepackage}\oarg{options}\Arg{ar}
\end{flushleft}
The available options are the following:
\begin{description}
\item[CM] is the default setting and produces glyphs that are consistent with Computer Modern, CM-super, EC, and Latin Modern fonts;
\item[TM] or \textbf{TX} selects glyphs that are suitable with the Times, and Times eXtended fonts, but they are generally suited with other narrow fonts as, for example, the Linux Libertine ones.
\item[PA] or \textbf{PX} selects glyphs that are suitable with Palatino, Palatino eXtended, Mathpazo, and other wide fonts.
\end{description}
These options have been defined to allow the consistent use of the aspect ratio symbol in documents typeset with fonts other than Computer Modern and its siblings. Hopefully additional options will be roved in the future for a more comprehensive font support.
\section{Package contents}
The package contains several files of different types:
\begin{itemize}
\item This \file{ar.tex} and its compiled result \file{ar.pdf}.
\item The \MF\ source files \file{ar}\meta{size}\file{.mf} in order to produce the bitmapped fonts, in case the user wanted to typeset a document in DVI format and wanted to preview it with a DVI viewer. The \meta{size} ranges from 5 to 12, in the usual sequence as any other math font in the \TeX\ system.
\item The \MF\ source files for sans serif medium and bold fonts in 10pt size and the only source file for the typewriter monospaced version of the font in 10 pt size.
\item The metric files \file{*.tfm} for both the bitmapped and vector fonts.
\item The vector font files for the Computer Modern version, \file{ar}\meta{series}\meta{size}\file{.pfb} includiong the single 10pt size sans serif and typewriter style fonts, and the vector font files for the Times and Palatino variants, \file{amar}\meta{series}\file{i.pfb} and \file{apar}\meta{series}\file{i.pfb}. The meta{series} is either empty or equal to `b', for medium and bold series respectively.
\item The \file{ar.sty} extension file through which the suitable settings are made and the \AR\ symbol made available to the user.
\item The \file{aspectratio.map} file necessary to configure your system for having the vector typesetting engines access the vector fonts.
\end{itemize}
\section{Installing the package and the fonts}
If the whole package is not already installed by your \TeX\ system distribution, we suggest to create an \file{AspectRatio} folder under these branches of your personal (for individual usage) or local (for machine wide usage) \file{TEXMF} tree\footnote{Here we use the UNIX folder separator \texttt{/}; change \texttt{/} with \texttt{\textbackslash} on Windows platforms. Moreover we do not specify the full path from the disk root, but we show only the end branches preceded by an ellipsis.}:
\begin{itemize}
\item \file{.../doc/latex/AspectRatio/}: move this file{ar.pdf} file to this folder;
\item \file{.../source/latex/AspectRatio/}: move the corresponding \file{ar.tex} file to this folder;
\item \file{.../tex/latex/AspectRatio/}: move the \file{ar.sty} to this folder;
\item \file{.../fonts/source/public/AspectRatio/}: move all the \file{*.mf} files to this folder;
\item \file{.../fonts/tfm/public/AspectRatio/}: move all the \file{*.tfm} files to this folder;
\item \file{.../fonts/type1/public/AspectRatio/}: move all the \file{*.,pub} files to this folder;
\item \file{.../fonts/map/AspectRatio/}: move all the \file{*.map} files to this folder.
\end{itemize}
If your TeX distribution requires it, refresh the file name database according to the procedure described in your system guide.
In order to use the Type~1 fonts, you must follow the instructions that came with your \TeX\ system distribution; the idea with all distributions is to let the configuration file \file{updmap.cfg} know about the new map file; you have to execute a \TeX\ system command in order to regenerate the map file suitable for the pdftex, dvips, and dvipdfm programs; after the operation is completed the \LaTeX\ engine will use the bitmapped fonts in the DVI output file, while the other vector font oriented engines will use the vector fonts in Type~1 format.
\section{Author guide}
The symbol \AR, as a means for indicating the aspect ratio, is meaningful only in mathematics, so that only the ``italic'' shape makes sense. As an isolated symbol the \AR\ ligature might be used also in text mode outside a math expression.
Keeping in mind that only the ``italic'' shape is available, with the Computer Modern and the Times or Palatino variants, the ligature produced by \cs{AR} is sensitive to the text series declarations or commands. In math mode the bold symbol is available only through the explicit command \cs{ARb}. This asymmetry arises in order to avoid defining yet another math font group with an alphabet for medium math and a variant for bold math. Remember that when typesetting math, the various typesetting engines can handle a maximum of 16 alphabets; it does not happen very often, but while testing this very package with different fonts we eventually exceeded this maxim number and we produced an immediate fatal error. Therefore we decided that for just one glyph it was totally useless to define a new math group and we went around this problem with a \cs{mathcoice} work-around that allows us to actually typeset in the proper math size within a box to be used in math mode; apparently complicated, but actually very simple. This, on the other side, forbids us to use the regular math font switching commands to adapt the font family to the requested one.
Table~\ref{tab:glyphs} shows the various ways to produce the \AR\ symbol in math and in text mode with the various families available with the Computer Modern compliant \AR\ symbol. With Times and Palatino there is only one family available, the serifed one, and the variant commands resort to the italic shape.
If you get the source file of this documentation (\file{ar.tex} and comment out certain lines in the preamble, while uncommenting certain other lines (clearly marked in the preamble itself) you can typeset again this same file with the Times o Palatino fonts and you can verify the outcome of the various commands in the above mentioned table and in the following examples.
\begin{table}
\vspace*{-\bigskipamount}
\caption{Symbols and commands in both math and text modes}\label{tab:glyphs}
\bigskip
{\centering
\begin{tabular}{lcccc}
\toprule
Description & Math & Math & Text & Text \\
& symbol& command &symbol & command \\
\midrule
Medium ligature & \AR & \cs{AR} & \AR & \cs{AR} \\
Bold ligature &$\ARb$
&\cs{ARb}
&\textbf{\AR}
&\cs{textbf}\Arg{\cs{AR}}\\
Medium sans serif ligature& & &\ARss
&\cs{ARss}\\
Bold sans serif ligature & & &\textbf{\ARss}
&\cs{textbf}\Arg{\cs{ARss}}\\
Monospaced ligature & & &\ARtt
&\cs{ARtt}\\
\bottomrule
\end{tabular}
\par}
\medskip
{\footnotesize
By means of the specified options it is possible to get the corresponding glyphs in Times or Palatino. Remember, though, that both Times and Palatino have only the proportional serifed medium and bold fonts, so there is no possibility of using the sans serif or the monospaced version. Avoid using the Computer Modern monospaced version; it is available, but certainly the complex ligature does not fit well into the narrow character width of the monospaced fonts.\par}
\end{table}
Here we show the various series, shapes, and sizes available with the package; with the CM fonts and their siblings you get the various families produce the proper variations; with the Times and Platino variants you get only the italics medium and bold variations always in serifed italics.
\begin{center}
\noindent\begin{tabular}{ll}
(\aroption Regular)& {\tiny \AR} {\scriptsize \AR} {\footnotesize \AR} {\small \AR} {\normalsize \AR} {\large \AR} {\Large \AR} {\LARGE \AR} {\huge \AR} {\Huge \AR}\\
(\aroption Bold)& {\bfseries{\tiny \AR} {\scriptsize \AR} {\footnotesize \AR} {\small \AR} {\normalsize \AR} {\large \AR} {\Large \AR} {\LARGE \AR} {\huge \AR} {\Huge \AR}}\\
(\aroption Sans serif regular)& {\tiny \ARss} {\scriptsize \ARss} {\footnotesize \ARss} {\small \ARss} {\normalsize \ARss} {\large \ARss} {\Large \ARss} {\LARGE \ARss} {\huge \ARss} {\Huge \ARss}\\
(\aroption Sans serif bold)& {\bfseries{\tiny \ARss} {\scriptsize \ARss} {\footnotesize \ARss} {\small \ARss} {\normalsize \ARss} {\large \ARss} {\Large \ARss} {\LARGE \ARss} {\huge \ARss} {\Huge \ARss}}\\
(\aroption Monospaced regular)& {\tiny \ARtt} {\scriptsize \ARtt} {\footnotesize \ARtt} {\small \ARtt} {\normalsize \ARtt} {\large \ARtt} {\Large \ARtt} {\LARGE \ARtt} {\huge \ARtt} {\Huge \ARtt}
\end{tabular}
\end{center}
\section{Legal issues}
This work is distributed under the LaTeX Project Public Licence, version 1.3 or later; the Licence text is distributed with any distribution of the TEX system.
This package is temporarily maintained by Claudio Beccari \texttt{claudio dot beccari at gmail dot com}.
\end{document}
|
\section*{The Einstein metrics with smooth scri}
\end{center}
-\centerline{J. Tafel}
\noindent
\centerline{Institute of Theoretical Physics, University of Warsaw,}
\centerline{Pasteura 5, 02-093 Warsaw, Poland, email: <EMAIL>}
\bigskip
\begin{abstract}
We consider solutions of the Einstein equations with cosmological constant $\Lambda\neq 0$ admitting conformal compactification with smooth scri $\mathscr{I^+}$. Metrics are written in the Bondi-Sachs coordinates and expanded into inverse powers of the affine distance $r$. Unlike in the case $\Lambda=0$ all free data are located on the scri. There are linear differential constraints on the Bondi mass and angular momentum aspects. All other components of metrics are defined in a recursive way.
\end{abstract}
\bigskip
\section{Introduction}
This paper is a continuation of \cite{t} where the case $\Lambda=0$ is considered. We assume that metric $\tilde g$ satisfying vacuum equations with $\Lambda\neq 0$ is conformally equivalent to metric $g$ which is smooth in a neighbourhood of the scri $\mathscr{I^+}$.
We consider its Bondi-Sachs form with the affine distance $r$. Because of this choice our approach slightly differs from the standard one using the luminosity distance (see \cite{bt,c} and references). If metric is expanded into inverse powers of $r$ almost all Einstein equations can be solved in an algebraic way with respect to metric coefficients. Only the mass aspect $M$ and the angular momentum aspect $L_A$ undergo some linear differential constraints on $\mathscr{I^+}$. All free data are defined on the scri. The constraints on $M$ and $L_A$ reduce to a single Laplace-Beltrami equation in special coordinates. If metric tends asymptotically to the (anti) de Sitter solution this equation implies preservation of the generalized Bondi energy and the oscilatory (for $\Lambda<0$) or
exponential (for $\Lambda>0$) behaviour of the linear momentum.
\section{Recursive solvability of the Einstein equations}
Let ($M,g$) be a partial conformal compactification of spacetime ($\tilde M,\tilde g$) with
a boundary $\mathscr{I^+}=R\times \Sigma$ foliated by 2-dimensional spacelike surfaces.
In the standard way \cite{t} in this neighbourhood we introduce a system of the Bondi-Sachs coordinates $u$, $\Omega=1/r$, $x^A$ such that
\begin{equation}\label{1}
g=du(g_{00}du-2d\Omega+2 g_{0A}dx^A)+g_{AB}dx^Adx^B\ ,
\end{equation}
where
\begin{equation}\label{2}
\hat g_{0A} =0
\end{equation}
(the hat denotes value on $\mathscr{I^+}$). We assume that components $g_{\mu\nu}$ admit the Taylor expansion in $\Omega$ of any order.
The physical metric is given by
\begin{equation}\label{3}
\tilde g=\Omega^{-2} g=du(\tilde g_{00}du+2dr+2\tilde g_{0A}dx^A)+\tilde g_{AB}dx^Adx^B\ ,
\end{equation}
where $\tilde g_{00},\tilde g_{AB}$ are of the order $r^2$ and $\tilde g_{0A}=O(r)$.
We impose on (\ref{3}) the Einstein equations with cosmological constant $\Lambda\neq 0$
\begin{equation}\label{6a}
\tilde R_{\mu\nu}+\Lambda\tilde g_{\mu\nu}=0\ .
\end{equation}
In terms of the compactified metric they are
\begin{equation}\label{7}
R_{\mu\nu}-2Y_{\mu\nu}-g_{\mu\nu}Y^{\ \alpha}_{\alpha}=0 \ ,
\end{equation}
where
\begin{equation}\label{8}
Y_{\mu\nu}=-\frac{1}{\Omega}\Omega_{|\mu\nu}+\frac {1}{2\Omega^{2}}(\Omega_{|\alpha}\Omega^{|\alpha}-\frac{\Lambda}{3})g_{\mu\nu}
\end{equation}
and ${}_{|\mu}$ denotes the covariant derivative defined by $g$.
Assume that $\tilde R_{\mu\nu}+\Lambda\tilde g_{\mu\nu}$ is finite on $\mathscr{I^+}$. Then $Y_{\mu\nu}$ must be also finite. In order to avoid a second order pole at $\Omega=0$ one has to assume that
\begin{equation}\label{10}
g_{00}=-\frac{\Lambda}{3}+a\Omega+b\Omega^2-2M\Omega^3+ O(\Omega^4)\ .
\end{equation}
Thus, metric induced by $g$ on the scri reads
\begin{equation}\label{8b}
\hat g=-\frac{\Lambda}{3}du^2+\hat g_{AB}dx^Adx^B
\end{equation}
and $\mathscr{I^+}$ is spacelike if $\Lambda>0$ and timelike if $\Lambda<0$. For the (A)dS metric there is $\hat g_{AB}=-s_{AB}$. Still we are free to change coordinates $u,x^A$ provided that $\hat g$ preserves form (\ref{8b}) modulo a conformal factor.
It means that we can impose one condition on $\hat g_{AB}$, e. g.
we can try to obtain
\begin{equation}\label{11f}
\det{\hat g_{AB}}=f\ ,\ \ f_{,0}=0
\end{equation}
with a prescribed function $f$.
For instance, if $\mathscr{I^+}=R\times S_2$, where $S_2$ is the 2-dimensional sphere, a natural condition would be
\begin{equation}\label{11g}
\det{\hat g_{AB}}=\det{s_{AB}}\ .
\end{equation}
Equation (\ref{11g}) is a reminiscent of the Bondi luminosity gauge
\begin{equation}\label{11h}
\det{g_{AB}}=\det{s_{AB}}\ .
\end{equation}
Regularity of $Y_{1A}$ yields
\begin{equation}\label{10a}
g_{0A}=q_A\Omega^2+2L_A\Omega^3+O(\Omega^4)\ .
\end{equation}
A lack of singularities in $Y_{AB}$ is equivalent to the relation
\begin{equation}\label{11}
\frac{\Lambda}{3}n_{AB}=\hat g_{AB,0}-a\hat g_{AB}
\end{equation}
between expansion coefficients of $g_{AB}$
\begin{equation}\label{12}
g_{AB}=\hat g_{AB}+n_{AB}\Omega +p_{AB}\Omega^2+O(\Omega^3)\ .
\end{equation}
All other components of $Y_{\mu\nu}$ are nonsingular at $\Omega=0$ under conditions (\ref{10})-(\ref{11}).
Unlike in the case $\Lambda=0$, equation (\ref{11}) defines $n_{AB}$ in terms of $\hat g_{AB}$ and $a$. This relation can be further simplified by means of a shift of $r$ leading to
\begin{equation}\label{12a}
a=0\ .
\end{equation}
We will not assume (\ref{12a}) at this stage since another condition
\begin{equation}\label{12b}
n=0\ ,
\end{equation}
where $n=\hat g^{AB}n_{AB}$, may be more convenient. Note that (\ref{12b}) is equivalent to
\begin{equation}\label{12c}
a=\frac 12(\ln{\det\hat g_{AB}})_{,0}
\end{equation}
thanks to (\ref{11}).
Note that under (\ref{11f}) conditions (\ref{12a}) and (\ref{12b}) are equivalent.
We will test usefulness of the above conditions after obtaining field equations.
Folowing \cite{t} we can reduce a number of the Einstein equations (\ref{6a}) by means of the Bianchi identity. To this end it is sufficent to replace $\tilde R_{\mu\nu}$ by $\tilde R_{\mu\nu}+\Lambda\tilde g_{\mu\nu}$ and $\tilde G_{\mu \nu}$ by $\tilde G_{\mu \nu}-\frac 12\Lambda\tilde g_{\mu\nu}$ in equation (24) in \cite{t}. In this way one obtains a minimal system of equations given by
\begin{equation}\label{40a}
\tilde R_{11}^{(k)}=\tilde R_{1A}^{(k)}=0\ ,\ \ (\tilde R_{AB}+\Lambda \tilde g_{AB})^{(k)}=0
\end{equation}
and
\begin{equation}\label{40b}
(\tilde R_{00}+\Lambda \tilde g_{00})^{(2)}=(\tilde R_{0A}+\Lambda \tilde g_{0A})^{(2)}=0\ ,
\end{equation}
where the subscript $(k)$ denotes coefficient of order $k$ in an expansion with respect to $\Omega$.
Equation $\tilde R_{11}=0$ and $\tilde R_{1A}=0$ are similar in character to those for $\Lambda=0$. They yield
\begin{equation}\label{a2a}
\hat g^{AB}g^{(k+2)}_{AB}=\langle g^{(l)}_{AB},l\leq k+1\rangle\ ,\ k\geq 0\ ,
\end{equation}
\begin{equation}\label{a4a}
g^{(k+2)}_{0A}=\langle g^{(l)}_{\mu\nu},l\leq k+1\rangle\ ,\ k\geq 0\ ,\ k\neq 1\ ,
\end{equation}
where brackets on the r. h. s. denote expressions depending on functions within the bracket.
For $k=0$ equation (\ref{a4a}) reads
\begin{equation}\label{a5}
q_A=-\frac 12n^B_{\ A|B}+\frac 12n_{,A}\ .
\end{equation}
Equation $\tilde R_{1A}^{(1)}=0$ does not define $g^{(3)}_{0A}$. It is equivalent to
\begin{equation}\label{a6}
(p^B_{\ A}+\frac 14 nn^A_{\ B}-\frac 12 n^{BC}n_{AC})_{|B}+\frac 18(n_{BC}n^{BC}- n^2)_{,A}=0\ ,
\end{equation}
where operations on indices and covariant derivatives are defined by $\hat g_{AB}$.
For $\Lambda=0$ we used in \cite{t} the topology of $\mathscr{I^+}=R\times S_2$ to show that (\ref{a6}) yields
\begin{equation}\label{a7}
p_{AB}=\frac 18(n_{AB}n^{AB}-n^2)\hat g_{AB}+\frac 14nn_{AB}\ .
\end{equation}
For $\Lambda\neq 0$
equation (\ref{a7}) follows algebraically from $(\tilde R_{AB}+\Lambda \tilde g_{AB})^{(0)}=0$. The latter equation also implies
\begin{equation}\label{22}
b=-\frac 12\hat R+\frac 12n_{,0}+\frac 14an+\frac{\Lambda}{8}(n_{AB}n^{AB}-\frac 13n^2)\ ,
\end{equation}
where $\hat R$ is the Ricci scalar of $\hat g_{AB}$. From the point of view of equations (\ref{a7}) and (\ref{22}) the most convenient gauge condition is (\ref{12b}). In this gauge $n_{AB}$ is proportional to the traceless part of the exterior curvature of the section $u=const$ of $\mathscr{I^+}$ and $p_{AB}$ is proportional to metric of this section.
Equation $(\tilde R_{AB}+\Lambda \tilde g_{AB})^{(k)}=0$ with $k=1$ does not contain any new information. For $k\geq 2$ it can be splitted into its trace (with respect to $\hat g_{AB}$) and a traceless part. The trace part is equivalent to
the following composition of equations
\begin{equation}\label{a10}
\hat g^{AB}(\tilde R_{AB}+\Lambda \tilde g_{AB})^{(k)}-\frac{\Lambda (k-3)}{3(k+1)}\tilde R^{(k)}_{11}=0\ ,\ k\geq 2\ .
\end{equation}
It allows to obtain $g_{00}^{(k+2)}$ in terms of lower order coefficients
\begin{equation}
g^{(k+2)}_{00}=\langle g^{(l)}_{\mu\nu},l\leq k+1\rangle\ ,\ k\geq 2\ .\label{a11}
\end{equation}
The traceless part of $(\tilde R_{AB}+\Lambda \tilde g_{AB})^{(k)}=0$ yields the traceless part of $g_{AB}^{(k)}$ (denoted by a ring)
\begin{equation}\label{a12}
\mathring g^{(k+2)}_{AB}=\langle g^{(l)}_{\mu\nu},l\leq k+1\rangle\ ,\ k\geq 2\ .
\end{equation}
Note that in the case $\Lambda=0$ instead of (\ref{a12}) one obtains an expression for $\mathring g^{(k+1)}_{AB,0}$ in terms of lower order coefficients \cite{t}. Thus, (\ref{a12}) is the second equation, after (\ref{11}), which makes a qualitative difference between $\Lambda=0$ and $\Lambda\neq 0$. For $\Lambda\neq 0$ there is no need of initial data for $g_{AB}$. Moreover, tensor
\begin{equation}\label{53a}
N_{AB}=\mathring g_{AB}^{(3)}
\end{equation}
(instead of $n_{AB}$) is arbitrary if $\Lambda\neq 0$.
The last equations to consider are (\ref{40b}). They are equivalent to equations
\begin{equation}\label{a22}
(\tilde R_{00}+\Lambda \tilde g_{00})^{(2)}-\frac{\Lambda}{6}\hat g^{AB}(\tilde R_{AB}+\Lambda \tilde g_{AB})^{(2)}-\frac{\Lambda^2}{18}\tilde R _{11}^{(2)}-\frac{\Lambda}{9}a\tilde R_{11}^{(1)}=0\ ,
\end{equation}
\begin{equation}\label{b22}
(\tilde R_{0A}+\Lambda \tilde g_{0A})^{(2)}-\frac{\Lambda}{3}\tilde R_{1A}^{(2)}+\frac {\Lambda}{6}\tilde R_{11,A}^{(1)}=0
\end{equation}
having the following structure
\begin{equation}\label{22a}
M_{,0}+\frac{3}{4}(\ln{|\hat g|})_{,0}M+\frac{\Lambda}{2}L^A_{\ |A}=\frac{\Lambda^2}{24}n^{AB} N_{AB}+\langle a,\hat g_{AB}\rangle\ ,
\end{equation}
\begin{equation}\label{22b}
L_{A,0}+\frac 12(\ln{|\hat g|})_{,0}L_A+\frac 13M_{,A}=\frac{\Lambda}{6} N_{AB}^{\ \ |B}+\langle a,\hat g_{AB}\rangle\ .
\end{equation}
Given $a,\ \hat g_{AB}$ and $N_{AB}$ equations (\ref{22a}) and (\ref{22b}) constitute a conjugate system of linear equations for $M$ and $L_A$.
Condition (\ref{11f}) allows to simplify them to
\begin{equation}\label{a22a}
M_{,0}+\frac{\Lambda}{2}L^A_{\ |A}=f\ ,
\end{equation}
\begin{equation}\label{a22b}
L_{A,0}+\frac 13M_{,A}=h_A\ ,
\end{equation}
where $f$ and $h_A$ are known functions. Let us introduce function $ M'$ such that
\begin{equation}\label{53}
M=M'_{,0}\ .
\end{equation}
Then (\ref{a22b}) yields
\begin{equation}\label{54}
L_A=-\frac 13 M'_{,A}+\int{h_Adu}
\end{equation}
and (\ref{a22a}) becomes
\begin{equation}\label{55}
-\frac{6}{\Lambda}M'_{,00}+\hat \Delta M'=-\frac{6}{\Lambda}f+3\nabla^A\int{h_Adu}\ ,
\end{equation}
where $\hat \Delta$ is the covariant Laplace operator with respect to $\hat g_{AB}$. Thus, equations (\ref{22a}) and (\ref{22b}) reduce to a hyperbolic (if $\Lambda<0$) or elliptic ($\Lambda>0$) equation for $M'$.
Note that the operator acting on $M'$ is not the Laplace-Beltrami operator of the induced metric (\ref{8b}) on $\mathscr {I}^+$ because of the ``wrong`` coefficient $\frac{6}{\Lambda}$.
Let us summarize consequences of the Einstein equations expanded into powers of $1/r$.
\begin{prop}
The Einstein metric with $\Lambda\neq 0$ and smooth scri $\mathscr{I^+}$ is given by
\begin{equation}\label{3a}
\tilde g=du(\tilde g_{00}du+2dr+2\tilde g_{0A}dx^A)+\tilde g_{AB}dx^Adx^B\ ,
\end{equation}
\begin{equation}\label{a10d}
\tilde g_{00}=-\frac{\Lambda}{3}r^2+ar+b-\frac{2M}{r}+\Sigma_2^{\infty} g_{00}^{(k+2)}r^{-k}\ ,
\end{equation}
\begin{equation}\label{a10a}
\tilde g_{0A}=q_A+\frac{2L_A}{r} +\Sigma_2^{\infty} g_{0A}^{(k+2)}r^{-k}\ ,
\end{equation}
\begin{equation}\label{a12c}
\tilde g_{AB}=r^2\hat g_{AB}+rn_{AB} +p_{AB}+\frac 1r(N_{AB}+N\hat g_{AB})+\Sigma_2^{\infty} g_{AB}^{(k+2)}r^{-k}\ ,
\end{equation}
where coefficients are defined in the following way:
\begin{itemize}
\item u-dependent metric $\hat g_{AB}$ and a traceless (with respect to $\hat g_{AB}$) tensor $N_{AB}$ can be arbitrarily chosen up to a gauge condition, e.g. (\ref{11f}).
\item Function $a$ is arbitrary but it can be gauged away, e.g. by means of (\ref{12c}).
\item Components $n_{AB}$, $p_{AB}$, $N$, $q_A$ and $b$ are defined by $a$ and $\hat g_{AB}$ according to (\ref{11}), (\ref{a7}), (\ref{a2a}), (\ref{a5}) and (\ref{22}).
\item Coefficients $M$ and $L_A$ are defined by their initial values at $u=u_0$ via equations (\ref{22a}) and (\ref{22b}). Under condition (\ref{11f}) these equations reduce to equation (\ref{55}) for $M'$ and relations (\ref{53}) and (\ref{54}).
\item All other components are defined in a recursive way by equations (\ref{a2a}), (\ref{a4a}), (\ref{a11}) and (\ref{a12}).
\end{itemize}
\end{prop}
The main difference between present situation and that for $\Lambda=0$ (see Proposition 2.1 in \cite{t}) is that now all free data are located at $\mathscr{I^+}$. It is interesting that for $\Lambda<0$ there is no need of the Cauchy data ($\mathscr{I^+}$ is timelike).
Evolution of the Cauchy data given on $\mathscr{I^+}$ for $\Lambda>0$ was first obtained by Friedrich \cite{f}.
\section{Solutions with the (A)dS boundary metric}
If metric $\tilde g$ tends asymptotically to the (anti) de Sitter solution then
\begin{equation}\label{56}
\hat g_{AB}=-s_{AB}\ .
\end{equation}
In this case it follows from equations (\ref{10})-(\ref{11}) and (\ref{a2a})-(\ref{22}) that in the gauge $a=0$ there is
\begin{equation}\label{56a}
\tilde g_{00}=-\frac{\Lambda}{3}r^2+1-\frac{2M}{r}+O(\frac{1}{r^2})\ ,
\end{equation}
\begin{equation}\label{56b}
\tilde g_{0A}=\frac{2L_A}{r}+O(\frac{1}{r^2})\ ,
\end{equation}
\begin{equation}\label{56c}
\tilde g_{AB}=-r^2s_{AB}+\frac 1r N_{AB}+O(\frac{1}{r^2})\ .
\end{equation}
The r. h. s. of equations (\ref{a22a}) and (\ref{a22b}) is given by
\begin{equation}\label{57}
h_A=\frac{\Lambda}{6} N^{\ \ |B}_{AB}\ ,\ f=0\ .
\end{equation}
Let us introduce a tensor $N'_{AB}$ such that
\begin{equation}\label{58}
N_{AB}=N'_{AB,0}\ ,\ \ s^{AB}N'_{AB}=0\ .
\end{equation}
Equation (\ref{54}) yields
\begin{equation}\label{59}
L_A=-\frac 13 M'_{,A}+\frac{\Lambda}{6}{N'}_{AB}^{\ \ |B}+\tilde L_A\ ,\ \ \tilde L_{A,0}=0\ .
\end{equation}
Using decomposition
\begin{equation}\label{59a}
\tilde L_A=f_{,A}+\eta^C_{\ A}h_{,C}
\end{equation}
we can incorporate $f$ into $M'$ and $h$ into $N'_{AB}$ except the dipole component of $h$ (see Lemma 3.1 in \cite{t}). Thus, without loss of generality we can assume that
\begin{equation}\label{59b}
\tilde L_A=\eta^C_{\ A}h_{,C}\ ,\ \ h=\alpha^kY_{1k}\ ,
\end{equation}
where $Y_{lk}$ are spherical harmonics and $\alpha^k=const$.
In spherical coordinates $\theta,\varphi$ related to the axis defined by $\alpha^k$ one obtains expression
\begin{equation}\label{60a}
\tilde L_Adx^A=\alpha\sin^2{\theta}d\varphi\ ,
\end{equation}
which is the angular momentum term known from the Kerr solution.
In view of (\ref{59}) and (\ref{59b}) equation (\ref{a22a}) transforms into
\begin{equation}\label{60}
\frac{6}{\Lambda}M'_{,00}+\Delta M'=-\frac {\Lambda}{2} {N'}^{AB}_{\ \ \ |AB}\ ,
\end{equation}
where $\Delta$ is the standard Laplace operator on the sphere.
If we integrate equation (\ref{60}) over the sphere we obtain the known result
\begin{equation}\label{81}
\oint_{S_2}{Md\sigma}=const.
\end{equation}
The l. h. s. of (\ref{81}) is, modulo $4\pi$, rather unique candidate for the total energy in this case. One can interpret (\ref{81})
as a lack of gravitational radiation.
If we integrate (\ref{60}) with $Y_{1k}$ we get equation
\begin{equation}\label{85}
\frac{6}{\Lambda}M'_{k,00}-2 M'_k=0
\end{equation}
for dipole moments
\begin{equation}\label{84}
M'_k=\oint_{S_2}{M'Y_{1k}d\sigma}\ .
\end{equation}
The Bondi linear momentum $P_k$ is proportional to the time derivatives of $M'_k$, so it satisfies equation
\begin{equation}\label{86}
P_{k,00}=\frac{\Lambda}{3}P_k\ .
\end{equation}
Hence, it either oscilates when $u$ changes (for $\Lambda<$) or behaves in an exponential way (for $\Lambda>0$). In the second case its square must exceed the total energy either for increasing or decreasing $u$, so a reasonable physical assumption would be $M_k=0$.
Higher moments of $M'$ satisfy nonhomogeneous equations coming from (\ref{60}). Their solutions are defined in quadratures up to oscillatory or exponential functions depending on sign of $\Lambda$.
An alternative approach to equation (\ref{60}) is to treat it as an equation for tensor $N'_{AB}$. We can represent this tensor
by two scalar functions $f$ and $h$ \cite{t}
\begin{equation}\label{62}
N'_{AB}=f_{|AB}-\frac 12 \Delta fs_{AB}+\eta^C_{\ (A}h_{|B)C}\ .
\end{equation}
Substituting (\ref{62}) into (\ref{60}) yields
\begin{equation}\label{63}
\frac{6}{\Lambda}M'_{,00}+\Delta M'=-\frac {\Lambda}{4} (\Delta+2)\Delta f\ .
\end{equation}
Function $f$ exists if conditions (\ref{81}) and (\ref{84}) are satisfied.
Then we can write
\begin{equation}\label{66a}
f=-\frac{4}{\Lambda}\Delta^{-1}(\Delta+2)^{-1}(\frac{6}{\Lambda}M'_{,00}+\Delta M')\ .
\end{equation}
Formula (\ref{66a}) can be realised by decomposition of $M'$ into $Y_{lk}$. In this approach functions $M'$ and $h$ are arbitrary modulo conditions (\ref{81}) and (\ref{84}).
We summarize above results in the following proposition.
\begin{prop}
The Einstein metric which has a smooth scri $\mathscr{I^+}$ and tends asymptotically to the (A)dS metric is given by (\ref{3a}) and (\ref{56a})-(\ref{56c}).
Components $M$, $N_{AB}$ and $L_A$ are defined by solution $(M',N'_{AB})$ of equation (\ref{60}) via relations (\ref{53}), (\ref{58}), (\ref{59}) and (\ref{60a}). Higher order components
are defined in a recursive way by equations (\ref{a2a}), (\ref{a4a}), (\ref{a11}) and (\ref{a12}). The Bondi energy is constant and the Bondi linear momentum satisfies (\ref{86}).
\end{prop}
\section{Stationary merics}
If metric considered in section 2 admits a timelike Killing vector $K$ then we can transform it to the form (\ref{3a}) with $u$-independent coefficients but we cannot assume (\ref{2}) (see Section 3 in \cite{t})
The KIlling vector is given by $K=\partial_0$.
Let us assume (\ref{12}) and expansions
\begin{equation}\label{62a}
g_{00}=\hat g_{00}+a\Omega+b\Omega^2-2M\Omega^3+ O(\Omega^4)\ ,
\end{equation}
\begin{equation}\label{62b}
g_{0A}=\hat g_{0A}+v_{A}\Omega+q_A\Omega^2+2L_A\Omega^3+O(\Omega^4)
\end{equation}
with coefficients depending on coordinates $x^A$.
Tensor (\ref{8}) is finite on the boundary iff
\begin{equation}\label{63}
\hat g_{00}=-\frac{\Lambda}{3}+\hat g_{0A}\hat g_0^{\ A}\ ,
\end{equation}
\begin{equation}\label{63a}
v_A=n_A^{\ B}\hat g_{0B}
\end{equation}
and
\begin{equation}\label{64}
\frac{\Lambda}{3}n_{AB}=-2\hat g_{0(A|B)}+(n_{CD}\hat g_0^{\ C}\hat g_0^{\ D}-a)\hat g_{AB}\ ,
\end{equation}
where indices $A,B$ are raised by means of $\hat g^{AB}$. Since $\hat g_{0A}\hat g_0^{\ A}\leq 0$ and we exclude $\Lambda=0$ it follows from (\ref{63})-(\ref{64}) and timelike property of $K=\partial_0$ that
\begin{equation}\label{72}
\Lambda<0\ ,\ \ |\hat g_{0A}\hat g_0^{\ A}|\leq \frac{|\Lambda|}{3}\ .
\end{equation}
Thus, metrics are of the AdS type.
It is convenient to shift coordinate $r$ in order to obtain
\begin{equation}\label{64c}
n^A_{\ A}=0\ .
\end{equation}
Equation (\ref{64}) is then equivalent to the following relations defining $n_{AB}$ and $a$
\begin{equation}\label{64b}
\frac{\Lambda}{3}n_{AB}=-2\hat g_{0(A|B)}+\hat g_{0\ |C}^{\ C}\hat g_{AB}\ ,
\end{equation}
\begin{equation}\label{64a}
a=n_{CD}\hat g_0^{\ C}\hat g_0^{\ D}-\hat g_{0\ |A}^{\ A}\ .
\end{equation}
Thus, again $n_{AB}$ is proportional to the traceless part of the exterior curvature of the section $u=const$ of $\mathscr{I^+}$.
Transformation of coordinates $x^A$ allows us to obtain
\begin{equation}\label{64d}
\hat g_{AB}=-\gamma^2s_{AB}\ ,
\end{equation}
where $\gamma$ is a positive function of $x^A$. Still there is a freedom of supertranslation
\begin{equation}
u\rightarrow u+f(x^A)
\end{equation}
which can be used to impose a condition on $\hat g_{0A}$, $\gamma$ or higher order coefficients in metric $g$.
All the Einstein equations except (\ref{40b})
can be solved algebraically as in the general case in Section 2.
Equations (\ref{40b})
are now a complicated system of constraints on $M$, $L_A$, $\gamma$, $\hat g_{0A}$ and $N_{AB}$, which is linear in $M$, $L_A$ and $N_{AB}$.
The space of stationary solutions is much bigger than in the case $\Lambda=0$ \cite{t}. In general, solutions do not tend asymptotically to the AdS metric. It is so if
the boundary metric
\begin{equation}\label{65}
\hat g=-\frac{\Lambda}{3}du^2-\gamma^2 s_{AB}(dx^A+\hat g_0^{\ A}du)(dx^B+\hat g_0^{\ B}du)
\end{equation}
is conformally equivalent to
\begin{equation}\label{66}
\hat g'=-\frac{\Lambda}{3}du'^2-s_{AB}dx'^Adx'^B\ .
\end{equation}
Metric (\ref{66}) is conformally flat (note that it coincides with the conformal compactification of the 3-dimensional Minkowski metric). Thus,
metric (\ref{65}) must be also conformally flat. Since $K=\partial_u$ is the Killing field of (\ref{65}) it is a conformal Killing vector of the flat metric. Such vectors compose the 10-dimensional algebra $so(2,3)$.
Given one of them in terms of the Cartesian coordinates of flat space one can find (at least in principle) coordinates $u,\ x^A$ such that $K=\partial_u$. Writing flat metric in these coordinates shows what metrics (\ref{65}) are available
if the physical metric $\tilde g$ is asymptotically AdS (in transformed coordinates).
\section{Summary}
In this paper we considered the Einstein metrics admitting a conformal compactification with smooth scri. For nonvanishing cosmological constant the scri is either spacelike ($\Lambda>0$) or timelike ($\Lambda<0$). Metrics can be transformed to the Bondi-Sachs form with components which can be expanded into powers of a radial distance. The Einstein equations imply that the expansion coefficients can be expressed by coefficients of lower order (see Proposition 2.1). Beside these recursive formulas there are 3 differential conditions constraining the analog of the Bondi mass and angular momentum aspect. For a special foliation of the scri they can be reduced to one second order equation for one function. All free data are located on the scri.
Asymptotic form of metric simplifies considerably if it tends to the (anty) de Sitter solution (see Proposition 3.1). In this case the total energy calculated according to the Bondi prescription is constant and the total momentum oscilates or behaves exponentially.
Finally we considered solutions with the timelike Killing field. In this case, in contrary to $\Lambda=0$, there is no big difference in solving the Einstein equations with respect to the general case.
However, existence of stationary solutions with nonvanishing asymptotic twist (then $\hat g_{0A}\neq 0$) can be important for a notion of the total energy in general case with $\Lambda< 0$.
We expect that a crucial ingredient in any definition of the energy should be the mass aspect $M$ related to coordinates such that equations (\ref{8b}) and (\ref{11h}) are satisfied. If
the timelike Killing field $K$ is present then transformation between coordinates used in (\ref{8b}) and those in (\ref{65}) must depend on time $u$. It seems unavoidable that the total energy also depend on time, what seems unacceptable in a stationary spacetime.
\null
\noindent
\textbf{Acknowledgments}
\null
\noindent
This work was partially supported by Project OPUS 2017/27/B/ST2/02806
of Polish National Science Centre (NCN).
|
\section{Introduction and Motivation \label{sec:intro}}
Naturalness reasoning strongly suggests that beyond the Standard Model (BSM) CP-odd phases should exist.
The 1964 experiment at BNL by Christenson et al~\cite{Christenson:1964fg} showed for the first time that CP is not a
symmetry of nature. Therefore, any interaction that we write down that can be complex, allowing for CP violation to occur,
simply has to be complex since there is no symmetry that requires it to be real.
But to discern the effects of
new physics (NP) from that of the Standard Model (SM) requires precision both in experimental measurements as well as in
theoretical predictions from the SM. More often than not the latter task can be very challenging because of the
influence of non-perturbative effects of QCD on the relevant channels for $K$, $D$ or $B$ decays involved in direct CP
violation studies.
This explicitly manifests itself in the well-known examples of $\epsilon'$ for
$K \to \pi \pi$~\cite{Abbott:2020hxn, Aebischer:2020jto, Cirigliano:2019ani},
$A_{CP}$ measurements in the four $B \to K, \pi$ modes~\cite{Zyla:2020zbs,Aaij:2020wnj}, and the more relevant example to this work, $\Delta A_{CP}$ in $D^0$ decays, where~\cite{Amhis:2019ckw, Aaij:2019kcg, BaBar:2007tfw, CDF:2012ous, LHCb:2014kcb, LHCb:2016csn, LHCb:2019hro}
\begin{align}
\Delta A_{CP} \approx \Delta a_{CP}^{\mathrm{dir}} &\equiv a_{CP}^{\mathrm{dir}}(D^0\rightarrow K^+K^-) - a_{CP}^{\mathrm{dir}}(D^0\rightarrow \pi^+\pi^-) \label{eq:DeltaACP}\\
&=-0.00161 \pm 0.00028\,,
\end{align}
and
\begin{align}
a_{CP}^{\mathrm{dir}}(f) &\equiv \frac{
|\mathcal{A}(D^0\rightarrow f) |^2 - |\mathcal{A}(\overline{D}^0\rightarrow f)|^2
}{
|\mathcal{A}(D^0\rightarrow f) |^2 + |\mathcal{A}(\overline{D}^0\rightarrow f)|^2
}\,.
\end{align}
Note that for charm decays to CP eigenstates the indirect contributions to CP violation $a_{CP}^{\mathrm{ind}}$ are universal to a very good approximation, and we can write the time-integrated CP asymmetry as~\cite{Grossman:2006jg}
\begin{align}
A_{CP} &\approx a_{CP}^{\mathrm{dir}} + a_{CP}^{\mathrm{ind}}\,, \label{eq:indirect} \\
a_{CP}^{\mathrm{ind}} &\equiv -\eta_f^{CP}\frac{y}{2} \left( \left|\frac{q}{p}\right| - \left|\frac{p}{q}\right| \right) \cos\phi +
\eta_f^{CP}\frac{x}{2} \left( \left|\frac{q}{p}\right| + \left|\frac{p}{q}\right| \right) \sin\phi\,,
\end{align}
with the weak mixing phase $\phi$ and using the standard notation for the mixing parameters~\cite{Grossman:2006jg}.
It follows that $a_{CP}^{\mathrm{ind}}$ cancels in the difference of CP asymmetries in Eq.~(\ref{eq:DeltaACP}).
Recently, LHCb and Belle also improved the precision of the measurement of the CP asymmetry of charm decays into neutral kaons, leading to the new average~\cite{Amhis:2019ckw, LHCb:2021rdn, Dash:2017heu, LHCb:2015ope, CLEO:2000opx}
\begin{align}
a_{CP}^{\mathrm{dir}}(D^0\rightarrow K_SK_S) &= -0.019\pm 0.010\,,
\end{align}
bringing the measurement closer to a possible second observation of CP violation in the charm system~\cite{Nierste:2015zra}.
For the case at hand of $\Delta A_{CP}$, there are theoretical frameworks~\cite{Chala:2019fdb, Khodjamirian:2017zdu} which seem to suggest that the SM Cabibbo-Kobayashi-Maskawa (CKM)
paradigm cannot account for the value measured by LHCb whereas others, such as ours, seem to model non-perturbative
effects in such a way that no significant difference exists at present between the experimental measurements and their estimated value based on the SM.
After the measurement presented in Ref.~\cite{Aaij:2019kcg}, different theoretical interpretations have been brought forward~\cite{Grossman:2019xcj,Chala:2019fdb, Li:2019hho, Soni:2019xko, Cheng:2019ggx, Buras:2021rdg, Acaroglu:2021qae} (see for earlier works Refs.~\cite{Einhorn:1975fw, Abbott:1979fw,Golden:1989qx, Brod:2012ud, Bhattacharya:2012ah, Franco:2012ck, Hiller:2012xm, Nierste:2017cua, Nierste:2015zra, Muller:2015rna, Grossman:2018ptn, Buccella:1994nf, Grossman:2006jg, Artuso:2008vf, Khodjamirian:2017zdu, Buccella:2013tya, Cheng:2012wr, Feldmann:2012js, Li:2012cfa, Atwood:2012ac, Grossman:2012ry, Buccella:2019kpn, Yu:2017oky, Brod:2011re,Pirtskhalava:2011va}). It is an open question if the observed amount of CP violation is to be interpreted as new physics or not.
Making precise theoretical predictions for the case of charm decays is particularly challenging because of the presence of many nearby resonances. These require a reliable non-perturbative framework. In recent years lattice methods have made some progress in this regard and are starting to show
limited success~\cite{Johnson:2020ilc, Rendon:2020rtw, Fischer:2020fvl, Bruno:2020kyl},
but are not quite ready yet to tackle $\Delta A_{CP}$ for charm decays. Therefore, we have to resort to phenomenological models.
As is well known, singly-Cabibbo suppressed charm decays are dominated by tree operators.
Through rescattering, penguin contractions of tree operators are formed, which have a relative weak and strong phase with respect
to the tree amplitude, leading to a non-vanishing CP asymmetry~\cite{Brod:2012ud, Grossman:2019xcj}.
An exception is the decay $D^0\rightarrow K_SK_S$, where the dominant contribution to CP violation comes from the interference of SU(3)$_F$-breaking and SU(3)$_F$-leading exchange diagrams,~\emph{i.e.},~CP violation is even present for
vanishing penguin annihilation~\cite{Nierste:2015zra, Nierste:2017cua, Hiller:2012xm, Atwood:2012ac}.
In this paper, we work out a specific dynamical mechanism of charm CP violation, namely the possible enhancement of rescattering through scalar resonances close to the $D^0$ mass. The possible role of the scalar resonances for charm CP violation has recently been pointed out in Ref.~\cite{Soni:2019xko}. In Ref.~\cite{Soni:2019xko} only the impact of the $f_0(1710)$ resonance has been studied. Here, we study the implications of the picture of Ref.~\cite{Soni:2019xko} in more detail and take also the implications of the $f_0(1790)$ resonance into account, which is even closer to $m_{D^0}$.\footnote{Thus this work supersedes Ref.~\cite{Soni:2019xko}.} Note that the $f_0(1790)$ is not yet fully established and so far not listed in Ref.~\cite{Zyla:2020zbs}.
It has first been seen by BESII~\cite{Ablikim:2004wn}, and has also been seen by LHCb~\cite{Aaij:2014emv}.\footnote{Note however, that the $f_0(1790)$ resonance has not been seen in the analysis of CLEO data in Ref.~\cite{Dobbs:2015dwa}.}
As the masses of these resonances are close to $m_{D^0}$, their effect might be significant. It is therefore reasonable to investigate their role for CP violation further. Unfortunately, available experimental information on the $f_0(1790)$ resonance is rather sparse and this will limit our ability to be precise.
The role of resonances for nonleptonic charm decays has been studied
for a long time~\cite{Donoghue:1979fp, Lipkin:1980es, Sorensen:1981vu, Kamal:1988ub, Buccella:1990sp, Buccella:1992ym, Buccella:1996uy, Fajfer:1999hh, Rosner:1999xd, Gronau:1999zt, Cheng:2010ry, Fusheng:2011tw, Biswas:2015aaa, Gronau:2015zga}, including also corresponding implications for
CP violation \cite{Eilam:1991yv, Atwood:1994zm, Li:2012cfa, Li:2013xsa, Buccella:1992sg, Buccella:2013tya, Buccella:2019kpn, Dery:2021mll}. The possible impact of scalar resonances on charm CP violation has also been mentioned in Ref.~\cite{Khodjamirian:2017zdu}.
Resonances play a role in the discussion of
charm mixing, too~\cite{Falk:1999ts,Golowich:1998pz, Bergmann:2000id, Falk:2001hx, Falk:2004wg}.
Note that with CP asymmetries measured in more channels in the future, the CPT theorem might have important implications for
the interplay of CP violation with final state interactions between different decay channels that share the same quantum numbers, i.e.~which can rescatter into each other, see Refs.~\cite{Gerard:1988jj, Wolfenstein:1990ks, Gerard:1990ni, Atwood:1997iw, Atwood:2000tu, Atwood:2012ac, Zwicky:2007vv, Bediaga:2020qxg}.
\section{Isospin decomposition \label{sec:isospin}}
Employing the notation of Refs.~\cite{Muller:2015lua, Muller:2015rna}, we write the amplitudes of
singly-Cabibbo suppressed (SCS) decays as
\begin{align}
\mathcal{A} &= \lambda_{sd} A_{sd} - \frac{\lambda_b}{2} A_b\,, \label{eq:general-decomposition}
\end{align}
where we use the CKM matrix element combinations
\begin{align}
\lambda_{sd} &= \frac{\lambda_s-\lambda_d}{2}\,, \quad
\lambda_q = V_{cq}^* V_{uq}\,.
\end{align}
The amplitudes are related to the branching ratio as
\begin{align}
\mathcal{B}(D\rightarrow P_1P_2) &= \vert \mathcal{A}\vert^2 \, \mathcal{P}\,,
\end{align}
with the phase space factor
\begin{align}
\mathcal{P} &= \frac{\tau_D}{16\pi m_D^3} \sqrt{(m_D^2 - (m_{P_1} - m_{P_2})^2 ) ( m_D^2 - (m_{P_1} + m_{P_2})^2)}\,.
\end{align}
Note that the phase space has mass dimension $[\text{GeV}^{-2}]$, such that the amplitudes have mass dimension $[\text{GeV}]$.
In this notation the direct CP asymmetries of SCS decays are given as
\begin{align}
a_{CP}^{\mathrm{dir}} &= \frac{\vert\mathcal{A}\vert^2 - \vert \overline{\mathcal{A}}\vert^2
}{
\vert\mathcal{A}\vert^2 + \vert \overline{\mathcal{A}}\vert^2
}
= \mathrm{Im}\left(\frac{\lambda_b}{\lambda_{sd}}\right) \mathrm{Im}\left( \frac{A_b}{A_{sd}}\right)\,,
\end{align}
where~\cite{Zyla:2020zbs}
\begin{align}
\mathrm{Im}\left(\frac{\lambda_b}{\lambda_{sd}}\right) &= -0.00059\pm 0.00003\,.
\end{align}
In the following, we consider the impact of resonances with definite isospin quantum numbers. Therefore, we parametrize the SCS charm decays in terms of isospin matrix elements.
We adapt the notation from Refs.~\cite{Atwood:2012ac, Franco:2012ck} in such a way as to make the dependence on the CKM matrix elements explicit. In this way, the hadronic matrix elements have no absorbed CKM matrix elements and carry a strong phase only:
\begin{align}
\mathcal{A}(D^+\rightarrow \pi^+\pi^0) &= \frac{\sqrt{3}}{2} \lambda_{sd} A^{\pi\pi}_{\frac{3}{2}, 2}\,, \label{eq:isospin-first} \\
\mathcal{A}(D^0\rightarrow \pi^+\pi^-) &= \frac{1}{\sqrt{6}} \lambda_{sd} A_{\frac{3}{2},2}^{\pi\pi} +
\frac{1}{\sqrt{3}} \left( \lambda_{sd} A^{\pi\pi}_{\frac{1}{2},0} -\frac{\lambda_b}{2} B^{\pi\pi}_{\frac{1}{2},0} \right) \,, \\
\mathcal{A}(D^0\rightarrow \pi^0\pi^0) &= \frac{1}{\sqrt{3}} \lambda_{sd} A_{\frac{3}{2},2}^{\pi\pi} -
\frac{1}{\sqrt{6}} \left( \lambda_{sd} A_{\frac{1}{2},0}^{\pi\pi} - \frac{\lambda_b}{2} B_{\frac{1}{2},0}^{\pi\pi} \right) \,,
\end{align}
and
\begin{align}
-\sqrt{2} \mathcal{A}(D^+\rightarrow K^+K_S) &= -\frac{1}{2} \lambda_{sd} A_{\frac{3}{2},1}^{KK} +
\left( \lambda_{sd} A_{\frac{1}{2},1}^{KK} -\frac{\lambda_b}{2} B_{\frac{1}{2},1}^{KK} \right)\,, \\
\mathcal{A}(D^0\rightarrow K^+K^-) &= \frac{1}{2} \lambda_{sd} A_{\frac{3}{2},1}^{KK} +
\frac{1}{2} \left( \lambda_{sd} A_{\frac{1}{2},1}^{KK} - \frac{\lambda_b}{2} B_{\frac{1}{2},1}^{KK} \right) +
\frac{1}{2} \left( \lambda_{sd} A_{\frac{1}{2},0}^{KK} - \frac{\lambda_b}{2} B_{\frac{1}{2},0}^{KK}\right)\,,\\
-\sqrt{2} \mathcal{A}(D^0\rightarrow K_SK_S) &= \frac{1}{2} \lambda_{sd} A_{\frac{3}{2},1}^{KK} +
\frac{1}{2} \left( \lambda_{sd} A_{\frac{1}{2},1}^{KK} - \frac{\lambda_b}{2} B_{\frac{1}{2},1}^{KK} \right) -
\frac{1}{2} \left( \lambda_{sd} A_{\frac{1}{2},0}^{KK} - \frac{\lambda_b}{2} B_{\frac{1}{2},0}^{KK} \right)\,. \label{eq:isospin-last}
\end{align}
Here, the superscripts of the hadronic matrix elements indicate the final state, the first subscript corresponds to the isospin of the operator in the Hamiltonian, and the second subscript indicates the isospin of the final state.
Note that the contribution of the $\Delta I=3/2$ Hamiltonian to the CKM-subleading amplitude $A_b$ is negligible in the SM~\cite{Grossman:2012eb, Franco:2012ck}.
Working in the isospin limit, we have \cite{Buccella:1992sg,Grossman:2012eb}
\begin{align}
a_{CP}^{\mathrm{dir}}(D^0\rightarrow \pi^+\pi^0) &= 0\,, \label{eq:pipi-isospin-1}
\end{align}
and the sum rule~\cite{Kwong:1993ri, Grossman:2012eb,Atwood:2012ac}
\begin{align}
- \mathcal{A}(D^+\rightarrow \pi^+\pi^0) +
\frac{1}{\sqrt{2}} \mathcal{A}(D^0\rightarrow \pi^+\pi^-) +
\mathcal{A}(D^0\rightarrow \pi^0\pi^0) = 0\,. \label{eq:pipi-isospin-2}
\end{align}
For the $D\rightarrow KK$ system there are no isospin relations analogous to Eqs.~(\ref{eq:pipi-isospin-1}) and (\ref{eq:pipi-isospin-2}).
As we are only sensitive to relative phases, we choose $A_{\frac{3}{2},2}^{\pi\pi}$ to be real and positive.
We employ therefore as real theory parameters of $D\rightarrow \pi\pi$:
\begin{align}
A_{\frac{3}{2},2}^{\pi\pi}\,,
\vert A_{\frac{1}{2},0}^{\pi\pi}\vert\,,
\mathrm{arg}\left( A_{\frac{1}{2},0}^{\pi\pi}\right),
\vert B_{\frac{1}{2},0}^{\pi\pi}\vert\,,
\mathrm{arg}\left( B_{\frac{1}{2},0}^{\pi\pi}\right).
\end{align}
Regarding the $D\rightarrow KK$ decays, we choose $A_{\frac{3}{2},1}^{KK}$ to be real and positive and employ the
real theory parameters
\begin{align}
A_{\frac{3}{2},1}^{KK},\,
\vert A_{\frac{1}{2},1}^{KK}\vert,\,
\vert A_{\frac{1}{2},0}^{KK}\vert,\,
\mathrm{arg} \left( A_{\frac{1}{2},1}^{KK}\right),\,
\mathrm{arg} \left( A_{\frac{1}{2},0}^{KK}\right),\, \nonumber\\
\vert B_{\frac{1}{2},1}^{KK}\vert,\,
\vert B_{\frac{1}{2},0}^{KK}\vert,
\mathrm{arg} \left( B_{\frac{1}{2},1}^{KK}\right),\,
\mathrm{arg} \left( B_{\frac{1}{2},0}^{KK}\right).
\end{align}
In our numerical analysis, we vary all strong phases in their complete range.
\section{Resonance Model \label{sec:resonance}}
On top of the isospin decomposition Eqs.~(\ref{eq:isospin-first})--(\ref{eq:isospin-last}), we make the following assumptions:
\begin{itemize}
\item The decay amplitudes to $I=0$ states
$A_{\frac{1}{2},0}^{\pi\pi}$, $B_{\frac{1}{2},0}^{\pi\pi}$, $A_{\frac{1}{2},0}^{KK}$, $B_{\frac{1}{2},0}^{KK}$ are dominated by scalar resonances, namely $f_0(1710)$ and/or $f_0(1790)$,
with quantum numbers $I=0$ and $J^{PC} = 0^{++}$.
\item The $I=1$ amplitudes are relatively suppressed:
\begin{align}
\left|\frac{A_{\frac{1}{2},1}^K}{A_{\frac{1}{2},0}^K}\right| &\lesssim 40\%\,, &
\left|\frac{A_{\frac{3}{2},1}^K}{A_{\frac{1}{2},0}^K}\right| &\lesssim 40\%\,, &
\left|\frac{B_{\frac{1}{2},1}^K}{B_{\frac{1}{2},0}^K}\right| &\lesssim 40\%\,, \label{eq:iso1-suppression}
\end{align}
reflecting the dominance of the scalar resonances.
\end{itemize}
In our numerical study we implement Eq.~(\ref{eq:iso1-suppression}) with a Gaussian error.
With more data, in the future the assumptions of our resonance model can be revisited and improved.
Our ansatz employs the strong effective couplings $g_{f_0PP'}$ of the resonance and the pseudoscalars, which we define through the equation
\begin{align}
\mathcal{B}(f_0\rightarrow PP') &\equiv g_{f_0 PP'}^2 \mathcal{P}(f_0\rightarrow PP')\,.
\end{align}
Furthermore, we define the hadronic matrix elements
\begin{align}
M_{f_0}^{sd} &\equiv \bra{f_0} \mathcal{O}^{\Delta I=1/2}_{sd}\ket{D^0}\,, \qquad
M_{f_0}^{b} \equiv \bra{f_0} \mathcal{O}^{\Delta I=1/2}_{b}\ket{D^0}\,, \label{eq:hadronic-MEs}
\end{align}
where $\mathcal{O}_{sd}^{\Delta I=1/2}$ is the $\Delta I=1/2$ operator contributing to $A_{sd}$ and
$\mathcal{O}_{b}^{\Delta I=1/2}$ is the $\Delta I=1/2$ operator contributing to $A_{b}$ in Eq.~(\ref{eq:general-decomposition}).
Due to the lack of precise data on the scalar resonances, including the hadronic matrix elements
Eq.~(\ref{eq:hadronic-MEs}), we consider three simplified scenarios:
\begin{enumerate}
\item[(1)] Dominance of $f_0(1710)$ only.
\item[(2)] Dominance of $f_0(1790)$ only.
\item[(3)] Contribution of both $f_0(1710)$ and $f_0(1790)$, but with the same hadronic matrix elements for both scalar $f_0$ resonances $M_{f_0}^{sd}$ and $M_{f_0}^{b}$.
\end{enumerate}
Our implementation of these scenarios does not depend on the details of the functional shape of the resonance
amplitude $R( m_{f_0}, \Gamma_{f_0}, m_{D^0})$. Specific examples can be found in
Refs.~\cite{Choi:1998yx,Fajfer:1999hh, Cheng:2002wu, Li:2019hho, Soni:2019xko, Li:2012cfa, Gronau:1999zt, Colangelo:2010bg, Ropertz:2018stk,Kou:2018nap}.
For scenario (1) and (2) our simplified resonance model ansatz reads:
\begin{align}
A_{\frac{1}{2},0}^{\pi\pi} &= g_{f_0\pi\pi} M_{f_0}^{sd} R( m_{f_0}, \Gamma_{f_0}, m_{D^0})\,, \label{eq:res-model-1} \\
B_{\frac{1}{2},0}^{\pi\pi} &= g_{f_0\pi\pi} M_{f_0}^{b} R( m_{f_0}, \Gamma_{f_0}, m_{D^0})\,, \label{eq:res-model-2} \\
A_{\frac{1}{2},0}^{KK} &= g_{f_0KK} M_{f_0}^{sd}R( m_{f_0}, \Gamma_{f_0}, m_{D^0})\,, \label{eq:res-model-3}\\
B_{\frac{1}{2},0}^{KK} &= g_{f_0KK} M_{f_0}^{b}R( m_{f_0}, \Gamma_{f_0}, m_{D^0})\,. \label{eq:res-model-4}
\end{align}
In Eqs.~(\ref{eq:res-model-1})--(\ref{eq:res-model-4}), $f_0$ is either $f_0(1710)$ or $f_0(1790)$.
In scenario (3), we have
\begin{align}
A_{\frac{1}{2},0}^{\pi\pi} &= M_{f_0}^{sd} \, R( g_{f_0(1710)\pi\pi}, g_{f_0(1790)\pi\pi}, m_{f_0(1710)}, \Gamma_{f_0(1710)}, m_{f_0(1790)}, \Gamma_{f_0(1790)}, m_{D^0} )\,, \label{eq:res-model-5} \\
B_{\frac{1}{2},0}^{\pi\pi} &= M_{f_0}^{b} \, R( g_{f_0(1710)\pi\pi}, g_{f_0(1790)\pi\pi}, m_{f_0(1710)}, \Gamma_{f_0(1710)}, m_{f_0(1790)}, \Gamma_{f_0(1790)}, m_{D^0} ) \,, \label{eq:res-model-6} \\
A_{\frac{1}{2},0}^{KK} &= M_{f_0}^{sd}\, R( g_{f_0(1710)KK}, g_{f_0(1790)KK}, m_{f_0(1710)}, \Gamma_{f_0(1710)}, m_{f_0(1790)}, \Gamma_{f_0(1790)}, m_{D^0} ) \,, \label{eq:res-model-7}\\
B_{\frac{1}{2},0}^{KK} &= M_{f_0}^{b}\, R( g_{f_0(1710)KK}, g_{f_0(1790)KK}, m_{f_0(1710)}, \Gamma_{f_0(1710)}, m_{f_0(1790)}, \Gamma_{f_0(1790)}, m_{D^0} )\,, \label{eq:res-model-8}
\end{align}
where the $R$-function denotes now the unknown interference of the $f_0(1710)$ and $f_0(1790)$ resonances including
possible mixing effects. The data in Table~\ref{tab:resonances} indicates that the resonances $f_0(1710)$ and $f_0(1790)$ are
overlapping and therefore, in general, mixing between the scalar resonances takes place.
In all three simplified scenarios, we express four isospin amplitudes by two hadronic matrix elements, which are
defined in Eq.~(\ref{eq:hadronic-MEs}). It follows
\begin{align}
\frac{ A_{\frac{1}{2},0}^{\pi\pi} }{ A_{\frac{1}{2},0}^{KK} } &=
\frac{ B_{\frac{1}{2},0}^{\pi\pi} }{ B_{\frac{1}{2},0}^{KK} } \equiv r_{f_0}\,, \label{eq:amplitude-ratio}
\end{align}
where $r_{f_0}$ does not depend anymore on $M_{f_0}^{sd}$ or $M_{f_0}^b$.
In scenarios (1) and (2), we have
\begin{align}
\vert r_{f_0}\vert &= \left|\frac{g_{f_0\pi\pi} }{ g_{f_0KK} }\right| = \sqrt{
\frac{\mathcal{B}( f_0 \rightarrow \pi\pi )}{\mathcal{B}( f_0\rightarrow KK )}
\frac{\mathcal{P}( f_0 \rightarrow KK )}{\mathcal{P}( f_0\rightarrow \pi\pi )}
}\,. \label{eq:resonance-ratio}
\end{align}
Only in scenario (1) we have experimental information on $\vert r_{f_0}\vert$, namely~\cite{Zyla:2020zbs}
\begin{align}
\frac{\mathcal{B}(f_0(1710)\rightarrow \pi\pi)}{\mathcal{B}(f_0(1710)\rightarrow KK)} &= 0.23\pm 0.05\,. \label{eq:ratio-1710}
\end{align}
Absolute branching ratios of $f_0(1710)$ or $f_0(1790)$ are not quoted by the PDG~\cite{Zyla:2020zbs}.
Clearly, any further experimental information would improve our results considerably.
Due to lack of experimental information, in scenarios (2) and (3) we leave $\vert r_{f_0}\vert$ unconstrained and extract the corresponding value needed in order to explain the charm decay data from our fit. In scenario (2) this leads to a prediction for the ratios of branching ratios to pions and kaons, analogous to Eq.~(\ref{eq:ratio-1710}), see Sec.~\ref{sec:predictions} below. In scenario~(3), the result can be interpreted with more information on the scalar resonances in the future.
In all three considered scenarios, from Eq.~(\ref{eq:amplitude-ratio}) it follows furthermore the following relation between the relevant strong phases:
\begin{align}
\mathrm{arg}(A_{\frac{1}{2},0}^{\pi\pi}) - \mathrm{arg}(B_{\frac{1}{2},0}^{\pi\pi})
= \mathrm{arg}(A_{\frac{1}{2},0}^{KK}) - \mathrm{arg}(B_{\frac{1}{2},0}^{KK}) \,. \label{eq:resonance-phase}
\end{align}
Consequently, Eq.~(\ref{eq:amplitude-ratio}) eliminates altogether two absolute values and one phase. It
introduces a correlation between the $D\rightarrow \pi\pi$ and $D\rightarrow KK$ systems of decays which is absent in the plain isospin parametrization.
We stress that for the implementation of the resonance models we employ
Eqs.~(\ref{eq:amplitude-ratio}, \ref{eq:resonance-phase}) only, i.e. we do not use a specific resonance shape.
However, we emphasize also that nevertheless we are considering simplified scenarios only,
which are most probably a rough approximation
of physical reality. The reason for considering the simplified scenarios is that, as long as the
hadronic matrix elements $M_{f_0}^{sd}$ and $M_{f_0}^{b}$ are unknown, a complete general two-resonance model
would not lead to a reduction of complexity. The four $I=0$ amplitudes would then be merely replaced by four unknown hadronic matrix elements of the resonances and the $D^0$ meson, amounting to a fit that is equivalent to the plain isospin fit.
Since the mass region near the charm meson mass is known from lattice studies to be very rich in gluonic content~\cite{MP99} through non-perturbative
effects we should expect this to become a source for large breaking of SU(3)$_F$ or for that matter $U$-spin~\cite{CS84}.
Thus this should help us understand the large breaking of SU(3)$_F$ seen in experiments since a long time ago; for example in the ratio of branching ratios for $D^0 \to K^+K^-$ versus $D^0 \to \pi^+ \pi^-$~\cite{Zyla:2020zbs}. Moreover, this characteristic of large SU(3)$_F$ breaking is also displayed in the decays of the two scalar resonances under study in this work.
\section{Numerical Results \label{sec:predictions}}
\subsection{Input Data}
The employed numerical input is summarized in Tables~\ref{tab:resonances} and \ref{tab:numericalinput}.
We calculate the direct CP asymmetries from the time-integrated ones according to Eq.~(\ref{eq:indirect}),
using the HFLAV average of indirect charm CP violation~\cite{Amhis:2019ckw, BaBar:2012bho, CDF:2014wyb, LHCb:2015xyd, Belle:2015etc, LHCb:2017ejh, Aaij:2019kcg, BaBar:2007tfw, CDF:2012ous, LHCb:2014kcb, LHCb:2016csn, LHCb:2019hro}
\begin{align}
a_{CP}^{\mathrm{ind}} &= -0.00010\pm 0.00012\,. \label{eq:indirect-charm-CPV}
\end{align}
The current world average of time-integrated CP violation in $D^0\rightarrow K_SK_S$ is given as
$A_{CP}(D^0\rightarrow K_SK_S) = -0.019\pm 0.010$~\cite{LHCb:2021rdn, Dash:2017heu, LHCb:2015ope, CLEO:2000opx}.
Compared to that we neglect charm and kaon indirect CP violation and
use the approximation
$A_{CP}(D^0\rightarrow K_SK_S) \approx a_{CP}^{\mathrm{dir}}(D^0\rightarrow K_SK_S)$.
The branching ratio input is taken from Ref.~\cite{Zyla:2020zbs}.
We do not take into account experimental correlations between CP asymmetries or branching ratios.
For example, once the LHCb~Run~1 measurement of $A_{CP}(D^0\rightarrow K^+K^-)$ provided in Ref.~\cite{LHCb:2016nxk} is updated in the future, it is desirable to include the corresponding correlation with the LHCb Run~1+2 result for $\Delta A_{CP}$~\cite{Aaij:2019kcg}.
\begin{table}[t]
\begin{center}
\begin{tabular}{c|c|c|c|c}
\hline \hline
Resonance & $I^G(J^{PC})$ & mass $m$ [MeV] & $\Gamma$ [MeV] & Ref. \\\hline
$f_0(1710)$ & $0^+(0^{++})$ & $1704\pm 12$ & $123\pm 18$ & \cite{Zyla:2020zbs} \\
$f_0(1790)$ & $0^+(0^{++})$ & $1790^{+40}_{-30}$ & $270^{+60}_{-30}$ & \cite{Aaij:2014emv, Ablikim:2004wn} \\\hline\hline
\end{tabular}
\caption{Employed experimental data for scalar unflavored resonances close to the $D^0$ mass.
\label{tab:resonances}}
\end{center}
\end{table}
\begin{table}[t]
\begin{center}
\begin{tabular}{c|c|c}
\hline \hline
Observable & Input & Ref. \\\hline
$\Delta a_{CP}^{\mathrm{dir}}$ & $-0.00161\pm 0.00028$ & \cite{Amhis:2019ckw, Aaij:2019kcg, BaBar:2007tfw, CDF:2012ous, LHCb:2014kcb, LHCb:2016csn, LHCb:2019hro} \\
$a_{CP}^{\mathrm{dir}}(D^0\rightarrow \pi^+\pi^-)$ & $ 0.005\pm 0.005$ & \cite{Belle:2008ddg, CLEO:2001lgl, FOCUS:2000ejh, E791:1997txw} \\
$a_{CP}^{\mathrm{dir}}(D^0\rightarrow \pi^0\pi^0)$ & $-0.0002 \pm 0.006$ & \cite{Belle:2014evd, CLEO:2000opx} \\
$a_{CP}^{\mathrm{dir}}(D^0\rightarrow K^+K^-)$ & $0.000 \pm 0.001$ & \cite{LHCb:2016nxk, Belle:2008ddg, CLEO:2001lgl, FOCUS:2000ejh, E791:1997txw} \\
$a_{CP}^{\mathrm{dir}}(D^+\rightarrow K^+K_S)$ & $-0.0001 \pm 0.0007$ & \cite{LHCb:2019dwr, BaBar:2012wep, Belle:2012ygx} \\
$a_{CP}^{\mathrm{dir}}(D^0\rightarrow K_SK_S)$ & $-0.019 \pm 0.010$ & \cite{LHCb:2021rdn, Dash:2017heu, LHCb:2015ope, CLEO:2000opx} \\\hline\hline
\end{tabular}
\caption{Numerical input for direct CP asymmetries from experiment.
In order to avoid double counting of data, we only use input for the single CP asymmetries
$a_{CP}^{\mathrm{dir}}(D^0\rightarrow \pi^+\pi^-)$ and $a_{CP}^{\mathrm{dir}}(D^0\rightarrow K^+K^-)$
which are not included in the HFLAV average for $\Delta a_{CP}^{\mathrm{dir}}$, or where the corresponding
correlation is small, namely for the LHCb Run~1 measurement of $A_{CP}(D^0\rightarrow K^+K^-)$ given in Ref.~\cite{LHCb:2016nxk}.
Note that for this reason we do not include the result of Ref.~\cite{LHCb:2016nxk} for $A_{CP}(D^0\rightarrow \pi^+\pi^-)$.
In these cases we calculate our own corresponding averages neglecting any correlations.
\label{tab:numericalinput}}
\end{center}
\end{table}
\subsection{Fit results}
\begin{table}[t]
\begin{center}
\begin{tabular}{c|c|c|c|c}
\hline \hline
Scenario &
Involved resonances &
dof &
$\Delta \chi^2$ &
Rejection/compatibility \\\hline
(1) & only $f_0(1710)$ & 3 & $\Delta \chi^2_{\mathrm{min}} = 62.5$ & $7.4\sigma$ \\
(2) & only $f_0(1790)$ & 2 & $\Delta \chi^2_{\mathrm{min}} = 3.6$ & $1.4\sigma$ \\
(3) & $f_0(1710)$ and $f_0(1790)$ & 2 & $\Delta \chi^2_{\mathrm{min}} = 3.6$ & $1.4\sigma$ \\\hline\hline
\end{tabular}
\caption{Results of the likelihood ratio tests of the simplified scenarios of the resonance model compared to the null hypothesis, which is the plain isospin fit. Due to the lack of $f_0(1790)$ branching ratio data, scenarios (2) and (3) are equivalent and give identical results. Both scenarios (2) and (3) are compatible with the data at $1.4\sigma$, whereas scenario (1) is rejected at $7.4\sigma$.
\label{tab:scenario-tests}}
\end{center}
\end{table}
\begin{table}[t]
\begin{center}
\begin{tabular}{c|c|c}
\hline \hline
Observable &
Isospin &
Scalar resonance model \\\hline
$\Delta a_{CP}^{\mathrm{dir}}$
& $ -0.0016\pm 0.0003$
& $ -0.0016\pm 0.0003$
\\
$a_{CP}^{\mathrm{dir}}(D^0\rightarrow \pi^+\pi^-)$
& $ 0.002 \pm 0.001$
& $ 0.002 \pm 0.001$
\\
$a_{CP}^{\mathrm{dir}}(D^0\rightarrow K^+K^-)$
& $ 0.000\pm 0.001$
& $ 0.0002^{+0.0009}_{-0.0010}$\\
$a_{CP}^{\mathrm{dir}}(D^0\rightarrow \pi^0\pi^0)$
& $ 0.000\pm 0.006$
& $-0.005\pm 0.003$
\\
$a_{CP}^{\mathrm{dir}}(D^+\rightarrow K^+K_S)$
& $ -0.0001\pm 0.0007$
& $ -0.0001\pm 0.0007$
\\
$a_{CP}^{\mathrm{dir}}(D^0\rightarrow K_SK_S)$
& $ -0.019\pm 0.010$
& $ -0.015^{+0.009}_{-0.010}$
\\\hline\hline
\end{tabular}
\caption{Theory results in the plain isospin fit as well as in the resonance model, for the simplified scenarios (2) and (3), which give identical results. Scenario~(1) is excluded by experimental data and therefore not shown.
\label{tab:numericaloutput}}
\end{center}
\end{table}
\begin{figure}[t]
\begin{center}
\includegraphics[width=\textwidth]{plot.pdf}
\caption{Comparison of input data (blue), see Table~\ref{tab:numericalinput}, plain isospin fit (green) and the resonance model (red) in case of the simplified scenarios (2) and (3), which give identical results. Shown are the central values and respective $1\sigma$ error bars.
The simplified scenario (1) is excluded and therefore not shown, because in this scenario Eqs.~(\ref{eq:amplitude-ratio})--(\ref{eq:ratio-1710}) are incompatible with the charm decay data.}
\label{fig:plot}
\end{center}
\end{figure}
For the model-independent plain isospin fit of Eqs.~(\ref{eq:isospin-first})--(\ref{eq:isospin-last}) we obtain the global minimum $\chi^2_{\mathrm{min}} = 0.4$.
We extract in this fit scenario the charm $\Delta I=1/2$ rule, i.e.~the ratio of
$\Delta I=1/2$ over $\Delta I=3/2$ matrix elements, as
\begin{align}
\left| \frac{
A_{\frac{1}{2},0}^{\pi\pi}
}{
A_{\frac{3}{2},2}^{\pi\pi}
} \right| = 2.44\pm 0.04\,, \label{eq:DeltaI12-rule}
\end{align}
in agreement with extractions in Refs.~\cite{Franco:2012ck, CLEO:2005mti}.
From the perspective of the resonance model, Eq.~(\ref{eq:DeltaI12-rule}) reflects the
dominance of charm decays by scalar resonances close to the $D^0$ mass.
The scalar resonance models are implemented by imposing Eqs.~(\ref{eq:amplitude-ratio}, \ref{eq:resonance-phase}) on the isospin fit. Thereby, the resonance models implicitly include an assumption about the mechanism of SU(3)$_F$ breaking.
The relative degrees of freedom of the resonance models compared to the plain isospin fit, \emph{i.e.},~the number of parameters fixed through Eqs.~(\ref{eq:amplitude-ratio}, \ref{eq:resonance-phase}), is $\nu=3$ in scenario (1), and $\nu=2$ for scenarios (2) and (3), as in these scenarios the ratio $r_{f_0}$ is left unconstrained.
The results of the likelihood ratio tests of the three scenarios compared to the null hypothesis, which is the
plain isospin fit, are shown in Table~\ref{tab:scenario-tests}.
The model based solely on the $f_0(1710)$ is excluded at $7.4\sigma$.
The scenarios including only the $f_0(1790)$ and both $f_0(1710)$ and $f_0(1790)$, are currently both a viable
description of the data.
Note however, that this is mainly because in case of the $f_0(1710)$ we have a measurement of the ratio of branching ratios
Eq.~(\ref{eq:ratio-1710}), which is not the case for the $f_0(1790)$ resonance.
In scenario (3), i.e.~when combining $f_0(1790)$ and $f_0(1710)$, we obtain
\begin{align}
\vert r_{f_0}\vert &= 0.74^{+0.06}_{-0.02}\,, \label{eq:rf0-sc3}
\end{align}
with $r_{f_0}$ defined in Eq.~(\ref{eq:amplitude-ratio}).
Note that the limit $\vert r_{f_0}\vert = 1$ corresponds to no SU(3)$_F$-breaking between the $I=0$ amplitudes.
Therefore, Eq.~(\ref{eq:rf0-sc3}) implies SU(3)$_F$-breaking of $\sim 30\%$, in agreement with
model-independent studies~\cite{Hiller:2012xm, Muller:2015lua}.
The further interpretation of Eq.~(\ref{eq:rf0-sc3}) depends on the details of the considered resonance model.
In scenario (2), i.e.~the model based on the $f_0(1790)$ resonance only, we can predict the ratio of branching ratios to pions and kaons needed in order to explain the charm decay data. We obtain
\begin{align}
\frac{\mathcal{B}(f_0(1790)\rightarrow \pi\pi)}{\mathcal{B}(f_0(1790)\rightarrow KK)} &= 0.65^{+0.12}_{-0.04}\,. \label{eq:ratio-prediction}
\end{align}
More branching ratio data for the $f_0(1790)$ resonance could in principle easily challenge scenario (2).
In the literature, we find only the qualitative observation, that $f_0(1790)\rightarrow KK$ decays are relatively suppressed compared to $f_0(1790)\rightarrow \pi\pi$~\cite{Ablikim:2004wn}. If confirmed, this could possibly lead to the exclusion of scenario (2) and make more complicated scenarios necessary, such as scenario (3).
We encourage dedicated measurements of the branching ratios of the $f_0(1790)$ resonance in order to test our model.
In Table~\ref{tab:numericaloutput} and Fig.~\ref{fig:plot} we show our theory results for the CP asymmetries
obtained from the isospin fit and the resonance model in scenarios (2) and (3), compared to the input data.
The isospin fit basically just reproduces the data.
As in both scenarios (2) and (3) the ratio $r_{f_0}$ is unknown, these two scenarios give currently identical results.
We obtain nontrivial theory results especially
in case of the decays $D^0\rightarrow K_SK_S$ and $D^0\rightarrow \pi^0\pi^0$. Notably we predict
the sign of $a_{CP}^{\mathrm{dir}}(D^0\rightarrow \pi^0\pi^0)$ to be negative.
Furthermore, the resonance model seems to indicate that out of the final states of $D$ decays (see Table~\ref{tab:numericaloutput}) that we are considering, $a_{CP}^{\mathrm{dir}}(D^0\rightarrow K_SK_S)$ is the largest CP asymmetry, in agreement with the tendency seen in model-independent predictions~\cite{Atwood:2012ac,Nierste:2015zra}, followed by $a_{CP}^{\mathrm{dir}}(D^0\rightarrow \pi^0\pi^0)$.
Note that due to the presence of resonances, in our model deviations from the $U$-spin symmetry are possible.
Due to the large uncertainties regarding the information about the scalar resonances, the theoretical results are currently very broad.
Still, with more precise CP asymmetry measurements, as well as with branching ratio measurements of the scalar resonances, in the future it will become feasible to test the model directly from data, and to distinguish it from other approaches.
\section{Conclusions \label{sec:conclusions}}
Modelling the $KK$--$\pi\pi$ rescattering in charm decays with scalar resonances close to the neutral charm meson mass,
we can reasonably well account for the recently measured charm CP violation within the Standard Model.
Thus, from the perspective of this resonance model there does not seem to be a compelling need for new physics at this time.
Our theory results for further charm CP asymmetries make it possible to test our proposed dynamical mechanism with
future measurements. More experimental information on the scalar $f_0$ resonances will
improve the theoretical results for charm CP violation.
In particular, it is extremely important to measure their branching ratios to $KK$ and $\pi\pi$ modes as precisely as possible.
Lattice studies in the charm region should also help.
We strongly encourage experimentalists to measure all CP asymmetries of singly-Cabibbo suppressed charm decays in order to test
the different theoretical scenarios being proposed and to search for new physics.
\begin{acknowledgments}
We thank Avital Dery, Marco Gersabeck, Tim Gershon, Yuval Grossman, Alexander Lenz, Michael Morello, Ulrich Nierste, Tommaso Pajero, Aleksey Rusov and Sheldon Stone
for discussions. We also want to thank Sheldon Stone for pointing out the resonance $f_0(1790)$.
S.S. is supported by a Stephen Hawking Fellowship from UKRI under reference EP/T01623X/1 and the Lancaster-Manchester-Sheffield Consortium for Fundamental Physics, under STFC research grant ST/T001038/1.
The work of A.S.~is supported in part by the US DOE Contract No. DE-SC 0012704.
\end{acknowledgments}
|
\section{Introduction}
\label{sec:introduction}
The fast development of Deep neural networks (DNNs) has promoted many artificial intelligence fields ranging from image classification\cite{he2016deep} to natural language translation\cite{zhang2019bridging} and self-driving cars\cite{prakash2021multi}.
However, lots of studies~\cite{szegedy2014intriguing,goodfellow2014explaining,carlini2017towards,lin2020nesterov} demonstrated that by adding human imperceptible adversarial perturbations to clean input data, even well-trained DNNs can be fooled with high probabilities.
This phenomenon indicates that DNNs are vulnerable to adversarial attacks and reveals their robustness issue.
To evaluate the robustness of trained DNNs, lots of adversarial attack generation methods were proposed in recent years.
According to how many victim model information the attacker can access, existing adversarial attack methods can mainly be categorized into the following three main stream settings: white-box~\cite{goodfellow2014explaining,moosavi2016deepfool}, query-based black-box~\cite{brendel2018decision-based,ilyas2018black,cheng2018query}, and query-free black-box attack~\cite{liu2016delving,papernot2017practical}, which can be summarized in Table \ref{tab:problem}.
In the white-box attack setting, the attacker can access all information of the victim model to generate adversarial attacks, which is often idealistic and impractical.
Query-based black-box attack assumes that except for querying the output or prediction of the victim model, all the other information of the victim model is hidden from the attacker.
Among the three settings, query-free black-box attack is the most challenging one because as Table \ref{tab:problem} shows, even querying the victim model is also forbidden to the attacker.
\begin{table*}[]
\centering
\footnotesize
\caption{The information that the attack methods can access.
}
\begin{tabular}{cc}
\toprule
Attack Setting & Information of the target model that can be accessed \\
\midrule
White-Box & \tabincell{c}{All information (network architecture, network weight, gradient, score, \\ prediction, input-resolution, output-dimension, label-set, \emph{etc.})} \\
\midrule
Query-based Black-Box & \tabincell{c}{ Limited information (prediction, score, input-resolution, label-set) } \\
\midrule
Query-free Black-Box & Limited information (input-resolution, label-set) \\
\midrule
Generalized Transferable Attack & Non information (/) \\
\bottomrule
\end{tabular}
\label{tab:problem}
\end{table*}
\begin{figure*}[h]
\centering
\includegraphics[width=0.95\textwidth]{figures/fig1_FFT.pdf}
\caption{
For each of the four datasets Cifar-10, Cifar-100, Tiered$_{V56}$, and Tiered$_{T84}$, the first raw shows the sampled clean image, the corresponding adversarial image and ICE noise.
The second raw: the three channels of the ICE noise.
The third raw: the spectrum diagram of each noise channel.
Note that all the original noises are constrained into the range of $-\epsilon$ to $\epsilon$, and each spectrum diagram are calculated on the original noise span, but for visualization, we normalize the visualized noises into the range of 0 to 255.
Resolutions of Cifar-10, Cifar-100, Tiered$_{V56}$, and Tiered$_{T84}$ are 32$\times$32, 32$\times$32, 56$\times$56, and 84$\times$84, respectively.
We resize all resulting adversarial images and ICE noises into a consistent larger resolution for clear visualization.
Tiered$_{V56}$ and Tiered$_{T84}$ are two sub-datasets of TieredImageNet and contain no overlapping image categories.
Section~\ref{sec:experiment} will detail Tiered$_{V56}$ and Tiered$_{T84}$.
}
\label{fig:fig1}
\end{figure*}
Unfortunately, the victim model is still vulnerable due to the demonstrated transferability of adversarial examples ~\cite{wu2018understanding,wu2020boosting,naseer2019cross-domain,demontis2019why}.
The transferability-based adversarial attack methods (shorten as transfer attack in the rest of the paper) usually leverage one or a few surrogate white-box models to construct adversarial examples and transfer the obtained adversarial examples to attack the victim model without querying.
Though previous works have demonstrated the effectiveness of transfer attacks~\cite{papernot2017practical,guo2020backpropagating} and methods have been developed to improve the transferability of adversarial examples~\cite{li2020learning,wang2021enhancing}, they usually implicitly assume that {\bf surrogate white-box models and the victim model are trained on the same dataset or are using the same label set}, and their experimental results commonly depend on this assumption.
This means the attacker actually knows some information (e.g., {\bf input resolution} and {\bf label set}) of the victim model under the query-free black-box attack setting.
For example, when the victim model and the victim images are from Cifar-10~\cite{krizhevsky2009learning}, they assume that the surrogate models are also trained on Cifar-10 rather than ImageNet~\cite{deng2009imagenet}.
Despite the three main-stream attack settings, the authors of \cite{naseer2019cross-domain} proposed a more challenging cross-domain attack setting where the surrogate models and the victim model are trained on different datasets, but they still implicitly assumed that the image resolutions of victim models and the victim model are consistent.
In practice, however, the victim image (and victim model) can randomly come from any dataset with arbitrary label and resolution, and the attacker won't know which dataset is being used in advance. It is also impractical to retrain new surrogate models for each new dataset.
Therefore, we need to assume the surrogate models and the victim model are trained on different datasets with different label sets and different image resolutions.
We denote this more practical setting as \textbf{g}eneralized \textbf{t}ransferable \textbf{a}ttack (\textbf{GTA}) because the attacker needs to be generalized to attack \textbf{images from unknown datasets} and attack \textbf{any unknown models} predicting these images. Table~\ref{tab:problem} illustrates the difference between GTA and the previous attack settings.
In this paper, we aim to investigate the robustness of DNNs under the GTA setting, and explore what kind of attacks can break DNNs under this setting.
Although most of the existing transfer attacks are not designed for attacking across datasets, with some careful modifications on the attack loss and rescaling techniques, it is possible to extend existing attacks~\cite{dong2019evading,xie2019improving,dong2018boosting}
to this new setting (we will discuss these modifications in Sections \ref{sec:evaluate of baselines} and Appendix). Unfortunately, as will be seen in the experimental results, even with these modifications the existing transfer attacks suffer from poor attack success rates due to the mis-match of label set and input size between source and target models.
To tackle these challenges, we sequentially propose two novel methods where the second one is highly inspired by the phenomenons exhibited by the first one.
The \textbf{first} method is inspired by meta-learning\cite{finn2017model,qin2020layer,SNAIL}.
It uses resources (\emph{i.e.}, some source datasets containing various label sets and input sizes, some white-box models trained on these source data) to meta-train~\cite{finn2017model,SNAIL,meta-teacher,liu2019metapruning} a universal surrogate model for the objective that classification information of each image from any category with any resolution can be erased by attacking the universal surrogate model.
After meta-training the universal surrogate model, we use it to erase classification information of new images that are outside the label space and image resolution space of the source datasets.
We call the first method Image Classification Eraser (\textbf{ICE}) and will describe the detail in Section~\ref{sec:method-ICE}.
The experimental results in Section~\ref{sec:experiment} will demonstrate the effectiveness of the proposed ICE under the GTA setting.
Despite the performance, here we want to emphasize an interesting observation shown by ICE.
To save space, we only visualize a few ICE noise examples in Fig.\ref{fig:fig1} and will visualize the noises generated by the other methods in Section \ref{sec:visualization}.
Fig.~\ref{fig:fig1} clearly shows that \textbf{ICE uses similar texture-like noise to perturb different images, even when the images are sampled from different datasets with different resolutions.}
Moreover, Section~\ref{sec:transfer noises}
will demonstrate that the ICE noise generated for each image can be directly used to attack other images from the same or different datasets.
In other words, ICE can be viewed as a universal noise generator.
Compared with existing universal adversarial attacks\cite{moosavi2017universal,zhang2021data} who commonly focus on constructing the adversarial noise that is universal only within a single dataset, ICE is more generalizable because the generated ICE noise is universal not only within a single dataset but also across datasets and resolutions.
As Fig.\ref{fig:fig1} shows, we also analyze noises generated by ICE via fast fourier transformation (FFT) and find that the main components in each ICE noise are three sine waves for the input R, G, and B channels.
Each sine wave is represented by the pair of the brightest points in each spectrum diagram.
Inspired by this interesting insight, we design the \textbf{second} method Sine Attack (\textbf{SA}) which directly optimizes the three sine waves for the R, G, and B channels.
Fig.\ref{fig:fig2} shows the optimized sine signals when we using Cifar-10, Cifar-100, and Tiered$_{V56}$ as source datasets.
The experiments in Section \ref{sec:experiment} demonstrate that SA performs comparably to ICE under the GTA setting, indicating that \textbf{three particular sine signals are effective adversarial signals under the GTA setting}.
\begin{figure}[t]
\centering
\includegraphics[width=0.43\textwidth]{figures/fig2_optimized_sines.pdf}
\caption{
The optimized three sine waves when we use Cifar-10, Cifar-100, and Tiered$_{V56}$ as source datasets.
The left most RGB noise is constructed by the three sine wave maps for the R, G, and B channel, respectively.
The image under each sine wave map is the corresponding spectrum diagram.
}
\label{fig:fig2}
\end{figure}
The main contributions of our paper are as follows.
\begin{itemize}
\item To the best of our knowledge, we are the first to propose the challenging and practical generalized transfer attack (GTA) setting.
Under GTA, the attacker is acquired to directly attack any images without knowing the image label, the dataset that the image belongs to, and the image resolution, in advance.
\item To generate adversarial example under the GTA setting, we propose a novel meta-learning inspired method ICE.
Throughout experiments in Section~\ref{sec:experiment} demonstrate the effectiveness of the proposed ICE.
\item Despite the excellent performance of ICE, we also analyze the noises generated by ICE via both experiments and FFT, and we find that ICE can be viewed as a generator for three universal sine signals for the GTA setting.
Our finding may inspire researchers to analyze and improve the robustness of DNNs.
\item Inspired by our finding, we propose another method SA to address the challenging GTA setting.
Throughout experiments in Section~\ref{sec:experiment} demonstrate that SA
performs comparably to ICE under the GTA setting.
\item Furthermore, the experiments in Sections \ref{sec:transfer noises} \ref{sec:attacking CUB}, \ref{sec:attacking ImageNet}, and \ref{sec:attacking object detection} demonstrate that the ICE noise and SA noise are universal perturbations across images, datasets, resolutions, and tasks.
For example, Section~\ref{sec:attacking object detection} shows that the SA noise constructed by three sine waves can be directly used to effectively attack object detection models.
\end{itemize}
\section{Background}
\subsection{Adversarial attack settings}
Existing adversarial attack methods~\cite{ganeshan2019fda,croce2020minimally,wu2020stronger,li2020practical,kaidi2019structured,maksym2020square,xiao2021graph,zhang2021data} can mainly be categorized into white-box, query-based black-box, and query-free black-box attacks.
Among the three kinds of attacks, white-box attack~\cite{kurakin2016adversarial} is the most effective one because all information of the target model can be leveraged to generate adversarial examples.
Query-based black-box attack assumes that some information of the target model is hidden from users and the users can only query the target model and access the hard-label or soft-label predictions.
Researchers have proposed many query-based black-box attack methods~\cite{chen2017zoo,cheng2020signopt,huang2020black,gao2020patch} and have shown that adversarial examples can still be effectively generated only based on predictions.
The most recently developed query-based methods mainly focus on improving the querying efficiency and reducing the query counts~\cite{cheng2019improving,li2020qeba,du2020query-efficient,wang2020spanning,yuan2021meta}.
The query-free black-box attack further assumes that the target model's prediction is also hidden from users.
In this challenging situation, researchers usually generate adversarial examples via attacking surrogate models~\cite{dong2019evading,xie2019improving}.
Then, by leveraging the transferability of adversarial examples~\cite{papernot2016transferability,tramer2017ensemble,nathan2020perturbing}, we can directly use the adversarial examples to attack the target model without querying~\cite{huang2019enhancing,zhou2018transferable,Lu_2020_CVPR}.
The level of information that the attacker can leverage reduces from white-box to query-free black-box attacks.
However, the three most popularly used attack settings still need to know some information of the victim model, which is summarized in Table \ref{tab:problem}.
For example, transfer-based attacks commonly assume the surrogate model and victim model are trained on the same dataset, which indicates that the attacker implicitly knows the dataset, input size, and label set of the victim model.
Beyond the three main stream settings, a few other attack settings have also been studied in recent works~\cite{li2020practical,nathan2020perturbing,inkawhich2021can,zhang2021data,naseer2019cross-domain}.
For instance, in no-box attack~\cite{li2020practical}, the attacker can access neither a large scale training data nor pre-trained models on it.
To attack a victim image, the no-box attacker needs firstly to collect a mini-training set that contains a small number of images and secondly to train a substitute model for generating adversarial noise on the victim image.
Due to the two steps in attacking each victim image, no-box attack is costly to attack a large amount of diverse images in practice.
The authors of \cite{inkawhich2021can} studied the problem where the white-box model and victim model share no label.
They firstly query the victim model by using the white-box dataset and obtain the mapping between white-box labels and black-box labels, and then use the white-box models and the resulting mapping to attack victim images.
Though their study is interesting, their method is sometimes unrealistic since before attacking, they have to build the mapping via extensively querying the victim model.
Universal adversarial perturbation (UAP) \cite{zhang2021data} aims to obtain a universal perturbation pattern within an image distribution so that adding the perturbation pattern to any clean image sampled from the distribution can possibly produce an adversarial example.
However, as will be demonstrated in our experiment, UAP performs not well enough to attack the images sampled from unknown distributions (datasets).
The authors of \cite{naseer2019cross-domain} proposed a cross-domain attack setting where the surrogate models and the victim model are trained on different datasets.
However, they also implicitly assumed that the resolutions of victim models and the victim model are consistent, which is still unrealistic in some challenging cases.
In this paper,
we consider a novel and more challenging problem called generalized transferable attack (GTA) and will detail the GTA setting in Section~\ref{sec:GTA}.
\subsection{Adversarial perturbation's frequency characteristic}
Some works \cite{geirhos2019imagenet-trained,ilyas2019adversarial,das2018shield} tried to explain why adversarial perturbation exists from the perspective of frequency.
It is demonstrated that DNNs' predictions are simultaneously determined by image's low-frequency and high-frequency (including human imperceptible) information, especially the high-frequency\cite{yin2019fourier}.
For example, by using synthesized texture-shape cue conflict images, the work \cite{geirhos2019imagenet-trained} demonstrated that ImageNet CNN models are usually more sensitive to image's texture information than to shape information.
Texture is high-frequency information while shape is low-frequency.
Via experiments and analysis, the work \cite{yin2019fourier} demonstrated that instance dependent adversarial perturbations are high frequency signals and used Auto-Augmentation\cite{Cubuk_2019_CVPR} to improve DNNs' robustness towards high frequency perturbations.
Similar high-frequency characteristic was also found in universal adversarial perturbations by \cite{zhang2021universal}.
Since DNNs commonly are sensitive to high-frequency information, their predictions will possibly be determined by adversarial perturbation containing rich high-frequency information rather than the original clean image.
Though these previous works have demonstrated the high-frequency characteristic of adversarial perturbation, they didn't analyze how high the frequency should be, especially under the challenging GTA setting.
In this work, inspired by the observation shown in Fig.\ref{fig:fig1}, we propose a novel Sine Attack method which directly optimizes the parameters determining not only the frequency but also the sine wave direction and initial phase, for generating effective adversarial attacks under the GTA setting.
\section{Generalized transferable attack}
\label{sec:GTA}
GTA can be described as the following attack scene.
We have the resources of:
1) $m$ source image classification datasets denoted as $\mathcal{D}_1$, $\mathcal{D}_2$, ..., $\mathcal{D}_m$.
Different datasets have different label sets, with potentially different label sizes and image shapes.
2) $N_k$ trained models denoted as $\mathbf{M}_{\mathcal{D}_k}$ = \{ $\mathbf{M}_{\mathcal{D}_k}^1$, $\mathbf{M}_{\mathcal{D}_k}^2$, ..., $\mathbf{M}_{\mathcal{D}_k}^{N_k}$ \} for each source dataset $\mathcal{D}_k$.
With these resources, we can prepare an adversarial attacker.
Then, given a randomly intercepted image $x$, we are required to directly leverage the resources or the prepared attacker to disturb the image and obtain an perturbed image $\hat{x}$ so that any unknown target model $\mathbb{M}$ that predicts this image will make wrong predictions for the disturbed image, which can be formulated as $\mathbb{M}(\hat{x}) \neq \mathbb{M}(x)$.
Note that the target image is randomly intercepted so it is normal that the source datasets do not contain the image category of the target image.
Further, the resolution of the target images and target models cannot be known in advance.
\section{Image Classification Eraser}
\label{sec:method-ICE}
GTA is a great challenging setting because we cannot know any information of the victim image and the victim model in advance but are required to directly perturb the victim image.
So how can we effectively perturb any possible victim images?
Ensemble-based attack\cite{dong2018boosting,xie2019improving}, which simply ensemble all source models and perturb the victim image by attacking all the source models with gradient-based methods (\emph{i.e.}, FGSM, PGD), is a popular direction for the transfer attack setting.
However, since source models can have different input shapes and label spaces in the GTA setting, it is nontrivial to ensemble them into a single model. Furthermore, a naive ensemble may not optimize the performance for GTA.
We propose a novel meta-learning \cite{finn2017model,qin2020layer} inspired framework called image classification eraser (ICE) to perturb images under the GTA setting.
It firstly leverages the resources mentioned in Section~\ref{sec:GTA} to meta-train a single model for the objective that classification information of any images sampled from any categories with any kinds of resolutions can be erased by confusing the model.
This trained model can be understood as a universal surrogate model for victim target models trained on any datasets.
The model will be denoted as
$\mathcal{U}_\theta$ in the rest of this paper, where $\theta$ is the model parameter.
After meta-training the model $\mathcal{U}_\theta$, it is expected that by confusing $\mathcal{U}_\theta$ with gradient ascent, we can obtain an effective adversarial example $\hat{x}$ for any unknown intercepted image $x$ to fool an unknown target model $\mathbb{M}$.
To meta-train the model $\mathcal{U}_\theta$ for GTA, we need to consider the following issues:
1) We cannot predict the category of a randomly encountered image $x$ in advance, so no ground-truth information can be leveraged. This means the commonly used cross-entropy loss which needs the ground-truth label cannot be directly used in GTA.
We therefore use entropy instead of cross-entropy for the attack.
Eq.~\ref{eq:entropy} shows the formulations of cross-entropy CE$(d, y)$ and entropy $\mathcal{L}(d)$, where
$d$ is a vector that denotes the prediction distribution and $y$ is the one-hot ground-truth label.
\begin{equation}
\left\{
\begin{array}{lr}
d &= \text{Softmax}(\text{logit}) \\
\mathcal{L}(d) &= -d^T \cdot \text{log}(d) \\
\text{CE}(d, y) &= -y^T \cdot \text{log}(d)
\end{array}
\right.
\label{eq:entropy}
\end{equation}
Entropy of any distribution denotes the degree of disorder, randomness, or uncertainty of the distribution.
Therefore, by maximizing the entropy, we can obtain a perturbation that makes the input image hard to be classified without using the ground-truth label.
Furthermore, the entropy loss also enables us to flexibly set the output dimension of model $\mathcal{U}_\theta$ without knowing the number of categories of any dataset.
In the experiments, we set the output dimension of $\mathcal{U}_\theta$ to 1000 by default and investigate the impact of the output dimension on its performance in Appendix.
2)
Randomly encountered images may have diverse shapes.
For example, the image-shapes from different datasets (\emph{e.g.}, Cifar-10 and ImageNet) differ from each other.
Therefore, the model $\mathcal{U}_\theta$ must be capable of directly handling images with different shapes without resizing images.
To guarantee this point, we set the network architecture of model $\mathcal{U}_\theta$ to a fully convolutional network without flatten and fully-connected layers.
All down-sampling operations are implemented by max-pooling and average-pooling.
We will show details of the network architecture in Appendix.
\subsection{Training the universal surrogate model}
\begin{figure*}[t]
\centering
\includegraphics[width=0.95\textwidth]{figures/framework4.pdf}
\caption{
(a) The training framework of the proposed ICE.
Different images sampled from different source datasets are used to mimic randomly encountered images and are simultaneously fed into $\mathcal{U}_\theta$.
By maximizing the prediction entropy of model $\mathcal{U}_\theta$ for the input images, we obtain adversarial images.
To evaluate how confusing the generated adversarial images are, we feed them into the source models and measure the cross-entropy loss.
Finally,
we optimize $\mathcal{U}_\theta$ by maximizing the source models' cross-entropy losses.
(b) The testing pipeline of the proposed ICE.
}
\label{fig:framework}
\end{figure*}
To optimize the model $\mathcal{U}_\theta$, we use all source models and images to simulate unknown victim models and unknown images, respectively, and develop a novel bi-level training framework~\cite{finn2017model,liu2019metapruning,ren2018learning,liu2019darts}.
Each bi-level training iteration contains an inner-loop and an outer-loop optimization.
\textbf{In the inner-loop}, we perturb any simulated unknown images under the GTA setting and obtain the resulting perturbed images.
Specifically, given any source dataset $D_k$ and any image $x_{\mathcal{D}_k} \in \mathcal{D}_k$, we simulate it as a randomly encountered unknown image and feed $x_{\mathcal{D}_k}$ into $\mathcal{U}_\theta$ and obtain the prediction $\mathcal{U}_\theta(x_{\mathcal{D}_k})$.
Then we generate adversarial example $\hat{x}_{\mathcal{D}_k}$ by maximizing the entropy of $\mathcal{U}_\theta(x_{\mathcal{D}_k})$.
To enable the gradient back-propagating in the bi-level optimization framework, we maximize the model $\mathcal{U}_\theta$'s prediction entropy via one-step Customized PGD (Customized FGSM)~\cite{qin2021training}.
The reason why we conduct one-step but not multi-step attack in the inner loop will be described in Appendix.
Then, we can formulate $\hat{x}_{\mathcal{D}_k}$ as
\begin{equation}
\left\{
\begin{array}{lr}
g(\theta) = \nabla_{x_{\mathcal{D}_k}}\mathcal{L}(\mathcal{U}_\theta(x_{\mathcal{D}_k})) \\
\hat{x}_{\mathcal{D}_k} \!\! = \! \text{Clip}\Big(x_{\mathcal{D}_k} \!\!+ \epsilon_c \! \cdot\! \big (\gamma_1 \!\cdot\! \frac{g(\theta)}{\text{sum}(\text{abs}(g(\theta)))} +\! \\
\qquad \qquad \qquad \qquad \quad \ \gamma_2\! \cdot\! \frac{2}{\pi}\! \cdot\! \text{arctan}(\frac{g(\theta)}{\text{mean}(\text{abs}(g(\theta)))}) +\! \\
\qquad \qquad \qquad \qquad \quad \ \text{sign}(g(\theta)) \big) \Big ),
\end{array}
\right.
\label{eq:inner-loop}
\end{equation}
where both $\gamma_1$ and $\gamma_2$ are set to 0.01 by default.
$\epsilon_c$ determines the perturbation scale.
$\mathcal{L}(\mathcal{U}_{\theta}(x_{\mathcal{D}_k}))$ is the entropy of $\mathcal{U}_\theta(x_{\mathcal{D}_k})$, and $g(\theta)$ is the gradient of the entropy \emph{w.r.t} $x_{\mathcal{D}_k}$ based on the current parameter $\theta$.
Clip is the function that clips each pixel value of the image into the range of $[0, 255]$.
\begin{algorithm}[t]
\caption{Training of Image Classification Eraser}
\label{algorithm:ICE_train}
{\bfseries input:} Source datasets $\mathbf{D} \!=\! \{\mathcal{D}_1, \mathcal{D}_2, ..., \mathcal{D}_m\}$, Source models $\mathbf{M}_{\mathcal{D}_k} \!=\! \{ \mathbf{M}_{\mathcal{D}_k}^1, \mathbf{M}_{\mathcal{D}_k}^2, ..., \mathbf{M}_{\mathcal{D}_k}^{N_k} \}$ for each dataset $\mathcal{D}_k$. \\
{\bfseries output:} Optimized weight $\theta$. \\
{\bfseries 1 $\;\!$:} {\bfseries while not} done {\bfseries do} \\
{\bfseries 2 $\;\!$:} \quad \textbf{for} each $\mathcal{D}_k \in \mathbb{D}$ \textbf{do}\\
{\bfseries 3 $\;\!$:} \qquad Sample a mini data batch $(X_{\mathcal{D}_k}, Y_{\mathcal{D}_k}) \in \mathcal{D}_k$ \\
{\bfseries 4 $\;\!$:} \qquad Obtain adversarial examples $\hat{X}_{\mathcal{D}_k}$ via Eq.~\ref{eq:inner-loop} \\
{\bfseries 5 $\;\!$:} \quad\quad \textbf{for} each $\mathbf{M}_{\mathcal{D}_k}^j \in \mathbf{M}_{\mathcal{D}_k}$ \textbf{do} \\
{\bfseries 6 $\;\!$:} \qquad\quad Obtain adversarial loss $\mathbf{L}_{\mathcal{D}_k}^{j}$ for $\hat{X}_{\mathcal{D}_k}$ via Eq.~\ref{eq:evaluation}. \\
{\bfseries 7 $\;\!$:} \qquad \textbf{end for} \\
{\bfseries 8 $\;\!$:} \quad \textbf{end for} \\
{\bfseries 9 $\;\!$:} \quad $\theta \!=\! \theta \!+\! \alpha\!\cdot\!\nabla_{\theta} \big( \frac{1}{m}\sum_{k=1}^{m} (\frac{1}{N_k} \sum_{j=1}^{N_k} \mathbf{L}_{\mathcal{D}_k}^{j}) \big)$ \\
{\bfseries 10:} {\bfseries end while} \\
{\bfseries 11:} {\bfseries return} $\theta$
\end{algorithm}
\textbf{In the outer-Loop},
we evaluate the attack success rate of the adversarial examples generated under the GTA setting and optimize the success rate by optimizing the model $\mathcal{U}_\theta$.
Specifically, we evaluate how the perturbed image $\hat{x}_{\mathcal{D}_k}$ generated in inner-loop fools each simulated unknown target model $\mathbf{M}_{\mathcal{D}_k}^j \in \mathbf{M}_{\mathcal{D}_k}$ by calculating the adversarial loss
\begin{equation}
\!\!\!\!\!\! \bm{l}_{\mathcal{D}_k}^{j} \!\!\!\!=\! \text{CE}(\mathbf{M}_{\mathcal{D}_k}^j \!\! (\hat{x}_{\mathcal{D}_k}\!), y_{\mathcal{D}_k}\!),
\label{eq:evaluation}
\end{equation}
where $j \in [1, N_k]$;
$y_{\mathcal{D}_k}$ is the groundtruth of $x_{\mathcal{D}_k}$; CE is cross-entropy; $\mathbf{M}_{\mathcal{D}_k}^j(\hat{x}_{\mathcal{D}_k})$ is the target model's prediction for $\hat{x}_{\mathcal{D}_k}$.
A larger adversarial loss $\bm{l}_{\mathcal{D}_k}^j$ indicates a higher possibility that the simulated unknown target model $\mathbf{M}_{\mathcal{D}_k}^j$ is fooled by $\hat{x}_{\mathcal{D}_k}$.
Note that the groundtruths of all simulated unknown images are accessible when we training the model $\mathcal{U}_\theta$, so we use cross-entropy instead of entropy used in inner-loop to calculate the loss in outer-loop.
To ensure that the classification information of each image $x_{D_k}\in \mathcal{D}_k$ can be erased by attacking model $\mathcal{U}_\theta$ and to ensure that the perturbed image $\hat{x}_{D_k}$ is confusing for the simulated unknown target model $\mathbf{M}_{\mathcal{D}_k}^j$ to predict, we optimize the model $\mathcal{U}_\theta$ by maximizing the adversarial loss $\bm{l}_{\mathcal{D}_k}^j$ by the following SGD update:
\begin{equation}
\theta = \theta + \alpha \cdot
\nabla_{\theta} \bm{l}_{\mathcal{D}_k}^j,
\end{equation}
where $\alpha$ is the learning rate.
$\bm{l}_{\mathcal{D}_k}^j$ is differentiable \emph{w.r.t} $\theta$ because $\bm{l}_{\mathcal{D}_k}^j$ depends on $\hat{x}_{\mathcal{D}_k}$ and $\hat{x}_{\mathcal{D}_k}$ depends on $\theta$.
In our experiments, we optimize the model $\mathcal{U}_\theta$ by simultaneously maximizing the adversarial losses on all source models in each iteration, which is summarized in Algorithm~\ref{algorithm:ICE_train}.
This procedure will enforce $\mathcal{U}_\theta$ having the property that for clean images from unknown datasets, the corresponding adversarial examples constructed by attacking $\mathcal{U}_\theta$ are effective to fool unknown target models.
\subsection{Inference and evaluation}
\label{sec:inference of ICE}
Given any clean unknown image $x$ that will be fed into an unknown target model $\mathbb{M}$, we evaluate the proposed ICE with the following steps.
\textbf{1)} Directly feed the image $x$ into model $\mathcal{U}_\theta$ and generate the adversarial example $x^{(T)}$ by maximizing the entropy for $T$ gradient ascent steps.
The $i$-th step is formulated as
\begin{equation}
\left\{
\begin{array}{lr}
\delta^{(i-1)} = \text{sign}\big (\nabla_{x^{(i-1)}}\mathcal{L}(\mathcal{U}_\theta(x^{(i-1)}) ), \\
x^{(i)} = \text{clip}(x^{(i-1)} + \frac{\epsilon}{T} \cdot \delta^{(i-1)}),
\end{array}
\right.
\label{eq:generation of the perturbation}
\end{equation}
where $x^{(0)}=x$ and $\delta^{(i-1)}$ is the perturbation generated in the $i$-th step.
$\epsilon/T$ is the $L_\infty$ perturbation scale in each step.
\textbf{2)} Generate the adversarial example $\hat{x}$ by the formulation
\begin{equation}
\hat{x} = \text{clip}(x+\epsilon \cdot \text{sign}(x^{(T)} - x))
\label{eq:sign projection}
\end{equation}
We call this step `Sign-Projection' (SP).
The reason why we use SP is that it enlarges the average distortion of each pixel without amplifying the $L_\infty$ norm of the perturbation, which improves the GTA success rate.
Ablation study in Section~\ref{sec:adlation-SP} will show that SP improves both ICE and the baselines introduced in Section~\ref{sec:evaluate of baselines}.
\textbf{3)} Feed the adversarial example $\hat{x}$ and the clean image $x$ into the unknown target model $\mathbb{M}$ and get its predictions $\mathbb{M}(\hat{x})$ and $\mathbb{M}(x)$.
\textbf{4)} The GTA process is successful if $\mathbb{M}(\hat{x}) \neq \mathbb{M}(x)$.
The evaluation pipeline is also illustrated in Figure~\ref{fig:framework}b, where the green arrows indicate the second step.
\section{Sine Attack}
\label{sec:method-SA}
Experiments in Section~\ref{sec:experiment} will demonstrate that ICE significantly outperforms the baselines under the GTA setting.
Despite the clearly improved GTA performance achieved by ICE, another great interesting phenomenon is that compared with the other methods, ICE generates similar texture-like noises for different images, even when the images are sampled from different categories and with different resolutions, as Fig.\ref{fig:fig1} and Section~\ref{sec:visualization} show.
To further investigate ICE noise, we visualize some ICE noises in frequency domain via FFT.
Fig~\ref{fig:fig1} shows some spectrum diagram examples of ICE noise and Section~\ref{sec:visualization} will show more spectrums.
Obviously, each ICE noise contains several sine signals and the most important components are three sine waves with certain frequencies for the three R, G, B image channels, respectively.
Inspired by this interesting finding, we design another simple yet effective method called Sine Attack (SA) which uses the three sine signals to build an attacker and optimizes the attacker's fooling rate under the GTA setting by optimizing the three sine waves.
We formulate each sine wave as
\begin{equation}
Z_j=Sine(a_j\cdot X_{map}+b_j\cdot Y_{map}+c_j),
\label{eq:sa1}
\end{equation}
where $j\in$ [R, G, B].
($X_{map}$, $Y_{map}$) denote the coordinate of all pixels, as Fig.~\ref{fig:X_Y} shows.
$a_j, b_j, c_j$ are three scalars of the sine wave in $j$-channel.
They together determine the sine wave frequency, direction, and initial phase.
That is to say, SA optimizes the three sine waves by optimizing
the nine scalars: $a_R$, $b_R$, $c_R$, $a_G$, $b_G$, $c_G$, $a_B$, $b_B$, $c_B$.
For simplicity, we use $\bm{\omega}$ to represent the vector of the nine scalars in the rest of the paper and formulate the attacker constructed by the three sine waves as
\begin{equation}
\mathcal{Z}(\bm{\omega}, X_{map}, Y_{map}) = \text{Concat}(Z_R, Z_G, Z_B).
\label{eq:SA-concat}
\end{equation}
\begin{figure}[t]
\centering
\includegraphics[width=0.5\linewidth]{figures/X_Y.pdf}
\caption{
The ($X_{map}$, $Y_{map}$) coordination map with resolution $n\times m$.
Each pixel's coordination is denoted as ($x_{map}$,$y_{map}$).
}
\label{fig:X_Y}
\end{figure}
\subsection{Training of SA}
We also optimize $\bm{\omega}$ on the resources introduced in Section~\ref{sec:GTA}: source datasets $\mathbf{D} \!=\! \{\mathcal{D}_1, \mathcal{D}_2, ..., \mathcal{D}_m\}$ and source models $\mathbf{M}_{\mathcal{D}_k} \!=\! \{ \mathbf{M}_{\mathcal{D}_k}^1, \mathbf{M}_{\mathcal{D}_k}^2, ..., \mathbf{M}_{\mathcal{D}_k}^{N_k} \}$ for each source dataset $\mathcal{D}_k$.
Given any source dataset $D_k$ and any image $x_{\mathcal{D}_k} \in \mathcal{D}_k$, we simulate it as a randomly encountered image and use the following three steps to optimize $\bm{\omega}$.
\textbf{First}, we create the coordinate map ($X_{map}$, $Y_{map}$) for $x_{\mathcal{D}_k}$ and use the noise $\mathcal{Z}(\bm{\omega}, X_{map}, Y_{map})$ to conduct a GTA process on $x_{\mathcal{D}_k}$ to obtain the resulting adversarial image.
We formulate this process as
\begin{equation}
\hat{x}_{\mathcal{D}_k} \!\! = \! \text{Clip}(x_{\mathcal{D}_k} \!\!+ \epsilon \cdot \mathcal{Z}(\bm{\omega}, X_{map}, Y_{map}) ).
\label{eq:SA-train-attack}
\end{equation}
Compared with ICE, the creation of adversarial images with SA is much simpler than that in ICE because SA crafts the adversarial image by directly adding the perturbation to the victim image without calculating gradient.
\textbf{Second}, we simulate each source model $\mathbf{M}_{\mathcal{D}_k}^j \in \mathbf{M}_{\mathcal{D}_k}$ as an unknown victim model and evaluate how the perturbed image $\hat{x}_{\mathcal{D}_k}$ fools each simulated unknown victim model by calculating the adversarial loss
\begin{equation}
\!\!\!\!\!\! \bm{l}_{\mathcal{D}_k}^{j} \!\!\!\!=\! \text{CE}(\mathbf{M}_{\mathcal{D}_k}^j \!\! (\hat{x}_{\mathcal{D}_k}\!), y_{\mathcal{D}_k}\!).
\label{eq:SA-train-evaluate}
\end{equation}
\textbf{Third}, the goal of SA is that the noise $\mathcal{Z}(\bm{\omega}, X_{map}, Y_{map})$ can effectively perturb the victim image so that the victim model can be fooled.
Therefore we optimize the attacker by maximizing the above adversarial loss, which can be formulated as
\begin{equation}
\begin{aligned}
\bm{\omega} \!&=\! \bm{\omega} \!+\! \alpha \!\cdot\!\nabla_{\bm{\omega}} \bm{l}_{\mathcal{D}_k}^j \\
\!&=\! \bm{\omega} \!+\! \alpha \!\cdot\!\nabla_{\bm{\omega}} \text{CE}(\mathbf{M}_{\mathcal{D}_k}^j \!\! (\hat{x}_{\mathcal{D}_k}\!), y_{\mathcal{D}_k}\!) \\
\!&=\! \bm{\omega} \!+\! \alpha \!\cdot\!\nabla_{\bm{\omega}}\hat{x}_{\mathcal{D}_k} \!\cdot\! \nabla_{\hat{x}_{\mathcal{D}_k}}\text{CE}(\mathbf{M}_{\mathcal{D}_k}^j \!\! (\hat{x}_{\mathcal{D}_k}\!), y_{\mathcal{D}_k}\!) \\
\!&=\! \bm{\omega} \!+\! \alpha \!\cdot\!\nabla_{\bm{\omega}}(\text{Clip}(x_{\mathcal{D}_k} \!\!+ \epsilon \!\cdot\! \mathcal{Z}(\bm{\omega}, X_{map}, Y_{map})) \cdot \\
&\qquad \qquad \qquad \qquad \qquad \qquad \nabla_{\hat{x}_{\mathcal{D}_k}}\text{CE}(\mathbf{M}_{\mathcal{D}_k}^j \!\! (\hat{x}_{\mathcal{D}_k}\!), y_{\mathcal{D}_k}\!),
\end{aligned}
\label{eq:SA-train-optimize}
\end{equation}
where
$\alpha$ is the learning rate.
Eq.\ref{eq:SA-train-optimize} shows that when we optimize $\bm{\omega}$, we need to calculate the gradient $\nabla_{\bm{\omega}}(\text{Clip}(x_{\mathcal{D}_k} \!\!+ \epsilon \!\cdot\! \mathcal{Z}(\bm{\omega}, X_{map}, Y_{map}))$, so we do not use sign function to discrete each pixel of $\mathcal{Z}(\bm{\omega}, X_{map}, Y_{map})$ to -1 or 1 in Eq.\ref{eq:SA-train-attack}.
Note that in our experiments, we optimize $\bm{\omega}$ by simultaneously maximizing the adversarial losses on all source models in each iteration, which is summarized in Algorithm~\ref{algorithm:SA_train}.
\begin{algorithm}[t]
\caption{Training of Sine Attacker}
\label{algorithm:SA_train}
{\bfseries input:} Source datasets $\mathbb{D} \!=\! \{\mathcal{D}_1, \mathcal{D}_2, ..., \mathcal{D}_m\}$, Source models $\mathbf{M}_{\mathcal{D}_k} \!=\! \{ \mathbf{M}_{\mathcal{D}_k}^1, \mathbf{M}_{\mathcal{D}_k}^2, ..., \mathbf{M}_{\mathcal{D}_k}^{N_k} \}$ for each dataset $\mathcal{D}_k$. \\
{\bfseries output:} Optimized perturbation parameter $\bm{\omega}$. \\
{\bfseries 1 $\;\!$:} {\bfseries while not} done {\bfseries do} \\
{\bfseries 2 $\;\!$:} \quad \textbf{for} each $\mathcal{D}_k \in \mathbb{D}$ \textbf{do}\\
{\bfseries 3 $\;\!$:} \qquad Sample a mini data batch $(X_{\mathcal{D}_k}, Y_{\mathcal{D}_k}) \in \mathcal{D}_k$ \\
{\bfseries 4 $\;\!$:} \qquad Create ($X_{map}$, $Y_{map}$) for all image $\in$ $X_{\mathcal{D}_k}$ \\
{\bfseries 5 $\;\!$:} \qquad Obtain adversarial examples $\hat{X}_{\mathcal{D}_k}$ via Eq.~\ref{eq:SA-train-attack} \\
{\bfseries 6 $\;\!$:} \quad\quad \textbf{for} each $\mathbf{M}_{\mathcal{D}_k}^j \in \mathbf{M}_{\mathcal{D}_k}$ \textbf{do} \\
{\bfseries 7 $\;\!$:} \qquad\quad Obtain adversarial loss $\mathbf{L}_{\mathcal{D}_k}^{j}$ on $\hat{X}_{\mathcal{D}_k}$ via Eq.~\ref{eq:SA-train-evaluate}. \\
{\bfseries 8 $\;\!$:} \qquad \textbf{end for} \\
{\bfseries 9 $\;\!$:} \quad \textbf{end for} \\
{\bfseries 10:} \quad $\bm{\omega} \!=\! \bm{\omega} \!+\! \alpha\!\cdot\!\nabla_{\bm{\omega}} \big( \frac{1}{m}\sum_{k=1}^{m} (\frac{1}{N_k} \sum_{j=1}^{N_k} \mathbf{L}_{\mathcal{D}_k}^{j}) \big)$ \\
{\bfseries 11:} {\bfseries end while} \\
{\bfseries 12:} {\bfseries return} ${\bm{\omega}}$
\end{algorithm}
\subsection{Inference and evaluation}
\textbf{Note that SA is a natural universal adversarial perturbation that use the optimized three sine waves to perturb all possible images.}
Therefore, compared with ICE which generates image-dependent perturbations, SA perturbs victim images simpler in inference.
Given any clean image $x$ that will be fed into an unknown target model $\mathbb{M}$, we evaluate the proposed SA with the following steps.
\textbf{1)} Create the coordination map ($X_{map}$, $Y_{map}$) for $x$.
\textbf{2)} Use the trained $\bm{\omega}$ to generate three sine waves and construct the noise $\mathcal{Z}(\bm{\omega}, X_{map}, Y_{map})$ via Eq.\ref{eq:SA-concat}.
\textbf{3)}
Generate the adversarial example $\hat{x}$ by the formulation
\begin{equation}
\hat{x} = \text{clip}\Big(x+\epsilon \cdot \text{sign}\big(\mathcal{Z}(\bm{\omega}, X_{map}, Y_{map})\big)\Big)
\label{eq:SA_sign projection}
\end{equation}
\textbf{4)} Feed the adversarial example $\hat{x}$ and the clean image $x$ into the unknown target model $\mathbb{M}$ and get its predictions $\mathbb{M}(\hat{x})$ and $\mathbb{M}(x)$.
\textbf{5)} The GTA process is successful if $\mathbb{M}(\hat{x}) \neq \mathbb{M}(x)$.
\section{Experimental Results}
\label{sec:experiment}
In this section, we conduct several experiments to evaluate the proposed methods for conducting generalized transferable attacks.
Four datasets Cifar-10~\cite{krizhevsky2009learning}, Cifar-100, Tiered$_{V56}$, and Tiered$_{T84}$ are used to build the basic GTA testing scenes.
More challenging GTA testing scenes on other datasets (\emph{e.g.}, ImageNet~\cite{deng2009imagenet}, CUB~\cite{WahCUB_200_2011}, MS-COCO\cite{lin2014microsoft}) will be introduced in Sections~\ref{sec:attacking CUB}, \ref{sec:attacking ImageNet}, and \ref{sec:attacking object detection}.
Both Cifar-10 and Cifar-100 contain 60,000 images with the resolution of 32$\times$32.
Tiered$_{V56}$ and Tiered$_{T84}$ are two sub-datasets of TieredImageNet~\cite{ren2018meta}.
TieredImageNet is a subset sampled from ImageNet~\cite{deng2009imagenet} and contains 351, 97, and 160 image classes in training, validation, and testing, respectively.
We split TieredImageNet into Tiered$_{V56}$ and Tiered$_{T84}$.
Tiered$_{T84}$ contains all images from the training set of TieredImageNet and all the images are resized into 84$\times$84 resolution.
Tiered$_{V56}$ contains all images from the validation and test sets of TieredImageNet and all the images are resized into 56$\times$56 resolution.
There is no overlapping image category between Tiered$_{V56}$ and Tiered$_{T84}$.
For either Tiered$_{T84}$ or Tiered$_{V56}$, we
use the first 1200 images of each category to compose the training set and use the last 100 images to compose the testing set.
More details of {Tiered$_{T84}$}, and {Tiered$_{V56}$} will be shown in Appendix.
\begin{figure}[t]
\centering
\includegraphics[width=0.95\linewidth]{figures/test_baselines4.pdf}
\caption{
Inference pipeline of PGD-based baselines.
}
\label{fig:test_pipeline}
\end{figure}
\subsection{Baselines}
\label{sec:baselines}
GTA is a novel adversarial attack problem and few existing methods can be directly used as baselines.
Considering that transfer attack is a similar problem to GTA, we deploy several transferable adversarial attack methods including MI~\cite{dong2018boosting}, DI~\cite{xie2019improving}, TI-DIM~\cite{dong2019evading}, AEG~\cite{bose2020adversarial}, IR~\cite{wang2021unified}, MTA~\cite{qin2021training}, and FDA+xent\cite{nathan2020perturbing} on GTA as baselines for the proposed ICE and SA.
Despite the transfer attack methods, we also consider some other methods as baselines, such as FGSM, PGD, UAP\cite{zhang2021data}, and the cross domain adversarial attack method RAP\cite{naseer2019cross-domain}.
\subsubsection{How to use baselines for GTA?}
\label{sec:evaluate of baselines}
The baselines FGSM, PGD, MI, DI, and TI-DIM are default implemented on all our GTA experiments (including the adversarial attack to MS-COCO in Section~\ref{sec:attacking object detection}) with the inference pipeline described below.
More kinds of implementation pipelines (\emph{e.g.}, using KL-divergence, optimizing a single adversarial image) will be shown in Appendix.
The detailed implementations of AEG, IR, MTA, FDA+xent, UAP, and RAP will also be introduced in Appendix.
1) Since the source models trained on different source datasets commonly have different input shapes, we firstly resize the victim image $x$ to the input shapes of all source models and then feed the resized images to the source models, respectively.
Then the inference of $x$ on each source model $\mathbf{M}_{D_k}^j \in \mathbf{M}_{D_k}$ can be formulated as $y_{D_k}^{(0, j)} = \mathbf{M}_{D_k}^j(x_{D_k}^{(0,j)})$, where $x_{D_k}^{(0,j)} = \text{resize}(x, \text{resolution}(\mathbf{M}_{D_k}^j))$.
$\text{resolution}(\mathbf{M}_{D_k}^j)$ means the resolution of $\mathbf{M}_{D_k}^j$.
2) Because we cannot access the category of the image $x$ in advance, no ground-truth label can be leveraged to perturb the resized image.
Therefore, for each source model, we generate adversarial perturbation by maximizing the entropy (as used in our ICE) for $T$ gradient ascent steps.
The $i$-th step is
\begin{equation}
\left\{
\begin{array}{lr}
y_{D_k}^{(i-1,j)} = \mathbf{M}_{D_k}^j(x_{D_k}^{(i-1,j)}), \\
\delta_{D_k}^{(i-1,j)} = \text{sign}\big (\nabla_{x_{D_k}^{(i-1,j)}}\mathcal{L}(y_{D_k}^{(i-1,j)}) ), \\
x_{D_k}^{(i,j)} = \text{clip}(x_{D_k}^{(i-1,j)} + \frac{\epsilon}{T} \cdot \delta_{D_k}^{(i-1,j)}),
\end{array}
\right.
\label{eq:generation of the perturbation for baseline}
\end{equation}
where $y_{D_k}^{(i-1,j)}$ is the source model's output and $x_{D_k}^{(i,j)}$ and
$\delta_{D_k}^{(i-1,j)}$ are the adversarial example and the perturbation generated in the $i$-th step, respectively.
3) We resize all the adversarial examples to the image $x$'s original shape and average fuse them to a single image $x_{adv}$ following the formulation $x_{adv} = \frac{1}{m} \cdot \sum_{k=1}^{m} \cdot \big ( \frac{1}{N_k} \cdot \sum_{j=1}^{N_k}\text{resize}(x_{D_k}^{(T,j)}, \text{resolution}(x)) \big)$.
4) We generate adversarial example $\hat{x}$ by the formulation $\hat{x} = \text{clip}(x+\epsilon \cdot \text{sign}(x_{adv} - x))$, which is the SP step defined in Section~\ref{sec:inference of ICE}.
5) We feed the adversarial example $\hat{x}$ and the clean image $x$ into the unknown target model $\mathbb{M}$ and get its predictions.
6) The GTA process is successful if $\mathbb{M}(\hat{x}) \neq \mathbb{M}(x)$.
The evaluation of the PGD-based baselines is also illustrated in Fig.~\ref{fig:test_pipeline}, where green arrows denote the third and fourth steps.
\subsection{Experimental settings}
\label{sec:experimental settings}
1) {\bf Source and Target models}.
On each dataset, we train several models including VGG-16~\cite{simonyan2015very}, ResNet-18~\cite{he2016deep}, MobileNet-V3~\cite{howard2019searching}, and DenseNet-26~\cite{huang2017densely}.
The last three models will be simplified as RN-18, MB-V3, and DN-26, respectively, in the rest of our paper.
All these models will be used as the source and target models in GTA testing experiments.
The training details and the architectures of these models will be introduced in Appendix.
2) {\bf Hyper-parameters in the training phase of ICE}.
In the training phase of ICE, we train $\mathcal{U}_\theta$ for 50,000 iterations with batch size of 64 for each source dataset.
Learning rate $\alpha$ is set to 0.01.
$\epsilon_c$ is set to 3000 and is periodically decayed by 0.9$\times$ for every 3000 iterations.
3) {\bf Hyper-parameters in the training phase of SA}.
We train $\bm{\omega}$ for 10,000 iterations with batch size of 64 for each source dataset.
Learning rate $\alpha$ is set to 0.01.
4) {\bf Evaluation}.
In the inference phase, we set $T$ to 10 by default for ICE and other PGD-based baselines, and set $\epsilon$ to 15 for all methods.
Section~\ref{sec:smaller_eps} will show the experimental results with $\epsilon=8$, where ICE and SA still outperform the baselines with significant margins.
Note that we use \textbf{attack success rate} to denote the GTA performance by default.
Because attacking the images that are wrongly classified by the target model is meaningless, we only attack the images that are correctly classified by the target model.
\subsection{Generalized transferable attack to Cifar-100}
\label{sec:attack cifar100}
In this subsection, we perform GTA experiments on Cifar-100, i.e., perturbing Cifar-100 images, to fool the target victim models.
The models MB-V3, VGG-16, RN-18, and DN-26 trained on Cifar-100 are used as victims to calculate the GTA success rate.
Different settings of source models and datasets are evaluated and experimental results are reported in Table \ref{tab:attack on cifar100}.
In the first experiment (the first row of Table~\ref{tab:attack on cifar100}), we use RN-18 and Cifar-10 as the source model and the source dataset, respectively, and conduct GTA on the testing images from Cifar-100.
The proposed ICE and SA outperform all existing methods on the GTA problem.
Compared with UAP, ICE and SA promote the average attack success rate on the four target models by about 1.1\% and 16.6\%, respectively.
Moreover, SA outperforms ICE with clear margins, indicates that the optimized three sine waves are enough to perturb Cifar-100 images.
We guess the reasons why SA outperforms ICE are \textbf{1)} GTA is a much challenging setting where the attacker must be generalized enough to perturb any unknown images; \textbf{2)} the optimized three sine waves in SA are the most crucial noises to solve GTA; \textbf{3)} despite the three sine waves, ICE uses more complex noises to solve GTA which may improve ICE's performance in training phase but damage the generalization.
In other words, \textbf{compared with ICE, SA which uses only three sine waves is more generalized to perturb unknown images.}
\begin{table}[t!]
\centering
\caption{GTA success rates on Cifar-100.
}
\setlength\tabcolsep{5pt}
\scriptsize
\begin{tabular}{cccccc}
\toprule
Resource & Method & MB-V3 & VGG-16 & RN-18 & DN-26\\
\midrule
\multirow{7}{*}{ \tabincell{c}{Cifar-10 \\ (Res-18)} }
&FGSM &47.7\% &64.2\% & 59.1\% & 73.1\% \\
&PGD &37.3\% &50.9\% & 43.2\% & 61.5\%\\
&DI &39.3\% &54.9\% & 47.2\% & 64.7\%\\
&MI &45.4\% &62.3\% & 56.3\% & 72.1\%\\
&TI-DIM &51.0\% &45.6\% & 48.0\% & 55.3\%\\
&IR &48.3\% &62.1\% & 60.5\% & 73.7\%\\
&AEG &51.3\% &61.7\% & 61.2\% & 66.4\%\\
&MTA &42.3\% &49.5\% & 45.0\% & 60.0\% \\
&UAP &53.9\% &69.7\% & 71.2\% & 79.5\% \\
&RAP &50.5\% &68.8\% & 70.7\% & 81.3\% \\
&FDA+xent &37.8\% &51.5\% & 44.3\% & 62.6\% \\
\rowcolor {gray!20} \cellcolor{gray!0} &\textbf{ICE} &{54.9\%} &{67.3\%} & {71.5\%} & {83.5\%}\\
\rowcolor {gray!20} \cellcolor{gray!0} &\textbf{SA} &\textbf{78.3\%} &\textbf{77.8\%} & \textbf{79.6\%} & \textbf{84.2\%}\\
\midrule
\multirow{6}{*}{\tabincell{c}{Cifar-10 \\ + Tiered$_{T84}$ \\ (Res-18)}}
&FGSM &33.0\% &47.3\% & 38.6\% & 58.7\% \\
&PGD &39.0\% &52.2\% & 41.7\% & 64.9\% \\
&DI &39.1\% &52.3\% & 41.7\% & 64.2\% \\
&MI &42.7\% &57.0\% & 48.2\% &68.1\%\\
&TI-DIM &55.1\% &48.2\% & 50.5\% & 58.1\% \\
&IR &44.0\% &58.1\% & 51.9\% & 69.5\%\\
&AEG &50.9\% &64.6\% & 58.2\% & 70.3\%\\
&MTA &43.6\% &56.7\% & 47.5\% & 68.0\% \\
&UAP &51.1\% &73.7\% & 74.2\% & 81.5\% \\
&RAP &46.2\% &70.3\% & 67.0\% & 78.6\% \\
&FDA+xent &38.8\% &52.0\% & 42.6\% & 65.2\% \\
\rowcolor {gray!20} \cellcolor{gray!0}&\textbf{ICE} &{57.0\%} &{76.2\%} & {77.5\%} & {83.3\%} \\
\rowcolor {gray!20} \cellcolor{gray!0} &\textbf{SA} &\textbf{68.8\%} &\textbf{78.2\%} & \textbf{80.7\%} & \textbf{85.2\%}\\
\midrule
\multirow{6}{*}{\tabincell{c}{Cifar-10 \\ + Tiered$_{V56}$ \\ (RN-18)}}
&FGSM &33.7\% &48.5\% & 38.9\% &58.2\% \\
&PGD &40.1\% &54.9\% & 44.3\% & 65.7\%\\
&DI &40.5\% &55.0\% & 44.7\% & 65.9\% \\
&MI &45.5\% &60.3\% & 50.2\% & 70.6\%\\
&TI-DIM &51.4\% &48.0\% & 49.4\% & 58.9\% \\
&IR &44.7\% &59.2\% & 54.0\% & 70.1\%\\
&AEG &51.5\% &62.3\% & 58.9\% & 68.3\% \\
&MTA &43.3\% &60.8\% & 50.0\% & 70.2\% \\
&UAP &51.2\% &71.8\% & 69.7\% & 80.6\% \\
&RAP &46.5\% &69.7\% & 67.5\% & 77.2\% \\
&FDA+xent &39.6\% &52.3\% & 42.9\% & 65.6\% \\
\rowcolor {gray!20} \cellcolor{gray!0}&\textbf{ICE} &{52.7\%} &{76.9\%} & {79.8\%} & {84.9\%}\\
\rowcolor {gray!20} \cellcolor{gray!0} &\textbf{SA} &\textbf{65.9\%} &\textbf{77.3\%} & \textbf{82.6\%} & \textbf{86.0\%}\\
\midrule
\multirow{6}{*}{\tabincell{c}{Cifar-10 \\ + Tiered$_{T84}$ \\ + Tiered$_{V56}$ \\ (RN-18)}}
&FGSM &45.3\% &59.6\% & 49.9\% & 70.0\% \\
&PGD &40.6\% &54.7\% & 42.8\% & 65.8\% \\
&DI &40.6\% &54.9\% & 42.9\% & 66.0\% \\
&MI &45.0\% &58.9\% & 48.1\% & 69.2\%\\
&TI-DIM &54.3\% &48.7\% & 50.1\% & 58.7\% \\
&IR &48.9\% &57.5\% & 51.2\% & 70.1\%\\
&AEG &47.7\% &63.8\% & 56.0\% & 71.7\% \\
&MTA &40.6\% &61.3\% & 49.2\% & 69.7\% \\
&UAP &46.9\% &65.6\% & 66.2\% & 77.3\% \\
&RAP &47.2\% &62.7\% & 56.1\% & 70.8\% \\
&FDA+xent &43.3\% &53.8\% & 45.0\% & 66.2\% \\
\rowcolor {gray!20} \cellcolor{gray!0}&\textbf{ICE} &{56.3\%} &\textbf{{83.2\%}} & \textbf{{90.1\%}} & \textbf{{92.4\%}} \\
\rowcolor {gray!20} \cellcolor{gray!0}&\textbf{ICE$_{A\_I}$} &{54.1\%} &{80.9\%} & {87.9\%} & {91.7\%} \\
\rowcolor {gray!20} \cellcolor{gray!0}&\textbf{ICE$_{C\_\text{Tiered}_{T84}}$} &{57.8\%} &{78.3\%} & {84.9\%} & {90.1\%} \\
\rowcolor {gray!20} \cellcolor{gray!0}&\textbf{ICE$_{C\_\text{Tiered}_{V56}}$} &{64.2\%} &{82.1\%} & {87.2\%} & {89.8\%} \\
\rowcolor {gray!20} \cellcolor{gray!0} &\textbf{SA} &\textbf{{66.9\%}} &{80.5\%} & {85.0\%} & {86.9\%}\\
\bottomrule
\end{tabular}
\label{tab:attack on cifar100}
\end{table}
In the second experiment, we
use Cifar-10 and Tiered$_{T84}$ as the source datasets, and use two RN-18 models respectively trained on source datasets as the source models.
Interestingly, the baselines' performances in this experiment are commonly worse than their performances in the first experiment.
The possible reason for this result is that the resolution of the images from Tiered$_{T84}$ is 84$\times$84, which differs greatly from the resolution of the target dataset Cifar-100.
In comparison, ICE's performance in this experiment is much better than its performance in the first experiment, which indicates that ICE can efficiently make use of all the resources to improve its performance despite the difference among the source datasets.
Moreover, UAP performs the best among all baselines and SA still outperforms ICE.
In the third experiment, we use two RN-18 models respectively trained on Cifar-10 and Tiered$_{T56}$ as the source models.
It is observed that most of the baselines' performances in this experiment are slightly better than their performances in the third experiment.
For instance, compared with the performances in the third experiment, the performance of FGSM in this experiment is improved by about 1.7\%.
The possible reason for this result is that compared with the resolution of Tiered$_{T84}$, the resolution of Tiered$_{V56}$ is closer to the resolution of Cifar-100.
Compared with UAP, the proposed ICE and SA promote the average attack success rate by about 7.7\% and 14.1\%, respectively.
In the fourth experiment, we use three RN-18 models respectively trained on Cifar-10, Tiered$_{T84}$ and Tiered$_{V56}$ as the source models.
Compared with the best baseline UAP, ICE and SA promotes the average attack success rate by about 25.8\% and 24.7\%.
The possible reason why ICE outperforms SA in this experiment are \textbf{1)} this experiment contains more source datasets than the three previous experiments, and \textbf{2)} \textbf{ICE is more efficient than SA to leverage additional source datasets to improve its generalization}.
The results of ICE$_{A\_I}$, ICE$_{C\_T84}$, and ICE$_{C\_V56}$ will be discussed in Section~\ref{sec:transfer noises}.
\begin{table}[t]
\centering
\caption{The GTA success rates on Cifar-10, Tiered$_{T84}$, and Tiered$_{V56}$.
}
\setlength\tabcolsep{5pt}
\footnotesize
\begin{tabular}{cccccc}
\toprule
Resource & Method & MN-V3 & VGG-16 & RN-18 & DN-26\\
\midrule
\multirow{6}{*}{\tabincell{c}{-Cifar-10 \\ (RN-18)}}
&FGSM &21.3\% &28.3\% & 33.0\% & 47.9\% \\
&MI &20.6\% &28.7\% & 33.2\% & 47.8\% \\
&IR &19.8\% &28.3\% & 32.3\% & 47.9\% \\
&AEG &24.4\% &38.5\% & 41.2\% & 52.1\% \\
&UAP &22.6\% &37.9\% & 49.2\% & 54.9\% \\
&RAP &23.0\% &38.2\% & 40.9\% & 52.2\% \\
\rowcolor {gray!20} \cellcolor{gray!0} &\textbf{ICE} &\textbf{36.3\%} &{45.5\%} & {61.0\%} & {65.5\%}\\
\rowcolor {gray!20} \cellcolor{gray!0} &\textbf{SA} &{23.2\%} &\textbf{52.5\%} & \textbf{68.8\%} & \textbf{70.9\%}\\
\midrule
\multirow{6}{*}{\tabincell{c}{-Tiered$_{T84}$ \\ (RN-18)}}
&FGSM &56.5\% &62.1\% & 83.5\% & 88.7\% \\
&MI & \textbf{58.2\%} &61.8\% & 82.6\% & 88.3\% \\
&IR &52.5\% &56.8\% & 79.3\% & 87.5\% \\
&AEG &47.8\% &45.6\% & 56.0\% & 64.5\% \\
&UAP &57.3\% &63.3\% & 82.8\% & 88.0\% \\
&RAP &47.7\% &49.6\% & 58.3\% & 64.6\% \\
\rowcolor {gray!20} \cellcolor{gray!0} &\textbf{ICE} &{52.3\%} &{73.0\%} & {90.8\%} & {93.5\%}\\
\rowcolor {gray!20} \cellcolor{gray!0} &\textbf{SA} &\textbf{63.5\%} &\textbf{87.2\%} & \textbf{97.5\%} & \textbf{97.7\%}\\
\midrule
\multirow{6}{*}{\tabincell{c}{-Tiered$_{V56}$ \\ (RN-18)}}
&FGSM &65.7\% &69.8\% & 78.1\% & 86.9\% \\
&MI &64.6\% &68.0\% & 76.2\% & 86.5\% \\
&IR &59.7\% &67.1\% & 74.3\% & 83.5\%\\
&AEG &57.0\% &67.2\% & 72.7\% & 78.6\% \\
&UAP &70.9\% &65.7\% & 72.3\% & 82.3\% \\
&RAP &57.1\% &67.3\% & 72.2\% & 78.7\% \\
\rowcolor {gray!20} \cellcolor{gray!0} &\textbf{ICE} &{72.0\%} &{83.5\%} & {91.7\%} & {93.3\%}\\
\rowcolor {gray!20} \cellcolor{gray!0} &\textbf{ICE$_{A\_I}$} &\textbf{79.7\%} &\textbf{90.7\%} & {95.3\%} & {96.0\%} \\
\rowcolor {gray!20} \cellcolor{gray!0} &\textbf{ICE$_{C\_T84}$} &{69.8\%} &{88.3\%} & \textbf{96.5\%} & \textbf{98.1\%} \\
\rowcolor {gray!20} \cellcolor{gray!0} &\textbf{SA} &{71.3\%} &{77.0\%} & {91.2\%} & {92.5\%}\\
\bottomrule
\end{tabular}
\label{tab:attack across}
\end{table}
\subsection{Generalized transferable attack to Cifar-10, Tiered$_{T84}$, and Tiered$_{V56}$}
\label{sec:exp_attacking_across_datasets}
Now we show ICE and SA still outperform baselines when using other datasets as target images.
Table \ref{tab:attack across} reports the experimental results.
There are four datasets in total (Cifar-10, Cifar-100, Tiered$_{T84}$, and Tiered$_{V56}$), and
each row (denoted as -target) shows the experiment where we conduct GTA on the target dataset by using RN-18 trained on the other three datasets as source models.
For example, the `-Cifar10' row denotes the experiment where we use the other three datasets and the three respectively trained RN-18 models as resources to conduct GTA to Cifar-10 images.
Table~\ref{tab:attack across} does not show the `-Cifar100' row because the corresponding results have been shown in the last row of Table~\ref{tab:attack on cifar100}.
Note that to save space, Table \ref{tab:attack across} ignores some unimportant baselines used in Table \ref{tab:attack on cifar100}.
It is clear that given three datasets and the models trained on the three datasets, ICE and SA outperform the baselines to attack unknown images from other datasets.
The results of ICE$_{A\_I}$ and ICE$_{C\_T84}$ will be discussed in Section~\ref{sec:transfer noises}.
\subsection{Universal perturbation across images, datasets, and resolutions}
\label{sec:transfer noises}
We have experienced with the transfer-ability of ICE and SA under the standard GTA experiments.
Fig.\ref{fig:fig1} shows that ICE generates similar texture-like noises for different images, even when the images are sampled from different datasets with different resolutions.
Inspired by this phenomenon, here we perform an experiment to study whether the ICE noise generated for each image can lead to universal perturbation across images, datasets, and resolutions.
To evaluate whether the ICE noise generated for each image is universal within the target dataset, we randomly sample 1000 testing images from the target dataset, and transfer the ICE noise generated for each sampled image to other 999 sampled images.
We conduct this experiment on Cifar-100 and Tiered$_{V56}$ and denote this experiment as ICE$_{A\_I}$.
Table~\ref{tab:attack on cifar100}'s fourth raw and Table~\ref{tab:attack across}'s third raw show the experimental results, where ICE$_{A\_I}$ performs comparable to ICE, demonstrating that ICE noise generated for each image is universal to perturb other images within the same dataset.
Note that for each image, we transfer its ICE noise to the other 999 images, so there are 1000$\times$999 attack results on each unknown target model, and the reported result on each model is the average of the 1000$\times$999 results.
We conduct cross-dataset experiments to evaluate whether the ICE noise generated for each image is universal to directly attack images from other datasets with different resolutions.
Let's use $\mathbb{D}_a$ and $\mathbb{D}_b$ to denote the two datasets in each cross-dataset experiment scene.
Because each original dataset in our paper contains at least 10,000 testing images, using all testing images to run the cross-dataset experiments is expensive.
So for each original dataset, we sample 1000 testing images to construct a sub-dataset.
Therefore we can obtain two sub-datasets $\mathcal{D}_a \in \mathbb{D}_a$ and $\mathcal{D}_b \in \mathbb{D}_b$.
In the cross-dataset experiment, we transfer the ICE noise of each image $x_a \in \mathcal{D}_a$ to each image $x_b \in \mathcal{D}_b$ with the formulation $\hat{x}_b = x_b + \epsilon\cdot\text{sign}(\text{crop}(\delta_a))$, where $\delta_a=\hat{x}_a-x_a$ is the ICE noise generated for image $x_a$.
We denote the ICE noise transferred to $x_b$ as ICE$_{C\_\mathbb{D}_a}$.
The experimental results shown in Table~\ref{tab:attack on cifar100}'s fourth raw and Table~\ref{tab:attack across}'s third raw demonstrate that ICE noise generated for each image is universal across datasets and resolutions.
Note that the image resolution of $\mathcal{D}_a$ should be no smaller than that of $\mathcal{D}_b$.
Either $\mathcal{D}_a$ or $\mathcal{D}_b$ contains 1000 sampled testing images, so there are 1000$\times$1000 cross-dataset testing results on each unknown target model, and the reported result on each model is the average of the 1000$\times$1000 results.
\begin{table}[t]
\centering
\caption{GTA results on CUB models. }
\footnotesize
\begin{tabular}{ccccc}
\toprule
Resource & Method & VGG-16 & RN-18 & DN-26 \\
\midrule
\multirow{6}{*}{\tabincell{c}{Cifar-10 \\ (RN-18)}}
&FGSM & 14.4\% &22.7\% &32.6\% \\
&DI & 31.8\% &49.2\% &72.5\% \\
&MI & 13.6\% &22.0\% &30.7\% \\
&TI-DIM & 40.9\% &62.8\% &73.4\% \\
&MTA &42.6\% &51.3\% &62.3\% \\
&UAP &18.2\% &31.0\% &47.5\%\\
&RAP &45.9\% &60.5\% &72.6\%\\
\rowcolor {gray!20} \cellcolor{gray!0} &\textbf{ICE} &{47.9\%} &{66.8\%} &{87.3\%} \\
\rowcolor {gray!20} \cellcolor{gray!0} &\textbf{SA} &\textbf{50.3\%} &\textbf{92.5\%} & \textbf{95.2\%}\\
\bottomrule
\end{tabular}
\label{tab:attack CUB}
\end{table}
\subsection{Attacking Fine-Grained classification models}
\label{sec:attacking CUB}
We evaluate whether the proposed ICE and SA can be directly transferred to disturb the images from fine-grained classification datasets.
In this experiment, the source dataset and source model are Cifar-10 and RN-18 respectively, and the target dataset is CUB~\cite{WahCUB_200_2011}, a fine-grained bird classification dataset that differs greatly from Cifar-10.
The target models are VGG-16, RN-18, and DN-26 trained on CUB.
The three target models use a consistent input resolution of 112$\times$112 and achieve approximately 49.2\%, 54.9\%, and 48.3\% testing accuracies.
The attack success rates on the three target models are reported in Table~\ref{tab:attack CUB}.
Obviously, SA outperforms ICE, and ICE outperforms the other methods.
\subsection{Attacking ImageNet models}
\label{sec:attacking ImageNet}
Here we transfer the proposed ICE and SA to perturb victim images from the 392 ImageNet categories that do not belonging to TieredImageNet.
In this experiment, we use three RN-18 models respectively trained on Cifar-10, Tiered$_{T84}$ and Tiered$_{V56}$ as the source models, and use the publicly trained\footnote{https://github.com/tensorflow/models/tree/r1.12.0/research/slim} ResNet-101, DenseNet-161, Inception-V1, and Inception-V3 as the target models.
The four models will be shortened as RN-101, DN-161, Inc-V1, and Inc-V3, respectively.
Table~\ref{tab:attack ImageNet} reports the experimental results and demonstrates that the proposed ICE and SA still outperform the baselines in perturbing images that are sampled from novel categories and are using much larger resolutions (224$\times$224, 299$\times$299) than the source datasets.
\begin{table}[t]
\centering
\caption{GTA results on 392 ImageNet categories. }
\footnotesize
\setlength\tabcolsep{5pt}
\begin{tabular}{cccccc}
\toprule
Resource & Method & RN-101 & DN-161 & Inc-V1 & Inc-V3 \\
\midrule
\multirow{8}{*}{\tabincell{c}{Cifar-10 \\ + Tiered$_{T84}$ \\ + Tiered$_{V56}$ \\ (RN-18)}}
&FGSM & 8.9\% & 6.6\% & 8.5\% & 6.7\% \\
&DI & 8.0\% & 6.5\% & 6.9\% & 5.8\% \\
&MI & 9.3\% & 7.6\% & 9.4\% & 7.9\% \\
&TI-DIM & 21.9\% & 18.7\% & 22.0\% & 16.9\%\\
&MTA & 19.6\% & 16.3\% & 22.5\% & 17.2\%\\
&RAP & 23.9\% & 19.5\% & 23.0\% & 18.3\%\\
&UAP & 21.2\% & 15.8\% & 17.1\% & 17.8\%\\
\rowcolor {gray!20} \cellcolor{gray!0} & \textbf{ICE} & \textbf{30.8\%} & \textbf{30.5\%} & {35.0\%} & {38.3\%}\\
\rowcolor {gray!20} \cellcolor{gray!0} & \textbf{SA} & {30.6\%} & {23.8\%} & \textbf{{43.3\%}} & \textbf{{42.5\%}}\\
\bottomrule
\end{tabular}
\label{tab:attack ImageNet}
\end{table}
\begin{table}[t]
\centering
\caption{Ablation GTA experiments on Cifar-100.
}
\footnotesize
\begin{tabular}{cccccc}
\toprule
Setting & Method & MN-V3 & VGG-16 & RN-18 & DN-26\\
\midrule
\multirow{5}{*}{w/o SP}
&PGD &7.0\% &13.9\% & 12.3\% & 19.1\% \\
&DI &6.9\% &13.0\% & 11.7\% & 18.5\% \\
&MI &\textbf{{33.8\%}} &\textbf{{50.9\%}} & {45.3\%} & {60.7\%} \\
&TI-DIM &10.0\% &8.0\% & 8.5\% & 12.1\%\\
\rowcolor {gray!20} \cellcolor{gray!0} &\textbf{ICE} &21.9\% & 23.6\% & 23.3\% & 35.1\% \\
\midrule
\multirow{6}{*}{Default}
&PGD &37.3\% &50.9\% & 43.2\% & 61.5\%\\
&DI &39.3\% &54.9\% & 47.2\% & 64.7\%\\
&MI &45.4\% &62.3\% & 56.3\% & 72.1\%\\
&TI-DIM &51.0\% &45.6\% & 48.0\% & 55.3\%\\
\rowcolor {gray!20} \cellcolor{gray!0} &\textbf{ICE} &{54.9\%} &{67.3\%} & {71.5\%} & {83.5\%}\\
\bottomrule
\end{tabular}
\label{tab:ablation}
\end{table}
\subsection{Ablation Study}
We verify the effect of each setting or component in our work via ablation experiments.
Note that the target dataset is Cifar-100 in all ablation experiments.
\subsubsection{The effect of SP}
\label{sec:adlation-SP}
When evaluating ICE and baselines on the GTA problem, we use the SP trick (the third step in Section \ref{sec:inference of ICE}) to improve their performances.
Here we validate the effectiveness of SP by removing it, which means for ICE, we directly use the adversarial example $x^{(T)}$ generated in the second step in Section \ref{sec:inference of ICE} to attack the target models.
Similarly, for the PGD-based baselines, we directly use $x_{adv}$ generated in the third step in Section \ref{sec:evaluate of baselines} to attack the target models.
Table~\ref{tab:ablation} reports the experimental results.
The source model and source dataset are RN-18 and Cifar-10, respectively.
It is clear that without SP, the performances of all methods are damaged.
The reason is that SP enlarges the average perturbation scale of each pixel, which possibly is an important factor for GTA performance.
\subsubsection{Smaller perturbation}
\label{sec:smaller_eps}
Here we test whether the proposed ICE and SA are sensitive to the perturbation scale by changing $\epsilon$ from the default 15 to 8.
The source datasets are Cifar-10, Tiered$_{T84}$, and Tiered$_{T56}$.
The source models are three RN-18 models respectively trained on the three source datasets.
Table~\ref{tab:more experiment on cifar100} shows the experimental results.
Both ICE and SA outperform baselines with significant margins.
In lots of cases, ICE and SA double the performances achieved by baselines, which further indicates the effectiveness of the proposed ICE and SA.
\begin{figure*}[t]
\centering
\includegraphics[width=0.95\textwidth]{figures/visualization3.pdf}
\caption{
Some adversarial examples for clean images from Tiered$_{V56}$ and Tiered$_{T84}$.
The adversarial examples are generated via FGSM, PGD, MI, DI, TI-DIM, RAP, UAP, MTA, ICE, and SA, with $\epsilon=15$.
For visualization, we normalize the visualized noises into the range of 0 to 255.
}
\label{fig:visualization}
\end{figure*}
\subsection{Attacking object detection images}
\label{sec:attacking object detection}
We conduct an experiment to test whether the proposed ICE and SA can be directly used to perturb object detection images.
In this experiment, the target dataset is MS-COCO\cite{lin2014microsoft}, a popularly used object detection dataset, and the victim models are RetinaNet\cite{lin2017focal}, Faster-RCNN\cite{ren2015faster}, and Mask-RCNN\cite{he2017mask}.
All the victim models use ResNet-50\cite{he2016deep} as backbone.
The source datasets are Cifar-10, Tiered$_{T84}$, and Tiered$_{V56}$, and the source models are three RN-18 models respectively trained on the three source datasets.
$\epsilon$ is also set to 15.
Note that for each method (include baselines), we still perturb MS-COCO images with the same pipelines as we do in Sections~\ref{sec:attack cifar100}, \ref{sec:exp_attacking_across_datasets}, \ref{sec:attacking CUB}, and \ref{sec:attacking ImageNet} without any special modification (\emph{i.e.}, just treat MS-COCO images as classification images).
For each MS-COCO image and each attack method, we create the adversarial image by the following steps: \textbf{1)} resize the clean image to 512$\times$512; \textbf{2)} use the attack method to perturb the resized MS-COCO image; \textbf{3)} resize the perturbed MS-COCO image back into the original shape.
The metric is average precision of predicted bounding boxes (AP$_{bbox}$).
Table~\ref{tab:attack_object detection} reports the experimental results.
Among all methods, SA performs the best to perturb MS-COCO images.
For instance, from RetinaNet's clean AP$_{bbox}$ 36.6\%, SA decreases AP$_{bbox}$ to 19.3\% while UAP can only decrease AP$_{bbox}$ to 22.5\%.
We will visualize the perturbed MS-COCO images and the detection results in Appendix.
\begin{table}[t!]
\centering
\caption{GTA success rates on Cifar-100 with $\epsilon$=8.
}
{\footnotesize
\begin{tabular}{cccccc}
\toprule
Resource & Method & MN-V3 & VGG-16 & RN-18 & DN-26\\
\midrule
\multirow{8}{*}{\tabincell{c}{Cifar-10 \\ + Tiered$_{T84}$ \\ + Tiered$_{V56}$ \\ (RN-18)}}
&FGSM &18.3\% &29.9\% & 22.3\% & 39.7\% \\
&MI &17.5\% &28.5\% & 21.6\% & 39.5\%\\
&AEG &21.1\% &29.7\% & 23.5\% & 38.0\%\\
&MTA &17.2\% &26.5\% & 21.0\% & 35.3\% \\
&UAP &20.9\% &37.5\% & 34.0\% & 48.3\% \\
&RAP &15.2\% &31.9\% & 24.5\% & 41.0\% \\
\rowcolor {gray!20} \cellcolor{gray!0} &\textbf{ICE} &{26.9\%} &\textbf{49.9\%} & \textbf{62.4\%} & \textbf{72.2\%}\\
\rowcolor {gray!20} \cellcolor{gray!0} &\textbf{SA} &\textbf{36.0\%} &{45.3\%} & {53.9\%} & {58.6\%}\\
\bottomrule
\end{tabular}
}
\label{tab:more experiment on cifar100}
\end{table}
\begin{table}[t]
\centering
\caption{GTA results on object detection images. }
\setlength\tabcolsep{3.5pt}
\scriptsize
\begin{tabular}{ccccc}
\toprule
Resource & Method & RetinaNet & Faster-RCNN & Mask-RCNN\\
\midrule
\multirow{10}{*}{\tabincell{c}{Cifar-10 \\ + Tiered$_{T84}$ \\ + Tiered$_{V56}$ \\ (RN-18)}}
&Clean &36.6\% &37.4\% & 38.0\% \\
\midrule
&FGSM &34.0\%($\downarrow$2.6\%) &34.7\%($\downarrow$2.7\%) & 35.7\%($\downarrow$2.3\%) \\
&MI &34.2\%($\downarrow$2.4\%) &34.9\%($\downarrow$2.5\%) & 35.5\%($\downarrow$2.5\%) \\
&TI-DIM &22.9\%($\downarrow$13.7\%) &23.0\%($\downarrow$14.4\%) & 23.7\%($\downarrow$14.3\%) \\
&MTA &26.0\%($\downarrow$10.6\%) &26.8\%($\downarrow$10.6\%) & 27.5\%($\downarrow$10.5\%) \\
&UAP &22.5\%($\downarrow$14.1\%) &23.2\%($\downarrow$14.2\%) & 23.8\%($\downarrow$14.2\%) \\
&RAP &24.6\%($\downarrow$12.0\%) &25.1\%($\downarrow$12.3\%) & 25.9\%($\downarrow$12.1\%) \\
\rowcolor {gray!20} \cellcolor{gray!0} &\textbf{ICE} &{27.7\%($\downarrow$8.9\%)} &{28.2\%($\downarrow$8.8\%)} & 28.8\%($\downarrow$9.2\%) \\
\rowcolor {gray!20} \cellcolor{gray!0} &\textbf{SA} &{19.3\%}\textbf{($\downarrow$17.3\%)} &{19.7\%}\textbf{($\downarrow$17.7\%)} & {20.3\%}\textbf{($\downarrow$17.7\%)} \\
\bottomrule
\end{tabular}
\label{tab:attack_object detection}
\end{table}
\section{Visualization}
\label{sec:visualization}
We visualize generated adversarial examples, the noise maps, and spectrum diagrams in Fig.\ref{fig:visualization}.
To save space, we only sample two clean images from Tiered$_{T84}$ and Tiered$_{V56}$, respectively, for visualization, and will show more visualizations in Appendix.
When attacking the sampled Tiered$_{T84}$ image, we use Cifar-10, Cifar-100, and Tiered$_{V56}$ as the source datasets, and use three RN-18 respectively trained on the source datasets as the source models.
When attacking the sampled Tiered$_{V56}$ image, we use Cifar-10, Cifar-100, and Tiered$_{T84}$ as the source datasets.
The raw FFT$_R$ denotes the spectrum of the noise map's R channel and Appendix will visualize FFT$_G$ and FFT$_B$.
By analyzing both Figs.\ref{fig:fig1} and \ref{fig:visualization}, we conclude that the perturbation noise generated by ICE differs greatly from those generated by the other methods.
Compared with other methods which use irregular noise patterns to perturb different images, ICE uses regular similar texture-like noises to perturb images from different datasets.
More visualizations in Appendix will also support this phenomenon.
The possible reason for this phenomenon is that ICE needs to disturb the input image without knowing any information in advance, which forces ICE to learn the perturbation pattern that correlates little with ground-truth information.
In our opinion, learning a generalizable perturbation pattern may be a straightforward way for ICE to solve GTA.
The spectrum diagrams demonstrate that the main components in each ICE noise are three sine waves for R, G, and B channels.
The proposed SA directly optimizes the three sine waves.
The visualization shows that the optimized sine waves use similar frequencies and directions with the three sine waves in ICE noise.
\section{Conclusion and future work}
In this paper, we propose the Generalized Transfer Attack (GTA) setting which is more challenging and more realistic than existing adversarial attack settings. We show that although existing attacks can be modified to solve this problem, they often perform poorly when the target dataset/model is different from sources. We then propose a novel Image Classification Eraser (ICE) method and show that ICE outperforms all the baselines in generalized transfer attack.
More importantly, visualization and fast fourier transform analysis demonstrate that ICE uses similar texture-like noise pattern to perturb all images and the main components in each ICE noise are three sine waves.
Inspired by this finding, we propose another novel method Sine Attack which directly optimizes three sine waves to solve the GTA problem.
Experiments also demonstrate that SA outperforms comparably to ICE.
Our work indicates that the three sine waves are enough and effective to perturb unknown images under the GTA setting.
{\small
\bibliographystyle{ieee_fullname}
|
\section{Introduction}
Convolutional neural networks (CNNs) achieve impressive results in a variety of applications, such as image classification~\cite{krizhevsky_imagenet_2017} or object detection~\cite{ren_faster_2017}. The performance of a CNN trained on a given training data set is typically assessed in terms of prediction accuracy on a held-out validation dataset. If the statistical distributions of the training and validation set differ, the high performance can drop precipitously. In that case, the model is not robust against the variability within the validation data. CNN robustness has been shown to be affected by image perturbations such as cropping~\cite{zheng2016improving}, blur~\cite{dodge2016understanding}, high-frequency noise~\cite{yin_fourier_2019}, or texture variations~\cite{geirhos_imagenet_2019}. Prominent examples of noise perturbation are adversarial attacks: Single pixels of an input image are changed slightly such that a CNN misclassifies it~\cite{szegedy_intriguing_2014}. To humans, these changes are typically imperceptible, and they would still assign the correct labels.
Robustness is a highly safety-critical aspect of CNNs in various applications, such as self-driving cars~\cite{lechner2020neural}. For example, researchers have demonstrated how targeted yet unsuspicious changes of traffic signs can cause CNNs to consistently miss stop signs~\cite{eykholt_robust_2018}. Understanding which factors influence the robustness of CNNs and, consequently, designing and evaluating more robust models are therefore research topics of central importance in the machine learning community~\cite{goodfellow2014explaining,szegedy_intriguing_2014,madry_towards_2019,lechner2021adversarial}.
To create a basis for tackling robustness, researchers aim to gain a better general understanding of which image features CNNs are most sensitive to and how this sensitivity differs from human visual perception \cite{babaiee2021off}. For example, Geirhos et al.~\cite{geirhos_imagenet_2019} investigated the influence of shape versus texture on humans and CNNs in image classification and found a strong bias for texture in CNNs. This texture bias was also confirmed by Zhang and Zhu~\cite{zhang_interpreting_2019} by evaluating the effect of image saturation or random shuffling of image patches. Through systematic analysis of image perturbation in the Fourier domain, Yin et al.~\cite{yin_fourier_2019} could show that CNNs are highly sensitive to high-frequency perturbations.
A common way to improve robustness is to train models on more variable input images, using data augmentation methods~\cite{perez2017effectiveness}, image stylization~\cite{geirhos_imagenet_2019}, or adversarial examples~\cite{madry_towards_2019}.
To evaluate the performance of these improved models, they are then compared to standard CNN models.
To perform these analyses, researchers automatically perturb images offline, log the predictions for each perturbed input image, and compare the results between models.
According to our collaborating domain experts, such experiments take several hours to set up and include time-consuming parameter searches. For some experiments, it is not even possible to determine network performance fully automatically as the ground truth of the input images also becomes unclear for humans because of the performed image perturbations. Clearly, these factors severely limit machine learning experts to quickly explore factors of image perturbation that may impact CNN performance.
We hypothesize that \emph{being able to interactively manipulate a synthetic input scene with a large and diverse set of visual perturbation parameters and observing the changing activations and predictions instantly will allow researchers to quickly generate hypotheses and build a stronger intuition for potential vulnerabilities of CNNs}.
In this work, we therefore introduce Perturber -- a novel real-time experimentation interface for exploratory analysis of model robustness (see Figure~\ref{fig:teaser}).
Our work provides the following contributions:
\begin{enumerate}
\item Interactive input perturbations of 3D scenes in combination with feature visualizations, activation maps, and model predictions as a novel approach to interactively explore model robustness.
\item A direct comparison interface for qualitative visual validation of more robust, fine-tuned model variants.
\item Implementation of a publicly accessible web-based VA tool\footnote{\url{http://perturber.stefansietzen.at/}} that supports a large variety of perturbation methods of 3D input scenes and visualizes the model responses in real-time.
\item Observations from case studies with machine learning experts demonstrating that live inspection of input perturbations allows experts to visually explore known vulnerabilities, to compare model behaviors, and to generate new hypotheses concerning model robustness.
\item Quantitative verification of selected user observations from the case study shedding new light on the robustness of adversarially trained models.
\end{enumerate}
\section{Related Work}
In recent years, a wide spectrum of deep learning visualization methods has emerged.
For a comprehensive overview of visual analytics (VA) for deep learning, we refer the reader to a survey by Hohman et al.~\cite{hohman_visual_2019}.
For example, graph structure visualizations, such as the \emph{TensorFlow Graph Visualizer}~\cite{wongsuphasawat_visualizing_2018}, help users to get a better understanding of their models' structure, which comprise numerous layers and connections.
Others, such as \emph{DeepEyes}~\cite{pezzotti_deepeyes_2018} and \emph{DeepTracker}~\cite{liu_deeptracker_2019}, track detailed metrics throughout the training process to facilitate the identification of model problems or anomalous iterations.
\emph{ExplAIner}~\cite{spinner_explainer_2020} goes beyond monitoring and also integrates different steering mechanisms to help users understand and optimize their models.
A case study using a VA system to assess a model's performance to detect and classify traffic lights has shown that interactive VA systems can successfully guide experts to improve their training data~\cite{gou2020vatld}.
While these examples all focus on the inspection of a single model, others support model comparisons.
For example, using \emph{REMAP}~\cite{cashman_ablate_2020}, users can rapidly create model architectures through ablations (i.e., removing single layers of an existing model) and variations (i.e., creating new models through layer replacements) and compare the created models by their structure and performance.
Ma et al.~\cite{ma_visual_2020} designed multiple coordinated views to help experts analyze network model behaviors after transfer learning.
\emph{CNN\-Comparator}~\cite{zeng2017cnncomparator} uses multiple coordinated views to compare the architecture and the prediction of a selected input image between two CNNs. Model comparison is also a crucial aspect of our work. However, the focus of the present work lies on fluid modification of input stimuli and instantaneous analysis and comparison of the network responses. Thus, we let users \emph{generate and perturb} input images from 3D scenes in a ``playground-like'' manner rather than letting the user select input images with a ground-truth class label.
Interactive ``playgrounds'' require relatively little underlying deep learning knowledge and can be used for educational purposes. For more informed users, they are valuable for building an intuition and validating knowledge from literature~\cite{kahng_gan_2019}.
Notable examples are \emph{TensorFlow Playground} \cite{smilkov_direct-manipulation_2017}, which supports the interactive modification and training of DNNs in the browser, and \emph{GANLab}~\cite{kahng_gan_2019}, which is designed in a similar style and supports experimentation with Generative Adversarial Networks. \emph{CNN Explainer}~\cite{wang_cnn_2020} provides a visual explanation of the inner workings of a CNN by showing connections between layers and activation maps, allowing the user to choose the input from a fixed collection of images.
More similarly to our work, Harley~\cite{bebis_interactive_2015} provides an online tool where users can draw digits onto a canvas. Then, the responses of all neurons in a simple MNIST-trained network are visualized in real-time. \emph{Adversarial Playground}~\cite{norton2017adversarial} allows users to interactively generate adversarial attacks and instantly observe the predictions of a simple MNIST-trained network.
To support interactive probing of model responses based on input modifications, \emph{Prospector}~\cite{krause_interacting_2016}, the \emph{what-if-tool}~\cite{wexler_what-if_2020}, and \emph{NLIZE}~\cite{liu2018nlize} allow users to interrogate the model by varying the input in the domains of tabular data and natural language processing, respectively.
These works inspired us to build a system that lets users \emph{interactively explore} model robustness through input perturbations. In contrast to prior work, Perturber operates on complex CNN models, such as Inception-V1 trained on ImageNet, and can be used to discover and explain vulnerabilities to complex input scenes, like animals or man-made objects in different environments.
To explain a CNN model, there are powerful methods to reveal the role of a network's hidden units by visualizing their learned features. \emph{Feature visualization} is an activation maximization
technique, which was improved by combining a variety of regularization techniques~\cite{yosinski_understanding_2015, olah_feature_2017}. Feature visualizations have been used to
identify and characterize causal connections of neurons in CNNs~\cite{cammarata_thread_2020}, and to comprehensively document the role of individual neurons in large CNNs \cite{OpenAIMi10:online}. Within VA tools, feature visualizations have been used to compare learned features before and after transfer learning~\cite{ma_visual_2020} or to visualize a graph of the most relevant neurons and their connections for a selected target class~\cite{hohman_summit_2020}. Similarly, \emph{Bluff}~\cite{das_bluff_2020} shows a graph containing the most relevant neurons explaining precomputed adversarial attacks, where neurons are represented by their feature visualizations.
Other powerful interpretability methods are \emph{saliency maps} (or \emph{attribution maps}), which show the saliency of the input image's regions with respect to the selected target class or network component~\cite{simonyan2013deep}.
Saliency maps and other gradient-based methods like \emph{LRP} \cite{lapuschkin_lrp_2016}, \emph{Integrated Gradients} \cite{sundararajan_axiomatic_2017}, or \emph{Grad-CAM} \cite{selvaraju_grad-cam_2017}, however, require a computationally expensive back-propagation pass. A very simple solution to reveal relevant image regions for a model's prediction is to directly visualize the forward-propagation activations of selected feature maps in intermediate layers. For example, the \emph{DeepVis Toolbox}~\cite{yosinski_understanding_2015} shows live visualizations of CNN activations from a webcam feed.
The goal is to get a general intuition what features a CNN has learned.
\emph{AEVis}~\cite{liu_analyzing_noise_2018} shows activations of neurons to a pre-defined set of input images along a ``datapath visualization''. This visualization allows users to trace the effects of adversarial attacks through the hidden layers of a network. Datapaths are formed by critical neurons and their connections that are responsible for the predictions.
Like the \emph{DeepVis Toolbox}~\cite{yosinski_understanding_2015}, Perturber visualizes activations and predictions based on live input.
The major difference is that Perturber generates input images from an interactive 3D scene and provides a rich palette of input perturbation methods to gradually explore potential vulnerabilities of a model. In addition, it facilitates direct comparison between a standard model and a more robust variation thereof to explore the benefits and limitations of model variations.
\section{Perturber Interface}
The high-level goal of Perturber is to interactively explore potential sources of vulnerabilities for CNNs to facilitate the design of more robust models.
Through constant exchange between visualization researchers and machine learning experts investigating model interpretability and robustness, we identified four central requirements of a VA application to support exploratory analysis and qualitative validation of model robustness:
\textbf{R1:} Rich \textbf{online input perturbation} of a representative scene is essential to quickly generate input images for which model responses can be investigated. The system should provide a large variety of possible perturbation parameters and allow a flexible combination of these perturbations.
\textbf{R2:} To understand what makes a model robust, it is not sufficient to understand \emph{how} a model responds to the perturbed input but also \emph{why} the model's responses change. To this end, looking not only at the input and output, but particularly at the numerous \textbf{hidden layers} is inevitable
when trying to understand what makes a model react unpredictably for humans
\textbf{R3:} Interactivity can help users to build intuitions through dynamic experiments~\cite{kahng2019does}. We thus aim to make the model responses to input perturbations \textbf{instantly} visible to support a fluid feedback loop in a playground-like environment.
\textbf{R4:}
Robustness can be achieved by training models on more versatile input images, using data augmentation methods, such as affine image transformations, image stylization, or adversarial training.
A direct visual \textbf{comparison} between the standard CNN model and its more robust fine-tuned version are necessary for a first qualitative validation whether a fine-tuned model is generally more robust to input perturbations or only selectively more robust to specific perturbations it has been trained on.
\begin{figure*}[h]
\centering
\includegraphics[width=0.9\textwidth]{images/interface.pdf}
\caption{Perturber interface: users manipulate the 3D scene (A) through a large variety of input perturbation methods that are grouped into functional categories (B). Per category, input perturbations can be seamlessly controlled through sliders (see inset (C)). Users can select two models to compare, as well as a set of neurons from dedicated layers (D). For the selected neurons and models, respectively, neuron responses are visualized live (E). Top-5 predictions for both models are shown as logits or probabilities along with an image example (F). }
\label{fig:interface}
\end{figure*}
Perturber supports these four core requirements through a highly interactive web-based playground consisting of the following components: the 3D input scene (Section~\ref{sec:inputScene}, Figure~\ref{fig:interface} A), the perturbation control (Section~\ref{sec:perturbationControl}, Figure~\ref{fig:interface} B-C), the prediction view (Section~\ref{sec:predictionView}, Figure~\ref{fig:interface} F), and the (comparative) neuron activation view (Section~\ref{sec:activationView}, Figure~\ref{fig:interface} E-D). These views allow for a qualitative inspection of the effects of input scene perturbations on one or two selected models (Section~\ref{sec:comparison}).
Conceptually, Perturber can handle any CNN architecture.
The current version supports models based on the Inception-V1 (also called GoogLeNet)~\cite{szegedy2015going} architecture. Our standard model was trained on ImageNet~\cite{russakovsky2015imagenet}. We use a pre-trained model with weights accessible through the \emph{Lucid}\footnote{https://github.com/tensorflow/lucid} feature visualization library~\cite{olah_feature_2017}.
\subsection{Input Scene}
\label{sec:inputScene}
The input image is generated from an interactive 3D scene that can be manipulated in numerous ways (Section~\ref{sec:perturbationControl}). In computer vision projects, rendered images are often used instead of -- or in addition to -- photographs for training CNNs~\cite{su2015render}.
Quantitative experiments have shown that, for high-level computer vision algorithms, the gap between synthetic and real images is fairly small~\cite{gaidon2016virtual}. CNN responses were shown to be consistent between simple rendered 3D models and natural images~\cite{aubry2015understanding}. Interactive exploration of CNN responses based on synthetic scenes therefore seems to be sufficiently representative of real-world applications.
The major advantage of a synthetic 3D scene is that perturbation factors can be easily disentangled. That means, input modifications can be flexibly applied independently from each other to assess their isolated effects as well as their interactions.
Out of the 1000 ImageNet classes, 90 of them are dog breeds. Therefore, there are numerous neurons in our Inception-V1 model that specialize in dog-related patterns, such as snouts, head-orientations, fur, eyes, ears, etc. As a consequence, we chose a dog as our main foreground 3D object to represent this special significance and to feed an input image that many hand-identified neuron circuits~\cite{cammarata_thread_2020} respond to strongly. As a second 3D model, we provide a vehicle, which has very different characteristics from a dog and is a commonly analyzed object in the machine learning community (e.g., \cite{gaidon2016virtual,aubry2015understanding}). \rev{In addition, users can upload custom 3D models.}
This way, users can verify observations they have made \rev{also with other models.
\subsection{Perturbation Control}
\label{sec:perturbationControl}
The core principle of Perturber is that the user can flexibly vary the perturbation factors they want to explore (R1) and observe their effects instantly (R3).
Below the input scene view, Perturber provides sliders for seamless control of the perturbations (Figure~\ref{fig:interface} C). Perturber provides more than 20 scene perturbation factors, as well as all possible combinations between these factors.
We group these perturbation methods into the following categories:
\textbf{Geometry} perturbations, such as rotation, translation, or cropping an image, are classic perturbations a CNN might be sensitive to~\cite{zheng2016improving,athalye_obfuscated_2018,engstrom2019exploring}. Therefore, affine image transformations are also a common data augmentation method for more robust training~\cite{perez2017effectiveness}. We support geometric perturbations through simple camera controls, which allow the user to arbitrarily orbit around the 3D model, as well as to freely zoom and pan the scene to create image cropping effects. In addition, users can rotate the camera around the z-axis to simulate image rotation.
\textbf{Scene} perturbations that may seem irrelevant for human observers may easily confuse a standard CNN model. For example, changing the background~\cite{xiao2020noise} can have a tremendous effect on the prediction. We, therefore, support various scene perturbation operations, such as changing the background image, as well as modifying the lighting and texture parameters of the main object (see Figure~\ref{fig:scenePerturbations}). Through combinations of these parameters, specialized perturbations, such as silhouette images, can be generated (see Figure~\ref{fig:scenePerturbations} bottom right).
\begin{figure}[h]
\centering
\includegraphics[width=1\columnwidth]{images/scenePerturbations.pdf}
\caption{Scene perturbations -- first row: original scene, no lighting influence, full background blur, different background image; second row: no texture influence, full texture blur, lowest background saturation, combination of background saturation, lighting influence, texture influence, and background blur. All perturbation parameters can be controlled seamlessly. }
\label{fig:scenePerturbations}
\end{figure}
\textbf{Shape} perturbations let users morph the shape of the main scene object into another one. Perturber currently supports morphing between dog and cat, as well as between a firetruck and a race car. By morphing the shape of an object independently from its texture, the texture-shape cue conflict~\cite{geirhos_imagenet_2019} can be investigated.
\textbf{Color} perturbations act on the rasterized 2d image. We support individual post-processing effects, such as alpha blending to black, hue shifting, and (de-)saturation.
In addition to these parameters, users are provided with a text field where they can write their own GLSL code, taking a single parameter which is controlled by the slider. The code snippet defaults to code for contrast adjustment. Using these post-processing effects, users can assess the models' surprisingly high robustness against low contrast~\cite{dodge2016understanding} and low image saturation~\cite{sturmfels2020visualizing}.
\textbf{Frequency} perturbations selectively modify different image frequencies. Perturber provides three-parameter frequency decomposition that splits the image into low and high-frequency bands, which is achieved by Laplacian decomposition. Through selective frequency suppression, users can investigate phenomena such as the one described by Yin et al.~\cite{yin_fourier_2019}, where the authors show that adversarially trained networks are robust against high-frequency perturbations but very sensitive to low-frequency perturbations.
\textbf{Spatial} perturbations are post-processing effects systematically changing the image's pixel order. Perturber supports patch shuffling, which was used by Zhang and Zhu~\cite{zhang_interpreting_2019} to reveal a model’s sensitivity to global structure, which gets highly disturbed for human observers by patch shuffling. The image is divided into a grid of $k \times k$ cells, which are then randomly re-ordered.
\textbf{Adversarial} perturbations, finally, are a core feature to understanding model robustness. Users can perform projected gradient descent (PGD) adversarial attacks~\cite{madry_towards_2019} on the current scene image. To do that, they choose a model to generate the attack from, a target class or the option to suppress the original prediction, as well as the attack $\epsilon$ and $L_p$ norm ($L_2$ or $L_{\infty}$). With each button press, the user performs one PGD step with a step length of $\epsilon/8$. This perturbation is costly and cannot be performed instantaneously. The first step typically takes around 10 seconds on a powerful consumer PC as the gradient function needs to be computed for the current input image first. To let users inspect the effect of an adversarial attack on the model fluidly like the other perturbation parameters, we overlay the current input image with the perturbation vector once it has been computed. This way, users can interactively fade the attack alpha using a slider and observe the system's response instantaneously. They can also fade the original image to inspect the perturbation vector itself.
\subsection{Prediction View}
\label{sec:predictionView}
The prediction view shows the top-5 classification results for each model (Figure~\ref{fig:interface} F) either as probability or as logits. This allows the user to observe the classification changes resulting from input perturbations in real-time (R3).
Each model's top result is represented by a class image example. Perturber shows the first image of the respective class in the ImageNet validation dataset.
\subsection{Neuron Activation View}
\label{sec:activationView}
The neuron activation view is the central interface for the analysis of how input perturbations affect the models' hidden layers (R2). Perturber represents neurons through feature visualizations.
Feature visualizations aim at providing a lens into networks to visualize what patterns certain network units respond to \cite{olah_feature_2017}. These patterns might be, for example, edges in a particular direction in earlier layers (cf., Figure~\ref{fig:neurons}), or specific objects, like dog heads (cf., Figure~\ref{fig:teaser}) or car windows, in later layers.
\begin{figure}[b]
\centering
\includegraphics[width=1\columnwidth]{images/activations_stdModel.pdf}
\caption{Feature visualizations for neurons associated with complex shapes and curvatures in layer mixed4a for the standard model, as well as their activation maps for the input on the left. Note how rotating the input model causes an activation change for oriented shape detectors (insets on the bottom). }
\label{fig:neurons}
\end{figure}
Activation maps show which regions of the input image cause the respective neuron to be activated. After experimenting with an implementation of the gradient-based visualization method \emph{Grad-CAM}~\cite{ribeiro_why_2016}, we decided to focus on the computationally far less expensive activation maps showing the neuron activations for a forward propagation pass. This way, neuron activations can be observed instantly (R3).
Input regions that highly activate the neuron are shown in red, while blue regions cause a negative activation (see Figure~\ref{fig:neurons}). As the user manipulates the input scene, the activation maps update instantaneously so that the user can observe in real-time to which image features a neuron responds in one of the models. For example, in Figure~\ref{fig:neurons}, neurons 412 and 418 of layer mixed4a respond to oriented patterns. Rotating the textured race car causes these two neurons to strongly change their activations.
Inception-V1 contains thousands of neurons. Clearly, it is impossible to show feature visualizations and interactive activation maps for all neurons concurrently. Instead, we provide pre-selected neuron groups and categories for each layer, which were characterized into semantically meaningful neuron families in the course of the \emph{Circuits} project~\cite{cammarata_thread_2020}. In addition to these previously presented neuron groups, we identified further sets of dog-, cat-, race car-, and firetruck-related neurons in later layers using \emph{Summit}~\cite{hohman_summit_2020}. These pre-selected neuron groups consisting of one to up to around 70 neurons can be selected from drop-down menus (Figure~\ref{fig:interface} D).
\subsection{Model Comparison}
\label{sec:comparison}
After exploring potential sources of vulnerability, model designers commonly fine-tune existing models using augmented training data covering the identified vulnerability. To leverage Perturber for direct comparison of robustness (R4), we employ a transfer learning approach, which is initialized with the weights from the standard model. This is necessary to guarantee that feature correspondence of individual network units is kept intact. Only this way, we can assume that units across models respond similarly to identical input images.
To illustrate what individual network units respond to, Perturber relies on feature visualizations. Feature visualizations are independent of the input image, but they differ between models. Indeed, Table~\ref{tab:featureVis} shows noticeable differences between a single neuron's feature visualizations of three different model variations.
To showcase model comparison, Perturber already includes two robust model variations of Inception-V1:
The first model was adversarially trained with projected gradient descent (PGD), which was described by Madry et al.~\cite{madry_towards_2019}. Adversarial training increases model robustness by incorporating strong adversarial attacks into the training procedure. The second model was trained by Stylized-ImageNet, which is a variation of ImageNet, where images were transformed into different painting styles using a style transfer algorithm~\cite{geirhos_imagenet_2019}. The purpose of Stylized-ImageNet was to induce a shape bias, while the standard ImageNet-trained models were found to be biased unproportionally towards texture~\cite{geirhos_imagenet_2019}. Users can choose two models for comparison from the interface (Figure~\ref{fig:interface} D). In addition, they can choose multiple checkpoints along the incremental fine-tuning process to analyze the development of the models during training. Corresponding neuron activation views are then juxtaposed in the Perturber interface for direct comparison (Figure~\ref{fig:interface} E).
\section{Web-Based Implementation}
\label{sec:implementation}
Perturber runs purely on the client-side in the user’s browser and without any server-side computations at runtime.
We precomputed all data that does not depend on interactively changeable elements to make the UI as efficient as possible.
For transfer learning, we initialized the weights with those of the standard model (i.e., Inception-V1 trained on ImageNet), which were obtained through the \emph{Lucid} library~\cite{olah_feature_2017}. For adversarial fine-tuning, we used the open-source implementation by Tsipras et al.~\cite{tsipras_robustness_2019}. No layers were frozen for transfer learning. For both fine-tuned models, we used a learning rate of 0.003, a batch size of 128, and we trained until the models reached a top-5 training error of around 0.5, which took approximately 50K iterations for the adversarial fine-tuning and around 90K iterations for the Stylized-ImageNet fine-tuning.
To generate feature visualization, we employed the implementation provided by the \emph{Lucid}~\cite{olah_feature_2017} library.
For each model, we generated feature visualizations for 5808 neurons from the most relevant layers (i.e., the first three convolutional layers and the concatenation layers at the end of each mixed-block).
During fine-tuning, we obtained feature visualizations for 17 checkpoints of adversarial fine-tuning and seven checkpoints of Stylized-ImageNet fine-tuning.
For each neuron, we computed feature visualizations using two parametrization methods -- naive pixel or Fourier basis~\cite{olah_feature_2017} (see Table~\ref{tab:featureVis}). The second performs gradient ascent in Fourier space and leads to a more equal distribution of frequencies, resulting in more naturally looking feature visualizations. We use \emph{transformation robustness}~\cite{olah_feature_2017} in addition to both methods. Without Fourier basis parametrization, the differences between the models are more visually distinct (Table~\ref{tab:featureVis} top row).
The computation of all $\sim$300K generated feature visualizations required around one month on a machine with two NVIDIA GTX 1070 GPUs. \footnote{All generated feature visualizations can be downloaded from \url{https://github.com/stefsietz/perturber/}%
}.
\begin{table}[]
\centering
\caption{Comparison between naive pixel (top) and Fourier basis (bottom) parametrized feature visualizations of neuron 222 in layer mixed4a for three model variations. }
\label{tab:featureVis}
\begin{tabular}{m{1.45cm}M{1.64cm} M{1.64cm} M{1.64cm}}
& \textbf{standard} & \textbf{adversarial} & \textbf{stylized} \\
\textbf{naive pixel param.} &
\includegraphics[scale=.4]{images/featureVis/mixed4a_222_standard_naive.jpg} & \includegraphics[scale=.4]{images/featureVis/mixed4a_222_adversarial_naive.jpg} &
\includegraphics[scale=.4]{images/featureVis/mixed4a_222_stylized_naive.jpg} \\
\textbf{Fourier basis param.} &
\includegraphics[scale=.4]{images/featureVis/mixed4a_222_standard.jpg} & \includegraphics[scale=.4]{images/featureVis/mixed4a_222_adversarial.jpg} & \includegraphics[scale=.4]{images/featureVis/mixed4a_222_stylized.jpg}\\
\end{tabular}
\end{table}
Manipulation of the 3D scene, model inference based on the rendered image, and computation of activation maps are performed live in the web browser.
The client relies on GPU acceleration for both, 3D scene rendering and CNN inference. We use the WebGL-based libraries \emph{Three.js} and \emph{TensorFlow.js}~\cite{smilkov_tensorflowjs_2019} for these tasks, respectively. The front-end GUI is based on \emph{React.js}.
Input perturbations based on post-processing effects are implemented as multiple sequential render passes with custom GLSL shaders.
For model inference, we use \emph{Tensorflow.js}~\cite{smilkov_tensorflowjs_2019}, which enables fast GPU-accelerated CNN inference. \emph{Tensorflow.js} is also used for computing adversarial attacks.
A major requirement of Perturber is that the effects of input perturbations can be observed instantaneously (R3).
To assess requirement R3, we measured the client's performance while constantly orbiting the camera around the object. Figure~\ref{fig:performance} shows the recorded frame rates for two client notebooks: a MacBook Pro 13” 2018 with Intel Iris Plus Graphics 655 (MBP) and an AORUS 15G Gaming Notebook with an NVIDIA GeForce GTX 2080 Super GPU (AORUS). It is clearly visible that the GPU of the client machine has a strong influence on the frame rate. The application performance also depends on the enabled visualizations, with the prediction view (Section~\ref{sec:predictionView}) being more computationally expensive than showing four live activation maps (Section~\ref{sec:activationView}). When using less powerful client machines, users can selectively switch off views.
\begin{figure}[t]
\centering
\includegraphics[width=0.8\columnwidth]{images/performance.pdf}
\caption{Performance benchmarks for four different output configurations, measured on two machines (MBP or AORUS): neuron activation view (NAV) only, or neuron activation view in combination with prediction view (NAV+PV) visualized for one or two visualized models concurrently. }
\label{fig:performance}
\end{figure}
\section{Exploration Scenario: Texture-Shape Conflict}
\label{sec:scenarios}
To demonstrate the capabilities of Perturber, we selected one important aspect of robustness: the influence of shape and texture on CNN classification results. It has been shown that CNNs are biased towards texture, which affects robustness~\cite{geirhos_imagenet_2019}. In this scenario, we first aim to investigate whether this effect could have been discovered using Perturber. Secondly, we aim to qualitatively validate whether a more robust model variation, which was fine-tuned using Stylized-ImageNet~\cite{geirhos_imagenet_2019}, can improve upon this bias.
For our first analysis, we explore the texture-shape cue conflict through shape and texture perturbations. Figure~\ref{fig:morphing} illustrates that shape perturbations alone do not cause the standard model to predict a cat breed. However, when morphing the texture, the model predictions switch to cat breeds -- even when the object shape remains unchanged. This is a first indicator that the model is indeed much more sensitive to texture than to shape perturbations.
\begin{figure}[h]
\centering
\includegraphics[width=0.7\columnwidth]{images/morphing.pdf}
\caption{Object morphing between dog (bottom left) and cat (top right) along the texture dimension (y axis) and shape dimension (x axis): Top-1 standard model predictions with their probabilities are shown as text labels. Input images leading to cat breeds within the top-5 or top-1 predicted classes are indicated by a yellow and orange frame, respectively (top row). }
\label{fig:morphing}
\end{figure}
To further investigate the texture sensitivity of the standard model, we replicated the patch shuffling experiment by Zhang and Zhu~\cite{zhang_interpreting_2019}. Using Perturber, we can inspect single neurons' activations during this experiment. We illustrate our observations on a hand-picked neuron in Figure~\ref{fig:patchShuffling}, which is strongly activated by dog faces looking to the left. Note how this neuron is activated by strong texture contrasts, especially around the mouth of the dog. Image regions containing ears do not activate this particular neuron. For this scene, the standard model still predicts a dog breed up to $k=7$ randomly shuffled image patches. This illustrates that the decomposed shape has indeed very little influence on the model prediction.
\begin{figure}[]
\centering
\includegraphics[width=0.8\columnwidth]{images/patchShuffling.pdf}
\caption{Randomly shuffling the dog scene into 3 (top), 4 (middle), and 8 (bottom) patches: A dog-related neuron in mixed4b of the standard model is activated by patches containing parts of the dog's face and textured body parts. }
\label{fig:patchShuffling}
\end{figure}
In the next step, we analyze if the standard model is indeed more sensitive to texture than the model fine-tuned on Stylized-ImageNet~\cite{geirhos_imagenet_2019} by gradually removing the texture of the main object. The predictions in Figure~\ref{fig:stylizedTexture} confirm that the Stylized-ImageNet trained model consistently predicts a dog breed, even in the absence of a texture. The standard model, on the other hand, seems to rely much more on the texture. The model trained with Stylized-ImageNet is also more sensitive to patch shuffling, which is another indication that it relies less on texture information than the standard model (see Section A in the Supplemental Document for image examples).
\begin{figure}[h]
\centering
\includegraphics[width=0.7\columnwidth]{images/stylizedTexture.pdf}
\caption{Inspecting the shape vs.~texture conflict through scene perturbations for the standard (left) and the Stylized-ImageNet trained model (right): while the standard model gets confused by missing textures, the Stylized-ImageNet trained model predicts a dog breed even if the 3D model is untextured.}
\label{fig:stylizedTexture}
\end{figure}
Finally, we compare shape sensitivity between the two models. To this end, we combine various scene perturbations to generate a silhouette image of the dog. We then gradually change the pose of the dog. Figure~\ref{fig:silhouette} shows the predictions of the standard model and the model fine-tuned with Stylized-ImageNet. Clearly, the predictions of the standard model are unstable, especially when the dog is directly facing the camera. The model trained with Stylized-ImageNet, however, reliably predicts a white wolf with high probability. The activations of relevant neurons indicate that the Stylized-ImageNet-based model has learned a more stable representation of a frontal dog head, which also get activated by a silhouette image.
\begin{figure}[b]
\centering
\includegraphics[width=1\columnwidth]{images/silhouette_activations.pdf}
\caption{Rotating a dog silhouette: the corresponding predictions of the standard model and the model trained on Stylized-ImageNet, as well as activations of dog-related neurons for the front-facing dog image in mixed4d of the standard model (left) and the robust model (right). }
\label{fig:silhouette}
\end{figure}
These examples illustrate that Perturber provides flexible ways to explore and better understand potential threats to robustness, such as the texture-shape conflict of CNNs. Section A in the supplemental document contains more exploration scenarios.
\section{Case Studies}
\label{sec:caseStudies}
While the previous scenario investigated a known vulnerability, we further explored whether Perturber can help to discover unknown threats to robustness through case studies with machine learning experts.
We conducted case studies with five machine learning researchers (four PhD students, one post-doctoral researcher; one female, four males). Except for one user, sessions were conducted individually through video conferencing and lasted approximately one hour each. One user preferred to perform the case study offline and sent a textual text report instead.
The researchers cover various topics of expertise in the field of AI interpretability and the design of robust machine learning models. The concrete research areas are listed in Section B of the supplemental document.
Two users were involved in the co-design process of Perturber and are co-authors of the paper. Three users were unfamiliar with the system before the evaluation. One of these users entered the co-design iteration process after the case study and is also a co-author. Paper co-authorship is a common collaboration role in design studies~\cite{sedlmair2012design}.
During the video conference, the participant used the online tool while sharing his/her screen with the first and last author. Every session was recorded while conversations and observations were transcribed on-the-fly or in retrospective through automatic speech-to-text.
Every video conferencing session started with a short introduction by the participant describing his/her research focus. Afterwards, we gave a short demonstration of Perturber's features. We then asked the participant to shortly comment on his/her first impression and his/her expected insights from the analysis. Then, the participant freely played around with Perturber while thinking aloud. In particular, we asked the user to always state his/her intent before performing an action and whether he/she would have any particular hypotheses about the response and behavior of the network based on the chosen input. If a user could not find the respective functionality of the tool, the first and/or last author provided oral assistance. At the end of the study, the user was encouraged to summarize his/her impressions, the potential benefits of the tool, and to provide suggestions for improvements.
\subsection{Observations and Feedback}
Users praised the fact that Perturber works ``live'' and therefore allows for ad-hoc \textbf{exploratory analyses}. They liked that Perturber allows to play around with simple examples to quickly find patterns and form hypotheses. Generally, the main focus of the exploratory analyses was trying to identify input perturbations where a model would respond unexpectedly. One user described this process as trying to answer the following questions: \emph{``How can I break a model? What do I need to do so that the resulting prediction is wrong?''} For example, three users were surprised to see how vulnerable the adversarially trained model seems to be to some geometric changes, such as zooming or rotating. Two users also discovered a high sensitivity of adversarially trained models to background modifications, as illustrated in Figure~\ref{fig:backgroundBlurPredictions}.
The live approach was praised in particular for cases without a clear ground truth, such as object morphing (Figure~\ref{fig:morphing}). Such scenarios would not be possible to assess quantitatively without human subject studies. Being able to quickly generate hypotheses that could then be formally tested in a more controlled setting was considered very useful.
\begin{figure}[h]
\centering
\includegraphics[width=0.9\columnwidth]{images/backgroundBlurPredictions.pdf}
\caption{Blurring the background of the race car has little influence on the standard model's predictions (left). The adversarially trained model (right) is very sensitive to the chosen background. The probability for ``racer'' is 11\% with a grass background and 7\% with a uniform green background. }
\label{fig:backgroundBlurPredictions}
\end{figure}
We also observed indications that Perturber is practically helpful for \textbf{visual confirmation}.
For example, using the model predictions and activations of selected neurons, one user verified that adversarial attacks only affect the standard model. He also observed how animal-related neurons of the standard model got activated during an attack with the target class ``badger'' (Figure~\ref{fig:adversarialAttack}).
Not all assumptions were actually confirmed. For example, one user expected that the model trained on Stylized-ImageNet would be noticeably more vulnerable to image blur than the standard model. Unexpectedly, the model's responses were not more sensitive to blur than the standard model's for the chosen input scene (see Section A in the supplemental document).
\begin{figure}[b]
\centering
\includegraphics[width=1\columnwidth]{images/adversarialAttack.pdf}
\caption{Adversarial attack with target class ``badger'' on the fire truck scene (top) and the effect on two cat-related neurons in layer mixed4b for the standard model (left) and adversarially trained model (right): As the attack strength is increased (bottom), activations of some standard model neurons increase while the activations of the adversarially trained model are hardly affected. The prediction probabilities for the fire engine decrease to 0.1\% for the standard model and to 6\% for the adversarially trained model after the attack. }
\label{fig:adversarialAttack}
\end{figure}
Finally, a user pointed out that playing around with Perturber would \textbf{let non-experts get an intuition} how easily CNNs are fooled. For example, one user demonstrated how the standard model sometimes rapidly changes its predictions during a simple translation of the dog. Another expert demonstrated that a rotation of the dog along the z-axis in combination with an unusual background (street) was sufficient to disturb the adversarially trained model. Accordingly, he stated that Perturber could be informative for ``people fearing that AI will take over the world''.
A comprehensive list of observations reported by the individual users can be found in Section B of the supplemental document.
Overall, the most frequently performed perturbations leading to the most interesting observations in our case studies were 1) geometric transformations, such as object rotation, zooming, and translation, 2) modification of the background, 3) combinations of (1) and (2), as well as 4) object morphing. The prediction view was perceived as giving instant, easily interpretable, and useful feedback. It was thus the primary view to observe model behavior. Feature visualizations were considered useful to characterize the difference between the models. Participants described feature visualizations of the adversarially trained model as more ``intuitive'' or ``cartoonish'' compared to the corresponding standard model's neurons. One participant found feature visualizations sometimes hard to interpret. One recommendation therefore was to additionally show strongly activating natural image examples from a dataset.
\subsection{Quantitative Evaluation of User Observations}
We performed quantitative measurements to see whether what users observed visually can be generalized beyond the given input scene, synthesized input images, and the Inception-V1 network architecture. To test the generalizability of the users' observations, we performed quantitative measurements using different models than the ones used in the online tool. Specifically, we used the pre-trained ResNet50 from the \emph{torchvision} library of \emph{PyTorch} \cite{pytorch} as the standard model. For comparison, we used weights of an adversarially trained version of the same model from the \emph{robustness} library~\cite{robustness} (ResNet50 ImageNet $L_2$-norm $\epsilon$ 3/255).
First, we investigated the adversarially trained model's sensitivity to background changes, which was reported by two users in the case study. For verification, we performed the \emph{Background Challenge} by Xiao et al.~\cite{xiao2020noise}, where a model is tested with (natural image) adversarial backgrounds.
The adversarially trained model can only correctly predict 12.3\% under background variations. The standard model achieves 22.3\% accuracy. Using this test dataset, random guessing would yield 11.1\% accuracy~\cite{xiao2020noise}. This shows that the robustified network is considerably more susceptible to adversarial backgrounds than the standard model.
Second, we tested if the adversarially trained model is indeed more sensitive to geometric scene transformations than the standard model (reported by three users). In particular, we looked into camera rotation. We generated a synthetic dataset, where we rendered the four 3D models provided in Perturber from seven yaw angles, ranging from -70$^{\circ}$ to 70$^{\circ}$\ (Figure~\ref{fig:yaw}), two pitch angles, and two distances of the camera to the object, as shown in Figure \ref{fig:prototypes}. In total, we generated 28 views for each of the four 3D models.
\begin{figure}[]
\centering
\subfloat[]{
\includegraphics[width=0.9\columnwidth]{images/yaw_examples.jpg}
\label{fig:yaw}
}
\vspace{-0.2cm}
\\
\subfloat[]{
\includegraphics[width=0.9\columnwidth]{images/pitch_dist_variation.jpg}
\label{fig:prototypes}
}
\caption{Seven yaw angles variations (a) tested for four prototype views (b) per model (two of the four models are shown here). The prototype view in (a) is highlighted.}
\label{fig:rotation}
\end{figure}
To compare how much the predictions fluctuate, we chose a prototype view with a yaw angle of -23.3$^{\circ}$ for each of the four pitch / distance combinations and 3D model (Figure \ref{fig:prototypes}). For each of the 16 resulting prototypes, the logits of the top-10 classes served as ground truth vector $\vec{l}_{10}^{\star}$.
For the 16 prototype views, we then calculated a fluctuation score $f_p$:
\begin{equation}
f_p = \sum_{y} \frac{\lVert \vec{l}_{10}^{\star} - \vec{l}_{10}^{y} \rVert_2}{s(\vec{l}_{n}^{\star})},
\end{equation}
where $\vec{l}_{10}^{\star}$ are the top-10 predictions of the prototype view, $\vec{l}_{10}^{y}$ is the logit vector of these 10 classes for the view associated with yaw $y$, and $s(\vec{l}_{n}^{\star})$ is the standard deviation of the logits of all $n$ classes in the prototype view. In other words, the fluctuation score measures how strongly the logits of the top-10 prototype classes diverge in the rotated input images.
Figure~\ref{fig:score} shows the average fluctuation score of all 28 prototype views.
The fluctuation scores are considerably higher for the adversarially trained model for three of the four 3D models. This verifies that the adversarially trained model can be indeed more vulnerable to rotations of the main object.
\begin{figure}[h]
\centering
\includegraphics[width=0.8\columnwidth]{images/yaw_diffs.pdf}
\caption{Average yaw fluctuation scores for the standard model and the adversarially trained model across the four 3D models. Error bars show the standard deviation. }
\label{fig:score}
\end{figure}
\section{Discussion \& Conclusions}
We showed that interactive perturbations in combination with live activations can be an effective method to explore potential vulnerabilities of CNNs and to perform qualitative evaluations of more robust model variations. In an exploration scenario, we could replicate the known texture-shape conflict~\cite{geirhos_imagenet_2019} through multiple perturbation examples. Machine learning experts participating in our case study observed a variety of known but also unexpected network behaviors. \rev{They could successfully replicate known CNN properties, such as models’ varying sensitivity to adversarial attacks or patch shuffling, using our synthetic input scenes. In addition, our quantitative post-study experiments of selected user observations could replicate their observations using a different network architecture and natural images. }
Through these experiments, we demonstrated vulnerabilities of adversarially trained models to background modifications and yaw rotations of the main object that, to the best of our knowledge, have not been discussed yet in the machine learning community.
The majority of insights reported by our users were based on observations how the model predictions changed when perturbing the image. This implies that the principle of live input perturbations could also be useful for pure black-box models. To get a better intuition about which image features are influential for the model's final decision, our exploration scenarios indicate that neuron activations are also essential. But due to the large number of logical neuron units distributed among multiple layers, it can be time-consuming to find a set of neurons that eventually is affected strongly by the performed input perturbation. In the future, we thus plan to investigate alternative methods to select the displayed neurons in the neuron activation view. Also visual guidance to support the discovery of potentially harmful perturbation factors would be helpful. However, traditional guidance mechanisms may require costly computation, which will hamper interactivity (R3). Effective guidance mechanisms can therefore be considered interesting future work. \rev{Another interesting line of future work would be the support for encoder-decoder architectures, used prominently for semantic segmentation and image translation tasks among others. This could be facilitated by replacing the prediction view with a continuously updating display of the generated image.}
\section*{Acknowledgments}
We thank Robert Geirhos and Roland Zimmermann for their participation in the case study and valuable feedback, Chris Olah and Nick Cammarata for valuable discussions in the early phase of the project, as well as the Distill Slack workspace as a platform for discussions.
M.L.~is supported in part by the Austrian Science Fund (FWF) under grant Z211-N23 (Wittgenstein Award).
J.B.~is supported by the German Federal Ministry of Education and Research (BMBF) through the Competence Center for Machine Learning (TUE.AI, FKZ 01IS18039A) and the International Max Planck Research School for Intelligent Systems (IMPRS-IS). R.H. is partially supported by Boeing and Horizon-2020 ECSEL (grant 783163, iDev40).
\bibliographystyle{eg-alpha-doi}
|
\section{Introduction}
A number of recent arguments seek to challenge the objectivity of measurement outcomes. More concretely, such arguments, involving scenarios in which entangled systems are shared by sets of Wigner’s friend arrangements, try to show that the assumption that the experiments performed by the friends yield definite outcomes, leads to a contradiction. From this, it is concluded that the results of (at least some) measurements, cannot be thought of as being actual or objective.
These new arguments employ experimental settings that have been used to derive Bell-type results, such as Bell's own, GHZ's or Hardy's. However, they place, on each wing of the experiments, a Wigner's friend arrangement, where the friend first measures inside a sealed room, with Wigner standing outside, and Wigner measures afterwards. This procedure allows to employ, on a single run, measurement settings that, in a standard Bell-type settings, would correspond to different possibilities. For this reason, these arguments are touted as having the added advantage of being formulated in terms of actual results of measurements---and not in terms of counterfactuals or potentially unreliable theoretical elements, such as Bell's $\lambda$'s or PBR's ontic states.
One such argument is reported in a talk by Pusey \citep{Pusey}, where it is ultimately traced to an informal argument offered by Masanes (see also \cite{H1}; \cite{H2}; \cite{H3}; \cite{Bub}; \cite{Leifer}). The proposed construction involves a Bell-type setting, in which the two spin-$\frac{1}{2}$ particles of a singlet are sent to two spatially separated laboratories. Each laboratory is then stipulated to contain a Wigner's friend arrangement where, first the friend, and then Wigner, perform measurements. With such an experiment, and by assuming that all measurements yield definite outcomes, an inconsistency is claimed to arise. From that, the objectivity of measurement results is challenged.
Another such argument is offered in \cite{Zuk}, where an extended Wigner's friend scenario is used, this time inspired by the GHZ arrangement---a three spin-$\frac{1}{2}$ particles variant of Bell's. The conclusion of the analysis is that any attempt to introduce an actual outcome for the friends' measurements, leads to a logical contradiction.
The previous works are both motivated, at least in part, by the more famous no-go theorems in \cite{FR}, and \cite{Bru1}. Ironically, though, while the Pusey-Masanes argument is described as easier to understand, but containing all the essential features of \cite{FR}, Żukowski and Markiewicz present their work as ``nullifying'' the ``apparent paradoxes'' of \cite{Bru1} and \cite{FR}, and correcting their mistakes.
Frauchiger and Renner consider an extended Wigner's friend scenario modeled after Hardy's paradox---yet another variant of Bell's arrangement---and a conflict is argued to arise from the assumption that quantum mechanics can be consistently applied to complex, macroscopic systems. They take their result as proving that ``quantum theory cannot consistently describe the use of itself''. Building on \cite{Bru2}, the first to employ a Bell-Wigner mash-up, \cite{Bru1} presents a ``no-go theorem for observer-independent facts''. The result is interpreted as showing that there can be no theory in which the results of both Wigner and the friend can jointly be considered as objective properties of the world.
This work offers a detailed examination of these arguments. In particular, I show that the first two arguments crucially depend upon a mistaken assumption, regarding the ``mixed'' correlations between the results of the friends and those of ``the Wigners''. It is not, then, that the assumption of definite outcomes leads to trouble; instead, such an assumption, \emph{coupled with the unwarranted assumption about mixed correlations}, leads to inconsistencies. I further argue that the unwarranted assumption regarding mixed correlations can be traced to a lack of recognition i) that in these arguments, hidden variables, with their inevitable contextual and non-local nature, are being (implicitly) postulated, and ii) that, in spite of such features of hidden variables, signaling is fully avoided in the scenarios considered. Regarding the Frauchiger-Renner theorem, I show that, although it seems to be presented in a different language, at the end, it also crucially depends on the mistaken assumption regarding mixed correlations. As a result, the theorem is invalid.
My manuscript is organized as follows. In section \ref{M}, I present the Pusey-Masanes and the Żukowski-Markiewicz arguments and, in section \ref{MCA}, I show that they depend on an unwarranted assumption regarding mixed correlations. After that, in section \ref{FR}, I explore the Frauchiger-Renner theorem, on which the previous arguments are motivated. I show that it depends on the same unwarranted assumption regarding mixed correlations. Along the way, I illustrate my claims with explicit calculations in the context of pilot-wave theory, the best-developed hidden-variable theory (in the appendix \ref{PWM}, I collect the details of such calculations). Finally, in section \ref{Co}, I state my conclusions.
\section{The Pusey-Masanes and Żukowski-Markiewicz arguments}
\label{M}
The Pusey-Masanes (PM) argument employs an Bell-type experiment that starts with the two spin-$\frac{1}{2}$ particles of a singlet state being sent to two spatially separated laboratories. There, observers $\mathcal{A}_1$ and $\mathcal{B}_1$ perform spin measurements along directions $a_1$ and $b_1$, respectively; we denote these measurement results, which take values $+1$ or $-1$, by $A_1$ and $B_1$. Next, observers $\mathcal{A}_2$ and $\mathcal{B}_2$, who are outside of the laboratories of $\mathcal{A}_1$ and $\mathcal{B}_1$, respectively, come and \emph{undo} these initial measurements. The idea behind this is that, from the point of view of, say, $\mathcal{A}_2$, all that happens when $\mathcal{A}_1$ measures, is some complicated unitary transformation. Therefore, $\mathcal{A}_2$ can, at least in principle, come and apply the inverse of such a unitary to $\mathcal{A}_1$'s laboratory to nullify the measurement. The last step is for $\mathcal{A}_2$ and $\mathcal{B}_2$ to perform spin measurements on the original particles of the singlet, along directions $a_2$ and $b_2$, respectively; we denote the corresponding results by $A_2$ and $B_2$ (a schematic representation of the proposed experiment is presented in Figure 1).
\begin{figure
\centering
\includegraphics[height=10cm]{fig.jpg}
\caption{Schematic representation of the Pusey-Masanes experiment.}
\end{figure}
To develop the argument, we start by assuming that all the observers involved obtain objective results when they measure. That is, we assume that all the measurements performed yield definite, objective outcomes. As a result of such an assumption, there must exist a \emph{joint probability distribution} for all four results, $p(A_1,B_1,A_2,B_2)$. Given such a joint probability distribution, one can calculate the marginal probabilities $p_{\mathcal{A}_1\mathcal{B}_1}(A_1,B_1)$, $p_{\mathcal{B}_1\mathcal{A}_2}(B_1,A_2)$, $p_{\mathcal{A}_2\mathcal{B}_2}(A_2,B_2)$, $p_{\mathcal{A}_1\mathcal{B}_2}(A_1,B_2)$ and, with them, the expectation values of products of pairs of results $E_{\mathcal{A}_1\mathcal{B}_1}$, $E_{\mathcal{B}_1\mathcal{A}_2}$, $E_{\mathcal{A}_2\mathcal{B}_2}$, $E_{\mathcal{A}_1\mathcal{B}_2}$. Now, as shown in \cite{Fine}, the expectation values computed by these marginals necessarily satisfy the CHSH inequality
\begin{equation}
\label{chsh}
|E_{\mathcal{A}_1\mathcal{B}_1} + E_{\mathcal{B}_1\mathcal{A}_2} + E_{\mathcal{A}_2\mathcal{B}_2} - E_{\mathcal{A}_1\mathcal{B}_2}| \le 2 .
\end{equation}
That is, to assume a joint distribution for all four outcomes, leads to the same consequences than imposing Bell's locality. Of course, an important difference is that, in this case, all four measurements are actually performed in each run of the experiment.
Next, we employ quantum mechanics to make predictions for the expectation values in Eq. (\ref{chsh}). We start with $E^q_{\mathcal{A}_1\mathcal{B}_1}$, for which we notice that the situation exactly corresponds to a standard Bell scenario. In consequence,
\begin{equation}
E^q_{\mathcal{A}_1\mathcal{B}_1} = - \cos(a_1-b_1).
\end{equation}
Similarly, $E^q_{\mathcal{A}_2\mathcal{B}_2}$ seems easy to compute. The state starts being a singlet, then comes a big identity in its evolution, followed by standard Bell-type measurements. As a result,
\begin{equation}
\label{E2}
E^q_{\mathcal{A}_2\mathcal{B}_2} = - \cos(a_2-b_2).
\end{equation}
The ``mixed'' terms, with one measurement by $\mathcal{A}_1$ ($\mathcal{B}_1$) and the other by $\mathcal{B}_2$ ($\mathcal{A}_2$) seem more challenging. However, one might reason as follows. Say, $\mathcal{A}_1$, measures her particle. Then, to predict the result of $\mathcal{B}_2$, she updates the state according to the result obtained, and evolves the state on the other side. However, that evolution simply amounts to the identity, corresponding to the doing and undoing of the measurement. As a consequence, according to $\mathcal{A}_1$, the relation of her particle to that of $\mathcal{B}_2$, is exactly the same than to the particle of $\mathcal{B}_1$, so the prediction for the mixed correlation must be the same as in the original Bell scenario.
Another way to put this is that, for $\mathcal{A}_1$ to compute the correlation between her measurement and that of $\mathcal{B}_2$, she observes that, since the evolution from $\mathcal{B}_1$ to $\mathcal{B}_2$ is the identity, then the state of her particle and that of $\mathcal{B}_2$ is simply a singlet. Therefore, a straightforward application of Born's rule leads to the standard quantum correlations. From all this, it follows that
\begin{equation}
\label{mix}
E^q_{\mathcal{A}_1\mathcal{B}_2} = - \cos(a_1-b_2), \qquad E^q_{\mathcal{B}_1\mathcal{A}_2} = - \cos(b_1-a_2).
\end{equation}
The problem, of course, is that we know that these quantum predictions can violate the CHSH inequality. That is, for adequate values of $a_1,b_1,a_2,b_2$,
\begin{equation}
\label{chshq}
|E^q_{\mathcal{A}_1\mathcal{B}_1} + E^q_{\mathcal{B}_1\mathcal{A}_2} + E^q_{\mathcal{A}_2\mathcal{B}_2} - E^q_{\mathcal{A}_1\mathcal{B}_2}| > 2 .
\end{equation}
We arrive, then, at a contradiction. In detail, the assumption that all measurements performed in the experiment yield definite, objective outcomes, which is what allowed for Eq. (\ref{chsh}) to be established, is found to be incompatible with the predictions of quantum mechanics in Eq. (\ref{chshq}).
From all this, it is concluded that quantum measurement outcomes cannot be taken to be objective, at least not in a straightforward way. That is, the argument is taken to entail that not every quantum measurement can be thought of as having a definite, objective, physical outcome. Moreover, the argument is read as implying that, at least in some sense, whatever maneuver is employed to deal with Bell's theorem, such a maneuver cannot be restricted to the microscopic level. That is, that the measure taken to address the CHSH violation in the Bell case, must also have an effect on actual, macroscopic experimental results.
The Żukowski-Markiewicz (ZM) argument \citep{Zuk}, is to the PM argument, what the GHZ scheme \citep{GHZ}, is to Bell's theorem. They consider a setting in which the three spin-$\frac{1}{2}$ particles of a GHZ state, given by
\begin{equation}
\label{GHZ}
|GHZ \rangle = \frac{1}{\sqrt{2}} \left[| +_z, +_z, +_z \rangle + | -_z, -_z, -_z \rangle \right] ,
\end{equation}
are distributed to three spatially separated laboratories, numbered $m=$1, 2, 3. In each of them, the particle is sent to a sealed room, where it is measured by the corresponding friend, $F_m$, in the $y$ basis. Such a measurement is assumed to involve a purely unitary evolution, in which the friend simply gets correlated with the measured particle. That is, in which $ | \text{ready} \rangle_{F_m} | \pm_y \rangle_m \rightarrow | \pm \rangle_{F_m} | \pm_y \rangle_m $. Still, the measurement is assumed to yield a definite outcome. After that, the corresponding ``Wigner'', $W_m$, measures the joint friend+particle system in the $x$ basis. This measurement is assumed to be irreversible.\footnote{\cite{Zuk} incorrectly states that \emph{decoherence} is able to explain such an irreversible process (see \cite{Less} for a detailed explanation of why this is not the case). In any case, such a misconception is immaterial for the description and evaluation of the argument.}
Before constructing the argument, we note that, unlike the PM experiment, this one does not involve the undoing of the friends' measurements. However, such a difference is not really substantial. This is because, in this case, ``the Wigners'' measure the composite system, friend+particle, and not only the particle, as in the PM case. This, together with the assumption that the friends' measurements only involve them getting correlated with the particles, implies that the procedures yield analogous results.
To develop the argument, the following fact is noted. Since the measurements of the friends are fully described by unitary transformations---in which the friends and the particles get entangled---then, the probabilities that ``the Wigners'' assign to their possible results are exactly the same than what they would assign to direct measurements of the particles, if the friends were fully removed from the arrangement. They take this to imply that the results of the friends' measurements must coincide with what ``the Wigners'' would have gotten, in case the friends were removed. This, in turn, is taken to imply, for example, that $P(r,s,t|n,n',n'')_{W_1,F_2,F_3} = P(r,s,t|n,n',n'')_{W_1,W_2,W_3}$ for results $r,s,t$ (with values $\pm 1$) and settings $n, n', n''$. That is, that in the case in which all friends and all Wigners measure, the mixed probabilities for the possible results of some friends and some Wigners, and thus, the mixed correlations of those results, must mimic the probabilities and correlations of the results of all three Wigners. This, however, is shown to lead to trouble.
As is well know, the GHZ state implies a number of perfect correlations. In particular, the expectation value of the product of the results for the three particles satisfies
\begin{equation}
\label{GHZE}
E(x,y,y) = E(y,x,y) = E(y,y,x) = -E(x,x,x) = -1.
\end{equation}
So, if we denote the results of the Wigners and the friends by $w_m$ and $f_m$, we must have $w_1 f_2 f_3 = -1$, $f_1 w_2 f_3 = -1$, $f_1 f_2 w_3 = -1$ and $w_1 w_2 w_3 = 1$. But this means that $(f_1 f_2 f_3)^2 = -1$, which is inconsistent with the fact that all results have values $\pm1$.
From this, it is concluded that the attempt to introduce actual outcomes for the measurements of the friends, leads to a logical contradiction. That is, that such measurements cannot be though of as being actual in any way.
\section{Mixed correlations}
\label{MCA}
What are we to make of these arguments? Do they really show that the objectivity of measurements must be called into question? Do we really have to throw away the sensible idea that well-conducted measurements yield definite, physical outcomes?
We start by noting that a key element of these arguments is the calculation of mixed correlations between results of the friends and those of ``the Wigners''. For instance, in the PM case, we had $E^q_{\mathcal{A}_1\mathcal{B}_2} = - \cos(a_1-b_2) $ and $E^q_{\mathcal{B}_1\mathcal{A}_2} = - \cos(b_1-a_2)$. Similarly, in the ZM case, it is argued that relations, such as $P(r,s,t|n,n',n'')_{W_1,F_2,F_3} = P(r,s,t|n,n',n'')_{W_1,W_2,W_3}$, must obtain. And, of course, such mixed correlations were essential for the conclusion that the postulation of definite outcomes is incompatible with the quantum predictions for these experiments (i.e., the mixed correlations were essential in showing that such predictions violate CHSH, in the first case, and satisfy Eq. (\ref{GHZE}), even for mixed observers, in the second one). That is, the assumption that the mixed correlations---between results of the friends and those of ``the Wigners''---are equal to the standard quantum correlations among only the friends, or only ``the Wigners'', is required for these arguments to get off the ground.
Two key questions arise:
\begin{enumerate}
\item Are these mixed correlations the correct quantum predictions for these experiments?
\item Is there, or could there be, empirical evidence for such mixed correlations?
\end{enumerate}
Clearly, if the first question has a negative answer, then the alleged inconsistencies would never arise. Moreover, if the answer to the second question is also negative, then the consequences of the assumption of objective outcomes, independently of what the quantum predictions actually are, could not lead to empirical trouble.
To answer these questions we start by pointing out the obvious fact that, for these arguments to run, one has to assume that, at least during the measurements of the friends, the evolution of the quantum state is purely unitary. This is explicitly stated in the ZM scenario and is implicitly understood in the PM argument; otherwise, in the second case, the whole idea of undoing the measurements clearly does not go through. That is, it is only because it is assumed that the measurements performed by the first observers could be fully described by a complicated unitary, that it is possible to argue that such measurements could be undone by the application of another unitary---namely, the inverse of the one describing the measurements. In contrast, if one assumes that measurements involve some sort of breakdown of unitarity, the arguments simply do not work. From this, it clearly follows that theories that take this latter position, such as objective collapse models (\cite{GRW}; \cite{CSL}), are simply not at all affected by these arguments.
It is clear, then, that these arguments require the assumption that the evolution of the quantum state is always unitary. Does that mean that, at least, we can conclude that \emph{purely unitary} quantum theory is inconsistent with the assumption that measurements have definite outcomes? Not really. As we saw, the arguments presuppose 1) that measurements always yield definite results, and 2) that the evolution of the quantum state is always unitary. Now, in \cite{Tim} it is shown that those two assumptions are incompatible with another assumption, namely 3) that the physical description given by the quantum state is complete. In other words, the assumptions of these arguments necessarily imply that the quantum description is incomplete, so it must be supplemented in some way.
The point is that, a theory in which time evolution is always unitary, even during measurements, and in which the quantum state is assumed to be complete, is simply incompatible with the claim that measurements yield objective outcomes. That is, if the initial state of the system is a superposition of eigenstates of the observable to be measured, and if the interaction of the system with the measuring apparatus is purely unitary, then the final state of the composite system+apparatus will be a superposition of the possible results. If, moreover, the quantum state is assumed to be complete, then such a final superposition is simply incompatible with the claim that only one of the terms of the superposition can be taken as the actual, objective result of the measurement (for an explanation of why decoherence does not alter this conclusion, see \cite{Less}).
One might complain that the incompatibility result in \cite{Tim} depends upon assuming some level of objectivity or realism that goes against the spirit of the new type of arguments under discussion. That is, that the whole point of these new arguments is precisely to stay clear of theoretical constructs, such as hidden variables. However, if the objective of these new results is to \emph{argue against} an objective position, it would amount to begging the question to assume from the onset that such a position should not be taken.
The conclusion, then, is that the assumptions of the arguments imply that the quantum state must be supplemented by so-called hidden variables. Regarding these hidden variables, we know from the Kochen-Specker theorem \citep{KS}, that they must be \emph{contextual}, meaning that the results of at least some measurements must depend on the measurement context. Therefore, the presuppositions of these arguments establish that the outcome of measurements may depend, not only on the initial state of the system, but also on the particular way in which measurements are conducted. In particular, the outcome of an experiment may depend on which other observables are also being measured, even if remotely (of course, this possible non-local effect of remote joint measurements is a direct consequence of the unavoidable non-locality unearthed by Bell (\cite{Bell1964}; \cite{Bell1971}; \cite{Bell1976}; \cite{Bell1990}); in fact, a violation of locality, in the sense of Bell's theorem, already implies a violation of contextuality, in the sense of Kochen-Specker).
With these issues cleared up, we can go back to the two key questions regarding mixed correlations. We start with the first one: are such mixed correlations the correct quantum predictions? In order to answer, let us review what are the arguments behind the correlations offered. In the PM case, it was argued that, in order for, say, $\mathcal{A}_1$, to predict the result of $\mathcal{B}_2$, she updates the state according to her result and evolves the state on the other side. And, since such an evolution amounts to an identity, she makes the same predictions as in the original Bell scenario. That is, the idea is that, since the state of the particle when $\mathcal{B}_2$ measures is the same than when $\mathcal{B}_1$ measures, then $\mathcal{A}_1$'s correlation with both must also be the same. In the ZM argument, on the other hand, the mixed correlations are arrived at by employing counterfactual reasoning. First it is noted that the friends' results must coincide with what ``the Wigners'' would have gotten, in case the friends were removed. From that counterfactual, it is argued that, in the case in which all friends and all Wigners do measure, the mixed correlations for the results of some friends and some Wigners must coincide with the correlations of the results of all three Wigners alone.
These arguments might sound reasonable, but the recognition that hidden variables, with their inevitable contextual and non-local nature, are being (implicitly) postulated, renders them invalid. As we saw, in the PM case, the argument depended on the claim that, since the state of the particle when $\mathcal{B}_2$ measures is the same than when $\mathcal{B}_1$ measures, then $\mathcal{A}_1$'s correlation with both must also be the same. However, as explained above, the contextual nature of the setting implies that the outcomes may depend, not only on the initial state of the system, but also on which other observables are also being measured. As a result, from the fact that the quantum state is the same when $\mathcal{B}_1$ and $\mathcal{B}_2$ measure, does not follow that the correlations of those results, with that of $\mathcal{A}_1$, must coincide; the context of each measurement must be taken into account.
Regarding the other argument, the counterfactual, namely, that the friends' results must coincide with what ``the Wigners'' would have gotten, in case the friends were removed, seems quite reasonable. The problem, is that from it, it does not follow that, when everybody measures, the mixed correlations obtain. The issue, once more, is that the contextuality of the hidden variables gets in the way: what other measurements are being performed elsewhere is fully relevant to compute probabilities and correlations between measurements.
To see all this in a concrete example, we consider the PM arrangement, from the point of view of the best-developed hidden-variable framework, the de Broglie-Bohm pilot-wave theory (see appendix \ref{PWM} for all the details). Such a framework proposes that a complete characterization of a particle system is given by its wave function, \emph{together} with the actual positions of the particles, which are taken to always possess well-defined values.
Now, according to the pilot-wave perspective, a particle can be assigned an objective, definite value of spin, in relation to the path through which it is deflected when traveling through a Stern-Gerlach device \citep{Nor2}. Moreover, such a value of spin can be conferred, independently of whether the particle is measured to be there. That is, the particle can be assigned a spin value, even if there is no process to correlate its position, after it passes the apparatus, with some sort of macroscopic measuring device. Finally, after sending a particle through a Stern-Gerlach apparatus, one could use reverse magnets to deflect the two separating components and recombine them to reconstitute the original beam. Even in that case, according to the pilot-wave perspective, a definite result can be associated with the measurement.
As a result of all this, from the pilot-wave perspective, the PM experiment amounts to sending the two particles of a singlet in opposite directions, measuring them along directions $a_1$ and $b_1$, respectively, recombining the beams of each particle, and measuring them along directions $a_2$ and $b_2$, respectively. It must be mentioned that, within pilot-wave, the order of operations, even if remote, is relevant for the calculation of the results. Therefore, in order to make concrete predictions, we must specify a particular order in which the operations are performed. Given the particular order described in the appendix (see Figure 3), the expectation values of pairs of results can be calculated to be
\begin{eqnarray}
\label{Es}
E^{p}_{\mathcal{A}_1\mathcal{B}_1} &=& -\cos(a_1-b_1) \nonumber \\
E^{p}_{\mathcal{A}_2\mathcal{B}_2} &=& -\cos(a_2-b_2) \nonumber \\
E^{p}_{\mathcal{A}_1\mathcal{B}_2} &=& 0 \nonumber \\
E^{p}_{\mathcal{B}_1\mathcal{A}_2} &=& -\left(1- \frac{2 \alpha}{\pi} \right)\cos(a_1-b_1) - \left(1-\frac{2 \beta}{\pi} \right) \cos(a_2-b_2) \nonumber \\
& & - \frac{2}{\pi} \left\lbrace \alpha \cos(a_1-b_1) |\cos(a_2-b_2)| + \beta |\cos(a_1-b_1)| \cos(a_2-b_2) \right\rbrace \nonumber \\
& & + 2 \cos(a_1-b_1) \cos(a_2-b_2) \theta(\cos(a_1-b_1), \cos(a_2-b_2)) \nonumber \\
& & - 2 \cos(a_1-b_1) \cos(a_2-b_2) \theta(-\cos(a_1-b_1), -\cos(a_2-b_2)),
\end{eqnarray}
with $\alpha $ the angle between $a_1$ and $a_2$, $\beta $ the angle between $b_1$ and $b_2$ and $\theta (x,y)$ the two-dimensional Heaviside step function, which is 1 only when both $x$ and $y$ are positive.
Given these results, several comments are in order. First, we note that, while the correlations between the first and second sets of measurements coincide with the standard quantum predictions, the mixed correlations do not satisfy Eq. (\ref{mix}). In fact, we know from Fine's theorem that the predictions of any framework that assigns definite values to all measurements, such as pilot-wave, must satisfy CHSH and, indeed, it can be checked that these results do so. We also note, from $E^{p}_{\mathcal{B}_1\mathcal{A}_2}$, the explicit non-locality and contextuality of the theory, as the correlation between the measurements of $\mathcal{B}_1$ and $\mathcal{A}_2$ depends on the angles employed by $\mathcal{A}_1$ and $\mathcal{B}_2$. Finally, we note that, even though these results would change if the order of operations changes, the new results would share with these results the relevant features mentioned above.
One might worry that the non-local character of $E^{p}_{\mathcal{B}_1\mathcal{A}_2}$ could offer the possibility of signaling. For instance, the correlations between the first and second measurements on each side are given by
\begin{eqnarray}
E^{p}_{\mathcal{A}_1\mathcal{A}_2} &=& \left(1-\frac{2 \alpha}{\pi} \right) \left( 1-|\cos(a_1-b_1)| \right) \nonumber \\
E^{p}_{\mathcal{B}_1\mathcal{B}_2} &=& \left(1-\frac{2 \beta}{\pi} \right) \left( 1-|\cos(a_2-b_2)| \right) .
\end{eqnarray}
Therefore, the correlation on each side gets modulated by measurements on the other. In particular, the correlation between, say, $\mathcal{A}_1$ and $\mathcal{A}_2$, can be made to appear or disappear, by adequate choices of $b_1$. It seems, then, that a protocol could be constructed in order for the $\mathcal{B}$-side to send a superluminal message to the $\mathcal{A}$-side. We must note, however, that, by construction, it is simply impossible to actually compare first and second measurements on a given side. The point is that, by stipulation, after the first measurements, the beams are recombined, which means that absolutely all records of such measurements are fully erased. We conclude that, after all, this scheme cannot be employed for superluminal signaling.
Going back to Eq. (\ref{Es}), it is clear that the pilot-wave predictions do not coincide with Eq. (\ref{mix}). Still, one could ask what are the ``correct'' standard, quantum predictions for the PM experiment. The truth is that \emph{standard quantum mechanics simply cannot make predictions for the mixed correlations in these experiments}. The point is that, as we saw above, these arguments presuppose, both, a fully unitary evolution during measurements (at least those of the friends), and the existence of definite results for all measurements. The problem is that, within the standard framework, for definite results to obtain, a breakdown of unitarity is required. Otherwise, one ends up with a superposition of possible outcomes and not a definite result. In other words, the presuppositions of the arguments are incompatible from the get-go with the standard framework and, as a result, it is simply meaningless to ask for the correct standard quantum predictions for these scenarios.
At this point, it seems useful to speculate on what led the proponents of these arguments to embrace the mistaken assumptions. That is, to speculate on what views on quantum theory, made them blind to the fact that the mixed correlations are, in fact, incorrect. I advance that the root of the problem can be found in three interrelated issues.
First, the fact that the proponents of these arguments, implicitly or explicitly, espouse an \emph{operational} reading of quantum theory, which makes attributing objectivity to measurement outcomes foreign, even meaningless. However, as mentioned above, if the whole point of these arguments is to argue against an objective stance, adopting a position which, effectively, does not allow for objective outcomes to obtain, amounts to begging the question.
The second issue is the fact that the operational position adopted, effectively ignores the so-called measurement problem, i.e., the inherent ambiguity of the standard framework regarding what happens during measurements. However, simply ignoring such a problem, does not make it go away. The point is that, it is this ambiguity of the standard framework, when applied to these complicated Wigner’s friends scenarios, which leads to seemingly inconsistent results. Then, by ignoring the measurement problem, the inconsistencies encountered are blamed on \emph{objectivity}, and not on the inconsistencies built into the perspective from the beginning. As explained above, standard quantum mechanics simply cannot make predictions for the mixed correlations in these experiments, and this is true even if one adopts an operational position.
Finally, even accepting the measurement problem, and conceding that hidden variables are presupposed, it could be argued that one could simply ignore them, and use the Born rule to make predictions. This, however, would be mistaken. The point is that, if one assumes i) purely unitary evolution during measurements and ii) that all measurements yield objective outcomes, then there are legitimate and illegitimate uses of the Born rule. The way to distinguish between them, is by employing a theory which is, in fact, consistent with those assumptions, i.e., a hidden-variable theory, such as pilot-wave.
Now, since for situations in which the standard framework is able to make concrete predictions, e.g., situations which do not involve Wigner’s friend scenarios, pilot-wave yields the same predictions as the standard framework, in those situations, the predictions of pilot-wave, coincide with the Born rule. There are, however, other scenarios, such as those leading to the mixed correlations in the PM setting, in which the standard framework is unable to make predictions; in those cases, the use of the Born rule would be illegitimate. As we saw above, Pilot-wave can make those predictions, but they do not have to conform to the Born rule. In fact, in \cite{Berndl}, using an extended Wigner's friend scenario modeled after Hardy's paradox, it is shown that, in pilot-wave (and related theories), the Born rule cannot be satisfied on all space-like hyperplanes.
Before moving on, it is instructive to consider the predictions, for the PM setting, of a an objective collapse model, in which all measurements are associated with an objective, irreversible collapse of the quantum state. In that case, the expectation values would be given by
\begin{eqnarray}
E^{c}_{\mathcal{A}_1\mathcal{B}_1} &=& -\cos(a_1-b_1) \nonumber \\
E^{c}_{\mathcal{A}_2\mathcal{B}_2} &=& -\cos\alpha \cos\beta\cos(a_1-b_1) \nonumber \\
E^{c}_{\mathcal{A}_1\mathcal{B}_2} &=& - \cos\beta\cos(a_1-b_1) \nonumber \\
E^{c}_{\mathcal{B}_1\mathcal{A}_2} &=& -\cos\alpha \cos(a_1-b_1) .
\end{eqnarray}
As expected, these values deviate from those in Eqs. (\ref{E2}) and (\ref{mix}), so the PM argument does not apply. These values also deviate from those in Eq. (\ref{Es}), so the PM experiment would discriminate between pilot-wave theory and objective collapse models (although simpler experiments could accomplish the same thing, e.g., \cite{Bassi}, \cite{Das}).
The second key question regarding the mixed correlations is whether there is, or there could be, empirical evidence for them. The issue is that, even if frameworks that assume objective outcomes for all measurements are theoretically safe, regarding not being incompatible with standard quantum predictions, they could be in trouble if the mixed correlations where empirically required. It is easy to see, though, that the mixed correlations cannot be empirically accessed. As we saw above, the undoing of the measurements implies erasure of all records so, by the time the second measurements are performed, there is nothing to compare them with. More generally, considering all possible time orders of the different measurements of, say, the PM arrangement, it is possible to see that, at least one expectation value cannot be measured (see \cite{Leifer} for an equivalent assertion). As a result, the corresponding CHSH inequality can never be shown to be violated empirically, so not satisfying the mixed correlations does not imply a clash with experiments.
To see that, for any time order of events, at least one expectation value cannot be measured, we denote by $ e_1 \prec e_2$ when $e_2$ is on, or inside, the future light cone at $e_1$. Now, for the correlations between $\mathcal{A}_1$ and $\mathcal{B}_2$ to be empirically accessible, we need the undoing of the measurement on the $\mathcal{A}$ side to be to the future of the measurement of $\mathcal{B}_2$, i.e., we need $e_{b_2} \prec e_{a_1^{-1}}$. Otherwise, $\mathcal{B}_2$ would not have access to $\mathcal{A}_1$'s result. However, by construction of the experiment, it must be the case that $e_{a_1^{-1}} \prec e_{a_2} $ and $ e_{b_1^{-1}} \prec e_{b_2}$. That is, the second measurement on each side must be preceded by the undoing of the first measurement on that same side. The problem is that these three time orders imply that $e_{b_1^{-1}} \prec e_{a_2}$, which means that the correlations between $\mathcal{B}_1$ and $\mathcal{A}_2$ cannot be measured. We conclude that, if one of the mixed correlations can be measured, the other one cannot.
It is important to point out that, for those correlations that can be empirically accessed, hidden-variable theories have enough means, through contextuality and non-locality, to ensure that the predictions in such cases are in accordance with the standard framework (which, in such cases, is able to make predictions). This, indeed, can be shown to be the case for pilot-wave, the prime example of a hidden-variable theory.
In sum, not satisfying the mixed correlations does not lead to any sort of trouble: it neither implies a clash with standard predictions, nor it could lead to a conflict with experiments. Therefore, it is not that the assumption of definite outcomes leads to trouble, but that such an assumption, \emph{coupled with the unwarranted assumption regarding mixed correlations}, leads to inconsistent results. We conclude that these arguments fully fail in placing restrictions to theories that stipulate all measurements to lead to objective outcomes.
\section{The Frauchiger-Renner no-go theorem}
\label{FR}
In this section, we explore the no-go theorem in \cite{FR}, on which the previous arguments are motivated, and about which a lot has been written (\cite{baumann}; \cite{sudbery1}; \cite{sudbery2}; \cite{Drezet}; \cite{lazarovici}; \cite{tausk}; \cite{WCF}). In their work, Frauchiger and Renner present a result, which they take as proving that ``quantum theory cannot consistently describe the use of itself''. I confess that I do not quite understand what they mean by this phrase; in any case, what they actually do is to consider an extended Wigner's friend scenario, this time based o Hardy's variant of Bell's theorem (\cite{Hardy1}; \cite{Hardy2}), and the conflict is argued to arise from the assumption that quantum mechanics can be consistently applied to complex, macroscopic systems.
In more detail, the result is cast in the form of a no-go theorem, arguing for the mutual incompatibility of three assumptions: C, demanding consistency between different observers, S, establishing that measurements have single outcomes and Q, capturing the universal validity of quantum mechanics.\footnote{In more detail, the assumptions employed are as follows. C: if observer $O$ concludes that $O'$ is certain about the value of a property, then $O$ must also be certain about such a value; S: if $O$ is certain that the value of a property is $\xi$, then she must deny that she is certain that the value of that property is not $\xi$; Q: if $O$ assigns a state to a system to be measured, and that state assigns a Born probability of 1 to the result $\xi$, then $O$ must be certain that, after the measurement, the value of the measured property is $\xi$.} As I explained above, the Frauchiger-Renner (FR) experimental arrangement is, once more, an extended Bell-type scenario, this time modeled after Hardy's setting. There are, however, important differences between this argument and the previous ones, so it is useful to describe it in detail.
The arrangement contains four different agents, F, $\overline{\text{F}}$, W and $\overline{\text{W}}$, and two labs, L and $\overline{\text{L}}$. F is inside L and $\overline{\text{F}}$ inside $\overline{\text{L}}$ and W and $\overline{\text{W}}$ are outside of their respective labs and can perform measurements on them. Finally, there is a communication channel from $\overline{\text{L}}$ to L. The experiment runs in steps as follows (in \cite{FR}, these steps are repeated in rounds, but that is an unnecessary complication):
\begin{description}[font=\normalfont,labelindent=.5cm]
\item[\textbf{Step 1}:] $\overline{\text{F}}$ prepares a quantum coin in the state $\sqrt{\frac{1}{3}}\ket{h}+ \sqrt{\frac{2}{3}}\ket{t}$ and measures it. If she finds \emph{h}, she prepares an electron in the state $\ket{\downarrow}$, if she finds \emph{t}, she prepares it in the state $\ket{\rightarrow}= \frac{1}{\sqrt{2}} \left( \ket{\uparrow}+\ket{\downarrow} \right)$. She then sends the prepared electron to F.
\item[\textbf{Step 2}:] F measures the electron in the $\{\ket{\uparrow},\ket{\downarrow}\}$ basis.
\item[\textbf{Step 3}:] $\overline{\text{W}}$ measures $\overline{\text{L}}$ in the basis
$
\{\ket{o}_{\overline{\text{L}}}=\frac{1}{\sqrt{2}}\left[\ket{h}_{\overline{\text{L}}}-\ket{t}_{\overline{\text{L}}}\right],
\ket{f}_{\overline{\text{L}}}=\frac{1}{\sqrt{2}}\left[\ket{h}_{\overline{\text{L}}}+\ket{t}_{\overline{\text{L}}}\right] \}
$
(with $\ket{h}_{\overline{\text{L}}}$ and $\ket{t}_{\overline{\text{L}}}$ the states of $\overline{\text{L}}$ after $\overline{\text{F}}$ measures the coin and finds the corresponding result). Then she announces her result.
\item[\textbf{Step 4}:] W measures L in the basis
$
\{ \ket{o}_{\text{L}}=\frac{1}{\sqrt{2}}\left[\ket{\downarrow}_{\text{L}}-\ket{\uparrow}_{\text{L}}\right],
\ket{f}_{\text{L}}=\frac{1}{\sqrt{2}}\left[\ket{\downarrow}_{\text{L}}+\ket{\uparrow}_{\text{L}}\right] \}
$
(with $\ket{\downarrow}_{\text{L}}$ and $\ket{\uparrow}_{\text{L}}$ the states of L after F measures the spin and finds the corresponding result).
\end{description}
The experiment is then analyzed assuming C, S and Q. By doing so, several implications are considered (see Table 3 in \cite{FR}):
\begin{enumerate}
\item If $\overline{\text{F}}$ finds \emph{t}, then she sends F the electron in state $\ket{\rightarrow}$. If so, $\overline{\text{F}}$ reasons, after F measures, for W the state of L will be $\ket{f}_{\overline{\text{L}}}$, so W's measurement will result for sure in \emph{f}. That is:
\begin{equation}\label{I1}
\text{If $\overline{\text{F}}$ finds \emph{t}, she knows that W will find \emph{f} with certainty.}
\end{equation}
\item If F finds $\uparrow$, then she reasons that $\overline{\text{F}}$ must have gotten \emph{t}. But, from (\ref{I1}), this implies that W will find \emph{f} with certainty. Then, using C:
\begin{equation}\label{I2}
\text{If F finds $\uparrow$, then she knows that W will find \emph{f} with certainty.}
\end{equation}
\item According to $\overline{\text{W}}$, after step 2, the system $\overline{\text{L}}$+L is in the state
$
\sqrt{\frac{1}{3}}\ket{h}_{\overline{\text{L}}}\ket{\downarrow}_{\text{L}}+ \sqrt{\frac{2}{3}}\ket{t}_{\overline{\text{L}}}\ket{\rightarrow}_{\text{L}} .
$
But that state is orthogonal to $\ket{o}_{\overline{\text{L}}}\ket{\downarrow}_{\text{L}}$. As a result:
\begin{equation}\label{Ix}
\text{If $\overline{\text{W}}$ finds \emph{o}, then she knows with certainty that F found $\uparrow$.}
\end{equation}
So, employing (\ref{I2}), (\ref{Ix}) and C:
\begin{equation}\label{I3}
\text{If $\overline{\text{W}}$ finds \emph{o}, then she knows that W will find \emph{f} with certainty.}
\end{equation}
\item Finally, using (\ref{I3}) and C:
\begin{equation}\label{I4}
\text{If $\overline{\text{W}}$ announces she got \emph{o}, then W knows she will find \emph{f} with certainty.}
\end{equation}
\end{enumerate}
The problem is that, at the same time, according to W, the state of L+$\overline{\text{L}}$, after step 2, is given by
$
\sqrt{\frac{1}{3}}\ket{h}_{\overline{\text{L}}}\ket{\downarrow}_{\text{L}}+ \sqrt{\frac{2}{3}}\ket{t}_{\overline{\text{L}}}\ket{\rightarrow}_{\text{L}} ,
$
which can also be written as
\begin{equation}
\frac{1}{2\sqrt{3}} \left[ \ket{o}_{\overline{\text{L}}}\ket{o}_{\text{L}} - \ket{o}_{\overline{\text{L}}} \ket{f}_{\text{L}}+\ket{f}_{\overline{\text{L}}} \ket{o}_{\text{L}} \right] + \frac{\sqrt{3}}{2}\ket{f}_{\overline{\text{L}}} \ket{f}_{\text{L}}.
\end{equation}
Given that $\ket{o}_{\overline{\text{L}}}\ket{o}_{\text{L}}$ has a non-zero coefficient, W concludes that it is possible for her to get the result \emph{o}, even if $\overline{\text{W}}$ also gets \emph{o}. But this is inconsistent with (\ref{I4}).
From all this, it is concluded that the conjunction of S, C and Q is inconsistent, which means that (at least) one of them, must be given up.
What to make of this argument? We start by noting that it is not presented in the same form as the previous ones. To begin with, it involves one of the friends using the result of a measurement to decide which state to send to the other. This complicated procedure, however, is easily seen to be equivalent to the friends sharing, as in the PM and ZM arguments, an entangled system. In this case, an entangled coin and electron, in the state
\begin{equation}
\frac{1}{\sqrt{3}} \left[ \ket{h}\ket{\downarrow} + \ket{t}\ket{\uparrow} + \ket{t}\ket{\downarrow} \right] ,
\end{equation}
which can be recognized as a particular instance of a Hardy pair (\cite{Hardy1}; \cite{Hardy2}). After that, both friends measure in the $\{h,t\}$ and $\{\uparrow,\downarrow\}$ bases, respectively and, finally, ``the Wigners'' measure in their respective $\{o,f\}$ bases. Put like this, the FR set-up can be seen to be the Hardy version of the PM and ZM set-ups. Note that, as in the ZM case, and unlike in the PM one, in this case ``the Wigners'' measure their whole labs, containing the system and the friend. However, we already saw that this is equivalent to the undoing of the measurements, followed by measurements of only the particles, considered in the PM argument.
Another important difference between the FR theorem and the PM and ZM arguments, is that the former is cast in terms of\emph{epistemic} elements, such agents expressing certainty on certain beliefs, agents announcing their results or agents changing their beliefs on hearing the results of others. The problem is that, such a description at an epistemic level, is not accompanied by a clear and explicit description of how, physically speaking, measurements are to be dealt with. For instance, in the PM and ZM arguments, it was made clear that the measurements of the friends were to involve purely unitary evolution, and to lead to well-defined results. In this case, in contrast, we are not directly informed about any of these issues. And, of course, the inherently vague nature of the the standard framework---its ambiguity as to what happens during \emph{measurements}---allows for a range of interpretations in these non-standard scenarios, so a detailed prescription is required. As a result, in order to dissect the FR argument, the rules actually employed are to be extracted from the reasoning presented.
Regarding the evolution during measurements, we note, for instance, that in order to derive (\ref{I1}), $\overline{\text{F}}$ must assume that W describes F's measurement in a purely unitary way. Similarly, to derive (\ref{Ix}), $ \overline{\text{W}}$ must describe the measurements of F and $\overline{\text{F}}$ in a unitary fashion. We conclude that, as in the PM and ZM arguments, the FR theorem (implicitly) assumes measurements to involve purely unitary evolution.
Regarding measurement results, even if it is not stated explicitly in \cite{FR}, all measurements are assumed to yield definite results. For instance, as we just saw, to derive (\ref{I1}), W describes F's measurement in a purely unitary way. However, in the next step, in order to derive (\ref{I2}), it is assumed that F's measurement yields a definite outcome.\footnote{The conjunction of S and C could be thought to lead to the same conclusion. However, the fact that they are written in a purely epistemic language (see the previous footnote), makes the connection insecure.} In sum, as in the PM and ZM arguments, the derivation of the FR theorem assumes, both, measurements to involve purely unitary evolution, and measurements to lead to definite outcomes.
After these observations, we are in position to evaluate the FR argument. The first comment is that, just as the arguments in section \ref{M}, the FR argument employs a Bell-type scenario---in this case Hardy's---and extends it with Wigner's friend arrangements to allow all possible measurement settings to be employed on a single run. The second comment is that, just as the arguments in section \ref{M}, the FR argument employs mixed correlations between the measurements of the Fs and the Ws, such as (\ref{I1}) and (\ref{Ix}), in order to derive a contradiction.
Now, are these mixed correlations legitimate? Of course not. For instance, to derive (\ref{I1}), it is assumed that F's measurement does not spoil the correlation between the results of $\overline{\text{F}}$ and W. Similarly, (\ref{Ix}) depends on assuming that $\overline{\text{F}}$'s measurement does not break the correlation between the results of F and $\overline{\text{W}}$. However, as we saw in detail in section \ref{MCA}, this is incompatible with the (implicit) assumptions behind the theorem. As a result, these correlations are simply invalid. The upshot, of course, is that the whole FR theorem is unsound.
As in the PM case, we can illustrate all this with explicit calculations within pilot-wave theory (see appendix \ref{PWM} for details). In particular, we can compute the full joint probability for all four results, which includes the fact that
\begin{equation}
P_{\overline{\text{F}}FW\overline{\text{W}}}(t,\downarrow,o,o)=\frac{1}{9} .
\end{equation}
But this clearly shows that, if one takes seriously the (implicit) assumptions behind the theorem, then both (\ref{I1}) and (\ref{Ix}) are simply not satisfied. This is because the fact that this probability is different from zero, directly implies that it is neither the case that $\overline{\text{F}}$ finding \emph{t} implies W finding \emph{f}, nor that $\overline{\text{W}}$ finding \emph{o} implies F finding $\uparrow$. And, of course, without these correlations, no contradiction arises.
As mentioned above, a lot has been written about the FR theorem. In particular, in works such as \cite{baumann}, \cite{sudbery1}, \cite{sudbery2}, \cite{Drezet}, \cite{lazarovici}, \cite{tausk} and \cite{WCF}, it has been pointed out that the FR argument depends upon a hidden assumption. There is, however, no consensus regarding the exact nature of such an assumption. For instance, it has been argued that the theorem rests on an assumption to the effect that, when a measurement is carried out inside of a closed lab, such a measurement leads to a collapse for inside observers, but it does not lead to a collapse for outside observers. While I take such observations to be on the right track, I believe that the characterization of the hidden assumption offered in this work, in terms of mixed correlations, manages to fully uncover the true nature of the required additional assumption.
Moreover, and more importantly, this work is the first to show that this hidden assumption, is not only needed for the arguments to run, but that it is, in fact, invalid. This, of course, makes for a crucial difference: \emph{the FR argument is not only found to be limited, it is found to be false}. Finally, by exploring the FR argument in conjunction with similar arguments, the present work is not only able to shed light on the FR result, but to reveal the underlying structure of the family of arguments employing extended Wigner's friend scenarios to challenge the objectivity of quantum mechanics. In doing so, it shows why such results fully fail in their mission.
\section{Conclusions}
\label{Co}
Recent arguments, involving entangled systems shared by sets of Wigner’s friend arrangements, aim at calling into question the objectivity of measurement results. In particular, they purportedly show that the assumption that quantum measurements yield definite, objective outcomes, is incompatible with quantum predictions.
In this work, I show that such arguments depend upon a mistaken assumption---regarding the mixed correlations between the results of the friends and those of ``the Wigners''---which leads to invalid predictions. It is not, then, that the assumption of definite outcomes leads to trouble, but that the results derived with such an assumption are contrasted with faulty predictions. I also explore the more famous no-go theorem by Frauchiger and Renner, on which these recent arguments are motivated. I argue that it depends upon the same mistaken assumption about mixed correlations. I conclude that none of these results is able to impose interesting constraints on the assumption that measurements lead to objective outcomes.
\begin{appendices}
\section{Pilot-wave analysis of the Pusey-Masanes and Frauchiger-Renner experiments}
\label{PWM}
In this appendix, I offer a detailed analysis of the Pusey-Masanes (PM) and the Frauchiger-Renner (FR) experimental settings, in the context of the de Broglie-Bohm pilot-wave theory \citep{Bohm}. For this, I rely heavily on the pilot-wave treatment of spin presented in \cite{Nor2}.
Pilot-wave theory is the best-developed example of a hidden-variable model. The framework proposes that a complete characterization of a $N$-particle system is given by its wave function $ \Psi (x, t) $, \emph{together} with the actual positions of the particles \linebreak $\{\mathbf{X}_1 (t),\mathbf{X}_2 (t),\dots,\mathbf{X}_N (t)\}$, which are taken to always possess well-defined values. The wave function is postulated to satisfy at all times the usual Schr\"odinger equation
\begin{equation}
i \hbar \frac{\partial \Psi}{ \partial t } = \hat{H} \Psi ,
\end{equation}
and the positions to evolve according to the deterministic ``guiding'' equation
\begin{equation}
\frac{d \mathbf{X}_k(t)}{dt} = \left. \frac{\sum_\nu \mathbf{j}_{k,\nu} (\mathbf{x}_1,\mathbf{x}_2,\dots,\mathbf{x}_N,t)}{\sum_\nu\rho_\nu(\mathbf{x}_1,\mathbf{x}_2,\dots,\mathbf{x}_N,t)} \right\rvert_{\mathbf{x}=\mathbf{X}(t)},
\label{guide}
\end{equation}
with $\mathbf{j}_{k,\nu} = (\hbar/2mi) (\Psi_\nu^* \nabla_k \Psi_\nu - \Psi_\nu \nabla_k \Psi_\nu^*) $, $\rho_\nu = \Psi_\nu^* \Psi_\nu$ and $\nu$ a spin index, if available. Finally, the initial particle positions are assumed to be random, with probability distribution at $t=0$ given by
\begin{equation}
\label{SP}
P (\mathbf{X} = \mathbf{x}) = \sum_\nu | \Psi_\nu (\mathbf{x}) |^2.
\end{equation}
Before analyzing the full PM and FR settings, we consider a single spin-$\frac{1}{2}$ particle, traveling along the $y$ axis, which is sent through a Stern-Gerlach apparatus that produces a magnetic field $\mathbf{B} \approx b z \hat{z}$. Following \cite{Nor2}, we assume that the magnetic field is very strong, but is non-zero only in an extremely small region around $y=0$. This allows for the standard Stern-Gerlach interaction Hamiltonian to be well-approximated by $\hat{H} = - \mu b z \delta(y) \sigma_z$.
We assume the particle's initial wave function to be given by what \cite{Nor1} calls a ``plane-wave packet''. Such an incident wave packet is assumed to be nearly constant over a tubular region of length $L$ along the $y$ direction and radius $r$ centered at the origin of the $x-z$ plane, and to be zero outside such a region. Moreover, the particle is assumed to initially propagate along the $+y$-direction, with a reasonably sharply defined wave number $k$, such that $\lambda = 2 \pi/k \ll L$. This guarantees that, to a very good approximation, the packet will maintain its shape and travel at the appropriate group velocity.
As shown in \cite{Nor2}, by assuming that the impulse imparted to the particle by the apparatus, $\kappa$, is small compared to $k$, and that the width $2 r$ of the wave packet is small compared to $1/\kappa$, one can write
\begin{equation}
\Psi(x,y,z) \propto
\begin{cases}
e^{i k y} (c_+ \chi_+ + c_- \chi_- ) & y<0 \\
e^{i k' y} (e^{+i \kappa y} c_+ \chi_+ + e^{-i \kappa y} c_- \chi_- ) & y>0
\end{cases}
\end{equation}
with $k' \approx k$, $\chi_\pm$ eigenstates of $S_z$ and $c_\pm$ the weights of such eigenstates in the initial state. Of course, the above expression for $y<0$ is valid only where the wave function has support and, for $y>0$, in the region where the two separating components of the incident wave function overlap.
This wavefunction, together with the guiding equation, can be used to calculate the particle trajectories. For $y<0$ one obtains
\begin{equation}
\label{G1}
\frac{d \mathbf{X}(t)}{dt} = \frac{\hbar k}{m} \hat{y},
\end{equation}
which means that, no matter where inside the incident packet the particle is located, it will simply travel with the packet's group velocity. For $y>0$ one obtains
\begin{equation}
\label{G2}
\frac{d \mathbf{X}(t)}{dt} = \frac{\hbar k'}{m} \hat{y} + \frac{\hbar \kappa}{m} (|c_+|^2 +|c_-|^2) \hat{z},
\end{equation}
which amounts to a velocity in the $z$-direction equal to the weighted average of the group velocities of the two components of the wave function emerging out of the Stern-Gerlach apparatus.
Eq. (\ref{SP}), together with the fact that the initial wave function is assumed to be constant where it does not vanish, implies that the particle is equally likely to be located at any point inside the packet. Moreover, the pilot-wave dynamics is easily shown to preserve the initial distribution of Eq. (\ref{SP}), which means that the probability for the particle to end up in the upper (lower) path, and thus to be assigned spin-up (spin-down) when detected, is given by $c_+(c_-)$. To demonstrate this explicitly, one can analyze in detail what regions of the cross-section of the incoming packet result in the particle deflecting up or down; the result of such an analysis is given in Figure 2 (see \cite{Nor2} for details).
\begin{figure
\centering
\includegraphics[height=8cm]{fig2.jpg}
\caption{Regions of the cross-section of the incoming packet resulting in the particle deflecting up or down, upon encountering a Stern-Gerlach apparatus oriented along $z$.}
\end{figure}
At this point, it is important to point out that, according to the pilot-wave perspective, a particle could be assigned an objective, definite value of spin, according to the path through which it is deflected, independently of whether it is detected, observed or measured to be there. That is, the particle can be assigned a spin value, even if there is no process to correlate its position, after it passes the Stern-Gerlach device, with some sort of macroscopic measuring apparatus. Moreover, it is clear that, after sending a particle through a Stern-Gerlach device, one could use reverse magnets to deflect the two separating components and recombine them to reconstitute the original beam. Even in that case, according to the pilot-wave perspective, a definite result can be associated with the measurement. Finally, it is clear from Eqs. (\ref{G1}) and (\ref{G2}) that, to the level of the approximations employed, if the recombination is performed, after it, the $(x,z)$ position of the particle inside the wave packet will be exactly the position it had before it went through the apparatus.
The last element we need to review, before exploring the PM and FR settings, is the pilot-wave treatment of pairs of entangled particles. The first point to mention is that, as any other empirically viable hidden-variable theory, pilot-wave is non-local and contextual. In particular, such features imply that the results of an experiment may depend on what other measurements are being performed elsewhere. Moreover, the explicit non-locality of the framework requires an objective notion of simultaneity, which allows for the order of space-like separated events to be relevant for the calculation of the results of an experiment.
Consider the two spin-$\frac{1}{2}$ particles of a singlet, propagating from the origin along the $y$ axis, in opposite directions. Assume that the particles' initial wave functions are the plane wave packets described above. Suppose, moreover, two Stern-Gerlach apparatuses positioned on the $y$ axis, along the paths of the two particles. As explained in \cite{Nor2}, if particle 1 encounters its Stern-Gerlach apparatus first, then, whether it is deflected up or down (relative to the Stern-Gerlach orientation), is fully determined by the position of particle 1: if it is in the upper half of the wave packet, it will deflect up and, if it is in the lower half of the wave packet, it will deflect down. What happens when particle 2 is then measured? It can be shown that the behavior of particle 2, when measured, depends on where particle 1 ended up. In particular, if particle 1 is found with a certain spin, then particle 2 will effectively be guided by a wave function that corresponds to a wave function that ``collapsed'' according to the result obtained for particle 1. For instance, if both measurements are along the same axis, and if particle 1 is deflected up, then particle 2 will be deflected down, independently of its position---that is, even if its position is such that, if measured first, it would have deflected up.
We are finally in position to analyze the experiments; we start with the PM case. We recall that it consists of the two spin-$\frac{1}{2}$ particles of a singlet being sent to two spatially separated laboratories, where observers $\mathcal{A}_1$ and $\mathcal{B}_1$ perform spin measurements along directions $a_1$ and $b_1$, respectively. Next, observers $\mathcal{A}_2$ and $\mathcal{B}_2$ \emph{undo} these initial measurements and perform measurements along directions $a_2$ and $b_2$, respectively. From the pilot-wave perspective, this experiment can be carried out without actually involving observers who perform and undo measurements. As I mentioned above, according to pilot-wave theory, when a particle goes through a Stern-Gerlach device, it can be though of as yielding a definite result for its spin, independently of whether the result is amplified or observed. Moreover, by recombining the beams, one achieves what is envisioned when the PM argument talks about the measurements being undone.
Therefore, from the pilot-wave perspective, the PM experiment amounts to sending the two particles of a singlet, $a$ and $b$, in opposite directions, measuring them along directions $a_1$ and $b_1$, respectively, recombining the beams, and measuring them along directions $a_2$ and $b_2$, respectively. We mentioned above, though, that the order of operations, even if remote, is relevant for the calculation of results. Therefore, in order to make concrete predictions, we must specify a particular order in which the operations are performed. For concreteness, we assume the following order: $t_1$) particle $a$ is measured along $a_1$; $t_2$) particle $b$ is measured along $b_1$; $t_3$) particle $b$ is recombined; $t_4$) particle $a$ is recombined; $t_5$) particle $b$ is measured along $b_2$ and $t_6$) particle $a$ is measured along $a_2$ (see Figure 3).
\begin{figure
\centering
\includegraphics[height=9cm]{fig1.jpg}
\caption{Schematic representation of the Pusey-Masanes experiment, with the particular order of operations used for the calculations.}
\end{figure}
Given what I have said so far, it can be shown that the possible results of the measurements depend on the initial positions of the particles according to Figure 4. Moreover, the sizes of the areas, corresponding to the different possible results, are given by
\begin{eqnarray}
a_{++} &=& \frac{r^2}{2} \left[ (\pi-\alpha) + \alpha |2 \omega_a^\pm-1| - \pi \theta(1-2 \omega_a^\pm) (1-2\omega_a^\pm) \right] \nonumber \\
a_{+-} &=& \frac{r^2}{2} \left[ \alpha - \alpha |2 \omega_a^\pm-1| + \pi \theta(1-2 \omega_a^\pm) (1-2\omega_a^\pm) \right] \nonumber \\
a_{-+} &=& \frac{r^2}{2} \left[ \alpha - \alpha |2 \omega_a^\pm-1| + \pi \theta(2 \omega_a^\pm-1) (2\omega_a^\pm-1) \right] \nonumber \\
a_{--} &=& \frac{r^2}{2} \left[ (\pi-\alpha) + \alpha |2 \omega_a^\pm-1| - \pi \theta(2 \omega_a^\pm-1) (2\omega_a^\pm-1) \right] \nonumber \\
b_{++} &=& \frac{r^2}{2} \left[ (\pi-\beta) + \beta |2 \omega_b^\pm-1| - \pi \theta(1-2 \omega_b^\pm) (1-2\omega_b^\pm) \right] \nonumber \\
b_{+-} &=& \frac{r^2}{2} \left[ \beta - \beta |2 \omega_b^\pm-1| + \pi \theta(2 \omega_b^\pm-1) (2\omega_b^\pm-1) \right] \nonumber \\
b_{-+} &=& \frac{r^2}{2} \left[ \beta - \beta |2 \omega_b^\pm-1| + \pi \theta(1-2 \omega_b^\pm) (1-2\omega_b^\pm) \right] \nonumber \\
b_{--} &=& \frac{r^2}{2} \left[ (\pi-\beta) + \beta |2 \omega_b^\pm-1| - \pi \theta(2 \omega_b^\pm-1) (2\omega_b^\pm-1) \right] ,
\end{eqnarray}
where $\theta (x,y)$ is the two-dimensional Heaviside step function, which is 1 only when both $x$ and $y$ are positive, $\alpha $ the angle between $a_1$ and $a_2$, $\beta $ the angle between $b_1$ and $b_2$, and with $\omega_a^\pm $ and $\omega_b^\pm $ determined by the results of the measurements along $b_2$ and $a_1$, respectively, according to
\begin{eqnarray}
\omega_a^+ &=& \sin^2(\frac{a_2-b_2}{2}) \nonumber \\
\omega_a^- &=& \cos^2(\frac{a_2-b_2}{2}) \nonumber \\
\omega_b^+ &=& \sin^2(\frac{a_1-b_1}{2}) \nonumber \\
\omega_b^- &=& \cos^2(\frac{a_1-b_1}{2}) .
\end{eqnarray}
\begin{figure
\centering
\includegraphics[height=9cm]{fig3.jpg}
\caption{Regions of the cross-sections of the incoming packets, resulting in the particles deflecting up or down in their pair of measurements.}
\end{figure}
From all this, one can compute the joint distribution for all four measurements, then compute the marginals for pairs of results and, finally, the expectation values. The results are given by
\begin{eqnarray}
E^{p}_{\mathcal{A}_1\mathcal{B}_1} &=& -\cos(a_1-b_1) \nonumber \\
E^{p}_{\mathcal{A}_2\mathcal{B}_2} &=& -\cos(a_2-b_2) \nonumber \\
E^{p}_{\mathcal{A}_1\mathcal{B}_2} &=& 0 \nonumber \\
E^{p}_{\mathcal{B}_1\mathcal{A}_2} &=& -\left(1- \frac{2 \alpha}{\pi} \right)\cos(a_1-b_1) - \left(1-\frac{2 \beta}{\pi} \right) \cos(a_2-b_2) \nonumber \\
& & - \frac{2}{\pi} \left\lbrace \alpha \cos(a_1-b_1) |\cos(a_2-b_2)| + \beta |\cos(a_1-b_1)| \cos(a_2-b_2) \right\rbrace \nonumber \\
& & + 2 \cos(a_1-b_1) \cos(a_2-b_2) \theta(\cos(a_1-b_1), \cos(a_2-b_2)) \nonumber \\
& & - 2 \cos(a_1-b_1) \cos(a_2-b_2) \theta(-\cos(a_1-b_1), -\cos(a_2-b_2)) ,
\end{eqnarray}
where the $p$ superscript indicates that these are the results, according to pilot-wave theory (given the particular order of operations chosen).
Regarding these results, we note that the first two correlation coefficients, exactly match the standard quantum prediction, which makes perfect sense. Such correlations can be calculated with the standard framework, and we know that, in such a case, the pilot-wave predictions reproduce the standard predictions. The third value, the one that is zero, also makes perfect sense. To begin with, the $a_1$ result is fully determined by the initial position of particle $a$. Moreover, by the time the $b_2$ measurement occurs, both the $a_1$ and $b_1$ measurements have been completely undone. Therefore, the $b_2$ measurement happens on a ``fresh'' singlet state, and its outcome is fully determined by the initial position of particle $b$. Finally, the initial positions of particles $a$ and $b$ are uncorrelated, from which it follows that the $a_1$ and $b_2$ results are uncorrelated. Regarding the fourth value, while it is relatively easy to see that such a correlation is neither zero, nor equal to the standard value, apart from the actual calculations, I do not have any insights regarding the actual value obtained.
In sum, we see that the expectation values for the first and second sets of measurements are, as expected, proportional to the cosine of the angle between the measurement axes. We see, however, that the mixed correlations depart from such a behaviour. This, of course, is a necessary consequence of Fine's theorem and, indeed, it can be checked that the first four values above, satisfy the CHSH inequality.
The correlations between the first and second measurements on each particle are given by
\begin{eqnarray}
E^{p}_{\mathcal{A}_1\mathcal{A}_2} &=& \left(1-\frac{2 \alpha}{\pi} \right) \left( 1-|\cos(b_1-a_1)| \right) \nonumber \\
E^{p}_{\mathcal{B}_1\mathcal{B}_2} &=& \left(1-\frac{2 \beta}{\pi} \right) \left( 1-|\cos(b_2-a_2)| \right) .
\end{eqnarray}
We see that these values explicitly show the non-locality and contextuality of the framework, in the sense that the correlation between the measurements on one side can be modulated by the selection of the orientation of a measurement on the other.
Next we analyze the FR arrangement. It contains four agents, F, $\overline{\text{F}}$, W and $\overline{\text{W}}$, and two labs, L and $\overline{\text{L}}$, with a communication channel between them. The experiment then proceeds as follows. First, Inside $\overline{\text{L}}$, $\overline{\text{F}}$ prepares a quantum coin in the state $\sqrt{\frac{1}{3}}\ket{h}+ \sqrt{\frac{2}{3}}\ket{t}$ and measures it. If she finds \emph{h}, she sends F an electron in the state $\ket{\downarrow}$, if she finds \emph{t}, she sends it in the state $\ket{\rightarrow}= \frac{1}{\sqrt{2}} \left[ \ket{\uparrow}+\ket{\downarrow} \right]$. Then, inside L, F measures the electron in the $\{\ket{\uparrow},\ket{\downarrow}\}$ basis. Next, $\overline{\text{W}}$ measures $\overline{\text{L}}$ in the basis
$
\{ \ket{o}_{\overline{\text{L}}}=\frac{1}{\sqrt{2}}\left[\ket{h}_{\overline{\text{L}}}-\ket{t}_{\overline{\text{L}}}\right], \ket{f}_{\overline{\text{L}}}=\frac{1}{\sqrt{2}}\left[\ket{h}_{\overline{\text{L}}}+\ket{t}_{\overline{\text{L}}}\right] \}
$
(with $\ket{h}_{\overline{\text{L}}}$ and $\ket{t}_{\overline{\text{L}}}$ the states of $\overline{\text{L}}$ after $\overline{\text{F}}$ measures the coin and finds the corresponding result), and announces her result. Finally, W measures L in the basis
$
\{ \ket{o}_{\text{L}}=\frac{1}{\sqrt{2}}\left[\ket{\downarrow}_{\text{L}}-\ket{\uparrow}_{\text{L}}\right], \ket{f}_{\text{L}}=\frac{1}{\sqrt{2}}\left[\ket{\downarrow}_{\text{L}}+\ket{\uparrow}_{\text{L}}\right] \}
$
(with $\ket{\downarrow}_{\text{L}}$ and $\ket{\uparrow}_{\text{L}}$ the states of L after F measures the spin and finds the corresponding result).
Now, as I explained in section \ref{FR}, the procedure in which $\overline{\text{F}}$ measures the coin, and sends the electron to F, is equivalent to $\overline{\text{F}}$ and F sharing an entangled coin-electron pair in the state
\begin{equation}
\label{ce}
\frac{1}{\sqrt{3}} \left[ \ket{h}\ket{\downarrow} + \ket{t}\ket{\uparrow} + \ket{t}\ket{\downarrow} \right] .
\end{equation}
Moreover, as I also explained above, given the assumption that measurements are to be described by a purely unitary evolution, the measurements of the Ws on their respective labs are equivalent to the initial measurements by the Fs being \emph{undone}, as in the PM case, with new measurements being performed directly on the coin and electron.
Given all this, the FR experiment is equivalent to the following procedure. $\overline{\text{F}}$ and F share an entangled coin-electron pair in the state (\ref{ce}). Then, $\overline{\text{F}}$ measures the coin in the $\{\ket{h},\ket{t}\}$ basis and F measures the electron in the $\{\ket{\uparrow},\ket{\downarrow}\}$ basis. Next, the coin and electron measurements are undone. Finally, they are measured again, the coin in the $\{\ket{f}_c=\frac{1}{\sqrt{2}}\left[\ket{h}+\ket{t}\right],\ket{o}_c=\frac{1}{\sqrt{2}}\left[\ket{h}-\ket{t}\right]\}$ basis, and the electron in the $\{\ket{f}_e=\frac{1}{\sqrt{2}}\left[\ket{\downarrow}+\ket{\uparrow}\right],\ket{o}_e=\frac{1}{\sqrt{2}}\left[\ket{\downarrow}-\ket{\uparrow}\right]\}$ basis.
Now, if the coin is substituted by an electron, given the techniques described above, this procedure can easily be modeled within pilot-wave theory. If one does so, and chooses the time order of operations analogous to the one used in the PM case, the full joint probability distribution over the results can be calculated. These probabilities are given by
\begin{eqnarray}
P_{\overline{\text{F}}FW\overline{\text{W}}}(h,\downarrow,f,o) &=& 1/9 \nonumber \\
P_{\overline{\text{F}}FW\overline{\text{W}}}(h,\downarrow,o,f) &=& 1/9 \nonumber \\
P_{\overline{\text{F}}FW\overline{\text{W}}}(h,\downarrow,o,o) &=& 1/9 \nonumber \\
P_{\overline{\text{F}}FW\overline{\text{W}}}(t,\uparrow,f,o) &=& 1/9 \nonumber \\
P_{\overline{\text{F}}FW\overline{\text{W}}}(t,\uparrow,o,f) &=& 1/9 \nonumber \\
P_{\overline{\text{F}}FW\overline{\text{W}}}(t,\uparrow,o,o) &=& 1/9 \nonumber \\
P_{\overline{\text{F}}FW\overline{\text{W}}}(t,\downarrow,f,o) &=& 1/9 \nonumber \\
P_{\overline{\text{F}}FW\overline{\text{W}}}(t,\downarrow,o,f) &=& 1/9 \nonumber \\
P_{\overline{\text{F}}FW\overline{\text{W}}}(t,\downarrow,o,o) &=& 1/9 ,
\end{eqnarray}
with all other terms equal to zero.
\end{appendices}
\section*{Acknowledgments}
I thank Ricardo Muciño and Travis Norsen for fruitful conversations, and two anonymous referees for valuable comments. I acknowledge support from UNAM-PAPIIT (grant IN102219) and CONACYT (grant 140630).
\begin{flushright}
\emph{
Elias Okon\\
Institute for Philosophical Research\\
National Autonomous University of Mexico\\
Circuito Maestro Mario de la Cueva s/n\\
C.U., 04510, CDMX, México.\\
<EMAIL>
}
\end{flushright}
|
\section{Introduction}
\label{sec:intro}
New sources of CP violation can be a key ingredient to explain the matter-antimatter asymmetry of the universe~\cite{Sakharov:1967dj,Kajantie:1996mn,Huet:1994jb}. Hence, the quest for new CP violating interactions is a clear target for beyond the Standard Model (SM) searches, being a critical component of the physics program of the LHC. A particularly interesting option is that the Higgs boson couplings present these new physics sources.
From the theoretical point of view, some Higgs interactions are more inclined to display CP violation effects than others. While the widely studied beyond the SM CP structure for the Higgs to vector boson couplings are loop suppressed, arising only at dimension-6 or higher~\cite{Buchmuller:1985jz,Grzadkowski:2010es}, CP violation in Higgs to fermion interactions can manifest already at the tree-level~\cite{Buckley:2015vsa}, being naturally larger. Owning to its magnitude, the top quark Yukawa coupling can play a significant role in this context and be most sensitive to new physics.
Whereas it is possible to access the Higgs-top coupling through loop induced processes~\cite{Brod:2013cka,Dolan:2014upa,Englert:2012xt,Kobakhidze:2016mfx,Bernlochner:2018opw,Englert:2019xhk,Gritsan:2020pib,Bahl:2020wee}, the direct Higgs-top production via $pp\to t\bar{t}h$ is crucial to disentangle possible new physics effects~\cite{Ellis:2013yxa,Boudjema:2015nda,Buckley:2015vsa,Buckley:2015ctj,Gritsan:2016hjl,Goncalves:2016qhh,Mileo:2016mxg,AmorDosSantos:2017ayi,Azevedo:2017qiz,Li:2017dyz,Goncalves:2018agy,ATLAS:2018mme,CMS:2018uxb,Ren:2019xhp,Bortolato:2020zcg,Cao:2020hhb,MammenAbraham:2021ssc,Martini:2021uey,Goncalves:2021dcu}. This channel was observed in 2018 by both ATLAS and CMS with significances of 6.3~$\sigma$ and 5.2~$\sigma$, respectively~\cite{ATLAS:2018mme,CMS:2018uxb}. The high-luminosity LHC (HL-LHC) studies indicate that the Higgs-top interaction will be probed to outstanding accuracy at the end of the LHC run, reaching $\delta y_t\lesssim 4\%$ when combining the HL-LHC ATLAS and CMS data~\cite{Cepeda:2019klc}. The same projections indicate that the $t\bar{t}h$ channel in the $h\to \gamma\gamma$ final state will display dominant sensitivities. While the di-photon final state presents limited statistics, it highly benefits from controlled backgrounds from side-bands. Recently, ATLAS and CMS have reported the first experimental Higgs-top CP studies, exploring the $t\bar{t}h$ channel~\cite{ATLAS:2020ior,CMS:2020cga}. Both analyses focus on the di-photon final state, $h\to \gamma\gamma$. ATLAS and CMS exclude Higgs-top CP-mixing angles above $43^\circ$ and $55^\circ$ at 95\%~CL, respectively.
In the present manuscript, we perform a detailed investigation of the Higgs-top CP sensitivity with the $pp\to t\bar{t}h$ channel at the HL-LHC, considering the most promising decay mode, $h\to \gamma\gamma$. We explore the complex multiparticle final state with a combination of machine learning techniques and efficient kinematic reconstruction methods. Since distinct Higgs-top CP-phases affect the net top and anti-top quark polarization, propagating the spin effects to the top quark final states, we devote special attention to include the top polarization observables in our study. In particular, these spin effects are used to define genuine CP-observables. After motivating and constructing the relevant kinematic observables, we evaluate how much information can be extracted with them. The convenient metric adopted to quantify this is given by the Fisher information. We show that the ability of probing the $pp\to t\bar{t}h$ channel exploring the complex multiparticle final state not only in terms of a raw rate, but as a polarized process, can offer a crucial pathway to probe the underlying production dynamics, accessing possible new physics effects.
The structure of this paper is as follows. In Section~\ref{sec:theory}, we present the theoretical parametrization for the top Yukawa coupling. We discuss the new physics effects to the top polarization, define the CP-sensitive observables, and quantify how much information on the CP-phase can be extracted using distinct observables. In Section~\ref{sec:kin_reconstruct}, we present the kinematic reconstruction methods, which are relevant to build prominent observables to new physics and maximally explore the $t\bar{t}h$ final state. Next, in Section~\ref{sec:analysis}, we move on to the detailed analysis, where we derive the projected sensitivities for the Higgs-top CP-phase at the HL-LHC. This study is inclusive in respect to the top pair final states, combining the leptonic, semi-leptonic, and hadronic channels. Finally, a summary of our key findings is delivered in Section~\ref{sec:summary}.
\section{CP Structure and Observables}
\label{sec:theory}
We parametrize the top quark Yukawa coupling with the following Lagrangian
\begin{equation} \begin{aligned}
\mathcal{L} = - \frac{m_{t}}{v} \kappa_t \bar{t}(\cos\alpha + i \gamma_5 \sin\alpha) t h\, ,
\label{eq:Higgs-top}
\end{aligned} \end{equation}
where $m_{t}$ is the mass of the top quark, $v$ is the vacuum expectation value in the SM~($v=246$~GeV), $\kappa_{t}$ is a real number, and $\alpha$ is the CP-phase. The interaction between the CP-even Higgs boson and the top quark in the SM is represented by $(\kappa_{t},\alpha)=(1,0)$, while $\alpha=\pi/2$ results in a pure CP-odd Higgs-top interaction. New physics contributions in \cref{eq:Higgs-top} will display effects both in the Higgs $t\bar{t}h$ production and decay, $h\rightarrow \gamma\gamma$. Whereas the Higgs decay will more relevantly change the total signal rate~\cite{Ellis:2013yxa}, we will devote special attention to probe the new physics effects in the Higgs production, exploring the top quarks' final state kinematics. This will be an essential ingredient to uplift the new physics sensitivity from CP-phase effects.
\begin{figure*}[!htb]
\centering
\includegraphics[height=4.55cm,width=5.8cm]{ptgaga_parton_level.pdf}
\includegraphics[height=4.55cm,width=5.8cm]{mtt_parton_level.pdf}
\includegraphics[height=4.55cm,width=5.8cm]{b4_parton_level.pdf} \\
\includegraphics[height=4.55cm,width=5.8cm]{ocp4_parton_level.pdf}
\includegraphics[height=4.55cm,width=5.8cm]{delphi_ll_parton_level.pdf}
\includegraphics[height=4.55cm,width=5.8cm]{delphi_ljsoft_parton_level.pdf}
\caption{Top panels: Distributions for the transverse momentum for the Higgs boson $p_{Th}$ (left), invariant mass for the top pair $m_{t\bar{t}}$ (center), and the product of the projections of the top and anti-top momentum $b_{4} = p_{t}^{z} p_{\bar{t}}^{z} / p_{t} p_{\bar{t}}$ (right). Bottom panels: Distributions for the Collins-Soper angle $\theta^{*}$ (left), the azimuthal angle between the two charged leptons in the top pair rest frame $\Delta\phi_{\ell \ell}^{t\bar{t}}$ for fully leptonic $t\bar{t}h$ events (center), and the same angle between the charged lepton and the softest light jet in the top rest frame $\Delta\phi_{\ell j_\text{soft}}^{t\bar{t}}$ for semi-leptonic $t\bar{t}h$ events (right). Each panel shows parton level results for the $t\bar{t}h$ sample for the SM Higgs ($\alpha = 0$), a CP-odd Higgs ($\alpha = \pi/2$) and mixed hypotheses ($\alpha = \pm\pi/4$). We also present the ratio between new physics and SM scenarios on the bottom panel of each figure. The results are presented for the 14~TeV LHC.}
\label{fig:semilep_parton_dist}
\end{figure*}
\subsection{Top Polarization}
Among the observables sensitive to the structure of the top quark Yukawa interaction, the spin correlations between the top and anti-top in $t\bar{t}h$ associated production offer a prominent pathway for precision studies~\cite{Bar-Shalom:1995quw,Gunion:1996xu, Atwood:2000tu,Berge:2008wi, Ellis:2013yxa, Buckley:2015vsa, Boudjema:2015nda, Gritsan:2016hjl, Buckley:2015ctj,Goncalves:2016qhh,Mileo:2016mxg, Goncalves:2018agy, Aad:2020ivc,Goncalves:2021dcu,Degrande:2021zpv}. Owing to its short lifetime ($\sim 10^{-25}$s)~\cite{ParticleDataGroup:2020ssz}, the top quark is expected to decay before hadronization occurs ($\sim 10^{-24}$s) and spin decorrelation effects take place ($\sim 10^{-21}$s)~\cite{Mahlon:2010gw}. Thus, the spin-spin correlations between $t$ and $\bar{t}$ can be traced back from the top quark decay products. In particular, it is possible to observe correlations between any two decay products, one from the top quark and the other from the anti-top quark. The correlations scale with the spin analyzing power associated with each top decay product.
More accurately, the top quark final states in the leptonic $t\to W^+ b\to \ell^+ \nu b$ and hadronic $t\to W^+ b\to \bar{d} u b$ channels are correlated with the top quark spin axis as follows:
\begin{equation} \begin{aligned}
\frac{1}{\Gamma}\frac{d\Gamma}{d\cos\xi_i}=\frac{1}{2}\left( 1+ \beta_i P_t\cos\xi_i \right)\,,
\end{aligned} \end{equation}
where $\Gamma$ is the partial decay width, $\xi_i$ is the angle between the $i$-th decay product and the top quark spin axis in the top quark rest frame, $P_t$ is the polarization of the decaying top $(-1\le P_t\le 1)$, and $\beta_i$ is the spin analyzing power of the final state particle $i$~\cite{Bernreuther:2010ny}. At leading order, the coefficient $\beta_i$ is $+1$ for charged lepton $\ell^+$ and $\bar{d}$-quark, $-0.3$ for $\bar{\nu}$ and $u$-quark, $-0.4$ for the $b$-quark, and $0.4$ for $W$-boson. The sign of the coefficient $\beta_i$ is flipped for anti-top decays.
Granted by the $V\!-\!A$ current structure of the weak interaction, the charged lepton will be a prominent spin analyzer, favoring studies with di-leptonic top pairs. Exploring this phenomenology, the $\Delta \phi_{\ell\ell}^\text{lab}$ observable, which is the azimuthal angle difference between the two charged leptons in the lab frame, is a good example of probe that has been found effective in accessing the Higgs-top CP-properties~\cite{Buckley:2015vsa,Goncalves:2018agy}. Remarkably, the sensitivity of $\Delta \phi_{\ell\ell}^\text{lab}$ improves further in the boosted Higgs regime due to the change in the net polarization for the top-pair at high energies.
Analogously to the charged lepton, the $d$-quark also presents maximal spin analyzing power. However, it is a challenging task to tag a $d$-quark jet in a collider environment. An efficient solution is to select the softest of the two light-quark jets, $j_\text{soft}$, in the top quark rest frame. This choice uplifts the spin analyzing power of $j_\text{soft}$ to 50\% of the lepton's~\cite{Jezabek:1994qs}. This approach boosts the spin correlation analyses for the semi-leptonic and hadronic top quark pairs. Several observables can be defined exploiting this fact, a particularly relevant example, that we will explore in this manuscript, is the azimuthal angle difference between the charged lepton and softest light jet in the top pair rest frame,~$\Delta\phi_{\ell j_\text{soft}}^{t\bar{t}}$.
\subsection{CP-sensitive Observables}
Various kinematic observables have been studied in the literature to access the Higgs-top CP structure in the $pp\to t\bar{t}h$ channel. Some illustrative distributions are presented in \cref{fig:semilep_parton_dist}, such as the transverse momentum for the Higgs boson $p_{Th}$ (top left)~\cite{Demartin:2014fia,Demartin:2015uha}, the invariant mass for the top pair $m_{tt}$ (top center), the product of projections of top and anti-top momentum $b_{4} = p_{t}^{z} p_{\bar{t}}^{z} / p_{t}p_{\bar{t}}$ (top right)~\cite{Gunion:1996xu}, and the angle between the top quark and the beam direction in the $t\bar{t}$ CM frame $\theta^{*}$ which is also known as Collins-Soper angle (bottom left)~\cite{Goncalves:2018agy}. These observables result in distinct profiles for different Higgs-top CP-phases. The pure CP-odd phase, $\alpha=\pm\pi/2$, leads to a shift to higher energies in the peak of the distributions compared to the SM scenario, $\alpha=0$. Different CP-phases interpolate between these two profiles without sensitivity for the sign of the phase.
The variables $p_{Th}$, $m_{t\bar{t}}$, $b_{4}$, and $\theta^{*}$ are CP-even observables, being sensitive to the squared terms: $\cos^{2}\alpha$ and $\sin^{2}\alpha$. Thus, these probes are indifferent to the CP-even and CP-odd Higgs-top interference terms, which are proportional to $\cos\alpha \sin\alpha$. In particular, they are not sensitive to variations from a relative sign-difference in the CP-phase. Genuine CP-sensitive observables can be constructed from antisymmetric tensor products that require four linearly independent four-momenta. Owing to the top polarization being carried out to the decays, it is possible to construct such observable using, for instance, the top, anti-top and their decay products~\cite{Boudjema:2015nda, Mileo:2016mxg, Goncalves:2018agy}. In general, the antisymmetric tensor product can be expressed as
\begin{equation} \begin{aligned}
\epsilon(p_t,p_{\bar{t}},p_{i},p_{k}) \equiv \epsilon_{\mu\nu\rho\sigma}p_{t}^{\mu}p_{\bar{t}}^{\nu}p_{i}^{\rho}p_{k}^{\sigma}\,,
\label{eq:eps_tensor_prod}
\end{aligned} \end{equation}
where $\epsilon_{0123} = 1$, and $\{i,k\}$ represent the final state particles produced from the top and the anti-top decays, respectively.
In the $t\bar{t}$ CM frame, \cref{eq:eps_tensor_prod} can be fortuitously simplified to $p_{t}\cdot(p_{i} \times p_{k})$. This mathematical relation can be used to define azimuthal angle differences between the decay products, in the $t\bar{t}$ CM frame, that are odd under CP-transformations~\cite{Goncalves:2018agy}:
\begin{equation} \begin{aligned}
\!\!\! \Delta \phi_{ik}^{t\bar{t}} \!=\!
\text{sgn} \left[\vec{p}_{t} \!\cdot\! (\vec{p}_{i} \!\times\! \vec{p}_{k})\right]
\arccos \!\left[ \frac{\vec{p}_{t} \!\times\! \vec{p}_{i}}{|\vec{p}_{t} \!\times\! \vec{p}_{i}|} \!\cdot\! \frac{\vec{p}_{t} \!\times\! \vec{p}_{k}}{|\vec{p}_{t} \!\times\! \vec{p}_{k}|}\right]\!.
\label{eq:CP_odd_obseravable}
\end{aligned} \end{equation}
For illustration, we present in \cref{fig:semilep_parton_dist} the azimuthal angle between the two charged leptons $\Delta\phi_{\ell \ell}^{t\bar{t}}$ in the fully leptonic case (bottom center) and between the charged lepton and the softest light jet in the top rest frame $\Delta\phi_{\ell j_\text{soft}}^{t\bar{t}}$ in the semi-leptonic case (bottom right). Two comments are in order. First, we notice that $\Delta \phi_{ik}^{t\bar{t}}$ is indeed sensitive to the sign of the CP-phase, as illustrated in a comparison between the distribution profiles for $\alpha=\pi/4$ against $-\pi/4$. Second, in light of the spin analyzing power of the charged lepton in relation to $j_\text{soft}$, the relative CP-sensitivity of the di-leptonic against the semi-leptonic correlation follows our expectation. Namely, the beyond the SM effects in the $\Delta\phi_{\ell j_{soft}}^{t\bar t}$ observable are $\sim 50\%$ weaker in respect to $\Delta\phi_{\ell\ell}^{t\bar t}$. This can be observed by comparing the bottom panel of these plots, where we display the BSM/SM ratio.
\subsection{Observable Information}
\begin{figure*}[!htb]
\centering
\includegraphics[width=0.48\textwidth]{FigureA.pdf}
\includegraphics[width=0.48\textwidth]{FigureB.pdf}
\caption{Comparison of sensitivity on the CP-phase $\alpha$ arising from different observables in terms of the Fisher Information $I$ for CP-odd observables probing linear new physics effects (left) and in terms of the modified Fisher Information $I'$ for CP-even observables probing the non-linear new physics effects (right).}
\label{fig:information}
\end{figure*}
Before proceeding to a full analysis, let us pause for a moment to better understand which distributions and channels are sensitive to the CP-phase $\alpha$. In particular, we would like to quantify and compare how much information on the CP-phase is available using the different observables in a parton level setup. This will provide some benchmarks and highlight the main ingredients required for an efficient analysis strategy that will be presented in Sec.~\ref{sec:analysis}.
Let us first consider the spin correlation observables $\Delta \phi_{ik}^{t\bar t}$ between two decay products from the top and anti-top, which probe the new physics effects linear in $\alpha$. A convenient metric to quantify the sensitivity of these observables to constrain the parameters of our model is given by the Fisher information~\cite{Brehmer:2016nyr, Brehmer:2017lrt}. Its component describing the sensitivity to the CP-phase $\alpha$ is defined as
\begin{equation} \begin{aligned}
I= \mathbb{E} \left[
\frac{\partial \log p(x|\kappa_t, \alpha)}{d \alpha}
\frac{\partial \log p(x|\kappa_t, \alpha)}{d \alpha}
\right] \ .
\label{eq:info}
\end{aligned} \end{equation}
Here, $p(x|\kappa_t, \alpha)$ is the likelihood function, which describes the probability to observe a set of events with corresponding observables $x$ as a function of the model parameter $\kappa_t$ and $\alpha$. $\mathbb{E} [\cdot]$ denotes the expectation value evaluated at the SM point, $(\kappa_t,\alpha)_{\rm SM}=(1,0)$. In the following we use the \texttt{MadMiner} package to calculate the Fisher information~\cite{Brehmer:2019xox}.
In the left panel of \cref{fig:information}, we show the Fisher information associated with the CP-sensitive spin correlation observables for the di-leptonic (red), semi-leptonic (gray), and hadronic (blue) channels. The bars on the left show the \textit{full} information, \emph{i.e.}, the information that could be accessed via a comprehensive multivariate analysis. This was estimated using the machine learning method based on the \texttt{SALLY} algorithm~\cite{Brehmer:2018eca, Brehmer:2018kdj, Brehmer:2018hga} trained with all possible spin correlation observables. The remaining bars show the information in individual observables $\Delta \phi_{ik}^{t\bar t}$, which were estimated using a histogram based approach.
Focusing first on the di-leptonic channel, the most sensitive among these observables is the spin correlation between the leptons, $\Delta\phi^{t\bar t}_{\ell\ell}$, since the spin analyzing power for the charged leptons are maximal. The next most sensitive observables are those where a charged lepton has been replaced with a $b$-jet or a $W$ boson. We observe that the corresponding Fisher information in $\Delta\phi^{t\bar t}_{\ell b}$ and $\Delta\phi^{t\bar t}_{\ell W}$ are suppressed compared to $\Delta\phi^{t\bar t}_{\ell\ell}$ by the square of the spin analyzing power $\beta_{b/W}^{2} \sim 0.4^{2}$, as expected. The information in the spin correlations observables between a pair of $b$-jet(s) and/or $W$ boson(s) is further suppressed by an additional factor of $\beta_{b/W}^{2}$.
Let us now also consider the other top decay channels. As the Fisher information is proportional to the rate~\cite{Brehmer:2016nyr}, we expect it to increase relative to the fully leptonic channel by a factor ${2\times\text{BR}_{W \!\to\text{had}}/\text{BR}_{W \!\to\text{lep}} \sim 6}$ for the semi-leptonic channel and $(\text{BR}_{W \!\to\text{had}} /\text{BR}_{W \!\to\text{lep}})^2\sim 9$ for the hadronic channel. Looking at the last three observables involving $b$-jets and $W$-bosons, this is indeed the case. For the other observables, we notice an additional loss of about a factor 2 in spin analyzing power, and hence a factor 4 in the Fisher information, which is caused by probing $j_\text{soft}$ instead of the $d$-quark.
Overall, we see that the different observables have distinct overall importance in the three channels. For di-leptonic top decays, most of the information is contained in the spin correlation between the leptons, while the information in other observables is significantly suppressed. In contrast, for the hadronic decay channel, all shown observables have almost similar information. In this case, the resulting full information, that can be obtained by combining the different spin correlation observables, significantly exceeds the information of any individual observable. Overall, all three channels carry a similar amount of information on the CP-phase $\alpha$, which suggest performing a combined analysis. \medskip
Due to the limited $t\bar{t}h$ event rate at the LHC, we expect the non-linear new physics effects to dominate over the linear ones. We therefore expect most of the sensitivity on the CP-structure of the top Yukawa coupling to arise from these non-linear terms, despite the fact that the corresponding observables are not genuine CP-sensitive. To quantify the sensitivity of these CP-even observables to the squared terms, we use modified version of the Fisher information that was introduced in Ref.~\cite{Brehmer:2019gmn}. In this approach, we simply consider the square of the coupling as our new model parameter and define
\begin{equation} \begin{aligned}
I' = \mathbb{E} \left[
\frac{\partial \log p(x|\kappa_t^2, \alpha^2)}{d \alpha^2}
\frac{\partial \log p(x|\kappa_t^2, \alpha^2)}{d \alpha^2}
\right] \ .
\end{aligned} \end{equation}
The result is shown in the right panel of \cref{fig:information}. Here, we show the information associated with a two-dimensional distribution of two observables, relative to the full information associated with a multivariate analysis using all observables. As none of the presented observables relies on the top quark final state kinematics, the results are identical for all three top quark decay channels.
The distribution of the invariant mass of the photon pair, $m_{\gamma\gamma}$, is only sensitive to the theory parameters through its normalization. Correlating it with itself, we obtain the information associated with the signal strength measurements, which accounts for $31\%$ of the information on the CP-phase. In the absence of background, the correlation of $m_{\gamma\gamma}$ and any other observable is equivalent to the information in a single differential distribution of that observable. This is shown in the bottom row. As expected, it is also identical to the information for the correlation of an observable with itself, which are shown in the diagonal. We can identify $\Delta \eta_{t\bar t}$ and $\theta^*$ as the two most sensitive observables, which individually carry about $60\%$ of the full information.
Combing two different observables further increases the information. The two most promising combinations are $\Delta \eta_{t\bar t}$ vs. $p_{Th}$ as well as $\theta^*$ vs. $m_{t\bar t h }$, which carry about $73\%$ of the full information. Successively adding more observables further increases the information. This shows that a multivariate analysis is vital to maximize the sensitivity on the CP-phase $\alpha$.
\section{Kinematic Reconstruction}
\label{sec:kin_reconstruct}
Most of the new physics probes discussed so far, $viz.$ $m_{tt}$, $\theta^{*}$, $b_{4}$, and $\Delta \phi_{ik}^{t\bar{t}}$, require a full reconstruction of the top and anti-top momenta. This is a challenging task at the LHC due to combinatorial ambiguities and the presence of up to two neutrinos in the $t\bar{t}(h\to \gamma\gamma)$ final state. In this section, we discuss the strategies adopted for the kinematic reconstruction of the semi-leptonic and hadronic channels, and the more complex di-leptonic mode.
\medskip
\noindent\textbf{Semi-leptonic channel}: In the semi-leptonic channel, the full reconstruction of the $t\bar{t}$ system requires the determination of the longitudinal momentum of the missing particle $\nu$. We compute it by constraining the invariant mass of the lepton and the neutrino to the $W$-boson mass. Typically, either two solutions or zero solutions are obtained. Around $35\%$ of events give zero solutions, and discarding all such events would lead to a significant reduction in event statistics. Therefore, in such events, we vary the transverse momentum of the missing system~(at most by $\pm 10\%$) while keeping the azimuth angle of $\nu$ unchanged until physical solutions of $p_{z,\nu}$ are obtained. Events which give zero solutions even after the aforesaid variation are ignored.
We perform the reconstruction for the top quarks iterating over all possible partitions of light jets~($j$) and $b$-jet forming the hadronic top~($jjb$) and leptons and $b$-jet for the leptonic top~($\ell \nu b$). The two possible neutrino solutions are separately accounted for, forming different partitions. We select the combination that minimizes
\begin{align}
(m_{jjb}-m_t)^2+(m_{\ell\nu b} - m_t)^2,
\end{align}
where $m_{t}$ is the on-shell mass of the top quark.
\medskip
\noindent\textbf{Hadronic channel:} We follow a similar mass minimization strategy in the hadronic channel. We reconstruct the two top quarks, $t_1$ and $t_2$, by iterating over all possible combinations of light jets and $b$-jets.
The combination which minimizes
\begin{align}
(m_{t_1}-m_t)^2+(m_{t_2} - m_t)^2,
\end{align}
is chosen.
\medskip
\noindent\textbf{Di-leptonic channel}: In the more complex di-leptonic $t\bar{t}h$ channel, the invisible system is constituted by two neutrinos. Therefore, in addition to determining the unknown longitudinal momentum of the missing particles, it is also indispensable to partition the four-momentum of the missing system into the two neutrinos in order to fully reconstruct the top and the anti-top. An additional combinatorial ambiguity arises from the tandem $b$-jet and $\ell$ pairing. The study in Ref.~\cite{Goncalves:2018agy} reconstructed the $t\bar{t}(h\to b\bar{b})$ system in di-leptonic mode using $M_{2}$ assisted reconstruction algorithm and a boosted $h\to b\bar{b}$, with jet substructure techniques, to suppress the additional combinatorics between the Higgs boson and top quark decays.
In contrast, the present analysis reconstructs the $t\bar{t}(h\to \gamma\gamma)$ system following the Recursive Jigsaw Reconstruction~(RJR) algorithm presented in Ref.~\cite{Jackson:2017gcy}.
The RJR approach utilizes a series of jigsaw rules optimized to estimate the unknown kinematic degrees of freedom in an event topology and resolve the combinatorial ambiguities between/within the final state visible and invisible objects. It results is a complete kinematic basis which can be used to define the four-momenta of all the final state and intermediate objects in an event decay tree.
The first step involves the resolution of combinatorial ambiguity between the $b$-jets and the leptons by using the ``Combinatorial Minimization" Jigsaw Rule~(JR)~\cite{Jackson:2017gcy}, identifying the ($b$-jet, $\ell$) pairs by minimizing
\begin{equation}
(m_{ b_j \ell^{+}}^{2} + m_{b_k \ell^{-}}^{2});~ j,k=1,2;~ j \neq k.
\end{equation}
After establishing the two visible hemispheres corresponding to the top and the anti-top, we apply the ``Invisible Mass" JR to estimate the invariant mass of the invisible system~($m_{I}$)~\cite{Jackson:2017gcy} defined as
\begin{equation}
m_{I}^{2} = m_{V}^{2} - 4m_{V_{a}}^{2}m_{V_{b}}^{2},
\end{equation}
where $m_{V}$ is the invariant mass of all the two $b$-tagged jets and the two leptons in the final state. $m_{V_{a}}$ and $m_{V_{b}}$ correspond to the invariant mass of the two visible hemispheres associated with the top and the anti-top that were reconstructed in the previous step. $m_{I}$ is chosen such that it is the smallest Lorentz invariant mass that ensures a non-tachyonic four-momenta for the individual neutrinos upon partitioning the invisible system. Next, we determine the longitudinal momentum of the invisible system, $\slashed{p}_{z}$, using the following relation given by the ``Invisible Rapidity" JR~\cite{Jackson:2017gcy}:
\begin{equation}
\slashed{p}_{z} = p_{z}^{V} \frac{\sqrt{{|\slashed{p}_{T}|}^{2} + m_{I}^{2}}}{\sqrt{{|p_{T}^{V}|}^{2} + m_{V}^{2}}}.
\end{equation}
Here, $p_{z}^{V}$ and $p_{T}^{V}$ represent the longitudinal and transverse momenta, respectively, of the visible system constituted by the two $b$-jets and the two leptons, while $\slashed{p}_{T}$ is the missing transverse momentum.
At this point, we have all the ingredients required to reconstruct the $t\bar{t}$ system. However, in order to reconstruct the top and the anti-top individually, the invisible four-momentum has to be correctly partitioned into the two neutrinos. This is achieved by using the ``Contraboost Invariant" JR specified in Ref.~\cite{Jackson:2017gcy} which estimates the four-momenta of the neutrinos produced from top and anti-top decay in the $t\bar{t}$ CM frame under the assumption that both $t$ and $\bar{t}$ have the same invariant mass. The resolved four-momenta of the neutrinos along with the correctly paired $b$-jets and leptons allows defining the $t$ and the $\bar{t}$ systems independently. The reconstruction efficiency of this method is about $80\%$, which is comparable with $M_{2}$ assisted reconstruction algorithm~\cite{Goncalves:2018agy}.
\medskip
With the fully-resolved $t\bar{t}h$ system, we can reconstruct a multitude of CP-even and CP-odd spin correlation observables defined in the $t\bar{t}$ CM frame and the lab frame. Several observables that do not depend on the spin-polarization of $t\bar{t}$ pair are also considered. Our goal here is to maximally explore the $t\bar{t}h$ multi-particle final state, augmenting the CP-sensitivity of the $pp \to t\bar{t}h$ channel at the HL-LHC.
\section{Analysis}
\label{sec:analysis}
\subsection{Simulation and Event Selection}
\begin{figure*}[!tb]
\centering
\includegraphics[width=0.32\textwidth]{ocp4_delphes_dilep.pdf}
\includegraphics[width=0.32\textwidth]{ocp4_delphes_semilep.pdf}
\includegraphics[width=0.32\textwidth]{ocp4_delphes_hadronic.pdf}
\caption{Reconstructed detector level distributions for the Collins-Soper angle $\theta^{*}$ for the di-leptonic channel (left), semi-leptonic channel (center), and hadronic channel (right). We present for the SM $(\alpha=0)$ and several beyond the SM Higgs-top CP-hypotheses $(\alpha=\pi/2,\pm \pi/4)$.}
\label{fig:theta-star_delphes}
\end{figure*}
In this section, we explore the \emph{direct} Higgs-top CP measurement combining machine learning techniques and efficient kinematic reconstruction methods. We consider $t\bar{t}h$ signal with $h \to \gamma\gamma$ in the di-leptonic, semi-leptonic, and hadronic top decay modes at the HL-LHC. The dominant background to this process is given by continuum $t\bar{t}\gamma\gamma$ production. We simulate both the signal and background event samples with \texttt{MadGraph5$\_$aMC@NLO}~\cite{ Alwall:2014hca} within the \texttt{MadMiner} framework~\cite{Brehmer:2019xox} at leading order~(LO) with a center-of-mass energy of $\sqrt{s} = 14~\text{TeV}$. Higher order effects to the signal rate are included via a flat next-to-leading order k-factor~\cite{deFlorian:2016spz, HXSWG}.
We use \texttt{NNPDF2.3QED} parton distribution function~\cite{Ball:2013hta}. No generation-level cuts have been applied for the signal events, while the backgrounds have been generated in the mass window $105~\text{GeV} \!<\! m_{\gamma\gamma} \!<\! 145~\text{GeV}$. Parton shower and hadronization effects have been included with \texttt{Pythia~8}~\cite{Sjostrand:2007gs} and fast detector simulation with the \texttt{Delphes3} package~\cite{deFavereau:2013fsa}, using the default HL-LHC detector card~\cite{HLLHC_card,Cepeda:2019klc}.
To obtain the cross section and likelihood function as a function of the theory parameters, we use the morphing technique that is already implemented in \texttt{MadMiner}. Here, we take into account the dependence of new physics theory parameters at both $t\bar{t}h$ production and $h \to \gamma\gamma$ decay, and therefore choose a quartic ansatz in the morphing setup, which is used to interpolate the event weights as a function of $\kappa_H=\kappa_{t} \cos{\alpha}$ and $\kappa_A= \kappa_t \sin\alpha$.
We start our analysis by selecting events consisting of two photons and at least two $b$-tagged jets. In addition, we require the final state to contain exactly two opposite-sign leptons for the di-leptonic channel, exactly one lepton and at least two light jets for the semi-leptonic channel, and at least four light jets for the hadronic channel. We demand the individual particles to pass the following identification cuts:
\begin{equation} \begin{aligned}
p_{T\ell} &\!>\! 15~\text{GeV}, \ |\eta_\ell| \!<\! 4,
\quad
p_{T\gamma} \!> \!15~\text{GeV}, \ |\eta_\gamma| \!<\! 4,
\\
p_{Tb} &\!>\! 25~\text{GeV}, \ |\eta_b| \!<\! 4,
\quad
p_{Tj} \!>\! 25~\text{GeV}, \ |\eta_j| \!<\! 5.
\end{aligned} \end{equation}
In addition, we require the di-photon invariant mass to satisfy $|m_{\gamma\gamma} \!-\! 125| \!<\! 10~\text{GeV}$.
We fully reconstruct the $t\bar{t}h$ system following the strategy described in \cref{sec:kin_reconstruct}. In particular, this allows to obtain both the lab frame and the $t\bar{t}$ CM frame observables. As an example for an observable that requires the top reconstruction, we present the distribution of the Collins-Soper angle $\theta^{*}$ in \cref{fig:theta-star_delphes}. When comparing these detector level distributions to the result at parton level, presented in \cref{fig:semilep_parton_dist}, we observe the robustness of our analysis in respect to the reconstruction strategy and detector effects. The distributions are found to retain the CP sensitivity at the detector level, albeit a reduction of about $20\%$ for the di-leptonic channel, $40\%$ for the semi-leptonic channel and $50\%$ for the hadronic channel, compared to parton level.
\subsection{Analysis Methodology}
As we have seen in \cref{sec:theory}, there is no single observable that carries all the information on the CP-structure of the top quark Yukawa. Instead, there is a variety of sensitive observables. Hence, a multi-variate analysis is needed to extract the maximal information on the theory parameters from the data. In the following, we will summarize the adopted observables and the analysis methodology.
In this analysis, we consider the following list of 80 observables to describe the kinematics of signal and background events.
\begin{equation} \begin{aligned}
\textbf{Observables: } &
\Delta\phi_{ik}^{t\bar{t}}, \
\Delta\phi_{hi(k)}^{t\bar{t}}, \
\Delta \phi_{ik}^\text{lab},\\
&
\
\Delta \phi_{h t/h\bar{t}}, \
\Delta\phi_{t\bar{t}}, \theta^{*}, \
b_{4},\ \\
&
m_{\gamma\gamma}, \
m_{t\bar{t}}, \
m_{t\bar{t}h},\Delta R^{\rm 2^{nd}min}_{\gamma j},\\
&
\Delta R_{j_\text{soft}j_\text{hard}}, \
\Delta R_{\ell\nu},\Delta R_{Wb}, \
\Delta R^\text{min}_{\gamma j}, \\ &
\Delta \eta_{t\bar{t}}, \
m_{ht/h\bar{t}}, \
H_{T},
\slashed{E}_{T}/\sqrt{H_{T}}, \\
&
\{p_{T},~\eta\}_{X} \text{ for } X =i, k, t, \bar{t},h .
%
\end{aligned} \end{equation}
We include the complete set of observables used by the ATLAS collaboration in a recent Higgs-top CP study~\cite{ATLAS:2020ior} and complement this set with additional CP-even observables that show strong sensitivity to the CP-phase $(\theta^*, b_4,m_{t\bar{t}}, m_{t\bar{t}h})$ together with the transverse momentum and pseudorapidity of all final state and reconstructed objects. We also incorporate a comprehensive list of spin correlations, as introduced in \cref{eq:CP_odd_obseravable}, which are constructed between all possible final state pairs. We include both observables constructed in the $t\bar{t}$ rest frame, $\Delta\phi_{ik}^{t\bar{t}}$, and in the lab frame, $\Delta\phi_{ik}^{\text{lab}}$. Finally, we account for the correlation observables $\Delta\phi_{hi(k)}^{t\bar{t}}$ that arise from the tensor products involving the Higgs boson momentum, $\epsilon (p_t,p_{\bar t}, p_h, p_{i(k)})$. The following pairs $\{i,k\}$ are considered for the different channels
\begin{equation} \begin{aligned}
\text{di-leptonic: }
i &= \ell^{+},\nu_{t} ,b_{t}, W_{t} \\
k &= \ell^{-},\nu_{\bar{t}} ,b_{\bar{t}}, W_{\bar{t}} \\
\text{semi-leptonic: }
i &= {\ell,\nu,b_{\ell},W_{\ell}} \\
k &= {j_\text{soft},j_\text{hard}, b_\text{had}, W_\text{had}}\\
\text{hadronic: }
i &= j_\text{soft}^{t_1}, j_\text{hard}^{t_1}, b_{t_1}, W_{t_1} \\
k &= j_\text{soft}^{t_2}, j_\text{hard}^{t_2}, b_{t_2}, W_{t_2}
\end{aligned} \end{equation}
In the semi-leptonic case, $b_{\ell}/W_{\ell}$ and $b_\text{had}/W_\text{had}$ represent the $b$-jets/$W$-bosons produced from the leptonically and hadronically decaying top quarks, respectively. In events with more than two $b$-tagged jets, the hardest two are considered while reconstructing the top and the anti-top quarks. $j_\text{hard}$ corresponds to the hardest light jet, from the hadronic top quark, in the top rest frame. \medskip
To interpret the results of our analysis and obtain projected sensitivities, we follow a likelihood-based approach. According to the Neyman-Pearson lemma, the most powerful test statistic to discriminate between two hypotheses, in our case a new physics model parameterized by $\theta = (\kappa_t,\alpha)$ and the SM with $\theta_\text{SM} = (1,0)$, is the likelihood ratio $r(x|\theta; \theta_\text{SM})$. Here, $x$ denotes the set of reconstructed observables introduced above.
\begin{figure*}
\centering
\includegraphics[scale=0.34]{combined_68CL_2_latest_3.pdf}\hspace{1cm}
\includegraphics[scale=0.34]{combined_projections_latest_3.pdf}
\caption{Projected sensitivity in the $(\alpha,\kappa_t)$ plane. In the left panel we show the projected $68\%$ CL contours from direct Higgs-top searches in the semi-leptonic~(blue), di-leptonic~(green), hadronic~(red) $t\bar{t}h$ channels, and their combination~(black), considering all input observables. The right panel shows the projected $68\%$~CL (dashed) and $95\%$~CL (solid) contours from the combination of the three channels considering all input observables. The color palette illustrates the expected p-value of the estimated log-likelihood ratio. The projections are derived for 14~TeV LHC assuming an integrated luminosity of $3~\text{ab}^{-1}$.}
\label{fig:combined_excl_limits}
\end{figure*}
Whereas the likelihood ratio involving detector level observables is intractable, meaning that it cannot be computed directly, it can be estimated using simulations. To address this issue, we use the machine learning based technique introduced in Refs.~\cite{Brehmer:2018eca, Brehmer:2018kdj, Brehmer:2018hga, Stoye:2018ovl, Brehmer:2019bvj, Brehmer:2020ako, Brehmer:2020zwh}, which has been implemented in the \texttt{MadMiner} tool~\cite{Brehmer:2019xox}. This approach uses both reconstructed observables and matrix-element information, which are then used to train neural network models that estimate the likelihood ratio. It therefore accounts for the effects of parton shower, hadronization, and detector effects, while the matrix-element information helps to significantly improve the performance of the neural network training. Using the estimated likelihood ratio function $r(x|\theta; \theta_\text{SM})$, which describes both the linear and non-linear new physics effects, we then perform a likelihood ratio test to obtain our projected sensitivities.
We simulate $10^6$ signal and $10^6$ background events before event selection. Using \texttt{MadMiner}, we train neural networks to estimate the likelihood ratio using the \texttt{ALICES} algorithm with its hyperparameter set to unity~\cite{Stoye:2018ovl}. We use fully connected neural networks with three hidden layers, each containing 100 nodes and $\tanh$ activation function. The neural network training is performed over 100 epochs using the Adam optimizer. To avoid overtraining, we evaluate the loss function on an independent validation set and employ an early stopping procedure. We use a batch size of 128, and an exponentially decaying learning rate~(from $10^{-4}$ to $10^{-5}$). The limit setting is performed with \texttt{MadMiner}'s \texttt{Likelihood} class.
\subsection{Results}
Let us now turn to the results of our study. In \cref{fig:combined_excl_limits} we show the projected sensitivity on the top Yukawa coupling in terms of $\kappa_t$ and $\alpha$ using the $t\bar{t}(h\to \gamma\gamma)$ measurement. In the left panel, we present the $68\%$ CL contours for the individual top decay channels as colored dashed lines. A combination of all channels is shown in the black solid line. The studied channels can be organized in ascending order of sensitivity as: di-leptonic, semi-leptonic, and hadronic modes. Since the leading observables display efficient reconstruction for all channels, as illustrated in Fig.~\ref{fig:theta-star_delphes}, the order of sensitivity among the final state modes follow their correspondent event rate.
In the right panel, we show the $68\%$ and $95\%$ CL contours as dashed and solid lines, respectively. The p-values in the $({\kappa_t,\alpha})$ parameter space are presented through the color palette. We observe that $|\kappa_t|$ can be constrained within $\mathcal{O}(8\%)$ of the SM value at $68\%$ CL through a combination of direct searches in the $t\bar{t}(h\to\gamma\gamma)$ channel at the HL-LHC. Assuming $\kappa_t = 1$, the combined search would be able to probe the Higgs-top CP phase up to $|\alpha| \lesssim 13^{\circ}$ at $68\%$ CL.
We also perform a separate analysis in which we train a neural network exclusively with the CP-even observables shown in the right panel of \cref{fig:information}. We observe that the projected sensitivity of such an analysis, using this smaller set of CP-even observables that are most-sensitive to the non-linear new physics effects, is almost comparable to the projected sensitivity of the combination study performed using the full set of observables. Overall, almost all the sensitivity to the Higgs-top CP-structure is provided by the non-linear terms in $\alpha$. The limited $t\bar{t}(h\to\gamma\gamma)$ event statistics renders sub-leading sensitivity to the observables which probe the linear terms.
\subsection{Systematic Effects}
In this section, we explore the implications from systematic uncertainties on the projected sensitivity of $\kappa_{t}$ and $\alpha$. In particular, we will consider two sources of uncertainty associated with the normalization of both signal and background.
In the statistical analysis, these uncertainties are parameterized through nuisance parameters $\nu_{s}$ and $\nu_{b}$ for the signal and background normalization, respectively. These nuisance parameters encode theoretical and experimental uncertainties on the normalization of distributions, neglecting possible shape uncertainties. As before, we train a neural network using the \texttt{ALICES} method in \texttt{MadMiner} to estimate the likelihood ratio $r\left(x | \theta, \nu; \theta_{SM}, \nu_{SM} \right)$. This is now a function of both the model parameters $\theta = (\kappa_t, \alpha)$ and the nuisance parameters $\nu = (\nu_{s},\nu_{b})$ which have a nominal value $\nu_{SM} = (0,0)$. Before setting limits, a constraint term describing our prior knowledge on the nuisance parameter is added. Adopting a conservative approach, we assume a prior constraint of $20\%$ and $50\%$ in the $t\bar{t}(h \to \gamma\gamma)$ signal and the $t\bar{t}\gamma\gamma$ background, respectively. Finally, we profile over the nuisance parameters following the procedure described in Ref.~\cite{Brehmer:2018eca}.
Before turning to the sensitivity contours, let us remind ourselves that the presented results are based on a multi-variate analysis. In particular, this includes the invariant mass of the di-photon pair. The considered range, $115~\text{GeV} < m_{\gamma\gamma} < 135~\text{GeV}$, was chosen sufficiently wide to contain both a signal dominated region at the Higgs resonance and a background dominated region around it. \texttt{MadMiner} uses this background dominated region to constrain the nuisance parameter associated with the background normalization $\nu_b$, and therefore effectively performs a data-driven side-band analysis. As we will see in a moment, the effective uncertainty of the background normalization is therefore significantly smaller than the 50\% which we assumed as a prior.
In the following, we analyze three scenarios to study the impact of systematic uncertainties on the projected sensitivity in the $\left(\alpha, \kappa_{t}\right)$ plane. In the first scenario, we study the impact associated with only the uncertainty on the background normalization. To do so, we fix the $\nu_{s}=0$ in the estimated likelihood ratio and profile over $\nu_{b}$. Similarly, in a second scenario, we fix $\nu_{b}=0$ and we profile over $\nu_{s}$ to study the impact of the signal uncertainty. Finally, in a third scenario, we obtain the limits after profiling the likelihood ratio over both $\nu_{s}$ and $\nu_{b}$. In \cref{fig:combined_systematics_excl} we present the projected sensitivity on $\alpha$ and $\kappa_{t}$ for all scenarios. The blue, green, and red contours correspond to the first, second, and third scenarios, respectively. The black contour represents the sensitivity assuming no systematic uncertainty and corresponds to the black-solid contours in \cref{fig:combined_excl_limits}.
\begin{figure}[t]
\centering
\includegraphics[scale=0.37,trim={5mm 0 0 0},clip]{combined_projections_nuisance_3.pdf}
\caption{Projected $68\%$ CL contours on $(\alpha,\kappa_{t})$ from the combination of the three channels considering all input observables, profiled over background~(B) uncertainties~(blue), signal~(S) uncertainties~(green), and both signal and background systematic uncertainties~(red). The projections are derived for 14 TeV LHC assuming $\mathcal{L} = 3~\text{ab}^{-1}$.}
\label{fig:combined_systematics_excl}
\end{figure}
At first, we observe that the sensitivity in $\alpha$ remains unaffected from systematic uncertainties~\cite{Goncalves:2021dcu}. This stems from the reason that at $\kappa_{t} = 1$, the sensitivity in $\alpha$ is dominantly controlled by the shape information from kinematic distributions and is largely independent of the event rate due to the combination of two competing effects. On the one hand, the signal cross section $\sigma_{t\bar{t}\left(h \to \gamma\gamma\right)}$ decreases with $\alpha$: for example at $\kappa_t = 1$ the cross section $\sigma_{t\bar{t}\left(h \to \gamma\gamma\right)}$ falls by $\mathcal{O}(25\%)$ from $\alpha = 0$ until $\alpha \sim \pi/3$ and then remains roughly unchanged until $\alpha = \pi/2$. On the other hand, the signal efficiency also improves with $\alpha$. These two effects roughly offset any overall dependence on the event rate, thereby leading to unchanged projection contours in the direction of $\alpha$ even after profiling over the nuisance parameters.
The situation is qualitatively different in the $\kappa_t$ direction. When $\alpha = 0$, the measurement is purely based on a rate information, implying that the Higgs coupling strength $\kappa_t$ and the signal normalization, as parameterized by $\nu_s$, are essentially degenerate. Therefore, our prior uncertainty of the signal normalization will directly propagate into a systematic uncertainty on $\kappa_t$. The new physics effects in the Higgs-top coupling manifest as $\sim \kappa_{t}^{2}$ at the $t\bar{t}h$ production level and as $ \sim \left(1.28 - 0.28\kappa_{t}\right)^{2}$ in $h \to \gamma\gamma$ decay~\cite{Brod:2013cka}. After combining these two factors, an uncertainty of $20\%$ in the $pp \to t\bar{t}(h \to \gamma\gamma)$ cross section translates to roughly $12\%$ uncertainty in $\kappa_{t}$. We observe this effect in \cref{fig:combined_systematics_excl}: for $\alpha = 0$ the projected sensitivity falls from $|\kappa_{t}|\lesssim 8\%$ in the absence of systematic uncertainties to $|\kappa_{t}| \lesssim 13\%$ on profiling over $\nu_{s}$. We observe that despite a prior $50\%$ uncertainty in the background normalization compared to $20\%$ in the signal, its impact on the projection contours in the $\left(\kappa_t, \alpha \right)$ plane is milder. As discussed above, this is a consequence of the side-band measurement and illustrates the robustness of our multi-variate analysis.
\section{Summary}
\label{sec:summary}
In this study, we derived the prospects to \emph{direct} measure the Higgs-top CP-structure in $t\bar{t}(h\to \gamma\gamma)$ channel at the HL-LHC. We show that a combination of machine learning techniques and efficient kinematic reconstruction methods can boost new physics sensitivity, effectively exploring the complex $t\bar{t}h$ multi-particle phase space.
Among the several probes included in our machine learning analysis, this study encompass a comprehensive set of spin correlation observables. Beyond the SM CP-phases steer the spin-polarization of the top pair, and the spin correlations are carried forward by their decay products. We harness the potential of the spin correlation observables via the full reconstruction of the top and anti-top, evaluating these particular observables in the $t\bar{t}$ CM frame, where the correlations are maximal. In the hadronic and semi-leptonic $t\bar{t}h$ channels, we used mass minimization to fully reconstruct the $t\bar{t}h$ system. In the more complex di-leptonic channel, we employed the Recursive Jigsaw Reconstruction technique to resolve the combinatorial ambiguities and determine the unknown degrees of freedom. In all channels, the effects of parton showering, hadronization, and detector resolution were included.
Exploring the intricate $t\bar{t}h$ multi-particle phase space with CP-odd and even observables defined in the laboratory frame and the $t\bar{t}$ CM frame, we obtain strong projections for the Higgs-top CP-phase. Through a combined semi-leptonic, hadronic, and di-leptonic $t\bar{t}(h\to\gamma\gamma)$ search, the HL-LHC can directly probe the Higgs-top coupling modifier and CP-phase respectively up to $|\kappa_t|\lesssim 8\%$ and $|\alpha|\lesssim 13^{\circ}$ at $68\%$~CL.
Possible improvements can be expected by including other relevant channels, such as $t\bar{t}(h\to b\bar{b})$~\cite{Buckley:2015vsa,AmorDosSantos:2017ayi,Goncalves:2018agy,Goncalves:2021dcu}. While this channel displays the bulk of the Higgs decay, $\mathcal{BR}(h\to b\bar{b})\sim 58\%$, it results in sub-leading limits in comparison to $t\bar{t}(h\to \gamma\gamma)$ as it endures a substantial QCD background that is associated with sizable uncertainties~\cite{ATLAS:2017fak,CMS:2018hnq}. Fast-moving precision calculations~\cite{Jezo:2018yaf,Denner:2020orv,Bevilacqua:2021cit} and possible combination of side-band analysis with $t\bar{t}h/t\bar{t}Z$ ratios~\cite{Mangano:2015aow,Goncalves:2021dcu} may change this scenario, controlling the respective uncertainties, and pushing further forward the sensitivity with this extra channel in the near future.
\acknowledgements
We thank Johann Brehmer and Sam Homiller for helpful discussions. RKB and DG thank the U.S.~Department of Energy for the financial support, under grant number DE-SC 0016013. The work of FK is supported by the U.S.~Department of Energy under Grant No.~DE-AC02-76SF00515 and by the Deutsche Forschungsgemeinschaft under Germany’s Excellence Strategy - EXC 2121 Quantum Universe - 390833306. Part of this work was performed at the Aspen Center for Physics, which is supported by National Science Foundation grant PHY-1607611. Some computing for this project was performed at the High Performance Computing Center at Oklahoma State University, supported in part through the National Science Foundation grant OAC-1531128.
|
\section{Introduction} \label{sec:intro}
\input{1-improved-sampling-for-TD-intro-arxiv-v2}
\section{Related Work} \label{sec:related-work}
\input{2-improved-sampling-for-TD-related-work-arxiv-v2}
\section{Preliminaries} \label{sec:preliminaries}
\input{3-improved-sampling-for-TD-preliminaries-arxiv-v2}
\section{Efficient Sampling for Tensor Decomposition} \label{sec:efficient-sampling-for-TD}
\input{4-improved-sampling-for-TD-proposal-arxiv-v2}
\section{Experiments} \label{sec:experiments}
\input{6-improved-sampling-for-TD-experiments-arxiv-v2}
\section{Discussion and Conclusion} \label{sec:conclusion}
\input{7-improved-sampling-for-TD-conclusion-arxiv-v2}
\section*{Acknowledgements}
We thank the anonymous reviewers for their feedback which helped improve the paper.
This work was supported by the Laboratory Directed Research and Development Program of Lawrence Berkeley National Laboratory under U.S.\ Department of Energy Contract No.\ DE-AC02-05CH11231.
\subsection{Tensor Decomposition}
We first introduce the CP decomposition.
Consider an $N$-way tensor $\Xe \in \Rb^{I_1 \times \cdots \times I_N}$.
A rank-$R$ \emph{CP decomposition} of $\Xe$ is of the form
\begin{equation} \label{eq:cp-decomposition}
\Xe(i_1, \ldots, i_N) = \sum_{r=1}^R \prod_{j=1}^N \Abf^{(j)}(i_j, r),
\end{equation}
where each $\Abf^{(j)} \in \Rb^{I_j \times R}$ is called a \emph{factor matrix}.
We use $\CP(\Abf^{(1)}, \ldots, \Abf^{(N)})$ to denote the tensor in \eqref{eq:cp-decomposition}.
The problem of computing a rank-$R$ CP decomposition of a data tensor $\Xe$ can be formulated as
\begin{equation} \label{eq:cp-optimization-problem}
\argmin_{\Abf^{(1)}, \ldots, \Abf^{(N)}} \| \CP(\Abf^{(1)}, \ldots, \Abf^{(N)}) - \Xe \|_\F.
\end{equation}
Unfortunately, this problem is non-convex and difficult to solve exactly.
ALS is the ``workhorse'' algorithm for solving this problem approximately \citep{kolda2009TensorDecompositions}.
With ALS, we consider the objective in \eqref{eq:cp-optimization-problem}, but only solve with respect to one of the factor matrices at a time while keeping the others fixed:
\begin{equation} \label{eq:cp-optimization-An}
\argmin_{\Abf^{(n)}} \| \CP(\Abf^{(1)}, \ldots, \Abf^{(N)}) - \Xe \|_\F.
\end{equation}
The problem in \eqref{eq:cp-optimization-An} can be rewritten as the linear least squares problem
\begin{equation} \label{eq:cp-optimization-An-matrix}
\argmin_{\Abf^{(n)}} \| \Abf^{\neq n} \Abf^{(n) \top} - \Xbf_{(n)}^\top \|_\F,
\end{equation}
where $\Abf^{\neq n} \in \Rb^{(\Pi_{j \neq n} I_j) \times R}$ is defined as
\begin{equation} \label{eq:cp-design-matrix}
\Abf^{\neq n} \defeq \Abf^{(N)} \odot \cdots \odot \Abf^{(n+1)} \odot \Abf^{(n-1)} \odot \cdots \odot \Abf^{(1)}.
\end{equation}
By repeatedly updating each factor matrix one at a time via \eqref{eq:cp-optimization-An-matrix}, we get the standard CP-ALS algorithm outlined in Algorithm~\ref{alg:CP-ALS}.
For further details on the CP decomposition, see \citet{kolda2009TensorDecompositions}.
\begin{algorithm}
\caption{CP-ALS}
\label{alg:CP-ALS}
\DontPrintSemicolon
\KwIn{$\Xe \in \Rb^{I_1 \times \cdots \times I_N}$, rank $R$}
\KwOut{Factor matrices $\Abf^{(1)}, \ldots, \Abf^{(N)}$}
Initialize factor matrices $\Abf^{(2)}, \ldots, \Abf^{(N)}$ \label{line:cp-als:initialize-cores}\;
\While{termination criteria not met}{
\For{$n = 1,\ldots, N$}{
$\Abf^{(n)} = \argmin_{\Abf} \| \Abf^{\neq n} \Abf^\top - \Xbf_{(n)}^\top \|_\F$ \label{line:cp-als:ls}\;
}
}
\Return{$\Abf^{(1)}, \ldots, \Abf^{(N)}$}\;
\end{algorithm}
Next, we introduce the tensor ring decomposition.
For $n \in [N]$, let $\Ge^{(n)} \in \Rb^{R_{n-1} \times I_n \times R_n}$ be 3-way tensors with $R_0 = R_N$.
A rank-$(R_1, \ldots, R_N)$ \emph{tensor ring decomposition} of $\Xe$ is of the form
\begin{equation} \label{eq:tr-decomposition}
\Xe(i_1, \ldots, i_N) = \sum_{r_1, \ldots, r_N} \prod_{n=1}^N \Ge^{(n)}(r_{n-1}, i_n, r_n),
\end{equation}
where each $r_n$ in the sum goes from $1$ to $R_n$ and $r_0 = r_N$, and each $\Ge^{(n)}$ is called a \emph{core tensor}.
We use $\TR(\Ge^{(1)},\ldots,\Ge^{(N)})$ to denote the tensor in \eqref{eq:tr-decomposition}.
Finding the best possible rank-$(R_1, \ldots, R_N)$ tensor ring decomposition of a tensor $\Xe$ is difficult.
With an ALS approach we can update a single core tensor at a time by solving the following problem:
\begin{equation} \label{eq:tr-optimization-Gn}
\argmin_{\Ge^{(n)}} \| \TR(\Ge^{(1)}, \ldots, \Ge^{(N)}) - \Xe \|_\F.
\end{equation}
To reformulate this problem into a linear least squares problem we will need the following definition.
\begin{definition} \label{def:subchain}
By merging all cores except the $n$th, we get a \emph{subchain tensor} $\Ge^{\neq n} \in \Rb^{R_n \times (\Pi_{j \neq n} I_j) \times R_{n-1}}$ defined elementwise via
\begin{equation}
\begin{aligned}
&\Ge^{\neq n}(r_n, \overline{i_{n+1} \ldots i_N i_1 \ldots i_{n-1}}, r_{n-1}) \\
&\defeq \sum_{\substack{r_1, \ldots, r_{n-2}\\r_{n+1}, \ldots, r_N}} \prod_{\substack{j = 1\\j \neq n}}^N \Ge^{(j)}(r_{j-1}, i_j, r_j).
\end{aligned}
\end{equation}
\end{definition}
The problem in \eqref{eq:tr-optimization-Gn} can now be written as the linear least squares problem
\begin{equation} \label{eq:tr-optimization-Gn-matrix}
\Ge^{(n)} = \argmin_{\Ge} \| \Gbf_{[2]}^{\neq n} \Gbf_{(2)}^\top - \Xbf_{[n]}^{\top} \|_\F.
\end{equation}
By repeatedly updating each core tensor one at a time via \eqref{eq:tr-optimization-Gn-matrix}, we get the standard TR-ALS algorithm outlined in Algorithm~\ref{alg:TR-ALS}.
For further details on the tensor ring decomposition, see \citet{zhao2016TensorRing}.
\begin{algorithm}
\caption{TR-ALS}
\label{alg:TR-ALS}
\DontPrintSemicolon
\KwIn{$\Xe \in \Rb^{I_1 \times \cdots \times I_N}$, ranks $R_1,\ldots,R_N$}
\KwOut{Core tensors $\Ge^{(1)}, \ldots, \Ge^{(N)}$}
Initialize core tensors $\Ge^{(2)}, \ldots, \Ge^{(N)}$ \label{line:tr-als:initialize-cores}\;
\While{termination criteria not met}{
\For{$n = 1,\ldots, N$}{
$\Ge^{(n)} = \argmin_{\Ge} \| \Gbf_{[2]}^{\neq n} \Gbf_{(2)}^\top - \Xbf_{[n]}^{\top} \|_\F$ \label{line:tr-als:ls}\;
}
}
\Return{$\Ge^{(1)}, \ldots, \Ge^{(N)}$}\;
\end{algorithm}
\subsection{Recursive Sketching} \label{sec:gaussian-and-recursive-sketches}
\citet{ahle2020ObliviousSketching} present two variants of their recursive sketch.
The first one, which we will use, combines CountSketch and TensorSketch into a single sketch which can be applied efficiently to Kronecker structured vectors.
CountSketch was first introduced by \citet{charikar2004FindingFrequent} and extended to the linear algebra setting by \citet{clarkson2017LowRankApproximation}.
Recall that a function $h : [I] \rightarrow [J]$ is said to be \emph{$k$-wise independent} if it is chosen from a family of functions such that for any $k$ distinct $i_1, \ldots, i_k \in [I]$ the values $h(i_1), \ldots, h(i_k)$ are independent random variables uniformly distributed in $[J]$ \citep{pagh2013CompressedMatrix}.
\begin{definition}
Let $h : [I] \rightarrow [J]$ and $s : [I] \rightarrow \{-1,+1\}$ be 3- and 4-wise independent functions, respectively.
The \emph{CountSketch} matrix $\Cbf \in \Rb^{J \times I}$ is defined elementwise via $\Cbf(j,i) \defeq s(i) \cdot \Ind\{ h(i) = j\}$.
\end{definition}
The TensorSketch was developed in a series of papers by \citet{pagh2013CompressedMatrix}, \citet{pham2013FastScalable}, \citet{avron2014SubspaceEmbeddings} and \citet{diao2018SketchingKronecker}.
\begin{definition}
Let $h_1, h_2 : [I] \rightarrow [J]$ and $s_1, s_2 : [I] \rightarrow \{-1,+1\}$ be 3- and 4-wise independent functions, respectively.
Define $h : [I] \times [I] \rightarrow [J]$ via
\begin{equation}
h(i_1, i_2) \defeq (h_1(i_1) + h_2(i_2) \mod J) + 1.
\end{equation}
The \emph{degree-two TensorSketch} matrix $\Tbf \in \Rb^{J \times I^2}$ is defined elementwise via
\begin{equation}
\Tbf(j,\overline{i_1 i_2}) \defeq s(i_1) s(i_2) \cdot \Ind\{ h(i_1, i_2) = j \}.
\end{equation}
\end{definition}
We are now ready to describe the recursive sketch of \citet{ahle2020ObliviousSketching}.
It is easiest to understand it if we consider its application to Kronecker structured vectors.
Consider $\xbf = \xbf_1 \otimes \cdots \otimes \xbf_N \in \Rb^{I_1 \cdots I_N}$, where each $\xbf_n \in \Rb^{I_n}$.%
\footnote{
\citet{ahle2020ObliviousSketching} consider the case when each $\xbf_n$ has the same length.
We consider a slightly more general definition here since this allows us to work with tensors whose modes are of different size.
}
Suppose first that $N = 2^q$ is a power of 2.
The first step of the recursive sketch is to apply an independent CountSketch matrix $\Cbf_n \in \Rb^{J \times I_n}$ to each $\xbf_n$:
\begin{equation} \label{eq:recursive-sketch-init}
\ybf^{(0)}_n \defeq \Cbf_n \xbf_n \in \Rb^J, \;\;\;\; n \in [N].
\end{equation}
The vectors $\ybf_n^{(0)}$ are then combined pairwise using independent degree-two TensorSketches $\Tbf^{(1)}_n \in \Rb^{J \times J^2}$:
\begin{equation} \label{eq:recursive-sketch-1}
\ybf^{(1)}_n \defeq \Tbf^{(1)}_n( \ybf^{(0)}_{2n-1} \otimes \ybf^{(0)}_{2n} ), \;\;\;\; n \in [N/2].
\end{equation}
This process is then repeated:
At each step, pairs of length-$J$ vectors are combined using independent TensorSketches of size $J \times J^2$.
The $m$th step is
\begin{equation} \label{eq:recursive-sketch-m}
\ybf^{(m)}_n \defeq \Tbf^{(m)}_n( \ybf^{(m-1)}_{2n-1} \otimes \ybf^{(m-1)}_{2n} ), \;\;\;\; n \in [N/2^m].
\end{equation}
When $m=q$, we are left with a single vector $\ybf^{(q)}_1 \in \Rb^J$.
The mapping $\xbf \mapsto \ybf_1^{(q)}$, which we denote by $\Psibf_{J, (I_n)_{n=1}^{2^q}}$, is the recursive sketch.
If $N$ is not a power of 2, we choose $q \defeq \lceil \log_2 N \rceil$ and define the recursive sketch as $\Psibf_{J, (I_n)_{n=1}^N} : \xbf \mapsto \Psibf_{J, (\tilde{I}_n)_{n=1}^{2^q} }(\xbf \otimes \ebf_1^{\otimes(2^q-N)})$, where $\ebf_1$ is the first canonical basis vector of length $I_{\max} \defeq \max_{n \in [N]} I_n$, and each $\tilde{I}_n \defeq I_{n}$ for $n \leq N$ and $\tilde{I}_n \defeq I_{\max}$ if $n > N$.
We refer to $\Psibf_{J, (I_n)_{n=1}^N}$ as a $(J, (I_n)_{n=1}^N)$-recursive sketch.
It is in fact linear, and when $N=2^q$ we can write $\Psibf_{J, (I_n)_{n=1}^N}$ as a product of $q+1$ matrices:
\begin{equation} \label{eq:recursive-sketch-factorization}
\Psibf_{J, (I_n)_{n=1}^N} = \Tbf^{(q)} \Tbf^{(q-1)} \cdots \Tbf^{(1)} \Cbf,
\end{equation}
where $\Cbf \defeq \bigotimes_{n=1}^N \Cbf_n$ is a $J^N \times \prod_{n} I_n$ matrix and $\Tbf^{(m)} \defeq \bigotimes_{n=1}^{2^{q-m}} \Tbf_n^{(m)}$ is a $J^{2^{q-m}} \times J^{2^{q-m+1}}$ matrix.
Figure~\ref{fig:recursiveSketch} illustrates the recursive sketch for $N=4$.
\begin{figure}[h]
\centering
\includegraphics[width=.45\columnwidth]{recursiveSketch.pdf}
\caption{
Illustration of the recursive sketch applied to a vector $\xbf_1 \otimes \xbf_2 \otimes \xbf_3 \otimes \xbf_4$.
This is an adaption of Figure~1 in \citet{ahle2020ObliviousSketching}.
}
\label{fig:recursiveSketch}
\begin{picture}(0,0)
\put(55,150){\footnotesize $\xbf_1$}
\put(55,121){\footnotesize $\xbf_2$}
\put(55,92){\footnotesize $\xbf_3$}
\put(55,63){\footnotesize $\xbf_4$}
\put(24,149){\footnotesize $\Cbf_1$}
\put(24,120){\footnotesize $\Cbf_2$}
\put(24,91){\footnotesize $\Cbf_3$}
\put(24,62){\footnotesize $\Cbf_4$}
\put(-8.5,134){\footnotesize $\Tbf^{(1)}_1$}
\put(-8.5,76){\footnotesize $\Tbf^{(1)}_2$}
\put(-37.5,105){\footnotesize $\Tbf^{(2)}_1$}
\put(-69,106){\footnotesize $\ybf^{(2)}_1$}
\end{picture}
\end{figure}
The recursive sketch is a subspace embedding with high probability.
\begin{definition}
A matrix $\Psibf \in \Rb^{J \times I}$ is called a \emph{$\gamma$-subspace embedding} for a matrix $\Abf \in \Rb^{I \times R}$ if
\begin{equation}
\big| \| \Psibf \Abf \xbf \|_2^2 - \|\Abf \xbf\|_2^2 \big| \leq \gamma \| \Abf \xbf \|_2^2 \;\;\;\; \text{for all } \xbf \in \Rb^R.
\end{equation}
\end{definition}
The recursive sketch has the remarkable feature that the embedding dimension required for subspace embedding guarantees does not depend exponentially on $N$.
See Theorem~1 in \citet{ahle2020ObliviousSketching} or Theorem~\ref{thm:recursive-sketch} for a precise statement.
\subsection{Leverage Score Sampling}
Leverage score sampling is a popular technique for a variety of problems in numerical linear algebra.
For an in-depth discussion, see \citet{mahoney2011RandomizedAlgorithms} and \citet{woodruff2014SketchingTool}.
\begin{definition} \label{def:leverage-score}
Let $\Abf \in \Rb^{I \times R}$ and suppose $\Ubf \in \Rb^{I \times \rank(\Abf)}$ contains the left singular vectors of $\Abf$.
The $i$th \emph{leverage score} of $\Abf$ is defined as $\ell_i(\Abf) \defeq \| \Ubf(i,:) \|_2^2$ for $i \in [I]$.
\end{definition}
\begin{definition} \label{def:leverage-score-sampling}
Let $\qbf \in \Rb^I$ be a probability distribution and let $f : [J] \rightarrow [I]$ be a random map such that each $f(j)$ is independent and distributed according to $\qbf$.
Define $\Sbf \in \Rb^{J \times I}$ elementwise via
\begin{equation}
\Sbf(j,i) \defeq \Ind\{ f(j) = i \} / \sqrt{J \qbf(f(j))}.
\end{equation}
We call $\Sbf$ a \emph{sampling matrix with parameters $(J,\qbf)$}, or $\Sbf \sim \Dc(J, \qbf)$ for short.
Let $\Abf \in \Rb^{I \times R}$ be nonzero and suppose $\beta \in (0,1]$.
Define the distribution $\pbf \in \Rb^{I}$ via $\pbf(i) \defeq \ell_i(\Abf)/\rank(\Abf)$.
We say that $\Sbf \sim \Dc(J, \qbf)$ is a \emph{leverage score sampling matrix for $(\Abf, \beta)$} if $\qbf(i) \geq \beta \pbf(i)$ for all $i \in [I]$.
\end{definition}
For a least squares problem $\min_{\xbf} \| \Abf \xbf - \ybf \|_2$ where $\Abf \in \Rb^{I \times R}$ has many more rows than columns, we can use sampling to reduce the size of the problem to $\min_{\xbf} \| \Sbf \Abf \xbf - \Sbf \ybf \|_2$.
We would ideally like to sample according to the distribution $\pbf$ in Definition~\ref{def:leverage-score-sampling}, but this requires computing $\Ubf$ in Definition~\ref{def:leverage-score} (e.g., via the SVD or QR decomposition) which costs $O(IR^2)$.
This is the same cost as solving the full least squares problem and is therefore too expensive.
However, as shown by \citet{drineas2012FastApproximation}, the leverage scores can be accurately estimated in less time.
Theorem~\ref{thm:leverage-score-estimation} is a variant of Lemma~9 by \citet{drineas2012FastApproximation}.
They consider the case when $\Psibf$ is a fast Johnson--Lindenstrauss transform instead of a subspace embedding.
\begin{theorem} \label{thm:leverage-score-estimation}
Let $\Abf \in \Rb^{I \times R}$ where $I > R$, $\gamma \in (0,1)$ and suppose $\Psibf \Abf = \Ubf_1 \Sigmabf_1 \Vbf_1^\top$ is a compact SVD.
Define
\begin{equation} \label{eq:u-tilde}
\tilde{\ell}_i(\Abf) \defeq \| \ebf_i^\top \Abf \Vbf_1 \Sigmabf_1^{-1} \|_2^2.
\end{equation}
Suppose that $\Psibf$ is a $\gamma$-subspace embedding for $\Abf$.
Then
\begin{equation}
| \ell_i(\Abf) - \tilde{\ell}_i(\Abf) | \leq \frac{\gamma}{1-\gamma} \ell_i(\Abf) \;\;\;\; \text{for all } i \in [I].
\end{equation}
\end{theorem}
A proof of Theorem~\ref{thm:leverage-score-estimation} appears in Section~\ref{sec:proof-leverage-score-estimation}.
\subsection{CP Decomposition} \label{sec:sampling-for-cp}
Each least squares solve on line~\ref{line:cp-als:ls} in Algorithm~\ref{alg:CP-ALS} involves all entries in $\Xe$.
To reduce the size of this problem, we sample rows according to an approximate leverage score distribution computed as in Theorem~\ref{thm:leverage-score-estimation} with $\Psibf$ chosen to be a recursive sketch.
Theorem~\ref{thm:cp-main} shows that such a sampling approach yields relative error guarantees for the CP-ALS least squares problem.
\begin{theorem} \label{thm:cp-main}
Let $\Abf^{\neq n}$ be defined as in \eqref{eq:cp-design-matrix}.
Define the vector $\vbf \defeq \begin{bmatrix} N, & \cdots & n+1, & n-1, & \cdots & 1\end{bmatrix}$ and suppose $\varepsilon, \delta \in (0,1)$.
Suppose the estimates $\tilde{\ell}_i(\Abf^{\neq n})$ are computed as in Theorem~\ref{thm:leverage-score-estimation}, with $\Psibf \in \Rb^{J_1 \times \Pi_{j \neq n} I_j}$ chosen to be a $(J_1, (I_{\vbf(j)})_{j=1}^{N-1})$-recursive sketch.
Moreover, suppose $\Sbf \in \Rb^{J_2 \times \Pi_{j \neq n} I_j}$ is a sampling matrix with parameters $(J_2, \qbf)$ where $\qbf(i) \propto \tilde{\ell}_i(\Abf^{\neq n})$.
If
\begin{align}
J_1 &\gtrsim N R^2/\delta, \label{eq:cp-J1} \\
J_2 &\gtrsim R \, \max \big( \log(R/\delta), 1/(\varepsilon \delta)\big), \label{eq:cp-J3}
\end{align}
then $\tilde{\Abf} \defeq \argmin_{\Abf} \| \Sbf \Abf^{\neq n} \Abf^\top - \Sbf \Xbf_{(n)}^\top \|_\F$ satisfies the following with probability at least $1-\delta$:
\begin{equation} \label{eq:cp-main-relative-error}
\| \Abf^{\neq n} \tilde{\Abf}^\top - \Xbf_{(n)}^\top \|_\F \leq (1+\varepsilon) \min_{\Abf} \| \Abf^{\neq n} \Abf^\top - \Xbf_{(n)}^\top \|_\F.
\end{equation}
\end{theorem}
A proof of Theorem~\ref{thm:cp-main} is provided in Section~\ref{sec:proof-cp-main}.
It combines well-known results for leverage score sampling with an efficient leverage score estimation procedure.
The estimation procedure follows ideas by \citet{drineas2012FastApproximation} but uses the recursive sketch by \citet{ahle2020ObliviousSketching} instead of the fast Johnson--Lindenstrauss transform that \citeauthor{drineas2012FastApproximation} use.
The dependence on $R$ in \eqref{eq:cp-J3} is optimal in the sense that it cannot be improved when rows are sampled i.i.d. \citep{derezinski2018ReverseIterative}.
It is a significant improvement over the current state-of-the-art sampling-based ALS method by \citet{larsen2020PracticalLeverageBased} which requires $O(R^{N-1} \max(\log(R/\delta), 1/(\delta \varepsilon)))$ samples to achieve relative error guarantees.
The method by \citet{cheng2016SPALSFast} requires $O(R^N \log(I_n/\delta)/\varepsilon^2)$ samples and only achieves weaker additive error guarantees.
In Sections~\ref{sec:cp-efficient-subspace-embedding} and \ref{sec:cp-sampling-scheme} we discuss how to compute the approximate solution $\tilde{\Abf}$ in Theorem~\ref{thm:cp-main} efficiently.
In Section~\ref{sec:cp-complexity-analysis} we compare the complexity of our method to that of other CP decomposition methods.
\subsubsection{Step 1: Computing $\Psibf \Abf^{\neq n}$} \label{sec:cp-efficient-subspace-embedding}
The columns of $\Abf^{\neq n}$ are Kronecker products, so applying the recursive sketch $\Psibf$ to $\Abf^{\neq n}$ efficiently is straightforward.
Let $q \defeq \lceil \log_2(N-1) \rceil$.
First, independent CountSketches $\Cbf_j$ with $J_1$ rows and an appropriate number of columns are applied:
\begin{equation} \label{eq:cp-recursive-init}
\begin{aligned}
\Ybf^{(0)}_j \defeq
\begin{cases}
\Cbf_j \Abf^{(\vbf(j))} & \text{if } 1 \leq j \leq N-1, \\
\Cbf_j \ebf_1 \onebf_{1 \times R} & \text{if } N-1 < j \leq 2^q,
\end{cases}
\end{aligned}
\end{equation}
where $\vbf$ is defined as in Theorem~\ref{thm:cp-main} and $\onebf_{1 \times R}$ is a length-$R$ row vector of ones.
Then, independent TensorSketches are applied recursively:
\begin{equation} \label{eq:cp-recursive-m}
\Ybf_j^{(m)} = \Tbf_j^{(m)} (\Ybf_{2j-1}^{(m-1)} \odot \Ybf_{2j}^{(m-1)}), \;\;\;\; j \in [2^{q-m}],
\end{equation}
for $m = 1, \ldots, q$, where each $\Tbf_j^{(m)} \in \Rb^{J_1 \times J_1^2}$.
The final output is $\Ybf^{(q)}_1 = \Psibf \Abf^{\neq n}$.
\subsubsection{Step 2: Drawing Samples Efficiently} \label{sec:cp-sampling-scheme}
Since a row index $i \in [\prod_{j \neq n} I_j]$ of $\Abf^{\neq n}$ can be written as $i = \overline{i_1 \cdots i_{n-1} i_{n+1} \cdots i_N}$ where each $i_j \in [I_j]$, we can sample an index $i \in [\prod_{j \neq n} I_j]$ by sampling subindices $i_j \in [I_j]$ for each $j \neq n$.
By sampling the subindices in sequence one after another we avoid computing all entries in $\qbf$ which otherwise would cost $\Omega(\prod_{j \neq n} I_j)$.
We use an abbreviated notation to denote the probability of drawing subsequences of indices.
For example, $\Pb(i_1)$ denotes the probability that the first index is $i_1$, and $\Pb( (i_j)_{\substack{j \leq m, j \neq n}} )$ denotes the probability that the first $m$ indices (excluding the $n$th) are $i_1, \ldots, i_{n-1}, i_{n+1}, \ldots, i_m$.
\begin{lemma} \label{lem:cp-sampling-results}
Let $\Psibf\Abf^{\neq n} = \Ubf_1 \Sigmabf_1 \Vbf_1^\top$ be a compact SVD and define $\Phibf \defeq \Vbf_1 \Sigmabf_1^{-1} (\Vbf_1 \Sigmabf_1^{-1})^\top$.
The normalization constant for the distribution $\qbf$ with $\qbf(i) \propto \tilde{\ell}_i(\Abf^{\neq n})$ is
\begin{equation} \label{eq:cp-sampling-results-C}
C
\defeq \sum_{i} \tilde{\ell}_i(\Abf^{\neq n})
= \sum_{r, k} \Phibf(r, k) \cdot \prod_{j \neq n} (\Abf^{(j)\top} \Abf^{(j)})(r,k).
\end{equation}
The marginal probability of drawing $(i_j)_{\substack{j \leq m, j \neq n}}$ is
\begin{equation} \label{eq:cp-sampling-results-joint-prob}
\begin{aligned}
&\Pb( (i_j)_{\substack{j \leq m, j \neq n}} ) = \frac{1}{C} \sum_{r, k} \Phibf(r, k) \\
& \cdot \Big( \prod_{\substack{j \leq m \\ j \neq n}} \Abf^{(j)}(i_j, r) \Abf^{(j)}(i_j, k) \Big) \Big( \prod_{\substack{j > m \\ j \neq n}} \big( \Abf^{(j)\top} \Abf^{(j)} \big)(r, k) \Big).
\end{aligned}
\end{equation}
In \eqref{eq:cp-sampling-results-C} and \eqref{eq:cp-sampling-results-joint-prob}, the summations are over $i \in [\prod_{j \neq n} I_j]$ and $r, k \in [R]$.
\end{lemma}
The proof of Lemma~\ref{lem:cp-sampling-results} is given in Section~\ref{sec:proof-cp-sampling-results}.
We now describe the sampling procedure by first describing how to sample the first index $i_1$ (or $i_2$, if $n = 1$), followed by all subsequent indices.
\paragraph{Sampling First Index}
Suppose $n \neq 1$.
We compute the probability of sampling $i_1$ for all $i_1 \in [I_1]$ via \eqref{eq:cp-sampling-results-joint-prob} and sample an index $i_1 \in [I_1]$ from that distribution.
If $n = 1$, we do this for the second index $i_2$ instead.
\paragraph{Sampling Subsequent Indices}
After drawing $i_1$ (or $i_2$, if $n=1$), all subsequent indices can be drawn one at a time \emph{conditionally} on the previous indices.
Suppose we have drawn indices $(i_j)_{j < m, j \neq n}$.
The conditional distribution of $i_m$ (or $i_{m+1}$, if $n = m$) given the previously drawn indices is then
\begin{equation} \label{eq:conditional-probability-im}
\Pb(i_m \mid (i_j)_{j < m, j \neq n}) = \frac{\Pb((i_j)_{j \leq m, j \neq n})}{\Pb((i_j)_{j < m, j \neq n})}.
\end{equation}
We compute the conditional probability in \eqref{eq:conditional-probability-im} for all $i_m \in [I_m]$ via \eqref{eq:cp-sampling-results-joint-prob} and draw a sample $i_m$ from that distribution.
Once the $J_2$ samples in $[\prod_{j \neq n} I_j]$ have been drawn, the matrix $\Sbf \Abf^{\neq n}$ can be computed without forming $\Abf^{\neq n}$.
The matrix $\Sbf \Xbf_{(n)}^{\top}$ can be computed by extracting only $J_2$ rows from $\Xbf_{(n)}^\top$.
\subsubsection{Complexity Analysis} \label{sec:cp-complexity-analysis}
If $J_1$ and $J_2$ are chosen as in \eqref{eq:cp-J1} and \eqref{eq:cp-J3}, and if we assume that $I_j = I$ for all $j \in [N]$ and ignore $\log$ factors, then the per-iteration complexity for our method CP-ALS-ES is
$\tilde{O}( N^2 R^3 (R + N I / \varepsilon) / \delta )$.
In Table~\ref{tab:cp-complexity-comparison}, we compare this to the complexity of other ALS-based methods for CP decomposition (see Section~\ref{sec:related-work}).
Our method is the only one that does not have an exponential per-iteration dependence on $N$.
See Section~\ref{sec:detailed-complexity} for a detailed complexity analysis.
\begin{table}[ht!]
\centering
\caption{
Comparison of leading order computational cost for various CP decomposition methods.
We ignore $\log$ factors and assume that $I_j = I$ for all $j \in [N]$.
$\noiter$ is the number of ALS iterations.
SPALS has an additional upfront cost of $\nnz(\Xe)$.
\label{tab:cp-complexity-comparison}
}
\begin{tabular}{ll}
\toprule
Method & Complexity \\
\midrule
CP-ALS & $\noiter \cdot N (N+I) I^{N-1} R$ \\
SPALS & $\noiter \cdot N (N+I) R^{N+1} / \varepsilon^2$ \\
CP-ARLS-LEV & $\noiter \cdot N ( R + I ) R^N / (\delta \varepsilon)$ \\
\textbf{CP-ALS-ES (our)} & $\noiter \cdot N^2 R^3 (R + N I / \varepsilon) / \delta$ \\
\bottomrule
\end{tabular}
\end{table}
\subsection{Tensor Ring Decomposition} \label{sec:sampling-for-tr}
The least squares problem for TR-ALS on line~\ref{line:tr-als:ls} in Algorithm~\ref{alg:TR-ALS} also involves all entries in $\Xe$.
We use an approach similar to that for the CP decomposition, which yields the following approximation guarantees for the TR-ALS least squares problem.
\begin{theorem} \label{thm:tr-main}
Let $\Gbf_{[2]}^{\neq n}$ be the mode-2 unfolding of the subchain tensor $\Ge^{\neq n}$ (see Definitions~\ref{def:unfolding} and \ref{def:subchain}).
Define the vector $\wbf \defeq \begin{bmatrix}n-1, & \cdots & 1, & N, & \cdots & n+1\end{bmatrix}$ and suppose $\varepsilon, \delta \in (0,1)$.
Suppose the estimates $\tilde{\ell}_i(\Gbf_{[2]}^{\neq n})$ are computed as in Theorem~\ref{thm:leverage-score-estimation}, with $\Psibf \in \Rb^{J_1 \times \Pi_{j \neq n} I_j}$ chosen to be a $(J_1, (I_{\wbf(j)})_{j=1}^{N-1})$-recursive sketch.
Moreover, suppose $\Sbf \in \Rb^{J_2 \times \Pi_{j \neq n} I_j}$ is a sampling matrix with parameters $(J_2, \qbf)$ where $\qbf(i) \propto \tilde{\ell}_i(\Gbf_{[2]}^{\neq n})$.
If
\begin{align}
J_1 &\gtrsim N (R_{n-1} R_n)^2/\delta, \label{eq:tr-J1} \\
J_2 &\gtrsim R_{n-1} R_n \, \max \big( \log(R_{n-1} R_n/\delta), 1/(\varepsilon \delta)\big), \label{eq:tr-J3}
\end{align}
then $\tilde{\Gbf} \defeq \argmin_{\Gbf} \| \Sbf \Gbf_{[2]}^{\neq n} \Gbf^\top - \Sbf \Xbf_{[n]}^\top \|_\F$ satisfies the following with probability at least $1-\delta$:
\begin{equation} \label{eq:tr-main-relative-error}
\| \Gbf_{[2]}^{\neq n} \tilde{\Gbf}^\top - \Xbf_{[n]}^\top\|_\F \leq (1+\varepsilon) \min_{\Gbf} \|\Gbf_{[2]}^{\neq n} \Gbf^\top - \Xbf_{[n]}^\top \|_\F.
\end{equation}
\end{theorem}
A proof of Theorem~\ref{thm:tr-main} is provided in Section~\ref{sec:proof-tr-main}.
The proof uses similar steps as the proof of Theorem~\ref{thm:cp-main}.
The main difference is the structure and number of columns of the least squares design matrix.
Since $\Gbf_{[2]}^{\neq n}$ has $R_{n-1} R_n$ columns, the sample complexity in \eqref{eq:tr-J3} has optimal rank dependence in the sense discussed in Section~\ref{sec:sampling-for-cp}.
This is a significant improvement over the current state-of-the-art sampling-based ALS method by \citet{malik2021SamplingBasedMethod} which requires $O((\prod_{j} R_j^2) \max(\log(R_{n-1} R_n / \delta)),1/(\varepsilon\delta))$ samples to achieve relative error guarantees.
In Sections~\ref{sec:tr-efficient-subspace-embedding} and \ref{sec:tr-sampling-scheme} we discuss how to compute the approximate solution $\tilde{\Gbf}$ in Theorem~\ref{thm:tr-main} efficiently.
In Section~\ref{sec:tr-complexity-analysis} we compare the complexity of our method to that of other tensor ring methods.
\subsubsection{Step 1: Computing $\Psibf \Gbf_{[2]}^{\neq n}$} \label{sec:tr-efficient-subspace-embedding}
Although $\Gbf_{[2]}^{\neq n}$ has a more complicated structure than $\Abf^{\neq n}$, $\Psibf$ can still be applied efficiently to $\Gbf_{[2]}^{\neq n}$.
We describe a scheme for computing the column $\Psibf \Gbf_{[2]}^{\neq n}(:, \overline{r_{n-1} r_n})$ below, and give a more detailed motivation in Section~\ref{sec:proof-correctness-of-tr-sketch}.
Let $q \defeq \lceil \log_2 (N-1) \rceil$.
Define matrices $\Hbf^{(j)}$ for $j \in [2^q]$ as follows:
Let $\Hbf^{(1)} \in \Rb^{I_{n-1} \times R_{n-2}}$ be a matrix with columns $\Hbf^{(1)}(:, k) \defeq \Gbf_{[2]}^{(n-1)}(:, \overline{r_{n-1} k})$ for $k \in [R_{n-2}]$.
Let $\Hbf^{(j)} \defeq \Gbf_{[2]}^{(\wbf(j))} \in \Rb^{I_{\wbf(j)} \times R_{\wbf(j)} R_{\wbf(j)-1}}$ for $2 \leq j \leq N-2$.
Let $\Hbf^{(N-1)} \in \Rb^{I_{n+1} \times R_{n+1}}$ be a matrix with columns $\Hbf^{(N-1)}(:, k) \defeq \Gbf_{[2]}^{(n+1)}(:, \overline{k r_n})$ for $k \in [R_{n+1}]$.
Let $\Hbf^{(j)} \defeq \ebf_1 \in \Rb^{\max_{j \neq n} I_j}$ be a column vector for $N \leq j \leq 2^q$.
Next, define
\begin{align}
\Ybf_j^{(0)} &\defeq \Cbf_j \Hbf^{(j)}, \;\; j \in [2^q], \label{eq:Y0j}\\
K_j^{(0)} &\defeq
\begin{cases}
R_{\wbf(j)} & \text{if } 2 \leq j \leq N-1, \\
1 & \text{if } j = 1 \text{ or } N \leq j \leq 2^q+1.
\end{cases}
\end{align}
The TensorSketch matrices are then applied recursively as follows.
For each $m = 1, \ldots, q$, compute
\begin{equation} \label{eq:Ymj}
\begin{aligned}
&\Ybf_j^{(m)}(:, \overline{k_{1} k_{3}}) = \\
&\sum_{k_{2} \in [K_{2j}^{(m-1)}]} \Tbf_j^{(m)} \big( \Ybf^{(m-1)}_{2j-1}(:, \overline{k_{1} k_{2}}) \otimes \Ybf^{(m-1)}_{2j}(:, \overline{k_{2} k_{3}}) \big)\\
\end{aligned}
\end{equation}
for each $k_1 \in [K^{(m-1)}_{2j-1}], k_3 \in [K^{(m-1)}_{2j+1}], j \in [2^{q-m}]$. For each $m = 1, \ldots, q$, also compute
\begin{equation}
K_{j}^{(m)} \defeq K_{2j-1}^{(m-1)}, \;\; j \in [2^{q-m}+1].
\end{equation}
We prove the following in Section~\ref{sec:proof-correctness-of-tr-sketch}.
\begin{lemma} \label{lem:correctness-of-tr-sketch}
$\Ybf_1^{(q)}$ satisfies $\Ybf_1^{(q)} = \Psibf \Gbf_{[2]}^{\neq n}(:, \overline{r_{n-1} r_n})$.
\end{lemma}
The entire matrix $\Psibf \Gbf_{[2]}^{\neq n}$ can be computed by repeating the steps above for each column $\overline{r_{n-1} r_n} \in [R_{n-1} R_n]$.
\subsubsection{Step 2: Drawing Samples Efficiently} \label{sec:tr-sampling-scheme}
The sampling approach for the tensor ring decomposition is similar to the approach for the CP decomposition which we described in Section~\ref{sec:cp-sampling-scheme}.
\begin{lemma} \label{lem:tr-sampling-results}
Let $\Psibf \Gbf_{[2]}^{\neq n} = \Ubf_1 \Sigmabf_1 \Vbf_1^\top$ be a compact SVD and define $\Phibf \defeq \Vbf_1 \Sigmabf_1^{-1} (\Vbf_1 \Sigmabf_1^{-1})^\top$.
The normalization constant for the distribution $\qbf$ with $\qbf(i) \propto \tilde{\ell}_i(\Gbf_{[2]}^{\neq n})$ is
\begin{equation} \label{eq:tr-sampling-results-C}
\begin{aligned}
&C \defeq \sum_{i} \tilde{\ell}_i(\Gbf_{[2]}^{\neq n}) = \sum_{\substack{r_1,\ldots,r_N \\ k_1, \ldots, k_N}} \Phibf(\overline{r_{n-1} r_n}, \overline{k_{n-1} k_n}) \\
&\cdot \prod_{j \neq n} \big(\Gbf_{[2]}^{(j)\top} \Gbf_{[2]}^{(j)}\big)(\overline{r_{j} r_{j-1}}, \overline{k_{j} k_{j-1}}).
\end{aligned}
\end{equation}
The marginal probability of drawing $(i_j)_{j \leq m, j \neq n}$ is
\begin{equation} \label{eq:tr-sampling-results-joint-prob}
\begin{aligned}
&\Pb( (i_j)_{j \leq m, j \neq n} ) = \frac{1}{C} \sum_{\substack{r_1, \ldots, r_N \\ k_1, \ldots, k_N}} \Phibf(\overline{r_{n-1} r_n}, \overline{k_{n-1} k_n}) \\
& \cdot \Big( \prod_{\substack{j \leq m \\ j \neq n}} \Gbf_{[2]}^{(j)}(i_j, \overline{r_{j} r_{j-1}}) \Gbf_{[2]}^{(j)}(i_j, \overline{k_{j} k_{j-1}}) \Big) \\
& \cdot \Big( \prod_{\substack{j > m \\ j \neq n}} \big( \Gbf_{[2]}^{(j)\top} \Gbf_{[2]}^{(j)} \big)(\overline{r_{j} r_{j-1}}, \overline{k_{j} k_{j-1}}) \Big).
\end{aligned}
\end{equation}
In \eqref{eq:tr-sampling-results-C} and \eqref{eq:tr-sampling-results-joint-prob}, the summations are over $i \in [\prod_{j \neq n} I_j]$ and $r_j, k_j \in [R_j]$ for each $j \in [N]$.
\end{lemma}
The proof of Lemma~\ref{lem:tr-sampling-results} is given in Section~\ref{sec:proof-tr-sampling-results}.
The sampling procedure itself is the same as for the CP decomposition.
The distribution $(\Pb(i_1))_{i_1=1}^{I_1}$ is computed via \eqref{eq:tr-sampling-results-joint-prob} and an index $i_1$ is sampled; if $n=1$, these computations are done for $i_2$ instead of $i_1$.
All subsequent indices are then sampled conditionally on the previous indices.
This is done by computing the conditional distribution in \eqref{eq:conditional-probability-im} by using \eqref{eq:tr-sampling-results-joint-prob}.
The expression in \eqref{eq:tr-sampling-results-joint-prob} can be computed efficiently despite the exponential number of terms in the summation; see Remark~\ref{remark:computing-tr-probabilities} for details.
Once the $J_2$ samples in $[\prod_{j \neq n} I_j]$ have been drawn, the matrix $\Sbf \Gbf_{[2]}^{\neq n}$ can be computed without forming $\Gbf_{[2]}^{\neq n}$.
We describe this in detail in Remark~\ref{remark:computing-sketched-tr-design-matrix}.
The matrix $\Sbf \Xbf_{[2]}^\top$ can be computed by extracting only $J_2$ rows from $\Xbf_{[2]}^\top$.
\subsubsection{Complexity Analysis} \label{sec:tr-complexity-analysis}
If $J_1$ and $J_2$ are chosen as in \eqref{eq:tr-J1} and \eqref{eq:tr-J3}, and if we assume that $R_j = R$ and $I_j = I$ for all $j \in [N]$ and ignore $\log$ factors, then the per-iteration complexity for our method TR-ALS-ES is $\tilde{O}( N^3 R^9/\delta + N^3 I R^8/(\varepsilon \delta) )$.
In Table~\ref{tab:tr-complexity-comparison}, we compare this to the complexity of several other methods for tensor ring decomposition (see Section~\ref{sec:related-work}).
rTR-ALS refers to the method by \citet{yuan2019RandomizedTensor} with each $I$ compressed to $K$ and with TR-ALS as the deterministic algorithm.
TR-SVD-Rand refers to Algorithm~7 in \citet{ahmadi-asl2020RandomizedAlgorithms}.
Our method is the only one that does not have an explicit exponential dependence on $N$.
See Section~\ref{sec:detailed-complexity} for a detailed complexity analysis.
\begin{table}[ht!]
\centering
\caption{
Comparison of leading order computational cost for various tensor ring decomposition methods.
We ignore $\log$ factors and assume that $R_j = R$ and $I_j = I$ for all $j \in [N]$.
$\noiter$ is the number of ALS iterations.
\label{tab:tr-complexity-comparison}
}
\begin{tabular}{ll}
\toprule
Method & Complexity \\
\midrule
TR-ALS & $\noiter \cdot N I^N R^2$ \\
rTR-ALS & $N I^N K + \noiter \cdot N K^N R^2$ \\
TR-SVD & $I^{N+1} + I^N R^3$ \\
TR-SVD-Rand & $I^N R^2$ \\
TR-ALS-Sampled & $\noiter \cdot N I R^{2N+2} / (\varepsilon \delta)$ \\
\textbf{TR-ALS-ES (our)} & $\noiter \cdot N^3 R^8 ( R + I / \varepsilon)/\delta$ \\
\bottomrule
\end{tabular}
\end{table}
\subsection{Sampling Distribution Comparison} \label{sec:sampling-ditribution-comparison}
We first compare the sampling distributions used by our methods with those used by the previous state-of-the-art---CP-ARLS-LEV by \citet{larsen2020PracticalLeverageBased} for the CP decomposition and TR-ALS-Sampled by \citet{malik2021SamplingBasedMethod} for the tensor ring decomposition---when solving the least squares problems in \eqref{eq:cp-optimization-An} and \eqref{eq:tr-optimization-Gn}.
We run standard CP-ALS and TR-ALS on a real data tensor to get realistic factor matrices and core tensors when defining the design matrices $\Abf^{\neq n}$ and $\Gbf_{[2]}^{\neq n}$.
We get the real data tensor $\Xe \in \Rb^{16 \times \cdots \times 16}$ by reshaping a $4096 \times 4096$ gray scale image of a tabby cat into a 6-way tensor and then appropriately permuting the modes, a process called visual data tensorization \citep{yuan2019HighorderTensor}.
We then consider the least squares problems corresponding to an update of the 6th factor matrix or core tensor.
As a performance measure, we compute the KL-divergence of the approximate distribution $\qbf$ from the exact leverage score sampling distribution $\pbf$ in Definition~\ref{def:leverage-score-sampling}.
Tables~\ref{tab:KL-div-CP} and \ref{tab:KL-div-TR} report the results for different $J_1$ and ranks.
The results show that our methods sample from a distribution much closer to the exact leverage score distribution when $J_1$ is as small as $J_1 = 1000$.
See Figures~\ref{fig:CP-rank-10-probability-comparison}--\ref{fig:TR-rank-5-probability-comparison} for a graphical comparison.
\begin{table}[ht!]
\centering
\caption{
KL-divergence (lower is better) of the approximated sampling distribution from the exact one for a CP-ALS least squares problem \eqref{eq:cp-optimization-An} with target rank $R$.
\label{tab:KL-div-CP}
}
\begin{tabular}{lll}
\toprule
Method & $R=10$ & $R=20$ \\
\midrule
CP-ARLS-LEV & 0.2342 & 0.1853 \\
CP-ALS-ES ($J_1 = \text{1e+4}$) & 0.0005 & 0.0006 \\
CP-ALS-ES ($J_1 = \text{1e+3}$) & 0.0151 & 0.0070 \\
CP-ALS-ES ($J_1 = \text{1e+2}$) & 0.1416 & 0.2173 \\
\bottomrule
\end{tabular}
\end{table}
\begin{table}[ht!]
\centering
\caption{
KL-divergence (lower is better) of the approximated sampling distribution from the exact one for a TR-ALS least squares problem \eqref{eq:tr-optimization-Gn-matrix} with target rank $(R,\ldots,R)$.
\label{tab:KL-div-TR}
}
\begin{tabular}{lll}
\toprule
Method & $R=3$ & $R=5$ \\
\midrule
TR-ALS-Sampled & 0.3087 & 0.1279 \\
TR-ALS-ES ($J_1 = \text{1e+4}$) & 0.0005 & 0.0007 \\
TR-ALS-ES ($J_1 = \text{1e+3}$) & 0.0076 & 0.0070 \\
TR-ALS-ES ($J_1 = \text{1e+2}$) & 0.1565 & 0.1831 \\
\bottomrule
\end{tabular}
\end{table}
\subsection{Feature Extraction} \label{sec:feature-extraction}
Next, we run a benchmark feature extraction experiment on the COIL-100 image dataset \citep{nene1996ColumbiaObject} with a setup similar to that in \citet{zhao2016TensorRing} and \citet{malik2021SamplingBasedMethod}.
The data consists of 7200 color images of size $128 \times 128$ pixels, each belonging to one of 100 different classes.
The data is arranged into a $128 \times 128 \times 3 \times 7200$ tensor which is decomposed using either a rank-25 CP decomposition or a rank-$(5,5,5,5)$ tensor ring decomposition.
The mode-4 factor matrix or core tensor is then used as a feature matrix in a $k$-NN algorithm with $k=1$ and 10-fold cross validation.
For our CP-ALS-ES, we use $J_1 = 10000$ and $J_2 = 2000$, and for our TR-ALS-ES we use $J_1 = 10000$ and $J_2 = 1000$.
For the CP decomposition, we compare against CP-ALS in Tensor Toolbox \citep{bader2006Algorithm862, bader2021MATLABTensor}; CPD-ALS, CPD-MINF and CPD-NLS in Tensorlab \citep{vervliet2016Tensorlab}; and our own implementation of CP-ARLS-LEV.
For the tensor ring decomposition, we compare against the implementations of TR-ALS and TR-ALS-Sampled provided by \citet{malik2021SamplingBasedMethod}.
For CP-ARLS-LEV and TR-ALS-Sampled we use 2000 and 1000 samples, respectively.
All iterative methods are run for 40 iterations.
Table~\ref{tab:classification-results} shows the average decomposition time, decomposition error, and classification accuracy for the various methods over 10 repetitions of the experiment\footnote{TR-ALS was only run once due to how long it takes to run.}.
For the CP decomposition, the two randomized methods are faster than the competing methods.
Our method takes about as long to run as CP-ARLS-LEV.
This does not contradict our earlier analysis, which was a \emph{worst-case} analysis.
Real-world datasets are typically better behaved than the worst case, which is why CP-ARLS-LEV requires no more samples than CP-ALS-ES in this example.
For the tensor ring decomposition, the randomized methods are substantially faster than the deterministic one.
Our method is a bit slower here than TR-ALS-Sampled.
All methods achieve good classification accuracy and similar decomposition errors.
\begin{table}[ht!]
\centering
\caption{
Run time, decomposition error and classification accuracy when using tensor decomposition for feature extraction.
\label{tab:classification-results}
}
\begin{tabular}{lrrr}
\toprule
Method & Time (s) & Err.\ & Acc.\ (\%) \\
\midrule
CP-ALS (Ten.\ Toolbox) & 43.6 & 0.31 & 99.2 \\
CPD-ALS (Tensorlab) & 68.4 & 0.31 & 99.0 \\
CPD-MINF (Tensorlab) & 102.3 & 0.34 & 99.7 \\
CPD-NLS (Tensorlab) & 107.8 & 0.31 & 92.1 \\
CP-ARLS-LEV & 28.7 & 0.32 & 98.5 \\
\textbf{CP-ALS-ES (our)} & 27.9 & 0.32 & 98.3 \\
\midrule
TR-ALS & 9813.7 & 0.31 & 99.3 \\
TR-ALS-Sampled & 9.9 & 0.33 & 98.5 \\
\textbf{TR-ALS-ES (our)} & 28.5 & 0.33 & 98.0 \\
\bottomrule
\end{tabular}
\end{table}
\begin{remark}
If $k$-NN was applied directly to the uncompressed images its cost would scale linearly or worse with the number of tensor entries.
Due to the sublinear per-iteration complexity of our proposed methods, the cost of the entire decomposition is sublinear if the number of iterations is chosen appropriately.
While fixing the number of iterations is not guaranteed to produce a good decomposition, we expect this to work well on typical datasets.
Once the decomposition is computed each image has a representation of much lower dimension which makes applying $k$-NN cheaper.
This leads to a reduction in the overall classification cost.
See Section~\ref{sec:supp-feature-extraction-experiments} for a discussion on handling new images that were not part of the initial decomposition.
\end{remark}
\subsection{Demonstration of Improved Complexity}
We construct a synthetic 10-way tensor that demonstrates the improved sampling and computational complexity of our proposed CP-ALS-ES over CP-ARLS-LEV.
It is constructed via \eqref{eq:cp-decomposition} from factor matrices $\Abf^{(n)} \in \Rb^{6 \times 4}$ for $n \in [10]$ with $\Abf^{(n)}(1,1) = 4$, each $\Abf^{(n)}(i,j)$ drawn i.i.d.\ from a Gaussian distribution for $2 \leq i,j \leq 6$, and all other entries zero.
Additionally, i.i.d.\ Gaussian noise with standard deviation $0.01$ is added to all entries of the tensor.
Both methods are run for 20 iterations with a target rank of 4 and are initialized using the randomized range finding approach proposed by \citet{larsen2020PracticalLeverageBased}, Appendix~F.
CP-ARLS-LEV requires $J=6^8 \approx 1.7\text{e+6}$ samples to get an accurate solution, taking 350 seconds.
By contrast, our CP-ALS-ES only requires a recursive sketch size of $J_1 = 1000$ and $J_2 = 50$ samples to get an accurate solution, taking only 4.8 seconds.
Our method improves the sampling complexity and compute time by 4 and almost 2 orders of magnitude, respectively.
A similar example for the tensor ring decomposition is provided in Section~\ref{sec:demonstration-improvement-TR}.
\subsection{Proof of Theorem~\ref{thm:leverage-score-estimation}} \label{sec:proof-leverage-score-estimation}
We first state and prove Lemma~\ref{lemma:subspace-embedding-properties}.
It is similar to Lemma~5 in \citet{drineas2012FastApproximation} and Lemma~4.1 in \citet{drineas2006SamplingAlgorithms} which consider the case when $\Psibf$ is a fast Johnson--Lindenstrauss transform and a sampling matrix, respectively, instead of a subspace embedding.
\begin{lemma} \label{lemma:subspace-embedding-properties}
Consider a matrix $\Abf \in \Rb^{I \times R}$ where $I > R$.
Let $\Abf = \Ubf \Sigmabf \Vbf^{\top}$ be a compact SVD of $\Abf$.
Suppose $\Psibf$ is a $\gamma$-subspace embedding for $\Abf$ with $\gamma \in (0,1)$, and let $\Psibf \Ubf = \Qbf \Lambdabf \Wbf^\top$ be a compact SVD.
Then, the following hold:
\begin{enumerate}[(i)]
\item $\rank(\Psibf \Abf) = \rank(\Psibf \Ubf) = \rank(\Abf)$,
\item $\| \Ibf - \Lambdabf^{-2} \|_2 \leq \gamma/(1-\gamma)$,
\item $(\Psibf \Abf)^\dagger = \Vbf \Sigmabf^{-1} (\Psibf \Ubf)^\dagger$.
\end{enumerate}
\end{lemma}
\begin{proof}
The proof follows similar arguments as those used in the proof of Lemma~4.1 in \citet{drineas2006SamplingAlgorithms}.
Since $\Psibf$ is a $\gamma$-subspace embedding for $\Abf$, we have that
\begin{equation}
(1-\gamma) \|\Ubf \Sigmabf \Vbf^\top \xbf\|_2^2 \leq \| \Psibf \Ubf \Sigmabf \Vbf^\top \xbf \|_2^2 \leq (1+\gamma) \| \Ubf \Sigmabf \Vbf^\top \xbf \|_2^2 \;\;\;\; \text{for all } \xbf \in \Rb^R.
\end{equation}
Let $r \defeq \rank(\Abf)$.
Since $\Sigmabf \Vbf^\top \in \Rb^{r \times R}$ is full rank, and using unitary invariance of the spectral norm, it follows that
\begin{equation}
(1-\gamma) \|\ybf\|_2^2 \leq \| \Psibf \Ubf \ybf \|_2^2 \leq (1+\gamma) \| \ybf \|_2^2 \;\;\;\; \text{for all } \ybf \in \Rb^r.
\end{equation}
Using Theorem~8.6.1 in \citet{golub2013MatrixComputations}, this in turn implies that
\begin{equation} \label{eq:singular-value-bound-Pi-U}
1-\gamma \leq \sigma_i^2(\Psibf \Ubf) \leq 1+\gamma \;\;\;\; \text{for all } i \in [r].
\end{equation}
Consequently, $\rank(\Psibf \Ubf) = r = \rank(\Abf)$.
Moreover, since $\rank(\Psibf \Abf) = \rank(\Psibf \Ubf \Sigmabf \Vbf^\top)$ and $\Sigmabf \Vbf^\top$ is full rank, it follows that $\rank(\Psibf \Abf) = \rank(\Psibf \Ubf)$.
This completes the proof of (i).
Next, note that
\begin{equation}
\| \Ibf - \Lambdabf^{-2} \|_2
= \max_{i \in [r]} \Big| 1 - \frac{1}{\sigma_i^2(\Psibf \Ubf)} \Big|
= \max_{i \in [r]} \Big| \frac{\sigma_i^2(\Psibf \Ubf) - 1}{\sigma_i^2(\Psibf \Ubf)} \Big| \leq \frac{\gamma}{1-\gamma},
\end{equation}
where the inequality follows from the bound in \eqref{eq:singular-value-bound-Pi-U}.
This completes the proof of (ii).
We may write
\begin{equation} \label{eq:SA-dagger}
(\Psibf \Abf)^\dagger
= (\Qbf \Lambdabf \Wbf^{\top} \Sigmabf \Vbf^\top)^\dagger
= \Vbf (\Lambdabf \Wbf^{\top} \Sigmabf)^\dagger \Qbf^\top
\end{equation}
where the second equality follows since $\Qbf$ and $\Vbf$ have orthonormal columns.
Since $\rank(\Psibf \Ubf) = \rank(\Abf)$ due to (i), the matrix $\Lambdabf \Wbf^\top \in \Rb^{r \times r}$ is invertible, and therefore $\Lambdabf \Wbf^\top \Sigmabf \in \Rb^{r \times r}$ is invertible, and hence
\begin{equation} \label{eq:invertible-matrix}
(\Lambdabf \Wbf^\top \Sigmabf)^\dagger
= \Sigmabf^{-1} \Wbf \Lambdabf^{-1}.
\end{equation}
Consequently, combining \eqref{eq:SA-dagger} and \eqref{eq:invertible-matrix} we have
\begin{equation}
(\Psibf \Abf)^\dagger
= \Vbf \Sigmabf^{-1} \Wbf \Lambdabf^{-1} \Qbf^\top
= \Vbf \Sigmabf^{-1} (\Psibf \Ubf)^\dagger.
\end{equation}
This completes the proof of (iii).
\end{proof}
We are now ready to prove the statement in Theorem~\ref{thm:leverage-score-estimation}.
\begin{proof}[Proof of Theorem~\ref{thm:leverage-score-estimation}]
Our proof is similar to the proof of Lemma~9 in \citet{drineas2012FastApproximation}.
Let $\Abf = \Ubf \Sigmabf \Vbf^\top$ be a compact SVD, $r \defeq \rank(\Abf)$ and suppose $i \in [I]$.
From Definition~\ref{def:leverage-score}, we have
\begin{equation} \label{eq:ell-i}
\ell_i(\Abf) = \|\Ubf(i,:)\|_2^2 = \ebf_i^\top \Ubf \Ubf^\top \ebf_i.
\end{equation}
Moreover,
\begin{equation} \label{eq:hat-ell-i}
\tilde{\ell}_i(\Abf)
= \| \ebf_i^\top \Abf \Vbf_1 \Sigmabf_1^{-1} \Ubf_1^\top\|_2^2
= \| \ebf_i^\top \Abf (\Psibf \Abf)^\dagger \|_2^2
= \| \ebf_i^\top \Ubf (\Psibf \Ubf)^\dagger \|_2^2
= \ebf_i^\top \Ubf (\Psibf \Ubf)^\dagger (\Psibf \Ubf)^{\dagger\top} \Ubf^\top \ebf_i,
\end{equation}
where the first equality follows from the definition of $\tilde{\ell}_i(\Abf)$ in \eqref{eq:u-tilde} and the unitary invariance of the spectral norm, and the third equality follows from Lemma~\ref{lemma:subspace-embedding-properties} (iii).
From \eqref{eq:ell-i} and \eqref{eq:hat-ell-i}, we have
\begin{equation}
\begin{aligned} \label{eq:1-pf-lemma-dist-U-u-hat}
| \ell_i(\Abf) - \tilde{\ell}_i(\Abf) |
&= | \ebf_i^\top \Ubf \big( \Ibf - (\Psibf \Ubf)^\dagger (\Psibf \Ubf)^{\dagger \top} \big) \Ubf^\top \ebf_i | \\
&\leq \| \ebf_i^\top \Ubf \|_2 \cdot \| ( \Ibf - (\Psibf \Ubf)^\dagger (\Psibf \Ubf)^{\dagger \top} ) \Ubf^\top \ebf_i \|_2 \\
&\leq \| \ebf_i^\top \Ubf \|_2 \cdot \| \Ibf - (\Psibf \Ubf)^\dagger (\Psibf \Ubf)^{\dagger \top} \|_2 \cdot \| \Ubf^\top \ebf_i \|_2 \\
&= \| \Ibf - (\Psibf \Ubf)^\dagger (\Psibf \Ubf)^{\dagger \top} \|_2 \cdot \ell_i(\Abf),
\end{aligned}
\end{equation}
where the first inequality follows from Cauchy--Schwarz inequality, and the second inequality follows from the definition of the matrix spectral norm.
Let $\Psibf \Ubf = \Qbf \Lambdabf \Wbf^\top$ be a compact SVD.
It follows that
\begin{equation} \label{eq:2-pf-lemma-dist-U-u-hat}
\| \Ibf - (\Psibf \Ubf)^\dagger (\Psibf \Ubf)^{\dagger \top} \|_2 = \| \Ibf - \Wbf \Lambdabf^{-2} \Wbf^\top \|_2.
\end{equation}
From Lemma~\ref{lemma:subspace-embedding-properties} (i), it follows that $\Wbf$ is $r \times r$, hence $\Wbf \Wbf^\top = \Ibf$.
Consequently, and using unitary invariance of the spectral norm,
\begin{equation} \label{eq:3-pf-lemma-dist-U-u-hat}
\| \Ibf - \Wbf \Lambdabf^{-2} \Wbf^\top \|_2 = \|\Ibf - \Lambdabf^{-2}\|_2.
\end{equation}
Combining \eqref{eq:1-pf-lemma-dist-U-u-hat}, \eqref{eq:2-pf-lemma-dist-U-u-hat} and \eqref{eq:3-pf-lemma-dist-U-u-hat}, we get
\begin{equation}
\big| \ell_i(\Abf) - \tilde{\ell}_i(\Abf) \big|
\leq \|\Ibf - \Lambdabf^{-2}\|_2 \cdot \ell_i(\Abf) \leq \frac{\gamma}{1-\gamma} \ell_i(\Abf),
\end{equation}
where the last inequality follows from Lemma~\ref{lemma:subspace-embedding-properties} (ii).
This completes the proof.
\end{proof}
\subsection{Proof of Theorem~\ref{thm:cp-main}} \label{sec:proof-cp-main}
We first state some results that we will need for this proof.
Lemma~\ref{lemma:kronecker-property-1} follows from Lemma~4.2.10 in \citet{horn1994TopicsMatrix}.
\begin{lemma} \label{lemma:kronecker-property-1}
For matrices $\Mbf_1, \ldots, \Mbf_n$ and $\Nbf_1, \ldots, \Nbf_n$ of appropriate sizes,
\begin{equation}
(\Mbf_1 \otimes \cdots \otimes \Mbf_n) \cdot (\Nbf_1 \otimes \cdots \otimes \Nbf_n) = (\Mbf_1 \Nbf_1) \otimes \cdots \otimes (\Mbf_n \Nbf_n).
\end{equation}
\end{lemma}
Theorem~\ref{thm:recursive-sketch} follows directly from Theorem~1 in \citet{ahle2020ObliviousSketching} and its proof.%
\footnote{
Since we are not considering regularized least squares problems, the statistical dimension $s_\lambda$ in \citet{ahle2020ObliviousSketching} just becomes equal to the number of columns of $\Abf$, which is $R$ in our case.
The statement of Theorem~1 in \citet{ahle2020ObliviousSketching} uses $\delta=1/10$, but the statement for general $\delta$ is easy to infer from their proof of the theorem.
}
\begin{theorem} \label{thm:recursive-sketch}
Let $\Abf \in \Rb^{I^N \times R}$.
Let $\Psibf \in \Rb^{J \times I^N}$ be the $(J, (I)_{j=1}^N)$-recursive sketch described in Section~\ref{sec:gaussian-and-recursive-sketches}.
If $J \gtrsim N R^2 /(\gamma^2 \delta)$, then $\Psibf$ is a $\gamma$-subspace embedding for $\Abf$ with probability at least $1-\delta$.
\end{theorem}
It is easy to generalize Theorem~\ref{thm:recursive-sketch} to the setting when $\Psibf$ is a $(J, (I_j)_{j=1}^N)$-recursive sketch where the $I_j$ are not necessarily all equal.
\begin{corr} \label{corr:recursive-sketch}
Let $\Abf \in \Rb^{\Pi_{j=1}^N I_j \times R}$.
Let $\Psibf \in \Rb^{J \times \Pi_{j=1}^N I_j}$ be the $(J, (I_j)_{j=1}^N)$-recursive sketch described in Section~\ref{sec:gaussian-and-recursive-sketches}.
If $J \gtrsim N R^2 / (\gamma^2 \delta)$, then $\Psibf$ is a $\gamma$-subspace embedding for $\Abf$ with probability at least $1-\delta$.
\end{corr}
\begin{proof}
Let $q \defeq \lceil \log_2(N) \rceil$, $I_{\max} \defeq \max_{j \in [N]} I_j$, and $\tilde{I}_j \defeq I_j$ for $j \leq N$ and $\tilde{I}_j \defeq I_{\max}$ for $j > N$.
Let $\onebf_{1 \times R}$ denote a length-$R$ row vector of all ones.
From the definition of the recursive sketch in Section~\ref{sec:gaussian-and-recursive-sketches} and the factorization in \eqref{eq:recursive-sketch-factorization}, we have
\begin{equation} \label{eq:Psibf-A}
\begin{aligned}
\Psibf \Abf
&= \Psibf_{J, (\tilde{I}_j)_{j=1}^{2^q}} \big( \Abf \odot (\ebf_1^{\otimes(2^q - N)} \onebf_{1 \times R}) \big) \\
&= \Tbf^{(q)} \Tbf^{(q-1)} \cdots \Tbf^{(1)} \Big( \bigotimes_{j=1}^{2^q} \Cbf_j \Big) \big( \Abf \odot (\ebf_1^{\otimes(2^q - N)} \onebf_{1 \times R}) \big) \\
&= \Tbf^{(q)} \Tbf^{(q-1)} \cdots \Tbf^{(1)} \bigg( \Big( \Big(\bigotimes_{j=1}^N \Cbf_{j}\Big) \Abf \Big) \odot \Big( \Big( \bigotimes_{j=N+1}^{2^q} \Cbf_{j} \ebf_1 \Big) \onebf_{1 \times R} \Big) \bigg),
\end{aligned}
\end{equation}
where the last equality follows from Lemma~\ref{lemma:kronecker-property-1}.
Define two index sets
\begin{equation}
\Ic \defeq [I_1] \times \cdots \times [I_N] \;\;\;\; \text{and} \;\;\;\; \Ic^c \defeq [I_{\max}]^N \setminus \Ic.
\end{equation}
Let $\hat{\Abf} \in \Rb^{I_{\max}^N \times R}$ be an augmented version of $\Abf$ defined as
\begin{equation} \label{eq:augmented-A}
\hat{\Abf}(\overline{i_N \cdots i_1}, :) =
\begin{cases}
\Abf(\overline{i_N \cdots i_1}, :) & \text{if } (i_1, \ldots, i_N) \in \Ic, \\
0 & \text{if } (i_1, \ldots, i_N) \in \Ic^c.
\end{cases}
\end{equation}
Let $\hat{\Psibf} \in \Rb^{J \times I_{\max}^N}$ be the $(J, (I_{\max})_{j=1}^{N})$-recursive sketch which uses independent CountSketches defined as
\begin{equation} \label{eq:C-splitting}
\hat{\Cbf}_j \in \Rb^{J \times I_{\max}}
\defeq
\begin{bmatrix}
\Cbf_j & \tilde{\Cbf}_j
\end{bmatrix}, \;\;\;\; j \in [2^q],
\end{equation}
where the matrices $\Cbf_j$ are the same as in \eqref{eq:Psibf-A}, and each $\tilde{\Cbf}_j$ is an independent CountSketch of size $J \times (I_{\max} - I_j)$.
Again using the factorization in \eqref{eq:recursive-sketch-factorization}, we have
\begin{equation} \label{eq:hat-Psibf-hat-A}
\begin{aligned}
\hat{\Psibf} \hat{\Abf}
&= \Psibf_{J, (I_{\max})_{j=1}^{2^q}} (\hat{\Abf} \odot \big(\ebf_1^{\otimes (2^q - N)} \onebf_{1 \times R})\big) \\
&= \Tbf^{(q)} \Tbf^{(q-1)} \cdots \Tbf^{(1)} \Big(\bigotimes_{j=1}^{2^q} \hat{\Cbf}_j \Big) \big(\hat{\Abf} \odot (\ebf_1^{\otimes (2^q - N)} \onebf_{1 \times R})\big) \\
&= \Tbf^{(q)} \Tbf^{(q-1)} \cdots \Tbf^{(1)} \bigg( \Big(\Big(\bigotimes_{j=1}^N \hat{\Cbf}_{j}\Big) \hat{\Abf}\Big) \odot \Big( \Big(\bigotimes_{j=N+1}^{2^q} \hat{\Cbf}_{j} \ebf_1 \Big) \onebf_{1 \times R} \Big) \bigg) ,
\end{aligned}
\end{equation}
where the last equality follows from Lemma~\ref{lemma:kronecker-property-1}.
From the definition of matrix multiplication, we have
\begin{equation} \label{eq:Kronecker-C-applied-to-aug-A}
\Big(\bigotimes_{j=1}^N \hat{\Cbf}_{j}\Big) \hat{\Abf}
= \sum_{(i_1, \ldots, i_N) \in \Ic \cup \Ic^c} \Big(\bigotimes_{j=1}^N \hat{\Cbf}_{j}\Big)(:, \overline{i_N \cdots i_1}) \hat{\Abf}(\overline{i_N \cdots i_1}, :).
\end{equation}
Due to \eqref{eq:C-splitting}, it follows that $\hat{\Cbf}_j(:, i_j) = \Cbf_j(:, i_j)$ when $i_j \in [I_j]$, and consequently
\begin{equation} \label{eq:C-splitting-effect}
\Big(\bigotimes_{j=1}^N \hat{\Cbf}_{j}\Big)(:, \overline{i_N \cdots i_1})
= \bigotimes_{j=1}^{N}\hat{\Cbf}_j(:, i_j)
= \bigotimes_{j=1}^{N}\Cbf_j(:, i_j)
= \Big(\bigotimes_{j=1}^N \Cbf_j \Big)(:, \overline{i_N \cdots i_1}) \;\;\;\; \text{for all } (i_1,\ldots,i_N) \in \Ic.
\end{equation}
Using \eqref{eq:augmented-A} and \eqref{eq:C-splitting-effect}, we can simplify \eqref{eq:Kronecker-C-applied-to-aug-A} to
\begin{equation} \label{eq:equality-in-dist-1}
\Big(\bigotimes_{j=1}^N \hat{\Cbf}_{j}\Big) \hat{\Abf}
= \sum_{(i_1, \ldots, i_N) \in \Ic} \Big(\bigotimes_{j=1}^N \Cbf_j\Big)(:, \overline{i_N \cdots i_1}) \Abf(\overline{i_N \cdots i_1}, :)
= \Big(\bigotimes_{j=1}^N \Cbf_{j} \Big) \Abf.
\end{equation}
Similarly, since the first column of each $\hat{\Cbf}_j$ and $\Cbf_j$ are the same,
\begin{equation} \label{eq:equality-in-dist-2}
\Big( \bigotimes_{j=N+1}^{2^q} \hat{\Cbf}_{j} \ebf_1 \Big) \onebf_{1 \times R} = \Big( \bigotimes_{j=N+1}^{2^q} \Cbf_{j} \ebf_1 \Big) \onebf_{1 \times R}.
\end{equation}
Equations \eqref{eq:Psibf-A}, \eqref{eq:hat-Psibf-hat-A}, \eqref{eq:equality-in-dist-1} and \eqref{eq:equality-in-dist-2} together now imply that
\begin{equation} \label{eq:equality-in-dist-3}
\Psibf \Abf = \hat{\Psibf} \hat{\Abf}.
\end{equation}
Moreover, it follows immediately from \eqref{eq:augmented-A} that
\begin{equation} \label{eq:equality-in-norm}
\| \Abf \xbf \|_2 = \| \hat{\Abf} \xbf \|_2 \;\;\;\; \text{for all } \xbf \in \Rb^R.
\end{equation}
Theorem~\ref{thm:recursive-sketch} implies that
\begin{equation}
\Pb\big( \big| \|\hat{\Psibf} \hat{\Abf} \xbf\|_2^2 - \| \hat{\Abf} \xbf \|_2^2 \big| \leq \gamma \|\hat{\Abf} \xbf\|_2^2 \;\; \text{for all } \xbf \in \Rb^R \big) \geq 1-\delta.
\end{equation}
Due to \eqref{eq:equality-in-dist-3} and \eqref{eq:equality-in-norm}, this implies that
\begin{equation}
\Pb\big( \big| \|\Psibf \Abf \xbf\|_2^2 - \| \Abf \xbf \|_2^2 \big| \leq \gamma \|\Abf \xbf\|_2^2 \;\; \text{for all } \xbf \in \Rb^R \big) \geq 1-\delta,
\end{equation}
which is what we wanted to show.
\end{proof}
Theorem~\ref{thm:least-squares-guarantees} is a well-known result.
Since slightly different variations of it have appeared in the literature \citep{drineas2006SamplingAlgorithms, drineas2008RelativeerrorCUR, drineas2011FasterLeast, larsen2020PracticalLeverageBased} we provide a proof sketch just to give the reader some idea of how to derive the version we use.
\begin{theorem} \label{thm:least-squares-guarantees}
Let $\Abf \in \Rb^{I \times R}$ be a matrix, and suppose $\Sbf \sim \Dc(J, \qbf)$ is a leverage score sampling matrix for $(\Abf, \beta)$ where $\beta \in (0,1]$, and that $\varepsilon, \delta \in (0,1)$.
Moreover, define $\OPT \defeq \min_{\Xbf} \|\Abf \Xbf - \Ybf\|_\F$ and $\tilde{\Xbf} \defeq \argmin_{\Xbf} \| \Sbf \Abf \Xbf - \Sbf \Ybf\|_\F$.
If
\begin{equation} \label{eq:least-squares-guarantees-sketch-rate}
J > \frac{4R}{\beta} \max\Big( \frac{4}{3(\sqrt{2}-1)^2}\ln\Big( \frac{4R}{\delta} \Big), \frac{1}{\varepsilon \delta} \Big),
\end{equation}
then the following holds with probability at least $1-\delta$:
\begin{equation} \label{eq:relative-error-guarantee}
\| \Abf \tilde{\Xbf} - \Ybf \|_\F \leq (1+\varepsilon) \OPT.
\end{equation}
\end{theorem}
\begin{proof}[Proof sketch]
Let $\Ubf \in \Rb^{I \times \rank(\Abf)}$ contain the left singular vectors of $\Abf$, and define $\Ybf^\perp \defeq (\Ibf - \Ubf\Ubf^\top) \Ybf$.
According to a matrix version\footnote{See Lemma~S1 in \citet{malik2021SamplingBasedMethod}.} of Lemma~1 by \citet{drineas2011FasterLeast}, the statement in \eqref{eq:relative-error-guarantee} holds if both
\begin{equation} \label{eq:cond-1-sketch}
\sigma_{\min}^2(\Sbf \Ubf) \geq \frac{1}{\sqrt{2}}
\end{equation}
and
\begin{equation} \label{eq:cond-2-sketch}
\| \Ubf^\top \Sbf^\top \Sbf \Ybf^\perp \|_\F^2 \leq \frac{\varepsilon}{2} \OPT^2.
\end{equation}
To complete the proof, it is therefore sufficient to show that $\Sbf$ satisfies both \eqref{eq:cond-1-sketch} and \eqref{eq:cond-2-sketch} with probability at least $1-\delta$.
Using Lemma~S2 in \citet{malik2021SamplingBasedMethod}, which is the same as Theorem~2.11 in \citet{woodruff2014SketchingTool} but with a slightly smaller constant, one can show that the condition \eqref{eq:cond-1-sketch} is satisfied with probability at least $1-\delta/2$ if
\begin{equation} \label{eq:J-partial-1}
J > \frac{16}{3 (\sqrt{2}-1)^2} \frac{R}{\beta} \ln\Big(\frac{4R}{\delta}\Big).
\end{equation}
Next, using Lemma~8 in \citet{drineas2006FastMonte}, it follows that
\begin{equation}
\Eb\| \Ubf^\top \Sbf^\top \Sbf \Ybf^\perp \|_\F^2 \leq \frac{1}{J \beta} R \cdot \OPT^2.
\end{equation}
Markov's inequality together with the assumption
\begin{equation} \label{eq:J-partial-2}
J > \frac{4R}{\beta \varepsilon \delta}
\end{equation}
then implies that
\begin{equation}
\Pb\Big(\|\Ubf^\top \Sbf^\top \Sbf \Ybf^\perp\|_\F^2 > \frac{\varepsilon}{2} \OPT^2\Big) \leq \frac{2R}{J\varepsilon\beta} < \frac{\delta}{2}.
\end{equation}
If \eqref{eq:least-squares-guarantees-sketch-rate} is satisfied, then both \eqref{eq:J-partial-1} and \eqref{eq:J-partial-2} are satisfied, and consequently \eqref{eq:cond-1-sketch} and \eqref{eq:cond-2-sketch} are both true with probability at least $1-\delta$.
\end{proof}
We are now ready to prove the statement in Theorem~\ref{thm:cp-main}.
\begin{proof}[Proof of Theorem~\ref{thm:cp-main}]
Let $E_1$ denote the event that $\Psibf$ is a $1/3$-subspace embedding for $\Abf^{\neq n}$.
Following the notation used in Theorem~\ref{thm:leverage-score-estimation}, let $\Psibf \Abf^{\neq n} = \Ubf_1 \Sigmabf_1 \Vbf_1^{\top}$ be a compact SVD.
Let $E_2$ denote the event that \eqref{eq:cp-main-relative-error} is true.
According to Corollary~\ref{corr:recursive-sketch}, we can guarantee that $\Pb(E_1) \geq 1-\delta/2$ if we choose $J_1$ as in \eqref{eq:cp-J1}.
With $\gamma=1/3$ in Theorem~\ref{thm:leverage-score-estimation}, the estimates $\tilde{\ell}_i(\Abf^{\neq n})$ satisfy
\begin{equation} \label{eq:bound-on-elli}
\frac{1}{2} \ell_i(\Abf^{\neq n}) \leq \tilde{\ell}_i(\Abf^{\neq n}) \leq \frac{3}{2} \ell_i(\Abf^{\neq n}).
\end{equation}
Consequently,
\begin{equation} \label{eq:elli-sum}
\sum_{i=1}^{\Pi_{j \neq n} I_j} \tilde{\ell}_i(\Abf^{\neq n}) \leq \frac{3}{2} \sum_{i=1}^{\Pi_{j \neq n} I_j} \ell_i(\Abf^{\neq n}) = \frac{3}{2} \rank(\Abf^{\neq n}).
\end{equation}
Therefore, since $\qbf(i) \propto \tilde{\ell}_i(\Abf^{\neq n})$, it follows by combining \eqref{eq:bound-on-elli} and \eqref{eq:elli-sum} that
\begin{equation}
\qbf(i) = \frac{\tilde{\ell}_i(\Abf^{\neq n})}{\sum_{i=1}^{\Pi_{j \neq n} I_j} \tilde{\ell}_i(\Abf^{\neq n})} \geq \frac{1}{3} \frac{\ell_i(\Abf^{\neq n})}{\rank(\Abf^{\neq n})}.
\end{equation}
In view of Definition~\ref{def:leverage-score-sampling}, Theorem~\ref{thm:leverage-score-estimation} therefore implies that $\Sbf \sim \Dc(J_2, \qbf)$ is a leverage score sampling matrix for $(\Abf^{\neq n}, 1/3)$ if the event $E_1$ is true.
From Theorem~\ref{thm:least-squares-guarantees}, it then follows that $\Pb(E_2 \mid E_1 ) \geq 1 - \delta/2$ if $J_2$ is chosen as in \eqref{eq:cp-J3}.
With the choices of $J_1$ and $J_2$ above we now have
\begin{equation} \label{eq:event-formula}
\Pb(E_2) \geq \Pb(E_1, E_2) = \Pb(E_1) \Pb(E_2 \mid E_1) \geq (1-\delta/2)^2 \geq 1-\delta
\end{equation}
which is what we wanted to show.
\end{proof}
\subsection{Proof of Lemma~\ref{lem:cp-sampling-results}} \label{sec:proof-cp-sampling-results}
Recall that $\Phibf \defeq \Vbf_1 \Sigmabf_1^{-1} (\Vbf_1 \Sigmabf_1^{-1})^\top$, where $\Psibf \Abf^{\neq n} = \Ubf_1 \Sigmabf_1 \Vbf_1^\top$ is a compact SVD.
From \eqref{eq:u-tilde} we have
\begin{equation} \label{eq:cp-lev-score-estimate}
\tilde{\ell_i}(\Abf^{\neq n})
= \ebf_i^\top \Abf^{\neq n} \Phibf \Abf^{\neq n \top} \ebf_i
= (\Abf^{\neq n} \Phibf \Abf^{\neq n \top})(i,i)
= \sum_{r, k} \Phibf(r,k) \cdot \prod_{j \neq n} \Abf^{(j)}(i_j,r) \Abf^{(j)}(i_j,k),
\end{equation}
where the last equality follows from the definition of $\Abf^{\neq n}$ in \eqref{eq:cp-design-matrix}, and $i = \overline{i_1 \cdots i_{n-1} i_{n+1} \cdots i_N}$.
Using \eqref{eq:cp-lev-score-estimate}, we can compute the normalization constant $C$ as
\begin{equation}
C
\defeq \sum_{i} \tilde{\ell}_i(\Abf^{\neq n})
= \sum_{r, k} \Phibf(r,k) \cdot \prod_{j \neq n} \sum_{i_j} \Abf^{(j)}(i_j,r) \Abf^{(j)}(i_j,k)
= \sum_{r, k} \Phibf(r,k) \cdot \prod_{j \neq n} (\Abf^{(j)\top} \Abf^{(j)})(r,k),
\end{equation}
which proves \eqref{eq:cp-sampling-results-C}.
To make notation a bit less cumbersome, we will use the abbreviated notation
\begin{equation}
\sum_{\{i_j\}_{j>m, j \neq n}} \;\;\;\; \text{to denote} \;\;\;\;
\begin{cases}
\sum_{i_{m+1}} \cdots \sum_{i_{n-1}} \sum_{i_{n+1}} \cdots \sum_{i_N} & \text{if } n > m, \\
\sum_{i_{m+1}} \cdots \sum_{i_N} & \text{otherwise}.
\end{cases}
\end{equation}
Similar abbreviated notation will also be used later on for other indices.
We can again use \eqref{eq:cp-lev-score-estimate} to compute the marginal probabilities of drawing $(i_j)_{j \leq m, j \neq n}$ as
\begin{equation} \label{eq:cp-sample-subsequent-indices}
\begin{aligned}
\Pb((i_j)_{j \leq m, j \neq n})
&= \frac{1}{C} \sum_{\{ i_j \}_{j > m, j \neq n}} \tilde{\ell}_i(\Abf^{\neq n}) \\
&= \frac{1}{C} \sum_{\{ i_j \}_{j > m, j \neq n}} \Big( \sum_{r, k} \Phibf(r,k) \prod_{j \neq n} \Abf^{(j)}(i_j, r) \Abf^{(j)}(i_j,k) \Big) \\
&= \frac{1}{C} \sum_{r,k} \Phibf(r,k) \Big(\prod_{\substack{j \leq m \\ j \neq n}} \Abf^{(j)}(i_j, r) \Abf^{(j)}(i_j,k) \Big) \Big(\prod_{\substack{j > m \\ j \neq n}} \big(\Abf^{(j)\top} \Abf^{(j)} \big)(r,k) \Big),
\end{aligned}
\end{equation}
which proves \eqref{eq:cp-sampling-results-joint-prob}.
\subsection{Proof of Theorem~\ref{thm:tr-main}} \label{sec:proof-tr-main}
The strategy of this proof is similar to that for the proof of Theorem~\ref{thm:cp-main} given in Section~\ref{sec:proof-cp-main}.
Let $E_1$ denote the event that $\Psibf$ is a $1/3$-subspace embedding for $\Gbf_{[2]}^{\neq n}$.
Following the notation used in Theorem~\ref{thm:leverage-score-estimation}, let $\Psibf \Gbf_{[2]}^{\neq n} = \Ubf_1 \Sigmabf_1 \Vbf_1^\top$ be a compact SVD.
Let $E_2$ denote the event that \eqref{eq:tr-main-relative-error} is true.
The matrix $\Gbf_{[2]}^{\neq n}$ is of size $\prod_{j \neq n} I_j \times R_{n-1} R_n$.
According to Corollary~\ref{corr:recursive-sketch}, we can therefore guarantee that $\Pb(E_1) \geq 1 - \delta/2$ if we choose $J_1$ as in \eqref{eq:tr-J1}.
Following the same line of reasoning as in the proof of Theorem~\ref{thm:cp-main}, we can show that the choice $\gamma=1/3$ in Theorem~\ref{thm:leverage-score-estimation} combined with the fact $\qbf(i) \propto \tilde{\ell}_i(\Gbf_{[2]}^{\neq n})$ implies that
\begin{equation}
\qbf(i) = \frac{\tilde{\ell}_i(\Gbf_{[2]}^{\neq n})}{\sum_{i=1}^{\Pi_{j \neq n} I_j} \tilde{\ell}_i(\Gbf_{[2]}^{\neq n})} \geq \frac{1}{3} \frac{\ell_i(\Gbf_{[2]}^{\neq n})}{\rank(\Gbf_{[2]}^{\neq n})}.
\end{equation}
In view of Definition~\ref{def:leverage-score-sampling}, Theorem~\ref{thm:leverage-score-estimation} therefore implies that $\Sbf \sim \Dc(J_2, \qbf)$ is a leverage score sampling matrix for $(\Gbf_{[2]}^{\neq n}, 1/3)$ if the event $E_1$ is true.
From Theorem~\ref{thm:least-squares-guarantees}, it then follows that $\Pb(E_2 \mid E_1) \geq 1 - \delta/2$ if $J_2$ is chosen as in \eqref{eq:tr-J3}.
With the choices of $J_1$ and $J_2$ above and the formula \eqref{eq:event-formula}, we have that $\Pb(E_2) \geq 1-\delta$, which is what we wanted to show.
\subsection{Proof of Lemma~\ref{lem:correctness-of-tr-sketch}} \label{sec:proof-correctness-of-tr-sketch}
It follows directly from Definitions~\ref{def:unfolding} and \ref{def:subchain} that
\begin{equation} \label{eq:unfolded-subchain-tensor-w-ordering}
\Gbf_{[2]}^{\neq n}(\overline{i_{n+1} \cdots i_N i_1 \cdots i_{n-1}}, \overline{r_{n-1} r_n}) = \sum_{\{r_j\}_{j \neq n-1, n}} \prod_{j=1}^{N-1} \Gbf_{[2]}^{(\wbf(j))}(i_{\wbf(j)}, \overline{r_{\wbf(j)} r_{\wbf(j)-1}}),
\end{equation}
and therefore the columns of $\Gbf_{[2]}^{\neq n}$ can be written as
\begin{equation}
\Gbf_{[2]}^{\neq n}(:, \overline{r_{n-1} r_n}) = \sum_{\{r_j\}_{j \neq n-1, n}} \bigotimes_{j=1}^{N-1} \Gbf_{[2]}^{(\wbf(j))}(:, \overline{r_{\wbf(j)} r_{\wbf(j)-1}}).
\end{equation}
Let $q \defeq \lceil \log_2 (N-1) \rceil$.
Using the definition of the recursive sketch in Section~\ref{sec:gaussian-and-recursive-sketches} and the factorization in \eqref{eq:recursive-sketch-factorization}, we have
\begin{equation} \label{eq:justification-tr-sketch-1}
\Psibf \Gbf_{[2]}^{\neq n}(:, \overline{r_{n-1} r_n})
= \Tbf^{(q)} \Tbf^{(q-1)} \cdots \Tbf^{(1)} \Cbf \sum_{\{r_j\}_{j \neq n-1, n}} \bigg(\Big(\bigotimes_{j=1}^{N-1} \Gbf_{[2]}^{(\wbf(j))}(:, \overline{r_{\wbf(j)} r_{\wbf(j)-1}}) \Big) \otimes \ebf_1^{\otimes(2^q - (N-1))} \bigg).
\end{equation}
The notation in the equation above is quite cumbersome.
In particular, the ordering of the matrices $\Gbf_{[2]}^{(j)}$ in the Kronecker product is somewhat awkward.
To alleviate the issue somewhat, we define $\Hbf^{(j)}$ for $j \in [2^q]$ as we did in Section~\ref{sec:tr-efficient-subspace-embedding}:
\begin{itemize}
\item Let $\Hbf^{(1)} \in \Rb^{I_{n-1} \times R_{n-2}}$ be a matrix with columns $\Hbf^{(1)}(:, k) \defeq \Gbf_{[2]}^{(n-1)}(:, \overline{r_{n-1} k})$ for $k \in [R_{n-2}]$.
\item Let $\Hbf^{(j)} \defeq \Gbf_{[2]}^{(\wbf(j))} \in \Rb^{I_{\wbf(j)} \times R_{\wbf(j)} R_{\wbf(j)-1}}$ for $2 \leq j \leq N-2$.
\item Let $\Hbf^{(N-1)} \in \Rb^{I_{n+1} \times R_{n+1}}$ be a matrix with columns $\Hbf^{(N-1)}(:, k) \defeq \Gbf_{[2]}^{(n+1)}(:, \overline{k r_n})$ for $k \in [R_{n+1}]$.
\item Let $\Hbf^{(j)} \defeq \ebf_1 \in \Rb^{\max_{j \neq n} I_j}$ be a column vector for $N \leq j \leq 2^q$.
\end{itemize}
Moreover, we also define the numbers $K^{(0)}_j$ for $j \in [2^q+1]$ as in Section~\ref{sec:tr-efficient-subspace-embedding}:
\begin{equation}
K_j^{(0)} \defeq
\begin{cases}
R_{\wbf(j)} & \text{if } 2 \leq j \leq N-1, \\
1 & \text{otherwise}.
\end{cases}
\end{equation}
With this new notation, we can write \eqref{eq:justification-tr-sketch-1} as
\begin{equation} \label{eq:justification-tr-sketch-2}
\Psibf \Gbf_{[2]}^{\neq n}(:, \overline{r_{n-1} r_n})
= \Tbf^{(q)} \Tbf^{(q-1)} \cdots \Tbf^{(1)} \Cbf \sum_{\{k_j\}_{j=1}^{2^{q}+1}} \bigotimes_{j=1}^{2^q} \Hbf^{(j)}(:, \overline{k_j k_{j+1}}),
\end{equation}
where each summation index $k_j$ goes over values $k_j \in [K_j^{(0)}]$.
Using Lemma~\ref{lemma:kronecker-property-1}, Equation \eqref{eq:justification-tr-sketch-2} can be written as
\begin{equation}
\begin{aligned} \label{eq:justification-tr-sketch-3}
\Psibf \Gbf_{[2]}^{\neq n}(:, \overline{r_{n-1} r_n})
&= \Tbf^{(q)} \Tbf^{(q-1)} \cdots \Tbf^{(1)} \sum_{\{k_j\}_{j=1}^{2^{q}+1}} \bigotimes_{j=1}^{2^q} \Cbf_j \Hbf^{(j)}(:, \overline{k_j k_{j+1}}) \\
&= \Tbf^{(q)} \Tbf^{(q-1)} \cdots \Tbf^{(1)} \sum_{\{k_j\}_{j=1}^{2^{q}+1}} \bigotimes_{j=1}^{2^q} \Ybf_{j}^{(0)}(:, \overline{k_j k_{j+1}}) , \\
\end{aligned}
\end{equation}
where $\Ybf_j^{(0)}$ was defined in \eqref{eq:Y0j}.
Recalling that $\Tbf^{(1)} \defeq \bigotimes_{j=1}^{2^{q-1}} \Tbf^{(1)}_j$, we may further rewrite \eqref{eq:justification-tr-sketch-3} as
\begin{equation}
\begin{aligned}
\Psibf \Gbf_{[2]}^{\neq n}(:, \overline{r_{n-1} r_n})
&= \Tbf^{(q)} \Tbf^{(q-1)} \cdots \Tbf^{(2)} \Big( \bigotimes_{j=1}^{2^{q-1}} \Tbf^{(1)}_j \Big) \sum_{\{k_j\}_{j=1}^{2^{q}+1}} \bigotimes_{j=1}^{2^{q-1}} (\Ybf^{(0)}_{2j-1}(:, \overline{k_{2j-1} k_{2j}}) \otimes \Ybf^{(0)}_{2j}(:, \overline{k_{2j} k_{2j+1}})) \\
&= \Tbf^{(q)} \Tbf^{(q-1)} \cdots \Tbf^{(2)} \sum_{\{k_{2j-1}\}_{j=1}^{2^{q-1}+1}} \bigotimes_{j=1}^{2^{q-1}} \sum_{k_{2j}} \Tbf^{(1)}_j (\Ybf^{(0)}_{2j-1}(:, \overline{k_{2j-1} k_{2j}}) \otimes \Ybf^{(0)}_{2j}(:, \overline{k_{2j} k_{2j+1}})) \\
&= \Tbf^{(q)} \Tbf^{(q-1)} \cdots \Tbf^{(2)} \sum_{\{k_{2j-1}\}_{j=1}^{2^{q-1}+1}} \bigotimes_{j=1}^{2^{q-1}} \Ybf^{(1)}_j(:, \overline{k_{2j-1} k_{2j+1}}), \\
\end{aligned}
\end{equation}
where the second equality follows from Lemma~\ref{lemma:kronecker-property-1}, and each $\Ybf^{(1)}_j$ is defined as in \eqref{eq:Ymj}.
Defining $K_j^{(1)} \defeq K_{2j-1}^{(0)}$ for $j \in [2^{q-1}+1]$, we can further rewrite the equation above as
\begin{equation} \label{eq:justification-tr-sketch-4}
\Psibf \Gbf_{[2]}^{\neq n}(:, \overline{r_{n-1} r_n}) = \Tbf^{(q)} \Tbf^{(q-1)} \cdots \Tbf^{(2)} \sum_{\{k_{j}\}_{j=1}^{2^{q-1}+1}} \bigotimes_{j=1}^{2^{q-1}} \Ybf^{(1)}_j(:, \overline{k_{j} k_{j+1}}),
\end{equation}
where each summation index $k_j$ now goes over the values $k_j \in [K_j^{(1)}]$.
In general, for $m \in [q]$, we have
\begin{equation} \label{eq:justification-tr-sketch-5}
\begin{aligned}
&\Tbf^{(q)} \Tbf^{(q-1)} \cdots \Tbf^{(m)} \sum_{\{k_{j}\}_{j=1}^{2^{q-m+1}+1}} \bigotimes_{j=1}^{2^{q-m+1}} \Ybf^{(m-1)}_j(:, \overline{k_{j} k_{j+1}}) \\
&= \Tbf^{(q)} \Tbf^{(q-1)} \cdots \Tbf^{(m+1)} \sum_{\{\ell_{j}\}_{j=1}^{2^{q-m}+1}} \bigotimes_{j=1}^{2^{q-m}} \Ybf^{(m)}_j(:, \overline{\ell_{j} \ell_{j+1}}),
\end{aligned}
\end{equation}
where the summation indices $k_j$ and $\ell_j$ take on values $k_j \in [K_j^{(m-1)}]$ and $\ell_j \in [K_j^{(m)}]$, respectively, where $K_j^{(m)} \defeq K_{2j-1}^{(m-1)}$ for $j \in [2^{q-m}+1]$, and where each $\Ybf^{(m)}_j$ is defined as in \eqref{eq:Ymj}.
Combining \eqref{eq:justification-tr-sketch-4} and \eqref{eq:justification-tr-sketch-5}, it follows by induction that
\begin{equation}
\Psibf \Gbf_{[2]}^{\neq n} (:, \overline{r_{n-1} r_n}) = \sum_{k_1 \in [K^{(q)}_1]} \sum_{k_2 \in [K^{(q)}_2]} \Ybf_{1}^{(q)}(:, \overline{k_1 k_2}) = \Ybf_{1}^{(q)},
\end{equation}
where the last equality follows since $K_1^{(q)} = K_2^{(q)} = 1$.
\subsection{Proof of Lemma~\ref{lem:tr-sampling-results}} \label{sec:proof-tr-sampling-results}
Throughout the following computations the summation indices go over $i = \overline{i_{n+1} \cdots i_N i_1 \cdots i_{n-1}} \in [\prod_{j \neq n} I_j]$ with $i_j \in [I_j]$ and $r_j, k_j \in [R_j]$ for each $j \in [N]$.
Recall that $\Phibf \defeq \Vbf_1 \Sigmabf_1^{-1} (\Vbf_1 \Sigmabf_1^{-1})^\top$, where $\Psibf \Gbf_{[2]}^{\neq n} = \Ubf_1 \Sigmabf_1 \Vbf_1^\top$ is a compact SVD.
From \eqref{eq:u-tilde} we have
\begin{equation} \label{eq:tr-lev-score-estimate}
\begin{aligned}
\tilde{\ell}_i(\Gbf_{[2]}^{\neq n})
&= \ebf_i^\top \Gbf_{[2]}^{\neq n} \Phibf \Gbf_{[2]}^{\neq n \top} \ebf_i
= \big( \Gbf_{[2]}^{\neq n} \Phibf \Gbf_{[2]}^{\neq n \top} \big)(i,i) \\
&= \sum_{\substack{r_{n-1}, r_n \\ k_{n-1}, k_n}} \Gbf_{[2]}^{\neq n}(i, \overline{r_{n-1} r_n}) \Phibf(\overline{r_{n-1} r_n}, \overline{k_{n-1} k_n}) \Gbf_{[2]}^{\neq n}(i, \overline{k_{n-1} k_n}).
\end{aligned}
\end{equation}
From Definitions~\ref{def:unfolding} and \ref{def:subchain} it follows that%
\footnote{
The only difference between \eqref{eq:unfolded-subchain-tensor-w-ordering} and \eqref{eq:unfolded-subchain-tensor} is that the terms in the product are arranged in a different order.
}
\begin{equation} \label{eq:unfolded-subchain-tensor}
\Gbf_{[2]}^{\neq n}(i, \overline{r_{n-1} r_{n}})
= \Gbf_{[2]}^{\neq n}(\overline{i_{n+1} \cdots i_{N} i_1 \cdots i_{n-1}}, \overline{r_{n-1} r_{n}})
= \sum_{\{r_j\}_{j \neq n-1, n}} \prod_{j \neq n} \Gbf_{[2]}^{(j)}(i_{j}, \overline{r_{j} r_{j-1}}).
\end{equation}
Using \eqref{eq:tr-lev-score-estimate} and \eqref{eq:unfolded-subchain-tensor}, we have
\begin{equation} \label{eq:tr-constant-1}
\begin{aligned}
C
&\defeq \sum_i \tilde{\ell}_i(\Gbf_{[2]}^{\neq n}) \\
&= \sum_i \sum_{\substack{r_{n-1}, r_n \\ k_{n-1}, k_n}} \Big(\sum_{\{r_j\}_{j \neq n-1, n}} \prod_{j \neq n} \Gbf_{[2]}^{(j)}(i_{j}, \overline{r_{j} r_{j-1}})\Big) \Phibf(\overline{r_{n-1} r_{n}}, \overline{k_{n-1} k_n}) \Big(\sum_{\{k_j\}_{j \neq n-1, n}} \prod_{j \neq n} \Gbf_{[2]}^{(j)}(i_{j}, \overline{k_{j} k_{j-1}})\Big) \\
&= \sum_{\substack{r_1, \ldots, r_N \\ k_1, \ldots, k_N}} \Phibf(\overline{r_{n-1} r_n}, \overline{k_{n-1} k_n}) \prod_{j \neq n} \Big( \sum_{i_j} \Gbf_{[2]}^{(j)}(i_{j}, \overline{r_{j} r_{j-1}}) \Gbf_{[2]}^{(j)}(i_{j}, \overline{k_{j} k_{j-1}}) \Big) \\
&= \sum_{\substack{r_1, \ldots, r_N \\ k_1, \ldots, k_N}} \Phibf(\overline{r_{n-1} r_n}, \overline{k_{n-1} k_n}) \prod_{j \neq n} \big(\Gbf_{[2]}^{(j) \top} \Gbf_{[2]}^{(j)}\big)(\overline{r_{j} r_{j-1}}, \overline{k_{j} k_{j-1}}),
\end{aligned}
\end{equation}
which proves the expression in \eqref{eq:tr-sampling-results-C}.
Moreover, using \eqref{eq:tr-lev-score-estimate} and \eqref{eq:unfolded-subchain-tensor} we have that the marginal probability of drawing $(i_j)_{j \leq m, j \neq n}$ is
\begin{equation}
\begin{aligned}
&\Pb((i_j)_{j \leq m, j \neq n})
= \frac{1}{C} \sum_{\{i_j\}_{j > m, j \neq n}} \tilde{\ell}_i(\Gbf_{[2]}^{\neq n}) \\
&= \frac{1}{C} \sum_{\{i_j\}_{j > m, j \neq n}} \sum_{\substack{r_{n-1}, r_n \\ k_{n-1}, k_n}} \Phibf(\overline{r_{n-1} r_n}, \overline{k_{n-1} k_n}) \Big( \sum_{\{r_j\}_{j \neq n-1, n}} \prod_{j \neq n} \Gbf_{[2]}^{(j)}(i_j, \overline{r_{j} r_{j-1}}) \Big) \Big( \sum_{\{k_j\}_{j \neq n-1, n}} \prod_{j \neq n} \Gbf_{[2]}^{(j)}(i_j, \overline{k_{j} k_{j-1}}) \Big) \\
&= \frac{1}{C} \sum_{\substack{r_1, \ldots, r_N \\ k_1, \ldots, k_N}} \Phibf(\overline{r_{n-1} r_n}, \overline{k_{n-1} k_n}) \Big( \prod_{\substack{j \leq m \\ j \neq n}} \Gbf_{[2]}^{(j)}(i_j, \overline{r_{j}r_{j-1}}) \Gbf_{[2]}^{(j)}(i_j, \overline{k_{j}k_{j-1}}) \Big) \Big( \prod_{\substack{j > m \\ j \neq n}} \big(\Gbf_{[2]}^{(j) \top} \Gbf_{[2]}^{(j)}\big)(\overline{r_{j} r_{j-1}}, \overline{k_{j} k_{j-1}}) \Big),
\end{aligned}
\end{equation}
which proves \eqref{eq:tr-sampling-results-joint-prob}.
\subsection{CP-ALS-ES: Proposed Sampling Scheme for CP Decomposition} \label{sec:detailed-complexity-cp}
In this section we derive the computational complexity of the scheme proposed in Section~\ref{sec:sampling-for-cp}.
\paragraph{Computing $\Psibf \Abf^{\neq n}$}
First, we consider the costs of computing $\Psibf \Abf^{\neq n}$ as described in Section~\ref{sec:cp-efficient-subspace-embedding}:
\begin{itemize}
\item Computing $\Ybf_j^{(0)}$ for all $j \in [2^q]$:
Each $\Cbf_j \Abf^{(\vbf(j))}$ costs at most $O(I_{\vbf(j)} R)$ to compute, and each $\Cbf_j (\ebf_1 \onebf_{1 \times R})$ costs $O(R)$ to compute.
Since $2^q \leq 2N$, the total cost for this step is therefore $O(R \sum_{j \neq n} I_j)$.
\item Computing $\Ybf_j^{(m)}$ for all $m \in [q]$ and all $j \in [2^{q-m}]$:
A single $J_1 \times J_1^2$ TensorSketch costs $O(R J_1 \log J_1)$ to apply to a matrix of the form $\Ybf_{2j-1}^{(m-1)} \odot \Ybf_{2j}^{(m-1)}$.
Such a TensorSketch is applied a total of $\sum_{m=1}^q 2^{q-m} = 2^q - 1 = O(N)$ times, so the total cost of this whole step is therefore $O(N R J_1 \log J_1)$.
\end{itemize}
The cost for computing $\Psibf \Abf^{\neq n}$ is therefore
\begin{equation} \label{eq:cp-cost-1}
O\bigg(R\Big(N J_1 \log J_1 + \sum_{j \neq n} I_j\Big)\bigg).
\end{equation}
\paragraph{Drawing $J_2$ Samples}
Second, we consider the cost of drawing $J_2$ samples in $[\prod_{j \neq n} I_j]$ from the distribution $\qbf$ as described in Section~\ref{sec:cp-sampling-scheme}:
\begin{itemize}
\item One-time costs:
Computing the SVD of $\Psibf \Abf^{\neq n}$ costs $O(J_1 R^2)$.
Computing $\Phibf = \Vbf_1 \Sigmabf_1^{-1} (\Vbf_1 \Sigmabf_1^{-1})^\top$ costs $O(R^3)$.
Moreover, we can compute all products $\Abf^{(j)\top} \Abf^{(j)}$ for $j \neq n$ upfront for a cost of $O(R^2 \sum_{j \neq n} I_j)$.
The sum of these one-time costs is $O(R^2(J_1 + R + \sum_{j \neq n} I_j))$.
\item Cost of sampling $J_2$ indices:
Since each $\Abf^{(j)\top} \Abf^{(j)}$ for $j \neq n$ has already been computed, the cost of computing the probability $\Pb(i_m \mid (i_j)_{j<m, j \neq n})$ for a single set $(i_j)_{j \leq m, j \neq n}$ via \eqref{eq:cp-sampling-results-joint-prob} and \eqref{eq:conditional-probability-im} is $O(R^2 N)$.
The total cost for computing the whole distribution for $i_m \in [I_m]$, for all $m \in [N] \setminus \{n\}$, is therefore $O(R^2 N \sum_{j \neq n} I_j)$.
Since the main cost of sampling an index $i = \overline{i_1 \cdots i_{n-1} i_{n+1} \cdots i_N}$ is computing the distribution for each subindex, and we need to sample a total of $J_2$ samples, it follows that the total cost of drawing $J_2$ samples is $O(J_2 R^2 N \sum_{j \neq n} I_j)$.
\end{itemize}
In total, when including both one-time and per-sample costs, we get a cost for drawing $J_2$ samples from $\qbf$ of
\begin{equation} \label{eq:cp-cost-2}
O \bigg( R^2 \Big( J_1 + R + J_2 N \sum_{j \neq n} I_j \Big) \bigg).
\end{equation}
\paragraph{Sampled Least Squares Problem}
Finally, we consider the cost of constructing and solving the sampled least squares problem once the $J_2$ samples in $[\prod_{j \neq n} I_j]$ have been drawn:
\begin{itemize}
\item Once the $J_2$ samples in $[\prod_{j \neq n} I_j]$ are drawn, it costs $O(J_2 R N)$ to form $\Sbf \Abf^{\neq n}$, and $O(J_2 I_n)$ to form $\Sbf \Xbf_{(n)}^{\top}$.
This can be done implicitly without forming the matrices $\Sbf$, $\Abf^{\neq n}$, and $\Xbf_{(n)}^\top$.
\item The cost of computing the solution $\tilde{\Abf}^\top = (\Sbf \Abf^{\neq n})^\dagger \Sbf \Xbf_{(n)}^\top$ using a standard method (e.g., via QR decomposition) is $O(J_2 R^2 + J_2 R I_n)$; see Section~5.3.3 in \citet{golub2013MatrixComputations} for details.
\end{itemize}
In total, the costs of constructing and solving the least squares problem is therefore
\begin{equation} \label{eq:cp-cost-3}
O(J_2 R(N + R + I_n)).
\end{equation}
\paragraph{Total Per-Iteration Cost for CP-ALS-ES}
Recall that for each iteration of CP-ALS, we need to solve $N-1$ least squares problems.
Consequently, adding the costs in \eqref{eq:cp-cost-1}, \eqref{eq:cp-cost-2}, \eqref{eq:cp-cost-3} and multiplying by $N-1$, we get a total cost per iteration of
\begin{equation} \label{eq:cp-full-complexity}
O\bigg(R N^2 J_1 \log J_1 + R^2 N \Big(J_1 + R + J_2 N \sum_{j \neq n} I_j \Big) + J_2 R N I_n\bigg).
\end{equation}
If the sketch rates $J_1$ and $J_2$ are chosen according to \eqref{eq:cp-J1} and \eqref{eq:cp-J3}, this per-iteration cost becomes
\begin{equation} \label{eq:cp-full-complexity-J-replaced}
O\bigg( \frac{R^3 N^3}{\delta} \log\Big( \frac{R^2 N}{\delta} \Big) + \frac{R^4 N^2}{\delta} + \Big( R^3 N^2 \sum_{j \neq n} I_j + R^2 N I_n \Big) \max\Big( \log\Big(\frac{R}{\delta}\Big), \frac{1}{\varepsilon\delta} \Big) \bigg).
\end{equation}
\subsection{TR-ALS-ES: Proposed Sampling Scheme for Tensor Ring Decomposition} \label{sec:detailed-complexity-tr}
In this section we derive the computational complexity of the scheme proposed in Section~\ref{sec:sampling-for-tr}.
\paragraph{Computing $\Psibf \Gbf_{[2]}^{\neq n}$}
First, we consider the computation of $\Psibf \Gbf_{[2]}^{\neq n}$ described in Section~\ref{sec:tr-efficient-subspace-embedding}:
\begin{itemize}
\item Computing $\Ybf_j^{(0)}$ for all $j \in [2^q]$:
Recall that computing $\Cbf_j \Hbf^{(j)}$ costs $\nnz(\Hbf^{(j)})$.
Consequently, the cost of computing all $\Ybf_j^{(0)}$ for $N \leq j \leq 2^q$ is just $O(1)$.
The total cost for this step is therefore $O(\sum_{j = 1}^N I_j R_{j-1} R_j)$.
\item Computing $\Ybf_j^{(m)}$ for all $m \in [q]$ and all $j \in [2^{q-m}]$:
Computing $\Tbf_j^{(m)}( \Ybf_{2j-1}^{(m-1)}(:, \overline{k_1 k_2}) \otimes \Ybf_{2j}^{(m-1)}(:, \overline{k_2 k_3}) )$ requires applying a $J_1 \times J_1^2$ TensorSketch to the Kronecker product of two vectors, which costs $O(J_1 \log J_1)$.
This needs to be done for each $k_2 \in [K_{2j}^{(m-1)}]$ when computing the sum in \eqref{eq:Ymj}.
This sum, in turn, needs to be computed for all $k_1 \in [K_{2j-1}^{(m-1)}]$, $k_3 \in [K_{2j+1}^{(m-1)}]$ and $j \in [2^{q-m}]$.
Doing this for each $m \in [q]$ brings the total cost of this step to
\begin{equation}
O\Big( \sum_{m=1}^q \sum_{j=1}^{2^{q-m}} \sum_{k_1=1}^{K_{2j-1}^{(m-1)}} \sum_{k_2=1}^{K_{2j}^{(m-1)}} \sum_{k_3=1}^{K_{2j+1}^{(m-1)}} J_1 \log J_1 \Big).
\end{equation}
As we will see further down, this expression simplifies considerably if all $R_i$ are assumed to be equal.
\end{itemize}
Adding up the per-column costs above and multiplying them by the number of columns $R_{n-1} R_n$, we get that the cost for computing $\Psibf \Gbf_{[2]}^{\neq n}$ is
\begin{equation} \label{eq:tr-cost-1}
O\bigg(R_{n-1} R_n \Big(\sum_{j = 1}^N I_j R_{j-1} R_j + \sum_{m=1}^q \sum_{j=1}^{2^{q-m}} \sum_{k_1=1}^{K_{2j-1}^{(m-1)}} \sum_{k_2=1}^{K_{2j}^{(m-1)}} \sum_{k_3=1}^{K_{2j+1}^{(m-1)}} J_1 \log J_1 \Big)\bigg).
\end{equation}
\paragraph{Drawing $J_2$ Samples}
Second, we consider the cost of drawing $J_2$ samples in $[\prod_{j \neq n} I_j]$ from the distribution $\qbf$ as described in Section~\ref{sec:tr-sampling-scheme}:
\begin{itemize}
\item One-time costs:
Computing the SVD of $\Psibf \Gbf_{[2]}^{\neq n}$ costs $O(J_1 (R_{n-1}R_n)^2)$.
Computing $\Phibf = \Vbf_1 \Sigmabf_1^{-1} (\Vbf_1 \Sigmabf_1^{-1})^\top$ costs $O((R_{n-1}R_n)^3)$.
Moreover, we can compute all products $\Gbf_{[2]}^{(j) \top} \Gbf_{[2]}^{(j)}$ for $j \neq n$ upfront for a cost of $O(\sum_{j \neq n} (R_{j-1} R_j)^2 I_j)$.
The sum of these one-time costs is $O(J_1 (R_{n-1} R_n)^2 + (R_{n-1} R_n)^3 + \sum_{j \neq n} (R_{j-1} R_j)^2 I_j)$.
\item Cost of sampling $J_2$ indices:
The main cost of drawing the samples is computing the sampling distributions.
Even though the number of terms in the sum of \eqref{eq:tr-sampling-results-joint-prob} is exponential in $N$, the joint probability distribution can be computed efficiently.
We discuss how to do this in Remark~\ref{remark:computing-tr-probabilities}.
The cost of doing this for one set of indices $(i_j)_{j \leq m, j \neq n}$ is given in \eqref{eq:tensor-evaluation-cost}.
Repeating this for all $i_j \in [I_j]$, which is required to get the distribution for the $j$th index, brings the cost to
\begin{equation}
O\Big( I_j R_N^2 \sum_{d=1}^{N-1} R_d^2 R_{d+1}^2 \Big).
\end{equation}
When this is repeated for all $N$ indices, and a total of $J_2$ times to get all samples, this brings the cost to
\begin{equation}
O\bigg( J_2 \Big( \sum_{j=1}^N I_j \Big) R_N^2 \sum_{d=1}^{N-1} R_d^2 R_{d+1}^2 \bigg).
\end{equation}
\end{itemize}
Adding the one-time costs and the costs associated to computing the distributions, we get the following total cost for drawing $J_2$ samples:
\begin{equation} \label{eq:tr-cost-2}
O\bigg( J_1 (R_{n-1} R_n)^2 + (R_{n-1} R_n)^3 + J_2 \Big(\sum_{j=1}^N I_j \Big) R_N^2 \sum_{d=1}^{N-1} R_d^2 R_{d+1}^2 \bigg).
\end{equation}
\paragraph{Sampled Least Squares Problem}
Finally, we consider the cost of constructing and solving the sampled least squares problem once the $J_2$ samples in $[\prod_{j \neq n} I_j]$ have been drawn:
\begin{itemize}
\item Once $J_2$ samples in $[\prod_{j \neq n} I_j]$ are drawn, the sketched design matrix $\Sbf \Gbf_{[2]}^{\neq n}$ can be computed efficiently without having to form the full matrix $\Gbf_{[2]}^{\neq n}$.
We provide further details in Remark~\ref{remark:computing-sketched-tr-design-matrix}.
With this approach, the cost of forming $\Sbf \Gbf_{[2]}^{\neq n}$ is
\begin{equation}
O\Big(J_2 R_n \sum_{j \in [N] \setminus \{n, n+1\}} R_{j-1} R_{j}\Big).
\end{equation}
Forming $\Sbf \Xbf_{[n]}^{\top}$ by sampling the appropriate rows costs $O(J_2 I_n)$.
\item The cost of computing the solution $\tilde{\Gbf}^\top = (\Sbf \Gbf_{[2]}^{\neq n})^\dagger \Sbf \Xbf_{[n]}^\top$ using a standard method (e.g., via QR decomposition) is $O(J_2 (R_{n-1} R_n)^2 + J_2 R_{n-1} R_n I_n)$; see Section~5.3.3 in \citet{golub2013MatrixComputations} for details.
\end{itemize}
In total, the cost of constructing and solving the least squares problem is therefore
\begin{equation} \label{eq:tr-cost-3}
O\bigg( J_2 \Big( R_n \sum_{j \in [N] \setminus \{ n, n+1 \}} R_{j-1} R_j + (R_{n-1} R_n)^2 + R_{n-1} R_n I_n \Big) \bigg).
\end{equation}
\paragraph{Total Per-Iteration Cost for TR-ALS-ES}
Recall that for each iteration of TR-ALS, we need to solve $N-1$ least squares problems.
Consequently, adding the costs in \eqref{eq:tr-cost-1}, \eqref{eq:tr-cost-2}, \eqref{eq:tr-cost-3} and multiplying by $N-1$, we get the total per-iteration cost.
If we assume that $R_j = R$ and $I_j = I$ for all $j \in [N]$, the expression simplifies considerably and we get a total per-iteration cost of
\begin{equation}
O(N^2 R^5 J_1 \log J_1 + N^3 I R^6 J_2).
\end{equation}
If the sketch rates $J_1$ and $J_2$ are chosen according to \eqref{eq:tr-J1} and \eqref{eq:tr-J3}, this per-iteration cost becomes
\begin{equation}
O\bigg( \frac{N^3 R^9}{\delta} \log\Big( \frac{N R^4}{\delta} \Big) + N^3 I R^8 \cdot \max\Big( \log\Big( \frac{R^2}{\delta} \Big), \frac{1}{\varepsilon \delta} \Big)\bigg).
\end{equation}
\begin{remark} \label{remark:computing-tr-probabilities}
At first sight, the joint probability computation in \eqref{eq:tr-sampling-results-joint-prob} looks expensive since the number of terms in the sum is exponential in $N$.
However, since not all summation indices $r_j$ and $k_j$ appear in every term, the summation can be done more efficiently.
In fact, the computation \eqref{eq:tr-sampling-results-joint-prob} can be viewed as the evaluation of a tensor ring, which can be done efficiently by contracting core tensors pairwise.
To see this, define core tensors $\Ce^{(j)}$ for $j \in [N]$ as follows:
\begin{itemize}
\item For $j \leq m$ and $j \neq n$, let $\Ce^{(j)} \in \Rb^{R_{j-1}^2 \times I_j \times R_{j}^2}$ be defined elementwise via
\begin{equation}
\Ce^{(j)}(\overline{r_{j-1} k_{j-1}}, i_j ,\overline{r_j k_j}) \defeq \Gbf_{[2]}^{(j)}(i_j, \overline{r_{j} r_{j-1}}) \Gbf_{[2]}^{(j)}(i_j, \overline{k_{j} k_{j-1}}).
\end{equation}
\item For $m < j \leq N$ and $j \neq n$, let $\Ce^{(j)} \in \Rb^{R_{j-1}^2 \times 1 \times R_j^2}$ be defined elementwise via
\begin{equation}
\Ce^{(j)}(\overline{r_{j-1} k_{j-1}}, 1,\overline{r_j k_j}) \defeq \big( \Gbf_{[2]}^{(j)\top} \Gbf_{[2]}^{(j)} \big)(\overline{r_{j} r_{j-1}}, \overline{k_{j} k_{j-1}}).
\end{equation}
\item For $j = n$, let $\Ce^{(j)} = \Ce^{(n)} \in \Rb^{R_{n-1}^2 \times 1 \times R_n^2}$ be defined elementwise via
\begin{equation}
\Ce^{(n)}(\overline{r_{n-1} k_{n-1}}, 1,\overline{r_n k_n}) \defeq \frac{1}{C} \Phibf (\overline{r_{n-1} r_n}, \overline{k_{n-1} k_n}).
\end{equation}
\end{itemize}
We can now rewrite the expression in \eqref{eq:tr-sampling-results-joint-prob} as
\begin{equation} \label{eq:joint-probability-as-contraction}
\Pb((i_j)_{j \leq m, j \neq n}) = \TR(\Ce^{(1)}, \ldots, \Ce^{(N)})_{\xi_1, \ldots, \xi_{N}},
\end{equation}
where
\begin{equation}
\xi_j \defeq
\begin{cases}
i_j & \text{if } j \leq m, j \neq n, \\
1 & \text{otherwise}.
\end{cases}
\end{equation}
As discussed in \citet{zhao2016TensorRing}, the value of an entry in a tensor ring can be computed via a sequence of matrix-matrix products follows by taking the matrix trace:
\begin{equation} \label{eq:trace-matrix-product}
\TR(\Ce^{(1)}, \ldots, \Ce^{(N)})_{\xi_1, \ldots, \xi_{N}} = \trace\big( \Ce^{(1)}(:, \xi_1, :) \cdot \Ce^{(2)}(:, \xi_2, :) \cdots \Ce^{(N)}(:, \xi_N, :) \big),
\end{equation}
where each $\Ce^{(j)}(:, \xi_j, :)$ is treated as a $R_{j-1}^2 \times R_{j}^2$ matrix.
If the matrix product in \eqref{eq:trace-matrix-product} is done left to right, evaluating the right hand side costs
\begin{equation} \label{eq:tensor-evaluation-cost}
O\Big( R_N^2 \sum_{j=1}^{N-1} R_j^2 R_{j+1}^2 \Big).
\end{equation}
\end{remark}
\begin{remark} \label{remark:computing-sketched-tr-design-matrix}
As described by \citet{malik2021SamplingBasedMethod}, it is possible to construct the sketched design matrix $\Sbf \Gbf_{[2]}^{\neq n}$ efficiently without first forming the full matrix $\Gbf_{[2]}^{\neq n}$.
To see how, note that each row $\Gbf_{[2]}^{\neq n}(i, :)$ is the vectorization of the tensor slice $\Ge^{\neq n}(:, i, :)$ due to Definition~\ref{def:unfolding}.
From Definition~\ref{def:subchain}, the tensor slice $\Ge^{\neq n}(:, i, :)$ is given by
\begin{equation}
\Ge^{\neq n}(:,\overline{i_{n+1} \cdots i_N i_1 \cdots i_{n-1}},:) = \Ge^{(n+1)}(:, i_{n+1}, :) \cdots \Ge^{(N)}(:, i_N, :) \cdot \Ge^{(1)}(:, i_1, :) \cdots \Ge^{(n-1)}(:, i_{n-1}, :).
\end{equation}
Suppose $\vbf \in [\prod_{j \neq n} I_j]^{J_2}$ contains the $J_2$ sampled indices corresponding to the sketch $\Sbf$.
Let $\tilde{\Ge}^{\neq n} \in \Rb^{R_n \times J_2 \times R_{n-1}}$ be a tensor which we define as follows:
For each $j \in [J_2]$, let $i = \overline{i_{n+1} \cdots i_N i_1 \cdots i_{n-1}} \defeq \vbf(j)$ and define
\begin{equation} \label{eq:computing-sketched-slice}
\tilde{\Ge}^{\neq n}(:,j,:) \defeq \frac{1}{\sqrt{J_2 \qbf(i)}} \Ge^{(n+1)}(:, i_{n+1}, :) \cdots \Ge^{(N)}(:, i_N, :) \cdot \Ge^{(1)}(:, i_1, :) \cdots \Ge^{(n-1)}(:, i_{n-1}, :).
\end{equation}
We now have $\Sbf \Gbf_{[2]}^{\neq n} = \tilde{\Gbf}_{[2]}^{\neq n}$.
If the matrix product in \eqref{eq:computing-sketched-slice} is computed from left to right, it costs $O(R_n \sum_{j \in [N] \setminus \{n, n+1\}} R_{j-1} R_{j})$.
Since this needs to be computed for each $j \in [J_2]$, the total cost for computing $\Sbf \Gbf_{[2]}^{\neq n}$ via this scheme is
\begin{equation}
O\Big(J_2 R_n \sum_{j \in [N] \setminus \{n, n+1\}} R_{j-1} R_{j}\Big).
\end{equation}
We refer the reader to \citet{malik2021SamplingBasedMethod} for further details.
\end{remark}
\subsection{Complexity Analysis of Competing Methods} \label{sec:detailed-complexity-other}
In this section we provide a few notes on how we computed the computational complexity of the other methods we compare with in Tables~\ref{tab:cp-complexity-comparison} and \ref{tab:tr-complexity-comparison}.
\subsubsection{CP-ALS}
The standard way to implement CP-ALS is given in Figure~3.3 in \citet{kolda2009TensorDecompositions}.
The leading order cost per least squares solve for that algorithm is
\begin{equation}
O(N I R^2 + R^3 + N I^{N-1} R + I^N R).
\end{equation}
Since $N$ such least squares problems need to be solved each iteration, the per-iteration cost is
\begin{equation}
O(N^2 I R^2 + N R^3 + N^2 I^{N-1} R + N I^N R).
\end{equation}
When $N$ is large, this becomes $O(N (N+I) I^{N-1} R)$ which is what we report in Table~\ref{tab:cp-complexity-comparison}.
\subsubsection{SPALS}
\citet{cheng2016SPALSFast} only give the sampling complexity for the case when $N=3$ in their paper.
For arbitrary $N$, and without any assumptions on the rank of the factor matrices or the Khatri--Rao product design matrix, their scheme requires $J \gtrsim R^N \log(I_n / \delta) / \varepsilon^2$ samples when solving for the $n$th factor matrix in order to achieve the additive error guarantees in Theorem~4.1 of their paper.%
\footnote{
If the Khatri--Rao product design matrix is full rank, which happens if all factor matrices are full rank, then $J \gtrsim R^{N-1} \log(I_n / \delta) / \varepsilon^2$ samples will suffice.
}
SPALS requires a one-time upfront cost of $\nnz(\Xe)$ in order to compute the second term in Equation~(5) in \citet{cheng2016SPALSFast}.
In SPALS, the $n$th factor is updated via
\begin{equation}
\Abf^{(n)} = \Xbf_{(n)} \Sbf^\top \Sbf \Big( \kr_{\substack{j=N \\ j \neq n}}^{1} \Abf^{(j)} \Big) \Big( \startimes_{\substack{j=1 \\ j \neq n}}^N \Abf^{(j) \top} \Abf^{(j)} \Big)^{-1},
\end{equation}
where $\Sbf$ is a sampling matrix and $\circledast$ denotes elementwise (Hadamard) product.
When this is computed in the appropriate order, and if $\log$ factors are ignored and we assume that $I_n = I$ for all $n \in [N]$, then the cost of computing $\Abf^{(n)}$ is
\begin{equation}
\tilde{O}( N I R^2 + (N + I) R^{N+1} / \varepsilon^2 ).
\end{equation}
Notice that the cost of computing the sampling distribution is dominated by the cost above.
Since $N$ factor matrices need to be updated per iteration, the total per-iteration cost is
\begin{equation}
\tilde{O} (N^2 I R^2 + N (N+I) R^{N+1} / \varepsilon^2).
\end{equation}
When $N$ is large, this becomes $\tilde{O} ( N (N+I) R^{N+1} / \varepsilon^2 )$, which is what we report in Table~\ref{tab:cp-complexity-comparison}.
\subsubsection{CP-ARLS-LEV}
From Theorem~8 in \citet{larsen2020PracticalLeverageBased}, the sampling complexity for CP-ARLS-LEV required to achieve relative error guarantees when solving for the $n$th factor matrix is $J \gtrsim R^{N-1} \max(\log(R/\delta), 1/(\delta \varepsilon))$.
Solving the sampled least squares problem, which has a design matrix of size $J \times R$ and $I_n$ right hand sides via e.g.\ QR decomposition (see Section~5.3.3 in \citet{golub2013MatrixComputations}) will therefore cost $O((R + I_n) R^N \max(\log(R/\delta), 1/(\delta \varepsilon)))$.
Each iteration requires solving $N$ such least squares problems.
If we assume that $I_n = I$ for all $n \in [N]$ and ignore $\log$ factors, the per-iteration cost becomes
\begin{equation}
\tilde{O} \big( N ( R + I ) R^N / (\delta \varepsilon) \big),
\end{equation}
which is what we report in Table~\ref{tab:cp-complexity-comparison}.
Consider the least squares problem in \eqref{eq:cp-optimization-An-matrix} with the design matrix $\Abf^{\neq n}$ defined in as in \eqref{eq:cp-design-matrix}.
When the leverage score sampling distribution is estimated as in CP-ARLS-LEV, the exponential dependence on $N$ in the sampling complexity cannot be improved.
The following example provides a concrete example when the exponential dependence is required.
\begin{example}
Without loss of generality, consider the case $n = N$ in which case the least squares design matrix in \eqref{eq:cp-design-matrix} is
\begin{equation}
\Abf^{\neq N} = \Abf^{(N-1)} \odot \cdots \odot \Abf^{(1)}.
\end{equation}
Suppose all $\Abf^{(j)}$ for $j \in [J-1]$ are of size $R \times R$ and defined as
\begin{equation}
\Abf^{(j)} \defeq
\begin{bmatrix}
1 & \zerobf \\
\zerobf & \Omegabf^{(j)}
\end{bmatrix}
\end{equation}
where each $\Omegabf^{(j)} \in \Rb^{(R-1) \times (R-1)}$ has i.i.d.\ standard Gaussian entries.
We assume the matrices $\Omegabf^{(j)}$ are all full-rank, which is true almost surely.
The first column and row of $\Abf^{\neq N}$ are $\ebf_1$ and $\ebf_1^\top$, respectively.
For $r \geq 2$ we have
\begin{equation}
\Abf^{\neq N}(:,r) =
\begin{bmatrix}
0 \\
\Omegabf^{(N-1)}(:,r)
\end{bmatrix}
\otimes \cdots \otimes
\begin{bmatrix}
0 \\
\Omegabf^{(1)}(:,r)
\end{bmatrix}.
\end{equation}
Since all the matrices $\Omegabf^{(j)}$ are full-rank it follows that their Kronecker product is full-rank (this follows from Theorem~4.2.15 in \citet{horn1994TopicsMatrix}).
Since the columns $\Abf^{\neq N}(:,r)$ for $r \geq 2$ are equal to columns of $\Omegabf^{(N-1)} \otimes \cdots \otimes \Omegabf^{(1)}$ with some added zeros, it follows that they are linearly independent, and therefore the submatrix $\Gammabf \defeq (\Abf^{\neq N}(i,j))_{i \geq 2, j \geq 2}$ is full-rank.
We may write
\begin{equation}
\Abf^{\neq N} =
\begin{bmatrix}
1 & \zerobf \\
\zerobf & \Gammabf
\end{bmatrix} \in \Rb^{R^{N-1} \times R}.
\end{equation}
If a sampling matrix $\Sbf$ does not sample the first row of $\Abf^{\neq N}$, then $\Sbf \Abf^{\neq N}$ will be rank-deficient and relative error guarantees therefore unachievable.
Since all $\Abf^{(j)}$ are square and full-rank, the sampling procedure used in CP-ARLS-LEV will sample rows of $\Abf^{\neq N}$ uniformly.
In order to sample the first row of $\Abf^{\neq N}$ with probability at least 0.5 with uniform sampling, we clearly need to sample at least half of all rows of $\Abf^{\neq N}$, i.e., we need $J \geq R^{N-1}/2$.
\end{example}
\subsubsection{Methods for Tensor Ring Decomposition}
The complexities we report in Table~\ref{tab:tr-complexity-comparison} for other methods where taken directly from Table~1 in \citet{malik2021SamplingBasedMethod}.
\subsection{Details on Algorithm Implementations}
Our implementation of CP-ARLS-LEV is based on Algorithm~3 in \citet{larsen2020PracticalLeverageBased}.
We do not use any hybrid-deterministic sampling, but we do combine repeated rows.
Some key functionality required for our CP-ALS-ES is written in C and incorporated into Matlab via the MEX interface.
Our own TR-ALS-ES is implemented by appropriately modifying the Matlab code for TR-ALS-Sampled by \citet{malik2021SamplingBasedMethod}.
\subsection{Datasets}
The photo used for the sampling distribution comparison was taken by Sebastian M\"{u}ller on Unsplash and is available at \url{https://unsplash.com/photos/l54ZALpH2_I}.
We converted this figure to gray scale by averaging the three color channels.
We also cropped the image slightly to make the width and height a power of 2.
The tensorization is done following the ideas for visual data tensorization discussed in \citet{yuan2019HighorderTensor}.
Please see our code for precise details.
The COIL-100 dataset was created by \citet{nene1996ColumbiaObject} and is available for download at \url{https://www.cs.columbia.edu/CAVE/software/softlib/coil-100.php}.
\subsection{Sampling Distribution Plots and Computational Time}
We have included figures below that compare the sampling distributions used by our methods with those used by the previous state-of-the-art methods in the least squares problem considered in the first experiment in Section~\ref{sec:experiments}.
For a rank-10 CP decomposition of the tabby cat tensor, Figure~\ref{fig:CP-rank-10-probability-comparison} shows the exact leverage score distribution ($\pbf$ in Definition~\ref{def:leverage-score-sampling}), the sampling distribution used by CP-ARLS-LEV, and a realization (for $J_1 = 1000$) of the distribution our CP-ALS-ES uses.
Figure~\ref{fig:CP-rank-20-probability-comparison} shows the same things as Figure~\ref{fig:CP-rank-10-probability-comparison}, but for a rank-20 CP decomposition.
For a rank-$(3,\ldots,3)$ tensor ring decomposition of the tabby cat tensor, Figure~\ref{fig:TR-rank-3-probability-comparison} shows the exact leverage score distribution, the sampling distribution used by TR-ALS-Sampled, and a realization (for $J_1=1000$) of the distribution our TR-ALS-ES uses.
Figure~\ref{fig:TR-rank-5-probability-comparison} shows the same things as Figure~\ref{fig:TR-rank-3-probability-comparison}, but for a rank-$(5,\ldots,5)$ tensor ring decomposition.
Notice that the sampling distribution that our methods use follow the exact leverage score sampling distribution closely.
The distributions used by CP-ARLS-LEV and TR-ALS-Sampled are less accurate.
In particular, when $R > I = 16$ for the CP decomposition (Figure~\ref{fig:CP-rank-20-probability-comparison}) or when $R_{n-1}R_n > I = 16$ for the tensor ring decomposition (Figure~\ref{fig:TR-rank-5-probability-comparison}), CP-ARLS-LEV and TR-ALS-Sampled sample from \emph{a uniform distribution}.
This is not an anomaly, but rather a direct consequence from how those methods estimate the leverage scores.
Our proposed methods, by contrast, handle those cases well.
\begin{figure}[h]
\centering
\includegraphics[width=1\columnwidth]{CP-rank-10-probability-comparison.pdf}
\caption{
Comparison of the exact leverage score distribution, the sampling distribution used by CP-ARLS-LEV, and a realization (for $J_1=1000$) of the distribution used by our CP-ALS-ES.
The least squares problem corresponds to solving for the 6th factor matrix in a rank-10 CP decomposition of the 6-way tabby cat tensor.
}
\label{fig:CP-rank-10-probability-comparison}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=1\columnwidth]{CP-rank-20-probability-comparison.pdf}
\caption{
Same as Figure~\ref{fig:CP-rank-10-probability-comparison}, but for a rank-20 decomposition.
}
\label{fig:CP-rank-20-probability-comparison}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=1\columnwidth]{TR-rank-3-probability-comparison.pdf}
\caption{
Comparison of the exact leverage score distribution, the sampling distribution used by TR-ALS-Sampled, and a realization (for $J_1 = 1000$) of the distribution used by our TR-ALS-ES.
The least squares problem corresponds to solving for the 6th core tensor in a rank-$(3,\ldots,3)$ tensor ring decomposition of the 6-way tabby cat tensor.
}
\label{fig:TR-rank-3-probability-comparison}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=1\columnwidth]{TR-rank-5-probability-comparison.pdf}
\caption{
Same as Figure~\ref{fig:TR-rank-3-probability-comparison}, but for a rank-$(5,\ldots,5)$ decomposition.
}
\label{fig:TR-rank-5-probability-comparison}
\end{figure}
Tables~\ref{tab:KL-div-CP-time} and \ref{tab:KL-div-TR-time} report the time it took to compute the distributions used in Tables~\ref{tab:KL-div-CP} and \ref{tab:KL-div-TR}, respectively.
Note that the different methods do not compute the full distributions the way we do in Tables~\ref{tab:KL-div-CP}--\ref{tab:KL-div-TR} and Figures~\ref{fig:CP-rank-10-probability-comparison}--\ref{fig:TR-rank-5-probability-comparison}, so these numbers are not representative of actual decomposition time and are only added here for completeness.
\begin{table}[ht!]
\centering
\caption{
Time in seconds it took to compute the distributions used in Table~\ref{tab:KL-div-CP}.
\label{tab:KL-div-CP-time}
}
\begin{tabular}{lll}
\toprule
Method & $R=10$ & $R=20$ \\
\midrule
CP-ARLS-LEV & 0.01 & 0.01 \\
CP-ALS-ES ($J_1 = \text{1e+4}$) & 0.06 & 0.12 \\
CP-ALS-ES ($J_1 = \text{1e+3}$) & 0.04 & 0.07 \\
CP-ALS-ES ($J_1 = \text{1e+2}$) & 0.03 & 0.07 \\
\bottomrule
\end{tabular}
\end{table}
\begin{table}[ht!]
\centering
\caption{
Time in seconds it took to compute the distributions used in Table~\ref{tab:KL-div-TR}.
\label{tab:KL-div-TR-time}
}
\begin{tabular}{lll}
\toprule
Method & $R=3$ & $R=5$ \\
\midrule
TR-ALS-Sampled & 0.01 & 0.01 \\
TR-ALS-ES ($J_1 = \text{1e+4}$) & 0.07 & 0.21 \\
TR-ALS-ES ($J_1 = \text{1e+3}$) & 0.03 & 0.10 \\
TR-ALS-ES ($J_1 = \text{1e+2}$) & 0.03 & 0.10 \\
\bottomrule
\end{tabular}
\end{table}
\subsection{Feature Extraction Experiments} \label{sec:supp-feature-extraction-experiments}
We provide some further details on the feature extraction experiments in Section~\ref{sec:feature-extraction} in this section.
For a rank-25 CP decomposition, the 4th factor matrix is of size $7200 \times 25$.
We directly use this factor matrix as the feature matrix we feed to the $k$-NN method in Matlab.
For the rank-$(5, \ldots, 5)$ tensor ring decomposition, the 4th core tensor is of size $5 \times 7200 \times 5$.
We turn this into a $7200 \times 25$ matrix via a classical mode-2 unfolding which we then use as the feature matrix in the $k$-NN algorithm.
In our feature extraction experiment we assume that both the labeled and unlabeled images are available when the tensor decomposition is computed.
This is a limitation since we might want to classify new unlabeled images that arrive after the decomposition has been computed without having to recompute the decomposition.
We now propose a potential approach to circumventing this limitation.
Adding a new image corresponds to adding new rows to $\Xe_{(4)}$ and $\Xe_{[4]}$.
The factor matrix $\Abf^{(4)}$ for the CP decomposition of this augmented tensor will have an additional row, while the number of rows will remain the same in the other factor matrices.
Similarly, the core tensor $\Ge^{(4)}$ for the tensor ring decomposition will have an additional lateral slice, while the number of lateral slices will remain the same for the other cores.
For the CP decomposition, a feature vector for the new image can therefore be computed via
\begin{equation} \label{eq:CP-feat-ext-new-samp}
\abf^* = \argmin_{\abf} \| \Abf^{\neq 4} \abf^\top - \xbf^\top \|_2,
\end{equation}
where $\abf^* \in \Rb^{1 \times R}$ is the new row in $\Abf^{(4)}$ and $\xbf \in \Rb^{1 \times 49152}$ is the new row in $\Xe_{(4)}$.
For the tensor train decomposition, a feature vector for the new image can similarly be computed via
\begin{equation} \label{eq:TR-feat-ext-new-samp}
\gbf^* = \argmin_{\gbf} \| \Gbf_{[2]}^{\neq 4} \gbf^\top - \hat{\xbf}^\top \|_2,
\end{equation}
where $\gbf^* \in \Rb^{1 \times R_3 R_4}$ is a reshaped version of the new lateral slice in $\Ge^{(4)}$ and $\hat{\xbf} \in \Rb^{1 \times 49152}$ is the new row in $\Xe_{[4]}$.
The sampling techniques in Section~\ref{sec:efficient-sampling-for-TD} can be used to compute approximate solutions to \eqref{eq:CP-feat-ext-new-samp} and \eqref{eq:TR-feat-ext-new-samp} efficiently.
The feature vectors $\abf^*$ and $\gbf^*$ can now be used to classify the new image.
\subsection{Demonstration of Improved Complexity for the Tensor Ring Decomposition} \label{sec:demonstration-improvement-TR}
We construct a synthetic 10-way tensor that demonstrates the improved sampling and computational complexity of our proposed TR-ALS-ES over TR-ALS-Sampled.
It is constructed via \eqref{eq:tr-decomposition} from core tensors $\Ge^{(n)} \in \Rb^{3 \times 6 \times 3}$ for $n \in [10]$ with $\Ge^{(n)}(1,1,1) = 3$ and all other entries zero.
Additionally, i.i.d.\ Gaussian noise with standard deviation $0.01$ is added to all entries of the tensor.
Both methods are run for 20 iterations with target ranks $(3, 3, \ldots, 3)$ and are initialized using a variant of the randomized range finding approach proposed by \citet{larsen2020PracticalLeverageBased}, Appendix~F, adapted to the tensor ring decomposition.
TR-ALS-Sampled fails even when as many as \emph{half} (i.e., $J = 6^9/2 \approx 5.0\text{e+6}$) of all rows are sampled, taking 966 seconds.
By contrast, our TR-ALS-ES only requires a recursive sketch size of $J_1 = 1\text{e+4}$ and $J_2 = 1\text{e+3}$ samples to get an accurate solution, taking 41 seconds.
Our method improves the sampling complexity and compute time by 3 and 1 orders of magnitude, respectively.
\subsection{Preliminary Results From Experiments on the Tensor Train Decomposition} \label{sec:tt-experiments}
In this section we provide some preliminary results from experiments on the tensor train (TT) decomposition.
We do these experiments by running the different tensor ring decomposition algorithms with $R_0 = R_N = 1$ which makes the resulting decomposition a TT.
We refer to the methods by the same names as the tensor ring decomposition methods, but with ``TR'' replaced by ``TT.''
\begin{table}[ht!]
\centering
\caption{
KL-divergence (lower is better) of the approximated sampling distribution from the exact one for a TT-ALS least squares problem with target TT-ranks $R_n = R$ for $1 \leq n \leq 5$.
The TT-ALS least squares problem is identical to the TR-ALS problem in \eqref{eq:tr-optimization-Gn-matrix} but with the restriction $R_0=R_N=1$.
\label{tab:KL-div-TT}
}
\begin{tabular}{lll}
\toprule
Method & $R=3$ & $R=5$ \\
\midrule
TT-ALS-Sampled & 0.4843 & 0.2469 \\
TT-ALS-ES ($J_1 = \text{1e+4}$) & 0.0004 & 0.0007 \\
TT-ALS-ES ($J_1 = \text{1e+3}$) & 0.0087 & 0.0065 \\
TT-ALS-ES ($J_1 = \text{1e+2}$) & 0.0425 & 0.0845 \\
\bottomrule
\end{tabular}
\end{table}
\begin{table}[ht!]
\centering
\caption{
Run time, decomposition error and classification accuracy when using the TT decomposition for feature extraction.
\label{tab:classification-results-TT}
}
\begin{tabular}{lrrr}
\toprule
Method & Time (s) & Error & Accuracy (\%) \\
\midrule
TT-ALS & 9440.1 & 0.44 & 95.2 \\
TT-ALS-Sampled & 12.3 & 0.44 & 94.1 \\
\textbf{TT-ALS-ES (our)} & 31.4 & 0.44 & 94.2 \\
\bottomrule
\end{tabular}
\end{table}
First, we repeat the experiments in Section~\ref{sec:sampling-ditribution-comparison} for the TT decomposition.
All settings are the same as for the tensor ring decomposition except that $R_0 = R_6 = 1$.
The results are shown in Table~\ref{tab:KL-div-TT}.
The discrepancy between the approximate leverage score sampling distribution that TT-ALS-Sampled samples from and the exact one is greater than it is for the tensor ring decomposition (compare with Table~\ref{tab:KL-div-TR}).
The discrepancy of TT-ALS-ES is similar to that of TR-ALS-ES for $J_1 = 1\text{e+4}$ and $J_1 = 1\text{e+3}$ and smaller for $J_1 = 1\text{e+2}$.
Since we are solving for the 6th core tensor out of 6, the theoretical bound on $J_1$ in \eqref{eq:tr-J1} becomes $J_1 \gtrsim N (R_5 R_6)^2/\delta = N R^2/\delta$ since $R_6=1$.
For the tensor ring decomposition with all $R_n=R$, the same bound is $J_1 \gtrsim N R^4 / \delta$.
A smaller discrepancy is therefore expected for TT-ALS-ES than for TR-ALS-ES.
Next, we repeat the feature extraction experiment in Section~\ref{sec:feature-extraction} for the TT decomposition.
In addition to the restriction $R_0=R_1=1$ we also increase the number of samples from 1000 to 3000 for both TT-ALS-Sampled and TT-ALS-ES since using fewer than 3000 samples yields poor results for both methods.
We also add a small Tikhonov regularization term (with regularization constant $10^{-2}$) in all least squares solves for both randomized methods in order to avoid numerical issues that otherwise appear for both.
The TT-ranks are $R_n = 5$ for $1 \leq n \leq 3$.
The results are reported in Table~\ref{tab:classification-results-TT}.
The run time for TT-ALS is slightly faster than that of TR-ALS which is expected since the TT has fewer parameters than the tensor ring (compare with Table~\ref{tab:classification-results}).
The two randomized algorithms are a bit slower than they are for the tensor ring decomposition due to the larger number of samples being drawn.
The decomposition error is higher and the classification accuracy lower than they are for the tensor ring decomposition.
This is also expected since the TT decomposition has fewer parameters.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.